xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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> UnrollAllowLoopNestsPeeling(
158     "unroll-allow-loop-nests-peeling", cl::init(false), cl::Hidden,
159     cl::desc("Allows loop nests to be peeled."));
160 
161 static cl::opt<bool> UnrollUnrollRemainder(
162   "unroll-remainder", cl::Hidden,
163   cl::desc("Allow the loop remainder to be unrolled."));
164 
165 // This option isn't ever intended to be enabled, it serves to allow
166 // experiments to check the assumptions about when this kind of revisit is
167 // necessary.
168 static cl::opt<bool> UnrollRevisitChildLoops(
169     "unroll-revisit-child-loops", cl::Hidden,
170     cl::desc("Enqueue and re-visit child loops in the loop PM after unrolling. "
171              "This shouldn't typically be needed as child loops (or their "
172              "clones) were already visited."));
173 
174 static cl::opt<unsigned> UnrollThresholdAggressive(
175     "unroll-threshold-aggressive", cl::init(300), cl::Hidden,
176     cl::desc("Threshold (max size of unrolled loop) to use in aggressive (O3) "
177              "optimizations"));
178 static cl::opt<unsigned>
179     UnrollThresholdDefault("unroll-threshold-default", cl::init(150),
180                            cl::Hidden,
181                            cl::desc("Default threshold (max size of unrolled "
182                                     "loop), used in all but O3 optimizations"));
183 
184 /// A magic value for use with the Threshold parameter to indicate
185 /// that the loop unroll should be performed regardless of how much
186 /// code expansion would result.
187 static const unsigned NoThreshold = std::numeric_limits<unsigned>::max();
188 
189 /// Gather the various unrolling parameters based on the defaults, compiler
190 /// flags, TTI overrides and user specified parameters.
191 TargetTransformInfo::UnrollingPreferences llvm::gatherUnrollingPreferences(
192     Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
193     BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, int OptLevel,
194     Optional<unsigned> UserThreshold, Optional<unsigned> UserCount,
195     Optional<bool> UserAllowPartial, Optional<bool> UserRuntime,
196     Optional<bool> UserUpperBound, Optional<unsigned> UserFullUnrollMaxCount) {
197   TargetTransformInfo::UnrollingPreferences UP;
198 
199   // Set up the defaults
200   UP.Threshold =
201       OptLevel > 2 ? UnrollThresholdAggressive : UnrollThresholdDefault;
202   UP.MaxPercentThresholdBoost = 400;
203   UP.OptSizeThreshold = 0;
204   UP.PartialThreshold = 150;
205   UP.PartialOptSizeThreshold = 0;
206   UP.Count = 0;
207   UP.DefaultUnrollRuntimeCount = 8;
208   UP.MaxCount = std::numeric_limits<unsigned>::max();
209   UP.FullUnrollMaxCount = std::numeric_limits<unsigned>::max();
210   UP.BEInsns = 2;
211   UP.Partial = false;
212   UP.Runtime = false;
213   UP.AllowRemainder = true;
214   UP.UnrollRemainder = false;
215   UP.AllowExpensiveTripCount = false;
216   UP.Force = false;
217   UP.UpperBound = false;
218   UP.UnrollAndJam = false;
219   UP.UnrollAndJamInnerLoopThreshold = 60;
220   UP.MaxIterationsCountToAnalyze = UnrollMaxIterationsCountToAnalyze;
221 
222   // Override with any target specific settings
223   TTI.getUnrollingPreferences(L, SE, UP);
224 
225   // Apply size attributes
226   bool OptForSize = L->getHeader()->getParent()->hasOptSize() ||
227                     llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,
228                                                 PGSOQueryType::IRPass);
229   if (OptForSize) {
230     UP.Threshold = UP.OptSizeThreshold;
231     UP.PartialThreshold = UP.PartialOptSizeThreshold;
232     UP.MaxPercentThresholdBoost = 100;
233   }
234 
235   // Apply any user values specified by cl::opt
236   if (UnrollThreshold.getNumOccurrences() > 0)
237     UP.Threshold = UnrollThreshold;
238   if (UnrollPartialThreshold.getNumOccurrences() > 0)
239     UP.PartialThreshold = UnrollPartialThreshold;
240   if (UnrollMaxPercentThresholdBoost.getNumOccurrences() > 0)
241     UP.MaxPercentThresholdBoost = UnrollMaxPercentThresholdBoost;
242   if (UnrollMaxCount.getNumOccurrences() > 0)
243     UP.MaxCount = UnrollMaxCount;
244   if (UnrollFullMaxCount.getNumOccurrences() > 0)
245     UP.FullUnrollMaxCount = UnrollFullMaxCount;
246   if (UnrollAllowPartial.getNumOccurrences() > 0)
247     UP.Partial = UnrollAllowPartial;
248   if (UnrollAllowRemainder.getNumOccurrences() > 0)
249     UP.AllowRemainder = UnrollAllowRemainder;
250   if (UnrollRuntime.getNumOccurrences() > 0)
251     UP.Runtime = UnrollRuntime;
252   if (UnrollMaxUpperBound == 0)
253     UP.UpperBound = false;
254   if (UnrollUnrollRemainder.getNumOccurrences() > 0)
255     UP.UnrollRemainder = UnrollUnrollRemainder;
256   if (UnrollMaxIterationsCountToAnalyze.getNumOccurrences() > 0)
257     UP.MaxIterationsCountToAnalyze = UnrollMaxIterationsCountToAnalyze;
258 
259   // Apply user values provided by argument
260   if (UserThreshold.hasValue()) {
261     UP.Threshold = *UserThreshold;
262     UP.PartialThreshold = *UserThreshold;
263   }
264   if (UserCount.hasValue())
265     UP.Count = *UserCount;
266   if (UserAllowPartial.hasValue())
267     UP.Partial = *UserAllowPartial;
268   if (UserRuntime.hasValue())
269     UP.Runtime = *UserRuntime;
270   if (UserUpperBound.hasValue())
271     UP.UpperBound = *UserUpperBound;
272   if (UserFullUnrollMaxCount.hasValue())
273     UP.FullUnrollMaxCount = *UserFullUnrollMaxCount;
274 
275   return UP;
276 }
277 
278 TargetTransformInfo::PeelingPreferences
279 llvm::gatherPeelingPreferences(Loop *L, ScalarEvolution &SE,
280                                const TargetTransformInfo &TTI,
281                                Optional<bool> UserAllowPeeling,
282                                Optional<bool> UserAllowProfileBasedPeeling) {
283   TargetTransformInfo::PeelingPreferences PP;
284 
285   // Default values
286   PP.PeelCount = 0;
287   PP.AllowPeeling = true;
288   PP.AllowLoopNestsPeeling = false;
289   PP.PeelProfiledIterations = true;
290 
291   // Get Target Specifc Values
292   TTI.getPeelingPreferences(L, SE, PP);
293 
294   // User Specified Values using cl::opt
295   if (UnrollPeelCount.getNumOccurrences() > 0)
296     PP.PeelCount = UnrollPeelCount;
297   if (UnrollAllowPeeling.getNumOccurrences() > 0)
298     PP.AllowPeeling = UnrollAllowPeeling;
299   if (UnrollAllowLoopNestsPeeling.getNumOccurrences() > 0)
300     PP.AllowLoopNestsPeeling = UnrollAllowLoopNestsPeeling;
301 
302   // User Specifed values provided by argument
303   if (UserAllowPeeling.hasValue())
304     PP.AllowPeeling = *UserAllowPeeling;
305   if (UserAllowProfileBasedPeeling.hasValue())
306     PP.PeelProfiledIterations = *UserAllowProfileBasedPeeling;
307 
308   return PP;
309 }
310 
311 namespace {
312 
313 /// A struct to densely store the state of an instruction after unrolling at
314 /// each iteration.
315 ///
316 /// This is designed to work like a tuple of <Instruction *, int> for the
317 /// purposes of hashing and lookup, but to be able to associate two boolean
318 /// states with each key.
319 struct UnrolledInstState {
320   Instruction *I;
321   int Iteration : 30;
322   unsigned IsFree : 1;
323   unsigned IsCounted : 1;
324 };
325 
326 /// Hashing and equality testing for a set of the instruction states.
327 struct UnrolledInstStateKeyInfo {
328   using PtrInfo = DenseMapInfo<Instruction *>;
329   using PairInfo = DenseMapInfo<std::pair<Instruction *, int>>;
330 
331   static inline UnrolledInstState getEmptyKey() {
332     return {PtrInfo::getEmptyKey(), 0, 0, 0};
333   }
334 
335   static inline UnrolledInstState getTombstoneKey() {
336     return {PtrInfo::getTombstoneKey(), 0, 0, 0};
337   }
338 
339   static inline unsigned getHashValue(const UnrolledInstState &S) {
340     return PairInfo::getHashValue({S.I, S.Iteration});
341   }
342 
343   static inline bool isEqual(const UnrolledInstState &LHS,
344                              const UnrolledInstState &RHS) {
345     return PairInfo::isEqual({LHS.I, LHS.Iteration}, {RHS.I, RHS.Iteration});
346   }
347 };
348 
349 struct EstimatedUnrollCost {
350   /// The estimated cost after unrolling.
351   unsigned UnrolledCost;
352 
353   /// The estimated dynamic cost of executing the instructions in the
354   /// rolled form.
355   unsigned RolledDynamicCost;
356 };
357 
358 } // end anonymous namespace
359 
360 /// Figure out if the loop is worth full unrolling.
361 ///
362 /// Complete loop unrolling can make some loads constant, and we need to know
363 /// if that would expose any further optimization opportunities.  This routine
364 /// estimates this optimization.  It computes cost of unrolled loop
365 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
366 /// dynamic cost we mean that we won't count costs of blocks that are known not
367 /// to be executed (i.e. if we have a branch in the loop and we know that at the
368 /// given iteration its condition would be resolved to true, we won't add up the
369 /// cost of the 'false'-block).
370 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
371 /// the analysis failed (no benefits expected from the unrolling, or the loop is
372 /// too big to analyze), the returned value is None.
373 static Optional<EstimatedUnrollCost> analyzeLoopUnrollCost(
374     const Loop *L, unsigned TripCount, DominatorTree &DT, ScalarEvolution &SE,
375     const SmallPtrSetImpl<const Value *> &EphValues,
376     const TargetTransformInfo &TTI, unsigned MaxUnrolledLoopSize,
377     unsigned MaxIterationsCountToAnalyze) {
378   // We want to be able to scale offsets by the trip count and add more offsets
379   // to them without checking for overflows, and we already don't want to
380   // analyze *massive* trip counts, so we force the max to be reasonably small.
381   assert(MaxIterationsCountToAnalyze <
382              (unsigned)(std::numeric_limits<int>::max() / 2) &&
383          "The unroll iterations max is too large!");
384 
385   // Only analyze inner loops. We can't properly estimate cost of nested loops
386   // and we won't visit inner loops again anyway.
387   if (!L->empty())
388     return None;
389 
390   // Don't simulate loops with a big or unknown tripcount
391   if (!TripCount || TripCount > MaxIterationsCountToAnalyze)
392     return None;
393 
394   SmallSetVector<BasicBlock *, 16> BBWorklist;
395   SmallSetVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitWorklist;
396   DenseMap<Value *, Constant *> SimplifiedValues;
397   SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues;
398 
399   // The estimated cost of the unrolled form of the loop. We try to estimate
400   // this by simplifying as much as we can while computing the estimate.
401   unsigned UnrolledCost = 0;
402 
403   // We also track the estimated dynamic (that is, actually executed) cost in
404   // the rolled form. This helps identify cases when the savings from unrolling
405   // aren't just exposing dead control flows, but actual reduced dynamic
406   // instructions due to the simplifications which we expect to occur after
407   // unrolling.
408   unsigned RolledDynamicCost = 0;
409 
410   // We track the simplification of each instruction in each iteration. We use
411   // this to recursively merge costs into the unrolled cost on-demand so that
412   // we don't count the cost of any dead code. This is essentially a map from
413   // <instruction, int> to <bool, bool>, but stored as a densely packed struct.
414   DenseSet<UnrolledInstState, UnrolledInstStateKeyInfo> InstCostMap;
415 
416   // A small worklist used to accumulate cost of instructions from each
417   // observable and reached root in the loop.
418   SmallVector<Instruction *, 16> CostWorklist;
419 
420   // PHI-used worklist used between iterations while accumulating cost.
421   SmallVector<Instruction *, 4> PHIUsedList;
422 
423   // Helper function to accumulate cost for instructions in the loop.
424   auto AddCostRecursively = [&](Instruction &RootI, int Iteration) {
425     assert(Iteration >= 0 && "Cannot have a negative iteration!");
426     assert(CostWorklist.empty() && "Must start with an empty cost list");
427     assert(PHIUsedList.empty() && "Must start with an empty phi used list");
428     CostWorklist.push_back(&RootI);
429     for (;; --Iteration) {
430       do {
431         Instruction *I = CostWorklist.pop_back_val();
432 
433         // InstCostMap only uses I and Iteration as a key, the other two values
434         // don't matter here.
435         auto CostIter = InstCostMap.find({I, Iteration, 0, 0});
436         if (CostIter == InstCostMap.end())
437           // If an input to a PHI node comes from a dead path through the loop
438           // we may have no cost data for it here. What that actually means is
439           // that it is free.
440           continue;
441         auto &Cost = *CostIter;
442         if (Cost.IsCounted)
443           // Already counted this instruction.
444           continue;
445 
446         // Mark that we are counting the cost of this instruction now.
447         Cost.IsCounted = true;
448 
449         // If this is a PHI node in the loop header, just add it to the PHI set.
450         if (auto *PhiI = dyn_cast<PHINode>(I))
451           if (PhiI->getParent() == L->getHeader()) {
452             assert(Cost.IsFree && "Loop PHIs shouldn't be evaluated as they "
453                                   "inherently simplify during unrolling.");
454             if (Iteration == 0)
455               continue;
456 
457             // Push the incoming value from the backedge into the PHI used list
458             // if it is an in-loop instruction. We'll use this to populate the
459             // cost worklist for the next iteration (as we count backwards).
460             if (auto *OpI = dyn_cast<Instruction>(
461                     PhiI->getIncomingValueForBlock(L->getLoopLatch())))
462               if (L->contains(OpI))
463                 PHIUsedList.push_back(OpI);
464             continue;
465           }
466 
467         // First accumulate the cost of this instruction.
468         if (!Cost.IsFree) {
469           UnrolledCost += TTI.getUserCost(I, TargetTransformInfo::TCK_CodeSize);
470           LLVM_DEBUG(dbgs() << "Adding cost of instruction (iteration "
471                             << Iteration << "): ");
472           LLVM_DEBUG(I->dump());
473         }
474 
475         // We must count the cost of every operand which is not free,
476         // recursively. If we reach a loop PHI node, simply add it to the set
477         // to be considered on the next iteration (backwards!).
478         for (Value *Op : I->operands()) {
479           // Check whether this operand is free due to being a constant or
480           // outside the loop.
481           auto *OpI = dyn_cast<Instruction>(Op);
482           if (!OpI || !L->contains(OpI))
483             continue;
484 
485           // Otherwise accumulate its cost.
486           CostWorklist.push_back(OpI);
487         }
488       } while (!CostWorklist.empty());
489 
490       if (PHIUsedList.empty())
491         // We've exhausted the search.
492         break;
493 
494       assert(Iteration > 0 &&
495              "Cannot track PHI-used values past the first iteration!");
496       CostWorklist.append(PHIUsedList.begin(), PHIUsedList.end());
497       PHIUsedList.clear();
498     }
499   };
500 
501   // Ensure that we don't violate the loop structure invariants relied on by
502   // this analysis.
503   assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
504   assert(L->isLCSSAForm(DT) &&
505          "Must have loops in LCSSA form to track live-out values.");
506 
507   LLVM_DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
508 
509   // Simulate execution of each iteration of the loop counting instructions,
510   // which would be simplified.
511   // Since the same load will take different values on different iterations,
512   // we literally have to go through all loop's iterations.
513   for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
514     LLVM_DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
515 
516     // Prepare for the iteration by collecting any simplified entry or backedge
517     // inputs.
518     for (Instruction &I : *L->getHeader()) {
519       auto *PHI = dyn_cast<PHINode>(&I);
520       if (!PHI)
521         break;
522 
523       // The loop header PHI nodes must have exactly two input: one from the
524       // loop preheader and one from the loop latch.
525       assert(
526           PHI->getNumIncomingValues() == 2 &&
527           "Must have an incoming value only for the preheader and the latch.");
528 
529       Value *V = PHI->getIncomingValueForBlock(
530           Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
531       Constant *C = dyn_cast<Constant>(V);
532       if (Iteration != 0 && !C)
533         C = SimplifiedValues.lookup(V);
534       if (C)
535         SimplifiedInputValues.push_back({PHI, C});
536     }
537 
538     // Now clear and re-populate the map for the next iteration.
539     SimplifiedValues.clear();
540     while (!SimplifiedInputValues.empty())
541       SimplifiedValues.insert(SimplifiedInputValues.pop_back_val());
542 
543     UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L);
544 
545     BBWorklist.clear();
546     BBWorklist.insert(L->getHeader());
547     // Note that we *must not* cache the size, this loop grows the worklist.
548     for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
549       BasicBlock *BB = BBWorklist[Idx];
550 
551       // Visit all instructions in the given basic block and try to simplify
552       // it.  We don't change the actual IR, just count optimization
553       // opportunities.
554       for (Instruction &I : *BB) {
555         // These won't get into the final code - don't even try calculating the
556         // cost for them.
557         if (isa<DbgInfoIntrinsic>(I) || EphValues.count(&I))
558           continue;
559 
560         // Track this instruction's expected baseline cost when executing the
561         // rolled loop form.
562         RolledDynamicCost += TTI.getUserCost(&I, TargetTransformInfo::TCK_CodeSize);
563 
564         // Visit the instruction to analyze its loop cost after unrolling,
565         // and if the visitor returns true, mark the instruction as free after
566         // unrolling and continue.
567         bool IsFree = Analyzer.visit(I);
568         bool Inserted = InstCostMap.insert({&I, (int)Iteration,
569                                            (unsigned)IsFree,
570                                            /*IsCounted*/ false}).second;
571         (void)Inserted;
572         assert(Inserted && "Cannot have a state for an unvisited instruction!");
573 
574         if (IsFree)
575           continue;
576 
577         // Can't properly model a cost of a call.
578         // FIXME: With a proper cost model we should be able to do it.
579         if (auto *CI = dyn_cast<CallInst>(&I)) {
580           const Function *Callee = CI->getCalledFunction();
581           if (!Callee || TTI.isLoweredToCall(Callee)) {
582             LLVM_DEBUG(dbgs() << "Can't analyze cost of loop with call\n");
583             return None;
584           }
585         }
586 
587         // If the instruction might have a side-effect recursively account for
588         // the cost of it and all the instructions leading up to it.
589         if (I.mayHaveSideEffects())
590           AddCostRecursively(I, Iteration);
591 
592         // If unrolled body turns out to be too big, bail out.
593         if (UnrolledCost > MaxUnrolledLoopSize) {
594           LLVM_DEBUG(dbgs() << "  Exceeded threshold.. exiting.\n"
595                             << "  UnrolledCost: " << UnrolledCost
596                             << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
597                             << "\n");
598           return None;
599         }
600       }
601 
602       Instruction *TI = BB->getTerminator();
603 
604       // Add in the live successors by first checking whether we have terminator
605       // that may be simplified based on the values simplified by this call.
606       BasicBlock *KnownSucc = nullptr;
607       if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
608         if (BI->isConditional()) {
609           if (Constant *SimpleCond =
610                   SimplifiedValues.lookup(BI->getCondition())) {
611             // Just take the first successor if condition is undef
612             if (isa<UndefValue>(SimpleCond))
613               KnownSucc = BI->getSuccessor(0);
614             else if (ConstantInt *SimpleCondVal =
615                          dyn_cast<ConstantInt>(SimpleCond))
616               KnownSucc = BI->getSuccessor(SimpleCondVal->isZero() ? 1 : 0);
617           }
618         }
619       } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
620         if (Constant *SimpleCond =
621                 SimplifiedValues.lookup(SI->getCondition())) {
622           // Just take the first successor if condition is undef
623           if (isa<UndefValue>(SimpleCond))
624             KnownSucc = SI->getSuccessor(0);
625           else if (ConstantInt *SimpleCondVal =
626                        dyn_cast<ConstantInt>(SimpleCond))
627             KnownSucc = SI->findCaseValue(SimpleCondVal)->getCaseSuccessor();
628         }
629       }
630       if (KnownSucc) {
631         if (L->contains(KnownSucc))
632           BBWorklist.insert(KnownSucc);
633         else
634           ExitWorklist.insert({BB, KnownSucc});
635         continue;
636       }
637 
638       // Add BB's successors to the worklist.
639       for (BasicBlock *Succ : successors(BB))
640         if (L->contains(Succ))
641           BBWorklist.insert(Succ);
642         else
643           ExitWorklist.insert({BB, Succ});
644       AddCostRecursively(*TI, Iteration);
645     }
646 
647     // If we found no optimization opportunities on the first iteration, we
648     // won't find them on later ones too.
649     if (UnrolledCost == RolledDynamicCost) {
650       LLVM_DEBUG(dbgs() << "  No opportunities found.. exiting.\n"
651                         << "  UnrolledCost: " << UnrolledCost << "\n");
652       return None;
653     }
654   }
655 
656   while (!ExitWorklist.empty()) {
657     BasicBlock *ExitingBB, *ExitBB;
658     std::tie(ExitingBB, ExitBB) = ExitWorklist.pop_back_val();
659 
660     for (Instruction &I : *ExitBB) {
661       auto *PN = dyn_cast<PHINode>(&I);
662       if (!PN)
663         break;
664 
665       Value *Op = PN->getIncomingValueForBlock(ExitingBB);
666       if (auto *OpI = dyn_cast<Instruction>(Op))
667         if (L->contains(OpI))
668           AddCostRecursively(*OpI, TripCount - 1);
669     }
670   }
671 
672   LLVM_DEBUG(dbgs() << "Analysis finished:\n"
673                     << "UnrolledCost: " << UnrolledCost << ", "
674                     << "RolledDynamicCost: " << RolledDynamicCost << "\n");
675   return {{UnrolledCost, RolledDynamicCost}};
676 }
677 
678 /// ApproximateLoopSize - Approximate the size of the loop.
679 unsigned llvm::ApproximateLoopSize(
680     const Loop *L, unsigned &NumCalls, bool &NotDuplicatable, bool &Convergent,
681     const TargetTransformInfo &TTI,
682     const SmallPtrSetImpl<const Value *> &EphValues, unsigned BEInsns) {
683   CodeMetrics Metrics;
684   for (BasicBlock *BB : L->blocks())
685     Metrics.analyzeBasicBlock(BB, TTI, EphValues);
686   NumCalls = Metrics.NumInlineCandidates;
687   NotDuplicatable = Metrics.notDuplicatable;
688   Convergent = Metrics.convergent;
689 
690   unsigned LoopSize = Metrics.NumInsts;
691 
692   // Don't allow an estimate of size zero.  This would allows unrolling of loops
693   // with huge iteration counts, which is a compile time problem even if it's
694   // not a problem for code quality. Also, the code using this size may assume
695   // that each loop has at least three instructions (likely a conditional
696   // branch, a comparison feeding that branch, and some kind of loop increment
697   // feeding that comparison instruction).
698   LoopSize = std::max(LoopSize, BEInsns + 1);
699 
700   return LoopSize;
701 }
702 
703 // Returns the loop hint metadata node with the given name (for example,
704 // "llvm.loop.unroll.count").  If no such metadata node exists, then nullptr is
705 // returned.
706 static MDNode *getUnrollMetadataForLoop(const Loop *L, StringRef Name) {
707   if (MDNode *LoopID = L->getLoopID())
708     return GetUnrollMetadata(LoopID, Name);
709   return nullptr;
710 }
711 
712 // Returns true if the loop has an unroll(full) pragma.
713 static bool hasUnrollFullPragma(const Loop *L) {
714   return getUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
715 }
716 
717 // Returns true if the loop has an unroll(enable) pragma. This metadata is used
718 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
719 static bool hasUnrollEnablePragma(const Loop *L) {
720   return getUnrollMetadataForLoop(L, "llvm.loop.unroll.enable");
721 }
722 
723 // Returns true if the loop has an runtime unroll(disable) pragma.
724 static bool hasRuntimeUnrollDisablePragma(const Loop *L) {
725   return getUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
726 }
727 
728 // If loop has an unroll_count pragma return the (necessarily
729 // positive) value from the pragma.  Otherwise return 0.
730 static unsigned unrollCountPragmaValue(const Loop *L) {
731   MDNode *MD = getUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
732   if (MD) {
733     assert(MD->getNumOperands() == 2 &&
734            "Unroll count hint metadata should have two operands.");
735     unsigned Count =
736         mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
737     assert(Count >= 1 && "Unroll count must be positive.");
738     return Count;
739   }
740   return 0;
741 }
742 
743 // Computes the boosting factor for complete unrolling.
744 // If fully unrolling the loop would save a lot of RolledDynamicCost, it would
745 // be beneficial to fully unroll the loop even if unrolledcost is large. We
746 // use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust
747 // the unroll threshold.
748 static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost,
749                                             unsigned MaxPercentThresholdBoost) {
750   if (Cost.RolledDynamicCost >= std::numeric_limits<unsigned>::max() / 100)
751     return 100;
752   else if (Cost.UnrolledCost != 0)
753     // The boosting factor is RolledDynamicCost / UnrolledCost
754     return std::min(100 * Cost.RolledDynamicCost / Cost.UnrolledCost,
755                     MaxPercentThresholdBoost);
756   else
757     return MaxPercentThresholdBoost;
758 }
759 
760 // Returns loop size estimation for unrolled loop.
761 static uint64_t getUnrolledLoopSize(
762     unsigned LoopSize,
763     TargetTransformInfo::UnrollingPreferences &UP) {
764   assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
765   return (uint64_t)(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns;
766 }
767 
768 // Returns true if unroll count was set explicitly.
769 // Calculates unroll count and writes it to UP.Count.
770 // Unless IgnoreUser is true, will also use metadata and command-line options
771 // that are specific to to the LoopUnroll pass (which, for instance, are
772 // irrelevant for the LoopUnrollAndJam pass).
773 // FIXME: This function is used by LoopUnroll and LoopUnrollAndJam, but consumes
774 // many LoopUnroll-specific options. The shared functionality should be
775 // refactored into it own function.
776 bool llvm::computeUnrollCount(
777     Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI,
778     ScalarEvolution &SE, const SmallPtrSetImpl<const Value *> &EphValues,
779     OptimizationRemarkEmitter *ORE, unsigned &TripCount, unsigned MaxTripCount,
780     bool MaxOrZero, unsigned &TripMultiple, unsigned LoopSize,
781     TargetTransformInfo::UnrollingPreferences &UP,
782     TargetTransformInfo::PeelingPreferences &PP, bool &UseUpperBound) {
783 
784   // Check for explicit Count.
785   // 1st priority is unroll count set by "unroll-count" option.
786   bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0;
787   if (UserUnrollCount) {
788     UP.Count = UnrollCount;
789     UP.AllowExpensiveTripCount = true;
790     UP.Force = true;
791     if (UP.AllowRemainder && getUnrolledLoopSize(LoopSize, UP) < UP.Threshold)
792       return true;
793   }
794 
795   // 2nd priority is unroll count set by pragma.
796   unsigned PragmaCount = unrollCountPragmaValue(L);
797   if (PragmaCount > 0) {
798     UP.Count = PragmaCount;
799     UP.Runtime = true;
800     UP.AllowExpensiveTripCount = true;
801     UP.Force = true;
802     if ((UP.AllowRemainder || (TripMultiple % PragmaCount == 0)) &&
803         getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
804       return true;
805   }
806   bool PragmaFullUnroll = hasUnrollFullPragma(L);
807   if (PragmaFullUnroll && TripCount != 0) {
808     UP.Count = TripCount;
809     if (getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
810       return false;
811   }
812 
813   bool PragmaEnableUnroll = hasUnrollEnablePragma(L);
814   bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll ||
815                         PragmaEnableUnroll || UserUnrollCount;
816 
817   if (ExplicitUnroll && TripCount != 0) {
818     // If the loop has an unrolling pragma, we want to be more aggressive with
819     // unrolling limits. Set thresholds to at least the PragmaUnrollThreshold
820     // value which is larger than the default limits.
821     UP.Threshold = std::max<unsigned>(UP.Threshold, PragmaUnrollThreshold);
822     UP.PartialThreshold =
823         std::max<unsigned>(UP.PartialThreshold, PragmaUnrollThreshold);
824   }
825 
826   // 3rd priority is full unroll count.
827   // Full unroll makes sense only when TripCount or its upper bound could be
828   // statically calculated.
829   // Also we need to check if we exceed FullUnrollMaxCount.
830   // If using the upper bound to unroll, TripMultiple should be set to 1 because
831   // we do not know when loop may exit.
832 
833   // We can unroll by the upper bound amount if it's generally allowed or if
834   // we know that the loop is executed either the upper bound or zero times.
835   // (MaxOrZero unrolling keeps only the first loop test, so the number of
836   // loop tests remains the same compared to the non-unrolled version, whereas
837   // the generic upper bound unrolling keeps all but the last loop test so the
838   // number of loop tests goes up which may end up being worse on targets with
839   // constrained branch predictor resources so is controlled by an option.)
840   // In addition we only unroll small upper bounds.
841   unsigned FullUnrollMaxTripCount = MaxTripCount;
842   if (!(UP.UpperBound || MaxOrZero) ||
843       FullUnrollMaxTripCount > UnrollMaxUpperBound)
844     FullUnrollMaxTripCount = 0;
845 
846   // UnrollByMaxCount and ExactTripCount cannot both be non zero since we only
847   // compute the former when the latter is zero.
848   unsigned ExactTripCount = TripCount;
849   assert((ExactTripCount == 0 || FullUnrollMaxTripCount == 0) &&
850          "ExtractTripCount and UnrollByMaxCount cannot both be non zero.");
851 
852   unsigned FullUnrollTripCount =
853       ExactTripCount ? ExactTripCount : FullUnrollMaxTripCount;
854   UP.Count = FullUnrollTripCount;
855   if (FullUnrollTripCount && FullUnrollTripCount <= UP.FullUnrollMaxCount) {
856     // When computing the unrolled size, note that BEInsns are not replicated
857     // like the rest of the loop body.
858     if (getUnrolledLoopSize(LoopSize, UP) < UP.Threshold) {
859       UseUpperBound = (FullUnrollMaxTripCount == FullUnrollTripCount);
860       TripCount = FullUnrollTripCount;
861       TripMultiple = UP.UpperBound ? 1 : TripMultiple;
862       return ExplicitUnroll;
863     } else {
864       // The loop isn't that small, but we still can fully unroll it if that
865       // helps to remove a significant number of instructions.
866       // To check that, run additional analysis on the loop.
867       if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
868               L, FullUnrollTripCount, DT, SE, EphValues, TTI,
869               UP.Threshold * UP.MaxPercentThresholdBoost / 100,
870               UP.MaxIterationsCountToAnalyze)) {
871         unsigned Boost =
872             getFullUnrollBoostingFactor(*Cost, UP.MaxPercentThresholdBoost);
873         if (Cost->UnrolledCost < UP.Threshold * Boost / 100) {
874           UseUpperBound = (FullUnrollMaxTripCount == FullUnrollTripCount);
875           TripCount = FullUnrollTripCount;
876           TripMultiple = UP.UpperBound ? 1 : TripMultiple;
877           return ExplicitUnroll;
878         }
879       }
880     }
881   }
882 
883   // 4th priority is loop peeling.
884   computePeelCount(L, LoopSize, UP, PP, TripCount, SE);
885   if (PP.PeelCount) {
886     UP.Runtime = false;
887     UP.Count = 1;
888     return ExplicitUnroll;
889   }
890 
891   // 5th priority is partial unrolling.
892   // Try partial unroll only when TripCount could be statically calculated.
893   if (TripCount) {
894     UP.Partial |= ExplicitUnroll;
895     if (!UP.Partial) {
896       LLVM_DEBUG(dbgs() << "  will not try to unroll partially because "
897                         << "-unroll-allow-partial not given\n");
898       UP.Count = 0;
899       return false;
900     }
901     if (UP.Count == 0)
902       UP.Count = TripCount;
903     if (UP.PartialThreshold != NoThreshold) {
904       // Reduce unroll count to be modulo of TripCount for partial unrolling.
905       if (getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
906         UP.Count =
907             (std::max(UP.PartialThreshold, UP.BEInsns + 1) - UP.BEInsns) /
908             (LoopSize - UP.BEInsns);
909       if (UP.Count > UP.MaxCount)
910         UP.Count = UP.MaxCount;
911       while (UP.Count != 0 && TripCount % UP.Count != 0)
912         UP.Count--;
913       if (UP.AllowRemainder && UP.Count <= 1) {
914         // If there is no Count that is modulo of TripCount, set Count to
915         // largest power-of-two factor that satisfies the threshold limit.
916         // As we'll create fixup loop, do the type of unrolling only if
917         // remainder loop is allowed.
918         UP.Count = UP.DefaultUnrollRuntimeCount;
919         while (UP.Count != 0 &&
920                getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
921           UP.Count >>= 1;
922       }
923       if (UP.Count < 2) {
924         if (PragmaEnableUnroll)
925           ORE->emit([&]() {
926             return OptimizationRemarkMissed(DEBUG_TYPE,
927                                             "UnrollAsDirectedTooLarge",
928                                             L->getStartLoc(), L->getHeader())
929                    << "Unable to unroll loop as directed by unroll(enable) "
930                       "pragma "
931                       "because unrolled size is too large.";
932           });
933         UP.Count = 0;
934       }
935     } else {
936       UP.Count = TripCount;
937     }
938     if (UP.Count > UP.MaxCount)
939       UP.Count = UP.MaxCount;
940     if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
941         UP.Count != TripCount)
942       ORE->emit([&]() {
943         return OptimizationRemarkMissed(DEBUG_TYPE,
944                                         "FullUnrollAsDirectedTooLarge",
945                                         L->getStartLoc(), L->getHeader())
946                << "Unable to fully unroll loop as directed by unroll pragma "
947                   "because "
948                   "unrolled size is too large.";
949       });
950     LLVM_DEBUG(dbgs() << "  partially unrolling with count: " << UP.Count
951                       << "\n");
952     return ExplicitUnroll;
953   }
954   assert(TripCount == 0 &&
955          "All cases when TripCount is constant should be covered here.");
956   if (PragmaFullUnroll)
957     ORE->emit([&]() {
958       return OptimizationRemarkMissed(
959                  DEBUG_TYPE, "CantFullUnrollAsDirectedRuntimeTripCount",
960                  L->getStartLoc(), L->getHeader())
961              << "Unable to fully unroll loop as directed by unroll(full) "
962                 "pragma "
963                 "because loop has a runtime trip count.";
964     });
965 
966   // 6th priority is runtime unrolling.
967   // Don't unroll a runtime trip count loop when it is disabled.
968   if (hasRuntimeUnrollDisablePragma(L)) {
969     UP.Count = 0;
970     return false;
971   }
972 
973   // Don't unroll a small upper bound loop unless user or TTI asked to do so.
974   if (MaxTripCount && !UP.Force && MaxTripCount < UnrollMaxUpperBound) {
975     UP.Count = 0;
976     return false;
977   }
978 
979   // Check if the runtime trip count is too small when profile is available.
980   if (L->getHeader()->getParent()->hasProfileData()) {
981     if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) {
982       if (*ProfileTripCount < FlatLoopTripCountThreshold)
983         return false;
984       else
985         UP.AllowExpensiveTripCount = true;
986     }
987   }
988 
989   // Reduce count based on the type of unrolling and the threshold values.
990   UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount;
991   if (!UP.Runtime) {
992     LLVM_DEBUG(
993         dbgs() << "  will not try to unroll loop with runtime trip count "
994                << "-unroll-runtime not given\n");
995     UP.Count = 0;
996     return false;
997   }
998   if (UP.Count == 0)
999     UP.Count = UP.DefaultUnrollRuntimeCount;
1000 
1001   // Reduce unroll count to be the largest power-of-two factor of
1002   // the original count which satisfies the threshold limit.
1003   while (UP.Count != 0 &&
1004          getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
1005     UP.Count >>= 1;
1006 
1007 #ifndef NDEBUG
1008   unsigned OrigCount = UP.Count;
1009 #endif
1010 
1011   if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) {
1012     while (UP.Count != 0 && TripMultiple % UP.Count != 0)
1013       UP.Count >>= 1;
1014     LLVM_DEBUG(
1015         dbgs() << "Remainder loop is restricted (that could architecture "
1016                   "specific or because the loop contains a convergent "
1017                   "instruction), so unroll count must divide the trip "
1018                   "multiple, "
1019                << TripMultiple << ".  Reducing unroll count from " << OrigCount
1020                << " to " << UP.Count << ".\n");
1021 
1022     using namespace ore;
1023 
1024     if (PragmaCount > 0 && !UP.AllowRemainder)
1025       ORE->emit([&]() {
1026         return OptimizationRemarkMissed(DEBUG_TYPE,
1027                                         "DifferentUnrollCountFromDirected",
1028                                         L->getStartLoc(), L->getHeader())
1029                << "Unable to unroll loop the number of times directed by "
1030                   "unroll_count pragma because remainder loop is restricted "
1031                   "(that could architecture specific or because the loop "
1032                   "contains a convergent instruction) and so must have an "
1033                   "unroll "
1034                   "count that divides the loop trip multiple of "
1035                << NV("TripMultiple", TripMultiple) << ".  Unrolling instead "
1036                << NV("UnrollCount", UP.Count) << " time(s).";
1037       });
1038   }
1039 
1040   if (UP.Count > UP.MaxCount)
1041     UP.Count = UP.MaxCount;
1042 
1043   if (MaxTripCount && UP.Count > MaxTripCount)
1044     UP.Count = MaxTripCount;
1045 
1046   LLVM_DEBUG(dbgs() << "  runtime unrolling with count: " << UP.Count
1047                     << "\n");
1048   if (UP.Count < 2)
1049     UP.Count = 0;
1050   return ExplicitUnroll;
1051 }
1052 
1053 static LoopUnrollResult tryToUnrollLoop(
1054     Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE,
1055     const TargetTransformInfo &TTI, AssumptionCache &AC,
1056     OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
1057     ProfileSummaryInfo *PSI, bool PreserveLCSSA, int OptLevel,
1058     bool OnlyWhenForced, bool ForgetAllSCEV, Optional<unsigned> ProvidedCount,
1059     Optional<unsigned> ProvidedThreshold, Optional<bool> ProvidedAllowPartial,
1060     Optional<bool> ProvidedRuntime, Optional<bool> ProvidedUpperBound,
1061     Optional<bool> ProvidedAllowPeeling,
1062     Optional<bool> ProvidedAllowProfileBasedPeeling,
1063     Optional<unsigned> ProvidedFullUnrollMaxCount) {
1064   LLVM_DEBUG(dbgs() << "Loop Unroll: F["
1065                     << L->getHeader()->getParent()->getName() << "] Loop %"
1066                     << L->getHeader()->getName() << "\n");
1067   TransformationMode TM = hasUnrollTransformation(L);
1068   if (TM & TM_Disable)
1069     return LoopUnrollResult::Unmodified;
1070   if (!L->isLoopSimplifyForm()) {
1071     LLVM_DEBUG(
1072         dbgs() << "  Not unrolling loop which is not in loop-simplify form.\n");
1073     return LoopUnrollResult::Unmodified;
1074   }
1075 
1076   // When automtatic unrolling is disabled, do not unroll unless overridden for
1077   // this loop.
1078   if (OnlyWhenForced && !(TM & TM_Enable))
1079     return LoopUnrollResult::Unmodified;
1080 
1081   bool OptForSize = L->getHeader()->getParent()->hasOptSize();
1082   unsigned NumInlineCandidates;
1083   bool NotDuplicatable;
1084   bool Convergent;
1085   TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
1086       L, SE, TTI, BFI, PSI, OptLevel, ProvidedThreshold, ProvidedCount,
1087       ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound,
1088       ProvidedFullUnrollMaxCount);
1089   TargetTransformInfo::PeelingPreferences PP = gatherPeelingPreferences(
1090       L, SE, TTI, ProvidedAllowPeeling, ProvidedAllowProfileBasedPeeling);
1091 
1092   // Exit early if unrolling is disabled. For OptForSize, we pick the loop size
1093   // as threshold later on.
1094   if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0) &&
1095       !OptForSize)
1096     return LoopUnrollResult::Unmodified;
1097 
1098   SmallPtrSet<const Value *, 32> EphValues;
1099   CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
1100 
1101   unsigned LoopSize =
1102       ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent,
1103                           TTI, EphValues, UP.BEInsns);
1104   LLVM_DEBUG(dbgs() << "  Loop Size = " << LoopSize << "\n");
1105   if (NotDuplicatable) {
1106     LLVM_DEBUG(dbgs() << "  Not unrolling loop which contains non-duplicatable"
1107                       << " instructions.\n");
1108     return LoopUnrollResult::Unmodified;
1109   }
1110 
1111   // When optimizing for size, use LoopSize + 1 as threshold (we use < Threshold
1112   // later), to (fully) unroll loops, if it does not increase code size.
1113   if (OptForSize)
1114     UP.Threshold = std::max(UP.Threshold, LoopSize + 1);
1115 
1116   if (NumInlineCandidates != 0) {
1117     LLVM_DEBUG(dbgs() << "  Not unrolling loop with inlinable calls.\n");
1118     return LoopUnrollResult::Unmodified;
1119   }
1120 
1121   // Find trip count and trip multiple if count is not available
1122   unsigned TripCount = 0;
1123   unsigned TripMultiple = 1;
1124   // If there are multiple exiting blocks but one of them is the latch, use the
1125   // latch for the trip count estimation. Otherwise insist on a single exiting
1126   // block for the trip count estimation.
1127   BasicBlock *ExitingBlock = L->getLoopLatch();
1128   if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
1129     ExitingBlock = L->getExitingBlock();
1130   if (ExitingBlock) {
1131     TripCount = SE.getSmallConstantTripCount(L, ExitingBlock);
1132     TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock);
1133   }
1134 
1135   // If the loop contains a convergent operation, the prelude we'd add
1136   // to do the first few instructions before we hit the unrolled loop
1137   // is unsafe -- it adds a control-flow dependency to the convergent
1138   // operation.  Therefore restrict remainder loop (try unrollig without).
1139   //
1140   // TODO: This is quite conservative.  In practice, convergent_op()
1141   // is likely to be called unconditionally in the loop.  In this
1142   // case, the program would be ill-formed (on most architectures)
1143   // unless n were the same on all threads in a thread group.
1144   // Assuming n is the same on all threads, any kind of unrolling is
1145   // safe.  But currently llvm's notion of convergence isn't powerful
1146   // enough to express this.
1147   if (Convergent)
1148     UP.AllowRemainder = false;
1149 
1150   // Try to find the trip count upper bound if we cannot find the exact trip
1151   // count.
1152   unsigned MaxTripCount = 0;
1153   bool MaxOrZero = false;
1154   if (!TripCount) {
1155     MaxTripCount = SE.getSmallConstantMaxTripCount(L);
1156     MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L);
1157   }
1158 
1159   // computeUnrollCount() decides whether it is beneficial to use upper bound to
1160   // fully unroll the loop.
1161   bool UseUpperBound = false;
1162   bool IsCountSetExplicitly = computeUnrollCount(
1163       L, TTI, DT, LI, SE, EphValues, &ORE, TripCount, MaxTripCount, MaxOrZero,
1164       TripMultiple, LoopSize, UP, PP, UseUpperBound);
1165   if (!UP.Count)
1166     return LoopUnrollResult::Unmodified;
1167   // Unroll factor (Count) must be less or equal to TripCount.
1168   if (TripCount && UP.Count > TripCount)
1169     UP.Count = TripCount;
1170 
1171   // Save loop properties before it is transformed.
1172   MDNode *OrigLoopID = L->getLoopID();
1173 
1174   // Unroll the loop.
1175   Loop *RemainderLoop = nullptr;
1176   LoopUnrollResult UnrollResult = UnrollLoop(
1177       L,
1178       {UP.Count, TripCount, UP.Force, UP.Runtime, UP.AllowExpensiveTripCount,
1179        UseUpperBound, MaxOrZero, TripMultiple, PP.PeelCount, UP.UnrollRemainder,
1180        ForgetAllSCEV},
1181       LI, &SE, &DT, &AC, &TTI, &ORE, PreserveLCSSA, &RemainderLoop);
1182   if (UnrollResult == LoopUnrollResult::Unmodified)
1183     return LoopUnrollResult::Unmodified;
1184 
1185   if (RemainderLoop) {
1186     Optional<MDNode *> RemainderLoopID =
1187         makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1188                                         LLVMLoopUnrollFollowupRemainder});
1189     if (RemainderLoopID.hasValue())
1190       RemainderLoop->setLoopID(RemainderLoopID.getValue());
1191   }
1192 
1193   if (UnrollResult != LoopUnrollResult::FullyUnrolled) {
1194     Optional<MDNode *> NewLoopID =
1195         makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1196                                         LLVMLoopUnrollFollowupUnrolled});
1197     if (NewLoopID.hasValue()) {
1198       L->setLoopID(NewLoopID.getValue());
1199 
1200       // Do not setLoopAlreadyUnrolled if loop attributes have been specified
1201       // explicitly.
1202       return UnrollResult;
1203     }
1204   }
1205 
1206   // If loop has an unroll count pragma or unrolled by explicitly set count
1207   // mark loop as unrolled to prevent unrolling beyond that requested.
1208   // If the loop was peeled, we already "used up" the profile information
1209   // we had, so we don't want to unroll or peel again.
1210   if (UnrollResult != LoopUnrollResult::FullyUnrolled &&
1211       (IsCountSetExplicitly || (PP.PeelProfiledIterations && PP.PeelCount)))
1212     L->setLoopAlreadyUnrolled();
1213 
1214   return UnrollResult;
1215 }
1216 
1217 namespace {
1218 
1219 class LoopUnroll : public LoopPass {
1220 public:
1221   static char ID; // Pass ID, replacement for typeid
1222 
1223   int OptLevel;
1224 
1225   /// If false, use a cost model to determine whether unrolling of a loop is
1226   /// profitable. If true, only loops that explicitly request unrolling via
1227   /// metadata are considered. All other loops are skipped.
1228   bool OnlyWhenForced;
1229 
1230   /// If false, when SCEV is invalidated, only forget everything in the
1231   /// top-most loop (call forgetTopMostLoop), of the loop being processed.
1232   /// Otherwise, forgetAllLoops and rebuild when needed next.
1233   bool ForgetAllSCEV;
1234 
1235   Optional<unsigned> ProvidedCount;
1236   Optional<unsigned> ProvidedThreshold;
1237   Optional<bool> ProvidedAllowPartial;
1238   Optional<bool> ProvidedRuntime;
1239   Optional<bool> ProvidedUpperBound;
1240   Optional<bool> ProvidedAllowPeeling;
1241   Optional<bool> ProvidedAllowProfileBasedPeeling;
1242   Optional<unsigned> ProvidedFullUnrollMaxCount;
1243 
1244   LoopUnroll(int OptLevel = 2, bool OnlyWhenForced = false,
1245              bool ForgetAllSCEV = false, Optional<unsigned> Threshold = None,
1246              Optional<unsigned> Count = None,
1247              Optional<bool> AllowPartial = None, Optional<bool> Runtime = None,
1248              Optional<bool> UpperBound = None,
1249              Optional<bool> AllowPeeling = None,
1250              Optional<bool> AllowProfileBasedPeeling = None,
1251              Optional<unsigned> ProvidedFullUnrollMaxCount = None)
1252       : LoopPass(ID), OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
1253         ForgetAllSCEV(ForgetAllSCEV), ProvidedCount(std::move(Count)),
1254         ProvidedThreshold(Threshold), ProvidedAllowPartial(AllowPartial),
1255         ProvidedRuntime(Runtime), ProvidedUpperBound(UpperBound),
1256         ProvidedAllowPeeling(AllowPeeling),
1257         ProvidedAllowProfileBasedPeeling(AllowProfileBasedPeeling),
1258         ProvidedFullUnrollMaxCount(ProvidedFullUnrollMaxCount) {
1259     initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
1260   }
1261 
1262   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1263     if (skipLoop(L))
1264       return false;
1265 
1266     Function &F = *L->getHeader()->getParent();
1267 
1268     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1269     LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1270     ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1271     const TargetTransformInfo &TTI =
1272         getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1273     auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1274     // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
1275     // pass.  Function analyses need to be preserved across loop transformations
1276     // but ORE cannot be preserved (see comment before the pass definition).
1277     OptimizationRemarkEmitter ORE(&F);
1278     bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1279 
1280     LoopUnrollResult Result = tryToUnrollLoop(
1281         L, DT, LI, SE, TTI, AC, ORE, nullptr, nullptr, PreserveLCSSA, OptLevel,
1282         OnlyWhenForced, ForgetAllSCEV, ProvidedCount, ProvidedThreshold,
1283         ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound,
1284         ProvidedAllowPeeling, ProvidedAllowProfileBasedPeeling,
1285         ProvidedFullUnrollMaxCount);
1286 
1287     if (Result == LoopUnrollResult::FullyUnrolled)
1288       LPM.markLoopAsDeleted(*L);
1289 
1290     return Result != LoopUnrollResult::Unmodified;
1291   }
1292 
1293   /// This transformation requires natural loop information & requires that
1294   /// loop preheaders be inserted into the CFG...
1295   void getAnalysisUsage(AnalysisUsage &AU) const override {
1296     AU.addRequired<AssumptionCacheTracker>();
1297     AU.addRequired<TargetTransformInfoWrapperPass>();
1298     // FIXME: Loop passes are required to preserve domtree, and for now we just
1299     // recreate dom info if anything gets unrolled.
1300     getLoopAnalysisUsage(AU);
1301   }
1302 };
1303 
1304 } // end anonymous namespace
1305 
1306 char LoopUnroll::ID = 0;
1307 
1308 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1309 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1310 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1311 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1312 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1313 
1314 Pass *llvm::createLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1315                                  bool ForgetAllSCEV, int Threshold, int Count,
1316                                  int AllowPartial, int Runtime, int UpperBound,
1317                                  int AllowPeeling) {
1318   // TODO: It would make more sense for this function to take the optionals
1319   // directly, but that's dangerous since it would silently break out of tree
1320   // callers.
1321   return new LoopUnroll(
1322       OptLevel, OnlyWhenForced, ForgetAllSCEV,
1323       Threshold == -1 ? None : Optional<unsigned>(Threshold),
1324       Count == -1 ? None : Optional<unsigned>(Count),
1325       AllowPartial == -1 ? None : Optional<bool>(AllowPartial),
1326       Runtime == -1 ? None : Optional<bool>(Runtime),
1327       UpperBound == -1 ? None : Optional<bool>(UpperBound),
1328       AllowPeeling == -1 ? None : Optional<bool>(AllowPeeling));
1329 }
1330 
1331 Pass *llvm::createSimpleLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1332                                        bool ForgetAllSCEV) {
1333   return createLoopUnrollPass(OptLevel, OnlyWhenForced, ForgetAllSCEV, -1, -1,
1334                               0, 0, 0, 0);
1335 }
1336 
1337 PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
1338                                           LoopStandardAnalysisResults &AR,
1339                                           LPMUpdater &Updater) {
1340   // For the new PM, we can't use OptimizationRemarkEmitter as an analysis
1341   // pass. Function analyses need to be preserved across loop transformations
1342   // but ORE cannot be preserved (see comment before the pass definition).
1343   OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
1344 
1345   // Keep track of the previous loop structure so we can identify new loops
1346   // created by unrolling.
1347   Loop *ParentL = L.getParentLoop();
1348   SmallPtrSet<Loop *, 4> OldLoops;
1349   if (ParentL)
1350     OldLoops.insert(ParentL->begin(), ParentL->end());
1351   else
1352     OldLoops.insert(AR.LI.begin(), AR.LI.end());
1353 
1354   std::string LoopName = std::string(L.getName());
1355 
1356   bool Changed = tryToUnrollLoop(&L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, ORE,
1357                                  /*BFI*/ nullptr, /*PSI*/ nullptr,
1358                                  /*PreserveLCSSA*/ true, OptLevel,
1359                                  OnlyWhenForced, ForgetSCEV, /*Count*/ None,
1360                                  /*Threshold*/ None, /*AllowPartial*/ false,
1361                                  /*Runtime*/ false, /*UpperBound*/ false,
1362                                  /*AllowPeeling*/ false,
1363                                  /*AllowProfileBasedPeeling*/ false,
1364                                  /*FullUnrollMaxCount*/ None) !=
1365                  LoopUnrollResult::Unmodified;
1366   if (!Changed)
1367     return PreservedAnalyses::all();
1368 
1369   // The parent must not be damaged by unrolling!
1370 #ifndef NDEBUG
1371   if (ParentL)
1372     ParentL->verifyLoop();
1373 #endif
1374 
1375   // Unrolling can do several things to introduce new loops into a loop nest:
1376   // - Full unrolling clones child loops within the current loop but then
1377   //   removes the current loop making all of the children appear to be new
1378   //   sibling loops.
1379   //
1380   // When a new loop appears as a sibling loop after fully unrolling,
1381   // its nesting structure has fundamentally changed and we want to revisit
1382   // it to reflect that.
1383   //
1384   // When unrolling has removed the current loop, we need to tell the
1385   // infrastructure that it is gone.
1386   //
1387   // Finally, we support a debugging/testing mode where we revisit child loops
1388   // as well. These are not expected to require further optimizations as either
1389   // they or the loop they were cloned from have been directly visited already.
1390   // But the debugging mode allows us to check this assumption.
1391   bool IsCurrentLoopValid = false;
1392   SmallVector<Loop *, 4> SibLoops;
1393   if (ParentL)
1394     SibLoops.append(ParentL->begin(), ParentL->end());
1395   else
1396     SibLoops.append(AR.LI.begin(), AR.LI.end());
1397   erase_if(SibLoops, [&](Loop *SibLoop) {
1398     if (SibLoop == &L) {
1399       IsCurrentLoopValid = true;
1400       return true;
1401     }
1402 
1403     // Otherwise erase the loop from the list if it was in the old loops.
1404     return OldLoops.count(SibLoop) != 0;
1405   });
1406   Updater.addSiblingLoops(SibLoops);
1407 
1408   if (!IsCurrentLoopValid) {
1409     Updater.markLoopAsDeleted(L, LoopName);
1410   } else {
1411     // We can only walk child loops if the current loop remained valid.
1412     if (UnrollRevisitChildLoops) {
1413       // Walk *all* of the child loops.
1414       SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end());
1415       Updater.addChildLoops(ChildLoops);
1416     }
1417   }
1418 
1419   return getLoopPassPreservedAnalyses();
1420 }
1421 
1422 PreservedAnalyses LoopUnrollPass::run(Function &F,
1423                                       FunctionAnalysisManager &AM) {
1424   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1425   auto &LI = AM.getResult<LoopAnalysis>(F);
1426   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1427   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1428   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1429   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1430 
1431   LoopAnalysisManager *LAM = nullptr;
1432   if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(F))
1433     LAM = &LAMProxy->getManager();
1434 
1435   auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
1436   ProfileSummaryInfo *PSI =
1437       MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
1438   auto *BFI = (PSI && PSI->hasProfileSummary()) ?
1439       &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
1440 
1441   bool Changed = false;
1442 
1443   // The unroller requires loops to be in simplified form, and also needs LCSSA.
1444   // Since simplification may add new inner loops, it has to run before the
1445   // legality and profitability checks. This means running the loop unroller
1446   // will simplify all loops, regardless of whether anything end up being
1447   // unrolled.
1448   for (auto &L : LI) {
1449     Changed |=
1450         simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */);
1451     Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1452   }
1453 
1454   // Add the loop nests in the reverse order of LoopInfo. See method
1455   // declaration.
1456   SmallPriorityWorklist<Loop *, 4> Worklist;
1457   appendLoopsToWorklist(LI, Worklist);
1458 
1459   while (!Worklist.empty()) {
1460     // Because the LoopInfo stores the loops in RPO, we walk the worklist
1461     // from back to front so that we work forward across the CFG, which
1462     // for unrolling is only needed to get optimization remarks emitted in
1463     // a forward order.
1464     Loop &L = *Worklist.pop_back_val();
1465 #ifndef NDEBUG
1466     Loop *ParentL = L.getParentLoop();
1467 #endif
1468 
1469     // Check if the profile summary indicates that the profiled application
1470     // has a huge working set size, in which case we disable peeling to avoid
1471     // bloating it further.
1472     Optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling;
1473     if (PSI && PSI->hasHugeWorkingSetSize())
1474       LocalAllowPeeling = false;
1475     std::string LoopName = std::string(L.getName());
1476     // The API here is quite complex to call and we allow to select some
1477     // flavors of unrolling during construction time (by setting UnrollOpts).
1478     LoopUnrollResult Result = tryToUnrollLoop(
1479         &L, DT, &LI, SE, TTI, AC, ORE, BFI, PSI,
1480         /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, UnrollOpts.OnlyWhenForced,
1481         UnrollOpts.ForgetSCEV, /*Count*/ None,
1482         /*Threshold*/ None, UnrollOpts.AllowPartial, UnrollOpts.AllowRuntime,
1483         UnrollOpts.AllowUpperBound, LocalAllowPeeling,
1484         UnrollOpts.AllowProfileBasedPeeling, UnrollOpts.FullUnrollMaxCount);
1485     Changed |= Result != LoopUnrollResult::Unmodified;
1486 
1487     // The parent must not be damaged by unrolling!
1488 #ifndef NDEBUG
1489     if (Result != LoopUnrollResult::Unmodified && ParentL)
1490       ParentL->verifyLoop();
1491 #endif
1492 
1493     // Clear any cached analysis results for L if we removed it completely.
1494     if (LAM && Result == LoopUnrollResult::FullyUnrolled)
1495       LAM->clear(L, LoopName);
1496   }
1497 
1498   if (!Changed)
1499     return PreservedAnalyses::all();
1500 
1501   return getLoopPassPreservedAnalyses();
1502 }
1503