xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp (revision 5956d97f4b3204318ceb6aa9c77bd0bc6ea87a41)
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/CFG.h"
20 #include "llvm/Analysis/CodeMetrics.h"
21 #include "llvm/Analysis/GuardUtils.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/LoopAnalysisManager.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/LoopIterator.h"
26 #include "llvm/Analysis/LoopPass.h"
27 #include "llvm/Analysis/MemorySSA.h"
28 #include "llvm/Analysis/MemorySSAUpdater.h"
29 #include "llvm/Analysis/MustExecute.h"
30 #include "llvm/Analysis/ScalarEvolution.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/Constant.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/IRBuilder.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Instruction.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/PatternMatch.h"
43 #include "llvm/IR/Use.h"
44 #include "llvm/IR/Value.h"
45 #include "llvm/InitializePasses.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/GenericDomTree.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
54 #include "llvm/Transforms/Utils/Cloning.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/LoopUtils.h"
57 #include "llvm/Transforms/Utils/ValueMapper.h"
58 #include <algorithm>
59 #include <cassert>
60 #include <iterator>
61 #include <numeric>
62 #include <utility>
63 
64 #define DEBUG_TYPE "simple-loop-unswitch"
65 
66 using namespace llvm;
67 using namespace llvm::PatternMatch;
68 
69 STATISTIC(NumBranches, "Number of branches unswitched");
70 STATISTIC(NumSwitches, "Number of switches unswitched");
71 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
72 STATISTIC(NumTrivial, "Number of unswitches that are trivial");
73 STATISTIC(
74     NumCostMultiplierSkipped,
75     "Number of unswitch candidates that had their cost multiplier skipped");
76 
77 static cl::opt<bool> EnableNonTrivialUnswitch(
78     "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
79     cl::desc("Forcibly enables non-trivial loop unswitching rather than "
80              "following the configuration passed into the pass."));
81 
82 static cl::opt<int>
83     UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
84                       cl::ZeroOrMore,
85                       cl::desc("The cost threshold for unswitching a loop."));
86 
87 static cl::opt<bool> EnableUnswitchCostMultiplier(
88     "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden,
89     cl::desc("Enable unswitch cost multiplier that prohibits exponential "
90              "explosion in nontrivial unswitch."));
91 static cl::opt<int> UnswitchSiblingsToplevelDiv(
92     "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden,
93     cl::desc("Toplevel siblings divisor for cost multiplier."));
94 static cl::opt<int> UnswitchNumInitialUnscaledCandidates(
95     "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden,
96     cl::desc("Number of unswitch candidates that are ignored when calculating "
97              "cost multiplier."));
98 static cl::opt<bool> UnswitchGuards(
99     "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
100     cl::desc("If enabled, simple loop unswitching will also consider "
101              "llvm.experimental.guard intrinsics as unswitch candidates."));
102 static cl::opt<bool> DropNonTrivialImplicitNullChecks(
103     "simple-loop-unswitch-drop-non-trivial-implicit-null-checks",
104     cl::init(false), cl::Hidden,
105     cl::desc("If enabled, drop make.implicit metadata in unswitched implicit "
106              "null checks to save time analyzing if we can keep it."));
107 static cl::opt<unsigned>
108     MSSAThreshold("simple-loop-unswitch-memoryssa-threshold",
109                   cl::desc("Max number of memory uses to explore during "
110                            "partial unswitching analysis"),
111                   cl::init(100), cl::Hidden);
112 static cl::opt<bool> FreezeLoopUnswitchCond(
113     "freeze-loop-unswitch-cond", cl::init(false), cl::Hidden,
114     cl::desc("If enabled, the freeze instruction will be added to condition "
115              "of loop unswitch to prevent miscompilation."));
116 
117 /// Collect all of the loop invariant input values transitively used by the
118 /// homogeneous instruction graph from a given root.
119 ///
120 /// This essentially walks from a root recursively through loop variant operands
121 /// which have the exact same opcode and finds all inputs which are loop
122 /// invariant. For some operations these can be re-associated and unswitched out
123 /// of the loop entirely.
124 static TinyPtrVector<Value *>
125 collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root,
126                                          LoopInfo &LI) {
127   assert(!L.isLoopInvariant(&Root) &&
128          "Only need to walk the graph if root itself is not invariant.");
129   TinyPtrVector<Value *> Invariants;
130 
131   bool IsRootAnd = match(&Root, m_LogicalAnd());
132   bool IsRootOr  = match(&Root, m_LogicalOr());
133 
134   // Build a worklist and recurse through operators collecting invariants.
135   SmallVector<Instruction *, 4> Worklist;
136   SmallPtrSet<Instruction *, 8> Visited;
137   Worklist.push_back(&Root);
138   Visited.insert(&Root);
139   do {
140     Instruction &I = *Worklist.pop_back_val();
141     for (Value *OpV : I.operand_values()) {
142       // Skip constants as unswitching isn't interesting for them.
143       if (isa<Constant>(OpV))
144         continue;
145 
146       // Add it to our result if loop invariant.
147       if (L.isLoopInvariant(OpV)) {
148         Invariants.push_back(OpV);
149         continue;
150       }
151 
152       // If not an instruction with the same opcode, nothing we can do.
153       Instruction *OpI = dyn_cast<Instruction>(OpV);
154 
155       if (OpI && ((IsRootAnd && match(OpI, m_LogicalAnd())) ||
156                   (IsRootOr  && match(OpI, m_LogicalOr())))) {
157         // Visit this operand.
158         if (Visited.insert(OpI).second)
159           Worklist.push_back(OpI);
160       }
161     }
162   } while (!Worklist.empty());
163 
164   return Invariants;
165 }
166 
167 static void replaceLoopInvariantUses(Loop &L, Value *Invariant,
168                                      Constant &Replacement) {
169   assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");
170 
171   // Replace uses of LIC in the loop with the given constant.
172   // We use make_early_inc_range as set invalidates the iterator.
173   for (Use &U : llvm::make_early_inc_range(Invariant->uses())) {
174     Instruction *UserI = dyn_cast<Instruction>(U.getUser());
175 
176     // Replace this use within the loop body.
177     if (UserI && L.contains(UserI))
178       U.set(&Replacement);
179   }
180 }
181 
182 /// Check that all the LCSSA PHI nodes in the loop exit block have trivial
183 /// incoming values along this edge.
184 static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB,
185                                          BasicBlock &ExitBB) {
186   for (Instruction &I : ExitBB) {
187     auto *PN = dyn_cast<PHINode>(&I);
188     if (!PN)
189       // No more PHIs to check.
190       return true;
191 
192     // If the incoming value for this edge isn't loop invariant the unswitch
193     // won't be trivial.
194     if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB)))
195       return false;
196   }
197   llvm_unreachable("Basic blocks should never be empty!");
198 }
199 
200 /// Copy a set of loop invariant values \p ToDuplicate and insert them at the
201 /// end of \p BB and conditionally branch on the copied condition. We only
202 /// branch on a single value.
203 static void buildPartialUnswitchConditionalBranch(
204     BasicBlock &BB, ArrayRef<Value *> Invariants, bool Direction,
205     BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, bool InsertFreeze) {
206   IRBuilder<> IRB(&BB);
207 
208   Value *Cond = Direction ? IRB.CreateOr(Invariants) :
209     IRB.CreateAnd(Invariants);
210   if (InsertFreeze)
211     Cond = IRB.CreateFreeze(Cond, Cond->getName() + ".fr");
212   IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
213                    Direction ? &NormalSucc : &UnswitchedSucc);
214 }
215 
216 /// Copy a set of loop invariant values, and conditionally branch on them.
217 static void buildPartialInvariantUnswitchConditionalBranch(
218     BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction,
219     BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L,
220     MemorySSAUpdater *MSSAU) {
221   ValueToValueMapTy VMap;
222   for (auto *Val : reverse(ToDuplicate)) {
223     Instruction *Inst = cast<Instruction>(Val);
224     Instruction *NewInst = Inst->clone();
225     BB.getInstList().insert(BB.end(), NewInst);
226     RemapInstruction(NewInst, VMap,
227                      RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
228     VMap[Val] = NewInst;
229 
230     if (!MSSAU)
231       continue;
232 
233     MemorySSA *MSSA = MSSAU->getMemorySSA();
234     if (auto *MemUse =
235             dyn_cast_or_null<MemoryUse>(MSSA->getMemoryAccess(Inst))) {
236       auto *DefiningAccess = MemUse->getDefiningAccess();
237       // Get the first defining access before the loop.
238       while (L.contains(DefiningAccess->getBlock())) {
239         // If the defining access is a MemoryPhi, get the incoming
240         // value for the pre-header as defining access.
241         if (auto *MemPhi = dyn_cast<MemoryPhi>(DefiningAccess))
242           DefiningAccess =
243               MemPhi->getIncomingValueForBlock(L.getLoopPreheader());
244         else
245           DefiningAccess = cast<MemoryDef>(DefiningAccess)->getDefiningAccess();
246       }
247       MSSAU->createMemoryAccessInBB(NewInst, DefiningAccess,
248                                     NewInst->getParent(),
249                                     MemorySSA::BeforeTerminator);
250     }
251   }
252 
253   IRBuilder<> IRB(&BB);
254   Value *Cond = VMap[ToDuplicate[0]];
255   IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
256                    Direction ? &NormalSucc : &UnswitchedSucc);
257 }
258 
259 /// Rewrite the PHI nodes in an unswitched loop exit basic block.
260 ///
261 /// Requires that the loop exit and unswitched basic block are the same, and
262 /// that the exiting block was a unique predecessor of that block. Rewrites the
263 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
264 /// PHI nodes from the old preheader that now contains the unswitched
265 /// terminator.
266 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
267                                                   BasicBlock &OldExitingBB,
268                                                   BasicBlock &OldPH) {
269   for (PHINode &PN : UnswitchedBB.phis()) {
270     // When the loop exit is directly unswitched we just need to update the
271     // incoming basic block. We loop to handle weird cases with repeated
272     // incoming blocks, but expect to typically only have one operand here.
273     for (auto i : seq<int>(0, PN.getNumOperands())) {
274       assert(PN.getIncomingBlock(i) == &OldExitingBB &&
275              "Found incoming block different from unique predecessor!");
276       PN.setIncomingBlock(i, &OldPH);
277     }
278   }
279 }
280 
281 /// Rewrite the PHI nodes in the loop exit basic block and the split off
282 /// unswitched block.
283 ///
284 /// Because the exit block remains an exit from the loop, this rewrites the
285 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
286 /// nodes into the unswitched basic block to select between the value in the
287 /// old preheader and the loop exit.
288 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
289                                                       BasicBlock &UnswitchedBB,
290                                                       BasicBlock &OldExitingBB,
291                                                       BasicBlock &OldPH,
292                                                       bool FullUnswitch) {
293   assert(&ExitBB != &UnswitchedBB &&
294          "Must have different loop exit and unswitched blocks!");
295   Instruction *InsertPt = &*UnswitchedBB.begin();
296   for (PHINode &PN : ExitBB.phis()) {
297     auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,
298                                   PN.getName() + ".split", InsertPt);
299 
300     // Walk backwards over the old PHI node's inputs to minimize the cost of
301     // removing each one. We have to do this weird loop manually so that we
302     // create the same number of new incoming edges in the new PHI as we expect
303     // each case-based edge to be included in the unswitched switch in some
304     // cases.
305     // FIXME: This is really, really gross. It would be much cleaner if LLVM
306     // allowed us to create a single entry for a predecessor block without
307     // having separate entries for each "edge" even though these edges are
308     // required to produce identical results.
309     for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
310       if (PN.getIncomingBlock(i) != &OldExitingBB)
311         continue;
312 
313       Value *Incoming = PN.getIncomingValue(i);
314       if (FullUnswitch)
315         // No more edge from the old exiting block to the exit block.
316         PN.removeIncomingValue(i);
317 
318       NewPN->addIncoming(Incoming, &OldPH);
319     }
320 
321     // Now replace the old PHI with the new one and wire the old one in as an
322     // input to the new one.
323     PN.replaceAllUsesWith(NewPN);
324     NewPN->addIncoming(&PN, &ExitBB);
325   }
326 }
327 
328 /// Hoist the current loop up to the innermost loop containing a remaining exit.
329 ///
330 /// Because we've removed an exit from the loop, we may have changed the set of
331 /// loops reachable and need to move the current loop up the loop nest or even
332 /// to an entirely separate nest.
333 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
334                                  DominatorTree &DT, LoopInfo &LI,
335                                  MemorySSAUpdater *MSSAU, ScalarEvolution *SE) {
336   // If the loop is already at the top level, we can't hoist it anywhere.
337   Loop *OldParentL = L.getParentLoop();
338   if (!OldParentL)
339     return;
340 
341   SmallVector<BasicBlock *, 4> Exits;
342   L.getExitBlocks(Exits);
343   Loop *NewParentL = nullptr;
344   for (auto *ExitBB : Exits)
345     if (Loop *ExitL = LI.getLoopFor(ExitBB))
346       if (!NewParentL || NewParentL->contains(ExitL))
347         NewParentL = ExitL;
348 
349   if (NewParentL == OldParentL)
350     return;
351 
352   // The new parent loop (if different) should always contain the old one.
353   if (NewParentL)
354     assert(NewParentL->contains(OldParentL) &&
355            "Can only hoist this loop up the nest!");
356 
357   // The preheader will need to move with the body of this loop. However,
358   // because it isn't in this loop we also need to update the primary loop map.
359   assert(OldParentL == LI.getLoopFor(&Preheader) &&
360          "Parent loop of this loop should contain this loop's preheader!");
361   LI.changeLoopFor(&Preheader, NewParentL);
362 
363   // Remove this loop from its old parent.
364   OldParentL->removeChildLoop(&L);
365 
366   // Add the loop either to the new parent or as a top-level loop.
367   if (NewParentL)
368     NewParentL->addChildLoop(&L);
369   else
370     LI.addTopLevelLoop(&L);
371 
372   // Remove this loops blocks from the old parent and every other loop up the
373   // nest until reaching the new parent. Also update all of these
374   // no-longer-containing loops to reflect the nesting change.
375   for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
376        OldContainingL = OldContainingL->getParentLoop()) {
377     llvm::erase_if(OldContainingL->getBlocksVector(),
378                    [&](const BasicBlock *BB) {
379                      return BB == &Preheader || L.contains(BB);
380                    });
381 
382     OldContainingL->getBlocksSet().erase(&Preheader);
383     for (BasicBlock *BB : L.blocks())
384       OldContainingL->getBlocksSet().erase(BB);
385 
386     // Because we just hoisted a loop out of this one, we have essentially
387     // created new exit paths from it. That means we need to form LCSSA PHI
388     // nodes for values used in the no-longer-nested loop.
389     formLCSSA(*OldContainingL, DT, &LI, SE);
390 
391     // We shouldn't need to form dedicated exits because the exit introduced
392     // here is the (just split by unswitching) preheader. However, after trivial
393     // unswitching it is possible to get new non-dedicated exits out of parent
394     // loop so let's conservatively form dedicated exit blocks and figure out
395     // if we can optimize later.
396     formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU,
397                             /*PreserveLCSSA*/ true);
398   }
399 }
400 
401 // Return the top-most loop containing ExitBB and having ExitBB as exiting block
402 // or the loop containing ExitBB, if there is no parent loop containing ExitBB
403 // as exiting block.
404 static Loop *getTopMostExitingLoop(BasicBlock *ExitBB, LoopInfo &LI) {
405   Loop *TopMost = LI.getLoopFor(ExitBB);
406   Loop *Current = TopMost;
407   while (Current) {
408     if (Current->isLoopExiting(ExitBB))
409       TopMost = Current;
410     Current = Current->getParentLoop();
411   }
412   return TopMost;
413 }
414 
415 /// Unswitch a trivial branch if the condition is loop invariant.
416 ///
417 /// This routine should only be called when loop code leading to the branch has
418 /// been validated as trivial (no side effects). This routine checks if the
419 /// condition is invariant and one of the successors is a loop exit. This
420 /// allows us to unswitch without duplicating the loop, making it trivial.
421 ///
422 /// If this routine fails to unswitch the branch it returns false.
423 ///
424 /// If the branch can be unswitched, this routine splits the preheader and
425 /// hoists the branch above that split. Preserves loop simplified form
426 /// (splitting the exit block as necessary). It simplifies the branch within
427 /// the loop to an unconditional branch but doesn't remove it entirely. Further
428 /// cleanup can be done with some simplifycfg like pass.
429 ///
430 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
431 /// invalidated by this.
432 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
433                                   LoopInfo &LI, ScalarEvolution *SE,
434                                   MemorySSAUpdater *MSSAU) {
435   assert(BI.isConditional() && "Can only unswitch a conditional branch!");
436   LLVM_DEBUG(dbgs() << "  Trying to unswitch branch: " << BI << "\n");
437 
438   // The loop invariant values that we want to unswitch.
439   TinyPtrVector<Value *> Invariants;
440 
441   // When true, we're fully unswitching the branch rather than just unswitching
442   // some input conditions to the branch.
443   bool FullUnswitch = false;
444 
445   if (L.isLoopInvariant(BI.getCondition())) {
446     Invariants.push_back(BI.getCondition());
447     FullUnswitch = true;
448   } else {
449     if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition()))
450       Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);
451     if (Invariants.empty()) {
452       LLVM_DEBUG(dbgs() << "   Couldn't find invariant inputs!\n");
453       return false;
454     }
455   }
456 
457   // Check that one of the branch's successors exits, and which one.
458   bool ExitDirection = true;
459   int LoopExitSuccIdx = 0;
460   auto *LoopExitBB = BI.getSuccessor(0);
461   if (L.contains(LoopExitBB)) {
462     ExitDirection = false;
463     LoopExitSuccIdx = 1;
464     LoopExitBB = BI.getSuccessor(1);
465     if (L.contains(LoopExitBB)) {
466       LLVM_DEBUG(dbgs() << "   Branch doesn't exit the loop!\n");
467       return false;
468     }
469   }
470   auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
471   auto *ParentBB = BI.getParent();
472   if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) {
473     LLVM_DEBUG(dbgs() << "   Loop exit PHI's aren't loop-invariant!\n");
474     return false;
475   }
476 
477   // When unswitching only part of the branch's condition, we need the exit
478   // block to be reached directly from the partially unswitched input. This can
479   // be done when the exit block is along the true edge and the branch condition
480   // is a graph of `or` operations, or the exit block is along the false edge
481   // and the condition is a graph of `and` operations.
482   if (!FullUnswitch) {
483     if (ExitDirection ? !match(BI.getCondition(), m_LogicalOr())
484                       : !match(BI.getCondition(), m_LogicalAnd())) {
485       LLVM_DEBUG(dbgs() << "   Branch condition is in improper form for "
486                            "non-full unswitch!\n");
487       return false;
488     }
489   }
490 
491   LLVM_DEBUG({
492     dbgs() << "    unswitching trivial invariant conditions for: " << BI
493            << "\n";
494     for (Value *Invariant : Invariants) {
495       dbgs() << "      " << *Invariant << " == true";
496       if (Invariant != Invariants.back())
497         dbgs() << " ||";
498       dbgs() << "\n";
499     }
500   });
501 
502   // If we have scalar evolutions, we need to invalidate them including this
503   // loop, the loop containing the exit block and the topmost parent loop
504   // exiting via LoopExitBB.
505   if (SE) {
506     if (Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI))
507       SE->forgetLoop(ExitL);
508     else
509       // Forget the entire nest as this exits the entire nest.
510       SE->forgetTopmostLoop(&L);
511   }
512 
513   if (MSSAU && VerifyMemorySSA)
514     MSSAU->getMemorySSA()->verifyMemorySSA();
515 
516   // Split the preheader, so that we know that there is a safe place to insert
517   // the conditional branch. We will change the preheader to have a conditional
518   // branch on LoopCond.
519   BasicBlock *OldPH = L.getLoopPreheader();
520   BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
521 
522   // Now that we have a place to insert the conditional branch, create a place
523   // to branch to: this is the exit block out of the loop that we are
524   // unswitching. We need to split this if there are other loop predecessors.
525   // Because the loop is in simplified form, *any* other predecessor is enough.
526   BasicBlock *UnswitchedBB;
527   if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
528     assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
529            "A branch's parent isn't a predecessor!");
530     UnswitchedBB = LoopExitBB;
531   } else {
532     UnswitchedBB =
533         SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI, MSSAU);
534   }
535 
536   if (MSSAU && VerifyMemorySSA)
537     MSSAU->getMemorySSA()->verifyMemorySSA();
538 
539   // Actually move the invariant uses into the unswitched position. If possible,
540   // we do this by moving the instructions, but when doing partial unswitching
541   // we do it by building a new merge of the values in the unswitched position.
542   OldPH->getTerminator()->eraseFromParent();
543   if (FullUnswitch) {
544     // If fully unswitching, we can use the existing branch instruction.
545     // Splice it into the old PH to gate reaching the new preheader and re-point
546     // its successors.
547     OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(),
548                                 BI);
549     if (MSSAU) {
550       // Temporarily clone the terminator, to make MSSA update cheaper by
551       // separating "insert edge" updates from "remove edge" ones.
552       ParentBB->getInstList().push_back(BI.clone());
553     } else {
554       // Create a new unconditional branch that will continue the loop as a new
555       // terminator.
556       BranchInst::Create(ContinueBB, ParentBB);
557     }
558     BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
559     BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
560   } else {
561     // Only unswitching a subset of inputs to the condition, so we will need to
562     // build a new branch that merges the invariant inputs.
563     if (ExitDirection)
564       assert(match(BI.getCondition(), m_LogicalOr()) &&
565              "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the "
566              "condition!");
567     else
568       assert(match(BI.getCondition(), m_LogicalAnd()) &&
569              "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the"
570              " condition!");
571     buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection,
572                                           *UnswitchedBB, *NewPH, false);
573   }
574 
575   // Update the dominator tree with the added edge.
576   DT.insertEdge(OldPH, UnswitchedBB);
577 
578   // After the dominator tree was updated with the added edge, update MemorySSA
579   // if available.
580   if (MSSAU) {
581     SmallVector<CFGUpdate, 1> Updates;
582     Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
583     MSSAU->applyInsertUpdates(Updates, DT);
584   }
585 
586   // Finish updating dominator tree and memory ssa for full unswitch.
587   if (FullUnswitch) {
588     if (MSSAU) {
589       // Remove the cloned branch instruction.
590       ParentBB->getTerminator()->eraseFromParent();
591       // Create unconditional branch now.
592       BranchInst::Create(ContinueBB, ParentBB);
593       MSSAU->removeEdge(ParentBB, LoopExitBB);
594     }
595     DT.deleteEdge(ParentBB, LoopExitBB);
596   }
597 
598   if (MSSAU && VerifyMemorySSA)
599     MSSAU->getMemorySSA()->verifyMemorySSA();
600 
601   // Rewrite the relevant PHI nodes.
602   if (UnswitchedBB == LoopExitBB)
603     rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
604   else
605     rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
606                                               *ParentBB, *OldPH, FullUnswitch);
607 
608   // The constant we can replace all of our invariants with inside the loop
609   // body. If any of the invariants have a value other than this the loop won't
610   // be entered.
611   ConstantInt *Replacement = ExitDirection
612                                  ? ConstantInt::getFalse(BI.getContext())
613                                  : ConstantInt::getTrue(BI.getContext());
614 
615   // Since this is an i1 condition we can also trivially replace uses of it
616   // within the loop with a constant.
617   for (Value *Invariant : Invariants)
618     replaceLoopInvariantUses(L, Invariant, *Replacement);
619 
620   // If this was full unswitching, we may have changed the nesting relationship
621   // for this loop so hoist it to its correct parent if needed.
622   if (FullUnswitch)
623     hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
624 
625   if (MSSAU && VerifyMemorySSA)
626     MSSAU->getMemorySSA()->verifyMemorySSA();
627 
628   LLVM_DEBUG(dbgs() << "    done: unswitching trivial branch...\n");
629   ++NumTrivial;
630   ++NumBranches;
631   return true;
632 }
633 
634 /// Unswitch a trivial switch if the condition is loop invariant.
635 ///
636 /// This routine should only be called when loop code leading to the switch has
637 /// been validated as trivial (no side effects). This routine checks if the
638 /// condition is invariant and that at least one of the successors is a loop
639 /// exit. This allows us to unswitch without duplicating the loop, making it
640 /// trivial.
641 ///
642 /// If this routine fails to unswitch the switch it returns false.
643 ///
644 /// If the switch can be unswitched, this routine splits the preheader and
645 /// copies the switch above that split. If the default case is one of the
646 /// exiting cases, it copies the non-exiting cases and points them at the new
647 /// preheader. If the default case is not exiting, it copies the exiting cases
648 /// and points the default at the preheader. It preserves loop simplified form
649 /// (splitting the exit blocks as necessary). It simplifies the switch within
650 /// the loop by removing now-dead cases. If the default case is one of those
651 /// unswitched, it replaces its destination with a new basic block containing
652 /// only unreachable. Such basic blocks, while technically loop exits, are not
653 /// considered for unswitching so this is a stable transform and the same
654 /// switch will not be revisited. If after unswitching there is only a single
655 /// in-loop successor, the switch is further simplified to an unconditional
656 /// branch. Still more cleanup can be done with some simplifycfg like pass.
657 ///
658 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
659 /// invalidated by this.
660 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
661                                   LoopInfo &LI, ScalarEvolution *SE,
662                                   MemorySSAUpdater *MSSAU) {
663   LLVM_DEBUG(dbgs() << "  Trying to unswitch switch: " << SI << "\n");
664   Value *LoopCond = SI.getCondition();
665 
666   // If this isn't switching on an invariant condition, we can't unswitch it.
667   if (!L.isLoopInvariant(LoopCond))
668     return false;
669 
670   auto *ParentBB = SI.getParent();
671 
672   // The same check must be used both for the default and the exit cases. We
673   // should never leave edges from the switch instruction to a basic block that
674   // we are unswitching, hence the condition used to determine the default case
675   // needs to also be used to populate ExitCaseIndices, which is then used to
676   // remove cases from the switch.
677   auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) {
678     // BBToCheck is not an exit block if it is inside loop L.
679     if (L.contains(&BBToCheck))
680       return false;
681     // BBToCheck is not trivial to unswitch if its phis aren't loop invariant.
682     if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck))
683       return false;
684     // We do not unswitch a block that only has an unreachable statement, as
685     // it's possible this is a previously unswitched block. Only unswitch if
686     // either the terminator is not unreachable, or, if it is, it's not the only
687     // instruction in the block.
688     auto *TI = BBToCheck.getTerminator();
689     bool isUnreachable = isa<UnreachableInst>(TI);
690     return !isUnreachable ||
691            (isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI));
692   };
693 
694   SmallVector<int, 4> ExitCaseIndices;
695   for (auto Case : SI.cases())
696     if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor()))
697       ExitCaseIndices.push_back(Case.getCaseIndex());
698   BasicBlock *DefaultExitBB = nullptr;
699   SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight =
700       SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0);
701   if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) {
702     DefaultExitBB = SI.getDefaultDest();
703   } else if (ExitCaseIndices.empty())
704     return false;
705 
706   LLVM_DEBUG(dbgs() << "    unswitching trivial switch...\n");
707 
708   if (MSSAU && VerifyMemorySSA)
709     MSSAU->getMemorySSA()->verifyMemorySSA();
710 
711   // We may need to invalidate SCEVs for the outermost loop reached by any of
712   // the exits.
713   Loop *OuterL = &L;
714 
715   if (DefaultExitBB) {
716     // Clear out the default destination temporarily to allow accurate
717     // predecessor lists to be examined below.
718     SI.setDefaultDest(nullptr);
719     // Check the loop containing this exit.
720     Loop *ExitL = LI.getLoopFor(DefaultExitBB);
721     if (!ExitL || ExitL->contains(OuterL))
722       OuterL = ExitL;
723   }
724 
725   // Store the exit cases into a separate data structure and remove them from
726   // the switch.
727   SmallVector<std::tuple<ConstantInt *, BasicBlock *,
728                          SwitchInstProfUpdateWrapper::CaseWeightOpt>,
729               4> ExitCases;
730   ExitCases.reserve(ExitCaseIndices.size());
731   SwitchInstProfUpdateWrapper SIW(SI);
732   // We walk the case indices backwards so that we remove the last case first
733   // and don't disrupt the earlier indices.
734   for (unsigned Index : reverse(ExitCaseIndices)) {
735     auto CaseI = SI.case_begin() + Index;
736     // Compute the outer loop from this exit.
737     Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor());
738     if (!ExitL || ExitL->contains(OuterL))
739       OuterL = ExitL;
740     // Save the value of this case.
741     auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex());
742     ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W);
743     // Delete the unswitched cases.
744     SIW.removeCase(CaseI);
745   }
746 
747   if (SE) {
748     if (OuterL)
749       SE->forgetLoop(OuterL);
750     else
751       SE->forgetTopmostLoop(&L);
752   }
753 
754   // Check if after this all of the remaining cases point at the same
755   // successor.
756   BasicBlock *CommonSuccBB = nullptr;
757   if (SI.getNumCases() > 0 &&
758       all_of(drop_begin(SI.cases()), [&SI](const SwitchInst::CaseHandle &Case) {
759         return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor();
760       }))
761     CommonSuccBB = SI.case_begin()->getCaseSuccessor();
762   if (!DefaultExitBB) {
763     // If we're not unswitching the default, we need it to match any cases to
764     // have a common successor or if we have no cases it is the common
765     // successor.
766     if (SI.getNumCases() == 0)
767       CommonSuccBB = SI.getDefaultDest();
768     else if (SI.getDefaultDest() != CommonSuccBB)
769       CommonSuccBB = nullptr;
770   }
771 
772   // Split the preheader, so that we know that there is a safe place to insert
773   // the switch.
774   BasicBlock *OldPH = L.getLoopPreheader();
775   BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
776   OldPH->getTerminator()->eraseFromParent();
777 
778   // Now add the unswitched switch.
779   auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
780   SwitchInstProfUpdateWrapper NewSIW(*NewSI);
781 
782   // Rewrite the IR for the unswitched basic blocks. This requires two steps.
783   // First, we split any exit blocks with remaining in-loop predecessors. Then
784   // we update the PHIs in one of two ways depending on if there was a split.
785   // We walk in reverse so that we split in the same order as the cases
786   // appeared. This is purely for convenience of reading the resulting IR, but
787   // it doesn't cost anything really.
788   SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
789   SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
790   // Handle the default exit if necessary.
791   // FIXME: It'd be great if we could merge this with the loop below but LLVM's
792   // ranges aren't quite powerful enough yet.
793   if (DefaultExitBB) {
794     if (pred_empty(DefaultExitBB)) {
795       UnswitchedExitBBs.insert(DefaultExitBB);
796       rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
797     } else {
798       auto *SplitBB =
799           SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI, MSSAU);
800       rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,
801                                                 *ParentBB, *OldPH,
802                                                 /*FullUnswitch*/ true);
803       DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
804     }
805   }
806   // Note that we must use a reference in the for loop so that we update the
807   // container.
808   for (auto &ExitCase : reverse(ExitCases)) {
809     // Grab a reference to the exit block in the pair so that we can update it.
810     BasicBlock *ExitBB = std::get<1>(ExitCase);
811 
812     // If this case is the last edge into the exit block, we can simply reuse it
813     // as it will no longer be a loop exit. No mapping necessary.
814     if (pred_empty(ExitBB)) {
815       // Only rewrite once.
816       if (UnswitchedExitBBs.insert(ExitBB).second)
817         rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
818       continue;
819     }
820 
821     // Otherwise we need to split the exit block so that we retain an exit
822     // block from the loop and a target for the unswitched condition.
823     BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
824     if (!SplitExitBB) {
825       // If this is the first time we see this, do the split and remember it.
826       SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
827       rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,
828                                                 *ParentBB, *OldPH,
829                                                 /*FullUnswitch*/ true);
830     }
831     // Update the case pair to point to the split block.
832     std::get<1>(ExitCase) = SplitExitBB;
833   }
834 
835   // Now add the unswitched cases. We do this in reverse order as we built them
836   // in reverse order.
837   for (auto &ExitCase : reverse(ExitCases)) {
838     ConstantInt *CaseVal = std::get<0>(ExitCase);
839     BasicBlock *UnswitchedBB = std::get<1>(ExitCase);
840 
841     NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase));
842   }
843 
844   // If the default was unswitched, re-point it and add explicit cases for
845   // entering the loop.
846   if (DefaultExitBB) {
847     NewSIW->setDefaultDest(DefaultExitBB);
848     NewSIW.setSuccessorWeight(0, DefaultCaseWeight);
849 
850     // We removed all the exit cases, so we just copy the cases to the
851     // unswitched switch.
852     for (const auto &Case : SI.cases())
853       NewSIW.addCase(Case.getCaseValue(), NewPH,
854                      SIW.getSuccessorWeight(Case.getSuccessorIndex()));
855   } else if (DefaultCaseWeight) {
856     // We have to set branch weight of the default case.
857     uint64_t SW = *DefaultCaseWeight;
858     for (const auto &Case : SI.cases()) {
859       auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex());
860       assert(W &&
861              "case weight must be defined as default case weight is defined");
862       SW += *W;
863     }
864     NewSIW.setSuccessorWeight(0, SW);
865   }
866 
867   // If we ended up with a common successor for every path through the switch
868   // after unswitching, rewrite it to an unconditional branch to make it easy
869   // to recognize. Otherwise we potentially have to recognize the default case
870   // pointing at unreachable and other complexity.
871   if (CommonSuccBB) {
872     BasicBlock *BB = SI.getParent();
873     // We may have had multiple edges to this common successor block, so remove
874     // them as predecessors. We skip the first one, either the default or the
875     // actual first case.
876     bool SkippedFirst = DefaultExitBB == nullptr;
877     for (auto Case : SI.cases()) {
878       assert(Case.getCaseSuccessor() == CommonSuccBB &&
879              "Non-common successor!");
880       (void)Case;
881       if (!SkippedFirst) {
882         SkippedFirst = true;
883         continue;
884       }
885       CommonSuccBB->removePredecessor(BB,
886                                       /*KeepOneInputPHIs*/ true);
887     }
888     // Now nuke the switch and replace it with a direct branch.
889     SIW.eraseFromParent();
890     BranchInst::Create(CommonSuccBB, BB);
891   } else if (DefaultExitBB) {
892     assert(SI.getNumCases() > 0 &&
893            "If we had no cases we'd have a common successor!");
894     // Move the last case to the default successor. This is valid as if the
895     // default got unswitched it cannot be reached. This has the advantage of
896     // being simple and keeping the number of edges from this switch to
897     // successors the same, and avoiding any PHI update complexity.
898     auto LastCaseI = std::prev(SI.case_end());
899 
900     SI.setDefaultDest(LastCaseI->getCaseSuccessor());
901     SIW.setSuccessorWeight(
902         0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex()));
903     SIW.removeCase(LastCaseI);
904   }
905 
906   // Walk the unswitched exit blocks and the unswitched split blocks and update
907   // the dominator tree based on the CFG edits. While we are walking unordered
908   // containers here, the API for applyUpdates takes an unordered list of
909   // updates and requires them to not contain duplicates.
910   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
911   for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
912     DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});
913     DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});
914   }
915   for (auto SplitUnswitchedPair : SplitExitBBMap) {
916     DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first});
917     DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second});
918   }
919 
920   if (MSSAU) {
921     MSSAU->applyUpdates(DTUpdates, DT, /*UpdateDT=*/true);
922     if (VerifyMemorySSA)
923       MSSAU->getMemorySSA()->verifyMemorySSA();
924   } else {
925     DT.applyUpdates(DTUpdates);
926   }
927 
928   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
929 
930   // We may have changed the nesting relationship for this loop so hoist it to
931   // its correct parent if needed.
932   hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
933 
934   if (MSSAU && VerifyMemorySSA)
935     MSSAU->getMemorySSA()->verifyMemorySSA();
936 
937   ++NumTrivial;
938   ++NumSwitches;
939   LLVM_DEBUG(dbgs() << "    done: unswitching trivial switch...\n");
940   return true;
941 }
942 
943 /// This routine scans the loop to find a branch or switch which occurs before
944 /// any side effects occur. These can potentially be unswitched without
945 /// duplicating the loop. If a branch or switch is successfully unswitched the
946 /// scanning continues to see if subsequent branches or switches have become
947 /// trivial. Once all trivial candidates have been unswitched, this routine
948 /// returns.
949 ///
950 /// The return value indicates whether anything was unswitched (and therefore
951 /// changed).
952 ///
953 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
954 /// invalidated by this.
955 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
956                                          LoopInfo &LI, ScalarEvolution *SE,
957                                          MemorySSAUpdater *MSSAU) {
958   bool Changed = false;
959 
960   // If loop header has only one reachable successor we should keep looking for
961   // trivial condition candidates in the successor as well. An alternative is
962   // to constant fold conditions and merge successors into loop header (then we
963   // only need to check header's terminator). The reason for not doing this in
964   // LoopUnswitch pass is that it could potentially break LoopPassManager's
965   // invariants. Folding dead branches could either eliminate the current loop
966   // or make other loops unreachable. LCSSA form might also not be preserved
967   // after deleting branches. The following code keeps traversing loop header's
968   // successors until it finds the trivial condition candidate (condition that
969   // is not a constant). Since unswitching generates branches with constant
970   // conditions, this scenario could be very common in practice.
971   BasicBlock *CurrentBB = L.getHeader();
972   SmallPtrSet<BasicBlock *, 8> Visited;
973   Visited.insert(CurrentBB);
974   do {
975     // Check if there are any side-effecting instructions (e.g. stores, calls,
976     // volatile loads) in the part of the loop that the code *would* execute
977     // without unswitching.
978     if (MSSAU) // Possible early exit with MSSA
979       if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB))
980         if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end()))
981           return Changed;
982     if (llvm::any_of(*CurrentBB,
983                      [](Instruction &I) { return I.mayHaveSideEffects(); }))
984       return Changed;
985 
986     Instruction *CurrentTerm = CurrentBB->getTerminator();
987 
988     if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
989       // Don't bother trying to unswitch past a switch with a constant
990       // condition. This should be removed prior to running this pass by
991       // simplifycfg.
992       if (isa<Constant>(SI->getCondition()))
993         return Changed;
994 
995       if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU))
996         // Couldn't unswitch this one so we're done.
997         return Changed;
998 
999       // Mark that we managed to unswitch something.
1000       Changed = true;
1001 
1002       // If unswitching turned the terminator into an unconditional branch then
1003       // we can continue. The unswitching logic specifically works to fold any
1004       // cases it can into an unconditional branch to make it easier to
1005       // recognize here.
1006       auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
1007       if (!BI || BI->isConditional())
1008         return Changed;
1009 
1010       CurrentBB = BI->getSuccessor(0);
1011       continue;
1012     }
1013 
1014     auto *BI = dyn_cast<BranchInst>(CurrentTerm);
1015     if (!BI)
1016       // We do not understand other terminator instructions.
1017       return Changed;
1018 
1019     // Don't bother trying to unswitch past an unconditional branch or a branch
1020     // with a constant value. These should be removed by simplifycfg prior to
1021     // running this pass.
1022     if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1023       return Changed;
1024 
1025     // Found a trivial condition candidate: non-foldable conditional branch. If
1026     // we fail to unswitch this, we can't do anything else that is trivial.
1027     if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU))
1028       return Changed;
1029 
1030     // Mark that we managed to unswitch something.
1031     Changed = true;
1032 
1033     // If we only unswitched some of the conditions feeding the branch, we won't
1034     // have collapsed it to a single successor.
1035     BI = cast<BranchInst>(CurrentBB->getTerminator());
1036     if (BI->isConditional())
1037       return Changed;
1038 
1039     // Follow the newly unconditional branch into its successor.
1040     CurrentBB = BI->getSuccessor(0);
1041 
1042     // When continuing, if we exit the loop or reach a previous visited block,
1043     // then we can not reach any trivial condition candidates (unfoldable
1044     // branch instructions or switch instructions) and no unswitch can happen.
1045   } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
1046 
1047   return Changed;
1048 }
1049 
1050 /// Build the cloned blocks for an unswitched copy of the given loop.
1051 ///
1052 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
1053 /// after the split block (`SplitBB`) that will be used to select between the
1054 /// cloned and original loop.
1055 ///
1056 /// This routine handles cloning all of the necessary loop blocks and exit
1057 /// blocks including rewriting their instructions and the relevant PHI nodes.
1058 /// Any loop blocks or exit blocks which are dominated by a different successor
1059 /// than the one for this clone of the loop blocks can be trivially skipped. We
1060 /// use the `DominatingSucc` map to determine whether a block satisfies that
1061 /// property with a simple map lookup.
1062 ///
1063 /// It also correctly creates the unconditional branch in the cloned
1064 /// unswitched parent block to only point at the unswitched successor.
1065 ///
1066 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit
1067 /// block splitting is correctly reflected in `LoopInfo`, essentially all of
1068 /// the cloned blocks (and their loops) are left without full `LoopInfo`
1069 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
1070 /// blocks to them but doesn't create the cloned `DominatorTree` structure and
1071 /// instead the caller must recompute an accurate DT. It *does* correctly
1072 /// update the `AssumptionCache` provided in `AC`.
1073 static BasicBlock *buildClonedLoopBlocks(
1074     Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
1075     ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
1076     BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
1077     const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
1078     ValueToValueMapTy &VMap,
1079     SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
1080     DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
1081   SmallVector<BasicBlock *, 4> NewBlocks;
1082   NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());
1083 
1084   // We will need to clone a bunch of blocks, wrap up the clone operation in
1085   // a helper.
1086   auto CloneBlock = [&](BasicBlock *OldBB) {
1087     // Clone the basic block and insert it before the new preheader.
1088     BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());
1089     NewBB->moveBefore(LoopPH);
1090 
1091     // Record this block and the mapping.
1092     NewBlocks.push_back(NewBB);
1093     VMap[OldBB] = NewBB;
1094 
1095     return NewBB;
1096   };
1097 
1098   // We skip cloning blocks when they have a dominating succ that is not the
1099   // succ we are cloning for.
1100   auto SkipBlock = [&](BasicBlock *BB) {
1101     auto It = DominatingSucc.find(BB);
1102     return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
1103   };
1104 
1105   // First, clone the preheader.
1106   auto *ClonedPH = CloneBlock(LoopPH);
1107 
1108   // Then clone all the loop blocks, skipping the ones that aren't necessary.
1109   for (auto *LoopBB : L.blocks())
1110     if (!SkipBlock(LoopBB))
1111       CloneBlock(LoopBB);
1112 
1113   // Split all the loop exit edges so that when we clone the exit blocks, if
1114   // any of the exit blocks are *also* a preheader for some other loop, we
1115   // don't create multiple predecessors entering the loop header.
1116   for (auto *ExitBB : ExitBlocks) {
1117     if (SkipBlock(ExitBB))
1118       continue;
1119 
1120     // When we are going to clone an exit, we don't need to clone all the
1121     // instructions in the exit block and we want to ensure we have an easy
1122     // place to merge the CFG, so split the exit first. This is always safe to
1123     // do because there cannot be any non-loop predecessors of a loop exit in
1124     // loop simplified form.
1125     auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU);
1126 
1127     // Rearrange the names to make it easier to write test cases by having the
1128     // exit block carry the suffix rather than the merge block carrying the
1129     // suffix.
1130     MergeBB->takeName(ExitBB);
1131     ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1132 
1133     // Now clone the original exit block.
1134     auto *ClonedExitBB = CloneBlock(ExitBB);
1135     assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1136            "Exit block should have been split to have one successor!");
1137     assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1138            "Cloned exit block has the wrong successor!");
1139 
1140     // Remap any cloned instructions and create a merge phi node for them.
1141     for (auto ZippedInsts : llvm::zip_first(
1142              llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),
1143              llvm::make_range(ClonedExitBB->begin(),
1144                               std::prev(ClonedExitBB->end())))) {
1145       Instruction &I = std::get<0>(ZippedInsts);
1146       Instruction &ClonedI = std::get<1>(ZippedInsts);
1147 
1148       // The only instructions in the exit block should be PHI nodes and
1149       // potentially a landing pad.
1150       assert(
1151           (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1152           "Bad instruction in exit block!");
1153       // We should have a value map between the instruction and its clone.
1154       assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1155 
1156       auto *MergePN =
1157           PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi",
1158                           &*MergeBB->getFirstInsertionPt());
1159       I.replaceAllUsesWith(MergePN);
1160       MergePN->addIncoming(&I, ExitBB);
1161       MergePN->addIncoming(&ClonedI, ClonedExitBB);
1162     }
1163   }
1164 
1165   // Rewrite the instructions in the cloned blocks to refer to the instructions
1166   // in the cloned blocks. We have to do this as a second pass so that we have
1167   // everything available. Also, we have inserted new instructions which may
1168   // include assume intrinsics, so we update the assumption cache while
1169   // processing this.
1170   for (auto *ClonedBB : NewBlocks)
1171     for (Instruction &I : *ClonedBB) {
1172       RemapInstruction(&I, VMap,
1173                        RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1174       if (auto *II = dyn_cast<AssumeInst>(&I))
1175         AC.registerAssumption(II);
1176     }
1177 
1178   // Update any PHI nodes in the cloned successors of the skipped blocks to not
1179   // have spurious incoming values.
1180   for (auto *LoopBB : L.blocks())
1181     if (SkipBlock(LoopBB))
1182       for (auto *SuccBB : successors(LoopBB))
1183         if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))
1184           for (PHINode &PN : ClonedSuccBB->phis())
1185             PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);
1186 
1187   // Remove the cloned parent as a predecessor of any successor we ended up
1188   // cloning other than the unswitched one.
1189   auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));
1190   for (auto *SuccBB : successors(ParentBB)) {
1191     if (SuccBB == UnswitchedSuccBB)
1192       continue;
1193 
1194     auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));
1195     if (!ClonedSuccBB)
1196       continue;
1197 
1198     ClonedSuccBB->removePredecessor(ClonedParentBB,
1199                                     /*KeepOneInputPHIs*/ true);
1200   }
1201 
1202   // Replace the cloned branch with an unconditional branch to the cloned
1203   // unswitched successor.
1204   auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));
1205   Instruction *ClonedTerminator = ClonedParentBB->getTerminator();
1206   // Trivial Simplification. If Terminator is a conditional branch and
1207   // condition becomes dead - erase it.
1208   Value *ClonedConditionToErase = nullptr;
1209   if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator))
1210     ClonedConditionToErase = BI->getCondition();
1211   else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator))
1212     ClonedConditionToErase = SI->getCondition();
1213 
1214   ClonedTerminator->eraseFromParent();
1215   BranchInst::Create(ClonedSuccBB, ClonedParentBB);
1216 
1217   if (ClonedConditionToErase)
1218     RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr,
1219                                                MSSAU);
1220 
1221   // If there are duplicate entries in the PHI nodes because of multiple edges
1222   // to the unswitched successor, we need to nuke all but one as we replaced it
1223   // with a direct branch.
1224   for (PHINode &PN : ClonedSuccBB->phis()) {
1225     bool Found = false;
1226     // Loop over the incoming operands backwards so we can easily delete as we
1227     // go without invalidating the index.
1228     for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1229       if (PN.getIncomingBlock(i) != ClonedParentBB)
1230         continue;
1231       if (!Found) {
1232         Found = true;
1233         continue;
1234       }
1235       PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);
1236     }
1237   }
1238 
1239   // Record the domtree updates for the new blocks.
1240   SmallPtrSet<BasicBlock *, 4> SuccSet;
1241   for (auto *ClonedBB : NewBlocks) {
1242     for (auto *SuccBB : successors(ClonedBB))
1243       if (SuccSet.insert(SuccBB).second)
1244         DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});
1245     SuccSet.clear();
1246   }
1247 
1248   return ClonedPH;
1249 }
1250 
1251 /// Recursively clone the specified loop and all of its children.
1252 ///
1253 /// The target parent loop for the clone should be provided, or can be null if
1254 /// the clone is a top-level loop. While cloning, all the blocks are mapped
1255 /// with the provided value map. The entire original loop must be present in
1256 /// the value map. The cloned loop is returned.
1257 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1258                            const ValueToValueMapTy &VMap, LoopInfo &LI) {
1259   auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1260     assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1261     ClonedL.reserveBlocks(OrigL.getNumBlocks());
1262     for (auto *BB : OrigL.blocks()) {
1263       auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));
1264       ClonedL.addBlockEntry(ClonedBB);
1265       if (LI.getLoopFor(BB) == &OrigL)
1266         LI.changeLoopFor(ClonedBB, &ClonedL);
1267     }
1268   };
1269 
1270   // We specially handle the first loop because it may get cloned into
1271   // a different parent and because we most commonly are cloning leaf loops.
1272   Loop *ClonedRootL = LI.AllocateLoop();
1273   if (RootParentL)
1274     RootParentL->addChildLoop(ClonedRootL);
1275   else
1276     LI.addTopLevelLoop(ClonedRootL);
1277   AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1278 
1279   if (OrigRootL.isInnermost())
1280     return ClonedRootL;
1281 
1282   // If we have a nest, we can quickly clone the entire loop nest using an
1283   // iterative approach because it is a tree. We keep the cloned parent in the
1284   // data structure to avoid repeatedly querying through a map to find it.
1285   SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1286   // Build up the loops to clone in reverse order as we'll clone them from the
1287   // back.
1288   for (Loop *ChildL : llvm::reverse(OrigRootL))
1289     LoopsToClone.push_back({ClonedRootL, ChildL});
1290   do {
1291     Loop *ClonedParentL, *L;
1292     std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();
1293     Loop *ClonedL = LI.AllocateLoop();
1294     ClonedParentL->addChildLoop(ClonedL);
1295     AddClonedBlocksToLoop(*L, *ClonedL);
1296     for (Loop *ChildL : llvm::reverse(*L))
1297       LoopsToClone.push_back({ClonedL, ChildL});
1298   } while (!LoopsToClone.empty());
1299 
1300   return ClonedRootL;
1301 }
1302 
1303 /// Build the cloned loops of an original loop from unswitching.
1304 ///
1305 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1306 /// operation. We need to re-verify that there even is a loop (as the backedge
1307 /// may not have been cloned), and even if there are remaining backedges the
1308 /// backedge set may be different. However, we know that each child loop is
1309 /// undisturbed, we only need to find where to place each child loop within
1310 /// either any parent loop or within a cloned version of the original loop.
1311 ///
1312 /// Because child loops may end up cloned outside of any cloned version of the
1313 /// original loop, multiple cloned sibling loops may be created. All of them
1314 /// are returned so that the newly introduced loop nest roots can be
1315 /// identified.
1316 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1317                              const ValueToValueMapTy &VMap, LoopInfo &LI,
1318                              SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1319   Loop *ClonedL = nullptr;
1320 
1321   auto *OrigPH = OrigL.getLoopPreheader();
1322   auto *OrigHeader = OrigL.getHeader();
1323 
1324   auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));
1325   auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));
1326 
1327   // We need to know the loops of the cloned exit blocks to even compute the
1328   // accurate parent loop. If we only clone exits to some parent of the
1329   // original parent, we want to clone into that outer loop. We also keep track
1330   // of the loops that our cloned exit blocks participate in.
1331   Loop *ParentL = nullptr;
1332   SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1333   SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1334   ClonedExitsInLoops.reserve(ExitBlocks.size());
1335   for (auto *ExitBB : ExitBlocks)
1336     if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))
1337       if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1338         ExitLoopMap[ClonedExitBB] = ExitL;
1339         ClonedExitsInLoops.push_back(ClonedExitBB);
1340         if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1341           ParentL = ExitL;
1342       }
1343   assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1344           ParentL->contains(OrigL.getParentLoop())) &&
1345          "The computed parent loop should always contain (or be) the parent of "
1346          "the original loop.");
1347 
1348   // We build the set of blocks dominated by the cloned header from the set of
1349   // cloned blocks out of the original loop. While not all of these will
1350   // necessarily be in the cloned loop, it is enough to establish that they
1351   // aren't in unreachable cycles, etc.
1352   SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1353   for (auto *BB : OrigL.blocks())
1354     if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))
1355       ClonedLoopBlocks.insert(ClonedBB);
1356 
1357   // Rebuild the set of blocks that will end up in the cloned loop. We may have
1358   // skipped cloning some region of this loop which can in turn skip some of
1359   // the backedges so we have to rebuild the blocks in the loop based on the
1360   // backedges that remain after cloning.
1361   SmallVector<BasicBlock *, 16> Worklist;
1362   SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1363   for (auto *Pred : predecessors(ClonedHeader)) {
1364     // The only possible non-loop header predecessor is the preheader because
1365     // we know we cloned the loop in simplified form.
1366     if (Pred == ClonedPH)
1367       continue;
1368 
1369     // Because the loop was in simplified form, the only non-loop predecessor
1370     // should be the preheader.
1371     assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1372                                            "header other than the preheader "
1373                                            "that is not part of the loop!");
1374 
1375     // Insert this block into the loop set and on the first visit (and if it
1376     // isn't the header we're currently walking) put it into the worklist to
1377     // recurse through.
1378     if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)
1379       Worklist.push_back(Pred);
1380   }
1381 
1382   // If we had any backedges then there *is* a cloned loop. Put the header into
1383   // the loop set and then walk the worklist backwards to find all the blocks
1384   // that remain within the loop after cloning.
1385   if (!BlocksInClonedLoop.empty()) {
1386     BlocksInClonedLoop.insert(ClonedHeader);
1387 
1388     while (!Worklist.empty()) {
1389       BasicBlock *BB = Worklist.pop_back_val();
1390       assert(BlocksInClonedLoop.count(BB) &&
1391              "Didn't put block into the loop set!");
1392 
1393       // Insert any predecessors that are in the possible set into the cloned
1394       // set, and if the insert is successful, add them to the worklist. Note
1395       // that we filter on the blocks that are definitely reachable via the
1396       // backedge to the loop header so we may prune out dead code within the
1397       // cloned loop.
1398       for (auto *Pred : predecessors(BB))
1399         if (ClonedLoopBlocks.count(Pred) &&
1400             BlocksInClonedLoop.insert(Pred).second)
1401           Worklist.push_back(Pred);
1402     }
1403 
1404     ClonedL = LI.AllocateLoop();
1405     if (ParentL) {
1406       ParentL->addBasicBlockToLoop(ClonedPH, LI);
1407       ParentL->addChildLoop(ClonedL);
1408     } else {
1409       LI.addTopLevelLoop(ClonedL);
1410     }
1411     NonChildClonedLoops.push_back(ClonedL);
1412 
1413     ClonedL->reserveBlocks(BlocksInClonedLoop.size());
1414     // We don't want to just add the cloned loop blocks based on how we
1415     // discovered them. The original order of blocks was carefully built in
1416     // a way that doesn't rely on predecessor ordering. Rather than re-invent
1417     // that logic, we just re-walk the original blocks (and those of the child
1418     // loops) and filter them as we add them into the cloned loop.
1419     for (auto *BB : OrigL.blocks()) {
1420       auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));
1421       if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))
1422         continue;
1423 
1424       // Directly add the blocks that are only in this loop.
1425       if (LI.getLoopFor(BB) == &OrigL) {
1426         ClonedL->addBasicBlockToLoop(ClonedBB, LI);
1427         continue;
1428       }
1429 
1430       // We want to manually add it to this loop and parents.
1431       // Registering it with LoopInfo will happen when we clone the top
1432       // loop for this block.
1433       for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1434         PL->addBlockEntry(ClonedBB);
1435     }
1436 
1437     // Now add each child loop whose header remains within the cloned loop. All
1438     // of the blocks within the loop must satisfy the same constraints as the
1439     // header so once we pass the header checks we can just clone the entire
1440     // child loop nest.
1441     for (Loop *ChildL : OrigL) {
1442       auto *ClonedChildHeader =
1443           cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1444       if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))
1445         continue;
1446 
1447 #ifndef NDEBUG
1448       // We should never have a cloned child loop header but fail to have
1449       // all of the blocks for that child loop.
1450       for (auto *ChildLoopBB : ChildL->blocks())
1451         assert(BlocksInClonedLoop.count(
1452                    cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1453                "Child cloned loop has a header within the cloned outer "
1454                "loop but not all of its blocks!");
1455 #endif
1456 
1457       cloneLoopNest(*ChildL, ClonedL, VMap, LI);
1458     }
1459   }
1460 
1461   // Now that we've handled all the components of the original loop that were
1462   // cloned into a new loop, we still need to handle anything from the original
1463   // loop that wasn't in a cloned loop.
1464 
1465   // Figure out what blocks are left to place within any loop nest containing
1466   // the unswitched loop. If we never formed a loop, the cloned PH is one of
1467   // them.
1468   SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1469   if (BlocksInClonedLoop.empty())
1470     UnloopedBlockSet.insert(ClonedPH);
1471   for (auto *ClonedBB : ClonedLoopBlocks)
1472     if (!BlocksInClonedLoop.count(ClonedBB))
1473       UnloopedBlockSet.insert(ClonedBB);
1474 
1475   // Copy the cloned exits and sort them in ascending loop depth, we'll work
1476   // backwards across these to process them inside out. The order shouldn't
1477   // matter as we're just trying to build up the map from inside-out; we use
1478   // the map in a more stably ordered way below.
1479   auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1480   llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1481     return ExitLoopMap.lookup(LHS)->getLoopDepth() <
1482            ExitLoopMap.lookup(RHS)->getLoopDepth();
1483   });
1484 
1485   // Populate the existing ExitLoopMap with everything reachable from each
1486   // exit, starting from the inner most exit.
1487   while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1488     assert(Worklist.empty() && "Didn't clear worklist!");
1489 
1490     BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1491     Loop *ExitL = ExitLoopMap.lookup(ExitBB);
1492 
1493     // Walk the CFG back until we hit the cloned PH adding everything reachable
1494     // and in the unlooped set to this exit block's loop.
1495     Worklist.push_back(ExitBB);
1496     do {
1497       BasicBlock *BB = Worklist.pop_back_val();
1498       // We can stop recursing at the cloned preheader (if we get there).
1499       if (BB == ClonedPH)
1500         continue;
1501 
1502       for (BasicBlock *PredBB : predecessors(BB)) {
1503         // If this pred has already been moved to our set or is part of some
1504         // (inner) loop, no update needed.
1505         if (!UnloopedBlockSet.erase(PredBB)) {
1506           assert(
1507               (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1508               "Predecessor not mapped to a loop!");
1509           continue;
1510         }
1511 
1512         // We just insert into the loop set here. We'll add these blocks to the
1513         // exit loop after we build up the set in an order that doesn't rely on
1514         // predecessor order (which in turn relies on use list order).
1515         bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;
1516         (void)Inserted;
1517         assert(Inserted && "Should only visit an unlooped block once!");
1518 
1519         // And recurse through to its predecessors.
1520         Worklist.push_back(PredBB);
1521       }
1522     } while (!Worklist.empty());
1523   }
1524 
1525   // Now that the ExitLoopMap gives as  mapping for all the non-looping cloned
1526   // blocks to their outer loops, walk the cloned blocks and the cloned exits
1527   // in their original order adding them to the correct loop.
1528 
1529   // We need a stable insertion order. We use the order of the original loop
1530   // order and map into the correct parent loop.
1531   for (auto *BB : llvm::concat<BasicBlock *const>(
1532            makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))
1533     if (Loop *OuterL = ExitLoopMap.lookup(BB))
1534       OuterL->addBasicBlockToLoop(BB, LI);
1535 
1536 #ifndef NDEBUG
1537   for (auto &BBAndL : ExitLoopMap) {
1538     auto *BB = BBAndL.first;
1539     auto *OuterL = BBAndL.second;
1540     assert(LI.getLoopFor(BB) == OuterL &&
1541            "Failed to put all blocks into outer loops!");
1542   }
1543 #endif
1544 
1545   // Now that all the blocks are placed into the correct containing loop in the
1546   // absence of child loops, find all the potentially cloned child loops and
1547   // clone them into whatever outer loop we placed their header into.
1548   for (Loop *ChildL : OrigL) {
1549     auto *ClonedChildHeader =
1550         cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1551     if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))
1552       continue;
1553 
1554 #ifndef NDEBUG
1555     for (auto *ChildLoopBB : ChildL->blocks())
1556       assert(VMap.count(ChildLoopBB) &&
1557              "Cloned a child loop header but not all of that loops blocks!");
1558 #endif
1559 
1560     NonChildClonedLoops.push_back(cloneLoopNest(
1561         *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));
1562   }
1563 }
1564 
1565 static void
1566 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1567                        ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1568                        DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1569   // Find all the dead clones, and remove them from their successors.
1570   SmallVector<BasicBlock *, 16> DeadBlocks;
1571   for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))
1572     for (auto &VMap : VMaps)
1573       if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))
1574         if (!DT.isReachableFromEntry(ClonedBB)) {
1575           for (BasicBlock *SuccBB : successors(ClonedBB))
1576             SuccBB->removePredecessor(ClonedBB);
1577           DeadBlocks.push_back(ClonedBB);
1578         }
1579 
1580   // Remove all MemorySSA in the dead blocks
1581   if (MSSAU) {
1582     SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(),
1583                                                  DeadBlocks.end());
1584     MSSAU->removeBlocks(DeadBlockSet);
1585   }
1586 
1587   // Drop any remaining references to break cycles.
1588   for (BasicBlock *BB : DeadBlocks)
1589     BB->dropAllReferences();
1590   // Erase them from the IR.
1591   for (BasicBlock *BB : DeadBlocks)
1592     BB->eraseFromParent();
1593 }
1594 
1595 static void
1596 deleteDeadBlocksFromLoop(Loop &L,
1597                          SmallVectorImpl<BasicBlock *> &ExitBlocks,
1598                          DominatorTree &DT, LoopInfo &LI,
1599                          MemorySSAUpdater *MSSAU,
1600                          function_ref<void(Loop &, StringRef)> DestroyLoopCB) {
1601   // Find all the dead blocks tied to this loop, and remove them from their
1602   // successors.
1603   SmallSetVector<BasicBlock *, 8> DeadBlockSet;
1604 
1605   // Start with loop/exit blocks and get a transitive closure of reachable dead
1606   // blocks.
1607   SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1608                                                 ExitBlocks.end());
1609   DeathCandidates.append(L.blocks().begin(), L.blocks().end());
1610   while (!DeathCandidates.empty()) {
1611     auto *BB = DeathCandidates.pop_back_val();
1612     if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {
1613       for (BasicBlock *SuccBB : successors(BB)) {
1614         SuccBB->removePredecessor(BB);
1615         DeathCandidates.push_back(SuccBB);
1616       }
1617       DeadBlockSet.insert(BB);
1618     }
1619   }
1620 
1621   // Remove all MemorySSA in the dead blocks
1622   if (MSSAU)
1623     MSSAU->removeBlocks(DeadBlockSet);
1624 
1625   // Filter out the dead blocks from the exit blocks list so that it can be
1626   // used in the caller.
1627   llvm::erase_if(ExitBlocks,
1628                  [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1629 
1630   // Walk from this loop up through its parents removing all of the dead blocks.
1631   for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1632     for (auto *BB : DeadBlockSet)
1633       ParentL->getBlocksSet().erase(BB);
1634     llvm::erase_if(ParentL->getBlocksVector(),
1635                    [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1636   }
1637 
1638   // Now delete the dead child loops. This raw delete will clear them
1639   // recursively.
1640   llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {
1641     if (!DeadBlockSet.count(ChildL->getHeader()))
1642       return false;
1643 
1644     assert(llvm::all_of(ChildL->blocks(),
1645                         [&](BasicBlock *ChildBB) {
1646                           return DeadBlockSet.count(ChildBB);
1647                         }) &&
1648            "If the child loop header is dead all blocks in the child loop must "
1649            "be dead as well!");
1650     DestroyLoopCB(*ChildL, ChildL->getName());
1651     LI.destroy(ChildL);
1652     return true;
1653   });
1654 
1655   // Remove the loop mappings for the dead blocks and drop all the references
1656   // from these blocks to others to handle cyclic references as we start
1657   // deleting the blocks themselves.
1658   for (auto *BB : DeadBlockSet) {
1659     // Check that the dominator tree has already been updated.
1660     assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1661     LI.changeLoopFor(BB, nullptr);
1662     // Drop all uses of the instructions to make sure we won't have dangling
1663     // uses in other blocks.
1664     for (auto &I : *BB)
1665       if (!I.use_empty())
1666         I.replaceAllUsesWith(UndefValue::get(I.getType()));
1667     BB->dropAllReferences();
1668   }
1669 
1670   // Actually delete the blocks now that they've been fully unhooked from the
1671   // IR.
1672   for (auto *BB : DeadBlockSet)
1673     BB->eraseFromParent();
1674 }
1675 
1676 /// Recompute the set of blocks in a loop after unswitching.
1677 ///
1678 /// This walks from the original headers predecessors to rebuild the loop. We
1679 /// take advantage of the fact that new blocks can't have been added, and so we
1680 /// filter by the original loop's blocks. This also handles potentially
1681 /// unreachable code that we don't want to explore but might be found examining
1682 /// the predecessors of the header.
1683 ///
1684 /// If the original loop is no longer a loop, this will return an empty set. If
1685 /// it remains a loop, all the blocks within it will be added to the set
1686 /// (including those blocks in inner loops).
1687 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1688                                                                  LoopInfo &LI) {
1689   SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1690 
1691   auto *PH = L.getLoopPreheader();
1692   auto *Header = L.getHeader();
1693 
1694   // A worklist to use while walking backwards from the header.
1695   SmallVector<BasicBlock *, 16> Worklist;
1696 
1697   // First walk the predecessors of the header to find the backedges. This will
1698   // form the basis of our walk.
1699   for (auto *Pred : predecessors(Header)) {
1700     // Skip the preheader.
1701     if (Pred == PH)
1702       continue;
1703 
1704     // Because the loop was in simplified form, the only non-loop predecessor
1705     // is the preheader.
1706     assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1707                                "than the preheader that is not part of the "
1708                                "loop!");
1709 
1710     // Insert this block into the loop set and on the first visit and, if it
1711     // isn't the header we're currently walking, put it into the worklist to
1712     // recurse through.
1713     if (LoopBlockSet.insert(Pred).second && Pred != Header)
1714       Worklist.push_back(Pred);
1715   }
1716 
1717   // If no backedges were found, we're done.
1718   if (LoopBlockSet.empty())
1719     return LoopBlockSet;
1720 
1721   // We found backedges, recurse through them to identify the loop blocks.
1722   while (!Worklist.empty()) {
1723     BasicBlock *BB = Worklist.pop_back_val();
1724     assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1725 
1726     // No need to walk past the header.
1727     if (BB == Header)
1728       continue;
1729 
1730     // Because we know the inner loop structure remains valid we can use the
1731     // loop structure to jump immediately across the entire nested loop.
1732     // Further, because it is in loop simplified form, we can directly jump
1733     // to its preheader afterward.
1734     if (Loop *InnerL = LI.getLoopFor(BB))
1735       if (InnerL != &L) {
1736         assert(L.contains(InnerL) &&
1737                "Should not reach a loop *outside* this loop!");
1738         // The preheader is the only possible predecessor of the loop so
1739         // insert it into the set and check whether it was already handled.
1740         auto *InnerPH = InnerL->getLoopPreheader();
1741         assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1742                                       "but not contain the inner loop "
1743                                       "preheader!");
1744         if (!LoopBlockSet.insert(InnerPH).second)
1745           // The only way to reach the preheader is through the loop body
1746           // itself so if it has been visited the loop is already handled.
1747           continue;
1748 
1749         // Insert all of the blocks (other than those already present) into
1750         // the loop set. We expect at least the block that led us to find the
1751         // inner loop to be in the block set, but we may also have other loop
1752         // blocks if they were already enqueued as predecessors of some other
1753         // outer loop block.
1754         for (auto *InnerBB : InnerL->blocks()) {
1755           if (InnerBB == BB) {
1756             assert(LoopBlockSet.count(InnerBB) &&
1757                    "Block should already be in the set!");
1758             continue;
1759           }
1760 
1761           LoopBlockSet.insert(InnerBB);
1762         }
1763 
1764         // Add the preheader to the worklist so we will continue past the
1765         // loop body.
1766         Worklist.push_back(InnerPH);
1767         continue;
1768       }
1769 
1770     // Insert any predecessors that were in the original loop into the new
1771     // set, and if the insert is successful, add them to the worklist.
1772     for (auto *Pred : predecessors(BB))
1773       if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)
1774         Worklist.push_back(Pred);
1775   }
1776 
1777   assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1778 
1779   // We've found all the blocks participating in the loop, return our completed
1780   // set.
1781   return LoopBlockSet;
1782 }
1783 
1784 /// Rebuild a loop after unswitching removes some subset of blocks and edges.
1785 ///
1786 /// The removal may have removed some child loops entirely but cannot have
1787 /// disturbed any remaining child loops. However, they may need to be hoisted
1788 /// to the parent loop (or to be top-level loops). The original loop may be
1789 /// completely removed.
1790 ///
1791 /// The sibling loops resulting from this update are returned. If the original
1792 /// loop remains a valid loop, it will be the first entry in this list with all
1793 /// of the newly sibling loops following it.
1794 ///
1795 /// Returns true if the loop remains a loop after unswitching, and false if it
1796 /// is no longer a loop after unswitching (and should not continue to be
1797 /// referenced).
1798 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1799                                      LoopInfo &LI,
1800                                      SmallVectorImpl<Loop *> &HoistedLoops) {
1801   auto *PH = L.getLoopPreheader();
1802 
1803   // Compute the actual parent loop from the exit blocks. Because we may have
1804   // pruned some exits the loop may be different from the original parent.
1805   Loop *ParentL = nullptr;
1806   SmallVector<Loop *, 4> ExitLoops;
1807   SmallVector<BasicBlock *, 4> ExitsInLoops;
1808   ExitsInLoops.reserve(ExitBlocks.size());
1809   for (auto *ExitBB : ExitBlocks)
1810     if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1811       ExitLoops.push_back(ExitL);
1812       ExitsInLoops.push_back(ExitBB);
1813       if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1814         ParentL = ExitL;
1815     }
1816 
1817   // Recompute the blocks participating in this loop. This may be empty if it
1818   // is no longer a loop.
1819   auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1820 
1821   // If we still have a loop, we need to re-set the loop's parent as the exit
1822   // block set changing may have moved it within the loop nest. Note that this
1823   // can only happen when this loop has a parent as it can only hoist the loop
1824   // *up* the nest.
1825   if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1826     // Remove this loop's (original) blocks from all of the intervening loops.
1827     for (Loop *IL = L.getParentLoop(); IL != ParentL;
1828          IL = IL->getParentLoop()) {
1829       IL->getBlocksSet().erase(PH);
1830       for (auto *BB : L.blocks())
1831         IL->getBlocksSet().erase(BB);
1832       llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {
1833         return BB == PH || L.contains(BB);
1834       });
1835     }
1836 
1837     LI.changeLoopFor(PH, ParentL);
1838     L.getParentLoop()->removeChildLoop(&L);
1839     if (ParentL)
1840       ParentL->addChildLoop(&L);
1841     else
1842       LI.addTopLevelLoop(&L);
1843   }
1844 
1845   // Now we update all the blocks which are no longer within the loop.
1846   auto &Blocks = L.getBlocksVector();
1847   auto BlocksSplitI =
1848       LoopBlockSet.empty()
1849           ? Blocks.begin()
1850           : std::stable_partition(
1851                 Blocks.begin(), Blocks.end(),
1852                 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });
1853 
1854   // Before we erase the list of unlooped blocks, build a set of them.
1855   SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1856   if (LoopBlockSet.empty())
1857     UnloopedBlocks.insert(PH);
1858 
1859   // Now erase these blocks from the loop.
1860   for (auto *BB : make_range(BlocksSplitI, Blocks.end()))
1861     L.getBlocksSet().erase(BB);
1862   Blocks.erase(BlocksSplitI, Blocks.end());
1863 
1864   // Sort the exits in ascending loop depth, we'll work backwards across these
1865   // to process them inside out.
1866   llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1867     return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);
1868   });
1869 
1870   // We'll build up a set for each exit loop.
1871   SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1872   Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1873 
1874   auto RemoveUnloopedBlocksFromLoop =
1875       [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1876         for (auto *BB : UnloopedBlocks)
1877           L.getBlocksSet().erase(BB);
1878         llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {
1879           return UnloopedBlocks.count(BB);
1880         });
1881       };
1882 
1883   SmallVector<BasicBlock *, 16> Worklist;
1884   while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1885     assert(Worklist.empty() && "Didn't clear worklist!");
1886     assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
1887 
1888     // Grab the next exit block, in decreasing loop depth order.
1889     BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1890     Loop &ExitL = *LI.getLoopFor(ExitBB);
1891     assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
1892 
1893     // Erase all of the unlooped blocks from the loops between the previous
1894     // exit loop and this exit loop. This works because the ExitInLoops list is
1895     // sorted in increasing order of loop depth and thus we visit loops in
1896     // decreasing order of loop depth.
1897     for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
1898       RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1899 
1900     // Walk the CFG back until we hit the cloned PH adding everything reachable
1901     // and in the unlooped set to this exit block's loop.
1902     Worklist.push_back(ExitBB);
1903     do {
1904       BasicBlock *BB = Worklist.pop_back_val();
1905       // We can stop recursing at the cloned preheader (if we get there).
1906       if (BB == PH)
1907         continue;
1908 
1909       for (BasicBlock *PredBB : predecessors(BB)) {
1910         // If this pred has already been moved to our set or is part of some
1911         // (inner) loop, no update needed.
1912         if (!UnloopedBlocks.erase(PredBB)) {
1913           assert((NewExitLoopBlocks.count(PredBB) ||
1914                   ExitL.contains(LI.getLoopFor(PredBB))) &&
1915                  "Predecessor not in a nested loop (or already visited)!");
1916           continue;
1917         }
1918 
1919         // We just insert into the loop set here. We'll add these blocks to the
1920         // exit loop after we build up the set in a deterministic order rather
1921         // than the predecessor-influenced visit order.
1922         bool Inserted = NewExitLoopBlocks.insert(PredBB).second;
1923         (void)Inserted;
1924         assert(Inserted && "Should only visit an unlooped block once!");
1925 
1926         // And recurse through to its predecessors.
1927         Worklist.push_back(PredBB);
1928       }
1929     } while (!Worklist.empty());
1930 
1931     // If blocks in this exit loop were directly part of the original loop (as
1932     // opposed to a child loop) update the map to point to this exit loop. This
1933     // just updates a map and so the fact that the order is unstable is fine.
1934     for (auto *BB : NewExitLoopBlocks)
1935       if (Loop *BBL = LI.getLoopFor(BB))
1936         if (BBL == &L || !L.contains(BBL))
1937           LI.changeLoopFor(BB, &ExitL);
1938 
1939     // We will remove the remaining unlooped blocks from this loop in the next
1940     // iteration or below.
1941     NewExitLoopBlocks.clear();
1942   }
1943 
1944   // Any remaining unlooped blocks are no longer part of any loop unless they
1945   // are part of some child loop.
1946   for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
1947     RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1948   for (auto *BB : UnloopedBlocks)
1949     if (Loop *BBL = LI.getLoopFor(BB))
1950       if (BBL == &L || !L.contains(BBL))
1951         LI.changeLoopFor(BB, nullptr);
1952 
1953   // Sink all the child loops whose headers are no longer in the loop set to
1954   // the parent (or to be top level loops). We reach into the loop and directly
1955   // update its subloop vector to make this batch update efficient.
1956   auto &SubLoops = L.getSubLoopsVector();
1957   auto SubLoopsSplitI =
1958       LoopBlockSet.empty()
1959           ? SubLoops.begin()
1960           : std::stable_partition(
1961                 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {
1962                   return LoopBlockSet.count(SubL->getHeader());
1963                 });
1964   for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {
1965     HoistedLoops.push_back(HoistedL);
1966     HoistedL->setParentLoop(nullptr);
1967 
1968     // To compute the new parent of this hoisted loop we look at where we
1969     // placed the preheader above. We can't lookup the header itself because we
1970     // retained the mapping from the header to the hoisted loop. But the
1971     // preheader and header should have the exact same new parent computed
1972     // based on the set of exit blocks from the original loop as the preheader
1973     // is a predecessor of the header and so reached in the reverse walk. And
1974     // because the loops were all in simplified form the preheader of the
1975     // hoisted loop can't be part of some *other* loop.
1976     if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))
1977       NewParentL->addChildLoop(HoistedL);
1978     else
1979       LI.addTopLevelLoop(HoistedL);
1980   }
1981   SubLoops.erase(SubLoopsSplitI, SubLoops.end());
1982 
1983   // Actually delete the loop if nothing remained within it.
1984   if (Blocks.empty()) {
1985     assert(SubLoops.empty() &&
1986            "Failed to remove all subloops from the original loop!");
1987     if (Loop *ParentL = L.getParentLoop())
1988       ParentL->removeChildLoop(llvm::find(*ParentL, &L));
1989     else
1990       LI.removeLoop(llvm::find(LI, &L));
1991     // markLoopAsDeleted for L should be triggered by the caller (it is typically
1992     // done by using the UnswitchCB callback).
1993     LI.destroy(&L);
1994     return false;
1995   }
1996 
1997   return true;
1998 }
1999 
2000 /// Helper to visit a dominator subtree, invoking a callable on each node.
2001 ///
2002 /// Returning false at any point will stop walking past that node of the tree.
2003 template <typename CallableT>
2004 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
2005   SmallVector<DomTreeNode *, 4> DomWorklist;
2006   DomWorklist.push_back(DT[BB]);
2007 #ifndef NDEBUG
2008   SmallPtrSet<DomTreeNode *, 4> Visited;
2009   Visited.insert(DT[BB]);
2010 #endif
2011   do {
2012     DomTreeNode *N = DomWorklist.pop_back_val();
2013 
2014     // Visit this node.
2015     if (!Callable(N->getBlock()))
2016       continue;
2017 
2018     // Accumulate the child nodes.
2019     for (DomTreeNode *ChildN : *N) {
2020       assert(Visited.insert(ChildN).second &&
2021              "Cannot visit a node twice when walking a tree!");
2022       DomWorklist.push_back(ChildN);
2023     }
2024   } while (!DomWorklist.empty());
2025 }
2026 
2027 static void unswitchNontrivialInvariants(
2028     Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
2029     SmallVectorImpl<BasicBlock *> &ExitBlocks, IVConditionInfo &PartialIVInfo,
2030     DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC,
2031     function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB,
2032     ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
2033     function_ref<void(Loop &, StringRef)> DestroyLoopCB) {
2034   auto *ParentBB = TI.getParent();
2035   BranchInst *BI = dyn_cast<BranchInst>(&TI);
2036   SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
2037 
2038   // We can only unswitch switches, conditional branches with an invariant
2039   // condition, or combining invariant conditions with an instruction or
2040   // partially invariant instructions.
2041   assert((SI || (BI && BI->isConditional())) &&
2042          "Can only unswitch switches and conditional branch!");
2043   bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty();
2044   bool FullUnswitch =
2045       SI || (BI->getCondition() == Invariants[0] && !PartiallyInvariant);
2046   if (FullUnswitch)
2047     assert(Invariants.size() == 1 &&
2048            "Cannot have other invariants with full unswitching!");
2049   else
2050     assert(isa<Instruction>(BI->getCondition()) &&
2051            "Partial unswitching requires an instruction as the condition!");
2052 
2053   if (MSSAU && VerifyMemorySSA)
2054     MSSAU->getMemorySSA()->verifyMemorySSA();
2055 
2056   // Constant and BBs tracking the cloned and continuing successor. When we are
2057   // unswitching the entire condition, this can just be trivially chosen to
2058   // unswitch towards `true`. However, when we are unswitching a set of
2059   // invariants combined with `and` or `or` or partially invariant instructions,
2060   // the combining operation determines the best direction to unswitch: we want
2061   // to unswitch the direction that will collapse the branch.
2062   bool Direction = true;
2063   int ClonedSucc = 0;
2064   if (!FullUnswitch) {
2065     Value *Cond = BI->getCondition();
2066     (void)Cond;
2067     assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) ||
2068             PartiallyInvariant) &&
2069            "Only `or`, `and`, an `select`, partially invariant instructions "
2070            "can combine invariants being unswitched.");
2071     if (!match(BI->getCondition(), m_LogicalOr())) {
2072       if (match(BI->getCondition(), m_LogicalAnd()) ||
2073           (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) {
2074         Direction = false;
2075         ClonedSucc = 1;
2076       }
2077     }
2078   }
2079 
2080   BasicBlock *RetainedSuccBB =
2081       BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();
2082   SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
2083   if (BI)
2084     UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));
2085   else
2086     for (auto Case : SI->cases())
2087       if (Case.getCaseSuccessor() != RetainedSuccBB)
2088         UnswitchedSuccBBs.insert(Case.getCaseSuccessor());
2089 
2090   assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
2091          "Should not unswitch the same successor we are retaining!");
2092 
2093   // The branch should be in this exact loop. Any inner loop's invariant branch
2094   // should be handled by unswitching that inner loop. The caller of this
2095   // routine should filter out any candidates that remain (but were skipped for
2096   // whatever reason).
2097   assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
2098 
2099   // Compute the parent loop now before we start hacking on things.
2100   Loop *ParentL = L.getParentLoop();
2101   // Get blocks in RPO order for MSSA update, before changing the CFG.
2102   LoopBlocksRPO LBRPO(&L);
2103   if (MSSAU)
2104     LBRPO.perform(&LI);
2105 
2106   // Compute the outer-most loop containing one of our exit blocks. This is the
2107   // furthest up our loopnest which can be mutated, which we will use below to
2108   // update things.
2109   Loop *OuterExitL = &L;
2110   for (auto *ExitBB : ExitBlocks) {
2111     Loop *NewOuterExitL = LI.getLoopFor(ExitBB);
2112     if (!NewOuterExitL) {
2113       // We exited the entire nest with this block, so we're done.
2114       OuterExitL = nullptr;
2115       break;
2116     }
2117     if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))
2118       OuterExitL = NewOuterExitL;
2119   }
2120 
2121   // At this point, we're definitely going to unswitch something so invalidate
2122   // any cached information in ScalarEvolution for the outer most loop
2123   // containing an exit block and all nested loops.
2124   if (SE) {
2125     if (OuterExitL)
2126       SE->forgetLoop(OuterExitL);
2127     else
2128       SE->forgetTopmostLoop(&L);
2129   }
2130 
2131   bool InsertFreeze = false;
2132   if (FreezeLoopUnswitchCond) {
2133     ICFLoopSafetyInfo SafetyInfo;
2134     SafetyInfo.computeLoopSafetyInfo(&L);
2135     InsertFreeze = !SafetyInfo.isGuaranteedToExecute(TI, &DT, &L);
2136   }
2137 
2138   // If the edge from this terminator to a successor dominates that successor,
2139   // store a map from each block in its dominator subtree to it. This lets us
2140   // tell when cloning for a particular successor if a block is dominated by
2141   // some *other* successor with a single data structure. We use this to
2142   // significantly reduce cloning.
2143   SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
2144   for (auto *SuccBB : llvm::concat<BasicBlock *const>(
2145            makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs))
2146     if (SuccBB->getUniquePredecessor() ||
2147         llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2148           return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);
2149         }))
2150       visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {
2151         DominatingSucc[BB] = SuccBB;
2152         return true;
2153       });
2154 
2155   // Split the preheader, so that we know that there is a safe place to insert
2156   // the conditional branch. We will change the preheader to have a conditional
2157   // branch on LoopCond. The original preheader will become the split point
2158   // between the unswitched versions, and we will have a new preheader for the
2159   // original loop.
2160   BasicBlock *SplitBB = L.getLoopPreheader();
2161   BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU);
2162 
2163   // Keep track of the dominator tree updates needed.
2164   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2165 
2166   // Clone the loop for each unswitched successor.
2167   SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
2168   VMaps.reserve(UnswitchedSuccBBs.size());
2169   SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
2170   for (auto *SuccBB : UnswitchedSuccBBs) {
2171     VMaps.emplace_back(new ValueToValueMapTy());
2172     ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2173         L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,
2174         DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU);
2175   }
2176 
2177   // Drop metadata if we may break its semantics by moving this instr into the
2178   // split block.
2179   if (TI.getMetadata(LLVMContext::MD_make_implicit)) {
2180     if (DropNonTrivialImplicitNullChecks)
2181       // Do not spend time trying to understand if we can keep it, just drop it
2182       // to save compile time.
2183       TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2184     else {
2185       // It is only legal to preserve make.implicit metadata if we are
2186       // guaranteed no reach implicit null check after following this branch.
2187       ICFLoopSafetyInfo SafetyInfo;
2188       SafetyInfo.computeLoopSafetyInfo(&L);
2189       if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))
2190         TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2191     }
2192   }
2193 
2194   // The stitching of the branched code back together depends on whether we're
2195   // doing full unswitching or not with the exception that we always want to
2196   // nuke the initial terminator placed in the split block.
2197   SplitBB->getTerminator()->eraseFromParent();
2198   if (FullUnswitch) {
2199     // Splice the terminator from the original loop and rewrite its
2200     // successors.
2201     SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI);
2202 
2203     // Keep a clone of the terminator for MSSA updates.
2204     Instruction *NewTI = TI.clone();
2205     ParentBB->getInstList().push_back(NewTI);
2206 
2207     // First wire up the moved terminator to the preheaders.
2208     if (BI) {
2209       BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2210       BI->setSuccessor(ClonedSucc, ClonedPH);
2211       BI->setSuccessor(1 - ClonedSucc, LoopPH);
2212       if (InsertFreeze) {
2213         auto Cond = BI->getCondition();
2214         if (!isGuaranteedNotToBeUndefOrPoison(Cond, &AC, BI, &DT))
2215           BI->setCondition(new FreezeInst(Cond, Cond->getName() + ".fr", BI));
2216       }
2217       DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2218     } else {
2219       assert(SI && "Must either be a branch or switch!");
2220 
2221       // Walk the cases and directly update their successors.
2222       assert(SI->getDefaultDest() == RetainedSuccBB &&
2223              "Not retaining default successor!");
2224       SI->setDefaultDest(LoopPH);
2225       for (auto &Case : SI->cases())
2226         if (Case.getCaseSuccessor() == RetainedSuccBB)
2227           Case.setSuccessor(LoopPH);
2228         else
2229           Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);
2230 
2231       if (InsertFreeze) {
2232         auto Cond = SI->getCondition();
2233         if (!isGuaranteedNotToBeUndefOrPoison(Cond, &AC, SI, &DT))
2234           SI->setCondition(new FreezeInst(Cond, Cond->getName() + ".fr", SI));
2235       }
2236       // We need to use the set to populate domtree updates as even when there
2237       // are multiple cases pointing at the same successor we only want to
2238       // remove and insert one edge in the domtree.
2239       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2240         DTUpdates.push_back(
2241             {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});
2242     }
2243 
2244     if (MSSAU) {
2245       DT.applyUpdates(DTUpdates);
2246       DTUpdates.clear();
2247 
2248       // Remove all but one edge to the retained block and all unswitched
2249       // blocks. This is to avoid having duplicate entries in the cloned Phis,
2250       // when we know we only keep a single edge for each case.
2251       MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB);
2252       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2253         MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB);
2254 
2255       for (auto &VMap : VMaps)
2256         MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2257                                    /*IgnoreIncomingWithNoClones=*/true);
2258       MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2259 
2260       // Remove all edges to unswitched blocks.
2261       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2262         MSSAU->removeEdge(ParentBB, SuccBB);
2263     }
2264 
2265     // Now unhook the successor relationship as we'll be replacing
2266     // the terminator with a direct branch. This is much simpler for branches
2267     // than switches so we handle those first.
2268     if (BI) {
2269       // Remove the parent as a predecessor of the unswitched successor.
2270       assert(UnswitchedSuccBBs.size() == 1 &&
2271              "Only one possible unswitched block for a branch!");
2272       BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2273       UnswitchedSuccBB->removePredecessor(ParentBB,
2274                                           /*KeepOneInputPHIs*/ true);
2275       DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2276     } else {
2277       // Note that we actually want to remove the parent block as a predecessor
2278       // of *every* case successor. The case successor is either unswitched,
2279       // completely eliminating an edge from the parent to that successor, or it
2280       // is a duplicate edge to the retained successor as the retained successor
2281       // is always the default successor and as we'll replace this with a direct
2282       // branch we no longer need the duplicate entries in the PHI nodes.
2283       SwitchInst *NewSI = cast<SwitchInst>(NewTI);
2284       assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2285              "Not retaining default successor!");
2286       for (auto &Case : NewSI->cases())
2287         Case.getCaseSuccessor()->removePredecessor(
2288             ParentBB,
2289             /*KeepOneInputPHIs*/ true);
2290 
2291       // We need to use the set to populate domtree updates as even when there
2292       // are multiple cases pointing at the same successor we only want to
2293       // remove and insert one edge in the domtree.
2294       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2295         DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});
2296     }
2297 
2298     // After MSSAU update, remove the cloned terminator instruction NewTI.
2299     ParentBB->getTerminator()->eraseFromParent();
2300 
2301     // Create a new unconditional branch to the continuing block (as opposed to
2302     // the one cloned).
2303     BranchInst::Create(RetainedSuccBB, ParentBB);
2304   } else {
2305     assert(BI && "Only branches have partial unswitching.");
2306     assert(UnswitchedSuccBBs.size() == 1 &&
2307            "Only one possible unswitched block for a branch!");
2308     BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2309     // When doing a partial unswitch, we have to do a bit more work to build up
2310     // the branch in the split block.
2311     if (PartiallyInvariant)
2312       buildPartialInvariantUnswitchConditionalBranch(
2313           *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, L, MSSAU);
2314     else
2315       buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction,
2316                                             *ClonedPH, *LoopPH, InsertFreeze);
2317     DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2318 
2319     if (MSSAU) {
2320       DT.applyUpdates(DTUpdates);
2321       DTUpdates.clear();
2322 
2323       // Perform MSSA cloning updates.
2324       for (auto &VMap : VMaps)
2325         MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2326                                    /*IgnoreIncomingWithNoClones=*/true);
2327       MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2328     }
2329   }
2330 
2331   // Apply the updates accumulated above to get an up-to-date dominator tree.
2332   DT.applyUpdates(DTUpdates);
2333 
2334   // Now that we have an accurate dominator tree, first delete the dead cloned
2335   // blocks so that we can accurately build any cloned loops. It is important to
2336   // not delete the blocks from the original loop yet because we still want to
2337   // reference the original loop to understand the cloned loop's structure.
2338   deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2339 
2340   // Build the cloned loop structure itself. This may be substantially
2341   // different from the original structure due to the simplified CFG. This also
2342   // handles inserting all the cloned blocks into the correct loops.
2343   SmallVector<Loop *, 4> NonChildClonedLoops;
2344   for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2345     buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);
2346 
2347   // Now that our cloned loops have been built, we can update the original loop.
2348   // First we delete the dead blocks from it and then we rebuild the loop
2349   // structure taking these deletions into account.
2350   deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, DestroyLoopCB);
2351 
2352   if (MSSAU && VerifyMemorySSA)
2353     MSSAU->getMemorySSA()->verifyMemorySSA();
2354 
2355   SmallVector<Loop *, 4> HoistedLoops;
2356   bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops);
2357 
2358   if (MSSAU && VerifyMemorySSA)
2359     MSSAU->getMemorySSA()->verifyMemorySSA();
2360 
2361   // This transformation has a high risk of corrupting the dominator tree, and
2362   // the below steps to rebuild loop structures will result in hard to debug
2363   // errors in that case so verify that the dominator tree is sane first.
2364   // FIXME: Remove this when the bugs stop showing up and rely on existing
2365   // verification steps.
2366   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2367 
2368   if (BI && !PartiallyInvariant) {
2369     // If we unswitched a branch which collapses the condition to a known
2370     // constant we want to replace all the uses of the invariants within both
2371     // the original and cloned blocks. We do this here so that we can use the
2372     // now updated dominator tree to identify which side the users are on.
2373     assert(UnswitchedSuccBBs.size() == 1 &&
2374            "Only one possible unswitched block for a branch!");
2375     BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2376 
2377     // When considering multiple partially-unswitched invariants
2378     // we cant just go replace them with constants in both branches.
2379     //
2380     // For 'AND' we infer that true branch ("continue") means true
2381     // for each invariant operand.
2382     // For 'OR' we can infer that false branch ("continue") means false
2383     // for each invariant operand.
2384     // So it happens that for multiple-partial case we dont replace
2385     // in the unswitched branch.
2386     bool ReplaceUnswitched =
2387         FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant;
2388 
2389     ConstantInt *UnswitchedReplacement =
2390         Direction ? ConstantInt::getTrue(BI->getContext())
2391                   : ConstantInt::getFalse(BI->getContext());
2392     ConstantInt *ContinueReplacement =
2393         Direction ? ConstantInt::getFalse(BI->getContext())
2394                   : ConstantInt::getTrue(BI->getContext());
2395     for (Value *Invariant : Invariants) {
2396       assert(!isa<Constant>(Invariant) &&
2397              "Should not be replacing constant values!");
2398       // Use make_early_inc_range here as set invalidates the iterator.
2399       for (Use &U : llvm::make_early_inc_range(Invariant->uses())) {
2400         Instruction *UserI = dyn_cast<Instruction>(U.getUser());
2401         if (!UserI)
2402           continue;
2403 
2404         // Replace it with the 'continue' side if in the main loop body, and the
2405         // unswitched if in the cloned blocks.
2406         if (DT.dominates(LoopPH, UserI->getParent()))
2407           U.set(ContinueReplacement);
2408         else if (ReplaceUnswitched &&
2409                  DT.dominates(ClonedPH, UserI->getParent()))
2410           U.set(UnswitchedReplacement);
2411       }
2412     }
2413   }
2414 
2415   // We can change which blocks are exit blocks of all the cloned sibling
2416   // loops, the current loop, and any parent loops which shared exit blocks
2417   // with the current loop. As a consequence, we need to re-form LCSSA for
2418   // them. But we shouldn't need to re-form LCSSA for any child loops.
2419   // FIXME: This could be made more efficient by tracking which exit blocks are
2420   // new, and focusing on them, but that isn't likely to be necessary.
2421   //
2422   // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2423   // loop nest and update every loop that could have had its exits changed. We
2424   // also need to cover any intervening loops. We add all of these loops to
2425   // a list and sort them by loop depth to achieve this without updating
2426   // unnecessary loops.
2427   auto UpdateLoop = [&](Loop &UpdateL) {
2428 #ifndef NDEBUG
2429     UpdateL.verifyLoop();
2430     for (Loop *ChildL : UpdateL) {
2431       ChildL->verifyLoop();
2432       assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2433              "Perturbed a child loop's LCSSA form!");
2434     }
2435 #endif
2436     // First build LCSSA for this loop so that we can preserve it when
2437     // forming dedicated exits. We don't want to perturb some other loop's
2438     // LCSSA while doing that CFG edit.
2439     formLCSSA(UpdateL, DT, &LI, SE);
2440 
2441     // For loops reached by this loop's original exit blocks we may
2442     // introduced new, non-dedicated exits. At least try to re-form dedicated
2443     // exits for these loops. This may fail if they couldn't have dedicated
2444     // exits to start with.
2445     formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true);
2446   };
2447 
2448   // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2449   // and we can do it in any order as they don't nest relative to each other.
2450   //
2451   // Also check if any of the loops we have updated have become top-level loops
2452   // as that will necessitate widening the outer loop scope.
2453   for (Loop *UpdatedL :
2454        llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {
2455     UpdateLoop(*UpdatedL);
2456     if (UpdatedL->isOutermost())
2457       OuterExitL = nullptr;
2458   }
2459   if (IsStillLoop) {
2460     UpdateLoop(L);
2461     if (L.isOutermost())
2462       OuterExitL = nullptr;
2463   }
2464 
2465   // If the original loop had exit blocks, walk up through the outer most loop
2466   // of those exit blocks to update LCSSA and form updated dedicated exits.
2467   if (OuterExitL != &L)
2468     for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2469          OuterL = OuterL->getParentLoop())
2470       UpdateLoop(*OuterL);
2471 
2472 #ifndef NDEBUG
2473   // Verify the entire loop structure to catch any incorrect updates before we
2474   // progress in the pass pipeline.
2475   LI.verify(DT);
2476 #endif
2477 
2478   // Now that we've unswitched something, make callbacks to report the changes.
2479   // For that we need to merge together the updated loops and the cloned loops
2480   // and check whether the original loop survived.
2481   SmallVector<Loop *, 4> SibLoops;
2482   for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
2483     if (UpdatedL->getParentLoop() == ParentL)
2484       SibLoops.push_back(UpdatedL);
2485   UnswitchCB(IsStillLoop, PartiallyInvariant, SibLoops);
2486 
2487   if (MSSAU && VerifyMemorySSA)
2488     MSSAU->getMemorySSA()->verifyMemorySSA();
2489 
2490   if (BI)
2491     ++NumBranches;
2492   else
2493     ++NumSwitches;
2494 }
2495 
2496 /// Recursively compute the cost of a dominator subtree based on the per-block
2497 /// cost map provided.
2498 ///
2499 /// The recursive computation is memozied into the provided DT-indexed cost map
2500 /// to allow querying it for most nodes in the domtree without it becoming
2501 /// quadratic.
2502 static InstructionCost computeDomSubtreeCost(
2503     DomTreeNode &N,
2504     const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap,
2505     SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) {
2506   // Don't accumulate cost (or recurse through) blocks not in our block cost
2507   // map and thus not part of the duplication cost being considered.
2508   auto BBCostIt = BBCostMap.find(N.getBlock());
2509   if (BBCostIt == BBCostMap.end())
2510     return 0;
2511 
2512   // Lookup this node to see if we already computed its cost.
2513   auto DTCostIt = DTCostMap.find(&N);
2514   if (DTCostIt != DTCostMap.end())
2515     return DTCostIt->second;
2516 
2517   // If not, we have to compute it. We can't use insert above and update
2518   // because computing the cost may insert more things into the map.
2519   InstructionCost Cost = std::accumulate(
2520       N.begin(), N.end(), BBCostIt->second,
2521       [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost {
2522         return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
2523       });
2524   bool Inserted = DTCostMap.insert({&N, Cost}).second;
2525   (void)Inserted;
2526   assert(Inserted && "Should not insert a node while visiting children!");
2527   return Cost;
2528 }
2529 
2530 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2531 /// making the following replacement:
2532 ///
2533 ///   --code before guard--
2534 ///   call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2535 ///   --code after guard--
2536 ///
2537 /// into
2538 ///
2539 ///   --code before guard--
2540 ///   br i1 %cond, label %guarded, label %deopt
2541 ///
2542 /// guarded:
2543 ///   --code after guard--
2544 ///
2545 /// deopt:
2546 ///   call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2547 ///   unreachable
2548 ///
2549 /// It also makes all relevant DT and LI updates, so that all structures are in
2550 /// valid state after this transform.
2551 static BranchInst *
2552 turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2553                     SmallVectorImpl<BasicBlock *> &ExitBlocks,
2554                     DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) {
2555   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2556   LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2557   BasicBlock *CheckBB = GI->getParent();
2558 
2559   if (MSSAU && VerifyMemorySSA)
2560      MSSAU->getMemorySSA()->verifyMemorySSA();
2561 
2562   // Remove all CheckBB's successors from DomTree. A block can be seen among
2563   // successors more than once, but for DomTree it should be added only once.
2564   SmallPtrSet<BasicBlock *, 4> Successors;
2565   for (auto *Succ : successors(CheckBB))
2566     if (Successors.insert(Succ).second)
2567       DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ});
2568 
2569   Instruction *DeoptBlockTerm =
2570       SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true);
2571   BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
2572   // SplitBlockAndInsertIfThen inserts control flow that branches to
2573   // DeoptBlockTerm if the condition is true.  We want the opposite.
2574   CheckBI->swapSuccessors();
2575 
2576   BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
2577   GuardedBlock->setName("guarded");
2578   CheckBI->getSuccessor(1)->setName("deopt");
2579   BasicBlock *DeoptBlock = CheckBI->getSuccessor(1);
2580 
2581   // We now have a new exit block.
2582   ExitBlocks.push_back(CheckBI->getSuccessor(1));
2583 
2584   if (MSSAU)
2585     MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI);
2586 
2587   GI->moveBefore(DeoptBlockTerm);
2588   GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));
2589 
2590   // Add new successors of CheckBB into DomTree.
2591   for (auto *Succ : successors(CheckBB))
2592     DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ});
2593 
2594   // Now the blocks that used to be CheckBB's successors are GuardedBlock's
2595   // successors.
2596   for (auto *Succ : Successors)
2597     DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ});
2598 
2599   // Make proper changes to DT.
2600   DT.applyUpdates(DTUpdates);
2601   // Inform LI of a new loop block.
2602   L.addBasicBlockToLoop(GuardedBlock, LI);
2603 
2604   if (MSSAU) {
2605     MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI));
2606     MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator);
2607     if (VerifyMemorySSA)
2608       MSSAU->getMemorySSA()->verifyMemorySSA();
2609   }
2610 
2611   ++NumGuards;
2612   return CheckBI;
2613 }
2614 
2615 /// Cost multiplier is a way to limit potentially exponential behavior
2616 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2617 /// candidates available. Also accounting for the number of "sibling" loops with
2618 /// the idea to account for previous unswitches that already happened on this
2619 /// cluster of loops. There was an attempt to keep this formula simple,
2620 /// just enough to limit the worst case behavior. Even if it is not that simple
2621 /// now it is still not an attempt to provide a detailed heuristic size
2622 /// prediction.
2623 ///
2624 /// TODO: Make a proper accounting of "explosion" effect for all kinds of
2625 /// unswitch candidates, making adequate predictions instead of wild guesses.
2626 /// That requires knowing not just the number of "remaining" candidates but
2627 /// also costs of unswitching for each of these candidates.
2628 static int CalculateUnswitchCostMultiplier(
2629     Instruction &TI, Loop &L, LoopInfo &LI, DominatorTree &DT,
2630     ArrayRef<std::pair<Instruction *, TinyPtrVector<Value *>>>
2631         UnswitchCandidates) {
2632 
2633   // Guards and other exiting conditions do not contribute to exponential
2634   // explosion as soon as they dominate the latch (otherwise there might be
2635   // another path to the latch remaining that does not allow to eliminate the
2636   // loop copy on unswitch).
2637   BasicBlock *Latch = L.getLoopLatch();
2638   BasicBlock *CondBlock = TI.getParent();
2639   if (DT.dominates(CondBlock, Latch) &&
2640       (isGuard(&TI) ||
2641        llvm::count_if(successors(&TI), [&L](BasicBlock *SuccBB) {
2642          return L.contains(SuccBB);
2643        }) <= 1)) {
2644     NumCostMultiplierSkipped++;
2645     return 1;
2646   }
2647 
2648   auto *ParentL = L.getParentLoop();
2649   int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2650                                : std::distance(LI.begin(), LI.end()));
2651   // Count amount of clones that all the candidates might cause during
2652   // unswitching. Branch/guard counts as 1, switch counts as log2 of its cases.
2653   int UnswitchedClones = 0;
2654   for (auto Candidate : UnswitchCandidates) {
2655     Instruction *CI = Candidate.first;
2656     BasicBlock *CondBlock = CI->getParent();
2657     bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch);
2658     if (isGuard(CI)) {
2659       if (!SkipExitingSuccessors)
2660         UnswitchedClones++;
2661       continue;
2662     }
2663     int NonExitingSuccessors = llvm::count_if(
2664         successors(CondBlock), [SkipExitingSuccessors, &L](BasicBlock *SuccBB) {
2665           return !SkipExitingSuccessors || L.contains(SuccBB);
2666         });
2667     UnswitchedClones += Log2_32(NonExitingSuccessors);
2668   }
2669 
2670   // Ignore up to the "unscaled candidates" number of unswitch candidates
2671   // when calculating the power-of-two scaling of the cost. The main idea
2672   // with this control is to allow a small number of unswitches to happen
2673   // and rely more on siblings multiplier (see below) when the number
2674   // of candidates is small.
2675   unsigned ClonesPower =
2676       std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0);
2677 
2678   // Allowing top-level loops to spread a bit more than nested ones.
2679   int SiblingsMultiplier =
2680       std::max((ParentL ? SiblingsCount
2681                         : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2682                1);
2683   // Compute the cost multiplier in a way that won't overflow by saturating
2684   // at an upper bound.
2685   int CostMultiplier;
2686   if (ClonesPower > Log2_32(UnswitchThreshold) ||
2687       SiblingsMultiplier > UnswitchThreshold)
2688     CostMultiplier = UnswitchThreshold;
2689   else
2690     CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower),
2691                               (int)UnswitchThreshold);
2692 
2693   LLVM_DEBUG(dbgs() << "  Computed multiplier  " << CostMultiplier
2694                     << " (siblings " << SiblingsMultiplier << " * clones "
2695                     << (1 << ClonesPower) << ")"
2696                     << " for unswitch candidate: " << TI << "\n");
2697   return CostMultiplier;
2698 }
2699 
2700 static bool unswitchBestCondition(
2701     Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC,
2702     AAResults &AA, TargetTransformInfo &TTI,
2703     function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB,
2704     ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
2705     function_ref<void(Loop &, StringRef)> DestroyLoopCB) {
2706   // Collect all invariant conditions within this loop (as opposed to an inner
2707   // loop which would be handled when visiting that inner loop).
2708   SmallVector<std::pair<Instruction *, TinyPtrVector<Value *>>, 4>
2709       UnswitchCandidates;
2710 
2711   // Whether or not we should also collect guards in the loop.
2712   bool CollectGuards = false;
2713   if (UnswitchGuards) {
2714     auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2715         Intrinsic::getName(Intrinsic::experimental_guard));
2716     if (GuardDecl && !GuardDecl->use_empty())
2717       CollectGuards = true;
2718   }
2719 
2720   IVConditionInfo PartialIVInfo;
2721   for (auto *BB : L.blocks()) {
2722     if (LI.getLoopFor(BB) != &L)
2723       continue;
2724 
2725     if (CollectGuards)
2726       for (auto &I : *BB)
2727         if (isGuard(&I)) {
2728           auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0);
2729           // TODO: Support AND, OR conditions and partial unswitching.
2730           if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
2731             UnswitchCandidates.push_back({&I, {Cond}});
2732         }
2733 
2734     if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
2735       // We can only consider fully loop-invariant switch conditions as we need
2736       // to completely eliminate the switch after unswitching.
2737       if (!isa<Constant>(SI->getCondition()) &&
2738           L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor())
2739         UnswitchCandidates.push_back({SI, {SI->getCondition()}});
2740       continue;
2741     }
2742 
2743     auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
2744     if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) ||
2745         BI->getSuccessor(0) == BI->getSuccessor(1))
2746       continue;
2747 
2748     // If BI's condition is 'select _, true, false', simplify it to confuse
2749     // matchers
2750     Value *Cond = BI->getCondition(), *CondNext;
2751     while (match(Cond, m_Select(m_Value(CondNext), m_One(), m_Zero())))
2752       Cond = CondNext;
2753     BI->setCondition(Cond);
2754 
2755     if (isa<Constant>(Cond))
2756       continue;
2757 
2758     if (L.isLoopInvariant(BI->getCondition())) {
2759       UnswitchCandidates.push_back({BI, {BI->getCondition()}});
2760       continue;
2761     }
2762 
2763     Instruction &CondI = *cast<Instruction>(BI->getCondition());
2764     if (match(&CondI, m_CombineOr(m_LogicalAnd(), m_LogicalOr()))) {
2765       TinyPtrVector<Value *> Invariants =
2766           collectHomogenousInstGraphLoopInvariants(L, CondI, LI);
2767       if (Invariants.empty())
2768         continue;
2769 
2770       UnswitchCandidates.push_back({BI, std::move(Invariants)});
2771       continue;
2772     }
2773   }
2774 
2775   Instruction *PartialIVCondBranch = nullptr;
2776   if (MSSAU && !findOptionMDForLoop(&L, "llvm.loop.unswitch.partial.disable") &&
2777       !any_of(UnswitchCandidates, [&L](auto &TerminatorAndInvariants) {
2778         return TerminatorAndInvariants.first == L.getHeader()->getTerminator();
2779       })) {
2780     MemorySSA *MSSA = MSSAU->getMemorySSA();
2781     if (auto Info = hasPartialIVCondition(L, MSSAThreshold, *MSSA, AA)) {
2782       LLVM_DEBUG(
2783           dbgs() << "simple-loop-unswitch: Found partially invariant condition "
2784                  << *Info->InstToDuplicate[0] << "\n");
2785       PartialIVInfo = *Info;
2786       PartialIVCondBranch = L.getHeader()->getTerminator();
2787       TinyPtrVector<Value *> ValsToDuplicate;
2788       for (auto *Inst : Info->InstToDuplicate)
2789         ValsToDuplicate.push_back(Inst);
2790       UnswitchCandidates.push_back(
2791           {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)});
2792     }
2793   }
2794 
2795   // If we didn't find any candidates, we're done.
2796   if (UnswitchCandidates.empty())
2797     return false;
2798 
2799   // Check if there are irreducible CFG cycles in this loop. If so, we cannot
2800   // easily unswitch non-trivial edges out of the loop. Doing so might turn the
2801   // irreducible control flow into reducible control flow and introduce new
2802   // loops "out of thin air". If we ever discover important use cases for doing
2803   // this, we can add support to loop unswitch, but it is a lot of complexity
2804   // for what seems little or no real world benefit.
2805   LoopBlocksRPO RPOT(&L);
2806   RPOT.perform(&LI);
2807   if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
2808     return false;
2809 
2810   SmallVector<BasicBlock *, 4> ExitBlocks;
2811   L.getUniqueExitBlocks(ExitBlocks);
2812 
2813   // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch
2814   // instruction as we don't know how to split those exit blocks.
2815   // FIXME: We should teach SplitBlock to handle this and remove this
2816   // restriction.
2817   for (auto *ExitBB : ExitBlocks) {
2818     auto *I = ExitBB->getFirstNonPHI();
2819     if (isa<CleanupPadInst>(I) || isa<CatchSwitchInst>(I)) {
2820       LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch "
2821                            "in exit block\n");
2822       return false;
2823     }
2824   }
2825 
2826   LLVM_DEBUG(
2827       dbgs() << "Considering " << UnswitchCandidates.size()
2828              << " non-trivial loop invariant conditions for unswitching.\n");
2829 
2830   // Given that unswitching these terminators will require duplicating parts of
2831   // the loop, so we need to be able to model that cost. Compute the ephemeral
2832   // values and set up a data structure to hold per-BB costs. We cache each
2833   // block's cost so that we don't recompute this when considering different
2834   // subsets of the loop for duplication during unswitching.
2835   SmallPtrSet<const Value *, 4> EphValues;
2836   CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);
2837   SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap;
2838 
2839   // Compute the cost of each block, as well as the total loop cost. Also, bail
2840   // out if we see instructions which are incompatible with loop unswitching
2841   // (convergent, noduplicate, or cross-basic-block tokens).
2842   // FIXME: We might be able to safely handle some of these in non-duplicated
2843   // regions.
2844   TargetTransformInfo::TargetCostKind CostKind =
2845       L.getHeader()->getParent()->hasMinSize()
2846       ? TargetTransformInfo::TCK_CodeSize
2847       : TargetTransformInfo::TCK_SizeAndLatency;
2848   InstructionCost LoopCost = 0;
2849   for (auto *BB : L.blocks()) {
2850     InstructionCost Cost = 0;
2851     for (auto &I : *BB) {
2852       if (EphValues.count(&I))
2853         continue;
2854 
2855       if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
2856         return false;
2857       if (auto *CB = dyn_cast<CallBase>(&I))
2858         if (CB->isConvergent() || CB->cannotDuplicate())
2859           return false;
2860 
2861       Cost += TTI.getUserCost(&I, CostKind);
2862     }
2863     assert(Cost >= 0 && "Must not have negative costs!");
2864     LoopCost += Cost;
2865     assert(LoopCost >= 0 && "Must not have negative loop costs!");
2866     BBCostMap[BB] = Cost;
2867   }
2868   LLVM_DEBUG(dbgs() << "  Total loop cost: " << LoopCost << "\n");
2869 
2870   // Now we find the best candidate by searching for the one with the following
2871   // properties in order:
2872   //
2873   // 1) An unswitching cost below the threshold
2874   // 2) The smallest number of duplicated unswitch candidates (to avoid
2875   //    creating redundant subsequent unswitching)
2876   // 3) The smallest cost after unswitching.
2877   //
2878   // We prioritize reducing fanout of unswitch candidates provided the cost
2879   // remains below the threshold because this has a multiplicative effect.
2880   //
2881   // This requires memoizing each dominator subtree to avoid redundant work.
2882   //
2883   // FIXME: Need to actually do the number of candidates part above.
2884   SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap;
2885   // Given a terminator which might be unswitched, computes the non-duplicated
2886   // cost for that terminator.
2887   auto ComputeUnswitchedCost = [&](Instruction &TI,
2888                                    bool FullUnswitch) -> InstructionCost {
2889     BasicBlock &BB = *TI.getParent();
2890     SmallPtrSet<BasicBlock *, 4> Visited;
2891 
2892     InstructionCost Cost = 0;
2893     for (BasicBlock *SuccBB : successors(&BB)) {
2894       // Don't count successors more than once.
2895       if (!Visited.insert(SuccBB).second)
2896         continue;
2897 
2898       // If this is a partial unswitch candidate, then it must be a conditional
2899       // branch with a condition of either `or`, `and`, their corresponding
2900       // select forms or partially invariant instructions. In that case, one of
2901       // the successors is necessarily duplicated, so don't even try to remove
2902       // its cost.
2903       if (!FullUnswitch) {
2904         auto &BI = cast<BranchInst>(TI);
2905         if (match(BI.getCondition(), m_LogicalAnd())) {
2906           if (SuccBB == BI.getSuccessor(1))
2907             continue;
2908         } else if (match(BI.getCondition(), m_LogicalOr())) {
2909           if (SuccBB == BI.getSuccessor(0))
2910             continue;
2911         } else if ((PartialIVInfo.KnownValue->isOneValue() &&
2912                     SuccBB == BI.getSuccessor(0)) ||
2913                    (!PartialIVInfo.KnownValue->isOneValue() &&
2914                     SuccBB == BI.getSuccessor(1)))
2915           continue;
2916       }
2917 
2918       // This successor's domtree will not need to be duplicated after
2919       // unswitching if the edge to the successor dominates it (and thus the
2920       // entire tree). This essentially means there is no other path into this
2921       // subtree and so it will end up live in only one clone of the loop.
2922       if (SuccBB->getUniquePredecessor() ||
2923           llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2924             return PredBB == &BB || DT.dominates(SuccBB, PredBB);
2925           })) {
2926         Cost += computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);
2927         assert(Cost <= LoopCost &&
2928                "Non-duplicated cost should never exceed total loop cost!");
2929       }
2930     }
2931 
2932     // Now scale the cost by the number of unique successors minus one. We
2933     // subtract one because there is already at least one copy of the entire
2934     // loop. This is computing the new cost of unswitching a condition.
2935     // Note that guards always have 2 unique successors that are implicit and
2936     // will be materialized if we decide to unswitch it.
2937     int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
2938     assert(SuccessorsCount > 1 &&
2939            "Cannot unswitch a condition without multiple distinct successors!");
2940     return (LoopCost - Cost) * (SuccessorsCount - 1);
2941   };
2942   Instruction *BestUnswitchTI = nullptr;
2943   InstructionCost BestUnswitchCost = 0;
2944   ArrayRef<Value *> BestUnswitchInvariants;
2945   for (auto &TerminatorAndInvariants : UnswitchCandidates) {
2946     Instruction &TI = *TerminatorAndInvariants.first;
2947     ArrayRef<Value *> Invariants = TerminatorAndInvariants.second;
2948     BranchInst *BI = dyn_cast<BranchInst>(&TI);
2949     InstructionCost CandidateCost = ComputeUnswitchedCost(
2950         TI, /*FullUnswitch*/ !BI || (Invariants.size() == 1 &&
2951                                      Invariants[0] == BI->getCondition()));
2952     // Calculate cost multiplier which is a tool to limit potentially
2953     // exponential behavior of loop-unswitch.
2954     if (EnableUnswitchCostMultiplier) {
2955       int CostMultiplier =
2956           CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
2957       assert(
2958           (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
2959           "cost multiplier needs to be in the range of 1..UnswitchThreshold");
2960       CandidateCost *= CostMultiplier;
2961       LLVM_DEBUG(dbgs() << "  Computed cost of " << CandidateCost
2962                         << " (multiplier: " << CostMultiplier << ")"
2963                         << " for unswitch candidate: " << TI << "\n");
2964     } else {
2965       LLVM_DEBUG(dbgs() << "  Computed cost of " << CandidateCost
2966                         << " for unswitch candidate: " << TI << "\n");
2967     }
2968 
2969     if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) {
2970       BestUnswitchTI = &TI;
2971       BestUnswitchCost = CandidateCost;
2972       BestUnswitchInvariants = Invariants;
2973     }
2974   }
2975   assert(BestUnswitchTI && "Failed to find loop unswitch candidate");
2976 
2977   if (BestUnswitchCost >= UnswitchThreshold) {
2978     LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: "
2979                       << BestUnswitchCost << "\n");
2980     return false;
2981   }
2982 
2983   if (BestUnswitchTI != PartialIVCondBranch)
2984     PartialIVInfo.InstToDuplicate.clear();
2985 
2986   // If the best candidate is a guard, turn it into a branch.
2987   if (isGuard(BestUnswitchTI))
2988     BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L,
2989                                          ExitBlocks, DT, LI, MSSAU);
2990 
2991   LLVM_DEBUG(dbgs() << "  Unswitching non-trivial (cost = "
2992                     << BestUnswitchCost << ") terminator: " << *BestUnswitchTI
2993                     << "\n");
2994   unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants,
2995                                ExitBlocks, PartialIVInfo, DT, LI, AC,
2996                                UnswitchCB, SE, MSSAU, DestroyLoopCB);
2997   return true;
2998 }
2999 
3000 /// Unswitch control flow predicated on loop invariant conditions.
3001 ///
3002 /// This first hoists all branches or switches which are trivial (IE, do not
3003 /// require duplicating any part of the loop) out of the loop body. It then
3004 /// looks at other loop invariant control flows and tries to unswitch those as
3005 /// well by cloning the loop if the result is small enough.
3006 ///
3007 /// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are
3008 /// also updated based on the unswitch. The `MSSA` analysis is also updated if
3009 /// valid (i.e. its use is enabled).
3010 ///
3011 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
3012 /// true, we will attempt to do non-trivial unswitching as well as trivial
3013 /// unswitching.
3014 ///
3015 /// The `UnswitchCB` callback provided will be run after unswitching is
3016 /// complete, with the first parameter set to `true` if the provided loop
3017 /// remains a loop, and a list of new sibling loops created.
3018 ///
3019 /// If `SE` is non-null, we will update that analysis based on the unswitching
3020 /// done.
3021 static bool
3022 unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC,
3023              AAResults &AA, TargetTransformInfo &TTI, bool Trivial,
3024              bool NonTrivial,
3025              function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB,
3026              ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
3027              function_ref<void(Loop &, StringRef)> DestroyLoopCB) {
3028   assert(L.isRecursivelyLCSSAForm(DT, LI) &&
3029          "Loops must be in LCSSA form before unswitching.");
3030 
3031   // Must be in loop simplified form: we need a preheader and dedicated exits.
3032   if (!L.isLoopSimplifyForm())
3033     return false;
3034 
3035   // Try trivial unswitch first before loop over other basic blocks in the loop.
3036   if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
3037     // If we unswitched successfully we will want to clean up the loop before
3038     // processing it further so just mark it as unswitched and return.
3039     UnswitchCB(/*CurrentLoopValid*/ true, false, {});
3040     return true;
3041   }
3042 
3043   // Check whether we should continue with non-trivial conditions.
3044   // EnableNonTrivialUnswitch: Global variable that forces non-trivial
3045   //                           unswitching for testing and debugging.
3046   // NonTrivial: Parameter that enables non-trivial unswitching for this
3047   //             invocation of the transform. But this should be allowed only
3048   //             for targets without branch divergence.
3049   //
3050   // FIXME: If divergence analysis becomes available to a loop
3051   // transform, we should allow unswitching for non-trivial uniform
3052   // branches even on targets that have divergence.
3053   // https://bugs.llvm.org/show_bug.cgi?id=48819
3054   bool ContinueWithNonTrivial =
3055       EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence());
3056   if (!ContinueWithNonTrivial)
3057     return false;
3058 
3059   // Skip non-trivial unswitching for optsize functions.
3060   if (L.getHeader()->getParent()->hasOptSize())
3061     return false;
3062 
3063   // Skip non-trivial unswitching for loops that cannot be cloned.
3064   if (!L.isSafeToClone())
3065     return false;
3066 
3067   // For non-trivial unswitching, because it often creates new loops, we rely on
3068   // the pass manager to iterate on the loops rather than trying to immediately
3069   // reach a fixed point. There is no substantial advantage to iterating
3070   // internally, and if any of the new loops are simplified enough to contain
3071   // trivial unswitching we want to prefer those.
3072 
3073   // Try to unswitch the best invariant condition. We prefer this full unswitch to
3074   // a partial unswitch when possible below the threshold.
3075   if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, UnswitchCB, SE, MSSAU,
3076                             DestroyLoopCB))
3077     return true;
3078 
3079   // No other opportunities to unswitch.
3080   return false;
3081 }
3082 
3083 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
3084                                               LoopStandardAnalysisResults &AR,
3085                                               LPMUpdater &U) {
3086   Function &F = *L.getHeader()->getParent();
3087   (void)F;
3088 
3089   LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
3090                     << "\n");
3091 
3092   // Save the current loop name in a variable so that we can report it even
3093   // after it has been deleted.
3094   std::string LoopName = std::string(L.getName());
3095 
3096   auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid,
3097                                         bool PartiallyInvariant,
3098                                         ArrayRef<Loop *> NewLoops) {
3099     // If we did a non-trivial unswitch, we have added new (cloned) loops.
3100     if (!NewLoops.empty())
3101       U.addSiblingLoops(NewLoops);
3102 
3103     // If the current loop remains valid, we should revisit it to catch any
3104     // other unswitch opportunities. Otherwise, we need to mark it as deleted.
3105     if (CurrentLoopValid) {
3106       if (PartiallyInvariant) {
3107         // Mark the new loop as partially unswitched, to avoid unswitching on
3108         // the same condition again.
3109         auto &Context = L.getHeader()->getContext();
3110         MDNode *DisableUnswitchMD = MDNode::get(
3111             Context,
3112             MDString::get(Context, "llvm.loop.unswitch.partial.disable"));
3113         MDNode *NewLoopID = makePostTransformationMetadata(
3114             Context, L.getLoopID(), {"llvm.loop.unswitch.partial"},
3115             {DisableUnswitchMD});
3116         L.setLoopID(NewLoopID);
3117       } else
3118         U.revisitCurrentLoop();
3119     } else
3120       U.markLoopAsDeleted(L, LoopName);
3121   };
3122 
3123   auto DestroyLoopCB = [&U](Loop &L, StringRef Name) {
3124     U.markLoopAsDeleted(L, Name);
3125   };
3126 
3127   Optional<MemorySSAUpdater> MSSAU;
3128   if (AR.MSSA) {
3129     MSSAU = MemorySSAUpdater(AR.MSSA);
3130     if (VerifyMemorySSA)
3131       AR.MSSA->verifyMemorySSA();
3132   }
3133   if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.AA, AR.TTI, Trivial, NonTrivial,
3134                     UnswitchCB, &AR.SE,
3135                     MSSAU.hasValue() ? MSSAU.getPointer() : nullptr,
3136                     DestroyLoopCB))
3137     return PreservedAnalyses::all();
3138 
3139   if (AR.MSSA && VerifyMemorySSA)
3140     AR.MSSA->verifyMemorySSA();
3141 
3142   // Historically this pass has had issues with the dominator tree so verify it
3143   // in asserts builds.
3144   assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));
3145 
3146   auto PA = getLoopPassPreservedAnalyses();
3147   if (AR.MSSA)
3148     PA.preserve<MemorySSAAnalysis>();
3149   return PA;
3150 }
3151 
3152 void SimpleLoopUnswitchPass::printPipeline(
3153     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
3154   static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline(
3155       OS, MapClassName2PassName);
3156 
3157   OS << "<";
3158   OS << (NonTrivial ? "" : "no-") << "nontrivial;";
3159   OS << (Trivial ? "" : "no-") << "trivial";
3160   OS << ">";
3161 }
3162 
3163 namespace {
3164 
3165 class SimpleLoopUnswitchLegacyPass : public LoopPass {
3166   bool NonTrivial;
3167 
3168 public:
3169   static char ID; // Pass ID, replacement for typeid
3170 
3171   explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false)
3172       : LoopPass(ID), NonTrivial(NonTrivial) {
3173     initializeSimpleLoopUnswitchLegacyPassPass(
3174         *PassRegistry::getPassRegistry());
3175   }
3176 
3177   bool runOnLoop(Loop *L, LPPassManager &LPM) override;
3178 
3179   void getAnalysisUsage(AnalysisUsage &AU) const override {
3180     AU.addRequired<AssumptionCacheTracker>();
3181     AU.addRequired<TargetTransformInfoWrapperPass>();
3182     AU.addRequired<MemorySSAWrapperPass>();
3183     AU.addPreserved<MemorySSAWrapperPass>();
3184     getLoopAnalysisUsage(AU);
3185   }
3186 };
3187 
3188 } // end anonymous namespace
3189 
3190 bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
3191   if (skipLoop(L))
3192     return false;
3193 
3194   Function &F = *L->getHeader()->getParent();
3195 
3196   LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *L
3197                     << "\n");
3198 
3199   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3200   auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
3201   auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3202   auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
3203   auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
3204   MemorySSA *MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
3205   MemorySSAUpdater MSSAU(MSSA);
3206 
3207   auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
3208   auto *SE = SEWP ? &SEWP->getSE() : nullptr;
3209 
3210   auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid, bool PartiallyInvariant,
3211                                ArrayRef<Loop *> NewLoops) {
3212     // If we did a non-trivial unswitch, we have added new (cloned) loops.
3213     for (auto *NewL : NewLoops)
3214       LPM.addLoop(*NewL);
3215 
3216     // If the current loop remains valid, re-add it to the queue. This is
3217     // a little wasteful as we'll finish processing the current loop as well,
3218     // but it is the best we can do in the old PM.
3219     if (CurrentLoopValid) {
3220       // If the current loop has been unswitched using a partially invariant
3221       // condition, we should not re-add the current loop to avoid unswitching
3222       // on the same condition again.
3223       if (!PartiallyInvariant)
3224         LPM.addLoop(*L);
3225     } else
3226       LPM.markLoopAsDeleted(*L);
3227   };
3228 
3229   auto DestroyLoopCB = [&LPM](Loop &L, StringRef /* Name */) {
3230     LPM.markLoopAsDeleted(L);
3231   };
3232 
3233   if (VerifyMemorySSA)
3234     MSSA->verifyMemorySSA();
3235 
3236   bool Changed = unswitchLoop(*L, DT, LI, AC, AA, TTI, true, NonTrivial,
3237                               UnswitchCB, SE, &MSSAU, DestroyLoopCB);
3238 
3239   if (VerifyMemorySSA)
3240     MSSA->verifyMemorySSA();
3241 
3242   // Historically this pass has had issues with the dominator tree so verify it
3243   // in asserts builds.
3244   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
3245 
3246   return Changed;
3247 }
3248 
3249 char SimpleLoopUnswitchLegacyPass::ID = 0;
3250 INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
3251                       "Simple unswitch loops", false, false)
3252 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3253 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3254 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
3255 INITIALIZE_PASS_DEPENDENCY(LoopPass)
3256 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
3257 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
3258 INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",
3259                     "Simple unswitch loops", false, false)
3260 
3261 Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) {
3262   return new SimpleLoopUnswitchLegacyPass(NonTrivial);
3263 }
3264