1 //===- LoopUnroll.cpp - Loop unroller 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 implements a simple loop unroller.  It works best when loops have
10 // been canonicalized by the -indvars pass, allowing it to determine the trip
11 // counts of loops easily.
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Scalar/LoopUnrollPass.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseMapInfo.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/StringRef.h"
25 #include "llvm/Analysis/AssumptionCache.h"
26 #include "llvm/Analysis/BlockFrequencyInfo.h"
27 #include "llvm/Analysis/CodeMetrics.h"
28 #include "llvm/Analysis/LazyBlockFrequencyInfo.h"
29 #include "llvm/Analysis/LoopAnalysisManager.h"
30 #include "llvm/Analysis/LoopInfo.h"
31 #include "llvm/Analysis/LoopPass.h"
32 #include "llvm/Analysis/LoopUnrollAnalyzer.h"
33 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
34 #include "llvm/Analysis/ProfileSummaryInfo.h"
35 #include "llvm/Analysis/ScalarEvolution.h"
36 #include "llvm/Analysis/TargetTransformInfo.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DiagnosticInfo.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Function.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Metadata.h"
48 #include "llvm/IR/PassManager.h"
49 #include "llvm/InitializePasses.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Support/Casting.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/ErrorHandling.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Scalar.h"
57 #include "llvm/Transforms/Scalar/LoopPassManager.h"
58 #include "llvm/Transforms/Utils.h"
59 #include "llvm/Transforms/Utils/LoopSimplify.h"
60 #include "llvm/Transforms/Utils/LoopUtils.h"
61 #include "llvm/Transforms/Utils/SizeOpts.h"
62 #include "llvm/Transforms/Utils/UnrollLoop.h"
63 #include <algorithm>
64 #include <cassert>
65 #include <cstdint>
66 #include <limits>
67 #include <string>
68 #include <tuple>
69 #include <utility>
70 
71 using namespace llvm;
72 
73 #define DEBUG_TYPE "loop-unroll"
74 
75 cl::opt<bool> llvm::ForgetSCEVInLoopUnroll(
76     "forget-scev-loop-unroll", cl::init(false), cl::Hidden,
77     cl::desc("Forget everything in SCEV when doing LoopUnroll, instead of just"
78              " the current top-most loop. This is somtimes preferred to reduce"
79              " compile time."));
80 
81 static cl::opt<unsigned>
82     UnrollThreshold("unroll-threshold", cl::Hidden,
83                     cl::desc("The cost threshold for loop unrolling"));
84 
85 static cl::opt<unsigned> UnrollPartialThreshold(
86     "unroll-partial-threshold", cl::Hidden,
87     cl::desc("The cost threshold for partial loop unrolling"));
88 
89 static cl::opt<unsigned> UnrollMaxPercentThresholdBoost(
90     "unroll-max-percent-threshold-boost", cl::init(400), cl::Hidden,
91     cl::desc("The maximum 'boost' (represented as a percentage >= 100) applied "
92              "to the threshold when aggressively unrolling a loop due to the "
93              "dynamic cost savings. If completely unrolling a loop will reduce "
94              "the total runtime from X to Y, we boost the loop unroll "
95              "threshold to DefaultThreshold*std::min(MaxPercentThresholdBoost, "
96              "X/Y). This limit avoids excessive code bloat."));
97 
98 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
99     "unroll-max-iteration-count-to-analyze", cl::init(10), cl::Hidden,
100     cl::desc("Don't allow loop unrolling to simulate more than this number of"
101              "iterations when checking full unroll profitability"));
102 
103 static cl::opt<unsigned> UnrollCount(
104     "unroll-count", cl::Hidden,
105     cl::desc("Use this unroll count for all loops including those with "
106              "unroll_count pragma values, for testing purposes"));
107 
108 static cl::opt<unsigned> UnrollMaxCount(
109     "unroll-max-count", cl::Hidden,
110     cl::desc("Set the max unroll count for partial and runtime unrolling, for"
111              "testing purposes"));
112 
113 static cl::opt<unsigned> UnrollFullMaxCount(
114     "unroll-full-max-count", cl::Hidden,
115     cl::desc(
116         "Set the max unroll count for full unrolling, for testing purposes"));
117 
118 static cl::opt<unsigned> UnrollPeelCount(
119     "unroll-peel-count", cl::Hidden,
120     cl::desc("Set the unroll peeling count, for testing purposes"));
121 
122 static cl::opt<bool>
123     UnrollAllowPartial("unroll-allow-partial", cl::Hidden,
124                        cl::desc("Allows loops to be partially unrolled until "
125                                 "-unroll-threshold loop size is reached."));
126 
127 static cl::opt<bool> UnrollAllowRemainder(
128     "unroll-allow-remainder", cl::Hidden,
129     cl::desc("Allow generation of a loop remainder (extra iterations) "
130              "when unrolling a loop."));
131 
132 static cl::opt<bool>
133     UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::Hidden,
134                   cl::desc("Unroll loops with run-time trip counts"));
135 
136 static cl::opt<unsigned> UnrollMaxUpperBound(
137     "unroll-max-upperbound", cl::init(8), cl::Hidden,
138     cl::desc(
139         "The max of trip count upper bound that is considered in unrolling"));
140 
141 static cl::opt<unsigned> PragmaUnrollThreshold(
142     "pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
143     cl::desc("Unrolled size limit for loops with an unroll(full) or "
144              "unroll_count pragma."));
145 
146 static cl::opt<unsigned> FlatLoopTripCountThreshold(
147     "flat-loop-tripcount-threshold", cl::init(5), cl::Hidden,
148     cl::desc("If the runtime tripcount for the loop is lower than the "
149              "threshold, the loop is considered as flat and will be less "
150              "aggressively unrolled."));
151 
152 static cl::opt<bool>
153     UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden,
154                        cl::desc("Allows loops to be peeled when the dynamic "
155                                 "trip count is known to be low."));
156 
157 static cl::opt<bool> UnrollUnrollRemainder(
158   "unroll-remainder", cl::Hidden,
159   cl::desc("Allow the loop remainder to be unrolled."));
160 
161 // This option isn't ever intended to be enabled, it serves to allow
162 // experiments to check the assumptions about when this kind of revisit is
163 // necessary.
164 static cl::opt<bool> UnrollRevisitChildLoops(
165     "unroll-revisit-child-loops", cl::Hidden,
166     cl::desc("Enqueue and re-visit child loops in the loop PM after unrolling. "
167              "This shouldn't typically be needed as child loops (or their "
168              "clones) were already visited."));
169 
170 /// A magic value for use with the Threshold parameter to indicate
171 /// that the loop unroll should be performed regardless of how much
172 /// code expansion would result.
173 static const unsigned NoThreshold = std::numeric_limits<unsigned>::max();
174 
175 /// Gather the various unrolling parameters based on the defaults, compiler
176 /// flags, TTI overrides and user specified parameters.
177 TargetTransformInfo::UnrollingPreferences llvm::gatherUnrollingPreferences(
178     Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
179     BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, int OptLevel,
180     Optional<unsigned> UserThreshold, Optional<unsigned> UserCount,
181     Optional<bool> UserAllowPartial, Optional<bool> UserRuntime,
182     Optional<bool> UserUpperBound, Optional<bool> UserAllowPeeling,
183     Optional<bool> UserAllowProfileBasedPeeling,
184     Optional<unsigned> UserFullUnrollMaxCount) {
185   TargetTransformInfo::UnrollingPreferences UP;
186 
187   // Set up the defaults
188   UP.Threshold = OptLevel > 2 ? 300 : 150;
189   UP.MaxPercentThresholdBoost = 400;
190   UP.OptSizeThreshold = 0;
191   UP.PartialThreshold = 150;
192   UP.PartialOptSizeThreshold = 0;
193   UP.Count = 0;
194   UP.PeelCount = 0;
195   UP.DefaultUnrollRuntimeCount = 8;
196   UP.MaxCount = std::numeric_limits<unsigned>::max();
197   UP.FullUnrollMaxCount = std::numeric_limits<unsigned>::max();
198   UP.BEInsns = 2;
199   UP.Partial = false;
200   UP.Runtime = false;
201   UP.AllowRemainder = true;
202   UP.UnrollRemainder = false;
203   UP.AllowExpensiveTripCount = false;
204   UP.Force = false;
205   UP.UpperBound = false;
206   UP.AllowPeeling = true;
207   UP.UnrollAndJam = false;
208   UP.PeelProfiledIterations = true;
209   UP.UnrollAndJamInnerLoopThreshold = 60;
210 
211   // Override with any target specific settings
212   TTI.getUnrollingPreferences(L, SE, UP);
213 
214   // Apply size attributes
215   bool OptForSize = L->getHeader()->getParent()->hasOptSize() ||
216                     llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,
217                                                 PGSOQueryType::IRPass);
218   if (OptForSize) {
219     UP.Threshold = UP.OptSizeThreshold;
220     UP.PartialThreshold = UP.PartialOptSizeThreshold;
221     UP.MaxPercentThresholdBoost = 100;
222   }
223 
224   // Apply any user values specified by cl::opt
225   if (UnrollThreshold.getNumOccurrences() > 0)
226     UP.Threshold = UnrollThreshold;
227   if (UnrollPartialThreshold.getNumOccurrences() > 0)
228     UP.PartialThreshold = UnrollPartialThreshold;
229   if (UnrollMaxPercentThresholdBoost.getNumOccurrences() > 0)
230     UP.MaxPercentThresholdBoost = UnrollMaxPercentThresholdBoost;
231   if (UnrollMaxCount.getNumOccurrences() > 0)
232     UP.MaxCount = UnrollMaxCount;
233   if (UnrollFullMaxCount.getNumOccurrences() > 0)
234     UP.FullUnrollMaxCount = UnrollFullMaxCount;
235   if (UnrollPeelCount.getNumOccurrences() > 0)
236     UP.PeelCount = UnrollPeelCount;
237   if (UnrollAllowPartial.getNumOccurrences() > 0)
238     UP.Partial = UnrollAllowPartial;
239   if (UnrollAllowRemainder.getNumOccurrences() > 0)
240     UP.AllowRemainder = UnrollAllowRemainder;
241   if (UnrollRuntime.getNumOccurrences() > 0)
242     UP.Runtime = UnrollRuntime;
243   if (UnrollMaxUpperBound == 0)
244     UP.UpperBound = false;
245   if (UnrollAllowPeeling.getNumOccurrences() > 0)
246     UP.AllowPeeling = UnrollAllowPeeling;
247   if (UnrollUnrollRemainder.getNumOccurrences() > 0)
248     UP.UnrollRemainder = UnrollUnrollRemainder;
249 
250   // Apply user values provided by argument
251   if (UserThreshold.hasValue()) {
252     UP.Threshold = *UserThreshold;
253     UP.PartialThreshold = *UserThreshold;
254   }
255   if (UserCount.hasValue())
256     UP.Count = *UserCount;
257   if (UserAllowPartial.hasValue())
258     UP.Partial = *UserAllowPartial;
259   if (UserRuntime.hasValue())
260     UP.Runtime = *UserRuntime;
261   if (UserUpperBound.hasValue())
262     UP.UpperBound = *UserUpperBound;
263   if (UserAllowPeeling.hasValue())
264     UP.AllowPeeling = *UserAllowPeeling;
265   if (UserAllowProfileBasedPeeling.hasValue())
266     UP.PeelProfiledIterations = *UserAllowProfileBasedPeeling;
267   if (UserFullUnrollMaxCount.hasValue())
268     UP.FullUnrollMaxCount = *UserFullUnrollMaxCount;
269 
270   return UP;
271 }
272 
273 namespace {
274 
275 /// A struct to densely store the state of an instruction after unrolling at
276 /// each iteration.
277 ///
278 /// This is designed to work like a tuple of <Instruction *, int> for the
279 /// purposes of hashing and lookup, but to be able to associate two boolean
280 /// states with each key.
281 struct UnrolledInstState {
282   Instruction *I;
283   int Iteration : 30;
284   unsigned IsFree : 1;
285   unsigned IsCounted : 1;
286 };
287 
288 /// Hashing and equality testing for a set of the instruction states.
289 struct UnrolledInstStateKeyInfo {
290   using PtrInfo = DenseMapInfo<Instruction *>;
291   using PairInfo = DenseMapInfo<std::pair<Instruction *, int>>;
292 
293   static inline UnrolledInstState getEmptyKey() {
294     return {PtrInfo::getEmptyKey(), 0, 0, 0};
295   }
296 
297   static inline UnrolledInstState getTombstoneKey() {
298     return {PtrInfo::getTombstoneKey(), 0, 0, 0};
299   }
300 
301   static inline unsigned getHashValue(const UnrolledInstState &S) {
302     return PairInfo::getHashValue({S.I, S.Iteration});
303   }
304 
305   static inline bool isEqual(const UnrolledInstState &LHS,
306                              const UnrolledInstState &RHS) {
307     return PairInfo::isEqual({LHS.I, LHS.Iteration}, {RHS.I, RHS.Iteration});
308   }
309 };
310 
311 struct EstimatedUnrollCost {
312   /// The estimated cost after unrolling.
313   unsigned UnrolledCost;
314 
315   /// The estimated dynamic cost of executing the instructions in the
316   /// rolled form.
317   unsigned RolledDynamicCost;
318 };
319 
320 } // end anonymous namespace
321 
322 /// Figure out if the loop is worth full unrolling.
323 ///
324 /// Complete loop unrolling can make some loads constant, and we need to know
325 /// if that would expose any further optimization opportunities.  This routine
326 /// estimates this optimization.  It computes cost of unrolled loop
327 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
328 /// dynamic cost we mean that we won't count costs of blocks that are known not
329 /// to be executed (i.e. if we have a branch in the loop and we know that at the
330 /// given iteration its condition would be resolved to true, we won't add up the
331 /// cost of the 'false'-block).
332 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
333 /// the analysis failed (no benefits expected from the unrolling, or the loop is
334 /// too big to analyze), the returned value is None.
335 static Optional<EstimatedUnrollCost> analyzeLoopUnrollCost(
336     const Loop *L, unsigned TripCount, DominatorTree &DT, ScalarEvolution &SE,
337     const SmallPtrSetImpl<const Value *> &EphValues,
338     const TargetTransformInfo &TTI, unsigned MaxUnrolledLoopSize) {
339   // We want to be able to scale offsets by the trip count and add more offsets
340   // to them without checking for overflows, and we already don't want to
341   // analyze *massive* trip counts, so we force the max to be reasonably small.
342   assert(UnrollMaxIterationsCountToAnalyze <
343              (unsigned)(std::numeric_limits<int>::max() / 2) &&
344          "The unroll iterations max is too large!");
345 
346   // Only analyze inner loops. We can't properly estimate cost of nested loops
347   // and we won't visit inner loops again anyway.
348   if (!L->empty())
349     return None;
350 
351   // Don't simulate loops with a big or unknown tripcount
352   if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
353       TripCount > UnrollMaxIterationsCountToAnalyze)
354     return None;
355 
356   SmallSetVector<BasicBlock *, 16> BBWorklist;
357   SmallSetVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitWorklist;
358   DenseMap<Value *, Constant *> SimplifiedValues;
359   SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues;
360 
361   // The estimated cost of the unrolled form of the loop. We try to estimate
362   // this by simplifying as much as we can while computing the estimate.
363   unsigned UnrolledCost = 0;
364 
365   // We also track the estimated dynamic (that is, actually executed) cost in
366   // the rolled form. This helps identify cases when the savings from unrolling
367   // aren't just exposing dead control flows, but actual reduced dynamic
368   // instructions due to the simplifications which we expect to occur after
369   // unrolling.
370   unsigned RolledDynamicCost = 0;
371 
372   // We track the simplification of each instruction in each iteration. We use
373   // this to recursively merge costs into the unrolled cost on-demand so that
374   // we don't count the cost of any dead code. This is essentially a map from
375   // <instruction, int> to <bool, bool>, but stored as a densely packed struct.
376   DenseSet<UnrolledInstState, UnrolledInstStateKeyInfo> InstCostMap;
377 
378   // A small worklist used to accumulate cost of instructions from each
379   // observable and reached root in the loop.
380   SmallVector<Instruction *, 16> CostWorklist;
381 
382   // PHI-used worklist used between iterations while accumulating cost.
383   SmallVector<Instruction *, 4> PHIUsedList;
384 
385   // Helper function to accumulate cost for instructions in the loop.
386   auto AddCostRecursively = [&](Instruction &RootI, int Iteration) {
387     assert(Iteration >= 0 && "Cannot have a negative iteration!");
388     assert(CostWorklist.empty() && "Must start with an empty cost list");
389     assert(PHIUsedList.empty() && "Must start with an empty phi used list");
390     CostWorklist.push_back(&RootI);
391     for (;; --Iteration) {
392       do {
393         Instruction *I = CostWorklist.pop_back_val();
394 
395         // InstCostMap only uses I and Iteration as a key, the other two values
396         // don't matter here.
397         auto CostIter = InstCostMap.find({I, Iteration, 0, 0});
398         if (CostIter == InstCostMap.end())
399           // If an input to a PHI node comes from a dead path through the loop
400           // we may have no cost data for it here. What that actually means is
401           // that it is free.
402           continue;
403         auto &Cost = *CostIter;
404         if (Cost.IsCounted)
405           // Already counted this instruction.
406           continue;
407 
408         // Mark that we are counting the cost of this instruction now.
409         Cost.IsCounted = true;
410 
411         // If this is a PHI node in the loop header, just add it to the PHI set.
412         if (auto *PhiI = dyn_cast<PHINode>(I))
413           if (PhiI->getParent() == L->getHeader()) {
414             assert(Cost.IsFree && "Loop PHIs shouldn't be evaluated as they "
415                                   "inherently simplify during unrolling.");
416             if (Iteration == 0)
417               continue;
418 
419             // Push the incoming value from the backedge into the PHI used list
420             // if it is an in-loop instruction. We'll use this to populate the
421             // cost worklist for the next iteration (as we count backwards).
422             if (auto *OpI = dyn_cast<Instruction>(
423                     PhiI->getIncomingValueForBlock(L->getLoopLatch())))
424               if (L->contains(OpI))
425                 PHIUsedList.push_back(OpI);
426             continue;
427           }
428 
429         // First accumulate the cost of this instruction.
430         if (!Cost.IsFree) {
431           UnrolledCost += TTI.getUserCost(I);
432           LLVM_DEBUG(dbgs() << "Adding cost of instruction (iteration "
433                             << Iteration << "): ");
434           LLVM_DEBUG(I->dump());
435         }
436 
437         // We must count the cost of every operand which is not free,
438         // recursively. If we reach a loop PHI node, simply add it to the set
439         // to be considered on the next iteration (backwards!).
440         for (Value *Op : I->operands()) {
441           // Check whether this operand is free due to being a constant or
442           // outside the loop.
443           auto *OpI = dyn_cast<Instruction>(Op);
444           if (!OpI || !L->contains(OpI))
445             continue;
446 
447           // Otherwise accumulate its cost.
448           CostWorklist.push_back(OpI);
449         }
450       } while (!CostWorklist.empty());
451 
452       if (PHIUsedList.empty())
453         // We've exhausted the search.
454         break;
455 
456       assert(Iteration > 0 &&
457              "Cannot track PHI-used values past the first iteration!");
458       CostWorklist.append(PHIUsedList.begin(), PHIUsedList.end());
459       PHIUsedList.clear();
460     }
461   };
462 
463   // Ensure that we don't violate the loop structure invariants relied on by
464   // this analysis.
465   assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
466   assert(L->isLCSSAForm(DT) &&
467          "Must have loops in LCSSA form to track live-out values.");
468 
469   LLVM_DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
470 
471   // Simulate execution of each iteration of the loop counting instructions,
472   // which would be simplified.
473   // Since the same load will take different values on different iterations,
474   // we literally have to go through all loop's iterations.
475   for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
476     LLVM_DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
477 
478     // Prepare for the iteration by collecting any simplified entry or backedge
479     // inputs.
480     for (Instruction &I : *L->getHeader()) {
481       auto *PHI = dyn_cast<PHINode>(&I);
482       if (!PHI)
483         break;
484 
485       // The loop header PHI nodes must have exactly two input: one from the
486       // loop preheader and one from the loop latch.
487       assert(
488           PHI->getNumIncomingValues() == 2 &&
489           "Must have an incoming value only for the preheader and the latch.");
490 
491       Value *V = PHI->getIncomingValueForBlock(
492           Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
493       Constant *C = dyn_cast<Constant>(V);
494       if (Iteration != 0 && !C)
495         C = SimplifiedValues.lookup(V);
496       if (C)
497         SimplifiedInputValues.push_back({PHI, C});
498     }
499 
500     // Now clear and re-populate the map for the next iteration.
501     SimplifiedValues.clear();
502     while (!SimplifiedInputValues.empty())
503       SimplifiedValues.insert(SimplifiedInputValues.pop_back_val());
504 
505     UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L);
506 
507     BBWorklist.clear();
508     BBWorklist.insert(L->getHeader());
509     // Note that we *must not* cache the size, this loop grows the worklist.
510     for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
511       BasicBlock *BB = BBWorklist[Idx];
512 
513       // Visit all instructions in the given basic block and try to simplify
514       // it.  We don't change the actual IR, just count optimization
515       // opportunities.
516       for (Instruction &I : *BB) {
517         // These won't get into the final code - don't even try calculating the
518         // cost for them.
519         if (isa<DbgInfoIntrinsic>(I) || EphValues.count(&I))
520           continue;
521 
522         // Track this instruction's expected baseline cost when executing the
523         // rolled loop form.
524         RolledDynamicCost += TTI.getUserCost(&I);
525 
526         // Visit the instruction to analyze its loop cost after unrolling,
527         // and if the visitor returns true, mark the instruction as free after
528         // unrolling and continue.
529         bool IsFree = Analyzer.visit(I);
530         bool Inserted = InstCostMap.insert({&I, (int)Iteration,
531                                            (unsigned)IsFree,
532                                            /*IsCounted*/ false}).second;
533         (void)Inserted;
534         assert(Inserted && "Cannot have a state for an unvisited instruction!");
535 
536         if (IsFree)
537           continue;
538 
539         // Can't properly model a cost of a call.
540         // FIXME: With a proper cost model we should be able to do it.
541         if (auto *CI = dyn_cast<CallInst>(&I)) {
542           const Function *Callee = CI->getCalledFunction();
543           if (!Callee || TTI.isLoweredToCall(Callee)) {
544             LLVM_DEBUG(dbgs() << "Can't analyze cost of loop with call\n");
545             return None;
546           }
547         }
548 
549         // If the instruction might have a side-effect recursively account for
550         // the cost of it and all the instructions leading up to it.
551         if (I.mayHaveSideEffects())
552           AddCostRecursively(I, Iteration);
553 
554         // If unrolled body turns out to be too big, bail out.
555         if (UnrolledCost > MaxUnrolledLoopSize) {
556           LLVM_DEBUG(dbgs() << "  Exceeded threshold.. exiting.\n"
557                             << "  UnrolledCost: " << UnrolledCost
558                             << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
559                             << "\n");
560           return None;
561         }
562       }
563 
564       Instruction *TI = BB->getTerminator();
565 
566       // Add in the live successors by first checking whether we have terminator
567       // that may be simplified based on the values simplified by this call.
568       BasicBlock *KnownSucc = nullptr;
569       if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
570         if (BI->isConditional()) {
571           if (Constant *SimpleCond =
572                   SimplifiedValues.lookup(BI->getCondition())) {
573             // Just take the first successor if condition is undef
574             if (isa<UndefValue>(SimpleCond))
575               KnownSucc = BI->getSuccessor(0);
576             else if (ConstantInt *SimpleCondVal =
577                          dyn_cast<ConstantInt>(SimpleCond))
578               KnownSucc = BI->getSuccessor(SimpleCondVal->isZero() ? 1 : 0);
579           }
580         }
581       } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
582         if (Constant *SimpleCond =
583                 SimplifiedValues.lookup(SI->getCondition())) {
584           // Just take the first successor if condition is undef
585           if (isa<UndefValue>(SimpleCond))
586             KnownSucc = SI->getSuccessor(0);
587           else if (ConstantInt *SimpleCondVal =
588                        dyn_cast<ConstantInt>(SimpleCond))
589             KnownSucc = SI->findCaseValue(SimpleCondVal)->getCaseSuccessor();
590         }
591       }
592       if (KnownSucc) {
593         if (L->contains(KnownSucc))
594           BBWorklist.insert(KnownSucc);
595         else
596           ExitWorklist.insert({BB, KnownSucc});
597         continue;
598       }
599 
600       // Add BB's successors to the worklist.
601       for (BasicBlock *Succ : successors(BB))
602         if (L->contains(Succ))
603           BBWorklist.insert(Succ);
604         else
605           ExitWorklist.insert({BB, Succ});
606       AddCostRecursively(*TI, Iteration);
607     }
608 
609     // If we found no optimization opportunities on the first iteration, we
610     // won't find them on later ones too.
611     if (UnrolledCost == RolledDynamicCost) {
612       LLVM_DEBUG(dbgs() << "  No opportunities found.. exiting.\n"
613                         << "  UnrolledCost: " << UnrolledCost << "\n");
614       return None;
615     }
616   }
617 
618   while (!ExitWorklist.empty()) {
619     BasicBlock *ExitingBB, *ExitBB;
620     std::tie(ExitingBB, ExitBB) = ExitWorklist.pop_back_val();
621 
622     for (Instruction &I : *ExitBB) {
623       auto *PN = dyn_cast<PHINode>(&I);
624       if (!PN)
625         break;
626 
627       Value *Op = PN->getIncomingValueForBlock(ExitingBB);
628       if (auto *OpI = dyn_cast<Instruction>(Op))
629         if (L->contains(OpI))
630           AddCostRecursively(*OpI, TripCount - 1);
631     }
632   }
633 
634   LLVM_DEBUG(dbgs() << "Analysis finished:\n"
635                     << "UnrolledCost: " << UnrolledCost << ", "
636                     << "RolledDynamicCost: " << RolledDynamicCost << "\n");
637   return {{UnrolledCost, RolledDynamicCost}};
638 }
639 
640 /// ApproximateLoopSize - Approximate the size of the loop.
641 unsigned llvm::ApproximateLoopSize(
642     const Loop *L, unsigned &NumCalls, bool &NotDuplicatable, bool &Convergent,
643     const TargetTransformInfo &TTI,
644     const SmallPtrSetImpl<const Value *> &EphValues, unsigned BEInsns) {
645   CodeMetrics Metrics;
646   for (BasicBlock *BB : L->blocks())
647     Metrics.analyzeBasicBlock(BB, TTI, EphValues);
648   NumCalls = Metrics.NumInlineCandidates;
649   NotDuplicatable = Metrics.notDuplicatable;
650   Convergent = Metrics.convergent;
651 
652   unsigned LoopSize = Metrics.NumInsts;
653 
654   // Don't allow an estimate of size zero.  This would allows unrolling of loops
655   // with huge iteration counts, which is a compile time problem even if it's
656   // not a problem for code quality. Also, the code using this size may assume
657   // that each loop has at least three instructions (likely a conditional
658   // branch, a comparison feeding that branch, and some kind of loop increment
659   // feeding that comparison instruction).
660   LoopSize = std::max(LoopSize, BEInsns + 1);
661 
662   return LoopSize;
663 }
664 
665 // Returns the loop hint metadata node with the given name (for example,
666 // "llvm.loop.unroll.count").  If no such metadata node exists, then nullptr is
667 // returned.
668 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
669   if (MDNode *LoopID = L->getLoopID())
670     return GetUnrollMetadata(LoopID, Name);
671   return nullptr;
672 }
673 
674 // Returns true if the loop has an unroll(full) pragma.
675 static bool HasUnrollFullPragma(const Loop *L) {
676   return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
677 }
678 
679 // Returns true if the loop has an unroll(enable) pragma. This metadata is used
680 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
681 static bool HasUnrollEnablePragma(const Loop *L) {
682   return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.enable");
683 }
684 
685 // Returns true if the loop has an runtime unroll(disable) pragma.
686 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
687   return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
688 }
689 
690 // If loop has an unroll_count pragma return the (necessarily
691 // positive) value from the pragma.  Otherwise return 0.
692 static unsigned UnrollCountPragmaValue(const Loop *L) {
693   MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
694   if (MD) {
695     assert(MD->getNumOperands() == 2 &&
696            "Unroll count hint metadata should have two operands.");
697     unsigned Count =
698         mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
699     assert(Count >= 1 && "Unroll count must be positive.");
700     return Count;
701   }
702   return 0;
703 }
704 
705 // Computes the boosting factor for complete unrolling.
706 // If fully unrolling the loop would save a lot of RolledDynamicCost, it would
707 // be beneficial to fully unroll the loop even if unrolledcost is large. We
708 // use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust
709 // the unroll threshold.
710 static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost,
711                                             unsigned MaxPercentThresholdBoost) {
712   if (Cost.RolledDynamicCost >= std::numeric_limits<unsigned>::max() / 100)
713     return 100;
714   else if (Cost.UnrolledCost != 0)
715     // The boosting factor is RolledDynamicCost / UnrolledCost
716     return std::min(100 * Cost.RolledDynamicCost / Cost.UnrolledCost,
717                     MaxPercentThresholdBoost);
718   else
719     return MaxPercentThresholdBoost;
720 }
721 
722 // Returns loop size estimation for unrolled loop.
723 static uint64_t getUnrolledLoopSize(
724     unsigned LoopSize,
725     TargetTransformInfo::UnrollingPreferences &UP) {
726   assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
727   return (uint64_t)(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns;
728 }
729 
730 // Returns true if unroll count was set explicitly.
731 // Calculates unroll count and writes it to UP.Count.
732 // Unless IgnoreUser is true, will also use metadata and command-line options
733 // that are specific to to the LoopUnroll pass (which, for instance, are
734 // irrelevant for the LoopUnrollAndJam pass).
735 // FIXME: This function is used by LoopUnroll and LoopUnrollAndJam, but consumes
736 // many LoopUnroll-specific options. The shared functionality should be
737 // refactored into it own function.
738 bool llvm::computeUnrollCount(
739     Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI,
740     ScalarEvolution &SE, const SmallPtrSetImpl<const Value *> &EphValues,
741     OptimizationRemarkEmitter *ORE, unsigned &TripCount, unsigned MaxTripCount,
742     bool MaxOrZero, unsigned &TripMultiple, unsigned LoopSize,
743     TargetTransformInfo::UnrollingPreferences &UP, bool &UseUpperBound) {
744 
745   // Check for explicit Count.
746   // 1st priority is unroll count set by "unroll-count" option.
747   bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0;
748   if (UserUnrollCount) {
749     UP.Count = UnrollCount;
750     UP.AllowExpensiveTripCount = true;
751     UP.Force = true;
752     if (UP.AllowRemainder && getUnrolledLoopSize(LoopSize, UP) < UP.Threshold)
753       return true;
754   }
755 
756   // 2nd priority is unroll count set by pragma.
757   unsigned PragmaCount = UnrollCountPragmaValue(L);
758   if (PragmaCount > 0) {
759     UP.Count = PragmaCount;
760     UP.Runtime = true;
761     UP.AllowExpensiveTripCount = true;
762     UP.Force = true;
763     if ((UP.AllowRemainder || (TripMultiple % PragmaCount == 0)) &&
764         getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
765       return true;
766   }
767   bool PragmaFullUnroll = HasUnrollFullPragma(L);
768   if (PragmaFullUnroll && TripCount != 0) {
769     UP.Count = TripCount;
770     if (getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
771       return false;
772   }
773 
774   bool PragmaEnableUnroll = HasUnrollEnablePragma(L);
775   bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll ||
776                         PragmaEnableUnroll || UserUnrollCount;
777 
778   if (ExplicitUnroll && TripCount != 0) {
779     // If the loop has an unrolling pragma, we want to be more aggressive with
780     // unrolling limits. Set thresholds to at least the PragmaUnrollThreshold
781     // value which is larger than the default limits.
782     UP.Threshold = std::max<unsigned>(UP.Threshold, PragmaUnrollThreshold);
783     UP.PartialThreshold =
784         std::max<unsigned>(UP.PartialThreshold, PragmaUnrollThreshold);
785   }
786 
787   // 3rd priority is full unroll count.
788   // Full unroll makes sense only when TripCount or its upper bound could be
789   // statically calculated.
790   // Also we need to check if we exceed FullUnrollMaxCount.
791   // If using the upper bound to unroll, TripMultiple should be set to 1 because
792   // we do not know when loop may exit.
793 
794   // We can unroll by the upper bound amount if it's generally allowed or if
795   // we know that the loop is executed either the upper bound or zero times.
796   // (MaxOrZero unrolling keeps only the first loop test, so the number of
797   // loop tests remains the same compared to the non-unrolled version, whereas
798   // the generic upper bound unrolling keeps all but the last loop test so the
799   // number of loop tests goes up which may end up being worse on targets with
800   // constrained branch predictor resources so is controlled by an option.)
801   // In addition we only unroll small upper bounds.
802   unsigned FullUnrollMaxTripCount = MaxTripCount;
803   if (!(UP.UpperBound || MaxOrZero) ||
804       FullUnrollMaxTripCount > UnrollMaxUpperBound)
805     FullUnrollMaxTripCount = 0;
806 
807   // UnrollByMaxCount and ExactTripCount cannot both be non zero since we only
808   // compute the former when the latter is zero.
809   unsigned ExactTripCount = TripCount;
810   assert((ExactTripCount == 0 || FullUnrollMaxTripCount == 0) &&
811          "ExtractTripCount and UnrollByMaxCount cannot both be non zero.");
812 
813   unsigned FullUnrollTripCount =
814       ExactTripCount ? ExactTripCount : FullUnrollMaxTripCount;
815   UP.Count = FullUnrollTripCount;
816   if (FullUnrollTripCount && FullUnrollTripCount <= UP.FullUnrollMaxCount) {
817     // When computing the unrolled size, note that BEInsns are not replicated
818     // like the rest of the loop body.
819     if (getUnrolledLoopSize(LoopSize, UP) < UP.Threshold) {
820       UseUpperBound = (FullUnrollMaxTripCount == FullUnrollTripCount);
821       TripCount = FullUnrollTripCount;
822       TripMultiple = UP.UpperBound ? 1 : TripMultiple;
823       return ExplicitUnroll;
824     } else {
825       // The loop isn't that small, but we still can fully unroll it if that
826       // helps to remove a significant number of instructions.
827       // To check that, run additional analysis on the loop.
828       if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
829               L, FullUnrollTripCount, DT, SE, EphValues, TTI,
830               UP.Threshold * UP.MaxPercentThresholdBoost / 100)) {
831         unsigned Boost =
832             getFullUnrollBoostingFactor(*Cost, UP.MaxPercentThresholdBoost);
833         if (Cost->UnrolledCost < UP.Threshold * Boost / 100) {
834           UseUpperBound = (FullUnrollMaxTripCount == FullUnrollTripCount);
835           TripCount = FullUnrollTripCount;
836           TripMultiple = UP.UpperBound ? 1 : TripMultiple;
837           return ExplicitUnroll;
838         }
839       }
840     }
841   }
842 
843   // 4th priority is loop peeling.
844   computePeelCount(L, LoopSize, UP, TripCount, SE);
845   if (UP.PeelCount) {
846     UP.Runtime = false;
847     UP.Count = 1;
848     return ExplicitUnroll;
849   }
850 
851   // 5th priority is partial unrolling.
852   // Try partial unroll only when TripCount could be statically calculated.
853   if (TripCount) {
854     UP.Partial |= ExplicitUnroll;
855     if (!UP.Partial) {
856       LLVM_DEBUG(dbgs() << "  will not try to unroll partially because "
857                         << "-unroll-allow-partial not given\n");
858       UP.Count = 0;
859       return false;
860     }
861     if (UP.Count == 0)
862       UP.Count = TripCount;
863     if (UP.PartialThreshold != NoThreshold) {
864       // Reduce unroll count to be modulo of TripCount for partial unrolling.
865       if (getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
866         UP.Count =
867             (std::max(UP.PartialThreshold, UP.BEInsns + 1) - UP.BEInsns) /
868             (LoopSize - UP.BEInsns);
869       if (UP.Count > UP.MaxCount)
870         UP.Count = UP.MaxCount;
871       while (UP.Count != 0 && TripCount % UP.Count != 0)
872         UP.Count--;
873       if (UP.AllowRemainder && UP.Count <= 1) {
874         // If there is no Count that is modulo of TripCount, set Count to
875         // largest power-of-two factor that satisfies the threshold limit.
876         // As we'll create fixup loop, do the type of unrolling only if
877         // remainder loop is allowed.
878         UP.Count = UP.DefaultUnrollRuntimeCount;
879         while (UP.Count != 0 &&
880                getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
881           UP.Count >>= 1;
882       }
883       if (UP.Count < 2) {
884         if (PragmaEnableUnroll)
885           ORE->emit([&]() {
886             return OptimizationRemarkMissed(DEBUG_TYPE,
887                                             "UnrollAsDirectedTooLarge",
888                                             L->getStartLoc(), L->getHeader())
889                    << "Unable to unroll loop as directed by unroll(enable) "
890                       "pragma "
891                       "because unrolled size is too large.";
892           });
893         UP.Count = 0;
894       }
895     } else {
896       UP.Count = TripCount;
897     }
898     if (UP.Count > UP.MaxCount)
899       UP.Count = UP.MaxCount;
900     if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
901         UP.Count != TripCount)
902       ORE->emit([&]() {
903         return OptimizationRemarkMissed(DEBUG_TYPE,
904                                         "FullUnrollAsDirectedTooLarge",
905                                         L->getStartLoc(), L->getHeader())
906                << "Unable to fully unroll loop as directed by unroll pragma "
907                   "because "
908                   "unrolled size is too large.";
909       });
910     LLVM_DEBUG(dbgs() << "  partially unrolling with count: " << UP.Count
911                       << "\n");
912     return ExplicitUnroll;
913   }
914   assert(TripCount == 0 &&
915          "All cases when TripCount is constant should be covered here.");
916   if (PragmaFullUnroll)
917     ORE->emit([&]() {
918       return OptimizationRemarkMissed(
919                  DEBUG_TYPE, "CantFullUnrollAsDirectedRuntimeTripCount",
920                  L->getStartLoc(), L->getHeader())
921              << "Unable to fully unroll loop as directed by unroll(full) "
922                 "pragma "
923                 "because loop has a runtime trip count.";
924     });
925 
926   // 6th priority is runtime unrolling.
927   // Don't unroll a runtime trip count loop when it is disabled.
928   if (HasRuntimeUnrollDisablePragma(L)) {
929     UP.Count = 0;
930     return false;
931   }
932 
933   // Don't unroll a small upper bound loop unless user or TTI asked to do so.
934   if (MaxTripCount && !UP.Force && MaxTripCount < UnrollMaxUpperBound) {
935     UP.Count = 0;
936     return false;
937   }
938 
939   // Check if the runtime trip count is too small when profile is available.
940   if (L->getHeader()->getParent()->hasProfileData()) {
941     if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) {
942       if (*ProfileTripCount < FlatLoopTripCountThreshold)
943         return false;
944       else
945         UP.AllowExpensiveTripCount = true;
946     }
947   }
948 
949   // Reduce count based on the type of unrolling and the threshold values.
950   UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount;
951   if (!UP.Runtime) {
952     LLVM_DEBUG(
953         dbgs() << "  will not try to unroll loop with runtime trip count "
954                << "-unroll-runtime not given\n");
955     UP.Count = 0;
956     return false;
957   }
958   if (UP.Count == 0)
959     UP.Count = UP.DefaultUnrollRuntimeCount;
960 
961   // Reduce unroll count to be the largest power-of-two factor of
962   // the original count which satisfies the threshold limit.
963   while (UP.Count != 0 &&
964          getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
965     UP.Count >>= 1;
966 
967 #ifndef NDEBUG
968   unsigned OrigCount = UP.Count;
969 #endif
970 
971   if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) {
972     while (UP.Count != 0 && TripMultiple % UP.Count != 0)
973       UP.Count >>= 1;
974     LLVM_DEBUG(
975         dbgs() << "Remainder loop is restricted (that could architecture "
976                   "specific or because the loop contains a convergent "
977                   "instruction), so unroll count must divide the trip "
978                   "multiple, "
979                << TripMultiple << ".  Reducing unroll count from " << OrigCount
980                << " to " << UP.Count << ".\n");
981 
982     using namespace ore;
983 
984     if (PragmaCount > 0 && !UP.AllowRemainder)
985       ORE->emit([&]() {
986         return OptimizationRemarkMissed(DEBUG_TYPE,
987                                         "DifferentUnrollCountFromDirected",
988                                         L->getStartLoc(), L->getHeader())
989                << "Unable to unroll loop the number of times directed by "
990                   "unroll_count pragma because remainder loop is restricted "
991                   "(that could architecture specific or because the loop "
992                   "contains a convergent instruction) and so must have an "
993                   "unroll "
994                   "count that divides the loop trip multiple of "
995                << NV("TripMultiple", TripMultiple) << ".  Unrolling instead "
996                << NV("UnrollCount", UP.Count) << " time(s).";
997       });
998   }
999 
1000   if (UP.Count > UP.MaxCount)
1001     UP.Count = UP.MaxCount;
1002 
1003   if (MaxTripCount && UP.Count > MaxTripCount)
1004     UP.Count = MaxTripCount;
1005 
1006   LLVM_DEBUG(dbgs() << "  runtime unrolling with count: " << UP.Count
1007                     << "\n");
1008   if (UP.Count < 2)
1009     UP.Count = 0;
1010   return ExplicitUnroll;
1011 }
1012 
1013 static LoopUnrollResult tryToUnrollLoop(
1014     Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE,
1015     const TargetTransformInfo &TTI, AssumptionCache &AC,
1016     OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
1017     ProfileSummaryInfo *PSI, bool PreserveLCSSA, int OptLevel,
1018     bool OnlyWhenForced, bool ForgetAllSCEV, Optional<unsigned> ProvidedCount,
1019     Optional<unsigned> ProvidedThreshold, Optional<bool> ProvidedAllowPartial,
1020     Optional<bool> ProvidedRuntime, Optional<bool> ProvidedUpperBound,
1021     Optional<bool> ProvidedAllowPeeling,
1022     Optional<bool> ProvidedAllowProfileBasedPeeling,
1023     Optional<unsigned> ProvidedFullUnrollMaxCount) {
1024   LLVM_DEBUG(dbgs() << "Loop Unroll: F["
1025                     << L->getHeader()->getParent()->getName() << "] Loop %"
1026                     << L->getHeader()->getName() << "\n");
1027   TransformationMode TM = hasUnrollTransformation(L);
1028   if (TM & TM_Disable)
1029     return LoopUnrollResult::Unmodified;
1030   if (!L->isLoopSimplifyForm()) {
1031     LLVM_DEBUG(
1032         dbgs() << "  Not unrolling loop which is not in loop-simplify form.\n");
1033     return LoopUnrollResult::Unmodified;
1034   }
1035 
1036   // When automtatic unrolling is disabled, do not unroll unless overridden for
1037   // this loop.
1038   if (OnlyWhenForced && !(TM & TM_Enable))
1039     return LoopUnrollResult::Unmodified;
1040 
1041   bool OptForSize = L->getHeader()->getParent()->hasOptSize();
1042   unsigned NumInlineCandidates;
1043   bool NotDuplicatable;
1044   bool Convergent;
1045   TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
1046       L, SE, TTI, BFI, PSI, OptLevel, ProvidedThreshold, ProvidedCount,
1047       ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound,
1048       ProvidedAllowPeeling, ProvidedAllowProfileBasedPeeling,
1049       ProvidedFullUnrollMaxCount);
1050 
1051   // Exit early if unrolling is disabled. For OptForSize, we pick the loop size
1052   // as threshold later on.
1053   if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0) &&
1054       !OptForSize)
1055     return LoopUnrollResult::Unmodified;
1056 
1057   SmallPtrSet<const Value *, 32> EphValues;
1058   CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
1059 
1060   unsigned LoopSize =
1061       ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent,
1062                           TTI, EphValues, UP.BEInsns);
1063   LLVM_DEBUG(dbgs() << "  Loop Size = " << LoopSize << "\n");
1064   if (NotDuplicatable) {
1065     LLVM_DEBUG(dbgs() << "  Not unrolling loop which contains non-duplicatable"
1066                       << " instructions.\n");
1067     return LoopUnrollResult::Unmodified;
1068   }
1069 
1070   // When optimizing for size, use LoopSize + 1 as threshold (we use < Threshold
1071   // later), to (fully) unroll loops, if it does not increase code size.
1072   if (OptForSize)
1073     UP.Threshold = std::max(UP.Threshold, LoopSize + 1);
1074 
1075   if (NumInlineCandidates != 0) {
1076     LLVM_DEBUG(dbgs() << "  Not unrolling loop with inlinable calls.\n");
1077     return LoopUnrollResult::Unmodified;
1078   }
1079 
1080   // Find trip count and trip multiple if count is not available
1081   unsigned TripCount = 0;
1082   unsigned TripMultiple = 1;
1083   // If there are multiple exiting blocks but one of them is the latch, use the
1084   // latch for the trip count estimation. Otherwise insist on a single exiting
1085   // block for the trip count estimation.
1086   BasicBlock *ExitingBlock = L->getLoopLatch();
1087   if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
1088     ExitingBlock = L->getExitingBlock();
1089   if (ExitingBlock) {
1090     TripCount = SE.getSmallConstantTripCount(L, ExitingBlock);
1091     TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock);
1092   }
1093 
1094   // If the loop contains a convergent operation, the prelude we'd add
1095   // to do the first few instructions before we hit the unrolled loop
1096   // is unsafe -- it adds a control-flow dependency to the convergent
1097   // operation.  Therefore restrict remainder loop (try unrollig without).
1098   //
1099   // TODO: This is quite conservative.  In practice, convergent_op()
1100   // is likely to be called unconditionally in the loop.  In this
1101   // case, the program would be ill-formed (on most architectures)
1102   // unless n were the same on all threads in a thread group.
1103   // Assuming n is the same on all threads, any kind of unrolling is
1104   // safe.  But currently llvm's notion of convergence isn't powerful
1105   // enough to express this.
1106   if (Convergent)
1107     UP.AllowRemainder = false;
1108 
1109   // Try to find the trip count upper bound if we cannot find the exact trip
1110   // count.
1111   unsigned MaxTripCount = 0;
1112   bool MaxOrZero = false;
1113   if (!TripCount) {
1114     MaxTripCount = SE.getSmallConstantMaxTripCount(L);
1115     MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L);
1116   }
1117 
1118   // computeUnrollCount() decides whether it is beneficial to use upper bound to
1119   // fully unroll the loop.
1120   bool UseUpperBound = false;
1121   bool IsCountSetExplicitly = computeUnrollCount(
1122       L, TTI, DT, LI, SE, EphValues, &ORE, TripCount, MaxTripCount, MaxOrZero,
1123       TripMultiple, LoopSize, UP, UseUpperBound);
1124   if (!UP.Count)
1125     return LoopUnrollResult::Unmodified;
1126   // Unroll factor (Count) must be less or equal to TripCount.
1127   if (TripCount && UP.Count > TripCount)
1128     UP.Count = TripCount;
1129 
1130   // Save loop properties before it is transformed.
1131   MDNode *OrigLoopID = L->getLoopID();
1132 
1133   // Unroll the loop.
1134   Loop *RemainderLoop = nullptr;
1135   LoopUnrollResult UnrollResult = UnrollLoop(
1136       L,
1137       {UP.Count, TripCount, UP.Force, UP.Runtime, UP.AllowExpensiveTripCount,
1138        UseUpperBound, MaxOrZero, TripMultiple, UP.PeelCount, UP.UnrollRemainder,
1139        ForgetAllSCEV},
1140       LI, &SE, &DT, &AC, &ORE, PreserveLCSSA, &RemainderLoop);
1141   if (UnrollResult == LoopUnrollResult::Unmodified)
1142     return LoopUnrollResult::Unmodified;
1143 
1144   if (RemainderLoop) {
1145     Optional<MDNode *> RemainderLoopID =
1146         makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1147                                         LLVMLoopUnrollFollowupRemainder});
1148     if (RemainderLoopID.hasValue())
1149       RemainderLoop->setLoopID(RemainderLoopID.getValue());
1150   }
1151 
1152   if (UnrollResult != LoopUnrollResult::FullyUnrolled) {
1153     Optional<MDNode *> NewLoopID =
1154         makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1155                                         LLVMLoopUnrollFollowupUnrolled});
1156     if (NewLoopID.hasValue()) {
1157       L->setLoopID(NewLoopID.getValue());
1158 
1159       // Do not setLoopAlreadyUnrolled if loop attributes have been specified
1160       // explicitly.
1161       return UnrollResult;
1162     }
1163   }
1164 
1165   // If loop has an unroll count pragma or unrolled by explicitly set count
1166   // mark loop as unrolled to prevent unrolling beyond that requested.
1167   // If the loop was peeled, we already "used up" the profile information
1168   // we had, so we don't want to unroll or peel again.
1169   if (UnrollResult != LoopUnrollResult::FullyUnrolled &&
1170       (IsCountSetExplicitly || (UP.PeelProfiledIterations && UP.PeelCount)))
1171     L->setLoopAlreadyUnrolled();
1172 
1173   return UnrollResult;
1174 }
1175 
1176 namespace {
1177 
1178 class LoopUnroll : public LoopPass {
1179 public:
1180   static char ID; // Pass ID, replacement for typeid
1181 
1182   int OptLevel;
1183 
1184   /// If false, use a cost model to determine whether unrolling of a loop is
1185   /// profitable. If true, only loops that explicitly request unrolling via
1186   /// metadata are considered. All other loops are skipped.
1187   bool OnlyWhenForced;
1188 
1189   /// If false, when SCEV is invalidated, only forget everything in the
1190   /// top-most loop (call forgetTopMostLoop), of the loop being processed.
1191   /// Otherwise, forgetAllLoops and rebuild when needed next.
1192   bool ForgetAllSCEV;
1193 
1194   Optional<unsigned> ProvidedCount;
1195   Optional<unsigned> ProvidedThreshold;
1196   Optional<bool> ProvidedAllowPartial;
1197   Optional<bool> ProvidedRuntime;
1198   Optional<bool> ProvidedUpperBound;
1199   Optional<bool> ProvidedAllowPeeling;
1200   Optional<bool> ProvidedAllowProfileBasedPeeling;
1201   Optional<unsigned> ProvidedFullUnrollMaxCount;
1202 
1203   LoopUnroll(int OptLevel = 2, bool OnlyWhenForced = false,
1204              bool ForgetAllSCEV = false, Optional<unsigned> Threshold = None,
1205              Optional<unsigned> Count = None,
1206              Optional<bool> AllowPartial = None, Optional<bool> Runtime = None,
1207              Optional<bool> UpperBound = None,
1208              Optional<bool> AllowPeeling = None,
1209              Optional<bool> AllowProfileBasedPeeling = None,
1210              Optional<unsigned> ProvidedFullUnrollMaxCount = None)
1211       : LoopPass(ID), OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
1212         ForgetAllSCEV(ForgetAllSCEV), ProvidedCount(std::move(Count)),
1213         ProvidedThreshold(Threshold), ProvidedAllowPartial(AllowPartial),
1214         ProvidedRuntime(Runtime), ProvidedUpperBound(UpperBound),
1215         ProvidedAllowPeeling(AllowPeeling),
1216         ProvidedAllowProfileBasedPeeling(AllowProfileBasedPeeling),
1217         ProvidedFullUnrollMaxCount(ProvidedFullUnrollMaxCount) {
1218     initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
1219   }
1220 
1221   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1222     if (skipLoop(L))
1223       return false;
1224 
1225     Function &F = *L->getHeader()->getParent();
1226 
1227     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1228     LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1229     ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1230     const TargetTransformInfo &TTI =
1231         getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1232     auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1233     // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
1234     // pass.  Function analyses need to be preserved across loop transformations
1235     // but ORE cannot be preserved (see comment before the pass definition).
1236     OptimizationRemarkEmitter ORE(&F);
1237     bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1238 
1239     LoopUnrollResult Result = tryToUnrollLoop(
1240         L, DT, LI, SE, TTI, AC, ORE, nullptr, nullptr, PreserveLCSSA, OptLevel,
1241         OnlyWhenForced, ForgetAllSCEV, ProvidedCount, ProvidedThreshold,
1242         ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound,
1243         ProvidedAllowPeeling, ProvidedAllowProfileBasedPeeling,
1244         ProvidedFullUnrollMaxCount);
1245 
1246     if (Result == LoopUnrollResult::FullyUnrolled)
1247       LPM.markLoopAsDeleted(*L);
1248 
1249     return Result != LoopUnrollResult::Unmodified;
1250   }
1251 
1252   /// This transformation requires natural loop information & requires that
1253   /// loop preheaders be inserted into the CFG...
1254   void getAnalysisUsage(AnalysisUsage &AU) const override {
1255     AU.addRequired<AssumptionCacheTracker>();
1256     AU.addRequired<TargetTransformInfoWrapperPass>();
1257     // FIXME: Loop passes are required to preserve domtree, and for now we just
1258     // recreate dom info if anything gets unrolled.
1259     getLoopAnalysisUsage(AU);
1260   }
1261 };
1262 
1263 } // end anonymous namespace
1264 
1265 char LoopUnroll::ID = 0;
1266 
1267 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1268 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1269 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1270 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1271 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1272 
1273 Pass *llvm::createLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1274                                  bool ForgetAllSCEV, int Threshold, int Count,
1275                                  int AllowPartial, int Runtime, int UpperBound,
1276                                  int AllowPeeling) {
1277   // TODO: It would make more sense for this function to take the optionals
1278   // directly, but that's dangerous since it would silently break out of tree
1279   // callers.
1280   return new LoopUnroll(
1281       OptLevel, OnlyWhenForced, ForgetAllSCEV,
1282       Threshold == -1 ? None : Optional<unsigned>(Threshold),
1283       Count == -1 ? None : Optional<unsigned>(Count),
1284       AllowPartial == -1 ? None : Optional<bool>(AllowPartial),
1285       Runtime == -1 ? None : Optional<bool>(Runtime),
1286       UpperBound == -1 ? None : Optional<bool>(UpperBound),
1287       AllowPeeling == -1 ? None : Optional<bool>(AllowPeeling));
1288 }
1289 
1290 Pass *llvm::createSimpleLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1291                                        bool ForgetAllSCEV) {
1292   return createLoopUnrollPass(OptLevel, OnlyWhenForced, ForgetAllSCEV, -1, -1,
1293                               0, 0, 0, 0);
1294 }
1295 
1296 PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
1297                                           LoopStandardAnalysisResults &AR,
1298                                           LPMUpdater &Updater) {
1299   const auto &FAM =
1300       AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
1301   Function *F = L.getHeader()->getParent();
1302 
1303   auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F);
1304   // FIXME: This should probably be optional rather than required.
1305   if (!ORE)
1306     report_fatal_error(
1307         "LoopFullUnrollPass: OptimizationRemarkEmitterAnalysis not "
1308         "cached at a higher level");
1309 
1310   // Keep track of the previous loop structure so we can identify new loops
1311   // created by unrolling.
1312   Loop *ParentL = L.getParentLoop();
1313   SmallPtrSet<Loop *, 4> OldLoops;
1314   if (ParentL)
1315     OldLoops.insert(ParentL->begin(), ParentL->end());
1316   else
1317     OldLoops.insert(AR.LI.begin(), AR.LI.end());
1318 
1319   std::string LoopName = L.getName();
1320 
1321   bool Changed = tryToUnrollLoop(&L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, *ORE,
1322                                  /*BFI*/ nullptr, /*PSI*/ nullptr,
1323                                  /*PreserveLCSSA*/ true, OptLevel,
1324                                  OnlyWhenForced, ForgetSCEV, /*Count*/ None,
1325                                  /*Threshold*/ None, /*AllowPartial*/ false,
1326                                  /*Runtime*/ false, /*UpperBound*/ false,
1327                                  /*AllowPeeling*/ false,
1328                                  /*AllowProfileBasedPeeling*/ false,
1329                                  /*FullUnrollMaxCount*/ None) !=
1330                  LoopUnrollResult::Unmodified;
1331   if (!Changed)
1332     return PreservedAnalyses::all();
1333 
1334   // The parent must not be damaged by unrolling!
1335 #ifndef NDEBUG
1336   if (ParentL)
1337     ParentL->verifyLoop();
1338 #endif
1339 
1340   // Unrolling can do several things to introduce new loops into a loop nest:
1341   // - Full unrolling clones child loops within the current loop but then
1342   //   removes the current loop making all of the children appear to be new
1343   //   sibling loops.
1344   //
1345   // When a new loop appears as a sibling loop after fully unrolling,
1346   // its nesting structure has fundamentally changed and we want to revisit
1347   // it to reflect that.
1348   //
1349   // When unrolling has removed the current loop, we need to tell the
1350   // infrastructure that it is gone.
1351   //
1352   // Finally, we support a debugging/testing mode where we revisit child loops
1353   // as well. These are not expected to require further optimizations as either
1354   // they or the loop they were cloned from have been directly visited already.
1355   // But the debugging mode allows us to check this assumption.
1356   bool IsCurrentLoopValid = false;
1357   SmallVector<Loop *, 4> SibLoops;
1358   if (ParentL)
1359     SibLoops.append(ParentL->begin(), ParentL->end());
1360   else
1361     SibLoops.append(AR.LI.begin(), AR.LI.end());
1362   erase_if(SibLoops, [&](Loop *SibLoop) {
1363     if (SibLoop == &L) {
1364       IsCurrentLoopValid = true;
1365       return true;
1366     }
1367 
1368     // Otherwise erase the loop from the list if it was in the old loops.
1369     return OldLoops.count(SibLoop) != 0;
1370   });
1371   Updater.addSiblingLoops(SibLoops);
1372 
1373   if (!IsCurrentLoopValid) {
1374     Updater.markLoopAsDeleted(L, LoopName);
1375   } else {
1376     // We can only walk child loops if the current loop remained valid.
1377     if (UnrollRevisitChildLoops) {
1378       // Walk *all* of the child loops.
1379       SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end());
1380       Updater.addChildLoops(ChildLoops);
1381     }
1382   }
1383 
1384   return getLoopPassPreservedAnalyses();
1385 }
1386 
1387 template <typename RangeT>
1388 static SmallVector<Loop *, 8> appendLoopsToWorklist(RangeT &&Loops) {
1389   SmallVector<Loop *, 8> Worklist;
1390   // We use an internal worklist to build up the preorder traversal without
1391   // recursion.
1392   SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist;
1393 
1394   for (Loop *RootL : Loops) {
1395     assert(PreOrderLoops.empty() && "Must start with an empty preorder walk.");
1396     assert(PreOrderWorklist.empty() &&
1397            "Must start with an empty preorder walk worklist.");
1398     PreOrderWorklist.push_back(RootL);
1399     do {
1400       Loop *L = PreOrderWorklist.pop_back_val();
1401       PreOrderWorklist.append(L->begin(), L->end());
1402       PreOrderLoops.push_back(L);
1403     } while (!PreOrderWorklist.empty());
1404 
1405     Worklist.append(PreOrderLoops.begin(), PreOrderLoops.end());
1406     PreOrderLoops.clear();
1407   }
1408   return Worklist;
1409 }
1410 
1411 PreservedAnalyses LoopUnrollPass::run(Function &F,
1412                                       FunctionAnalysisManager &AM) {
1413   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1414   auto &LI = AM.getResult<LoopAnalysis>(F);
1415   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1416   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1417   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1418   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1419 
1420   LoopAnalysisManager *LAM = nullptr;
1421   if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(F))
1422     LAM = &LAMProxy->getManager();
1423 
1424   const ModuleAnalysisManager &MAM =
1425       AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
1426   ProfileSummaryInfo *PSI =
1427       MAM.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
1428   auto *BFI = (PSI && PSI->hasProfileSummary()) ?
1429       &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
1430 
1431   bool Changed = false;
1432 
1433   // The unroller requires loops to be in simplified form, and also needs LCSSA.
1434   // Since simplification may add new inner loops, it has to run before the
1435   // legality and profitability checks. This means running the loop unroller
1436   // will simplify all loops, regardless of whether anything end up being
1437   // unrolled.
1438   for (auto &L : LI) {
1439     Changed |=
1440         simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */);
1441     Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1442   }
1443 
1444   SmallVector<Loop *, 8> Worklist = appendLoopsToWorklist(LI);
1445 
1446   while (!Worklist.empty()) {
1447     // Because the LoopInfo stores the loops in RPO, we walk the worklist
1448     // from back to front so that we work forward across the CFG, which
1449     // for unrolling is only needed to get optimization remarks emitted in
1450     // a forward order.
1451     Loop &L = *Worklist.pop_back_val();
1452 #ifndef NDEBUG
1453     Loop *ParentL = L.getParentLoop();
1454 #endif
1455 
1456     // Check if the profile summary indicates that the profiled application
1457     // has a huge working set size, in which case we disable peeling to avoid
1458     // bloating it further.
1459     Optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling;
1460     if (PSI && PSI->hasHugeWorkingSetSize())
1461       LocalAllowPeeling = false;
1462     std::string LoopName = L.getName();
1463     // The API here is quite complex to call and we allow to select some
1464     // flavors of unrolling during construction time (by setting UnrollOpts).
1465     LoopUnrollResult Result = tryToUnrollLoop(
1466         &L, DT, &LI, SE, TTI, AC, ORE, BFI, PSI,
1467         /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, UnrollOpts.OnlyWhenForced,
1468         UnrollOpts.ForgetSCEV, /*Count*/ None,
1469         /*Threshold*/ None, UnrollOpts.AllowPartial, UnrollOpts.AllowRuntime,
1470         UnrollOpts.AllowUpperBound, LocalAllowPeeling,
1471         UnrollOpts.AllowProfileBasedPeeling, UnrollOpts.FullUnrollMaxCount);
1472     Changed |= Result != LoopUnrollResult::Unmodified;
1473 
1474     // The parent must not be damaged by unrolling!
1475 #ifndef NDEBUG
1476     if (Result != LoopUnrollResult::Unmodified && ParentL)
1477       ParentL->verifyLoop();
1478 #endif
1479 
1480     // Clear any cached analysis results for L if we removed it completely.
1481     if (LAM && Result == LoopUnrollResult::FullyUnrolled)
1482       LAM->clear(L, LoopName);
1483   }
1484 
1485   if (!Changed)
1486     return PreservedAnalyses::all();
1487 
1488   return getLoopPassPreservedAnalyses();
1489 }
1490