1	///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===//
2	//
3	// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	// See https://llvm.org/LICENSE.txt for license information.
5	// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	//
7	//===----------------------------------------------------------------------===//
8
9	#include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
10	#include "llvm/ADT/DenseMap.h"
11	#include "llvm/ADT/STLExtras.h"
12	#include "llvm/ADT/Sequence.h"
13	#include "llvm/ADT/SetVector.h"
14	#include "llvm/ADT/SmallPtrSet.h"
15	#include "llvm/ADT/SmallVector.h"
16	#include "llvm/ADT/Statistic.h"
17	#include "llvm/ADT/Twine.h"
18	#include "llvm/Analysis/AssumptionCache.h"
19	#include "llvm/Analysis/BlockFrequencyInfo.h"
20	#include "llvm/Analysis/CFG.h"
21	#include "llvm/Analysis/CodeMetrics.h"
22	#include "llvm/Analysis/DomTreeUpdater.h"
23	#include "llvm/Analysis/GuardUtils.h"
24	#include "llvm/Analysis/LoopAnalysisManager.h"
25	#include "llvm/Analysis/LoopInfo.h"
26	#include "llvm/Analysis/LoopIterator.h"
27	#include "llvm/Analysis/MemorySSA.h"
28	#include "llvm/Analysis/MemorySSAUpdater.h"
29	#include "llvm/Analysis/MustExecute.h"
30	#include "llvm/Analysis/ProfileSummaryInfo.h"
31	#include "llvm/Analysis/ScalarEvolution.h"
32	#include "llvm/Analysis/TargetTransformInfo.h"
33	#include "llvm/Analysis/ValueTracking.h"
34	#include "llvm/IR/BasicBlock.h"
35	#include "llvm/IR/Constant.h"
36	#include "llvm/IR/Constants.h"
37	#include "llvm/IR/Dominators.h"
38	#include "llvm/IR/Function.h"
39	#include "llvm/IR/IRBuilder.h"
40	#include "llvm/IR/InstrTypes.h"
41	#include "llvm/IR/Instruction.h"
42	#include "llvm/IR/Instructions.h"
43	#include "llvm/IR/IntrinsicInst.h"
44	#include "llvm/IR/PatternMatch.h"
45	#include "llvm/IR/ProfDataUtils.h"
46	#include "llvm/IR/Use.h"
47	#include "llvm/IR/Value.h"
48	#include "llvm/Support/Casting.h"
49	#include "llvm/Support/CommandLine.h"
50	#include "llvm/Support/Debug.h"
51	#include "llvm/Support/ErrorHandling.h"
52	#include "llvm/Support/GenericDomTree.h"
53	#include "llvm/Support/InstructionCost.h"
54	#include "llvm/Support/raw_ostream.h"
55	#include "llvm/Transforms/Scalar/LoopPassManager.h"
56	#include "llvm/Transforms/Utils/BasicBlockUtils.h"
57	#include "llvm/Transforms/Utils/Cloning.h"
58	#include "llvm/Transforms/Utils/Local.h"
59	#include "llvm/Transforms/Utils/LoopUtils.h"
60	#include "llvm/Transforms/Utils/ValueMapper.h"
61	#include <algorithm>
62	#include <cassert>
63	#include <iterator>
64	#include <numeric>
65	#include <optional>
66	#include <utility>
67
68	#define DEBUG_TYPE "simple-loop-unswitch"
69
70	using namespace llvm;
71	using namespace llvm::PatternMatch;
72
73	STATISTIC(NumBranches, "Number of branches unswitched");
74	STATISTIC(NumSwitches, "Number of switches unswitched");
75	STATISTIC(NumSelects, "Number of selects turned into branches for unswitching");
76	STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
77	STATISTIC(NumTrivial, "Number of unswitches that are trivial");
78	STATISTIC(
79	NumCostMultiplierSkipped,
80	"Number of unswitch candidates that had their cost multiplier skipped");
81	STATISTIC(NumInvariantConditionsInjected,
82	"Number of invariant conditions injected and unswitched");
83
84	static cl::opt<bool> EnableNonTrivialUnswitch(
85	"enable-nontrivial-unswitch", cl::init(Val: false), cl::Hidden,
86	cl::desc("Forcibly enables non-trivial loop unswitching rather than "
87	"following the configuration passed into the pass."));
88
89	static cl::opt<int>
90	UnswitchThreshold("unswitch-threshold", cl::init(Val: 50), cl::Hidden,
91	cl::desc("The cost threshold for unswitching a loop."));
92
93	static cl::opt<bool> EnableUnswitchCostMultiplier(
94	"enable-unswitch-cost-multiplier", cl::init(Val: true), cl::Hidden,
95	cl::desc("Enable unswitch cost multiplier that prohibits exponential "
96	"explosion in nontrivial unswitch."));
97	static cl::opt<int> UnswitchSiblingsToplevelDiv(
98	"unswitch-siblings-toplevel-div", cl::init(Val: 2), cl::Hidden,
99	cl::desc("Toplevel siblings divisor for cost multiplier."));
100	static cl::opt<int> UnswitchNumInitialUnscaledCandidates(
101	"unswitch-num-initial-unscaled-candidates", cl::init(Val: 8), cl::Hidden,
102	cl::desc("Number of unswitch candidates that are ignored when calculating "
103	"cost multiplier."));
104	static cl::opt<bool> UnswitchGuards(
105	"simple-loop-unswitch-guards", cl::init(Val: true), cl::Hidden,
106	cl::desc("If enabled, simple loop unswitching will also consider "
107	"llvm.experimental.guard intrinsics as unswitch candidates."));
108	static cl::opt<bool> DropNonTrivialImplicitNullChecks(
109	"simple-loop-unswitch-drop-non-trivial-implicit-null-checks",
110	cl::init(Val: false), cl::Hidden,
111	cl::desc("If enabled, drop make.implicit metadata in unswitched implicit "
112	"null checks to save time analyzing if we can keep it."));
113	static cl::opt<unsigned>
114	MSSAThreshold("simple-loop-unswitch-memoryssa-threshold",
115	cl::desc("Max number of memory uses to explore during "
116	"partial unswitching analysis"),
117	cl::init(Val: 100), cl::Hidden);
118	static cl::opt<bool> FreezeLoopUnswitchCond(
119	"freeze-loop-unswitch-cond", cl::init(Val: true), cl::Hidden,
120	cl::desc("If enabled, the freeze instruction will be added to condition "
121	"of loop unswitch to prevent miscompilation."));
122
123	static cl::opt<bool> InjectInvariantConditions(
124	"simple-loop-unswitch-inject-invariant-conditions", cl::Hidden,
125	cl::desc("Whether we should inject new invariants and unswitch them to "
126	"eliminate some existing (non-invariant) conditions."),
127	cl::init(Val: true));
128
129	static cl::opt<unsigned> InjectInvariantConditionHotnesThreshold(
130	"simple-loop-unswitch-inject-invariant-condition-hotness-threshold",
131	cl::Hidden, cl::desc("Only try to inject loop invariant conditions and "
132	"unswitch on them to eliminate branches that are "
133	"not-taken 1/<this option> times or less."),
134	cl::init(Val: 16));
135
136	AnalysisKey ShouldRunExtraSimpleLoopUnswitch::Key;
137	namespace {
138	struct CompareDesc {
139	BranchInst *Term;
140	Value *Invariant;
141	BasicBlock *InLoopSucc;
142
143	CompareDesc(BranchInst *Term, Value *Invariant, BasicBlock *InLoopSucc)
144	: Term(Term), Invariant(Invariant), InLoopSucc(InLoopSucc) {}
145	};
146
147	struct InjectedInvariant {
148	ICmpInst::Predicate Pred;
149	Value *LHS;
150	Value *RHS;
151	BasicBlock *InLoopSucc;
152
153	InjectedInvariant(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
154	BasicBlock *InLoopSucc)
155	: Pred(Pred), LHS(LHS), RHS(RHS), InLoopSucc(InLoopSucc) {}
156	};
157
158	struct NonTrivialUnswitchCandidate {
159	Instruction *TI = nullptr;
160	TinyPtrVector<Value *> Invariants;
161	std::optional<InstructionCost> Cost;
162	std::optional<InjectedInvariant> PendingInjection;
163	NonTrivialUnswitchCandidate(
164	Instruction *TI, ArrayRef<Value *> Invariants,
165	std::optional<InstructionCost> Cost = std::nullopt,
166	std::optional<InjectedInvariant> PendingInjection = std::nullopt)
167	: TI(TI), Invariants(Invariants), Cost(Cost),
168	PendingInjection(PendingInjection) {};
169
170	bool hasPendingInjection() const { return PendingInjection.has_value(); }
171	};
172	} // end anonymous namespace.
173
174	// Helper to skip (select x, true, false), which matches both a logical AND and
175	// OR and can confuse code that tries to determine if \p Cond is either a
176	// logical AND or OR but not both.
177	static Value *skipTrivialSelect(Value *Cond) {
178	Value *CondNext;
179	while (match(V: Cond, P: m_Select(C: m_Value(V&: CondNext), L: m_One(), R: m_Zero())))
180	Cond = CondNext;
181	return Cond;
182	}
183
184	/// Collect all of the loop invariant input values transitively used by the
185	/// homogeneous instruction graph from a given root.
186	///
187	/// This essentially walks from a root recursively through loop variant operands
188	/// which have perform the same logical operation (AND or OR) and finds all
189	/// inputs which are loop invariant. For some operations these can be
190	/// re-associated and unswitched out of the loop entirely.
191	static TinyPtrVector<Value *>
192	collectHomogenousInstGraphLoopInvariants(const Loop &L, Instruction &Root,
193	const LoopInfo &LI) {
194	assert(!L.isLoopInvariant(&Root) &&
195	"Only need to walk the graph if root itself is not invariant.");
196	TinyPtrVector<Value *> Invariants;
197
198	bool IsRootAnd = match(V: &Root, P: m_LogicalAnd());
199	bool IsRootOr = match(V: &Root, P: m_LogicalOr());
200
201	// Build a worklist and recurse through operators collecting invariants.
202	SmallVector<Instruction *, 4> Worklist;
203	SmallPtrSet<Instruction *, 8> Visited;
204	Worklist.push_back(Elt: &Root);
205	Visited.insert(Ptr: &Root);
206	do {
207	Instruction &I = *Worklist.pop_back_val();
208	for (Value *OpV : I.operand_values()) {
209	// Skip constants as unswitching isn't interesting for them.
210	if (isa<Constant>(Val: OpV))
211	continue;
212
213	// Add it to our result if loop invariant.
214	if (L.isLoopInvariant(V: OpV)) {
215	Invariants.push_back(NewVal: OpV);
216	continue;
217	}
218
219	// If not an instruction with the same opcode, nothing we can do.
220	Instruction *OpI = dyn_cast<Instruction>(Val: skipTrivialSelect(Cond: OpV));
221
222	if (OpI && ((IsRootAnd && match(V: OpI, P: m_LogicalAnd())) ||
223	(IsRootOr && match(V: OpI, P: m_LogicalOr())))) {
224	// Visit this operand.
225	if (Visited.insert(Ptr: OpI).second)
226	Worklist.push_back(Elt: OpI);
227	}
228	}
229	} while (!Worklist.empty());
230
231	return Invariants;
232	}
233
234	static void replaceLoopInvariantUses(const Loop &L, Value *Invariant,
235	Constant &Replacement) {
236	assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");
237
238	// Replace uses of LIC in the loop with the given constant.
239	// We use make_early_inc_range as set invalidates the iterator.
240	for (Use &U : llvm::make_early_inc_range(Range: Invariant->uses())) {
241	Instruction *UserI = dyn_cast<Instruction>(Val: U.getUser());
242
243	// Replace this use within the loop body.
244	if (UserI && L.contains(Inst: UserI))
245	U.set(&Replacement);
246	}
247	}
248
249	/// Check that all the LCSSA PHI nodes in the loop exit block have trivial
250	/// incoming values along this edge.
251	static bool areLoopExitPHIsLoopInvariant(const Loop &L,
252	const BasicBlock &ExitingBB,
253	const BasicBlock &ExitBB) {
254	for (const Instruction &I : ExitBB) {
255	auto *PN = dyn_cast<PHINode>(Val: &I);
256	if (!PN)
257	// No more PHIs to check.
258	return true;
259
260	// If the incoming value for this edge isn't loop invariant the unswitch
261	// won't be trivial.
262	if (!L.isLoopInvariant(V: PN->getIncomingValueForBlock(BB: &ExitingBB)))
263	return false;
264	}
265	llvm_unreachable("Basic blocks should never be empty!");
266	}
267
268	/// Copy a set of loop invariant values \p ToDuplicate and insert them at the
269	/// end of \p BB and conditionally branch on the copied condition. We only
270	/// branch on a single value.
271	static void buildPartialUnswitchConditionalBranch(
272	BasicBlock &BB, ArrayRef<Value *> Invariants, bool Direction,
273	BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, bool InsertFreeze,
274	const Instruction *I, AssumptionCache *AC, const DominatorTree &DT) {
275	IRBuilder<> IRB(&BB);
276
277	SmallVector<Value *> FrozenInvariants;
278	for (Value *Inv : Invariants) {
279	if (InsertFreeze && !isGuaranteedNotToBeUndefOrPoison(V: Inv, AC, CtxI: I, DT: &DT))
280	Inv = IRB.CreateFreeze(V: Inv, Name: Inv->getName() + ".fr");
281	FrozenInvariants.push_back(Elt: Inv);
282	}
283
284	Value *Cond = Direction ? IRB.CreateOr(Ops: FrozenInvariants)
285	: IRB.CreateAnd(Ops: FrozenInvariants);
286	IRB.CreateCondBr(Cond, True: Direction ? &UnswitchedSucc : &NormalSucc,
287	False: Direction ? &NormalSucc : &UnswitchedSucc);
288	}
289
290	/// Copy a set of loop invariant values, and conditionally branch on them.
291	static void buildPartialInvariantUnswitchConditionalBranch(
292	BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction,
293	BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L,
294	MemorySSAUpdater *MSSAU) {
295	ValueToValueMapTy VMap;
296	for (auto *Val : reverse(C&: ToDuplicate)) {
297	Instruction *Inst = cast<Instruction>(Val);
298	Instruction *NewInst = Inst->clone();
299	NewInst->insertInto(ParentBB: &BB, It: BB.end());
300	RemapInstruction(I: NewInst, VM&: VMap,
301	Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
302	VMap[Val] = NewInst;
303
304	if (!MSSAU)
305	continue;
306
307	MemorySSA *MSSA = MSSAU->getMemorySSA();
308	if (auto *MemUse =
309	dyn_cast_or_null<MemoryUse>(Val: MSSA->getMemoryAccess(I: Inst))) {
310	auto *DefiningAccess = MemUse->getDefiningAccess();
311	// Get the first defining access before the loop.
312	while (L.contains(BB: DefiningAccess->getBlock())) {
313	// If the defining access is a MemoryPhi, get the incoming
314	// value for the pre-header as defining access.
315	if (auto *MemPhi = dyn_cast<MemoryPhi>(Val: DefiningAccess))
316	DefiningAccess =
317	MemPhi->getIncomingValueForBlock(BB: L.getLoopPreheader());
318	else
319	DefiningAccess = cast<MemoryDef>(Val: DefiningAccess)->getDefiningAccess();
320	}
321	MSSAU->createMemoryAccessInBB(I: NewInst, Definition: DefiningAccess,
322	BB: NewInst->getParent(),
323	Point: MemorySSA::BeforeTerminator);
324	}
325	}
326
327	IRBuilder<> IRB(&BB);
328	Value *Cond = VMap[ToDuplicate[0]];
329	IRB.CreateCondBr(Cond, True: Direction ? &UnswitchedSucc : &NormalSucc,
330	False: Direction ? &NormalSucc : &UnswitchedSucc);
331	}
332
333	/// Rewrite the PHI nodes in an unswitched loop exit basic block.
334	///
335	/// Requires that the loop exit and unswitched basic block are the same, and
336	/// that the exiting block was a unique predecessor of that block. Rewrites the
337	/// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
338	/// PHI nodes from the old preheader that now contains the unswitched
339	/// terminator.
340	static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
341	BasicBlock &OldExitingBB,
342	BasicBlock &OldPH) {
343	for (PHINode &PN : UnswitchedBB.phis()) {
344	// When the loop exit is directly unswitched we just need to update the
345	// incoming basic block. We loop to handle weird cases with repeated
346	// incoming blocks, but expect to typically only have one operand here.
347	for (auto i : seq<int>(Begin: 0, End: PN.getNumOperands())) {
348	assert(PN.getIncomingBlock(i) == &OldExitingBB &&
349	"Found incoming block different from unique predecessor!");
350	PN.setIncomingBlock(i, BB: &OldPH);
351	}
352	}
353	}
354
355	/// Rewrite the PHI nodes in the loop exit basic block and the split off
356	/// unswitched block.
357	///
358	/// Because the exit block remains an exit from the loop, this rewrites the
359	/// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
360	/// nodes into the unswitched basic block to select between the value in the
361	/// old preheader and the loop exit.
362	static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
363	BasicBlock &UnswitchedBB,
364	BasicBlock &OldExitingBB,
365	BasicBlock &OldPH,
366	bool FullUnswitch) {
367	assert(&ExitBB != &UnswitchedBB &&
368	"Must have different loop exit and unswitched blocks!");
369	BasicBlock::iterator InsertPt = UnswitchedBB.begin();
370	for (PHINode &PN : ExitBB.phis()) {
371	auto *NewPN = PHINode::Create(Ty: PN.getType(), /*NumReservedValues*/ 2,
372	NameStr: PN.getName() + ".split");
373	NewPN->insertBefore(InsertPos: InsertPt);
374
375	// Walk backwards over the old PHI node's inputs to minimize the cost of
376	// removing each one. We have to do this weird loop manually so that we
377	// create the same number of new incoming edges in the new PHI as we expect
378	// each case-based edge to be included in the unswitched switch in some
379	// cases.
380	// FIXME: This is really, really gross. It would be much cleaner if LLVM
381	// allowed us to create a single entry for a predecessor block without
382	// having separate entries for each "edge" even though these edges are
383	// required to produce identical results.
384	for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
385	if (PN.getIncomingBlock(i) != &OldExitingBB)
386	continue;
387
388	Value *Incoming = PN.getIncomingValue(i);
389	if (FullUnswitch)
390	// No more edge from the old exiting block to the exit block.
391	PN.removeIncomingValue(Idx: i);
392
393	NewPN->addIncoming(V: Incoming, BB: &OldPH);
394	}
395
396	// Now replace the old PHI with the new one and wire the old one in as an
397	// input to the new one.
398	PN.replaceAllUsesWith(V: NewPN);
399	NewPN->addIncoming(V: &PN, BB: &ExitBB);
400	}
401	}
402
403	/// Hoist the current loop up to the innermost loop containing a remaining exit.
404	///
405	/// Because we've removed an exit from the loop, we may have changed the set of
406	/// loops reachable and need to move the current loop up the loop nest or even
407	/// to an entirely separate nest.
408	static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
409	DominatorTree &DT, LoopInfo &LI,
410	MemorySSAUpdater *MSSAU, ScalarEvolution *SE) {
411	// If the loop is already at the top level, we can't hoist it anywhere.
412	Loop *OldParentL = L.getParentLoop();
413	if (!OldParentL)
414	return;
415
416	SmallVector<BasicBlock *, 4> Exits;
417	L.getExitBlocks(ExitBlocks&: Exits);
418	Loop *NewParentL = nullptr;
419	for (auto *ExitBB : Exits)
420	if (Loop *ExitL = LI.getLoopFor(BB: ExitBB))
421	if (!NewParentL || NewParentL->contains(L: ExitL))
422	NewParentL = ExitL;
423
424	if (NewParentL == OldParentL)
425	return;
426
427	// The new parent loop (if different) should always contain the old one.
428	if (NewParentL)
429	assert(NewParentL->contains(OldParentL) &&
430	"Can only hoist this loop up the nest!");
431
432	// The preheader will need to move with the body of this loop. However,
433	// because it isn't in this loop we also need to update the primary loop map.
434	assert(OldParentL == LI.getLoopFor(&Preheader) &&
435	"Parent loop of this loop should contain this loop's preheader!");
436	LI.changeLoopFor(BB: &Preheader, L: NewParentL);
437
438	// Remove this loop from its old parent.
439	OldParentL->removeChildLoop(Child: &L);
440
441	// Add the loop either to the new parent or as a top-level loop.
442	if (NewParentL)
443	NewParentL->addChildLoop(NewChild: &L);
444	else
445	LI.addTopLevelLoop(New: &L);
446
447	// Remove this loops blocks from the old parent and every other loop up the
448	// nest until reaching the new parent. Also update all of these
449	// no-longer-containing loops to reflect the nesting change.
450	for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
451	OldContainingL = OldContainingL->getParentLoop()) {
452	llvm::erase_if(C&: OldContainingL->getBlocksVector(),
453	P: [&](const BasicBlock *BB) {
454	return BB == &Preheader || L.contains(BB);
455	});
456
457	OldContainingL->getBlocksSet().erase(Ptr: &Preheader);
458	for (BasicBlock *BB : L.blocks())
459	OldContainingL->getBlocksSet().erase(Ptr: BB);
460
461	// Because we just hoisted a loop out of this one, we have essentially
462	// created new exit paths from it. That means we need to form LCSSA PHI
463	// nodes for values used in the no-longer-nested loop.
464	formLCSSA(L&: *OldContainingL, DT, LI: &LI, SE);
465
466	// We shouldn't need to form dedicated exits because the exit introduced
467	// here is the (just split by unswitching) preheader. However, after trivial
468	// unswitching it is possible to get new non-dedicated exits out of parent
469	// loop so let's conservatively form dedicated exit blocks and figure out
470	// if we can optimize later.
471	formDedicatedExitBlocks(L: OldContainingL, DT: &DT, LI: &LI, MSSAU,
472	/*PreserveLCSSA*/ true);
473	}
474	}
475
476	// Return the top-most loop containing ExitBB and having ExitBB as exiting block
477	// or the loop containing ExitBB, if there is no parent loop containing ExitBB
478	// as exiting block.
479	static Loop *getTopMostExitingLoop(const BasicBlock *ExitBB,
480	const LoopInfo &LI) {
481	Loop *TopMost = LI.getLoopFor(BB: ExitBB);
482	Loop *Current = TopMost;
483	while (Current) {
484	if (Current->isLoopExiting(BB: ExitBB))
485	TopMost = Current;
486	Current = Current->getParentLoop();
487	}
488	return TopMost;
489	}
490
491	/// Unswitch a trivial branch if the condition is loop invariant.
492	///
493	/// This routine should only be called when loop code leading to the branch has
494	/// been validated as trivial (no side effects). This routine checks if the
495	/// condition is invariant and one of the successors is a loop exit. This
496	/// allows us to unswitch without duplicating the loop, making it trivial.
497	///
498	/// If this routine fails to unswitch the branch it returns false.
499	///
500	/// If the branch can be unswitched, this routine splits the preheader and
501	/// hoists the branch above that split. Preserves loop simplified form
502	/// (splitting the exit block as necessary). It simplifies the branch within
503	/// the loop to an unconditional branch but doesn't remove it entirely. Further
504	/// cleanup can be done with some simplifycfg like pass.
505	///
506	/// If `SE` is not null, it will be updated based on the potential loop SCEVs
507	/// invalidated by this.
508	static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
509	LoopInfo &LI, ScalarEvolution *SE,
510	MemorySSAUpdater *MSSAU) {
511	assert(BI.isConditional() && "Can only unswitch a conditional branch!");
512	LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n");
513
514	// The loop invariant values that we want to unswitch.
515	TinyPtrVector<Value *> Invariants;
516
517	// When true, we're fully unswitching the branch rather than just unswitching
518	// some input conditions to the branch.
519	bool FullUnswitch = false;
520
521	Value *Cond = skipTrivialSelect(Cond: BI.getCondition());
522	if (L.isLoopInvariant(V: Cond)) {
523	Invariants.push_back(NewVal: Cond);
524	FullUnswitch = true;
525	} else {
526	if (auto *CondInst = dyn_cast<Instruction>(Val: Cond))
527	Invariants = collectHomogenousInstGraphLoopInvariants(L, Root&: *CondInst, LI);
528	if (Invariants.empty()) {
529	LLVM_DEBUG(dbgs() << " Couldn't find invariant inputs!\n");
530	return false;
531	}
532	}
533
534	// Check that one of the branch's successors exits, and which one.
535	bool ExitDirection = true;
536	int LoopExitSuccIdx = 0;
537	auto *LoopExitBB = BI.getSuccessor(i: 0);
538	if (L.contains(BB: LoopExitBB)) {
539	ExitDirection = false;
540	LoopExitSuccIdx = 1;
541	LoopExitBB = BI.getSuccessor(i: 1);
542	if (L.contains(BB: LoopExitBB)) {
543	LLVM_DEBUG(dbgs() << " Branch doesn't exit the loop!\n");
544	return false;
545	}
546	}
547	auto *ContinueBB = BI.getSuccessor(i: 1 - LoopExitSuccIdx);
548	auto *ParentBB = BI.getParent();
549	if (!areLoopExitPHIsLoopInvariant(L, ExitingBB: *ParentBB, ExitBB: *LoopExitBB)) {
550	LLVM_DEBUG(dbgs() << " Loop exit PHI's aren't loop-invariant!\n");
551	return false;
552	}
553
554	// When unswitching only part of the branch's condition, we need the exit
555	// block to be reached directly from the partially unswitched input. This can
556	// be done when the exit block is along the true edge and the branch condition
557	// is a graph of `or` operations, or the exit block is along the false edge
558	// and the condition is a graph of `and` operations.
559	if (!FullUnswitch) {
560	if (ExitDirection ? !match(V: Cond, P: m_LogicalOr())
561	: !match(V: Cond, P: m_LogicalAnd())) {
562	LLVM_DEBUG(dbgs() << " Branch condition is in improper form for "
563	"non-full unswitch!\n");
564	return false;
565	}
566	}
567
568	LLVM_DEBUG({
569	dbgs() << " unswitching trivial invariant conditions for: " << BI
570	<< "\n";
571	for (Value *Invariant : Invariants) {
572	dbgs() << " " << *Invariant << " == true";
573	if (Invariant != Invariants.back())
574	dbgs() << " ||";
575	dbgs() << "\n";
576	}
577	});
578
579	// If we have scalar evolutions, we need to invalidate them including this
580	// loop, the loop containing the exit block and the topmost parent loop
581	// exiting via LoopExitBB.
582	if (SE) {
583	if (const Loop *ExitL = getTopMostExitingLoop(ExitBB: LoopExitBB, LI))
584	SE->forgetLoop(L: ExitL);
585	else
586	// Forget the entire nest as this exits the entire nest.
587	SE->forgetTopmostLoop(L: &L);
588	SE->forgetBlockAndLoopDispositions();
589	}
590
591	if (MSSAU && VerifyMemorySSA)
592	MSSAU->getMemorySSA()->verifyMemorySSA();
593
594	// Split the preheader, so that we know that there is a safe place to insert
595	// the conditional branch. We will change the preheader to have a conditional
596	// branch on LoopCond.
597	BasicBlock *OldPH = L.getLoopPreheader();
598	BasicBlock *NewPH = SplitEdge(From: OldPH, To: L.getHeader(), DT: &DT, LI: &LI, MSSAU);
599
600	// Now that we have a place to insert the conditional branch, create a place
601	// to branch to: this is the exit block out of the loop that we are
602	// unswitching. We need to split this if there are other loop predecessors.
603	// Because the loop is in simplified form, *any* other predecessor is enough.
604	BasicBlock *UnswitchedBB;
605	if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
606	assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
607	"A branch's parent isn't a predecessor!");
608	UnswitchedBB = LoopExitBB;
609	} else {
610	UnswitchedBB =
611	SplitBlock(Old: LoopExitBB, SplitPt: LoopExitBB->begin(), DT: &DT, LI: &LI, MSSAU, BBName: "", Before: false);
612	}
613
614	if (MSSAU && VerifyMemorySSA)
615	MSSAU->getMemorySSA()->verifyMemorySSA();
616
617	// Actually move the invariant uses into the unswitched position. If possible,
618	// we do this by moving the instructions, but when doing partial unswitching
619	// we do it by building a new merge of the values in the unswitched position.
620	OldPH->getTerminator()->eraseFromParent();
621	if (FullUnswitch) {
622	// If fully unswitching, we can use the existing branch instruction.
623	// Splice it into the old PH to gate reaching the new preheader and re-point
624	// its successors.
625	BI.moveBefore(BB&: *OldPH, I: OldPH->end());
626	BI.setCondition(Cond);
627	if (MSSAU) {
628	// Temporarily clone the terminator, to make MSSA update cheaper by
629	// separating "insert edge" updates from "remove edge" ones.
630	BI.clone()->insertInto(ParentBB, It: ParentBB->end());
631	} else {
632	// Create a new unconditional branch that will continue the loop as a new
633	// terminator.
634	BranchInst::Create(IfTrue: ContinueBB, InsertAtEnd: ParentBB);
635	}
636	BI.setSuccessor(idx: LoopExitSuccIdx, NewSucc: UnswitchedBB);
637	BI.setSuccessor(idx: 1 - LoopExitSuccIdx, NewSucc: NewPH);
638	} else {
639	// Only unswitching a subset of inputs to the condition, so we will need to
640	// build a new branch that merges the invariant inputs.
641	if (ExitDirection)
642	assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalOr()) &&
643	"Must have an `or` of `i1`s or `select i1 X, true, Y`s for the "
644	"condition!");
645	else
646	assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalAnd()) &&
647	"Must have an `and` of `i1`s or `select i1 X, Y, false`s for the"
648	" condition!");
649	buildPartialUnswitchConditionalBranch(
650	BB&: *OldPH, Invariants, Direction: ExitDirection, UnswitchedSucc&: *UnswitchedBB, NormalSucc&: *NewPH,
651	InsertFreeze: FreezeLoopUnswitchCond, I: OldPH->getTerminator(), AC: nullptr, DT);
652	}
653
654	// Update the dominator tree with the added edge.
655	DT.insertEdge(From: OldPH, To: UnswitchedBB);
656
657	// After the dominator tree was updated with the added edge, update MemorySSA
658	// if available.
659	if (MSSAU) {
660	SmallVector<CFGUpdate, 1> Updates;
661	Updates.push_back(Elt: {cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
662	MSSAU->applyInsertUpdates(Updates, DT);
663	}
664
665	// Finish updating dominator tree and memory ssa for full unswitch.
666	if (FullUnswitch) {
667	if (MSSAU) {
668	// Remove the cloned branch instruction.
669	ParentBB->getTerminator()->eraseFromParent();
670	// Create unconditional branch now.
671	BranchInst::Create(IfTrue: ContinueBB, InsertAtEnd: ParentBB);
672	MSSAU->removeEdge(From: ParentBB, To: LoopExitBB);
673	}
674	DT.deleteEdge(From: ParentBB, To: LoopExitBB);
675	}
676
677	if (MSSAU && VerifyMemorySSA)
678	MSSAU->getMemorySSA()->verifyMemorySSA();
679
680	// Rewrite the relevant PHI nodes.
681	if (UnswitchedBB == LoopExitBB)
682	rewritePHINodesForUnswitchedExitBlock(UnswitchedBB&: *UnswitchedBB, OldExitingBB&: *ParentBB, OldPH&: *OldPH);
683	else
684	rewritePHINodesForExitAndUnswitchedBlocks(ExitBB&: *LoopExitBB, UnswitchedBB&: *UnswitchedBB,
685	OldExitingBB&: *ParentBB, OldPH&: *OldPH, FullUnswitch);
686
687	// The constant we can replace all of our invariants with inside the loop
688	// body. If any of the invariants have a value other than this the loop won't
689	// be entered.
690	ConstantInt *Replacement = ExitDirection
691	? ConstantInt::getFalse(Context&: BI.getContext())
692	: ConstantInt::getTrue(Context&: BI.getContext());
693
694	// Since this is an i1 condition we can also trivially replace uses of it
695	// within the loop with a constant.
696	for (Value *Invariant : Invariants)
697	replaceLoopInvariantUses(L, Invariant, Replacement&: *Replacement);
698
699	// If this was full unswitching, we may have changed the nesting relationship
700	// for this loop so hoist it to its correct parent if needed.
701	if (FullUnswitch)
702	hoistLoopToNewParent(L, Preheader&: *NewPH, DT, LI, MSSAU, SE);
703
704	if (MSSAU && VerifyMemorySSA)
705	MSSAU->getMemorySSA()->verifyMemorySSA();
706
707	LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n");
708	++NumTrivial;
709	++NumBranches;
710	return true;
711	}
712
713	/// Unswitch a trivial switch if the condition is loop invariant.
714	///
715	/// This routine should only be called when loop code leading to the switch has
716	/// been validated as trivial (no side effects). This routine checks if the
717	/// condition is invariant and that at least one of the successors is a loop
718	/// exit. This allows us to unswitch without duplicating the loop, making it
719	/// trivial.
720	///
721	/// If this routine fails to unswitch the switch it returns false.
722	///
723	/// If the switch can be unswitched, this routine splits the preheader and
724	/// copies the switch above that split. If the default case is one of the
725	/// exiting cases, it copies the non-exiting cases and points them at the new
726	/// preheader. If the default case is not exiting, it copies the exiting cases
727	/// and points the default at the preheader. It preserves loop simplified form
728	/// (splitting the exit blocks as necessary). It simplifies the switch within
729	/// the loop by removing now-dead cases. If the default case is one of those
730	/// unswitched, it replaces its destination with a new basic block containing
731	/// only unreachable. Such basic blocks, while technically loop exits, are not
732	/// considered for unswitching so this is a stable transform and the same
733	/// switch will not be revisited. If after unswitching there is only a single
734	/// in-loop successor, the switch is further simplified to an unconditional
735	/// branch. Still more cleanup can be done with some simplifycfg like pass.
736	///
737	/// If `SE` is not null, it will be updated based on the potential loop SCEVs
738	/// invalidated by this.
739	static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
740	LoopInfo &LI, ScalarEvolution *SE,
741	MemorySSAUpdater *MSSAU) {
742	LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n");
743	Value *LoopCond = SI.getCondition();
744
745	// If this isn't switching on an invariant condition, we can't unswitch it.
746	if (!L.isLoopInvariant(V: LoopCond))
747	return false;
748
749	auto *ParentBB = SI.getParent();
750
751	// The same check must be used both for the default and the exit cases. We
752	// should never leave edges from the switch instruction to a basic block that
753	// we are unswitching, hence the condition used to determine the default case
754	// needs to also be used to populate ExitCaseIndices, which is then used to
755	// remove cases from the switch.
756	auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) {
757	// BBToCheck is not an exit block if it is inside loop L.
758	if (L.contains(BB: &BBToCheck))
759	return false;
760	// BBToCheck is not trivial to unswitch if its phis aren't loop invariant.
761	if (!areLoopExitPHIsLoopInvariant(L, ExitingBB: *ParentBB, ExitBB: BBToCheck))
762	return false;
763	// We do not unswitch a block that only has an unreachable statement, as
764	// it's possible this is a previously unswitched block. Only unswitch if
765	// either the terminator is not unreachable, or, if it is, it's not the only
766	// instruction in the block.
767	auto *TI = BBToCheck.getTerminator();
768	bool isUnreachable = isa<UnreachableInst>(Val: TI);
769	return !isUnreachable ||
770	(isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI));
771	};
772
773	SmallVector<int, 4> ExitCaseIndices;
774	for (auto Case : SI.cases())
775	if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor()))
776	ExitCaseIndices.push_back(Elt: Case.getCaseIndex());
777	BasicBlock *DefaultExitBB = nullptr;
778	SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight =
779	SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, idx: 0);
780	if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) {
781	DefaultExitBB = SI.getDefaultDest();
782	} else if (ExitCaseIndices.empty())
783	return false;
784
785	LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n");
786
787	if (MSSAU && VerifyMemorySSA)
788	MSSAU->getMemorySSA()->verifyMemorySSA();
789
790	// We may need to invalidate SCEVs for the outermost loop reached by any of
791	// the exits.
792	Loop *OuterL = &L;
793
794	if (DefaultExitBB) {
795	// Check the loop containing this exit.
796	Loop *ExitL = getTopMostExitingLoop(ExitBB: DefaultExitBB, LI);
797	if (!ExitL || ExitL->contains(L: OuterL))
798	OuterL = ExitL;
799	}
800	for (unsigned Index : ExitCaseIndices) {
801	auto CaseI = SI.case_begin() + Index;
802	// Compute the outer loop from this exit.
803	Loop *ExitL = getTopMostExitingLoop(ExitBB: CaseI->getCaseSuccessor(), LI);
804	if (!ExitL || ExitL->contains(L: OuterL))
805	OuterL = ExitL;
806	}
807
808	if (SE) {
809	if (OuterL)
810	SE->forgetLoop(L: OuterL);
811	else
812	SE->forgetTopmostLoop(L: &L);
813	}
814
815	if (DefaultExitBB) {
816	// Clear out the default destination temporarily to allow accurate
817	// predecessor lists to be examined below.
818	SI.setDefaultDest(nullptr);
819	}
820
821	// Store the exit cases into a separate data structure and remove them from
822	// the switch.
823	SmallVector<std::tuple<ConstantInt *, BasicBlock *,
824	SwitchInstProfUpdateWrapper::CaseWeightOpt>,
825	4> ExitCases;
826	ExitCases.reserve(N: ExitCaseIndices.size());
827	SwitchInstProfUpdateWrapper SIW(SI);
828	// We walk the case indices backwards so that we remove the last case first
829	// and don't disrupt the earlier indices.
830	for (unsigned Index : reverse(C&: ExitCaseIndices)) {
831	auto CaseI = SI.case_begin() + Index;
832	// Save the value of this case.
833	auto W = SIW.getSuccessorWeight(idx: CaseI->getSuccessorIndex());
834	ExitCases.emplace_back(Args: CaseI->getCaseValue(), Args: CaseI->getCaseSuccessor(), Args&: W);
835	// Delete the unswitched cases.
836	SIW.removeCase(I: CaseI);
837	}
838
839	// Check if after this all of the remaining cases point at the same
840	// successor.
841	BasicBlock *CommonSuccBB = nullptr;
842	if (SI.getNumCases() > 0 &&
843	all_of(Range: drop_begin(RangeOrContainer: SI.cases()), P: [&SI](const SwitchInst::CaseHandle &Case) {
844	return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor();
845	}))
846	CommonSuccBB = SI.case_begin()->getCaseSuccessor();
847	if (!DefaultExitBB) {
848	// If we're not unswitching the default, we need it to match any cases to
849	// have a common successor or if we have no cases it is the common
850	// successor.
851	if (SI.getNumCases() == 0)
852	CommonSuccBB = SI.getDefaultDest();
853	else if (SI.getDefaultDest() != CommonSuccBB)
854	CommonSuccBB = nullptr;
855	}
856
857	// Split the preheader, so that we know that there is a safe place to insert
858	// the switch.
859	BasicBlock *OldPH = L.getLoopPreheader();
860	BasicBlock *NewPH = SplitEdge(From: OldPH, To: L.getHeader(), DT: &DT, LI: &LI, MSSAU);
861	OldPH->getTerminator()->eraseFromParent();
862
863	// Now add the unswitched switch.
864	auto *NewSI = SwitchInst::Create(Value: LoopCond, Default: NewPH, NumCases: ExitCases.size(), InsertAtEnd: OldPH);
865	SwitchInstProfUpdateWrapper NewSIW(*NewSI);
866
867	// Rewrite the IR for the unswitched basic blocks. This requires two steps.
868	// First, we split any exit blocks with remaining in-loop predecessors. Then
869	// we update the PHIs in one of two ways depending on if there was a split.
870	// We walk in reverse so that we split in the same order as the cases
871	// appeared. This is purely for convenience of reading the resulting IR, but
872	// it doesn't cost anything really.
873	SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
874	SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
875	// Handle the default exit if necessary.
876	// FIXME: It'd be great if we could merge this with the loop below but LLVM's
877	// ranges aren't quite powerful enough yet.
878	if (DefaultExitBB) {
879	if (pred_empty(BB: DefaultExitBB)) {
880	UnswitchedExitBBs.insert(Ptr: DefaultExitBB);
881	rewritePHINodesForUnswitchedExitBlock(UnswitchedBB&: *DefaultExitBB, OldExitingBB&: *ParentBB, OldPH&: *OldPH);
882	} else {
883	auto *SplitBB =
884	SplitBlock(Old: DefaultExitBB, SplitPt: DefaultExitBB->begin(), DT: &DT, LI: &LI, MSSAU);
885	rewritePHINodesForExitAndUnswitchedBlocks(ExitBB&: *DefaultExitBB, UnswitchedBB&: *SplitBB,
886	OldExitingBB&: *ParentBB, OldPH&: *OldPH,
887	/*FullUnswitch*/ true);
888	DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
889	}
890	}
891	// Note that we must use a reference in the for loop so that we update the
892	// container.
893	for (auto &ExitCase : reverse(C&: ExitCases)) {
894	// Grab a reference to the exit block in the pair so that we can update it.
895	BasicBlock *ExitBB = std::get<1>(t&: ExitCase);
896
897	// If this case is the last edge into the exit block, we can simply reuse it
898	// as it will no longer be a loop exit. No mapping necessary.
899	if (pred_empty(BB: ExitBB)) {
900	// Only rewrite once.
901	if (UnswitchedExitBBs.insert(Ptr: ExitBB).second)
902	rewritePHINodesForUnswitchedExitBlock(UnswitchedBB&: *ExitBB, OldExitingBB&: *ParentBB, OldPH&: *OldPH);
903	continue;
904	}
905
906	// Otherwise we need to split the exit block so that we retain an exit
907	// block from the loop and a target for the unswitched condition.
908	BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
909	if (!SplitExitBB) {
910	// If this is the first time we see this, do the split and remember it.
911	SplitExitBB = SplitBlock(Old: ExitBB, SplitPt: ExitBB->begin(), DT: &DT, LI: &LI, MSSAU);
912	rewritePHINodesForExitAndUnswitchedBlocks(ExitBB&: *ExitBB, UnswitchedBB&: *SplitExitBB,
913	OldExitingBB&: *ParentBB, OldPH&: *OldPH,
914	/*FullUnswitch*/ true);
915	}
916	// Update the case pair to point to the split block.
917	std::get<1>(t&: ExitCase) = SplitExitBB;
918	}
919
920	// Now add the unswitched cases. We do this in reverse order as we built them
921	// in reverse order.
922	for (auto &ExitCase : reverse(C&: ExitCases)) {
923	ConstantInt *CaseVal = std::get<0>(t&: ExitCase);
924	BasicBlock *UnswitchedBB = std::get<1>(t&: ExitCase);
925
926	NewSIW.addCase(OnVal: CaseVal, Dest: UnswitchedBB, W: std::get<2>(t&: ExitCase));
927	}
928
929	// If the default was unswitched, re-point it and add explicit cases for
930	// entering the loop.
931	if (DefaultExitBB) {
932	NewSIW->setDefaultDest(DefaultExitBB);
933	NewSIW.setSuccessorWeight(idx: 0, W: DefaultCaseWeight);
934
935	// We removed all the exit cases, so we just copy the cases to the
936	// unswitched switch.
937	for (const auto &Case : SI.cases())
938	NewSIW.addCase(OnVal: Case.getCaseValue(), Dest: NewPH,
939	W: SIW.getSuccessorWeight(idx: Case.getSuccessorIndex()));
940	} else if (DefaultCaseWeight) {
941	// We have to set branch weight of the default case.
942	uint64_t SW = *DefaultCaseWeight;
943	for (const auto &Case : SI.cases()) {
944	auto W = SIW.getSuccessorWeight(idx: Case.getSuccessorIndex());
945	assert(W &&
946	"case weight must be defined as default case weight is defined");
947	SW += *W;
948	}
949	NewSIW.setSuccessorWeight(idx: 0, W: SW);
950	}
951
952	// If we ended up with a common successor for every path through the switch
953	// after unswitching, rewrite it to an unconditional branch to make it easy
954	// to recognize. Otherwise we potentially have to recognize the default case
955	// pointing at unreachable and other complexity.
956	if (CommonSuccBB) {
957	BasicBlock *BB = SI.getParent();
958	// We may have had multiple edges to this common successor block, so remove
959	// them as predecessors. We skip the first one, either the default or the
960	// actual first case.
961	bool SkippedFirst = DefaultExitBB == nullptr;
962	for (auto Case : SI.cases()) {
963	assert(Case.getCaseSuccessor() == CommonSuccBB &&
964	"Non-common successor!");
965	(void)Case;
966	if (!SkippedFirst) {
967	SkippedFirst = true;
968	continue;
969	}
970	CommonSuccBB->removePredecessor(Pred: BB,
971	/*KeepOneInputPHIs*/ true);
972	}
973	// Now nuke the switch and replace it with a direct branch.
974	SIW.eraseFromParent();
975	BranchInst::Create(IfTrue: CommonSuccBB, InsertAtEnd: BB);
976	} else if (DefaultExitBB) {
977	assert(SI.getNumCases() > 0 &&
978	"If we had no cases we'd have a common successor!");
979	// Move the last case to the default successor. This is valid as if the
980	// default got unswitched it cannot be reached. This has the advantage of
981	// being simple and keeping the number of edges from this switch to
982	// successors the same, and avoiding any PHI update complexity.
983	auto LastCaseI = std::prev(x: SI.case_end());
984
985	SI.setDefaultDest(LastCaseI->getCaseSuccessor());
986	SIW.setSuccessorWeight(
987	idx: 0, W: SIW.getSuccessorWeight(idx: LastCaseI->getSuccessorIndex()));
988	SIW.removeCase(I: LastCaseI);
989	}
990
991	// Walk the unswitched exit blocks and the unswitched split blocks and update
992	// the dominator tree based on the CFG edits. While we are walking unordered
993	// containers here, the API for applyUpdates takes an unordered list of
994	// updates and requires them to not contain duplicates.
995	SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
996	for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
997	DTUpdates.push_back(Elt: {DT.Delete, ParentBB, UnswitchedExitBB});
998	DTUpdates.push_back(Elt: {DT.Insert, OldPH, UnswitchedExitBB});
999	}
1000	for (auto SplitUnswitchedPair : SplitExitBBMap) {
1001	DTUpdates.push_back(Elt: {DT.Delete, ParentBB, SplitUnswitchedPair.first});
1002	DTUpdates.push_back(Elt: {DT.Insert, OldPH, SplitUnswitchedPair.second});
1003	}
1004
1005	if (MSSAU) {
1006	MSSAU->applyUpdates(Updates: DTUpdates, DT, /*UpdateDT=*/UpdateDTFirst: true);
1007	if (VerifyMemorySSA)
1008	MSSAU->getMemorySSA()->verifyMemorySSA();
1009	} else {
1010	DT.applyUpdates(Updates: DTUpdates);
1011	}
1012
1013	assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1014
1015	// We may have changed the nesting relationship for this loop so hoist it to
1016	// its correct parent if needed.
1017	hoistLoopToNewParent(L, Preheader&: *NewPH, DT, LI, MSSAU, SE);
1018
1019	if (MSSAU && VerifyMemorySSA)
1020	MSSAU->getMemorySSA()->verifyMemorySSA();
1021
1022	++NumTrivial;
1023	++NumSwitches;
1024	LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n");
1025	return true;
1026	}
1027
1028	/// This routine scans the loop to find a branch or switch which occurs before
1029	/// any side effects occur. These can potentially be unswitched without
1030	/// duplicating the loop. If a branch or switch is successfully unswitched the
1031	/// scanning continues to see if subsequent branches or switches have become
1032	/// trivial. Once all trivial candidates have been unswitched, this routine
1033	/// returns.
1034	///
1035	/// The return value indicates whether anything was unswitched (and therefore
1036	/// changed).
1037	///
1038	/// If `SE` is not null, it will be updated based on the potential loop SCEVs
1039	/// invalidated by this.
1040	static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
1041	LoopInfo &LI, ScalarEvolution *SE,
1042	MemorySSAUpdater *MSSAU) {
1043	bool Changed = false;
1044
1045	// If loop header has only one reachable successor we should keep looking for
1046	// trivial condition candidates in the successor as well. An alternative is
1047	// to constant fold conditions and merge successors into loop header (then we
1048	// only need to check header's terminator). The reason for not doing this in
1049	// LoopUnswitch pass is that it could potentially break LoopPassManager's
1050	// invariants. Folding dead branches could either eliminate the current loop
1051	// or make other loops unreachable. LCSSA form might also not be preserved
1052	// after deleting branches. The following code keeps traversing loop header's
1053	// successors until it finds the trivial condition candidate (condition that
1054	// is not a constant). Since unswitching generates branches with constant
1055	// conditions, this scenario could be very common in practice.
1056	BasicBlock *CurrentBB = L.getHeader();
1057	SmallPtrSet<BasicBlock *, 8> Visited;
1058	Visited.insert(Ptr: CurrentBB);
1059	do {
1060	// Check if there are any side-effecting instructions (e.g. stores, calls,
1061	// volatile loads) in the part of the loop that the code *would* execute
1062	// without unswitching.
1063	if (MSSAU) // Possible early exit with MSSA
1064	if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(BB: CurrentBB))
1065	if (!isa<MemoryPhi>(Val: *Defs->begin()) || (++Defs->begin() != Defs->end()))
1066	return Changed;
1067	if (llvm::any_of(Range&: *CurrentBB,
1068	P: [](Instruction &I) { return I.mayHaveSideEffects(); }))
1069	return Changed;
1070
1071	Instruction *CurrentTerm = CurrentBB->getTerminator();
1072
1073	if (auto *SI = dyn_cast<SwitchInst>(Val: CurrentTerm)) {
1074	// Don't bother trying to unswitch past a switch with a constant
1075	// condition. This should be removed prior to running this pass by
1076	// simplifycfg.
1077	if (isa<Constant>(Val: SI->getCondition()))
1078	return Changed;
1079
1080	if (!unswitchTrivialSwitch(L, SI&: *SI, DT, LI, SE, MSSAU))
1081	// Couldn't unswitch this one so we're done.
1082	return Changed;
1083
1084	// Mark that we managed to unswitch something.
1085	Changed = true;
1086
1087	// If unswitching turned the terminator into an unconditional branch then
1088	// we can continue. The unswitching logic specifically works to fold any
1089	// cases it can into an unconditional branch to make it easier to
1090	// recognize here.
1091	auto *BI = dyn_cast<BranchInst>(Val: CurrentBB->getTerminator());
1092	if (!BI || BI->isConditional())
1093	return Changed;
1094
1095	CurrentBB = BI->getSuccessor(i: 0);
1096	continue;
1097	}
1098
1099	auto *BI = dyn_cast<BranchInst>(Val: CurrentTerm);
1100	if (!BI)
1101	// We do not understand other terminator instructions.
1102	return Changed;
1103
1104	// Don't bother trying to unswitch past an unconditional branch or a branch
1105	// with a constant value. These should be removed by simplifycfg prior to
1106	// running this pass.
1107	if (!BI->isConditional() ||
1108	isa<Constant>(Val: skipTrivialSelect(Cond: BI->getCondition())))
1109	return Changed;
1110
1111	// Found a trivial condition candidate: non-foldable conditional branch. If
1112	// we fail to unswitch this, we can't do anything else that is trivial.
1113	if (!unswitchTrivialBranch(L, BI&: *BI, DT, LI, SE, MSSAU))
1114	return Changed;
1115
1116	// Mark that we managed to unswitch something.
1117	Changed = true;
1118
1119	// If we only unswitched some of the conditions feeding the branch, we won't
1120	// have collapsed it to a single successor.
1121	BI = cast<BranchInst>(Val: CurrentBB->getTerminator());
1122	if (BI->isConditional())
1123	return Changed;
1124
1125	// Follow the newly unconditional branch into its successor.
1126	CurrentBB = BI->getSuccessor(i: 0);
1127
1128	// When continuing, if we exit the loop or reach a previous visited block,
1129	// then we can not reach any trivial condition candidates (unfoldable
1130	// branch instructions or switch instructions) and no unswitch can happen.
1131	} while (L.contains(BB: CurrentBB) && Visited.insert(Ptr: CurrentBB).second);
1132
1133	return Changed;
1134	}
1135
1136	/// Build the cloned blocks for an unswitched copy of the given loop.
1137	///
1138	/// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
1139	/// after the split block (`SplitBB`) that will be used to select between the
1140	/// cloned and original loop.
1141	///
1142	/// This routine handles cloning all of the necessary loop blocks and exit
1143	/// blocks including rewriting their instructions and the relevant PHI nodes.
1144	/// Any loop blocks or exit blocks which are dominated by a different successor
1145	/// than the one for this clone of the loop blocks can be trivially skipped. We
1146	/// use the `DominatingSucc` map to determine whether a block satisfies that
1147	/// property with a simple map lookup.
1148	///
1149	/// It also correctly creates the unconditional branch in the cloned
1150	/// unswitched parent block to only point at the unswitched successor.
1151	///
1152	/// This does not handle most of the necessary updates to `LoopInfo`. Only exit
1153	/// block splitting is correctly reflected in `LoopInfo`, essentially all of
1154	/// the cloned blocks (and their loops) are left without full `LoopInfo`
1155	/// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
1156	/// blocks to them but doesn't create the cloned `DominatorTree` structure and
1157	/// instead the caller must recompute an accurate DT. It *does* correctly
1158	/// update the `AssumptionCache` provided in `AC`.
1159	static BasicBlock *buildClonedLoopBlocks(
1160	Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
1161	ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
1162	BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
1163	const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
1164	ValueToValueMapTy &VMap,
1165	SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
1166	DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU,
1167	ScalarEvolution *SE) {
1168	SmallVector<BasicBlock *, 4> NewBlocks;
1169	NewBlocks.reserve(N: L.getNumBlocks() + ExitBlocks.size());
1170
1171	// We will need to clone a bunch of blocks, wrap up the clone operation in
1172	// a helper.
1173	auto CloneBlock = [&](BasicBlock *OldBB) {
1174	// Clone the basic block and insert it before the new preheader.
1175	BasicBlock *NewBB = CloneBasicBlock(BB: OldBB, VMap, NameSuffix: ".us", F: OldBB->getParent());
1176	NewBB->moveBefore(MovePos: LoopPH);
1177
1178	// Record this block and the mapping.
1179	NewBlocks.push_back(Elt: NewBB);
1180	VMap[OldBB] = NewBB;
1181
1182	return NewBB;
1183	};
1184
1185	// We skip cloning blocks when they have a dominating succ that is not the
1186	// succ we are cloning for.
1187	auto SkipBlock = [&](BasicBlock *BB) {
1188	auto It = DominatingSucc.find(Val: BB);
1189	return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
1190	};
1191
1192	// First, clone the preheader.
1193	auto *ClonedPH = CloneBlock(LoopPH);
1194
1195	// Then clone all the loop blocks, skipping the ones that aren't necessary.
1196	for (auto *LoopBB : L.blocks())
1197	if (!SkipBlock(LoopBB))
1198	CloneBlock(LoopBB);
1199
1200	// Split all the loop exit edges so that when we clone the exit blocks, if
1201	// any of the exit blocks are *also* a preheader for some other loop, we
1202	// don't create multiple predecessors entering the loop header.
1203	for (auto *ExitBB : ExitBlocks) {
1204	if (SkipBlock(ExitBB))
1205	continue;
1206
1207	// When we are going to clone an exit, we don't need to clone all the
1208	// instructions in the exit block and we want to ensure we have an easy
1209	// place to merge the CFG, so split the exit first. This is always safe to
1210	// do because there cannot be any non-loop predecessors of a loop exit in
1211	// loop simplified form.
1212	auto *MergeBB = SplitBlock(Old: ExitBB, SplitPt: ExitBB->begin(), DT: &DT, LI: &LI, MSSAU);
1213
1214	// Rearrange the names to make it easier to write test cases by having the
1215	// exit block carry the suffix rather than the merge block carrying the
1216	// suffix.
1217	MergeBB->takeName(V: ExitBB);
1218	ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1219
1220	// Now clone the original exit block.
1221	auto *ClonedExitBB = CloneBlock(ExitBB);
1222	assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1223	"Exit block should have been split to have one successor!");
1224	assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1225	"Cloned exit block has the wrong successor!");
1226
1227	// Remap any cloned instructions and create a merge phi node for them.
1228	for (auto ZippedInsts : llvm::zip_first(
1229	t: llvm::make_range(x: ExitBB->begin(), y: std::prev(x: ExitBB->end())),
1230	u: llvm::make_range(x: ClonedExitBB->begin(),
1231	y: std::prev(x: ClonedExitBB->end())))) {
1232	Instruction &I = std::get<0>(t&: ZippedInsts);
1233	Instruction &ClonedI = std::get<1>(t&: ZippedInsts);
1234
1235	// The only instructions in the exit block should be PHI nodes and
1236	// potentially a landing pad.
1237	assert(
1238	(isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1239	"Bad instruction in exit block!");
1240	// We should have a value map between the instruction and its clone.
1241	assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1242
1243	// Forget SCEVs based on exit phis in case SCEV looked through the phi.
1244	if (SE && isa<PHINode>(Val: I))
1245	SE->forgetValue(V: &I);
1246
1247	auto *MergePN =
1248	PHINode::Create(Ty: I.getType(), /*NumReservedValues*/ 2, NameStr: ".us-phi");
1249	MergePN->insertBefore(InsertPos: MergeBB->getFirstInsertionPt());
1250	I.replaceAllUsesWith(V: MergePN);
1251	MergePN->addIncoming(V: &I, BB: ExitBB);
1252	MergePN->addIncoming(V: &ClonedI, BB: ClonedExitBB);
1253	}
1254	}
1255
1256	// Rewrite the instructions in the cloned blocks to refer to the instructions
1257	// in the cloned blocks. We have to do this as a second pass so that we have
1258	// everything available. Also, we have inserted new instructions which may
1259	// include assume intrinsics, so we update the assumption cache while
1260	// processing this.
1261	Module *M = ClonedPH->getParent()->getParent();
1262	for (auto *ClonedBB : NewBlocks)
1263	for (Instruction &I : *ClonedBB) {
1264	RemapDbgVariableRecordRange(M, Range: I.getDbgRecordRange(), VM&: VMap,
1265	Flags: RF_NoModuleLevelChanges |
1266	RF_IgnoreMissingLocals);
1267	RemapInstruction(I: &I, VM&: VMap,
1268	Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1269	if (auto *II = dyn_cast<AssumeInst>(Val: &I))
1270	AC.registerAssumption(CI: II);
1271	}
1272
1273	// Update any PHI nodes in the cloned successors of the skipped blocks to not
1274	// have spurious incoming values.
1275	for (auto *LoopBB : L.blocks())
1276	if (SkipBlock(LoopBB))
1277	for (auto *SuccBB : successors(BB: LoopBB))
1278	if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: SuccBB)))
1279	for (PHINode &PN : ClonedSuccBB->phis())
1280	PN.removeIncomingValue(BB: LoopBB, /*DeletePHIIfEmpty*/ false);
1281
1282	// Remove the cloned parent as a predecessor of any successor we ended up
1283	// cloning other than the unswitched one.
1284	auto *ClonedParentBB = cast<BasicBlock>(Val: VMap.lookup(Val: ParentBB));
1285	for (auto *SuccBB : successors(BB: ParentBB)) {
1286	if (SuccBB == UnswitchedSuccBB)
1287	continue;
1288
1289	auto *ClonedSuccBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: SuccBB));
1290	if (!ClonedSuccBB)
1291	continue;
1292
1293	ClonedSuccBB->removePredecessor(Pred: ClonedParentBB,
1294	/*KeepOneInputPHIs*/ true);
1295	}
1296
1297	// Replace the cloned branch with an unconditional branch to the cloned
1298	// unswitched successor.
1299	auto *ClonedSuccBB = cast<BasicBlock>(Val: VMap.lookup(Val: UnswitchedSuccBB));
1300	Instruction *ClonedTerminator = ClonedParentBB->getTerminator();
1301	// Trivial Simplification. If Terminator is a conditional branch and
1302	// condition becomes dead - erase it.
1303	Value *ClonedConditionToErase = nullptr;
1304	if (auto *BI = dyn_cast<BranchInst>(Val: ClonedTerminator))
1305	ClonedConditionToErase = BI->getCondition();
1306	else if (auto *SI = dyn_cast<SwitchInst>(Val: ClonedTerminator))
1307	ClonedConditionToErase = SI->getCondition();
1308
1309	ClonedTerminator->eraseFromParent();
1310	BranchInst::Create(IfTrue: ClonedSuccBB, InsertAtEnd: ClonedParentBB);
1311
1312	if (ClonedConditionToErase)
1313	RecursivelyDeleteTriviallyDeadInstructions(V: ClonedConditionToErase, TLI: nullptr,
1314	MSSAU);
1315
1316	// If there are duplicate entries in the PHI nodes because of multiple edges
1317	// to the unswitched successor, we need to nuke all but one as we replaced it
1318	// with a direct branch.
1319	for (PHINode &PN : ClonedSuccBB->phis()) {
1320	bool Found = false;
1321	// Loop over the incoming operands backwards so we can easily delete as we
1322	// go without invalidating the index.
1323	for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1324	if (PN.getIncomingBlock(i) != ClonedParentBB)
1325	continue;
1326	if (!Found) {
1327	Found = true;
1328	continue;
1329	}
1330	PN.removeIncomingValue(Idx: i, /*DeletePHIIfEmpty*/ false);
1331	}
1332	}
1333
1334	// Record the domtree updates for the new blocks.
1335	SmallPtrSet<BasicBlock *, 4> SuccSet;
1336	for (auto *ClonedBB : NewBlocks) {
1337	for (auto *SuccBB : successors(BB: ClonedBB))
1338	if (SuccSet.insert(Ptr: SuccBB).second)
1339	DTUpdates.push_back(Elt: {DominatorTree::Insert, ClonedBB, SuccBB});
1340	SuccSet.clear();
1341	}
1342
1343	return ClonedPH;
1344	}
1345
1346	/// Recursively clone the specified loop and all of its children.
1347	///
1348	/// The target parent loop for the clone should be provided, or can be null if
1349	/// the clone is a top-level loop. While cloning, all the blocks are mapped
1350	/// with the provided value map. The entire original loop must be present in
1351	/// the value map. The cloned loop is returned.
1352	static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1353	const ValueToValueMapTy &VMap, LoopInfo &LI) {
1354	auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1355	assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1356	ClonedL.reserveBlocks(size: OrigL.getNumBlocks());
1357	for (auto *BB : OrigL.blocks()) {
1358	auto *ClonedBB = cast<BasicBlock>(Val: VMap.lookup(Val: BB));
1359	ClonedL.addBlockEntry(BB: ClonedBB);
1360	if (LI.getLoopFor(BB) == &OrigL)
1361	LI.changeLoopFor(BB: ClonedBB, L: &ClonedL);
1362	}
1363	};
1364
1365	// We specially handle the first loop because it may get cloned into
1366	// a different parent and because we most commonly are cloning leaf loops.
1367	Loop *ClonedRootL = LI.AllocateLoop();
1368	if (RootParentL)
1369	RootParentL->addChildLoop(NewChild: ClonedRootL);
1370	else
1371	LI.addTopLevelLoop(New: ClonedRootL);
1372	AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1373
1374	if (OrigRootL.isInnermost())
1375	return ClonedRootL;
1376
1377	// If we have a nest, we can quickly clone the entire loop nest using an
1378	// iterative approach because it is a tree. We keep the cloned parent in the
1379	// data structure to avoid repeatedly querying through a map to find it.
1380	SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1381	// Build up the loops to clone in reverse order as we'll clone them from the
1382	// back.
1383	for (Loop *ChildL : llvm::reverse(C&: OrigRootL))
1384	LoopsToClone.push_back(Elt: {ClonedRootL, ChildL});
1385	do {
1386	Loop *ClonedParentL, *L;
1387	std::tie(args&: ClonedParentL, args&: L) = LoopsToClone.pop_back_val();
1388	Loop *ClonedL = LI.AllocateLoop();
1389	ClonedParentL->addChildLoop(NewChild: ClonedL);
1390	AddClonedBlocksToLoop(*L, *ClonedL);
1391	for (Loop *ChildL : llvm::reverse(C&: *L))
1392	LoopsToClone.push_back(Elt: {ClonedL, ChildL});
1393	} while (!LoopsToClone.empty());
1394
1395	return ClonedRootL;
1396	}
1397
1398	/// Build the cloned loops of an original loop from unswitching.
1399	///
1400	/// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1401	/// operation. We need to re-verify that there even is a loop (as the backedge
1402	/// may not have been cloned), and even if there are remaining backedges the
1403	/// backedge set may be different. However, we know that each child loop is
1404	/// undisturbed, we only need to find where to place each child loop within
1405	/// either any parent loop or within a cloned version of the original loop.
1406	///
1407	/// Because child loops may end up cloned outside of any cloned version of the
1408	/// original loop, multiple cloned sibling loops may be created. All of them
1409	/// are returned so that the newly introduced loop nest roots can be
1410	/// identified.
1411	static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1412	const ValueToValueMapTy &VMap, LoopInfo &LI,
1413	SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1414	Loop *ClonedL = nullptr;
1415
1416	auto *OrigPH = OrigL.getLoopPreheader();
1417	auto *OrigHeader = OrigL.getHeader();
1418
1419	auto *ClonedPH = cast<BasicBlock>(Val: VMap.lookup(Val: OrigPH));
1420	auto *ClonedHeader = cast<BasicBlock>(Val: VMap.lookup(Val: OrigHeader));
1421
1422	// We need to know the loops of the cloned exit blocks to even compute the
1423	// accurate parent loop. If we only clone exits to some parent of the
1424	// original parent, we want to clone into that outer loop. We also keep track
1425	// of the loops that our cloned exit blocks participate in.
1426	Loop *ParentL = nullptr;
1427	SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1428	SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1429	ClonedExitsInLoops.reserve(N: ExitBlocks.size());
1430	for (auto *ExitBB : ExitBlocks)
1431	if (auto *ClonedExitBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: ExitBB)))
1432	if (Loop *ExitL = LI.getLoopFor(BB: ExitBB)) {
1433	ExitLoopMap[ClonedExitBB] = ExitL;
1434	ClonedExitsInLoops.push_back(Elt: ClonedExitBB);
1435	if (!ParentL || (ParentL != ExitL && ParentL->contains(L: ExitL)))
1436	ParentL = ExitL;
1437	}
1438	assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1439	ParentL->contains(OrigL.getParentLoop())) &&
1440	"The computed parent loop should always contain (or be) the parent of "
1441	"the original loop.");
1442
1443	// We build the set of blocks dominated by the cloned header from the set of
1444	// cloned blocks out of the original loop. While not all of these will
1445	// necessarily be in the cloned loop, it is enough to establish that they
1446	// aren't in unreachable cycles, etc.
1447	SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1448	for (auto *BB : OrigL.blocks())
1449	if (auto *ClonedBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: BB)))
1450	ClonedLoopBlocks.insert(X: ClonedBB);
1451
1452	// Rebuild the set of blocks that will end up in the cloned loop. We may have
1453	// skipped cloning some region of this loop which can in turn skip some of
1454	// the backedges so we have to rebuild the blocks in the loop based on the
1455	// backedges that remain after cloning.
1456	SmallVector<BasicBlock *, 16> Worklist;
1457	SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1458	for (auto *Pred : predecessors(BB: ClonedHeader)) {
1459	// The only possible non-loop header predecessor is the preheader because
1460	// we know we cloned the loop in simplified form.
1461	if (Pred == ClonedPH)
1462	continue;
1463
1464	// Because the loop was in simplified form, the only non-loop predecessor
1465	// should be the preheader.
1466	assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1467	"header other than the preheader "
1468	"that is not part of the loop!");
1469
1470	// Insert this block into the loop set and on the first visit (and if it
1471	// isn't the header we're currently walking) put it into the worklist to
1472	// recurse through.
1473	if (BlocksInClonedLoop.insert(Ptr: Pred).second && Pred != ClonedHeader)
1474	Worklist.push_back(Elt: Pred);
1475	}
1476
1477	// If we had any backedges then there *is* a cloned loop. Put the header into
1478	// the loop set and then walk the worklist backwards to find all the blocks
1479	// that remain within the loop after cloning.
1480	if (!BlocksInClonedLoop.empty()) {
1481	BlocksInClonedLoop.insert(Ptr: ClonedHeader);
1482
1483	while (!Worklist.empty()) {
1484	BasicBlock *BB = Worklist.pop_back_val();
1485	assert(BlocksInClonedLoop.count(BB) &&
1486	"Didn't put block into the loop set!");
1487
1488	// Insert any predecessors that are in the possible set into the cloned
1489	// set, and if the insert is successful, add them to the worklist. Note
1490	// that we filter on the blocks that are definitely reachable via the
1491	// backedge to the loop header so we may prune out dead code within the
1492	// cloned loop.
1493	for (auto *Pred : predecessors(BB))
1494	if (ClonedLoopBlocks.count(key: Pred) &&
1495	BlocksInClonedLoop.insert(Ptr: Pred).second)
1496	Worklist.push_back(Elt: Pred);
1497	}
1498
1499	ClonedL = LI.AllocateLoop();
1500	if (ParentL) {
1501	ParentL->addBasicBlockToLoop(NewBB: ClonedPH, LI);
1502	ParentL->addChildLoop(NewChild: ClonedL);
1503	} else {
1504	LI.addTopLevelLoop(New: ClonedL);
1505	}
1506	NonChildClonedLoops.push_back(Elt: ClonedL);
1507
1508	ClonedL->reserveBlocks(size: BlocksInClonedLoop.size());
1509	// We don't want to just add the cloned loop blocks based on how we
1510	// discovered them. The original order of blocks was carefully built in
1511	// a way that doesn't rely on predecessor ordering. Rather than re-invent
1512	// that logic, we just re-walk the original blocks (and those of the child
1513	// loops) and filter them as we add them into the cloned loop.
1514	for (auto *BB : OrigL.blocks()) {
1515	auto *ClonedBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: BB));
1516	if (!ClonedBB || !BlocksInClonedLoop.count(Ptr: ClonedBB))
1517	continue;
1518
1519	// Directly add the blocks that are only in this loop.
1520	if (LI.getLoopFor(BB) == &OrigL) {
1521	ClonedL->addBasicBlockToLoop(NewBB: ClonedBB, LI);
1522	continue;
1523	}
1524
1525	// We want to manually add it to this loop and parents.
1526	// Registering it with LoopInfo will happen when we clone the top
1527	// loop for this block.
1528	for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1529	PL->addBlockEntry(BB: ClonedBB);
1530	}
1531
1532	// Now add each child loop whose header remains within the cloned loop. All
1533	// of the blocks within the loop must satisfy the same constraints as the
1534	// header so once we pass the header checks we can just clone the entire
1535	// child loop nest.
1536	for (Loop *ChildL : OrigL) {
1537	auto *ClonedChildHeader =
1538	cast_or_null<BasicBlock>(Val: VMap.lookup(Val: ChildL->getHeader()));
1539	if (!ClonedChildHeader || !BlocksInClonedLoop.count(Ptr: ClonedChildHeader))
1540	continue;
1541
1542	#ifndef NDEBUG
1543	// We should never have a cloned child loop header but fail to have
1544	// all of the blocks for that child loop.
1545	for (auto *ChildLoopBB : ChildL->blocks())
1546	assert(BlocksInClonedLoop.count(
1547	cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1548	"Child cloned loop has a header within the cloned outer "
1549	"loop but not all of its blocks!");
1550	#endif
1551
1552	cloneLoopNest(OrigRootL&: *ChildL, RootParentL: ClonedL, VMap, LI);
1553	}
1554	}
1555
1556	// Now that we've handled all the components of the original loop that were
1557	// cloned into a new loop, we still need to handle anything from the original
1558	// loop that wasn't in a cloned loop.
1559
1560	// Figure out what blocks are left to place within any loop nest containing
1561	// the unswitched loop. If we never formed a loop, the cloned PH is one of
1562	// them.
1563	SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1564	if (BlocksInClonedLoop.empty())
1565	UnloopedBlockSet.insert(Ptr: ClonedPH);
1566	for (auto *ClonedBB : ClonedLoopBlocks)
1567	if (!BlocksInClonedLoop.count(Ptr: ClonedBB))
1568	UnloopedBlockSet.insert(Ptr: ClonedBB);
1569
1570	// Copy the cloned exits and sort them in ascending loop depth, we'll work
1571	// backwards across these to process them inside out. The order shouldn't
1572	// matter as we're just trying to build up the map from inside-out; we use
1573	// the map in a more stably ordered way below.
1574	auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1575	llvm::sort(C&: OrderedClonedExitsInLoops, Comp: [&](BasicBlock *LHS, BasicBlock *RHS) {
1576	return ExitLoopMap.lookup(Val: LHS)->getLoopDepth() <
1577	ExitLoopMap.lookup(Val: RHS)->getLoopDepth();
1578	});
1579
1580	// Populate the existing ExitLoopMap with everything reachable from each
1581	// exit, starting from the inner most exit.
1582	while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1583	assert(Worklist.empty() && "Didn't clear worklist!");
1584
1585	BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1586	Loop *ExitL = ExitLoopMap.lookup(Val: ExitBB);
1587
1588	// Walk the CFG back until we hit the cloned PH adding everything reachable
1589	// and in the unlooped set to this exit block's loop.
1590	Worklist.push_back(Elt: ExitBB);
1591	do {
1592	BasicBlock *BB = Worklist.pop_back_val();
1593	// We can stop recursing at the cloned preheader (if we get there).
1594	if (BB == ClonedPH)
1595	continue;
1596
1597	for (BasicBlock *PredBB : predecessors(BB)) {
1598	// If this pred has already been moved to our set or is part of some
1599	// (inner) loop, no update needed.
1600	if (!UnloopedBlockSet.erase(Ptr: PredBB)) {
1601	assert(
1602	(BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1603	"Predecessor not mapped to a loop!");
1604	continue;
1605	}
1606
1607	// We just insert into the loop set here. We'll add these blocks to the
1608	// exit loop after we build up the set in an order that doesn't rely on
1609	// predecessor order (which in turn relies on use list order).
1610	bool Inserted = ExitLoopMap.insert(KV: {PredBB, ExitL}).second;
1611	(void)Inserted;
1612	assert(Inserted && "Should only visit an unlooped block once!");
1613
1614	// And recurse through to its predecessors.
1615	Worklist.push_back(Elt: PredBB);
1616	}
1617	} while (!Worklist.empty());
1618	}
1619
1620	// Now that the ExitLoopMap gives as mapping for all the non-looping cloned
1621	// blocks to their outer loops, walk the cloned blocks and the cloned exits
1622	// in their original order adding them to the correct loop.
1623
1624	// We need a stable insertion order. We use the order of the original loop
1625	// order and map into the correct parent loop.
1626	for (auto *BB : llvm::concat<BasicBlock *const>(
1627	Ranges: ArrayRef(ClonedPH), Ranges&: ClonedLoopBlocks, Ranges&: ClonedExitsInLoops))
1628	if (Loop *OuterL = ExitLoopMap.lookup(Val: BB))
1629	OuterL->addBasicBlockToLoop(NewBB: BB, LI);
1630
1631	#ifndef NDEBUG
1632	for (auto &BBAndL : ExitLoopMap) {
1633	auto *BB = BBAndL.first;
1634	auto *OuterL = BBAndL.second;
1635	assert(LI.getLoopFor(BB) == OuterL &&
1636	"Failed to put all blocks into outer loops!");
1637	}
1638	#endif
1639
1640	// Now that all the blocks are placed into the correct containing loop in the
1641	// absence of child loops, find all the potentially cloned child loops and
1642	// clone them into whatever outer loop we placed their header into.
1643	for (Loop *ChildL : OrigL) {
1644	auto *ClonedChildHeader =
1645	cast_or_null<BasicBlock>(Val: VMap.lookup(Val: ChildL->getHeader()));
1646	if (!ClonedChildHeader || BlocksInClonedLoop.count(Ptr: ClonedChildHeader))
1647	continue;
1648
1649	#ifndef NDEBUG
1650	for (auto *ChildLoopBB : ChildL->blocks())
1651	assert(VMap.count(ChildLoopBB) &&
1652	"Cloned a child loop header but not all of that loops blocks!");
1653	#endif
1654
1655	NonChildClonedLoops.push_back(Elt: cloneLoopNest(
1656	OrigRootL&: *ChildL, RootParentL: ExitLoopMap.lookup(Val: ClonedChildHeader), VMap, LI));
1657	}
1658	}
1659
1660	static void
1661	deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1662	ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1663	DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1664	// Find all the dead clones, and remove them from their successors.
1665	SmallVector<BasicBlock *, 16> DeadBlocks;
1666	for (BasicBlock *BB : llvm::concat<BasicBlock *const>(Ranges: L.blocks(), Ranges&: ExitBlocks))
1667	for (const auto &VMap : VMaps)
1668	if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(Val: VMap->lookup(Val: BB)))
1669	if (!DT.isReachableFromEntry(A: ClonedBB)) {
1670	for (BasicBlock *SuccBB : successors(BB: ClonedBB))
1671	SuccBB->removePredecessor(Pred: ClonedBB);
1672	DeadBlocks.push_back(Elt: ClonedBB);
1673	}
1674
1675	// Remove all MemorySSA in the dead blocks
1676	if (MSSAU) {
1677	SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(),
1678	DeadBlocks.end());
1679	MSSAU->removeBlocks(DeadBlocks: DeadBlockSet);
1680	}
1681
1682	// Drop any remaining references to break cycles.
1683	for (BasicBlock *BB : DeadBlocks)
1684	BB->dropAllReferences();
1685	// Erase them from the IR.
1686	for (BasicBlock *BB : DeadBlocks)
1687	BB->eraseFromParent();
1688	}
1689
1690	static void deleteDeadBlocksFromLoop(Loop &L,
1691	SmallVectorImpl<BasicBlock *> &ExitBlocks,
1692	DominatorTree &DT, LoopInfo &LI,
1693	MemorySSAUpdater *MSSAU,
1694	ScalarEvolution *SE,
1695	LPMUpdater &LoopUpdater) {
1696	// Find all the dead blocks tied to this loop, and remove them from their
1697	// successors.
1698	SmallSetVector<BasicBlock *, 8> DeadBlockSet;
1699
1700	// Start with loop/exit blocks and get a transitive closure of reachable dead
1701	// blocks.
1702	SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1703	ExitBlocks.end());
1704	DeathCandidates.append(in_start: L.blocks().begin(), in_end: L.blocks().end());
1705	while (!DeathCandidates.empty()) {
1706	auto *BB = DeathCandidates.pop_back_val();
1707	if (!DeadBlockSet.count(key: BB) && !DT.isReachableFromEntry(A: BB)) {
1708	for (BasicBlock *SuccBB : successors(BB)) {
1709	SuccBB->removePredecessor(Pred: BB);
1710	DeathCandidates.push_back(Elt: SuccBB);
1711	}
1712	DeadBlockSet.insert(X: BB);
1713	}
1714	}
1715
1716	// Remove all MemorySSA in the dead blocks
1717	if (MSSAU)
1718	MSSAU->removeBlocks(DeadBlocks: DeadBlockSet);
1719
1720	// Filter out the dead blocks from the exit blocks list so that it can be
1721	// used in the caller.
1722	llvm::erase_if(C&: ExitBlocks,
1723	P: [&](BasicBlock *BB) { return DeadBlockSet.count(key: BB); });
1724
1725	// Walk from this loop up through its parents removing all of the dead blocks.
1726	for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1727	for (auto *BB : DeadBlockSet)
1728	ParentL->getBlocksSet().erase(Ptr: BB);
1729	llvm::erase_if(C&: ParentL->getBlocksVector(),
1730	P: [&](BasicBlock *BB) { return DeadBlockSet.count(key: BB); });
1731	}
1732
1733	// Now delete the dead child loops. This raw delete will clear them
1734	// recursively.
1735	llvm::erase_if(C&: L.getSubLoopsVector(), P: [&](Loop *ChildL) {
1736	if (!DeadBlockSet.count(key: ChildL->getHeader()))
1737	return false;
1738
1739	assert(llvm::all_of(ChildL->blocks(),
1740	[&](BasicBlock *ChildBB) {
1741	return DeadBlockSet.count(ChildBB);
1742	}) &&
1743	"If the child loop header is dead all blocks in the child loop must "
1744	"be dead as well!");
1745	LoopUpdater.markLoopAsDeleted(L&: *ChildL, Name: ChildL->getName());
1746	if (SE)
1747	SE->forgetBlockAndLoopDispositions();
1748	LI.destroy(L: ChildL);
1749	return true;
1750	});
1751
1752	// Remove the loop mappings for the dead blocks and drop all the references
1753	// from these blocks to others to handle cyclic references as we start
1754	// deleting the blocks themselves.
1755	for (auto *BB : DeadBlockSet) {
1756	// Check that the dominator tree has already been updated.
1757	assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1758	LI.changeLoopFor(BB, L: nullptr);
1759	// Drop all uses of the instructions to make sure we won't have dangling
1760	// uses in other blocks.
1761	for (auto &I : *BB)
1762	if (!I.use_empty())
1763	I.replaceAllUsesWith(V: PoisonValue::get(T: I.getType()));
1764	BB->dropAllReferences();
1765	}
1766
1767	// Actually delete the blocks now that they've been fully unhooked from the
1768	// IR.
1769	for (auto *BB : DeadBlockSet)
1770	BB->eraseFromParent();
1771	}
1772
1773	/// Recompute the set of blocks in a loop after unswitching.
1774	///
1775	/// This walks from the original headers predecessors to rebuild the loop. We
1776	/// take advantage of the fact that new blocks can't have been added, and so we
1777	/// filter by the original loop's blocks. This also handles potentially
1778	/// unreachable code that we don't want to explore but might be found examining
1779	/// the predecessors of the header.
1780	///
1781	/// If the original loop is no longer a loop, this will return an empty set. If
1782	/// it remains a loop, all the blocks within it will be added to the set
1783	/// (including those blocks in inner loops).
1784	static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1785	LoopInfo &LI) {
1786	SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1787
1788	auto *PH = L.getLoopPreheader();
1789	auto *Header = L.getHeader();
1790
1791	// A worklist to use while walking backwards from the header.
1792	SmallVector<BasicBlock *, 16> Worklist;
1793
1794	// First walk the predecessors of the header to find the backedges. This will
1795	// form the basis of our walk.
1796	for (auto *Pred : predecessors(BB: Header)) {
1797	// Skip the preheader.
1798	if (Pred == PH)
1799	continue;
1800
1801	// Because the loop was in simplified form, the only non-loop predecessor
1802	// is the preheader.
1803	assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1804	"than the preheader that is not part of the "
1805	"loop!");
1806
1807	// Insert this block into the loop set and on the first visit and, if it
1808	// isn't the header we're currently walking, put it into the worklist to
1809	// recurse through.
1810	if (LoopBlockSet.insert(Ptr: Pred).second && Pred != Header)
1811	Worklist.push_back(Elt: Pred);
1812	}
1813
1814	// If no backedges were found, we're done.
1815	if (LoopBlockSet.empty())
1816	return LoopBlockSet;
1817
1818	// We found backedges, recurse through them to identify the loop blocks.
1819	while (!Worklist.empty()) {
1820	BasicBlock *BB = Worklist.pop_back_val();
1821	assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1822
1823	// No need to walk past the header.
1824	if (BB == Header)
1825	continue;
1826
1827	// Because we know the inner loop structure remains valid we can use the
1828	// loop structure to jump immediately across the entire nested loop.
1829	// Further, because it is in loop simplified form, we can directly jump
1830	// to its preheader afterward.
1831	if (Loop *InnerL = LI.getLoopFor(BB))
1832	if (InnerL != &L) {
1833	assert(L.contains(InnerL) &&
1834	"Should not reach a loop *outside* this loop!");
1835	// The preheader is the only possible predecessor of the loop so
1836	// insert it into the set and check whether it was already handled.
1837	auto *InnerPH = InnerL->getLoopPreheader();
1838	assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1839	"but not contain the inner loop "
1840	"preheader!");
1841	if (!LoopBlockSet.insert(Ptr: InnerPH).second)
1842	// The only way to reach the preheader is through the loop body
1843	// itself so if it has been visited the loop is already handled.
1844	continue;
1845
1846	// Insert all of the blocks (other than those already present) into
1847	// the loop set. We expect at least the block that led us to find the
1848	// inner loop to be in the block set, but we may also have other loop
1849	// blocks if they were already enqueued as predecessors of some other
1850	// outer loop block.
1851	for (auto *InnerBB : InnerL->blocks()) {
1852	if (InnerBB == BB) {
1853	assert(LoopBlockSet.count(InnerBB) &&
1854	"Block should already be in the set!");
1855	continue;
1856	}
1857
1858	LoopBlockSet.insert(Ptr: InnerBB);
1859	}
1860
1861	// Add the preheader to the worklist so we will continue past the
1862	// loop body.
1863	Worklist.push_back(Elt: InnerPH);
1864	continue;
1865	}
1866
1867	// Insert any predecessors that were in the original loop into the new
1868	// set, and if the insert is successful, add them to the worklist.
1869	for (auto *Pred : predecessors(BB))
1870	if (L.contains(BB: Pred) && LoopBlockSet.insert(Ptr: Pred).second)
1871	Worklist.push_back(Elt: Pred);
1872	}
1873
1874	assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1875
1876	// We've found all the blocks participating in the loop, return our completed
1877	// set.
1878	return LoopBlockSet;
1879	}
1880
1881	/// Rebuild a loop after unswitching removes some subset of blocks and edges.
1882	///
1883	/// The removal may have removed some child loops entirely but cannot have
1884	/// disturbed any remaining child loops. However, they may need to be hoisted
1885	/// to the parent loop (or to be top-level loops). The original loop may be
1886	/// completely removed.
1887	///
1888	/// The sibling loops resulting from this update are returned. If the original
1889	/// loop remains a valid loop, it will be the first entry in this list with all
1890	/// of the newly sibling loops following it.
1891	///
1892	/// Returns true if the loop remains a loop after unswitching, and false if it
1893	/// is no longer a loop after unswitching (and should not continue to be
1894	/// referenced).
1895	static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1896	LoopInfo &LI,
1897	SmallVectorImpl<Loop *> &HoistedLoops,
1898	ScalarEvolution *SE) {
1899	auto *PH = L.getLoopPreheader();
1900
1901	// Compute the actual parent loop from the exit blocks. Because we may have
1902	// pruned some exits the loop may be different from the original parent.
1903	Loop *ParentL = nullptr;
1904	SmallVector<Loop *, 4> ExitLoops;
1905	SmallVector<BasicBlock *, 4> ExitsInLoops;
1906	ExitsInLoops.reserve(N: ExitBlocks.size());
1907	for (auto *ExitBB : ExitBlocks)
1908	if (Loop *ExitL = LI.getLoopFor(BB: ExitBB)) {
1909	ExitLoops.push_back(Elt: ExitL);
1910	ExitsInLoops.push_back(Elt: ExitBB);
1911	if (!ParentL || (ParentL != ExitL && ParentL->contains(L: ExitL)))
1912	ParentL = ExitL;
1913	}
1914
1915	// Recompute the blocks participating in this loop. This may be empty if it
1916	// is no longer a loop.
1917	auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1918
1919	// If we still have a loop, we need to re-set the loop's parent as the exit
1920	// block set changing may have moved it within the loop nest. Note that this
1921	// can only happen when this loop has a parent as it can only hoist the loop
1922	// *up* the nest.
1923	if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1924	// Remove this loop's (original) blocks from all of the intervening loops.
1925	for (Loop *IL = L.getParentLoop(); IL != ParentL;
1926	IL = IL->getParentLoop()) {
1927	IL->getBlocksSet().erase(Ptr: PH);
1928	for (auto *BB : L.blocks())
1929	IL->getBlocksSet().erase(Ptr: BB);
1930	llvm::erase_if(C&: IL->getBlocksVector(), P: [&](BasicBlock *BB) {
1931	return BB == PH || L.contains(BB);
1932	});
1933	}
1934
1935	LI.changeLoopFor(BB: PH, L: ParentL);
1936	L.getParentLoop()->removeChildLoop(Child: &L);
1937	if (ParentL)
1938	ParentL->addChildLoop(NewChild: &L);
1939	else
1940	LI.addTopLevelLoop(New: &L);
1941	}
1942
1943	// Now we update all the blocks which are no longer within the loop.
1944	auto &Blocks = L.getBlocksVector();
1945	auto BlocksSplitI =
1946	LoopBlockSet.empty()
1947	? Blocks.begin()
1948	: std::stable_partition(
1949	first: Blocks.begin(), last: Blocks.end(),
1950	pred: [&](BasicBlock *BB) { return LoopBlockSet.count(Ptr: BB); });
1951
1952	// Before we erase the list of unlooped blocks, build a set of them.
1953	SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1954	if (LoopBlockSet.empty())
1955	UnloopedBlocks.insert(Ptr: PH);
1956
1957	// Now erase these blocks from the loop.
1958	for (auto *BB : make_range(x: BlocksSplitI, y: Blocks.end()))
1959	L.getBlocksSet().erase(Ptr: BB);
1960	Blocks.erase(first: BlocksSplitI, last: Blocks.end());
1961
1962	// Sort the exits in ascending loop depth, we'll work backwards across these
1963	// to process them inside out.
1964	llvm::stable_sort(Range&: ExitsInLoops, C: [&](BasicBlock *LHS, BasicBlock *RHS) {
1965	return LI.getLoopDepth(BB: LHS) < LI.getLoopDepth(BB: RHS);
1966	});
1967
1968	// We'll build up a set for each exit loop.
1969	SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1970	Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1971
1972	auto RemoveUnloopedBlocksFromLoop =
1973	[](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1974	for (auto *BB : UnloopedBlocks)
1975	L.getBlocksSet().erase(Ptr: BB);
1976	llvm::erase_if(C&: L.getBlocksVector(), P: [&](BasicBlock *BB) {
1977	return UnloopedBlocks.count(Ptr: BB);
1978	});
1979	};
1980
1981	SmallVector<BasicBlock *, 16> Worklist;
1982	while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1983	assert(Worklist.empty() && "Didn't clear worklist!");
1984	assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
1985
1986	// Grab the next exit block, in decreasing loop depth order.
1987	BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1988	Loop &ExitL = *LI.getLoopFor(BB: ExitBB);
1989	assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
1990
1991	// Erase all of the unlooped blocks from the loops between the previous
1992	// exit loop and this exit loop. This works because the ExitInLoops list is
1993	// sorted in increasing order of loop depth and thus we visit loops in
1994	// decreasing order of loop depth.
1995	for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
1996	RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
1997
1998	// Walk the CFG back until we hit the cloned PH adding everything reachable
1999	// and in the unlooped set to this exit block's loop.
2000	Worklist.push_back(Elt: ExitBB);
2001	do {
2002	BasicBlock *BB = Worklist.pop_back_val();
2003	// We can stop recursing at the cloned preheader (if we get there).
2004	if (BB == PH)
2005	continue;
2006
2007	for (BasicBlock *PredBB : predecessors(BB)) {
2008	// If this pred has already been moved to our set or is part of some
2009	// (inner) loop, no update needed.
2010	if (!UnloopedBlocks.erase(Ptr: PredBB)) {
2011	assert((NewExitLoopBlocks.count(PredBB) ||
2012	ExitL.contains(LI.getLoopFor(PredBB))) &&
2013	"Predecessor not in a nested loop (or already visited)!");
2014	continue;
2015	}
2016
2017	// We just insert into the loop set here. We'll add these blocks to the
2018	// exit loop after we build up the set in a deterministic order rather
2019	// than the predecessor-influenced visit order.
2020	bool Inserted = NewExitLoopBlocks.insert(Ptr: PredBB).second;
2021	(void)Inserted;
2022	assert(Inserted && "Should only visit an unlooped block once!");
2023
2024	// And recurse through to its predecessors.
2025	Worklist.push_back(Elt: PredBB);
2026	}
2027	} while (!Worklist.empty());
2028
2029	// If blocks in this exit loop were directly part of the original loop (as
2030	// opposed to a child loop) update the map to point to this exit loop. This
2031	// just updates a map and so the fact that the order is unstable is fine.
2032	for (auto *BB : NewExitLoopBlocks)
2033	if (Loop *BBL = LI.getLoopFor(BB))
2034	if (BBL == &L || !L.contains(L: BBL))
2035	LI.changeLoopFor(BB, L: &ExitL);
2036
2037	// We will remove the remaining unlooped blocks from this loop in the next
2038	// iteration or below.
2039	NewExitLoopBlocks.clear();
2040	}
2041
2042	// Any remaining unlooped blocks are no longer part of any loop unless they
2043	// are part of some child loop.
2044	for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
2045	RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
2046	for (auto *BB : UnloopedBlocks)
2047	if (Loop *BBL = LI.getLoopFor(BB))
2048	if (BBL == &L || !L.contains(L: BBL))
2049	LI.changeLoopFor(BB, L: nullptr);
2050
2051	// Sink all the child loops whose headers are no longer in the loop set to
2052	// the parent (or to be top level loops). We reach into the loop and directly
2053	// update its subloop vector to make this batch update efficient.
2054	auto &SubLoops = L.getSubLoopsVector();
2055	auto SubLoopsSplitI =
2056	LoopBlockSet.empty()
2057	? SubLoops.begin()
2058	: std::stable_partition(
2059	first: SubLoops.begin(), last: SubLoops.end(), pred: [&](Loop *SubL) {
2060	return LoopBlockSet.count(Ptr: SubL->getHeader());
2061	});
2062	for (auto *HoistedL : make_range(x: SubLoopsSplitI, y: SubLoops.end())) {
2063	HoistedLoops.push_back(Elt: HoistedL);
2064	HoistedL->setParentLoop(nullptr);
2065
2066	// To compute the new parent of this hoisted loop we look at where we
2067	// placed the preheader above. We can't lookup the header itself because we
2068	// retained the mapping from the header to the hoisted loop. But the
2069	// preheader and header should have the exact same new parent computed
2070	// based on the set of exit blocks from the original loop as the preheader
2071	// is a predecessor of the header and so reached in the reverse walk. And
2072	// because the loops were all in simplified form the preheader of the
2073	// hoisted loop can't be part of some *other* loop.
2074	if (auto *NewParentL = LI.getLoopFor(BB: HoistedL->getLoopPreheader()))
2075	NewParentL->addChildLoop(NewChild: HoistedL);
2076	else
2077	LI.addTopLevelLoop(New: HoistedL);
2078	}
2079	SubLoops.erase(first: SubLoopsSplitI, last: SubLoops.end());
2080
2081	// Actually delete the loop if nothing remained within it.
2082	if (Blocks.empty()) {
2083	assert(SubLoops.empty() &&
2084	"Failed to remove all subloops from the original loop!");
2085	if (Loop *ParentL = L.getParentLoop())
2086	ParentL->removeChildLoop(I: llvm::find(Range&: *ParentL, Val: &L));
2087	else
2088	LI.removeLoop(I: llvm::find(Range&: LI, Val: &L));
2089	// markLoopAsDeleted for L should be triggered by the caller (it is
2090	// typically done within postUnswitch).
2091	if (SE)
2092	SE->forgetBlockAndLoopDispositions();
2093	LI.destroy(L: &L);
2094	return false;
2095	}
2096
2097	return true;
2098	}
2099
2100	/// Helper to visit a dominator subtree, invoking a callable on each node.
2101	///
2102	/// Returning false at any point will stop walking past that node of the tree.
2103	template <typename CallableT>
2104	void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
2105	SmallVector<DomTreeNode *, 4> DomWorklist;
2106	DomWorklist.push_back(Elt: DT[BB]);
2107	#ifndef NDEBUG
2108	SmallPtrSet<DomTreeNode *, 4> Visited;
2109	Visited.insert(Ptr: DT[BB]);
2110	#endif
2111	do {
2112	DomTreeNode *N = DomWorklist.pop_back_val();
2113
2114	// Visit this node.
2115	if (!Callable(N->getBlock()))
2116	continue;
2117
2118	// Accumulate the child nodes.
2119	for (DomTreeNode *ChildN : *N) {
2120	assert(Visited.insert(ChildN).second &&
2121	"Cannot visit a node twice when walking a tree!");
2122	DomWorklist.push_back(Elt: ChildN);
2123	}
2124	} while (!DomWorklist.empty());
2125	}
2126
2127	void postUnswitch(Loop &L, LPMUpdater &U, StringRef LoopName,
2128	bool CurrentLoopValid, bool PartiallyInvariant,
2129	bool InjectedCondition, ArrayRef<Loop *> NewLoops) {
2130	// If we did a non-trivial unswitch, we have added new (cloned) loops.
2131	if (!NewLoops.empty())
2132	U.addSiblingLoops(NewSibLoops: NewLoops);
2133
2134	// If the current loop remains valid, we should revisit it to catch any
2135	// other unswitch opportunities. Otherwise, we need to mark it as deleted.
2136	if (CurrentLoopValid) {
2137	if (PartiallyInvariant) {
2138	// Mark the new loop as partially unswitched, to avoid unswitching on
2139	// the same condition again.
2140	auto &Context = L.getHeader()->getContext();
2141	MDNode *DisableUnswitchMD = MDNode::get(
2142	Context,
2143	MDs: MDString::get(Context, Str: "llvm.loop.unswitch.partial.disable"));
2144	MDNode *NewLoopID = makePostTransformationMetadata(
2145	Context, OrigLoopID: L.getLoopID(), RemovePrefixes: {"llvm.loop.unswitch.partial"},
2146	AddAttrs: {DisableUnswitchMD});
2147	L.setLoopID(NewLoopID);
2148	} else if (InjectedCondition) {
2149	// Do the same for injection of invariant conditions.
2150	auto &Context = L.getHeader()->getContext();
2151	MDNode *DisableUnswitchMD = MDNode::get(
2152	Context,
2153	MDs: MDString::get(Context, Str: "llvm.loop.unswitch.injection.disable"));
2154	MDNode *NewLoopID = makePostTransformationMetadata(
2155	Context, OrigLoopID: L.getLoopID(), RemovePrefixes: {"llvm.loop.unswitch.injection"},
2156	AddAttrs: {DisableUnswitchMD});
2157	L.setLoopID(NewLoopID);
2158	} else
2159	U.revisitCurrentLoop();
2160	} else
2161	U.markLoopAsDeleted(L, Name: LoopName);
2162	}
2163
2164	static void unswitchNontrivialInvariants(
2165	Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
2166	IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI,
2167	AssumptionCache &AC, ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
2168	LPMUpdater &LoopUpdater, bool InsertFreeze, bool InjectedCondition) {
2169	auto *ParentBB = TI.getParent();
2170	BranchInst *BI = dyn_cast<BranchInst>(Val: &TI);
2171	SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(Val: &TI);
2172
2173	// Save the current loop name in a variable so that we can report it even
2174	// after it has been deleted.
2175	std::string LoopName(L.getName());
2176
2177	// We can only unswitch switches, conditional branches with an invariant
2178	// condition, or combining invariant conditions with an instruction or
2179	// partially invariant instructions.
2180	assert((SI || (BI && BI->isConditional())) &&
2181	"Can only unswitch switches and conditional branch!");
2182	bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty();
2183	bool FullUnswitch =
2184	SI || (skipTrivialSelect(Cond: BI->getCondition()) == Invariants[0] &&
2185	!PartiallyInvariant);
2186	if (FullUnswitch)
2187	assert(Invariants.size() == 1 &&
2188	"Cannot have other invariants with full unswitching!");
2189	else
2190	assert(isa<Instruction>(skipTrivialSelect(BI->getCondition())) &&
2191	"Partial unswitching requires an instruction as the condition!");
2192
2193	if (MSSAU && VerifyMemorySSA)
2194	MSSAU->getMemorySSA()->verifyMemorySSA();
2195
2196	// Constant and BBs tracking the cloned and continuing successor. When we are
2197	// unswitching the entire condition, this can just be trivially chosen to
2198	// unswitch towards `true`. However, when we are unswitching a set of
2199	// invariants combined with `and` or `or` or partially invariant instructions,
2200	// the combining operation determines the best direction to unswitch: we want
2201	// to unswitch the direction that will collapse the branch.
2202	bool Direction = true;
2203	int ClonedSucc = 0;
2204	if (!FullUnswitch) {
2205	Value *Cond = skipTrivialSelect(Cond: BI->getCondition());
2206	(void)Cond;
2207	assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) ||
2208	PartiallyInvariant) &&
2209	"Only `or`, `and`, an `select`, partially invariant instructions "
2210	"can combine invariants being unswitched.");
2211	if (!match(V: Cond, P: m_LogicalOr())) {
2212	if (match(V: Cond, P: m_LogicalAnd()) ||
2213	(PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) {
2214	Direction = false;
2215	ClonedSucc = 1;
2216	}
2217	}
2218	}
2219
2220	BasicBlock *RetainedSuccBB =
2221	BI ? BI->getSuccessor(i: 1 - ClonedSucc) : SI->getDefaultDest();
2222	SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
2223	if (BI)
2224	UnswitchedSuccBBs.insert(X: BI->getSuccessor(i: ClonedSucc));
2225	else
2226	for (auto Case : SI->cases())
2227	if (Case.getCaseSuccessor() != RetainedSuccBB)
2228	UnswitchedSuccBBs.insert(X: Case.getCaseSuccessor());
2229
2230	assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
2231	"Should not unswitch the same successor we are retaining!");
2232
2233	// The branch should be in this exact loop. Any inner loop's invariant branch
2234	// should be handled by unswitching that inner loop. The caller of this
2235	// routine should filter out any candidates that remain (but were skipped for
2236	// whatever reason).
2237	assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
2238
2239	// Compute the parent loop now before we start hacking on things.
2240	Loop *ParentL = L.getParentLoop();
2241	// Get blocks in RPO order for MSSA update, before changing the CFG.
2242	LoopBlocksRPO LBRPO(&L);
2243	if (MSSAU)
2244	LBRPO.perform(LI: &LI);
2245
2246	// Compute the outer-most loop containing one of our exit blocks. This is the
2247	// furthest up our loopnest which can be mutated, which we will use below to
2248	// update things.
2249	Loop *OuterExitL = &L;
2250	SmallVector<BasicBlock *, 4> ExitBlocks;
2251	L.getUniqueExitBlocks(ExitBlocks);
2252	for (auto *ExitBB : ExitBlocks) {
2253	// ExitBB can be an exit block for several levels in the loop nest. Make
2254	// sure we find the top most.
2255	Loop *NewOuterExitL = getTopMostExitingLoop(ExitBB, LI);
2256	if (!NewOuterExitL) {
2257	// We exited the entire nest with this block, so we're done.
2258	OuterExitL = nullptr;
2259	break;
2260	}
2261	if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(L: OuterExitL))
2262	OuterExitL = NewOuterExitL;
2263	}
2264
2265	// At this point, we're definitely going to unswitch something so invalidate
2266	// any cached information in ScalarEvolution for the outer most loop
2267	// containing an exit block and all nested loops.
2268	if (SE) {
2269	if (OuterExitL)
2270	SE->forgetLoop(L: OuterExitL);
2271	else
2272	SE->forgetTopmostLoop(L: &L);
2273	SE->forgetBlockAndLoopDispositions();
2274	}
2275
2276	// If the edge from this terminator to a successor dominates that successor,
2277	// store a map from each block in its dominator subtree to it. This lets us
2278	// tell when cloning for a particular successor if a block is dominated by
2279	// some *other* successor with a single data structure. We use this to
2280	// significantly reduce cloning.
2281	SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
2282	for (auto *SuccBB : llvm::concat<BasicBlock *const>(Ranges: ArrayRef(RetainedSuccBB),
2283	Ranges&: UnswitchedSuccBBs))
2284	if (SuccBB->getUniquePredecessor() ||
2285	llvm::all_of(Range: predecessors(BB: SuccBB), P: [&](BasicBlock *PredBB) {
2286	return PredBB == ParentBB || DT.dominates(A: SuccBB, B: PredBB);
2287	}))
2288	visitDomSubTree(DT, BB: SuccBB, Callable: [&](BasicBlock *BB) {
2289	DominatingSucc[BB] = SuccBB;
2290	return true;
2291	});
2292
2293	// Split the preheader, so that we know that there is a safe place to insert
2294	// the conditional branch. We will change the preheader to have a conditional
2295	// branch on LoopCond. The original preheader will become the split point
2296	// between the unswitched versions, and we will have a new preheader for the
2297	// original loop.
2298	BasicBlock *SplitBB = L.getLoopPreheader();
2299	BasicBlock *LoopPH = SplitEdge(From: SplitBB, To: L.getHeader(), DT: &DT, LI: &LI, MSSAU);
2300
2301	// Keep track of the dominator tree updates needed.
2302	SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2303
2304	// Clone the loop for each unswitched successor.
2305	SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
2306	VMaps.reserve(N: UnswitchedSuccBBs.size());
2307	SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
2308	for (auto *SuccBB : UnswitchedSuccBBs) {
2309	VMaps.emplace_back(Args: new ValueToValueMapTy());
2310	ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2311	L, LoopPH, SplitBB, ExitBlocks, ParentBB, UnswitchedSuccBB: SuccBB, ContinueSuccBB: RetainedSuccBB,
2312	DominatingSucc, VMap&: *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU, SE);
2313	}
2314
2315	// Drop metadata if we may break its semantics by moving this instr into the
2316	// split block.
2317	if (TI.getMetadata(KindID: LLVMContext::MD_make_implicit)) {
2318	if (DropNonTrivialImplicitNullChecks)
2319	// Do not spend time trying to understand if we can keep it, just drop it
2320	// to save compile time.
2321	TI.setMetadata(KindID: LLVMContext::MD_make_implicit, Node: nullptr);
2322	else {
2323	// It is only legal to preserve make.implicit metadata if we are
2324	// guaranteed no reach implicit null check after following this branch.
2325	ICFLoopSafetyInfo SafetyInfo;
2326	SafetyInfo.computeLoopSafetyInfo(CurLoop: &L);
2327	if (!SafetyInfo.isGuaranteedToExecute(Inst: TI, DT: &DT, CurLoop: &L))
2328	TI.setMetadata(KindID: LLVMContext::MD_make_implicit, Node: nullptr);
2329	}
2330	}
2331
2332	// The stitching of the branched code back together depends on whether we're
2333	// doing full unswitching or not with the exception that we always want to
2334	// nuke the initial terminator placed in the split block.
2335	SplitBB->getTerminator()->eraseFromParent();
2336	if (FullUnswitch) {
2337	// Splice the terminator from the original loop and rewrite its
2338	// successors.
2339	TI.moveBefore(BB&: *SplitBB, I: SplitBB->end());
2340
2341	// Keep a clone of the terminator for MSSA updates.
2342	Instruction *NewTI = TI.clone();
2343	NewTI->insertInto(ParentBB, It: ParentBB->end());
2344
2345	// First wire up the moved terminator to the preheaders.
2346	if (BI) {
2347	BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2348	BI->setSuccessor(idx: ClonedSucc, NewSucc: ClonedPH);
2349	BI->setSuccessor(idx: 1 - ClonedSucc, NewSucc: LoopPH);
2350	Value *Cond = skipTrivialSelect(Cond: BI->getCondition());
2351	if (InsertFreeze)
2352	Cond = new FreezeInst(Cond, Cond->getName() + ".fr", BI->getIterator());
2353	BI->setCondition(Cond);
2354	DTUpdates.push_back(Elt: {DominatorTree::Insert, SplitBB, ClonedPH});
2355	} else {
2356	assert(SI && "Must either be a branch or switch!");
2357
2358	// Walk the cases and directly update their successors.
2359	assert(SI->getDefaultDest() == RetainedSuccBB &&
2360	"Not retaining default successor!");
2361	SI->setDefaultDest(LoopPH);
2362	for (const auto &Case : SI->cases())
2363	if (Case.getCaseSuccessor() == RetainedSuccBB)
2364	Case.setSuccessor(LoopPH);
2365	else
2366	Case.setSuccessor(ClonedPHs.find(Val: Case.getCaseSuccessor())->second);
2367
2368	if (InsertFreeze)
2369	SI->setCondition(new FreezeInst(SI->getCondition(),
2370	SI->getCondition()->getName() + ".fr",
2371	SI->getIterator()));
2372
2373	// We need to use the set to populate domtree updates as even when there
2374	// are multiple cases pointing at the same successor we only want to
2375	// remove and insert one edge in the domtree.
2376	for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2377	DTUpdates.push_back(
2378	Elt: {DominatorTree::Insert, SplitBB, ClonedPHs.find(Val: SuccBB)->second});
2379	}
2380
2381	if (MSSAU) {
2382	DT.applyUpdates(Updates: DTUpdates);
2383	DTUpdates.clear();
2384
2385	// Remove all but one edge to the retained block and all unswitched
2386	// blocks. This is to avoid having duplicate entries in the cloned Phis,
2387	// when we know we only keep a single edge for each case.
2388	MSSAU->removeDuplicatePhiEdgesBetween(From: ParentBB, To: RetainedSuccBB);
2389	for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2390	MSSAU->removeDuplicatePhiEdgesBetween(From: ParentBB, To: SuccBB);
2391
2392	for (auto &VMap : VMaps)
2393	MSSAU->updateForClonedLoop(LoopBlocks: LBRPO, ExitBlocks, VM: *VMap,
2394	/*IgnoreIncomingWithNoClones=*/true);
2395	MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2396
2397	// Remove all edges to unswitched blocks.
2398	for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2399	MSSAU->removeEdge(From: ParentBB, To: SuccBB);
2400	}
2401
2402	// Now unhook the successor relationship as we'll be replacing
2403	// the terminator with a direct branch. This is much simpler for branches
2404	// than switches so we handle those first.
2405	if (BI) {
2406	// Remove the parent as a predecessor of the unswitched successor.
2407	assert(UnswitchedSuccBBs.size() == 1 &&
2408	"Only one possible unswitched block for a branch!");
2409	BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2410	UnswitchedSuccBB->removePredecessor(Pred: ParentBB,
2411	/*KeepOneInputPHIs*/ true);
2412	DTUpdates.push_back(Elt: {DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2413	} else {
2414	// Note that we actually want to remove the parent block as a predecessor
2415	// of *every* case successor. The case successor is either unswitched,
2416	// completely eliminating an edge from the parent to that successor, or it
2417	// is a duplicate edge to the retained successor as the retained successor
2418	// is always the default successor and as we'll replace this with a direct
2419	// branch we no longer need the duplicate entries in the PHI nodes.
2420	SwitchInst *NewSI = cast<SwitchInst>(Val: NewTI);
2421	assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2422	"Not retaining default successor!");
2423	for (const auto &Case : NewSI->cases())
2424	Case.getCaseSuccessor()->removePredecessor(
2425	Pred: ParentBB,
2426	/*KeepOneInputPHIs*/ true);
2427
2428	// We need to use the set to populate domtree updates as even when there
2429	// are multiple cases pointing at the same successor we only want to
2430	// remove and insert one edge in the domtree.
2431	for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2432	DTUpdates.push_back(Elt: {DominatorTree::Delete, ParentBB, SuccBB});
2433	}
2434
2435	// After MSSAU update, remove the cloned terminator instruction NewTI.
2436	ParentBB->getTerminator()->eraseFromParent();
2437
2438	// Create a new unconditional branch to the continuing block (as opposed to
2439	// the one cloned).
2440	BranchInst::Create(IfTrue: RetainedSuccBB, InsertAtEnd: ParentBB);
2441	} else {
2442	assert(BI && "Only branches have partial unswitching.");
2443	assert(UnswitchedSuccBBs.size() == 1 &&
2444	"Only one possible unswitched block for a branch!");
2445	BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2446	// When doing a partial unswitch, we have to do a bit more work to build up
2447	// the branch in the split block.
2448	if (PartiallyInvariant)
2449	buildPartialInvariantUnswitchConditionalBranch(
2450	BB&: *SplitBB, ToDuplicate: Invariants, Direction, UnswitchedSucc&: *ClonedPH, NormalSucc&: *LoopPH, L, MSSAU);
2451	else {
2452	buildPartialUnswitchConditionalBranch(
2453	BB&: *SplitBB, Invariants, Direction, UnswitchedSucc&: *ClonedPH, NormalSucc&: *LoopPH,
2454	InsertFreeze: FreezeLoopUnswitchCond, I: BI, AC: &AC, DT);
2455	}
2456	DTUpdates.push_back(Elt: {DominatorTree::Insert, SplitBB, ClonedPH});
2457
2458	if (MSSAU) {
2459	DT.applyUpdates(Updates: DTUpdates);
2460	DTUpdates.clear();
2461
2462	// Perform MSSA cloning updates.
2463	for (auto &VMap : VMaps)
2464	MSSAU->updateForClonedLoop(LoopBlocks: LBRPO, ExitBlocks, VM: *VMap,
2465	/*IgnoreIncomingWithNoClones=*/true);
2466	MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2467	}
2468	}
2469
2470	// Apply the updates accumulated above to get an up-to-date dominator tree.
2471	DT.applyUpdates(Updates: DTUpdates);
2472
2473	// Now that we have an accurate dominator tree, first delete the dead cloned
2474	// blocks so that we can accurately build any cloned loops. It is important to
2475	// not delete the blocks from the original loop yet because we still want to
2476	// reference the original loop to understand the cloned loop's structure.
2477	deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2478
2479	// Build the cloned loop structure itself. This may be substantially
2480	// different from the original structure due to the simplified CFG. This also
2481	// handles inserting all the cloned blocks into the correct loops.
2482	SmallVector<Loop *, 4> NonChildClonedLoops;
2483	for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2484	buildClonedLoops(OrigL&: L, ExitBlocks, VMap: *VMap, LI, NonChildClonedLoops);
2485
2486	// Now that our cloned loops have been built, we can update the original loop.
2487	// First we delete the dead blocks from it and then we rebuild the loop
2488	// structure taking these deletions into account.
2489	deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, SE, LoopUpdater);
2490
2491	if (MSSAU && VerifyMemorySSA)
2492	MSSAU->getMemorySSA()->verifyMemorySSA();
2493
2494	SmallVector<Loop *, 4> HoistedLoops;
2495	bool IsStillLoop =
2496	rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops, SE);
2497
2498	if (MSSAU && VerifyMemorySSA)
2499	MSSAU->getMemorySSA()->verifyMemorySSA();
2500
2501	// This transformation has a high risk of corrupting the dominator tree, and
2502	// the below steps to rebuild loop structures will result in hard to debug
2503	// errors in that case so verify that the dominator tree is sane first.
2504	// FIXME: Remove this when the bugs stop showing up and rely on existing
2505	// verification steps.
2506	assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2507
2508	if (BI && !PartiallyInvariant) {
2509	// If we unswitched a branch which collapses the condition to a known
2510	// constant we want to replace all the uses of the invariants within both
2511	// the original and cloned blocks. We do this here so that we can use the
2512	// now updated dominator tree to identify which side the users are on.
2513	assert(UnswitchedSuccBBs.size() == 1 &&
2514	"Only one possible unswitched block for a branch!");
2515	BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2516
2517	// When considering multiple partially-unswitched invariants
2518	// we cant just go replace them with constants in both branches.
2519	//
2520	// For 'AND' we infer that true branch ("continue") means true
2521	// for each invariant operand.
2522	// For 'OR' we can infer that false branch ("continue") means false
2523	// for each invariant operand.
2524	// So it happens that for multiple-partial case we dont replace
2525	// in the unswitched branch.
2526	bool ReplaceUnswitched =
2527	FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant;
2528
2529	ConstantInt *UnswitchedReplacement =
2530	Direction ? ConstantInt::getTrue(Context&: BI->getContext())
2531	: ConstantInt::getFalse(Context&: BI->getContext());
2532	ConstantInt *ContinueReplacement =
2533	Direction ? ConstantInt::getFalse(Context&: BI->getContext())
2534	: ConstantInt::getTrue(Context&: BI->getContext());
2535	for (Value *Invariant : Invariants) {
2536	assert(!isa<Constant>(Invariant) &&
2537	"Should not be replacing constant values!");
2538	// Use make_early_inc_range here as set invalidates the iterator.
2539	for (Use &U : llvm::make_early_inc_range(Range: Invariant->uses())) {
2540	Instruction *UserI = dyn_cast<Instruction>(Val: U.getUser());
2541	if (!UserI)
2542	continue;
2543
2544	// Replace it with the 'continue' side if in the main loop body, and the
2545	// unswitched if in the cloned blocks.
2546	if (DT.dominates(A: LoopPH, B: UserI->getParent()))
2547	U.set(ContinueReplacement);
2548	else if (ReplaceUnswitched &&
2549	DT.dominates(A: ClonedPH, B: UserI->getParent()))
2550	U.set(UnswitchedReplacement);
2551	}
2552	}
2553	}
2554
2555	// We can change which blocks are exit blocks of all the cloned sibling
2556	// loops, the current loop, and any parent loops which shared exit blocks
2557	// with the current loop. As a consequence, we need to re-form LCSSA for
2558	// them. But we shouldn't need to re-form LCSSA for any child loops.
2559	// FIXME: This could be made more efficient by tracking which exit blocks are
2560	// new, and focusing on them, but that isn't likely to be necessary.
2561	//
2562	// In order to reasonably rebuild LCSSA we need to walk inside-out across the
2563	// loop nest and update every loop that could have had its exits changed. We
2564	// also need to cover any intervening loops. We add all of these loops to
2565	// a list and sort them by loop depth to achieve this without updating
2566	// unnecessary loops.
2567	auto UpdateLoop = [&](Loop &UpdateL) {
2568	#ifndef NDEBUG
2569	UpdateL.verifyLoop();
2570	for (Loop *ChildL : UpdateL) {
2571	ChildL->verifyLoop();
2572	assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2573	"Perturbed a child loop's LCSSA form!");
2574	}
2575	#endif
2576	// First build LCSSA for this loop so that we can preserve it when
2577	// forming dedicated exits. We don't want to perturb some other loop's
2578	// LCSSA while doing that CFG edit.
2579	formLCSSA(L&: UpdateL, DT, LI: &LI, SE);
2580
2581	// For loops reached by this loop's original exit blocks we may
2582	// introduced new, non-dedicated exits. At least try to re-form dedicated
2583	// exits for these loops. This may fail if they couldn't have dedicated
2584	// exits to start with.
2585	formDedicatedExitBlocks(L: &UpdateL, DT: &DT, LI: &LI, MSSAU, /*PreserveLCSSA*/ true);
2586	};
2587
2588	// For non-child cloned loops and hoisted loops, we just need to update LCSSA
2589	// and we can do it in any order as they don't nest relative to each other.
2590	//
2591	// Also check if any of the loops we have updated have become top-level loops
2592	// as that will necessitate widening the outer loop scope.
2593	for (Loop *UpdatedL :
2594	llvm::concat<Loop *>(Ranges&: NonChildClonedLoops, Ranges&: HoistedLoops)) {
2595	UpdateLoop(*UpdatedL);
2596	if (UpdatedL->isOutermost())
2597	OuterExitL = nullptr;
2598	}
2599	if (IsStillLoop) {
2600	UpdateLoop(L);
2601	if (L.isOutermost())
2602	OuterExitL = nullptr;
2603	}
2604
2605	// If the original loop had exit blocks, walk up through the outer most loop
2606	// of those exit blocks to update LCSSA and form updated dedicated exits.
2607	if (OuterExitL != &L)
2608	for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2609	OuterL = OuterL->getParentLoop())
2610	UpdateLoop(*OuterL);
2611
2612	#ifndef NDEBUG
2613	// Verify the entire loop structure to catch any incorrect updates before we
2614	// progress in the pass pipeline.
2615	LI.verify(DomTree: DT);
2616	#endif
2617
2618	// Now that we've unswitched something, make callbacks to report the changes.
2619	// For that we need to merge together the updated loops and the cloned loops
2620	// and check whether the original loop survived.
2621	SmallVector<Loop *, 4> SibLoops;
2622	for (Loop *UpdatedL : llvm::concat<Loop *>(Ranges&: NonChildClonedLoops, Ranges&: HoistedLoops))
2623	if (UpdatedL->getParentLoop() == ParentL)
2624	SibLoops.push_back(Elt: UpdatedL);
2625	postUnswitch(L, U&: LoopUpdater, LoopName, CurrentLoopValid: IsStillLoop, PartiallyInvariant,
2626	InjectedCondition, NewLoops: SibLoops);
2627
2628	if (MSSAU && VerifyMemorySSA)
2629	MSSAU->getMemorySSA()->verifyMemorySSA();
2630
2631	if (BI)
2632	++NumBranches;
2633	else
2634	++NumSwitches;
2635	}
2636
2637	/// Recursively compute the cost of a dominator subtree based on the per-block
2638	/// cost map provided.
2639	///
2640	/// The recursive computation is memozied into the provided DT-indexed cost map
2641	/// to allow querying it for most nodes in the domtree without it becoming
2642	/// quadratic.
2643	static InstructionCost computeDomSubtreeCost(
2644	DomTreeNode &N,
2645	const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap,
2646	SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) {
2647	// Don't accumulate cost (or recurse through) blocks not in our block cost
2648	// map and thus not part of the duplication cost being considered.
2649	auto BBCostIt = BBCostMap.find(Val: N.getBlock());
2650	if (BBCostIt == BBCostMap.end())
2651	return 0;
2652
2653	// Lookup this node to see if we already computed its cost.
2654	auto DTCostIt = DTCostMap.find(Val: &N);
2655	if (DTCostIt != DTCostMap.end())
2656	return DTCostIt->second;
2657
2658	// If not, we have to compute it. We can't use insert above and update
2659	// because computing the cost may insert more things into the map.
2660	InstructionCost Cost = std::accumulate(
2661	first: N.begin(), last: N.end(), init: BBCostIt->second,
2662	binary_op: [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost {
2663	return Sum + computeDomSubtreeCost(N&: *ChildN, BBCostMap, DTCostMap);
2664	});
2665	bool Inserted = DTCostMap.insert(KV: {&N, Cost}).second;
2666	(void)Inserted;
2667	assert(Inserted && "Should not insert a node while visiting children!");
2668	return Cost;
2669	}
2670
2671	/// Turns a select instruction into implicit control flow branch,
2672	/// making the following replacement:
2673	///
2674	/// head:
2675	/// --code before select--
2676	/// select %cond, %trueval, %falseval
2677	/// --code after select--
2678	///
2679	/// into
2680	///
2681	/// head:
2682	/// --code before select--
2683	/// br i1 %cond, label %then, label %tail
2684	///
2685	/// then:
2686	/// br %tail
2687	///
2688	/// tail:
2689	/// phi [ %trueval, %then ], [ %falseval, %head]
2690	/// unreachable
2691	///
2692	/// It also makes all relevant DT and LI updates, so that all structures are in
2693	/// valid state after this transform.
2694	static BranchInst *turnSelectIntoBranch(SelectInst *SI, DominatorTree &DT,
2695	LoopInfo &LI, MemorySSAUpdater *MSSAU,
2696	AssumptionCache *AC) {
2697	LLVM_DEBUG(dbgs() << "Turning " << *SI << " into a branch.\n");
2698	BasicBlock *HeadBB = SI->getParent();
2699
2700	DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2701	SplitBlockAndInsertIfThen(Cond: SI->getCondition(), SplitBefore: SI, Unreachable: false,
2702	BranchWeights: SI->getMetadata(KindID: LLVMContext::MD_prof), DTU: &DTU, LI: &LI);
2703	auto *CondBr = cast<BranchInst>(Val: HeadBB->getTerminator());
2704	BasicBlock *ThenBB = CondBr->getSuccessor(i: 0),
2705	*TailBB = CondBr->getSuccessor(i: 1);
2706	if (MSSAU)
2707	MSSAU->moveAllAfterSpliceBlocks(From: HeadBB, To: TailBB, Start: SI);
2708
2709	PHINode *Phi =
2710	PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "unswitched.select", InsertBefore: SI->getIterator());
2711	Phi->addIncoming(V: SI->getTrueValue(), BB: ThenBB);
2712	Phi->addIncoming(V: SI->getFalseValue(), BB: HeadBB);
2713	SI->replaceAllUsesWith(V: Phi);
2714	SI->eraseFromParent();
2715
2716	if (MSSAU && VerifyMemorySSA)
2717	MSSAU->getMemorySSA()->verifyMemorySSA();
2718
2719	++NumSelects;
2720	return CondBr;
2721	}
2722
2723	/// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2724	/// making the following replacement:
2725	///
2726	/// --code before guard--
2727	/// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2728	/// --code after guard--
2729	///
2730	/// into
2731	///
2732	/// --code before guard--
2733	/// br i1 %cond, label %guarded, label %deopt
2734	///
2735	/// guarded:
2736	/// --code after guard--
2737	///
2738	/// deopt:
2739	/// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2740	/// unreachable
2741	///
2742	/// It also makes all relevant DT and LI updates, so that all structures are in
2743	/// valid state after this transform.
2744	static BranchInst *turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2745	DominatorTree &DT, LoopInfo &LI,
2746	MemorySSAUpdater *MSSAU) {
2747	SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2748	LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2749	BasicBlock *CheckBB = GI->getParent();
2750
2751	if (MSSAU && VerifyMemorySSA)
2752	MSSAU->getMemorySSA()->verifyMemorySSA();
2753
2754	DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2755	Instruction *DeoptBlockTerm =
2756	SplitBlockAndInsertIfThen(Cond: GI->getArgOperand(i: 0), SplitBefore: GI, Unreachable: true,
2757	BranchWeights: GI->getMetadata(KindID: LLVMContext::MD_prof), DTU: &DTU, LI: &LI);
2758	BranchInst *CheckBI = cast<BranchInst>(Val: CheckBB->getTerminator());
2759	// SplitBlockAndInsertIfThen inserts control flow that branches to
2760	// DeoptBlockTerm if the condition is true. We want the opposite.
2761	CheckBI->swapSuccessors();
2762
2763	BasicBlock *GuardedBlock = CheckBI->getSuccessor(i: 0);
2764	GuardedBlock->setName("guarded");
2765	CheckBI->getSuccessor(i: 1)->setName("deopt");
2766	BasicBlock *DeoptBlock = CheckBI->getSuccessor(i: 1);
2767
2768	if (MSSAU)
2769	MSSAU->moveAllAfterSpliceBlocks(From: CheckBB, To: GuardedBlock, Start: GI);
2770
2771	GI->moveBefore(MovePos: DeoptBlockTerm);
2772	GI->setArgOperand(i: 0, v: ConstantInt::getFalse(Context&: GI->getContext()));
2773
2774	if (MSSAU) {
2775	MemoryDef *MD = cast<MemoryDef>(Val: MSSAU->getMemorySSA()->getMemoryAccess(I: GI));
2776	MSSAU->moveToPlace(What: MD, BB: DeoptBlock, Where: MemorySSA::BeforeTerminator);
2777	if (VerifyMemorySSA)
2778	MSSAU->getMemorySSA()->verifyMemorySSA();
2779	}
2780
2781	if (VerifyLoopInfo)
2782	LI.verify(DomTree: DT);
2783	++NumGuards;
2784	return CheckBI;
2785	}
2786
2787	/// Cost multiplier is a way to limit potentially exponential behavior
2788	/// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2789	/// candidates available. Also accounting for the number of "sibling" loops with
2790	/// the idea to account for previous unswitches that already happened on this
2791	/// cluster of loops. There was an attempt to keep this formula simple,
2792	/// just enough to limit the worst case behavior. Even if it is not that simple
2793	/// now it is still not an attempt to provide a detailed heuristic size
2794	/// prediction.
2795	///
2796	/// TODO: Make a proper accounting of "explosion" effect for all kinds of
2797	/// unswitch candidates, making adequate predictions instead of wild guesses.
2798	/// That requires knowing not just the number of "remaining" candidates but
2799	/// also costs of unswitching for each of these candidates.
2800	static int CalculateUnswitchCostMultiplier(
2801	const Instruction &TI, const Loop &L, const LoopInfo &LI,
2802	const DominatorTree &DT,
2803	ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates) {
2804
2805	// Guards and other exiting conditions do not contribute to exponential
2806	// explosion as soon as they dominate the latch (otherwise there might be
2807	// another path to the latch remaining that does not allow to eliminate the
2808	// loop copy on unswitch).
2809	const BasicBlock *Latch = L.getLoopLatch();
2810	const BasicBlock *CondBlock = TI.getParent();
2811	if (DT.dominates(A: CondBlock, B: Latch) &&
2812	(isGuard(U: &TI) ||
2813	(TI.isTerminator() &&
2814	llvm::count_if(Range: successors(I: &TI), P: [&L](const BasicBlock *SuccBB) {
2815	return L.contains(BB: SuccBB);
2816	}) <= 1))) {
2817	NumCostMultiplierSkipped++;
2818	return 1;
2819	}
2820
2821	auto *ParentL = L.getParentLoop();
2822	int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2823	: std::distance(first: LI.begin(), last: LI.end()));
2824	// Count amount of clones that all the candidates might cause during
2825	// unswitching. Branch/guard/select counts as 1, switch counts as log2 of its
2826	// cases.
2827	int UnswitchedClones = 0;
2828	for (const auto &Candidate : UnswitchCandidates) {
2829	const Instruction *CI = Candidate.TI;
2830	const BasicBlock *CondBlock = CI->getParent();
2831	bool SkipExitingSuccessors = DT.dominates(A: CondBlock, B: Latch);
2832	if (isa<SelectInst>(Val: CI)) {
2833	UnswitchedClones++;
2834	continue;
2835	}
2836	if (isGuard(U: CI)) {
2837	if (!SkipExitingSuccessors)
2838	UnswitchedClones++;
2839	continue;
2840	}
2841	int NonExitingSuccessors =
2842	llvm::count_if(Range: successors(BB: CondBlock),
2843	P: [SkipExitingSuccessors, &L](const BasicBlock *SuccBB) {
2844	return !SkipExitingSuccessors || L.contains(BB: SuccBB);
2845	});
2846	UnswitchedClones += Log2_32(Value: NonExitingSuccessors);
2847	}
2848
2849	// Ignore up to the "unscaled candidates" number of unswitch candidates
2850	// when calculating the power-of-two scaling of the cost. The main idea
2851	// with this control is to allow a small number of unswitches to happen
2852	// and rely more on siblings multiplier (see below) when the number
2853	// of candidates is small.
2854	unsigned ClonesPower =
2855	std::max(a: UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, b: 0);
2856
2857	// Allowing top-level loops to spread a bit more than nested ones.
2858	int SiblingsMultiplier =
2859	std::max(a: (ParentL ? SiblingsCount
2860	: SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2861	b: 1);
2862	// Compute the cost multiplier in a way that won't overflow by saturating
2863	// at an upper bound.
2864	int CostMultiplier;
2865	if (ClonesPower > Log2_32(Value: UnswitchThreshold) ||
2866	SiblingsMultiplier > UnswitchThreshold)
2867	CostMultiplier = UnswitchThreshold;
2868	else
2869	CostMultiplier = std::min(a: SiblingsMultiplier * (1 << ClonesPower),
2870	b: (int)UnswitchThreshold);
2871
2872	LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier
2873	<< " (siblings " << SiblingsMultiplier << " * clones "
2874	<< (1 << ClonesPower) << ")"
2875	<< " for unswitch candidate: " << TI << "\n");
2876	return CostMultiplier;
2877	}
2878
2879	static bool collectUnswitchCandidates(
2880	SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
2881	IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch,
2882	const Loop &L, const LoopInfo &LI, AAResults &AA,
2883	const MemorySSAUpdater *MSSAU) {
2884	assert(UnswitchCandidates.empty() && "Should be!");
2885
2886	auto AddUnswitchCandidatesForInst = [&](Instruction *I, Value *Cond) {
2887	Cond = skipTrivialSelect(Cond);
2888	if (isa<Constant>(Val: Cond))
2889	return;
2890	if (L.isLoopInvariant(V: Cond)) {
2891	UnswitchCandidates.push_back(Elt: {I, {Cond}});
2892	return;
2893	}
2894	if (match(V: Cond, P: m_CombineOr(L: m_LogicalAnd(), R: m_LogicalOr()))) {
2895	TinyPtrVector<Value *> Invariants =
2896	collectHomogenousInstGraphLoopInvariants(
2897	L, Root&: *static_cast<Instruction *>(Cond), LI);
2898	if (!Invariants.empty())
2899	UnswitchCandidates.push_back(Elt: {I, std::move(Invariants)});
2900	}
2901	};
2902
2903	// Whether or not we should also collect guards in the loop.
2904	bool CollectGuards = false;
2905	if (UnswitchGuards) {
2906	auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2907	Intrinsic::getName(Intrinsic::experimental_guard));
2908	if (GuardDecl && !GuardDecl->use_empty())
2909	CollectGuards = true;
2910	}
2911
2912	for (auto *BB : L.blocks()) {
2913	if (LI.getLoopFor(BB) != &L)
2914	continue;
2915
2916	for (auto &I : *BB) {
2917	if (auto *SI = dyn_cast<SelectInst>(Val: &I)) {
2918	auto *Cond = SI->getCondition();
2919	// Do not unswitch vector selects and logical and/or selects
2920	if (Cond->getType()->isIntegerTy(Bitwidth: 1) && !SI->getType()->isIntegerTy(Bitwidth: 1))
2921	AddUnswitchCandidatesForInst(SI, Cond);
2922	} else if (CollectGuards && isGuard(U: &I)) {
2923	auto *Cond =
2924	skipTrivialSelect(Cond: cast<IntrinsicInst>(Val: &I)->getArgOperand(i: 0));
2925	// TODO: Support AND, OR conditions and partial unswitching.
2926	if (!isa<Constant>(Val: Cond) && L.isLoopInvariant(V: Cond))
2927	UnswitchCandidates.push_back(Elt: {&I, {Cond}});
2928	}
2929	}
2930
2931	if (auto *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator())) {
2932	// We can only consider fully loop-invariant switch conditions as we need
2933	// to completely eliminate the switch after unswitching.
2934	if (!isa<Constant>(Val: SI->getCondition()) &&
2935	L.isLoopInvariant(V: SI->getCondition()) && !BB->getUniqueSuccessor())
2936	UnswitchCandidates.push_back(Elt: {SI, {SI->getCondition()}});
2937	continue;
2938	}
2939
2940	auto *BI = dyn_cast<BranchInst>(Val: BB->getTerminator());
2941	if (!BI || !BI->isConditional() ||
2942	BI->getSuccessor(i: 0) == BI->getSuccessor(i: 1))
2943	continue;
2944
2945	AddUnswitchCandidatesForInst(BI, BI->getCondition());
2946	}
2947
2948	if (MSSAU && !findOptionMDForLoop(TheLoop: &L, Name: "llvm.loop.unswitch.partial.disable") &&
2949	!any_of(Range&: UnswitchCandidates, P: [&L](auto &TerminatorAndInvariants) {
2950	return TerminatorAndInvariants.TI == L.getHeader()->getTerminator();
2951	})) {
2952	MemorySSA *MSSA = MSSAU->getMemorySSA();
2953	if (auto Info = hasPartialIVCondition(L, MSSAThreshold, MSSA: *MSSA, AA)) {
2954	LLVM_DEBUG(
2955	dbgs() << "simple-loop-unswitch: Found partially invariant condition "
2956	<< *Info->InstToDuplicate[0] << "\n");
2957	PartialIVInfo = *Info;
2958	PartialIVCondBranch = L.getHeader()->getTerminator();
2959	TinyPtrVector<Value *> ValsToDuplicate;
2960	llvm::append_range(C&: ValsToDuplicate, R&: Info->InstToDuplicate);
2961	UnswitchCandidates.push_back(
2962	Elt: {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)});
2963	}
2964	}
2965	return !UnswitchCandidates.empty();
2966	}
2967
2968	/// Tries to canonicalize condition described by:
2969	///
2970	/// br (LHS pred RHS), label IfTrue, label IfFalse
2971	///
2972	/// into its equivalent where `Pred` is something that we support for injected
2973	/// invariants (so far it is limited to ult), LHS in canonicalized form is
2974	/// non-invariant and RHS is an invariant.
2975	static void canonicalizeForInvariantConditionInjection(
2976	ICmpInst::Predicate &Pred, Value *&LHS, Value *&RHS, BasicBlock *&IfTrue,
2977	BasicBlock *&IfFalse, const Loop &L) {
2978	if (!L.contains(BB: IfTrue)) {
2979	Pred = ICmpInst::getInversePredicate(pred: Pred);
2980	std::swap(a&: IfTrue, b&: IfFalse);
2981	}
2982
2983	// Move loop-invariant argument to RHS position.
2984	if (L.isLoopInvariant(V: LHS)) {
2985	Pred = ICmpInst::getSwappedPredicate(pred: Pred);
2986	std::swap(a&: LHS, b&: RHS);
2987	}
2988
2989	if (Pred == ICmpInst::ICMP_SGE && match(V: RHS, P: m_Zero())) {
2990	// Turn "x >=s 0" into "x <u UMIN_INT"
2991	Pred = ICmpInst::ICMP_ULT;
2992	RHS = ConstantInt::get(
2993	Context&: RHS->getContext(),
2994	V: APInt::getSignedMinValue(numBits: RHS->getType()->getIntegerBitWidth()));
2995	}
2996	}
2997
2998	/// Returns true, if predicate described by ( \p Pred, \p LHS, \p RHS )
2999	/// succeeding into blocks ( \p IfTrue, \p IfFalse) can be optimized by
3000	/// injecting a loop-invariant condition.
3001	static bool shouldTryInjectInvariantCondition(
3002	const ICmpInst::Predicate Pred, const Value *LHS, const Value *RHS,
3003	const BasicBlock *IfTrue, const BasicBlock *IfFalse, const Loop &L) {
3004	if (L.isLoopInvariant(V: LHS) || !L.isLoopInvariant(V: RHS))
3005	return false;
3006	// TODO: Support other predicates.
3007	if (Pred != ICmpInst::ICMP_ULT)
3008	return false;
3009	// TODO: Support non-loop-exiting branches?
3010	if (!L.contains(BB: IfTrue) || L.contains(BB: IfFalse))
3011	return false;
3012	// FIXME: For some reason this causes problems with MSSA updates, need to
3013	// investigate why. So far, just don't unswitch latch.
3014	if (L.getHeader() == IfTrue)
3015	return false;
3016	return true;
3017	}
3018
3019	/// Returns true, if metadata on \p BI allows us to optimize branching into \p
3020	/// TakenSucc via injection of invariant conditions. The branch should be not
3021	/// enough and not previously unswitched, the information about this comes from
3022	/// the metadata.
3023	bool shouldTryInjectBasingOnMetadata(const BranchInst *BI,
3024	const BasicBlock *TakenSucc) {
3025	SmallVector<uint32_t> Weights;
3026	if (!extractBranchWeights(I: *BI, Weights))
3027	return false;
3028	unsigned T = InjectInvariantConditionHotnesThreshold;
3029	BranchProbability LikelyTaken(T - 1, T);
3030
3031	assert(Weights.size() == 2 && "Unexpected profile data!");
3032	size_t Idx = BI->getSuccessor(i: 0) == TakenSucc ? 0 : 1;
3033	auto Num = Weights[Idx];
3034	auto Denom = Weights[0] + Weights[1];
3035	// Degenerate or overflowed metadata.
3036	if (Denom == 0 || Num > Denom)
3037	return false;
3038	BranchProbability ActualTaken(Num, Denom);
3039	if (LikelyTaken > ActualTaken)
3040	return false;
3041	return true;
3042	}
3043
3044	/// Materialize pending invariant condition of the given candidate into IR. The
3045	/// injected loop-invariant condition implies the original loop-variant branch
3046	/// condition, so the materialization turns
3047	///
3048	/// loop_block:
3049	/// ...
3050	/// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3051	///
3052	/// into
3053	///
3054	/// preheader:
3055	/// %invariant_cond = LHS pred RHS
3056	/// ...
3057	/// loop_block:
3058	/// br i1 %invariant_cond, label InLoopSucc, label OriginalCheck
3059	/// OriginalCheck:
3060	/// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3061	/// ...
3062	static NonTrivialUnswitchCandidate
3063	injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L,
3064	DominatorTree &DT, LoopInfo &LI,
3065	AssumptionCache &AC, MemorySSAUpdater *MSSAU) {
3066	assert(Candidate.hasPendingInjection() && "Nothing to inject!");
3067	BasicBlock *Preheader = L.getLoopPreheader();
3068	assert(Preheader && "Loop is not in simplified form?");
3069	assert(LI.getLoopFor(Candidate.TI->getParent()) == &L &&
3070	"Unswitching branch of inner loop!");
3071
3072	auto Pred = Candidate.PendingInjection->Pred;
3073	auto *LHS = Candidate.PendingInjection->LHS;
3074	auto *RHS = Candidate.PendingInjection->RHS;
3075	auto *InLoopSucc = Candidate.PendingInjection->InLoopSucc;
3076	auto *TI = cast<BranchInst>(Val: Candidate.TI);
3077	auto *BB = Candidate.TI->getParent();
3078	auto *OutOfLoopSucc = InLoopSucc == TI->getSuccessor(i: 0) ? TI->getSuccessor(i: 1)
3079	: TI->getSuccessor(i: 0);
3080	// FIXME: Remove this once limitation on successors is lifted.
3081	assert(L.contains(InLoopSucc) && "Not supported yet!");
3082	assert(!L.contains(OutOfLoopSucc) && "Not supported yet!");
3083	auto &Ctx = BB->getContext();
3084
3085	IRBuilder<> Builder(Preheader->getTerminator());
3086	assert(ICmpInst::isUnsigned(Pred) && "Not supported yet!");
3087	if (LHS->getType() != RHS->getType()) {
3088	if (LHS->getType()->getIntegerBitWidth() <
3089	RHS->getType()->getIntegerBitWidth())
3090	LHS = Builder.CreateZExt(V: LHS, DestTy: RHS->getType(), Name: LHS->getName() + ".wide");
3091	else
3092	RHS = Builder.CreateZExt(V: RHS, DestTy: LHS->getType(), Name: RHS->getName() + ".wide");
3093	}
3094	// Do not use builder here: CreateICmp may simplify this into a constant and
3095	// unswitching will break. Better optimize it away later.
3096	auto *InjectedCond =
3097	ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: LHS, S2: RHS, Name: "injected.cond",
3098	InsertBefore: Preheader->getTerminator()->getIterator());
3099
3100	BasicBlock *CheckBlock = BasicBlock::Create(Context&: Ctx, Name: BB->getName() + ".check",
3101	Parent: BB->getParent(), InsertBefore: InLoopSucc);
3102	Builder.SetInsertPoint(TI);
3103	auto *InvariantBr =
3104	Builder.CreateCondBr(Cond: InjectedCond, True: InLoopSucc, False: CheckBlock);
3105
3106	Builder.SetInsertPoint(CheckBlock);
3107	Builder.CreateCondBr(Cond: TI->getCondition(), True: TI->getSuccessor(i: 0),
3108	False: TI->getSuccessor(i: 1));
3109	TI->eraseFromParent();
3110
3111	// Fixup phis.
3112	for (auto &I : *InLoopSucc) {
3113	auto *PN = dyn_cast<PHINode>(Val: &I);
3114	if (!PN)
3115	break;
3116	auto *Inc = PN->getIncomingValueForBlock(BB);
3117	PN->addIncoming(V: Inc, BB: CheckBlock);
3118	}
3119	OutOfLoopSucc->replacePhiUsesWith(Old: BB, New: CheckBlock);
3120
3121	SmallVector<DominatorTree::UpdateType, 4> DTUpdates = {
3122	{ DominatorTree::Insert, BB, CheckBlock },
3123	{ DominatorTree::Insert, CheckBlock, InLoopSucc },
3124	{ DominatorTree::Insert, CheckBlock, OutOfLoopSucc },
3125	{ DominatorTree::Delete, BB, OutOfLoopSucc }
3126	};
3127
3128	DT.applyUpdates(Updates: DTUpdates);
3129	if (MSSAU)
3130	MSSAU->applyUpdates(Updates: DTUpdates, DT);
3131	L.addBasicBlockToLoop(NewBB: CheckBlock, LI);
3132
3133	#ifndef NDEBUG
3134	DT.verify();
3135	LI.verify(DomTree: DT);
3136	if (MSSAU && VerifyMemorySSA)
3137	MSSAU->getMemorySSA()->verifyMemorySSA();
3138	#endif
3139
3140	// TODO: In fact, cost of unswitching a new invariant candidate is *slightly*
3141	// higher because we have just inserted a new block. Need to think how to
3142	// adjust the cost of injected candidates when it was first computed.
3143	LLVM_DEBUG(dbgs() << "Injected a new loop-invariant branch " << *InvariantBr
3144	<< " and considering it for unswitching.");
3145	++NumInvariantConditionsInjected;
3146	return NonTrivialUnswitchCandidate(InvariantBr, { InjectedCond },
3147	Candidate.Cost);
3148	}
3149
3150	/// Given chain of loop branch conditions looking like:
3151	/// br (Variant < Invariant1)
3152	/// br (Variant < Invariant2)
3153	/// br (Variant < Invariant3)
3154	/// ...
3155	/// collect set of invariant conditions on which we want to unswitch, which
3156	/// look like:
3157	/// Invariant1 <= Invariant2
3158	/// Invariant2 <= Invariant3
3159	/// ...
3160	/// Though they might not immediately exist in the IR, we can still inject them.
3161	static bool insertCandidatesWithPendingInjections(
3162	SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, Loop &L,
3163	ICmpInst::Predicate Pred, ArrayRef<CompareDesc> Compares,
3164	const DominatorTree &DT) {
3165
3166	assert(ICmpInst::isRelational(Pred));
3167	assert(ICmpInst::isStrictPredicate(Pred));
3168	if (Compares.size() < 2)
3169	return false;
3170	ICmpInst::Predicate NonStrictPred = ICmpInst::getNonStrictPredicate(pred: Pred);
3171	for (auto Prev = Compares.begin(), Next = Compares.begin() + 1;
3172	Next != Compares.end(); ++Prev, ++Next) {
3173	Value *LHS = Next->Invariant;
3174	Value *RHS = Prev->Invariant;
3175	BasicBlock *InLoopSucc = Prev->InLoopSucc;
3176	InjectedInvariant ToInject(NonStrictPred, LHS, RHS, InLoopSucc);
3177	NonTrivialUnswitchCandidate Candidate(Prev->Term, { LHS, RHS },
3178	std::nullopt, std::move(ToInject));
3179	UnswitchCandidates.push_back(Elt: std::move(Candidate));
3180	}
3181	return true;
3182	}
3183
3184	/// Collect unswitch candidates by invariant conditions that are not immediately
3185	/// present in the loop. However, they can be injected into the code if we
3186	/// decide it's profitable.
3187	/// An example of such conditions is following:
3188	///
3189	/// for (...) {
3190	/// x = load ...
3191	/// if (! x <u C1) break;
3192	/// if (! x <u C2) break;
3193	/// <do something>
3194	/// }
3195	///
3196	/// We can unswitch by condition "C1 <=u C2". If that is true, then "x <u C1 <=
3197	/// C2" automatically implies "x <u C2", so we can get rid of one of
3198	/// loop-variant checks in unswitched loop version.
3199	static bool collectUnswitchCandidatesWithInjections(
3200	SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
3201	IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, Loop &L,
3202	const DominatorTree &DT, const LoopInfo &LI, AAResults &AA,
3203	const MemorySSAUpdater *MSSAU) {
3204	if (!InjectInvariantConditions)
3205	return false;
3206
3207	if (!DT.isReachableFromEntry(A: L.getHeader()))
3208	return false;
3209	auto *Latch = L.getLoopLatch();
3210	// Need to have a single latch and a preheader.
3211	if (!Latch)
3212	return false;
3213	assert(L.getLoopPreheader() && "Must have a preheader!");
3214
3215	DenseMap<Value *, SmallVector<CompareDesc, 4> > CandidatesULT;
3216	// Traverse the conditions that dominate latch (and therefore dominate each
3217	// other).
3218	for (auto *DTN = DT.getNode(BB: Latch); L.contains(BB: DTN->getBlock());
3219	DTN = DTN->getIDom()) {
3220	ICmpInst::Predicate Pred;
3221	Value *LHS = nullptr, *RHS = nullptr;
3222	BasicBlock *IfTrue = nullptr, *IfFalse = nullptr;
3223	auto *BB = DTN->getBlock();
3224	// Ignore inner loops.
3225	if (LI.getLoopFor(BB) != &L)
3226	continue;
3227	auto *Term = BB->getTerminator();
3228	if (!match(V: Term, P: m_Br(C: m_ICmp(Pred, L: m_Value(V&: LHS), R: m_Value(V&: RHS)),
3229	T: m_BasicBlock(V&: IfTrue), F: m_BasicBlock(V&: IfFalse))))
3230	continue;
3231	if (!LHS->getType()->isIntegerTy())
3232	continue;
3233	canonicalizeForInvariantConditionInjection(Pred, LHS, RHS, IfTrue, IfFalse,
3234	L);
3235	if (!shouldTryInjectInvariantCondition(Pred, LHS, RHS, IfTrue, IfFalse, L))
3236	continue;
3237	if (!shouldTryInjectBasingOnMetadata(BI: cast<BranchInst>(Val: Term), TakenSucc: IfTrue))
3238	continue;
3239	// Strip ZEXT for unsigned predicate.
3240	// TODO: once signed predicates are supported, also strip SEXT.
3241	CompareDesc Desc(cast<BranchInst>(Val: Term), RHS, IfTrue);
3242	while (auto *Zext = dyn_cast<ZExtInst>(Val: LHS))
3243	LHS = Zext->getOperand(i_nocapture: 0);
3244	CandidatesULT[LHS].push_back(Elt: Desc);
3245	}
3246
3247	bool Found = false;
3248	for (auto &It : CandidatesULT)
3249	Found |= insertCandidatesWithPendingInjections(
3250	UnswitchCandidates, L, Pred: ICmpInst::ICMP_ULT, Compares: It.second, DT);
3251	return Found;
3252	}
3253
3254	static bool isSafeForNoNTrivialUnswitching(Loop &L, LoopInfo &LI) {
3255	if (!L.isSafeToClone())
3256	return false;
3257	for (auto *BB : L.blocks())
3258	for (auto &I : *BB) {
3259	if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
3260	return false;
3261	if (auto *CB = dyn_cast<CallBase>(Val: &I)) {
3262	assert(!CB->cannotDuplicate() && "Checked by L.isSafeToClone().");
3263	if (CB->isConvergent())
3264	return false;
3265	}
3266	}
3267
3268	// Check if there are irreducible CFG cycles in this loop. If so, we cannot
3269	// easily unswitch non-trivial edges out of the loop. Doing so might turn the
3270	// irreducible control flow into reducible control flow and introduce new
3271	// loops "out of thin air". If we ever discover important use cases for doing
3272	// this, we can add support to loop unswitch, but it is a lot of complexity
3273	// for what seems little or no real world benefit.
3274	LoopBlocksRPO RPOT(&L);
3275	RPOT.perform(LI: &LI);
3276	if (containsIrreducibleCFG<const BasicBlock *>(RPOTraversal&: RPOT, LI))
3277	return false;
3278
3279	SmallVector<BasicBlock *, 4> ExitBlocks;
3280	L.getUniqueExitBlocks(ExitBlocks);
3281	// We cannot unswitch if exit blocks contain a cleanuppad/catchswitch
3282	// instruction as we don't know how to split those exit blocks.
3283	// FIXME: We should teach SplitBlock to handle this and remove this
3284	// restriction.
3285	for (auto *ExitBB : ExitBlocks) {
3286	auto *I = ExitBB->getFirstNonPHI();
3287	if (isa<CleanupPadInst>(Val: I) || isa<CatchSwitchInst>(Val: I)) {
3288	LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch "
3289	"in exit block\n");
3290	return false;
3291	}
3292	}
3293
3294	return true;
3295	}
3296
3297	static NonTrivialUnswitchCandidate findBestNonTrivialUnswitchCandidate(
3298	ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates, const Loop &L,
3299	const DominatorTree &DT, const LoopInfo &LI, AssumptionCache &AC,
3300	const TargetTransformInfo &TTI, const IVConditionInfo &PartialIVInfo) {
3301	// Given that unswitching these terminators will require duplicating parts of
3302	// the loop, so we need to be able to model that cost. Compute the ephemeral
3303	// values and set up a data structure to hold per-BB costs. We cache each
3304	// block's cost so that we don't recompute this when considering different
3305	// subsets of the loop for duplication during unswitching.
3306	SmallPtrSet<const Value *, 4> EphValues;
3307	CodeMetrics::collectEphemeralValues(L: &L, AC: &AC, EphValues);
3308	SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap;
3309
3310	// Compute the cost of each block, as well as the total loop cost. Also, bail
3311	// out if we see instructions which are incompatible with loop unswitching
3312	// (convergent, noduplicate, or cross-basic-block tokens).
3313	// FIXME: We might be able to safely handle some of these in non-duplicated
3314	// regions.
3315	TargetTransformInfo::TargetCostKind CostKind =
3316	L.getHeader()->getParent()->hasMinSize()
3317	? TargetTransformInfo::TCK_CodeSize
3318	: TargetTransformInfo::TCK_SizeAndLatency;
3319	InstructionCost LoopCost = 0;
3320	for (auto *BB : L.blocks()) {
3321	InstructionCost Cost = 0;
3322	for (auto &I : *BB) {
3323	if (EphValues.count(Ptr: &I))
3324	continue;
3325	Cost += TTI.getInstructionCost(U: &I, CostKind);
3326	}
3327	assert(Cost >= 0 && "Must not have negative costs!");
3328	LoopCost += Cost;
3329	assert(LoopCost >= 0 && "Must not have negative loop costs!");
3330	BBCostMap[BB] = Cost;
3331	}
3332	LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n");
3333
3334	// Now we find the best candidate by searching for the one with the following
3335	// properties in order:
3336	//
3337	// 1) An unswitching cost below the threshold
3338	// 2) The smallest number of duplicated unswitch candidates (to avoid
3339	// creating redundant subsequent unswitching)
3340	// 3) The smallest cost after unswitching.
3341	//
3342	// We prioritize reducing fanout of unswitch candidates provided the cost
3343	// remains below the threshold because this has a multiplicative effect.
3344	//
3345	// This requires memoizing each dominator subtree to avoid redundant work.
3346	//
3347	// FIXME: Need to actually do the number of candidates part above.
3348	SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap;
3349	// Given a terminator which might be unswitched, computes the non-duplicated
3350	// cost for that terminator.
3351	auto ComputeUnswitchedCost = [&](Instruction &TI,
3352	bool FullUnswitch) -> InstructionCost {
3353	// Unswitching selects unswitches the entire loop.
3354	if (isa<SelectInst>(Val: TI))
3355	return LoopCost;
3356
3357	BasicBlock &BB = *TI.getParent();
3358	SmallPtrSet<BasicBlock *, 4> Visited;
3359
3360	InstructionCost Cost = 0;
3361	for (BasicBlock *SuccBB : successors(BB: &BB)) {
3362	// Don't count successors more than once.
3363	if (!Visited.insert(Ptr: SuccBB).second)
3364	continue;
3365
3366	// If this is a partial unswitch candidate, then it must be a conditional
3367	// branch with a condition of either `or`, `and`, their corresponding
3368	// select forms or partially invariant instructions. In that case, one of
3369	// the successors is necessarily duplicated, so don't even try to remove
3370	// its cost.
3371	if (!FullUnswitch) {
3372	auto &BI = cast<BranchInst>(Val&: TI);
3373	Value *Cond = skipTrivialSelect(Cond: BI.getCondition());
3374	if (match(V: Cond, P: m_LogicalAnd())) {
3375	if (SuccBB == BI.getSuccessor(i: 1))
3376	continue;
3377	} else if (match(V: Cond, P: m_LogicalOr())) {
3378	if (SuccBB == BI.getSuccessor(i: 0))
3379	continue;
3380	} else if ((PartialIVInfo.KnownValue->isOneValue() &&
3381	SuccBB == BI.getSuccessor(i: 0)) ||
3382	(!PartialIVInfo.KnownValue->isOneValue() &&
3383	SuccBB == BI.getSuccessor(i: 1)))
3384	continue;
3385	}
3386
3387	// This successor's domtree will not need to be duplicated after
3388	// unswitching if the edge to the successor dominates it (and thus the
3389	// entire tree). This essentially means there is no other path into this
3390	// subtree and so it will end up live in only one clone of the loop.
3391	if (SuccBB->getUniquePredecessor() ||
3392	llvm::all_of(Range: predecessors(BB: SuccBB), P: [&](BasicBlock *PredBB) {
3393	return PredBB == &BB || DT.dominates(A: SuccBB, B: PredBB);
3394	})) {
3395	Cost += computeDomSubtreeCost(N&: *DT[SuccBB], BBCostMap, DTCostMap);
3396	assert(Cost <= LoopCost &&
3397	"Non-duplicated cost should never exceed total loop cost!");
3398	}
3399	}
3400
3401	// Now scale the cost by the number of unique successors minus one. We
3402	// subtract one because there is already at least one copy of the entire
3403	// loop. This is computing the new cost of unswitching a condition.
3404	// Note that guards always have 2 unique successors that are implicit and
3405	// will be materialized if we decide to unswitch it.
3406	int SuccessorsCount = isGuard(U: &TI) ? 2 : Visited.size();
3407	assert(SuccessorsCount > 1 &&
3408	"Cannot unswitch a condition without multiple distinct successors!");
3409	return (LoopCost - Cost) * (SuccessorsCount - 1);
3410	};
3411
3412	std::optional<NonTrivialUnswitchCandidate> Best;
3413	for (auto &Candidate : UnswitchCandidates) {
3414	Instruction &TI = *Candidate.TI;
3415	ArrayRef<Value *> Invariants = Candidate.Invariants;
3416	BranchInst *BI = dyn_cast<BranchInst>(Val: &TI);
3417	bool FullUnswitch =
3418	!BI || Candidate.hasPendingInjection() ||
3419	(Invariants.size() == 1 &&
3420	Invariants[0] == skipTrivialSelect(Cond: BI->getCondition()));
3421	InstructionCost CandidateCost = ComputeUnswitchedCost(TI, FullUnswitch);
3422	// Calculate cost multiplier which is a tool to limit potentially
3423	// exponential behavior of loop-unswitch.
3424	if (EnableUnswitchCostMultiplier) {
3425	int CostMultiplier =
3426	CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
3427	assert(
3428	(CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
3429	"cost multiplier needs to be in the range of 1..UnswitchThreshold");
3430	CandidateCost *= CostMultiplier;
3431	LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
3432	<< " (multiplier: " << CostMultiplier << ")"
3433	<< " for unswitch candidate: " << TI << "\n");
3434	} else {
3435	LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
3436	<< " for unswitch candidate: " << TI << "\n");
3437	}
3438
3439	if (!Best || CandidateCost < Best->Cost) {
3440	Best = Candidate;
3441	Best->Cost = CandidateCost;
3442	}
3443	}
3444	assert(Best && "Must be!");
3445	return *Best;
3446	}
3447
3448	// Insert a freeze on an unswitched branch if all is true:
3449	// 1. freeze-loop-unswitch-cond option is true
3450	// 2. The branch may not execute in the loop pre-transformation. If a branch may
3451	// not execute and could cause UB, it would always cause UB if it is hoisted outside
3452	// of the loop. Insert a freeze to prevent this case.
3453	// 3. The branch condition may be poison or undef
3454	static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT,
3455	AssumptionCache &AC) {
3456	assert(isa<BranchInst>(TI) || isa<SwitchInst>(TI));
3457	if (!FreezeLoopUnswitchCond)
3458	return false;
3459
3460	ICFLoopSafetyInfo SafetyInfo;
3461	SafetyInfo.computeLoopSafetyInfo(CurLoop: &L);
3462	if (SafetyInfo.isGuaranteedToExecute(Inst: TI, DT: &DT, CurLoop: &L))
3463	return false;
3464
3465	Value *Cond;
3466	if (BranchInst *BI = dyn_cast<BranchInst>(Val: &TI))
3467	Cond = skipTrivialSelect(Cond: BI->getCondition());
3468	else
3469	Cond = skipTrivialSelect(Cond: cast<SwitchInst>(Val: &TI)->getCondition());
3470	return !isGuaranteedNotToBeUndefOrPoison(
3471	V: Cond, AC: &AC, CtxI: L.getLoopPreheader()->getTerminator(), DT: &DT);
3472	}
3473
3474	static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
3475	AssumptionCache &AC, AAResults &AA,
3476	TargetTransformInfo &TTI, ScalarEvolution *SE,
3477	MemorySSAUpdater *MSSAU,
3478	LPMUpdater &LoopUpdater) {
3479	// Collect all invariant conditions within this loop (as opposed to an inner
3480	// loop which would be handled when visiting that inner loop).
3481	SmallVector<NonTrivialUnswitchCandidate, 4> UnswitchCandidates;
3482	IVConditionInfo PartialIVInfo;
3483	Instruction *PartialIVCondBranch = nullptr;
3484	collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo,
3485	PartialIVCondBranch, L, LI, AA, MSSAU);
3486	if (!findOptionMDForLoop(TheLoop: &L, Name: "llvm.loop.unswitch.injection.disable"))
3487	collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo,
3488	PartialIVCondBranch, L, DT, LI, AA,
3489	MSSAU);
3490	// If we didn't find any candidates, we're done.
3491	if (UnswitchCandidates.empty())
3492	return false;
3493
3494	LLVM_DEBUG(
3495	dbgs() << "Considering " << UnswitchCandidates.size()
3496	<< " non-trivial loop invariant conditions for unswitching.\n");
3497
3498	NonTrivialUnswitchCandidate Best = findBestNonTrivialUnswitchCandidate(
3499	UnswitchCandidates, L, DT, LI, AC, TTI, PartialIVInfo);
3500
3501	assert(Best.TI && "Failed to find loop unswitch candidate");
3502	assert(Best.Cost && "Failed to compute cost");
3503
3504	if (*Best.Cost >= UnswitchThreshold) {
3505	LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << *Best.Cost
3506	<< "\n");
3507	return false;
3508	}
3509
3510	bool InjectedCondition = false;
3511	if (Best.hasPendingInjection()) {
3512	Best = injectPendingInvariantConditions(Candidate: Best, L, DT, LI, AC, MSSAU);
3513	InjectedCondition = true;
3514	}
3515	assert(!Best.hasPendingInjection() &&
3516	"All injections should have been done by now!");
3517
3518	if (Best.TI != PartialIVCondBranch)
3519	PartialIVInfo.InstToDuplicate.clear();
3520
3521	bool InsertFreeze;
3522	if (auto *SI = dyn_cast<SelectInst>(Val: Best.TI)) {
3523	// If the best candidate is a select, turn it into a branch. Select
3524	// instructions with a poison conditional do not propagate poison, but
3525	// branching on poison causes UB. Insert a freeze on the select
3526	// conditional to prevent UB after turning the select into a branch.
3527	InsertFreeze = !isGuaranteedNotToBeUndefOrPoison(
3528	V: SI->getCondition(), AC: &AC, CtxI: L.getLoopPreheader()->getTerminator(), DT: &DT);
3529	Best.TI = turnSelectIntoBranch(SI, DT, LI, MSSAU, AC: &AC);
3530	} else {
3531	// If the best candidate is a guard, turn it into a branch.
3532	if (isGuard(U: Best.TI))
3533	Best.TI =
3534	turnGuardIntoBranch(GI: cast<IntrinsicInst>(Val: Best.TI), L, DT, LI, MSSAU);
3535	InsertFreeze = shouldInsertFreeze(L, TI&: *Best.TI, DT, AC);
3536	}
3537
3538	LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = " << Best.Cost
3539	<< ") terminator: " << *Best.TI << "\n");
3540	unswitchNontrivialInvariants(L, TI&: *Best.TI, Invariants: Best.Invariants, PartialIVInfo, DT,
3541	LI, AC, SE, MSSAU, LoopUpdater, InsertFreeze,
3542	InjectedCondition);
3543	return true;
3544	}
3545
3546	/// Unswitch control flow predicated on loop invariant conditions.
3547	///
3548	/// This first hoists all branches or switches which are trivial (IE, do not
3549	/// require duplicating any part of the loop) out of the loop body. It then
3550	/// looks at other loop invariant control flows and tries to unswitch those as
3551	/// well by cloning the loop if the result is small enough.
3552	///
3553	/// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are
3554	/// also updated based on the unswitch. The `MSSA` analysis is also updated if
3555	/// valid (i.e. its use is enabled).
3556	///
3557	/// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
3558	/// true, we will attempt to do non-trivial unswitching as well as trivial
3559	/// unswitching.
3560	///
3561	/// The `postUnswitch` function will be run after unswitching is complete
3562	/// with information on whether or not the provided loop remains a loop and
3563	/// a list of new sibling loops created.
3564	///
3565	/// If `SE` is non-null, we will update that analysis based on the unswitching
3566	/// done.
3567	static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
3568	AssumptionCache &AC, AAResults &AA,
3569	TargetTransformInfo &TTI, bool Trivial,
3570	bool NonTrivial, ScalarEvolution *SE,
3571	MemorySSAUpdater *MSSAU, ProfileSummaryInfo *PSI,
3572	BlockFrequencyInfo *BFI, LPMUpdater &LoopUpdater) {
3573	assert(L.isRecursivelyLCSSAForm(DT, LI) &&
3574	"Loops must be in LCSSA form before unswitching.");
3575
3576	// Must be in loop simplified form: we need a preheader and dedicated exits.
3577	if (!L.isLoopSimplifyForm())
3578	return false;
3579
3580	// Try trivial unswitch first before loop over other basic blocks in the loop.
3581	if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
3582	// If we unswitched successfully we will want to clean up the loop before
3583	// processing it further so just mark it as unswitched and return.
3584	postUnswitch(L, U&: LoopUpdater, LoopName: L.getName(),
3585	/*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false,
3586	/*InjectedCondition*/ false, NewLoops: {});
3587	return true;
3588	}
3589
3590	const Function *F = L.getHeader()->getParent();
3591
3592	// Check whether we should continue with non-trivial conditions.
3593	// EnableNonTrivialUnswitch: Global variable that forces non-trivial
3594	// unswitching for testing and debugging.
3595	// NonTrivial: Parameter that enables non-trivial unswitching for this
3596	// invocation of the transform. But this should be allowed only
3597	// for targets without branch divergence.
3598	//
3599	// FIXME: If divergence analysis becomes available to a loop
3600	// transform, we should allow unswitching for non-trivial uniform
3601	// branches even on targets that have divergence.
3602	// https://bugs.llvm.org/show_bug.cgi?id=48819
3603	bool ContinueWithNonTrivial =
3604	EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence(F));
3605	if (!ContinueWithNonTrivial)
3606	return false;
3607
3608	// Skip non-trivial unswitching for optsize functions.
3609	if (F->hasOptSize())
3610	return false;
3611
3612	// Returns true if Loop L's loop nest is cold, i.e. if the headers of L,
3613	// of the loops L is nested in, and of the loops nested in L are all cold.
3614	auto IsLoopNestCold = [&](const Loop *L) {
3615	// Check L and all of its parent loops.
3616	auto *Parent = L;
3617	while (Parent) {
3618	if (!PSI->isColdBlock(BB: Parent->getHeader(), BFI))
3619	return false;
3620	Parent = Parent->getParentLoop();
3621	}
3622	// Next check all loops nested within L.
3623	SmallVector<const Loop *, 4> Worklist;
3624	Worklist.insert(I: Worklist.end(), From: L->getSubLoops().begin(),
3625	To: L->getSubLoops().end());
3626	while (!Worklist.empty()) {
3627	auto *CurLoop = Worklist.pop_back_val();
3628	if (!PSI->isColdBlock(BB: CurLoop->getHeader(), BFI))
3629	return false;
3630	Worklist.insert(I: Worklist.end(), From: CurLoop->getSubLoops().begin(),
3631	To: CurLoop->getSubLoops().end());
3632	}
3633	return true;
3634	};
3635
3636	// Skip cold loops in cold loop nests, as unswitching them brings little
3637	// benefit but increases the code size
3638	if (PSI && PSI->hasProfileSummary() && BFI && IsLoopNestCold(&L)) {
3639	LLVM_DEBUG(dbgs() << " Skip cold loop: " << L << "\n");
3640	return false;
3641	}
3642
3643	// Perform legality checks.
3644	if (!isSafeForNoNTrivialUnswitching(L, LI))
3645	return false;
3646
3647	// For non-trivial unswitching, because it often creates new loops, we rely on
3648	// the pass manager to iterate on the loops rather than trying to immediately
3649	// reach a fixed point. There is no substantial advantage to iterating
3650	// internally, and if any of the new loops are simplified enough to contain
3651	// trivial unswitching we want to prefer those.
3652
3653	// Try to unswitch the best invariant condition. We prefer this full unswitch to
3654	// a partial unswitch when possible below the threshold.
3655	if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, SE, MSSAU, LoopUpdater))
3656	return true;
3657
3658	// No other opportunities to unswitch.
3659	return false;
3660	}
3661
3662	PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
3663	LoopStandardAnalysisResults &AR,
3664	LPMUpdater &U) {
3665	Function &F = *L.getHeader()->getParent();
3666	(void)F;
3667	ProfileSummaryInfo *PSI = nullptr;
3668	if (auto OuterProxy =
3669	AM.getResult<FunctionAnalysisManagerLoopProxy>(IR&: L, ExtraArgs&: AR)
3670	.getCachedResult<ModuleAnalysisManagerFunctionProxy>(IR&: F))
3671	PSI = OuterProxy->getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
3672	LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
3673	<< "\n");
3674
3675	std::optional<MemorySSAUpdater> MSSAU;
3676	if (AR.MSSA) {
3677	MSSAU = MemorySSAUpdater(AR.MSSA);
3678	if (VerifyMemorySSA)
3679	AR.MSSA->verifyMemorySSA();
3680	}
3681	if (!unswitchLoop(L, DT&: AR.DT, LI&: AR.LI, AC&: AR.AC, AA&: AR.AA, TTI&: AR.TTI, Trivial, NonTrivial,
3682	SE: &AR.SE, MSSAU: MSSAU ? &*MSSAU : nullptr, PSI, BFI: AR.BFI, LoopUpdater&: U))
3683	return PreservedAnalyses::all();
3684
3685	if (AR.MSSA && VerifyMemorySSA)
3686	AR.MSSA->verifyMemorySSA();
3687
3688	// Historically this pass has had issues with the dominator tree so verify it
3689	// in asserts builds.
3690	assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));
3691
3692	auto PA = getLoopPassPreservedAnalyses();
3693	if (AR.MSSA)
3694	PA.preserve<MemorySSAAnalysis>();
3695	return PA;
3696	}
3697
3698	void SimpleLoopUnswitchPass::printPipeline(
3699	raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
3700	static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline(
3701	OS, MapClassName2PassName);
3702
3703	OS << '<';
3704	OS << (NonTrivial ? "" : "no-") << "nontrivial;";
3705	OS << (Trivial ? "" : "no-") << "trivial";
3706	OS << '>';
3707	}
3708