1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements some loop unrolling utilities. It does not define any
10 // actual pass or policy, but provides a single function to perform loop
11 // unrolling.
12 //
13 // The process of unrolling can produce extraneous basic blocks linked with
14 // unconditional branches.  This will be corrected in the future.
15 //
16 //===----------------------------------------------------------------------===//
17 
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/ADT/Twine.h"
27 #include "llvm/ADT/ilist_iterator.h"
28 #include "llvm/ADT/iterator_range.h"
29 #include "llvm/Analysis/AssumptionCache.h"
30 #include "llvm/Analysis/DomTreeUpdater.h"
31 #include "llvm/Analysis/InstructionSimplify.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/LoopIterator.h"
34 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
35 #include "llvm/Analysis/ScalarEvolution.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CFG.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DebugInfoMetadata.h"
40 #include "llvm/IR/DebugLoc.h"
41 #include "llvm/IR/DiagnosticInfo.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Function.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Metadata.h"
48 #include "llvm/IR/Module.h"
49 #include "llvm/IR/Use.h"
50 #include "llvm/IR/User.h"
51 #include "llvm/IR/ValueHandle.h"
52 #include "llvm/IR/ValueMap.h"
53 #include "llvm/Support/Casting.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/GenericDomTree.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
60 #include "llvm/Transforms/Utils/Cloning.h"
61 #include "llvm/Transforms/Utils/Local.h"
62 #include "llvm/Transforms/Utils/LoopSimplify.h"
63 #include "llvm/Transforms/Utils/LoopUtils.h"
64 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
65 #include "llvm/Transforms/Utils/UnrollLoop.h"
66 #include "llvm/Transforms/Utils/ValueMapper.h"
67 #include <algorithm>
68 #include <assert.h>
69 #include <type_traits>
70 #include <vector>
71 
72 namespace llvm {
73 class DataLayout;
74 class Value;
75 } // namespace llvm
76 
77 using namespace llvm;
78 
79 #define DEBUG_TYPE "loop-unroll"
80 
81 // TODO: Should these be here or in LoopUnroll?
82 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
83 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
84 STATISTIC(NumUnrolledNotLatch, "Number of loops unrolled without a conditional "
85                                "latch (completely or otherwise)");
86 
87 static cl::opt<bool>
88 UnrollRuntimeEpilog("unroll-runtime-epilog", cl::init(false), cl::Hidden,
89                     cl::desc("Allow runtime unrolled loops to be unrolled "
90                              "with epilog instead of prolog."));
91 
92 static cl::opt<bool>
93 UnrollVerifyDomtree("unroll-verify-domtree", cl::Hidden,
94                     cl::desc("Verify domtree after unrolling"),
95 #ifdef EXPENSIVE_CHECKS
96     cl::init(true)
97 #else
98     cl::init(false)
99 #endif
100                     );
101 
102 static cl::opt<bool>
103 UnrollVerifyLoopInfo("unroll-verify-loopinfo", cl::Hidden,
104                     cl::desc("Verify loopinfo after unrolling"),
105 #ifdef EXPENSIVE_CHECKS
106     cl::init(true)
107 #else
108     cl::init(false)
109 #endif
110                     );
111 
112 
113 /// Check if unrolling created a situation where we need to insert phi nodes to
114 /// preserve LCSSA form.
115 /// \param Blocks is a vector of basic blocks representing unrolled loop.
116 /// \param L is the outer loop.
117 /// It's possible that some of the blocks are in L, and some are not. In this
118 /// case, if there is a use is outside L, and definition is inside L, we need to
119 /// insert a phi-node, otherwise LCSSA will be broken.
120 /// The function is just a helper function for llvm::UnrollLoop that returns
121 /// true if this situation occurs, indicating that LCSSA needs to be fixed.
122 static bool needToInsertPhisForLCSSA(Loop *L,
123                                      const std::vector<BasicBlock *> &Blocks,
124                                      LoopInfo *LI) {
125   for (BasicBlock *BB : Blocks) {
126     if (LI->getLoopFor(BB) == L)
127       continue;
128     for (Instruction &I : *BB) {
129       for (Use &U : I.operands()) {
130         if (const auto *Def = dyn_cast<Instruction>(U)) {
131           Loop *DefLoop = LI->getLoopFor(Def->getParent());
132           if (!DefLoop)
133             continue;
134           if (DefLoop->contains(L))
135             return true;
136         }
137       }
138     }
139   }
140   return false;
141 }
142 
143 /// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary
144 /// and adds a mapping from the original loop to the new loop to NewLoops.
145 /// Returns nullptr if no new loop was created and a pointer to the
146 /// original loop OriginalBB was part of otherwise.
147 const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB,
148                                            BasicBlock *ClonedBB, LoopInfo *LI,
149                                            NewLoopsMap &NewLoops) {
150   // Figure out which loop New is in.
151   const Loop *OldLoop = LI->getLoopFor(OriginalBB);
152   assert(OldLoop && "Should (at least) be in the loop being unrolled!");
153 
154   Loop *&NewLoop = NewLoops[OldLoop];
155   if (!NewLoop) {
156     // Found a new sub-loop.
157     assert(OriginalBB == OldLoop->getHeader() &&
158            "Header should be first in RPO");
159 
160     NewLoop = LI->AllocateLoop();
161     Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop());
162 
163     if (NewLoopParent)
164       NewLoopParent->addChildLoop(NewLoop);
165     else
166       LI->addTopLevelLoop(NewLoop);
167 
168     NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
169     return OldLoop;
170   } else {
171     NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
172     return nullptr;
173   }
174 }
175 
176 /// The function chooses which type of unroll (epilog or prolog) is more
177 /// profitabale.
178 /// Epilog unroll is more profitable when there is PHI that starts from
179 /// constant.  In this case epilog will leave PHI start from constant,
180 /// but prolog will convert it to non-constant.
181 ///
182 /// loop:
183 ///   PN = PHI [I, Latch], [CI, PreHeader]
184 ///   I = foo(PN)
185 ///   ...
186 ///
187 /// Epilog unroll case.
188 /// loop:
189 ///   PN = PHI [I2, Latch], [CI, PreHeader]
190 ///   I1 = foo(PN)
191 ///   I2 = foo(I1)
192 ///   ...
193 /// Prolog unroll case.
194 ///   NewPN = PHI [PrologI, Prolog], [CI, PreHeader]
195 /// loop:
196 ///   PN = PHI [I2, Latch], [NewPN, PreHeader]
197 ///   I1 = foo(PN)
198 ///   I2 = foo(I1)
199 ///   ...
200 ///
201 static bool isEpilogProfitable(Loop *L) {
202   BasicBlock *PreHeader = L->getLoopPreheader();
203   BasicBlock *Header = L->getHeader();
204   assert(PreHeader && Header);
205   for (const PHINode &PN : Header->phis()) {
206     if (isa<ConstantInt>(PN.getIncomingValueForBlock(PreHeader)))
207       return true;
208   }
209   return false;
210 }
211 
212 /// Perform some cleanup and simplifications on loops after unrolling. It is
213 /// useful to simplify the IV's in the new loop, as well as do a quick
214 /// simplify/dce pass of the instructions.
215 void llvm::simplifyLoopAfterUnroll(Loop *L, bool SimplifyIVs, LoopInfo *LI,
216                                    ScalarEvolution *SE, DominatorTree *DT,
217                                    AssumptionCache *AC,
218                                    const TargetTransformInfo *TTI) {
219   // Simplify any new induction variables in the partially unrolled loop.
220   if (SE && SimplifyIVs) {
221     SmallVector<WeakTrackingVH, 16> DeadInsts;
222     simplifyLoopIVs(L, SE, DT, LI, TTI, DeadInsts);
223 
224     // Aggressively clean up dead instructions that simplifyLoopIVs already
225     // identified. Any remaining should be cleaned up below.
226     while (!DeadInsts.empty()) {
227       Value *V = DeadInsts.pop_back_val();
228       if (Instruction *Inst = dyn_cast_or_null<Instruction>(V))
229         RecursivelyDeleteTriviallyDeadInstructions(Inst);
230     }
231   }
232 
233   // At this point, the code is well formed.  Perform constprop, instsimplify,
234   // and dce.
235   const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
236   SmallVector<WeakTrackingVH, 16> DeadInsts;
237   for (BasicBlock *BB : L->getBlocks()) {
238     for (Instruction &Inst : llvm::make_early_inc_range(*BB)) {
239       if (Value *V = simplifyInstruction(&Inst, {DL, nullptr, DT, AC}))
240         if (LI->replacementPreservesLCSSAForm(&Inst, V))
241           Inst.replaceAllUsesWith(V);
242       if (isInstructionTriviallyDead(&Inst))
243         DeadInsts.emplace_back(&Inst);
244     }
245     // We can't do recursive deletion until we're done iterating, as we might
246     // have a phi which (potentially indirectly) uses instructions later in
247     // the block we're iterating through.
248     RecursivelyDeleteTriviallyDeadInstructions(DeadInsts);
249   }
250 }
251 
252 /// Unroll the given loop by Count. The loop must be in LCSSA form.  Unrolling
253 /// can only fail when the loop's latch block is not terminated by a conditional
254 /// branch instruction. However, if the trip count (and multiple) are not known,
255 /// loop unrolling will mostly produce more code that is no faster.
256 ///
257 /// If Runtime is true then UnrollLoop will try to insert a prologue or
258 /// epilogue that ensures the latch has a trip multiple of Count. UnrollLoop
259 /// will not runtime-unroll the loop if computing the run-time trip count will
260 /// be expensive and AllowExpensiveTripCount is false.
261 ///
262 /// The LoopInfo Analysis that is passed will be kept consistent.
263 ///
264 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
265 /// DominatorTree if they are non-null.
266 ///
267 /// If RemainderLoop is non-null, it will receive the remainder loop (if
268 /// required and not fully unrolled).
269 LoopUnrollResult llvm::UnrollLoop(Loop *L, UnrollLoopOptions ULO, LoopInfo *LI,
270                                   ScalarEvolution *SE, DominatorTree *DT,
271                                   AssumptionCache *AC,
272                                   const TargetTransformInfo *TTI,
273                                   OptimizationRemarkEmitter *ORE,
274                                   bool PreserveLCSSA, Loop **RemainderLoop) {
275   assert(DT && "DomTree is required");
276 
277   if (!L->getLoopPreheader()) {
278     LLVM_DEBUG(dbgs() << "  Can't unroll; loop preheader-insertion failed.\n");
279     return LoopUnrollResult::Unmodified;
280   }
281 
282   if (!L->getLoopLatch()) {
283     LLVM_DEBUG(dbgs() << "  Can't unroll; loop exit-block-insertion failed.\n");
284     return LoopUnrollResult::Unmodified;
285   }
286 
287   // Loops with indirectbr cannot be cloned.
288   if (!L->isSafeToClone()) {
289     LLVM_DEBUG(dbgs() << "  Can't unroll; Loop body cannot be cloned.\n");
290     return LoopUnrollResult::Unmodified;
291   }
292 
293   if (L->getHeader()->hasAddressTaken()) {
294     // The loop-rotate pass can be helpful to avoid this in many cases.
295     LLVM_DEBUG(
296         dbgs() << "  Won't unroll loop: address of header block is taken.\n");
297     return LoopUnrollResult::Unmodified;
298   }
299 
300   assert(ULO.Count > 0);
301 
302   // All these values should be taken only after peeling because they might have
303   // changed.
304   BasicBlock *Preheader = L->getLoopPreheader();
305   BasicBlock *Header = L->getHeader();
306   BasicBlock *LatchBlock = L->getLoopLatch();
307   SmallVector<BasicBlock *, 4> ExitBlocks;
308   L->getExitBlocks(ExitBlocks);
309   std::vector<BasicBlock *> OriginalLoopBlocks = L->getBlocks();
310 
311   const unsigned MaxTripCount = SE->getSmallConstantMaxTripCount(L);
312   const bool MaxOrZero = SE->isBackedgeTakenCountMaxOrZero(L);
313 
314   // Effectively "DCE" unrolled iterations that are beyond the max tripcount
315   // and will never be executed.
316   if (MaxTripCount && ULO.Count > MaxTripCount)
317     ULO.Count = MaxTripCount;
318 
319   struct ExitInfo {
320     unsigned TripCount;
321     unsigned TripMultiple;
322     unsigned BreakoutTrip;
323     bool ExitOnTrue;
324     SmallVector<BasicBlock *> ExitingBlocks;
325   };
326   DenseMap<BasicBlock *, ExitInfo> ExitInfos;
327   SmallVector<BasicBlock *, 4> ExitingBlocks;
328   L->getExitingBlocks(ExitingBlocks);
329   for (auto *ExitingBlock : ExitingBlocks) {
330     // The folding code is not prepared to deal with non-branch instructions
331     // right now.
332     auto *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
333     if (!BI)
334       continue;
335 
336     ExitInfo &Info = ExitInfos.try_emplace(ExitingBlock).first->second;
337     Info.TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
338     Info.TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
339     if (Info.TripCount != 0) {
340       Info.BreakoutTrip = Info.TripCount % ULO.Count;
341       Info.TripMultiple = 0;
342     } else {
343       Info.BreakoutTrip = Info.TripMultiple =
344           (unsigned)GreatestCommonDivisor64(ULO.Count, Info.TripMultiple);
345     }
346     Info.ExitOnTrue = !L->contains(BI->getSuccessor(0));
347     Info.ExitingBlocks.push_back(ExitingBlock);
348     LLVM_DEBUG(dbgs() << "  Exiting block %" << ExitingBlock->getName()
349                       << ": TripCount=" << Info.TripCount
350                       << ", TripMultiple=" << Info.TripMultiple
351                       << ", BreakoutTrip=" << Info.BreakoutTrip << "\n");
352   }
353 
354   // Are we eliminating the loop control altogether?  Note that we can know
355   // we're eliminating the backedge without knowing exactly which iteration
356   // of the unrolled body exits.
357   const bool CompletelyUnroll = ULO.Count == MaxTripCount;
358 
359   const bool PreserveOnlyFirst = CompletelyUnroll && MaxOrZero;
360 
361   // There's no point in performing runtime unrolling if this unroll count
362   // results in a full unroll.
363   if (CompletelyUnroll)
364     ULO.Runtime = false;
365 
366   // Go through all exits of L and see if there are any phi-nodes there. We just
367   // conservatively assume that they're inserted to preserve LCSSA form, which
368   // means that complete unrolling might break this form. We need to either fix
369   // it in-place after the transformation, or entirely rebuild LCSSA. TODO: For
370   // now we just recompute LCSSA for the outer loop, but it should be possible
371   // to fix it in-place.
372   bool NeedToFixLCSSA =
373       PreserveLCSSA && CompletelyUnroll &&
374       any_of(ExitBlocks,
375              [](const BasicBlock *BB) { return isa<PHINode>(BB->begin()); });
376 
377   // The current loop unroll pass can unroll loops that have
378   // (1) single latch; and
379   // (2a) latch is unconditional; or
380   // (2b) latch is conditional and is an exiting block
381   // FIXME: The implementation can be extended to work with more complicated
382   // cases, e.g. loops with multiple latches.
383   BranchInst *LatchBI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
384 
385   // A conditional branch which exits the loop, which can be optimized to an
386   // unconditional branch in the unrolled loop in some cases.
387   bool LatchIsExiting = L->isLoopExiting(LatchBlock);
388   if (!LatchBI || (LatchBI->isConditional() && !LatchIsExiting)) {
389     LLVM_DEBUG(
390         dbgs() << "Can't unroll; a conditional latch must exit the loop");
391     return LoopUnrollResult::Unmodified;
392   }
393 
394   // Loops containing convergent instructions cannot use runtime unrolling,
395   // as the prologue/epilogue may add additional control-dependencies to
396   // convergent operations.
397   LLVM_DEBUG(
398       {
399         bool HasConvergent = false;
400         for (auto &BB : L->blocks())
401           for (auto &I : *BB)
402             if (auto *CB = dyn_cast<CallBase>(&I))
403               HasConvergent |= CB->isConvergent();
404         assert((!HasConvergent || !ULO.Runtime) &&
405                "Can't runtime unroll if loop contains a convergent operation.");
406       });
407 
408   bool EpilogProfitability =
409       UnrollRuntimeEpilog.getNumOccurrences() ? UnrollRuntimeEpilog
410                                               : isEpilogProfitable(L);
411 
412   if (ULO.Runtime &&
413       !UnrollRuntimeLoopRemainder(L, ULO.Count, ULO.AllowExpensiveTripCount,
414                                   EpilogProfitability, ULO.UnrollRemainder,
415                                   ULO.ForgetAllSCEV, LI, SE, DT, AC, TTI,
416                                   PreserveLCSSA, RemainderLoop)) {
417     if (ULO.Force)
418       ULO.Runtime = false;
419     else {
420       LLVM_DEBUG(dbgs() << "Won't unroll; remainder loop could not be "
421                            "generated when assuming runtime trip count\n");
422       return LoopUnrollResult::Unmodified;
423     }
424   }
425 
426   using namespace ore;
427   // Report the unrolling decision.
428   if (CompletelyUnroll) {
429     LLVM_DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
430                       << " with trip count " << ULO.Count << "!\n");
431     if (ORE)
432       ORE->emit([&]() {
433         return OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(),
434                                   L->getHeader())
435                << "completely unrolled loop with "
436                << NV("UnrollCount", ULO.Count) << " iterations";
437       });
438   } else {
439     LLVM_DEBUG(dbgs() << "UNROLLING loop %" << Header->getName() << " by "
440                       << ULO.Count);
441     if (ULO.Runtime)
442       LLVM_DEBUG(dbgs() << " with run-time trip count");
443     LLVM_DEBUG(dbgs() << "!\n");
444 
445     if (ORE)
446       ORE->emit([&]() {
447         OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(),
448                                 L->getHeader());
449         Diag << "unrolled loop by a factor of " << NV("UnrollCount", ULO.Count);
450         if (ULO.Runtime)
451           Diag << " with run-time trip count";
452         return Diag;
453       });
454   }
455 
456   // We are going to make changes to this loop. SCEV may be keeping cached info
457   // about it, in particular about backedge taken count. The changes we make
458   // are guaranteed to invalidate this information for our loop. It is tempting
459   // to only invalidate the loop being unrolled, but it is incorrect as long as
460   // all exiting branches from all inner loops have impact on the outer loops,
461   // and if something changes inside them then any of outer loops may also
462   // change. When we forget outermost loop, we also forget all contained loops
463   // and this is what we need here.
464   if (SE) {
465     if (ULO.ForgetAllSCEV)
466       SE->forgetAllLoops();
467     else
468       SE->forgetTopmostLoop(L);
469   }
470 
471   if (!LatchIsExiting)
472     ++NumUnrolledNotLatch;
473 
474   // For the first iteration of the loop, we should use the precloned values for
475   // PHI nodes.  Insert associations now.
476   ValueToValueMapTy LastValueMap;
477   std::vector<PHINode*> OrigPHINode;
478   for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
479     OrigPHINode.push_back(cast<PHINode>(I));
480   }
481 
482   std::vector<BasicBlock *> Headers;
483   std::vector<BasicBlock *> Latches;
484   Headers.push_back(Header);
485   Latches.push_back(LatchBlock);
486 
487   // The current on-the-fly SSA update requires blocks to be processed in
488   // reverse postorder so that LastValueMap contains the correct value at each
489   // exit.
490   LoopBlocksDFS DFS(L);
491   DFS.perform(LI);
492 
493   // Stash the DFS iterators before adding blocks to the loop.
494   LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
495   LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
496 
497   std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks();
498 
499   // Loop Unrolling might create new loops. While we do preserve LoopInfo, we
500   // might break loop-simplified form for these loops (as they, e.g., would
501   // share the same exit blocks). We'll keep track of loops for which we can
502   // break this so that later we can re-simplify them.
503   SmallSetVector<Loop *, 4> LoopsToSimplify;
504   for (Loop *SubLoop : *L)
505     LoopsToSimplify.insert(SubLoop);
506 
507   // When a FSDiscriminator is enabled, we don't need to add the multiply
508   // factors to the discriminators.
509   if (Header->getParent()->isDebugInfoForProfiling() && !EnableFSDiscriminator)
510     for (BasicBlock *BB : L->getBlocks())
511       for (Instruction &I : *BB)
512         if (!isa<DbgInfoIntrinsic>(&I))
513           if (const DILocation *DIL = I.getDebugLoc()) {
514             auto NewDIL = DIL->cloneByMultiplyingDuplicationFactor(ULO.Count);
515             if (NewDIL)
516               I.setDebugLoc(*NewDIL);
517             else
518               LLVM_DEBUG(dbgs()
519                          << "Failed to create new discriminator: "
520                          << DIL->getFilename() << " Line: " << DIL->getLine());
521           }
522 
523   // Identify what noalias metadata is inside the loop: if it is inside the
524   // loop, the associated metadata must be cloned for each iteration.
525   SmallVector<MDNode *, 6> LoopLocalNoAliasDeclScopes;
526   identifyNoAliasScopesToClone(L->getBlocks(), LoopLocalNoAliasDeclScopes);
527 
528   // We place the unrolled iterations immediately after the original loop
529   // latch.  This is a reasonable default placement if we don't have block
530   // frequencies, and if we do, well the layout will be adjusted later.
531   auto BlockInsertPt = std::next(LatchBlock->getIterator());
532   for (unsigned It = 1; It != ULO.Count; ++It) {
533     SmallVector<BasicBlock *, 8> NewBlocks;
534     SmallDenseMap<const Loop *, Loop *, 4> NewLoops;
535     NewLoops[L] = L;
536 
537     for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
538       ValueToValueMapTy VMap;
539       BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
540       Header->getParent()->getBasicBlockList().insert(BlockInsertPt, New);
541 
542       assert((*BB != Header || LI->getLoopFor(*BB) == L) &&
543              "Header should not be in a sub-loop");
544       // Tell LI about New.
545       const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops);
546       if (OldLoop)
547         LoopsToSimplify.insert(NewLoops[OldLoop]);
548 
549       if (*BB == Header)
550         // Loop over all of the PHI nodes in the block, changing them to use
551         // the incoming values from the previous block.
552         for (PHINode *OrigPHI : OrigPHINode) {
553           PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]);
554           Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
555           if (Instruction *InValI = dyn_cast<Instruction>(InVal))
556             if (It > 1 && L->contains(InValI))
557               InVal = LastValueMap[InValI];
558           VMap[OrigPHI] = InVal;
559           New->getInstList().erase(NewPHI);
560         }
561 
562       // Update our running map of newest clones
563       LastValueMap[*BB] = New;
564       for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
565            VI != VE; ++VI)
566         LastValueMap[VI->first] = VI->second;
567 
568       // Add phi entries for newly created values to all exit blocks.
569       for (BasicBlock *Succ : successors(*BB)) {
570         if (L->contains(Succ))
571           continue;
572         for (PHINode &PHI : Succ->phis()) {
573           Value *Incoming = PHI.getIncomingValueForBlock(*BB);
574           ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
575           if (It != LastValueMap.end())
576             Incoming = It->second;
577           PHI.addIncoming(Incoming, New);
578         }
579       }
580       // Keep track of new headers and latches as we create them, so that
581       // we can insert the proper branches later.
582       if (*BB == Header)
583         Headers.push_back(New);
584       if (*BB == LatchBlock)
585         Latches.push_back(New);
586 
587       // Keep track of the exiting block and its successor block contained in
588       // the loop for the current iteration.
589       auto ExitInfoIt = ExitInfos.find(*BB);
590       if (ExitInfoIt != ExitInfos.end())
591         ExitInfoIt->second.ExitingBlocks.push_back(New);
592 
593       NewBlocks.push_back(New);
594       UnrolledLoopBlocks.push_back(New);
595 
596       // Update DomTree: since we just copy the loop body, and each copy has a
597       // dedicated entry block (copy of the header block), this header's copy
598       // dominates all copied blocks. That means, dominance relations in the
599       // copied body are the same as in the original body.
600       if (*BB == Header)
601         DT->addNewBlock(New, Latches[It - 1]);
602       else {
603         auto BBDomNode = DT->getNode(*BB);
604         auto BBIDom = BBDomNode->getIDom();
605         BasicBlock *OriginalBBIDom = BBIDom->getBlock();
606         DT->addNewBlock(
607             New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
608       }
609     }
610 
611     // Remap all instructions in the most recent iteration
612     remapInstructionsInBlocks(NewBlocks, LastValueMap);
613     for (BasicBlock *NewBlock : NewBlocks)
614       for (Instruction &I : *NewBlock)
615         if (auto *II = dyn_cast<AssumeInst>(&I))
616           AC->registerAssumption(II);
617 
618     {
619       // Identify what other metadata depends on the cloned version. After
620       // cloning, replace the metadata with the corrected version for both
621       // memory instructions and noalias intrinsics.
622       std::string ext = (Twine("It") + Twine(It)).str();
623       cloneAndAdaptNoAliasScopes(LoopLocalNoAliasDeclScopes, NewBlocks,
624                                  Header->getContext(), ext);
625     }
626   }
627 
628   // Loop over the PHI nodes in the original block, setting incoming values.
629   for (PHINode *PN : OrigPHINode) {
630     if (CompletelyUnroll) {
631       PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
632       Header->getInstList().erase(PN);
633     } else if (ULO.Count > 1) {
634       Value *InVal = PN->removeIncomingValue(LatchBlock, false);
635       // If this value was defined in the loop, take the value defined by the
636       // last iteration of the loop.
637       if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
638         if (L->contains(InValI))
639           InVal = LastValueMap[InVal];
640       }
641       assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
642       PN->addIncoming(InVal, Latches.back());
643     }
644   }
645 
646   // Connect latches of the unrolled iterations to the headers of the next
647   // iteration. Currently they point to the header of the same iteration.
648   for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
649     unsigned j = (i + 1) % e;
650     Latches[i]->getTerminator()->replaceSuccessorWith(Headers[i], Headers[j]);
651   }
652 
653   // Update dominators of blocks we might reach through exits.
654   // Immediate dominator of such block might change, because we add more
655   // routes which can lead to the exit: we can now reach it from the copied
656   // iterations too.
657   if (ULO.Count > 1) {
658     for (auto *BB : OriginalLoopBlocks) {
659       auto *BBDomNode = DT->getNode(BB);
660       SmallVector<BasicBlock *, 16> ChildrenToUpdate;
661       for (auto *ChildDomNode : BBDomNode->children()) {
662         auto *ChildBB = ChildDomNode->getBlock();
663         if (!L->contains(ChildBB))
664           ChildrenToUpdate.push_back(ChildBB);
665       }
666       // The new idom of the block will be the nearest common dominator
667       // of all copies of the previous idom. This is equivalent to the
668       // nearest common dominator of the previous idom and the first latch,
669       // which dominates all copies of the previous idom.
670       BasicBlock *NewIDom = DT->findNearestCommonDominator(BB, LatchBlock);
671       for (auto *ChildBB : ChildrenToUpdate)
672         DT->changeImmediateDominator(ChildBB, NewIDom);
673     }
674   }
675 
676   assert(!UnrollVerifyDomtree ||
677          DT->verify(DominatorTree::VerificationLevel::Fast));
678 
679   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
680 
681   auto SetDest = [&](BasicBlock *Src, bool WillExit, bool ExitOnTrue) {
682     auto *Term = cast<BranchInst>(Src->getTerminator());
683     const unsigned Idx = ExitOnTrue ^ WillExit;
684     BasicBlock *Dest = Term->getSuccessor(Idx);
685     BasicBlock *DeadSucc = Term->getSuccessor(1-Idx);
686 
687     // Remove predecessors from all non-Dest successors.
688     DeadSucc->removePredecessor(Src, /* KeepOneInputPHIs */ true);
689 
690     // Replace the conditional branch with an unconditional one.
691     BranchInst::Create(Dest, Term);
692     Term->eraseFromParent();
693 
694     DTU.applyUpdates({{DominatorTree::Delete, Src, DeadSucc}});
695   };
696 
697   auto WillExit = [&](const ExitInfo &Info, unsigned i, unsigned j,
698                       bool IsLatch) -> Optional<bool> {
699     if (CompletelyUnroll) {
700       if (PreserveOnlyFirst) {
701         if (i == 0)
702           return None;
703         return j == 0;
704       }
705       // Complete (but possibly inexact) unrolling
706       if (j == 0)
707         return true;
708       if (Info.TripCount && j != Info.TripCount)
709         return false;
710       return None;
711     }
712 
713     if (ULO.Runtime) {
714       // If runtime unrolling inserts a prologue, information about non-latch
715       // exits may be stale.
716       if (IsLatch && j != 0)
717         return false;
718       return None;
719     }
720 
721     if (j != Info.BreakoutTrip &&
722         (Info.TripMultiple == 0 || j % Info.TripMultiple != 0)) {
723       // If we know the trip count or a multiple of it, we can safely use an
724       // unconditional branch for some iterations.
725       return false;
726     }
727     return None;
728   };
729 
730   // Fold branches for iterations where we know that they will exit or not
731   // exit.
732   for (const auto &Pair : ExitInfos) {
733     const ExitInfo &Info = Pair.second;
734     for (unsigned i = 0, e = Info.ExitingBlocks.size(); i != e; ++i) {
735       // The branch destination.
736       unsigned j = (i + 1) % e;
737       bool IsLatch = Pair.first == LatchBlock;
738       Optional<bool> KnownWillExit = WillExit(Info, i, j, IsLatch);
739       if (!KnownWillExit)
740         continue;
741 
742       // We don't fold known-exiting branches for non-latch exits here,
743       // because this ensures that both all loop blocks and all exit blocks
744       // remain reachable in the CFG.
745       // TODO: We could fold these branches, but it would require much more
746       // sophisticated updates to LoopInfo.
747       if (*KnownWillExit && !IsLatch)
748         continue;
749 
750       SetDest(Info.ExitingBlocks[i], *KnownWillExit, Info.ExitOnTrue);
751     }
752   }
753 
754   // When completely unrolling, the last latch becomes unreachable.
755   if (!LatchIsExiting && CompletelyUnroll)
756     changeToUnreachable(Latches.back()->getTerminator(), PreserveLCSSA, &DTU);
757 
758   // Merge adjacent basic blocks, if possible.
759   for (BasicBlock *Latch : Latches) {
760     BranchInst *Term = dyn_cast<BranchInst>(Latch->getTerminator());
761     assert((Term ||
762             (CompletelyUnroll && !LatchIsExiting && Latch == Latches.back())) &&
763            "Need a branch as terminator, except when fully unrolling with "
764            "unconditional latch");
765     if (Term && Term->isUnconditional()) {
766       BasicBlock *Dest = Term->getSuccessor(0);
767       BasicBlock *Fold = Dest->getUniquePredecessor();
768       if (MergeBlockIntoPredecessor(Dest, &DTU, LI)) {
769         // Dest has been folded into Fold. Update our worklists accordingly.
770         std::replace(Latches.begin(), Latches.end(), Dest, Fold);
771         llvm::erase_value(UnrolledLoopBlocks, Dest);
772       }
773     }
774   }
775   // Apply updates to the DomTree.
776   DT = &DTU.getDomTree();
777 
778   assert(!UnrollVerifyDomtree ||
779          DT->verify(DominatorTree::VerificationLevel::Fast));
780 
781   // At this point, the code is well formed.  We now simplify the unrolled loop,
782   // doing constant propagation and dead code elimination as we go.
783   simplifyLoopAfterUnroll(L, !CompletelyUnroll && ULO.Count > 1, LI, SE, DT, AC,
784                           TTI);
785 
786   NumCompletelyUnrolled += CompletelyUnroll;
787   ++NumUnrolled;
788 
789   Loop *OuterL = L->getParentLoop();
790   // Update LoopInfo if the loop is completely removed.
791   if (CompletelyUnroll)
792     LI->erase(L);
793 
794   // LoopInfo should not be valid, confirm that.
795   if (UnrollVerifyLoopInfo)
796     LI->verify(*DT);
797 
798   // After complete unrolling most of the blocks should be contained in OuterL.
799   // However, some of them might happen to be out of OuterL (e.g. if they
800   // precede a loop exit). In this case we might need to insert PHI nodes in
801   // order to preserve LCSSA form.
802   // We don't need to check this if we already know that we need to fix LCSSA
803   // form.
804   // TODO: For now we just recompute LCSSA for the outer loop in this case, but
805   // it should be possible to fix it in-place.
806   if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA)
807     NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI);
808 
809   // Make sure that loop-simplify form is preserved. We want to simplify
810   // at least one layer outside of the loop that was unrolled so that any
811   // changes to the parent loop exposed by the unrolling are considered.
812   if (OuterL) {
813     // OuterL includes all loops for which we can break loop-simplify, so
814     // it's sufficient to simplify only it (it'll recursively simplify inner
815     // loops too).
816     if (NeedToFixLCSSA) {
817       // LCSSA must be performed on the outermost affected loop. The unrolled
818       // loop's last loop latch is guaranteed to be in the outermost loop
819       // after LoopInfo's been updated by LoopInfo::erase.
820       Loop *LatchLoop = LI->getLoopFor(Latches.back());
821       Loop *FixLCSSALoop = OuterL;
822       if (!FixLCSSALoop->contains(LatchLoop))
823         while (FixLCSSALoop->getParentLoop() != LatchLoop)
824           FixLCSSALoop = FixLCSSALoop->getParentLoop();
825 
826       formLCSSARecursively(*FixLCSSALoop, *DT, LI, SE);
827     } else if (PreserveLCSSA) {
828       assert(OuterL->isLCSSAForm(*DT) &&
829              "Loops should be in LCSSA form after loop-unroll.");
830     }
831 
832     // TODO: That potentially might be compile-time expensive. We should try
833     // to fix the loop-simplified form incrementally.
834     simplifyLoop(OuterL, DT, LI, SE, AC, nullptr, PreserveLCSSA);
835   } else {
836     // Simplify loops for which we might've broken loop-simplify form.
837     for (Loop *SubLoop : LoopsToSimplify)
838       simplifyLoop(SubLoop, DT, LI, SE, AC, nullptr, PreserveLCSSA);
839   }
840 
841   return CompletelyUnroll ? LoopUnrollResult::FullyUnrolled
842                           : LoopUnrollResult::PartiallyUnrolled;
843 }
844 
845 /// Given an llvm.loop loop id metadata node, returns the loop hint metadata
846 /// node with the given name (for example, "llvm.loop.unroll.count"). If no
847 /// such metadata node exists, then nullptr is returned.
848 MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) {
849   // First operand should refer to the loop id itself.
850   assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
851   assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
852 
853   for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
854     MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
855     if (!MD)
856       continue;
857 
858     MDString *S = dyn_cast<MDString>(MD->getOperand(0));
859     if (!S)
860       continue;
861 
862     if (Name.equals(S->getString()))
863       return MD;
864   }
865   return nullptr;
866 }
867