xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
1 ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===//
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 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
10 #include "llvm/ADT/DenseMap.h"
11 #include "llvm/ADT/STLExtras.h"
12 #include "llvm/ADT/Sequence.h"
13 #include "llvm/ADT/SetVector.h"
14 #include "llvm/ADT/SmallPtrSet.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/ADT/Twine.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/BlockFrequencyInfo.h"
20 #include "llvm/Analysis/CFG.h"
21 #include "llvm/Analysis/CodeMetrics.h"
22 #include "llvm/Analysis/DomTreeUpdater.h"
23 #include "llvm/Analysis/GuardUtils.h"
24 #include "llvm/Analysis/LoopAnalysisManager.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/LoopIterator.h"
27 #include "llvm/Analysis/MemorySSA.h"
28 #include "llvm/Analysis/MemorySSAUpdater.h"
29 #include "llvm/Analysis/MustExecute.h"
30 #include "llvm/Analysis/ProfileSummaryInfo.h"
31 #include "llvm/Analysis/ScalarEvolution.h"
32 #include "llvm/Analysis/TargetTransformInfo.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/Constant.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/InstrTypes.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/PatternMatch.h"
45 #include "llvm/IR/ProfDataUtils.h"
46 #include "llvm/IR/Use.h"
47 #include "llvm/IR/Value.h"
48 #include "llvm/Support/Casting.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/GenericDomTree.h"
53 #include "llvm/Support/InstructionCost.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/Scalar/LoopPassManager.h"
56 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
57 #include "llvm/Transforms/Utils/Cloning.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include "llvm/Transforms/Utils/LoopUtils.h"
60 #include "llvm/Transforms/Utils/ValueMapper.h"
61 #include <algorithm>
62 #include <cassert>
63 #include <iterator>
64 #include <numeric>
65 #include <optional>
66 #include <utility>
67 
68 #define DEBUG_TYPE "simple-loop-unswitch"
69 
70 using namespace llvm;
71 using namespace llvm::PatternMatch;
72 
73 STATISTIC(NumBranches, "Number of branches unswitched");
74 STATISTIC(NumSwitches, "Number of switches unswitched");
75 STATISTIC(NumSelects, "Number of selects turned into branches for unswitching");
76 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
77 STATISTIC(NumTrivial, "Number of unswitches that are trivial");
78 STATISTIC(
79     NumCostMultiplierSkipped,
80     "Number of unswitch candidates that had their cost multiplier skipped");
81 STATISTIC(NumInvariantConditionsInjected,
82           "Number of invariant conditions injected and unswitched");
83 
84 static cl::opt<bool> EnableNonTrivialUnswitch(
85     "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
86     cl::desc("Forcibly enables non-trivial loop unswitching rather than "
87              "following the configuration passed into the pass."));
88 
89 static cl::opt<int>
90     UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
91                       cl::desc("The cost threshold for unswitching a loop."));
92 
93 static cl::opt<bool> EnableUnswitchCostMultiplier(
94     "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden,
95     cl::desc("Enable unswitch cost multiplier that prohibits exponential "
96              "explosion in nontrivial unswitch."));
97 static cl::opt<int> UnswitchSiblingsToplevelDiv(
98     "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden,
99     cl::desc("Toplevel siblings divisor for cost multiplier."));
100 static cl::opt<int> UnswitchNumInitialUnscaledCandidates(
101     "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden,
102     cl::desc("Number of unswitch candidates that are ignored when calculating "
103              "cost multiplier."));
104 static cl::opt<bool> UnswitchGuards(
105     "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
106     cl::desc("If enabled, simple loop unswitching will also consider "
107              "llvm.experimental.guard intrinsics as unswitch candidates."));
108 static cl::opt<bool> DropNonTrivialImplicitNullChecks(
109     "simple-loop-unswitch-drop-non-trivial-implicit-null-checks",
110     cl::init(false), cl::Hidden,
111     cl::desc("If enabled, drop make.implicit metadata in unswitched implicit "
112              "null checks to save time analyzing if we can keep it."));
113 static cl::opt<unsigned>
114     MSSAThreshold("simple-loop-unswitch-memoryssa-threshold",
115                   cl::desc("Max number of memory uses to explore during "
116                            "partial unswitching analysis"),
117                   cl::init(100), cl::Hidden);
118 static cl::opt<bool> FreezeLoopUnswitchCond(
119     "freeze-loop-unswitch-cond", cl::init(true), cl::Hidden,
120     cl::desc("If enabled, the freeze instruction will be added to condition "
121              "of loop unswitch to prevent miscompilation."));
122 
123 static cl::opt<bool> InjectInvariantConditions(
124     "simple-loop-unswitch-inject-invariant-conditions", cl::Hidden,
125     cl::desc("Whether we should inject new invariants and unswitch them to "
126              "eliminate some existing (non-invariant) conditions."),
127     cl::init(true));
128 
129 static cl::opt<unsigned> InjectInvariantConditionHotnesThreshold(
130     "simple-loop-unswitch-inject-invariant-condition-hotness-threshold",
131     cl::Hidden, cl::desc("Only try to inject loop invariant conditions and "
132                          "unswitch on them to eliminate branches that are "
133                          "not-taken 1/<this option> times or less."),
134     cl::init(16));
135 
136 namespace {
137 struct CompareDesc {
138   BranchInst *Term;
139   Value *Invariant;
140   BasicBlock *InLoopSucc;
141 
142   CompareDesc(BranchInst *Term, Value *Invariant, BasicBlock *InLoopSucc)
143       : Term(Term), Invariant(Invariant), InLoopSucc(InLoopSucc) {}
144 };
145 
146 struct InjectedInvariant {
147   ICmpInst::Predicate Pred;
148   Value *LHS;
149   Value *RHS;
150   BasicBlock *InLoopSucc;
151 
152   InjectedInvariant(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
153                     BasicBlock *InLoopSucc)
154       : Pred(Pred), LHS(LHS), RHS(RHS), InLoopSucc(InLoopSucc) {}
155 };
156 
157 struct NonTrivialUnswitchCandidate {
158   Instruction *TI = nullptr;
159   TinyPtrVector<Value *> Invariants;
160   std::optional<InstructionCost> Cost;
161   std::optional<InjectedInvariant> PendingInjection;
162   NonTrivialUnswitchCandidate(
163       Instruction *TI, ArrayRef<Value *> Invariants,
164       std::optional<InstructionCost> Cost = std::nullopt,
165       std::optional<InjectedInvariant> PendingInjection = std::nullopt)
166       : TI(TI), Invariants(Invariants), Cost(Cost),
167         PendingInjection(PendingInjection) {};
168 
169   bool hasPendingInjection() const { return PendingInjection.has_value(); }
170 };
171 } // end anonymous namespace.
172 
173 // Helper to skip (select x, true, false), which matches both a logical AND and
174 // OR and can confuse code that tries to determine if \p Cond is either a
175 // logical AND or OR but not both.
176 static Value *skipTrivialSelect(Value *Cond) {
177   Value *CondNext;
178   while (match(Cond, m_Select(m_Value(CondNext), m_One(), m_Zero())))
179     Cond = CondNext;
180   return Cond;
181 }
182 
183 /// Collect all of the loop invariant input values transitively used by the
184 /// homogeneous instruction graph from a given root.
185 ///
186 /// This essentially walks from a root recursively through loop variant operands
187 /// which have perform the same logical operation (AND or OR) and finds all
188 /// inputs which are loop invariant. For some operations these can be
189 /// re-associated and unswitched out of the loop entirely.
190 static TinyPtrVector<Value *>
191 collectHomogenousInstGraphLoopInvariants(const Loop &L, Instruction &Root,
192                                          const LoopInfo &LI) {
193   assert(!L.isLoopInvariant(&Root) &&
194          "Only need to walk the graph if root itself is not invariant.");
195   TinyPtrVector<Value *> Invariants;
196 
197   bool IsRootAnd = match(&Root, m_LogicalAnd());
198   bool IsRootOr  = match(&Root, m_LogicalOr());
199 
200   // Build a worklist and recurse through operators collecting invariants.
201   SmallVector<Instruction *, 4> Worklist;
202   SmallPtrSet<Instruction *, 8> Visited;
203   Worklist.push_back(&Root);
204   Visited.insert(&Root);
205   do {
206     Instruction &I = *Worklist.pop_back_val();
207     for (Value *OpV : I.operand_values()) {
208       // Skip constants as unswitching isn't interesting for them.
209       if (isa<Constant>(OpV))
210         continue;
211 
212       // Add it to our result if loop invariant.
213       if (L.isLoopInvariant(OpV)) {
214         Invariants.push_back(OpV);
215         continue;
216       }
217 
218       // If not an instruction with the same opcode, nothing we can do.
219       Instruction *OpI = dyn_cast<Instruction>(skipTrivialSelect(OpV));
220 
221       if (OpI && ((IsRootAnd && match(OpI, m_LogicalAnd())) ||
222                   (IsRootOr  && match(OpI, m_LogicalOr())))) {
223         // Visit this operand.
224         if (Visited.insert(OpI).second)
225           Worklist.push_back(OpI);
226       }
227     }
228   } while (!Worklist.empty());
229 
230   return Invariants;
231 }
232 
233 static void replaceLoopInvariantUses(const Loop &L, Value *Invariant,
234                                      Constant &Replacement) {
235   assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");
236 
237   // Replace uses of LIC in the loop with the given constant.
238   // We use make_early_inc_range as set invalidates the iterator.
239   for (Use &U : llvm::make_early_inc_range(Invariant->uses())) {
240     Instruction *UserI = dyn_cast<Instruction>(U.getUser());
241 
242     // Replace this use within the loop body.
243     if (UserI && L.contains(UserI))
244       U.set(&Replacement);
245   }
246 }
247 
248 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial
249 /// incoming values along this edge.
250 static bool areLoopExitPHIsLoopInvariant(const Loop &L,
251                                          const BasicBlock &ExitingBB,
252                                          const BasicBlock &ExitBB) {
253   for (const Instruction &I : ExitBB) {
254     auto *PN = dyn_cast<PHINode>(&I);
255     if (!PN)
256       // No more PHIs to check.
257       return true;
258 
259     // If the incoming value for this edge isn't loop invariant the unswitch
260     // won't be trivial.
261     if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))
262       return false;
263   }
264   llvm_unreachable("Basic blocks should never be empty!");
265 }
266 
267 /// Copy a set of loop invariant values \p ToDuplicate and insert them at the
268 /// end of \p BB and conditionally branch on the copied condition. We only
269 /// branch on a single value.
270 static void buildPartialUnswitchConditionalBranch(
271     BasicBlock &BB, ArrayRef<Value *> Invariants, bool Direction,
272     BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, bool InsertFreeze,
273     const Instruction *I, AssumptionCache *AC, const DominatorTree &DT) {
274   IRBuilder<> IRB(&BB);
275 
276   SmallVector<Value *> FrozenInvariants;
277   for (Value *Inv : Invariants) {
278     if (InsertFreeze && !isGuaranteedNotToBeUndefOrPoison(Inv, AC, I, &DT))
279       Inv = IRB.CreateFreeze(Inv, Inv->getName() + ".fr");
280     FrozenInvariants.push_back(Inv);
281   }
282 
283   Value *Cond = Direction ? IRB.CreateOr(FrozenInvariants)
284                           : IRB.CreateAnd(FrozenInvariants);
285   IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
286                    Direction ? &NormalSucc : &UnswitchedSucc);
287 }
288 
289 /// Copy a set of loop invariant values, and conditionally branch on them.
290 static void buildPartialInvariantUnswitchConditionalBranch(
291     BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction,
292     BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L,
293     MemorySSAUpdater *MSSAU) {
294   ValueToValueMapTy VMap;
295   for (auto *Val : reverse(ToDuplicate)) {
296     Instruction *Inst = cast<Instruction>(Val);
297     Instruction *NewInst = Inst->clone();
298     NewInst->insertInto(&BB, BB.end());
299     RemapInstruction(NewInst, VMap,
300                      RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
301     VMap[Val] = NewInst;
302 
303     if (!MSSAU)
304       continue;
305 
306     MemorySSA *MSSA = MSSAU->getMemorySSA();
307     if (auto *MemUse =
308             dyn_cast_or_null<MemoryUse>(MSSA->getMemoryAccess(Inst))) {
309       auto *DefiningAccess = MemUse->getDefiningAccess();
310       // Get the first defining access before the loop.
311       while (L.contains(DefiningAccess->getBlock())) {
312         // If the defining access is a MemoryPhi, get the incoming
313         // value for the pre-header as defining access.
314         if (auto *MemPhi = dyn_cast<MemoryPhi>(DefiningAccess))
315           DefiningAccess =
316               MemPhi->getIncomingValueForBlock(L.getLoopPreheader());
317         else
318           DefiningAccess = cast<MemoryDef>(DefiningAccess)->getDefiningAccess();
319       }
320       MSSAU->createMemoryAccessInBB(NewInst, DefiningAccess,
321                                     NewInst->getParent(),
322                                     MemorySSA::BeforeTerminator);
323     }
324   }
325 
326   IRBuilder<> IRB(&BB);
327   Value *Cond = VMap[ToDuplicate[0]];
328   IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
329                    Direction ? &NormalSucc : &UnswitchedSucc);
330 }
331 
332 /// Rewrite the PHI nodes in an unswitched loop exit basic block.
333 ///
334 /// Requires that the loop exit and unswitched basic block are the same, and
335 /// that the exiting block was a unique predecessor of that block. Rewrites the
336 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
337 /// PHI nodes from the old preheader that now contains the unswitched
338 /// terminator.
339 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
340                                                   BasicBlock &OldExitingBB,
341                                                   BasicBlock &OldPH) {
342   for (PHINode &PN : UnswitchedBB.phis()) {
343     // When the loop exit is directly unswitched we just need to update the
344     // incoming basic block. We loop to handle weird cases with repeated
345     // incoming blocks, but expect to typically only have one operand here.
346     for (auto i : seq<int>(0, PN.getNumOperands())) {
347       assert(PN.getIncomingBlock(i) == &OldExitingBB &&
348              "Found incoming block different from unique predecessor!");
349       PN.setIncomingBlock(i, &OldPH);
350     }
351   }
352 }
353 
354 /// Rewrite the PHI nodes in the loop exit basic block and the split off
355 /// unswitched block.
356 ///
357 /// Because the exit block remains an exit from the loop, this rewrites the
358 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
359 /// nodes into the unswitched basic block to select between the value in the
360 /// old preheader and the loop exit.
361 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
362                                                       BasicBlock &UnswitchedBB,
363                                                       BasicBlock &OldExitingBB,
364                                                       BasicBlock &OldPH,
365                                                       bool FullUnswitch) {
366   assert(&ExitBB != &UnswitchedBB &&
367          "Must have different loop exit and unswitched blocks!");
368   BasicBlock::iterator InsertPt = UnswitchedBB.begin();
369   for (PHINode &PN : ExitBB.phis()) {
370     auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,
371                                   PN.getName() + ".split");
372     NewPN->insertBefore(InsertPt);
373 
374     // Walk backwards over the old PHI node's inputs to minimize the cost of
375     // removing each one. We have to do this weird loop manually so that we
376     // create the same number of new incoming edges in the new PHI as we expect
377     // each case-based edge to be included in the unswitched switch in some
378     // cases.
379     // FIXME: This is really, really gross. It would be much cleaner if LLVM
380     // allowed us to create a single entry for a predecessor block without
381     // having separate entries for each "edge" even though these edges are
382     // required to produce identical results.
383     for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
384       if (PN.getIncomingBlock(i) != &OldExitingBB)
385         continue;
386 
387       Value *Incoming = PN.getIncomingValue(i);
388       if (FullUnswitch)
389         // No more edge from the old exiting block to the exit block.
390         PN.removeIncomingValue(i);
391 
392       NewPN->addIncoming(Incoming, &OldPH);
393     }
394 
395     // Now replace the old PHI with the new one and wire the old one in as an
396     // input to the new one.
397     PN.replaceAllUsesWith(NewPN);
398     NewPN->addIncoming(&PN, &ExitBB);
399   }
400 }
401 
402 /// Hoist the current loop up to the innermost loop containing a remaining exit.
403 ///
404 /// Because we've removed an exit from the loop, we may have changed the set of
405 /// loops reachable and need to move the current loop up the loop nest or even
406 /// to an entirely separate nest.
407 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
408                                  DominatorTree &DT, LoopInfo &LI,
409                                  MemorySSAUpdater *MSSAU, ScalarEvolution *SE) {
410   // If the loop is already at the top level, we can't hoist it anywhere.
411   Loop *OldParentL = L.getParentLoop();
412   if (!OldParentL)
413     return;
414 
415   SmallVector<BasicBlock *, 4> Exits;
416   L.getExitBlocks(Exits);
417   Loop *NewParentL = nullptr;
418   for (auto *ExitBB : Exits)
419     if (Loop *ExitL = LI.getLoopFor(ExitBB))
420       if (!NewParentL || NewParentL->contains(ExitL))
421         NewParentL = ExitL;
422 
423   if (NewParentL == OldParentL)
424     return;
425 
426   // The new parent loop (if different) should always contain the old one.
427   if (NewParentL)
428     assert(NewParentL->contains(OldParentL) &&
429            "Can only hoist this loop up the nest!");
430 
431   // The preheader will need to move with the body of this loop. However,
432   // because it isn't in this loop we also need to update the primary loop map.
433   assert(OldParentL == LI.getLoopFor(&Preheader) &&
434          "Parent loop of this loop should contain this loop's preheader!");
435   LI.changeLoopFor(&Preheader, NewParentL);
436 
437   // Remove this loop from its old parent.
438   OldParentL->removeChildLoop(&L);
439 
440   // Add the loop either to the new parent or as a top-level loop.
441   if (NewParentL)
442     NewParentL->addChildLoop(&L);
443   else
444     LI.addTopLevelLoop(&L);
445 
446   // Remove this loops blocks from the old parent and every other loop up the
447   // nest until reaching the new parent. Also update all of these
448   // no-longer-containing loops to reflect the nesting change.
449   for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
450        OldContainingL = OldContainingL->getParentLoop()) {
451     llvm::erase_if(OldContainingL->getBlocksVector(),
452                    [&](const BasicBlock *BB) {
453                      return BB == &Preheader || L.contains(BB);
454                    });
455 
456     OldContainingL->getBlocksSet().erase(&Preheader);
457     for (BasicBlock *BB : L.blocks())
458       OldContainingL->getBlocksSet().erase(BB);
459 
460     // Because we just hoisted a loop out of this one, we have essentially
461     // created new exit paths from it. That means we need to form LCSSA PHI
462     // nodes for values used in the no-longer-nested loop.
463     formLCSSA(*OldContainingL, DT, &LI, SE);
464 
465     // We shouldn't need to form dedicated exits because the exit introduced
466     // here is the (just split by unswitching) preheader. However, after trivial
467     // unswitching it is possible to get new non-dedicated exits out of parent
468     // loop so let's conservatively form dedicated exit blocks and figure out
469     // if we can optimize later.
470     formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU,
471                             /*PreserveLCSSA*/ true);
472   }
473 }
474 
475 // Return the top-most loop containing ExitBB and having ExitBB as exiting block
476 // or the loop containing ExitBB, if there is no parent loop containing ExitBB
477 // as exiting block.
478 static Loop *getTopMostExitingLoop(const BasicBlock *ExitBB,
479                                    const LoopInfo &LI) {
480   Loop *TopMost = LI.getLoopFor(ExitBB);
481   Loop *Current = TopMost;
482   while (Current) {
483     if (Current->isLoopExiting(ExitBB))
484       TopMost = Current;
485     Current = Current->getParentLoop();
486   }
487   return TopMost;
488 }
489 
490 /// Unswitch a trivial branch if the condition is loop invariant.
491 ///
492 /// This routine should only be called when loop code leading to the branch has
493 /// been validated as trivial (no side effects). This routine checks if the
494 /// condition is invariant and one of the successors is a loop exit. This
495 /// allows us to unswitch without duplicating the loop, making it trivial.
496 ///
497 /// If this routine fails to unswitch the branch it returns false.
498 ///
499 /// If the branch can be unswitched, this routine splits the preheader and
500 /// hoists the branch above that split. Preserves loop simplified form
501 /// (splitting the exit block as necessary). It simplifies the branch within
502 /// the loop to an unconditional branch but doesn't remove it entirely. Further
503 /// cleanup can be done with some simplifycfg like pass.
504 ///
505 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
506 /// invalidated by this.
507 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
508                                   LoopInfo &LI, ScalarEvolution *SE,
509                                   MemorySSAUpdater *MSSAU) {
510   assert(BI.isConditional() && "Can only unswitch a conditional branch!");
511   LLVM_DEBUG(dbgs() << "  Trying to unswitch branch: " << BI << "\n");
512 
513   // The loop invariant values that we want to unswitch.
514   TinyPtrVector<Value *> Invariants;
515 
516   // When true, we're fully unswitching the branch rather than just unswitching
517   // some input conditions to the branch.
518   bool FullUnswitch = false;
519 
520   Value *Cond = skipTrivialSelect(BI.getCondition());
521   if (L.isLoopInvariant(Cond)) {
522     Invariants.push_back(Cond);
523     FullUnswitch = true;
524   } else {
525     if (auto *CondInst = dyn_cast<Instruction>(Cond))
526       Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);
527     if (Invariants.empty()) {
528       LLVM_DEBUG(dbgs() << "   Couldn't find invariant inputs!\n");
529       return false;
530     }
531   }
532 
533   // Check that one of the branch's successors exits, and which one.
534   bool ExitDirection = true;
535   int LoopExitSuccIdx = 0;
536   auto *LoopExitBB = BI.getSuccessor(0);
537   if (L.contains(LoopExitBB)) {
538     ExitDirection = false;
539     LoopExitSuccIdx = 1;
540     LoopExitBB = BI.getSuccessor(1);
541     if (L.contains(LoopExitBB)) {
542       LLVM_DEBUG(dbgs() << "   Branch doesn't exit the loop!\n");
543       return false;
544     }
545   }
546   auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
547   auto *ParentBB = BI.getParent();
548   if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) {
549     LLVM_DEBUG(dbgs() << "   Loop exit PHI's aren't loop-invariant!\n");
550     return false;
551   }
552 
553   // When unswitching only part of the branch's condition, we need the exit
554   // block to be reached directly from the partially unswitched input. This can
555   // be done when the exit block is along the true edge and the branch condition
556   // is a graph of `or` operations, or the exit block is along the false edge
557   // and the condition is a graph of `and` operations.
558   if (!FullUnswitch) {
559     if (ExitDirection ? !match(Cond, m_LogicalOr())
560                       : !match(Cond, m_LogicalAnd())) {
561       LLVM_DEBUG(dbgs() << "   Branch condition is in improper form for "
562                            "non-full unswitch!\n");
563       return false;
564     }
565   }
566 
567   LLVM_DEBUG({
568     dbgs() << "    unswitching trivial invariant conditions for: " << BI
569            << "\n";
570     for (Value *Invariant : Invariants) {
571       dbgs() << "      " << *Invariant << " == true";
572       if (Invariant != Invariants.back())
573         dbgs() << " ||";
574       dbgs() << "\n";
575     }
576   });
577 
578   // If we have scalar evolutions, we need to invalidate them including this
579   // loop, the loop containing the exit block and the topmost parent loop
580   // exiting via LoopExitBB.
581   if (SE) {
582     if (const Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI))
583       SE->forgetLoop(ExitL);
584     else
585       // Forget the entire nest as this exits the entire nest.
586       SE->forgetTopmostLoop(&L);
587     SE->forgetBlockAndLoopDispositions();
588   }
589 
590   if (MSSAU && VerifyMemorySSA)
591     MSSAU->getMemorySSA()->verifyMemorySSA();
592 
593   // Split the preheader, so that we know that there is a safe place to insert
594   // the conditional branch. We will change the preheader to have a conditional
595   // branch on LoopCond.
596   BasicBlock *OldPH = L.getLoopPreheader();
597   BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
598 
599   // Now that we have a place to insert the conditional branch, create a place
600   // to branch to: this is the exit block out of the loop that we are
601   // unswitching. We need to split this if there are other loop predecessors.
602   // Because the loop is in simplified form, *any* other predecessor is enough.
603   BasicBlock *UnswitchedBB;
604   if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
605     assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
606            "A branch's parent isn't a predecessor!");
607     UnswitchedBB = LoopExitBB;
608   } else {
609     UnswitchedBB =
610         SplitBlock(LoopExitBB, LoopExitBB->begin(), &DT, &LI, MSSAU, "", false);
611   }
612 
613   if (MSSAU && VerifyMemorySSA)
614     MSSAU->getMemorySSA()->verifyMemorySSA();
615 
616   // Actually move the invariant uses into the unswitched position. If possible,
617   // we do this by moving the instructions, but when doing partial unswitching
618   // we do it by building a new merge of the values in the unswitched position.
619   OldPH->getTerminator()->eraseFromParent();
620   if (FullUnswitch) {
621     // If fully unswitching, we can use the existing branch instruction.
622     // Splice it into the old PH to gate reaching the new preheader and re-point
623     // its successors.
624     BI.moveBefore(*OldPH, OldPH->end());
625     BI.setCondition(Cond);
626     if (MSSAU) {
627       // Temporarily clone the terminator, to make MSSA update cheaper by
628       // separating "insert edge" updates from "remove edge" ones.
629       BI.clone()->insertInto(ParentBB, ParentBB->end());
630     } else {
631       // Create a new unconditional branch that will continue the loop as a new
632       // terminator.
633       BranchInst::Create(ContinueBB, ParentBB);
634     }
635     BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
636     BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
637   } else {
638     // Only unswitching a subset of inputs to the condition, so we will need to
639     // build a new branch that merges the invariant inputs.
640     if (ExitDirection)
641       assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalOr()) &&
642              "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the "
643              "condition!");
644     else
645       assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalAnd()) &&
646              "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the"
647              " condition!");
648     buildPartialUnswitchConditionalBranch(
649         *OldPH, Invariants, ExitDirection, *UnswitchedBB, *NewPH,
650         FreezeLoopUnswitchCond, OldPH->getTerminator(), nullptr, DT);
651   }
652 
653   // Update the dominator tree with the added edge.
654   DT.insertEdge(OldPH, UnswitchedBB);
655 
656   // After the dominator tree was updated with the added edge, update MemorySSA
657   // if available.
658   if (MSSAU) {
659     SmallVector<CFGUpdate, 1> Updates;
660     Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
661     MSSAU->applyInsertUpdates(Updates, DT);
662   }
663 
664   // Finish updating dominator tree and memory ssa for full unswitch.
665   if (FullUnswitch) {
666     if (MSSAU) {
667       // Remove the cloned branch instruction.
668       ParentBB->getTerminator()->eraseFromParent();
669       // Create unconditional branch now.
670       BranchInst::Create(ContinueBB, ParentBB);
671       MSSAU->removeEdge(ParentBB, LoopExitBB);
672     }
673     DT.deleteEdge(ParentBB, LoopExitBB);
674   }
675 
676   if (MSSAU && VerifyMemorySSA)
677     MSSAU->getMemorySSA()->verifyMemorySSA();
678 
679   // Rewrite the relevant PHI nodes.
680   if (UnswitchedBB == LoopExitBB)
681     rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
682   else
683     rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
684                                               *ParentBB, *OldPH, FullUnswitch);
685 
686   // The constant we can replace all of our invariants with inside the loop
687   // body. If any of the invariants have a value other than this the loop won't
688   // be entered.
689   ConstantInt *Replacement = ExitDirection
690                                  ? ConstantInt::getFalse(BI.getContext())
691                                  : ConstantInt::getTrue(BI.getContext());
692 
693   // Since this is an i1 condition we can also trivially replace uses of it
694   // within the loop with a constant.
695   for (Value *Invariant : Invariants)
696     replaceLoopInvariantUses(L, Invariant, *Replacement);
697 
698   // If this was full unswitching, we may have changed the nesting relationship
699   // for this loop so hoist it to its correct parent if needed.
700   if (FullUnswitch)
701     hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
702 
703   if (MSSAU && VerifyMemorySSA)
704     MSSAU->getMemorySSA()->verifyMemorySSA();
705 
706   LLVM_DEBUG(dbgs() << "    done: unswitching trivial branch...\n");
707   ++NumTrivial;
708   ++NumBranches;
709   return true;
710 }
711 
712 /// Unswitch a trivial switch if the condition is loop invariant.
713 ///
714 /// This routine should only be called when loop code leading to the switch has
715 /// been validated as trivial (no side effects). This routine checks if the
716 /// condition is invariant and that at least one of the successors is a loop
717 /// exit. This allows us to unswitch without duplicating the loop, making it
718 /// trivial.
719 ///
720 /// If this routine fails to unswitch the switch it returns false.
721 ///
722 /// If the switch can be unswitched, this routine splits the preheader and
723 /// copies the switch above that split. If the default case is one of the
724 /// exiting cases, it copies the non-exiting cases and points them at the new
725 /// preheader. If the default case is not exiting, it copies the exiting cases
726 /// and points the default at the preheader. It preserves loop simplified form
727 /// (splitting the exit blocks as necessary). It simplifies the switch within
728 /// the loop by removing now-dead cases. If the default case is one of those
729 /// unswitched, it replaces its destination with a new basic block containing
730 /// only unreachable. Such basic blocks, while technically loop exits, are not
731 /// considered for unswitching so this is a stable transform and the same
732 /// switch will not be revisited. If after unswitching there is only a single
733 /// in-loop successor, the switch is further simplified to an unconditional
734 /// branch. Still more cleanup can be done with some simplifycfg like pass.
735 ///
736 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
737 /// invalidated by this.
738 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
739                                   LoopInfo &LI, ScalarEvolution *SE,
740                                   MemorySSAUpdater *MSSAU) {
741   LLVM_DEBUG(dbgs() << "  Trying to unswitch switch: " << SI << "\n");
742   Value *LoopCond = SI.getCondition();
743 
744   // If this isn't switching on an invariant condition, we can't unswitch it.
745   if (!L.isLoopInvariant(LoopCond))
746     return false;
747 
748   auto *ParentBB = SI.getParent();
749 
750   // The same check must be used both for the default and the exit cases. We
751   // should never leave edges from the switch instruction to a basic block that
752   // we are unswitching, hence the condition used to determine the default case
753   // needs to also be used to populate ExitCaseIndices, which is then used to
754   // remove cases from the switch.
755   auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) {
756     // BBToCheck is not an exit block if it is inside loop L.
757     if (L.contains(&BBToCheck))
758       return false;
759     // BBToCheck is not trivial to unswitch if its phis aren't loop invariant.
760     if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck))
761       return false;
762     // We do not unswitch a block that only has an unreachable statement, as
763     // it's possible this is a previously unswitched block. Only unswitch if
764     // either the terminator is not unreachable, or, if it is, it's not the only
765     // instruction in the block.
766     auto *TI = BBToCheck.getTerminator();
767     bool isUnreachable = isa<UnreachableInst>(TI);
768     return !isUnreachable ||
769            (isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI));
770   };
771 
772   SmallVector<int, 4> ExitCaseIndices;
773   for (auto Case : SI.cases())
774     if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor()))
775       ExitCaseIndices.push_back(Case.getCaseIndex());
776   BasicBlock *DefaultExitBB = nullptr;
777   SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight =
778       SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0);
779   if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) {
780     DefaultExitBB = SI.getDefaultDest();
781   } else if (ExitCaseIndices.empty())
782     return false;
783 
784   LLVM_DEBUG(dbgs() << "    unswitching trivial switch...\n");
785 
786   if (MSSAU && VerifyMemorySSA)
787     MSSAU->getMemorySSA()->verifyMemorySSA();
788 
789   // We may need to invalidate SCEVs for the outermost loop reached by any of
790   // the exits.
791   Loop *OuterL = &L;
792 
793   if (DefaultExitBB) {
794     // Check the loop containing this exit.
795     Loop *ExitL = getTopMostExitingLoop(DefaultExitBB, LI);
796     if (!ExitL || ExitL->contains(OuterL))
797       OuterL = ExitL;
798   }
799   for (unsigned Index : ExitCaseIndices) {
800     auto CaseI = SI.case_begin() + Index;
801     // Compute the outer loop from this exit.
802     Loop *ExitL = getTopMostExitingLoop(CaseI->getCaseSuccessor(), LI);
803     if (!ExitL || ExitL->contains(OuterL))
804       OuterL = ExitL;
805   }
806 
807   if (SE) {
808     if (OuterL)
809       SE->forgetLoop(OuterL);
810     else
811       SE->forgetTopmostLoop(&L);
812   }
813 
814   if (DefaultExitBB) {
815     // Clear out the default destination temporarily to allow accurate
816     // predecessor lists to be examined below.
817     SI.setDefaultDest(nullptr);
818   }
819 
820   // Store the exit cases into a separate data structure and remove them from
821   // the switch.
822   SmallVector<std::tuple<ConstantInt *, BasicBlock *,
823                          SwitchInstProfUpdateWrapper::CaseWeightOpt>,
824               4> ExitCases;
825   ExitCases.reserve(ExitCaseIndices.size());
826   SwitchInstProfUpdateWrapper SIW(SI);
827   // We walk the case indices backwards so that we remove the last case first
828   // and don't disrupt the earlier indices.
829   for (unsigned Index : reverse(ExitCaseIndices)) {
830     auto CaseI = SI.case_begin() + Index;
831     // Save the value of this case.
832     auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex());
833     ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W);
834     // Delete the unswitched cases.
835     SIW.removeCase(CaseI);
836   }
837 
838   // Check if after this all of the remaining cases point at the same
839   // successor.
840   BasicBlock *CommonSuccBB = nullptr;
841   if (SI.getNumCases() > 0 &&
842       all_of(drop_begin(SI.cases()), [&SI](const SwitchInst::CaseHandle &Case) {
843         return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor();
844       }))
845     CommonSuccBB = SI.case_begin()->getCaseSuccessor();
846   if (!DefaultExitBB) {
847     // If we're not unswitching the default, we need it to match any cases to
848     // have a common successor or if we have no cases it is the common
849     // successor.
850     if (SI.getNumCases() == 0)
851       CommonSuccBB = SI.getDefaultDest();
852     else if (SI.getDefaultDest() != CommonSuccBB)
853       CommonSuccBB = nullptr;
854   }
855 
856   // Split the preheader, so that we know that there is a safe place to insert
857   // the switch.
858   BasicBlock *OldPH = L.getLoopPreheader();
859   BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
860   OldPH->getTerminator()->eraseFromParent();
861 
862   // Now add the unswitched switch.
863   auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
864   SwitchInstProfUpdateWrapper NewSIW(*NewSI);
865 
866   // Rewrite the IR for the unswitched basic blocks. This requires two steps.
867   // First, we split any exit blocks with remaining in-loop predecessors. Then
868   // we update the PHIs in one of two ways depending on if there was a split.
869   // We walk in reverse so that we split in the same order as the cases
870   // appeared. This is purely for convenience of reading the resulting IR, but
871   // it doesn't cost anything really.
872   SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
873   SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
874   // Handle the default exit if necessary.
875   // FIXME: It'd be great if we could merge this with the loop below but LLVM's
876   // ranges aren't quite powerful enough yet.
877   if (DefaultExitBB) {
878     if (pred_empty(DefaultExitBB)) {
879       UnswitchedExitBBs.insert(DefaultExitBB);
880       rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
881     } else {
882       auto *SplitBB =
883           SplitBlock(DefaultExitBB, DefaultExitBB->begin(), &DT, &LI, MSSAU);
884       rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,
885                                                 *ParentBB, *OldPH,
886                                                 /*FullUnswitch*/ true);
887       DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
888     }
889   }
890   // Note that we must use a reference in the for loop so that we update the
891   // container.
892   for (auto &ExitCase : reverse(ExitCases)) {
893     // Grab a reference to the exit block in the pair so that we can update it.
894     BasicBlock *ExitBB = std::get<1>(ExitCase);
895 
896     // If this case is the last edge into the exit block, we can simply reuse it
897     // as it will no longer be a loop exit. No mapping necessary.
898     if (pred_empty(ExitBB)) {
899       // Only rewrite once.
900       if (UnswitchedExitBBs.insert(ExitBB).second)
901         rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
902       continue;
903     }
904 
905     // Otherwise we need to split the exit block so that we retain an exit
906     // block from the loop and a target for the unswitched condition.
907     BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
908     if (!SplitExitBB) {
909       // If this is the first time we see this, do the split and remember it.
910       SplitExitBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU);
911       rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,
912                                                 *ParentBB, *OldPH,
913                                                 /*FullUnswitch*/ true);
914     }
915     // Update the case pair to point to the split block.
916     std::get<1>(ExitCase) = SplitExitBB;
917   }
918 
919   // Now add the unswitched cases. We do this in reverse order as we built them
920   // in reverse order.
921   for (auto &ExitCase : reverse(ExitCases)) {
922     ConstantInt *CaseVal = std::get<0>(ExitCase);
923     BasicBlock *UnswitchedBB = std::get<1>(ExitCase);
924 
925     NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase));
926   }
927 
928   // If the default was unswitched, re-point it and add explicit cases for
929   // entering the loop.
930   if (DefaultExitBB) {
931     NewSIW->setDefaultDest(DefaultExitBB);
932     NewSIW.setSuccessorWeight(0, DefaultCaseWeight);
933 
934     // We removed all the exit cases, so we just copy the cases to the
935     // unswitched switch.
936     for (const auto &Case : SI.cases())
937       NewSIW.addCase(Case.getCaseValue(), NewPH,
938                      SIW.getSuccessorWeight(Case.getSuccessorIndex()));
939   } else if (DefaultCaseWeight) {
940     // We have to set branch weight of the default case.
941     uint64_t SW = *DefaultCaseWeight;
942     for (const auto &Case : SI.cases()) {
943       auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex());
944       assert(W &&
945              "case weight must be defined as default case weight is defined");
946       SW += *W;
947     }
948     NewSIW.setSuccessorWeight(0, SW);
949   }
950 
951   // If we ended up with a common successor for every path through the switch
952   // after unswitching, rewrite it to an unconditional branch to make it easy
953   // to recognize. Otherwise we potentially have to recognize the default case
954   // pointing at unreachable and other complexity.
955   if (CommonSuccBB) {
956     BasicBlock *BB = SI.getParent();
957     // We may have had multiple edges to this common successor block, so remove
958     // them as predecessors. We skip the first one, either the default or the
959     // actual first case.
960     bool SkippedFirst = DefaultExitBB == nullptr;
961     for (auto Case : SI.cases()) {
962       assert(Case.getCaseSuccessor() == CommonSuccBB &&
963              "Non-common successor!");
964       (void)Case;
965       if (!SkippedFirst) {
966         SkippedFirst = true;
967         continue;
968       }
969       CommonSuccBB->removePredecessor(BB,
970                                       /*KeepOneInputPHIs*/ true);
971     }
972     // Now nuke the switch and replace it with a direct branch.
973     SIW.eraseFromParent();
974     BranchInst::Create(CommonSuccBB, BB);
975   } else if (DefaultExitBB) {
976     assert(SI.getNumCases() > 0 &&
977            "If we had no cases we'd have a common successor!");
978     // Move the last case to the default successor. This is valid as if the
979     // default got unswitched it cannot be reached. This has the advantage of
980     // being simple and keeping the number of edges from this switch to
981     // successors the same, and avoiding any PHI update complexity.
982     auto LastCaseI = std::prev(SI.case_end());
983 
984     SI.setDefaultDest(LastCaseI->getCaseSuccessor());
985     SIW.setSuccessorWeight(
986         0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex()));
987     SIW.removeCase(LastCaseI);
988   }
989 
990   // Walk the unswitched exit blocks and the unswitched split blocks and update
991   // the dominator tree based on the CFG edits. While we are walking unordered
992   // containers here, the API for applyUpdates takes an unordered list of
993   // updates and requires them to not contain duplicates.
994   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
995   for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
996     DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});
997     DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});
998   }
999   for (auto SplitUnswitchedPair : SplitExitBBMap) {
1000     DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first});
1001     DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second});
1002   }
1003 
1004   if (MSSAU) {
1005     MSSAU->applyUpdates(DTUpdates, DT, /*UpdateDT=*/true);
1006     if (VerifyMemorySSA)
1007       MSSAU->getMemorySSA()->verifyMemorySSA();
1008   } else {
1009     DT.applyUpdates(DTUpdates);
1010   }
1011 
1012   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1013 
1014   // We may have changed the nesting relationship for this loop so hoist it to
1015   // its correct parent if needed.
1016   hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
1017 
1018   if (MSSAU && VerifyMemorySSA)
1019     MSSAU->getMemorySSA()->verifyMemorySSA();
1020 
1021   ++NumTrivial;
1022   ++NumSwitches;
1023   LLVM_DEBUG(dbgs() << "    done: unswitching trivial switch...\n");
1024   return true;
1025 }
1026 
1027 /// This routine scans the loop to find a branch or switch which occurs before
1028 /// any side effects occur. These can potentially be unswitched without
1029 /// duplicating the loop. If a branch or switch is successfully unswitched the
1030 /// scanning continues to see if subsequent branches or switches have become
1031 /// trivial. Once all trivial candidates have been unswitched, this routine
1032 /// returns.
1033 ///
1034 /// The return value indicates whether anything was unswitched (and therefore
1035 /// changed).
1036 ///
1037 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
1038 /// invalidated by this.
1039 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
1040                                          LoopInfo &LI, ScalarEvolution *SE,
1041                                          MemorySSAUpdater *MSSAU) {
1042   bool Changed = false;
1043 
1044   // If loop header has only one reachable successor we should keep looking for
1045   // trivial condition candidates in the successor as well. An alternative is
1046   // to constant fold conditions and merge successors into loop header (then we
1047   // only need to check header's terminator). The reason for not doing this in
1048   // LoopUnswitch pass is that it could potentially break LoopPassManager's
1049   // invariants. Folding dead branches could either eliminate the current loop
1050   // or make other loops unreachable. LCSSA form might also not be preserved
1051   // after deleting branches. The following code keeps traversing loop header's
1052   // successors until it finds the trivial condition candidate (condition that
1053   // is not a constant). Since unswitching generates branches with constant
1054   // conditions, this scenario could be very common in practice.
1055   BasicBlock *CurrentBB = L.getHeader();
1056   SmallPtrSet<BasicBlock *, 8> Visited;
1057   Visited.insert(CurrentBB);
1058   do {
1059     // Check if there are any side-effecting instructions (e.g. stores, calls,
1060     // volatile loads) in the part of the loop that the code *would* execute
1061     // without unswitching.
1062     if (MSSAU) // Possible early exit with MSSA
1063       if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB))
1064         if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end()))
1065           return Changed;
1066     if (llvm::any_of(*CurrentBB,
1067                      [](Instruction &I) { return I.mayHaveSideEffects(); }))
1068       return Changed;
1069 
1070     Instruction *CurrentTerm = CurrentBB->getTerminator();
1071 
1072     if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
1073       // Don't bother trying to unswitch past a switch with a constant
1074       // condition. This should be removed prior to running this pass by
1075       // simplifycfg.
1076       if (isa<Constant>(SI->getCondition()))
1077         return Changed;
1078 
1079       if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU))
1080         // Couldn't unswitch this one so we're done.
1081         return Changed;
1082 
1083       // Mark that we managed to unswitch something.
1084       Changed = true;
1085 
1086       // If unswitching turned the terminator into an unconditional branch then
1087       // we can continue. The unswitching logic specifically works to fold any
1088       // cases it can into an unconditional branch to make it easier to
1089       // recognize here.
1090       auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
1091       if (!BI || BI->isConditional())
1092         return Changed;
1093 
1094       CurrentBB = BI->getSuccessor(0);
1095       continue;
1096     }
1097 
1098     auto *BI = dyn_cast<BranchInst>(CurrentTerm);
1099     if (!BI)
1100       // We do not understand other terminator instructions.
1101       return Changed;
1102 
1103     // Don't bother trying to unswitch past an unconditional branch or a branch
1104     // with a constant value. These should be removed by simplifycfg prior to
1105     // running this pass.
1106     if (!BI->isConditional() ||
1107         isa<Constant>(skipTrivialSelect(BI->getCondition())))
1108       return Changed;
1109 
1110     // Found a trivial condition candidate: non-foldable conditional branch. If
1111     // we fail to unswitch this, we can't do anything else that is trivial.
1112     if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU))
1113       return Changed;
1114 
1115     // Mark that we managed to unswitch something.
1116     Changed = true;
1117 
1118     // If we only unswitched some of the conditions feeding the branch, we won't
1119     // have collapsed it to a single successor.
1120     BI = cast<BranchInst>(CurrentBB->getTerminator());
1121     if (BI->isConditional())
1122       return Changed;
1123 
1124     // Follow the newly unconditional branch into its successor.
1125     CurrentBB = BI->getSuccessor(0);
1126 
1127     // When continuing, if we exit the loop or reach a previous visited block,
1128     // then we can not reach any trivial condition candidates (unfoldable
1129     // branch instructions or switch instructions) and no unswitch can happen.
1130   } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
1131 
1132   return Changed;
1133 }
1134 
1135 /// Build the cloned blocks for an unswitched copy of the given loop.
1136 ///
1137 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
1138 /// after the split block (`SplitBB`) that will be used to select between the
1139 /// cloned and original loop.
1140 ///
1141 /// This routine handles cloning all of the necessary loop blocks and exit
1142 /// blocks including rewriting their instructions and the relevant PHI nodes.
1143 /// Any loop blocks or exit blocks which are dominated by a different successor
1144 /// than the one for this clone of the loop blocks can be trivially skipped. We
1145 /// use the `DominatingSucc` map to determine whether a block satisfies that
1146 /// property with a simple map lookup.
1147 ///
1148 /// It also correctly creates the unconditional branch in the cloned
1149 /// unswitched parent block to only point at the unswitched successor.
1150 ///
1151 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit
1152 /// block splitting is correctly reflected in `LoopInfo`, essentially all of
1153 /// the cloned blocks (and their loops) are left without full `LoopInfo`
1154 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
1155 /// blocks to them but doesn't create the cloned `DominatorTree` structure and
1156 /// instead the caller must recompute an accurate DT. It *does* correctly
1157 /// update the `AssumptionCache` provided in `AC`.
1158 static BasicBlock *buildClonedLoopBlocks(
1159     Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
1160     ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
1161     BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
1162     const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
1163     ValueToValueMapTy &VMap,
1164     SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
1165     DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU,
1166     ScalarEvolution *SE) {
1167   SmallVector<BasicBlock *, 4> NewBlocks;
1168   NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());
1169 
1170   // We will need to clone a bunch of blocks, wrap up the clone operation in
1171   // a helper.
1172   auto CloneBlock = [&](BasicBlock *OldBB) {
1173     // Clone the basic block and insert it before the new preheader.
1174     BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());
1175     NewBB->moveBefore(LoopPH);
1176 
1177     // Record this block and the mapping.
1178     NewBlocks.push_back(NewBB);
1179     VMap[OldBB] = NewBB;
1180 
1181     return NewBB;
1182   };
1183 
1184   // We skip cloning blocks when they have a dominating succ that is not the
1185   // succ we are cloning for.
1186   auto SkipBlock = [&](BasicBlock *BB) {
1187     auto It = DominatingSucc.find(BB);
1188     return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
1189   };
1190 
1191   // First, clone the preheader.
1192   auto *ClonedPH = CloneBlock(LoopPH);
1193 
1194   // Then clone all the loop blocks, skipping the ones that aren't necessary.
1195   for (auto *LoopBB : L.blocks())
1196     if (!SkipBlock(LoopBB))
1197       CloneBlock(LoopBB);
1198 
1199   // Split all the loop exit edges so that when we clone the exit blocks, if
1200   // any of the exit blocks are *also* a preheader for some other loop, we
1201   // don't create multiple predecessors entering the loop header.
1202   for (auto *ExitBB : ExitBlocks) {
1203     if (SkipBlock(ExitBB))
1204       continue;
1205 
1206     // When we are going to clone an exit, we don't need to clone all the
1207     // instructions in the exit block and we want to ensure we have an easy
1208     // place to merge the CFG, so split the exit first. This is always safe to
1209     // do because there cannot be any non-loop predecessors of a loop exit in
1210     // loop simplified form.
1211     auto *MergeBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU);
1212 
1213     // Rearrange the names to make it easier to write test cases by having the
1214     // exit block carry the suffix rather than the merge block carrying the
1215     // suffix.
1216     MergeBB->takeName(ExitBB);
1217     ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1218 
1219     // Now clone the original exit block.
1220     auto *ClonedExitBB = CloneBlock(ExitBB);
1221     assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1222            "Exit block should have been split to have one successor!");
1223     assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1224            "Cloned exit block has the wrong successor!");
1225 
1226     // Remap any cloned instructions and create a merge phi node for them.
1227     for (auto ZippedInsts : llvm::zip_first(
1228              llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),
1229              llvm::make_range(ClonedExitBB->begin(),
1230                               std::prev(ClonedExitBB->end())))) {
1231       Instruction &I = std::get<0>(ZippedInsts);
1232       Instruction &ClonedI = std::get<1>(ZippedInsts);
1233 
1234       // The only instructions in the exit block should be PHI nodes and
1235       // potentially a landing pad.
1236       assert(
1237           (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1238           "Bad instruction in exit block!");
1239       // We should have a value map between the instruction and its clone.
1240       assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1241 
1242       // Forget SCEVs based on exit phis in case SCEV looked through the phi.
1243       if (SE && isa<PHINode>(I))
1244         SE->forgetValue(&I);
1245 
1246       auto *MergePN =
1247           PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi");
1248       MergePN->insertBefore(MergeBB->getFirstInsertionPt());
1249       I.replaceAllUsesWith(MergePN);
1250       MergePN->addIncoming(&I, ExitBB);
1251       MergePN->addIncoming(&ClonedI, ClonedExitBB);
1252     }
1253   }
1254 
1255   // Rewrite the instructions in the cloned blocks to refer to the instructions
1256   // in the cloned blocks. We have to do this as a second pass so that we have
1257   // everything available. Also, we have inserted new instructions which may
1258   // include assume intrinsics, so we update the assumption cache while
1259   // processing this.
1260   Module *M = ClonedPH->getParent()->getParent();
1261   for (auto *ClonedBB : NewBlocks)
1262     for (Instruction &I : *ClonedBB) {
1263       RemapDPValueRange(M, I.getDbgValueRange(), VMap,
1264                         RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1265       RemapInstruction(&I, VMap,
1266                        RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1267       if (auto *II = dyn_cast<AssumeInst>(&I))
1268         AC.registerAssumption(II);
1269     }
1270 
1271   // Update any PHI nodes in the cloned successors of the skipped blocks to not
1272   // have spurious incoming values.
1273   for (auto *LoopBB : L.blocks())
1274     if (SkipBlock(LoopBB))
1275       for (auto *SuccBB : successors(LoopBB))
1276         if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))
1277           for (PHINode &PN : ClonedSuccBB->phis())
1278             PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);
1279 
1280   // Remove the cloned parent as a predecessor of any successor we ended up
1281   // cloning other than the unswitched one.
1282   auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));
1283   for (auto *SuccBB : successors(ParentBB)) {
1284     if (SuccBB == UnswitchedSuccBB)
1285       continue;
1286 
1287     auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));
1288     if (!ClonedSuccBB)
1289       continue;
1290 
1291     ClonedSuccBB->removePredecessor(ClonedParentBB,
1292                                     /*KeepOneInputPHIs*/ true);
1293   }
1294 
1295   // Replace the cloned branch with an unconditional branch to the cloned
1296   // unswitched successor.
1297   auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));
1298   Instruction *ClonedTerminator = ClonedParentBB->getTerminator();
1299   // Trivial Simplification. If Terminator is a conditional branch and
1300   // condition becomes dead - erase it.
1301   Value *ClonedConditionToErase = nullptr;
1302   if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator))
1303     ClonedConditionToErase = BI->getCondition();
1304   else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator))
1305     ClonedConditionToErase = SI->getCondition();
1306 
1307   ClonedTerminator->eraseFromParent();
1308   BranchInst::Create(ClonedSuccBB, ClonedParentBB);
1309 
1310   if (ClonedConditionToErase)
1311     RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr,
1312                                                MSSAU);
1313 
1314   // If there are duplicate entries in the PHI nodes because of multiple edges
1315   // to the unswitched successor, we need to nuke all but one as we replaced it
1316   // with a direct branch.
1317   for (PHINode &PN : ClonedSuccBB->phis()) {
1318     bool Found = false;
1319     // Loop over the incoming operands backwards so we can easily delete as we
1320     // go without invalidating the index.
1321     for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1322       if (PN.getIncomingBlock(i) != ClonedParentBB)
1323         continue;
1324       if (!Found) {
1325         Found = true;
1326         continue;
1327       }
1328       PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);
1329     }
1330   }
1331 
1332   // Record the domtree updates for the new blocks.
1333   SmallPtrSet<BasicBlock *, 4> SuccSet;
1334   for (auto *ClonedBB : NewBlocks) {
1335     for (auto *SuccBB : successors(ClonedBB))
1336       if (SuccSet.insert(SuccBB).second)
1337         DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});
1338     SuccSet.clear();
1339   }
1340 
1341   return ClonedPH;
1342 }
1343 
1344 /// Recursively clone the specified loop and all of its children.
1345 ///
1346 /// The target parent loop for the clone should be provided, or can be null if
1347 /// the clone is a top-level loop. While cloning, all the blocks are mapped
1348 /// with the provided value map. The entire original loop must be present in
1349 /// the value map. The cloned loop is returned.
1350 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1351                            const ValueToValueMapTy &VMap, LoopInfo &LI) {
1352   auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1353     assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1354     ClonedL.reserveBlocks(OrigL.getNumBlocks());
1355     for (auto *BB : OrigL.blocks()) {
1356       auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));
1357       ClonedL.addBlockEntry(ClonedBB);
1358       if (LI.getLoopFor(BB) == &OrigL)
1359         LI.changeLoopFor(ClonedBB, &ClonedL);
1360     }
1361   };
1362 
1363   // We specially handle the first loop because it may get cloned into
1364   // a different parent and because we most commonly are cloning leaf loops.
1365   Loop *ClonedRootL = LI.AllocateLoop();
1366   if (RootParentL)
1367     RootParentL->addChildLoop(ClonedRootL);
1368   else
1369     LI.addTopLevelLoop(ClonedRootL);
1370   AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1371 
1372   if (OrigRootL.isInnermost())
1373     return ClonedRootL;
1374 
1375   // If we have a nest, we can quickly clone the entire loop nest using an
1376   // iterative approach because it is a tree. We keep the cloned parent in the
1377   // data structure to avoid repeatedly querying through a map to find it.
1378   SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1379   // Build up the loops to clone in reverse order as we'll clone them from the
1380   // back.
1381   for (Loop *ChildL : llvm::reverse(OrigRootL))
1382     LoopsToClone.push_back({ClonedRootL, ChildL});
1383   do {
1384     Loop *ClonedParentL, *L;
1385     std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();
1386     Loop *ClonedL = LI.AllocateLoop();
1387     ClonedParentL->addChildLoop(ClonedL);
1388     AddClonedBlocksToLoop(*L, *ClonedL);
1389     for (Loop *ChildL : llvm::reverse(*L))
1390       LoopsToClone.push_back({ClonedL, ChildL});
1391   } while (!LoopsToClone.empty());
1392 
1393   return ClonedRootL;
1394 }
1395 
1396 /// Build the cloned loops of an original loop from unswitching.
1397 ///
1398 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1399 /// operation. We need to re-verify that there even is a loop (as the backedge
1400 /// may not have been cloned), and even if there are remaining backedges the
1401 /// backedge set may be different. However, we know that each child loop is
1402 /// undisturbed, we only need to find where to place each child loop within
1403 /// either any parent loop or within a cloned version of the original loop.
1404 ///
1405 /// Because child loops may end up cloned outside of any cloned version of the
1406 /// original loop, multiple cloned sibling loops may be created. All of them
1407 /// are returned so that the newly introduced loop nest roots can be
1408 /// identified.
1409 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1410                              const ValueToValueMapTy &VMap, LoopInfo &LI,
1411                              SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1412   Loop *ClonedL = nullptr;
1413 
1414   auto *OrigPH = OrigL.getLoopPreheader();
1415   auto *OrigHeader = OrigL.getHeader();
1416 
1417   auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));
1418   auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));
1419 
1420   // We need to know the loops of the cloned exit blocks to even compute the
1421   // accurate parent loop. If we only clone exits to some parent of the
1422   // original parent, we want to clone into that outer loop. We also keep track
1423   // of the loops that our cloned exit blocks participate in.
1424   Loop *ParentL = nullptr;
1425   SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1426   SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1427   ClonedExitsInLoops.reserve(ExitBlocks.size());
1428   for (auto *ExitBB : ExitBlocks)
1429     if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))
1430       if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1431         ExitLoopMap[ClonedExitBB] = ExitL;
1432         ClonedExitsInLoops.push_back(ClonedExitBB);
1433         if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1434           ParentL = ExitL;
1435       }
1436   assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1437           ParentL->contains(OrigL.getParentLoop())) &&
1438          "The computed parent loop should always contain (or be) the parent of "
1439          "the original loop.");
1440 
1441   // We build the set of blocks dominated by the cloned header from the set of
1442   // cloned blocks out of the original loop. While not all of these will
1443   // necessarily be in the cloned loop, it is enough to establish that they
1444   // aren't in unreachable cycles, etc.
1445   SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1446   for (auto *BB : OrigL.blocks())
1447     if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))
1448       ClonedLoopBlocks.insert(ClonedBB);
1449 
1450   // Rebuild the set of blocks that will end up in the cloned loop. We may have
1451   // skipped cloning some region of this loop which can in turn skip some of
1452   // the backedges so we have to rebuild the blocks in the loop based on the
1453   // backedges that remain after cloning.
1454   SmallVector<BasicBlock *, 16> Worklist;
1455   SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1456   for (auto *Pred : predecessors(ClonedHeader)) {
1457     // The only possible non-loop header predecessor is the preheader because
1458     // we know we cloned the loop in simplified form.
1459     if (Pred == ClonedPH)
1460       continue;
1461 
1462     // Because the loop was in simplified form, the only non-loop predecessor
1463     // should be the preheader.
1464     assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1465                                            "header other than the preheader "
1466                                            "that is not part of the loop!");
1467 
1468     // Insert this block into the loop set and on the first visit (and if it
1469     // isn't the header we're currently walking) put it into the worklist to
1470     // recurse through.
1471     if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)
1472       Worklist.push_back(Pred);
1473   }
1474 
1475   // If we had any backedges then there *is* a cloned loop. Put the header into
1476   // the loop set and then walk the worklist backwards to find all the blocks
1477   // that remain within the loop after cloning.
1478   if (!BlocksInClonedLoop.empty()) {
1479     BlocksInClonedLoop.insert(ClonedHeader);
1480 
1481     while (!Worklist.empty()) {
1482       BasicBlock *BB = Worklist.pop_back_val();
1483       assert(BlocksInClonedLoop.count(BB) &&
1484              "Didn't put block into the loop set!");
1485 
1486       // Insert any predecessors that are in the possible set into the cloned
1487       // set, and if the insert is successful, add them to the worklist. Note
1488       // that we filter on the blocks that are definitely reachable via the
1489       // backedge to the loop header so we may prune out dead code within the
1490       // cloned loop.
1491       for (auto *Pred : predecessors(BB))
1492         if (ClonedLoopBlocks.count(Pred) &&
1493             BlocksInClonedLoop.insert(Pred).second)
1494           Worklist.push_back(Pred);
1495     }
1496 
1497     ClonedL = LI.AllocateLoop();
1498     if (ParentL) {
1499       ParentL->addBasicBlockToLoop(ClonedPH, LI);
1500       ParentL->addChildLoop(ClonedL);
1501     } else {
1502       LI.addTopLevelLoop(ClonedL);
1503     }
1504     NonChildClonedLoops.push_back(ClonedL);
1505 
1506     ClonedL->reserveBlocks(BlocksInClonedLoop.size());
1507     // We don't want to just add the cloned loop blocks based on how we
1508     // discovered them. The original order of blocks was carefully built in
1509     // a way that doesn't rely on predecessor ordering. Rather than re-invent
1510     // that logic, we just re-walk the original blocks (and those of the child
1511     // loops) and filter them as we add them into the cloned loop.
1512     for (auto *BB : OrigL.blocks()) {
1513       auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));
1514       if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))
1515         continue;
1516 
1517       // Directly add the blocks that are only in this loop.
1518       if (LI.getLoopFor(BB) == &OrigL) {
1519         ClonedL->addBasicBlockToLoop(ClonedBB, LI);
1520         continue;
1521       }
1522 
1523       // We want to manually add it to this loop and parents.
1524       // Registering it with LoopInfo will happen when we clone the top
1525       // loop for this block.
1526       for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1527         PL->addBlockEntry(ClonedBB);
1528     }
1529 
1530     // Now add each child loop whose header remains within the cloned loop. All
1531     // of the blocks within the loop must satisfy the same constraints as the
1532     // header so once we pass the header checks we can just clone the entire
1533     // child loop nest.
1534     for (Loop *ChildL : OrigL) {
1535       auto *ClonedChildHeader =
1536           cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1537       if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))
1538         continue;
1539 
1540 #ifndef NDEBUG
1541       // We should never have a cloned child loop header but fail to have
1542       // all of the blocks for that child loop.
1543       for (auto *ChildLoopBB : ChildL->blocks())
1544         assert(BlocksInClonedLoop.count(
1545                    cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1546                "Child cloned loop has a header within the cloned outer "
1547                "loop but not all of its blocks!");
1548 #endif
1549 
1550       cloneLoopNest(*ChildL, ClonedL, VMap, LI);
1551     }
1552   }
1553 
1554   // Now that we've handled all the components of the original loop that were
1555   // cloned into a new loop, we still need to handle anything from the original
1556   // loop that wasn't in a cloned loop.
1557 
1558   // Figure out what blocks are left to place within any loop nest containing
1559   // the unswitched loop. If we never formed a loop, the cloned PH is one of
1560   // them.
1561   SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1562   if (BlocksInClonedLoop.empty())
1563     UnloopedBlockSet.insert(ClonedPH);
1564   for (auto *ClonedBB : ClonedLoopBlocks)
1565     if (!BlocksInClonedLoop.count(ClonedBB))
1566       UnloopedBlockSet.insert(ClonedBB);
1567 
1568   // Copy the cloned exits and sort them in ascending loop depth, we'll work
1569   // backwards across these to process them inside out. The order shouldn't
1570   // matter as we're just trying to build up the map from inside-out; we use
1571   // the map in a more stably ordered way below.
1572   auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1573   llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1574     return ExitLoopMap.lookup(LHS)->getLoopDepth() <
1575            ExitLoopMap.lookup(RHS)->getLoopDepth();
1576   });
1577 
1578   // Populate the existing ExitLoopMap with everything reachable from each
1579   // exit, starting from the inner most exit.
1580   while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1581     assert(Worklist.empty() && "Didn't clear worklist!");
1582 
1583     BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1584     Loop *ExitL = ExitLoopMap.lookup(ExitBB);
1585 
1586     // Walk the CFG back until we hit the cloned PH adding everything reachable
1587     // and in the unlooped set to this exit block's loop.
1588     Worklist.push_back(ExitBB);
1589     do {
1590       BasicBlock *BB = Worklist.pop_back_val();
1591       // We can stop recursing at the cloned preheader (if we get there).
1592       if (BB == ClonedPH)
1593         continue;
1594 
1595       for (BasicBlock *PredBB : predecessors(BB)) {
1596         // If this pred has already been moved to our set or is part of some
1597         // (inner) loop, no update needed.
1598         if (!UnloopedBlockSet.erase(PredBB)) {
1599           assert(
1600               (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1601               "Predecessor not mapped to a loop!");
1602           continue;
1603         }
1604 
1605         // We just insert into the loop set here. We'll add these blocks to the
1606         // exit loop after we build up the set in an order that doesn't rely on
1607         // predecessor order (which in turn relies on use list order).
1608         bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;
1609         (void)Inserted;
1610         assert(Inserted && "Should only visit an unlooped block once!");
1611 
1612         // And recurse through to its predecessors.
1613         Worklist.push_back(PredBB);
1614       }
1615     } while (!Worklist.empty());
1616   }
1617 
1618   // Now that the ExitLoopMap gives as  mapping for all the non-looping cloned
1619   // blocks to their outer loops, walk the cloned blocks and the cloned exits
1620   // in their original order adding them to the correct loop.
1621 
1622   // We need a stable insertion order. We use the order of the original loop
1623   // order and map into the correct parent loop.
1624   for (auto *BB : llvm::concat<BasicBlock *const>(
1625            ArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))
1626     if (Loop *OuterL = ExitLoopMap.lookup(BB))
1627       OuterL->addBasicBlockToLoop(BB, LI);
1628 
1629 #ifndef NDEBUG
1630   for (auto &BBAndL : ExitLoopMap) {
1631     auto *BB = BBAndL.first;
1632     auto *OuterL = BBAndL.second;
1633     assert(LI.getLoopFor(BB) == OuterL &&
1634            "Failed to put all blocks into outer loops!");
1635   }
1636 #endif
1637 
1638   // Now that all the blocks are placed into the correct containing loop in the
1639   // absence of child loops, find all the potentially cloned child loops and
1640   // clone them into whatever outer loop we placed their header into.
1641   for (Loop *ChildL : OrigL) {
1642     auto *ClonedChildHeader =
1643         cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1644     if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))
1645       continue;
1646 
1647 #ifndef NDEBUG
1648     for (auto *ChildLoopBB : ChildL->blocks())
1649       assert(VMap.count(ChildLoopBB) &&
1650              "Cloned a child loop header but not all of that loops blocks!");
1651 #endif
1652 
1653     NonChildClonedLoops.push_back(cloneLoopNest(
1654         *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));
1655   }
1656 }
1657 
1658 static void
1659 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1660                        ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1661                        DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1662   // Find all the dead clones, and remove them from their successors.
1663   SmallVector<BasicBlock *, 16> DeadBlocks;
1664   for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))
1665     for (const auto &VMap : VMaps)
1666       if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))
1667         if (!DT.isReachableFromEntry(ClonedBB)) {
1668           for (BasicBlock *SuccBB : successors(ClonedBB))
1669             SuccBB->removePredecessor(ClonedBB);
1670           DeadBlocks.push_back(ClonedBB);
1671         }
1672 
1673   // Remove all MemorySSA in the dead blocks
1674   if (MSSAU) {
1675     SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(),
1676                                                  DeadBlocks.end());
1677     MSSAU->removeBlocks(DeadBlockSet);
1678   }
1679 
1680   // Drop any remaining references to break cycles.
1681   for (BasicBlock *BB : DeadBlocks)
1682     BB->dropAllReferences();
1683   // Erase them from the IR.
1684   for (BasicBlock *BB : DeadBlocks)
1685     BB->eraseFromParent();
1686 }
1687 
1688 static void deleteDeadBlocksFromLoop(Loop &L,
1689                                      SmallVectorImpl<BasicBlock *> &ExitBlocks,
1690                                      DominatorTree &DT, LoopInfo &LI,
1691                                      MemorySSAUpdater *MSSAU,
1692                                      ScalarEvolution *SE,
1693                                      LPMUpdater &LoopUpdater) {
1694   // Find all the dead blocks tied to this loop, and remove them from their
1695   // successors.
1696   SmallSetVector<BasicBlock *, 8> DeadBlockSet;
1697 
1698   // Start with loop/exit blocks and get a transitive closure of reachable dead
1699   // blocks.
1700   SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1701                                                 ExitBlocks.end());
1702   DeathCandidates.append(L.blocks().begin(), L.blocks().end());
1703   while (!DeathCandidates.empty()) {
1704     auto *BB = DeathCandidates.pop_back_val();
1705     if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {
1706       for (BasicBlock *SuccBB : successors(BB)) {
1707         SuccBB->removePredecessor(BB);
1708         DeathCandidates.push_back(SuccBB);
1709       }
1710       DeadBlockSet.insert(BB);
1711     }
1712   }
1713 
1714   // Remove all MemorySSA in the dead blocks
1715   if (MSSAU)
1716     MSSAU->removeBlocks(DeadBlockSet);
1717 
1718   // Filter out the dead blocks from the exit blocks list so that it can be
1719   // used in the caller.
1720   llvm::erase_if(ExitBlocks,
1721                  [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1722 
1723   // Walk from this loop up through its parents removing all of the dead blocks.
1724   for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1725     for (auto *BB : DeadBlockSet)
1726       ParentL->getBlocksSet().erase(BB);
1727     llvm::erase_if(ParentL->getBlocksVector(),
1728                    [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1729   }
1730 
1731   // Now delete the dead child loops. This raw delete will clear them
1732   // recursively.
1733   llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {
1734     if (!DeadBlockSet.count(ChildL->getHeader()))
1735       return false;
1736 
1737     assert(llvm::all_of(ChildL->blocks(),
1738                         [&](BasicBlock *ChildBB) {
1739                           return DeadBlockSet.count(ChildBB);
1740                         }) &&
1741            "If the child loop header is dead all blocks in the child loop must "
1742            "be dead as well!");
1743     LoopUpdater.markLoopAsDeleted(*ChildL, ChildL->getName());
1744     if (SE)
1745       SE->forgetBlockAndLoopDispositions();
1746     LI.destroy(ChildL);
1747     return true;
1748   });
1749 
1750   // Remove the loop mappings for the dead blocks and drop all the references
1751   // from these blocks to others to handle cyclic references as we start
1752   // deleting the blocks themselves.
1753   for (auto *BB : DeadBlockSet) {
1754     // Check that the dominator tree has already been updated.
1755     assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1756     LI.changeLoopFor(BB, nullptr);
1757     // Drop all uses of the instructions to make sure we won't have dangling
1758     // uses in other blocks.
1759     for (auto &I : *BB)
1760       if (!I.use_empty())
1761         I.replaceAllUsesWith(PoisonValue::get(I.getType()));
1762     BB->dropAllReferences();
1763   }
1764 
1765   // Actually delete the blocks now that they've been fully unhooked from the
1766   // IR.
1767   for (auto *BB : DeadBlockSet)
1768     BB->eraseFromParent();
1769 }
1770 
1771 /// Recompute the set of blocks in a loop after unswitching.
1772 ///
1773 /// This walks from the original headers predecessors to rebuild the loop. We
1774 /// take advantage of the fact that new blocks can't have been added, and so we
1775 /// filter by the original loop's blocks. This also handles potentially
1776 /// unreachable code that we don't want to explore but might be found examining
1777 /// the predecessors of the header.
1778 ///
1779 /// If the original loop is no longer a loop, this will return an empty set. If
1780 /// it remains a loop, all the blocks within it will be added to the set
1781 /// (including those blocks in inner loops).
1782 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1783                                                                  LoopInfo &LI) {
1784   SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1785 
1786   auto *PH = L.getLoopPreheader();
1787   auto *Header = L.getHeader();
1788 
1789   // A worklist to use while walking backwards from the header.
1790   SmallVector<BasicBlock *, 16> Worklist;
1791 
1792   // First walk the predecessors of the header to find the backedges. This will
1793   // form the basis of our walk.
1794   for (auto *Pred : predecessors(Header)) {
1795     // Skip the preheader.
1796     if (Pred == PH)
1797       continue;
1798 
1799     // Because the loop was in simplified form, the only non-loop predecessor
1800     // is the preheader.
1801     assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1802                                "than the preheader that is not part of the "
1803                                "loop!");
1804 
1805     // Insert this block into the loop set and on the first visit and, if it
1806     // isn't the header we're currently walking, put it into the worklist to
1807     // recurse through.
1808     if (LoopBlockSet.insert(Pred).second && Pred != Header)
1809       Worklist.push_back(Pred);
1810   }
1811 
1812   // If no backedges were found, we're done.
1813   if (LoopBlockSet.empty())
1814     return LoopBlockSet;
1815 
1816   // We found backedges, recurse through them to identify the loop blocks.
1817   while (!Worklist.empty()) {
1818     BasicBlock *BB = Worklist.pop_back_val();
1819     assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1820 
1821     // No need to walk past the header.
1822     if (BB == Header)
1823       continue;
1824 
1825     // Because we know the inner loop structure remains valid we can use the
1826     // loop structure to jump immediately across the entire nested loop.
1827     // Further, because it is in loop simplified form, we can directly jump
1828     // to its preheader afterward.
1829     if (Loop *InnerL = LI.getLoopFor(BB))
1830       if (InnerL != &L) {
1831         assert(L.contains(InnerL) &&
1832                "Should not reach a loop *outside* this loop!");
1833         // The preheader is the only possible predecessor of the loop so
1834         // insert it into the set and check whether it was already handled.
1835         auto *InnerPH = InnerL->getLoopPreheader();
1836         assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1837                                       "but not contain the inner loop "
1838                                       "preheader!");
1839         if (!LoopBlockSet.insert(InnerPH).second)
1840           // The only way to reach the preheader is through the loop body
1841           // itself so if it has been visited the loop is already handled.
1842           continue;
1843 
1844         // Insert all of the blocks (other than those already present) into
1845         // the loop set. We expect at least the block that led us to find the
1846         // inner loop to be in the block set, but we may also have other loop
1847         // blocks if they were already enqueued as predecessors of some other
1848         // outer loop block.
1849         for (auto *InnerBB : InnerL->blocks()) {
1850           if (InnerBB == BB) {
1851             assert(LoopBlockSet.count(InnerBB) &&
1852                    "Block should already be in the set!");
1853             continue;
1854           }
1855 
1856           LoopBlockSet.insert(InnerBB);
1857         }
1858 
1859         // Add the preheader to the worklist so we will continue past the
1860         // loop body.
1861         Worklist.push_back(InnerPH);
1862         continue;
1863       }
1864 
1865     // Insert any predecessors that were in the original loop into the new
1866     // set, and if the insert is successful, add them to the worklist.
1867     for (auto *Pred : predecessors(BB))
1868       if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)
1869         Worklist.push_back(Pred);
1870   }
1871 
1872   assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1873 
1874   // We've found all the blocks participating in the loop, return our completed
1875   // set.
1876   return LoopBlockSet;
1877 }
1878 
1879 /// Rebuild a loop after unswitching removes some subset of blocks and edges.
1880 ///
1881 /// The removal may have removed some child loops entirely but cannot have
1882 /// disturbed any remaining child loops. However, they may need to be hoisted
1883 /// to the parent loop (or to be top-level loops). The original loop may be
1884 /// completely removed.
1885 ///
1886 /// The sibling loops resulting from this update are returned. If the original
1887 /// loop remains a valid loop, it will be the first entry in this list with all
1888 /// of the newly sibling loops following it.
1889 ///
1890 /// Returns true if the loop remains a loop after unswitching, and false if it
1891 /// is no longer a loop after unswitching (and should not continue to be
1892 /// referenced).
1893 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1894                                      LoopInfo &LI,
1895                                      SmallVectorImpl<Loop *> &HoistedLoops,
1896                                      ScalarEvolution *SE) {
1897   auto *PH = L.getLoopPreheader();
1898 
1899   // Compute the actual parent loop from the exit blocks. Because we may have
1900   // pruned some exits the loop may be different from the original parent.
1901   Loop *ParentL = nullptr;
1902   SmallVector<Loop *, 4> ExitLoops;
1903   SmallVector<BasicBlock *, 4> ExitsInLoops;
1904   ExitsInLoops.reserve(ExitBlocks.size());
1905   for (auto *ExitBB : ExitBlocks)
1906     if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1907       ExitLoops.push_back(ExitL);
1908       ExitsInLoops.push_back(ExitBB);
1909       if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1910         ParentL = ExitL;
1911     }
1912 
1913   // Recompute the blocks participating in this loop. This may be empty if it
1914   // is no longer a loop.
1915   auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1916 
1917   // If we still have a loop, we need to re-set the loop's parent as the exit
1918   // block set changing may have moved it within the loop nest. Note that this
1919   // can only happen when this loop has a parent as it can only hoist the loop
1920   // *up* the nest.
1921   if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1922     // Remove this loop's (original) blocks from all of the intervening loops.
1923     for (Loop *IL = L.getParentLoop(); IL != ParentL;
1924          IL = IL->getParentLoop()) {
1925       IL->getBlocksSet().erase(PH);
1926       for (auto *BB : L.blocks())
1927         IL->getBlocksSet().erase(BB);
1928       llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {
1929         return BB == PH || L.contains(BB);
1930       });
1931     }
1932 
1933     LI.changeLoopFor(PH, ParentL);
1934     L.getParentLoop()->removeChildLoop(&L);
1935     if (ParentL)
1936       ParentL->addChildLoop(&L);
1937     else
1938       LI.addTopLevelLoop(&L);
1939   }
1940 
1941   // Now we update all the blocks which are no longer within the loop.
1942   auto &Blocks = L.getBlocksVector();
1943   auto BlocksSplitI =
1944       LoopBlockSet.empty()
1945           ? Blocks.begin()
1946           : std::stable_partition(
1947                 Blocks.begin(), Blocks.end(),
1948                 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });
1949 
1950   // Before we erase the list of unlooped blocks, build a set of them.
1951   SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1952   if (LoopBlockSet.empty())
1953     UnloopedBlocks.insert(PH);
1954 
1955   // Now erase these blocks from the loop.
1956   for (auto *BB : make_range(BlocksSplitI, Blocks.end()))
1957     L.getBlocksSet().erase(BB);
1958   Blocks.erase(BlocksSplitI, Blocks.end());
1959 
1960   // Sort the exits in ascending loop depth, we'll work backwards across these
1961   // to process them inside out.
1962   llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1963     return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);
1964   });
1965 
1966   // We'll build up a set for each exit loop.
1967   SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1968   Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1969 
1970   auto RemoveUnloopedBlocksFromLoop =
1971       [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1972         for (auto *BB : UnloopedBlocks)
1973           L.getBlocksSet().erase(BB);
1974         llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {
1975           return UnloopedBlocks.count(BB);
1976         });
1977       };
1978 
1979   SmallVector<BasicBlock *, 16> Worklist;
1980   while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1981     assert(Worklist.empty() && "Didn't clear worklist!");
1982     assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
1983 
1984     // Grab the next exit block, in decreasing loop depth order.
1985     BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1986     Loop &ExitL = *LI.getLoopFor(ExitBB);
1987     assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
1988 
1989     // Erase all of the unlooped blocks from the loops between the previous
1990     // exit loop and this exit loop. This works because the ExitInLoops list is
1991     // sorted in increasing order of loop depth and thus we visit loops in
1992     // decreasing order of loop depth.
1993     for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
1994       RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1995 
1996     // Walk the CFG back until we hit the cloned PH adding everything reachable
1997     // and in the unlooped set to this exit block's loop.
1998     Worklist.push_back(ExitBB);
1999     do {
2000       BasicBlock *BB = Worklist.pop_back_val();
2001       // We can stop recursing at the cloned preheader (if we get there).
2002       if (BB == PH)
2003         continue;
2004 
2005       for (BasicBlock *PredBB : predecessors(BB)) {
2006         // If this pred has already been moved to our set or is part of some
2007         // (inner) loop, no update needed.
2008         if (!UnloopedBlocks.erase(PredBB)) {
2009           assert((NewExitLoopBlocks.count(PredBB) ||
2010                   ExitL.contains(LI.getLoopFor(PredBB))) &&
2011                  "Predecessor not in a nested loop (or already visited)!");
2012           continue;
2013         }
2014 
2015         // We just insert into the loop set here. We'll add these blocks to the
2016         // exit loop after we build up the set in a deterministic order rather
2017         // than the predecessor-influenced visit order.
2018         bool Inserted = NewExitLoopBlocks.insert(PredBB).second;
2019         (void)Inserted;
2020         assert(Inserted && "Should only visit an unlooped block once!");
2021 
2022         // And recurse through to its predecessors.
2023         Worklist.push_back(PredBB);
2024       }
2025     } while (!Worklist.empty());
2026 
2027     // If blocks in this exit loop were directly part of the original loop (as
2028     // opposed to a child loop) update the map to point to this exit loop. This
2029     // just updates a map and so the fact that the order is unstable is fine.
2030     for (auto *BB : NewExitLoopBlocks)
2031       if (Loop *BBL = LI.getLoopFor(BB))
2032         if (BBL == &L || !L.contains(BBL))
2033           LI.changeLoopFor(BB, &ExitL);
2034 
2035     // We will remove the remaining unlooped blocks from this loop in the next
2036     // iteration or below.
2037     NewExitLoopBlocks.clear();
2038   }
2039 
2040   // Any remaining unlooped blocks are no longer part of any loop unless they
2041   // are part of some child loop.
2042   for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
2043     RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
2044   for (auto *BB : UnloopedBlocks)
2045     if (Loop *BBL = LI.getLoopFor(BB))
2046       if (BBL == &L || !L.contains(BBL))
2047         LI.changeLoopFor(BB, nullptr);
2048 
2049   // Sink all the child loops whose headers are no longer in the loop set to
2050   // the parent (or to be top level loops). We reach into the loop and directly
2051   // update its subloop vector to make this batch update efficient.
2052   auto &SubLoops = L.getSubLoopsVector();
2053   auto SubLoopsSplitI =
2054       LoopBlockSet.empty()
2055           ? SubLoops.begin()
2056           : std::stable_partition(
2057                 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {
2058                   return LoopBlockSet.count(SubL->getHeader());
2059                 });
2060   for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {
2061     HoistedLoops.push_back(HoistedL);
2062     HoistedL->setParentLoop(nullptr);
2063 
2064     // To compute the new parent of this hoisted loop we look at where we
2065     // placed the preheader above. We can't lookup the header itself because we
2066     // retained the mapping from the header to the hoisted loop. But the
2067     // preheader and header should have the exact same new parent computed
2068     // based on the set of exit blocks from the original loop as the preheader
2069     // is a predecessor of the header and so reached in the reverse walk. And
2070     // because the loops were all in simplified form the preheader of the
2071     // hoisted loop can't be part of some *other* loop.
2072     if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))
2073       NewParentL->addChildLoop(HoistedL);
2074     else
2075       LI.addTopLevelLoop(HoistedL);
2076   }
2077   SubLoops.erase(SubLoopsSplitI, SubLoops.end());
2078 
2079   // Actually delete the loop if nothing remained within it.
2080   if (Blocks.empty()) {
2081     assert(SubLoops.empty() &&
2082            "Failed to remove all subloops from the original loop!");
2083     if (Loop *ParentL = L.getParentLoop())
2084       ParentL->removeChildLoop(llvm::find(*ParentL, &L));
2085     else
2086       LI.removeLoop(llvm::find(LI, &L));
2087     // markLoopAsDeleted for L should be triggered by the caller (it is
2088     // typically done within postUnswitch).
2089     if (SE)
2090       SE->forgetBlockAndLoopDispositions();
2091     LI.destroy(&L);
2092     return false;
2093   }
2094 
2095   return true;
2096 }
2097 
2098 /// Helper to visit a dominator subtree, invoking a callable on each node.
2099 ///
2100 /// Returning false at any point will stop walking past that node of the tree.
2101 template <typename CallableT>
2102 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
2103   SmallVector<DomTreeNode *, 4> DomWorklist;
2104   DomWorklist.push_back(DT[BB]);
2105 #ifndef NDEBUG
2106   SmallPtrSet<DomTreeNode *, 4> Visited;
2107   Visited.insert(DT[BB]);
2108 #endif
2109   do {
2110     DomTreeNode *N = DomWorklist.pop_back_val();
2111 
2112     // Visit this node.
2113     if (!Callable(N->getBlock()))
2114       continue;
2115 
2116     // Accumulate the child nodes.
2117     for (DomTreeNode *ChildN : *N) {
2118       assert(Visited.insert(ChildN).second &&
2119              "Cannot visit a node twice when walking a tree!");
2120       DomWorklist.push_back(ChildN);
2121     }
2122   } while (!DomWorklist.empty());
2123 }
2124 
2125 void postUnswitch(Loop &L, LPMUpdater &U, StringRef LoopName,
2126                   bool CurrentLoopValid, bool PartiallyInvariant,
2127                   bool InjectedCondition, ArrayRef<Loop *> NewLoops) {
2128   // If we did a non-trivial unswitch, we have added new (cloned) loops.
2129   if (!NewLoops.empty())
2130     U.addSiblingLoops(NewLoops);
2131 
2132   // If the current loop remains valid, we should revisit it to catch any
2133   // other unswitch opportunities. Otherwise, we need to mark it as deleted.
2134   if (CurrentLoopValid) {
2135     if (PartiallyInvariant) {
2136       // Mark the new loop as partially unswitched, to avoid unswitching on
2137       // the same condition again.
2138       auto &Context = L.getHeader()->getContext();
2139       MDNode *DisableUnswitchMD = MDNode::get(
2140           Context,
2141           MDString::get(Context, "llvm.loop.unswitch.partial.disable"));
2142       MDNode *NewLoopID = makePostTransformationMetadata(
2143           Context, L.getLoopID(), {"llvm.loop.unswitch.partial"},
2144           {DisableUnswitchMD});
2145       L.setLoopID(NewLoopID);
2146     } else if (InjectedCondition) {
2147       // Do the same for injection of invariant conditions.
2148       auto &Context = L.getHeader()->getContext();
2149       MDNode *DisableUnswitchMD = MDNode::get(
2150           Context,
2151           MDString::get(Context, "llvm.loop.unswitch.injection.disable"));
2152       MDNode *NewLoopID = makePostTransformationMetadata(
2153           Context, L.getLoopID(), {"llvm.loop.unswitch.injection"},
2154           {DisableUnswitchMD});
2155       L.setLoopID(NewLoopID);
2156     } else
2157       U.revisitCurrentLoop();
2158   } else
2159     U.markLoopAsDeleted(L, LoopName);
2160 }
2161 
2162 static void unswitchNontrivialInvariants(
2163     Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
2164     IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI,
2165     AssumptionCache &AC, ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
2166     LPMUpdater &LoopUpdater, bool InsertFreeze, bool InjectedCondition) {
2167   auto *ParentBB = TI.getParent();
2168   BranchInst *BI = dyn_cast<BranchInst>(&TI);
2169   SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
2170 
2171   // Save the current loop name in a variable so that we can report it even
2172   // after it has been deleted.
2173   std::string LoopName(L.getName());
2174 
2175   // We can only unswitch switches, conditional branches with an invariant
2176   // condition, or combining invariant conditions with an instruction or
2177   // partially invariant instructions.
2178   assert((SI || (BI && BI->isConditional())) &&
2179          "Can only unswitch switches and conditional branch!");
2180   bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty();
2181   bool FullUnswitch =
2182       SI || (skipTrivialSelect(BI->getCondition()) == Invariants[0] &&
2183              !PartiallyInvariant);
2184   if (FullUnswitch)
2185     assert(Invariants.size() == 1 &&
2186            "Cannot have other invariants with full unswitching!");
2187   else
2188     assert(isa<Instruction>(skipTrivialSelect(BI->getCondition())) &&
2189            "Partial unswitching requires an instruction as the condition!");
2190 
2191   if (MSSAU && VerifyMemorySSA)
2192     MSSAU->getMemorySSA()->verifyMemorySSA();
2193 
2194   // Constant and BBs tracking the cloned and continuing successor. When we are
2195   // unswitching the entire condition, this can just be trivially chosen to
2196   // unswitch towards `true`. However, when we are unswitching a set of
2197   // invariants combined with `and` or `or` or partially invariant instructions,
2198   // the combining operation determines the best direction to unswitch: we want
2199   // to unswitch the direction that will collapse the branch.
2200   bool Direction = true;
2201   int ClonedSucc = 0;
2202   if (!FullUnswitch) {
2203     Value *Cond = skipTrivialSelect(BI->getCondition());
2204     (void)Cond;
2205     assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) ||
2206             PartiallyInvariant) &&
2207            "Only `or`, `and`, an `select`, partially invariant instructions "
2208            "can combine invariants being unswitched.");
2209     if (!match(Cond, m_LogicalOr())) {
2210       if (match(Cond, m_LogicalAnd()) ||
2211           (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) {
2212         Direction = false;
2213         ClonedSucc = 1;
2214       }
2215     }
2216   }
2217 
2218   BasicBlock *RetainedSuccBB =
2219       BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();
2220   SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
2221   if (BI)
2222     UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));
2223   else
2224     for (auto Case : SI->cases())
2225       if (Case.getCaseSuccessor() != RetainedSuccBB)
2226         UnswitchedSuccBBs.insert(Case.getCaseSuccessor());
2227 
2228   assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
2229          "Should not unswitch the same successor we are retaining!");
2230 
2231   // The branch should be in this exact loop. Any inner loop's invariant branch
2232   // should be handled by unswitching that inner loop. The caller of this
2233   // routine should filter out any candidates that remain (but were skipped for
2234   // whatever reason).
2235   assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
2236 
2237   // Compute the parent loop now before we start hacking on things.
2238   Loop *ParentL = L.getParentLoop();
2239   // Get blocks in RPO order for MSSA update, before changing the CFG.
2240   LoopBlocksRPO LBRPO(&L);
2241   if (MSSAU)
2242     LBRPO.perform(&LI);
2243 
2244   // Compute the outer-most loop containing one of our exit blocks. This is the
2245   // furthest up our loopnest which can be mutated, which we will use below to
2246   // update things.
2247   Loop *OuterExitL = &L;
2248   SmallVector<BasicBlock *, 4> ExitBlocks;
2249   L.getUniqueExitBlocks(ExitBlocks);
2250   for (auto *ExitBB : ExitBlocks) {
2251     // ExitBB can be an exit block for several levels in the loop nest. Make
2252     // sure we find the top most.
2253     Loop *NewOuterExitL = getTopMostExitingLoop(ExitBB, LI);
2254     if (!NewOuterExitL) {
2255       // We exited the entire nest with this block, so we're done.
2256       OuterExitL = nullptr;
2257       break;
2258     }
2259     if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))
2260       OuterExitL = NewOuterExitL;
2261   }
2262 
2263   // At this point, we're definitely going to unswitch something so invalidate
2264   // any cached information in ScalarEvolution for the outer most loop
2265   // containing an exit block and all nested loops.
2266   if (SE) {
2267     if (OuterExitL)
2268       SE->forgetLoop(OuterExitL);
2269     else
2270       SE->forgetTopmostLoop(&L);
2271     SE->forgetBlockAndLoopDispositions();
2272   }
2273 
2274   // If the edge from this terminator to a successor dominates that successor,
2275   // store a map from each block in its dominator subtree to it. This lets us
2276   // tell when cloning for a particular successor if a block is dominated by
2277   // some *other* successor with a single data structure. We use this to
2278   // significantly reduce cloning.
2279   SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
2280   for (auto *SuccBB : llvm::concat<BasicBlock *const>(ArrayRef(RetainedSuccBB),
2281                                                       UnswitchedSuccBBs))
2282     if (SuccBB->getUniquePredecessor() ||
2283         llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2284           return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);
2285         }))
2286       visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {
2287         DominatingSucc[BB] = SuccBB;
2288         return true;
2289       });
2290 
2291   // Split the preheader, so that we know that there is a safe place to insert
2292   // the conditional branch. We will change the preheader to have a conditional
2293   // branch on LoopCond. The original preheader will become the split point
2294   // between the unswitched versions, and we will have a new preheader for the
2295   // original loop.
2296   BasicBlock *SplitBB = L.getLoopPreheader();
2297   BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU);
2298 
2299   // Keep track of the dominator tree updates needed.
2300   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2301 
2302   // Clone the loop for each unswitched successor.
2303   SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
2304   VMaps.reserve(UnswitchedSuccBBs.size());
2305   SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
2306   for (auto *SuccBB : UnswitchedSuccBBs) {
2307     VMaps.emplace_back(new ValueToValueMapTy());
2308     ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2309         L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,
2310         DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU, SE);
2311   }
2312 
2313   // Drop metadata if we may break its semantics by moving this instr into the
2314   // split block.
2315   if (TI.getMetadata(LLVMContext::MD_make_implicit)) {
2316     if (DropNonTrivialImplicitNullChecks)
2317       // Do not spend time trying to understand if we can keep it, just drop it
2318       // to save compile time.
2319       TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2320     else {
2321       // It is only legal to preserve make.implicit metadata if we are
2322       // guaranteed no reach implicit null check after following this branch.
2323       ICFLoopSafetyInfo SafetyInfo;
2324       SafetyInfo.computeLoopSafetyInfo(&L);
2325       if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))
2326         TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2327     }
2328   }
2329 
2330   // The stitching of the branched code back together depends on whether we're
2331   // doing full unswitching or not with the exception that we always want to
2332   // nuke the initial terminator placed in the split block.
2333   SplitBB->getTerminator()->eraseFromParent();
2334   if (FullUnswitch) {
2335     // Splice the terminator from the original loop and rewrite its
2336     // successors.
2337     TI.moveBefore(*SplitBB, SplitBB->end());
2338 
2339     // Keep a clone of the terminator for MSSA updates.
2340     Instruction *NewTI = TI.clone();
2341     NewTI->insertInto(ParentBB, ParentBB->end());
2342 
2343     // First wire up the moved terminator to the preheaders.
2344     if (BI) {
2345       BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2346       BI->setSuccessor(ClonedSucc, ClonedPH);
2347       BI->setSuccessor(1 - ClonedSucc, LoopPH);
2348       Value *Cond = skipTrivialSelect(BI->getCondition());
2349       if (InsertFreeze)
2350         Cond = new FreezeInst(
2351             Cond, Cond->getName() + ".fr", BI);
2352       BI->setCondition(Cond);
2353       DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2354     } else {
2355       assert(SI && "Must either be a branch or switch!");
2356 
2357       // Walk the cases and directly update their successors.
2358       assert(SI->getDefaultDest() == RetainedSuccBB &&
2359              "Not retaining default successor!");
2360       SI->setDefaultDest(LoopPH);
2361       for (const auto &Case : SI->cases())
2362         if (Case.getCaseSuccessor() == RetainedSuccBB)
2363           Case.setSuccessor(LoopPH);
2364         else
2365           Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);
2366 
2367       if (InsertFreeze)
2368         SI->setCondition(new FreezeInst(
2369             SI->getCondition(), SI->getCondition()->getName() + ".fr", SI));
2370 
2371       // We need to use the set to populate domtree updates as even when there
2372       // are multiple cases pointing at the same successor we only want to
2373       // remove and insert one edge in the domtree.
2374       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2375         DTUpdates.push_back(
2376             {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});
2377     }
2378 
2379     if (MSSAU) {
2380       DT.applyUpdates(DTUpdates);
2381       DTUpdates.clear();
2382 
2383       // Remove all but one edge to the retained block and all unswitched
2384       // blocks. This is to avoid having duplicate entries in the cloned Phis,
2385       // when we know we only keep a single edge for each case.
2386       MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB);
2387       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2388         MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB);
2389 
2390       for (auto &VMap : VMaps)
2391         MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2392                                    /*IgnoreIncomingWithNoClones=*/true);
2393       MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2394 
2395       // Remove all edges to unswitched blocks.
2396       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2397         MSSAU->removeEdge(ParentBB, SuccBB);
2398     }
2399 
2400     // Now unhook the successor relationship as we'll be replacing
2401     // the terminator with a direct branch. This is much simpler for branches
2402     // than switches so we handle those first.
2403     if (BI) {
2404       // Remove the parent as a predecessor of the unswitched successor.
2405       assert(UnswitchedSuccBBs.size() == 1 &&
2406              "Only one possible unswitched block for a branch!");
2407       BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2408       UnswitchedSuccBB->removePredecessor(ParentBB,
2409                                           /*KeepOneInputPHIs*/ true);
2410       DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2411     } else {
2412       // Note that we actually want to remove the parent block as a predecessor
2413       // of *every* case successor. The case successor is either unswitched,
2414       // completely eliminating an edge from the parent to that successor, or it
2415       // is a duplicate edge to the retained successor as the retained successor
2416       // is always the default successor and as we'll replace this with a direct
2417       // branch we no longer need the duplicate entries in the PHI nodes.
2418       SwitchInst *NewSI = cast<SwitchInst>(NewTI);
2419       assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2420              "Not retaining default successor!");
2421       for (const auto &Case : NewSI->cases())
2422         Case.getCaseSuccessor()->removePredecessor(
2423             ParentBB,
2424             /*KeepOneInputPHIs*/ true);
2425 
2426       // We need to use the set to populate domtree updates as even when there
2427       // are multiple cases pointing at the same successor we only want to
2428       // remove and insert one edge in the domtree.
2429       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2430         DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});
2431     }
2432 
2433     // After MSSAU update, remove the cloned terminator instruction NewTI.
2434     ParentBB->getTerminator()->eraseFromParent();
2435 
2436     // Create a new unconditional branch to the continuing block (as opposed to
2437     // the one cloned).
2438     BranchInst::Create(RetainedSuccBB, ParentBB);
2439   } else {
2440     assert(BI && "Only branches have partial unswitching.");
2441     assert(UnswitchedSuccBBs.size() == 1 &&
2442            "Only one possible unswitched block for a branch!");
2443     BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2444     // When doing a partial unswitch, we have to do a bit more work to build up
2445     // the branch in the split block.
2446     if (PartiallyInvariant)
2447       buildPartialInvariantUnswitchConditionalBranch(
2448           *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, L, MSSAU);
2449     else {
2450       buildPartialUnswitchConditionalBranch(
2451           *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH,
2452           FreezeLoopUnswitchCond, BI, &AC, DT);
2453     }
2454     DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2455 
2456     if (MSSAU) {
2457       DT.applyUpdates(DTUpdates);
2458       DTUpdates.clear();
2459 
2460       // Perform MSSA cloning updates.
2461       for (auto &VMap : VMaps)
2462         MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2463                                    /*IgnoreIncomingWithNoClones=*/true);
2464       MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2465     }
2466   }
2467 
2468   // Apply the updates accumulated above to get an up-to-date dominator tree.
2469   DT.applyUpdates(DTUpdates);
2470 
2471   // Now that we have an accurate dominator tree, first delete the dead cloned
2472   // blocks so that we can accurately build any cloned loops. It is important to
2473   // not delete the blocks from the original loop yet because we still want to
2474   // reference the original loop to understand the cloned loop's structure.
2475   deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2476 
2477   // Build the cloned loop structure itself. This may be substantially
2478   // different from the original structure due to the simplified CFG. This also
2479   // handles inserting all the cloned blocks into the correct loops.
2480   SmallVector<Loop *, 4> NonChildClonedLoops;
2481   for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2482     buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);
2483 
2484   // Now that our cloned loops have been built, we can update the original loop.
2485   // First we delete the dead blocks from it and then we rebuild the loop
2486   // structure taking these deletions into account.
2487   deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, SE, LoopUpdater);
2488 
2489   if (MSSAU && VerifyMemorySSA)
2490     MSSAU->getMemorySSA()->verifyMemorySSA();
2491 
2492   SmallVector<Loop *, 4> HoistedLoops;
2493   bool IsStillLoop =
2494       rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops, SE);
2495 
2496   if (MSSAU && VerifyMemorySSA)
2497     MSSAU->getMemorySSA()->verifyMemorySSA();
2498 
2499   // This transformation has a high risk of corrupting the dominator tree, and
2500   // the below steps to rebuild loop structures will result in hard to debug
2501   // errors in that case so verify that the dominator tree is sane first.
2502   // FIXME: Remove this when the bugs stop showing up and rely on existing
2503   // verification steps.
2504   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2505 
2506   if (BI && !PartiallyInvariant) {
2507     // If we unswitched a branch which collapses the condition to a known
2508     // constant we want to replace all the uses of the invariants within both
2509     // the original and cloned blocks. We do this here so that we can use the
2510     // now updated dominator tree to identify which side the users are on.
2511     assert(UnswitchedSuccBBs.size() == 1 &&
2512            "Only one possible unswitched block for a branch!");
2513     BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2514 
2515     // When considering multiple partially-unswitched invariants
2516     // we cant just go replace them with constants in both branches.
2517     //
2518     // For 'AND' we infer that true branch ("continue") means true
2519     // for each invariant operand.
2520     // For 'OR' we can infer that false branch ("continue") means false
2521     // for each invariant operand.
2522     // So it happens that for multiple-partial case we dont replace
2523     // in the unswitched branch.
2524     bool ReplaceUnswitched =
2525         FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant;
2526 
2527     ConstantInt *UnswitchedReplacement =
2528         Direction ? ConstantInt::getTrue(BI->getContext())
2529                   : ConstantInt::getFalse(BI->getContext());
2530     ConstantInt *ContinueReplacement =
2531         Direction ? ConstantInt::getFalse(BI->getContext())
2532                   : ConstantInt::getTrue(BI->getContext());
2533     for (Value *Invariant : Invariants) {
2534       assert(!isa<Constant>(Invariant) &&
2535              "Should not be replacing constant values!");
2536       // Use make_early_inc_range here as set invalidates the iterator.
2537       for (Use &U : llvm::make_early_inc_range(Invariant->uses())) {
2538         Instruction *UserI = dyn_cast<Instruction>(U.getUser());
2539         if (!UserI)
2540           continue;
2541 
2542         // Replace it with the 'continue' side if in the main loop body, and the
2543         // unswitched if in the cloned blocks.
2544         if (DT.dominates(LoopPH, UserI->getParent()))
2545           U.set(ContinueReplacement);
2546         else if (ReplaceUnswitched &&
2547                  DT.dominates(ClonedPH, UserI->getParent()))
2548           U.set(UnswitchedReplacement);
2549       }
2550     }
2551   }
2552 
2553   // We can change which blocks are exit blocks of all the cloned sibling
2554   // loops, the current loop, and any parent loops which shared exit blocks
2555   // with the current loop. As a consequence, we need to re-form LCSSA for
2556   // them. But we shouldn't need to re-form LCSSA for any child loops.
2557   // FIXME: This could be made more efficient by tracking which exit blocks are
2558   // new, and focusing on them, but that isn't likely to be necessary.
2559   //
2560   // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2561   // loop nest and update every loop that could have had its exits changed. We
2562   // also need to cover any intervening loops. We add all of these loops to
2563   // a list and sort them by loop depth to achieve this without updating
2564   // unnecessary loops.
2565   auto UpdateLoop = [&](Loop &UpdateL) {
2566 #ifndef NDEBUG
2567     UpdateL.verifyLoop();
2568     for (Loop *ChildL : UpdateL) {
2569       ChildL->verifyLoop();
2570       assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2571              "Perturbed a child loop's LCSSA form!");
2572     }
2573 #endif
2574     // First build LCSSA for this loop so that we can preserve it when
2575     // forming dedicated exits. We don't want to perturb some other loop's
2576     // LCSSA while doing that CFG edit.
2577     formLCSSA(UpdateL, DT, &LI, SE);
2578 
2579     // For loops reached by this loop's original exit blocks we may
2580     // introduced new, non-dedicated exits. At least try to re-form dedicated
2581     // exits for these loops. This may fail if they couldn't have dedicated
2582     // exits to start with.
2583     formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true);
2584   };
2585 
2586   // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2587   // and we can do it in any order as they don't nest relative to each other.
2588   //
2589   // Also check if any of the loops we have updated have become top-level loops
2590   // as that will necessitate widening the outer loop scope.
2591   for (Loop *UpdatedL :
2592        llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {
2593     UpdateLoop(*UpdatedL);
2594     if (UpdatedL->isOutermost())
2595       OuterExitL = nullptr;
2596   }
2597   if (IsStillLoop) {
2598     UpdateLoop(L);
2599     if (L.isOutermost())
2600       OuterExitL = nullptr;
2601   }
2602 
2603   // If the original loop had exit blocks, walk up through the outer most loop
2604   // of those exit blocks to update LCSSA and form updated dedicated exits.
2605   if (OuterExitL != &L)
2606     for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2607          OuterL = OuterL->getParentLoop())
2608       UpdateLoop(*OuterL);
2609 
2610 #ifndef NDEBUG
2611   // Verify the entire loop structure to catch any incorrect updates before we
2612   // progress in the pass pipeline.
2613   LI.verify(DT);
2614 #endif
2615 
2616   // Now that we've unswitched something, make callbacks to report the changes.
2617   // For that we need to merge together the updated loops and the cloned loops
2618   // and check whether the original loop survived.
2619   SmallVector<Loop *, 4> SibLoops;
2620   for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
2621     if (UpdatedL->getParentLoop() == ParentL)
2622       SibLoops.push_back(UpdatedL);
2623   postUnswitch(L, LoopUpdater, LoopName, IsStillLoop, PartiallyInvariant,
2624                InjectedCondition, SibLoops);
2625 
2626   if (MSSAU && VerifyMemorySSA)
2627     MSSAU->getMemorySSA()->verifyMemorySSA();
2628 
2629   if (BI)
2630     ++NumBranches;
2631   else
2632     ++NumSwitches;
2633 }
2634 
2635 /// Recursively compute the cost of a dominator subtree based on the per-block
2636 /// cost map provided.
2637 ///
2638 /// The recursive computation is memozied into the provided DT-indexed cost map
2639 /// to allow querying it for most nodes in the domtree without it becoming
2640 /// quadratic.
2641 static InstructionCost computeDomSubtreeCost(
2642     DomTreeNode &N,
2643     const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap,
2644     SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) {
2645   // Don't accumulate cost (or recurse through) blocks not in our block cost
2646   // map and thus not part of the duplication cost being considered.
2647   auto BBCostIt = BBCostMap.find(N.getBlock());
2648   if (BBCostIt == BBCostMap.end())
2649     return 0;
2650 
2651   // Lookup this node to see if we already computed its cost.
2652   auto DTCostIt = DTCostMap.find(&N);
2653   if (DTCostIt != DTCostMap.end())
2654     return DTCostIt->second;
2655 
2656   // If not, we have to compute it. We can't use insert above and update
2657   // because computing the cost may insert more things into the map.
2658   InstructionCost Cost = std::accumulate(
2659       N.begin(), N.end(), BBCostIt->second,
2660       [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost {
2661         return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
2662       });
2663   bool Inserted = DTCostMap.insert({&N, Cost}).second;
2664   (void)Inserted;
2665   assert(Inserted && "Should not insert a node while visiting children!");
2666   return Cost;
2667 }
2668 
2669 /// Turns a select instruction into implicit control flow branch,
2670 /// making the following replacement:
2671 ///
2672 /// head:
2673 ///   --code before select--
2674 ///   select %cond, %trueval, %falseval
2675 ///   --code after select--
2676 ///
2677 /// into
2678 ///
2679 /// head:
2680 ///   --code before select--
2681 ///   br i1 %cond, label %then, label %tail
2682 ///
2683 /// then:
2684 ///   br %tail
2685 ///
2686 /// tail:
2687 ///   phi [ %trueval, %then ], [ %falseval, %head]
2688 ///   unreachable
2689 ///
2690 /// It also makes all relevant DT and LI updates, so that all structures are in
2691 /// valid state after this transform.
2692 static BranchInst *turnSelectIntoBranch(SelectInst *SI, DominatorTree &DT,
2693                                         LoopInfo &LI, MemorySSAUpdater *MSSAU,
2694                                         AssumptionCache *AC) {
2695   LLVM_DEBUG(dbgs() << "Turning " << *SI << " into a branch.\n");
2696   BasicBlock *HeadBB = SI->getParent();
2697 
2698   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2699   SplitBlockAndInsertIfThen(SI->getCondition(), SI, false,
2700                             SI->getMetadata(LLVMContext::MD_prof), &DTU, &LI);
2701   auto *CondBr = cast<BranchInst>(HeadBB->getTerminator());
2702   BasicBlock *ThenBB = CondBr->getSuccessor(0),
2703              *TailBB = CondBr->getSuccessor(1);
2704   if (MSSAU)
2705     MSSAU->moveAllAfterSpliceBlocks(HeadBB, TailBB, SI);
2706 
2707   PHINode *Phi = PHINode::Create(SI->getType(), 2, "unswitched.select", SI);
2708   Phi->addIncoming(SI->getTrueValue(), ThenBB);
2709   Phi->addIncoming(SI->getFalseValue(), HeadBB);
2710   SI->replaceAllUsesWith(Phi);
2711   SI->eraseFromParent();
2712 
2713   if (MSSAU && VerifyMemorySSA)
2714     MSSAU->getMemorySSA()->verifyMemorySSA();
2715 
2716   ++NumSelects;
2717   return CondBr;
2718 }
2719 
2720 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2721 /// making the following replacement:
2722 ///
2723 ///   --code before guard--
2724 ///   call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2725 ///   --code after guard--
2726 ///
2727 /// into
2728 ///
2729 ///   --code before guard--
2730 ///   br i1 %cond, label %guarded, label %deopt
2731 ///
2732 /// guarded:
2733 ///   --code after guard--
2734 ///
2735 /// deopt:
2736 ///   call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2737 ///   unreachable
2738 ///
2739 /// It also makes all relevant DT and LI updates, so that all structures are in
2740 /// valid state after this transform.
2741 static BranchInst *turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2742                                        DominatorTree &DT, LoopInfo &LI,
2743                                        MemorySSAUpdater *MSSAU) {
2744   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2745   LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2746   BasicBlock *CheckBB = GI->getParent();
2747 
2748   if (MSSAU && VerifyMemorySSA)
2749      MSSAU->getMemorySSA()->verifyMemorySSA();
2750 
2751   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2752   Instruction *DeoptBlockTerm =
2753       SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true,
2754                                 GI->getMetadata(LLVMContext::MD_prof), &DTU, &LI);
2755   BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
2756   // SplitBlockAndInsertIfThen inserts control flow that branches to
2757   // DeoptBlockTerm if the condition is true.  We want the opposite.
2758   CheckBI->swapSuccessors();
2759 
2760   BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
2761   GuardedBlock->setName("guarded");
2762   CheckBI->getSuccessor(1)->setName("deopt");
2763   BasicBlock *DeoptBlock = CheckBI->getSuccessor(1);
2764 
2765   if (MSSAU)
2766     MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI);
2767 
2768   GI->moveBefore(DeoptBlockTerm);
2769   GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));
2770 
2771   if (MSSAU) {
2772     MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI));
2773     MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator);
2774     if (VerifyMemorySSA)
2775       MSSAU->getMemorySSA()->verifyMemorySSA();
2776   }
2777 
2778   if (VerifyLoopInfo)
2779     LI.verify(DT);
2780   ++NumGuards;
2781   return CheckBI;
2782 }
2783 
2784 /// Cost multiplier is a way to limit potentially exponential behavior
2785 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2786 /// candidates available. Also accounting for the number of "sibling" loops with
2787 /// the idea to account for previous unswitches that already happened on this
2788 /// cluster of loops. There was an attempt to keep this formula simple,
2789 /// just enough to limit the worst case behavior. Even if it is not that simple
2790 /// now it is still not an attempt to provide a detailed heuristic size
2791 /// prediction.
2792 ///
2793 /// TODO: Make a proper accounting of "explosion" effect for all kinds of
2794 /// unswitch candidates, making adequate predictions instead of wild guesses.
2795 /// That requires knowing not just the number of "remaining" candidates but
2796 /// also costs of unswitching for each of these candidates.
2797 static int CalculateUnswitchCostMultiplier(
2798     const Instruction &TI, const Loop &L, const LoopInfo &LI,
2799     const DominatorTree &DT,
2800     ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates) {
2801 
2802   // Guards and other exiting conditions do not contribute to exponential
2803   // explosion as soon as they dominate the latch (otherwise there might be
2804   // another path to the latch remaining that does not allow to eliminate the
2805   // loop copy on unswitch).
2806   const BasicBlock *Latch = L.getLoopLatch();
2807   const BasicBlock *CondBlock = TI.getParent();
2808   if (DT.dominates(CondBlock, Latch) &&
2809       (isGuard(&TI) ||
2810        (TI.isTerminator() &&
2811         llvm::count_if(successors(&TI), [&L](const BasicBlock *SuccBB) {
2812           return L.contains(SuccBB);
2813         }) <= 1))) {
2814     NumCostMultiplierSkipped++;
2815     return 1;
2816   }
2817 
2818   auto *ParentL = L.getParentLoop();
2819   int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2820                                : std::distance(LI.begin(), LI.end()));
2821   // Count amount of clones that all the candidates might cause during
2822   // unswitching. Branch/guard/select counts as 1, switch counts as log2 of its
2823   // cases.
2824   int UnswitchedClones = 0;
2825   for (const auto &Candidate : UnswitchCandidates) {
2826     const Instruction *CI = Candidate.TI;
2827     const BasicBlock *CondBlock = CI->getParent();
2828     bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch);
2829     if (isa<SelectInst>(CI)) {
2830       UnswitchedClones++;
2831       continue;
2832     }
2833     if (isGuard(CI)) {
2834       if (!SkipExitingSuccessors)
2835         UnswitchedClones++;
2836       continue;
2837     }
2838     int NonExitingSuccessors =
2839         llvm::count_if(successors(CondBlock),
2840                        [SkipExitingSuccessors, &L](const BasicBlock *SuccBB) {
2841           return !SkipExitingSuccessors || L.contains(SuccBB);
2842         });
2843     UnswitchedClones += Log2_32(NonExitingSuccessors);
2844   }
2845 
2846   // Ignore up to the "unscaled candidates" number of unswitch candidates
2847   // when calculating the power-of-two scaling of the cost. The main idea
2848   // with this control is to allow a small number of unswitches to happen
2849   // and rely more on siblings multiplier (see below) when the number
2850   // of candidates is small.
2851   unsigned ClonesPower =
2852       std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0);
2853 
2854   // Allowing top-level loops to spread a bit more than nested ones.
2855   int SiblingsMultiplier =
2856       std::max((ParentL ? SiblingsCount
2857                         : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2858                1);
2859   // Compute the cost multiplier in a way that won't overflow by saturating
2860   // at an upper bound.
2861   int CostMultiplier;
2862   if (ClonesPower > Log2_32(UnswitchThreshold) ||
2863       SiblingsMultiplier > UnswitchThreshold)
2864     CostMultiplier = UnswitchThreshold;
2865   else
2866     CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower),
2867                               (int)UnswitchThreshold);
2868 
2869   LLVM_DEBUG(dbgs() << "  Computed multiplier  " << CostMultiplier
2870                     << " (siblings " << SiblingsMultiplier << " * clones "
2871                     << (1 << ClonesPower) << ")"
2872                     << " for unswitch candidate: " << TI << "\n");
2873   return CostMultiplier;
2874 }
2875 
2876 static bool collectUnswitchCandidates(
2877     SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
2878     IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch,
2879     const Loop &L, const LoopInfo &LI, AAResults &AA,
2880     const MemorySSAUpdater *MSSAU) {
2881   assert(UnswitchCandidates.empty() && "Should be!");
2882 
2883   auto AddUnswitchCandidatesForInst = [&](Instruction *I, Value *Cond) {
2884     Cond = skipTrivialSelect(Cond);
2885     if (isa<Constant>(Cond))
2886       return;
2887     if (L.isLoopInvariant(Cond)) {
2888       UnswitchCandidates.push_back({I, {Cond}});
2889       return;
2890     }
2891     if (match(Cond, m_CombineOr(m_LogicalAnd(), m_LogicalOr()))) {
2892       TinyPtrVector<Value *> Invariants =
2893           collectHomogenousInstGraphLoopInvariants(
2894               L, *static_cast<Instruction *>(Cond), LI);
2895       if (!Invariants.empty())
2896         UnswitchCandidates.push_back({I, std::move(Invariants)});
2897     }
2898   };
2899 
2900   // Whether or not we should also collect guards in the loop.
2901   bool CollectGuards = false;
2902   if (UnswitchGuards) {
2903     auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2904         Intrinsic::getName(Intrinsic::experimental_guard));
2905     if (GuardDecl && !GuardDecl->use_empty())
2906       CollectGuards = true;
2907   }
2908 
2909   for (auto *BB : L.blocks()) {
2910     if (LI.getLoopFor(BB) != &L)
2911       continue;
2912 
2913     for (auto &I : *BB) {
2914       if (auto *SI = dyn_cast<SelectInst>(&I)) {
2915         auto *Cond = SI->getCondition();
2916         // Do not unswitch vector selects and logical and/or selects
2917         if (Cond->getType()->isIntegerTy(1) && !SI->getType()->isIntegerTy(1))
2918           AddUnswitchCandidatesForInst(SI, Cond);
2919       } else if (CollectGuards && isGuard(&I)) {
2920         auto *Cond =
2921             skipTrivialSelect(cast<IntrinsicInst>(&I)->getArgOperand(0));
2922         // TODO: Support AND, OR conditions and partial unswitching.
2923         if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
2924           UnswitchCandidates.push_back({&I, {Cond}});
2925       }
2926     }
2927 
2928     if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
2929       // We can only consider fully loop-invariant switch conditions as we need
2930       // to completely eliminate the switch after unswitching.
2931       if (!isa<Constant>(SI->getCondition()) &&
2932           L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor())
2933         UnswitchCandidates.push_back({SI, {SI->getCondition()}});
2934       continue;
2935     }
2936 
2937     auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
2938     if (!BI || !BI->isConditional() ||
2939         BI->getSuccessor(0) == BI->getSuccessor(1))
2940       continue;
2941 
2942     AddUnswitchCandidatesForInst(BI, BI->getCondition());
2943   }
2944 
2945   if (MSSAU && !findOptionMDForLoop(&L, "llvm.loop.unswitch.partial.disable") &&
2946       !any_of(UnswitchCandidates, [&L](auto &TerminatorAndInvariants) {
2947          return TerminatorAndInvariants.TI == L.getHeader()->getTerminator();
2948        })) {
2949     MemorySSA *MSSA = MSSAU->getMemorySSA();
2950     if (auto Info = hasPartialIVCondition(L, MSSAThreshold, *MSSA, AA)) {
2951       LLVM_DEBUG(
2952           dbgs() << "simple-loop-unswitch: Found partially invariant condition "
2953                  << *Info->InstToDuplicate[0] << "\n");
2954       PartialIVInfo = *Info;
2955       PartialIVCondBranch = L.getHeader()->getTerminator();
2956       TinyPtrVector<Value *> ValsToDuplicate;
2957       llvm::append_range(ValsToDuplicate, Info->InstToDuplicate);
2958       UnswitchCandidates.push_back(
2959           {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)});
2960     }
2961   }
2962   return !UnswitchCandidates.empty();
2963 }
2964 
2965 /// Tries to canonicalize condition described by:
2966 ///
2967 ///   br (LHS pred RHS), label IfTrue, label IfFalse
2968 ///
2969 /// into its equivalent where `Pred` is something that we support for injected
2970 /// invariants (so far it is limited to ult), LHS in canonicalized form is
2971 /// non-invariant and RHS is an invariant.
2972 static void canonicalizeForInvariantConditionInjection(
2973     ICmpInst::Predicate &Pred, Value *&LHS, Value *&RHS, BasicBlock *&IfTrue,
2974     BasicBlock *&IfFalse, const Loop &L) {
2975   if (!L.contains(IfTrue)) {
2976     Pred = ICmpInst::getInversePredicate(Pred);
2977     std::swap(IfTrue, IfFalse);
2978   }
2979 
2980   // Move loop-invariant argument to RHS position.
2981   if (L.isLoopInvariant(LHS)) {
2982     Pred = ICmpInst::getSwappedPredicate(Pred);
2983     std::swap(LHS, RHS);
2984   }
2985 
2986   if (Pred == ICmpInst::ICMP_SGE && match(RHS, m_Zero())) {
2987     // Turn "x >=s 0" into "x <u UMIN_INT"
2988     Pred = ICmpInst::ICMP_ULT;
2989     RHS = ConstantInt::get(
2990         RHS->getContext(),
2991         APInt::getSignedMinValue(RHS->getType()->getIntegerBitWidth()));
2992   }
2993 }
2994 
2995 /// Returns true, if predicate described by ( \p Pred, \p LHS, \p RHS )
2996 /// succeeding into blocks ( \p IfTrue, \p IfFalse) can be optimized by
2997 /// injecting a loop-invariant condition.
2998 static bool shouldTryInjectInvariantCondition(
2999     const ICmpInst::Predicate Pred, const Value *LHS, const Value *RHS,
3000     const BasicBlock *IfTrue, const BasicBlock *IfFalse, const Loop &L) {
3001   if (L.isLoopInvariant(LHS) || !L.isLoopInvariant(RHS))
3002     return false;
3003   // TODO: Support other predicates.
3004   if (Pred != ICmpInst::ICMP_ULT)
3005     return false;
3006   // TODO: Support non-loop-exiting branches?
3007   if (!L.contains(IfTrue) || L.contains(IfFalse))
3008     return false;
3009   // FIXME: For some reason this causes problems with MSSA updates, need to
3010   // investigate why. So far, just don't unswitch latch.
3011   if (L.getHeader() == IfTrue)
3012     return false;
3013   return true;
3014 }
3015 
3016 /// Returns true, if metadata on \p BI allows us to optimize branching into \p
3017 /// TakenSucc via injection of invariant conditions. The branch should be not
3018 /// enough and not previously unswitched, the information about this comes from
3019 /// the metadata.
3020 bool shouldTryInjectBasingOnMetadata(const BranchInst *BI,
3021                                      const BasicBlock *TakenSucc) {
3022   SmallVector<uint32_t> Weights;
3023   if (!extractBranchWeights(*BI, Weights))
3024     return false;
3025   unsigned T = InjectInvariantConditionHotnesThreshold;
3026   BranchProbability LikelyTaken(T - 1, T);
3027 
3028   assert(Weights.size() == 2 && "Unexpected profile data!");
3029   size_t Idx = BI->getSuccessor(0) == TakenSucc ? 0 : 1;
3030   auto Num = Weights[Idx];
3031   auto Denom = Weights[0] + Weights[1];
3032   // Degenerate or overflowed metadata.
3033   if (Denom == 0 || Num > Denom)
3034     return false;
3035   BranchProbability ActualTaken(Num, Denom);
3036   if (LikelyTaken > ActualTaken)
3037     return false;
3038   return true;
3039 }
3040 
3041 /// Materialize pending invariant condition of the given candidate into IR. The
3042 /// injected loop-invariant condition implies the original loop-variant branch
3043 /// condition, so the materialization turns
3044 ///
3045 /// loop_block:
3046 ///   ...
3047 ///   br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3048 ///
3049 /// into
3050 ///
3051 /// preheader:
3052 ///   %invariant_cond = LHS pred RHS
3053 /// ...
3054 /// loop_block:
3055 ///   br i1 %invariant_cond, label InLoopSucc, label OriginalCheck
3056 /// OriginalCheck:
3057 ///   br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3058 /// ...
3059 static NonTrivialUnswitchCandidate
3060 injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L,
3061                                  DominatorTree &DT, LoopInfo &LI,
3062                                  AssumptionCache &AC, MemorySSAUpdater *MSSAU) {
3063   assert(Candidate.hasPendingInjection() && "Nothing to inject!");
3064   BasicBlock *Preheader = L.getLoopPreheader();
3065   assert(Preheader && "Loop is not in simplified form?");
3066   assert(LI.getLoopFor(Candidate.TI->getParent()) == &L &&
3067          "Unswitching branch of inner loop!");
3068 
3069   auto Pred = Candidate.PendingInjection->Pred;
3070   auto *LHS = Candidate.PendingInjection->LHS;
3071   auto *RHS = Candidate.PendingInjection->RHS;
3072   auto *InLoopSucc = Candidate.PendingInjection->InLoopSucc;
3073   auto *TI = cast<BranchInst>(Candidate.TI);
3074   auto *BB = Candidate.TI->getParent();
3075   auto *OutOfLoopSucc = InLoopSucc == TI->getSuccessor(0) ? TI->getSuccessor(1)
3076                                                           : TI->getSuccessor(0);
3077   // FIXME: Remove this once limitation on successors is lifted.
3078   assert(L.contains(InLoopSucc) && "Not supported yet!");
3079   assert(!L.contains(OutOfLoopSucc) && "Not supported yet!");
3080   auto &Ctx = BB->getContext();
3081 
3082   IRBuilder<> Builder(Preheader->getTerminator());
3083   assert(ICmpInst::isUnsigned(Pred) && "Not supported yet!");
3084   if (LHS->getType() != RHS->getType()) {
3085     if (LHS->getType()->getIntegerBitWidth() <
3086         RHS->getType()->getIntegerBitWidth())
3087       LHS = Builder.CreateZExt(LHS, RHS->getType(), LHS->getName() + ".wide");
3088     else
3089       RHS = Builder.CreateZExt(RHS, LHS->getType(), RHS->getName() + ".wide");
3090   }
3091   // Do not use builder here: CreateICmp may simplify this into a constant and
3092   // unswitching will break. Better optimize it away later.
3093   auto *InjectedCond =
3094       ICmpInst::Create(Instruction::ICmp, Pred, LHS, RHS, "injected.cond",
3095                        Preheader->getTerminator());
3096 
3097   BasicBlock *CheckBlock = BasicBlock::Create(Ctx, BB->getName() + ".check",
3098                                               BB->getParent(), InLoopSucc);
3099   Builder.SetInsertPoint(TI);
3100   auto *InvariantBr =
3101       Builder.CreateCondBr(InjectedCond, InLoopSucc, CheckBlock);
3102 
3103   Builder.SetInsertPoint(CheckBlock);
3104   Builder.CreateCondBr(TI->getCondition(), TI->getSuccessor(0),
3105                        TI->getSuccessor(1));
3106   TI->eraseFromParent();
3107 
3108   // Fixup phis.
3109   for (auto &I : *InLoopSucc) {
3110     auto *PN = dyn_cast<PHINode>(&I);
3111     if (!PN)
3112       break;
3113     auto *Inc = PN->getIncomingValueForBlock(BB);
3114     PN->addIncoming(Inc, CheckBlock);
3115   }
3116   OutOfLoopSucc->replacePhiUsesWith(BB, CheckBlock);
3117 
3118   SmallVector<DominatorTree::UpdateType, 4> DTUpdates = {
3119     { DominatorTree::Insert, BB, CheckBlock },
3120     { DominatorTree::Insert, CheckBlock, InLoopSucc },
3121     { DominatorTree::Insert, CheckBlock, OutOfLoopSucc },
3122     { DominatorTree::Delete, BB, OutOfLoopSucc }
3123   };
3124 
3125   DT.applyUpdates(DTUpdates);
3126   if (MSSAU)
3127     MSSAU->applyUpdates(DTUpdates, DT);
3128   L.addBasicBlockToLoop(CheckBlock, LI);
3129 
3130 #ifndef NDEBUG
3131   DT.verify();
3132   LI.verify(DT);
3133   if (MSSAU && VerifyMemorySSA)
3134     MSSAU->getMemorySSA()->verifyMemorySSA();
3135 #endif
3136 
3137   // TODO: In fact, cost of unswitching a new invariant candidate is *slightly*
3138   // higher because we have just inserted a new block. Need to think how to
3139   // adjust the cost of injected candidates when it was first computed.
3140   LLVM_DEBUG(dbgs() << "Injected a new loop-invariant branch " << *InvariantBr
3141                     << " and considering it for unswitching.");
3142   ++NumInvariantConditionsInjected;
3143   return NonTrivialUnswitchCandidate(InvariantBr, { InjectedCond },
3144                                      Candidate.Cost);
3145 }
3146 
3147 /// Given chain of loop branch conditions looking like:
3148 ///   br (Variant < Invariant1)
3149 ///   br (Variant < Invariant2)
3150 ///   br (Variant < Invariant3)
3151 ///   ...
3152 /// collect set of invariant conditions on which we want to unswitch, which
3153 /// look like:
3154 ///   Invariant1 <= Invariant2
3155 ///   Invariant2 <= Invariant3
3156 ///   ...
3157 /// Though they might not immediately exist in the IR, we can still inject them.
3158 static bool insertCandidatesWithPendingInjections(
3159     SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, Loop &L,
3160     ICmpInst::Predicate Pred, ArrayRef<CompareDesc> Compares,
3161     const DominatorTree &DT) {
3162 
3163   assert(ICmpInst::isRelational(Pred));
3164   assert(ICmpInst::isStrictPredicate(Pred));
3165   if (Compares.size() < 2)
3166     return false;
3167   ICmpInst::Predicate NonStrictPred = ICmpInst::getNonStrictPredicate(Pred);
3168   for (auto Prev = Compares.begin(), Next = Compares.begin() + 1;
3169        Next != Compares.end(); ++Prev, ++Next) {
3170     Value *LHS = Next->Invariant;
3171     Value *RHS = Prev->Invariant;
3172     BasicBlock *InLoopSucc = Prev->InLoopSucc;
3173     InjectedInvariant ToInject(NonStrictPred, LHS, RHS, InLoopSucc);
3174     NonTrivialUnswitchCandidate Candidate(Prev->Term, { LHS, RHS },
3175                                           std::nullopt, std::move(ToInject));
3176     UnswitchCandidates.push_back(std::move(Candidate));
3177   }
3178   return true;
3179 }
3180 
3181 /// Collect unswitch candidates by invariant conditions that are not immediately
3182 /// present in the loop. However, they can be injected into the code if we
3183 /// decide it's profitable.
3184 /// An example of such conditions is following:
3185 ///
3186 ///   for (...) {
3187 ///     x = load ...
3188 ///     if (! x <u C1) break;
3189 ///     if (! x <u C2) break;
3190 ///     <do something>
3191 ///   }
3192 ///
3193 /// We can unswitch by condition "C1 <=u C2". If that is true, then "x <u C1 <=
3194 /// C2" automatically implies "x <u C2", so we can get rid of one of
3195 /// loop-variant checks in unswitched loop version.
3196 static bool collectUnswitchCandidatesWithInjections(
3197     SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
3198     IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, Loop &L,
3199     const DominatorTree &DT, const LoopInfo &LI, AAResults &AA,
3200     const MemorySSAUpdater *MSSAU) {
3201   if (!InjectInvariantConditions)
3202     return false;
3203 
3204   if (!DT.isReachableFromEntry(L.getHeader()))
3205     return false;
3206   auto *Latch = L.getLoopLatch();
3207   // Need to have a single latch and a preheader.
3208   if (!Latch)
3209     return false;
3210   assert(L.getLoopPreheader() && "Must have a preheader!");
3211 
3212   DenseMap<Value *, SmallVector<CompareDesc, 4> > CandidatesULT;
3213   // Traverse the conditions that dominate latch (and therefore dominate each
3214   // other).
3215   for (auto *DTN = DT.getNode(Latch); L.contains(DTN->getBlock());
3216        DTN = DTN->getIDom()) {
3217     ICmpInst::Predicate Pred;
3218     Value *LHS = nullptr, *RHS = nullptr;
3219     BasicBlock *IfTrue = nullptr, *IfFalse = nullptr;
3220     auto *BB = DTN->getBlock();
3221     // Ignore inner loops.
3222     if (LI.getLoopFor(BB) != &L)
3223       continue;
3224     auto *Term = BB->getTerminator();
3225     if (!match(Term, m_Br(m_ICmp(Pred, m_Value(LHS), m_Value(RHS)),
3226                           m_BasicBlock(IfTrue), m_BasicBlock(IfFalse))))
3227       continue;
3228     if (!LHS->getType()->isIntegerTy())
3229       continue;
3230     canonicalizeForInvariantConditionInjection(Pred, LHS, RHS, IfTrue, IfFalse,
3231                                                L);
3232     if (!shouldTryInjectInvariantCondition(Pred, LHS, RHS, IfTrue, IfFalse, L))
3233       continue;
3234     if (!shouldTryInjectBasingOnMetadata(cast<BranchInst>(Term), IfTrue))
3235       continue;
3236     // Strip ZEXT for unsigned predicate.
3237     // TODO: once signed predicates are supported, also strip SEXT.
3238     CompareDesc Desc(cast<BranchInst>(Term), RHS, IfTrue);
3239     while (auto *Zext = dyn_cast<ZExtInst>(LHS))
3240       LHS = Zext->getOperand(0);
3241     CandidatesULT[LHS].push_back(Desc);
3242   }
3243 
3244   bool Found = false;
3245   for (auto &It : CandidatesULT)
3246     Found |= insertCandidatesWithPendingInjections(
3247         UnswitchCandidates, L, ICmpInst::ICMP_ULT, It.second, DT);
3248   return Found;
3249 }
3250 
3251 static bool isSafeForNoNTrivialUnswitching(Loop &L, LoopInfo &LI) {
3252   if (!L.isSafeToClone())
3253     return false;
3254   for (auto *BB : L.blocks())
3255     for (auto &I : *BB) {
3256       if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
3257         return false;
3258       if (auto *CB = dyn_cast<CallBase>(&I)) {
3259         assert(!CB->cannotDuplicate() && "Checked by L.isSafeToClone().");
3260         if (CB->isConvergent())
3261           return false;
3262       }
3263     }
3264 
3265   // Check if there are irreducible CFG cycles in this loop. If so, we cannot
3266   // easily unswitch non-trivial edges out of the loop. Doing so might turn the
3267   // irreducible control flow into reducible control flow and introduce new
3268   // loops "out of thin air". If we ever discover important use cases for doing
3269   // this, we can add support to loop unswitch, but it is a lot of complexity
3270   // for what seems little or no real world benefit.
3271   LoopBlocksRPO RPOT(&L);
3272   RPOT.perform(&LI);
3273   if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
3274     return false;
3275 
3276   SmallVector<BasicBlock *, 4> ExitBlocks;
3277   L.getUniqueExitBlocks(ExitBlocks);
3278   // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch
3279   // instruction as we don't know how to split those exit blocks.
3280   // FIXME: We should teach SplitBlock to handle this and remove this
3281   // restriction.
3282   for (auto *ExitBB : ExitBlocks) {
3283     auto *I = ExitBB->getFirstNonPHI();
3284     if (isa<CleanupPadInst>(I) || isa<CatchSwitchInst>(I)) {
3285       LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch "
3286                            "in exit block\n");
3287       return false;
3288     }
3289   }
3290 
3291   return true;
3292 }
3293 
3294 static NonTrivialUnswitchCandidate findBestNonTrivialUnswitchCandidate(
3295     ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates, const Loop &L,
3296     const DominatorTree &DT, const LoopInfo &LI, AssumptionCache &AC,
3297     const TargetTransformInfo &TTI, const IVConditionInfo &PartialIVInfo) {
3298   // Given that unswitching these terminators will require duplicating parts of
3299   // the loop, so we need to be able to model that cost. Compute the ephemeral
3300   // values and set up a data structure to hold per-BB costs. We cache each
3301   // block's cost so that we don't recompute this when considering different
3302   // subsets of the loop for duplication during unswitching.
3303   SmallPtrSet<const Value *, 4> EphValues;
3304   CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);
3305   SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap;
3306 
3307   // Compute the cost of each block, as well as the total loop cost. Also, bail
3308   // out if we see instructions which are incompatible with loop unswitching
3309   // (convergent, noduplicate, or cross-basic-block tokens).
3310   // FIXME: We might be able to safely handle some of these in non-duplicated
3311   // regions.
3312   TargetTransformInfo::TargetCostKind CostKind =
3313       L.getHeader()->getParent()->hasMinSize()
3314       ? TargetTransformInfo::TCK_CodeSize
3315       : TargetTransformInfo::TCK_SizeAndLatency;
3316   InstructionCost LoopCost = 0;
3317   for (auto *BB : L.blocks()) {
3318     InstructionCost Cost = 0;
3319     for (auto &I : *BB) {
3320       if (EphValues.count(&I))
3321         continue;
3322       Cost += TTI.getInstructionCost(&I, CostKind);
3323     }
3324     assert(Cost >= 0 && "Must not have negative costs!");
3325     LoopCost += Cost;
3326     assert(LoopCost >= 0 && "Must not have negative loop costs!");
3327     BBCostMap[BB] = Cost;
3328   }
3329   LLVM_DEBUG(dbgs() << "  Total loop cost: " << LoopCost << "\n");
3330 
3331   // Now we find the best candidate by searching for the one with the following
3332   // properties in order:
3333   //
3334   // 1) An unswitching cost below the threshold
3335   // 2) The smallest number of duplicated unswitch candidates (to avoid
3336   //    creating redundant subsequent unswitching)
3337   // 3) The smallest cost after unswitching.
3338   //
3339   // We prioritize reducing fanout of unswitch candidates provided the cost
3340   // remains below the threshold because this has a multiplicative effect.
3341   //
3342   // This requires memoizing each dominator subtree to avoid redundant work.
3343   //
3344   // FIXME: Need to actually do the number of candidates part above.
3345   SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap;
3346   // Given a terminator which might be unswitched, computes the non-duplicated
3347   // cost for that terminator.
3348   auto ComputeUnswitchedCost = [&](Instruction &TI,
3349                                    bool FullUnswitch) -> InstructionCost {
3350     // Unswitching selects unswitches the entire loop.
3351     if (isa<SelectInst>(TI))
3352       return LoopCost;
3353 
3354     BasicBlock &BB = *TI.getParent();
3355     SmallPtrSet<BasicBlock *, 4> Visited;
3356 
3357     InstructionCost Cost = 0;
3358     for (BasicBlock *SuccBB : successors(&BB)) {
3359       // Don't count successors more than once.
3360       if (!Visited.insert(SuccBB).second)
3361         continue;
3362 
3363       // If this is a partial unswitch candidate, then it must be a conditional
3364       // branch with a condition of either `or`, `and`, their corresponding
3365       // select forms or partially invariant instructions. In that case, one of
3366       // the successors is necessarily duplicated, so don't even try to remove
3367       // its cost.
3368       if (!FullUnswitch) {
3369         auto &BI = cast<BranchInst>(TI);
3370         Value *Cond = skipTrivialSelect(BI.getCondition());
3371         if (match(Cond, m_LogicalAnd())) {
3372           if (SuccBB == BI.getSuccessor(1))
3373             continue;
3374         } else if (match(Cond, m_LogicalOr())) {
3375           if (SuccBB == BI.getSuccessor(0))
3376             continue;
3377         } else if ((PartialIVInfo.KnownValue->isOneValue() &&
3378                     SuccBB == BI.getSuccessor(0)) ||
3379                    (!PartialIVInfo.KnownValue->isOneValue() &&
3380                     SuccBB == BI.getSuccessor(1)))
3381           continue;
3382       }
3383 
3384       // This successor's domtree will not need to be duplicated after
3385       // unswitching if the edge to the successor dominates it (and thus the
3386       // entire tree). This essentially means there is no other path into this
3387       // subtree and so it will end up live in only one clone of the loop.
3388       if (SuccBB->getUniquePredecessor() ||
3389           llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
3390             return PredBB == &BB || DT.dominates(SuccBB, PredBB);
3391           })) {
3392         Cost += computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);
3393         assert(Cost <= LoopCost &&
3394                "Non-duplicated cost should never exceed total loop cost!");
3395       }
3396     }
3397 
3398     // Now scale the cost by the number of unique successors minus one. We
3399     // subtract one because there is already at least one copy of the entire
3400     // loop. This is computing the new cost of unswitching a condition.
3401     // Note that guards always have 2 unique successors that are implicit and
3402     // will be materialized if we decide to unswitch it.
3403     int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
3404     assert(SuccessorsCount > 1 &&
3405            "Cannot unswitch a condition without multiple distinct successors!");
3406     return (LoopCost - Cost) * (SuccessorsCount - 1);
3407   };
3408 
3409   std::optional<NonTrivialUnswitchCandidate> Best;
3410   for (auto &Candidate : UnswitchCandidates) {
3411     Instruction &TI = *Candidate.TI;
3412     ArrayRef<Value *> Invariants = Candidate.Invariants;
3413     BranchInst *BI = dyn_cast<BranchInst>(&TI);
3414     bool FullUnswitch =
3415         !BI || Candidate.hasPendingInjection() ||
3416         (Invariants.size() == 1 &&
3417          Invariants[0] == skipTrivialSelect(BI->getCondition()));
3418     InstructionCost CandidateCost = ComputeUnswitchedCost(TI, FullUnswitch);
3419     // Calculate cost multiplier which is a tool to limit potentially
3420     // exponential behavior of loop-unswitch.
3421     if (EnableUnswitchCostMultiplier) {
3422       int CostMultiplier =
3423           CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
3424       assert(
3425           (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
3426           "cost multiplier needs to be in the range of 1..UnswitchThreshold");
3427       CandidateCost *= CostMultiplier;
3428       LLVM_DEBUG(dbgs() << "  Computed cost of " << CandidateCost
3429                         << " (multiplier: " << CostMultiplier << ")"
3430                         << " for unswitch candidate: " << TI << "\n");
3431     } else {
3432       LLVM_DEBUG(dbgs() << "  Computed cost of " << CandidateCost
3433                         << " for unswitch candidate: " << TI << "\n");
3434     }
3435 
3436     if (!Best || CandidateCost < Best->Cost) {
3437       Best = Candidate;
3438       Best->Cost = CandidateCost;
3439     }
3440   }
3441   assert(Best && "Must be!");
3442   return *Best;
3443 }
3444 
3445 // Insert a freeze on an unswitched branch if all is true:
3446 // 1. freeze-loop-unswitch-cond option is true
3447 // 2. The branch may not execute in the loop pre-transformation. If a branch may
3448 // not execute and could cause UB, it would always cause UB if it is hoisted outside
3449 // of the loop. Insert a freeze to prevent this case.
3450 // 3. The branch condition may be poison or undef
3451 static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT,
3452                                AssumptionCache &AC) {
3453   assert(isa<BranchInst>(TI) || isa<SwitchInst>(TI));
3454   if (!FreezeLoopUnswitchCond)
3455     return false;
3456 
3457   ICFLoopSafetyInfo SafetyInfo;
3458   SafetyInfo.computeLoopSafetyInfo(&L);
3459   if (SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))
3460     return false;
3461 
3462   Value *Cond;
3463   if (BranchInst *BI = dyn_cast<BranchInst>(&TI))
3464     Cond = skipTrivialSelect(BI->getCondition());
3465   else
3466     Cond = skipTrivialSelect(cast<SwitchInst>(&TI)->getCondition());
3467   return !isGuaranteedNotToBeUndefOrPoison(
3468       Cond, &AC, L.getLoopPreheader()->getTerminator(), &DT);
3469 }
3470 
3471 static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
3472                                   AssumptionCache &AC, AAResults &AA,
3473                                   TargetTransformInfo &TTI, ScalarEvolution *SE,
3474                                   MemorySSAUpdater *MSSAU,
3475                                   LPMUpdater &LoopUpdater) {
3476   // Collect all invariant conditions within this loop (as opposed to an inner
3477   // loop which would be handled when visiting that inner loop).
3478   SmallVector<NonTrivialUnswitchCandidate, 4> UnswitchCandidates;
3479   IVConditionInfo PartialIVInfo;
3480   Instruction *PartialIVCondBranch = nullptr;
3481   collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo,
3482                             PartialIVCondBranch, L, LI, AA, MSSAU);
3483   if (!findOptionMDForLoop(&L, "llvm.loop.unswitch.injection.disable"))
3484     collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo,
3485                                             PartialIVCondBranch, L, DT, LI, AA,
3486                                             MSSAU);
3487   // If we didn't find any candidates, we're done.
3488   if (UnswitchCandidates.empty())
3489     return false;
3490 
3491   LLVM_DEBUG(
3492       dbgs() << "Considering " << UnswitchCandidates.size()
3493              << " non-trivial loop invariant conditions for unswitching.\n");
3494 
3495   NonTrivialUnswitchCandidate Best = findBestNonTrivialUnswitchCandidate(
3496       UnswitchCandidates, L, DT, LI, AC, TTI, PartialIVInfo);
3497 
3498   assert(Best.TI && "Failed to find loop unswitch candidate");
3499   assert(Best.Cost && "Failed to compute cost");
3500 
3501   if (*Best.Cost >= UnswitchThreshold) {
3502     LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << *Best.Cost
3503                       << "\n");
3504     return false;
3505   }
3506 
3507   bool InjectedCondition = false;
3508   if (Best.hasPendingInjection()) {
3509     Best = injectPendingInvariantConditions(Best, L, DT, LI, AC, MSSAU);
3510     InjectedCondition = true;
3511   }
3512   assert(!Best.hasPendingInjection() &&
3513          "All injections should have been done by now!");
3514 
3515   if (Best.TI != PartialIVCondBranch)
3516     PartialIVInfo.InstToDuplicate.clear();
3517 
3518   bool InsertFreeze;
3519   if (auto *SI = dyn_cast<SelectInst>(Best.TI)) {
3520     // If the best candidate is a select, turn it into a branch. Select
3521     // instructions with a poison conditional do not propagate poison, but
3522     // branching on poison causes UB. Insert a freeze on the select
3523     // conditional to prevent UB after turning the select into a branch.
3524     InsertFreeze = !isGuaranteedNotToBeUndefOrPoison(
3525         SI->getCondition(), &AC, L.getLoopPreheader()->getTerminator(), &DT);
3526     Best.TI = turnSelectIntoBranch(SI, DT, LI, MSSAU, &AC);
3527   } else {
3528     // If the best candidate is a guard, turn it into a branch.
3529     if (isGuard(Best.TI))
3530       Best.TI =
3531           turnGuardIntoBranch(cast<IntrinsicInst>(Best.TI), L, DT, LI, MSSAU);
3532     InsertFreeze = shouldInsertFreeze(L, *Best.TI, DT, AC);
3533   }
3534 
3535   LLVM_DEBUG(dbgs() << "  Unswitching non-trivial (cost = " << Best.Cost
3536                     << ") terminator: " << *Best.TI << "\n");
3537   unswitchNontrivialInvariants(L, *Best.TI, Best.Invariants, PartialIVInfo, DT,
3538                                LI, AC, SE, MSSAU, LoopUpdater, InsertFreeze,
3539                                InjectedCondition);
3540   return true;
3541 }
3542 
3543 /// Unswitch control flow predicated on loop invariant conditions.
3544 ///
3545 /// This first hoists all branches or switches which are trivial (IE, do not
3546 /// require duplicating any part of the loop) out of the loop body. It then
3547 /// looks at other loop invariant control flows and tries to unswitch those as
3548 /// well by cloning the loop if the result is small enough.
3549 ///
3550 /// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are
3551 /// also updated based on the unswitch. The `MSSA` analysis is also updated if
3552 /// valid (i.e. its use is enabled).
3553 ///
3554 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
3555 /// true, we will attempt to do non-trivial unswitching as well as trivial
3556 /// unswitching.
3557 ///
3558 /// The `postUnswitch` function will be run after unswitching is complete
3559 /// with information on whether or not the provided loop remains a loop and
3560 /// a list of new sibling loops created.
3561 ///
3562 /// If `SE` is non-null, we will update that analysis based on the unswitching
3563 /// done.
3564 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
3565                          AssumptionCache &AC, AAResults &AA,
3566                          TargetTransformInfo &TTI, bool Trivial,
3567                          bool NonTrivial, ScalarEvolution *SE,
3568                          MemorySSAUpdater *MSSAU, ProfileSummaryInfo *PSI,
3569                          BlockFrequencyInfo *BFI, LPMUpdater &LoopUpdater) {
3570   assert(L.isRecursivelyLCSSAForm(DT, LI) &&
3571          "Loops must be in LCSSA form before unswitching.");
3572 
3573   // Must be in loop simplified form: we need a preheader and dedicated exits.
3574   if (!L.isLoopSimplifyForm())
3575     return false;
3576 
3577   // Try trivial unswitch first before loop over other basic blocks in the loop.
3578   if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
3579     // If we unswitched successfully we will want to clean up the loop before
3580     // processing it further so just mark it as unswitched and return.
3581     postUnswitch(L, LoopUpdater, L.getName(),
3582                  /*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false,
3583                  /*InjectedCondition*/ false, {});
3584     return true;
3585   }
3586 
3587   const Function *F = L.getHeader()->getParent();
3588 
3589   // Check whether we should continue with non-trivial conditions.
3590   // EnableNonTrivialUnswitch: Global variable that forces non-trivial
3591   //                           unswitching for testing and debugging.
3592   // NonTrivial: Parameter that enables non-trivial unswitching for this
3593   //             invocation of the transform. But this should be allowed only
3594   //             for targets without branch divergence.
3595   //
3596   // FIXME: If divergence analysis becomes available to a loop
3597   // transform, we should allow unswitching for non-trivial uniform
3598   // branches even on targets that have divergence.
3599   // https://bugs.llvm.org/show_bug.cgi?id=48819
3600   bool ContinueWithNonTrivial =
3601       EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence(F));
3602   if (!ContinueWithNonTrivial)
3603     return false;
3604 
3605   // Skip non-trivial unswitching for optsize functions.
3606   if (F->hasOptSize())
3607     return false;
3608 
3609   // Returns true if Loop L's loop nest is cold, i.e. if the headers of L,
3610   // of the loops L is nested in, and of the loops nested in L are all cold.
3611   auto IsLoopNestCold = [&](const Loop *L) {
3612     // Check L and all of its parent loops.
3613     auto *Parent = L;
3614     while (Parent) {
3615       if (!PSI->isColdBlock(Parent->getHeader(), BFI))
3616         return false;
3617       Parent = Parent->getParentLoop();
3618     }
3619     // Next check all loops nested within L.
3620     SmallVector<const Loop *, 4> Worklist;
3621     Worklist.insert(Worklist.end(), L->getSubLoops().begin(),
3622                     L->getSubLoops().end());
3623     while (!Worklist.empty()) {
3624       auto *CurLoop = Worklist.pop_back_val();
3625       if (!PSI->isColdBlock(CurLoop->getHeader(), BFI))
3626         return false;
3627       Worklist.insert(Worklist.end(), CurLoop->getSubLoops().begin(),
3628                       CurLoop->getSubLoops().end());
3629     }
3630     return true;
3631   };
3632 
3633   // Skip cold loops in cold loop nests, as unswitching them brings little
3634   // benefit but increases the code size
3635   if (PSI && PSI->hasProfileSummary() && BFI && IsLoopNestCold(&L)) {
3636     LLVM_DEBUG(dbgs() << " Skip cold loop: " << L << "\n");
3637     return false;
3638   }
3639 
3640   // Perform legality checks.
3641   if (!isSafeForNoNTrivialUnswitching(L, LI))
3642     return false;
3643 
3644   // For non-trivial unswitching, because it often creates new loops, we rely on
3645   // the pass manager to iterate on the loops rather than trying to immediately
3646   // reach a fixed point. There is no substantial advantage to iterating
3647   // internally, and if any of the new loops are simplified enough to contain
3648   // trivial unswitching we want to prefer those.
3649 
3650   // Try to unswitch the best invariant condition. We prefer this full unswitch to
3651   // a partial unswitch when possible below the threshold.
3652   if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, SE, MSSAU, LoopUpdater))
3653     return true;
3654 
3655   // No other opportunities to unswitch.
3656   return false;
3657 }
3658 
3659 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
3660                                               LoopStandardAnalysisResults &AR,
3661                                               LPMUpdater &U) {
3662   Function &F = *L.getHeader()->getParent();
3663   (void)F;
3664   ProfileSummaryInfo *PSI = nullptr;
3665   if (auto OuterProxy =
3666           AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR)
3667               .getCachedResult<ModuleAnalysisManagerFunctionProxy>(F))
3668     PSI = OuterProxy->getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
3669   LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
3670                     << "\n");
3671 
3672   std::optional<MemorySSAUpdater> MSSAU;
3673   if (AR.MSSA) {
3674     MSSAU = MemorySSAUpdater(AR.MSSA);
3675     if (VerifyMemorySSA)
3676       AR.MSSA->verifyMemorySSA();
3677   }
3678   if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.AA, AR.TTI, Trivial, NonTrivial,
3679                     &AR.SE, MSSAU ? &*MSSAU : nullptr, PSI, AR.BFI, U))
3680     return PreservedAnalyses::all();
3681 
3682   if (AR.MSSA && VerifyMemorySSA)
3683     AR.MSSA->verifyMemorySSA();
3684 
3685   // Historically this pass has had issues with the dominator tree so verify it
3686   // in asserts builds.
3687   assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));
3688 
3689   auto PA = getLoopPassPreservedAnalyses();
3690   if (AR.MSSA)
3691     PA.preserve<MemorySSAAnalysis>();
3692   return PA;
3693 }
3694 
3695 void SimpleLoopUnswitchPass::printPipeline(
3696     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
3697   static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline(
3698       OS, MapClassName2PassName);
3699 
3700   OS << '<';
3701   OS << (NonTrivial ? "" : "no-") << "nontrivial;";
3702   OS << (Trivial ? "" : "no-") << "trivial";
3703   OS << '>';
3704 }
3705