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/Module.h"
45#include "llvm/IR/PatternMatch.h"
46#include "llvm/IR/ProfDataUtils.h"
47#include "llvm/IR/Use.h"
48#include "llvm/IR/Value.h"
49#include "llvm/Support/Casting.h"
50#include "llvm/Support/CommandLine.h"
51#include "llvm/Support/Debug.h"
52#include "llvm/Support/ErrorHandling.h"
53#include "llvm/Support/GenericDomTree.h"
54#include "llvm/Support/InstructionCost.h"
55#include "llvm/Support/raw_ostream.h"
56#include "llvm/Transforms/Scalar/LoopPassManager.h"
57#include "llvm/Transforms/Utils/BasicBlockUtils.h"
58#include "llvm/Transforms/Utils/Cloning.h"
59#include "llvm/Transforms/Utils/Local.h"
60#include "llvm/Transforms/Utils/LoopUtils.h"
61#include "llvm/Transforms/Utils/ValueMapper.h"
62#include <algorithm>
63#include <cassert>
64#include <iterator>
65#include <numeric>
66#include <optional>
67#include <utility>
68
69#define DEBUG_TYPE "simple-loop-unswitch"
70
71using namespace llvm;
72using namespace llvm::PatternMatch;
73
74STATISTIC(NumBranches, "Number of branches unswitched");
75STATISTIC(NumSwitches, "Number of switches unswitched");
76STATISTIC(NumSelects, "Number of selects turned into branches for unswitching");
77STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
78STATISTIC(NumTrivial, "Number of unswitches that are trivial");
79STATISTIC(
80 NumCostMultiplierSkipped,
81 "Number of unswitch candidates that had their cost multiplier skipped");
82STATISTIC(NumInvariantConditionsInjected,
83 "Number of invariant conditions injected and unswitched");
84
85static cl::opt<bool> EnableNonTrivialUnswitch(
86 "enable-nontrivial-unswitch", cl::init(Val: false), cl::Hidden,
87 cl::desc("Forcibly enables non-trivial loop unswitching rather than "
88 "following the configuration passed into the pass."));
89
90static cl::opt<int>
91 UnswitchThreshold("unswitch-threshold", cl::init(Val: 50), cl::Hidden,
92 cl::desc("The cost threshold for unswitching a loop."));
93
94static cl::opt<bool> EnableUnswitchCostMultiplier(
95 "enable-unswitch-cost-multiplier", cl::init(Val: true), cl::Hidden,
96 cl::desc("Enable unswitch cost multiplier that prohibits exponential "
97 "explosion in nontrivial unswitch."));
98static cl::opt<int> UnswitchSiblingsToplevelDiv(
99 "unswitch-siblings-toplevel-div", cl::init(Val: 2), cl::Hidden,
100 cl::desc("Toplevel siblings divisor for cost multiplier."));
101static cl::opt<int> UnswitchNumInitialUnscaledCandidates(
102 "unswitch-num-initial-unscaled-candidates", cl::init(Val: 8), cl::Hidden,
103 cl::desc("Number of unswitch candidates that are ignored when calculating "
104 "cost multiplier."));
105static cl::opt<bool> UnswitchGuards(
106 "simple-loop-unswitch-guards", cl::init(Val: true), cl::Hidden,
107 cl::desc("If enabled, simple loop unswitching will also consider "
108 "llvm.experimental.guard intrinsics as unswitch candidates."));
109static cl::opt<bool> DropNonTrivialImplicitNullChecks(
110 "simple-loop-unswitch-drop-non-trivial-implicit-null-checks",
111 cl::init(Val: false), cl::Hidden,
112 cl::desc("If enabled, drop make.implicit metadata in unswitched implicit "
113 "null checks to save time analyzing if we can keep it."));
114static cl::opt<unsigned>
115 MSSAThreshold("simple-loop-unswitch-memoryssa-threshold",
116 cl::desc("Max number of memory uses to explore during "
117 "partial unswitching analysis"),
118 cl::init(Val: 100), cl::Hidden);
119static cl::opt<bool> FreezeLoopUnswitchCond(
120 "freeze-loop-unswitch-cond", cl::init(Val: true), cl::Hidden,
121 cl::desc("If enabled, the freeze instruction will be added to condition "
122 "of loop unswitch to prevent miscompilation."));
123
124static cl::opt<bool> InjectInvariantConditions(
125 "simple-loop-unswitch-inject-invariant-conditions", cl::Hidden,
126 cl::desc("Whether we should inject new invariants and unswitch them to "
127 "eliminate some existing (non-invariant) conditions."),
128 cl::init(Val: true));
129
130static cl::opt<unsigned> InjectInvariantConditionHotnesThreshold(
131 "simple-loop-unswitch-inject-invariant-condition-hotness-threshold",
132 cl::Hidden, cl::desc("Only try to inject loop invariant conditions and "
133 "unswitch on them to eliminate branches that are "
134 "not-taken 1/<this option> times or less."),
135 cl::init(Val: 16));
136
137AnalysisKey ShouldRunExtraSimpleLoopUnswitch::Key;
138namespace {
139struct CompareDesc {
140 BranchInst *Term;
141 Value *Invariant;
142 BasicBlock *InLoopSucc;
143
144 CompareDesc(BranchInst *Term, Value *Invariant, BasicBlock *InLoopSucc)
145 : Term(Term), Invariant(Invariant), InLoopSucc(InLoopSucc) {}
146};
147
148struct InjectedInvariant {
149 ICmpInst::Predicate Pred;
150 Value *LHS;
151 Value *RHS;
152 BasicBlock *InLoopSucc;
153
154 InjectedInvariant(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
155 BasicBlock *InLoopSucc)
156 : Pred(Pred), LHS(LHS), RHS(RHS), InLoopSucc(InLoopSucc) {}
157};
158
159struct NonTrivialUnswitchCandidate {
160 Instruction *TI = nullptr;
161 TinyPtrVector<Value *> Invariants;
162 std::optional<InstructionCost> Cost;
163 std::optional<InjectedInvariant> PendingInjection;
164 NonTrivialUnswitchCandidate(
165 Instruction *TI, ArrayRef<Value *> Invariants,
166 std::optional<InstructionCost> Cost = std::nullopt,
167 std::optional<InjectedInvariant> PendingInjection = std::nullopt)
168 : TI(TI), Invariants(Invariants), Cost(Cost),
169 PendingInjection(PendingInjection) {};
170
171 bool hasPendingInjection() const { return PendingInjection.has_value(); }
172};
173} // end anonymous namespace.
174
175// Helper to skip (select x, true, false), which matches both a logical AND and
176// OR and can confuse code that tries to determine if \p Cond is either a
177// logical AND or OR but not both.
178static Value *skipTrivialSelect(Value *Cond) {
179 Value *CondNext;
180 while (match(V: Cond, P: m_Select(C: m_Value(V&: CondNext), L: m_One(), R: m_Zero())))
181 Cond = CondNext;
182 return Cond;
183}
184
185/// Collect all of the loop invariant input values transitively used by the
186/// homogeneous instruction graph from a given root.
187///
188/// This essentially walks from a root recursively through loop variant operands
189/// which have perform the same logical operation (AND or OR) and finds all
190/// inputs which are loop invariant. For some operations these can be
191/// re-associated and unswitched out of the loop entirely.
192static TinyPtrVector<Value *>
193collectHomogenousInstGraphLoopInvariants(const Loop &L, Instruction &Root,
194 const LoopInfo &LI) {
195 assert(!L.isLoopInvariant(&Root) &&
196 "Only need to walk the graph if root itself is not invariant.");
197 TinyPtrVector<Value *> Invariants;
198
199 bool IsRootAnd = match(V: &Root, P: m_LogicalAnd());
200 bool IsRootOr = match(V: &Root, P: m_LogicalOr());
201
202 // Build a worklist and recurse through operators collecting invariants.
203 SmallVector<Instruction *, 4> Worklist;
204 SmallPtrSet<Instruction *, 8> Visited;
205 Worklist.push_back(Elt: &Root);
206 Visited.insert(Ptr: &Root);
207 do {
208 Instruction &I = *Worklist.pop_back_val();
209 for (Value *OpV : I.operand_values()) {
210 // Skip constants as unswitching isn't interesting for them.
211 if (isa<Constant>(Val: OpV))
212 continue;
213
214 // Add it to our result if loop invariant.
215 if (L.isLoopInvariant(V: OpV)) {
216 Invariants.push_back(NewVal: OpV);
217 continue;
218 }
219
220 // If not an instruction with the same opcode, nothing we can do.
221 Instruction *OpI = dyn_cast<Instruction>(Val: skipTrivialSelect(Cond: OpV));
222
223 if (OpI && ((IsRootAnd && match(V: OpI, P: m_LogicalAnd())) ||
224 (IsRootOr && match(V: OpI, P: m_LogicalOr())))) {
225 // Visit this operand.
226 if (Visited.insert(Ptr: OpI).second)
227 Worklist.push_back(Elt: OpI);
228 }
229 }
230 } while (!Worklist.empty());
231
232 return Invariants;
233}
234
235static void replaceLoopInvariantUses(const Loop &L, Value *Invariant,
236 Constant &Replacement) {
237 assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?");
238
239 // Replace uses of LIC in the loop with the given constant.
240 // We use make_early_inc_range as set invalidates the iterator.
241 for (Use &U : llvm::make_early_inc_range(Range: Invariant->uses())) {
242 Instruction *UserI = dyn_cast<Instruction>(Val: U.getUser());
243
244 // Replace this use within the loop body.
245 if (UserI && L.contains(Inst: UserI))
246 U.set(&Replacement);
247 }
248}
249
250/// Check that all the LCSSA PHI nodes in the loop exit block have trivial
251/// incoming values along this edge.
252static bool areLoopExitPHIsLoopInvariant(const Loop &L,
253 const BasicBlock &ExitingBB,
254 const BasicBlock &ExitBB) {
255 for (const Instruction &I : ExitBB) {
256 auto *PN = dyn_cast<PHINode>(Val: &I);
257 if (!PN)
258 // No more PHIs to check.
259 return true;
260
261 // If the incoming value for this edge isn't loop invariant the unswitch
262 // won't be trivial.
263 if (!L.isLoopInvariant(V: PN->getIncomingValueForBlock(BB: &ExitingBB)))
264 return false;
265 }
266 llvm_unreachable("Basic blocks should never be empty!");
267}
268
269/// Copy a set of loop invariant values \p ToDuplicate and insert them at the
270/// end of \p BB and conditionally branch on the copied condition. We only
271/// branch on a single value.
272static void buildPartialUnswitchConditionalBranch(
273 BasicBlock &BB, ArrayRef<Value *> Invariants, bool Direction,
274 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, bool InsertFreeze,
275 const Instruction *I, AssumptionCache *AC, const DominatorTree &DT) {
276 IRBuilder<> IRB(&BB);
277 IRB.SetCurrentDebugLocation(DebugLoc::getCompilerGenerated());
278
279 SmallVector<Value *> FrozenInvariants;
280 for (Value *Inv : Invariants) {
281 if (InsertFreeze && !isGuaranteedNotToBeUndefOrPoison(V: Inv, AC, CtxI: I, DT: &DT))
282 Inv = IRB.CreateFreeze(V: Inv, Name: Inv->getName() + ".fr");
283 FrozenInvariants.push_back(Elt: Inv);
284 }
285
286 Value *Cond = Direction ? IRB.CreateOr(Ops: FrozenInvariants)
287 : IRB.CreateAnd(Ops: FrozenInvariants);
288 IRB.CreateCondBr(Cond, True: Direction ? &UnswitchedSucc : &NormalSucc,
289 False: Direction ? &NormalSucc : &UnswitchedSucc);
290}
291
292/// Copy a set of loop invariant values, and conditionally branch on them.
293static void buildPartialInvariantUnswitchConditionalBranch(
294 BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction,
295 BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L,
296 MemorySSAUpdater *MSSAU) {
297 ValueToValueMapTy VMap;
298 for (auto *Val : reverse(C&: ToDuplicate)) {
299 Instruction *Inst = cast<Instruction>(Val);
300 Instruction *NewInst = Inst->clone();
301
302 if (const DebugLoc &DL = Inst->getDebugLoc())
303 mapAtomInstance(DL, VMap);
304
305 NewInst->insertInto(ParentBB: &BB, It: BB.end());
306 RemapInstruction(I: NewInst, VM&: VMap,
307 Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
308 VMap[Val] = NewInst;
309
310 if (!MSSAU)
311 continue;
312
313 MemorySSA *MSSA = MSSAU->getMemorySSA();
314 if (auto *MemUse =
315 dyn_cast_or_null<MemoryUse>(Val: MSSA->getMemoryAccess(I: Inst))) {
316 auto *DefiningAccess = MemUse->getDefiningAccess();
317 // Get the first defining access before the loop.
318 while (L.contains(BB: DefiningAccess->getBlock())) {
319 // If the defining access is a MemoryPhi, get the incoming
320 // value for the pre-header as defining access.
321 if (auto *MemPhi = dyn_cast<MemoryPhi>(Val: DefiningAccess))
322 DefiningAccess =
323 MemPhi->getIncomingValueForBlock(BB: L.getLoopPreheader());
324 else
325 DefiningAccess = cast<MemoryDef>(Val: DefiningAccess)->getDefiningAccess();
326 }
327 MSSAU->createMemoryAccessInBB(I: NewInst, Definition: DefiningAccess,
328 BB: NewInst->getParent(),
329 Point: MemorySSA::BeforeTerminator);
330 }
331 }
332
333 IRBuilder<> IRB(&BB);
334 IRB.SetCurrentDebugLocation(DebugLoc::getCompilerGenerated());
335 Value *Cond = VMap[ToDuplicate[0]];
336 IRB.CreateCondBr(Cond, True: Direction ? &UnswitchedSucc : &NormalSucc,
337 False: Direction ? &NormalSucc : &UnswitchedSucc);
338}
339
340/// Rewrite the PHI nodes in an unswitched loop exit basic block.
341///
342/// Requires that the loop exit and unswitched basic block are the same, and
343/// that the exiting block was a unique predecessor of that block. Rewrites the
344/// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
345/// PHI nodes from the old preheader that now contains the unswitched
346/// terminator.
347static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
348 BasicBlock &OldExitingBB,
349 BasicBlock &OldPH) {
350 for (PHINode &PN : UnswitchedBB.phis()) {
351 // When the loop exit is directly unswitched we just need to update the
352 // incoming basic block. We loop to handle weird cases with repeated
353 // incoming blocks, but expect to typically only have one operand here.
354 for (auto i : seq<int>(Begin: 0, End: PN.getNumOperands())) {
355 assert(PN.getIncomingBlock(i) == &OldExitingBB &&
356 "Found incoming block different from unique predecessor!");
357 PN.setIncomingBlock(i, BB: &OldPH);
358 }
359 }
360}
361
362/// Rewrite the PHI nodes in the loop exit basic block and the split off
363/// unswitched block.
364///
365/// Because the exit block remains an exit from the loop, this rewrites the
366/// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
367/// nodes into the unswitched basic block to select between the value in the
368/// old preheader and the loop exit.
369static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
370 BasicBlock &UnswitchedBB,
371 BasicBlock &OldExitingBB,
372 BasicBlock &OldPH,
373 bool FullUnswitch) {
374 assert(&ExitBB != &UnswitchedBB &&
375 "Must have different loop exit and unswitched blocks!");
376 BasicBlock::iterator InsertPt = UnswitchedBB.begin();
377 for (PHINode &PN : ExitBB.phis()) {
378 auto *NewPN = PHINode::Create(Ty: PN.getType(), /*NumReservedValues*/ 2,
379 NameStr: PN.getName() + ".split");
380 NewPN->insertBefore(InsertPos: InsertPt);
381
382 // Walk backwards over the old PHI node's inputs to minimize the cost of
383 // removing each one. We have to do this weird loop manually so that we
384 // create the same number of new incoming edges in the new PHI as we expect
385 // each case-based edge to be included in the unswitched switch in some
386 // cases.
387 // FIXME: This is really, really gross. It would be much cleaner if LLVM
388 // allowed us to create a single entry for a predecessor block without
389 // having separate entries for each "edge" even though these edges are
390 // required to produce identical results.
391 for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
392 if (PN.getIncomingBlock(i) != &OldExitingBB)
393 continue;
394
395 Value *Incoming = PN.getIncomingValue(i);
396 if (FullUnswitch)
397 // No more edge from the old exiting block to the exit block.
398 PN.removeIncomingValue(Idx: i);
399
400 NewPN->addIncoming(V: Incoming, BB: &OldPH);
401 }
402
403 // Now replace the old PHI with the new one and wire the old one in as an
404 // input to the new one.
405 PN.replaceAllUsesWith(V: NewPN);
406 NewPN->addIncoming(V: &PN, BB: &ExitBB);
407 }
408}
409
410/// Hoist the current loop up to the innermost loop containing a remaining exit.
411///
412/// Because we've removed an exit from the loop, we may have changed the set of
413/// loops reachable and need to move the current loop up the loop nest or even
414/// to an entirely separate nest.
415static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
416 DominatorTree &DT, LoopInfo &LI,
417 MemorySSAUpdater *MSSAU, ScalarEvolution *SE) {
418 // If the loop is already at the top level, we can't hoist it anywhere.
419 Loop *OldParentL = L.getParentLoop();
420 if (!OldParentL)
421 return;
422
423 SmallVector<BasicBlock *, 4> Exits;
424 L.getExitBlocks(ExitBlocks&: Exits);
425 Loop *NewParentL = nullptr;
426 for (auto *ExitBB : Exits)
427 if (Loop *ExitL = LI.getLoopFor(BB: ExitBB))
428 if (!NewParentL || NewParentL->contains(L: ExitL))
429 NewParentL = ExitL;
430
431 if (NewParentL == OldParentL)
432 return;
433
434 // The new parent loop (if different) should always contain the old one.
435 if (NewParentL)
436 assert(NewParentL->contains(OldParentL) &&
437 "Can only hoist this loop up the nest!");
438
439 // The preheader will need to move with the body of this loop. However,
440 // because it isn't in this loop we also need to update the primary loop map.
441 assert(OldParentL == LI.getLoopFor(&Preheader) &&
442 "Parent loop of this loop should contain this loop's preheader!");
443 LI.changeLoopFor(BB: &Preheader, L: NewParentL);
444
445 // Remove this loop from its old parent.
446 OldParentL->removeChildLoop(Child: &L);
447
448 // Add the loop either to the new parent or as a top-level loop.
449 if (NewParentL)
450 NewParentL->addChildLoop(NewChild: &L);
451 else
452 LI.addTopLevelLoop(New: &L);
453
454 // Remove this loops blocks from the old parent and every other loop up the
455 // nest until reaching the new parent. Also update all of these
456 // no-longer-containing loops to reflect the nesting change.
457 for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
458 OldContainingL = OldContainingL->getParentLoop()) {
459 llvm::erase_if(C&: OldContainingL->getBlocksVector(),
460 P: [&](const BasicBlock *BB) {
461 return BB == &Preheader || L.contains(BB);
462 });
463
464 OldContainingL->getBlocksSet().erase(Ptr: &Preheader);
465 for (BasicBlock *BB : L.blocks())
466 OldContainingL->getBlocksSet().erase(Ptr: BB);
467
468 // Because we just hoisted a loop out of this one, we have essentially
469 // created new exit paths from it. That means we need to form LCSSA PHI
470 // nodes for values used in the no-longer-nested loop.
471 formLCSSA(L&: *OldContainingL, DT, LI: &LI, SE);
472
473 // We shouldn't need to form dedicated exits because the exit introduced
474 // here is the (just split by unswitching) preheader. However, after trivial
475 // unswitching it is possible to get new non-dedicated exits out of parent
476 // loop so let's conservatively form dedicated exit blocks and figure out
477 // if we can optimize later.
478 formDedicatedExitBlocks(L: OldContainingL, DT: &DT, LI: &LI, MSSAU,
479 /*PreserveLCSSA*/ true);
480 }
481}
482
483// Return the top-most loop containing ExitBB and having ExitBB as exiting block
484// or the loop containing ExitBB, if there is no parent loop containing ExitBB
485// as exiting block.
486static Loop *getTopMostExitingLoop(const BasicBlock *ExitBB,
487 const LoopInfo &LI) {
488 Loop *TopMost = LI.getLoopFor(BB: ExitBB);
489 Loop *Current = TopMost;
490 while (Current) {
491 if (Current->isLoopExiting(BB: ExitBB))
492 TopMost = Current;
493 Current = Current->getParentLoop();
494 }
495 return TopMost;
496}
497
498/// Unswitch a trivial branch if the condition is loop invariant.
499///
500/// This routine should only be called when loop code leading to the branch has
501/// been validated as trivial (no side effects). This routine checks if the
502/// condition is invariant and one of the successors is a loop exit. This
503/// allows us to unswitch without duplicating the loop, making it trivial.
504///
505/// If this routine fails to unswitch the branch it returns false.
506///
507/// If the branch can be unswitched, this routine splits the preheader and
508/// hoists the branch above that split. Preserves loop simplified form
509/// (splitting the exit block as necessary). It simplifies the branch within
510/// the loop to an unconditional branch but doesn't remove it entirely. Further
511/// cleanup can be done with some simplifycfg like pass.
512///
513/// If `SE` is not null, it will be updated based on the potential loop SCEVs
514/// invalidated by this.
515static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
516 LoopInfo &LI, ScalarEvolution *SE,
517 MemorySSAUpdater *MSSAU) {
518 assert(BI.isConditional() && "Can only unswitch a conditional branch!");
519 LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n");
520
521 // The loop invariant values that we want to unswitch.
522 TinyPtrVector<Value *> Invariants;
523
524 // When true, we're fully unswitching the branch rather than just unswitching
525 // some input conditions to the branch.
526 bool FullUnswitch = false;
527
528 Value *Cond = skipTrivialSelect(Cond: BI.getCondition());
529 if (L.isLoopInvariant(V: Cond)) {
530 Invariants.push_back(NewVal: Cond);
531 FullUnswitch = true;
532 } else {
533 if (auto *CondInst = dyn_cast<Instruction>(Val: Cond))
534 Invariants = collectHomogenousInstGraphLoopInvariants(L, Root&: *CondInst, LI);
535 if (Invariants.empty()) {
536 LLVM_DEBUG(dbgs() << " Couldn't find invariant inputs!\n");
537 return false;
538 }
539 }
540
541 // Check that one of the branch's successors exits, and which one.
542 bool ExitDirection = true;
543 int LoopExitSuccIdx = 0;
544 auto *LoopExitBB = BI.getSuccessor(i: 0);
545 if (L.contains(BB: LoopExitBB)) {
546 ExitDirection = false;
547 LoopExitSuccIdx = 1;
548 LoopExitBB = BI.getSuccessor(i: 1);
549 if (L.contains(BB: LoopExitBB)) {
550 LLVM_DEBUG(dbgs() << " Branch doesn't exit the loop!\n");
551 return false;
552 }
553 }
554 auto *ContinueBB = BI.getSuccessor(i: 1 - LoopExitSuccIdx);
555 auto *ParentBB = BI.getParent();
556 if (!areLoopExitPHIsLoopInvariant(L, ExitingBB: *ParentBB, ExitBB: *LoopExitBB)) {
557 LLVM_DEBUG(dbgs() << " Loop exit PHI's aren't loop-invariant!\n");
558 return false;
559 }
560
561 // When unswitching only part of the branch's condition, we need the exit
562 // block to be reached directly from the partially unswitched input. This can
563 // be done when the exit block is along the true edge and the branch condition
564 // is a graph of `or` operations, or the exit block is along the false edge
565 // and the condition is a graph of `and` operations.
566 if (!FullUnswitch) {
567 if (ExitDirection ? !match(V: Cond, P: m_LogicalOr())
568 : !match(V: Cond, P: m_LogicalAnd())) {
569 LLVM_DEBUG(dbgs() << " Branch condition is in improper form for "
570 "non-full unswitch!\n");
571 return false;
572 }
573 }
574
575 LLVM_DEBUG({
576 dbgs() << " unswitching trivial invariant conditions for: " << BI
577 << "\n";
578 for (Value *Invariant : Invariants) {
579 dbgs() << " " << *Invariant << " == true";
580 if (Invariant != Invariants.back())
581 dbgs() << " ||";
582 dbgs() << "\n";
583 }
584 });
585
586 // If we have scalar evolutions, we need to invalidate them including this
587 // loop, the loop containing the exit block and the topmost parent loop
588 // exiting via LoopExitBB.
589 if (SE) {
590 if (const Loop *ExitL = getTopMostExitingLoop(ExitBB: LoopExitBB, LI))
591 SE->forgetLoop(L: ExitL);
592 else
593 // Forget the entire nest as this exits the entire nest.
594 SE->forgetTopmostLoop(L: &L);
595 SE->forgetBlockAndLoopDispositions();
596 }
597
598 if (MSSAU && VerifyMemorySSA)
599 MSSAU->getMemorySSA()->verifyMemorySSA();
600
601 // Split the preheader, so that we know that there is a safe place to insert
602 // the conditional branch. We will change the preheader to have a conditional
603 // branch on LoopCond.
604 BasicBlock *OldPH = L.getLoopPreheader();
605 BasicBlock *NewPH = SplitEdge(From: OldPH, To: L.getHeader(), DT: &DT, LI: &LI, MSSAU);
606
607 // Now that we have a place to insert the conditional branch, create a place
608 // to branch to: this is the exit block out of the loop that we are
609 // unswitching. We need to split this if there are other loop predecessors.
610 // Because the loop is in simplified form, *any* other predecessor is enough.
611 BasicBlock *UnswitchedBB;
612 if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
613 assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
614 "A branch's parent isn't a predecessor!");
615 UnswitchedBB = LoopExitBB;
616 } else {
617 UnswitchedBB =
618 SplitBlock(Old: LoopExitBB, SplitPt: LoopExitBB->begin(), DT: &DT, LI: &LI, MSSAU, BBName: "", Before: false);
619 }
620
621 if (MSSAU && VerifyMemorySSA)
622 MSSAU->getMemorySSA()->verifyMemorySSA();
623
624 // Actually move the invariant uses into the unswitched position. If possible,
625 // we do this by moving the instructions, but when doing partial unswitching
626 // we do it by building a new merge of the values in the unswitched position.
627 OldPH->getTerminator()->eraseFromParent();
628 if (FullUnswitch) {
629 // If fully unswitching, we can use the existing branch instruction.
630 // Splice it into the old PH to gate reaching the new preheader and re-point
631 // its successors.
632 BI.moveBefore(BB&: *OldPH, I: OldPH->end());
633 BI.setCondition(Cond);
634 if (MSSAU) {
635 // Temporarily clone the terminator, to make MSSA update cheaper by
636 // separating "insert edge" updates from "remove edge" ones.
637 BI.clone()->insertInto(ParentBB, It: ParentBB->end());
638 } else {
639 // Create a new unconditional branch that will continue the loop as a new
640 // terminator.
641 Instruction *NewBI = BranchInst::Create(IfTrue: ContinueBB, InsertBefore: ParentBB);
642 NewBI->setDebugLoc(BI.getDebugLoc());
643 }
644 BI.setSuccessor(idx: LoopExitSuccIdx, NewSucc: UnswitchedBB);
645 BI.setSuccessor(idx: 1 - LoopExitSuccIdx, NewSucc: NewPH);
646 } else {
647 // Only unswitching a subset of inputs to the condition, so we will need to
648 // build a new branch that merges the invariant inputs.
649 if (ExitDirection)
650 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalOr()) &&
651 "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the "
652 "condition!");
653 else
654 assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalAnd()) &&
655 "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the"
656 " condition!");
657 buildPartialUnswitchConditionalBranch(
658 BB&: *OldPH, Invariants, Direction: ExitDirection, UnswitchedSucc&: *UnswitchedBB, NormalSucc&: *NewPH,
659 InsertFreeze: FreezeLoopUnswitchCond, I: OldPH->getTerminator(), AC: nullptr, DT);
660 }
661
662 // Update the dominator tree with the added edge.
663 DT.insertEdge(From: OldPH, To: UnswitchedBB);
664
665 // After the dominator tree was updated with the added edge, update MemorySSA
666 // if available.
667 if (MSSAU) {
668 SmallVector<CFGUpdate, 1> Updates;
669 Updates.push_back(Elt: {cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
670 MSSAU->applyInsertUpdates(Updates, DT);
671 }
672
673 // Finish updating dominator tree and memory ssa for full unswitch.
674 if (FullUnswitch) {
675 if (MSSAU) {
676 Instruction *Term = ParentBB->getTerminator();
677 // Remove the cloned branch instruction and create unconditional branch
678 // now.
679 Instruction *NewBI = BranchInst::Create(IfTrue: ContinueBB, InsertBefore: ParentBB);
680 NewBI->setDebugLoc(Term->getDebugLoc());
681 Term->eraseFromParent();
682 MSSAU->removeEdge(From: ParentBB, To: LoopExitBB);
683 }
684 DT.deleteEdge(From: ParentBB, To: LoopExitBB);
685 }
686
687 if (MSSAU && VerifyMemorySSA)
688 MSSAU->getMemorySSA()->verifyMemorySSA();
689
690 // Rewrite the relevant PHI nodes.
691 if (UnswitchedBB == LoopExitBB)
692 rewritePHINodesForUnswitchedExitBlock(UnswitchedBB&: *UnswitchedBB, OldExitingBB&: *ParentBB, OldPH&: *OldPH);
693 else
694 rewritePHINodesForExitAndUnswitchedBlocks(ExitBB&: *LoopExitBB, UnswitchedBB&: *UnswitchedBB,
695 OldExitingBB&: *ParentBB, OldPH&: *OldPH, FullUnswitch);
696
697 // The constant we can replace all of our invariants with inside the loop
698 // body. If any of the invariants have a value other than this the loop won't
699 // be entered.
700 ConstantInt *Replacement = ExitDirection
701 ? ConstantInt::getFalse(Context&: BI.getContext())
702 : ConstantInt::getTrue(Context&: BI.getContext());
703
704 // Since this is an i1 condition we can also trivially replace uses of it
705 // within the loop with a constant.
706 for (Value *Invariant : Invariants)
707 replaceLoopInvariantUses(L, Invariant, Replacement&: *Replacement);
708
709 // If this was full unswitching, we may have changed the nesting relationship
710 // for this loop so hoist it to its correct parent if needed.
711 if (FullUnswitch)
712 hoistLoopToNewParent(L, Preheader&: *NewPH, DT, LI, MSSAU, SE);
713
714 if (MSSAU && VerifyMemorySSA)
715 MSSAU->getMemorySSA()->verifyMemorySSA();
716
717 LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n");
718 ++NumTrivial;
719 ++NumBranches;
720 return true;
721}
722
723/// Unswitch a trivial switch if the condition is loop invariant.
724///
725/// This routine should only be called when loop code leading to the switch has
726/// been validated as trivial (no side effects). This routine checks if the
727/// condition is invariant and that at least one of the successors is a loop
728/// exit. This allows us to unswitch without duplicating the loop, making it
729/// trivial.
730///
731/// If this routine fails to unswitch the switch it returns false.
732///
733/// If the switch can be unswitched, this routine splits the preheader and
734/// copies the switch above that split. If the default case is one of the
735/// exiting cases, it copies the non-exiting cases and points them at the new
736/// preheader. If the default case is not exiting, it copies the exiting cases
737/// and points the default at the preheader. It preserves loop simplified form
738/// (splitting the exit blocks as necessary). It simplifies the switch within
739/// the loop by removing now-dead cases. If the default case is one of those
740/// unswitched, it replaces its destination with a new basic block containing
741/// only unreachable. Such basic blocks, while technically loop exits, are not
742/// considered for unswitching so this is a stable transform and the same
743/// switch will not be revisited. If after unswitching there is only a single
744/// in-loop successor, the switch is further simplified to an unconditional
745/// branch. Still more cleanup can be done with some simplifycfg like pass.
746///
747/// If `SE` is not null, it will be updated based on the potential loop SCEVs
748/// invalidated by this.
749static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
750 LoopInfo &LI, ScalarEvolution *SE,
751 MemorySSAUpdater *MSSAU) {
752 LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n");
753 Value *LoopCond = SI.getCondition();
754
755 // If this isn't switching on an invariant condition, we can't unswitch it.
756 if (!L.isLoopInvariant(V: LoopCond))
757 return false;
758
759 auto *ParentBB = SI.getParent();
760
761 // The same check must be used both for the default and the exit cases. We
762 // should never leave edges from the switch instruction to a basic block that
763 // we are unswitching, hence the condition used to determine the default case
764 // needs to also be used to populate ExitCaseIndices, which is then used to
765 // remove cases from the switch.
766 auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) {
767 // BBToCheck is not an exit block if it is inside loop L.
768 if (L.contains(BB: &BBToCheck))
769 return false;
770 // BBToCheck is not trivial to unswitch if its phis aren't loop invariant.
771 if (!areLoopExitPHIsLoopInvariant(L, ExitingBB: *ParentBB, ExitBB: BBToCheck))
772 return false;
773 // We do not unswitch a block that only has an unreachable statement, as
774 // it's possible this is a previously unswitched block. Only unswitch if
775 // either the terminator is not unreachable, or, if it is, it's not the only
776 // instruction in the block.
777 auto *TI = BBToCheck.getTerminator();
778 bool isUnreachable = isa<UnreachableInst>(Val: TI);
779 return !isUnreachable || &*BBToCheck.getFirstNonPHIOrDbg() != TI;
780 };
781
782 SmallVector<int, 4> ExitCaseIndices;
783 for (auto Case : SI.cases())
784 if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor()))
785 ExitCaseIndices.push_back(Elt: Case.getCaseIndex());
786 BasicBlock *DefaultExitBB = nullptr;
787 SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight =
788 SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, idx: 0);
789 if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) {
790 DefaultExitBB = SI.getDefaultDest();
791 } else if (ExitCaseIndices.empty())
792 return false;
793
794 LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n");
795
796 if (MSSAU && VerifyMemorySSA)
797 MSSAU->getMemorySSA()->verifyMemorySSA();
798
799 // We may need to invalidate SCEVs for the outermost loop reached by any of
800 // the exits.
801 Loop *OuterL = &L;
802
803 if (DefaultExitBB) {
804 // Check the loop containing this exit.
805 Loop *ExitL = getTopMostExitingLoop(ExitBB: DefaultExitBB, LI);
806 if (!ExitL || ExitL->contains(L: OuterL))
807 OuterL = ExitL;
808 }
809 for (unsigned Index : ExitCaseIndices) {
810 auto CaseI = SI.case_begin() + Index;
811 // Compute the outer loop from this exit.
812 Loop *ExitL = getTopMostExitingLoop(ExitBB: CaseI->getCaseSuccessor(), LI);
813 if (!ExitL || ExitL->contains(L: OuterL))
814 OuterL = ExitL;
815 }
816
817 if (SE) {
818 if (OuterL)
819 SE->forgetLoop(L: OuterL);
820 else
821 SE->forgetTopmostLoop(L: &L);
822 }
823
824 if (DefaultExitBB) {
825 // Clear out the default destination temporarily to allow accurate
826 // predecessor lists to be examined below.
827 SI.setDefaultDest(nullptr);
828 }
829
830 // Store the exit cases into a separate data structure and remove them from
831 // the switch.
832 SmallVector<std::tuple<ConstantInt *, BasicBlock *,
833 SwitchInstProfUpdateWrapper::CaseWeightOpt>,
834 4> ExitCases;
835 ExitCases.reserve(N: ExitCaseIndices.size());
836 SwitchInstProfUpdateWrapper SIW(SI);
837 // We walk the case indices backwards so that we remove the last case first
838 // and don't disrupt the earlier indices.
839 for (unsigned Index : reverse(C&: ExitCaseIndices)) {
840 auto CaseI = SI.case_begin() + Index;
841 // Save the value of this case.
842 auto W = SIW.getSuccessorWeight(idx: CaseI->getSuccessorIndex());
843 ExitCases.emplace_back(Args: CaseI->getCaseValue(), Args: CaseI->getCaseSuccessor(), Args&: W);
844 // Delete the unswitched cases.
845 SIW.removeCase(I: CaseI);
846 }
847
848 // Check if after this all of the remaining cases point at the same
849 // successor.
850 BasicBlock *CommonSuccBB = nullptr;
851 if (SI.getNumCases() > 0 &&
852 all_of(Range: drop_begin(RangeOrContainer: SI.cases()), P: [&SI](const SwitchInst::CaseHandle &Case) {
853 return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor();
854 }))
855 CommonSuccBB = SI.case_begin()->getCaseSuccessor();
856 if (!DefaultExitBB) {
857 // If we're not unswitching the default, we need it to match any cases to
858 // have a common successor or if we have no cases it is the common
859 // successor.
860 if (SI.getNumCases() == 0)
861 CommonSuccBB = SI.getDefaultDest();
862 else if (SI.getDefaultDest() != CommonSuccBB)
863 CommonSuccBB = nullptr;
864 }
865
866 // Split the preheader, so that we know that there is a safe place to insert
867 // the switch.
868 BasicBlock *OldPH = L.getLoopPreheader();
869 BasicBlock *NewPH = SplitEdge(From: OldPH, To: L.getHeader(), DT: &DT, LI: &LI, MSSAU);
870 OldPH->getTerminator()->eraseFromParent();
871
872 // Now add the unswitched switch. This new switch instruction inherits the
873 // debug location of the old switch, because it semantically replace the old
874 // one.
875 auto *NewSI = SwitchInst::Create(Value: LoopCond, Default: NewPH, NumCases: ExitCases.size(), InsertBefore: OldPH);
876 NewSI->setDebugLoc(SIW->getDebugLoc());
877 SwitchInstProfUpdateWrapper NewSIW(*NewSI);
878
879 // Rewrite the IR for the unswitched basic blocks. This requires two steps.
880 // First, we split any exit blocks with remaining in-loop predecessors. Then
881 // we update the PHIs in one of two ways depending on if there was a split.
882 // We walk in reverse so that we split in the same order as the cases
883 // appeared. This is purely for convenience of reading the resulting IR, but
884 // it doesn't cost anything really.
885 SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
886 SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
887 // Handle the default exit if necessary.
888 // FIXME: It'd be great if we could merge this with the loop below but LLVM's
889 // ranges aren't quite powerful enough yet.
890 if (DefaultExitBB) {
891 if (pred_empty(BB: DefaultExitBB)) {
892 UnswitchedExitBBs.insert(Ptr: DefaultExitBB);
893 rewritePHINodesForUnswitchedExitBlock(UnswitchedBB&: *DefaultExitBB, OldExitingBB&: *ParentBB, OldPH&: *OldPH);
894 } else {
895 auto *SplitBB =
896 SplitBlock(Old: DefaultExitBB, SplitPt: DefaultExitBB->begin(), DT: &DT, LI: &LI, MSSAU);
897 rewritePHINodesForExitAndUnswitchedBlocks(ExitBB&: *DefaultExitBB, UnswitchedBB&: *SplitBB,
898 OldExitingBB&: *ParentBB, OldPH&: *OldPH,
899 /*FullUnswitch*/ true);
900 DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
901 }
902 }
903 // Note that we must use a reference in the for loop so that we update the
904 // container.
905 for (auto &ExitCase : reverse(C&: ExitCases)) {
906 // Grab a reference to the exit block in the pair so that we can update it.
907 BasicBlock *ExitBB = std::get<1>(t&: ExitCase);
908
909 // If this case is the last edge into the exit block, we can simply reuse it
910 // as it will no longer be a loop exit. No mapping necessary.
911 if (pred_empty(BB: ExitBB)) {
912 // Only rewrite once.
913 if (UnswitchedExitBBs.insert(Ptr: ExitBB).second)
914 rewritePHINodesForUnswitchedExitBlock(UnswitchedBB&: *ExitBB, OldExitingBB&: *ParentBB, OldPH&: *OldPH);
915 continue;
916 }
917
918 // Otherwise we need to split the exit block so that we retain an exit
919 // block from the loop and a target for the unswitched condition.
920 BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
921 if (!SplitExitBB) {
922 // If this is the first time we see this, do the split and remember it.
923 SplitExitBB = SplitBlock(Old: ExitBB, SplitPt: ExitBB->begin(), DT: &DT, LI: &LI, MSSAU);
924 rewritePHINodesForExitAndUnswitchedBlocks(ExitBB&: *ExitBB, UnswitchedBB&: *SplitExitBB,
925 OldExitingBB&: *ParentBB, OldPH&: *OldPH,
926 /*FullUnswitch*/ true);
927 }
928 // Update the case pair to point to the split block.
929 std::get<1>(t&: ExitCase) = SplitExitBB;
930 }
931
932 // Now add the unswitched cases. We do this in reverse order as we built them
933 // in reverse order.
934 for (auto &ExitCase : reverse(C&: ExitCases)) {
935 ConstantInt *CaseVal = std::get<0>(t&: ExitCase);
936 BasicBlock *UnswitchedBB = std::get<1>(t&: ExitCase);
937
938 NewSIW.addCase(OnVal: CaseVal, Dest: UnswitchedBB, W: std::get<2>(t&: ExitCase));
939 }
940
941 // If the default was unswitched, re-point it and add explicit cases for
942 // entering the loop.
943 if (DefaultExitBB) {
944 NewSIW->setDefaultDest(DefaultExitBB);
945 NewSIW.setSuccessorWeight(idx: 0, W: DefaultCaseWeight);
946
947 // We removed all the exit cases, so we just copy the cases to the
948 // unswitched switch.
949 for (const auto &Case : SI.cases())
950 NewSIW.addCase(OnVal: Case.getCaseValue(), Dest: NewPH,
951 W: SIW.getSuccessorWeight(idx: Case.getSuccessorIndex()));
952 } else if (DefaultCaseWeight) {
953 // We have to set branch weight of the default case.
954 uint64_t SW = *DefaultCaseWeight;
955 for (const auto &Case : SI.cases()) {
956 auto W = SIW.getSuccessorWeight(idx: Case.getSuccessorIndex());
957 assert(W &&
958 "case weight must be defined as default case weight is defined");
959 SW += *W;
960 }
961 NewSIW.setSuccessorWeight(idx: 0, W: SW);
962 }
963
964 // If we ended up with a common successor for every path through the switch
965 // after unswitching, rewrite it to an unconditional branch to make it easy
966 // to recognize. Otherwise we potentially have to recognize the default case
967 // pointing at unreachable and other complexity.
968 if (CommonSuccBB) {
969 BasicBlock *BB = SI.getParent();
970 // We may have had multiple edges to this common successor block, so remove
971 // them as predecessors. We skip the first one, either the default or the
972 // actual first case.
973 bool SkippedFirst = DefaultExitBB == nullptr;
974 for (auto Case : SI.cases()) {
975 assert(Case.getCaseSuccessor() == CommonSuccBB &&
976 "Non-common successor!");
977 (void)Case;
978 if (!SkippedFirst) {
979 SkippedFirst = true;
980 continue;
981 }
982 CommonSuccBB->removePredecessor(Pred: BB,
983 /*KeepOneInputPHIs*/ true);
984 }
985 // Now nuke the switch and replace it with a direct branch.
986 Instruction *NewBI = BranchInst::Create(IfTrue: CommonSuccBB, InsertBefore: BB);
987 NewBI->setDebugLoc(SIW->getDebugLoc());
988 SIW.eraseFromParent();
989 } else if (DefaultExitBB) {
990 assert(SI.getNumCases() > 0 &&
991 "If we had no cases we'd have a common successor!");
992 // Move the last case to the default successor. This is valid as if the
993 // default got unswitched it cannot be reached. This has the advantage of
994 // being simple and keeping the number of edges from this switch to
995 // successors the same, and avoiding any PHI update complexity.
996 auto LastCaseI = std::prev(x: SI.case_end());
997
998 SI.setDefaultDest(LastCaseI->getCaseSuccessor());
999 SIW.setSuccessorWeight(
1000 idx: 0, W: SIW.getSuccessorWeight(idx: LastCaseI->getSuccessorIndex()));
1001 SIW.removeCase(I: LastCaseI);
1002 }
1003
1004 // Walk the unswitched exit blocks and the unswitched split blocks and update
1005 // the dominator tree based on the CFG edits. While we are walking unordered
1006 // containers here, the API for applyUpdates takes an unordered list of
1007 // updates and requires them to not contain duplicates.
1008 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
1009 for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
1010 DTUpdates.push_back(Elt: {DT.Delete, ParentBB, UnswitchedExitBB});
1011 DTUpdates.push_back(Elt: {DT.Insert, OldPH, UnswitchedExitBB});
1012 }
1013 for (auto SplitUnswitchedPair : SplitExitBBMap) {
1014 DTUpdates.push_back(Elt: {DT.Delete, ParentBB, SplitUnswitchedPair.first});
1015 DTUpdates.push_back(Elt: {DT.Insert, OldPH, SplitUnswitchedPair.second});
1016 }
1017
1018 if (MSSAU) {
1019 MSSAU->applyUpdates(Updates: DTUpdates, DT, /*UpdateDT=*/UpdateDTFirst: true);
1020 if (VerifyMemorySSA)
1021 MSSAU->getMemorySSA()->verifyMemorySSA();
1022 } else {
1023 DT.applyUpdates(Updates: DTUpdates);
1024 }
1025
1026 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1027
1028 // We may have changed the nesting relationship for this loop so hoist it to
1029 // its correct parent if needed.
1030 hoistLoopToNewParent(L, Preheader&: *NewPH, DT, LI, MSSAU, SE);
1031
1032 if (MSSAU && VerifyMemorySSA)
1033 MSSAU->getMemorySSA()->verifyMemorySSA();
1034
1035 ++NumTrivial;
1036 ++NumSwitches;
1037 LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n");
1038 return true;
1039}
1040
1041/// This routine scans the loop to find a branch or switch which occurs before
1042/// any side effects occur. These can potentially be unswitched without
1043/// duplicating the loop. If a branch or switch is successfully unswitched the
1044/// scanning continues to see if subsequent branches or switches have become
1045/// trivial. Once all trivial candidates have been unswitched, this routine
1046/// returns.
1047///
1048/// The return value indicates whether anything was unswitched (and therefore
1049/// changed).
1050///
1051/// If `SE` is not null, it will be updated based on the potential loop SCEVs
1052/// invalidated by this.
1053static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
1054 LoopInfo &LI, ScalarEvolution *SE,
1055 MemorySSAUpdater *MSSAU) {
1056 bool Changed = false;
1057
1058 // If loop header has only one reachable successor we should keep looking for
1059 // trivial condition candidates in the successor as well. An alternative is
1060 // to constant fold conditions and merge successors into loop header (then we
1061 // only need to check header's terminator). The reason for not doing this in
1062 // LoopUnswitch pass is that it could potentially break LoopPassManager's
1063 // invariants. Folding dead branches could either eliminate the current loop
1064 // or make other loops unreachable. LCSSA form might also not be preserved
1065 // after deleting branches. The following code keeps traversing loop header's
1066 // successors until it finds the trivial condition candidate (condition that
1067 // is not a constant). Since unswitching generates branches with constant
1068 // conditions, this scenario could be very common in practice.
1069 BasicBlock *CurrentBB = L.getHeader();
1070 SmallPtrSet<BasicBlock *, 8> Visited;
1071 Visited.insert(Ptr: CurrentBB);
1072 do {
1073 // Check if there are any side-effecting instructions (e.g. stores, calls,
1074 // volatile loads) in the part of the loop that the code *would* execute
1075 // without unswitching.
1076 if (MSSAU) // Possible early exit with MSSA
1077 if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(BB: CurrentBB))
1078 if (!isa<MemoryPhi>(Val: *Defs->begin()) || (++Defs->begin() != Defs->end()))
1079 return Changed;
1080 if (llvm::any_of(Range&: *CurrentBB,
1081 P: [](Instruction &I) { return I.mayHaveSideEffects(); }))
1082 return Changed;
1083
1084 Instruction *CurrentTerm = CurrentBB->getTerminator();
1085
1086 if (auto *SI = dyn_cast<SwitchInst>(Val: CurrentTerm)) {
1087 // Don't bother trying to unswitch past a switch with a constant
1088 // condition. This should be removed prior to running this pass by
1089 // simplifycfg.
1090 if (isa<Constant>(Val: SI->getCondition()))
1091 return Changed;
1092
1093 if (!unswitchTrivialSwitch(L, SI&: *SI, DT, LI, SE, MSSAU))
1094 // Couldn't unswitch this one so we're done.
1095 return Changed;
1096
1097 // Mark that we managed to unswitch something.
1098 Changed = true;
1099
1100 // If unswitching turned the terminator into an unconditional branch then
1101 // we can continue. The unswitching logic specifically works to fold any
1102 // cases it can into an unconditional branch to make it easier to
1103 // recognize here.
1104 auto *BI = dyn_cast<BranchInst>(Val: CurrentBB->getTerminator());
1105 if (!BI || BI->isConditional())
1106 return Changed;
1107
1108 CurrentBB = BI->getSuccessor(i: 0);
1109 continue;
1110 }
1111
1112 auto *BI = dyn_cast<BranchInst>(Val: CurrentTerm);
1113 if (!BI)
1114 // We do not understand other terminator instructions.
1115 return Changed;
1116
1117 // Don't bother trying to unswitch past an unconditional branch or a branch
1118 // with a constant value. These should be removed by simplifycfg prior to
1119 // running this pass.
1120 if (!BI->isConditional() ||
1121 isa<Constant>(Val: skipTrivialSelect(Cond: BI->getCondition())))
1122 return Changed;
1123
1124 // Found a trivial condition candidate: non-foldable conditional branch. If
1125 // we fail to unswitch this, we can't do anything else that is trivial.
1126 if (!unswitchTrivialBranch(L, BI&: *BI, DT, LI, SE, MSSAU))
1127 return Changed;
1128
1129 // Mark that we managed to unswitch something.
1130 Changed = true;
1131
1132 // If we only unswitched some of the conditions feeding the branch, we won't
1133 // have collapsed it to a single successor.
1134 BI = cast<BranchInst>(Val: CurrentBB->getTerminator());
1135 if (BI->isConditional())
1136 return Changed;
1137
1138 // Follow the newly unconditional branch into its successor.
1139 CurrentBB = BI->getSuccessor(i: 0);
1140
1141 // When continuing, if we exit the loop or reach a previous visited block,
1142 // then we can not reach any trivial condition candidates (unfoldable
1143 // branch instructions or switch instructions) and no unswitch can happen.
1144 } while (L.contains(BB: CurrentBB) && Visited.insert(Ptr: CurrentBB).second);
1145
1146 return Changed;
1147}
1148
1149/// Build the cloned blocks for an unswitched copy of the given loop.
1150///
1151/// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
1152/// after the split block (`SplitBB`) that will be used to select between the
1153/// cloned and original loop.
1154///
1155/// This routine handles cloning all of the necessary loop blocks and exit
1156/// blocks including rewriting their instructions and the relevant PHI nodes.
1157/// Any loop blocks or exit blocks which are dominated by a different successor
1158/// than the one for this clone of the loop blocks can be trivially skipped. We
1159/// use the `DominatingSucc` map to determine whether a block satisfies that
1160/// property with a simple map lookup.
1161///
1162/// It also correctly creates the unconditional branch in the cloned
1163/// unswitched parent block to only point at the unswitched successor.
1164///
1165/// This does not handle most of the necessary updates to `LoopInfo`. Only exit
1166/// block splitting is correctly reflected in `LoopInfo`, essentially all of
1167/// the cloned blocks (and their loops) are left without full `LoopInfo`
1168/// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
1169/// blocks to them but doesn't create the cloned `DominatorTree` structure and
1170/// instead the caller must recompute an accurate DT. It *does* correctly
1171/// update the `AssumptionCache` provided in `AC`.
1172static BasicBlock *buildClonedLoopBlocks(
1173 Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
1174 ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
1175 BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
1176 const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
1177 ValueToValueMapTy &VMap,
1178 SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
1179 DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU,
1180 ScalarEvolution *SE) {
1181 SmallVector<BasicBlock *, 4> NewBlocks;
1182 NewBlocks.reserve(N: L.getNumBlocks() + ExitBlocks.size());
1183
1184 // We will need to clone a bunch of blocks, wrap up the clone operation in
1185 // a helper.
1186 auto CloneBlock = [&](BasicBlock *OldBB) {
1187 // Clone the basic block and insert it before the new preheader.
1188 BasicBlock *NewBB = CloneBasicBlock(BB: OldBB, VMap, NameSuffix: ".us", F: OldBB->getParent());
1189 NewBB->moveBefore(MovePos: LoopPH);
1190
1191 // Record this block and the mapping.
1192 NewBlocks.push_back(Elt: NewBB);
1193 VMap[OldBB] = NewBB;
1194
1195 return NewBB;
1196 };
1197
1198 // We skip cloning blocks when they have a dominating succ that is not the
1199 // succ we are cloning for.
1200 auto SkipBlock = [&](BasicBlock *BB) {
1201 auto It = DominatingSucc.find(Val: BB);
1202 return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
1203 };
1204
1205 // First, clone the preheader.
1206 auto *ClonedPH = CloneBlock(LoopPH);
1207
1208 // Then clone all the loop blocks, skipping the ones that aren't necessary.
1209 for (auto *LoopBB : L.blocks())
1210 if (!SkipBlock(LoopBB))
1211 CloneBlock(LoopBB);
1212
1213 // Split all the loop exit edges so that when we clone the exit blocks, if
1214 // any of the exit blocks are *also* a preheader for some other loop, we
1215 // don't create multiple predecessors entering the loop header.
1216 for (auto *ExitBB : ExitBlocks) {
1217 if (SkipBlock(ExitBB))
1218 continue;
1219
1220 // When we are going to clone an exit, we don't need to clone all the
1221 // instructions in the exit block and we want to ensure we have an easy
1222 // place to merge the CFG, so split the exit first. This is always safe to
1223 // do because there cannot be any non-loop predecessors of a loop exit in
1224 // loop simplified form.
1225 auto *MergeBB = SplitBlock(Old: ExitBB, SplitPt: ExitBB->begin(), DT: &DT, LI: &LI, MSSAU);
1226
1227 // Rearrange the names to make it easier to write test cases by having the
1228 // exit block carry the suffix rather than the merge block carrying the
1229 // suffix.
1230 MergeBB->takeName(V: ExitBB);
1231 ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1232
1233 // Now clone the original exit block.
1234 auto *ClonedExitBB = CloneBlock(ExitBB);
1235 assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1236 "Exit block should have been split to have one successor!");
1237 assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1238 "Cloned exit block has the wrong successor!");
1239
1240 // Remap any cloned instructions and create a merge phi node for them.
1241 for (auto ZippedInsts : llvm::zip_first(
1242 t: llvm::make_range(x: ExitBB->begin(), y: std::prev(x: ExitBB->end())),
1243 u: llvm::make_range(x: ClonedExitBB->begin(),
1244 y: std::prev(x: ClonedExitBB->end())))) {
1245 Instruction &I = std::get<0>(t&: ZippedInsts);
1246 Instruction &ClonedI = std::get<1>(t&: ZippedInsts);
1247
1248 // The only instructions in the exit block should be PHI nodes and
1249 // potentially a landing pad.
1250 assert(
1251 (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1252 "Bad instruction in exit block!");
1253 // We should have a value map between the instruction and its clone.
1254 assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1255
1256 // Forget SCEVs based on exit phis in case SCEV looked through the phi.
1257 if (SE)
1258 if (auto *PN = dyn_cast<PHINode>(Val: &I))
1259 SE->forgetLcssaPhiWithNewPredecessor(L: &L, V: PN);
1260
1261 BasicBlock::iterator InsertPt = MergeBB->getFirstInsertionPt();
1262
1263 auto *MergePN =
1264 PHINode::Create(Ty: I.getType(), /*NumReservedValues*/ 2, NameStr: ".us-phi");
1265 MergePN->insertBefore(InsertPos: InsertPt);
1266 MergePN->setDebugLoc(InsertPt->getDebugLoc());
1267 I.replaceAllUsesWith(V: MergePN);
1268 MergePN->addIncoming(V: &I, BB: ExitBB);
1269 MergePN->addIncoming(V: &ClonedI, BB: ClonedExitBB);
1270 }
1271 }
1272
1273 // Rewrite the instructions in the cloned blocks to refer to the instructions
1274 // in the cloned blocks. We have to do this as a second pass so that we have
1275 // everything available. Also, we have inserted new instructions which may
1276 // include assume intrinsics, so we update the assumption cache while
1277 // processing this.
1278 Module *M = ClonedPH->getParent()->getParent();
1279 for (auto *ClonedBB : NewBlocks)
1280 for (Instruction &I : *ClonedBB) {
1281 RemapDbgRecordRange(M, Range: I.getDbgRecordRange(), VM&: VMap,
1282 Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1283 RemapInstruction(I: &I, VM&: VMap,
1284 Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1285 if (auto *II = dyn_cast<AssumeInst>(Val: &I))
1286 AC.registerAssumption(CI: II);
1287 }
1288
1289 // Update any PHI nodes in the cloned successors of the skipped blocks to not
1290 // have spurious incoming values.
1291 for (auto *LoopBB : L.blocks())
1292 if (SkipBlock(LoopBB))
1293 for (auto *SuccBB : successors(BB: LoopBB))
1294 if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: SuccBB)))
1295 for (PHINode &PN : ClonedSuccBB->phis())
1296 PN.removeIncomingValue(BB: LoopBB, /*DeletePHIIfEmpty*/ false);
1297
1298 // Remove the cloned parent as a predecessor of any successor we ended up
1299 // cloning other than the unswitched one.
1300 auto *ClonedParentBB = cast<BasicBlock>(Val: VMap.lookup(Val: ParentBB));
1301 for (auto *SuccBB : successors(BB: ParentBB)) {
1302 if (SuccBB == UnswitchedSuccBB)
1303 continue;
1304
1305 auto *ClonedSuccBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: SuccBB));
1306 if (!ClonedSuccBB)
1307 continue;
1308
1309 ClonedSuccBB->removePredecessor(Pred: ClonedParentBB,
1310 /*KeepOneInputPHIs*/ true);
1311 }
1312
1313 // Replace the cloned branch with an unconditional branch to the cloned
1314 // unswitched successor.
1315 auto *ClonedSuccBB = cast<BasicBlock>(Val: VMap.lookup(Val: UnswitchedSuccBB));
1316 Instruction *ClonedTerminator = ClonedParentBB->getTerminator();
1317 // Trivial Simplification. If Terminator is a conditional branch and
1318 // condition becomes dead - erase it.
1319 Value *ClonedConditionToErase = nullptr;
1320 if (auto *BI = dyn_cast<BranchInst>(Val: ClonedTerminator))
1321 ClonedConditionToErase = BI->getCondition();
1322 else if (auto *SI = dyn_cast<SwitchInst>(Val: ClonedTerminator))
1323 ClonedConditionToErase = SI->getCondition();
1324
1325 Instruction *BI = BranchInst::Create(IfTrue: ClonedSuccBB, InsertBefore: ClonedParentBB);
1326 BI->setDebugLoc(ClonedTerminator->getDebugLoc());
1327 ClonedTerminator->eraseFromParent();
1328
1329 if (ClonedConditionToErase)
1330 RecursivelyDeleteTriviallyDeadInstructions(V: ClonedConditionToErase, TLI: nullptr,
1331 MSSAU);
1332
1333 // If there are duplicate entries in the PHI nodes because of multiple edges
1334 // to the unswitched successor, we need to nuke all but one as we replaced it
1335 // with a direct branch.
1336 for (PHINode &PN : ClonedSuccBB->phis()) {
1337 bool Found = false;
1338 // Loop over the incoming operands backwards so we can easily delete as we
1339 // go without invalidating the index.
1340 for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1341 if (PN.getIncomingBlock(i) != ClonedParentBB)
1342 continue;
1343 if (!Found) {
1344 Found = true;
1345 continue;
1346 }
1347 PN.removeIncomingValue(Idx: i, /*DeletePHIIfEmpty*/ false);
1348 }
1349 }
1350
1351 // Record the domtree updates for the new blocks.
1352 SmallPtrSet<BasicBlock *, 4> SuccSet;
1353 for (auto *ClonedBB : NewBlocks) {
1354 for (auto *SuccBB : successors(BB: ClonedBB))
1355 if (SuccSet.insert(Ptr: SuccBB).second)
1356 DTUpdates.push_back(Elt: {DominatorTree::Insert, ClonedBB, SuccBB});
1357 SuccSet.clear();
1358 }
1359
1360 return ClonedPH;
1361}
1362
1363/// Recursively clone the specified loop and all of its children.
1364///
1365/// The target parent loop for the clone should be provided, or can be null if
1366/// the clone is a top-level loop. While cloning, all the blocks are mapped
1367/// with the provided value map. The entire original loop must be present in
1368/// the value map. The cloned loop is returned.
1369static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1370 const ValueToValueMapTy &VMap, LoopInfo &LI) {
1371 auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1372 assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1373 ClonedL.reserveBlocks(size: OrigL.getNumBlocks());
1374 for (auto *BB : OrigL.blocks()) {
1375 auto *ClonedBB = cast<BasicBlock>(Val: VMap.lookup(Val: BB));
1376 ClonedL.addBlockEntry(BB: ClonedBB);
1377 if (LI.getLoopFor(BB) == &OrigL)
1378 LI.changeLoopFor(BB: ClonedBB, L: &ClonedL);
1379 }
1380 };
1381
1382 // We specially handle the first loop because it may get cloned into
1383 // a different parent and because we most commonly are cloning leaf loops.
1384 Loop *ClonedRootL = LI.AllocateLoop();
1385 if (RootParentL)
1386 RootParentL->addChildLoop(NewChild: ClonedRootL);
1387 else
1388 LI.addTopLevelLoop(New: ClonedRootL);
1389 AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1390
1391 if (OrigRootL.isInnermost())
1392 return ClonedRootL;
1393
1394 // If we have a nest, we can quickly clone the entire loop nest using an
1395 // iterative approach because it is a tree. We keep the cloned parent in the
1396 // data structure to avoid repeatedly querying through a map to find it.
1397 SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1398 // Build up the loops to clone in reverse order as we'll clone them from the
1399 // back.
1400 for (Loop *ChildL : llvm::reverse(C&: OrigRootL))
1401 LoopsToClone.push_back(Elt: {ClonedRootL, ChildL});
1402 do {
1403 Loop *ClonedParentL, *L;
1404 std::tie(args&: ClonedParentL, args&: L) = LoopsToClone.pop_back_val();
1405 Loop *ClonedL = LI.AllocateLoop();
1406 ClonedParentL->addChildLoop(NewChild: ClonedL);
1407 AddClonedBlocksToLoop(*L, *ClonedL);
1408 for (Loop *ChildL : llvm::reverse(C&: *L))
1409 LoopsToClone.push_back(Elt: {ClonedL, ChildL});
1410 } while (!LoopsToClone.empty());
1411
1412 return ClonedRootL;
1413}
1414
1415/// Build the cloned loops of an original loop from unswitching.
1416///
1417/// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1418/// operation. We need to re-verify that there even is a loop (as the backedge
1419/// may not have been cloned), and even if there are remaining backedges the
1420/// backedge set may be different. However, we know that each child loop is
1421/// undisturbed, we only need to find where to place each child loop within
1422/// either any parent loop or within a cloned version of the original loop.
1423///
1424/// Because child loops may end up cloned outside of any cloned version of the
1425/// original loop, multiple cloned sibling loops may be created. All of them
1426/// are returned so that the newly introduced loop nest roots can be
1427/// identified.
1428static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1429 const ValueToValueMapTy &VMap, LoopInfo &LI,
1430 SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1431 Loop *ClonedL = nullptr;
1432
1433 auto *OrigPH = OrigL.getLoopPreheader();
1434 auto *OrigHeader = OrigL.getHeader();
1435
1436 auto *ClonedPH = cast<BasicBlock>(Val: VMap.lookup(Val: OrigPH));
1437 auto *ClonedHeader = cast<BasicBlock>(Val: VMap.lookup(Val: OrigHeader));
1438
1439 // We need to know the loops of the cloned exit blocks to even compute the
1440 // accurate parent loop. If we only clone exits to some parent of the
1441 // original parent, we want to clone into that outer loop. We also keep track
1442 // of the loops that our cloned exit blocks participate in.
1443 Loop *ParentL = nullptr;
1444 SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1445 SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1446 ClonedExitsInLoops.reserve(N: ExitBlocks.size());
1447 for (auto *ExitBB : ExitBlocks)
1448 if (auto *ClonedExitBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: ExitBB)))
1449 if (Loop *ExitL = LI.getLoopFor(BB: ExitBB)) {
1450 ExitLoopMap[ClonedExitBB] = ExitL;
1451 ClonedExitsInLoops.push_back(Elt: ClonedExitBB);
1452 if (!ParentL || (ParentL != ExitL && ParentL->contains(L: ExitL)))
1453 ParentL = ExitL;
1454 }
1455 assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1456 ParentL->contains(OrigL.getParentLoop())) &&
1457 "The computed parent loop should always contain (or be) the parent of "
1458 "the original loop.");
1459
1460 // We build the set of blocks dominated by the cloned header from the set of
1461 // cloned blocks out of the original loop. While not all of these will
1462 // necessarily be in the cloned loop, it is enough to establish that they
1463 // aren't in unreachable cycles, etc.
1464 SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1465 for (auto *BB : OrigL.blocks())
1466 if (auto *ClonedBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: BB)))
1467 ClonedLoopBlocks.insert(X: ClonedBB);
1468
1469 // Rebuild the set of blocks that will end up in the cloned loop. We may have
1470 // skipped cloning some region of this loop which can in turn skip some of
1471 // the backedges so we have to rebuild the blocks in the loop based on the
1472 // backedges that remain after cloning.
1473 SmallVector<BasicBlock *, 16> Worklist;
1474 SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1475 for (auto *Pred : predecessors(BB: ClonedHeader)) {
1476 // The only possible non-loop header predecessor is the preheader because
1477 // we know we cloned the loop in simplified form.
1478 if (Pred == ClonedPH)
1479 continue;
1480
1481 // Because the loop was in simplified form, the only non-loop predecessor
1482 // should be the preheader.
1483 assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1484 "header other than the preheader "
1485 "that is not part of the loop!");
1486
1487 // Insert this block into the loop set and on the first visit (and if it
1488 // isn't the header we're currently walking) put it into the worklist to
1489 // recurse through.
1490 if (BlocksInClonedLoop.insert(Ptr: Pred).second && Pred != ClonedHeader)
1491 Worklist.push_back(Elt: Pred);
1492 }
1493
1494 // If we had any backedges then there *is* a cloned loop. Put the header into
1495 // the loop set and then walk the worklist backwards to find all the blocks
1496 // that remain within the loop after cloning.
1497 if (!BlocksInClonedLoop.empty()) {
1498 BlocksInClonedLoop.insert(Ptr: ClonedHeader);
1499
1500 while (!Worklist.empty()) {
1501 BasicBlock *BB = Worklist.pop_back_val();
1502 assert(BlocksInClonedLoop.count(BB) &&
1503 "Didn't put block into the loop set!");
1504
1505 // Insert any predecessors that are in the possible set into the cloned
1506 // set, and if the insert is successful, add them to the worklist. Note
1507 // that we filter on the blocks that are definitely reachable via the
1508 // backedge to the loop header so we may prune out dead code within the
1509 // cloned loop.
1510 for (auto *Pred : predecessors(BB))
1511 if (ClonedLoopBlocks.count(key: Pred) &&
1512 BlocksInClonedLoop.insert(Ptr: Pred).second)
1513 Worklist.push_back(Elt: Pred);
1514 }
1515
1516 ClonedL = LI.AllocateLoop();
1517 if (ParentL) {
1518 ParentL->addBasicBlockToLoop(NewBB: ClonedPH, LI);
1519 ParentL->addChildLoop(NewChild: ClonedL);
1520 } else {
1521 LI.addTopLevelLoop(New: ClonedL);
1522 }
1523 NonChildClonedLoops.push_back(Elt: ClonedL);
1524
1525 ClonedL->reserveBlocks(size: BlocksInClonedLoop.size());
1526 // We don't want to just add the cloned loop blocks based on how we
1527 // discovered them. The original order of blocks was carefully built in
1528 // a way that doesn't rely on predecessor ordering. Rather than re-invent
1529 // that logic, we just re-walk the original blocks (and those of the child
1530 // loops) and filter them as we add them into the cloned loop.
1531 for (auto *BB : OrigL.blocks()) {
1532 auto *ClonedBB = cast_or_null<BasicBlock>(Val: VMap.lookup(Val: BB));
1533 if (!ClonedBB || !BlocksInClonedLoop.count(Ptr: ClonedBB))
1534 continue;
1535
1536 // Directly add the blocks that are only in this loop.
1537 if (LI.getLoopFor(BB) == &OrigL) {
1538 ClonedL->addBasicBlockToLoop(NewBB: ClonedBB, LI);
1539 continue;
1540 }
1541
1542 // We want to manually add it to this loop and parents.
1543 // Registering it with LoopInfo will happen when we clone the top
1544 // loop for this block.
1545 for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1546 PL->addBlockEntry(BB: ClonedBB);
1547 }
1548
1549 // Now add each child loop whose header remains within the cloned loop. All
1550 // of the blocks within the loop must satisfy the same constraints as the
1551 // header so once we pass the header checks we can just clone the entire
1552 // child loop nest.
1553 for (Loop *ChildL : OrigL) {
1554 auto *ClonedChildHeader =
1555 cast_or_null<BasicBlock>(Val: VMap.lookup(Val: ChildL->getHeader()));
1556 if (!ClonedChildHeader || !BlocksInClonedLoop.count(Ptr: ClonedChildHeader))
1557 continue;
1558
1559#ifndef NDEBUG
1560 // We should never have a cloned child loop header but fail to have
1561 // all of the blocks for that child loop.
1562 for (auto *ChildLoopBB : ChildL->blocks())
1563 assert(BlocksInClonedLoop.count(
1564 cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1565 "Child cloned loop has a header within the cloned outer "
1566 "loop but not all of its blocks!");
1567#endif
1568
1569 cloneLoopNest(OrigRootL&: *ChildL, RootParentL: ClonedL, VMap, LI);
1570 }
1571 }
1572
1573 // Now that we've handled all the components of the original loop that were
1574 // cloned into a new loop, we still need to handle anything from the original
1575 // loop that wasn't in a cloned loop.
1576
1577 // Figure out what blocks are left to place within any loop nest containing
1578 // the unswitched loop. If we never formed a loop, the cloned PH is one of
1579 // them.
1580 SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1581 if (BlocksInClonedLoop.empty())
1582 UnloopedBlockSet.insert(Ptr: ClonedPH);
1583 for (auto *ClonedBB : ClonedLoopBlocks)
1584 if (!BlocksInClonedLoop.count(Ptr: ClonedBB))
1585 UnloopedBlockSet.insert(Ptr: ClonedBB);
1586
1587 // Copy the cloned exits and sort them in ascending loop depth, we'll work
1588 // backwards across these to process them inside out. The order shouldn't
1589 // matter as we're just trying to build up the map from inside-out; we use
1590 // the map in a more stably ordered way below.
1591 auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1592 llvm::sort(C&: OrderedClonedExitsInLoops, Comp: [&](BasicBlock *LHS, BasicBlock *RHS) {
1593 return ExitLoopMap.lookup(Val: LHS)->getLoopDepth() <
1594 ExitLoopMap.lookup(Val: RHS)->getLoopDepth();
1595 });
1596
1597 // Populate the existing ExitLoopMap with everything reachable from each
1598 // exit, starting from the inner most exit.
1599 while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1600 assert(Worklist.empty() && "Didn't clear worklist!");
1601
1602 BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1603 Loop *ExitL = ExitLoopMap.lookup(Val: ExitBB);
1604
1605 // Walk the CFG back until we hit the cloned PH adding everything reachable
1606 // and in the unlooped set to this exit block's loop.
1607 Worklist.push_back(Elt: ExitBB);
1608 do {
1609 BasicBlock *BB = Worklist.pop_back_val();
1610 // We can stop recursing at the cloned preheader (if we get there).
1611 if (BB == ClonedPH)
1612 continue;
1613
1614 for (BasicBlock *PredBB : predecessors(BB)) {
1615 // If this pred has already been moved to our set or is part of some
1616 // (inner) loop, no update needed.
1617 if (!UnloopedBlockSet.erase(Ptr: PredBB)) {
1618 assert(
1619 (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1620 "Predecessor not mapped to a loop!");
1621 continue;
1622 }
1623
1624 // We just insert into the loop set here. We'll add these blocks to the
1625 // exit loop after we build up the set in an order that doesn't rely on
1626 // predecessor order (which in turn relies on use list order).
1627 bool Inserted = ExitLoopMap.insert(KV: {PredBB, ExitL}).second;
1628 (void)Inserted;
1629 assert(Inserted && "Should only visit an unlooped block once!");
1630
1631 // And recurse through to its predecessors.
1632 Worklist.push_back(Elt: PredBB);
1633 }
1634 } while (!Worklist.empty());
1635 }
1636
1637 // Now that the ExitLoopMap gives as mapping for all the non-looping cloned
1638 // blocks to their outer loops, walk the cloned blocks and the cloned exits
1639 // in their original order adding them to the correct loop.
1640
1641 // We need a stable insertion order. We use the order of the original loop
1642 // order and map into the correct parent loop.
1643 for (auto *BB : llvm::concat<BasicBlock *const>(
1644 Ranges: ArrayRef(ClonedPH), Ranges&: ClonedLoopBlocks, Ranges&: ClonedExitsInLoops))
1645 if (Loop *OuterL = ExitLoopMap.lookup(Val: BB))
1646 OuterL->addBasicBlockToLoop(NewBB: BB, LI);
1647
1648#ifndef NDEBUG
1649 for (auto &BBAndL : ExitLoopMap) {
1650 auto *BB = BBAndL.first;
1651 auto *OuterL = BBAndL.second;
1652 assert(LI.getLoopFor(BB) == OuterL &&
1653 "Failed to put all blocks into outer loops!");
1654 }
1655#endif
1656
1657 // Now that all the blocks are placed into the correct containing loop in the
1658 // absence of child loops, find all the potentially cloned child loops and
1659 // clone them into whatever outer loop we placed their header into.
1660 for (Loop *ChildL : OrigL) {
1661 auto *ClonedChildHeader =
1662 cast_or_null<BasicBlock>(Val: VMap.lookup(Val: ChildL->getHeader()));
1663 if (!ClonedChildHeader || BlocksInClonedLoop.count(Ptr: ClonedChildHeader))
1664 continue;
1665
1666#ifndef NDEBUG
1667 for (auto *ChildLoopBB : ChildL->blocks())
1668 assert(VMap.count(ChildLoopBB) &&
1669 "Cloned a child loop header but not all of that loops blocks!");
1670#endif
1671
1672 NonChildClonedLoops.push_back(Elt: cloneLoopNest(
1673 OrigRootL&: *ChildL, RootParentL: ExitLoopMap.lookup(Val: ClonedChildHeader), VMap, LI));
1674 }
1675}
1676
1677static void
1678deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1679 ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1680 DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1681 // Find all the dead clones, and remove them from their successors.
1682 SmallVector<BasicBlock *, 16> DeadBlocks;
1683 for (BasicBlock *BB : llvm::concat<BasicBlock *const>(Ranges: L.blocks(), Ranges&: ExitBlocks))
1684 for (const auto &VMap : VMaps)
1685 if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(Val: VMap->lookup(Val: BB)))
1686 if (!DT.isReachableFromEntry(A: ClonedBB)) {
1687 for (BasicBlock *SuccBB : successors(BB: ClonedBB))
1688 SuccBB->removePredecessor(Pred: ClonedBB);
1689 DeadBlocks.push_back(Elt: ClonedBB);
1690 }
1691
1692 // Remove all MemorySSA in the dead blocks
1693 if (MSSAU) {
1694 SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(),
1695 DeadBlocks.end());
1696 MSSAU->removeBlocks(DeadBlocks: DeadBlockSet);
1697 }
1698
1699 // Drop any remaining references to break cycles.
1700 for (BasicBlock *BB : DeadBlocks)
1701 BB->dropAllReferences();
1702 // Erase them from the IR.
1703 for (BasicBlock *BB : DeadBlocks)
1704 BB->eraseFromParent();
1705}
1706
1707static void deleteDeadBlocksFromLoop(Loop &L,
1708 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1709 DominatorTree &DT, LoopInfo &LI,
1710 MemorySSAUpdater *MSSAU,
1711 ScalarEvolution *SE,
1712 LPMUpdater &LoopUpdater) {
1713 // Find all the dead blocks tied to this loop, and remove them from their
1714 // successors.
1715 SmallSetVector<BasicBlock *, 8> DeadBlockSet;
1716
1717 // Start with loop/exit blocks and get a transitive closure of reachable dead
1718 // blocks.
1719 SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1720 ExitBlocks.end());
1721 DeathCandidates.append(in_start: L.blocks().begin(), in_end: L.blocks().end());
1722 while (!DeathCandidates.empty()) {
1723 auto *BB = DeathCandidates.pop_back_val();
1724 if (!DeadBlockSet.count(key: BB) && !DT.isReachableFromEntry(A: BB)) {
1725 for (BasicBlock *SuccBB : successors(BB)) {
1726 SuccBB->removePredecessor(Pred: BB);
1727 DeathCandidates.push_back(Elt: SuccBB);
1728 }
1729 DeadBlockSet.insert(X: BB);
1730 }
1731 }
1732
1733 // Remove all MemorySSA in the dead blocks
1734 if (MSSAU)
1735 MSSAU->removeBlocks(DeadBlocks: DeadBlockSet);
1736
1737 // Filter out the dead blocks from the exit blocks list so that it can be
1738 // used in the caller.
1739 llvm::erase_if(C&: ExitBlocks,
1740 P: [&](BasicBlock *BB) { return DeadBlockSet.count(key: BB); });
1741
1742 // Walk from this loop up through its parents removing all of the dead blocks.
1743 for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1744 for (auto *BB : DeadBlockSet)
1745 ParentL->getBlocksSet().erase(Ptr: BB);
1746 llvm::erase_if(C&: ParentL->getBlocksVector(),
1747 P: [&](BasicBlock *BB) { return DeadBlockSet.count(key: BB); });
1748 }
1749
1750 // Now delete the dead child loops. This raw delete will clear them
1751 // recursively.
1752 llvm::erase_if(C&: L.getSubLoopsVector(), P: [&](Loop *ChildL) {
1753 if (!DeadBlockSet.count(key: ChildL->getHeader()))
1754 return false;
1755
1756 assert(llvm::all_of(ChildL->blocks(),
1757 [&](BasicBlock *ChildBB) {
1758 return DeadBlockSet.count(ChildBB);
1759 }) &&
1760 "If the child loop header is dead all blocks in the child loop must "
1761 "be dead as well!");
1762 LoopUpdater.markLoopAsDeleted(L&: *ChildL, Name: ChildL->getName());
1763 if (SE)
1764 SE->forgetBlockAndLoopDispositions();
1765 LI.destroy(L: ChildL);
1766 return true;
1767 });
1768
1769 // Remove the loop mappings for the dead blocks and drop all the references
1770 // from these blocks to others to handle cyclic references as we start
1771 // deleting the blocks themselves.
1772 for (auto *BB : DeadBlockSet) {
1773 // Check that the dominator tree has already been updated.
1774 assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1775 LI.changeLoopFor(BB, L: nullptr);
1776 // Drop all uses of the instructions to make sure we won't have dangling
1777 // uses in other blocks.
1778 for (auto &I : *BB)
1779 if (!I.use_empty())
1780 I.replaceAllUsesWith(V: PoisonValue::get(T: I.getType()));
1781 BB->dropAllReferences();
1782 }
1783
1784 // Actually delete the blocks now that they've been fully unhooked from the
1785 // IR.
1786 for (auto *BB : DeadBlockSet)
1787 BB->eraseFromParent();
1788}
1789
1790/// Recompute the set of blocks in a loop after unswitching.
1791///
1792/// This walks from the original headers predecessors to rebuild the loop. We
1793/// take advantage of the fact that new blocks can't have been added, and so we
1794/// filter by the original loop's blocks. This also handles potentially
1795/// unreachable code that we don't want to explore but might be found examining
1796/// the predecessors of the header.
1797///
1798/// If the original loop is no longer a loop, this will return an empty set. If
1799/// it remains a loop, all the blocks within it will be added to the set
1800/// (including those blocks in inner loops).
1801static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1802 LoopInfo &LI) {
1803 SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1804
1805 auto *PH = L.getLoopPreheader();
1806 auto *Header = L.getHeader();
1807
1808 // A worklist to use while walking backwards from the header.
1809 SmallVector<BasicBlock *, 16> Worklist;
1810
1811 // First walk the predecessors of the header to find the backedges. This will
1812 // form the basis of our walk.
1813 for (auto *Pred : predecessors(BB: Header)) {
1814 // Skip the preheader.
1815 if (Pred == PH)
1816 continue;
1817
1818 // Because the loop was in simplified form, the only non-loop predecessor
1819 // is the preheader.
1820 assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1821 "than the preheader that is not part of the "
1822 "loop!");
1823
1824 // Insert this block into the loop set and on the first visit and, if it
1825 // isn't the header we're currently walking, put it into the worklist to
1826 // recurse through.
1827 if (LoopBlockSet.insert(Ptr: Pred).second && Pred != Header)
1828 Worklist.push_back(Elt: Pred);
1829 }
1830
1831 // If no backedges were found, we're done.
1832 if (LoopBlockSet.empty())
1833 return LoopBlockSet;
1834
1835 // We found backedges, recurse through them to identify the loop blocks.
1836 while (!Worklist.empty()) {
1837 BasicBlock *BB = Worklist.pop_back_val();
1838 assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1839
1840 // No need to walk past the header.
1841 if (BB == Header)
1842 continue;
1843
1844 // Because we know the inner loop structure remains valid we can use the
1845 // loop structure to jump immediately across the entire nested loop.
1846 // Further, because it is in loop simplified form, we can directly jump
1847 // to its preheader afterward.
1848 if (Loop *InnerL = LI.getLoopFor(BB))
1849 if (InnerL != &L) {
1850 assert(L.contains(InnerL) &&
1851 "Should not reach a loop *outside* this loop!");
1852 // The preheader is the only possible predecessor of the loop so
1853 // insert it into the set and check whether it was already handled.
1854 auto *InnerPH = InnerL->getLoopPreheader();
1855 assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1856 "but not contain the inner loop "
1857 "preheader!");
1858 if (!LoopBlockSet.insert(Ptr: InnerPH).second)
1859 // The only way to reach the preheader is through the loop body
1860 // itself so if it has been visited the loop is already handled.
1861 continue;
1862
1863 // Insert all of the blocks (other than those already present) into
1864 // the loop set. We expect at least the block that led us to find the
1865 // inner loop to be in the block set, but we may also have other loop
1866 // blocks if they were already enqueued as predecessors of some other
1867 // outer loop block.
1868 for (auto *InnerBB : InnerL->blocks()) {
1869 if (InnerBB == BB) {
1870 assert(LoopBlockSet.count(InnerBB) &&
1871 "Block should already be in the set!");
1872 continue;
1873 }
1874
1875 LoopBlockSet.insert(Ptr: InnerBB);
1876 }
1877
1878 // Add the preheader to the worklist so we will continue past the
1879 // loop body.
1880 Worklist.push_back(Elt: InnerPH);
1881 continue;
1882 }
1883
1884 // Insert any predecessors that were in the original loop into the new
1885 // set, and if the insert is successful, add them to the worklist.
1886 for (auto *Pred : predecessors(BB))
1887 if (L.contains(BB: Pred) && LoopBlockSet.insert(Ptr: Pred).second)
1888 Worklist.push_back(Elt: Pred);
1889 }
1890
1891 assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1892
1893 // We've found all the blocks participating in the loop, return our completed
1894 // set.
1895 return LoopBlockSet;
1896}
1897
1898/// Rebuild a loop after unswitching removes some subset of blocks and edges.
1899///
1900/// The removal may have removed some child loops entirely but cannot have
1901/// disturbed any remaining child loops. However, they may need to be hoisted
1902/// to the parent loop (or to be top-level loops). The original loop may be
1903/// completely removed.
1904///
1905/// The sibling loops resulting from this update are returned. If the original
1906/// loop remains a valid loop, it will be the first entry in this list with all
1907/// of the newly sibling loops following it.
1908///
1909/// Returns true if the loop remains a loop after unswitching, and false if it
1910/// is no longer a loop after unswitching (and should not continue to be
1911/// referenced).
1912static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1913 LoopInfo &LI,
1914 SmallVectorImpl<Loop *> &HoistedLoops,
1915 ScalarEvolution *SE) {
1916 auto *PH = L.getLoopPreheader();
1917
1918 // Compute the actual parent loop from the exit blocks. Because we may have
1919 // pruned some exits the loop may be different from the original parent.
1920 Loop *ParentL = nullptr;
1921 SmallVector<Loop *, 4> ExitLoops;
1922 SmallVector<BasicBlock *, 4> ExitsInLoops;
1923 ExitsInLoops.reserve(N: ExitBlocks.size());
1924 for (auto *ExitBB : ExitBlocks)
1925 if (Loop *ExitL = LI.getLoopFor(BB: ExitBB)) {
1926 ExitLoops.push_back(Elt: ExitL);
1927 ExitsInLoops.push_back(Elt: ExitBB);
1928 if (!ParentL || (ParentL != ExitL && ParentL->contains(L: ExitL)))
1929 ParentL = ExitL;
1930 }
1931
1932 // Recompute the blocks participating in this loop. This may be empty if it
1933 // is no longer a loop.
1934 auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1935
1936 // If we still have a loop, we need to re-set the loop's parent as the exit
1937 // block set changing may have moved it within the loop nest. Note that this
1938 // can only happen when this loop has a parent as it can only hoist the loop
1939 // *up* the nest.
1940 if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1941 // Remove this loop's (original) blocks from all of the intervening loops.
1942 for (Loop *IL = L.getParentLoop(); IL != ParentL;
1943 IL = IL->getParentLoop()) {
1944 IL->getBlocksSet().erase(Ptr: PH);
1945 for (auto *BB : L.blocks())
1946 IL->getBlocksSet().erase(Ptr: BB);
1947 llvm::erase_if(C&: IL->getBlocksVector(), P: [&](BasicBlock *BB) {
1948 return BB == PH || L.contains(BB);
1949 });
1950 }
1951
1952 LI.changeLoopFor(BB: PH, L: ParentL);
1953 L.getParentLoop()->removeChildLoop(Child: &L);
1954 if (ParentL)
1955 ParentL->addChildLoop(NewChild: &L);
1956 else
1957 LI.addTopLevelLoop(New: &L);
1958 }
1959
1960 // Now we update all the blocks which are no longer within the loop.
1961 auto &Blocks = L.getBlocksVector();
1962 auto BlocksSplitI =
1963 LoopBlockSet.empty()
1964 ? Blocks.begin()
1965 : std::stable_partition(
1966 first: Blocks.begin(), last: Blocks.end(),
1967 pred: [&](BasicBlock *BB) { return LoopBlockSet.count(Ptr: BB); });
1968
1969 // Before we erase the list of unlooped blocks, build a set of them.
1970 SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1971 if (LoopBlockSet.empty())
1972 UnloopedBlocks.insert(Ptr: PH);
1973
1974 // Now erase these blocks from the loop.
1975 for (auto *BB : make_range(x: BlocksSplitI, y: Blocks.end()))
1976 L.getBlocksSet().erase(Ptr: BB);
1977 Blocks.erase(first: BlocksSplitI, last: Blocks.end());
1978
1979 // Sort the exits in ascending loop depth, we'll work backwards across these
1980 // to process them inside out.
1981 llvm::stable_sort(Range&: ExitsInLoops, C: [&](BasicBlock *LHS, BasicBlock *RHS) {
1982 return LI.getLoopDepth(BB: LHS) < LI.getLoopDepth(BB: RHS);
1983 });
1984
1985 // We'll build up a set for each exit loop.
1986 SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1987 Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1988
1989 auto RemoveUnloopedBlocksFromLoop =
1990 [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1991 for (auto *BB : UnloopedBlocks)
1992 L.getBlocksSet().erase(Ptr: BB);
1993 llvm::erase_if(C&: L.getBlocksVector(), P: [&](BasicBlock *BB) {
1994 return UnloopedBlocks.count(Ptr: BB);
1995 });
1996 };
1997
1998 SmallVector<BasicBlock *, 16> Worklist;
1999 while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
2000 assert(Worklist.empty() && "Didn't clear worklist!");
2001 assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
2002
2003 // Grab the next exit block, in decreasing loop depth order.
2004 BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
2005 Loop &ExitL = *LI.getLoopFor(BB: ExitBB);
2006 assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
2007
2008 // Erase all of the unlooped blocks from the loops between the previous
2009 // exit loop and this exit loop. This works because the ExitInLoops list is
2010 // sorted in increasing order of loop depth and thus we visit loops in
2011 // decreasing order of loop depth.
2012 for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
2013 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
2014
2015 // Walk the CFG back until we hit the cloned PH adding everything reachable
2016 // and in the unlooped set to this exit block's loop.
2017 Worklist.push_back(Elt: ExitBB);
2018 do {
2019 BasicBlock *BB = Worklist.pop_back_val();
2020 // We can stop recursing at the cloned preheader (if we get there).
2021 if (BB == PH)
2022 continue;
2023
2024 for (BasicBlock *PredBB : predecessors(BB)) {
2025 // If this pred has already been moved to our set or is part of some
2026 // (inner) loop, no update needed.
2027 if (!UnloopedBlocks.erase(Ptr: PredBB)) {
2028 assert((NewExitLoopBlocks.count(PredBB) ||
2029 ExitL.contains(LI.getLoopFor(PredBB))) &&
2030 "Predecessor not in a nested loop (or already visited)!");
2031 continue;
2032 }
2033
2034 // We just insert into the loop set here. We'll add these blocks to the
2035 // exit loop after we build up the set in a deterministic order rather
2036 // than the predecessor-influenced visit order.
2037 bool Inserted = NewExitLoopBlocks.insert(Ptr: PredBB).second;
2038 (void)Inserted;
2039 assert(Inserted && "Should only visit an unlooped block once!");
2040
2041 // And recurse through to its predecessors.
2042 Worklist.push_back(Elt: PredBB);
2043 }
2044 } while (!Worklist.empty());
2045
2046 // If blocks in this exit loop were directly part of the original loop (as
2047 // opposed to a child loop) update the map to point to this exit loop. This
2048 // just updates a map and so the fact that the order is unstable is fine.
2049 for (auto *BB : NewExitLoopBlocks)
2050 if (Loop *BBL = LI.getLoopFor(BB))
2051 if (BBL == &L || !L.contains(L: BBL))
2052 LI.changeLoopFor(BB, L: &ExitL);
2053
2054 // We will remove the remaining unlooped blocks from this loop in the next
2055 // iteration or below.
2056 NewExitLoopBlocks.clear();
2057 }
2058
2059 // Any remaining unlooped blocks are no longer part of any loop unless they
2060 // are part of some child loop.
2061 for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
2062 RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
2063 for (auto *BB : UnloopedBlocks)
2064 if (Loop *BBL = LI.getLoopFor(BB))
2065 if (BBL == &L || !L.contains(L: BBL))
2066 LI.changeLoopFor(BB, L: nullptr);
2067
2068 // Sink all the child loops whose headers are no longer in the loop set to
2069 // the parent (or to be top level loops). We reach into the loop and directly
2070 // update its subloop vector to make this batch update efficient.
2071 auto &SubLoops = L.getSubLoopsVector();
2072 auto SubLoopsSplitI =
2073 LoopBlockSet.empty()
2074 ? SubLoops.begin()
2075 : std::stable_partition(
2076 first: SubLoops.begin(), last: SubLoops.end(), pred: [&](Loop *SubL) {
2077 return LoopBlockSet.count(Ptr: SubL->getHeader());
2078 });
2079 for (auto *HoistedL : make_range(x: SubLoopsSplitI, y: SubLoops.end())) {
2080 HoistedLoops.push_back(Elt: HoistedL);
2081 HoistedL->setParentLoop(nullptr);
2082
2083 // To compute the new parent of this hoisted loop we look at where we
2084 // placed the preheader above. We can't lookup the header itself because we
2085 // retained the mapping from the header to the hoisted loop. But the
2086 // preheader and header should have the exact same new parent computed
2087 // based on the set of exit blocks from the original loop as the preheader
2088 // is a predecessor of the header and so reached in the reverse walk. And
2089 // because the loops were all in simplified form the preheader of the
2090 // hoisted loop can't be part of some *other* loop.
2091 if (auto *NewParentL = LI.getLoopFor(BB: HoistedL->getLoopPreheader()))
2092 NewParentL->addChildLoop(NewChild: HoistedL);
2093 else
2094 LI.addTopLevelLoop(New: HoistedL);
2095 }
2096 SubLoops.erase(first: SubLoopsSplitI, last: SubLoops.end());
2097
2098 // Actually delete the loop if nothing remained within it.
2099 if (Blocks.empty()) {
2100 assert(SubLoops.empty() &&
2101 "Failed to remove all subloops from the original loop!");
2102 if (Loop *ParentL = L.getParentLoop())
2103 ParentL->removeChildLoop(I: llvm::find(Range&: *ParentL, Val: &L));
2104 else
2105 LI.removeLoop(I: llvm::find(Range&: LI, Val: &L));
2106 // markLoopAsDeleted for L should be triggered by the caller (it is
2107 // typically done within postUnswitch).
2108 if (SE)
2109 SE->forgetBlockAndLoopDispositions();
2110 LI.destroy(L: &L);
2111 return false;
2112 }
2113
2114 return true;
2115}
2116
2117/// Helper to visit a dominator subtree, invoking a callable on each node.
2118///
2119/// Returning false at any point will stop walking past that node of the tree.
2120template <typename CallableT>
2121void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
2122 SmallVector<DomTreeNode *, 4> DomWorklist;
2123 DomWorklist.push_back(Elt: DT[BB]);
2124#ifndef NDEBUG
2125 SmallPtrSet<DomTreeNode *, 4> Visited;
2126 Visited.insert(DT[BB]);
2127#endif
2128 do {
2129 DomTreeNode *N = DomWorklist.pop_back_val();
2130
2131 // Visit this node.
2132 if (!Callable(N->getBlock()))
2133 continue;
2134
2135 // Accumulate the child nodes.
2136 for (DomTreeNode *ChildN : *N) {
2137 assert(Visited.insert(ChildN).second &&
2138 "Cannot visit a node twice when walking a tree!");
2139 DomWorklist.push_back(Elt: ChildN);
2140 }
2141 } while (!DomWorklist.empty());
2142}
2143
2144void postUnswitch(Loop &L, LPMUpdater &U, StringRef LoopName,
2145 bool CurrentLoopValid, bool PartiallyInvariant,
2146 bool InjectedCondition, ArrayRef<Loop *> NewLoops) {
2147 // If we did a non-trivial unswitch, we have added new (cloned) loops.
2148 if (!NewLoops.empty())
2149 U.addSiblingLoops(NewSibLoops: NewLoops);
2150
2151 // If the current loop remains valid, we should revisit it to catch any
2152 // other unswitch opportunities. Otherwise, we need to mark it as deleted.
2153 if (CurrentLoopValid) {
2154 if (PartiallyInvariant) {
2155 // Mark the new loop as partially unswitched, to avoid unswitching on
2156 // the same condition again.
2157 auto &Context = L.getHeader()->getContext();
2158 MDNode *DisableUnswitchMD = MDNode::get(
2159 Context,
2160 MDs: MDString::get(Context, Str: "llvm.loop.unswitch.partial.disable"));
2161 MDNode *NewLoopID = makePostTransformationMetadata(
2162 Context, OrigLoopID: L.getLoopID(), RemovePrefixes: {"llvm.loop.unswitch.partial"},
2163 AddAttrs: {DisableUnswitchMD});
2164 L.setLoopID(NewLoopID);
2165 } else if (InjectedCondition) {
2166 // Do the same for injection of invariant conditions.
2167 auto &Context = L.getHeader()->getContext();
2168 MDNode *DisableUnswitchMD = MDNode::get(
2169 Context,
2170 MDs: MDString::get(Context, Str: "llvm.loop.unswitch.injection.disable"));
2171 MDNode *NewLoopID = makePostTransformationMetadata(
2172 Context, OrigLoopID: L.getLoopID(), RemovePrefixes: {"llvm.loop.unswitch.injection"},
2173 AddAttrs: {DisableUnswitchMD});
2174 L.setLoopID(NewLoopID);
2175 } else
2176 U.revisitCurrentLoop();
2177 } else
2178 U.markLoopAsDeleted(L, Name: LoopName);
2179}
2180
2181static void unswitchNontrivialInvariants(
2182 Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
2183 IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI,
2184 AssumptionCache &AC, ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
2185 LPMUpdater &LoopUpdater, bool InsertFreeze, bool InjectedCondition) {
2186 auto *ParentBB = TI.getParent();
2187 BranchInst *BI = dyn_cast<BranchInst>(Val: &TI);
2188 SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(Val: &TI);
2189
2190 // Save the current loop name in a variable so that we can report it even
2191 // after it has been deleted.
2192 std::string LoopName(L.getName());
2193
2194 // We can only unswitch switches, conditional branches with an invariant
2195 // condition, or combining invariant conditions with an instruction or
2196 // partially invariant instructions.
2197 assert((SI || (BI && BI->isConditional())) &&
2198 "Can only unswitch switches and conditional branch!");
2199 bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty();
2200 bool FullUnswitch =
2201 SI || (skipTrivialSelect(Cond: BI->getCondition()) == Invariants[0] &&
2202 !PartiallyInvariant);
2203 if (FullUnswitch)
2204 assert(Invariants.size() == 1 &&
2205 "Cannot have other invariants with full unswitching!");
2206 else
2207 assert(isa<Instruction>(skipTrivialSelect(BI->getCondition())) &&
2208 "Partial unswitching requires an instruction as the condition!");
2209
2210 if (MSSAU && VerifyMemorySSA)
2211 MSSAU->getMemorySSA()->verifyMemorySSA();
2212
2213 // Constant and BBs tracking the cloned and continuing successor. When we are
2214 // unswitching the entire condition, this can just be trivially chosen to
2215 // unswitch towards `true`. However, when we are unswitching a set of
2216 // invariants combined with `and` or `or` or partially invariant instructions,
2217 // the combining operation determines the best direction to unswitch: we want
2218 // to unswitch the direction that will collapse the branch.
2219 bool Direction = true;
2220 int ClonedSucc = 0;
2221 if (!FullUnswitch) {
2222 Value *Cond = skipTrivialSelect(Cond: BI->getCondition());
2223 (void)Cond;
2224 assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) ||
2225 PartiallyInvariant) &&
2226 "Only `or`, `and`, an `select`, partially invariant instructions "
2227 "can combine invariants being unswitched.");
2228 if (!match(V: Cond, P: m_LogicalOr())) {
2229 if (match(V: Cond, P: m_LogicalAnd()) ||
2230 (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) {
2231 Direction = false;
2232 ClonedSucc = 1;
2233 }
2234 }
2235 }
2236
2237 BasicBlock *RetainedSuccBB =
2238 BI ? BI->getSuccessor(i: 1 - ClonedSucc) : SI->getDefaultDest();
2239 SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
2240 if (BI)
2241 UnswitchedSuccBBs.insert(X: BI->getSuccessor(i: ClonedSucc));
2242 else
2243 for (auto Case : SI->cases())
2244 if (Case.getCaseSuccessor() != RetainedSuccBB)
2245 UnswitchedSuccBBs.insert(X: Case.getCaseSuccessor());
2246
2247 assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
2248 "Should not unswitch the same successor we are retaining!");
2249
2250 // The branch should be in this exact loop. Any inner loop's invariant branch
2251 // should be handled by unswitching that inner loop. The caller of this
2252 // routine should filter out any candidates that remain (but were skipped for
2253 // whatever reason).
2254 assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
2255
2256 // Compute the parent loop now before we start hacking on things.
2257 Loop *ParentL = L.getParentLoop();
2258 // Get blocks in RPO order for MSSA update, before changing the CFG.
2259 LoopBlocksRPO LBRPO(&L);
2260 if (MSSAU)
2261 LBRPO.perform(LI: &LI);
2262
2263 // Compute the outer-most loop containing one of our exit blocks. This is the
2264 // furthest up our loopnest which can be mutated, which we will use below to
2265 // update things.
2266 Loop *OuterExitL = &L;
2267 SmallVector<BasicBlock *, 4> ExitBlocks;
2268 L.getUniqueExitBlocks(ExitBlocks);
2269 for (auto *ExitBB : ExitBlocks) {
2270 // ExitBB can be an exit block for several levels in the loop nest. Make
2271 // sure we find the top most.
2272 Loop *NewOuterExitL = getTopMostExitingLoop(ExitBB, LI);
2273 if (!NewOuterExitL) {
2274 // We exited the entire nest with this block, so we're done.
2275 OuterExitL = nullptr;
2276 break;
2277 }
2278 if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(L: OuterExitL))
2279 OuterExitL = NewOuterExitL;
2280 }
2281
2282 // At this point, we're definitely going to unswitch something so invalidate
2283 // any cached information in ScalarEvolution for the outer most loop
2284 // containing an exit block and all nested loops.
2285 if (SE) {
2286 if (OuterExitL)
2287 SE->forgetLoop(L: OuterExitL);
2288 else
2289 SE->forgetTopmostLoop(L: &L);
2290 SE->forgetBlockAndLoopDispositions();
2291 }
2292
2293 // If the edge from this terminator to a successor dominates that successor,
2294 // store a map from each block in its dominator subtree to it. This lets us
2295 // tell when cloning for a particular successor if a block is dominated by
2296 // some *other* successor with a single data structure. We use this to
2297 // significantly reduce cloning.
2298 SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
2299 for (auto *SuccBB : llvm::concat<BasicBlock *const>(Ranges: ArrayRef(RetainedSuccBB),
2300 Ranges&: UnswitchedSuccBBs))
2301 if (SuccBB->getUniquePredecessor() ||
2302 llvm::all_of(Range: predecessors(BB: SuccBB), P: [&](BasicBlock *PredBB) {
2303 return PredBB == ParentBB || DT.dominates(A: SuccBB, B: PredBB);
2304 }))
2305 visitDomSubTree(DT, BB: SuccBB, Callable: [&](BasicBlock *BB) {
2306 DominatingSucc[BB] = SuccBB;
2307 return true;
2308 });
2309
2310 // Split the preheader, so that we know that there is a safe place to insert
2311 // the conditional branch. We will change the preheader to have a conditional
2312 // branch on LoopCond. The original preheader will become the split point
2313 // between the unswitched versions, and we will have a new preheader for the
2314 // original loop.
2315 BasicBlock *SplitBB = L.getLoopPreheader();
2316 BasicBlock *LoopPH = SplitEdge(From: SplitBB, To: L.getHeader(), DT: &DT, LI: &LI, MSSAU);
2317
2318 // Keep track of the dominator tree updates needed.
2319 SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2320
2321 // Clone the loop for each unswitched successor.
2322 SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
2323 VMaps.reserve(N: UnswitchedSuccBBs.size());
2324 SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
2325 for (auto *SuccBB : UnswitchedSuccBBs) {
2326 VMaps.emplace_back(Args: new ValueToValueMapTy());
2327 ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2328 L, LoopPH, SplitBB, ExitBlocks, ParentBB, UnswitchedSuccBB: SuccBB, ContinueSuccBB: RetainedSuccBB,
2329 DominatingSucc, VMap&: *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU, SE);
2330 }
2331
2332 // Drop metadata if we may break its semantics by moving this instr into the
2333 // split block.
2334 if (TI.getMetadata(KindID: LLVMContext::MD_make_implicit)) {
2335 if (DropNonTrivialImplicitNullChecks)
2336 // Do not spend time trying to understand if we can keep it, just drop it
2337 // to save compile time.
2338 TI.setMetadata(KindID: LLVMContext::MD_make_implicit, Node: nullptr);
2339 else {
2340 // It is only legal to preserve make.implicit metadata if we are
2341 // guaranteed no reach implicit null check after following this branch.
2342 ICFLoopSafetyInfo SafetyInfo;
2343 SafetyInfo.computeLoopSafetyInfo(CurLoop: &L);
2344 if (!SafetyInfo.isGuaranteedToExecute(Inst: TI, DT: &DT, CurLoop: &L))
2345 TI.setMetadata(KindID: LLVMContext::MD_make_implicit, Node: nullptr);
2346 }
2347 }
2348
2349 // The stitching of the branched code back together depends on whether we're
2350 // doing full unswitching or not with the exception that we always want to
2351 // nuke the initial terminator placed in the split block.
2352 SplitBB->getTerminator()->eraseFromParent();
2353 if (FullUnswitch) {
2354 // Keep a clone of the terminator for MSSA updates.
2355 Instruction *NewTI = TI.clone();
2356 NewTI->insertInto(ParentBB, It: ParentBB->end());
2357
2358 // Splice the terminator from the original loop and rewrite its
2359 // successors.
2360 TI.moveBefore(BB&: *SplitBB, I: SplitBB->end());
2361 TI.dropLocation();
2362
2363 // First wire up the moved terminator to the preheaders.
2364 if (BI) {
2365 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2366 BI->setSuccessor(idx: ClonedSucc, NewSucc: ClonedPH);
2367 BI->setSuccessor(idx: 1 - ClonedSucc, NewSucc: LoopPH);
2368 Value *Cond = skipTrivialSelect(Cond: BI->getCondition());
2369 if (InsertFreeze) {
2370 // We don't give any debug location to the new freeze, because the
2371 // BI (`dyn_cast<BranchInst>(TI)`) is an in-loop instruction hoisted
2372 // out of the loop.
2373 Cond = new FreezeInst(Cond, Cond->getName() + ".fr", BI->getIterator());
2374 cast<Instruction>(Val: Cond)->setDebugLoc(DebugLoc::getDropped());
2375 }
2376 BI->setCondition(Cond);
2377 DTUpdates.push_back(Elt: {DominatorTree::Insert, SplitBB, ClonedPH});
2378 } else {
2379 assert(SI && "Must either be a branch or switch!");
2380
2381 // Walk the cases and directly update their successors.
2382 assert(SI->getDefaultDest() == RetainedSuccBB &&
2383 "Not retaining default successor!");
2384 SI->setDefaultDest(LoopPH);
2385 for (const auto &Case : SI->cases())
2386 if (Case.getCaseSuccessor() == RetainedSuccBB)
2387 Case.setSuccessor(LoopPH);
2388 else
2389 Case.setSuccessor(ClonedPHs.find(Val: Case.getCaseSuccessor())->second);
2390
2391 if (InsertFreeze)
2392 SI->setCondition(new FreezeInst(SI->getCondition(),
2393 SI->getCondition()->getName() + ".fr",
2394 SI->getIterator()));
2395
2396 // We need to use the set to populate domtree updates as even when there
2397 // are multiple cases pointing at the same successor we only want to
2398 // remove and insert one edge in the domtree.
2399 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2400 DTUpdates.push_back(
2401 Elt: {DominatorTree::Insert, SplitBB, ClonedPHs.find(Val: SuccBB)->second});
2402 }
2403
2404 if (MSSAU) {
2405 DT.applyUpdates(Updates: DTUpdates);
2406 DTUpdates.clear();
2407
2408 // Remove all but one edge to the retained block and all unswitched
2409 // blocks. This is to avoid having duplicate entries in the cloned Phis,
2410 // when we know we only keep a single edge for each case.
2411 MSSAU->removeDuplicatePhiEdgesBetween(From: ParentBB, To: RetainedSuccBB);
2412 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2413 MSSAU->removeDuplicatePhiEdgesBetween(From: ParentBB, To: SuccBB);
2414
2415 for (auto &VMap : VMaps)
2416 MSSAU->updateForClonedLoop(LoopBlocks: LBRPO, ExitBlocks, VM: *VMap,
2417 /*IgnoreIncomingWithNoClones=*/true);
2418 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2419
2420 // Remove all edges to unswitched blocks.
2421 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2422 MSSAU->removeEdge(From: ParentBB, To: SuccBB);
2423 }
2424
2425 // Now unhook the successor relationship as we'll be replacing
2426 // the terminator with a direct branch. This is much simpler for branches
2427 // than switches so we handle those first.
2428 if (BI) {
2429 // Remove the parent as a predecessor of the unswitched successor.
2430 assert(UnswitchedSuccBBs.size() == 1 &&
2431 "Only one possible unswitched block for a branch!");
2432 BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2433 UnswitchedSuccBB->removePredecessor(Pred: ParentBB,
2434 /*KeepOneInputPHIs*/ true);
2435 DTUpdates.push_back(Elt: {DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2436 } else {
2437 // Note that we actually want to remove the parent block as a predecessor
2438 // of *every* case successor. The case successor is either unswitched,
2439 // completely eliminating an edge from the parent to that successor, or it
2440 // is a duplicate edge to the retained successor as the retained successor
2441 // is always the default successor and as we'll replace this with a direct
2442 // branch we no longer need the duplicate entries in the PHI nodes.
2443 SwitchInst *NewSI = cast<SwitchInst>(Val: NewTI);
2444 assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2445 "Not retaining default successor!");
2446 for (const auto &Case : NewSI->cases())
2447 Case.getCaseSuccessor()->removePredecessor(
2448 Pred: ParentBB,
2449 /*KeepOneInputPHIs*/ true);
2450
2451 // We need to use the set to populate domtree updates as even when there
2452 // are multiple cases pointing at the same successor we only want to
2453 // remove and insert one edge in the domtree.
2454 for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2455 DTUpdates.push_back(Elt: {DominatorTree::Delete, ParentBB, SuccBB});
2456 }
2457
2458 // Create a new unconditional branch to the continuing block (as opposed to
2459 // the one cloned).
2460 Instruction *NewBI = BranchInst::Create(IfTrue: RetainedSuccBB, InsertBefore: ParentBB);
2461 NewBI->setDebugLoc(NewTI->getDebugLoc());
2462
2463 // After MSSAU update, remove the cloned terminator instruction NewTI.
2464 NewTI->eraseFromParent();
2465 } else {
2466 assert(BI && "Only branches have partial unswitching.");
2467 assert(UnswitchedSuccBBs.size() == 1 &&
2468 "Only one possible unswitched block for a branch!");
2469 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2470 // When doing a partial unswitch, we have to do a bit more work to build up
2471 // the branch in the split block.
2472 if (PartiallyInvariant)
2473 buildPartialInvariantUnswitchConditionalBranch(
2474 BB&: *SplitBB, ToDuplicate: Invariants, Direction, UnswitchedSucc&: *ClonedPH, NormalSucc&: *LoopPH, L, MSSAU);
2475 else {
2476 buildPartialUnswitchConditionalBranch(
2477 BB&: *SplitBB, Invariants, Direction, UnswitchedSucc&: *ClonedPH, NormalSucc&: *LoopPH,
2478 InsertFreeze: FreezeLoopUnswitchCond, I: BI, AC: &AC, DT);
2479 }
2480 DTUpdates.push_back(Elt: {DominatorTree::Insert, SplitBB, ClonedPH});
2481
2482 if (MSSAU) {
2483 DT.applyUpdates(Updates: DTUpdates);
2484 DTUpdates.clear();
2485
2486 // Perform MSSA cloning updates.
2487 for (auto &VMap : VMaps)
2488 MSSAU->updateForClonedLoop(LoopBlocks: LBRPO, ExitBlocks, VM: *VMap,
2489 /*IgnoreIncomingWithNoClones=*/true);
2490 MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2491 }
2492 }
2493
2494 // Apply the updates accumulated above to get an up-to-date dominator tree.
2495 DT.applyUpdates(Updates: DTUpdates);
2496
2497 // Now that we have an accurate dominator tree, first delete the dead cloned
2498 // blocks so that we can accurately build any cloned loops. It is important to
2499 // not delete the blocks from the original loop yet because we still want to
2500 // reference the original loop to understand the cloned loop's structure.
2501 deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2502
2503 // Build the cloned loop structure itself. This may be substantially
2504 // different from the original structure due to the simplified CFG. This also
2505 // handles inserting all the cloned blocks into the correct loops.
2506 SmallVector<Loop *, 4> NonChildClonedLoops;
2507 for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2508 buildClonedLoops(OrigL&: L, ExitBlocks, VMap: *VMap, LI, NonChildClonedLoops);
2509
2510 // Now that our cloned loops have been built, we can update the original loop.
2511 // First we delete the dead blocks from it and then we rebuild the loop
2512 // structure taking these deletions into account.
2513 deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, SE, LoopUpdater);
2514
2515 if (MSSAU && VerifyMemorySSA)
2516 MSSAU->getMemorySSA()->verifyMemorySSA();
2517
2518 SmallVector<Loop *, 4> HoistedLoops;
2519 bool IsStillLoop =
2520 rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops, SE);
2521
2522 if (MSSAU && VerifyMemorySSA)
2523 MSSAU->getMemorySSA()->verifyMemorySSA();
2524
2525 // This transformation has a high risk of corrupting the dominator tree, and
2526 // the below steps to rebuild loop structures will result in hard to debug
2527 // errors in that case so verify that the dominator tree is sane first.
2528 // FIXME: Remove this when the bugs stop showing up and rely on existing
2529 // verification steps.
2530 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2531
2532 if (BI && !PartiallyInvariant) {
2533 // If we unswitched a branch which collapses the condition to a known
2534 // constant we want to replace all the uses of the invariants within both
2535 // the original and cloned blocks. We do this here so that we can use the
2536 // now updated dominator tree to identify which side the users are on.
2537 assert(UnswitchedSuccBBs.size() == 1 &&
2538 "Only one possible unswitched block for a branch!");
2539 BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2540
2541 // When considering multiple partially-unswitched invariants
2542 // we cant just go replace them with constants in both branches.
2543 //
2544 // For 'AND' we infer that true branch ("continue") means true
2545 // for each invariant operand.
2546 // For 'OR' we can infer that false branch ("continue") means false
2547 // for each invariant operand.
2548 // So it happens that for multiple-partial case we dont replace
2549 // in the unswitched branch.
2550 bool ReplaceUnswitched =
2551 FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant;
2552
2553 ConstantInt *UnswitchedReplacement =
2554 Direction ? ConstantInt::getTrue(Context&: BI->getContext())
2555 : ConstantInt::getFalse(Context&: BI->getContext());
2556 ConstantInt *ContinueReplacement =
2557 Direction ? ConstantInt::getFalse(Context&: BI->getContext())
2558 : ConstantInt::getTrue(Context&: BI->getContext());
2559 for (Value *Invariant : Invariants) {
2560 assert(!isa<Constant>(Invariant) &&
2561 "Should not be replacing constant values!");
2562 // Use make_early_inc_range here as set invalidates the iterator.
2563 for (Use &U : llvm::make_early_inc_range(Range: Invariant->uses())) {
2564 Instruction *UserI = dyn_cast<Instruction>(Val: U.getUser());
2565 if (!UserI)
2566 continue;
2567
2568 // Replace it with the 'continue' side if in the main loop body, and the
2569 // unswitched if in the cloned blocks.
2570 if (DT.dominates(A: LoopPH, B: UserI->getParent()))
2571 U.set(ContinueReplacement);
2572 else if (ReplaceUnswitched &&
2573 DT.dominates(A: ClonedPH, B: UserI->getParent()))
2574 U.set(UnswitchedReplacement);
2575 }
2576 }
2577 }
2578
2579 // We can change which blocks are exit blocks of all the cloned sibling
2580 // loops, the current loop, and any parent loops which shared exit blocks
2581 // with the current loop. As a consequence, we need to re-form LCSSA for
2582 // them. But we shouldn't need to re-form LCSSA for any child loops.
2583 // FIXME: This could be made more efficient by tracking which exit blocks are
2584 // new, and focusing on them, but that isn't likely to be necessary.
2585 //
2586 // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2587 // loop nest and update every loop that could have had its exits changed. We
2588 // also need to cover any intervening loops. We add all of these loops to
2589 // a list and sort them by loop depth to achieve this without updating
2590 // unnecessary loops.
2591 auto UpdateLoop = [&](Loop &UpdateL) {
2592#ifndef NDEBUG
2593 UpdateL.verifyLoop();
2594 for (Loop *ChildL : UpdateL) {
2595 ChildL->verifyLoop();
2596 assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2597 "Perturbed a child loop's LCSSA form!");
2598 }
2599#endif
2600 // First build LCSSA for this loop so that we can preserve it when
2601 // forming dedicated exits. We don't want to perturb some other loop's
2602 // LCSSA while doing that CFG edit.
2603 formLCSSA(L&: UpdateL, DT, LI: &LI, SE);
2604
2605 // For loops reached by this loop's original exit blocks we may
2606 // introduced new, non-dedicated exits. At least try to re-form dedicated
2607 // exits for these loops. This may fail if they couldn't have dedicated
2608 // exits to start with.
2609 formDedicatedExitBlocks(L: &UpdateL, DT: &DT, LI: &LI, MSSAU, /*PreserveLCSSA*/ true);
2610 };
2611
2612 // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2613 // and we can do it in any order as they don't nest relative to each other.
2614 //
2615 // Also check if any of the loops we have updated have become top-level loops
2616 // as that will necessitate widening the outer loop scope.
2617 for (Loop *UpdatedL :
2618 llvm::concat<Loop *>(Ranges&: NonChildClonedLoops, Ranges&: HoistedLoops)) {
2619 UpdateLoop(*UpdatedL);
2620 if (UpdatedL->isOutermost())
2621 OuterExitL = nullptr;
2622 }
2623 if (IsStillLoop) {
2624 UpdateLoop(L);
2625 if (L.isOutermost())
2626 OuterExitL = nullptr;
2627 }
2628
2629 // If the original loop had exit blocks, walk up through the outer most loop
2630 // of those exit blocks to update LCSSA and form updated dedicated exits.
2631 if (OuterExitL != &L)
2632 for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2633 OuterL = OuterL->getParentLoop())
2634 UpdateLoop(*OuterL);
2635
2636#ifndef NDEBUG
2637 // Verify the entire loop structure to catch any incorrect updates before we
2638 // progress in the pass pipeline.
2639 LI.verify(DT);
2640#endif
2641
2642 // Now that we've unswitched something, make callbacks to report the changes.
2643 // For that we need to merge together the updated loops and the cloned loops
2644 // and check whether the original loop survived.
2645 SmallVector<Loop *, 4> SibLoops;
2646 for (Loop *UpdatedL : llvm::concat<Loop *>(Ranges&: NonChildClonedLoops, Ranges&: HoistedLoops))
2647 if (UpdatedL->getParentLoop() == ParentL)
2648 SibLoops.push_back(Elt: UpdatedL);
2649 postUnswitch(L, U&: LoopUpdater, LoopName, CurrentLoopValid: IsStillLoop, PartiallyInvariant,
2650 InjectedCondition, NewLoops: SibLoops);
2651
2652 if (MSSAU && VerifyMemorySSA)
2653 MSSAU->getMemorySSA()->verifyMemorySSA();
2654
2655 if (BI)
2656 ++NumBranches;
2657 else
2658 ++NumSwitches;
2659}
2660
2661/// Recursively compute the cost of a dominator subtree based on the per-block
2662/// cost map provided.
2663///
2664/// The recursive computation is memozied into the provided DT-indexed cost map
2665/// to allow querying it for most nodes in the domtree without it becoming
2666/// quadratic.
2667static InstructionCost computeDomSubtreeCost(
2668 DomTreeNode &N,
2669 const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap,
2670 SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) {
2671 // Don't accumulate cost (or recurse through) blocks not in our block cost
2672 // map and thus not part of the duplication cost being considered.
2673 auto BBCostIt = BBCostMap.find(Val: N.getBlock());
2674 if (BBCostIt == BBCostMap.end())
2675 return 0;
2676
2677 // Lookup this node to see if we already computed its cost.
2678 auto DTCostIt = DTCostMap.find(Val: &N);
2679 if (DTCostIt != DTCostMap.end())
2680 return DTCostIt->second;
2681
2682 // If not, we have to compute it. We can't use insert above and update
2683 // because computing the cost may insert more things into the map.
2684 InstructionCost Cost = std::accumulate(
2685 first: N.begin(), last: N.end(), init: BBCostIt->second,
2686 binary_op: [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost {
2687 return Sum + computeDomSubtreeCost(N&: *ChildN, BBCostMap, DTCostMap);
2688 });
2689 bool Inserted = DTCostMap.insert(KV: {&N, Cost}).second;
2690 (void)Inserted;
2691 assert(Inserted && "Should not insert a node while visiting children!");
2692 return Cost;
2693}
2694
2695/// Turns a select instruction into implicit control flow branch,
2696/// making the following replacement:
2697///
2698/// head:
2699/// --code before select--
2700/// select %cond, %trueval, %falseval
2701/// --code after select--
2702///
2703/// into
2704///
2705/// head:
2706/// --code before select--
2707/// br i1 %cond, label %then, label %tail
2708///
2709/// then:
2710/// br %tail
2711///
2712/// tail:
2713/// phi [ %trueval, %then ], [ %falseval, %head]
2714/// unreachable
2715///
2716/// It also makes all relevant DT and LI updates, so that all structures are in
2717/// valid state after this transform.
2718static BranchInst *turnSelectIntoBranch(SelectInst *SI, DominatorTree &DT,
2719 LoopInfo &LI, MemorySSAUpdater *MSSAU,
2720 AssumptionCache *AC) {
2721 LLVM_DEBUG(dbgs() << "Turning " << *SI << " into a branch.\n");
2722 BasicBlock *HeadBB = SI->getParent();
2723
2724 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2725 SplitBlockAndInsertIfThen(Cond: SI->getCondition(), SplitBefore: SI, Unreachable: false,
2726 BranchWeights: SI->getMetadata(KindID: LLVMContext::MD_prof), DTU: &DTU, LI: &LI);
2727 auto *CondBr = cast<BranchInst>(Val: HeadBB->getTerminator());
2728 BasicBlock *ThenBB = CondBr->getSuccessor(i: 0),
2729 *TailBB = CondBr->getSuccessor(i: 1);
2730 if (MSSAU)
2731 MSSAU->moveAllAfterSpliceBlocks(From: HeadBB, To: TailBB, Start: SI);
2732
2733 PHINode *Phi =
2734 PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "unswitched.select", InsertBefore: SI->getIterator());
2735 Phi->addIncoming(V: SI->getTrueValue(), BB: ThenBB);
2736 Phi->addIncoming(V: SI->getFalseValue(), BB: HeadBB);
2737 Phi->setDebugLoc(SI->getDebugLoc());
2738 SI->replaceAllUsesWith(V: Phi);
2739 SI->eraseFromParent();
2740
2741 if (MSSAU && VerifyMemorySSA)
2742 MSSAU->getMemorySSA()->verifyMemorySSA();
2743
2744 ++NumSelects;
2745 return CondBr;
2746}
2747
2748/// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2749/// making the following replacement:
2750///
2751/// --code before guard--
2752/// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2753/// --code after guard--
2754///
2755/// into
2756///
2757/// --code before guard--
2758/// br i1 %cond, label %guarded, label %deopt
2759///
2760/// guarded:
2761/// --code after guard--
2762///
2763/// deopt:
2764/// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2765/// unreachable
2766///
2767/// It also makes all relevant DT and LI updates, so that all structures are in
2768/// valid state after this transform.
2769static BranchInst *turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2770 DominatorTree &DT, LoopInfo &LI,
2771 MemorySSAUpdater *MSSAU) {
2772 LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2773 BasicBlock *CheckBB = GI->getParent();
2774
2775 if (MSSAU && VerifyMemorySSA)
2776 MSSAU->getMemorySSA()->verifyMemorySSA();
2777
2778 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2779 Instruction *DeoptBlockTerm =
2780 SplitBlockAndInsertIfThen(Cond: GI->getArgOperand(i: 0), SplitBefore: GI, Unreachable: true,
2781 BranchWeights: GI->getMetadata(KindID: LLVMContext::MD_prof), DTU: &DTU, LI: &LI);
2782 BranchInst *CheckBI = cast<BranchInst>(Val: CheckBB->getTerminator());
2783 // SplitBlockAndInsertIfThen inserts control flow that branches to
2784 // DeoptBlockTerm if the condition is true. We want the opposite.
2785 CheckBI->swapSuccessors();
2786
2787 BasicBlock *GuardedBlock = CheckBI->getSuccessor(i: 0);
2788 GuardedBlock->setName("guarded");
2789 CheckBI->getSuccessor(i: 1)->setName("deopt");
2790 BasicBlock *DeoptBlock = CheckBI->getSuccessor(i: 1);
2791
2792 if (MSSAU)
2793 MSSAU->moveAllAfterSpliceBlocks(From: CheckBB, To: GuardedBlock, Start: GI);
2794
2795 GI->moveBefore(InsertPos: DeoptBlockTerm->getIterator());
2796 GI->setArgOperand(i: 0, v: ConstantInt::getFalse(Context&: GI->getContext()));
2797
2798 if (MSSAU) {
2799 MemoryDef *MD = cast<MemoryDef>(Val: MSSAU->getMemorySSA()->getMemoryAccess(I: GI));
2800 MSSAU->moveToPlace(What: MD, BB: DeoptBlock, Where: MemorySSA::BeforeTerminator);
2801 if (VerifyMemorySSA)
2802 MSSAU->getMemorySSA()->verifyMemorySSA();
2803 }
2804
2805 if (VerifyLoopInfo)
2806 LI.verify(DomTree: DT);
2807 ++NumGuards;
2808 return CheckBI;
2809}
2810
2811/// Cost multiplier is a way to limit potentially exponential behavior
2812/// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2813/// candidates available. Also accounting for the number of "sibling" loops with
2814/// the idea to account for previous unswitches that already happened on this
2815/// cluster of loops. There was an attempt to keep this formula simple,
2816/// just enough to limit the worst case behavior. Even if it is not that simple
2817/// now it is still not an attempt to provide a detailed heuristic size
2818/// prediction.
2819///
2820/// TODO: Make a proper accounting of "explosion" effect for all kinds of
2821/// unswitch candidates, making adequate predictions instead of wild guesses.
2822/// That requires knowing not just the number of "remaining" candidates but
2823/// also costs of unswitching for each of these candidates.
2824static int CalculateUnswitchCostMultiplier(
2825 const Instruction &TI, const Loop &L, const LoopInfo &LI,
2826 const DominatorTree &DT,
2827 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates) {
2828
2829 // Guards and other exiting conditions do not contribute to exponential
2830 // explosion as soon as they dominate the latch (otherwise there might be
2831 // another path to the latch remaining that does not allow to eliminate the
2832 // loop copy on unswitch).
2833 const BasicBlock *Latch = L.getLoopLatch();
2834 const BasicBlock *CondBlock = TI.getParent();
2835 if (DT.dominates(A: CondBlock, B: Latch) &&
2836 (isGuard(U: &TI) ||
2837 (TI.isTerminator() &&
2838 llvm::count_if(Range: successors(I: &TI), P: [&L](const BasicBlock *SuccBB) {
2839 return L.contains(BB: SuccBB);
2840 }) <= 1))) {
2841 NumCostMultiplierSkipped++;
2842 return 1;
2843 }
2844
2845 auto *ParentL = L.getParentLoop();
2846 int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2847 : std::distance(first: LI.begin(), last: LI.end()));
2848 // Count amount of clones that all the candidates might cause during
2849 // unswitching. Branch/guard/select counts as 1, switch counts as log2 of its
2850 // cases.
2851 int UnswitchedClones = 0;
2852 for (const auto &Candidate : UnswitchCandidates) {
2853 const Instruction *CI = Candidate.TI;
2854 const BasicBlock *CondBlock = CI->getParent();
2855 bool SkipExitingSuccessors = DT.dominates(A: CondBlock, B: Latch);
2856 if (isa<SelectInst>(Val: CI)) {
2857 UnswitchedClones++;
2858 continue;
2859 }
2860 if (isGuard(U: CI)) {
2861 if (!SkipExitingSuccessors)
2862 UnswitchedClones++;
2863 continue;
2864 }
2865 int NonExitingSuccessors =
2866 llvm::count_if(Range: successors(BB: CondBlock),
2867 P: [SkipExitingSuccessors, &L](const BasicBlock *SuccBB) {
2868 return !SkipExitingSuccessors || L.contains(BB: SuccBB);
2869 });
2870 UnswitchedClones += Log2_32(Value: NonExitingSuccessors);
2871 }
2872
2873 // Ignore up to the "unscaled candidates" number of unswitch candidates
2874 // when calculating the power-of-two scaling of the cost. The main idea
2875 // with this control is to allow a small number of unswitches to happen
2876 // and rely more on siblings multiplier (see below) when the number
2877 // of candidates is small.
2878 unsigned ClonesPower =
2879 std::max(a: UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, b: 0);
2880
2881 // Allowing top-level loops to spread a bit more than nested ones.
2882 int SiblingsMultiplier =
2883 std::max(a: (ParentL ? SiblingsCount
2884 : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2885 b: 1);
2886 // Compute the cost multiplier in a way that won't overflow by saturating
2887 // at an upper bound.
2888 int CostMultiplier;
2889 if (ClonesPower > Log2_32(Value: UnswitchThreshold) ||
2890 SiblingsMultiplier > UnswitchThreshold)
2891 CostMultiplier = UnswitchThreshold;
2892 else
2893 CostMultiplier = std::min(a: SiblingsMultiplier * (1 << ClonesPower),
2894 b: (int)UnswitchThreshold);
2895
2896 LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplier
2897 << " (siblings " << SiblingsMultiplier << " * clones "
2898 << (1 << ClonesPower) << ")"
2899 << " for unswitch candidate: " << TI << "\n");
2900 return CostMultiplier;
2901}
2902
2903static bool collectUnswitchCandidates(
2904 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
2905 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch,
2906 const Loop &L, const LoopInfo &LI, AAResults &AA,
2907 const MemorySSAUpdater *MSSAU) {
2908 assert(UnswitchCandidates.empty() && "Should be!");
2909
2910 auto AddUnswitchCandidatesForInst = [&](Instruction *I, Value *Cond) {
2911 Cond = skipTrivialSelect(Cond);
2912 if (isa<Constant>(Val: Cond))
2913 return;
2914 if (L.isLoopInvariant(V: Cond)) {
2915 UnswitchCandidates.push_back(Elt: {I, {Cond}});
2916 return;
2917 }
2918 if (match(V: Cond, P: m_CombineOr(L: m_LogicalAnd(), R: m_LogicalOr()))) {
2919 TinyPtrVector<Value *> Invariants =
2920 collectHomogenousInstGraphLoopInvariants(
2921 L, Root&: *static_cast<Instruction *>(Cond), LI);
2922 if (!Invariants.empty())
2923 UnswitchCandidates.push_back(Elt: {I, std::move(Invariants)});
2924 }
2925 };
2926
2927 // Whether or not we should also collect guards in the loop.
2928 bool CollectGuards = false;
2929 if (UnswitchGuards) {
2930 auto *GuardDecl = Intrinsic::getDeclarationIfExists(
2931 M: L.getHeader()->getParent()->getParent(), id: Intrinsic::experimental_guard);
2932 if (GuardDecl && !GuardDecl->use_empty())
2933 CollectGuards = true;
2934 }
2935
2936 for (auto *BB : L.blocks()) {
2937 if (LI.getLoopFor(BB) != &L)
2938 continue;
2939
2940 for (auto &I : *BB) {
2941 if (auto *SI = dyn_cast<SelectInst>(Val: &I)) {
2942 auto *Cond = SI->getCondition();
2943 // Do not unswitch vector selects and logical and/or selects
2944 if (Cond->getType()->isIntegerTy(Bitwidth: 1) && !SI->getType()->isIntegerTy(Bitwidth: 1))
2945 AddUnswitchCandidatesForInst(SI, Cond);
2946 } else if (CollectGuards && isGuard(U: &I)) {
2947 auto *Cond =
2948 skipTrivialSelect(Cond: cast<IntrinsicInst>(Val: &I)->getArgOperand(i: 0));
2949 // TODO: Support AND, OR conditions and partial unswitching.
2950 if (!isa<Constant>(Val: Cond) && L.isLoopInvariant(V: Cond))
2951 UnswitchCandidates.push_back(Elt: {&I, {Cond}});
2952 }
2953 }
2954
2955 if (auto *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator())) {
2956 // We can only consider fully loop-invariant switch conditions as we need
2957 // to completely eliminate the switch after unswitching.
2958 if (!isa<Constant>(Val: SI->getCondition()) &&
2959 L.isLoopInvariant(V: SI->getCondition()) && !BB->getUniqueSuccessor())
2960 UnswitchCandidates.push_back(Elt: {SI, {SI->getCondition()}});
2961 continue;
2962 }
2963
2964 auto *BI = dyn_cast<BranchInst>(Val: BB->getTerminator());
2965 if (!BI || !BI->isConditional() ||
2966 BI->getSuccessor(i: 0) == BI->getSuccessor(i: 1))
2967 continue;
2968
2969 AddUnswitchCandidatesForInst(BI, BI->getCondition());
2970 }
2971
2972 if (MSSAU && !findOptionMDForLoop(TheLoop: &L, Name: "llvm.loop.unswitch.partial.disable") &&
2973 !any_of(Range&: UnswitchCandidates, P: [&L](auto &TerminatorAndInvariants) {
2974 return TerminatorAndInvariants.TI == L.getHeader()->getTerminator();
2975 })) {
2976 MemorySSA *MSSA = MSSAU->getMemorySSA();
2977 if (auto Info = hasPartialIVCondition(L, MSSAThreshold, MSSA: *MSSA, AA)) {
2978 LLVM_DEBUG(
2979 dbgs() << "simple-loop-unswitch: Found partially invariant condition "
2980 << *Info->InstToDuplicate[0] << "\n");
2981 PartialIVInfo = *Info;
2982 PartialIVCondBranch = L.getHeader()->getTerminator();
2983 TinyPtrVector<Value *> ValsToDuplicate;
2984 llvm::append_range(C&: ValsToDuplicate, R&: Info->InstToDuplicate);
2985 UnswitchCandidates.push_back(
2986 Elt: {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)});
2987 }
2988 }
2989 return !UnswitchCandidates.empty();
2990}
2991
2992/// Tries to canonicalize condition described by:
2993///
2994/// br (LHS pred RHS), label IfTrue, label IfFalse
2995///
2996/// into its equivalent where `Pred` is something that we support for injected
2997/// invariants (so far it is limited to ult), LHS in canonicalized form is
2998/// non-invariant and RHS is an invariant.
2999static void canonicalizeForInvariantConditionInjection(CmpPredicate &Pred,
3000 Value *&LHS, Value *&RHS,
3001 BasicBlock *&IfTrue,
3002 BasicBlock *&IfFalse,
3003 const Loop &L) {
3004 if (!L.contains(BB: IfTrue)) {
3005 Pred = ICmpInst::getInversePredicate(pred: Pred);
3006 std::swap(a&: IfTrue, b&: IfFalse);
3007 }
3008
3009 // Move loop-invariant argument to RHS position.
3010 if (L.isLoopInvariant(V: LHS)) {
3011 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
3012 std::swap(a&: LHS, b&: RHS);
3013 }
3014
3015 if (Pred == ICmpInst::ICMP_SGE && match(V: RHS, P: m_Zero())) {
3016 // Turn "x >=s 0" into "x <u UMIN_INT"
3017 Pred = ICmpInst::ICMP_ULT;
3018 RHS = ConstantInt::get(
3019 Context&: RHS->getContext(),
3020 V: APInt::getSignedMinValue(numBits: RHS->getType()->getIntegerBitWidth()));
3021 }
3022}
3023
3024/// Returns true, if predicate described by ( \p Pred, \p LHS, \p RHS )
3025/// succeeding into blocks ( \p IfTrue, \p IfFalse) can be optimized by
3026/// injecting a loop-invariant condition.
3027static bool shouldTryInjectInvariantCondition(
3028 const ICmpInst::Predicate Pred, const Value *LHS, const Value *RHS,
3029 const BasicBlock *IfTrue, const BasicBlock *IfFalse, const Loop &L) {
3030 if (L.isLoopInvariant(V: LHS) || !L.isLoopInvariant(V: RHS))
3031 return false;
3032 // TODO: Support other predicates.
3033 if (Pred != ICmpInst::ICMP_ULT)
3034 return false;
3035 // TODO: Support non-loop-exiting branches?
3036 if (!L.contains(BB: IfTrue) || L.contains(BB: IfFalse))
3037 return false;
3038 // FIXME: For some reason this causes problems with MSSA updates, need to
3039 // investigate why. So far, just don't unswitch latch.
3040 if (L.getHeader() == IfTrue)
3041 return false;
3042 return true;
3043}
3044
3045/// Returns true, if metadata on \p BI allows us to optimize branching into \p
3046/// TakenSucc via injection of invariant conditions. The branch should be not
3047/// enough and not previously unswitched, the information about this comes from
3048/// the metadata.
3049bool shouldTryInjectBasingOnMetadata(const BranchInst *BI,
3050 const BasicBlock *TakenSucc) {
3051 SmallVector<uint32_t> Weights;
3052 if (!extractBranchWeights(I: *BI, Weights))
3053 return false;
3054 unsigned T = InjectInvariantConditionHotnesThreshold;
3055 BranchProbability LikelyTaken(T - 1, T);
3056
3057 assert(Weights.size() == 2 && "Unexpected profile data!");
3058 size_t Idx = BI->getSuccessor(i: 0) == TakenSucc ? 0 : 1;
3059 auto Num = Weights[Idx];
3060 auto Denom = Weights[0] + Weights[1];
3061 // Degenerate or overflowed metadata.
3062 if (Denom == 0 || Num > Denom)
3063 return false;
3064 BranchProbability ActualTaken(Num, Denom);
3065 if (LikelyTaken > ActualTaken)
3066 return false;
3067 return true;
3068}
3069
3070/// Materialize pending invariant condition of the given candidate into IR. The
3071/// injected loop-invariant condition implies the original loop-variant branch
3072/// condition, so the materialization turns
3073///
3074/// loop_block:
3075/// ...
3076/// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3077///
3078/// into
3079///
3080/// preheader:
3081/// %invariant_cond = LHS pred RHS
3082/// ...
3083/// loop_block:
3084/// br i1 %invariant_cond, label InLoopSucc, label OriginalCheck
3085/// OriginalCheck:
3086/// br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3087/// ...
3088static NonTrivialUnswitchCandidate
3089injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L,
3090 DominatorTree &DT, LoopInfo &LI,
3091 AssumptionCache &AC, MemorySSAUpdater *MSSAU) {
3092 assert(Candidate.hasPendingInjection() && "Nothing to inject!");
3093 BasicBlock *Preheader = L.getLoopPreheader();
3094 assert(Preheader && "Loop is not in simplified form?");
3095 assert(LI.getLoopFor(Candidate.TI->getParent()) == &L &&
3096 "Unswitching branch of inner loop!");
3097
3098 auto Pred = Candidate.PendingInjection->Pred;
3099 auto *LHS = Candidate.PendingInjection->LHS;
3100 auto *RHS = Candidate.PendingInjection->RHS;
3101 auto *InLoopSucc = Candidate.PendingInjection->InLoopSucc;
3102 auto *TI = cast<BranchInst>(Val: Candidate.TI);
3103 auto *BB = Candidate.TI->getParent();
3104 auto *OutOfLoopSucc = InLoopSucc == TI->getSuccessor(i: 0) ? TI->getSuccessor(i: 1)
3105 : TI->getSuccessor(i: 0);
3106 // FIXME: Remove this once limitation on successors is lifted.
3107 assert(L.contains(InLoopSucc) && "Not supported yet!");
3108 assert(!L.contains(OutOfLoopSucc) && "Not supported yet!");
3109 auto &Ctx = BB->getContext();
3110
3111 IRBuilder<> Builder(Preheader->getTerminator());
3112 assert(ICmpInst::isUnsigned(Pred) && "Not supported yet!");
3113 if (LHS->getType() != RHS->getType()) {
3114 if (LHS->getType()->getIntegerBitWidth() <
3115 RHS->getType()->getIntegerBitWidth())
3116 LHS = Builder.CreateZExt(V: LHS, DestTy: RHS->getType(), Name: LHS->getName() + ".wide");
3117 else
3118 RHS = Builder.CreateZExt(V: RHS, DestTy: LHS->getType(), Name: RHS->getName() + ".wide");
3119 }
3120 // Do not use builder here: CreateICmp may simplify this into a constant and
3121 // unswitching will break. Better optimize it away later.
3122 auto *InjectedCond =
3123 ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: LHS, S2: RHS, Name: "injected.cond",
3124 InsertBefore: Preheader->getTerminator()->getIterator());
3125
3126 BasicBlock *CheckBlock = BasicBlock::Create(Context&: Ctx, Name: BB->getName() + ".check",
3127 Parent: BB->getParent(), InsertBefore: InLoopSucc);
3128 Builder.SetInsertPoint(TI);
3129 auto *InvariantBr =
3130 Builder.CreateCondBr(Cond: InjectedCond, True: InLoopSucc, False: CheckBlock);
3131
3132 Builder.SetInsertPoint(CheckBlock);
3133 Builder.CreateCondBr(Cond: TI->getCondition(), True: TI->getSuccessor(i: 0),
3134 False: TI->getSuccessor(i: 1));
3135 TI->eraseFromParent();
3136
3137 // Fixup phis.
3138 for (auto &I : *InLoopSucc) {
3139 auto *PN = dyn_cast<PHINode>(Val: &I);
3140 if (!PN)
3141 break;
3142 auto *Inc = PN->getIncomingValueForBlock(BB);
3143 PN->addIncoming(V: Inc, BB: CheckBlock);
3144 }
3145 OutOfLoopSucc->replacePhiUsesWith(Old: BB, New: CheckBlock);
3146
3147 SmallVector<DominatorTree::UpdateType, 4> DTUpdates = {
3148 { DominatorTree::Insert, BB, CheckBlock },
3149 { DominatorTree::Insert, CheckBlock, InLoopSucc },
3150 { DominatorTree::Insert, CheckBlock, OutOfLoopSucc },
3151 { DominatorTree::Delete, BB, OutOfLoopSucc }
3152 };
3153
3154 DT.applyUpdates(Updates: DTUpdates);
3155 if (MSSAU)
3156 MSSAU->applyUpdates(Updates: DTUpdates, DT);
3157 L.addBasicBlockToLoop(NewBB: CheckBlock, LI);
3158
3159#ifndef NDEBUG
3160 DT.verify();
3161 LI.verify(DT);
3162 if (MSSAU && VerifyMemorySSA)
3163 MSSAU->getMemorySSA()->verifyMemorySSA();
3164#endif
3165
3166 // TODO: In fact, cost of unswitching a new invariant candidate is *slightly*
3167 // higher because we have just inserted a new block. Need to think how to
3168 // adjust the cost of injected candidates when it was first computed.
3169 LLVM_DEBUG(dbgs() << "Injected a new loop-invariant branch " << *InvariantBr
3170 << " and considering it for unswitching.");
3171 ++NumInvariantConditionsInjected;
3172 return NonTrivialUnswitchCandidate(InvariantBr, { InjectedCond },
3173 Candidate.Cost);
3174}
3175
3176/// Given chain of loop branch conditions looking like:
3177/// br (Variant < Invariant1)
3178/// br (Variant < Invariant2)
3179/// br (Variant < Invariant3)
3180/// ...
3181/// collect set of invariant conditions on which we want to unswitch, which
3182/// look like:
3183/// Invariant1 <= Invariant2
3184/// Invariant2 <= Invariant3
3185/// ...
3186/// Though they might not immediately exist in the IR, we can still inject them.
3187static bool insertCandidatesWithPendingInjections(
3188 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, Loop &L,
3189 ICmpInst::Predicate Pred, ArrayRef<CompareDesc> Compares,
3190 const DominatorTree &DT) {
3191
3192 assert(ICmpInst::isRelational(Pred));
3193 assert(ICmpInst::isStrictPredicate(Pred));
3194 if (Compares.size() < 2)
3195 return false;
3196 ICmpInst::Predicate NonStrictPred = ICmpInst::getNonStrictPredicate(pred: Pred);
3197 for (auto Prev = Compares.begin(), Next = Compares.begin() + 1;
3198 Next != Compares.end(); ++Prev, ++Next) {
3199 Value *LHS = Next->Invariant;
3200 Value *RHS = Prev->Invariant;
3201 BasicBlock *InLoopSucc = Prev->InLoopSucc;
3202 InjectedInvariant ToInject(NonStrictPred, LHS, RHS, InLoopSucc);
3203 NonTrivialUnswitchCandidate Candidate(Prev->Term, { LHS, RHS },
3204 std::nullopt, std::move(ToInject));
3205 UnswitchCandidates.push_back(Elt: std::move(Candidate));
3206 }
3207 return true;
3208}
3209
3210/// Collect unswitch candidates by invariant conditions that are not immediately
3211/// present in the loop. However, they can be injected into the code if we
3212/// decide it's profitable.
3213/// An example of such conditions is following:
3214///
3215/// for (...) {
3216/// x = load ...
3217/// if (! x <u C1) break;
3218/// if (! x <u C2) break;
3219/// <do something>
3220/// }
3221///
3222/// We can unswitch by condition "C1 <=u C2". If that is true, then "x <u C1 <=
3223/// C2" automatically implies "x <u C2", so we can get rid of one of
3224/// loop-variant checks in unswitched loop version.
3225static bool collectUnswitchCandidatesWithInjections(
3226 SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
3227 IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, Loop &L,
3228 const DominatorTree &DT, const LoopInfo &LI, AAResults &AA,
3229 const MemorySSAUpdater *MSSAU) {
3230 if (!InjectInvariantConditions)
3231 return false;
3232
3233 if (!DT.isReachableFromEntry(A: L.getHeader()))
3234 return false;
3235 auto *Latch = L.getLoopLatch();
3236 // Need to have a single latch and a preheader.
3237 if (!Latch)
3238 return false;
3239 assert(L.getLoopPreheader() && "Must have a preheader!");
3240
3241 DenseMap<Value *, SmallVector<CompareDesc, 4> > CandidatesULT;
3242 // Traverse the conditions that dominate latch (and therefore dominate each
3243 // other).
3244 for (auto *DTN = DT.getNode(BB: Latch); L.contains(BB: DTN->getBlock());
3245 DTN = DTN->getIDom()) {
3246 CmpPredicate Pred;
3247 Value *LHS = nullptr, *RHS = nullptr;
3248 BasicBlock *IfTrue = nullptr, *IfFalse = nullptr;
3249 auto *BB = DTN->getBlock();
3250 // Ignore inner loops.
3251 if (LI.getLoopFor(BB) != &L)
3252 continue;
3253 auto *Term = BB->getTerminator();
3254 if (!match(V: Term, P: m_Br(C: m_ICmp(Pred, L: m_Value(V&: LHS), R: m_Value(V&: RHS)),
3255 T: m_BasicBlock(V&: IfTrue), F: m_BasicBlock(V&: IfFalse))))
3256 continue;
3257 if (!LHS->getType()->isIntegerTy())
3258 continue;
3259 canonicalizeForInvariantConditionInjection(Pred, LHS, RHS, IfTrue, IfFalse,
3260 L);
3261 if (!shouldTryInjectInvariantCondition(Pred, LHS, RHS, IfTrue, IfFalse, L))
3262 continue;
3263 if (!shouldTryInjectBasingOnMetadata(BI: cast<BranchInst>(Val: Term), TakenSucc: IfTrue))
3264 continue;
3265 // Strip ZEXT for unsigned predicate.
3266 // TODO: once signed predicates are supported, also strip SEXT.
3267 CompareDesc Desc(cast<BranchInst>(Val: Term), RHS, IfTrue);
3268 while (auto *Zext = dyn_cast<ZExtInst>(Val: LHS))
3269 LHS = Zext->getOperand(i_nocapture: 0);
3270 CandidatesULT[LHS].push_back(Elt: Desc);
3271 }
3272
3273 bool Found = false;
3274 for (auto &It : CandidatesULT)
3275 Found |= insertCandidatesWithPendingInjections(
3276 UnswitchCandidates, L, Pred: ICmpInst::ICMP_ULT, Compares: It.second, DT);
3277 return Found;
3278}
3279
3280static bool isSafeForNoNTrivialUnswitching(Loop &L, LoopInfo &LI) {
3281 if (!L.isSafeToClone())
3282 return false;
3283 for (auto *BB : L.blocks())
3284 for (auto &I : *BB) {
3285 if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
3286 return false;
3287 if (auto *CB = dyn_cast<CallBase>(Val: &I)) {
3288 assert(!CB->cannotDuplicate() && "Checked by L.isSafeToClone().");
3289 if (CB->isConvergent())
3290 return false;
3291 }
3292 }
3293
3294 // Check if there are irreducible CFG cycles in this loop. If so, we cannot
3295 // easily unswitch non-trivial edges out of the loop. Doing so might turn the
3296 // irreducible control flow into reducible control flow and introduce new
3297 // loops "out of thin air". If we ever discover important use cases for doing
3298 // this, we can add support to loop unswitch, but it is a lot of complexity
3299 // for what seems little or no real world benefit.
3300 LoopBlocksRPO RPOT(&L);
3301 RPOT.perform(LI: &LI);
3302 if (containsIrreducibleCFG<const BasicBlock *>(RPOTraversal&: RPOT, LI))
3303 return false;
3304
3305 SmallVector<BasicBlock *, 4> ExitBlocks;
3306 L.getUniqueExitBlocks(ExitBlocks);
3307 // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch
3308 // instruction as we don't know how to split those exit blocks.
3309 // FIXME: We should teach SplitBlock to handle this and remove this
3310 // restriction.
3311 for (auto *ExitBB : ExitBlocks) {
3312 auto It = ExitBB->getFirstNonPHIIt();
3313 if (isa<CleanupPadInst>(Val: It) || isa<CatchSwitchInst>(Val: It)) {
3314 LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch "
3315 "in exit block\n");
3316 return false;
3317 }
3318 }
3319
3320 return true;
3321}
3322
3323static NonTrivialUnswitchCandidate findBestNonTrivialUnswitchCandidate(
3324 ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates, const Loop &L,
3325 const DominatorTree &DT, const LoopInfo &LI, AssumptionCache &AC,
3326 const TargetTransformInfo &TTI, const IVConditionInfo &PartialIVInfo) {
3327 // Given that unswitching these terminators will require duplicating parts of
3328 // the loop, so we need to be able to model that cost. Compute the ephemeral
3329 // values and set up a data structure to hold per-BB costs. We cache each
3330 // block's cost so that we don't recompute this when considering different
3331 // subsets of the loop for duplication during unswitching.
3332 SmallPtrSet<const Value *, 4> EphValues;
3333 CodeMetrics::collectEphemeralValues(L: &L, AC: &AC, EphValues);
3334 SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap;
3335
3336 // Compute the cost of each block, as well as the total loop cost. Also, bail
3337 // out if we see instructions which are incompatible with loop unswitching
3338 // (convergent, noduplicate, or cross-basic-block tokens).
3339 // FIXME: We might be able to safely handle some of these in non-duplicated
3340 // regions.
3341 TargetTransformInfo::TargetCostKind CostKind =
3342 L.getHeader()->getParent()->hasMinSize()
3343 ? TargetTransformInfo::TCK_CodeSize
3344 : TargetTransformInfo::TCK_SizeAndLatency;
3345 InstructionCost LoopCost = 0;
3346 for (auto *BB : L.blocks()) {
3347 InstructionCost Cost = 0;
3348 for (auto &I : *BB) {
3349 if (EphValues.count(Ptr: &I))
3350 continue;
3351 Cost += TTI.getInstructionCost(U: &I, CostKind);
3352 }
3353 assert(Cost >= 0 && "Must not have negative costs!");
3354 LoopCost += Cost;
3355 assert(LoopCost >= 0 && "Must not have negative loop costs!");
3356 BBCostMap[BB] = Cost;
3357 }
3358 LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n");
3359
3360 // Now we find the best candidate by searching for the one with the following
3361 // properties in order:
3362 //
3363 // 1) An unswitching cost below the threshold
3364 // 2) The smallest number of duplicated unswitch candidates (to avoid
3365 // creating redundant subsequent unswitching)
3366 // 3) The smallest cost after unswitching.
3367 //
3368 // We prioritize reducing fanout of unswitch candidates provided the cost
3369 // remains below the threshold because this has a multiplicative effect.
3370 //
3371 // This requires memoizing each dominator subtree to avoid redundant work.
3372 //
3373 // FIXME: Need to actually do the number of candidates part above.
3374 SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap;
3375 // Given a terminator which might be unswitched, computes the non-duplicated
3376 // cost for that terminator.
3377 auto ComputeUnswitchedCost = [&](Instruction &TI,
3378 bool FullUnswitch) -> InstructionCost {
3379 // Unswitching selects unswitches the entire loop.
3380 if (isa<SelectInst>(Val: TI))
3381 return LoopCost;
3382
3383 BasicBlock &BB = *TI.getParent();
3384 SmallPtrSet<BasicBlock *, 4> Visited;
3385
3386 InstructionCost Cost = 0;
3387 for (BasicBlock *SuccBB : successors(BB: &BB)) {
3388 // Don't count successors more than once.
3389 if (!Visited.insert(Ptr: SuccBB).second)
3390 continue;
3391
3392 // If this is a partial unswitch candidate, then it must be a conditional
3393 // branch with a condition of either `or`, `and`, their corresponding
3394 // select forms or partially invariant instructions. In that case, one of
3395 // the successors is necessarily duplicated, so don't even try to remove
3396 // its cost.
3397 if (!FullUnswitch) {
3398 auto &BI = cast<BranchInst>(Val&: TI);
3399 Value *Cond = skipTrivialSelect(Cond: BI.getCondition());
3400 if (match(V: Cond, P: m_LogicalAnd())) {
3401 if (SuccBB == BI.getSuccessor(i: 1))
3402 continue;
3403 } else if (match(V: Cond, P: m_LogicalOr())) {
3404 if (SuccBB == BI.getSuccessor(i: 0))
3405 continue;
3406 } else if ((PartialIVInfo.KnownValue->isOneValue() &&
3407 SuccBB == BI.getSuccessor(i: 0)) ||
3408 (!PartialIVInfo.KnownValue->isOneValue() &&
3409 SuccBB == BI.getSuccessor(i: 1)))
3410 continue;
3411 }
3412
3413 // This successor's domtree will not need to be duplicated after
3414 // unswitching if the edge to the successor dominates it (and thus the
3415 // entire tree). This essentially means there is no other path into this
3416 // subtree and so it will end up live in only one clone of the loop.
3417 if (SuccBB->getUniquePredecessor() ||
3418 llvm::all_of(Range: predecessors(BB: SuccBB), P: [&](BasicBlock *PredBB) {
3419 return PredBB == &BB || DT.dominates(A: SuccBB, B: PredBB);
3420 })) {
3421 Cost += computeDomSubtreeCost(N&: *DT[SuccBB], BBCostMap, DTCostMap);
3422 assert(Cost <= LoopCost &&
3423 "Non-duplicated cost should never exceed total loop cost!");
3424 }
3425 }
3426
3427 // Now scale the cost by the number of unique successors minus one. We
3428 // subtract one because there is already at least one copy of the entire
3429 // loop. This is computing the new cost of unswitching a condition.
3430 // Note that guards always have 2 unique successors that are implicit and
3431 // will be materialized if we decide to unswitch it.
3432 int SuccessorsCount = isGuard(U: &TI) ? 2 : Visited.size();
3433 assert(SuccessorsCount > 1 &&
3434 "Cannot unswitch a condition without multiple distinct successors!");
3435 return (LoopCost - Cost) * (SuccessorsCount - 1);
3436 };
3437
3438 std::optional<NonTrivialUnswitchCandidate> Best;
3439 for (auto &Candidate : UnswitchCandidates) {
3440 Instruction &TI = *Candidate.TI;
3441 ArrayRef<Value *> Invariants = Candidate.Invariants;
3442 BranchInst *BI = dyn_cast<BranchInst>(Val: &TI);
3443 bool FullUnswitch =
3444 !BI || Candidate.hasPendingInjection() ||
3445 (Invariants.size() == 1 &&
3446 Invariants[0] == skipTrivialSelect(Cond: BI->getCondition()));
3447 InstructionCost CandidateCost = ComputeUnswitchedCost(TI, FullUnswitch);
3448 // Calculate cost multiplier which is a tool to limit potentially
3449 // exponential behavior of loop-unswitch.
3450 if (EnableUnswitchCostMultiplier) {
3451 int CostMultiplier =
3452 CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
3453 assert(
3454 (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
3455 "cost multiplier needs to be in the range of 1..UnswitchThreshold");
3456 CandidateCost *= CostMultiplier;
3457 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
3458 << " (multiplier: " << CostMultiplier << ")"
3459 << " for unswitch candidate: " << TI << "\n");
3460 } else {
3461 LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCost
3462 << " for unswitch candidate: " << TI << "\n");
3463 }
3464
3465 if (!Best || CandidateCost < Best->Cost) {
3466 Best = Candidate;
3467 Best->Cost = CandidateCost;
3468 }
3469 }
3470 assert(Best && "Must be!");
3471 return *Best;
3472}
3473
3474// Insert a freeze on an unswitched branch if all is true:
3475// 1. freeze-loop-unswitch-cond option is true
3476// 2. The branch may not execute in the loop pre-transformation. If a branch may
3477// not execute and could cause UB, it would always cause UB if it is hoisted outside
3478// of the loop. Insert a freeze to prevent this case.
3479// 3. The branch condition may be poison or undef
3480static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT,
3481 AssumptionCache &AC) {
3482 assert(isa<BranchInst>(TI) || isa<SwitchInst>(TI));
3483 if (!FreezeLoopUnswitchCond)
3484 return false;
3485
3486 ICFLoopSafetyInfo SafetyInfo;
3487 SafetyInfo.computeLoopSafetyInfo(CurLoop: &L);
3488 if (SafetyInfo.isGuaranteedToExecute(Inst: TI, DT: &DT, CurLoop: &L))
3489 return false;
3490
3491 Value *Cond;
3492 if (BranchInst *BI = dyn_cast<BranchInst>(Val: &TI))
3493 Cond = skipTrivialSelect(Cond: BI->getCondition());
3494 else
3495 Cond = skipTrivialSelect(Cond: cast<SwitchInst>(Val: &TI)->getCondition());
3496 return !isGuaranteedNotToBeUndefOrPoison(
3497 V: Cond, AC: &AC, CtxI: L.getLoopPreheader()->getTerminator(), DT: &DT);
3498}
3499
3500static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
3501 AssumptionCache &AC, AAResults &AA,
3502 TargetTransformInfo &TTI, ScalarEvolution *SE,
3503 MemorySSAUpdater *MSSAU,
3504 LPMUpdater &LoopUpdater) {
3505 // Collect all invariant conditions within this loop (as opposed to an inner
3506 // loop which would be handled when visiting that inner loop).
3507 SmallVector<NonTrivialUnswitchCandidate, 4> UnswitchCandidates;
3508 IVConditionInfo PartialIVInfo;
3509 Instruction *PartialIVCondBranch = nullptr;
3510 collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo,
3511 PartialIVCondBranch, L, LI, AA, MSSAU);
3512 if (!findOptionMDForLoop(TheLoop: &L, Name: "llvm.loop.unswitch.injection.disable"))
3513 collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo,
3514 PartialIVCondBranch, L, DT, LI, AA,
3515 MSSAU);
3516 // If we didn't find any candidates, we're done.
3517 if (UnswitchCandidates.empty())
3518 return false;
3519
3520 LLVM_DEBUG(
3521 dbgs() << "Considering " << UnswitchCandidates.size()
3522 << " non-trivial loop invariant conditions for unswitching.\n");
3523
3524 NonTrivialUnswitchCandidate Best = findBestNonTrivialUnswitchCandidate(
3525 UnswitchCandidates, L, DT, LI, AC, TTI, PartialIVInfo);
3526
3527 assert(Best.TI && "Failed to find loop unswitch candidate");
3528 assert(Best.Cost && "Failed to compute cost");
3529
3530 if (*Best.Cost >= UnswitchThreshold) {
3531 LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << *Best.Cost
3532 << "\n");
3533 return false;
3534 }
3535
3536 bool InjectedCondition = false;
3537 if (Best.hasPendingInjection()) {
3538 Best = injectPendingInvariantConditions(Candidate: Best, L, DT, LI, AC, MSSAU);
3539 InjectedCondition = true;
3540 }
3541 assert(!Best.hasPendingInjection() &&
3542 "All injections should have been done by now!");
3543
3544 if (Best.TI != PartialIVCondBranch)
3545 PartialIVInfo.InstToDuplicate.clear();
3546
3547 bool InsertFreeze;
3548 if (auto *SI = dyn_cast<SelectInst>(Val: Best.TI)) {
3549 // If the best candidate is a select, turn it into a branch. Select
3550 // instructions with a poison conditional do not propagate poison, but
3551 // branching on poison causes UB. Insert a freeze on the select
3552 // conditional to prevent UB after turning the select into a branch.
3553 InsertFreeze = !isGuaranteedNotToBeUndefOrPoison(
3554 V: SI->getCondition(), AC: &AC, CtxI: L.getLoopPreheader()->getTerminator(), DT: &DT);
3555 Best.TI = turnSelectIntoBranch(SI, DT, LI, MSSAU, AC: &AC);
3556 } else {
3557 // If the best candidate is a guard, turn it into a branch.
3558 if (isGuard(U: Best.TI))
3559 Best.TI =
3560 turnGuardIntoBranch(GI: cast<IntrinsicInst>(Val: Best.TI), L, DT, LI, MSSAU);
3561 InsertFreeze = shouldInsertFreeze(L, TI&: *Best.TI, DT, AC);
3562 }
3563
3564 LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = " << Best.Cost
3565 << ") terminator: " << *Best.TI << "\n");
3566 unswitchNontrivialInvariants(L, TI&: *Best.TI, Invariants: Best.Invariants, PartialIVInfo, DT,
3567 LI, AC, SE, MSSAU, LoopUpdater, InsertFreeze,
3568 InjectedCondition);
3569 return true;
3570}
3571
3572/// Unswitch control flow predicated on loop invariant conditions.
3573///
3574/// This first hoists all branches or switches which are trivial (IE, do not
3575/// require duplicating any part of the loop) out of the loop body. It then
3576/// looks at other loop invariant control flows and tries to unswitch those as
3577/// well by cloning the loop if the result is small enough.
3578///
3579/// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are
3580/// also updated based on the unswitch. The `MSSA` analysis is also updated if
3581/// valid (i.e. its use is enabled).
3582///
3583/// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
3584/// true, we will attempt to do non-trivial unswitching as well as trivial
3585/// unswitching.
3586///
3587/// The `postUnswitch` function will be run after unswitching is complete
3588/// with information on whether or not the provided loop remains a loop and
3589/// a list of new sibling loops created.
3590///
3591/// If `SE` is non-null, we will update that analysis based on the unswitching
3592/// done.
3593static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
3594 AssumptionCache &AC, AAResults &AA,
3595 TargetTransformInfo &TTI, bool Trivial,
3596 bool NonTrivial, ScalarEvolution *SE,
3597 MemorySSAUpdater *MSSAU, ProfileSummaryInfo *PSI,
3598 BlockFrequencyInfo *BFI, LPMUpdater &LoopUpdater) {
3599 assert(L.isRecursivelyLCSSAForm(DT, LI) &&
3600 "Loops must be in LCSSA form before unswitching.");
3601
3602 // Must be in loop simplified form: we need a preheader and dedicated exits.
3603 if (!L.isLoopSimplifyForm())
3604 return false;
3605
3606 // Try trivial unswitch first before loop over other basic blocks in the loop.
3607 if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
3608 // If we unswitched successfully we will want to clean up the loop before
3609 // processing it further so just mark it as unswitched and return.
3610 postUnswitch(L, U&: LoopUpdater, LoopName: L.getName(),
3611 /*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false,
3612 /*InjectedCondition*/ false, NewLoops: {});
3613 return true;
3614 }
3615
3616 const Function *F = L.getHeader()->getParent();
3617
3618 // Check whether we should continue with non-trivial conditions.
3619 // EnableNonTrivialUnswitch: Global variable that forces non-trivial
3620 // unswitching for testing and debugging.
3621 // NonTrivial: Parameter that enables non-trivial unswitching for this
3622 // invocation of the transform. But this should be allowed only
3623 // for targets without branch divergence.
3624 //
3625 // FIXME: If divergence analysis becomes available to a loop
3626 // transform, we should allow unswitching for non-trivial uniform
3627 // branches even on targets that have divergence.
3628 // https://bugs.llvm.org/show_bug.cgi?id=48819
3629 bool ContinueWithNonTrivial =
3630 EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence(F));
3631 if (!ContinueWithNonTrivial)
3632 return false;
3633
3634 // Skip non-trivial unswitching for optsize functions.
3635 if (F->hasOptSize())
3636 return false;
3637
3638 // Returns true if Loop L's loop nest is cold, i.e. if the headers of L,
3639 // of the loops L is nested in, and of the loops nested in L are all cold.
3640 auto IsLoopNestCold = [&](const Loop *L) {
3641 // Check L and all of its parent loops.
3642 auto *Parent = L;
3643 while (Parent) {
3644 if (!PSI->isColdBlock(BB: Parent->getHeader(), BFI))
3645 return false;
3646 Parent = Parent->getParentLoop();
3647 }
3648 // Next check all loops nested within L.
3649 SmallVector<const Loop *, 4> Worklist;
3650 llvm::append_range(C&: Worklist, R: L->getSubLoops());
3651 while (!Worklist.empty()) {
3652 auto *CurLoop = Worklist.pop_back_val();
3653 if (!PSI->isColdBlock(BB: CurLoop->getHeader(), BFI))
3654 return false;
3655 llvm::append_range(C&: Worklist, R: CurLoop->getSubLoops());
3656 }
3657 return true;
3658 };
3659
3660 // Skip cold loops in cold loop nests, as unswitching them brings little
3661 // benefit but increases the code size
3662 if (PSI && PSI->hasProfileSummary() && BFI && IsLoopNestCold(&L)) {
3663 LLVM_DEBUG(dbgs() << " Skip cold loop: " << L << "\n");
3664 return false;
3665 }
3666
3667 // Perform legality checks.
3668 if (!isSafeForNoNTrivialUnswitching(L, LI))
3669 return false;
3670
3671 // For non-trivial unswitching, because it often creates new loops, we rely on
3672 // the pass manager to iterate on the loops rather than trying to immediately
3673 // reach a fixed point. There is no substantial advantage to iterating
3674 // internally, and if any of the new loops are simplified enough to contain
3675 // trivial unswitching we want to prefer those.
3676
3677 // Try to unswitch the best invariant condition. We prefer this full unswitch to
3678 // a partial unswitch when possible below the threshold.
3679 if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, SE, MSSAU, LoopUpdater))
3680 return true;
3681
3682 // No other opportunities to unswitch.
3683 return false;
3684}
3685
3686PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
3687 LoopStandardAnalysisResults &AR,
3688 LPMUpdater &U) {
3689 Function &F = *L.getHeader()->getParent();
3690 (void)F;
3691 ProfileSummaryInfo *PSI = nullptr;
3692 if (auto OuterProxy =
3693 AM.getResult<FunctionAnalysisManagerLoopProxy>(IR&: L, ExtraArgs&: AR)
3694 .getCachedResult<ModuleAnalysisManagerFunctionProxy>(IR&: F))
3695 PSI = OuterProxy->getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
3696 LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
3697 << "\n");
3698
3699 std::optional<MemorySSAUpdater> MSSAU;
3700 if (AR.MSSA) {
3701 MSSAU = MemorySSAUpdater(AR.MSSA);
3702 if (VerifyMemorySSA)
3703 AR.MSSA->verifyMemorySSA();
3704 }
3705 if (!unswitchLoop(L, DT&: AR.DT, LI&: AR.LI, AC&: AR.AC, AA&: AR.AA, TTI&: AR.TTI, Trivial, NonTrivial,
3706 SE: &AR.SE, MSSAU: MSSAU ? &*MSSAU : nullptr, PSI, BFI: AR.BFI, LoopUpdater&: U))
3707 return PreservedAnalyses::all();
3708
3709 if (AR.MSSA && VerifyMemorySSA)
3710 AR.MSSA->verifyMemorySSA();
3711
3712 // Historically this pass has had issues with the dominator tree so verify it
3713 // in asserts builds.
3714 assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));
3715
3716 auto PA = getLoopPassPreservedAnalyses();
3717 if (AR.MSSA)
3718 PA.preserve<MemorySSAAnalysis>();
3719 return PA;
3720}
3721
3722void SimpleLoopUnswitchPass::printPipeline(
3723 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
3724 static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline(
3725 OS, MapClassName2PassName);
3726
3727 OS << '<';
3728 OS << (NonTrivial ? "" : "no-") << "nontrivial;";
3729 OS << (Trivial ? "" : "no-") << "trivial";
3730 OS << '>';
3731}
3732