xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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 
71 using namespace llvm;
72 using namespace llvm::PatternMatch;
73 
74 STATISTIC(NumBranches, "Number of branches unswitched");
75 STATISTIC(NumSwitches, "Number of switches unswitched");
76 STATISTIC(NumSelects, "Number of selects turned into branches for unswitching");
77 STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
78 STATISTIC(NumTrivial, "Number of unswitches that are trivial");
79 STATISTIC(
80     NumCostMultiplierSkipped,
81     "Number of unswitch candidates that had their cost multiplier skipped");
82 STATISTIC(NumInvariantConditionsInjected,
83           "Number of invariant conditions injected and unswitched");
84 
85 static cl::opt<bool> EnableNonTrivialUnswitch(
86     "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
87     cl::desc("Forcibly enables non-trivial loop unswitching rather than "
88              "following the configuration passed into the pass."));
89 
90 static cl::opt<int>
91     UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
92                       cl::desc("The cost threshold for unswitching a loop."));
93 
94 static cl::opt<bool> EnableUnswitchCostMultiplier(
95     "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden,
96     cl::desc("Enable unswitch cost multiplier that prohibits exponential "
97              "explosion in nontrivial unswitch."));
98 static cl::opt<int> UnswitchSiblingsToplevelDiv(
99     "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden,
100     cl::desc("Toplevel siblings divisor for cost multiplier."));
101 static cl::opt<int> UnswitchNumInitialUnscaledCandidates(
102     "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden,
103     cl::desc("Number of unswitch candidates that are ignored when calculating "
104              "cost multiplier."));
105 static cl::opt<bool> UnswitchGuards(
106     "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
107     cl::desc("If enabled, simple loop unswitching will also consider "
108              "llvm.experimental.guard intrinsics as unswitch candidates."));
109 static cl::opt<bool> DropNonTrivialImplicitNullChecks(
110     "simple-loop-unswitch-drop-non-trivial-implicit-null-checks",
111     cl::init(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."));
114 static 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(100), cl::Hidden);
119 static cl::opt<bool> FreezeLoopUnswitchCond(
120     "freeze-loop-unswitch-cond", cl::init(true), cl::Hidden,
121     cl::desc("If enabled, the freeze instruction will be added to condition "
122              "of loop unswitch to prevent miscompilation."));
123 
124 static 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(true));
129 
130 static 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(16));
136 
137 AnalysisKey ShouldRunExtraSimpleLoopUnswitch::Key;
138 namespace {
139 struct CompareDesc {
140   BranchInst *Term;
141   Value *Invariant;
142   BasicBlock *InLoopSucc;
143 
CompareDesc__anondb665f8c0111::CompareDesc144   CompareDesc(BranchInst *Term, Value *Invariant, BasicBlock *InLoopSucc)
145       : Term(Term), Invariant(Invariant), InLoopSucc(InLoopSucc) {}
146 };
147 
148 struct InjectedInvariant {
149   ICmpInst::Predicate Pred;
150   Value *LHS;
151   Value *RHS;
152   BasicBlock *InLoopSucc;
153 
InjectedInvariant__anondb665f8c0111::InjectedInvariant154   InjectedInvariant(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
155                     BasicBlock *InLoopSucc)
156       : Pred(Pred), LHS(LHS), RHS(RHS), InLoopSucc(InLoopSucc) {}
157 };
158 
159 struct NonTrivialUnswitchCandidate {
160   Instruction *TI = nullptr;
161   TinyPtrVector<Value *> Invariants;
162   std::optional<InstructionCost> Cost;
163   std::optional<InjectedInvariant> PendingInjection;
NonTrivialUnswitchCandidate__anondb665f8c0111::NonTrivialUnswitchCandidate164   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 
hasPendingInjection__anondb665f8c0111::NonTrivialUnswitchCandidate171   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.
skipTrivialSelect(Value * Cond)178 static Value *skipTrivialSelect(Value *Cond) {
179   Value *CondNext;
180   while (match(Cond, m_Select(m_Value(CondNext), m_One(), 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.
192 static TinyPtrVector<Value *>
collectHomogenousInstGraphLoopInvariants(const Loop & L,Instruction & Root,const LoopInfo & LI)193 collectHomogenousInstGraphLoopInvariants(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(&Root, m_LogicalAnd());
200   bool IsRootOr  = match(&Root, 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(&Root);
206   Visited.insert(&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>(OpV))
212         continue;
213 
214       // Add it to our result if loop invariant.
215       if (L.isLoopInvariant(OpV)) {
216         Invariants.push_back(OpV);
217         continue;
218       }
219 
220       // If not an instruction with the same opcode, nothing we can do.
221       Instruction *OpI = dyn_cast<Instruction>(skipTrivialSelect(OpV));
222 
223       if (OpI && ((IsRootAnd && match(OpI, m_LogicalAnd())) ||
224                   (IsRootOr  && match(OpI, m_LogicalOr())))) {
225         // Visit this operand.
226         if (Visited.insert(OpI).second)
227           Worklist.push_back(OpI);
228       }
229     }
230   } while (!Worklist.empty());
231 
232   return Invariants;
233 }
234 
replaceLoopInvariantUses(const Loop & L,Value * Invariant,Constant & Replacement)235 static 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(Invariant->uses())) {
242     Instruction *UserI = dyn_cast<Instruction>(U.getUser());
243 
244     // Replace this use within the loop body.
245     if (UserI && L.contains(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.
areLoopExitPHIsLoopInvariant(const Loop & L,const BasicBlock & ExitingBB,const BasicBlock & ExitBB)252 static 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>(&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(PN->getIncomingValueForBlock(&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.
buildPartialUnswitchConditionalBranch(BasicBlock & BB,ArrayRef<Value * > Invariants,bool Direction,BasicBlock & UnswitchedSucc,BasicBlock & NormalSucc,bool InsertFreeze,const Instruction * I,AssumptionCache * AC,const DominatorTree & DT)272 static 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 
278   SmallVector<Value *> FrozenInvariants;
279   for (Value *Inv : Invariants) {
280     if (InsertFreeze && !isGuaranteedNotToBeUndefOrPoison(Inv, AC, I, &DT))
281       Inv = IRB.CreateFreeze(Inv, Inv->getName() + ".fr");
282     FrozenInvariants.push_back(Inv);
283   }
284 
285   Value *Cond = Direction ? IRB.CreateOr(FrozenInvariants)
286                           : IRB.CreateAnd(FrozenInvariants);
287   IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
288                    Direction ? &NormalSucc : &UnswitchedSucc);
289 }
290 
291 /// Copy a set of loop invariant values, and conditionally branch on them.
buildPartialInvariantUnswitchConditionalBranch(BasicBlock & BB,ArrayRef<Value * > ToDuplicate,bool Direction,BasicBlock & UnswitchedSucc,BasicBlock & NormalSucc,Loop & L,MemorySSAUpdater * MSSAU)292 static void buildPartialInvariantUnswitchConditionalBranch(
293     BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction,
294     BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L,
295     MemorySSAUpdater *MSSAU) {
296   ValueToValueMapTy VMap;
297   for (auto *Val : reverse(ToDuplicate)) {
298     Instruction *Inst = cast<Instruction>(Val);
299     Instruction *NewInst = Inst->clone();
300     NewInst->insertInto(&BB, BB.end());
301     RemapInstruction(NewInst, VMap,
302                      RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
303     VMap[Val] = NewInst;
304 
305     if (!MSSAU)
306       continue;
307 
308     MemorySSA *MSSA = MSSAU->getMemorySSA();
309     if (auto *MemUse =
310             dyn_cast_or_null<MemoryUse>(MSSA->getMemoryAccess(Inst))) {
311       auto *DefiningAccess = MemUse->getDefiningAccess();
312       // Get the first defining access before the loop.
313       while (L.contains(DefiningAccess->getBlock())) {
314         // If the defining access is a MemoryPhi, get the incoming
315         // value for the pre-header as defining access.
316         if (auto *MemPhi = dyn_cast<MemoryPhi>(DefiningAccess))
317           DefiningAccess =
318               MemPhi->getIncomingValueForBlock(L.getLoopPreheader());
319         else
320           DefiningAccess = cast<MemoryDef>(DefiningAccess)->getDefiningAccess();
321       }
322       MSSAU->createMemoryAccessInBB(NewInst, DefiningAccess,
323                                     NewInst->getParent(),
324                                     MemorySSA::BeforeTerminator);
325     }
326   }
327 
328   IRBuilder<> IRB(&BB);
329   Value *Cond = VMap[ToDuplicate[0]];
330   IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc,
331                    Direction ? &NormalSucc : &UnswitchedSucc);
332 }
333 
334 /// Rewrite the PHI nodes in an unswitched loop exit basic block.
335 ///
336 /// Requires that the loop exit and unswitched basic block are the same, and
337 /// that the exiting block was a unique predecessor of that block. Rewrites the
338 /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial
339 /// PHI nodes from the old preheader that now contains the unswitched
340 /// terminator.
rewritePHINodesForUnswitchedExitBlock(BasicBlock & UnswitchedBB,BasicBlock & OldExitingBB,BasicBlock & OldPH)341 static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB,
342                                                   BasicBlock &OldExitingBB,
343                                                   BasicBlock &OldPH) {
344   for (PHINode &PN : UnswitchedBB.phis()) {
345     // When the loop exit is directly unswitched we just need to update the
346     // incoming basic block. We loop to handle weird cases with repeated
347     // incoming blocks, but expect to typically only have one operand here.
348     for (auto i : seq<int>(0, PN.getNumOperands())) {
349       assert(PN.getIncomingBlock(i) == &OldExitingBB &&
350              "Found incoming block different from unique predecessor!");
351       PN.setIncomingBlock(i, &OldPH);
352     }
353   }
354 }
355 
356 /// Rewrite the PHI nodes in the loop exit basic block and the split off
357 /// unswitched block.
358 ///
359 /// Because the exit block remains an exit from the loop, this rewrites the
360 /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI
361 /// nodes into the unswitched basic block to select between the value in the
362 /// old preheader and the loop exit.
rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock & ExitBB,BasicBlock & UnswitchedBB,BasicBlock & OldExitingBB,BasicBlock & OldPH,bool FullUnswitch)363 static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB,
364                                                       BasicBlock &UnswitchedBB,
365                                                       BasicBlock &OldExitingBB,
366                                                       BasicBlock &OldPH,
367                                                       bool FullUnswitch) {
368   assert(&ExitBB != &UnswitchedBB &&
369          "Must have different loop exit and unswitched blocks!");
370   BasicBlock::iterator InsertPt = UnswitchedBB.begin();
371   for (PHINode &PN : ExitBB.phis()) {
372     auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2,
373                                   PN.getName() + ".split");
374     NewPN->insertBefore(InsertPt);
375 
376     // Walk backwards over the old PHI node's inputs to minimize the cost of
377     // removing each one. We have to do this weird loop manually so that we
378     // create the same number of new incoming edges in the new PHI as we expect
379     // each case-based edge to be included in the unswitched switch in some
380     // cases.
381     // FIXME: This is really, really gross. It would be much cleaner if LLVM
382     // allowed us to create a single entry for a predecessor block without
383     // having separate entries for each "edge" even though these edges are
384     // required to produce identical results.
385     for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) {
386       if (PN.getIncomingBlock(i) != &OldExitingBB)
387         continue;
388 
389       Value *Incoming = PN.getIncomingValue(i);
390       if (FullUnswitch)
391         // No more edge from the old exiting block to the exit block.
392         PN.removeIncomingValue(i);
393 
394       NewPN->addIncoming(Incoming, &OldPH);
395     }
396 
397     // Now replace the old PHI with the new one and wire the old one in as an
398     // input to the new one.
399     PN.replaceAllUsesWith(NewPN);
400     NewPN->addIncoming(&PN, &ExitBB);
401   }
402 }
403 
404 /// Hoist the current loop up to the innermost loop containing a remaining exit.
405 ///
406 /// Because we've removed an exit from the loop, we may have changed the set of
407 /// loops reachable and need to move the current loop up the loop nest or even
408 /// to an entirely separate nest.
hoistLoopToNewParent(Loop & L,BasicBlock & Preheader,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU,ScalarEvolution * SE)409 static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader,
410                                  DominatorTree &DT, LoopInfo &LI,
411                                  MemorySSAUpdater *MSSAU, ScalarEvolution *SE) {
412   // If the loop is already at the top level, we can't hoist it anywhere.
413   Loop *OldParentL = L.getParentLoop();
414   if (!OldParentL)
415     return;
416 
417   SmallVector<BasicBlock *, 4> Exits;
418   L.getExitBlocks(Exits);
419   Loop *NewParentL = nullptr;
420   for (auto *ExitBB : Exits)
421     if (Loop *ExitL = LI.getLoopFor(ExitBB))
422       if (!NewParentL || NewParentL->contains(ExitL))
423         NewParentL = ExitL;
424 
425   if (NewParentL == OldParentL)
426     return;
427 
428   // The new parent loop (if different) should always contain the old one.
429   if (NewParentL)
430     assert(NewParentL->contains(OldParentL) &&
431            "Can only hoist this loop up the nest!");
432 
433   // The preheader will need to move with the body of this loop. However,
434   // because it isn't in this loop we also need to update the primary loop map.
435   assert(OldParentL == LI.getLoopFor(&Preheader) &&
436          "Parent loop of this loop should contain this loop's preheader!");
437   LI.changeLoopFor(&Preheader, NewParentL);
438 
439   // Remove this loop from its old parent.
440   OldParentL->removeChildLoop(&L);
441 
442   // Add the loop either to the new parent or as a top-level loop.
443   if (NewParentL)
444     NewParentL->addChildLoop(&L);
445   else
446     LI.addTopLevelLoop(&L);
447 
448   // Remove this loops blocks from the old parent and every other loop up the
449   // nest until reaching the new parent. Also update all of these
450   // no-longer-containing loops to reflect the nesting change.
451   for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL;
452        OldContainingL = OldContainingL->getParentLoop()) {
453     llvm::erase_if(OldContainingL->getBlocksVector(),
454                    [&](const BasicBlock *BB) {
455                      return BB == &Preheader || L.contains(BB);
456                    });
457 
458     OldContainingL->getBlocksSet().erase(&Preheader);
459     for (BasicBlock *BB : L.blocks())
460       OldContainingL->getBlocksSet().erase(BB);
461 
462     // Because we just hoisted a loop out of this one, we have essentially
463     // created new exit paths from it. That means we need to form LCSSA PHI
464     // nodes for values used in the no-longer-nested loop.
465     formLCSSA(*OldContainingL, DT, &LI, SE);
466 
467     // We shouldn't need to form dedicated exits because the exit introduced
468     // here is the (just split by unswitching) preheader. However, after trivial
469     // unswitching it is possible to get new non-dedicated exits out of parent
470     // loop so let's conservatively form dedicated exit blocks and figure out
471     // if we can optimize later.
472     formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU,
473                             /*PreserveLCSSA*/ true);
474   }
475 }
476 
477 // Return the top-most loop containing ExitBB and having ExitBB as exiting block
478 // or the loop containing ExitBB, if there is no parent loop containing ExitBB
479 // as exiting block.
getTopMostExitingLoop(const BasicBlock * ExitBB,const LoopInfo & LI)480 static Loop *getTopMostExitingLoop(const BasicBlock *ExitBB,
481                                    const LoopInfo &LI) {
482   Loop *TopMost = LI.getLoopFor(ExitBB);
483   Loop *Current = TopMost;
484   while (Current) {
485     if (Current->isLoopExiting(ExitBB))
486       TopMost = Current;
487     Current = Current->getParentLoop();
488   }
489   return TopMost;
490 }
491 
492 /// Unswitch a trivial branch if the condition is loop invariant.
493 ///
494 /// This routine should only be called when loop code leading to the branch has
495 /// been validated as trivial (no side effects). This routine checks if the
496 /// condition is invariant and one of the successors is a loop exit. This
497 /// allows us to unswitch without duplicating the loop, making it trivial.
498 ///
499 /// If this routine fails to unswitch the branch it returns false.
500 ///
501 /// If the branch can be unswitched, this routine splits the preheader and
502 /// hoists the branch above that split. Preserves loop simplified form
503 /// (splitting the exit block as necessary). It simplifies the branch within
504 /// the loop to an unconditional branch but doesn't remove it entirely. Further
505 /// cleanup can be done with some simplifycfg like pass.
506 ///
507 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
508 /// invalidated by this.
unswitchTrivialBranch(Loop & L,BranchInst & BI,DominatorTree & DT,LoopInfo & LI,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)509 static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT,
510                                   LoopInfo &LI, ScalarEvolution *SE,
511                                   MemorySSAUpdater *MSSAU) {
512   assert(BI.isConditional() && "Can only unswitch a conditional branch!");
513   LLVM_DEBUG(dbgs() << "  Trying to unswitch branch: " << BI << "\n");
514 
515   // The loop invariant values that we want to unswitch.
516   TinyPtrVector<Value *> Invariants;
517 
518   // When true, we're fully unswitching the branch rather than just unswitching
519   // some input conditions to the branch.
520   bool FullUnswitch = false;
521 
522   Value *Cond = skipTrivialSelect(BI.getCondition());
523   if (L.isLoopInvariant(Cond)) {
524     Invariants.push_back(Cond);
525     FullUnswitch = true;
526   } else {
527     if (auto *CondInst = dyn_cast<Instruction>(Cond))
528       Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI);
529     if (Invariants.empty()) {
530       LLVM_DEBUG(dbgs() << "   Couldn't find invariant inputs!\n");
531       return false;
532     }
533   }
534 
535   // Check that one of the branch's successors exits, and which one.
536   bool ExitDirection = true;
537   int LoopExitSuccIdx = 0;
538   auto *LoopExitBB = BI.getSuccessor(0);
539   if (L.contains(LoopExitBB)) {
540     ExitDirection = false;
541     LoopExitSuccIdx = 1;
542     LoopExitBB = BI.getSuccessor(1);
543     if (L.contains(LoopExitBB)) {
544       LLVM_DEBUG(dbgs() << "   Branch doesn't exit the loop!\n");
545       return false;
546     }
547   }
548   auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx);
549   auto *ParentBB = BI.getParent();
550   if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) {
551     LLVM_DEBUG(dbgs() << "   Loop exit PHI's aren't loop-invariant!\n");
552     return false;
553   }
554 
555   // When unswitching only part of the branch's condition, we need the exit
556   // block to be reached directly from the partially unswitched input. This can
557   // be done when the exit block is along the true edge and the branch condition
558   // is a graph of `or` operations, or the exit block is along the false edge
559   // and the condition is a graph of `and` operations.
560   if (!FullUnswitch) {
561     if (ExitDirection ? !match(Cond, m_LogicalOr())
562                       : !match(Cond, m_LogicalAnd())) {
563       LLVM_DEBUG(dbgs() << "   Branch condition is in improper form for "
564                            "non-full unswitch!\n");
565       return false;
566     }
567   }
568 
569   LLVM_DEBUG({
570     dbgs() << "    unswitching trivial invariant conditions for: " << BI
571            << "\n";
572     for (Value *Invariant : Invariants) {
573       dbgs() << "      " << *Invariant << " == true";
574       if (Invariant != Invariants.back())
575         dbgs() << " ||";
576       dbgs() << "\n";
577     }
578   });
579 
580   // If we have scalar evolutions, we need to invalidate them including this
581   // loop, the loop containing the exit block and the topmost parent loop
582   // exiting via LoopExitBB.
583   if (SE) {
584     if (const Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI))
585       SE->forgetLoop(ExitL);
586     else
587       // Forget the entire nest as this exits the entire nest.
588       SE->forgetTopmostLoop(&L);
589     SE->forgetBlockAndLoopDispositions();
590   }
591 
592   if (MSSAU && VerifyMemorySSA)
593     MSSAU->getMemorySSA()->verifyMemorySSA();
594 
595   // Split the preheader, so that we know that there is a safe place to insert
596   // the conditional branch. We will change the preheader to have a conditional
597   // branch on LoopCond.
598   BasicBlock *OldPH = L.getLoopPreheader();
599   BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
600 
601   // Now that we have a place to insert the conditional branch, create a place
602   // to branch to: this is the exit block out of the loop that we are
603   // unswitching. We need to split this if there are other loop predecessors.
604   // Because the loop is in simplified form, *any* other predecessor is enough.
605   BasicBlock *UnswitchedBB;
606   if (FullUnswitch && LoopExitBB->getUniquePredecessor()) {
607     assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&
608            "A branch's parent isn't a predecessor!");
609     UnswitchedBB = LoopExitBB;
610   } else {
611     UnswitchedBB =
612         SplitBlock(LoopExitBB, LoopExitBB->begin(), &DT, &LI, MSSAU, "", false);
613   }
614 
615   if (MSSAU && VerifyMemorySSA)
616     MSSAU->getMemorySSA()->verifyMemorySSA();
617 
618   // Actually move the invariant uses into the unswitched position. If possible,
619   // we do this by moving the instructions, but when doing partial unswitching
620   // we do it by building a new merge of the values in the unswitched position.
621   OldPH->getTerminator()->eraseFromParent();
622   if (FullUnswitch) {
623     // If fully unswitching, we can use the existing branch instruction.
624     // Splice it into the old PH to gate reaching the new preheader and re-point
625     // its successors.
626     BI.moveBefore(*OldPH, OldPH->end());
627     BI.setCondition(Cond);
628     if (MSSAU) {
629       // Temporarily clone the terminator, to make MSSA update cheaper by
630       // separating "insert edge" updates from "remove edge" ones.
631       BI.clone()->insertInto(ParentBB, ParentBB->end());
632     } else {
633       // Create a new unconditional branch that will continue the loop as a new
634       // terminator.
635       Instruction *NewBI = BranchInst::Create(ContinueBB, ParentBB);
636       NewBI->setDebugLoc(BI.getDebugLoc());
637     }
638     BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB);
639     BI.setSuccessor(1 - LoopExitSuccIdx, NewPH);
640   } else {
641     // Only unswitching a subset of inputs to the condition, so we will need to
642     // build a new branch that merges the invariant inputs.
643     if (ExitDirection)
644       assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalOr()) &&
645              "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the "
646              "condition!");
647     else
648       assert(match(skipTrivialSelect(BI.getCondition()), m_LogicalAnd()) &&
649              "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the"
650              " condition!");
651     buildPartialUnswitchConditionalBranch(
652         *OldPH, Invariants, ExitDirection, *UnswitchedBB, *NewPH,
653         FreezeLoopUnswitchCond, OldPH->getTerminator(), nullptr, DT);
654   }
655 
656   // Update the dominator tree with the added edge.
657   DT.insertEdge(OldPH, UnswitchedBB);
658 
659   // After the dominator tree was updated with the added edge, update MemorySSA
660   // if available.
661   if (MSSAU) {
662     SmallVector<CFGUpdate, 1> Updates;
663     Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB});
664     MSSAU->applyInsertUpdates(Updates, DT);
665   }
666 
667   // Finish updating dominator tree and memory ssa for full unswitch.
668   if (FullUnswitch) {
669     if (MSSAU) {
670       Instruction *Term = ParentBB->getTerminator();
671       // Remove the cloned branch instruction and create unconditional branch
672       // now.
673       Instruction *NewBI = BranchInst::Create(ContinueBB, ParentBB);
674       NewBI->setDebugLoc(Term->getDebugLoc());
675       Term->eraseFromParent();
676       MSSAU->removeEdge(ParentBB, LoopExitBB);
677     }
678     DT.deleteEdge(ParentBB, LoopExitBB);
679   }
680 
681   if (MSSAU && VerifyMemorySSA)
682     MSSAU->getMemorySSA()->verifyMemorySSA();
683 
684   // Rewrite the relevant PHI nodes.
685   if (UnswitchedBB == LoopExitBB)
686     rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH);
687   else
688     rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB,
689                                               *ParentBB, *OldPH, FullUnswitch);
690 
691   // The constant we can replace all of our invariants with inside the loop
692   // body. If any of the invariants have a value other than this the loop won't
693   // be entered.
694   ConstantInt *Replacement = ExitDirection
695                                  ? ConstantInt::getFalse(BI.getContext())
696                                  : ConstantInt::getTrue(BI.getContext());
697 
698   // Since this is an i1 condition we can also trivially replace uses of it
699   // within the loop with a constant.
700   for (Value *Invariant : Invariants)
701     replaceLoopInvariantUses(L, Invariant, *Replacement);
702 
703   // If this was full unswitching, we may have changed the nesting relationship
704   // for this loop so hoist it to its correct parent if needed.
705   if (FullUnswitch)
706     hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
707 
708   if (MSSAU && VerifyMemorySSA)
709     MSSAU->getMemorySSA()->verifyMemorySSA();
710 
711   LLVM_DEBUG(dbgs() << "    done: unswitching trivial branch...\n");
712   ++NumTrivial;
713   ++NumBranches;
714   return true;
715 }
716 
717 /// Unswitch a trivial switch if the condition is loop invariant.
718 ///
719 /// This routine should only be called when loop code leading to the switch has
720 /// been validated as trivial (no side effects). This routine checks if the
721 /// condition is invariant and that at least one of the successors is a loop
722 /// exit. This allows us to unswitch without duplicating the loop, making it
723 /// trivial.
724 ///
725 /// If this routine fails to unswitch the switch it returns false.
726 ///
727 /// If the switch can be unswitched, this routine splits the preheader and
728 /// copies the switch above that split. If the default case is one of the
729 /// exiting cases, it copies the non-exiting cases and points them at the new
730 /// preheader. If the default case is not exiting, it copies the exiting cases
731 /// and points the default at the preheader. It preserves loop simplified form
732 /// (splitting the exit blocks as necessary). It simplifies the switch within
733 /// the loop by removing now-dead cases. If the default case is one of those
734 /// unswitched, it replaces its destination with a new basic block containing
735 /// only unreachable. Such basic blocks, while technically loop exits, are not
736 /// considered for unswitching so this is a stable transform and the same
737 /// switch will not be revisited. If after unswitching there is only a single
738 /// in-loop successor, the switch is further simplified to an unconditional
739 /// branch. Still more cleanup can be done with some simplifycfg like pass.
740 ///
741 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
742 /// invalidated by this.
unswitchTrivialSwitch(Loop & L,SwitchInst & SI,DominatorTree & DT,LoopInfo & LI,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)743 static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT,
744                                   LoopInfo &LI, ScalarEvolution *SE,
745                                   MemorySSAUpdater *MSSAU) {
746   LLVM_DEBUG(dbgs() << "  Trying to unswitch switch: " << SI << "\n");
747   Value *LoopCond = SI.getCondition();
748 
749   // If this isn't switching on an invariant condition, we can't unswitch it.
750   if (!L.isLoopInvariant(LoopCond))
751     return false;
752 
753   auto *ParentBB = SI.getParent();
754 
755   // The same check must be used both for the default and the exit cases. We
756   // should never leave edges from the switch instruction to a basic block that
757   // we are unswitching, hence the condition used to determine the default case
758   // needs to also be used to populate ExitCaseIndices, which is then used to
759   // remove cases from the switch.
760   auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) {
761     // BBToCheck is not an exit block if it is inside loop L.
762     if (L.contains(&BBToCheck))
763       return false;
764     // BBToCheck is not trivial to unswitch if its phis aren't loop invariant.
765     if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck))
766       return false;
767     // We do not unswitch a block that only has an unreachable statement, as
768     // it's possible this is a previously unswitched block. Only unswitch if
769     // either the terminator is not unreachable, or, if it is, it's not the only
770     // instruction in the block.
771     auto *TI = BBToCheck.getTerminator();
772     bool isUnreachable = isa<UnreachableInst>(TI);
773     return !isUnreachable ||
774            (isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI));
775   };
776 
777   SmallVector<int, 4> ExitCaseIndices;
778   for (auto Case : SI.cases())
779     if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor()))
780       ExitCaseIndices.push_back(Case.getCaseIndex());
781   BasicBlock *DefaultExitBB = nullptr;
782   SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight =
783       SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0);
784   if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) {
785     DefaultExitBB = SI.getDefaultDest();
786   } else if (ExitCaseIndices.empty())
787     return false;
788 
789   LLVM_DEBUG(dbgs() << "    unswitching trivial switch...\n");
790 
791   if (MSSAU && VerifyMemorySSA)
792     MSSAU->getMemorySSA()->verifyMemorySSA();
793 
794   // We may need to invalidate SCEVs for the outermost loop reached by any of
795   // the exits.
796   Loop *OuterL = &L;
797 
798   if (DefaultExitBB) {
799     // Check the loop containing this exit.
800     Loop *ExitL = getTopMostExitingLoop(DefaultExitBB, LI);
801     if (!ExitL || ExitL->contains(OuterL))
802       OuterL = ExitL;
803   }
804   for (unsigned Index : ExitCaseIndices) {
805     auto CaseI = SI.case_begin() + Index;
806     // Compute the outer loop from this exit.
807     Loop *ExitL = getTopMostExitingLoop(CaseI->getCaseSuccessor(), LI);
808     if (!ExitL || ExitL->contains(OuterL))
809       OuterL = ExitL;
810   }
811 
812   if (SE) {
813     if (OuterL)
814       SE->forgetLoop(OuterL);
815     else
816       SE->forgetTopmostLoop(&L);
817   }
818 
819   if (DefaultExitBB) {
820     // Clear out the default destination temporarily to allow accurate
821     // predecessor lists to be examined below.
822     SI.setDefaultDest(nullptr);
823   }
824 
825   // Store the exit cases into a separate data structure and remove them from
826   // the switch.
827   SmallVector<std::tuple<ConstantInt *, BasicBlock *,
828                          SwitchInstProfUpdateWrapper::CaseWeightOpt>,
829               4> ExitCases;
830   ExitCases.reserve(ExitCaseIndices.size());
831   SwitchInstProfUpdateWrapper SIW(SI);
832   // We walk the case indices backwards so that we remove the last case first
833   // and don't disrupt the earlier indices.
834   for (unsigned Index : reverse(ExitCaseIndices)) {
835     auto CaseI = SI.case_begin() + Index;
836     // Save the value of this case.
837     auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex());
838     ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W);
839     // Delete the unswitched cases.
840     SIW.removeCase(CaseI);
841   }
842 
843   // Check if after this all of the remaining cases point at the same
844   // successor.
845   BasicBlock *CommonSuccBB = nullptr;
846   if (SI.getNumCases() > 0 &&
847       all_of(drop_begin(SI.cases()), [&SI](const SwitchInst::CaseHandle &Case) {
848         return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor();
849       }))
850     CommonSuccBB = SI.case_begin()->getCaseSuccessor();
851   if (!DefaultExitBB) {
852     // If we're not unswitching the default, we need it to match any cases to
853     // have a common successor or if we have no cases it is the common
854     // successor.
855     if (SI.getNumCases() == 0)
856       CommonSuccBB = SI.getDefaultDest();
857     else if (SI.getDefaultDest() != CommonSuccBB)
858       CommonSuccBB = nullptr;
859   }
860 
861   // Split the preheader, so that we know that there is a safe place to insert
862   // the switch.
863   BasicBlock *OldPH = L.getLoopPreheader();
864   BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU);
865   OldPH->getTerminator()->eraseFromParent();
866 
867   // Now add the unswitched switch. This new switch instruction inherits the
868   // debug location of the old switch, because it semantically replace the old
869   // one.
870   auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH);
871   NewSI->setDebugLoc(SIW->getDebugLoc());
872   SwitchInstProfUpdateWrapper NewSIW(*NewSI);
873 
874   // Rewrite the IR for the unswitched basic blocks. This requires two steps.
875   // First, we split any exit blocks with remaining in-loop predecessors. Then
876   // we update the PHIs in one of two ways depending on if there was a split.
877   // We walk in reverse so that we split in the same order as the cases
878   // appeared. This is purely for convenience of reading the resulting IR, but
879   // it doesn't cost anything really.
880   SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs;
881   SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap;
882   // Handle the default exit if necessary.
883   // FIXME: It'd be great if we could merge this with the loop below but LLVM's
884   // ranges aren't quite powerful enough yet.
885   if (DefaultExitBB) {
886     if (pred_empty(DefaultExitBB)) {
887       UnswitchedExitBBs.insert(DefaultExitBB);
888       rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH);
889     } else {
890       auto *SplitBB =
891           SplitBlock(DefaultExitBB, DefaultExitBB->begin(), &DT, &LI, MSSAU);
892       rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB,
893                                                 *ParentBB, *OldPH,
894                                                 /*FullUnswitch*/ true);
895       DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB;
896     }
897   }
898   // Note that we must use a reference in the for loop so that we update the
899   // container.
900   for (auto &ExitCase : reverse(ExitCases)) {
901     // Grab a reference to the exit block in the pair so that we can update it.
902     BasicBlock *ExitBB = std::get<1>(ExitCase);
903 
904     // If this case is the last edge into the exit block, we can simply reuse it
905     // as it will no longer be a loop exit. No mapping necessary.
906     if (pred_empty(ExitBB)) {
907       // Only rewrite once.
908       if (UnswitchedExitBBs.insert(ExitBB).second)
909         rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH);
910       continue;
911     }
912 
913     // Otherwise we need to split the exit block so that we retain an exit
914     // block from the loop and a target for the unswitched condition.
915     BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB];
916     if (!SplitExitBB) {
917       // If this is the first time we see this, do the split and remember it.
918       SplitExitBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU);
919       rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB,
920                                                 *ParentBB, *OldPH,
921                                                 /*FullUnswitch*/ true);
922     }
923     // Update the case pair to point to the split block.
924     std::get<1>(ExitCase) = SplitExitBB;
925   }
926 
927   // Now add the unswitched cases. We do this in reverse order as we built them
928   // in reverse order.
929   for (auto &ExitCase : reverse(ExitCases)) {
930     ConstantInt *CaseVal = std::get<0>(ExitCase);
931     BasicBlock *UnswitchedBB = std::get<1>(ExitCase);
932 
933     NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase));
934   }
935 
936   // If the default was unswitched, re-point it and add explicit cases for
937   // entering the loop.
938   if (DefaultExitBB) {
939     NewSIW->setDefaultDest(DefaultExitBB);
940     NewSIW.setSuccessorWeight(0, DefaultCaseWeight);
941 
942     // We removed all the exit cases, so we just copy the cases to the
943     // unswitched switch.
944     for (const auto &Case : SI.cases())
945       NewSIW.addCase(Case.getCaseValue(), NewPH,
946                      SIW.getSuccessorWeight(Case.getSuccessorIndex()));
947   } else if (DefaultCaseWeight) {
948     // We have to set branch weight of the default case.
949     uint64_t SW = *DefaultCaseWeight;
950     for (const auto &Case : SI.cases()) {
951       auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex());
952       assert(W &&
953              "case weight must be defined as default case weight is defined");
954       SW += *W;
955     }
956     NewSIW.setSuccessorWeight(0, SW);
957   }
958 
959   // If we ended up with a common successor for every path through the switch
960   // after unswitching, rewrite it to an unconditional branch to make it easy
961   // to recognize. Otherwise we potentially have to recognize the default case
962   // pointing at unreachable and other complexity.
963   if (CommonSuccBB) {
964     BasicBlock *BB = SI.getParent();
965     // We may have had multiple edges to this common successor block, so remove
966     // them as predecessors. We skip the first one, either the default or the
967     // actual first case.
968     bool SkippedFirst = DefaultExitBB == nullptr;
969     for (auto Case : SI.cases()) {
970       assert(Case.getCaseSuccessor() == CommonSuccBB &&
971              "Non-common successor!");
972       (void)Case;
973       if (!SkippedFirst) {
974         SkippedFirst = true;
975         continue;
976       }
977       CommonSuccBB->removePredecessor(BB,
978                                       /*KeepOneInputPHIs*/ true);
979     }
980     // Now nuke the switch and replace it with a direct branch.
981     Instruction *NewBI = BranchInst::Create(CommonSuccBB, BB);
982     NewBI->setDebugLoc(SIW->getDebugLoc());
983     SIW.eraseFromParent();
984   } else if (DefaultExitBB) {
985     assert(SI.getNumCases() > 0 &&
986            "If we had no cases we'd have a common successor!");
987     // Move the last case to the default successor. This is valid as if the
988     // default got unswitched it cannot be reached. This has the advantage of
989     // being simple and keeping the number of edges from this switch to
990     // successors the same, and avoiding any PHI update complexity.
991     auto LastCaseI = std::prev(SI.case_end());
992 
993     SI.setDefaultDest(LastCaseI->getCaseSuccessor());
994     SIW.setSuccessorWeight(
995         0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex()));
996     SIW.removeCase(LastCaseI);
997   }
998 
999   // Walk the unswitched exit blocks and the unswitched split blocks and update
1000   // the dominator tree based on the CFG edits. While we are walking unordered
1001   // containers here, the API for applyUpdates takes an unordered list of
1002   // updates and requires them to not contain duplicates.
1003   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
1004   for (auto *UnswitchedExitBB : UnswitchedExitBBs) {
1005     DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB});
1006     DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB});
1007   }
1008   for (auto SplitUnswitchedPair : SplitExitBBMap) {
1009     DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first});
1010     DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second});
1011   }
1012 
1013   if (MSSAU) {
1014     MSSAU->applyUpdates(DTUpdates, DT, /*UpdateDT=*/true);
1015     if (VerifyMemorySSA)
1016       MSSAU->getMemorySSA()->verifyMemorySSA();
1017   } else {
1018     DT.applyUpdates(DTUpdates);
1019   }
1020 
1021   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1022 
1023   // We may have changed the nesting relationship for this loop so hoist it to
1024   // its correct parent if needed.
1025   hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE);
1026 
1027   if (MSSAU && VerifyMemorySSA)
1028     MSSAU->getMemorySSA()->verifyMemorySSA();
1029 
1030   ++NumTrivial;
1031   ++NumSwitches;
1032   LLVM_DEBUG(dbgs() << "    done: unswitching trivial switch...\n");
1033   return true;
1034 }
1035 
1036 /// This routine scans the loop to find a branch or switch which occurs before
1037 /// any side effects occur. These can potentially be unswitched without
1038 /// duplicating the loop. If a branch or switch is successfully unswitched the
1039 /// scanning continues to see if subsequent branches or switches have become
1040 /// trivial. Once all trivial candidates have been unswitched, this routine
1041 /// returns.
1042 ///
1043 /// The return value indicates whether anything was unswitched (and therefore
1044 /// changed).
1045 ///
1046 /// If `SE` is not null, it will be updated based on the potential loop SCEVs
1047 /// invalidated by this.
unswitchAllTrivialConditions(Loop & L,DominatorTree & DT,LoopInfo & LI,ScalarEvolution * SE,MemorySSAUpdater * MSSAU)1048 static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT,
1049                                          LoopInfo &LI, ScalarEvolution *SE,
1050                                          MemorySSAUpdater *MSSAU) {
1051   bool Changed = false;
1052 
1053   // If loop header has only one reachable successor we should keep looking for
1054   // trivial condition candidates in the successor as well. An alternative is
1055   // to constant fold conditions and merge successors into loop header (then we
1056   // only need to check header's terminator). The reason for not doing this in
1057   // LoopUnswitch pass is that it could potentially break LoopPassManager's
1058   // invariants. Folding dead branches could either eliminate the current loop
1059   // or make other loops unreachable. LCSSA form might also not be preserved
1060   // after deleting branches. The following code keeps traversing loop header's
1061   // successors until it finds the trivial condition candidate (condition that
1062   // is not a constant). Since unswitching generates branches with constant
1063   // conditions, this scenario could be very common in practice.
1064   BasicBlock *CurrentBB = L.getHeader();
1065   SmallPtrSet<BasicBlock *, 8> Visited;
1066   Visited.insert(CurrentBB);
1067   do {
1068     // Check if there are any side-effecting instructions (e.g. stores, calls,
1069     // volatile loads) in the part of the loop that the code *would* execute
1070     // without unswitching.
1071     if (MSSAU) // Possible early exit with MSSA
1072       if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB))
1073         if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end()))
1074           return Changed;
1075     if (llvm::any_of(*CurrentBB,
1076                      [](Instruction &I) { return I.mayHaveSideEffects(); }))
1077       return Changed;
1078 
1079     Instruction *CurrentTerm = CurrentBB->getTerminator();
1080 
1081     if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
1082       // Don't bother trying to unswitch past a switch with a constant
1083       // condition. This should be removed prior to running this pass by
1084       // simplifycfg.
1085       if (isa<Constant>(SI->getCondition()))
1086         return Changed;
1087 
1088       if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU))
1089         // Couldn't unswitch this one so we're done.
1090         return Changed;
1091 
1092       // Mark that we managed to unswitch something.
1093       Changed = true;
1094 
1095       // If unswitching turned the terminator into an unconditional branch then
1096       // we can continue. The unswitching logic specifically works to fold any
1097       // cases it can into an unconditional branch to make it easier to
1098       // recognize here.
1099       auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator());
1100       if (!BI || BI->isConditional())
1101         return Changed;
1102 
1103       CurrentBB = BI->getSuccessor(0);
1104       continue;
1105     }
1106 
1107     auto *BI = dyn_cast<BranchInst>(CurrentTerm);
1108     if (!BI)
1109       // We do not understand other terminator instructions.
1110       return Changed;
1111 
1112     // Don't bother trying to unswitch past an unconditional branch or a branch
1113     // with a constant value. These should be removed by simplifycfg prior to
1114     // running this pass.
1115     if (!BI->isConditional() ||
1116         isa<Constant>(skipTrivialSelect(BI->getCondition())))
1117       return Changed;
1118 
1119     // Found a trivial condition candidate: non-foldable conditional branch. If
1120     // we fail to unswitch this, we can't do anything else that is trivial.
1121     if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU))
1122       return Changed;
1123 
1124     // Mark that we managed to unswitch something.
1125     Changed = true;
1126 
1127     // If we only unswitched some of the conditions feeding the branch, we won't
1128     // have collapsed it to a single successor.
1129     BI = cast<BranchInst>(CurrentBB->getTerminator());
1130     if (BI->isConditional())
1131       return Changed;
1132 
1133     // Follow the newly unconditional branch into its successor.
1134     CurrentBB = BI->getSuccessor(0);
1135 
1136     // When continuing, if we exit the loop or reach a previous visited block,
1137     // then we can not reach any trivial condition candidates (unfoldable
1138     // branch instructions or switch instructions) and no unswitch can happen.
1139   } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second);
1140 
1141   return Changed;
1142 }
1143 
1144 /// Build the cloned blocks for an unswitched copy of the given loop.
1145 ///
1146 /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and
1147 /// after the split block (`SplitBB`) that will be used to select between the
1148 /// cloned and original loop.
1149 ///
1150 /// This routine handles cloning all of the necessary loop blocks and exit
1151 /// blocks including rewriting their instructions and the relevant PHI nodes.
1152 /// Any loop blocks or exit blocks which are dominated by a different successor
1153 /// than the one for this clone of the loop blocks can be trivially skipped. We
1154 /// use the `DominatingSucc` map to determine whether a block satisfies that
1155 /// property with a simple map lookup.
1156 ///
1157 /// It also correctly creates the unconditional branch in the cloned
1158 /// unswitched parent block to only point at the unswitched successor.
1159 ///
1160 /// This does not handle most of the necessary updates to `LoopInfo`. Only exit
1161 /// block splitting is correctly reflected in `LoopInfo`, essentially all of
1162 /// the cloned blocks (and their loops) are left without full `LoopInfo`
1163 /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned
1164 /// blocks to them but doesn't create the cloned `DominatorTree` structure and
1165 /// instead the caller must recompute an accurate DT. It *does* correctly
1166 /// update the `AssumptionCache` provided in `AC`.
buildClonedLoopBlocks(Loop & L,BasicBlock * LoopPH,BasicBlock * SplitBB,ArrayRef<BasicBlock * > ExitBlocks,BasicBlock * ParentBB,BasicBlock * UnswitchedSuccBB,BasicBlock * ContinueSuccBB,const SmallDenseMap<BasicBlock *,BasicBlock *,16> & DominatingSucc,ValueToValueMapTy & VMap,SmallVectorImpl<DominatorTree::UpdateType> & DTUpdates,AssumptionCache & AC,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU,ScalarEvolution * SE)1167 static BasicBlock *buildClonedLoopBlocks(
1168     Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB,
1169     ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB,
1170     BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB,
1171     const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc,
1172     ValueToValueMapTy &VMap,
1173     SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC,
1174     DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU,
1175     ScalarEvolution *SE) {
1176   SmallVector<BasicBlock *, 4> NewBlocks;
1177   NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size());
1178 
1179   // We will need to clone a bunch of blocks, wrap up the clone operation in
1180   // a helper.
1181   auto CloneBlock = [&](BasicBlock *OldBB) {
1182     // Clone the basic block and insert it before the new preheader.
1183     BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent());
1184     NewBB->moveBefore(LoopPH);
1185 
1186     // Record this block and the mapping.
1187     NewBlocks.push_back(NewBB);
1188     VMap[OldBB] = NewBB;
1189 
1190     return NewBB;
1191   };
1192 
1193   // We skip cloning blocks when they have a dominating succ that is not the
1194   // succ we are cloning for.
1195   auto SkipBlock = [&](BasicBlock *BB) {
1196     auto It = DominatingSucc.find(BB);
1197     return It != DominatingSucc.end() && It->second != UnswitchedSuccBB;
1198   };
1199 
1200   // First, clone the preheader.
1201   auto *ClonedPH = CloneBlock(LoopPH);
1202 
1203   // Then clone all the loop blocks, skipping the ones that aren't necessary.
1204   for (auto *LoopBB : L.blocks())
1205     if (!SkipBlock(LoopBB))
1206       CloneBlock(LoopBB);
1207 
1208   // Split all the loop exit edges so that when we clone the exit blocks, if
1209   // any of the exit blocks are *also* a preheader for some other loop, we
1210   // don't create multiple predecessors entering the loop header.
1211   for (auto *ExitBB : ExitBlocks) {
1212     if (SkipBlock(ExitBB))
1213       continue;
1214 
1215     // When we are going to clone an exit, we don't need to clone all the
1216     // instructions in the exit block and we want to ensure we have an easy
1217     // place to merge the CFG, so split the exit first. This is always safe to
1218     // do because there cannot be any non-loop predecessors of a loop exit in
1219     // loop simplified form.
1220     auto *MergeBB = SplitBlock(ExitBB, ExitBB->begin(), &DT, &LI, MSSAU);
1221 
1222     // Rearrange the names to make it easier to write test cases by having the
1223     // exit block carry the suffix rather than the merge block carrying the
1224     // suffix.
1225     MergeBB->takeName(ExitBB);
1226     ExitBB->setName(Twine(MergeBB->getName()) + ".split");
1227 
1228     // Now clone the original exit block.
1229     auto *ClonedExitBB = CloneBlock(ExitBB);
1230     assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&
1231            "Exit block should have been split to have one successor!");
1232     assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&
1233            "Cloned exit block has the wrong successor!");
1234 
1235     // Remap any cloned instructions and create a merge phi node for them.
1236     for (auto ZippedInsts : llvm::zip_first(
1237              llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())),
1238              llvm::make_range(ClonedExitBB->begin(),
1239                               std::prev(ClonedExitBB->end())))) {
1240       Instruction &I = std::get<0>(ZippedInsts);
1241       Instruction &ClonedI = std::get<1>(ZippedInsts);
1242 
1243       // The only instructions in the exit block should be PHI nodes and
1244       // potentially a landing pad.
1245       assert(
1246           (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&
1247           "Bad instruction in exit block!");
1248       // We should have a value map between the instruction and its clone.
1249       assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!");
1250 
1251       // Forget SCEVs based on exit phis in case SCEV looked through the phi.
1252       if (SE && isa<PHINode>(I))
1253         SE->forgetValue(&I);
1254 
1255       BasicBlock::iterator InsertPt = MergeBB->getFirstInsertionPt();
1256 
1257       auto *MergePN =
1258           PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi");
1259       MergePN->insertBefore(InsertPt);
1260       MergePN->setDebugLoc(InsertPt->getDebugLoc());
1261       I.replaceAllUsesWith(MergePN);
1262       MergePN->addIncoming(&I, ExitBB);
1263       MergePN->addIncoming(&ClonedI, ClonedExitBB);
1264     }
1265   }
1266 
1267   // Rewrite the instructions in the cloned blocks to refer to the instructions
1268   // in the cloned blocks. We have to do this as a second pass so that we have
1269   // everything available. Also, we have inserted new instructions which may
1270   // include assume intrinsics, so we update the assumption cache while
1271   // processing this.
1272   Module *M = ClonedPH->getParent()->getParent();
1273   for (auto *ClonedBB : NewBlocks)
1274     for (Instruction &I : *ClonedBB) {
1275       RemapDbgRecordRange(M, I.getDbgRecordRange(), VMap,
1276                           RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1277       RemapInstruction(&I, VMap,
1278                        RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1279       if (auto *II = dyn_cast<AssumeInst>(&I))
1280         AC.registerAssumption(II);
1281     }
1282 
1283   // Update any PHI nodes in the cloned successors of the skipped blocks to not
1284   // have spurious incoming values.
1285   for (auto *LoopBB : L.blocks())
1286     if (SkipBlock(LoopBB))
1287       for (auto *SuccBB : successors(LoopBB))
1288         if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)))
1289           for (PHINode &PN : ClonedSuccBB->phis())
1290             PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false);
1291 
1292   // Remove the cloned parent as a predecessor of any successor we ended up
1293   // cloning other than the unswitched one.
1294   auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB));
1295   for (auto *SuccBB : successors(ParentBB)) {
1296     if (SuccBB == UnswitchedSuccBB)
1297       continue;
1298 
1299     auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB));
1300     if (!ClonedSuccBB)
1301       continue;
1302 
1303     ClonedSuccBB->removePredecessor(ClonedParentBB,
1304                                     /*KeepOneInputPHIs*/ true);
1305   }
1306 
1307   // Replace the cloned branch with an unconditional branch to the cloned
1308   // unswitched successor.
1309   auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB));
1310   Instruction *ClonedTerminator = ClonedParentBB->getTerminator();
1311   // Trivial Simplification. If Terminator is a conditional branch and
1312   // condition becomes dead - erase it.
1313   Value *ClonedConditionToErase = nullptr;
1314   if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator))
1315     ClonedConditionToErase = BI->getCondition();
1316   else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator))
1317     ClonedConditionToErase = SI->getCondition();
1318 
1319   Instruction *BI = BranchInst::Create(ClonedSuccBB, ClonedParentBB);
1320   BI->setDebugLoc(ClonedTerminator->getDebugLoc());
1321   ClonedTerminator->eraseFromParent();
1322 
1323   if (ClonedConditionToErase)
1324     RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr,
1325                                                MSSAU);
1326 
1327   // If there are duplicate entries in the PHI nodes because of multiple edges
1328   // to the unswitched successor, we need to nuke all but one as we replaced it
1329   // with a direct branch.
1330   for (PHINode &PN : ClonedSuccBB->phis()) {
1331     bool Found = false;
1332     // Loop over the incoming operands backwards so we can easily delete as we
1333     // go without invalidating the index.
1334     for (int i = PN.getNumOperands() - 1; i >= 0; --i) {
1335       if (PN.getIncomingBlock(i) != ClonedParentBB)
1336         continue;
1337       if (!Found) {
1338         Found = true;
1339         continue;
1340       }
1341       PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false);
1342     }
1343   }
1344 
1345   // Record the domtree updates for the new blocks.
1346   SmallPtrSet<BasicBlock *, 4> SuccSet;
1347   for (auto *ClonedBB : NewBlocks) {
1348     for (auto *SuccBB : successors(ClonedBB))
1349       if (SuccSet.insert(SuccBB).second)
1350         DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB});
1351     SuccSet.clear();
1352   }
1353 
1354   return ClonedPH;
1355 }
1356 
1357 /// Recursively clone the specified loop and all of its children.
1358 ///
1359 /// The target parent loop for the clone should be provided, or can be null if
1360 /// the clone is a top-level loop. While cloning, all the blocks are mapped
1361 /// with the provided value map. The entire original loop must be present in
1362 /// the value map. The cloned loop is returned.
cloneLoopNest(Loop & OrigRootL,Loop * RootParentL,const ValueToValueMapTy & VMap,LoopInfo & LI)1363 static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL,
1364                            const ValueToValueMapTy &VMap, LoopInfo &LI) {
1365   auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) {
1366     assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!");
1367     ClonedL.reserveBlocks(OrigL.getNumBlocks());
1368     for (auto *BB : OrigL.blocks()) {
1369       auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB));
1370       ClonedL.addBlockEntry(ClonedBB);
1371       if (LI.getLoopFor(BB) == &OrigL)
1372         LI.changeLoopFor(ClonedBB, &ClonedL);
1373     }
1374   };
1375 
1376   // We specially handle the first loop because it may get cloned into
1377   // a different parent and because we most commonly are cloning leaf loops.
1378   Loop *ClonedRootL = LI.AllocateLoop();
1379   if (RootParentL)
1380     RootParentL->addChildLoop(ClonedRootL);
1381   else
1382     LI.addTopLevelLoop(ClonedRootL);
1383   AddClonedBlocksToLoop(OrigRootL, *ClonedRootL);
1384 
1385   if (OrigRootL.isInnermost())
1386     return ClonedRootL;
1387 
1388   // If we have a nest, we can quickly clone the entire loop nest using an
1389   // iterative approach because it is a tree. We keep the cloned parent in the
1390   // data structure to avoid repeatedly querying through a map to find it.
1391   SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone;
1392   // Build up the loops to clone in reverse order as we'll clone them from the
1393   // back.
1394   for (Loop *ChildL : llvm::reverse(OrigRootL))
1395     LoopsToClone.push_back({ClonedRootL, ChildL});
1396   do {
1397     Loop *ClonedParentL, *L;
1398     std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val();
1399     Loop *ClonedL = LI.AllocateLoop();
1400     ClonedParentL->addChildLoop(ClonedL);
1401     AddClonedBlocksToLoop(*L, *ClonedL);
1402     for (Loop *ChildL : llvm::reverse(*L))
1403       LoopsToClone.push_back({ClonedL, ChildL});
1404   } while (!LoopsToClone.empty());
1405 
1406   return ClonedRootL;
1407 }
1408 
1409 /// Build the cloned loops of an original loop from unswitching.
1410 ///
1411 /// Because unswitching simplifies the CFG of the loop, this isn't a trivial
1412 /// operation. We need to re-verify that there even is a loop (as the backedge
1413 /// may not have been cloned), and even if there are remaining backedges the
1414 /// backedge set may be different. However, we know that each child loop is
1415 /// undisturbed, we only need to find where to place each child loop within
1416 /// either any parent loop or within a cloned version of the original loop.
1417 ///
1418 /// Because child loops may end up cloned outside of any cloned version of the
1419 /// original loop, multiple cloned sibling loops may be created. All of them
1420 /// are returned so that the newly introduced loop nest roots can be
1421 /// identified.
buildClonedLoops(Loop & OrigL,ArrayRef<BasicBlock * > ExitBlocks,const ValueToValueMapTy & VMap,LoopInfo & LI,SmallVectorImpl<Loop * > & NonChildClonedLoops)1422 static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks,
1423                              const ValueToValueMapTy &VMap, LoopInfo &LI,
1424                              SmallVectorImpl<Loop *> &NonChildClonedLoops) {
1425   Loop *ClonedL = nullptr;
1426 
1427   auto *OrigPH = OrigL.getLoopPreheader();
1428   auto *OrigHeader = OrigL.getHeader();
1429 
1430   auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH));
1431   auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader));
1432 
1433   // We need to know the loops of the cloned exit blocks to even compute the
1434   // accurate parent loop. If we only clone exits to some parent of the
1435   // original parent, we want to clone into that outer loop. We also keep track
1436   // of the loops that our cloned exit blocks participate in.
1437   Loop *ParentL = nullptr;
1438   SmallVector<BasicBlock *, 4> ClonedExitsInLoops;
1439   SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap;
1440   ClonedExitsInLoops.reserve(ExitBlocks.size());
1441   for (auto *ExitBB : ExitBlocks)
1442     if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB)))
1443       if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1444         ExitLoopMap[ClonedExitBB] = ExitL;
1445         ClonedExitsInLoops.push_back(ClonedExitBB);
1446         if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1447           ParentL = ExitL;
1448       }
1449   assert((!ParentL || ParentL == OrigL.getParentLoop() ||
1450           ParentL->contains(OrigL.getParentLoop())) &&
1451          "The computed parent loop should always contain (or be) the parent of "
1452          "the original loop.");
1453 
1454   // We build the set of blocks dominated by the cloned header from the set of
1455   // cloned blocks out of the original loop. While not all of these will
1456   // necessarily be in the cloned loop, it is enough to establish that they
1457   // aren't in unreachable cycles, etc.
1458   SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks;
1459   for (auto *BB : OrigL.blocks())
1460     if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)))
1461       ClonedLoopBlocks.insert(ClonedBB);
1462 
1463   // Rebuild the set of blocks that will end up in the cloned loop. We may have
1464   // skipped cloning some region of this loop which can in turn skip some of
1465   // the backedges so we have to rebuild the blocks in the loop based on the
1466   // backedges that remain after cloning.
1467   SmallVector<BasicBlock *, 16> Worklist;
1468   SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop;
1469   for (auto *Pred : predecessors(ClonedHeader)) {
1470     // The only possible non-loop header predecessor is the preheader because
1471     // we know we cloned the loop in simplified form.
1472     if (Pred == ClonedPH)
1473       continue;
1474 
1475     // Because the loop was in simplified form, the only non-loop predecessor
1476     // should be the preheader.
1477     assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "
1478                                            "header other than the preheader "
1479                                            "that is not part of the loop!");
1480 
1481     // Insert this block into the loop set and on the first visit (and if it
1482     // isn't the header we're currently walking) put it into the worklist to
1483     // recurse through.
1484     if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader)
1485       Worklist.push_back(Pred);
1486   }
1487 
1488   // If we had any backedges then there *is* a cloned loop. Put the header into
1489   // the loop set and then walk the worklist backwards to find all the blocks
1490   // that remain within the loop after cloning.
1491   if (!BlocksInClonedLoop.empty()) {
1492     BlocksInClonedLoop.insert(ClonedHeader);
1493 
1494     while (!Worklist.empty()) {
1495       BasicBlock *BB = Worklist.pop_back_val();
1496       assert(BlocksInClonedLoop.count(BB) &&
1497              "Didn't put block into the loop set!");
1498 
1499       // Insert any predecessors that are in the possible set into the cloned
1500       // set, and if the insert is successful, add them to the worklist. Note
1501       // that we filter on the blocks that are definitely reachable via the
1502       // backedge to the loop header so we may prune out dead code within the
1503       // cloned loop.
1504       for (auto *Pred : predecessors(BB))
1505         if (ClonedLoopBlocks.count(Pred) &&
1506             BlocksInClonedLoop.insert(Pred).second)
1507           Worklist.push_back(Pred);
1508     }
1509 
1510     ClonedL = LI.AllocateLoop();
1511     if (ParentL) {
1512       ParentL->addBasicBlockToLoop(ClonedPH, LI);
1513       ParentL->addChildLoop(ClonedL);
1514     } else {
1515       LI.addTopLevelLoop(ClonedL);
1516     }
1517     NonChildClonedLoops.push_back(ClonedL);
1518 
1519     ClonedL->reserveBlocks(BlocksInClonedLoop.size());
1520     // We don't want to just add the cloned loop blocks based on how we
1521     // discovered them. The original order of blocks was carefully built in
1522     // a way that doesn't rely on predecessor ordering. Rather than re-invent
1523     // that logic, we just re-walk the original blocks (and those of the child
1524     // loops) and filter them as we add them into the cloned loop.
1525     for (auto *BB : OrigL.blocks()) {
1526       auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB));
1527       if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB))
1528         continue;
1529 
1530       // Directly add the blocks that are only in this loop.
1531       if (LI.getLoopFor(BB) == &OrigL) {
1532         ClonedL->addBasicBlockToLoop(ClonedBB, LI);
1533         continue;
1534       }
1535 
1536       // We want to manually add it to this loop and parents.
1537       // Registering it with LoopInfo will happen when we clone the top
1538       // loop for this block.
1539       for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop())
1540         PL->addBlockEntry(ClonedBB);
1541     }
1542 
1543     // Now add each child loop whose header remains within the cloned loop. All
1544     // of the blocks within the loop must satisfy the same constraints as the
1545     // header so once we pass the header checks we can just clone the entire
1546     // child loop nest.
1547     for (Loop *ChildL : OrigL) {
1548       auto *ClonedChildHeader =
1549           cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1550       if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader))
1551         continue;
1552 
1553 #ifndef NDEBUG
1554       // We should never have a cloned child loop header but fail to have
1555       // all of the blocks for that child loop.
1556       for (auto *ChildLoopBB : ChildL->blocks())
1557         assert(BlocksInClonedLoop.count(
1558                    cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&
1559                "Child cloned loop has a header within the cloned outer "
1560                "loop but not all of its blocks!");
1561 #endif
1562 
1563       cloneLoopNest(*ChildL, ClonedL, VMap, LI);
1564     }
1565   }
1566 
1567   // Now that we've handled all the components of the original loop that were
1568   // cloned into a new loop, we still need to handle anything from the original
1569   // loop that wasn't in a cloned loop.
1570 
1571   // Figure out what blocks are left to place within any loop nest containing
1572   // the unswitched loop. If we never formed a loop, the cloned PH is one of
1573   // them.
1574   SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet;
1575   if (BlocksInClonedLoop.empty())
1576     UnloopedBlockSet.insert(ClonedPH);
1577   for (auto *ClonedBB : ClonedLoopBlocks)
1578     if (!BlocksInClonedLoop.count(ClonedBB))
1579       UnloopedBlockSet.insert(ClonedBB);
1580 
1581   // Copy the cloned exits and sort them in ascending loop depth, we'll work
1582   // backwards across these to process them inside out. The order shouldn't
1583   // matter as we're just trying to build up the map from inside-out; we use
1584   // the map in a more stably ordered way below.
1585   auto OrderedClonedExitsInLoops = ClonedExitsInLoops;
1586   llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1587     return ExitLoopMap.lookup(LHS)->getLoopDepth() <
1588            ExitLoopMap.lookup(RHS)->getLoopDepth();
1589   });
1590 
1591   // Populate the existing ExitLoopMap with everything reachable from each
1592   // exit, starting from the inner most exit.
1593   while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) {
1594     assert(Worklist.empty() && "Didn't clear worklist!");
1595 
1596     BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val();
1597     Loop *ExitL = ExitLoopMap.lookup(ExitBB);
1598 
1599     // Walk the CFG back until we hit the cloned PH adding everything reachable
1600     // and in the unlooped set to this exit block's loop.
1601     Worklist.push_back(ExitBB);
1602     do {
1603       BasicBlock *BB = Worklist.pop_back_val();
1604       // We can stop recursing at the cloned preheader (if we get there).
1605       if (BB == ClonedPH)
1606         continue;
1607 
1608       for (BasicBlock *PredBB : predecessors(BB)) {
1609         // If this pred has already been moved to our set or is part of some
1610         // (inner) loop, no update needed.
1611         if (!UnloopedBlockSet.erase(PredBB)) {
1612           assert(
1613               (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&
1614               "Predecessor not mapped to a loop!");
1615           continue;
1616         }
1617 
1618         // We just insert into the loop set here. We'll add these blocks to the
1619         // exit loop after we build up the set in an order that doesn't rely on
1620         // predecessor order (which in turn relies on use list order).
1621         bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second;
1622         (void)Inserted;
1623         assert(Inserted && "Should only visit an unlooped block once!");
1624 
1625         // And recurse through to its predecessors.
1626         Worklist.push_back(PredBB);
1627       }
1628     } while (!Worklist.empty());
1629   }
1630 
1631   // Now that the ExitLoopMap gives as  mapping for all the non-looping cloned
1632   // blocks to their outer loops, walk the cloned blocks and the cloned exits
1633   // in their original order adding them to the correct loop.
1634 
1635   // We need a stable insertion order. We use the order of the original loop
1636   // order and map into the correct parent loop.
1637   for (auto *BB : llvm::concat<BasicBlock *const>(
1638            ArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops))
1639     if (Loop *OuterL = ExitLoopMap.lookup(BB))
1640       OuterL->addBasicBlockToLoop(BB, LI);
1641 
1642 #ifndef NDEBUG
1643   for (auto &BBAndL : ExitLoopMap) {
1644     auto *BB = BBAndL.first;
1645     auto *OuterL = BBAndL.second;
1646     assert(LI.getLoopFor(BB) == OuterL &&
1647            "Failed to put all blocks into outer loops!");
1648   }
1649 #endif
1650 
1651   // Now that all the blocks are placed into the correct containing loop in the
1652   // absence of child loops, find all the potentially cloned child loops and
1653   // clone them into whatever outer loop we placed their header into.
1654   for (Loop *ChildL : OrigL) {
1655     auto *ClonedChildHeader =
1656         cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader()));
1657     if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader))
1658       continue;
1659 
1660 #ifndef NDEBUG
1661     for (auto *ChildLoopBB : ChildL->blocks())
1662       assert(VMap.count(ChildLoopBB) &&
1663              "Cloned a child loop header but not all of that loops blocks!");
1664 #endif
1665 
1666     NonChildClonedLoops.push_back(cloneLoopNest(
1667         *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI));
1668   }
1669 }
1670 
1671 static void
deleteDeadClonedBlocks(Loop & L,ArrayRef<BasicBlock * > ExitBlocks,ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,DominatorTree & DT,MemorySSAUpdater * MSSAU)1672 deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1673                        ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps,
1674                        DominatorTree &DT, MemorySSAUpdater *MSSAU) {
1675   // Find all the dead clones, and remove them from their successors.
1676   SmallVector<BasicBlock *, 16> DeadBlocks;
1677   for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks))
1678     for (const auto &VMap : VMaps)
1679       if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB)))
1680         if (!DT.isReachableFromEntry(ClonedBB)) {
1681           for (BasicBlock *SuccBB : successors(ClonedBB))
1682             SuccBB->removePredecessor(ClonedBB);
1683           DeadBlocks.push_back(ClonedBB);
1684         }
1685 
1686   // Remove all MemorySSA in the dead blocks
1687   if (MSSAU) {
1688     SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(),
1689                                                  DeadBlocks.end());
1690     MSSAU->removeBlocks(DeadBlockSet);
1691   }
1692 
1693   // Drop any remaining references to break cycles.
1694   for (BasicBlock *BB : DeadBlocks)
1695     BB->dropAllReferences();
1696   // Erase them from the IR.
1697   for (BasicBlock *BB : DeadBlocks)
1698     BB->eraseFromParent();
1699 }
1700 
deleteDeadBlocksFromLoop(Loop & L,SmallVectorImpl<BasicBlock * > & ExitBlocks,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU,ScalarEvolution * SE,LPMUpdater & LoopUpdater)1701 static void deleteDeadBlocksFromLoop(Loop &L,
1702                                      SmallVectorImpl<BasicBlock *> &ExitBlocks,
1703                                      DominatorTree &DT, LoopInfo &LI,
1704                                      MemorySSAUpdater *MSSAU,
1705                                      ScalarEvolution *SE,
1706                                      LPMUpdater &LoopUpdater) {
1707   // Find all the dead blocks tied to this loop, and remove them from their
1708   // successors.
1709   SmallSetVector<BasicBlock *, 8> DeadBlockSet;
1710 
1711   // Start with loop/exit blocks and get a transitive closure of reachable dead
1712   // blocks.
1713   SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(),
1714                                                 ExitBlocks.end());
1715   DeathCandidates.append(L.blocks().begin(), L.blocks().end());
1716   while (!DeathCandidates.empty()) {
1717     auto *BB = DeathCandidates.pop_back_val();
1718     if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) {
1719       for (BasicBlock *SuccBB : successors(BB)) {
1720         SuccBB->removePredecessor(BB);
1721         DeathCandidates.push_back(SuccBB);
1722       }
1723       DeadBlockSet.insert(BB);
1724     }
1725   }
1726 
1727   // Remove all MemorySSA in the dead blocks
1728   if (MSSAU)
1729     MSSAU->removeBlocks(DeadBlockSet);
1730 
1731   // Filter out the dead blocks from the exit blocks list so that it can be
1732   // used in the caller.
1733   llvm::erase_if(ExitBlocks,
1734                  [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1735 
1736   // Walk from this loop up through its parents removing all of the dead blocks.
1737   for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) {
1738     for (auto *BB : DeadBlockSet)
1739       ParentL->getBlocksSet().erase(BB);
1740     llvm::erase_if(ParentL->getBlocksVector(),
1741                    [&](BasicBlock *BB) { return DeadBlockSet.count(BB); });
1742   }
1743 
1744   // Now delete the dead child loops. This raw delete will clear them
1745   // recursively.
1746   llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) {
1747     if (!DeadBlockSet.count(ChildL->getHeader()))
1748       return false;
1749 
1750     assert(llvm::all_of(ChildL->blocks(),
1751                         [&](BasicBlock *ChildBB) {
1752                           return DeadBlockSet.count(ChildBB);
1753                         }) &&
1754            "If the child loop header is dead all blocks in the child loop must "
1755            "be dead as well!");
1756     LoopUpdater.markLoopAsDeleted(*ChildL, ChildL->getName());
1757     if (SE)
1758       SE->forgetBlockAndLoopDispositions();
1759     LI.destroy(ChildL);
1760     return true;
1761   });
1762 
1763   // Remove the loop mappings for the dead blocks and drop all the references
1764   // from these blocks to others to handle cyclic references as we start
1765   // deleting the blocks themselves.
1766   for (auto *BB : DeadBlockSet) {
1767     // Check that the dominator tree has already been updated.
1768     assert(!DT.getNode(BB) && "Should already have cleared domtree!");
1769     LI.changeLoopFor(BB, nullptr);
1770     // Drop all uses of the instructions to make sure we won't have dangling
1771     // uses in other blocks.
1772     for (auto &I : *BB)
1773       if (!I.use_empty())
1774         I.replaceAllUsesWith(PoisonValue::get(I.getType()));
1775     BB->dropAllReferences();
1776   }
1777 
1778   // Actually delete the blocks now that they've been fully unhooked from the
1779   // IR.
1780   for (auto *BB : DeadBlockSet)
1781     BB->eraseFromParent();
1782 }
1783 
1784 /// Recompute the set of blocks in a loop after unswitching.
1785 ///
1786 /// This walks from the original headers predecessors to rebuild the loop. We
1787 /// take advantage of the fact that new blocks can't have been added, and so we
1788 /// filter by the original loop's blocks. This also handles potentially
1789 /// unreachable code that we don't want to explore but might be found examining
1790 /// the predecessors of the header.
1791 ///
1792 /// If the original loop is no longer a loop, this will return an empty set. If
1793 /// it remains a loop, all the blocks within it will be added to the set
1794 /// (including those blocks in inner loops).
recomputeLoopBlockSet(Loop & L,LoopInfo & LI)1795 static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L,
1796                                                                  LoopInfo &LI) {
1797   SmallPtrSet<const BasicBlock *, 16> LoopBlockSet;
1798 
1799   auto *PH = L.getLoopPreheader();
1800   auto *Header = L.getHeader();
1801 
1802   // A worklist to use while walking backwards from the header.
1803   SmallVector<BasicBlock *, 16> Worklist;
1804 
1805   // First walk the predecessors of the header to find the backedges. This will
1806   // form the basis of our walk.
1807   for (auto *Pred : predecessors(Header)) {
1808     // Skip the preheader.
1809     if (Pred == PH)
1810       continue;
1811 
1812     // Because the loop was in simplified form, the only non-loop predecessor
1813     // is the preheader.
1814     assert(L.contains(Pred) && "Found a predecessor of the loop header other "
1815                                "than the preheader that is not part of the "
1816                                "loop!");
1817 
1818     // Insert this block into the loop set and on the first visit and, if it
1819     // isn't the header we're currently walking, put it into the worklist to
1820     // recurse through.
1821     if (LoopBlockSet.insert(Pred).second && Pred != Header)
1822       Worklist.push_back(Pred);
1823   }
1824 
1825   // If no backedges were found, we're done.
1826   if (LoopBlockSet.empty())
1827     return LoopBlockSet;
1828 
1829   // We found backedges, recurse through them to identify the loop blocks.
1830   while (!Worklist.empty()) {
1831     BasicBlock *BB = Worklist.pop_back_val();
1832     assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!");
1833 
1834     // No need to walk past the header.
1835     if (BB == Header)
1836       continue;
1837 
1838     // Because we know the inner loop structure remains valid we can use the
1839     // loop structure to jump immediately across the entire nested loop.
1840     // Further, because it is in loop simplified form, we can directly jump
1841     // to its preheader afterward.
1842     if (Loop *InnerL = LI.getLoopFor(BB))
1843       if (InnerL != &L) {
1844         assert(L.contains(InnerL) &&
1845                "Should not reach a loop *outside* this loop!");
1846         // The preheader is the only possible predecessor of the loop so
1847         // insert it into the set and check whether it was already handled.
1848         auto *InnerPH = InnerL->getLoopPreheader();
1849         assert(L.contains(InnerPH) && "Cannot contain an inner loop block "
1850                                       "but not contain the inner loop "
1851                                       "preheader!");
1852         if (!LoopBlockSet.insert(InnerPH).second)
1853           // The only way to reach the preheader is through the loop body
1854           // itself so if it has been visited the loop is already handled.
1855           continue;
1856 
1857         // Insert all of the blocks (other than those already present) into
1858         // the loop set. We expect at least the block that led us to find the
1859         // inner loop to be in the block set, but we may also have other loop
1860         // blocks if they were already enqueued as predecessors of some other
1861         // outer loop block.
1862         for (auto *InnerBB : InnerL->blocks()) {
1863           if (InnerBB == BB) {
1864             assert(LoopBlockSet.count(InnerBB) &&
1865                    "Block should already be in the set!");
1866             continue;
1867           }
1868 
1869           LoopBlockSet.insert(InnerBB);
1870         }
1871 
1872         // Add the preheader to the worklist so we will continue past the
1873         // loop body.
1874         Worklist.push_back(InnerPH);
1875         continue;
1876       }
1877 
1878     // Insert any predecessors that were in the original loop into the new
1879     // set, and if the insert is successful, add them to the worklist.
1880     for (auto *Pred : predecessors(BB))
1881       if (L.contains(Pred) && LoopBlockSet.insert(Pred).second)
1882         Worklist.push_back(Pred);
1883   }
1884 
1885   assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!");
1886 
1887   // We've found all the blocks participating in the loop, return our completed
1888   // set.
1889   return LoopBlockSet;
1890 }
1891 
1892 /// Rebuild a loop after unswitching removes some subset of blocks and edges.
1893 ///
1894 /// The removal may have removed some child loops entirely but cannot have
1895 /// disturbed any remaining child loops. However, they may need to be hoisted
1896 /// to the parent loop (or to be top-level loops). The original loop may be
1897 /// completely removed.
1898 ///
1899 /// The sibling loops resulting from this update are returned. If the original
1900 /// loop remains a valid loop, it will be the first entry in this list with all
1901 /// of the newly sibling loops following it.
1902 ///
1903 /// Returns true if the loop remains a loop after unswitching, and false if it
1904 /// is no longer a loop after unswitching (and should not continue to be
1905 /// referenced).
rebuildLoopAfterUnswitch(Loop & L,ArrayRef<BasicBlock * > ExitBlocks,LoopInfo & LI,SmallVectorImpl<Loop * > & HoistedLoops,ScalarEvolution * SE)1906 static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks,
1907                                      LoopInfo &LI,
1908                                      SmallVectorImpl<Loop *> &HoistedLoops,
1909                                      ScalarEvolution *SE) {
1910   auto *PH = L.getLoopPreheader();
1911 
1912   // Compute the actual parent loop from the exit blocks. Because we may have
1913   // pruned some exits the loop may be different from the original parent.
1914   Loop *ParentL = nullptr;
1915   SmallVector<Loop *, 4> ExitLoops;
1916   SmallVector<BasicBlock *, 4> ExitsInLoops;
1917   ExitsInLoops.reserve(ExitBlocks.size());
1918   for (auto *ExitBB : ExitBlocks)
1919     if (Loop *ExitL = LI.getLoopFor(ExitBB)) {
1920       ExitLoops.push_back(ExitL);
1921       ExitsInLoops.push_back(ExitBB);
1922       if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL)))
1923         ParentL = ExitL;
1924     }
1925 
1926   // Recompute the blocks participating in this loop. This may be empty if it
1927   // is no longer a loop.
1928   auto LoopBlockSet = recomputeLoopBlockSet(L, LI);
1929 
1930   // If we still have a loop, we need to re-set the loop's parent as the exit
1931   // block set changing may have moved it within the loop nest. Note that this
1932   // can only happen when this loop has a parent as it can only hoist the loop
1933   // *up* the nest.
1934   if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) {
1935     // Remove this loop's (original) blocks from all of the intervening loops.
1936     for (Loop *IL = L.getParentLoop(); IL != ParentL;
1937          IL = IL->getParentLoop()) {
1938       IL->getBlocksSet().erase(PH);
1939       for (auto *BB : L.blocks())
1940         IL->getBlocksSet().erase(BB);
1941       llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) {
1942         return BB == PH || L.contains(BB);
1943       });
1944     }
1945 
1946     LI.changeLoopFor(PH, ParentL);
1947     L.getParentLoop()->removeChildLoop(&L);
1948     if (ParentL)
1949       ParentL->addChildLoop(&L);
1950     else
1951       LI.addTopLevelLoop(&L);
1952   }
1953 
1954   // Now we update all the blocks which are no longer within the loop.
1955   auto &Blocks = L.getBlocksVector();
1956   auto BlocksSplitI =
1957       LoopBlockSet.empty()
1958           ? Blocks.begin()
1959           : std::stable_partition(
1960                 Blocks.begin(), Blocks.end(),
1961                 [&](BasicBlock *BB) { return LoopBlockSet.count(BB); });
1962 
1963   // Before we erase the list of unlooped blocks, build a set of them.
1964   SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end());
1965   if (LoopBlockSet.empty())
1966     UnloopedBlocks.insert(PH);
1967 
1968   // Now erase these blocks from the loop.
1969   for (auto *BB : make_range(BlocksSplitI, Blocks.end()))
1970     L.getBlocksSet().erase(BB);
1971   Blocks.erase(BlocksSplitI, Blocks.end());
1972 
1973   // Sort the exits in ascending loop depth, we'll work backwards across these
1974   // to process them inside out.
1975   llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) {
1976     return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS);
1977   });
1978 
1979   // We'll build up a set for each exit loop.
1980   SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks;
1981   Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop.
1982 
1983   auto RemoveUnloopedBlocksFromLoop =
1984       [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) {
1985         for (auto *BB : UnloopedBlocks)
1986           L.getBlocksSet().erase(BB);
1987         llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) {
1988           return UnloopedBlocks.count(BB);
1989         });
1990       };
1991 
1992   SmallVector<BasicBlock *, 16> Worklist;
1993   while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) {
1994     assert(Worklist.empty() && "Didn't clear worklist!");
1995     assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!");
1996 
1997     // Grab the next exit block, in decreasing loop depth order.
1998     BasicBlock *ExitBB = ExitsInLoops.pop_back_val();
1999     Loop &ExitL = *LI.getLoopFor(ExitBB);
2000     assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!");
2001 
2002     // Erase all of the unlooped blocks from the loops between the previous
2003     // exit loop and this exit loop. This works because the ExitInLoops list is
2004     // sorted in increasing order of loop depth and thus we visit loops in
2005     // decreasing order of loop depth.
2006     for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop())
2007       RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
2008 
2009     // Walk the CFG back until we hit the cloned PH adding everything reachable
2010     // and in the unlooped set to this exit block's loop.
2011     Worklist.push_back(ExitBB);
2012     do {
2013       BasicBlock *BB = Worklist.pop_back_val();
2014       // We can stop recursing at the cloned preheader (if we get there).
2015       if (BB == PH)
2016         continue;
2017 
2018       for (BasicBlock *PredBB : predecessors(BB)) {
2019         // If this pred has already been moved to our set or is part of some
2020         // (inner) loop, no update needed.
2021         if (!UnloopedBlocks.erase(PredBB)) {
2022           assert((NewExitLoopBlocks.count(PredBB) ||
2023                   ExitL.contains(LI.getLoopFor(PredBB))) &&
2024                  "Predecessor not in a nested loop (or already visited)!");
2025           continue;
2026         }
2027 
2028         // We just insert into the loop set here. We'll add these blocks to the
2029         // exit loop after we build up the set in a deterministic order rather
2030         // than the predecessor-influenced visit order.
2031         bool Inserted = NewExitLoopBlocks.insert(PredBB).second;
2032         (void)Inserted;
2033         assert(Inserted && "Should only visit an unlooped block once!");
2034 
2035         // And recurse through to its predecessors.
2036         Worklist.push_back(PredBB);
2037       }
2038     } while (!Worklist.empty());
2039 
2040     // If blocks in this exit loop were directly part of the original loop (as
2041     // opposed to a child loop) update the map to point to this exit loop. This
2042     // just updates a map and so the fact that the order is unstable is fine.
2043     for (auto *BB : NewExitLoopBlocks)
2044       if (Loop *BBL = LI.getLoopFor(BB))
2045         if (BBL == &L || !L.contains(BBL))
2046           LI.changeLoopFor(BB, &ExitL);
2047 
2048     // We will remove the remaining unlooped blocks from this loop in the next
2049     // iteration or below.
2050     NewExitLoopBlocks.clear();
2051   }
2052 
2053   // Any remaining unlooped blocks are no longer part of any loop unless they
2054   // are part of some child loop.
2055   for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop())
2056     RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks);
2057   for (auto *BB : UnloopedBlocks)
2058     if (Loop *BBL = LI.getLoopFor(BB))
2059       if (BBL == &L || !L.contains(BBL))
2060         LI.changeLoopFor(BB, nullptr);
2061 
2062   // Sink all the child loops whose headers are no longer in the loop set to
2063   // the parent (or to be top level loops). We reach into the loop and directly
2064   // update its subloop vector to make this batch update efficient.
2065   auto &SubLoops = L.getSubLoopsVector();
2066   auto SubLoopsSplitI =
2067       LoopBlockSet.empty()
2068           ? SubLoops.begin()
2069           : std::stable_partition(
2070                 SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) {
2071                   return LoopBlockSet.count(SubL->getHeader());
2072                 });
2073   for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) {
2074     HoistedLoops.push_back(HoistedL);
2075     HoistedL->setParentLoop(nullptr);
2076 
2077     // To compute the new parent of this hoisted loop we look at where we
2078     // placed the preheader above. We can't lookup the header itself because we
2079     // retained the mapping from the header to the hoisted loop. But the
2080     // preheader and header should have the exact same new parent computed
2081     // based on the set of exit blocks from the original loop as the preheader
2082     // is a predecessor of the header and so reached in the reverse walk. And
2083     // because the loops were all in simplified form the preheader of the
2084     // hoisted loop can't be part of some *other* loop.
2085     if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader()))
2086       NewParentL->addChildLoop(HoistedL);
2087     else
2088       LI.addTopLevelLoop(HoistedL);
2089   }
2090   SubLoops.erase(SubLoopsSplitI, SubLoops.end());
2091 
2092   // Actually delete the loop if nothing remained within it.
2093   if (Blocks.empty()) {
2094     assert(SubLoops.empty() &&
2095            "Failed to remove all subloops from the original loop!");
2096     if (Loop *ParentL = L.getParentLoop())
2097       ParentL->removeChildLoop(llvm::find(*ParentL, &L));
2098     else
2099       LI.removeLoop(llvm::find(LI, &L));
2100     // markLoopAsDeleted for L should be triggered by the caller (it is
2101     // typically done within postUnswitch).
2102     if (SE)
2103       SE->forgetBlockAndLoopDispositions();
2104     LI.destroy(&L);
2105     return false;
2106   }
2107 
2108   return true;
2109 }
2110 
2111 /// Helper to visit a dominator subtree, invoking a callable on each node.
2112 ///
2113 /// Returning false at any point will stop walking past that node of the tree.
2114 template <typename CallableT>
visitDomSubTree(DominatorTree & DT,BasicBlock * BB,CallableT Callable)2115 void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) {
2116   SmallVector<DomTreeNode *, 4> DomWorklist;
2117   DomWorklist.push_back(DT[BB]);
2118 #ifndef NDEBUG
2119   SmallPtrSet<DomTreeNode *, 4> Visited;
2120   Visited.insert(DT[BB]);
2121 #endif
2122   do {
2123     DomTreeNode *N = DomWorklist.pop_back_val();
2124 
2125     // Visit this node.
2126     if (!Callable(N->getBlock()))
2127       continue;
2128 
2129     // Accumulate the child nodes.
2130     for (DomTreeNode *ChildN : *N) {
2131       assert(Visited.insert(ChildN).second &&
2132              "Cannot visit a node twice when walking a tree!");
2133       DomWorklist.push_back(ChildN);
2134     }
2135   } while (!DomWorklist.empty());
2136 }
2137 
postUnswitch(Loop & L,LPMUpdater & U,StringRef LoopName,bool CurrentLoopValid,bool PartiallyInvariant,bool InjectedCondition,ArrayRef<Loop * > NewLoops)2138 void postUnswitch(Loop &L, LPMUpdater &U, StringRef LoopName,
2139                   bool CurrentLoopValid, bool PartiallyInvariant,
2140                   bool InjectedCondition, ArrayRef<Loop *> NewLoops) {
2141   // If we did a non-trivial unswitch, we have added new (cloned) loops.
2142   if (!NewLoops.empty())
2143     U.addSiblingLoops(NewLoops);
2144 
2145   // If the current loop remains valid, we should revisit it to catch any
2146   // other unswitch opportunities. Otherwise, we need to mark it as deleted.
2147   if (CurrentLoopValid) {
2148     if (PartiallyInvariant) {
2149       // Mark the new loop as partially unswitched, to avoid unswitching on
2150       // the same condition again.
2151       auto &Context = L.getHeader()->getContext();
2152       MDNode *DisableUnswitchMD = MDNode::get(
2153           Context,
2154           MDString::get(Context, "llvm.loop.unswitch.partial.disable"));
2155       MDNode *NewLoopID = makePostTransformationMetadata(
2156           Context, L.getLoopID(), {"llvm.loop.unswitch.partial"},
2157           {DisableUnswitchMD});
2158       L.setLoopID(NewLoopID);
2159     } else if (InjectedCondition) {
2160       // Do the same for injection of invariant conditions.
2161       auto &Context = L.getHeader()->getContext();
2162       MDNode *DisableUnswitchMD = MDNode::get(
2163           Context,
2164           MDString::get(Context, "llvm.loop.unswitch.injection.disable"));
2165       MDNode *NewLoopID = makePostTransformationMetadata(
2166           Context, L.getLoopID(), {"llvm.loop.unswitch.injection"},
2167           {DisableUnswitchMD});
2168       L.setLoopID(NewLoopID);
2169     } else
2170       U.revisitCurrentLoop();
2171   } else
2172     U.markLoopAsDeleted(L, LoopName);
2173 }
2174 
unswitchNontrivialInvariants(Loop & L,Instruction & TI,ArrayRef<Value * > Invariants,IVConditionInfo & PartialIVInfo,DominatorTree & DT,LoopInfo & LI,AssumptionCache & AC,ScalarEvolution * SE,MemorySSAUpdater * MSSAU,LPMUpdater & LoopUpdater,bool InsertFreeze,bool InjectedCondition)2175 static void unswitchNontrivialInvariants(
2176     Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
2177     IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI,
2178     AssumptionCache &AC, ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
2179     LPMUpdater &LoopUpdater, bool InsertFreeze, bool InjectedCondition) {
2180   auto *ParentBB = TI.getParent();
2181   BranchInst *BI = dyn_cast<BranchInst>(&TI);
2182   SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
2183 
2184   // Save the current loop name in a variable so that we can report it even
2185   // after it has been deleted.
2186   std::string LoopName(L.getName());
2187 
2188   // We can only unswitch switches, conditional branches with an invariant
2189   // condition, or combining invariant conditions with an instruction or
2190   // partially invariant instructions.
2191   assert((SI || (BI && BI->isConditional())) &&
2192          "Can only unswitch switches and conditional branch!");
2193   bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty();
2194   bool FullUnswitch =
2195       SI || (skipTrivialSelect(BI->getCondition()) == Invariants[0] &&
2196              !PartiallyInvariant);
2197   if (FullUnswitch)
2198     assert(Invariants.size() == 1 &&
2199            "Cannot have other invariants with full unswitching!");
2200   else
2201     assert(isa<Instruction>(skipTrivialSelect(BI->getCondition())) &&
2202            "Partial unswitching requires an instruction as the condition!");
2203 
2204   if (MSSAU && VerifyMemorySSA)
2205     MSSAU->getMemorySSA()->verifyMemorySSA();
2206 
2207   // Constant and BBs tracking the cloned and continuing successor. When we are
2208   // unswitching the entire condition, this can just be trivially chosen to
2209   // unswitch towards `true`. However, when we are unswitching a set of
2210   // invariants combined with `and` or `or` or partially invariant instructions,
2211   // the combining operation determines the best direction to unswitch: we want
2212   // to unswitch the direction that will collapse the branch.
2213   bool Direction = true;
2214   int ClonedSucc = 0;
2215   if (!FullUnswitch) {
2216     Value *Cond = skipTrivialSelect(BI->getCondition());
2217     (void)Cond;
2218     assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) ||
2219             PartiallyInvariant) &&
2220            "Only `or`, `and`, an `select`, partially invariant instructions "
2221            "can combine invariants being unswitched.");
2222     if (!match(Cond, m_LogicalOr())) {
2223       if (match(Cond, m_LogicalAnd()) ||
2224           (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) {
2225         Direction = false;
2226         ClonedSucc = 1;
2227       }
2228     }
2229   }
2230 
2231   BasicBlock *RetainedSuccBB =
2232       BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest();
2233   SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs;
2234   if (BI)
2235     UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc));
2236   else
2237     for (auto Case : SI->cases())
2238       if (Case.getCaseSuccessor() != RetainedSuccBB)
2239         UnswitchedSuccBBs.insert(Case.getCaseSuccessor());
2240 
2241   assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&
2242          "Should not unswitch the same successor we are retaining!");
2243 
2244   // The branch should be in this exact loop. Any inner loop's invariant branch
2245   // should be handled by unswitching that inner loop. The caller of this
2246   // routine should filter out any candidates that remain (but were skipped for
2247   // whatever reason).
2248   assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!");
2249 
2250   // Compute the parent loop now before we start hacking on things.
2251   Loop *ParentL = L.getParentLoop();
2252   // Get blocks in RPO order for MSSA update, before changing the CFG.
2253   LoopBlocksRPO LBRPO(&L);
2254   if (MSSAU)
2255     LBRPO.perform(&LI);
2256 
2257   // Compute the outer-most loop containing one of our exit blocks. This is the
2258   // furthest up our loopnest which can be mutated, which we will use below to
2259   // update things.
2260   Loop *OuterExitL = &L;
2261   SmallVector<BasicBlock *, 4> ExitBlocks;
2262   L.getUniqueExitBlocks(ExitBlocks);
2263   for (auto *ExitBB : ExitBlocks) {
2264     // ExitBB can be an exit block for several levels in the loop nest. Make
2265     // sure we find the top most.
2266     Loop *NewOuterExitL = getTopMostExitingLoop(ExitBB, LI);
2267     if (!NewOuterExitL) {
2268       // We exited the entire nest with this block, so we're done.
2269       OuterExitL = nullptr;
2270       break;
2271     }
2272     if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL))
2273       OuterExitL = NewOuterExitL;
2274   }
2275 
2276   // At this point, we're definitely going to unswitch something so invalidate
2277   // any cached information in ScalarEvolution for the outer most loop
2278   // containing an exit block and all nested loops.
2279   if (SE) {
2280     if (OuterExitL)
2281       SE->forgetLoop(OuterExitL);
2282     else
2283       SE->forgetTopmostLoop(&L);
2284     SE->forgetBlockAndLoopDispositions();
2285   }
2286 
2287   // If the edge from this terminator to a successor dominates that successor,
2288   // store a map from each block in its dominator subtree to it. This lets us
2289   // tell when cloning for a particular successor if a block is dominated by
2290   // some *other* successor with a single data structure. We use this to
2291   // significantly reduce cloning.
2292   SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc;
2293   for (auto *SuccBB : llvm::concat<BasicBlock *const>(ArrayRef(RetainedSuccBB),
2294                                                       UnswitchedSuccBBs))
2295     if (SuccBB->getUniquePredecessor() ||
2296         llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
2297           return PredBB == ParentBB || DT.dominates(SuccBB, PredBB);
2298         }))
2299       visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) {
2300         DominatingSucc[BB] = SuccBB;
2301         return true;
2302       });
2303 
2304   // Split the preheader, so that we know that there is a safe place to insert
2305   // the conditional branch. We will change the preheader to have a conditional
2306   // branch on LoopCond. The original preheader will become the split point
2307   // between the unswitched versions, and we will have a new preheader for the
2308   // original loop.
2309   BasicBlock *SplitBB = L.getLoopPreheader();
2310   BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU);
2311 
2312   // Keep track of the dominator tree updates needed.
2313   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2314 
2315   // Clone the loop for each unswitched successor.
2316   SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps;
2317   VMaps.reserve(UnswitchedSuccBBs.size());
2318   SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs;
2319   for (auto *SuccBB : UnswitchedSuccBBs) {
2320     VMaps.emplace_back(new ValueToValueMapTy());
2321     ClonedPHs[SuccBB] = buildClonedLoopBlocks(
2322         L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB,
2323         DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU, SE);
2324   }
2325 
2326   // Drop metadata if we may break its semantics by moving this instr into the
2327   // split block.
2328   if (TI.getMetadata(LLVMContext::MD_make_implicit)) {
2329     if (DropNonTrivialImplicitNullChecks)
2330       // Do not spend time trying to understand if we can keep it, just drop it
2331       // to save compile time.
2332       TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2333     else {
2334       // It is only legal to preserve make.implicit metadata if we are
2335       // guaranteed no reach implicit null check after following this branch.
2336       ICFLoopSafetyInfo SafetyInfo;
2337       SafetyInfo.computeLoopSafetyInfo(&L);
2338       if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))
2339         TI.setMetadata(LLVMContext::MD_make_implicit, nullptr);
2340     }
2341   }
2342 
2343   // The stitching of the branched code back together depends on whether we're
2344   // doing full unswitching or not with the exception that we always want to
2345   // nuke the initial terminator placed in the split block.
2346   SplitBB->getTerminator()->eraseFromParent();
2347   if (FullUnswitch) {
2348     // Keep a clone of the terminator for MSSA updates.
2349     Instruction *NewTI = TI.clone();
2350     NewTI->insertInto(ParentBB, ParentBB->end());
2351 
2352     // Splice the terminator from the original loop and rewrite its
2353     // successors.
2354     TI.moveBefore(*SplitBB, SplitBB->end());
2355     TI.dropLocation();
2356 
2357     // First wire up the moved terminator to the preheaders.
2358     if (BI) {
2359       BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2360       BI->setSuccessor(ClonedSucc, ClonedPH);
2361       BI->setSuccessor(1 - ClonedSucc, LoopPH);
2362       Value *Cond = skipTrivialSelect(BI->getCondition());
2363       if (InsertFreeze) {
2364         // We don't give any debug location to the new freeze, because the
2365         // BI (`dyn_cast<BranchInst>(TI)`) is an in-loop instruction hoisted
2366         // out of the loop.
2367         Cond = new FreezeInst(Cond, Cond->getName() + ".fr", BI->getIterator());
2368       }
2369       BI->setCondition(Cond);
2370       DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2371     } else {
2372       assert(SI && "Must either be a branch or switch!");
2373 
2374       // Walk the cases and directly update their successors.
2375       assert(SI->getDefaultDest() == RetainedSuccBB &&
2376              "Not retaining default successor!");
2377       SI->setDefaultDest(LoopPH);
2378       for (const auto &Case : SI->cases())
2379         if (Case.getCaseSuccessor() == RetainedSuccBB)
2380           Case.setSuccessor(LoopPH);
2381         else
2382           Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second);
2383 
2384       if (InsertFreeze)
2385         SI->setCondition(new FreezeInst(SI->getCondition(),
2386                                         SI->getCondition()->getName() + ".fr",
2387                                         SI->getIterator()));
2388 
2389       // We need to use the set to populate domtree updates as even when there
2390       // are multiple cases pointing at the same successor we only want to
2391       // remove and insert one edge in the domtree.
2392       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2393         DTUpdates.push_back(
2394             {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second});
2395     }
2396 
2397     if (MSSAU) {
2398       DT.applyUpdates(DTUpdates);
2399       DTUpdates.clear();
2400 
2401       // Remove all but one edge to the retained block and all unswitched
2402       // blocks. This is to avoid having duplicate entries in the cloned Phis,
2403       // when we know we only keep a single edge for each case.
2404       MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB);
2405       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2406         MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB);
2407 
2408       for (auto &VMap : VMaps)
2409         MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2410                                    /*IgnoreIncomingWithNoClones=*/true);
2411       MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2412 
2413       // Remove all edges to unswitched blocks.
2414       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2415         MSSAU->removeEdge(ParentBB, SuccBB);
2416     }
2417 
2418     // Now unhook the successor relationship as we'll be replacing
2419     // the terminator with a direct branch. This is much simpler for branches
2420     // than switches so we handle those first.
2421     if (BI) {
2422       // Remove the parent as a predecessor of the unswitched successor.
2423       assert(UnswitchedSuccBBs.size() == 1 &&
2424              "Only one possible unswitched block for a branch!");
2425       BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin();
2426       UnswitchedSuccBB->removePredecessor(ParentBB,
2427                                           /*KeepOneInputPHIs*/ true);
2428       DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB});
2429     } else {
2430       // Note that we actually want to remove the parent block as a predecessor
2431       // of *every* case successor. The case successor is either unswitched,
2432       // completely eliminating an edge from the parent to that successor, or it
2433       // is a duplicate edge to the retained successor as the retained successor
2434       // is always the default successor and as we'll replace this with a direct
2435       // branch we no longer need the duplicate entries in the PHI nodes.
2436       SwitchInst *NewSI = cast<SwitchInst>(NewTI);
2437       assert(NewSI->getDefaultDest() == RetainedSuccBB &&
2438              "Not retaining default successor!");
2439       for (const auto &Case : NewSI->cases())
2440         Case.getCaseSuccessor()->removePredecessor(
2441             ParentBB,
2442             /*KeepOneInputPHIs*/ true);
2443 
2444       // We need to use the set to populate domtree updates as even when there
2445       // are multiple cases pointing at the same successor we only want to
2446       // remove and insert one edge in the domtree.
2447       for (BasicBlock *SuccBB : UnswitchedSuccBBs)
2448         DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB});
2449     }
2450 
2451     // Create a new unconditional branch to the continuing block (as opposed to
2452     // the one cloned).
2453     Instruction *NewBI = BranchInst::Create(RetainedSuccBB, ParentBB);
2454     NewBI->setDebugLoc(NewTI->getDebugLoc());
2455 
2456     // After MSSAU update, remove the cloned terminator instruction NewTI.
2457     NewTI->eraseFromParent();
2458   } else {
2459     assert(BI && "Only branches have partial unswitching.");
2460     assert(UnswitchedSuccBBs.size() == 1 &&
2461            "Only one possible unswitched block for a branch!");
2462     BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2463     // When doing a partial unswitch, we have to do a bit more work to build up
2464     // the branch in the split block.
2465     if (PartiallyInvariant)
2466       buildPartialInvariantUnswitchConditionalBranch(
2467           *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, L, MSSAU);
2468     else {
2469       buildPartialUnswitchConditionalBranch(
2470           *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH,
2471           FreezeLoopUnswitchCond, BI, &AC, DT);
2472     }
2473     DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH});
2474 
2475     if (MSSAU) {
2476       DT.applyUpdates(DTUpdates);
2477       DTUpdates.clear();
2478 
2479       // Perform MSSA cloning updates.
2480       for (auto &VMap : VMaps)
2481         MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap,
2482                                    /*IgnoreIncomingWithNoClones=*/true);
2483       MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT);
2484     }
2485   }
2486 
2487   // Apply the updates accumulated above to get an up-to-date dominator tree.
2488   DT.applyUpdates(DTUpdates);
2489 
2490   // Now that we have an accurate dominator tree, first delete the dead cloned
2491   // blocks so that we can accurately build any cloned loops. It is important to
2492   // not delete the blocks from the original loop yet because we still want to
2493   // reference the original loop to understand the cloned loop's structure.
2494   deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU);
2495 
2496   // Build the cloned loop structure itself. This may be substantially
2497   // different from the original structure due to the simplified CFG. This also
2498   // handles inserting all the cloned blocks into the correct loops.
2499   SmallVector<Loop *, 4> NonChildClonedLoops;
2500   for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps)
2501     buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops);
2502 
2503   // Now that our cloned loops have been built, we can update the original loop.
2504   // First we delete the dead blocks from it and then we rebuild the loop
2505   // structure taking these deletions into account.
2506   deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, SE, LoopUpdater);
2507 
2508   if (MSSAU && VerifyMemorySSA)
2509     MSSAU->getMemorySSA()->verifyMemorySSA();
2510 
2511   SmallVector<Loop *, 4> HoistedLoops;
2512   bool IsStillLoop =
2513       rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops, SE);
2514 
2515   if (MSSAU && VerifyMemorySSA)
2516     MSSAU->getMemorySSA()->verifyMemorySSA();
2517 
2518   // This transformation has a high risk of corrupting the dominator tree, and
2519   // the below steps to rebuild loop structures will result in hard to debug
2520   // errors in that case so verify that the dominator tree is sane first.
2521   // FIXME: Remove this when the bugs stop showing up and rely on existing
2522   // verification steps.
2523   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2524 
2525   if (BI && !PartiallyInvariant) {
2526     // If we unswitched a branch which collapses the condition to a known
2527     // constant we want to replace all the uses of the invariants within both
2528     // the original and cloned blocks. We do this here so that we can use the
2529     // now updated dominator tree to identify which side the users are on.
2530     assert(UnswitchedSuccBBs.size() == 1 &&
2531            "Only one possible unswitched block for a branch!");
2532     BasicBlock *ClonedPH = ClonedPHs.begin()->second;
2533 
2534     // When considering multiple partially-unswitched invariants
2535     // we cant just go replace them with constants in both branches.
2536     //
2537     // For 'AND' we infer that true branch ("continue") means true
2538     // for each invariant operand.
2539     // For 'OR' we can infer that false branch ("continue") means false
2540     // for each invariant operand.
2541     // So it happens that for multiple-partial case we dont replace
2542     // in the unswitched branch.
2543     bool ReplaceUnswitched =
2544         FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant;
2545 
2546     ConstantInt *UnswitchedReplacement =
2547         Direction ? ConstantInt::getTrue(BI->getContext())
2548                   : ConstantInt::getFalse(BI->getContext());
2549     ConstantInt *ContinueReplacement =
2550         Direction ? ConstantInt::getFalse(BI->getContext())
2551                   : ConstantInt::getTrue(BI->getContext());
2552     for (Value *Invariant : Invariants) {
2553       assert(!isa<Constant>(Invariant) &&
2554              "Should not be replacing constant values!");
2555       // Use make_early_inc_range here as set invalidates the iterator.
2556       for (Use &U : llvm::make_early_inc_range(Invariant->uses())) {
2557         Instruction *UserI = dyn_cast<Instruction>(U.getUser());
2558         if (!UserI)
2559           continue;
2560 
2561         // Replace it with the 'continue' side if in the main loop body, and the
2562         // unswitched if in the cloned blocks.
2563         if (DT.dominates(LoopPH, UserI->getParent()))
2564           U.set(ContinueReplacement);
2565         else if (ReplaceUnswitched &&
2566                  DT.dominates(ClonedPH, UserI->getParent()))
2567           U.set(UnswitchedReplacement);
2568       }
2569     }
2570   }
2571 
2572   // We can change which blocks are exit blocks of all the cloned sibling
2573   // loops, the current loop, and any parent loops which shared exit blocks
2574   // with the current loop. As a consequence, we need to re-form LCSSA for
2575   // them. But we shouldn't need to re-form LCSSA for any child loops.
2576   // FIXME: This could be made more efficient by tracking which exit blocks are
2577   // new, and focusing on them, but that isn't likely to be necessary.
2578   //
2579   // In order to reasonably rebuild LCSSA we need to walk inside-out across the
2580   // loop nest and update every loop that could have had its exits changed. We
2581   // also need to cover any intervening loops. We add all of these loops to
2582   // a list and sort them by loop depth to achieve this without updating
2583   // unnecessary loops.
2584   auto UpdateLoop = [&](Loop &UpdateL) {
2585 #ifndef NDEBUG
2586     UpdateL.verifyLoop();
2587     for (Loop *ChildL : UpdateL) {
2588       ChildL->verifyLoop();
2589       assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&
2590              "Perturbed a child loop's LCSSA form!");
2591     }
2592 #endif
2593     // First build LCSSA for this loop so that we can preserve it when
2594     // forming dedicated exits. We don't want to perturb some other loop's
2595     // LCSSA while doing that CFG edit.
2596     formLCSSA(UpdateL, DT, &LI, SE);
2597 
2598     // For loops reached by this loop's original exit blocks we may
2599     // introduced new, non-dedicated exits. At least try to re-form dedicated
2600     // exits for these loops. This may fail if they couldn't have dedicated
2601     // exits to start with.
2602     formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true);
2603   };
2604 
2605   // For non-child cloned loops and hoisted loops, we just need to update LCSSA
2606   // and we can do it in any order as they don't nest relative to each other.
2607   //
2608   // Also check if any of the loops we have updated have become top-level loops
2609   // as that will necessitate widening the outer loop scope.
2610   for (Loop *UpdatedL :
2611        llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) {
2612     UpdateLoop(*UpdatedL);
2613     if (UpdatedL->isOutermost())
2614       OuterExitL = nullptr;
2615   }
2616   if (IsStillLoop) {
2617     UpdateLoop(L);
2618     if (L.isOutermost())
2619       OuterExitL = nullptr;
2620   }
2621 
2622   // If the original loop had exit blocks, walk up through the outer most loop
2623   // of those exit blocks to update LCSSA and form updated dedicated exits.
2624   if (OuterExitL != &L)
2625     for (Loop *OuterL = ParentL; OuterL != OuterExitL;
2626          OuterL = OuterL->getParentLoop())
2627       UpdateLoop(*OuterL);
2628 
2629 #ifndef NDEBUG
2630   // Verify the entire loop structure to catch any incorrect updates before we
2631   // progress in the pass pipeline.
2632   LI.verify(DT);
2633 #endif
2634 
2635   // Now that we've unswitched something, make callbacks to report the changes.
2636   // For that we need to merge together the updated loops and the cloned loops
2637   // and check whether the original loop survived.
2638   SmallVector<Loop *, 4> SibLoops;
2639   for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
2640     if (UpdatedL->getParentLoop() == ParentL)
2641       SibLoops.push_back(UpdatedL);
2642   postUnswitch(L, LoopUpdater, LoopName, IsStillLoop, PartiallyInvariant,
2643                InjectedCondition, SibLoops);
2644 
2645   if (MSSAU && VerifyMemorySSA)
2646     MSSAU->getMemorySSA()->verifyMemorySSA();
2647 
2648   if (BI)
2649     ++NumBranches;
2650   else
2651     ++NumSwitches;
2652 }
2653 
2654 /// Recursively compute the cost of a dominator subtree based on the per-block
2655 /// cost map provided.
2656 ///
2657 /// The recursive computation is memozied into the provided DT-indexed cost map
2658 /// to allow querying it for most nodes in the domtree without it becoming
2659 /// quadratic.
computeDomSubtreeCost(DomTreeNode & N,const SmallDenseMap<BasicBlock *,InstructionCost,4> & BBCostMap,SmallDenseMap<DomTreeNode *,InstructionCost,4> & DTCostMap)2660 static InstructionCost computeDomSubtreeCost(
2661     DomTreeNode &N,
2662     const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap,
2663     SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) {
2664   // Don't accumulate cost (or recurse through) blocks not in our block cost
2665   // map and thus not part of the duplication cost being considered.
2666   auto BBCostIt = BBCostMap.find(N.getBlock());
2667   if (BBCostIt == BBCostMap.end())
2668     return 0;
2669 
2670   // Lookup this node to see if we already computed its cost.
2671   auto DTCostIt = DTCostMap.find(&N);
2672   if (DTCostIt != DTCostMap.end())
2673     return DTCostIt->second;
2674 
2675   // If not, we have to compute it. We can't use insert above and update
2676   // because computing the cost may insert more things into the map.
2677   InstructionCost Cost = std::accumulate(
2678       N.begin(), N.end(), BBCostIt->second,
2679       [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost {
2680         return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
2681       });
2682   bool Inserted = DTCostMap.insert({&N, Cost}).second;
2683   (void)Inserted;
2684   assert(Inserted && "Should not insert a node while visiting children!");
2685   return Cost;
2686 }
2687 
2688 /// Turns a select instruction into implicit control flow branch,
2689 /// making the following replacement:
2690 ///
2691 /// head:
2692 ///   --code before select--
2693 ///   select %cond, %trueval, %falseval
2694 ///   --code after select--
2695 ///
2696 /// into
2697 ///
2698 /// head:
2699 ///   --code before select--
2700 ///   br i1 %cond, label %then, label %tail
2701 ///
2702 /// then:
2703 ///   br %tail
2704 ///
2705 /// tail:
2706 ///   phi [ %trueval, %then ], [ %falseval, %head]
2707 ///   unreachable
2708 ///
2709 /// It also makes all relevant DT and LI updates, so that all structures are in
2710 /// valid state after this transform.
turnSelectIntoBranch(SelectInst * SI,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU,AssumptionCache * AC)2711 static BranchInst *turnSelectIntoBranch(SelectInst *SI, DominatorTree &DT,
2712                                         LoopInfo &LI, MemorySSAUpdater *MSSAU,
2713                                         AssumptionCache *AC) {
2714   LLVM_DEBUG(dbgs() << "Turning " << *SI << " into a branch.\n");
2715   BasicBlock *HeadBB = SI->getParent();
2716 
2717   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2718   SplitBlockAndInsertIfThen(SI->getCondition(), SI, false,
2719                             SI->getMetadata(LLVMContext::MD_prof), &DTU, &LI);
2720   auto *CondBr = cast<BranchInst>(HeadBB->getTerminator());
2721   BasicBlock *ThenBB = CondBr->getSuccessor(0),
2722              *TailBB = CondBr->getSuccessor(1);
2723   if (MSSAU)
2724     MSSAU->moveAllAfterSpliceBlocks(HeadBB, TailBB, SI);
2725 
2726   PHINode *Phi =
2727       PHINode::Create(SI->getType(), 2, "unswitched.select", SI->getIterator());
2728   Phi->addIncoming(SI->getTrueValue(), ThenBB);
2729   Phi->addIncoming(SI->getFalseValue(), HeadBB);
2730   Phi->setDebugLoc(SI->getDebugLoc());
2731   SI->replaceAllUsesWith(Phi);
2732   SI->eraseFromParent();
2733 
2734   if (MSSAU && VerifyMemorySSA)
2735     MSSAU->getMemorySSA()->verifyMemorySSA();
2736 
2737   ++NumSelects;
2738   return CondBr;
2739 }
2740 
2741 /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
2742 /// making the following replacement:
2743 ///
2744 ///   --code before guard--
2745 ///   call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
2746 ///   --code after guard--
2747 ///
2748 /// into
2749 ///
2750 ///   --code before guard--
2751 ///   br i1 %cond, label %guarded, label %deopt
2752 ///
2753 /// guarded:
2754 ///   --code after guard--
2755 ///
2756 /// deopt:
2757 ///   call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
2758 ///   unreachable
2759 ///
2760 /// It also makes all relevant DT and LI updates, so that all structures are in
2761 /// valid state after this transform.
turnGuardIntoBranch(IntrinsicInst * GI,Loop & L,DominatorTree & DT,LoopInfo & LI,MemorySSAUpdater * MSSAU)2762 static BranchInst *turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
2763                                        DominatorTree &DT, LoopInfo &LI,
2764                                        MemorySSAUpdater *MSSAU) {
2765   SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
2766   LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
2767   BasicBlock *CheckBB = GI->getParent();
2768 
2769   if (MSSAU && VerifyMemorySSA)
2770      MSSAU->getMemorySSA()->verifyMemorySSA();
2771 
2772   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2773   Instruction *DeoptBlockTerm =
2774       SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true,
2775                                 GI->getMetadata(LLVMContext::MD_prof), &DTU, &LI);
2776   BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
2777   // SplitBlockAndInsertIfThen inserts control flow that branches to
2778   // DeoptBlockTerm if the condition is true.  We want the opposite.
2779   CheckBI->swapSuccessors();
2780 
2781   BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
2782   GuardedBlock->setName("guarded");
2783   CheckBI->getSuccessor(1)->setName("deopt");
2784   BasicBlock *DeoptBlock = CheckBI->getSuccessor(1);
2785 
2786   if (MSSAU)
2787     MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI);
2788 
2789   GI->moveBefore(DeoptBlockTerm);
2790   GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));
2791 
2792   if (MSSAU) {
2793     MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI));
2794     MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator);
2795     if (VerifyMemorySSA)
2796       MSSAU->getMemorySSA()->verifyMemorySSA();
2797   }
2798 
2799   if (VerifyLoopInfo)
2800     LI.verify(DT);
2801   ++NumGuards;
2802   return CheckBI;
2803 }
2804 
2805 /// Cost multiplier is a way to limit potentially exponential behavior
2806 /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch
2807 /// candidates available. Also accounting for the number of "sibling" loops with
2808 /// the idea to account for previous unswitches that already happened on this
2809 /// cluster of loops. There was an attempt to keep this formula simple,
2810 /// just enough to limit the worst case behavior. Even if it is not that simple
2811 /// now it is still not an attempt to provide a detailed heuristic size
2812 /// prediction.
2813 ///
2814 /// TODO: Make a proper accounting of "explosion" effect for all kinds of
2815 /// unswitch candidates, making adequate predictions instead of wild guesses.
2816 /// That requires knowing not just the number of "remaining" candidates but
2817 /// also costs of unswitching for each of these candidates.
CalculateUnswitchCostMultiplier(const Instruction & TI,const Loop & L,const LoopInfo & LI,const DominatorTree & DT,ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates)2818 static int CalculateUnswitchCostMultiplier(
2819     const Instruction &TI, const Loop &L, const LoopInfo &LI,
2820     const DominatorTree &DT,
2821     ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates) {
2822 
2823   // Guards and other exiting conditions do not contribute to exponential
2824   // explosion as soon as they dominate the latch (otherwise there might be
2825   // another path to the latch remaining that does not allow to eliminate the
2826   // loop copy on unswitch).
2827   const BasicBlock *Latch = L.getLoopLatch();
2828   const BasicBlock *CondBlock = TI.getParent();
2829   if (DT.dominates(CondBlock, Latch) &&
2830       (isGuard(&TI) ||
2831        (TI.isTerminator() &&
2832         llvm::count_if(successors(&TI), [&L](const BasicBlock *SuccBB) {
2833           return L.contains(SuccBB);
2834         }) <= 1))) {
2835     NumCostMultiplierSkipped++;
2836     return 1;
2837   }
2838 
2839   auto *ParentL = L.getParentLoop();
2840   int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size()
2841                                : std::distance(LI.begin(), LI.end()));
2842   // Count amount of clones that all the candidates might cause during
2843   // unswitching. Branch/guard/select counts as 1, switch counts as log2 of its
2844   // cases.
2845   int UnswitchedClones = 0;
2846   for (const auto &Candidate : UnswitchCandidates) {
2847     const Instruction *CI = Candidate.TI;
2848     const BasicBlock *CondBlock = CI->getParent();
2849     bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch);
2850     if (isa<SelectInst>(CI)) {
2851       UnswitchedClones++;
2852       continue;
2853     }
2854     if (isGuard(CI)) {
2855       if (!SkipExitingSuccessors)
2856         UnswitchedClones++;
2857       continue;
2858     }
2859     int NonExitingSuccessors =
2860         llvm::count_if(successors(CondBlock),
2861                        [SkipExitingSuccessors, &L](const BasicBlock *SuccBB) {
2862           return !SkipExitingSuccessors || L.contains(SuccBB);
2863         });
2864     UnswitchedClones += Log2_32(NonExitingSuccessors);
2865   }
2866 
2867   // Ignore up to the "unscaled candidates" number of unswitch candidates
2868   // when calculating the power-of-two scaling of the cost. The main idea
2869   // with this control is to allow a small number of unswitches to happen
2870   // and rely more on siblings multiplier (see below) when the number
2871   // of candidates is small.
2872   unsigned ClonesPower =
2873       std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0);
2874 
2875   // Allowing top-level loops to spread a bit more than nested ones.
2876   int SiblingsMultiplier =
2877       std::max((ParentL ? SiblingsCount
2878                         : SiblingsCount / (int)UnswitchSiblingsToplevelDiv),
2879                1);
2880   // Compute the cost multiplier in a way that won't overflow by saturating
2881   // at an upper bound.
2882   int CostMultiplier;
2883   if (ClonesPower > Log2_32(UnswitchThreshold) ||
2884       SiblingsMultiplier > UnswitchThreshold)
2885     CostMultiplier = UnswitchThreshold;
2886   else
2887     CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower),
2888                               (int)UnswitchThreshold);
2889 
2890   LLVM_DEBUG(dbgs() << "  Computed multiplier  " << CostMultiplier
2891                     << " (siblings " << SiblingsMultiplier << " * clones "
2892                     << (1 << ClonesPower) << ")"
2893                     << " for unswitch candidate: " << TI << "\n");
2894   return CostMultiplier;
2895 }
2896 
collectUnswitchCandidates(SmallVectorImpl<NonTrivialUnswitchCandidate> & UnswitchCandidates,IVConditionInfo & PartialIVInfo,Instruction * & PartialIVCondBranch,const Loop & L,const LoopInfo & LI,AAResults & AA,const MemorySSAUpdater * MSSAU)2897 static bool collectUnswitchCandidates(
2898     SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
2899     IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch,
2900     const Loop &L, const LoopInfo &LI, AAResults &AA,
2901     const MemorySSAUpdater *MSSAU) {
2902   assert(UnswitchCandidates.empty() && "Should be!");
2903 
2904   auto AddUnswitchCandidatesForInst = [&](Instruction *I, Value *Cond) {
2905     Cond = skipTrivialSelect(Cond);
2906     if (isa<Constant>(Cond))
2907       return;
2908     if (L.isLoopInvariant(Cond)) {
2909       UnswitchCandidates.push_back({I, {Cond}});
2910       return;
2911     }
2912     if (match(Cond, m_CombineOr(m_LogicalAnd(), m_LogicalOr()))) {
2913       TinyPtrVector<Value *> Invariants =
2914           collectHomogenousInstGraphLoopInvariants(
2915               L, *static_cast<Instruction *>(Cond), LI);
2916       if (!Invariants.empty())
2917         UnswitchCandidates.push_back({I, std::move(Invariants)});
2918     }
2919   };
2920 
2921   // Whether or not we should also collect guards in the loop.
2922   bool CollectGuards = false;
2923   if (UnswitchGuards) {
2924     auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
2925         Intrinsic::getName(Intrinsic::experimental_guard));
2926     if (GuardDecl && !GuardDecl->use_empty())
2927       CollectGuards = true;
2928   }
2929 
2930   for (auto *BB : L.blocks()) {
2931     if (LI.getLoopFor(BB) != &L)
2932       continue;
2933 
2934     for (auto &I : *BB) {
2935       if (auto *SI = dyn_cast<SelectInst>(&I)) {
2936         auto *Cond = SI->getCondition();
2937         // Do not unswitch vector selects and logical and/or selects
2938         if (Cond->getType()->isIntegerTy(1) && !SI->getType()->isIntegerTy(1))
2939           AddUnswitchCandidatesForInst(SI, Cond);
2940       } else if (CollectGuards && isGuard(&I)) {
2941         auto *Cond =
2942             skipTrivialSelect(cast<IntrinsicInst>(&I)->getArgOperand(0));
2943         // TODO: Support AND, OR conditions and partial unswitching.
2944         if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
2945           UnswitchCandidates.push_back({&I, {Cond}});
2946       }
2947     }
2948 
2949     if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
2950       // We can only consider fully loop-invariant switch conditions as we need
2951       // to completely eliminate the switch after unswitching.
2952       if (!isa<Constant>(SI->getCondition()) &&
2953           L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor())
2954         UnswitchCandidates.push_back({SI, {SI->getCondition()}});
2955       continue;
2956     }
2957 
2958     auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
2959     if (!BI || !BI->isConditional() ||
2960         BI->getSuccessor(0) == BI->getSuccessor(1))
2961       continue;
2962 
2963     AddUnswitchCandidatesForInst(BI, BI->getCondition());
2964   }
2965 
2966   if (MSSAU && !findOptionMDForLoop(&L, "llvm.loop.unswitch.partial.disable") &&
2967       !any_of(UnswitchCandidates, [&L](auto &TerminatorAndInvariants) {
2968          return TerminatorAndInvariants.TI == L.getHeader()->getTerminator();
2969        })) {
2970     MemorySSA *MSSA = MSSAU->getMemorySSA();
2971     if (auto Info = hasPartialIVCondition(L, MSSAThreshold, *MSSA, AA)) {
2972       LLVM_DEBUG(
2973           dbgs() << "simple-loop-unswitch: Found partially invariant condition "
2974                  << *Info->InstToDuplicate[0] << "\n");
2975       PartialIVInfo = *Info;
2976       PartialIVCondBranch = L.getHeader()->getTerminator();
2977       TinyPtrVector<Value *> ValsToDuplicate;
2978       llvm::append_range(ValsToDuplicate, Info->InstToDuplicate);
2979       UnswitchCandidates.push_back(
2980           {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)});
2981     }
2982   }
2983   return !UnswitchCandidates.empty();
2984 }
2985 
2986 /// Tries to canonicalize condition described by:
2987 ///
2988 ///   br (LHS pred RHS), label IfTrue, label IfFalse
2989 ///
2990 /// into its equivalent where `Pred` is something that we support for injected
2991 /// invariants (so far it is limited to ult), LHS in canonicalized form is
2992 /// non-invariant and RHS is an invariant.
canonicalizeForInvariantConditionInjection(ICmpInst::Predicate & Pred,Value * & LHS,Value * & RHS,BasicBlock * & IfTrue,BasicBlock * & IfFalse,const Loop & L)2993 static void canonicalizeForInvariantConditionInjection(
2994     ICmpInst::Predicate &Pred, Value *&LHS, Value *&RHS, BasicBlock *&IfTrue,
2995     BasicBlock *&IfFalse, const Loop &L) {
2996   if (!L.contains(IfTrue)) {
2997     Pred = ICmpInst::getInversePredicate(Pred);
2998     std::swap(IfTrue, IfFalse);
2999   }
3000 
3001   // Move loop-invariant argument to RHS position.
3002   if (L.isLoopInvariant(LHS)) {
3003     Pred = ICmpInst::getSwappedPredicate(Pred);
3004     std::swap(LHS, RHS);
3005   }
3006 
3007   if (Pred == ICmpInst::ICMP_SGE && match(RHS, m_Zero())) {
3008     // Turn "x >=s 0" into "x <u UMIN_INT"
3009     Pred = ICmpInst::ICMP_ULT;
3010     RHS = ConstantInt::get(
3011         RHS->getContext(),
3012         APInt::getSignedMinValue(RHS->getType()->getIntegerBitWidth()));
3013   }
3014 }
3015 
3016 /// Returns true, if predicate described by ( \p Pred, \p LHS, \p RHS )
3017 /// succeeding into blocks ( \p IfTrue, \p IfFalse) can be optimized by
3018 /// injecting a loop-invariant condition.
shouldTryInjectInvariantCondition(const ICmpInst::Predicate Pred,const Value * LHS,const Value * RHS,const BasicBlock * IfTrue,const BasicBlock * IfFalse,const Loop & L)3019 static bool shouldTryInjectInvariantCondition(
3020     const ICmpInst::Predicate Pred, const Value *LHS, const Value *RHS,
3021     const BasicBlock *IfTrue, const BasicBlock *IfFalse, const Loop &L) {
3022   if (L.isLoopInvariant(LHS) || !L.isLoopInvariant(RHS))
3023     return false;
3024   // TODO: Support other predicates.
3025   if (Pred != ICmpInst::ICMP_ULT)
3026     return false;
3027   // TODO: Support non-loop-exiting branches?
3028   if (!L.contains(IfTrue) || L.contains(IfFalse))
3029     return false;
3030   // FIXME: For some reason this causes problems with MSSA updates, need to
3031   // investigate why. So far, just don't unswitch latch.
3032   if (L.getHeader() == IfTrue)
3033     return false;
3034   return true;
3035 }
3036 
3037 /// Returns true, if metadata on \p BI allows us to optimize branching into \p
3038 /// TakenSucc via injection of invariant conditions. The branch should be not
3039 /// enough and not previously unswitched, the information about this comes from
3040 /// the metadata.
shouldTryInjectBasingOnMetadata(const BranchInst * BI,const BasicBlock * TakenSucc)3041 bool shouldTryInjectBasingOnMetadata(const BranchInst *BI,
3042                                      const BasicBlock *TakenSucc) {
3043   SmallVector<uint32_t> Weights;
3044   if (!extractBranchWeights(*BI, Weights))
3045     return false;
3046   unsigned T = InjectInvariantConditionHotnesThreshold;
3047   BranchProbability LikelyTaken(T - 1, T);
3048 
3049   assert(Weights.size() == 2 && "Unexpected profile data!");
3050   size_t Idx = BI->getSuccessor(0) == TakenSucc ? 0 : 1;
3051   auto Num = Weights[Idx];
3052   auto Denom = Weights[0] + Weights[1];
3053   // Degenerate or overflowed metadata.
3054   if (Denom == 0 || Num > Denom)
3055     return false;
3056   BranchProbability ActualTaken(Num, Denom);
3057   if (LikelyTaken > ActualTaken)
3058     return false;
3059   return true;
3060 }
3061 
3062 /// Materialize pending invariant condition of the given candidate into IR. The
3063 /// injected loop-invariant condition implies the original loop-variant branch
3064 /// condition, so the materialization turns
3065 ///
3066 /// loop_block:
3067 ///   ...
3068 ///   br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3069 ///
3070 /// into
3071 ///
3072 /// preheader:
3073 ///   %invariant_cond = LHS pred RHS
3074 /// ...
3075 /// loop_block:
3076 ///   br i1 %invariant_cond, label InLoopSucc, label OriginalCheck
3077 /// OriginalCheck:
3078 ///   br i1 %variant_cond, label InLoopSucc, label OutOfLoopSucc
3079 /// ...
3080 static NonTrivialUnswitchCandidate
injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate,Loop & L,DominatorTree & DT,LoopInfo & LI,AssumptionCache & AC,MemorySSAUpdater * MSSAU)3081 injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L,
3082                                  DominatorTree &DT, LoopInfo &LI,
3083                                  AssumptionCache &AC, MemorySSAUpdater *MSSAU) {
3084   assert(Candidate.hasPendingInjection() && "Nothing to inject!");
3085   BasicBlock *Preheader = L.getLoopPreheader();
3086   assert(Preheader && "Loop is not in simplified form?");
3087   assert(LI.getLoopFor(Candidate.TI->getParent()) == &L &&
3088          "Unswitching branch of inner loop!");
3089 
3090   auto Pred = Candidate.PendingInjection->Pred;
3091   auto *LHS = Candidate.PendingInjection->LHS;
3092   auto *RHS = Candidate.PendingInjection->RHS;
3093   auto *InLoopSucc = Candidate.PendingInjection->InLoopSucc;
3094   auto *TI = cast<BranchInst>(Candidate.TI);
3095   auto *BB = Candidate.TI->getParent();
3096   auto *OutOfLoopSucc = InLoopSucc == TI->getSuccessor(0) ? TI->getSuccessor(1)
3097                                                           : TI->getSuccessor(0);
3098   // FIXME: Remove this once limitation on successors is lifted.
3099   assert(L.contains(InLoopSucc) && "Not supported yet!");
3100   assert(!L.contains(OutOfLoopSucc) && "Not supported yet!");
3101   auto &Ctx = BB->getContext();
3102 
3103   IRBuilder<> Builder(Preheader->getTerminator());
3104   assert(ICmpInst::isUnsigned(Pred) && "Not supported yet!");
3105   if (LHS->getType() != RHS->getType()) {
3106     if (LHS->getType()->getIntegerBitWidth() <
3107         RHS->getType()->getIntegerBitWidth())
3108       LHS = Builder.CreateZExt(LHS, RHS->getType(), LHS->getName() + ".wide");
3109     else
3110       RHS = Builder.CreateZExt(RHS, LHS->getType(), RHS->getName() + ".wide");
3111   }
3112   // Do not use builder here: CreateICmp may simplify this into a constant and
3113   // unswitching will break. Better optimize it away later.
3114   auto *InjectedCond =
3115       ICmpInst::Create(Instruction::ICmp, Pred, LHS, RHS, "injected.cond",
3116                        Preheader->getTerminator()->getIterator());
3117 
3118   BasicBlock *CheckBlock = BasicBlock::Create(Ctx, BB->getName() + ".check",
3119                                               BB->getParent(), InLoopSucc);
3120   Builder.SetInsertPoint(TI);
3121   auto *InvariantBr =
3122       Builder.CreateCondBr(InjectedCond, InLoopSucc, CheckBlock);
3123 
3124   Builder.SetInsertPoint(CheckBlock);
3125   Builder.CreateCondBr(TI->getCondition(), TI->getSuccessor(0),
3126                        TI->getSuccessor(1));
3127   TI->eraseFromParent();
3128 
3129   // Fixup phis.
3130   for (auto &I : *InLoopSucc) {
3131     auto *PN = dyn_cast<PHINode>(&I);
3132     if (!PN)
3133       break;
3134     auto *Inc = PN->getIncomingValueForBlock(BB);
3135     PN->addIncoming(Inc, CheckBlock);
3136   }
3137   OutOfLoopSucc->replacePhiUsesWith(BB, CheckBlock);
3138 
3139   SmallVector<DominatorTree::UpdateType, 4> DTUpdates = {
3140     { DominatorTree::Insert, BB, CheckBlock },
3141     { DominatorTree::Insert, CheckBlock, InLoopSucc },
3142     { DominatorTree::Insert, CheckBlock, OutOfLoopSucc },
3143     { DominatorTree::Delete, BB, OutOfLoopSucc }
3144   };
3145 
3146   DT.applyUpdates(DTUpdates);
3147   if (MSSAU)
3148     MSSAU->applyUpdates(DTUpdates, DT);
3149   L.addBasicBlockToLoop(CheckBlock, LI);
3150 
3151 #ifndef NDEBUG
3152   DT.verify();
3153   LI.verify(DT);
3154   if (MSSAU && VerifyMemorySSA)
3155     MSSAU->getMemorySSA()->verifyMemorySSA();
3156 #endif
3157 
3158   // TODO: In fact, cost of unswitching a new invariant candidate is *slightly*
3159   // higher because we have just inserted a new block. Need to think how to
3160   // adjust the cost of injected candidates when it was first computed.
3161   LLVM_DEBUG(dbgs() << "Injected a new loop-invariant branch " << *InvariantBr
3162                     << " and considering it for unswitching.");
3163   ++NumInvariantConditionsInjected;
3164   return NonTrivialUnswitchCandidate(InvariantBr, { InjectedCond },
3165                                      Candidate.Cost);
3166 }
3167 
3168 /// Given chain of loop branch conditions looking like:
3169 ///   br (Variant < Invariant1)
3170 ///   br (Variant < Invariant2)
3171 ///   br (Variant < Invariant3)
3172 ///   ...
3173 /// collect set of invariant conditions on which we want to unswitch, which
3174 /// look like:
3175 ///   Invariant1 <= Invariant2
3176 ///   Invariant2 <= Invariant3
3177 ///   ...
3178 /// Though they might not immediately exist in the IR, we can still inject them.
insertCandidatesWithPendingInjections(SmallVectorImpl<NonTrivialUnswitchCandidate> & UnswitchCandidates,Loop & L,ICmpInst::Predicate Pred,ArrayRef<CompareDesc> Compares,const DominatorTree & DT)3179 static bool insertCandidatesWithPendingInjections(
3180     SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates, Loop &L,
3181     ICmpInst::Predicate Pred, ArrayRef<CompareDesc> Compares,
3182     const DominatorTree &DT) {
3183 
3184   assert(ICmpInst::isRelational(Pred));
3185   assert(ICmpInst::isStrictPredicate(Pred));
3186   if (Compares.size() < 2)
3187     return false;
3188   ICmpInst::Predicate NonStrictPred = ICmpInst::getNonStrictPredicate(Pred);
3189   for (auto Prev = Compares.begin(), Next = Compares.begin() + 1;
3190        Next != Compares.end(); ++Prev, ++Next) {
3191     Value *LHS = Next->Invariant;
3192     Value *RHS = Prev->Invariant;
3193     BasicBlock *InLoopSucc = Prev->InLoopSucc;
3194     InjectedInvariant ToInject(NonStrictPred, LHS, RHS, InLoopSucc);
3195     NonTrivialUnswitchCandidate Candidate(Prev->Term, { LHS, RHS },
3196                                           std::nullopt, std::move(ToInject));
3197     UnswitchCandidates.push_back(std::move(Candidate));
3198   }
3199   return true;
3200 }
3201 
3202 /// Collect unswitch candidates by invariant conditions that are not immediately
3203 /// present in the loop. However, they can be injected into the code if we
3204 /// decide it's profitable.
3205 /// An example of such conditions is following:
3206 ///
3207 ///   for (...) {
3208 ///     x = load ...
3209 ///     if (! x <u C1) break;
3210 ///     if (! x <u C2) break;
3211 ///     <do something>
3212 ///   }
3213 ///
3214 /// We can unswitch by condition "C1 <=u C2". If that is true, then "x <u C1 <=
3215 /// C2" automatically implies "x <u C2", so we can get rid of one of
3216 /// loop-variant checks in unswitched loop version.
collectUnswitchCandidatesWithInjections(SmallVectorImpl<NonTrivialUnswitchCandidate> & UnswitchCandidates,IVConditionInfo & PartialIVInfo,Instruction * & PartialIVCondBranch,Loop & L,const DominatorTree & DT,const LoopInfo & LI,AAResults & AA,const MemorySSAUpdater * MSSAU)3217 static bool collectUnswitchCandidatesWithInjections(
3218     SmallVectorImpl<NonTrivialUnswitchCandidate> &UnswitchCandidates,
3219     IVConditionInfo &PartialIVInfo, Instruction *&PartialIVCondBranch, Loop &L,
3220     const DominatorTree &DT, const LoopInfo &LI, AAResults &AA,
3221     const MemorySSAUpdater *MSSAU) {
3222   if (!InjectInvariantConditions)
3223     return false;
3224 
3225   if (!DT.isReachableFromEntry(L.getHeader()))
3226     return false;
3227   auto *Latch = L.getLoopLatch();
3228   // Need to have a single latch and a preheader.
3229   if (!Latch)
3230     return false;
3231   assert(L.getLoopPreheader() && "Must have a preheader!");
3232 
3233   DenseMap<Value *, SmallVector<CompareDesc, 4> > CandidatesULT;
3234   // Traverse the conditions that dominate latch (and therefore dominate each
3235   // other).
3236   for (auto *DTN = DT.getNode(Latch); L.contains(DTN->getBlock());
3237        DTN = DTN->getIDom()) {
3238     ICmpInst::Predicate Pred;
3239     Value *LHS = nullptr, *RHS = nullptr;
3240     BasicBlock *IfTrue = nullptr, *IfFalse = nullptr;
3241     auto *BB = DTN->getBlock();
3242     // Ignore inner loops.
3243     if (LI.getLoopFor(BB) != &L)
3244       continue;
3245     auto *Term = BB->getTerminator();
3246     if (!match(Term, m_Br(m_ICmp(Pred, m_Value(LHS), m_Value(RHS)),
3247                           m_BasicBlock(IfTrue), m_BasicBlock(IfFalse))))
3248       continue;
3249     if (!LHS->getType()->isIntegerTy())
3250       continue;
3251     canonicalizeForInvariantConditionInjection(Pred, LHS, RHS, IfTrue, IfFalse,
3252                                                L);
3253     if (!shouldTryInjectInvariantCondition(Pred, LHS, RHS, IfTrue, IfFalse, L))
3254       continue;
3255     if (!shouldTryInjectBasingOnMetadata(cast<BranchInst>(Term), IfTrue))
3256       continue;
3257     // Strip ZEXT for unsigned predicate.
3258     // TODO: once signed predicates are supported, also strip SEXT.
3259     CompareDesc Desc(cast<BranchInst>(Term), RHS, IfTrue);
3260     while (auto *Zext = dyn_cast<ZExtInst>(LHS))
3261       LHS = Zext->getOperand(0);
3262     CandidatesULT[LHS].push_back(Desc);
3263   }
3264 
3265   bool Found = false;
3266   for (auto &It : CandidatesULT)
3267     Found |= insertCandidatesWithPendingInjections(
3268         UnswitchCandidates, L, ICmpInst::ICMP_ULT, It.second, DT);
3269   return Found;
3270 }
3271 
isSafeForNoNTrivialUnswitching(Loop & L,LoopInfo & LI)3272 static bool isSafeForNoNTrivialUnswitching(Loop &L, LoopInfo &LI) {
3273   if (!L.isSafeToClone())
3274     return false;
3275   for (auto *BB : L.blocks())
3276     for (auto &I : *BB) {
3277       if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
3278         return false;
3279       if (auto *CB = dyn_cast<CallBase>(&I)) {
3280         assert(!CB->cannotDuplicate() && "Checked by L.isSafeToClone().");
3281         if (CB->isConvergent())
3282           return false;
3283       }
3284     }
3285 
3286   // Check if there are irreducible CFG cycles in this loop. If so, we cannot
3287   // easily unswitch non-trivial edges out of the loop. Doing so might turn the
3288   // irreducible control flow into reducible control flow and introduce new
3289   // loops "out of thin air". If we ever discover important use cases for doing
3290   // this, we can add support to loop unswitch, but it is a lot of complexity
3291   // for what seems little or no real world benefit.
3292   LoopBlocksRPO RPOT(&L);
3293   RPOT.perform(&LI);
3294   if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
3295     return false;
3296 
3297   SmallVector<BasicBlock *, 4> ExitBlocks;
3298   L.getUniqueExitBlocks(ExitBlocks);
3299   // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch
3300   // instruction as we don't know how to split those exit blocks.
3301   // FIXME: We should teach SplitBlock to handle this and remove this
3302   // restriction.
3303   for (auto *ExitBB : ExitBlocks) {
3304     auto *I = ExitBB->getFirstNonPHI();
3305     if (isa<CleanupPadInst>(I) || isa<CatchSwitchInst>(I)) {
3306       LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch "
3307                            "in exit block\n");
3308       return false;
3309     }
3310   }
3311 
3312   return true;
3313 }
3314 
findBestNonTrivialUnswitchCandidate(ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates,const Loop & L,const DominatorTree & DT,const LoopInfo & LI,AssumptionCache & AC,const TargetTransformInfo & TTI,const IVConditionInfo & PartialIVInfo)3315 static NonTrivialUnswitchCandidate findBestNonTrivialUnswitchCandidate(
3316     ArrayRef<NonTrivialUnswitchCandidate> UnswitchCandidates, const Loop &L,
3317     const DominatorTree &DT, const LoopInfo &LI, AssumptionCache &AC,
3318     const TargetTransformInfo &TTI, const IVConditionInfo &PartialIVInfo) {
3319   // Given that unswitching these terminators will require duplicating parts of
3320   // the loop, so we need to be able to model that cost. Compute the ephemeral
3321   // values and set up a data structure to hold per-BB costs. We cache each
3322   // block's cost so that we don't recompute this when considering different
3323   // subsets of the loop for duplication during unswitching.
3324   SmallPtrSet<const Value *, 4> EphValues;
3325   CodeMetrics::collectEphemeralValues(&L, &AC, EphValues);
3326   SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap;
3327 
3328   // Compute the cost of each block, as well as the total loop cost. Also, bail
3329   // out if we see instructions which are incompatible with loop unswitching
3330   // (convergent, noduplicate, or cross-basic-block tokens).
3331   // FIXME: We might be able to safely handle some of these in non-duplicated
3332   // regions.
3333   TargetTransformInfo::TargetCostKind CostKind =
3334       L.getHeader()->getParent()->hasMinSize()
3335       ? TargetTransformInfo::TCK_CodeSize
3336       : TargetTransformInfo::TCK_SizeAndLatency;
3337   InstructionCost LoopCost = 0;
3338   for (auto *BB : L.blocks()) {
3339     InstructionCost Cost = 0;
3340     for (auto &I : *BB) {
3341       if (EphValues.count(&I))
3342         continue;
3343       Cost += TTI.getInstructionCost(&I, CostKind);
3344     }
3345     assert(Cost >= 0 && "Must not have negative costs!");
3346     LoopCost += Cost;
3347     assert(LoopCost >= 0 && "Must not have negative loop costs!");
3348     BBCostMap[BB] = Cost;
3349   }
3350   LLVM_DEBUG(dbgs() << "  Total loop cost: " << LoopCost << "\n");
3351 
3352   // Now we find the best candidate by searching for the one with the following
3353   // properties in order:
3354   //
3355   // 1) An unswitching cost below the threshold
3356   // 2) The smallest number of duplicated unswitch candidates (to avoid
3357   //    creating redundant subsequent unswitching)
3358   // 3) The smallest cost after unswitching.
3359   //
3360   // We prioritize reducing fanout of unswitch candidates provided the cost
3361   // remains below the threshold because this has a multiplicative effect.
3362   //
3363   // This requires memoizing each dominator subtree to avoid redundant work.
3364   //
3365   // FIXME: Need to actually do the number of candidates part above.
3366   SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap;
3367   // Given a terminator which might be unswitched, computes the non-duplicated
3368   // cost for that terminator.
3369   auto ComputeUnswitchedCost = [&](Instruction &TI,
3370                                    bool FullUnswitch) -> InstructionCost {
3371     // Unswitching selects unswitches the entire loop.
3372     if (isa<SelectInst>(TI))
3373       return LoopCost;
3374 
3375     BasicBlock &BB = *TI.getParent();
3376     SmallPtrSet<BasicBlock *, 4> Visited;
3377 
3378     InstructionCost Cost = 0;
3379     for (BasicBlock *SuccBB : successors(&BB)) {
3380       // Don't count successors more than once.
3381       if (!Visited.insert(SuccBB).second)
3382         continue;
3383 
3384       // If this is a partial unswitch candidate, then it must be a conditional
3385       // branch with a condition of either `or`, `and`, their corresponding
3386       // select forms or partially invariant instructions. In that case, one of
3387       // the successors is necessarily duplicated, so don't even try to remove
3388       // its cost.
3389       if (!FullUnswitch) {
3390         auto &BI = cast<BranchInst>(TI);
3391         Value *Cond = skipTrivialSelect(BI.getCondition());
3392         if (match(Cond, m_LogicalAnd())) {
3393           if (SuccBB == BI.getSuccessor(1))
3394             continue;
3395         } else if (match(Cond, m_LogicalOr())) {
3396           if (SuccBB == BI.getSuccessor(0))
3397             continue;
3398         } else if ((PartialIVInfo.KnownValue->isOneValue() &&
3399                     SuccBB == BI.getSuccessor(0)) ||
3400                    (!PartialIVInfo.KnownValue->isOneValue() &&
3401                     SuccBB == BI.getSuccessor(1)))
3402           continue;
3403       }
3404 
3405       // This successor's domtree will not need to be duplicated after
3406       // unswitching if the edge to the successor dominates it (and thus the
3407       // entire tree). This essentially means there is no other path into this
3408       // subtree and so it will end up live in only one clone of the loop.
3409       if (SuccBB->getUniquePredecessor() ||
3410           llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) {
3411             return PredBB == &BB || DT.dominates(SuccBB, PredBB);
3412           })) {
3413         Cost += computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap);
3414         assert(Cost <= LoopCost &&
3415                "Non-duplicated cost should never exceed total loop cost!");
3416       }
3417     }
3418 
3419     // Now scale the cost by the number of unique successors minus one. We
3420     // subtract one because there is already at least one copy of the entire
3421     // loop. This is computing the new cost of unswitching a condition.
3422     // Note that guards always have 2 unique successors that are implicit and
3423     // will be materialized if we decide to unswitch it.
3424     int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
3425     assert(SuccessorsCount > 1 &&
3426            "Cannot unswitch a condition without multiple distinct successors!");
3427     return (LoopCost - Cost) * (SuccessorsCount - 1);
3428   };
3429 
3430   std::optional<NonTrivialUnswitchCandidate> Best;
3431   for (auto &Candidate : UnswitchCandidates) {
3432     Instruction &TI = *Candidate.TI;
3433     ArrayRef<Value *> Invariants = Candidate.Invariants;
3434     BranchInst *BI = dyn_cast<BranchInst>(&TI);
3435     bool FullUnswitch =
3436         !BI || Candidate.hasPendingInjection() ||
3437         (Invariants.size() == 1 &&
3438          Invariants[0] == skipTrivialSelect(BI->getCondition()));
3439     InstructionCost CandidateCost = ComputeUnswitchedCost(TI, FullUnswitch);
3440     // Calculate cost multiplier which is a tool to limit potentially
3441     // exponential behavior of loop-unswitch.
3442     if (EnableUnswitchCostMultiplier) {
3443       int CostMultiplier =
3444           CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates);
3445       assert(
3446           (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&
3447           "cost multiplier needs to be in the range of 1..UnswitchThreshold");
3448       CandidateCost *= CostMultiplier;
3449       LLVM_DEBUG(dbgs() << "  Computed cost of " << CandidateCost
3450                         << " (multiplier: " << CostMultiplier << ")"
3451                         << " for unswitch candidate: " << TI << "\n");
3452     } else {
3453       LLVM_DEBUG(dbgs() << "  Computed cost of " << CandidateCost
3454                         << " for unswitch candidate: " << TI << "\n");
3455     }
3456 
3457     if (!Best || CandidateCost < Best->Cost) {
3458       Best = Candidate;
3459       Best->Cost = CandidateCost;
3460     }
3461   }
3462   assert(Best && "Must be!");
3463   return *Best;
3464 }
3465 
3466 // Insert a freeze on an unswitched branch if all is true:
3467 // 1. freeze-loop-unswitch-cond option is true
3468 // 2. The branch may not execute in the loop pre-transformation. If a branch may
3469 // not execute and could cause UB, it would always cause UB if it is hoisted outside
3470 // of the loop. Insert a freeze to prevent this case.
3471 // 3. The branch condition may be poison or undef
shouldInsertFreeze(Loop & L,Instruction & TI,DominatorTree & DT,AssumptionCache & AC)3472 static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT,
3473                                AssumptionCache &AC) {
3474   assert(isa<BranchInst>(TI) || isa<SwitchInst>(TI));
3475   if (!FreezeLoopUnswitchCond)
3476     return false;
3477 
3478   ICFLoopSafetyInfo SafetyInfo;
3479   SafetyInfo.computeLoopSafetyInfo(&L);
3480   if (SafetyInfo.isGuaranteedToExecute(TI, &DT, &L))
3481     return false;
3482 
3483   Value *Cond;
3484   if (BranchInst *BI = dyn_cast<BranchInst>(&TI))
3485     Cond = skipTrivialSelect(BI->getCondition());
3486   else
3487     Cond = skipTrivialSelect(cast<SwitchInst>(&TI)->getCondition());
3488   return !isGuaranteedNotToBeUndefOrPoison(
3489       Cond, &AC, L.getLoopPreheader()->getTerminator(), &DT);
3490 }
3491 
unswitchBestCondition(Loop & L,DominatorTree & DT,LoopInfo & LI,AssumptionCache & AC,AAResults & AA,TargetTransformInfo & TTI,ScalarEvolution * SE,MemorySSAUpdater * MSSAU,LPMUpdater & LoopUpdater)3492 static bool unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
3493                                   AssumptionCache &AC, AAResults &AA,
3494                                   TargetTransformInfo &TTI, ScalarEvolution *SE,
3495                                   MemorySSAUpdater *MSSAU,
3496                                   LPMUpdater &LoopUpdater) {
3497   // Collect all invariant conditions within this loop (as opposed to an inner
3498   // loop which would be handled when visiting that inner loop).
3499   SmallVector<NonTrivialUnswitchCandidate, 4> UnswitchCandidates;
3500   IVConditionInfo PartialIVInfo;
3501   Instruction *PartialIVCondBranch = nullptr;
3502   collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo,
3503                             PartialIVCondBranch, L, LI, AA, MSSAU);
3504   if (!findOptionMDForLoop(&L, "llvm.loop.unswitch.injection.disable"))
3505     collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo,
3506                                             PartialIVCondBranch, L, DT, LI, AA,
3507                                             MSSAU);
3508   // If we didn't find any candidates, we're done.
3509   if (UnswitchCandidates.empty())
3510     return false;
3511 
3512   LLVM_DEBUG(
3513       dbgs() << "Considering " << UnswitchCandidates.size()
3514              << " non-trivial loop invariant conditions for unswitching.\n");
3515 
3516   NonTrivialUnswitchCandidate Best = findBestNonTrivialUnswitchCandidate(
3517       UnswitchCandidates, L, DT, LI, AC, TTI, PartialIVInfo);
3518 
3519   assert(Best.TI && "Failed to find loop unswitch candidate");
3520   assert(Best.Cost && "Failed to compute cost");
3521 
3522   if (*Best.Cost >= UnswitchThreshold) {
3523     LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: " << *Best.Cost
3524                       << "\n");
3525     return false;
3526   }
3527 
3528   bool InjectedCondition = false;
3529   if (Best.hasPendingInjection()) {
3530     Best = injectPendingInvariantConditions(Best, L, DT, LI, AC, MSSAU);
3531     InjectedCondition = true;
3532   }
3533   assert(!Best.hasPendingInjection() &&
3534          "All injections should have been done by now!");
3535 
3536   if (Best.TI != PartialIVCondBranch)
3537     PartialIVInfo.InstToDuplicate.clear();
3538 
3539   bool InsertFreeze;
3540   if (auto *SI = dyn_cast<SelectInst>(Best.TI)) {
3541     // If the best candidate is a select, turn it into a branch. Select
3542     // instructions with a poison conditional do not propagate poison, but
3543     // branching on poison causes UB. Insert a freeze on the select
3544     // conditional to prevent UB after turning the select into a branch.
3545     InsertFreeze = !isGuaranteedNotToBeUndefOrPoison(
3546         SI->getCondition(), &AC, L.getLoopPreheader()->getTerminator(), &DT);
3547     Best.TI = turnSelectIntoBranch(SI, DT, LI, MSSAU, &AC);
3548   } else {
3549     // If the best candidate is a guard, turn it into a branch.
3550     if (isGuard(Best.TI))
3551       Best.TI =
3552           turnGuardIntoBranch(cast<IntrinsicInst>(Best.TI), L, DT, LI, MSSAU);
3553     InsertFreeze = shouldInsertFreeze(L, *Best.TI, DT, AC);
3554   }
3555 
3556   LLVM_DEBUG(dbgs() << "  Unswitching non-trivial (cost = " << Best.Cost
3557                     << ") terminator: " << *Best.TI << "\n");
3558   unswitchNontrivialInvariants(L, *Best.TI, Best.Invariants, PartialIVInfo, DT,
3559                                LI, AC, SE, MSSAU, LoopUpdater, InsertFreeze,
3560                                InjectedCondition);
3561   return true;
3562 }
3563 
3564 /// Unswitch control flow predicated on loop invariant conditions.
3565 ///
3566 /// This first hoists all branches or switches which are trivial (IE, do not
3567 /// require duplicating any part of the loop) out of the loop body. It then
3568 /// looks at other loop invariant control flows and tries to unswitch those as
3569 /// well by cloning the loop if the result is small enough.
3570 ///
3571 /// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are
3572 /// also updated based on the unswitch. The `MSSA` analysis is also updated if
3573 /// valid (i.e. its use is enabled).
3574 ///
3575 /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is
3576 /// true, we will attempt to do non-trivial unswitching as well as trivial
3577 /// unswitching.
3578 ///
3579 /// The `postUnswitch` function will be run after unswitching is complete
3580 /// with information on whether or not the provided loop remains a loop and
3581 /// a list of new sibling loops created.
3582 ///
3583 /// If `SE` is non-null, we will update that analysis based on the unswitching
3584 /// done.
unswitchLoop(Loop & L,DominatorTree & DT,LoopInfo & LI,AssumptionCache & AC,AAResults & AA,TargetTransformInfo & TTI,bool Trivial,bool NonTrivial,ScalarEvolution * SE,MemorySSAUpdater * MSSAU,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI,LPMUpdater & LoopUpdater)3585 static bool unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI,
3586                          AssumptionCache &AC, AAResults &AA,
3587                          TargetTransformInfo &TTI, bool Trivial,
3588                          bool NonTrivial, ScalarEvolution *SE,
3589                          MemorySSAUpdater *MSSAU, ProfileSummaryInfo *PSI,
3590                          BlockFrequencyInfo *BFI, LPMUpdater &LoopUpdater) {
3591   assert(L.isRecursivelyLCSSAForm(DT, LI) &&
3592          "Loops must be in LCSSA form before unswitching.");
3593 
3594   // Must be in loop simplified form: we need a preheader and dedicated exits.
3595   if (!L.isLoopSimplifyForm())
3596     return false;
3597 
3598   // Try trivial unswitch first before loop over other basic blocks in the loop.
3599   if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
3600     // If we unswitched successfully we will want to clean up the loop before
3601     // processing it further so just mark it as unswitched and return.
3602     postUnswitch(L, LoopUpdater, L.getName(),
3603                  /*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false,
3604                  /*InjectedCondition*/ false, {});
3605     return true;
3606   }
3607 
3608   const Function *F = L.getHeader()->getParent();
3609 
3610   // Check whether we should continue with non-trivial conditions.
3611   // EnableNonTrivialUnswitch: Global variable that forces non-trivial
3612   //                           unswitching for testing and debugging.
3613   // NonTrivial: Parameter that enables non-trivial unswitching for this
3614   //             invocation of the transform. But this should be allowed only
3615   //             for targets without branch divergence.
3616   //
3617   // FIXME: If divergence analysis becomes available to a loop
3618   // transform, we should allow unswitching for non-trivial uniform
3619   // branches even on targets that have divergence.
3620   // https://bugs.llvm.org/show_bug.cgi?id=48819
3621   bool ContinueWithNonTrivial =
3622       EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence(F));
3623   if (!ContinueWithNonTrivial)
3624     return false;
3625 
3626   // Skip non-trivial unswitching for optsize functions.
3627   if (F->hasOptSize())
3628     return false;
3629 
3630   // Returns true if Loop L's loop nest is cold, i.e. if the headers of L,
3631   // of the loops L is nested in, and of the loops nested in L are all cold.
3632   auto IsLoopNestCold = [&](const Loop *L) {
3633     // Check L and all of its parent loops.
3634     auto *Parent = L;
3635     while (Parent) {
3636       if (!PSI->isColdBlock(Parent->getHeader(), BFI))
3637         return false;
3638       Parent = Parent->getParentLoop();
3639     }
3640     // Next check all loops nested within L.
3641     SmallVector<const Loop *, 4> Worklist;
3642     Worklist.insert(Worklist.end(), L->getSubLoops().begin(),
3643                     L->getSubLoops().end());
3644     while (!Worklist.empty()) {
3645       auto *CurLoop = Worklist.pop_back_val();
3646       if (!PSI->isColdBlock(CurLoop->getHeader(), BFI))
3647         return false;
3648       Worklist.insert(Worklist.end(), CurLoop->getSubLoops().begin(),
3649                       CurLoop->getSubLoops().end());
3650     }
3651     return true;
3652   };
3653 
3654   // Skip cold loops in cold loop nests, as unswitching them brings little
3655   // benefit but increases the code size
3656   if (PSI && PSI->hasProfileSummary() && BFI && IsLoopNestCold(&L)) {
3657     LLVM_DEBUG(dbgs() << " Skip cold loop: " << L << "\n");
3658     return false;
3659   }
3660 
3661   // Perform legality checks.
3662   if (!isSafeForNoNTrivialUnswitching(L, LI))
3663     return false;
3664 
3665   // For non-trivial unswitching, because it often creates new loops, we rely on
3666   // the pass manager to iterate on the loops rather than trying to immediately
3667   // reach a fixed point. There is no substantial advantage to iterating
3668   // internally, and if any of the new loops are simplified enough to contain
3669   // trivial unswitching we want to prefer those.
3670 
3671   // Try to unswitch the best invariant condition. We prefer this full unswitch to
3672   // a partial unswitch when possible below the threshold.
3673   if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, SE, MSSAU, LoopUpdater))
3674     return true;
3675 
3676   // No other opportunities to unswitch.
3677   return false;
3678 }
3679 
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)3680 PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
3681                                               LoopStandardAnalysisResults &AR,
3682                                               LPMUpdater &U) {
3683   Function &F = *L.getHeader()->getParent();
3684   (void)F;
3685   ProfileSummaryInfo *PSI = nullptr;
3686   if (auto OuterProxy =
3687           AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR)
3688               .getCachedResult<ModuleAnalysisManagerFunctionProxy>(F))
3689     PSI = OuterProxy->getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
3690   LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << L
3691                     << "\n");
3692 
3693   std::optional<MemorySSAUpdater> MSSAU;
3694   if (AR.MSSA) {
3695     MSSAU = MemorySSAUpdater(AR.MSSA);
3696     if (VerifyMemorySSA)
3697       AR.MSSA->verifyMemorySSA();
3698   }
3699   if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.AA, AR.TTI, Trivial, NonTrivial,
3700                     &AR.SE, MSSAU ? &*MSSAU : nullptr, PSI, AR.BFI, U))
3701     return PreservedAnalyses::all();
3702 
3703   if (AR.MSSA && VerifyMemorySSA)
3704     AR.MSSA->verifyMemorySSA();
3705 
3706   // Historically this pass has had issues with the dominator tree so verify it
3707   // in asserts builds.
3708   assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast));
3709 
3710   auto PA = getLoopPassPreservedAnalyses();
3711   if (AR.MSSA)
3712     PA.preserve<MemorySSAAnalysis>();
3713   return PA;
3714 }
3715 
printPipeline(raw_ostream & OS,function_ref<StringRef (StringRef)> MapClassName2PassName)3716 void SimpleLoopUnswitchPass::printPipeline(
3717     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
3718   static_cast<PassInfoMixin<SimpleLoopUnswitchPass> *>(this)->printPipeline(
3719       OS, MapClassName2PassName);
3720 
3721   OS << '<';
3722   OS << (NonTrivial ? "" : "no-") << "nontrivial;";
3723   OS << (Trivial ? "" : "no-") << "trivial";
3724   OS << '>';
3725 }
3726