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/STLExtras.h"
19#include "llvm/ADT/SetVector.h"
20#include "llvm/ADT/SmallPtrSet.h"
21#include "llvm/ADT/SmallVector.h"
22#include "llvm/ADT/StringRef.h"
23#include "llvm/Analysis/AssumptionCache.h"
24#include "llvm/Analysis/BlockFrequencyInfo.h"
25#include "llvm/Analysis/CodeMetrics.h"
26#include "llvm/Analysis/LoopAnalysisManager.h"
27#include "llvm/Analysis/LoopInfo.h"
28#include "llvm/Analysis/LoopPass.h"
29#include "llvm/Analysis/LoopUnrollAnalyzer.h"
30#include "llvm/Analysis/MemorySSA.h"
31#include "llvm/Analysis/OptimizationRemarkEmitter.h"
32#include "llvm/Analysis/ProfileSummaryInfo.h"
33#include "llvm/Analysis/ScalarEvolution.h"
34#include "llvm/Analysis/TargetTransformInfo.h"
35#include "llvm/IR/BasicBlock.h"
36#include "llvm/IR/CFG.h"
37#include "llvm/IR/Constant.h"
38#include "llvm/IR/Constants.h"
39#include "llvm/IR/DiagnosticInfo.h"
40#include "llvm/IR/Dominators.h"
41#include "llvm/IR/Function.h"
42#include "llvm/IR/Instruction.h"
43#include "llvm/IR/Instructions.h"
44#include "llvm/IR/Metadata.h"
45#include "llvm/IR/PassManager.h"
46#include "llvm/InitializePasses.h"
47#include "llvm/Pass.h"
48#include "llvm/Support/Casting.h"
49#include "llvm/Support/CommandLine.h"
50#include "llvm/Support/Debug.h"
51#include "llvm/Support/ErrorHandling.h"
52#include "llvm/Support/raw_ostream.h"
53#include "llvm/Transforms/Scalar.h"
54#include "llvm/Transforms/Scalar/LoopPassManager.h"
55#include "llvm/Transforms/Utils.h"
56#include "llvm/Transforms/Utils/LoopPeel.h"
57#include "llvm/Transforms/Utils/LoopSimplify.h"
58#include "llvm/Transforms/Utils/LoopUtils.h"
59#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
60#include "llvm/Transforms/Utils/SizeOpts.h"
61#include "llvm/Transforms/Utils/UnrollLoop.h"
62#include <algorithm>
63#include <cassert>
64#include <cstdint>
65#include <limits>
66#include <optional>
67#include <string>
68#include <tuple>
69#include <utility>
70
71using namespace llvm;
72
73#define DEBUG_TYPE "loop-unroll"
74
75cl::opt<bool> llvm::ForgetSCEVInLoopUnroll(
76 "forget-scev-loop-unroll", cl::init(Val: false), cl::Hidden,
77 cl::desc("Forget everything in SCEV when doing LoopUnroll, instead of just"
78 " the current top-most loop. This is sometimes preferred to reduce"
79 " compile time."));
80
81static cl::opt<unsigned>
82 UnrollThreshold("unroll-threshold", cl::Hidden,
83 cl::desc("The cost threshold for loop unrolling"));
84
85static cl::opt<unsigned>
86 UnrollOptSizeThreshold(
87 "unroll-optsize-threshold", cl::init(Val: 0), cl::Hidden,
88 cl::desc("The cost threshold for loop unrolling when optimizing for "
89 "size"));
90
91static cl::opt<unsigned> UnrollPartialThreshold(
92 "unroll-partial-threshold", cl::Hidden,
93 cl::desc("The cost threshold for partial loop unrolling"));
94
95static cl::opt<unsigned> UnrollMaxPercentThresholdBoost(
96 "unroll-max-percent-threshold-boost", cl::init(Val: 400), cl::Hidden,
97 cl::desc("The maximum 'boost' (represented as a percentage >= 100) applied "
98 "to the threshold when aggressively unrolling a loop due to the "
99 "dynamic cost savings. If completely unrolling a loop will reduce "
100 "the total runtime from X to Y, we boost the loop unroll "
101 "threshold to DefaultThreshold*std::min(MaxPercentThresholdBoost, "
102 "X/Y). This limit avoids excessive code bloat."));
103
104static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
105 "unroll-max-iteration-count-to-analyze", cl::init(Val: 10), cl::Hidden,
106 cl::desc("Don't allow loop unrolling to simulate more than this number of "
107 "iterations when checking full unroll profitability"));
108
109static cl::opt<unsigned> UnrollCount(
110 "unroll-count", cl::Hidden,
111 cl::desc("Use this unroll count for all loops including those with "
112 "unroll_count pragma values, for testing purposes"));
113
114static cl::opt<unsigned> UnrollMaxCount(
115 "unroll-max-count", cl::Hidden,
116 cl::desc("Set the max unroll count for partial and runtime unrolling, for"
117 "testing purposes"));
118
119static cl::opt<unsigned> UnrollFullMaxCount(
120 "unroll-full-max-count", cl::Hidden,
121 cl::desc(
122 "Set the max unroll count for full unrolling, for testing purposes"));
123
124static cl::opt<bool>
125 UnrollAllowPartial("unroll-allow-partial", cl::Hidden,
126 cl::desc("Allows loops to be partially unrolled until "
127 "-unroll-threshold loop size is reached."));
128
129static cl::opt<bool> UnrollAllowRemainder(
130 "unroll-allow-remainder", cl::Hidden,
131 cl::desc("Allow generation of a loop remainder (extra iterations) "
132 "when unrolling a loop."));
133
134static cl::opt<bool>
135 UnrollRuntime("unroll-runtime", cl::Hidden,
136 cl::desc("Unroll loops with run-time trip counts"));
137
138static cl::opt<unsigned> UnrollMaxUpperBound(
139 "unroll-max-upperbound", cl::init(Val: 8), cl::Hidden,
140 cl::desc(
141 "The max of trip count upper bound that is considered in unrolling"));
142
143static cl::opt<unsigned> PragmaUnrollThreshold(
144 "pragma-unroll-threshold", cl::init(Val: 16 * 1024), cl::Hidden,
145 cl::desc("Unrolled size limit for loops with an unroll(full) or "
146 "unroll_count pragma."));
147
148static cl::opt<unsigned> FlatLoopTripCountThreshold(
149 "flat-loop-tripcount-threshold", cl::init(Val: 5), cl::Hidden,
150 cl::desc("If the runtime tripcount for the loop is lower than the "
151 "threshold, the loop is considered as flat and will be less "
152 "aggressively unrolled."));
153
154static cl::opt<bool> UnrollUnrollRemainder(
155 "unroll-remainder", cl::Hidden,
156 cl::desc("Allow the loop remainder to be unrolled."));
157
158// This option isn't ever intended to be enabled, it serves to allow
159// experiments to check the assumptions about when this kind of revisit is
160// necessary.
161static cl::opt<bool> UnrollRevisitChildLoops(
162 "unroll-revisit-child-loops", cl::Hidden,
163 cl::desc("Enqueue and re-visit child loops in the loop PM after unrolling. "
164 "This shouldn't typically be needed as child loops (or their "
165 "clones) were already visited."));
166
167static cl::opt<unsigned> UnrollThresholdAggressive(
168 "unroll-threshold-aggressive", cl::init(Val: 300), cl::Hidden,
169 cl::desc("Threshold (max size of unrolled loop) to use in aggressive (O3) "
170 "optimizations"));
171static cl::opt<unsigned>
172 UnrollThresholdDefault("unroll-threshold-default", cl::init(Val: 150),
173 cl::Hidden,
174 cl::desc("Default threshold (max size of unrolled "
175 "loop), used in all but O3 optimizations"));
176
177static cl::opt<unsigned> PragmaUnrollFullMaxIterations(
178 "pragma-unroll-full-max-iterations", cl::init(Val: 1'000'000), cl::Hidden,
179 cl::desc("Maximum allowed iterations to unroll under pragma unroll full."));
180
181/// A magic value for use with the Threshold parameter to indicate
182/// that the loop unroll should be performed regardless of how much
183/// code expansion would result.
184static const unsigned NoThreshold = std::numeric_limits<unsigned>::max();
185
186/// Gather the various unrolling parameters based on the defaults, compiler
187/// flags, TTI overrides and user specified parameters.
188TargetTransformInfo::UnrollingPreferences llvm::gatherUnrollingPreferences(
189 Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
190 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
191 OptimizationRemarkEmitter &ORE, int OptLevel,
192 std::optional<unsigned> UserThreshold, std::optional<unsigned> UserCount,
193 std::optional<bool> UserAllowPartial, std::optional<bool> UserRuntime,
194 std::optional<bool> UserUpperBound,
195 std::optional<unsigned> UserFullUnrollMaxCount) {
196 TargetTransformInfo::UnrollingPreferences UP;
197
198 // Set up the defaults
199 UP.Threshold =
200 OptLevel > 2 ? UnrollThresholdAggressive : UnrollThresholdDefault;
201 UP.MaxPercentThresholdBoost = 400;
202 UP.OptSizeThreshold = UnrollOptSizeThreshold;
203 UP.PartialThreshold = 150;
204 UP.PartialOptSizeThreshold = UnrollOptSizeThreshold;
205 UP.Count = 0;
206 UP.DefaultUnrollRuntimeCount = 8;
207 UP.MaxCount = std::numeric_limits<unsigned>::max();
208 UP.MaxUpperBound = UnrollMaxUpperBound;
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 UP.SCEVExpansionBudget = SCEVCheapExpansionBudget;
222 UP.RuntimeUnrollMultiExit = false;
223 UP.AddAdditionalAccumulators = false;
224
225 // Override with any target specific settings
226 TTI.getUnrollingPreferences(L, SE, UP, ORE: &ORE);
227
228 // Apply size attributes
229 bool OptForSize = L->getHeader()->getParent()->hasOptSize() ||
230 // Let unroll hints / pragmas take precedence over PGSO.
231 (hasUnrollTransformation(L) != TM_ForcedByUser &&
232 llvm::shouldOptimizeForSize(BB: L->getHeader(), PSI, BFI,
233 QueryType: PGSOQueryType::IRPass));
234 if (OptForSize) {
235 UP.Threshold = UP.OptSizeThreshold;
236 UP.PartialThreshold = UP.PartialOptSizeThreshold;
237 UP.MaxPercentThresholdBoost = 100;
238 }
239
240 // Apply any user values specified by cl::opt
241 if (UnrollThreshold.getNumOccurrences() > 0)
242 UP.Threshold = UnrollThreshold;
243 if (UnrollPartialThreshold.getNumOccurrences() > 0)
244 UP.PartialThreshold = UnrollPartialThreshold;
245 if (UnrollMaxPercentThresholdBoost.getNumOccurrences() > 0)
246 UP.MaxPercentThresholdBoost = UnrollMaxPercentThresholdBoost;
247 if (UnrollMaxCount.getNumOccurrences() > 0)
248 UP.MaxCount = UnrollMaxCount;
249 if (UnrollMaxUpperBound.getNumOccurrences() > 0)
250 UP.MaxUpperBound = UnrollMaxUpperBound;
251 if (UnrollFullMaxCount.getNumOccurrences() > 0)
252 UP.FullUnrollMaxCount = UnrollFullMaxCount;
253 if (UnrollAllowPartial.getNumOccurrences() > 0)
254 UP.Partial = UnrollAllowPartial;
255 if (UnrollAllowRemainder.getNumOccurrences() > 0)
256 UP.AllowRemainder = UnrollAllowRemainder;
257 if (UnrollRuntime.getNumOccurrences() > 0)
258 UP.Runtime = UnrollRuntime;
259 if (UnrollMaxUpperBound == 0)
260 UP.UpperBound = false;
261 if (UnrollUnrollRemainder.getNumOccurrences() > 0)
262 UP.UnrollRemainder = UnrollUnrollRemainder;
263 if (UnrollMaxIterationsCountToAnalyze.getNumOccurrences() > 0)
264 UP.MaxIterationsCountToAnalyze = UnrollMaxIterationsCountToAnalyze;
265
266 // Apply user values provided by argument
267 if (UserThreshold) {
268 UP.Threshold = *UserThreshold;
269 UP.PartialThreshold = *UserThreshold;
270 }
271 if (UserCount)
272 UP.Count = *UserCount;
273 if (UserAllowPartial)
274 UP.Partial = *UserAllowPartial;
275 if (UserRuntime)
276 UP.Runtime = *UserRuntime;
277 if (UserUpperBound)
278 UP.UpperBound = *UserUpperBound;
279 if (UserFullUnrollMaxCount)
280 UP.FullUnrollMaxCount = *UserFullUnrollMaxCount;
281
282 return UP;
283}
284
285namespace {
286
287/// A struct to densely store the state of an instruction after unrolling at
288/// each iteration.
289///
290/// This is designed to work like a tuple of <Instruction *, int> for the
291/// purposes of hashing and lookup, but to be able to associate two boolean
292/// states with each key.
293struct UnrolledInstState {
294 Instruction *I;
295 int Iteration : 30;
296 unsigned IsFree : 1;
297 unsigned IsCounted : 1;
298};
299
300/// Hashing and equality testing for a set of the instruction states.
301struct UnrolledInstStateKeyInfo {
302 using PtrInfo = DenseMapInfo<Instruction *>;
303 using PairInfo = DenseMapInfo<std::pair<Instruction *, int>>;
304
305 static inline UnrolledInstState getEmptyKey() {
306 return {.I: PtrInfo::getEmptyKey(), .Iteration: 0, .IsFree: 0, .IsCounted: 0};
307 }
308
309 static inline UnrolledInstState getTombstoneKey() {
310 return {.I: PtrInfo::getTombstoneKey(), .Iteration: 0, .IsFree: 0, .IsCounted: 0};
311 }
312
313 static inline unsigned getHashValue(const UnrolledInstState &S) {
314 return PairInfo::getHashValue(PairVal: {S.I, S.Iteration});
315 }
316
317 static inline bool isEqual(const UnrolledInstState &LHS,
318 const UnrolledInstState &RHS) {
319 return PairInfo::isEqual(LHS: {LHS.I, LHS.Iteration}, RHS: {RHS.I, RHS.Iteration});
320 }
321};
322
323struct EstimatedUnrollCost {
324 /// The estimated cost after unrolling.
325 unsigned UnrolledCost;
326
327 /// The estimated dynamic cost of executing the instructions in the
328 /// rolled form.
329 unsigned RolledDynamicCost;
330};
331
332struct PragmaInfo {
333 PragmaInfo(bool UUC, bool PFU, unsigned PC, bool PEU)
334 : UserUnrollCount(UUC), PragmaFullUnroll(PFU), PragmaCount(PC),
335 PragmaEnableUnroll(PEU) {}
336 const bool UserUnrollCount;
337 const bool PragmaFullUnroll;
338 const unsigned PragmaCount;
339 const bool PragmaEnableUnroll;
340};
341
342} // end anonymous namespace
343
344/// Figure out if the loop is worth full unrolling.
345///
346/// Complete loop unrolling can make some loads constant, and we need to know
347/// if that would expose any further optimization opportunities. This routine
348/// estimates this optimization. It computes cost of unrolled loop
349/// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
350/// dynamic cost we mean that we won't count costs of blocks that are known not
351/// to be executed (i.e. if we have a branch in the loop and we know that at the
352/// given iteration its condition would be resolved to true, we won't add up the
353/// cost of the 'false'-block).
354/// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
355/// the analysis failed (no benefits expected from the unrolling, or the loop is
356/// too big to analyze), the returned value is std::nullopt.
357static std::optional<EstimatedUnrollCost> analyzeLoopUnrollCost(
358 const Loop *L, unsigned TripCount, DominatorTree &DT, ScalarEvolution &SE,
359 const SmallPtrSetImpl<const Value *> &EphValues,
360 const TargetTransformInfo &TTI, unsigned MaxUnrolledLoopSize,
361 unsigned MaxIterationsCountToAnalyze) {
362 // We want to be able to scale offsets by the trip count and add more offsets
363 // to them without checking for overflows, and we already don't want to
364 // analyze *massive* trip counts, so we force the max to be reasonably small.
365 assert(MaxIterationsCountToAnalyze <
366 (unsigned)(std::numeric_limits<int>::max() / 2) &&
367 "The unroll iterations max is too large!");
368
369 // Only analyze inner loops. We can't properly estimate cost of nested loops
370 // and we won't visit inner loops again anyway.
371 if (!L->isInnermost())
372 return std::nullopt;
373
374 // Don't simulate loops with a big or unknown tripcount
375 if (!TripCount || TripCount > MaxIterationsCountToAnalyze)
376 return std::nullopt;
377
378 SmallSetVector<BasicBlock *, 16> BBWorklist;
379 SmallSetVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitWorklist;
380 DenseMap<Value *, Value *> SimplifiedValues;
381 SmallVector<std::pair<Value *, Value *>, 4> SimplifiedInputValues;
382
383 // The estimated cost of the unrolled form of the loop. We try to estimate
384 // this by simplifying as much as we can while computing the estimate.
385 InstructionCost UnrolledCost = 0;
386
387 // We also track the estimated dynamic (that is, actually executed) cost in
388 // the rolled form. This helps identify cases when the savings from unrolling
389 // aren't just exposing dead control flows, but actual reduced dynamic
390 // instructions due to the simplifications which we expect to occur after
391 // unrolling.
392 InstructionCost RolledDynamicCost = 0;
393
394 // We track the simplification of each instruction in each iteration. We use
395 // this to recursively merge costs into the unrolled cost on-demand so that
396 // we don't count the cost of any dead code. This is essentially a map from
397 // <instruction, int> to <bool, bool>, but stored as a densely packed struct.
398 DenseSet<UnrolledInstState, UnrolledInstStateKeyInfo> InstCostMap;
399
400 // A small worklist used to accumulate cost of instructions from each
401 // observable and reached root in the loop.
402 SmallVector<Instruction *, 16> CostWorklist;
403
404 // PHI-used worklist used between iterations while accumulating cost.
405 SmallVector<Instruction *, 4> PHIUsedList;
406
407 // Helper function to accumulate cost for instructions in the loop.
408 auto AddCostRecursively = [&](Instruction &RootI, int Iteration) {
409 assert(Iteration >= 0 && "Cannot have a negative iteration!");
410 assert(CostWorklist.empty() && "Must start with an empty cost list");
411 assert(PHIUsedList.empty() && "Must start with an empty phi used list");
412 CostWorklist.push_back(Elt: &RootI);
413 TargetTransformInfo::TargetCostKind CostKind =
414 RootI.getFunction()->hasMinSize() ?
415 TargetTransformInfo::TCK_CodeSize :
416 TargetTransformInfo::TCK_SizeAndLatency;
417 for (;; --Iteration) {
418 do {
419 Instruction *I = CostWorklist.pop_back_val();
420
421 // InstCostMap only uses I and Iteration as a key, the other two values
422 // don't matter here.
423 auto CostIter = InstCostMap.find(V: {.I: I, .Iteration: Iteration, .IsFree: 0, .IsCounted: 0});
424 if (CostIter == InstCostMap.end())
425 // If an input to a PHI node comes from a dead path through the loop
426 // we may have no cost data for it here. What that actually means is
427 // that it is free.
428 continue;
429 auto &Cost = *CostIter;
430 if (Cost.IsCounted)
431 // Already counted this instruction.
432 continue;
433
434 // Mark that we are counting the cost of this instruction now.
435 Cost.IsCounted = true;
436
437 // If this is a PHI node in the loop header, just add it to the PHI set.
438 if (auto *PhiI = dyn_cast<PHINode>(Val: I))
439 if (PhiI->getParent() == L->getHeader()) {
440 assert(Cost.IsFree && "Loop PHIs shouldn't be evaluated as they "
441 "inherently simplify during unrolling.");
442 if (Iteration == 0)
443 continue;
444
445 // Push the incoming value from the backedge into the PHI used list
446 // if it is an in-loop instruction. We'll use this to populate the
447 // cost worklist for the next iteration (as we count backwards).
448 if (auto *OpI = dyn_cast<Instruction>(
449 Val: PhiI->getIncomingValueForBlock(BB: L->getLoopLatch())))
450 if (L->contains(Inst: OpI))
451 PHIUsedList.push_back(Elt: OpI);
452 continue;
453 }
454
455 // First accumulate the cost of this instruction.
456 if (!Cost.IsFree) {
457 // Consider simplified operands in instruction cost.
458 SmallVector<Value *, 4> Operands;
459 transform(Range: I->operands(), d_first: std::back_inserter(x&: Operands),
460 F: [&](Value *Op) {
461 if (auto Res = SimplifiedValues.lookup(Val: Op))
462 return Res;
463 return Op;
464 });
465 UnrolledCost += TTI.getInstructionCost(U: I, Operands, CostKind);
466 LLVM_DEBUG(dbgs().indent(4)
467 << "Adding cost of instruction (iteration " << Iteration
468 << "): ");
469 LLVM_DEBUG(I->dump());
470 }
471
472 // We must count the cost of every operand which is not free,
473 // recursively. If we reach a loop PHI node, simply add it to the set
474 // to be considered on the next iteration (backwards!).
475 for (Value *Op : I->operands()) {
476 // Check whether this operand is free due to being a constant or
477 // outside the loop.
478 auto *OpI = dyn_cast<Instruction>(Val: Op);
479 if (!OpI || !L->contains(Inst: OpI))
480 continue;
481
482 // Otherwise accumulate its cost.
483 CostWorklist.push_back(Elt: OpI);
484 }
485 } while (!CostWorklist.empty());
486
487 if (PHIUsedList.empty())
488 // We've exhausted the search.
489 break;
490
491 assert(Iteration > 0 &&
492 "Cannot track PHI-used values past the first iteration!");
493 CostWorklist.append(in_start: PHIUsedList.begin(), in_end: PHIUsedList.end());
494 PHIUsedList.clear();
495 }
496 };
497
498 // Ensure that we don't violate the loop structure invariants relied on by
499 // this analysis.
500 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
501 assert(L->isLCSSAForm(DT) &&
502 "Must have loops in LCSSA form to track live-out values.");
503
504 LLVM_DEBUG(dbgs().indent(4)
505 << "Starting LoopUnroll profitability analysis...\n");
506
507 TargetTransformInfo::TargetCostKind CostKind =
508 L->getHeader()->getParent()->hasMinSize() ?
509 TargetTransformInfo::TCK_CodeSize : TargetTransformInfo::TCK_SizeAndLatency;
510 // Simulate execution of each iteration of the loop counting instructions,
511 // which would be simplified.
512 // Since the same load will take different values on different iterations,
513 // we literally have to go through all loop's iterations.
514 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
515 LLVM_DEBUG(dbgs().indent(4) << "Analyzing iteration " << Iteration << "\n");
516
517 // Prepare for the iteration by collecting any simplified entry or backedge
518 // inputs.
519 for (Instruction &I : *L->getHeader()) {
520 auto *PHI = dyn_cast<PHINode>(Val: &I);
521 if (!PHI)
522 break;
523
524 // The loop header PHI nodes must have exactly two input: one from the
525 // loop preheader and one from the loop latch.
526 assert(
527 PHI->getNumIncomingValues() == 2 &&
528 "Must have an incoming value only for the preheader and the latch.");
529
530 Value *V = PHI->getIncomingValueForBlock(
531 BB: Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
532 if (Iteration != 0 && SimplifiedValues.count(Val: V))
533 V = SimplifiedValues.lookup(Val: V);
534 SimplifiedInputValues.push_back(Elt: {PHI, V});
535 }
536
537 // Now clear and re-populate the map for the next iteration.
538 SimplifiedValues.clear();
539 while (!SimplifiedInputValues.empty())
540 SimplifiedValues.insert(KV: SimplifiedInputValues.pop_back_val());
541
542 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L);
543
544 BBWorklist.clear();
545 BBWorklist.insert(X: L->getHeader());
546 // Note that we *must not* cache the size, this loop grows the worklist.
547 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
548 BasicBlock *BB = BBWorklist[Idx];
549
550 // Visit all instructions in the given basic block and try to simplify
551 // it. We don't change the actual IR, just count optimization
552 // opportunities.
553 for (Instruction &I : *BB) {
554 // These won't get into the final code - don't even try calculating the
555 // cost for them.
556 if (EphValues.count(Ptr: &I))
557 continue;
558
559 // Track this instruction's expected baseline cost when executing the
560 // rolled loop form.
561 RolledDynamicCost += TTI.getInstructionCost(U: &I, CostKind);
562
563 // Visit the instruction to analyze its loop cost after unrolling,
564 // and if the visitor returns true, mark the instruction as free after
565 // unrolling and continue.
566 bool IsFree = Analyzer.visit(I);
567 bool Inserted = InstCostMap.insert(V: {.I: &I, .Iteration: (int)Iteration,
568 .IsFree: (unsigned)IsFree,
569 /*IsCounted*/ false}).second;
570 (void)Inserted;
571 assert(Inserted && "Cannot have a state for an unvisited instruction!");
572
573 if (IsFree)
574 continue;
575
576 // Can't properly model a cost of a call.
577 // FIXME: With a proper cost model we should be able to do it.
578 if (auto *CI = dyn_cast<CallInst>(Val: &I)) {
579 const Function *Callee = CI->getCalledFunction();
580 if (!Callee || TTI.isLoweredToCall(F: Callee)) {
581 LLVM_DEBUG(dbgs().indent(4)
582 << "Can't analyze cost of loop with call\n");
583 return std::nullopt;
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({
595 dbgs().indent(4) << "Exceeded threshold.. exiting.\n";
596 dbgs().indent(4)
597 << "UnrolledCost: " << UnrolledCost
598 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize << "\n";
599 });
600 return std::nullopt;
601 }
602 }
603
604 Instruction *TI = BB->getTerminator();
605
606 auto getSimplifiedConstant = [&](Value *V) -> Constant * {
607 if (SimplifiedValues.count(Val: V))
608 V = SimplifiedValues.lookup(Val: V);
609 return dyn_cast<Constant>(Val: V);
610 };
611
612 // Add in the live successors by first checking whether we have terminator
613 // that may be simplified based on the values simplified by this call.
614 BasicBlock *KnownSucc = nullptr;
615 if (BranchInst *BI = dyn_cast<BranchInst>(Val: TI)) {
616 if (BI->isConditional()) {
617 if (auto *SimpleCond = getSimplifiedConstant(BI->getCondition())) {
618 // Just take the first successor if condition is undef
619 if (isa<UndefValue>(Val: SimpleCond))
620 KnownSucc = BI->getSuccessor(i: 0);
621 else if (ConstantInt *SimpleCondVal =
622 dyn_cast<ConstantInt>(Val: SimpleCond))
623 KnownSucc = BI->getSuccessor(i: SimpleCondVal->isZero() ? 1 : 0);
624 }
625 }
626 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: TI)) {
627 if (auto *SimpleCond = getSimplifiedConstant(SI->getCondition())) {
628 // Just take the first successor if condition is undef
629 if (isa<UndefValue>(Val: SimpleCond))
630 KnownSucc = SI->getSuccessor(idx: 0);
631 else if (ConstantInt *SimpleCondVal =
632 dyn_cast<ConstantInt>(Val: SimpleCond))
633 KnownSucc = SI->findCaseValue(C: SimpleCondVal)->getCaseSuccessor();
634 }
635 }
636 if (KnownSucc) {
637 if (L->contains(BB: KnownSucc))
638 BBWorklist.insert(X: KnownSucc);
639 else
640 ExitWorklist.insert(X: {BB, KnownSucc});
641 continue;
642 }
643
644 // Add BB's successors to the worklist.
645 for (BasicBlock *Succ : successors(BB))
646 if (L->contains(BB: Succ))
647 BBWorklist.insert(X: Succ);
648 else
649 ExitWorklist.insert(X: {BB, Succ});
650 AddCostRecursively(*TI, Iteration);
651 }
652
653 // If we found no optimization opportunities on the first iteration, we
654 // won't find them on later ones too.
655 if (UnrolledCost == RolledDynamicCost) {
656 LLVM_DEBUG({
657 dbgs().indent(4) << "No opportunities found.. exiting.\n";
658 dbgs().indent(4) << "UnrolledCost: " << UnrolledCost << "\n";
659 });
660 return std::nullopt;
661 }
662 }
663
664 while (!ExitWorklist.empty()) {
665 BasicBlock *ExitingBB, *ExitBB;
666 std::tie(args&: ExitingBB, args&: ExitBB) = ExitWorklist.pop_back_val();
667
668 for (Instruction &I : *ExitBB) {
669 auto *PN = dyn_cast<PHINode>(Val: &I);
670 if (!PN)
671 break;
672
673 Value *Op = PN->getIncomingValueForBlock(BB: ExitingBB);
674 if (auto *OpI = dyn_cast<Instruction>(Val: Op))
675 if (L->contains(Inst: OpI))
676 AddCostRecursively(*OpI, TripCount - 1);
677 }
678 }
679
680 assert(UnrolledCost.isValid() && RolledDynamicCost.isValid() &&
681 "All instructions must have a valid cost, whether the "
682 "loop is rolled or unrolled.");
683
684 LLVM_DEBUG({
685 dbgs().indent(4) << "Analysis finished:\n";
686 dbgs().indent(4) << "UnrolledCost: " << UnrolledCost
687 << ", RolledDynamicCost: " << RolledDynamicCost << "\n";
688 });
689 return {{.UnrolledCost: unsigned(UnrolledCost.getValue()),
690 .RolledDynamicCost: unsigned(RolledDynamicCost.getValue())}};
691}
692
693UnrollCostEstimator::UnrollCostEstimator(
694 const Loop *L, const TargetTransformInfo &TTI,
695 const SmallPtrSetImpl<const Value *> &EphValues, unsigned BEInsns) {
696 CodeMetrics Metrics;
697 for (BasicBlock *BB : L->blocks())
698 Metrics.analyzeBasicBlock(BB, TTI, EphValues, /* PrepareForLTO= */ false,
699 L);
700 NumInlineCandidates = Metrics.NumInlineCandidates;
701 NotDuplicatable = Metrics.notDuplicatable;
702 Convergence = Metrics.Convergence;
703 LoopSize = Metrics.NumInsts;
704 ConvergenceAllowsRuntime =
705 Metrics.Convergence != ConvergenceKind::Uncontrolled &&
706 !getLoopConvergenceHeart(TheLoop: L);
707
708 // Don't allow an estimate of size zero. This would allows unrolling of loops
709 // with huge iteration counts, which is a compile time problem even if it's
710 // not a problem for code quality. Also, the code using this size may assume
711 // that each loop has at least three instructions (likely a conditional
712 // branch, a comparison feeding that branch, and some kind of loop increment
713 // feeding that comparison instruction).
714 if (LoopSize.isValid() && LoopSize < BEInsns + 1)
715 // This is an open coded max() on InstructionCost
716 LoopSize = BEInsns + 1;
717}
718
719bool UnrollCostEstimator::canUnroll() const {
720 if (Convergence == ConvergenceKind::ExtendedLoop) {
721 LLVM_DEBUG(dbgs().indent(1)
722 << "Not unrolling: contains convergent operations.\n");
723 return false;
724 }
725 if (!LoopSize.isValid()) {
726 LLVM_DEBUG(dbgs().indent(1)
727 << "Not unrolling: loop size could not be computed.\n");
728 return false;
729 }
730 if (NotDuplicatable) {
731 LLVM_DEBUG(dbgs().indent(1)
732 << "Not unrolling: contains non-duplicatable instructions.\n");
733 return false;
734 }
735 return true;
736}
737
738uint64_t UnrollCostEstimator::getUnrolledLoopSize(
739 const TargetTransformInfo::UnrollingPreferences &UP,
740 unsigned CountOverwrite) const {
741 unsigned LS = LoopSize.getValue();
742 assert(LS >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
743 if (CountOverwrite)
744 return static_cast<uint64_t>(LS - UP.BEInsns) * CountOverwrite + UP.BEInsns;
745 else
746 return static_cast<uint64_t>(LS - UP.BEInsns) * UP.Count + UP.BEInsns;
747}
748
749// Returns the loop hint metadata node with the given name (for example,
750// "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
751// returned.
752static MDNode *getUnrollMetadataForLoop(const Loop *L, StringRef Name) {
753 if (MDNode *LoopID = L->getLoopID())
754 return GetUnrollMetadata(LoopID, Name);
755 return nullptr;
756}
757
758// Returns true if the loop has an unroll(full) pragma.
759static bool hasUnrollFullPragma(const Loop *L) {
760 return getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.full");
761}
762
763// Returns true if the loop has an unroll(enable) pragma. This metadata is used
764// for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
765static bool hasUnrollEnablePragma(const Loop *L) {
766 return getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.enable");
767}
768
769// Returns true if the loop has an runtime unroll(disable) pragma.
770static bool hasRuntimeUnrollDisablePragma(const Loop *L) {
771 return getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.runtime.disable");
772}
773
774// If loop has an unroll_count pragma return the (necessarily
775// positive) value from the pragma. Otherwise return 0.
776static unsigned unrollCountPragmaValue(const Loop *L) {
777 MDNode *MD = getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.count");
778 if (MD) {
779 assert(MD->getNumOperands() == 2 &&
780 "Unroll count hint metadata should have two operands.");
781 unsigned Count =
782 mdconst::extract<ConstantInt>(MD: MD->getOperand(I: 1))->getZExtValue();
783 assert(Count >= 1 && "Unroll count must be positive.");
784 return Count;
785 }
786 return 0;
787}
788
789// Computes the boosting factor for complete unrolling.
790// If fully unrolling the loop would save a lot of RolledDynamicCost, it would
791// be beneficial to fully unroll the loop even if unrolledcost is large. We
792// use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust
793// the unroll threshold.
794static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost,
795 unsigned MaxPercentThresholdBoost) {
796 if (Cost.RolledDynamicCost >= std::numeric_limits<unsigned>::max() / 100)
797 return 100;
798 else if (Cost.UnrolledCost != 0)
799 // The boosting factor is RolledDynamicCost / UnrolledCost
800 return std::min(a: 100 * Cost.RolledDynamicCost / Cost.UnrolledCost,
801 b: MaxPercentThresholdBoost);
802 else
803 return MaxPercentThresholdBoost;
804}
805
806static std::optional<unsigned>
807shouldPragmaUnroll(Loop *L, const PragmaInfo &PInfo,
808 const unsigned TripMultiple, const unsigned TripCount,
809 unsigned MaxTripCount, const UnrollCostEstimator UCE,
810 const TargetTransformInfo::UnrollingPreferences &UP) {
811
812 // Using unroll pragma
813 // 1st priority is unroll count set by "unroll-count" option.
814
815 if (PInfo.UserUnrollCount) {
816 if (UP.AllowRemainder &&
817 UCE.getUnrolledLoopSize(UP, CountOverwrite: (unsigned)UnrollCount) < UP.Threshold)
818 return (unsigned)UnrollCount;
819 }
820
821 // 2nd priority is unroll count set by pragma.
822 if (PInfo.PragmaCount > 0) {
823 if ((UP.AllowRemainder || (TripMultiple % PInfo.PragmaCount == 0)))
824 return PInfo.PragmaCount;
825 }
826
827 if (PInfo.PragmaFullUnroll && TripCount != 0) {
828 // Certain cases with UBSAN can cause trip count to be calculated as
829 // INT_MAX, Block full unrolling at a reasonable limit so that the compiler
830 // doesn't hang trying to unroll the loop. See PR77842
831 if (TripCount > PragmaUnrollFullMaxIterations) {
832 LLVM_DEBUG(dbgs().indent(3) << "Won't unroll; trip count is too large\n");
833 return std::nullopt;
834 }
835
836 return TripCount;
837 }
838
839 if (PInfo.PragmaEnableUnroll && !TripCount && MaxTripCount &&
840 MaxTripCount <= UP.MaxUpperBound)
841 return MaxTripCount;
842
843 return std::nullopt;
844}
845
846static std::optional<unsigned> shouldFullUnroll(
847 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT,
848 ScalarEvolution &SE, const SmallPtrSetImpl<const Value *> &EphValues,
849 const unsigned FullUnrollTripCount, const UnrollCostEstimator UCE,
850 const TargetTransformInfo::UnrollingPreferences &UP) {
851 assert(FullUnrollTripCount && "should be non-zero!");
852
853 if (FullUnrollTripCount > UP.FullUnrollMaxCount)
854 return std::nullopt;
855
856 // When computing the unrolled size, note that BEInsns are not replicated
857 // like the rest of the loop body.
858 if (UCE.getUnrolledLoopSize(UP) < UP.Threshold)
859 return FullUnrollTripCount;
860
861 // The loop isn't that small, but we still can fully unroll it if that
862 // helps to remove a significant number of instructions.
863 // To check that, run additional analysis on the loop.
864 if (std::optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
865 L, TripCount: FullUnrollTripCount, DT, SE, EphValues, TTI,
866 MaxUnrolledLoopSize: UP.Threshold * UP.MaxPercentThresholdBoost / 100,
867 MaxIterationsCountToAnalyze: UP.MaxIterationsCountToAnalyze)) {
868 unsigned Boost =
869 getFullUnrollBoostingFactor(Cost: *Cost, MaxPercentThresholdBoost: UP.MaxPercentThresholdBoost);
870 if (Cost->UnrolledCost < UP.Threshold * Boost / 100)
871 return FullUnrollTripCount;
872 }
873 return std::nullopt;
874}
875
876static std::optional<unsigned>
877shouldPartialUnroll(const unsigned LoopSize, const unsigned TripCount,
878 const UnrollCostEstimator UCE,
879 const TargetTransformInfo::UnrollingPreferences &UP) {
880
881 if (!TripCount)
882 return std::nullopt;
883
884 if (!UP.Partial) {
885 LLVM_DEBUG(dbgs().indent(3) << "Will not try to unroll partially because "
886 << "-unroll-allow-partial not given\n");
887 return 0;
888 }
889 unsigned count = UP.Count;
890 if (count == 0)
891 count = TripCount;
892 if (UP.PartialThreshold != NoThreshold) {
893 // Reduce unroll count to be modulo of TripCount for partial unrolling.
894 if (UCE.getUnrolledLoopSize(UP, CountOverwrite: count) > UP.PartialThreshold)
895 count = (std::max(a: UP.PartialThreshold, b: UP.BEInsns + 1) - UP.BEInsns) /
896 (LoopSize - UP.BEInsns);
897 if (count > UP.MaxCount)
898 count = UP.MaxCount;
899 while (count != 0 && TripCount % count != 0)
900 count--;
901 if (UP.AllowRemainder && count <= 1) {
902 // If there is no Count that is modulo of TripCount, set Count to
903 // largest power-of-two factor that satisfies the threshold limit.
904 // As we'll create fixup loop, do the type of unrolling only if
905 // remainder loop is allowed.
906 // Note: DefaultUnrollRuntimeCount is used as a reasonable starting point
907 // even though this is partial unrolling (not runtime unrolling).
908 count = UP.DefaultUnrollRuntimeCount;
909 while (count != 0 &&
910 UCE.getUnrolledLoopSize(UP, CountOverwrite: count) > UP.PartialThreshold)
911 count >>= 1;
912 }
913 if (count < 2) {
914 count = 0;
915 }
916 } else {
917 count = TripCount;
918 }
919 if (count > UP.MaxCount)
920 count = UP.MaxCount;
921
922 LLVM_DEBUG(dbgs().indent(3)
923 << "Partially unrolling with count: " << count << "\n");
924
925 return count;
926}
927// Returns true if unroll count was set explicitly.
928// Calculates unroll count and writes it to UP.Count.
929// Unless IgnoreUser is true, will also use metadata and command-line options
930// that are specific to the LoopUnroll pass (which, for instance, are
931// irrelevant for the LoopUnrollAndJam pass).
932// FIXME: This function is used by LoopUnroll and LoopUnrollAndJam, but consumes
933// many LoopUnroll-specific options. The shared functionality should be
934// refactored into it own function.
935bool llvm::computeUnrollCount(
936 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI,
937 AssumptionCache *AC, ScalarEvolution &SE,
938 const SmallPtrSetImpl<const Value *> &EphValues,
939 OptimizationRemarkEmitter *ORE, unsigned TripCount, unsigned MaxTripCount,
940 bool MaxOrZero, unsigned TripMultiple, const UnrollCostEstimator &UCE,
941 TargetTransformInfo::UnrollingPreferences &UP,
942 TargetTransformInfo::PeelingPreferences &PP, bool &UseUpperBound) {
943
944 unsigned LoopSize = UCE.getRolledLoopSize();
945
946 const bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0;
947 const bool PragmaFullUnroll = hasUnrollFullPragma(L);
948 const unsigned PragmaCount = unrollCountPragmaValue(L);
949 const bool PragmaEnableUnroll = hasUnrollEnablePragma(L);
950
951 const bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll ||
952 PragmaEnableUnroll || UserUnrollCount;
953
954 PragmaInfo PInfo(UserUnrollCount, PragmaFullUnroll, PragmaCount,
955 PragmaEnableUnroll);
956 // Use an explicit peel count that has been specified for testing. In this
957 // case it's not permitted to also specify an explicit unroll count.
958 if (PP.PeelCount) {
959 if (UnrollCount.getNumOccurrences() > 0) {
960 reportFatalUsageError(reason: "Cannot specify both explicit peel count and "
961 "explicit unroll count");
962 }
963 UP.Count = 1;
964 UP.Runtime = false;
965 return true;
966 }
967 // Check for explicit Count.
968 // 1st priority is unroll count set by "unroll-count" option.
969 // 2nd priority is unroll count set by pragma.
970 if (auto UnrollFactor = shouldPragmaUnroll(L, PInfo, TripMultiple, TripCount,
971 MaxTripCount, UCE, UP)) {
972 UP.Count = *UnrollFactor;
973
974 if (UserUnrollCount || (PragmaCount > 0)) {
975 UP.AllowExpensiveTripCount = true;
976 UP.Force = true;
977 }
978 UP.Runtime |= (PragmaCount > 0);
979 return ExplicitUnroll;
980 } else {
981 if (ExplicitUnroll && TripCount != 0) {
982 // If the loop has an unrolling pragma, we want to be more aggressive with
983 // unrolling limits. Set thresholds to at least the PragmaUnrollThreshold
984 // value which is larger than the default limits.
985 UP.Threshold = std::max<unsigned>(a: UP.Threshold, b: PragmaUnrollThreshold);
986 UP.PartialThreshold =
987 std::max<unsigned>(a: UP.PartialThreshold, b: PragmaUnrollThreshold);
988 }
989 }
990
991 // 3rd priority is exact full unrolling. This will eliminate all copies
992 // of some exit test.
993 UP.Count = 0;
994 if (TripCount) {
995 UP.Count = TripCount;
996 if (auto UnrollFactor = shouldFullUnroll(L, TTI, DT, SE, EphValues,
997 FullUnrollTripCount: TripCount, UCE, UP)) {
998 UP.Count = *UnrollFactor;
999 UseUpperBound = false;
1000 return ExplicitUnroll;
1001 }
1002 }
1003
1004 // 4th priority is bounded unrolling.
1005 // We can unroll by the upper bound amount if it's generally allowed or if
1006 // we know that the loop is executed either the upper bound or zero times.
1007 // (MaxOrZero unrolling keeps only the first loop test, so the number of
1008 // loop tests remains the same compared to the non-unrolled version, whereas
1009 // the generic upper bound unrolling keeps all but the last loop test so the
1010 // number of loop tests goes up which may end up being worse on targets with
1011 // constrained branch predictor resources so is controlled by an option.)
1012 // In addition we only unroll small upper bounds.
1013 // Note that the cost of bounded unrolling is always strictly greater than
1014 // cost of exact full unrolling. As such, if we have an exact count and
1015 // found it unprofitable, we'll never chose to bounded unroll.
1016 if (!TripCount && MaxTripCount && (UP.UpperBound || MaxOrZero) &&
1017 MaxTripCount <= UP.MaxUpperBound) {
1018 UP.Count = MaxTripCount;
1019 if (auto UnrollFactor = shouldFullUnroll(L, TTI, DT, SE, EphValues,
1020 FullUnrollTripCount: MaxTripCount, UCE, UP)) {
1021 UP.Count = *UnrollFactor;
1022 UseUpperBound = true;
1023 return ExplicitUnroll;
1024 }
1025 }
1026
1027 // 5th priority is loop peeling.
1028 computePeelCount(L, LoopSize, PP, TripCount, DT, SE, TTI, AC, Threshold: UP.Threshold);
1029 if (PP.PeelCount) {
1030 UP.Runtime = false;
1031 UP.Count = 1;
1032 return ExplicitUnroll;
1033 }
1034
1035 // Before starting partial unrolling, set up.partial to true,
1036 // if user explicitly asked for unrolling
1037 if (TripCount)
1038 UP.Partial |= ExplicitUnroll;
1039
1040 // 6th priority is partial unrolling.
1041 // Try partial unroll only when TripCount could be statically calculated.
1042 if (auto UnrollFactor = shouldPartialUnroll(LoopSize, TripCount, UCE, UP)) {
1043 UP.Count = *UnrollFactor;
1044
1045 if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
1046 UP.Count != TripCount)
1047 ORE->emit(RemarkBuilder: [&]() {
1048 return OptimizationRemarkMissed(DEBUG_TYPE,
1049 "FullUnrollAsDirectedTooLarge",
1050 L->getStartLoc(), L->getHeader())
1051 << "Unable to fully unroll loop as directed by unroll pragma "
1052 "because "
1053 "unrolled size is too large.";
1054 });
1055
1056 if (UP.PartialThreshold != NoThreshold) {
1057 if (UP.Count == 0) {
1058 if (PragmaEnableUnroll)
1059 ORE->emit(RemarkBuilder: [&]() {
1060 return OptimizationRemarkMissed(DEBUG_TYPE,
1061 "UnrollAsDirectedTooLarge",
1062 L->getStartLoc(), L->getHeader())
1063 << "Unable to unroll loop as directed by unroll(enable) "
1064 "pragma "
1065 "because unrolled size is too large.";
1066 });
1067 }
1068 }
1069 return ExplicitUnroll;
1070 }
1071 assert(TripCount == 0 &&
1072 "All cases when TripCount is constant should be covered here.");
1073 if (PragmaFullUnroll)
1074 ORE->emit(RemarkBuilder: [&]() {
1075 return OptimizationRemarkMissed(
1076 DEBUG_TYPE, "CantFullUnrollAsDirectedRuntimeTripCount",
1077 L->getStartLoc(), L->getHeader())
1078 << "Unable to fully unroll loop as directed by unroll(full) "
1079 "pragma "
1080 "because loop has a runtime trip count.";
1081 });
1082
1083 // 7th priority is runtime unrolling.
1084 // Don't unroll a runtime trip count loop when it is disabled.
1085 if (hasRuntimeUnrollDisablePragma(L)) {
1086 UP.Count = 0;
1087 return false;
1088 }
1089
1090 // Don't unroll a small upper bound loop unless user or TTI asked to do so.
1091 if (MaxTripCount && !UP.Force && MaxTripCount < UP.MaxUpperBound) {
1092 UP.Count = 0;
1093 return false;
1094 }
1095
1096 // Check if the runtime trip count is too small when profile is available.
1097 if (L->getHeader()->getParent()->hasProfileData()) {
1098 if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) {
1099 if (*ProfileTripCount < FlatLoopTripCountThreshold)
1100 return false;
1101 else
1102 UP.AllowExpensiveTripCount = true;
1103 }
1104 }
1105 UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount;
1106 if (!UP.Runtime) {
1107 LLVM_DEBUG(dbgs().indent(2)
1108 << "Will not try to unroll loop with runtime trip count "
1109 << "because -unroll-runtime not given\n");
1110 UP.Count = 0;
1111 return false;
1112 }
1113 if (UP.Count == 0)
1114 UP.Count = UP.DefaultUnrollRuntimeCount;
1115
1116 // Reduce unroll count to be the largest power-of-two factor of
1117 // the original count which satisfies the threshold limit.
1118 while (UP.Count != 0 &&
1119 UCE.getUnrolledLoopSize(UP) > UP.PartialThreshold)
1120 UP.Count >>= 1;
1121
1122#ifndef NDEBUG
1123 unsigned OrigCount = UP.Count;
1124#endif
1125
1126 if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) {
1127 while (UP.Count != 0 && TripMultiple % UP.Count != 0)
1128 UP.Count >>= 1;
1129 LLVM_DEBUG(dbgs().indent(2)
1130 << "Remainder loop is restricted (that could architecture "
1131 "specific or because the loop contains a convergent "
1132 "instruction), so unroll count must divide the trip "
1133 "multiple, "
1134 << TripMultiple << ". Reducing unroll count from " << OrigCount
1135 << " to " << UP.Count << ".\n");
1136
1137 using namespace ore;
1138
1139 if (PragmaCount > 0 && !UP.AllowRemainder)
1140 ORE->emit(RemarkBuilder: [&]() {
1141 return OptimizationRemarkMissed(DEBUG_TYPE,
1142 "DifferentUnrollCountFromDirected",
1143 L->getStartLoc(), L->getHeader())
1144 << "Unable to unroll loop the number of times directed by "
1145 "unroll_count pragma because remainder loop is restricted "
1146 "(that could architecture specific or because the loop "
1147 "contains a convergent instruction) and so must have an "
1148 "unroll "
1149 "count that divides the loop trip multiple of "
1150 << NV("TripMultiple", TripMultiple) << ". Unrolling instead "
1151 << NV("UnrollCount", UP.Count) << " time(s).";
1152 });
1153 }
1154
1155 if (UP.Count > UP.MaxCount)
1156 UP.Count = UP.MaxCount;
1157
1158 if (MaxTripCount && UP.Count > MaxTripCount)
1159 UP.Count = MaxTripCount;
1160
1161 LLVM_DEBUG(dbgs().indent(2)
1162 << "Runtime unrolling with count: " << UP.Count << "\n");
1163 if (UP.Count < 2)
1164 UP.Count = 0;
1165 return ExplicitUnroll;
1166}
1167
1168static LoopUnrollResult
1169tryToUnrollLoop(Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE,
1170 const TargetTransformInfo &TTI, AssumptionCache &AC,
1171 OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
1172 ProfileSummaryInfo *PSI, bool PreserveLCSSA, int OptLevel,
1173 bool OnlyFullUnroll, bool OnlyWhenForced, bool ForgetAllSCEV,
1174 std::optional<unsigned> ProvidedCount,
1175 std::optional<unsigned> ProvidedThreshold,
1176 std::optional<bool> ProvidedAllowPartial,
1177 std::optional<bool> ProvidedRuntime,
1178 std::optional<bool> ProvidedUpperBound,
1179 std::optional<bool> ProvidedAllowPeeling,
1180 std::optional<bool> ProvidedAllowProfileBasedPeeling,
1181 std::optional<unsigned> ProvidedFullUnrollMaxCount,
1182 AAResults *AA = nullptr) {
1183
1184 LLVM_DEBUG(dbgs() << "Loop Unroll: F["
1185 << L->getHeader()->getParent()->getName() << "] Loop %"
1186 << L->getHeader()->getName() << "\n");
1187 TransformationMode TM = hasUnrollTransformation(L);
1188 if (TM & TM_Disable)
1189 return LoopUnrollResult::Unmodified;
1190
1191 // If this loop isn't forced to be unrolled, avoid unrolling it when the
1192 // parent loop has an explicit unroll-and-jam pragma. This is to prevent
1193 // automatic unrolling from interfering with the user requested
1194 // transformation.
1195 Loop *ParentL = L->getParentLoop();
1196 if (ParentL != nullptr &&
1197 hasUnrollAndJamTransformation(L: ParentL) == TM_ForcedByUser &&
1198 hasUnrollTransformation(L) != TM_ForcedByUser) {
1199 LLVM_DEBUG(dbgs().indent(1) << "Not unrolling loop since parent loop has"
1200 << " llvm.loop.unroll_and_jam.\n");
1201 return LoopUnrollResult::Unmodified;
1202 }
1203
1204 // If this loop isn't forced to be unrolled, avoid unrolling it when the
1205 // loop has an explicit unroll-and-jam pragma. This is to prevent automatic
1206 // unrolling from interfering with the user requested transformation.
1207 if (hasUnrollAndJamTransformation(L) == TM_ForcedByUser &&
1208 hasUnrollTransformation(L) != TM_ForcedByUser) {
1209 LLVM_DEBUG(
1210 dbgs().indent(1)
1211 << "Not unrolling loop since it has llvm.loop.unroll_and_jam.\n");
1212 return LoopUnrollResult::Unmodified;
1213 }
1214
1215 if (!L->isLoopSimplifyForm()) {
1216 LLVM_DEBUG(dbgs().indent(1)
1217 << "Not unrolling loop which is not in loop-simplify form.\n");
1218 return LoopUnrollResult::Unmodified;
1219 }
1220
1221 // When automatic unrolling is disabled, do not unroll unless overridden for
1222 // this loop.
1223 if (OnlyWhenForced && !(TM & TM_Enable))
1224 return LoopUnrollResult::Unmodified;
1225
1226 bool OptForSize = L->getHeader()->getParent()->hasOptSize();
1227 TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
1228 L, SE, TTI, BFI, PSI, ORE, OptLevel, UserThreshold: ProvidedThreshold, UserCount: ProvidedCount,
1229 UserAllowPartial: ProvidedAllowPartial, UserRuntime: ProvidedRuntime, UserUpperBound: ProvidedUpperBound,
1230 UserFullUnrollMaxCount: ProvidedFullUnrollMaxCount);
1231 TargetTransformInfo::PeelingPreferences PP = gatherPeelingPreferences(
1232 L, SE, TTI, UserAllowPeeling: ProvidedAllowPeeling, UserAllowProfileBasedPeeling: ProvidedAllowProfileBasedPeeling, UnrollingSpecficValues: true);
1233
1234 // Exit early if unrolling is disabled. For OptForSize, we pick the loop size
1235 // as threshold later on.
1236 if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0) &&
1237 !OptForSize)
1238 return LoopUnrollResult::Unmodified;
1239
1240 SmallPtrSet<const Value *, 32> EphValues;
1241 CodeMetrics::collectEphemeralValues(L, AC: &AC, EphValues);
1242
1243 UnrollCostEstimator UCE(L, TTI, EphValues, UP.BEInsns);
1244 if (!UCE.canUnroll())
1245 return LoopUnrollResult::Unmodified;
1246
1247 unsigned LoopSize = UCE.getRolledLoopSize();
1248 LLVM_DEBUG(dbgs() << "Loop Size = " << LoopSize << "\n");
1249
1250 // When optimizing for size, use LoopSize + 1 as threshold (we use < Threshold
1251 // later), to (fully) unroll loops, if it does not increase code size.
1252 if (OptForSize)
1253 UP.Threshold = std::max(a: UP.Threshold, b: LoopSize + 1);
1254
1255 if (UCE.NumInlineCandidates != 0) {
1256 LLVM_DEBUG(dbgs().indent(1)
1257 << "Not unrolling loop with inlinable calls.\n");
1258 return LoopUnrollResult::Unmodified;
1259 }
1260
1261 // Find the smallest exact trip count for any exit. This is an upper bound
1262 // on the loop trip count, but an exit at an earlier iteration is still
1263 // possible. An unroll by the smallest exact trip count guarantees that all
1264 // branches relating to at least one exit can be eliminated. This is unlike
1265 // the max trip count, which only guarantees that the backedge can be broken.
1266 unsigned TripCount = 0;
1267 unsigned TripMultiple = 1;
1268 SmallVector<BasicBlock *, 8> ExitingBlocks;
1269 L->getExitingBlocks(ExitingBlocks);
1270 for (BasicBlock *ExitingBlock : ExitingBlocks)
1271 if (unsigned TC = SE.getSmallConstantTripCount(L, ExitingBlock))
1272 if (!TripCount || TC < TripCount)
1273 TripCount = TripMultiple = TC;
1274
1275 if (!TripCount) {
1276 // If no exact trip count is known, determine the trip multiple of either
1277 // the loop latch or the single exiting block.
1278 // TODO: Relax for multiple exits.
1279 BasicBlock *ExitingBlock = L->getLoopLatch();
1280 if (!ExitingBlock || !L->isLoopExiting(BB: ExitingBlock))
1281 ExitingBlock = L->getExitingBlock();
1282 if (ExitingBlock)
1283 TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock);
1284 }
1285
1286 // If the loop contains a convergent operation, the prelude we'd add
1287 // to do the first few instructions before we hit the unrolled loop
1288 // is unsafe -- it adds a control-flow dependency to the convergent
1289 // operation. Therefore restrict remainder loop (try unrolling without).
1290 //
1291 // TODO: This is somewhat conservative; we could allow the remainder if the
1292 // trip count is uniform.
1293 UP.AllowRemainder &= UCE.ConvergenceAllowsRuntime;
1294
1295 // Try to find the trip count upper bound if we cannot find the exact trip
1296 // count.
1297 unsigned MaxTripCount = 0;
1298 bool MaxOrZero = false;
1299 if (!TripCount) {
1300 MaxTripCount = SE.getSmallConstantMaxTripCount(L);
1301 MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L);
1302 }
1303
1304 // computeUnrollCount() decides whether it is beneficial to use upper bound to
1305 // fully unroll the loop.
1306 bool UseUpperBound = false;
1307 bool IsCountSetExplicitly = computeUnrollCount(
1308 L, TTI, DT, LI, AC: &AC, SE, EphValues, ORE: &ORE, TripCount, MaxTripCount,
1309 MaxOrZero, TripMultiple, UCE, UP, PP, UseUpperBound);
1310 if (!UP.Count)
1311 return LoopUnrollResult::Unmodified;
1312
1313 UP.Runtime &= UCE.ConvergenceAllowsRuntime;
1314
1315 if (PP.PeelCount) {
1316 assert(UP.Count == 1 && "Cannot perform peel and unroll in the same step");
1317 LLVM_DEBUG(dbgs() << "PEELING loop %" << L->getHeader()->getName()
1318 << " with iteration count " << PP.PeelCount << "!\n");
1319 ORE.emit(RemarkBuilder: [&]() {
1320 return OptimizationRemark(DEBUG_TYPE, "Peeled", L->getStartLoc(),
1321 L->getHeader())
1322 << "peeled loop by " << ore::NV("PeelCount", PP.PeelCount)
1323 << " iterations";
1324 });
1325
1326 ValueToValueMapTy VMap;
1327 peelLoop(L, PeelCount: PP.PeelCount, PeelLast: PP.PeelLast, LI, SE: &SE, DT, AC: &AC, PreserveLCSSA,
1328 VMap);
1329 simplifyLoopAfterUnroll(L, SimplifyIVs: true, LI, SE: &SE, DT: &DT, AC: &AC, TTI: &TTI, AA: nullptr);
1330 // If the loop was peeled, we already "used up" the profile information
1331 // we had, so we don't want to unroll or peel again.
1332 if (PP.PeelProfiledIterations)
1333 L->setLoopAlreadyUnrolled();
1334 return LoopUnrollResult::PartiallyUnrolled;
1335 }
1336
1337 // Do not attempt partial/runtime unrolling in FullLoopUnrolling
1338 if (OnlyFullUnroll && ((!TripCount && !MaxTripCount) ||
1339 UP.Count < TripCount || UP.Count < MaxTripCount)) {
1340 LLVM_DEBUG(dbgs().indent(1)
1341 << "Not attempting partial/runtime unroll in FullLoopUnroll.\n");
1342 return LoopUnrollResult::Unmodified;
1343 }
1344
1345 // At this point, UP.Runtime indicates that run-time unrolling is allowed.
1346 // However, we only want to actually perform it if we don't know the trip
1347 // count and the unroll count doesn't divide the known trip multiple.
1348 // TODO: This decision should probably be pushed up into
1349 // computeUnrollCount().
1350 UP.Runtime &= TripCount == 0 && TripMultiple % UP.Count != 0;
1351
1352 // Save loop properties before it is transformed.
1353 MDNode *OrigLoopID = L->getLoopID();
1354
1355 // Unroll the loop.
1356 Loop *RemainderLoop = nullptr;
1357 UnrollLoopOptions ULO;
1358 ULO.Count = UP.Count;
1359 ULO.Force = UP.Force;
1360 ULO.AllowExpensiveTripCount = UP.AllowExpensiveTripCount;
1361 ULO.UnrollRemainder = UP.UnrollRemainder;
1362 ULO.Runtime = UP.Runtime;
1363 ULO.ForgetAllSCEV = ForgetAllSCEV;
1364 ULO.Heart = getLoopConvergenceHeart(TheLoop: L);
1365 ULO.SCEVExpansionBudget = UP.SCEVExpansionBudget;
1366 ULO.RuntimeUnrollMultiExit = UP.RuntimeUnrollMultiExit;
1367 ULO.AddAdditionalAccumulators = UP.AddAdditionalAccumulators;
1368 LoopUnrollResult UnrollResult = UnrollLoop(
1369 L, ULO, LI, SE: &SE, DT: &DT, AC: &AC, TTI: &TTI, ORE: &ORE, PreserveLCSSA, RemainderLoop: &RemainderLoop, AA);
1370 if (UnrollResult == LoopUnrollResult::Unmodified)
1371 return LoopUnrollResult::Unmodified;
1372
1373 if (RemainderLoop) {
1374 std::optional<MDNode *> RemainderLoopID =
1375 makeFollowupLoopID(OrigLoopID, FollowupAttrs: {LLVMLoopUnrollFollowupAll,
1376 LLVMLoopUnrollFollowupRemainder});
1377 if (RemainderLoopID)
1378 RemainderLoop->setLoopID(*RemainderLoopID);
1379 }
1380
1381 if (UnrollResult != LoopUnrollResult::FullyUnrolled) {
1382 std::optional<MDNode *> NewLoopID =
1383 makeFollowupLoopID(OrigLoopID, FollowupAttrs: {LLVMLoopUnrollFollowupAll,
1384 LLVMLoopUnrollFollowupUnrolled});
1385 if (NewLoopID) {
1386 L->setLoopID(*NewLoopID);
1387
1388 // Do not setLoopAlreadyUnrolled if loop attributes have been specified
1389 // explicitly.
1390 return UnrollResult;
1391 }
1392 }
1393
1394 // If loop has an unroll count pragma or unrolled by explicitly set count
1395 // mark loop as unrolled to prevent unrolling beyond that requested.
1396 if (UnrollResult != LoopUnrollResult::FullyUnrolled && IsCountSetExplicitly)
1397 L->setLoopAlreadyUnrolled();
1398
1399 return UnrollResult;
1400}
1401
1402namespace {
1403
1404class LoopUnroll : public LoopPass {
1405public:
1406 static char ID; // Pass ID, replacement for typeid
1407
1408 int OptLevel;
1409
1410 /// If false, use a cost model to determine whether unrolling of a loop is
1411 /// profitable. If true, only loops that explicitly request unrolling via
1412 /// metadata are considered. All other loops are skipped.
1413 bool OnlyWhenForced;
1414
1415 /// If false, when SCEV is invalidated, only forget everything in the
1416 /// top-most loop (call forgetTopMostLoop), of the loop being processed.
1417 /// Otherwise, forgetAllLoops and rebuild when needed next.
1418 bool ForgetAllSCEV;
1419
1420 std::optional<unsigned> ProvidedCount;
1421 std::optional<unsigned> ProvidedThreshold;
1422 std::optional<bool> ProvidedAllowPartial;
1423 std::optional<bool> ProvidedRuntime;
1424 std::optional<bool> ProvidedUpperBound;
1425 std::optional<bool> ProvidedAllowPeeling;
1426 std::optional<bool> ProvidedAllowProfileBasedPeeling;
1427 std::optional<unsigned> ProvidedFullUnrollMaxCount;
1428
1429 LoopUnroll(int OptLevel = 2, bool OnlyWhenForced = false,
1430 bool ForgetAllSCEV = false,
1431 std::optional<unsigned> Threshold = std::nullopt,
1432 std::optional<unsigned> Count = std::nullopt,
1433 std::optional<bool> AllowPartial = std::nullopt,
1434 std::optional<bool> Runtime = std::nullopt,
1435 std::optional<bool> UpperBound = std::nullopt,
1436 std::optional<bool> AllowPeeling = std::nullopt,
1437 std::optional<bool> AllowProfileBasedPeeling = std::nullopt,
1438 std::optional<unsigned> ProvidedFullUnrollMaxCount = std::nullopt)
1439 : LoopPass(ID), OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
1440 ForgetAllSCEV(ForgetAllSCEV), ProvidedCount(std::move(Count)),
1441 ProvidedThreshold(Threshold), ProvidedAllowPartial(AllowPartial),
1442 ProvidedRuntime(Runtime), ProvidedUpperBound(UpperBound),
1443 ProvidedAllowPeeling(AllowPeeling),
1444 ProvidedAllowProfileBasedPeeling(AllowProfileBasedPeeling),
1445 ProvidedFullUnrollMaxCount(ProvidedFullUnrollMaxCount) {
1446 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
1447 }
1448
1449 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1450 if (skipLoop(L))
1451 return false;
1452
1453 Function &F = *L->getHeader()->getParent();
1454
1455 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1456 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1457 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1458 const TargetTransformInfo &TTI =
1459 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1460 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1461 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
1462 // pass. Function analyses need to be preserved across loop transformations
1463 // but ORE cannot be preserved (see comment before the pass definition).
1464 OptimizationRemarkEmitter ORE(&F);
1465 bool PreserveLCSSA = mustPreserveAnalysisID(AID&: LCSSAID);
1466
1467 LoopUnrollResult Result = tryToUnrollLoop(
1468 L, DT, LI, SE, TTI, AC, ORE, BFI: nullptr, PSI: nullptr, PreserveLCSSA, OptLevel,
1469 /*OnlyFullUnroll*/ false, OnlyWhenForced, ForgetAllSCEV, ProvidedCount,
1470 ProvidedThreshold, ProvidedAllowPartial, ProvidedRuntime,
1471 ProvidedUpperBound, ProvidedAllowPeeling,
1472 ProvidedAllowProfileBasedPeeling, ProvidedFullUnrollMaxCount);
1473
1474 if (Result == LoopUnrollResult::FullyUnrolled)
1475 LPM.markLoopAsDeleted(L&: *L);
1476
1477 return Result != LoopUnrollResult::Unmodified;
1478 }
1479
1480 /// This transformation requires natural loop information & requires that
1481 /// loop preheaders be inserted into the CFG...
1482 void getAnalysisUsage(AnalysisUsage &AU) const override {
1483 AU.addRequired<AssumptionCacheTracker>();
1484 AU.addRequired<TargetTransformInfoWrapperPass>();
1485 // FIXME: Loop passes are required to preserve domtree, and for now we just
1486 // recreate dom info if anything gets unrolled.
1487 getLoopAnalysisUsage(AU);
1488 }
1489};
1490
1491} // end anonymous namespace
1492
1493char LoopUnroll::ID = 0;
1494
1495INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1496INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1497INITIALIZE_PASS_DEPENDENCY(LoopPass)
1498INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1499INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1500
1501Pass *llvm::createLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1502 bool ForgetAllSCEV, int Threshold, int Count,
1503 int AllowPartial, int Runtime, int UpperBound,
1504 int AllowPeeling) {
1505 // TODO: It would make more sense for this function to take the optionals
1506 // directly, but that's dangerous since it would silently break out of tree
1507 // callers.
1508 return new LoopUnroll(
1509 OptLevel, OnlyWhenForced, ForgetAllSCEV,
1510 Threshold == -1 ? std::nullopt : std::optional<unsigned>(Threshold),
1511 Count == -1 ? std::nullopt : std::optional<unsigned>(Count),
1512 AllowPartial == -1 ? std::nullopt : std::optional<bool>(AllowPartial),
1513 Runtime == -1 ? std::nullopt : std::optional<bool>(Runtime),
1514 UpperBound == -1 ? std::nullopt : std::optional<bool>(UpperBound),
1515 AllowPeeling == -1 ? std::nullopt : std::optional<bool>(AllowPeeling));
1516}
1517
1518PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
1519 LoopStandardAnalysisResults &AR,
1520 LPMUpdater &Updater) {
1521 // For the new PM, we can't use OptimizationRemarkEmitter as an analysis
1522 // pass. Function analyses need to be preserved across loop transformations
1523 // but ORE cannot be preserved (see comment before the pass definition).
1524 OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
1525
1526 // Keep track of the previous loop structure so we can identify new loops
1527 // created by unrolling.
1528 Loop *ParentL = L.getParentLoop();
1529 SmallPtrSet<Loop *, 4> OldLoops;
1530 if (ParentL)
1531 OldLoops.insert_range(R&: *ParentL);
1532 else
1533 OldLoops.insert_range(R&: AR.LI);
1534
1535 std::string LoopName = std::string(L.getName());
1536
1537 bool Changed =
1538 tryToUnrollLoop(L: &L, DT&: AR.DT, LI: &AR.LI, SE&: AR.SE, TTI: AR.TTI, AC&: AR.AC, ORE,
1539 /*BFI*/ nullptr, /*PSI*/ nullptr,
1540 /*PreserveLCSSA*/ true, OptLevel, /*OnlyFullUnroll*/ true,
1541 OnlyWhenForced, ForgetAllSCEV: ForgetSCEV, /*Count*/ ProvidedCount: std::nullopt,
1542 /*Threshold*/ ProvidedThreshold: std::nullopt, /*AllowPartial*/ ProvidedAllowPartial: false,
1543 /*Runtime*/ ProvidedRuntime: false, /*UpperBound*/ ProvidedUpperBound: false,
1544 /*AllowPeeling*/ ProvidedAllowPeeling: true,
1545 /*AllowProfileBasedPeeling*/ ProvidedAllowProfileBasedPeeling: false,
1546 /*FullUnrollMaxCount*/ ProvidedFullUnrollMaxCount: std::nullopt) !=
1547 LoopUnrollResult::Unmodified;
1548 if (!Changed)
1549 return PreservedAnalyses::all();
1550
1551 // The parent must not be damaged by unrolling!
1552#ifndef NDEBUG
1553 if (ParentL)
1554 ParentL->verifyLoop();
1555#endif
1556
1557 // Unrolling can do several things to introduce new loops into a loop nest:
1558 // - Full unrolling clones child loops within the current loop but then
1559 // removes the current loop making all of the children appear to be new
1560 // sibling loops.
1561 //
1562 // When a new loop appears as a sibling loop after fully unrolling,
1563 // its nesting structure has fundamentally changed and we want to revisit
1564 // it to reflect that.
1565 //
1566 // When unrolling has removed the current loop, we need to tell the
1567 // infrastructure that it is gone.
1568 //
1569 // Finally, we support a debugging/testing mode where we revisit child loops
1570 // as well. These are not expected to require further optimizations as either
1571 // they or the loop they were cloned from have been directly visited already.
1572 // But the debugging mode allows us to check this assumption.
1573 bool IsCurrentLoopValid = false;
1574 SmallVector<Loop *, 4> SibLoops;
1575 if (ParentL)
1576 SibLoops.append(in_start: ParentL->begin(), in_end: ParentL->end());
1577 else
1578 SibLoops.append(in_start: AR.LI.begin(), in_end: AR.LI.end());
1579 erase_if(C&: SibLoops, P: [&](Loop *SibLoop) {
1580 if (SibLoop == &L) {
1581 IsCurrentLoopValid = true;
1582 return true;
1583 }
1584
1585 // Otherwise erase the loop from the list if it was in the old loops.
1586 return OldLoops.contains(Ptr: SibLoop);
1587 });
1588 Updater.addSiblingLoops(NewSibLoops: SibLoops);
1589
1590 if (!IsCurrentLoopValid) {
1591 Updater.markLoopAsDeleted(L, Name: LoopName);
1592 } else {
1593 // We can only walk child loops if the current loop remained valid.
1594 if (UnrollRevisitChildLoops) {
1595 // Walk *all* of the child loops.
1596 SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end());
1597 Updater.addChildLoops(NewChildLoops: ChildLoops);
1598 }
1599 }
1600
1601 return getLoopPassPreservedAnalyses();
1602}
1603
1604PreservedAnalyses LoopUnrollPass::run(Function &F,
1605 FunctionAnalysisManager &AM) {
1606 auto &LI = AM.getResult<LoopAnalysis>(IR&: F);
1607 // There are no loops in the function. Return before computing other expensive
1608 // analyses.
1609 if (LI.empty())
1610 return PreservedAnalyses::all();
1611 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(IR&: F);
1612 auto &TTI = AM.getResult<TargetIRAnalysis>(IR&: F);
1613 auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F);
1614 auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F);
1615 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F);
1616 AAResults &AA = AM.getResult<AAManager>(IR&: F);
1617
1618 LoopAnalysisManager *LAM = nullptr;
1619 if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(IR&: F))
1620 LAM = &LAMProxy->getManager();
1621
1622 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F);
1623 ProfileSummaryInfo *PSI =
1624 MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
1625 auto *BFI = (PSI && PSI->hasProfileSummary()) ?
1626 &AM.getResult<BlockFrequencyAnalysis>(IR&: F) : nullptr;
1627
1628 bool Changed = false;
1629
1630 // The unroller requires loops to be in simplified form, and also needs LCSSA.
1631 // Since simplification may add new inner loops, it has to run before the
1632 // legality and profitability checks. This means running the loop unroller
1633 // will simplify all loops, regardless of whether anything end up being
1634 // unrolled.
1635 for (const auto &L : LI) {
1636 Changed |=
1637 simplifyLoop(L, DT: &DT, LI: &LI, SE: &SE, AC: &AC, MSSAU: nullptr, PreserveLCSSA: false /* PreserveLCSSA */);
1638 Changed |= formLCSSARecursively(L&: *L, DT, LI: &LI, SE: &SE);
1639 }
1640
1641 // Add the loop nests in the reverse order of LoopInfo. See method
1642 // declaration.
1643 SmallPriorityWorklist<Loop *, 4> Worklist;
1644 appendLoopsToWorklist(LI, Worklist);
1645
1646 while (!Worklist.empty()) {
1647 // Because the LoopInfo stores the loops in RPO, we walk the worklist
1648 // from back to front so that we work forward across the CFG, which
1649 // for unrolling is only needed to get optimization remarks emitted in
1650 // a forward order.
1651 Loop &L = *Worklist.pop_back_val();
1652#ifndef NDEBUG
1653 Loop *ParentL = L.getParentLoop();
1654#endif
1655
1656 // Check if the profile summary indicates that the profiled application
1657 // has a huge working set size, in which case we disable peeling to avoid
1658 // bloating it further.
1659 std::optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling;
1660 if (PSI && PSI->hasHugeWorkingSetSize())
1661 LocalAllowPeeling = false;
1662 std::string LoopName = std::string(L.getName());
1663 // The API here is quite complex to call and we allow to select some
1664 // flavors of unrolling during construction time (by setting UnrollOpts).
1665 LoopUnrollResult Result = tryToUnrollLoop(
1666 L: &L, DT, LI: &LI, SE, TTI, AC, ORE, BFI, PSI,
1667 /*PreserveLCSSA*/ true, OptLevel: UnrollOpts.OptLevel, /*OnlyFullUnroll*/ false,
1668 OnlyWhenForced: UnrollOpts.OnlyWhenForced, ForgetAllSCEV: UnrollOpts.ForgetSCEV,
1669 /*Count*/ ProvidedCount: std::nullopt,
1670 /*Threshold*/ ProvidedThreshold: std::nullopt, ProvidedAllowPartial: UnrollOpts.AllowPartial,
1671 ProvidedRuntime: UnrollOpts.AllowRuntime, ProvidedUpperBound: UnrollOpts.AllowUpperBound, ProvidedAllowPeeling: LocalAllowPeeling,
1672 ProvidedAllowProfileBasedPeeling: UnrollOpts.AllowProfileBasedPeeling, ProvidedFullUnrollMaxCount: UnrollOpts.FullUnrollMaxCount,
1673 AA: &AA);
1674 Changed |= Result != LoopUnrollResult::Unmodified;
1675
1676 // The parent must not be damaged by unrolling!
1677#ifndef NDEBUG
1678 if (Result != LoopUnrollResult::Unmodified && ParentL)
1679 ParentL->verifyLoop();
1680#endif
1681
1682 // Clear any cached analysis results for L if we removed it completely.
1683 if (LAM && Result == LoopUnrollResult::FullyUnrolled)
1684 LAM->clear(IR&: L, Name: LoopName);
1685 }
1686
1687 if (!Changed)
1688 return PreservedAnalyses::all();
1689
1690 return getLoopPassPreservedAnalyses();
1691}
1692
1693void LoopUnrollPass::printPipeline(
1694 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1695 static_cast<PassInfoMixin<LoopUnrollPass> *>(this)->printPipeline(
1696 OS, MapClassName2PassName);
1697 OS << '<';
1698 if (UnrollOpts.AllowPartial != std::nullopt)
1699 OS << (*UnrollOpts.AllowPartial ? "" : "no-") << "partial;";
1700 if (UnrollOpts.AllowPeeling != std::nullopt)
1701 OS << (*UnrollOpts.AllowPeeling ? "" : "no-") << "peeling;";
1702 if (UnrollOpts.AllowRuntime != std::nullopt)
1703 OS << (*UnrollOpts.AllowRuntime ? "" : "no-") << "runtime;";
1704 if (UnrollOpts.AllowUpperBound != std::nullopt)
1705 OS << (*UnrollOpts.AllowUpperBound ? "" : "no-") << "upperbound;";
1706 if (UnrollOpts.AllowProfileBasedPeeling != std::nullopt)
1707 OS << (*UnrollOpts.AllowProfileBasedPeeling ? "" : "no-")
1708 << "profile-peeling;";
1709 if (UnrollOpts.FullUnrollMaxCount != std::nullopt)
1710 OS << "full-unroll-max=" << UnrollOpts.FullUnrollMaxCount << ';';
1711 OS << 'O' << UnrollOpts.OptLevel;
1712 OS << '>';
1713}
1714