xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LICM.cpp (revision cc68614da8232d8baaca0ae0d0dd8f890f06623e)
1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
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 // This pass performs loop invariant code motion, attempting to remove as much
10 // code from the body of a loop as possible.  It does this by either hoisting
11 // code into the preheader block, or by sinking code to the exit blocks if it is
12 // safe.  This pass also promotes must-aliased memory locations in the loop to
13 // live in registers, thus hoisting and sinking "invariant" loads and stores.
14 //
15 // Hoisting operations out of loops is a canonicalization transform.  It
16 // enables and simplifies subsequent optimizations in the middle-end.
17 // Rematerialization of hoisted instructions to reduce register pressure is the
18 // responsibility of the back-end, which has more accurate information about
19 // register pressure and also handles other optimizations than LICM that
20 // increase live-ranges.
21 //
22 // This pass uses alias analysis for two purposes:
23 //
24 //  1. Moving loop invariant loads and calls out of loops.  If we can determine
25 //     that a load or call inside of a loop never aliases anything stored to,
26 //     we can hoist it or sink it like any other instruction.
27 //  2. Scalar Promotion of Memory - If there is a store instruction inside of
28 //     the loop, we try to move the store to happen AFTER the loop instead of
29 //     inside of the loop.  This can only happen if a few conditions are true:
30 //       A. The pointer stored through is loop invariant
31 //       B. There are no stores or loads in the loop which _may_ alias the
32 //          pointer.  There are no calls in the loop which mod/ref the pointer.
33 //     If these conditions are true, we can promote the loads and stores in the
34 //     loop of the pointer to use a temporary alloca'd variable.  We then use
35 //     the SSAUpdater to construct the appropriate SSA form for the value.
36 //
37 //===----------------------------------------------------------------------===//
38 
39 #include "llvm/Transforms/Scalar/LICM.h"
40 #include "llvm/ADT/SetOperations.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/Analysis/AliasAnalysis.h"
43 #include "llvm/Analysis/AliasSetTracker.h"
44 #include "llvm/Analysis/BasicAliasAnalysis.h"
45 #include "llvm/Analysis/BlockFrequencyInfo.h"
46 #include "llvm/Analysis/CaptureTracking.h"
47 #include "llvm/Analysis/ConstantFolding.h"
48 #include "llvm/Analysis/GlobalsModRef.h"
49 #include "llvm/Analysis/GuardUtils.h"
50 #include "llvm/Analysis/LazyBlockFrequencyInfo.h"
51 #include "llvm/Analysis/Loads.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopIterator.h"
54 #include "llvm/Analysis/LoopPass.h"
55 #include "llvm/Analysis/MemoryBuiltins.h"
56 #include "llvm/Analysis/MemorySSA.h"
57 #include "llvm/Analysis/MemorySSAUpdater.h"
58 #include "llvm/Analysis/MustExecute.h"
59 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
60 #include "llvm/Analysis/ScalarEvolution.h"
61 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
62 #include "llvm/Analysis/TargetLibraryInfo.h"
63 #include "llvm/Analysis/ValueTracking.h"
64 #include "llvm/IR/CFG.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DebugInfoMetadata.h"
68 #include "llvm/IR/DerivedTypes.h"
69 #include "llvm/IR/Dominators.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/IntrinsicInst.h"
72 #include "llvm/IR/LLVMContext.h"
73 #include "llvm/IR/Metadata.h"
74 #include "llvm/IR/PatternMatch.h"
75 #include "llvm/IR/PredIteratorCache.h"
76 #include "llvm/InitializePasses.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/Debug.h"
79 #include "llvm/Support/raw_ostream.h"
80 #include "llvm/Transforms/Scalar.h"
81 #include "llvm/Transforms/Scalar/LoopPassManager.h"
82 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
83 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
84 #include "llvm/Transforms/Utils/Local.h"
85 #include "llvm/Transforms/Utils/LoopUtils.h"
86 #include "llvm/Transforms/Utils/SSAUpdater.h"
87 #include <algorithm>
88 #include <utility>
89 using namespace llvm;
90 
91 #define DEBUG_TYPE "licm"
92 
93 STATISTIC(NumCreatedBlocks, "Number of blocks created");
94 STATISTIC(NumClonedBranches, "Number of branches cloned");
95 STATISTIC(NumSunk, "Number of instructions sunk out of loop");
96 STATISTIC(NumHoisted, "Number of instructions hoisted out of loop");
97 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
98 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
99 STATISTIC(NumPromoted, "Number of memory locations promoted to registers");
100 
101 /// Memory promotion is enabled by default.
102 static cl::opt<bool>
103     DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false),
104                      cl::desc("Disable memory promotion in LICM pass"));
105 
106 static cl::opt<bool> ControlFlowHoisting(
107     "licm-control-flow-hoisting", cl::Hidden, cl::init(false),
108     cl::desc("Enable control flow (and PHI) hoisting in LICM"));
109 
110 static cl::opt<unsigned> HoistSinkColdnessThreshold(
111     "licm-coldness-threshold", cl::Hidden, cl::init(4),
112     cl::desc("Relative coldness Threshold of hoisting/sinking destination "
113              "block for LICM to be considered beneficial"));
114 
115 static cl::opt<uint32_t> MaxNumUsesTraversed(
116     "licm-max-num-uses-traversed", cl::Hidden, cl::init(8),
117     cl::desc("Max num uses visited for identifying load "
118              "invariance in loop using invariant start (default = 8)"));
119 
120 // Default value of zero implies we use the regular alias set tracker mechanism
121 // instead of the cross product using AA to identify aliasing of the memory
122 // location we are interested in.
123 static cl::opt<int>
124 LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0),
125                cl::desc("How many instruction to cross product using AA"));
126 
127 // Experimental option to allow imprecision in LICM in pathological cases, in
128 // exchange for faster compile. This is to be removed if MemorySSA starts to
129 // address the same issue. This flag applies only when LICM uses MemorySSA
130 // instead on AliasSetTracker. LICM calls MemorySSAWalker's
131 // getClobberingMemoryAccess, up to the value of the Cap, getting perfect
132 // accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess,
133 // which may not be precise, since optimizeUses is capped. The result is
134 // correct, but we may not get as "far up" as possible to get which access is
135 // clobbering the one queried.
136 cl::opt<unsigned> llvm::SetLicmMssaOptCap(
137     "licm-mssa-optimization-cap", cl::init(100), cl::Hidden,
138     cl::desc("Enable imprecision in LICM in pathological cases, in exchange "
139              "for faster compile. Caps the MemorySSA clobbering calls."));
140 
141 // Experimentally, memory promotion carries less importance than sinking and
142 // hoisting. Limit when we do promotion when using MemorySSA, in order to save
143 // compile time.
144 cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap(
145     "licm-mssa-max-acc-promotion", cl::init(250), cl::Hidden,
146     cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no "
147              "effect. When MSSA in LICM is enabled, then this is the maximum "
148              "number of accesses allowed to be present in a loop in order to "
149              "enable memory promotion."));
150 
151 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
152 static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
153                                   const LoopSafetyInfo *SafetyInfo,
154                                   TargetTransformInfo *TTI, bool &FreeInLoop);
155 static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
156                   BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
157                   MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
158                   OptimizationRemarkEmitter *ORE);
159 static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
160                  BlockFrequencyInfo *BFI, const Loop *CurLoop,
161                  ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU,
162                  OptimizationRemarkEmitter *ORE);
163 static bool isSafeToExecuteUnconditionally(Instruction &Inst,
164                                            const DominatorTree *DT,
165                                            const TargetLibraryInfo *TLI,
166                                            const Loop *CurLoop,
167                                            const LoopSafetyInfo *SafetyInfo,
168                                            OptimizationRemarkEmitter *ORE,
169                                            const Instruction *CtxI = nullptr);
170 static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
171                                      AliasSetTracker *CurAST, Loop *CurLoop,
172                                      AAResults *AA);
173 static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
174                                              Loop *CurLoop, Instruction &I,
175                                              SinkAndHoistLICMFlags &Flags);
176 static bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA,
177                                               MemoryUse &MU);
178 static Instruction *cloneInstructionInExitBlock(
179     Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
180     const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU);
181 
182 static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
183                              AliasSetTracker *AST, MemorySSAUpdater *MSSAU);
184 
185 static void moveInstructionBefore(Instruction &I, Instruction &Dest,
186                                   ICFLoopSafetyInfo &SafetyInfo,
187                                   MemorySSAUpdater *MSSAU, ScalarEvolution *SE);
188 
189 static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L,
190                                 function_ref<void(Instruction *)> Fn);
191 static SmallVector<SmallSetVector<Value *, 8>, 0>
192 collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L);
193 
194 namespace {
195 struct LoopInvariantCodeMotion {
196   bool runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT,
197                  BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI,
198                  TargetTransformInfo *TTI, ScalarEvolution *SE, MemorySSA *MSSA,
199                  OptimizationRemarkEmitter *ORE, bool LoopNestMode = false);
200 
201   LoopInvariantCodeMotion(unsigned LicmMssaOptCap,
202                           unsigned LicmMssaNoAccForPromotionCap)
203       : LicmMssaOptCap(LicmMssaOptCap),
204         LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {}
205 
206 private:
207   unsigned LicmMssaOptCap;
208   unsigned LicmMssaNoAccForPromotionCap;
209 
210   std::unique_ptr<AliasSetTracker>
211   collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AAResults *AA);
212 };
213 
214 struct LegacyLICMPass : public LoopPass {
215   static char ID; // Pass identification, replacement for typeid
216   LegacyLICMPass(
217       unsigned LicmMssaOptCap = SetLicmMssaOptCap,
218       unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap)
219       : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) {
220     initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
221   }
222 
223   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
224     if (skipLoop(L))
225       return false;
226 
227     LLVM_DEBUG(dbgs() << "Perform LICM on Loop with header at block "
228                       << L->getHeader()->getNameOrAsOperand() << "\n");
229 
230     auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
231     MemorySSA *MSSA = EnableMSSALoopDependency
232                           ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA())
233                           : nullptr;
234     bool hasProfileData = L->getHeader()->getParent()->hasProfileData();
235     BlockFrequencyInfo *BFI =
236         hasProfileData ? &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI()
237                        : nullptr;
238     // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
239     // pass. Function analyses need to be preserved across loop transformations
240     // but ORE cannot be preserved (see comment before the pass definition).
241     OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
242     return LICM.runOnLoop(
243         L, &getAnalysis<AAResultsWrapperPass>().getAAResults(),
244         &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
245         &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), BFI,
246         &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
247             *L->getHeader()->getParent()),
248         &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
249             *L->getHeader()->getParent()),
250         SE ? &SE->getSE() : nullptr, MSSA, &ORE);
251   }
252 
253   /// This transformation requires natural loop information & requires that
254   /// loop preheaders be inserted into the CFG...
255   ///
256   void getAnalysisUsage(AnalysisUsage &AU) const override {
257     AU.addPreserved<DominatorTreeWrapperPass>();
258     AU.addPreserved<LoopInfoWrapperPass>();
259     AU.addRequired<TargetLibraryInfoWrapperPass>();
260     if (EnableMSSALoopDependency) {
261       AU.addRequired<MemorySSAWrapperPass>();
262       AU.addPreserved<MemorySSAWrapperPass>();
263     }
264     AU.addRequired<TargetTransformInfoWrapperPass>();
265     getLoopAnalysisUsage(AU);
266     LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
267     AU.addPreserved<LazyBlockFrequencyInfoPass>();
268     AU.addPreserved<LazyBranchProbabilityInfoPass>();
269   }
270 
271 private:
272   LoopInvariantCodeMotion LICM;
273 };
274 } // namespace
275 
276 PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM,
277                                 LoopStandardAnalysisResults &AR, LPMUpdater &) {
278   // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
279   // pass.  Function analyses need to be preserved across loop transformations
280   // but ORE cannot be preserved (see comment before the pass definition).
281   OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
282 
283   LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
284   if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, AR.BFI, &AR.TLI, &AR.TTI,
285                       &AR.SE, AR.MSSA, &ORE))
286     return PreservedAnalyses::all();
287 
288   auto PA = getLoopPassPreservedAnalyses();
289 
290   PA.preserve<DominatorTreeAnalysis>();
291   PA.preserve<LoopAnalysis>();
292   if (AR.MSSA)
293     PA.preserve<MemorySSAAnalysis>();
294 
295   return PA;
296 }
297 
298 PreservedAnalyses LNICMPass::run(LoopNest &LN, LoopAnalysisManager &AM,
299                                  LoopStandardAnalysisResults &AR,
300                                  LPMUpdater &) {
301   // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
302   // pass.  Function analyses need to be preserved across loop transformations
303   // but ORE cannot be preserved (see comment before the pass definition).
304   OptimizationRemarkEmitter ORE(LN.getParent());
305 
306   LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
307 
308   Loop &OutermostLoop = LN.getOutermostLoop();
309   bool Changed = LICM.runOnLoop(&OutermostLoop, &AR.AA, &AR.LI, &AR.DT, AR.BFI,
310                                 &AR.TLI, &AR.TTI, &AR.SE, AR.MSSA, &ORE, true);
311 
312   if (!Changed)
313     return PreservedAnalyses::all();
314 
315   auto PA = getLoopPassPreservedAnalyses();
316 
317   PA.preserve<DominatorTreeAnalysis>();
318   PA.preserve<LoopAnalysis>();
319   if (AR.MSSA)
320     PA.preserve<MemorySSAAnalysis>();
321 
322   return PA;
323 }
324 
325 char LegacyLICMPass::ID = 0;
326 INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",
327                       false, false)
328 INITIALIZE_PASS_DEPENDENCY(LoopPass)
329 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
330 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
331 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
332 INITIALIZE_PASS_DEPENDENCY(LazyBFIPass)
333 INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,
334                     false)
335 
336 Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
337 Pass *llvm::createLICMPass(unsigned LicmMssaOptCap,
338                            unsigned LicmMssaNoAccForPromotionCap) {
339   return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
340 }
341 
342 llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(bool IsSink, Loop *L,
343                                                    MemorySSA *MSSA)
344     : SinkAndHoistLICMFlags(SetLicmMssaOptCap, SetLicmMssaNoAccForPromotionCap,
345                             IsSink, L, MSSA) {}
346 
347 llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(
348     unsigned LicmMssaOptCap, unsigned LicmMssaNoAccForPromotionCap, bool IsSink,
349     Loop *L, MemorySSA *MSSA)
350     : LicmMssaOptCap(LicmMssaOptCap),
351       LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap),
352       IsSink(IsSink) {
353   assert(((L != nullptr) == (MSSA != nullptr)) &&
354          "Unexpected values for SinkAndHoistLICMFlags");
355   if (!MSSA)
356     return;
357 
358   unsigned AccessCapCount = 0;
359   for (auto *BB : L->getBlocks())
360     if (const auto *Accesses = MSSA->getBlockAccesses(BB))
361       for (const auto &MA : *Accesses) {
362         (void)MA;
363         ++AccessCapCount;
364         if (AccessCapCount > LicmMssaNoAccForPromotionCap) {
365           NoOfMemAccTooLarge = true;
366           return;
367         }
368       }
369 }
370 
371 /// Hoist expressions out of the specified loop. Note, alias info for inner
372 /// loop is not preserved so it is not a good idea to run LICM multiple
373 /// times on one loop.
374 bool LoopInvariantCodeMotion::runOnLoop(
375     Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT,
376     BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
377     ScalarEvolution *SE, MemorySSA *MSSA, OptimizationRemarkEmitter *ORE,
378     bool LoopNestMode) {
379   bool Changed = false;
380 
381   assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
382 
383   // If this loop has metadata indicating that LICM is not to be performed then
384   // just exit.
385   if (hasDisableLICMTransformsHint(L)) {
386     return false;
387   }
388 
389   std::unique_ptr<AliasSetTracker> CurAST;
390   std::unique_ptr<MemorySSAUpdater> MSSAU;
391   std::unique_ptr<SinkAndHoistLICMFlags> Flags;
392 
393   // Don't sink stores from loops with coroutine suspend instructions.
394   // LICM would sink instructions into the default destination of
395   // the coroutine switch. The default destination of the switch is to
396   // handle the case where the coroutine is suspended, by which point the
397   // coroutine frame may have been destroyed. No instruction can be sunk there.
398   // FIXME: This would unfortunately hurt the performance of coroutines, however
399   // there is currently no general solution for this. Similar issues could also
400   // potentially happen in other passes where instructions are being moved
401   // across that edge.
402   bool HasCoroSuspendInst = llvm::any_of(L->getBlocks(), [](BasicBlock *BB) {
403     return llvm::any_of(*BB, [](Instruction &I) {
404       IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I);
405       return II && II->getIntrinsicID() == Intrinsic::coro_suspend;
406     });
407   });
408 
409   if (!MSSA) {
410     LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n");
411     CurAST = collectAliasInfoForLoop(L, LI, AA);
412     Flags = std::make_unique<SinkAndHoistLICMFlags>(
413         LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true);
414   } else {
415     LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n");
416     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
417     Flags = std::make_unique<SinkAndHoistLICMFlags>(
418         LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true, L, MSSA);
419   }
420 
421   // Get the preheader block to move instructions into...
422   BasicBlock *Preheader = L->getLoopPreheader();
423 
424   // Compute loop safety information.
425   ICFLoopSafetyInfo SafetyInfo;
426   SafetyInfo.computeLoopSafetyInfo(L);
427 
428   // We want to visit all of the instructions in this loop... that are not parts
429   // of our subloops (they have already had their invariants hoisted out of
430   // their loop, into this loop, so there is no need to process the BODIES of
431   // the subloops).
432   //
433   // Traverse the body of the loop in depth first order on the dominator tree so
434   // that we are guaranteed to see definitions before we see uses.  This allows
435   // us to sink instructions in one pass, without iteration.  After sinking
436   // instructions, we perform another pass to hoist them out of the loop.
437   if (L->hasDedicatedExits())
438     Changed |=
439         sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, TTI, L,
440                    CurAST.get(), MSSAU.get(), &SafetyInfo, *Flags.get(), ORE);
441   Flags->setIsSink(false);
442   if (Preheader)
443     Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, L,
444                            CurAST.get(), MSSAU.get(), SE, &SafetyInfo,
445                            *Flags.get(), ORE, LoopNestMode);
446 
447   // Now that all loop invariants have been removed from the loop, promote any
448   // memory references to scalars that we can.
449   // Don't sink stores from loops without dedicated block exits. Exits
450   // containing indirect branches are not transformed by loop simplify,
451   // make sure we catch that. An additional load may be generated in the
452   // preheader for SSA updater, so also avoid sinking when no preheader
453   // is available.
454   if (!DisablePromotion && Preheader && L->hasDedicatedExits() &&
455       !Flags->tooManyMemoryAccesses() && !HasCoroSuspendInst) {
456     // Figure out the loop exits and their insertion points
457     SmallVector<BasicBlock *, 8> ExitBlocks;
458     L->getUniqueExitBlocks(ExitBlocks);
459 
460     // We can't insert into a catchswitch.
461     bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) {
462       return isa<CatchSwitchInst>(Exit->getTerminator());
463     });
464 
465     if (!HasCatchSwitch) {
466       SmallVector<Instruction *, 8> InsertPts;
467       SmallVector<MemoryAccess *, 8> MSSAInsertPts;
468       InsertPts.reserve(ExitBlocks.size());
469       if (MSSAU)
470         MSSAInsertPts.reserve(ExitBlocks.size());
471       for (BasicBlock *ExitBlock : ExitBlocks) {
472         InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
473         if (MSSAU)
474           MSSAInsertPts.push_back(nullptr);
475       }
476 
477       PredIteratorCache PIC;
478 
479       bool Promoted = false;
480       if (CurAST.get()) {
481         // Loop over all of the alias sets in the tracker object.
482         for (AliasSet &AS : *CurAST) {
483           // We can promote this alias set if it has a store, if it is a "Must"
484           // alias set, if the pointer is loop invariant, and if we are not
485           // eliminating any volatile loads or stores.
486           if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
487               !L->isLoopInvariant(AS.begin()->getValue()))
488             continue;
489 
490           assert(
491               !AS.empty() &&
492               "Must alias set should have at least one pointer element in it!");
493 
494           SmallSetVector<Value *, 8> PointerMustAliases;
495           for (const auto &ASI : AS)
496             PointerMustAliases.insert(ASI.getValue());
497 
498           Promoted |= promoteLoopAccessesToScalars(
499               PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI,
500               DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE);
501         }
502       } else {
503         // Promoting one set of accesses may make the pointers for another set
504         // loop invariant, so run this in a loop (with the MaybePromotable set
505         // decreasing in size over time).
506         bool LocalPromoted;
507         do {
508           LocalPromoted = false;
509           for (const SmallSetVector<Value *, 8> &PointerMustAliases :
510                collectPromotionCandidates(MSSA, AA, L)) {
511             LocalPromoted |= promoteLoopAccessesToScalars(
512                 PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC,
513                 LI, DT, TLI, L, /*AST*/nullptr, MSSAU.get(), &SafetyInfo, ORE);
514           }
515           Promoted |= LocalPromoted;
516         } while (LocalPromoted);
517       }
518 
519       // Once we have promoted values across the loop body we have to
520       // recursively reform LCSSA as any nested loop may now have values defined
521       // within the loop used in the outer loop.
522       // FIXME: This is really heavy handed. It would be a bit better to use an
523       // SSAUpdater strategy during promotion that was LCSSA aware and reformed
524       // it as it went.
525       if (Promoted)
526         formLCSSARecursively(*L, *DT, LI, SE);
527 
528       Changed |= Promoted;
529     }
530   }
531 
532   // Check that neither this loop nor its parent have had LCSSA broken. LICM is
533   // specifically moving instructions across the loop boundary and so it is
534   // especially in need of sanity checking here.
535   assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
536   assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) &&
537          "Parent loop not left in LCSSA form after LICM!");
538 
539   if (MSSAU.get() && VerifyMemorySSA)
540     MSSAU->getMemorySSA()->verifyMemorySSA();
541 
542   if (Changed && SE)
543     SE->forgetLoopDispositions(L);
544   return Changed;
545 }
546 
547 /// Walk the specified region of the CFG (defined by all blocks dominated by
548 /// the specified block, and that are in the current loop) in reverse depth
549 /// first order w.r.t the DominatorTree.  This allows us to visit uses before
550 /// definitions, allowing us to sink a loop body in one pass without iteration.
551 ///
552 bool llvm::sinkRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI,
553                       DominatorTree *DT, BlockFrequencyInfo *BFI,
554                       TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
555                       Loop *CurLoop, AliasSetTracker *CurAST,
556                       MemorySSAUpdater *MSSAU, ICFLoopSafetyInfo *SafetyInfo,
557                       SinkAndHoistLICMFlags &Flags,
558                       OptimizationRemarkEmitter *ORE) {
559 
560   // Verify inputs.
561   assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
562          CurLoop != nullptr && SafetyInfo != nullptr &&
563          "Unexpected input to sinkRegion.");
564   assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
565          "Either AliasSetTracker or MemorySSA should be initialized.");
566 
567   // We want to visit children before parents. We will enque all the parents
568   // before their children in the worklist and process the worklist in reverse
569   // order.
570   SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop);
571 
572   bool Changed = false;
573   for (DomTreeNode *DTN : reverse(Worklist)) {
574     BasicBlock *BB = DTN->getBlock();
575     // Only need to process the contents of this block if it is not part of a
576     // subloop (which would already have been processed).
577     if (inSubLoop(BB, CurLoop, LI))
578       continue;
579 
580     for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
581       Instruction &I = *--II;
582 
583       // The instruction is not used in the loop if it is dead.  In this case,
584       // we just delete it instead of sinking it.
585       if (isInstructionTriviallyDead(&I, TLI)) {
586         LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
587         salvageKnowledge(&I);
588         salvageDebugInfo(I);
589         ++II;
590         eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
591         Changed = true;
592         continue;
593       }
594 
595       // Check to see if we can sink this instruction to the exit blocks
596       // of the loop.  We can do this if the all users of the instruction are
597       // outside of the loop.  In this case, it doesn't even matter if the
598       // operands of the instruction are loop invariant.
599       //
600       bool FreeInLoop = false;
601       if (!I.mayHaveSideEffects() &&
602           isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) &&
603           canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
604                              ORE)) {
605         if (sink(I, LI, DT, BFI, CurLoop, SafetyInfo, MSSAU, ORE)) {
606           if (!FreeInLoop) {
607             ++II;
608             salvageDebugInfo(I);
609             eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
610           }
611           Changed = true;
612         }
613       }
614     }
615   }
616   if (MSSAU && VerifyMemorySSA)
617     MSSAU->getMemorySSA()->verifyMemorySSA();
618   return Changed;
619 }
620 
621 namespace {
622 // This is a helper class for hoistRegion to make it able to hoist control flow
623 // in order to be able to hoist phis. The way this works is that we initially
624 // start hoisting to the loop preheader, and when we see a loop invariant branch
625 // we make note of this. When we then come to hoist an instruction that's
626 // conditional on such a branch we duplicate the branch and the relevant control
627 // flow, then hoist the instruction into the block corresponding to its original
628 // block in the duplicated control flow.
629 class ControlFlowHoister {
630 private:
631   // Information about the loop we are hoisting from
632   LoopInfo *LI;
633   DominatorTree *DT;
634   Loop *CurLoop;
635   MemorySSAUpdater *MSSAU;
636 
637   // A map of blocks in the loop to the block their instructions will be hoisted
638   // to.
639   DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap;
640 
641   // The branches that we can hoist, mapped to the block that marks a
642   // convergence point of their control flow.
643   DenseMap<BranchInst *, BasicBlock *> HoistableBranches;
644 
645 public:
646   ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop,
647                      MemorySSAUpdater *MSSAU)
648       : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {}
649 
650   void registerPossiblyHoistableBranch(BranchInst *BI) {
651     // We can only hoist conditional branches with loop invariant operands.
652     if (!ControlFlowHoisting || !BI->isConditional() ||
653         !CurLoop->hasLoopInvariantOperands(BI))
654       return;
655 
656     // The branch destinations need to be in the loop, and we don't gain
657     // anything by duplicating conditional branches with duplicate successors,
658     // as it's essentially the same as an unconditional branch.
659     BasicBlock *TrueDest = BI->getSuccessor(0);
660     BasicBlock *FalseDest = BI->getSuccessor(1);
661     if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) ||
662         TrueDest == FalseDest)
663       return;
664 
665     // We can hoist BI if one branch destination is the successor of the other,
666     // or both have common successor which we check by seeing if the
667     // intersection of their successors is non-empty.
668     // TODO: This could be expanded to allowing branches where both ends
669     // eventually converge to a single block.
670     SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc;
671     TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest));
672     FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest));
673     BasicBlock *CommonSucc = nullptr;
674     if (TrueDestSucc.count(FalseDest)) {
675       CommonSucc = FalseDest;
676     } else if (FalseDestSucc.count(TrueDest)) {
677       CommonSucc = TrueDest;
678     } else {
679       set_intersect(TrueDestSucc, FalseDestSucc);
680       // If there's one common successor use that.
681       if (TrueDestSucc.size() == 1)
682         CommonSucc = *TrueDestSucc.begin();
683       // If there's more than one pick whichever appears first in the block list
684       // (we can't use the value returned by TrueDestSucc.begin() as it's
685       // unpredicatable which element gets returned).
686       else if (!TrueDestSucc.empty()) {
687         Function *F = TrueDest->getParent();
688         auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); };
689         auto It = llvm::find_if(*F, IsSucc);
690         assert(It != F->end() && "Could not find successor in function");
691         CommonSucc = &*It;
692       }
693     }
694     // The common successor has to be dominated by the branch, as otherwise
695     // there will be some other path to the successor that will not be
696     // controlled by this branch so any phi we hoist would be controlled by the
697     // wrong condition. This also takes care of avoiding hoisting of loop back
698     // edges.
699     // TODO: In some cases this could be relaxed if the successor is dominated
700     // by another block that's been hoisted and we can guarantee that the
701     // control flow has been replicated exactly.
702     if (CommonSucc && DT->dominates(BI, CommonSucc))
703       HoistableBranches[BI] = CommonSucc;
704   }
705 
706   bool canHoistPHI(PHINode *PN) {
707     // The phi must have loop invariant operands.
708     if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN))
709       return false;
710     // We can hoist phis if the block they are in is the target of hoistable
711     // branches which cover all of the predecessors of the block.
712     SmallPtrSet<BasicBlock *, 8> PredecessorBlocks;
713     BasicBlock *BB = PN->getParent();
714     for (BasicBlock *PredBB : predecessors(BB))
715       PredecessorBlocks.insert(PredBB);
716     // If we have less predecessor blocks than predecessors then the phi will
717     // have more than one incoming value for the same block which we can't
718     // handle.
719     // TODO: This could be handled be erasing some of the duplicate incoming
720     // values.
721     if (PredecessorBlocks.size() != pred_size(BB))
722       return false;
723     for (auto &Pair : HoistableBranches) {
724       if (Pair.second == BB) {
725         // Which blocks are predecessors via this branch depends on if the
726         // branch is triangle-like or diamond-like.
727         if (Pair.first->getSuccessor(0) == BB) {
728           PredecessorBlocks.erase(Pair.first->getParent());
729           PredecessorBlocks.erase(Pair.first->getSuccessor(1));
730         } else if (Pair.first->getSuccessor(1) == BB) {
731           PredecessorBlocks.erase(Pair.first->getParent());
732           PredecessorBlocks.erase(Pair.first->getSuccessor(0));
733         } else {
734           PredecessorBlocks.erase(Pair.first->getSuccessor(0));
735           PredecessorBlocks.erase(Pair.first->getSuccessor(1));
736         }
737       }
738     }
739     // PredecessorBlocks will now be empty if for every predecessor of BB we
740     // found a hoistable branch source.
741     return PredecessorBlocks.empty();
742   }
743 
744   BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) {
745     if (!ControlFlowHoisting)
746       return CurLoop->getLoopPreheader();
747     // If BB has already been hoisted, return that
748     if (HoistDestinationMap.count(BB))
749       return HoistDestinationMap[BB];
750 
751     // Check if this block is conditional based on a pending branch
752     auto HasBBAsSuccessor =
753         [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) {
754           return BB != Pair.second && (Pair.first->getSuccessor(0) == BB ||
755                                        Pair.first->getSuccessor(1) == BB);
756         };
757     auto It = llvm::find_if(HoistableBranches, HasBBAsSuccessor);
758 
759     // If not involved in a pending branch, hoist to preheader
760     BasicBlock *InitialPreheader = CurLoop->getLoopPreheader();
761     if (It == HoistableBranches.end()) {
762       LLVM_DEBUG(dbgs() << "LICM using "
763                         << InitialPreheader->getNameOrAsOperand()
764                         << " as hoist destination for "
765                         << BB->getNameOrAsOperand() << "\n");
766       HoistDestinationMap[BB] = InitialPreheader;
767       return InitialPreheader;
768     }
769     BranchInst *BI = It->first;
770     assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==
771                HoistableBranches.end() &&
772            "BB is expected to be the target of at most one branch");
773 
774     LLVMContext &C = BB->getContext();
775     BasicBlock *TrueDest = BI->getSuccessor(0);
776     BasicBlock *FalseDest = BI->getSuccessor(1);
777     BasicBlock *CommonSucc = HoistableBranches[BI];
778     BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent());
779 
780     // Create hoisted versions of blocks that currently don't have them
781     auto CreateHoistedBlock = [&](BasicBlock *Orig) {
782       if (HoistDestinationMap.count(Orig))
783         return HoistDestinationMap[Orig];
784       BasicBlock *New =
785           BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent());
786       HoistDestinationMap[Orig] = New;
787       DT->addNewBlock(New, HoistTarget);
788       if (CurLoop->getParentLoop())
789         CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI);
790       ++NumCreatedBlocks;
791       LLVM_DEBUG(dbgs() << "LICM created " << New->getName()
792                         << " as hoist destination for " << Orig->getName()
793                         << "\n");
794       return New;
795     };
796     BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest);
797     BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest);
798     BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc);
799 
800     // Link up these blocks with branches.
801     if (!HoistCommonSucc->getTerminator()) {
802       // The new common successor we've generated will branch to whatever that
803       // hoist target branched to.
804       BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor();
805       assert(TargetSucc && "Expected hoist target to have a single successor");
806       HoistCommonSucc->moveBefore(TargetSucc);
807       BranchInst::Create(TargetSucc, HoistCommonSucc);
808     }
809     if (!HoistTrueDest->getTerminator()) {
810       HoistTrueDest->moveBefore(HoistCommonSucc);
811       BranchInst::Create(HoistCommonSucc, HoistTrueDest);
812     }
813     if (!HoistFalseDest->getTerminator()) {
814       HoistFalseDest->moveBefore(HoistCommonSucc);
815       BranchInst::Create(HoistCommonSucc, HoistFalseDest);
816     }
817 
818     // If BI is being cloned to what was originally the preheader then
819     // HoistCommonSucc will now be the new preheader.
820     if (HoistTarget == InitialPreheader) {
821       // Phis in the loop header now need to use the new preheader.
822       InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc);
823       if (MSSAU)
824         MSSAU->wireOldPredecessorsToNewImmediatePredecessor(
825             HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget});
826       // The new preheader dominates the loop header.
827       DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc);
828       DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader());
829       DT->changeImmediateDominator(HeaderNode, PreheaderNode);
830       // The preheader hoist destination is now the new preheader, with the
831       // exception of the hoist destination of this branch.
832       for (auto &Pair : HoistDestinationMap)
833         if (Pair.second == InitialPreheader && Pair.first != BI->getParent())
834           Pair.second = HoistCommonSucc;
835     }
836 
837     // Now finally clone BI.
838     ReplaceInstWithInst(
839         HoistTarget->getTerminator(),
840         BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition()));
841     ++NumClonedBranches;
842 
843     assert(CurLoop->getLoopPreheader() &&
844            "Hoisting blocks should not have destroyed preheader");
845     return HoistDestinationMap[BB];
846   }
847 };
848 } // namespace
849 
850 // Hoisting/sinking instruction out of a loop isn't always beneficial. It's only
851 // only worthwhile if the destination block is actually colder than current
852 // block.
853 static bool worthSinkOrHoistInst(Instruction &I, BasicBlock *DstBlock,
854                                  OptimizationRemarkEmitter *ORE,
855                                  BlockFrequencyInfo *BFI) {
856   // Check block frequency only when runtime profile is available
857   // to avoid pathological cases. With static profile, lean towards
858   // hosting because it helps canonicalize the loop for vectorizer.
859   if (!DstBlock->getParent()->hasProfileData())
860     return true;
861 
862   if (!HoistSinkColdnessThreshold || !BFI)
863     return true;
864 
865   BasicBlock *SrcBlock = I.getParent();
866   if (BFI->getBlockFreq(DstBlock).getFrequency() / HoistSinkColdnessThreshold >
867       BFI->getBlockFreq(SrcBlock).getFrequency()) {
868     ORE->emit([&]() {
869       return OptimizationRemarkMissed(DEBUG_TYPE, "SinkHoistInst", &I)
870              << "failed to sink or hoist instruction because containing block "
871                 "has lower frequency than destination block";
872     });
873     return false;
874   }
875 
876   return true;
877 }
878 
879 /// Walk the specified region of the CFG (defined by all blocks dominated by
880 /// the specified block, and that are in the current loop) in depth first
881 /// order w.r.t the DominatorTree.  This allows us to visit definitions before
882 /// uses, allowing us to hoist a loop body in one pass without iteration.
883 ///
884 bool llvm::hoistRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI,
885                        DominatorTree *DT, BlockFrequencyInfo *BFI,
886                        TargetLibraryInfo *TLI, Loop *CurLoop,
887                        AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
888                        ScalarEvolution *SE, ICFLoopSafetyInfo *SafetyInfo,
889                        SinkAndHoistLICMFlags &Flags,
890                        OptimizationRemarkEmitter *ORE, bool LoopNestMode) {
891   // Verify inputs.
892   assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
893          CurLoop != nullptr && SafetyInfo != nullptr &&
894          "Unexpected input to hoistRegion.");
895   assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
896          "Either AliasSetTracker or MemorySSA should be initialized.");
897 
898   ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU);
899 
900   // Keep track of instructions that have been hoisted, as they may need to be
901   // re-hoisted if they end up not dominating all of their uses.
902   SmallVector<Instruction *, 16> HoistedInstructions;
903 
904   // For PHI hoisting to work we need to hoist blocks before their successors.
905   // We can do this by iterating through the blocks in the loop in reverse
906   // post-order.
907   LoopBlocksRPO Worklist(CurLoop);
908   Worklist.perform(LI);
909   bool Changed = false;
910   for (BasicBlock *BB : Worklist) {
911     // Only need to process the contents of this block if it is not part of a
912     // subloop (which would already have been processed).
913     if (!LoopNestMode && inSubLoop(BB, CurLoop, LI))
914       continue;
915 
916     for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
917       Instruction &I = *II++;
918       // Try constant folding this instruction.  If all the operands are
919       // constants, it is technically hoistable, but it would be better to
920       // just fold it.
921       if (Constant *C = ConstantFoldInstruction(
922               &I, I.getModule()->getDataLayout(), TLI)) {
923         LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << "  --> " << *C
924                           << '\n');
925         if (CurAST)
926           CurAST->copyValue(&I, C);
927         // FIXME MSSA: Such replacements may make accesses unoptimized (D51960).
928         I.replaceAllUsesWith(C);
929         if (isInstructionTriviallyDead(&I, TLI))
930           eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
931         Changed = true;
932         continue;
933       }
934 
935       // Try hoisting the instruction out to the preheader.  We can only do
936       // this if all of the operands of the instruction are loop invariant and
937       // if it is safe to hoist the instruction. We also check block frequency
938       // to make sure instruction only gets hoisted into colder blocks.
939       // TODO: It may be safe to hoist if we are hoisting to a conditional block
940       // and we have accurately duplicated the control flow from the loop header
941       // to that block.
942       if (CurLoop->hasLoopInvariantOperands(&I) &&
943           canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
944                              ORE) &&
945           worthSinkOrHoistInst(I, CurLoop->getLoopPreheader(), ORE, BFI) &&
946           isSafeToExecuteUnconditionally(
947               I, DT, TLI, CurLoop, SafetyInfo, ORE,
948               CurLoop->getLoopPreheader()->getTerminator())) {
949         hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
950               MSSAU, SE, ORE);
951         HoistedInstructions.push_back(&I);
952         Changed = true;
953         continue;
954       }
955 
956       // Attempt to remove floating point division out of the loop by
957       // converting it to a reciprocal multiplication.
958       if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() &&
959           CurLoop->isLoopInvariant(I.getOperand(1))) {
960         auto Divisor = I.getOperand(1);
961         auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0);
962         auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor);
963         ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags());
964         SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent());
965         ReciprocalDivisor->insertBefore(&I);
966 
967         auto Product =
968             BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor);
969         Product->setFastMathFlags(I.getFastMathFlags());
970         SafetyInfo->insertInstructionTo(Product, I.getParent());
971         Product->insertAfter(&I);
972         I.replaceAllUsesWith(Product);
973         eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
974 
975         hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB),
976               SafetyInfo, MSSAU, SE, ORE);
977         HoistedInstructions.push_back(ReciprocalDivisor);
978         Changed = true;
979         continue;
980       }
981 
982       auto IsInvariantStart = [&](Instruction &I) {
983         using namespace PatternMatch;
984         return I.use_empty() &&
985                match(&I, m_Intrinsic<Intrinsic::invariant_start>());
986       };
987       auto MustExecuteWithoutWritesBefore = [&](Instruction &I) {
988         return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) &&
989                SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop);
990       };
991       if ((IsInvariantStart(I) || isGuard(&I)) &&
992           CurLoop->hasLoopInvariantOperands(&I) &&
993           MustExecuteWithoutWritesBefore(I)) {
994         hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
995               MSSAU, SE, ORE);
996         HoistedInstructions.push_back(&I);
997         Changed = true;
998         continue;
999       }
1000 
1001       if (PHINode *PN = dyn_cast<PHINode>(&I)) {
1002         if (CFH.canHoistPHI(PN)) {
1003           // Redirect incoming blocks first to ensure that we create hoisted
1004           // versions of those blocks before we hoist the phi.
1005           for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i)
1006             PN->setIncomingBlock(
1007                 i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i)));
1008           hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
1009                 MSSAU, SE, ORE);
1010           assert(DT->dominates(PN, BB) && "Conditional PHIs not expected");
1011           Changed = true;
1012           continue;
1013         }
1014       }
1015 
1016       // Remember possibly hoistable branches so we can actually hoist them
1017       // later if needed.
1018       if (BranchInst *BI = dyn_cast<BranchInst>(&I))
1019         CFH.registerPossiblyHoistableBranch(BI);
1020     }
1021   }
1022 
1023   // If we hoisted instructions to a conditional block they may not dominate
1024   // their uses that weren't hoisted (such as phis where some operands are not
1025   // loop invariant). If so make them unconditional by moving them to their
1026   // immediate dominator. We iterate through the instructions in reverse order
1027   // which ensures that when we rehoist an instruction we rehoist its operands,
1028   // and also keep track of where in the block we are rehoisting to to make sure
1029   // that we rehoist instructions before the instructions that use them.
1030   Instruction *HoistPoint = nullptr;
1031   if (ControlFlowHoisting) {
1032     for (Instruction *I : reverse(HoistedInstructions)) {
1033       if (!llvm::all_of(I->uses(),
1034                         [&](Use &U) { return DT->dominates(I, U); })) {
1035         BasicBlock *Dominator =
1036             DT->getNode(I->getParent())->getIDom()->getBlock();
1037         if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) {
1038           if (HoistPoint)
1039             assert(DT->dominates(Dominator, HoistPoint->getParent()) &&
1040                    "New hoist point expected to dominate old hoist point");
1041           HoistPoint = Dominator->getTerminator();
1042         }
1043         LLVM_DEBUG(dbgs() << "LICM rehoisting to "
1044                           << HoistPoint->getParent()->getNameOrAsOperand()
1045                           << ": " << *I << "\n");
1046         moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU, SE);
1047         HoistPoint = I;
1048         Changed = true;
1049       }
1050     }
1051   }
1052   if (MSSAU && VerifyMemorySSA)
1053     MSSAU->getMemorySSA()->verifyMemorySSA();
1054 
1055     // Now that we've finished hoisting make sure that LI and DT are still
1056     // valid.
1057 #ifdef EXPENSIVE_CHECKS
1058   if (Changed) {
1059     assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&
1060            "Dominator tree verification failed");
1061     LI->verify(*DT);
1062   }
1063 #endif
1064 
1065   return Changed;
1066 }
1067 
1068 // Return true if LI is invariant within scope of the loop. LI is invariant if
1069 // CurLoop is dominated by an invariant.start representing the same memory
1070 // location and size as the memory location LI loads from, and also the
1071 // invariant.start has no uses.
1072 static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT,
1073                                   Loop *CurLoop) {
1074   Value *Addr = LI->getOperand(0);
1075   const DataLayout &DL = LI->getModule()->getDataLayout();
1076   const TypeSize LocSizeInBits = DL.getTypeSizeInBits(LI->getType());
1077 
1078   // It is not currently possible for clang to generate an invariant.start
1079   // intrinsic with scalable vector types because we don't support thread local
1080   // sizeless types and we don't permit sizeless types in structs or classes.
1081   // Furthermore, even if support is added for this in future the intrinsic
1082   // itself is defined to have a size of -1 for variable sized objects. This
1083   // makes it impossible to verify if the intrinsic envelops our region of
1084   // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8>
1085   // types would have a -1 parameter, but the former is clearly double the size
1086   // of the latter.
1087   if (LocSizeInBits.isScalable())
1088     return false;
1089 
1090   // if the type is i8 addrspace(x)*, we know this is the type of
1091   // llvm.invariant.start operand
1092   auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()),
1093                                      LI->getPointerAddressSpace());
1094   unsigned BitcastsVisited = 0;
1095   // Look through bitcasts until we reach the i8* type (this is invariant.start
1096   // operand type).
1097   while (Addr->getType() != PtrInt8Ty) {
1098     auto *BC = dyn_cast<BitCastInst>(Addr);
1099     // Avoid traversing high number of bitcast uses.
1100     if (++BitcastsVisited > MaxNumUsesTraversed || !BC)
1101       return false;
1102     Addr = BC->getOperand(0);
1103   }
1104 
1105   unsigned UsesVisited = 0;
1106   // Traverse all uses of the load operand value, to see if invariant.start is
1107   // one of the uses, and whether it dominates the load instruction.
1108   for (auto *U : Addr->users()) {
1109     // Avoid traversing for Load operand with high number of users.
1110     if (++UsesVisited > MaxNumUsesTraversed)
1111       return false;
1112     IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1113     // If there are escaping uses of invariant.start instruction, the load maybe
1114     // non-invariant.
1115     if (!II || II->getIntrinsicID() != Intrinsic::invariant_start ||
1116         !II->use_empty())
1117       continue;
1118     ConstantInt *InvariantSize = cast<ConstantInt>(II->getArgOperand(0));
1119     // The intrinsic supports having a -1 argument for variable sized objects
1120     // so we should check for that here.
1121     if (InvariantSize->isNegative())
1122       continue;
1123     uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8;
1124     // Confirm the invariant.start location size contains the load operand size
1125     // in bits. Also, the invariant.start should dominate the load, and we
1126     // should not hoist the load out of a loop that contains this dominating
1127     // invariant.start.
1128     if (LocSizeInBits.getFixedSize() <= InvariantSizeInBits &&
1129         DT->properlyDominates(II->getParent(), CurLoop->getHeader()))
1130       return true;
1131   }
1132 
1133   return false;
1134 }
1135 
1136 namespace {
1137 /// Return true if-and-only-if we know how to (mechanically) both hoist and
1138 /// sink a given instruction out of a loop.  Does not address legality
1139 /// concerns such as aliasing or speculation safety.
1140 bool isHoistableAndSinkableInst(Instruction &I) {
1141   // Only these instructions are hoistable/sinkable.
1142   return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) ||
1143           isa<FenceInst>(I) || isa<CastInst>(I) || isa<UnaryOperator>(I) ||
1144           isa<BinaryOperator>(I) || isa<SelectInst>(I) ||
1145           isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
1146           isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
1147           isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) ||
1148           isa<InsertValueInst>(I) || isa<FreezeInst>(I));
1149 }
1150 /// Return true if all of the alias sets within this AST are known not to
1151 /// contain a Mod, or if MSSA knows there are no MemoryDefs in the loop.
1152 bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU,
1153                 const Loop *L) {
1154   if (CurAST) {
1155     for (AliasSet &AS : *CurAST) {
1156       if (!AS.isForwardingAliasSet() && AS.isMod()) {
1157         return false;
1158       }
1159     }
1160     return true;
1161   } else { /*MSSAU*/
1162     for (auto *BB : L->getBlocks())
1163       if (MSSAU->getMemorySSA()->getBlockDefs(BB))
1164         return false;
1165     return true;
1166   }
1167 }
1168 
1169 /// Return true if I is the only Instruction with a MemoryAccess in L.
1170 bool isOnlyMemoryAccess(const Instruction *I, const Loop *L,
1171                         const MemorySSAUpdater *MSSAU) {
1172   for (auto *BB : L->getBlocks())
1173     if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) {
1174       int NotAPhi = 0;
1175       for (const auto &Acc : *Accs) {
1176         if (isa<MemoryPhi>(&Acc))
1177           continue;
1178         const auto *MUD = cast<MemoryUseOrDef>(&Acc);
1179         if (MUD->getMemoryInst() != I || NotAPhi++ == 1)
1180           return false;
1181       }
1182     }
1183   return true;
1184 }
1185 }
1186 
1187 bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT,
1188                               Loop *CurLoop, AliasSetTracker *CurAST,
1189                               MemorySSAUpdater *MSSAU,
1190                               bool TargetExecutesOncePerLoop,
1191                               SinkAndHoistLICMFlags *Flags,
1192                               OptimizationRemarkEmitter *ORE) {
1193   assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
1194          "Either AliasSetTracker or MemorySSA should be initialized.");
1195 
1196   // If we don't understand the instruction, bail early.
1197   if (!isHoistableAndSinkableInst(I))
1198     return false;
1199 
1200   MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr;
1201   if (MSSA)
1202     assert(Flags != nullptr && "Flags cannot be null.");
1203 
1204   // Loads have extra constraints we have to verify before we can hoist them.
1205   if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
1206     if (!LI->isUnordered())
1207       return false; // Don't sink/hoist volatile or ordered atomic loads!
1208 
1209     // Loads from constant memory are always safe to move, even if they end up
1210     // in the same alias set as something that ends up being modified.
1211     if (AA->pointsToConstantMemory(LI->getOperand(0)))
1212       return true;
1213     if (LI->hasMetadata(LLVMContext::MD_invariant_load))
1214       return true;
1215 
1216     if (LI->isAtomic() && !TargetExecutesOncePerLoop)
1217       return false; // Don't risk duplicating unordered loads
1218 
1219     // This checks for an invariant.start dominating the load.
1220     if (isLoadInvariantInLoop(LI, DT, CurLoop))
1221       return true;
1222 
1223     bool Invalidated;
1224     if (CurAST)
1225       Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST,
1226                                              CurLoop, AA);
1227     else
1228       Invalidated = pointerInvalidatedByLoopWithMSSA(
1229           MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, I, *Flags);
1230     // Check loop-invariant address because this may also be a sinkable load
1231     // whose address is not necessarily loop-invariant.
1232     if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1233       ORE->emit([&]() {
1234         return OptimizationRemarkMissed(
1235                    DEBUG_TYPE, "LoadWithLoopInvariantAddressInvalidated", LI)
1236                << "failed to move load with loop-invariant address "
1237                   "because the loop may invalidate its value";
1238       });
1239 
1240     return !Invalidated;
1241   } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1242     // Don't sink or hoist dbg info; it's legal, but not useful.
1243     if (isa<DbgInfoIntrinsic>(I))
1244       return false;
1245 
1246     // Don't sink calls which can throw.
1247     if (CI->mayThrow())
1248       return false;
1249 
1250     // Convergent attribute has been used on operations that involve
1251     // inter-thread communication which results are implicitly affected by the
1252     // enclosing control flows. It is not safe to hoist or sink such operations
1253     // across control flow.
1254     if (CI->isConvergent())
1255       return false;
1256 
1257     using namespace PatternMatch;
1258     if (match(CI, m_Intrinsic<Intrinsic::assume>()))
1259       // Assumes don't actually alias anything or throw
1260       return true;
1261 
1262     if (match(CI, m_Intrinsic<Intrinsic::experimental_widenable_condition>()))
1263       // Widenable conditions don't actually alias anything or throw
1264       return true;
1265 
1266     // Handle simple cases by querying alias analysis.
1267     FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
1268     if (Behavior == FMRB_DoesNotAccessMemory)
1269       return true;
1270     if (AAResults::onlyReadsMemory(Behavior)) {
1271       // A readonly argmemonly function only reads from memory pointed to by
1272       // it's arguments with arbitrary offsets.  If we can prove there are no
1273       // writes to this memory in the loop, we can hoist or sink.
1274       if (AAResults::onlyAccessesArgPointees(Behavior)) {
1275         // TODO: expand to writeable arguments
1276         for (Value *Op : CI->arg_operands())
1277           if (Op->getType()->isPointerTy()) {
1278             bool Invalidated;
1279             if (CurAST)
1280               Invalidated = pointerInvalidatedByLoop(
1281                   MemoryLocation::getBeforeOrAfter(Op), CurAST, CurLoop, AA);
1282             else
1283               Invalidated = pointerInvalidatedByLoopWithMSSA(
1284                   MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop, I,
1285                   *Flags);
1286             if (Invalidated)
1287               return false;
1288           }
1289         return true;
1290       }
1291 
1292       // If this call only reads from memory and there are no writes to memory
1293       // in the loop, we can hoist or sink the call as appropriate.
1294       if (isReadOnly(CurAST, MSSAU, CurLoop))
1295         return true;
1296     }
1297 
1298     // FIXME: This should use mod/ref information to see if we can hoist or
1299     // sink the call.
1300 
1301     return false;
1302   } else if (auto *FI = dyn_cast<FenceInst>(&I)) {
1303     // Fences alias (most) everything to provide ordering.  For the moment,
1304     // just give up if there are any other memory operations in the loop.
1305     if (CurAST) {
1306       auto Begin = CurAST->begin();
1307       assert(Begin != CurAST->end() && "must contain FI");
1308       if (std::next(Begin) != CurAST->end())
1309         // constant memory for instance, TODO: handle better
1310         return false;
1311       auto *UniqueI = Begin->getUniqueInstruction();
1312       if (!UniqueI)
1313         // other memory op, give up
1314         return false;
1315       (void)FI; // suppress unused variable warning
1316       assert(UniqueI == FI && "AS must contain FI");
1317       return true;
1318     } else // MSSAU
1319       return isOnlyMemoryAccess(FI, CurLoop, MSSAU);
1320   } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
1321     if (!SI->isUnordered())
1322       return false; // Don't sink/hoist volatile or ordered atomic store!
1323 
1324     // We can only hoist a store that we can prove writes a value which is not
1325     // read or overwritten within the loop.  For those cases, we fallback to
1326     // load store promotion instead.  TODO: We can extend this to cases where
1327     // there is exactly one write to the location and that write dominates an
1328     // arbitrary number of reads in the loop.
1329     if (CurAST) {
1330       auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI));
1331 
1332       if (AS.isRef() || !AS.isMustAlias())
1333         // Quick exit test, handled by the full path below as well.
1334         return false;
1335       auto *UniqueI = AS.getUniqueInstruction();
1336       if (!UniqueI)
1337         // other memory op, give up
1338         return false;
1339       assert(UniqueI == SI && "AS must contain SI");
1340       return true;
1341     } else { // MSSAU
1342       if (isOnlyMemoryAccess(SI, CurLoop, MSSAU))
1343         return true;
1344       // If there are more accesses than the Promotion cap or no "quota" to
1345       // check clobber, then give up as we're not walking a list that long.
1346       if (Flags->tooManyMemoryAccesses() || Flags->tooManyClobberingCalls())
1347         return false;
1348       // If there are interfering Uses (i.e. their defining access is in the
1349       // loop), or ordered loads (stored as Defs!), don't move this store.
1350       // Could do better here, but this is conservatively correct.
1351       // TODO: Cache set of Uses on the first walk in runOnLoop, update when
1352       // moving accesses. Can also extend to dominating uses.
1353       auto *SIMD = MSSA->getMemoryAccess(SI);
1354       for (auto *BB : CurLoop->getBlocks())
1355         if (auto *Accesses = MSSA->getBlockAccesses(BB)) {
1356           for (const auto &MA : *Accesses)
1357             if (const auto *MU = dyn_cast<MemoryUse>(&MA)) {
1358               auto *MD = MU->getDefiningAccess();
1359               if (!MSSA->isLiveOnEntryDef(MD) &&
1360                   CurLoop->contains(MD->getBlock()))
1361                 return false;
1362               // Disable hoisting past potentially interfering loads. Optimized
1363               // Uses may point to an access outside the loop, as getClobbering
1364               // checks the previous iteration when walking the backedge.
1365               // FIXME: More precise: no Uses that alias SI.
1366               if (!Flags->getIsSink() && !MSSA->dominates(SIMD, MU))
1367                 return false;
1368             } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) {
1369               if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) {
1370                 (void)LI; // Silence warning.
1371                 assert(!LI->isUnordered() && "Expected unordered load");
1372                 return false;
1373               }
1374               // Any call, while it may not be clobbering SI, it may be a use.
1375               if (auto *CI = dyn_cast<CallInst>(MD->getMemoryInst())) {
1376                 // Check if the call may read from the memory location written
1377                 // to by SI. Check CI's attributes and arguments; the number of
1378                 // such checks performed is limited above by NoOfMemAccTooLarge.
1379                 ModRefInfo MRI = AA->getModRefInfo(CI, MemoryLocation::get(SI));
1380                 if (isModOrRefSet(MRI))
1381                   return false;
1382               }
1383             }
1384         }
1385       auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI);
1386       Flags->incrementClobberingCalls();
1387       // If there are no clobbering Defs in the loop, store is safe to hoist.
1388       return MSSA->isLiveOnEntryDef(Source) ||
1389              !CurLoop->contains(Source->getBlock());
1390     }
1391   }
1392 
1393   assert(!I.mayReadOrWriteMemory() && "unhandled aliasing");
1394 
1395   // We've established mechanical ability and aliasing, it's up to the caller
1396   // to check fault safety
1397   return true;
1398 }
1399 
1400 /// Returns true if a PHINode is a trivially replaceable with an
1401 /// Instruction.
1402 /// This is true when all incoming values are that instruction.
1403 /// This pattern occurs most often with LCSSA PHI nodes.
1404 ///
1405 static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) {
1406   for (const Value *IncValue : PN.incoming_values())
1407     if (IncValue != &I)
1408       return false;
1409 
1410   return true;
1411 }
1412 
1413 /// Return true if the instruction is free in the loop.
1414 static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop,
1415                          const TargetTransformInfo *TTI) {
1416 
1417   if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1418     if (TTI->getUserCost(GEP, TargetTransformInfo::TCK_SizeAndLatency) !=
1419         TargetTransformInfo::TCC_Free)
1420       return false;
1421     // For a GEP, we cannot simply use getUserCost because currently it
1422     // optimistically assume that a GEP will fold into addressing mode
1423     // regardless of its users.
1424     const BasicBlock *BB = GEP->getParent();
1425     for (const User *U : GEP->users()) {
1426       const Instruction *UI = cast<Instruction>(U);
1427       if (CurLoop->contains(UI) &&
1428           (BB != UI->getParent() ||
1429            (!isa<StoreInst>(UI) && !isa<LoadInst>(UI))))
1430         return false;
1431     }
1432     return true;
1433   } else
1434     return TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency) ==
1435            TargetTransformInfo::TCC_Free;
1436 }
1437 
1438 /// Return true if the only users of this instruction are outside of
1439 /// the loop. If this is true, we can sink the instruction to the exit
1440 /// blocks of the loop.
1441 ///
1442 /// We also return true if the instruction could be folded away in lowering.
1443 /// (e.g.,  a GEP can be folded into a load as an addressing mode in the loop).
1444 static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
1445                                   const LoopSafetyInfo *SafetyInfo,
1446                                   TargetTransformInfo *TTI, bool &FreeInLoop) {
1447   const auto &BlockColors = SafetyInfo->getBlockColors();
1448   bool IsFree = isFreeInLoop(I, CurLoop, TTI);
1449   for (const User *U : I.users()) {
1450     const Instruction *UI = cast<Instruction>(U);
1451     if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1452       const BasicBlock *BB = PN->getParent();
1453       // We cannot sink uses in catchswitches.
1454       if (isa<CatchSwitchInst>(BB->getTerminator()))
1455         return false;
1456 
1457       // We need to sink a callsite to a unique funclet.  Avoid sinking if the
1458       // phi use is too muddled.
1459       if (isa<CallInst>(I))
1460         if (!BlockColors.empty() &&
1461             BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
1462           return false;
1463     }
1464 
1465     if (CurLoop->contains(UI)) {
1466       if (IsFree) {
1467         FreeInLoop = true;
1468         continue;
1469       }
1470       return false;
1471     }
1472   }
1473   return true;
1474 }
1475 
1476 static Instruction *cloneInstructionInExitBlock(
1477     Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
1478     const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) {
1479   Instruction *New;
1480   if (auto *CI = dyn_cast<CallInst>(&I)) {
1481     const auto &BlockColors = SafetyInfo->getBlockColors();
1482 
1483     // Sinking call-sites need to be handled differently from other
1484     // instructions.  The cloned call-site needs a funclet bundle operand
1485     // appropriate for its location in the CFG.
1486     SmallVector<OperandBundleDef, 1> OpBundles;
1487     for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
1488          BundleIdx != BundleEnd; ++BundleIdx) {
1489       OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
1490       if (Bundle.getTagID() == LLVMContext::OB_funclet)
1491         continue;
1492 
1493       OpBundles.emplace_back(Bundle);
1494     }
1495 
1496     if (!BlockColors.empty()) {
1497       const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
1498       assert(CV.size() == 1 && "non-unique color for exit block!");
1499       BasicBlock *BBColor = CV.front();
1500       Instruction *EHPad = BBColor->getFirstNonPHI();
1501       if (EHPad->isEHPad())
1502         OpBundles.emplace_back("funclet", EHPad);
1503     }
1504 
1505     New = CallInst::Create(CI, OpBundles);
1506   } else {
1507     New = I.clone();
1508   }
1509 
1510   ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
1511   if (!I.getName().empty())
1512     New->setName(I.getName() + ".le");
1513 
1514   if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
1515     // Create a new MemoryAccess and let MemorySSA set its defining access.
1516     MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1517         New, nullptr, New->getParent(), MemorySSA::Beginning);
1518     if (NewMemAcc) {
1519       if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
1520         MSSAU->insertDef(MemDef, /*RenameUses=*/true);
1521       else {
1522         auto *MemUse = cast<MemoryUse>(NewMemAcc);
1523         MSSAU->insertUse(MemUse, /*RenameUses=*/true);
1524       }
1525     }
1526   }
1527 
1528   // Build LCSSA PHI nodes for any in-loop operands (if legal).  Note that
1529   // this is particularly cheap because we can rip off the PHI node that we're
1530   // replacing for the number and blocks of the predecessors.
1531   // OPT: If this shows up in a profile, we can instead finish sinking all
1532   // invariant instructions, and then walk their operands to re-establish
1533   // LCSSA. That will eliminate creating PHI nodes just to nuke them when
1534   // sinking bottom-up.
1535   for (Use &Op : New->operands())
1536     if (LI->wouldBeOutOfLoopUseRequiringLCSSA(Op.get(), PN.getParent())) {
1537       auto *OInst = cast<Instruction>(Op.get());
1538       PHINode *OpPN =
1539         PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
1540                         OInst->getName() + ".lcssa", &ExitBlock.front());
1541       for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1542         OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
1543       Op = OpPN;
1544     }
1545   return New;
1546 }
1547 
1548 static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
1549                              AliasSetTracker *AST, MemorySSAUpdater *MSSAU) {
1550   if (AST)
1551     AST->deleteValue(&I);
1552   if (MSSAU)
1553     MSSAU->removeMemoryAccess(&I);
1554   SafetyInfo.removeInstruction(&I);
1555   I.eraseFromParent();
1556 }
1557 
1558 static void moveInstructionBefore(Instruction &I, Instruction &Dest,
1559                                   ICFLoopSafetyInfo &SafetyInfo,
1560                                   MemorySSAUpdater *MSSAU,
1561                                   ScalarEvolution *SE) {
1562   SafetyInfo.removeInstruction(&I);
1563   SafetyInfo.insertInstructionTo(&I, Dest.getParent());
1564   I.moveBefore(&Dest);
1565   if (MSSAU)
1566     if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
1567             MSSAU->getMemorySSA()->getMemoryAccess(&I)))
1568       MSSAU->moveToPlace(OldMemAcc, Dest.getParent(),
1569                          MemorySSA::BeforeTerminator);
1570   if (SE)
1571     SE->forgetValue(&I);
1572 }
1573 
1574 static Instruction *sinkThroughTriviallyReplaceablePHI(
1575     PHINode *TPN, Instruction *I, LoopInfo *LI,
1576     SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies,
1577     const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop,
1578     MemorySSAUpdater *MSSAU) {
1579   assert(isTriviallyReplaceablePHI(*TPN, *I) &&
1580          "Expect only trivially replaceable PHI");
1581   BasicBlock *ExitBlock = TPN->getParent();
1582   Instruction *New;
1583   auto It = SunkCopies.find(ExitBlock);
1584   if (It != SunkCopies.end())
1585     New = It->second;
1586   else
1587     New = SunkCopies[ExitBlock] = cloneInstructionInExitBlock(
1588         *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU);
1589   return New;
1590 }
1591 
1592 static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) {
1593   BasicBlock *BB = PN->getParent();
1594   if (!BB->canSplitPredecessors())
1595     return false;
1596   // It's not impossible to split EHPad blocks, but if BlockColors already exist
1597   // it require updating BlockColors for all offspring blocks accordingly. By
1598   // skipping such corner case, we can make updating BlockColors after splitting
1599   // predecessor fairly simple.
1600   if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad())
1601     return false;
1602   for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1603     BasicBlock *BBPred = *PI;
1604     if (isa<IndirectBrInst>(BBPred->getTerminator()) ||
1605         isa<CallBrInst>(BBPred->getTerminator()))
1606       return false;
1607   }
1608   return true;
1609 }
1610 
1611 static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT,
1612                                         LoopInfo *LI, const Loop *CurLoop,
1613                                         LoopSafetyInfo *SafetyInfo,
1614                                         MemorySSAUpdater *MSSAU) {
1615 #ifndef NDEBUG
1616   SmallVector<BasicBlock *, 32> ExitBlocks;
1617   CurLoop->getUniqueExitBlocks(ExitBlocks);
1618   SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1619                                              ExitBlocks.end());
1620 #endif
1621   BasicBlock *ExitBB = PN->getParent();
1622   assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.");
1623 
1624   // Split predecessors of the loop exit to make instructions in the loop are
1625   // exposed to exit blocks through trivially replaceable PHIs while keeping the
1626   // loop in the canonical form where each predecessor of each exit block should
1627   // be contained within the loop. For example, this will convert the loop below
1628   // from
1629   //
1630   // LB1:
1631   //   %v1 =
1632   //   br %LE, %LB2
1633   // LB2:
1634   //   %v2 =
1635   //   br %LE, %LB1
1636   // LE:
1637   //   %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable
1638   //
1639   // to
1640   //
1641   // LB1:
1642   //   %v1 =
1643   //   br %LE.split, %LB2
1644   // LB2:
1645   //   %v2 =
1646   //   br %LE.split2, %LB1
1647   // LE.split:
1648   //   %p1 = phi [%v1, %LB1]  <-- trivially replaceable
1649   //   br %LE
1650   // LE.split2:
1651   //   %p2 = phi [%v2, %LB2]  <-- trivially replaceable
1652   //   br %LE
1653   // LE:
1654   //   %p = phi [%p1, %LE.split], [%p2, %LE.split2]
1655   //
1656   const auto &BlockColors = SafetyInfo->getBlockColors();
1657   SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB));
1658   while (!PredBBs.empty()) {
1659     BasicBlock *PredBB = *PredBBs.begin();
1660     assert(CurLoop->contains(PredBB) &&
1661            "Expect all predecessors are in the loop");
1662     if (PN->getBasicBlockIndex(PredBB) >= 0) {
1663       BasicBlock *NewPred = SplitBlockPredecessors(
1664           ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true);
1665       // Since we do not allow splitting EH-block with BlockColors in
1666       // canSplitPredecessors(), we can simply assign predecessor's color to
1667       // the new block.
1668       if (!BlockColors.empty())
1669         // Grab a reference to the ColorVector to be inserted before getting the
1670         // reference to the vector we are copying because inserting the new
1671         // element in BlockColors might cause the map to be reallocated.
1672         SafetyInfo->copyColors(NewPred, PredBB);
1673     }
1674     PredBBs.remove(PredBB);
1675   }
1676 }
1677 
1678 /// When an instruction is found to only be used outside of the loop, this
1679 /// function moves it to the exit blocks and patches up SSA form as needed.
1680 /// This method is guaranteed to remove the original instruction from its
1681 /// position, and may either delete it or move it to outside of the loop.
1682 ///
1683 static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
1684                  BlockFrequencyInfo *BFI, const Loop *CurLoop,
1685                  ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU,
1686                  OptimizationRemarkEmitter *ORE) {
1687   bool Changed = false;
1688   LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
1689 
1690   // Iterate over users to be ready for actual sinking. Replace users via
1691   // unreachable blocks with undef and make all user PHIs trivially replaceable.
1692   SmallPtrSet<Instruction *, 8> VisitedUsers;
1693   for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) {
1694     auto *User = cast<Instruction>(*UI);
1695     Use &U = UI.getUse();
1696     ++UI;
1697 
1698     if (VisitedUsers.count(User) || CurLoop->contains(User))
1699       continue;
1700 
1701     if (!DT->isReachableFromEntry(User->getParent())) {
1702       U = UndefValue::get(I.getType());
1703       Changed = true;
1704       continue;
1705     }
1706 
1707     // The user must be a PHI node.
1708     PHINode *PN = cast<PHINode>(User);
1709 
1710     // Surprisingly, instructions can be used outside of loops without any
1711     // exits.  This can only happen in PHI nodes if the incoming block is
1712     // unreachable.
1713     BasicBlock *BB = PN->getIncomingBlock(U);
1714     if (!DT->isReachableFromEntry(BB)) {
1715       U = UndefValue::get(I.getType());
1716       Changed = true;
1717       continue;
1718     }
1719 
1720     VisitedUsers.insert(PN);
1721     if (isTriviallyReplaceablePHI(*PN, I))
1722       continue;
1723 
1724     if (!canSplitPredecessors(PN, SafetyInfo))
1725       return Changed;
1726 
1727     // Split predecessors of the PHI so that we can make users trivially
1728     // replaceable.
1729     splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU);
1730 
1731     // Should rebuild the iterators, as they may be invalidated by
1732     // splitPredecessorsOfLoopExit().
1733     UI = I.user_begin();
1734     UE = I.user_end();
1735   }
1736 
1737   if (VisitedUsers.empty())
1738     return Changed;
1739 
1740   ORE->emit([&]() {
1741     return OptimizationRemark(DEBUG_TYPE, "InstSunk", &I)
1742            << "sinking " << ore::NV("Inst", &I);
1743   });
1744   if (isa<LoadInst>(I))
1745     ++NumMovedLoads;
1746   else if (isa<CallInst>(I))
1747     ++NumMovedCalls;
1748   ++NumSunk;
1749 
1750 #ifndef NDEBUG
1751   SmallVector<BasicBlock *, 32> ExitBlocks;
1752   CurLoop->getUniqueExitBlocks(ExitBlocks);
1753   SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1754                                              ExitBlocks.end());
1755 #endif
1756 
1757   // Clones of this instruction. Don't create more than one per exit block!
1758   SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
1759 
1760   // If this instruction is only used outside of the loop, then all users are
1761   // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
1762   // the instruction.
1763   // First check if I is worth sinking for all uses. Sink only when it is worth
1764   // across all uses.
1765   SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end());
1766   SmallVector<PHINode *, 8> ExitPNs;
1767   for (auto *UI : Users) {
1768     auto *User = cast<Instruction>(UI);
1769 
1770     if (CurLoop->contains(User))
1771       continue;
1772 
1773     PHINode *PN = cast<PHINode>(User);
1774     assert(ExitBlockSet.count(PN->getParent()) &&
1775            "The LCSSA PHI is not in an exit block!");
1776     if (!worthSinkOrHoistInst(I, PN->getParent(), ORE, BFI)) {
1777       return Changed;
1778     }
1779 
1780     ExitPNs.push_back(PN);
1781   }
1782 
1783   for (auto *PN : ExitPNs) {
1784 
1785     // The PHI must be trivially replaceable.
1786     Instruction *New = sinkThroughTriviallyReplaceablePHI(
1787         PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU);
1788     PN->replaceAllUsesWith(New);
1789     eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr);
1790     Changed = true;
1791   }
1792   return Changed;
1793 }
1794 
1795 /// When an instruction is found to only use loop invariant operands that
1796 /// is safe to hoist, this instruction is called to do the dirty work.
1797 ///
1798 static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
1799                   BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
1800                   MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
1801                   OptimizationRemarkEmitter *ORE) {
1802   LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getNameOrAsOperand() << ": "
1803                     << I << "\n");
1804   ORE->emit([&]() {
1805     return OptimizationRemark(DEBUG_TYPE, "Hoisted", &I) << "hoisting "
1806                                                          << ore::NV("Inst", &I);
1807   });
1808 
1809   // Metadata can be dependent on conditions we are hoisting above.
1810   // Conservatively strip all metadata on the instruction unless we were
1811   // guaranteed to execute I if we entered the loop, in which case the metadata
1812   // is valid in the loop preheader.
1813   // Similarly, If I is a call and it is not guaranteed to execute in the loop,
1814   // then moving to the preheader means we should strip attributes on the call
1815   // that can cause UB since we may be hoisting above conditions that allowed
1816   // inferring those attributes. They may not be valid at the preheader.
1817   if ((I.hasMetadataOtherThanDebugLoc() || isa<CallInst>(I)) &&
1818       // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
1819       // time in isGuaranteedToExecute if we don't actually have anything to
1820       // drop.  It is a compile time optimization, not required for correctness.
1821       !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop))
1822     I.dropUndefImplyingAttrsAndUnknownMetadata();
1823 
1824   if (isa<PHINode>(I))
1825     // Move the new node to the end of the phi list in the destination block.
1826     moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU, SE);
1827   else
1828     // Move the new node to the destination block, before its terminator.
1829     moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU, SE);
1830 
1831   I.updateLocationAfterHoist();
1832 
1833   if (isa<LoadInst>(I))
1834     ++NumMovedLoads;
1835   else if (isa<CallInst>(I))
1836     ++NumMovedCalls;
1837   ++NumHoisted;
1838 }
1839 
1840 /// Only sink or hoist an instruction if it is not a trapping instruction,
1841 /// or if the instruction is known not to trap when moved to the preheader.
1842 /// or if it is a trapping instruction and is guaranteed to execute.
1843 static bool isSafeToExecuteUnconditionally(Instruction &Inst,
1844                                            const DominatorTree *DT,
1845                                            const TargetLibraryInfo *TLI,
1846                                            const Loop *CurLoop,
1847                                            const LoopSafetyInfo *SafetyInfo,
1848                                            OptimizationRemarkEmitter *ORE,
1849                                            const Instruction *CtxI) {
1850   if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT, TLI))
1851     return true;
1852 
1853   bool GuaranteedToExecute =
1854       SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop);
1855 
1856   if (!GuaranteedToExecute) {
1857     auto *LI = dyn_cast<LoadInst>(&Inst);
1858     if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1859       ORE->emit([&]() {
1860         return OptimizationRemarkMissed(
1861                    DEBUG_TYPE, "LoadWithLoopInvariantAddressCondExecuted", LI)
1862                << "failed to hoist load with loop-invariant address "
1863                   "because load is conditionally executed";
1864       });
1865   }
1866 
1867   return GuaranteedToExecute;
1868 }
1869 
1870 namespace {
1871 class LoopPromoter : public LoadAndStorePromoter {
1872   Value *SomePtr; // Designated pointer to store to.
1873   const SmallSetVector<Value *, 8> &PointerMustAliases;
1874   SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
1875   SmallVectorImpl<Instruction *> &LoopInsertPts;
1876   SmallVectorImpl<MemoryAccess *> &MSSAInsertPts;
1877   PredIteratorCache &PredCache;
1878   AliasSetTracker *AST;
1879   MemorySSAUpdater *MSSAU;
1880   LoopInfo &LI;
1881   DebugLoc DL;
1882   int Alignment;
1883   bool UnorderedAtomic;
1884   AAMDNodes AATags;
1885   ICFLoopSafetyInfo &SafetyInfo;
1886 
1887   // We're about to add a use of V in a loop exit block.  Insert an LCSSA phi
1888   // (if legal) if doing so would add an out-of-loop use to an instruction
1889   // defined in-loop.
1890   Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
1891     if (!LI.wouldBeOutOfLoopUseRequiringLCSSA(V, BB))
1892       return V;
1893 
1894     Instruction *I = cast<Instruction>(V);
1895     // We need to create an LCSSA PHI node for the incoming value and
1896     // store that.
1897     PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
1898                                   I->getName() + ".lcssa", &BB->front());
1899     for (BasicBlock *Pred : PredCache.get(BB))
1900       PN->addIncoming(I, Pred);
1901     return PN;
1902   }
1903 
1904 public:
1905   LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
1906                const SmallSetVector<Value *, 8> &PMA,
1907                SmallVectorImpl<BasicBlock *> &LEB,
1908                SmallVectorImpl<Instruction *> &LIP,
1909                SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC,
1910                AliasSetTracker *ast, MemorySSAUpdater *MSSAU, LoopInfo &li,
1911                DebugLoc dl, int alignment, bool UnorderedAtomic,
1912                const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo)
1913       : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
1914         LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP),
1915         PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)),
1916         Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags),
1917         SafetyInfo(SafetyInfo) {}
1918 
1919   bool isInstInList(Instruction *I,
1920                     const SmallVectorImpl<Instruction *> &) const override {
1921     Value *Ptr;
1922     if (LoadInst *LI = dyn_cast<LoadInst>(I))
1923       Ptr = LI->getOperand(0);
1924     else
1925       Ptr = cast<StoreInst>(I)->getPointerOperand();
1926     return PointerMustAliases.count(Ptr);
1927   }
1928 
1929   void doExtraRewritesBeforeFinalDeletion() override {
1930     // Insert stores after in the loop exit blocks.  Each exit block gets a
1931     // store of the live-out values that feed them.  Since we've already told
1932     // the SSA updater about the defs in the loop and the preheader
1933     // definition, it is all set and we can start using it.
1934     for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
1935       BasicBlock *ExitBlock = LoopExitBlocks[i];
1936       Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
1937       LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
1938       Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
1939       Instruction *InsertPos = LoopInsertPts[i];
1940       StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
1941       if (UnorderedAtomic)
1942         NewSI->setOrdering(AtomicOrdering::Unordered);
1943       NewSI->setAlignment(Align(Alignment));
1944       NewSI->setDebugLoc(DL);
1945       if (AATags)
1946         NewSI->setAAMetadata(AATags);
1947 
1948       if (MSSAU) {
1949         MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i];
1950         MemoryAccess *NewMemAcc;
1951         if (!MSSAInsertPoint) {
1952           NewMemAcc = MSSAU->createMemoryAccessInBB(
1953               NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning);
1954         } else {
1955           NewMemAcc =
1956               MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint);
1957         }
1958         MSSAInsertPts[i] = NewMemAcc;
1959         MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1960         // FIXME: true for safety, false may still be correct.
1961       }
1962     }
1963   }
1964 
1965   void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
1966     // Update alias analysis.
1967     if (AST)
1968       AST->copyValue(LI, V);
1969   }
1970   void instructionDeleted(Instruction *I) const override {
1971     SafetyInfo.removeInstruction(I);
1972     if (AST)
1973       AST->deleteValue(I);
1974     if (MSSAU)
1975       MSSAU->removeMemoryAccess(I);
1976   }
1977 };
1978 
1979 bool isNotCapturedBeforeOrInLoop(const Value *V, const Loop *L,
1980                                  DominatorTree *DT) {
1981   // We can perform the captured-before check against any instruction in the
1982   // loop header, as the loop header is reachable from any instruction inside
1983   // the loop.
1984   // TODO: ReturnCaptures=true shouldn't be necessary here.
1985   return !PointerMayBeCapturedBefore(V, /* ReturnCaptures */ true,
1986                                      /* StoreCaptures */ true,
1987                                      L->getHeader()->getTerminator(), DT);
1988 }
1989 
1990 /// Return true iff we can prove that a caller of this function can not inspect
1991 /// the contents of the provided object in a well defined program.
1992 bool isKnownNonEscaping(Value *Object, const Loop *L,
1993                         const TargetLibraryInfo *TLI, DominatorTree *DT) {
1994   if (isa<AllocaInst>(Object))
1995     // Since the alloca goes out of scope, we know the caller can't retain a
1996     // reference to it and be well defined.  Thus, we don't need to check for
1997     // capture.
1998     return true;
1999 
2000   // For all other objects we need to know that the caller can't possibly
2001   // have gotten a reference to the object.  There are two components of
2002   // that:
2003   //   1) Object can't be escaped by this function.  This is what
2004   //      PointerMayBeCaptured checks.
2005   //   2) Object can't have been captured at definition site.  For this, we
2006   //      need to know the return value is noalias.  At the moment, we use a
2007   //      weaker condition and handle only AllocLikeFunctions (which are
2008   //      known to be noalias).  TODO
2009   return isAllocLikeFn(Object, TLI) &&
2010          isNotCapturedBeforeOrInLoop(Object, L, DT);
2011 }
2012 
2013 } // namespace
2014 
2015 /// Try to promote memory values to scalars by sinking stores out of the
2016 /// loop and moving loads to before the loop.  We do this by looping over
2017 /// the stores in the loop, looking for stores to Must pointers which are
2018 /// loop invariant.
2019 ///
2020 bool llvm::promoteLoopAccessesToScalars(
2021     const SmallSetVector<Value *, 8> &PointerMustAliases,
2022     SmallVectorImpl<BasicBlock *> &ExitBlocks,
2023     SmallVectorImpl<Instruction *> &InsertPts,
2024     SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC,
2025     LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
2026     Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
2027     ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) {
2028   // Verify inputs.
2029   assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&
2030          SafetyInfo != nullptr &&
2031          "Unexpected Input to promoteLoopAccessesToScalars");
2032 
2033   Value *SomePtr = *PointerMustAliases.begin();
2034   BasicBlock *Preheader = CurLoop->getLoopPreheader();
2035 
2036   // It is not safe to promote a load/store from the loop if the load/store is
2037   // conditional.  For example, turning:
2038   //
2039   //    for () { if (c) *P += 1; }
2040   //
2041   // into:
2042   //
2043   //    tmp = *P;  for () { if (c) tmp +=1; } *P = tmp;
2044   //
2045   // is not safe, because *P may only be valid to access if 'c' is true.
2046   //
2047   // The safety property divides into two parts:
2048   // p1) The memory may not be dereferenceable on entry to the loop.  In this
2049   //    case, we can't insert the required load in the preheader.
2050   // p2) The memory model does not allow us to insert a store along any dynamic
2051   //    path which did not originally have one.
2052   //
2053   // If at least one store is guaranteed to execute, both properties are
2054   // satisfied, and promotion is legal.
2055   //
2056   // This, however, is not a necessary condition. Even if no store/load is
2057   // guaranteed to execute, we can still establish these properties.
2058   // We can establish (p1) by proving that hoisting the load into the preheader
2059   // is safe (i.e. proving dereferenceability on all paths through the loop). We
2060   // can use any access within the alias set to prove dereferenceability,
2061   // since they're all must alias.
2062   //
2063   // There are two ways establish (p2):
2064   // a) Prove the location is thread-local. In this case the memory model
2065   // requirement does not apply, and stores are safe to insert.
2066   // b) Prove a store dominates every exit block. In this case, if an exit
2067   // blocks is reached, the original dynamic path would have taken us through
2068   // the store, so inserting a store into the exit block is safe. Note that this
2069   // is different from the store being guaranteed to execute. For instance,
2070   // if an exception is thrown on the first iteration of the loop, the original
2071   // store is never executed, but the exit blocks are not executed either.
2072 
2073   bool DereferenceableInPH = false;
2074   bool SafeToInsertStore = false;
2075 
2076   SmallVector<Instruction *, 64> LoopUses;
2077 
2078   // We start with an alignment of one and try to find instructions that allow
2079   // us to prove better alignment.
2080   Align Alignment;
2081   // Keep track of which types of access we see
2082   bool SawUnorderedAtomic = false;
2083   bool SawNotAtomic = false;
2084   AAMDNodes AATags;
2085 
2086   const DataLayout &MDL = Preheader->getModule()->getDataLayout();
2087 
2088   bool IsKnownThreadLocalObject = false;
2089   if (SafetyInfo->anyBlockMayThrow()) {
2090     // If a loop can throw, we have to insert a store along each unwind edge.
2091     // That said, we can't actually make the unwind edge explicit. Therefore,
2092     // we have to prove that the store is dead along the unwind edge.  We do
2093     // this by proving that the caller can't have a reference to the object
2094     // after return and thus can't possibly load from the object.
2095     Value *Object = getUnderlyingObject(SomePtr);
2096     if (!isKnownNonEscaping(Object, CurLoop, TLI, DT))
2097       return false;
2098     // Subtlety: Alloca's aren't visible to callers, but *are* potentially
2099     // visible to other threads if captured and used during their lifetimes.
2100     IsKnownThreadLocalObject = !isa<AllocaInst>(Object);
2101   }
2102 
2103   // Check that all of the pointers in the alias set have the same type.  We
2104   // cannot (yet) promote a memory location that is loaded and stored in
2105   // different sizes.  While we are at it, collect alignment and AA info.
2106   for (Value *ASIV : PointerMustAliases) {
2107     // Check that all of the pointers in the alias set have the same type.  We
2108     // cannot (yet) promote a memory location that is loaded and stored in
2109     // different sizes.
2110     if (SomePtr->getType() != ASIV->getType())
2111       return false;
2112 
2113     for (User *U : ASIV->users()) {
2114       // Ignore instructions that are outside the loop.
2115       Instruction *UI = dyn_cast<Instruction>(U);
2116       if (!UI || !CurLoop->contains(UI))
2117         continue;
2118 
2119       // If there is an non-load/store instruction in the loop, we can't promote
2120       // it.
2121       if (LoadInst *Load = dyn_cast<LoadInst>(UI)) {
2122         if (!Load->isUnordered())
2123           return false;
2124 
2125         SawUnorderedAtomic |= Load->isAtomic();
2126         SawNotAtomic |= !Load->isAtomic();
2127 
2128         Align InstAlignment = Load->getAlign();
2129 
2130         // Note that proving a load safe to speculate requires proving
2131         // sufficient alignment at the target location.  Proving it guaranteed
2132         // to execute does as well.  Thus we can increase our guaranteed
2133         // alignment as well.
2134         if (!DereferenceableInPH || (InstAlignment > Alignment))
2135           if (isSafeToExecuteUnconditionally(*Load, DT, TLI, CurLoop,
2136                                              SafetyInfo, ORE,
2137                                              Preheader->getTerminator())) {
2138             DereferenceableInPH = true;
2139             Alignment = std::max(Alignment, InstAlignment);
2140           }
2141       } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
2142         // Stores *of* the pointer are not interesting, only stores *to* the
2143         // pointer.
2144         if (UI->getOperand(1) != ASIV)
2145           continue;
2146         if (!Store->isUnordered())
2147           return false;
2148 
2149         SawUnorderedAtomic |= Store->isAtomic();
2150         SawNotAtomic |= !Store->isAtomic();
2151 
2152         // If the store is guaranteed to execute, both properties are satisfied.
2153         // We may want to check if a store is guaranteed to execute even if we
2154         // already know that promotion is safe, since it may have higher
2155         // alignment than any other guaranteed stores, in which case we can
2156         // raise the alignment on the promoted store.
2157         Align InstAlignment = Store->getAlign();
2158 
2159         if (!DereferenceableInPH || !SafeToInsertStore ||
2160             (InstAlignment > Alignment)) {
2161           if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) {
2162             DereferenceableInPH = true;
2163             SafeToInsertStore = true;
2164             Alignment = std::max(Alignment, InstAlignment);
2165           }
2166         }
2167 
2168         // If a store dominates all exit blocks, it is safe to sink.
2169         // As explained above, if an exit block was executed, a dominating
2170         // store must have been executed at least once, so we are not
2171         // introducing stores on paths that did not have them.
2172         // Note that this only looks at explicit exit blocks. If we ever
2173         // start sinking stores into unwind edges (see above), this will break.
2174         if (!SafeToInsertStore)
2175           SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) {
2176             return DT->dominates(Store->getParent(), Exit);
2177           });
2178 
2179         // If the store is not guaranteed to execute, we may still get
2180         // deref info through it.
2181         if (!DereferenceableInPH) {
2182           DereferenceableInPH = isDereferenceableAndAlignedPointer(
2183               Store->getPointerOperand(), Store->getValueOperand()->getType(),
2184               Store->getAlign(), MDL, Preheader->getTerminator(), DT, TLI);
2185         }
2186       } else
2187         return false; // Not a load or store.
2188 
2189       // Merge the AA tags.
2190       if (LoopUses.empty()) {
2191         // On the first load/store, just take its AA tags.
2192         UI->getAAMetadata(AATags);
2193       } else if (AATags) {
2194         UI->getAAMetadata(AATags, /* Merge = */ true);
2195       }
2196 
2197       LoopUses.push_back(UI);
2198     }
2199   }
2200 
2201   // If we found both an unordered atomic instruction and a non-atomic memory
2202   // access, bail.  We can't blindly promote non-atomic to atomic since we
2203   // might not be able to lower the result.  We can't downgrade since that
2204   // would violate memory model.  Also, align 0 is an error for atomics.
2205   if (SawUnorderedAtomic && SawNotAtomic)
2206     return false;
2207 
2208   // If we're inserting an atomic load in the preheader, we must be able to
2209   // lower it.  We're only guaranteed to be able to lower naturally aligned
2210   // atomics.
2211   auto *SomePtrElemType = SomePtr->getType()->getPointerElementType();
2212   if (SawUnorderedAtomic &&
2213       Alignment < MDL.getTypeStoreSize(SomePtrElemType))
2214     return false;
2215 
2216   // If we couldn't prove we can hoist the load, bail.
2217   if (!DereferenceableInPH)
2218     return false;
2219 
2220   // We know we can hoist the load, but don't have a guaranteed store.
2221   // Check whether the location is thread-local. If it is, then we can insert
2222   // stores along paths which originally didn't have them without violating the
2223   // memory model.
2224   if (!SafeToInsertStore) {
2225     if (IsKnownThreadLocalObject)
2226       SafeToInsertStore = true;
2227     else {
2228       Value *Object = getUnderlyingObject(SomePtr);
2229       SafeToInsertStore =
2230           (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) &&
2231           isNotCapturedBeforeOrInLoop(Object, CurLoop, DT);
2232     }
2233   }
2234 
2235   // If we've still failed to prove we can sink the store, give up.
2236   if (!SafeToInsertStore)
2237     return false;
2238 
2239   // Otherwise, this is safe to promote, lets do it!
2240   LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr
2241                     << '\n');
2242   ORE->emit([&]() {
2243     return OptimizationRemark(DEBUG_TYPE, "PromoteLoopAccessesToScalar",
2244                               LoopUses[0])
2245            << "Moving accesses to memory location out of the loop";
2246   });
2247   ++NumPromoted;
2248 
2249   // Look at all the loop uses, and try to merge their locations.
2250   std::vector<const DILocation *> LoopUsesLocs;
2251   for (auto U : LoopUses)
2252     LoopUsesLocs.push_back(U->getDebugLoc().get());
2253   auto DL = DebugLoc(DILocation::getMergedLocations(LoopUsesLocs));
2254 
2255   // We use the SSAUpdater interface to insert phi nodes as required.
2256   SmallVector<PHINode *, 16> NewPHIs;
2257   SSAUpdater SSA(&NewPHIs);
2258   LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
2259                         InsertPts, MSSAInsertPts, PIC, CurAST, MSSAU, *LI, DL,
2260                         Alignment.value(), SawUnorderedAtomic, AATags,
2261                         *SafetyInfo);
2262 
2263   // Set up the preheader to have a definition of the value.  It is the live-out
2264   // value from the preheader that uses in the loop will use.
2265   LoadInst *PreheaderLoad = new LoadInst(
2266       SomePtr->getType()->getPointerElementType(), SomePtr,
2267       SomePtr->getName() + ".promoted", Preheader->getTerminator());
2268   if (SawUnorderedAtomic)
2269     PreheaderLoad->setOrdering(AtomicOrdering::Unordered);
2270   PreheaderLoad->setAlignment(Alignment);
2271   PreheaderLoad->setDebugLoc(DebugLoc());
2272   if (AATags)
2273     PreheaderLoad->setAAMetadata(AATags);
2274   SSA.AddAvailableValue(Preheader, PreheaderLoad);
2275 
2276   if (MSSAU) {
2277     MemoryAccess *PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB(
2278         PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End);
2279     MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess);
2280     MSSAU->insertUse(NewMemUse, /*RenameUses=*/true);
2281   }
2282 
2283   if (MSSAU && VerifyMemorySSA)
2284     MSSAU->getMemorySSA()->verifyMemorySSA();
2285   // Rewrite all the loads in the loop and remember all the definitions from
2286   // stores in the loop.
2287   Promoter.run(LoopUses);
2288 
2289   if (MSSAU && VerifyMemorySSA)
2290     MSSAU->getMemorySSA()->verifyMemorySSA();
2291   // If the SSAUpdater didn't use the load in the preheader, just zap it now.
2292   if (PreheaderLoad->use_empty())
2293     eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU);
2294 
2295   return true;
2296 }
2297 
2298 static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L,
2299                                 function_ref<void(Instruction *)> Fn) {
2300   for (const BasicBlock *BB : L->blocks())
2301     if (const auto *Accesses = MSSA->getBlockAccesses(BB))
2302       for (const auto &Access : *Accesses)
2303         if (const auto *MUD = dyn_cast<MemoryUseOrDef>(&Access))
2304           Fn(MUD->getMemoryInst());
2305 }
2306 
2307 static SmallVector<SmallSetVector<Value *, 8>, 0>
2308 collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L) {
2309   AliasSetTracker AST(*AA);
2310 
2311   auto IsPotentiallyPromotable = [L](const Instruction *I) {
2312     if (const auto *SI = dyn_cast<StoreInst>(I))
2313       return L->isLoopInvariant(SI->getPointerOperand());
2314     if (const auto *LI = dyn_cast<LoadInst>(I))
2315       return L->isLoopInvariant(LI->getPointerOperand());
2316     return false;
2317   };
2318 
2319   // Populate AST with potentially promotable accesses and remove them from
2320   // MaybePromotable, so they will not be checked again on the next iteration.
2321   SmallPtrSet<Value *, 16> AttemptingPromotion;
2322   foreachMemoryAccess(MSSA, L, [&](Instruction *I) {
2323     if (IsPotentiallyPromotable(I)) {
2324       AttemptingPromotion.insert(I);
2325       AST.add(I);
2326     }
2327   });
2328 
2329   // We're only interested in must-alias sets that contain a mod.
2330   SmallVector<const AliasSet *, 8> Sets;
2331   for (AliasSet &AS : AST)
2332     if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias())
2333       Sets.push_back(&AS);
2334 
2335   if (Sets.empty())
2336     return {}; // Nothing to promote...
2337 
2338   // Discard any sets for which there is an aliasing non-promotable access.
2339   foreachMemoryAccess(MSSA, L, [&](Instruction *I) {
2340     if (AttemptingPromotion.contains(I))
2341       return;
2342 
2343     llvm::erase_if(Sets, [&](const AliasSet *AS) {
2344       return AS->aliasesUnknownInst(I, *AA);
2345     });
2346   });
2347 
2348   SmallVector<SmallSetVector<Value *, 8>, 0> Result;
2349   for (const AliasSet *Set : Sets) {
2350     SmallSetVector<Value *, 8> PointerMustAliases;
2351     for (const auto &ASI : *Set)
2352       PointerMustAliases.insert(ASI.getValue());
2353     Result.push_back(std::move(PointerMustAliases));
2354   }
2355 
2356   return Result;
2357 }
2358 
2359 /// Returns an owning pointer to an alias set which incorporates aliasing info
2360 /// from L and all subloops of L.
2361 std::unique_ptr<AliasSetTracker>
2362 LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
2363                                                  AAResults *AA) {
2364   auto CurAST = std::make_unique<AliasSetTracker>(*AA);
2365 
2366   // Add everything from all the sub loops.
2367   for (Loop *InnerL : L->getSubLoops())
2368     for (BasicBlock *BB : InnerL->blocks())
2369       CurAST->add(*BB);
2370 
2371   // And merge in this loop (without anything from inner loops).
2372   for (BasicBlock *BB : L->blocks())
2373     if (LI->getLoopFor(BB) == L)
2374       CurAST->add(*BB);
2375 
2376   return CurAST;
2377 }
2378 
2379 static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
2380                                      AliasSetTracker *CurAST, Loop *CurLoop,
2381                                      AAResults *AA) {
2382   // First check to see if any of the basic blocks in CurLoop invalidate *V.
2383   bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod();
2384 
2385   if (!isInvalidatedAccordingToAST || !LICMN2Theshold)
2386     return isInvalidatedAccordingToAST;
2387 
2388   // Check with a diagnostic analysis if we can refine the information above.
2389   // This is to identify the limitations of using the AST.
2390   // The alias set mechanism used by LICM has a major weakness in that it
2391   // combines all things which may alias into a single set *before* asking
2392   // modref questions. As a result, a single readonly call within a loop will
2393   // collapse all loads and stores into a single alias set and report
2394   // invalidation if the loop contains any store. For example, readonly calls
2395   // with deopt states have this form and create a general alias set with all
2396   // loads and stores.  In order to get any LICM in loops containing possible
2397   // deopt states we need a more precise invalidation of checking the mod ref
2398   // info of each instruction within the loop and LI. This has a complexity of
2399   // O(N^2), so currently, it is used only as a diagnostic tool since the
2400   // default value of LICMN2Threshold is zero.
2401 
2402   // Don't look at nested loops.
2403   if (CurLoop->begin() != CurLoop->end())
2404     return true;
2405 
2406   int N = 0;
2407   for (BasicBlock *BB : CurLoop->getBlocks())
2408     for (Instruction &I : *BB) {
2409       if (N >= LICMN2Theshold) {
2410         LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "
2411                           << *(MemLoc.Ptr) << "\n");
2412         return true;
2413       }
2414       N++;
2415       auto Res = AA->getModRefInfo(&I, MemLoc);
2416       if (isModSet(Res)) {
2417         LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "
2418                           << *(MemLoc.Ptr) << "\n");
2419         return true;
2420       }
2421     }
2422   LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n");
2423   return false;
2424 }
2425 
2426 bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
2427                                       Loop *CurLoop, Instruction &I,
2428                                       SinkAndHoistLICMFlags &Flags) {
2429   // For hoisting, use the walker to determine safety
2430   if (!Flags.getIsSink()) {
2431     MemoryAccess *Source;
2432     // See declaration of SetLicmMssaOptCap for usage details.
2433     if (Flags.tooManyClobberingCalls())
2434       Source = MU->getDefiningAccess();
2435     else {
2436       Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU);
2437       Flags.incrementClobberingCalls();
2438     }
2439     return !MSSA->isLiveOnEntryDef(Source) &&
2440            CurLoop->contains(Source->getBlock());
2441   }
2442 
2443   // For sinking, we'd need to check all Defs below this use. The getClobbering
2444   // call will look on the backedge of the loop, but will check aliasing with
2445   // the instructions on the previous iteration.
2446   // For example:
2447   // for (i ... )
2448   //   load a[i] ( Use (LoE)
2449   //   store a[i] ( 1 = Def (2), with 2 = Phi for the loop.
2450   //   i++;
2451   // The load sees no clobbering inside the loop, as the backedge alias check
2452   // does phi translation, and will check aliasing against store a[i-1].
2453   // However sinking the load outside the loop, below the store is incorrect.
2454 
2455   // For now, only sink if there are no Defs in the loop, and the existing ones
2456   // precede the use and are in the same block.
2457   // FIXME: Increase precision: Safe to sink if Use post dominates the Def;
2458   // needs PostDominatorTreeAnalysis.
2459   // FIXME: More precise: no Defs that alias this Use.
2460   if (Flags.tooManyMemoryAccesses())
2461     return true;
2462   for (auto *BB : CurLoop->getBlocks())
2463     if (pointerInvalidatedByBlockWithMSSA(*BB, *MSSA, *MU))
2464       return true;
2465   // When sinking, the source block may not be part of the loop so check it.
2466   if (!CurLoop->contains(&I))
2467     return pointerInvalidatedByBlockWithMSSA(*I.getParent(), *MSSA, *MU);
2468 
2469   return false;
2470 }
2471 
2472 bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA,
2473                                        MemoryUse &MU) {
2474   if (const auto *Accesses = MSSA.getBlockDefs(&BB))
2475     for (const auto &MA : *Accesses)
2476       if (const auto *MD = dyn_cast<MemoryDef>(&MA))
2477         if (MU.getBlock() != MD->getBlock() || !MSSA.locallyDominates(MD, &MU))
2478           return true;
2479   return false;
2480 }
2481 
2482 /// Little predicate that returns true if the specified basic block is in
2483 /// a subloop of the current one, not the current one itself.
2484 ///
2485 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
2486   assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
2487   return LI->getLoopFor(BB) != CurLoop;
2488 }
2489