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() << "Adding cost of instruction (iteration "
467 << Iteration << "): ");
468 LLVM_DEBUG(I->dump());
469 }
470
471 // We must count the cost of every operand which is not free,
472 // recursively. If we reach a loop PHI node, simply add it to the set
473 // to be considered on the next iteration (backwards!).
474 for (Value *Op : I->operands()) {
475 // Check whether this operand is free due to being a constant or
476 // outside the loop.
477 auto *OpI = dyn_cast<Instruction>(Val: Op);
478 if (!OpI || !L->contains(Inst: OpI))
479 continue;
480
481 // Otherwise accumulate its cost.
482 CostWorklist.push_back(Elt: OpI);
483 }
484 } while (!CostWorklist.empty());
485
486 if (PHIUsedList.empty())
487 // We've exhausted the search.
488 break;
489
490 assert(Iteration > 0 &&
491 "Cannot track PHI-used values past the first iteration!");
492 CostWorklist.append(in_start: PHIUsedList.begin(), in_end: PHIUsedList.end());
493 PHIUsedList.clear();
494 }
495 };
496
497 // Ensure that we don't violate the loop structure invariants relied on by
498 // this analysis.
499 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
500 assert(L->isLCSSAForm(DT) &&
501 "Must have loops in LCSSA form to track live-out values.");
502
503 LLVM_DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
504
505 TargetTransformInfo::TargetCostKind CostKind =
506 L->getHeader()->getParent()->hasMinSize() ?
507 TargetTransformInfo::TCK_CodeSize : TargetTransformInfo::TCK_SizeAndLatency;
508 // Simulate execution of each iteration of the loop counting instructions,
509 // which would be simplified.
510 // Since the same load will take different values on different iterations,
511 // we literally have to go through all loop's iterations.
512 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
513 LLVM_DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
514
515 // Prepare for the iteration by collecting any simplified entry or backedge
516 // inputs.
517 for (Instruction &I : *L->getHeader()) {
518 auto *PHI = dyn_cast<PHINode>(Val: &I);
519 if (!PHI)
520 break;
521
522 // The loop header PHI nodes must have exactly two input: one from the
523 // loop preheader and one from the loop latch.
524 assert(
525 PHI->getNumIncomingValues() == 2 &&
526 "Must have an incoming value only for the preheader and the latch.");
527
528 Value *V = PHI->getIncomingValueForBlock(
529 BB: Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
530 if (Iteration != 0 && SimplifiedValues.count(Val: V))
531 V = SimplifiedValues.lookup(Val: V);
532 SimplifiedInputValues.push_back(Elt: {PHI, V});
533 }
534
535 // Now clear and re-populate the map for the next iteration.
536 SimplifiedValues.clear();
537 while (!SimplifiedInputValues.empty())
538 SimplifiedValues.insert(KV: SimplifiedInputValues.pop_back_val());
539
540 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L);
541
542 BBWorklist.clear();
543 BBWorklist.insert(X: L->getHeader());
544 // Note that we *must not* cache the size, this loop grows the worklist.
545 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
546 BasicBlock *BB = BBWorklist[Idx];
547
548 // Visit all instructions in the given basic block and try to simplify
549 // it. We don't change the actual IR, just count optimization
550 // opportunities.
551 for (Instruction &I : *BB) {
552 // These won't get into the final code - don't even try calculating the
553 // cost for them.
554 if (EphValues.count(Ptr: &I))
555 continue;
556
557 // Track this instruction's expected baseline cost when executing the
558 // rolled loop form.
559 RolledDynamicCost += TTI.getInstructionCost(U: &I, CostKind);
560
561 // Visit the instruction to analyze its loop cost after unrolling,
562 // and if the visitor returns true, mark the instruction as free after
563 // unrolling and continue.
564 bool IsFree = Analyzer.visit(I);
565 bool Inserted = InstCostMap.insert(V: {.I: &I, .Iteration: (int)Iteration,
566 .IsFree: (unsigned)IsFree,
567 /*IsCounted*/ false}).second;
568 (void)Inserted;
569 assert(Inserted && "Cannot have a state for an unvisited instruction!");
570
571 if (IsFree)
572 continue;
573
574 // Can't properly model a cost of a call.
575 // FIXME: With a proper cost model we should be able to do it.
576 if (auto *CI = dyn_cast<CallInst>(Val: &I)) {
577 const Function *Callee = CI->getCalledFunction();
578 if (!Callee || TTI.isLoweredToCall(F: Callee)) {
579 LLVM_DEBUG(dbgs() << "Can't analyze cost of loop with call\n");
580 return std::nullopt;
581 }
582 }
583
584 // If the instruction might have a side-effect recursively account for
585 // the cost of it and all the instructions leading up to it.
586 if (I.mayHaveSideEffects())
587 AddCostRecursively(I, Iteration);
588
589 // If unrolled body turns out to be too big, bail out.
590 if (UnrolledCost > MaxUnrolledLoopSize) {
591 LLVM_DEBUG(dbgs() << " Exceeded threshold.. exiting.\n"
592 << " UnrolledCost: " << UnrolledCost
593 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
594 << "\n");
595 return std::nullopt;
596 }
597 }
598
599 Instruction *TI = BB->getTerminator();
600
601 auto getSimplifiedConstant = [&](Value *V) -> Constant * {
602 if (SimplifiedValues.count(Val: V))
603 V = SimplifiedValues.lookup(Val: V);
604 return dyn_cast<Constant>(Val: V);
605 };
606
607 // Add in the live successors by first checking whether we have terminator
608 // that may be simplified based on the values simplified by this call.
609 BasicBlock *KnownSucc = nullptr;
610 if (BranchInst *BI = dyn_cast<BranchInst>(Val: TI)) {
611 if (BI->isConditional()) {
612 if (auto *SimpleCond = getSimplifiedConstant(BI->getCondition())) {
613 // Just take the first successor if condition is undef
614 if (isa<UndefValue>(Val: SimpleCond))
615 KnownSucc = BI->getSuccessor(i: 0);
616 else if (ConstantInt *SimpleCondVal =
617 dyn_cast<ConstantInt>(Val: SimpleCond))
618 KnownSucc = BI->getSuccessor(i: SimpleCondVal->isZero() ? 1 : 0);
619 }
620 }
621 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: TI)) {
622 if (auto *SimpleCond = getSimplifiedConstant(SI->getCondition())) {
623 // Just take the first successor if condition is undef
624 if (isa<UndefValue>(Val: SimpleCond))
625 KnownSucc = SI->getSuccessor(idx: 0);
626 else if (ConstantInt *SimpleCondVal =
627 dyn_cast<ConstantInt>(Val: SimpleCond))
628 KnownSucc = SI->findCaseValue(C: SimpleCondVal)->getCaseSuccessor();
629 }
630 }
631 if (KnownSucc) {
632 if (L->contains(BB: KnownSucc))
633 BBWorklist.insert(X: KnownSucc);
634 else
635 ExitWorklist.insert(X: {BB, KnownSucc});
636 continue;
637 }
638
639 // Add BB's successors to the worklist.
640 for (BasicBlock *Succ : successors(BB))
641 if (L->contains(BB: Succ))
642 BBWorklist.insert(X: Succ);
643 else
644 ExitWorklist.insert(X: {BB, Succ});
645 AddCostRecursively(*TI, Iteration);
646 }
647
648 // If we found no optimization opportunities on the first iteration, we
649 // won't find them on later ones too.
650 if (UnrolledCost == RolledDynamicCost) {
651 LLVM_DEBUG(dbgs() << " No opportunities found.. exiting.\n"
652 << " UnrolledCost: " << UnrolledCost << "\n");
653 return std::nullopt;
654 }
655 }
656
657 while (!ExitWorklist.empty()) {
658 BasicBlock *ExitingBB, *ExitBB;
659 std::tie(args&: ExitingBB, args&: ExitBB) = ExitWorklist.pop_back_val();
660
661 for (Instruction &I : *ExitBB) {
662 auto *PN = dyn_cast<PHINode>(Val: &I);
663 if (!PN)
664 break;
665
666 Value *Op = PN->getIncomingValueForBlock(BB: ExitingBB);
667 if (auto *OpI = dyn_cast<Instruction>(Val: Op))
668 if (L->contains(Inst: OpI))
669 AddCostRecursively(*OpI, TripCount - 1);
670 }
671 }
672
673 assert(UnrolledCost.isValid() && RolledDynamicCost.isValid() &&
674 "All instructions must have a valid cost, whether the "
675 "loop is rolled or unrolled.");
676
677 LLVM_DEBUG(dbgs() << "Analysis finished:\n"
678 << "UnrolledCost: " << UnrolledCost << ", "
679 << "RolledDynamicCost: " << RolledDynamicCost << "\n");
680 return {{.UnrolledCost: unsigned(UnrolledCost.getValue()),
681 .RolledDynamicCost: unsigned(RolledDynamicCost.getValue())}};
682}
683
684UnrollCostEstimator::UnrollCostEstimator(
685 const Loop *L, const TargetTransformInfo &TTI,
686 const SmallPtrSetImpl<const Value *> &EphValues, unsigned BEInsns) {
687 CodeMetrics Metrics;
688 for (BasicBlock *BB : L->blocks())
689 Metrics.analyzeBasicBlock(BB, TTI, EphValues, /* PrepareForLTO= */ false,
690 L);
691 NumInlineCandidates = Metrics.NumInlineCandidates;
692 NotDuplicatable = Metrics.notDuplicatable;
693 Convergence = Metrics.Convergence;
694 LoopSize = Metrics.NumInsts;
695 ConvergenceAllowsRuntime =
696 Metrics.Convergence != ConvergenceKind::Uncontrolled &&
697 !getLoopConvergenceHeart(TheLoop: L);
698
699 // Don't allow an estimate of size zero. This would allows unrolling of loops
700 // with huge iteration counts, which is a compile time problem even if it's
701 // not a problem for code quality. Also, the code using this size may assume
702 // that each loop has at least three instructions (likely a conditional
703 // branch, a comparison feeding that branch, and some kind of loop increment
704 // feeding that comparison instruction).
705 if (LoopSize.isValid() && LoopSize < BEInsns + 1)
706 // This is an open coded max() on InstructionCost
707 LoopSize = BEInsns + 1;
708}
709
710bool UnrollCostEstimator::canUnroll() const {
711 switch (Convergence) {
712 case ConvergenceKind::ExtendedLoop:
713 LLVM_DEBUG(dbgs() << " Convergence prevents unrolling.\n");
714 return false;
715 default:
716 break;
717 }
718 if (!LoopSize.isValid()) {
719 LLVM_DEBUG(dbgs() << " Invalid loop size prevents unrolling.\n");
720 return false;
721 }
722 if (NotDuplicatable) {
723 LLVM_DEBUG(dbgs() << " Non-duplicatable blocks prevent unrolling.\n");
724 return false;
725 }
726 return true;
727}
728
729uint64_t UnrollCostEstimator::getUnrolledLoopSize(
730 const TargetTransformInfo::UnrollingPreferences &UP,
731 unsigned CountOverwrite) const {
732 unsigned LS = LoopSize.getValue();
733 assert(LS >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
734 if (CountOverwrite)
735 return static_cast<uint64_t>(LS - UP.BEInsns) * CountOverwrite + UP.BEInsns;
736 else
737 return static_cast<uint64_t>(LS - UP.BEInsns) * UP.Count + UP.BEInsns;
738}
739
740// Returns the loop hint metadata node with the given name (for example,
741// "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
742// returned.
743static MDNode *getUnrollMetadataForLoop(const Loop *L, StringRef Name) {
744 if (MDNode *LoopID = L->getLoopID())
745 return GetUnrollMetadata(LoopID, Name);
746 return nullptr;
747}
748
749// Returns true if the loop has an unroll(full) pragma.
750static bool hasUnrollFullPragma(const Loop *L) {
751 return getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.full");
752}
753
754// Returns true if the loop has an unroll(enable) pragma. This metadata is used
755// for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
756static bool hasUnrollEnablePragma(const Loop *L) {
757 return getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.enable");
758}
759
760// Returns true if the loop has an runtime unroll(disable) pragma.
761static bool hasRuntimeUnrollDisablePragma(const Loop *L) {
762 return getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.runtime.disable");
763}
764
765// If loop has an unroll_count pragma return the (necessarily
766// positive) value from the pragma. Otherwise return 0.
767static unsigned unrollCountPragmaValue(const Loop *L) {
768 MDNode *MD = getUnrollMetadataForLoop(L, Name: "llvm.loop.unroll.count");
769 if (MD) {
770 assert(MD->getNumOperands() == 2 &&
771 "Unroll count hint metadata should have two operands.");
772 unsigned Count =
773 mdconst::extract<ConstantInt>(MD: MD->getOperand(I: 1))->getZExtValue();
774 assert(Count >= 1 && "Unroll count must be positive.");
775 return Count;
776 }
777 return 0;
778}
779
780// Computes the boosting factor for complete unrolling.
781// If fully unrolling the loop would save a lot of RolledDynamicCost, it would
782// be beneficial to fully unroll the loop even if unrolledcost is large. We
783// use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust
784// the unroll threshold.
785static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost,
786 unsigned MaxPercentThresholdBoost) {
787 if (Cost.RolledDynamicCost >= std::numeric_limits<unsigned>::max() / 100)
788 return 100;
789 else if (Cost.UnrolledCost != 0)
790 // The boosting factor is RolledDynamicCost / UnrolledCost
791 return std::min(a: 100 * Cost.RolledDynamicCost / Cost.UnrolledCost,
792 b: MaxPercentThresholdBoost);
793 else
794 return MaxPercentThresholdBoost;
795}
796
797static std::optional<unsigned>
798shouldPragmaUnroll(Loop *L, const PragmaInfo &PInfo,
799 const unsigned TripMultiple, const unsigned TripCount,
800 unsigned MaxTripCount, const UnrollCostEstimator UCE,
801 const TargetTransformInfo::UnrollingPreferences &UP) {
802
803 // Using unroll pragma
804 // 1st priority is unroll count set by "unroll-count" option.
805
806 if (PInfo.UserUnrollCount) {
807 if (UP.AllowRemainder &&
808 UCE.getUnrolledLoopSize(UP, CountOverwrite: (unsigned)UnrollCount) < UP.Threshold)
809 return (unsigned)UnrollCount;
810 }
811
812 // 2nd priority is unroll count set by pragma.
813 if (PInfo.PragmaCount > 0) {
814 if ((UP.AllowRemainder || (TripMultiple % PInfo.PragmaCount == 0)))
815 return PInfo.PragmaCount;
816 }
817
818 if (PInfo.PragmaFullUnroll && TripCount != 0) {
819 // Certain cases with UBSAN can cause trip count to be calculated as
820 // INT_MAX, Block full unrolling at a reasonable limit so that the compiler
821 // doesn't hang trying to unroll the loop. See PR77842
822 if (TripCount > PragmaUnrollFullMaxIterations) {
823 LLVM_DEBUG(dbgs() << "Won't unroll; trip count is too large\n");
824 return std::nullopt;
825 }
826
827 return TripCount;
828 }
829
830 if (PInfo.PragmaEnableUnroll && !TripCount && MaxTripCount &&
831 MaxTripCount <= UP.MaxUpperBound)
832 return MaxTripCount;
833
834 return std::nullopt;
835}
836
837static std::optional<unsigned> shouldFullUnroll(
838 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT,
839 ScalarEvolution &SE, const SmallPtrSetImpl<const Value *> &EphValues,
840 const unsigned FullUnrollTripCount, const UnrollCostEstimator UCE,
841 const TargetTransformInfo::UnrollingPreferences &UP) {
842 assert(FullUnrollTripCount && "should be non-zero!");
843
844 if (FullUnrollTripCount > UP.FullUnrollMaxCount)
845 return std::nullopt;
846
847 // When computing the unrolled size, note that BEInsns are not replicated
848 // like the rest of the loop body.
849 if (UCE.getUnrolledLoopSize(UP) < UP.Threshold)
850 return FullUnrollTripCount;
851
852 // The loop isn't that small, but we still can fully unroll it if that
853 // helps to remove a significant number of instructions.
854 // To check that, run additional analysis on the loop.
855 if (std::optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
856 L, TripCount: FullUnrollTripCount, DT, SE, EphValues, TTI,
857 MaxUnrolledLoopSize: UP.Threshold * UP.MaxPercentThresholdBoost / 100,
858 MaxIterationsCountToAnalyze: UP.MaxIterationsCountToAnalyze)) {
859 unsigned Boost =
860 getFullUnrollBoostingFactor(Cost: *Cost, MaxPercentThresholdBoost: UP.MaxPercentThresholdBoost);
861 if (Cost->UnrolledCost < UP.Threshold * Boost / 100)
862 return FullUnrollTripCount;
863 }
864 return std::nullopt;
865}
866
867static std::optional<unsigned>
868shouldPartialUnroll(const unsigned LoopSize, const unsigned TripCount,
869 const UnrollCostEstimator UCE,
870 const TargetTransformInfo::UnrollingPreferences &UP) {
871
872 if (!TripCount)
873 return std::nullopt;
874
875 if (!UP.Partial) {
876 LLVM_DEBUG(dbgs() << " will not try to unroll partially because "
877 << "-unroll-allow-partial not given\n");
878 return 0;
879 }
880 unsigned count = UP.Count;
881 if (count == 0)
882 count = TripCount;
883 if (UP.PartialThreshold != NoThreshold) {
884 // Reduce unroll count to be modulo of TripCount for partial unrolling.
885 if (UCE.getUnrolledLoopSize(UP, CountOverwrite: count) > UP.PartialThreshold)
886 count = (std::max(a: UP.PartialThreshold, b: UP.BEInsns + 1) - UP.BEInsns) /
887 (LoopSize - UP.BEInsns);
888 if (count > UP.MaxCount)
889 count = UP.MaxCount;
890 while (count != 0 && TripCount % count != 0)
891 count--;
892 if (UP.AllowRemainder && count <= 1) {
893 // If there is no Count that is modulo of TripCount, set Count to
894 // largest power-of-two factor that satisfies the threshold limit.
895 // As we'll create fixup loop, do the type of unrolling only if
896 // remainder loop is allowed.
897 // Note: DefaultUnrollRuntimeCount is used as a reasonable starting point
898 // even though this is partial unrolling (not runtime unrolling).
899 count = UP.DefaultUnrollRuntimeCount;
900 while (count != 0 &&
901 UCE.getUnrolledLoopSize(UP, CountOverwrite: count) > UP.PartialThreshold)
902 count >>= 1;
903 }
904 if (count < 2) {
905 count = 0;
906 }
907 } else {
908 count = TripCount;
909 }
910 if (count > UP.MaxCount)
911 count = UP.MaxCount;
912
913 LLVM_DEBUG(dbgs() << " partially unrolling with count: " << count << "\n");
914
915 return count;
916}
917// Returns true if unroll count was set explicitly.
918// Calculates unroll count and writes it to UP.Count.
919// Unless IgnoreUser is true, will also use metadata and command-line options
920// that are specific to the LoopUnroll pass (which, for instance, are
921// irrelevant for the LoopUnrollAndJam pass).
922// FIXME: This function is used by LoopUnroll and LoopUnrollAndJam, but consumes
923// many LoopUnroll-specific options. The shared functionality should be
924// refactored into it own function.
925bool llvm::computeUnrollCount(
926 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI,
927 AssumptionCache *AC, ScalarEvolution &SE,
928 const SmallPtrSetImpl<const Value *> &EphValues,
929 OptimizationRemarkEmitter *ORE, unsigned TripCount, unsigned MaxTripCount,
930 bool MaxOrZero, unsigned TripMultiple, const UnrollCostEstimator &UCE,
931 TargetTransformInfo::UnrollingPreferences &UP,
932 TargetTransformInfo::PeelingPreferences &PP, bool &UseUpperBound) {
933
934 unsigned LoopSize = UCE.getRolledLoopSize();
935
936 const bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0;
937 const bool PragmaFullUnroll = hasUnrollFullPragma(L);
938 const unsigned PragmaCount = unrollCountPragmaValue(L);
939 const bool PragmaEnableUnroll = hasUnrollEnablePragma(L);
940
941 const bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll ||
942 PragmaEnableUnroll || UserUnrollCount;
943
944 PragmaInfo PInfo(UserUnrollCount, PragmaFullUnroll, PragmaCount,
945 PragmaEnableUnroll);
946 // Use an explicit peel count that has been specified for testing. In this
947 // case it's not permitted to also specify an explicit unroll count.
948 if (PP.PeelCount) {
949 if (UnrollCount.getNumOccurrences() > 0) {
950 reportFatalUsageError(reason: "Cannot specify both explicit peel count and "
951 "explicit unroll count");
952 }
953 UP.Count = 1;
954 UP.Runtime = false;
955 return true;
956 }
957 // Check for explicit Count.
958 // 1st priority is unroll count set by "unroll-count" option.
959 // 2nd priority is unroll count set by pragma.
960 if (auto UnrollFactor = shouldPragmaUnroll(L, PInfo, TripMultiple, TripCount,
961 MaxTripCount, UCE, UP)) {
962 UP.Count = *UnrollFactor;
963
964 if (UserUnrollCount || (PragmaCount > 0)) {
965 UP.AllowExpensiveTripCount = true;
966 UP.Force = true;
967 }
968 UP.Runtime |= (PragmaCount > 0);
969 return ExplicitUnroll;
970 } else {
971 if (ExplicitUnroll && TripCount != 0) {
972 // If the loop has an unrolling pragma, we want to be more aggressive with
973 // unrolling limits. Set thresholds to at least the PragmaUnrollThreshold
974 // value which is larger than the default limits.
975 UP.Threshold = std::max<unsigned>(a: UP.Threshold, b: PragmaUnrollThreshold);
976 UP.PartialThreshold =
977 std::max<unsigned>(a: UP.PartialThreshold, b: PragmaUnrollThreshold);
978 }
979 }
980
981 // 3rd priority is exact full unrolling. This will eliminate all copies
982 // of some exit test.
983 UP.Count = 0;
984 if (TripCount) {
985 UP.Count = TripCount;
986 if (auto UnrollFactor = shouldFullUnroll(L, TTI, DT, SE, EphValues,
987 FullUnrollTripCount: TripCount, UCE, UP)) {
988 UP.Count = *UnrollFactor;
989 UseUpperBound = false;
990 return ExplicitUnroll;
991 }
992 }
993
994 // 4th priority is bounded unrolling.
995 // We can unroll by the upper bound amount if it's generally allowed or if
996 // we know that the loop is executed either the upper bound or zero times.
997 // (MaxOrZero unrolling keeps only the first loop test, so the number of
998 // loop tests remains the same compared to the non-unrolled version, whereas
999 // the generic upper bound unrolling keeps all but the last loop test so the
1000 // number of loop tests goes up which may end up being worse on targets with
1001 // constrained branch predictor resources so is controlled by an option.)
1002 // In addition we only unroll small upper bounds.
1003 // Note that the cost of bounded unrolling is always strictly greater than
1004 // cost of exact full unrolling. As such, if we have an exact count and
1005 // found it unprofitable, we'll never chose to bounded unroll.
1006 if (!TripCount && MaxTripCount && (UP.UpperBound || MaxOrZero) &&
1007 MaxTripCount <= UP.MaxUpperBound) {
1008 UP.Count = MaxTripCount;
1009 if (auto UnrollFactor = shouldFullUnroll(L, TTI, DT, SE, EphValues,
1010 FullUnrollTripCount: MaxTripCount, UCE, UP)) {
1011 UP.Count = *UnrollFactor;
1012 UseUpperBound = true;
1013 return ExplicitUnroll;
1014 }
1015 }
1016
1017 // 5th priority is loop peeling.
1018 computePeelCount(L, LoopSize, PP, TripCount, DT, SE, TTI, AC, Threshold: UP.Threshold);
1019 if (PP.PeelCount) {
1020 UP.Runtime = false;
1021 UP.Count = 1;
1022 return ExplicitUnroll;
1023 }
1024
1025 // Before starting partial unrolling, set up.partial to true,
1026 // if user explicitly asked for unrolling
1027 if (TripCount)
1028 UP.Partial |= ExplicitUnroll;
1029
1030 // 6th priority is partial unrolling.
1031 // Try partial unroll only when TripCount could be statically calculated.
1032 if (auto UnrollFactor = shouldPartialUnroll(LoopSize, TripCount, UCE, UP)) {
1033 UP.Count = *UnrollFactor;
1034
1035 if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
1036 UP.Count != TripCount)
1037 ORE->emit(RemarkBuilder: [&]() {
1038 return OptimizationRemarkMissed(DEBUG_TYPE,
1039 "FullUnrollAsDirectedTooLarge",
1040 L->getStartLoc(), L->getHeader())
1041 << "Unable to fully unroll loop as directed by unroll pragma "
1042 "because "
1043 "unrolled size is too large.";
1044 });
1045
1046 if (UP.PartialThreshold != NoThreshold) {
1047 if (UP.Count == 0) {
1048 if (PragmaEnableUnroll)
1049 ORE->emit(RemarkBuilder: [&]() {
1050 return OptimizationRemarkMissed(DEBUG_TYPE,
1051 "UnrollAsDirectedTooLarge",
1052 L->getStartLoc(), L->getHeader())
1053 << "Unable to unroll loop as directed by unroll(enable) "
1054 "pragma "
1055 "because unrolled size is too large.";
1056 });
1057 }
1058 }
1059 return ExplicitUnroll;
1060 }
1061 assert(TripCount == 0 &&
1062 "All cases when TripCount is constant should be covered here.");
1063 if (PragmaFullUnroll)
1064 ORE->emit(RemarkBuilder: [&]() {
1065 return OptimizationRemarkMissed(
1066 DEBUG_TYPE, "CantFullUnrollAsDirectedRuntimeTripCount",
1067 L->getStartLoc(), L->getHeader())
1068 << "Unable to fully unroll loop as directed by unroll(full) "
1069 "pragma "
1070 "because loop has a runtime trip count.";
1071 });
1072
1073 // 7th priority is runtime unrolling.
1074 // Don't unroll a runtime trip count loop when it is disabled.
1075 if (hasRuntimeUnrollDisablePragma(L)) {
1076 UP.Count = 0;
1077 return false;
1078 }
1079
1080 // Don't unroll a small upper bound loop unless user or TTI asked to do so.
1081 if (MaxTripCount && !UP.Force && MaxTripCount < UP.MaxUpperBound) {
1082 UP.Count = 0;
1083 return false;
1084 }
1085
1086 // Check if the runtime trip count is too small when profile is available.
1087 if (L->getHeader()->getParent()->hasProfileData()) {
1088 if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) {
1089 if (*ProfileTripCount < FlatLoopTripCountThreshold)
1090 return false;
1091 else
1092 UP.AllowExpensiveTripCount = true;
1093 }
1094 }
1095 UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount;
1096 if (!UP.Runtime) {
1097 LLVM_DEBUG(
1098 dbgs() << " will not try to unroll loop with runtime trip count "
1099 << "-unroll-runtime not given\n");
1100 UP.Count = 0;
1101 return false;
1102 }
1103 if (UP.Count == 0)
1104 UP.Count = UP.DefaultUnrollRuntimeCount;
1105
1106 // Reduce unroll count to be the largest power-of-two factor of
1107 // the original count which satisfies the threshold limit.
1108 while (UP.Count != 0 &&
1109 UCE.getUnrolledLoopSize(UP) > UP.PartialThreshold)
1110 UP.Count >>= 1;
1111
1112#ifndef NDEBUG
1113 unsigned OrigCount = UP.Count;
1114#endif
1115
1116 if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) {
1117 while (UP.Count != 0 && TripMultiple % UP.Count != 0)
1118 UP.Count >>= 1;
1119 LLVM_DEBUG(
1120 dbgs() << "Remainder loop is restricted (that could architecture "
1121 "specific or because the loop contains a convergent "
1122 "instruction), so unroll count must divide the trip "
1123 "multiple, "
1124 << TripMultiple << ". Reducing unroll count from " << OrigCount
1125 << " to " << UP.Count << ".\n");
1126
1127 using namespace ore;
1128
1129 if (unrollCountPragmaValue(L) > 0 && !UP.AllowRemainder)
1130 ORE->emit(RemarkBuilder: [&]() {
1131 return OptimizationRemarkMissed(DEBUG_TYPE,
1132 "DifferentUnrollCountFromDirected",
1133 L->getStartLoc(), L->getHeader())
1134 << "Unable to unroll loop the number of times directed by "
1135 "unroll_count pragma because remainder loop is restricted "
1136 "(that could architecture specific or because the loop "
1137 "contains a convergent instruction) and so must have an "
1138 "unroll "
1139 "count that divides the loop trip multiple of "
1140 << NV("TripMultiple", TripMultiple) << ". Unrolling instead "
1141 << NV("UnrollCount", UP.Count) << " time(s).";
1142 });
1143 }
1144
1145 if (UP.Count > UP.MaxCount)
1146 UP.Count = UP.MaxCount;
1147
1148 if (MaxTripCount && UP.Count > MaxTripCount)
1149 UP.Count = MaxTripCount;
1150
1151 LLVM_DEBUG(dbgs() << " runtime unrolling with count: " << UP.Count
1152 << "\n");
1153 if (UP.Count < 2)
1154 UP.Count = 0;
1155 return ExplicitUnroll;
1156}
1157
1158static LoopUnrollResult
1159tryToUnrollLoop(Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE,
1160 const TargetTransformInfo &TTI, AssumptionCache &AC,
1161 OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
1162 ProfileSummaryInfo *PSI, bool PreserveLCSSA, int OptLevel,
1163 bool OnlyFullUnroll, bool OnlyWhenForced, bool ForgetAllSCEV,
1164 std::optional<unsigned> ProvidedCount,
1165 std::optional<unsigned> ProvidedThreshold,
1166 std::optional<bool> ProvidedAllowPartial,
1167 std::optional<bool> ProvidedRuntime,
1168 std::optional<bool> ProvidedUpperBound,
1169 std::optional<bool> ProvidedAllowPeeling,
1170 std::optional<bool> ProvidedAllowProfileBasedPeeling,
1171 std::optional<unsigned> ProvidedFullUnrollMaxCount,
1172 AAResults *AA = nullptr) {
1173
1174 LLVM_DEBUG(dbgs() << "Loop Unroll: F["
1175 << L->getHeader()->getParent()->getName() << "] Loop %"
1176 << L->getHeader()->getName() << "\n");
1177 TransformationMode TM = hasUnrollTransformation(L);
1178 if (TM & TM_Disable)
1179 return LoopUnrollResult::Unmodified;
1180
1181 // If this loop isn't forced to be unrolled, avoid unrolling it when the
1182 // parent loop has an explicit unroll-and-jam pragma. This is to prevent
1183 // automatic unrolling from interfering with the user requested
1184 // transformation.
1185 Loop *ParentL = L->getParentLoop();
1186 if (ParentL != nullptr &&
1187 hasUnrollAndJamTransformation(L: ParentL) == TM_ForcedByUser &&
1188 hasUnrollTransformation(L) != TM_ForcedByUser) {
1189 LLVM_DEBUG(dbgs() << "Not unrolling loop since parent loop has"
1190 << " llvm.loop.unroll_and_jam.\n");
1191 return LoopUnrollResult::Unmodified;
1192 }
1193
1194 // If this loop isn't forced to be unrolled, avoid unrolling it when the
1195 // loop has an explicit unroll-and-jam pragma. This is to prevent automatic
1196 // unrolling from interfering with the user requested transformation.
1197 if (hasUnrollAndJamTransformation(L) == TM_ForcedByUser &&
1198 hasUnrollTransformation(L) != TM_ForcedByUser) {
1199 LLVM_DEBUG(
1200 dbgs()
1201 << " Not unrolling loop since it has llvm.loop.unroll_and_jam.\n");
1202 return LoopUnrollResult::Unmodified;
1203 }
1204
1205 if (!L->isLoopSimplifyForm()) {
1206 LLVM_DEBUG(
1207 dbgs() << " Not unrolling loop which is not in loop-simplify form.\n");
1208 return LoopUnrollResult::Unmodified;
1209 }
1210
1211 // When automatic unrolling is disabled, do not unroll unless overridden for
1212 // this loop.
1213 if (OnlyWhenForced && !(TM & TM_Enable))
1214 return LoopUnrollResult::Unmodified;
1215
1216 bool OptForSize = L->getHeader()->getParent()->hasOptSize();
1217 TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
1218 L, SE, TTI, BFI, PSI, ORE, OptLevel, UserThreshold: ProvidedThreshold, UserCount: ProvidedCount,
1219 UserAllowPartial: ProvidedAllowPartial, UserRuntime: ProvidedRuntime, UserUpperBound: ProvidedUpperBound,
1220 UserFullUnrollMaxCount: ProvidedFullUnrollMaxCount);
1221 TargetTransformInfo::PeelingPreferences PP = gatherPeelingPreferences(
1222 L, SE, TTI, UserAllowPeeling: ProvidedAllowPeeling, UserAllowProfileBasedPeeling: ProvidedAllowProfileBasedPeeling, UnrollingSpecficValues: true);
1223
1224 // Exit early if unrolling is disabled. For OptForSize, we pick the loop size
1225 // as threshold later on.
1226 if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0) &&
1227 !OptForSize)
1228 return LoopUnrollResult::Unmodified;
1229
1230 SmallPtrSet<const Value *, 32> EphValues;
1231 CodeMetrics::collectEphemeralValues(L, AC: &AC, EphValues);
1232
1233 UnrollCostEstimator UCE(L, TTI, EphValues, UP.BEInsns);
1234 if (!UCE.canUnroll()) {
1235 LLVM_DEBUG(dbgs() << " Loop not considered unrollable.\n");
1236 return LoopUnrollResult::Unmodified;
1237 }
1238
1239 unsigned LoopSize = UCE.getRolledLoopSize();
1240 LLVM_DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
1241
1242 // When optimizing for size, use LoopSize + 1 as threshold (we use < Threshold
1243 // later), to (fully) unroll loops, if it does not increase code size.
1244 if (OptForSize)
1245 UP.Threshold = std::max(a: UP.Threshold, b: LoopSize + 1);
1246
1247 if (UCE.NumInlineCandidates != 0) {
1248 LLVM_DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
1249 return LoopUnrollResult::Unmodified;
1250 }
1251
1252 // Find the smallest exact trip count for any exit. This is an upper bound
1253 // on the loop trip count, but an exit at an earlier iteration is still
1254 // possible. An unroll by the smallest exact trip count guarantees that all
1255 // branches relating to at least one exit can be eliminated. This is unlike
1256 // the max trip count, which only guarantees that the backedge can be broken.
1257 unsigned TripCount = 0;
1258 unsigned TripMultiple = 1;
1259 SmallVector<BasicBlock *, 8> ExitingBlocks;
1260 L->getExitingBlocks(ExitingBlocks);
1261 for (BasicBlock *ExitingBlock : ExitingBlocks)
1262 if (unsigned TC = SE.getSmallConstantTripCount(L, ExitingBlock))
1263 if (!TripCount || TC < TripCount)
1264 TripCount = TripMultiple = TC;
1265
1266 if (!TripCount) {
1267 // If no exact trip count is known, determine the trip multiple of either
1268 // the loop latch or the single exiting block.
1269 // TODO: Relax for multiple exits.
1270 BasicBlock *ExitingBlock = L->getLoopLatch();
1271 if (!ExitingBlock || !L->isLoopExiting(BB: ExitingBlock))
1272 ExitingBlock = L->getExitingBlock();
1273 if (ExitingBlock)
1274 TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock);
1275 }
1276
1277 // If the loop contains a convergent operation, the prelude we'd add
1278 // to do the first few instructions before we hit the unrolled loop
1279 // is unsafe -- it adds a control-flow dependency to the convergent
1280 // operation. Therefore restrict remainder loop (try unrolling without).
1281 //
1282 // TODO: This is somewhat conservative; we could allow the remainder if the
1283 // trip count is uniform.
1284 UP.AllowRemainder &= UCE.ConvergenceAllowsRuntime;
1285
1286 // Try to find the trip count upper bound if we cannot find the exact trip
1287 // count.
1288 unsigned MaxTripCount = 0;
1289 bool MaxOrZero = false;
1290 if (!TripCount) {
1291 MaxTripCount = SE.getSmallConstantMaxTripCount(L);
1292 MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L);
1293 }
1294
1295 // computeUnrollCount() decides whether it is beneficial to use upper bound to
1296 // fully unroll the loop.
1297 bool UseUpperBound = false;
1298 bool IsCountSetExplicitly = computeUnrollCount(
1299 L, TTI, DT, LI, AC: &AC, SE, EphValues, ORE: &ORE, TripCount, MaxTripCount,
1300 MaxOrZero, TripMultiple, UCE, UP, PP, UseUpperBound);
1301 if (!UP.Count)
1302 return LoopUnrollResult::Unmodified;
1303
1304 UP.Runtime &= UCE.ConvergenceAllowsRuntime;
1305
1306 if (PP.PeelCount) {
1307 assert(UP.Count == 1 && "Cannot perform peel and unroll in the same step");
1308 LLVM_DEBUG(dbgs() << "PEELING loop %" << L->getHeader()->getName()
1309 << " with iteration count " << PP.PeelCount << "!\n");
1310 ORE.emit(RemarkBuilder: [&]() {
1311 return OptimizationRemark(DEBUG_TYPE, "Peeled", L->getStartLoc(),
1312 L->getHeader())
1313 << "peeled loop by " << ore::NV("PeelCount", PP.PeelCount)
1314 << " iterations";
1315 });
1316
1317 ValueToValueMapTy VMap;
1318 peelLoop(L, PeelCount: PP.PeelCount, PeelLast: PP.PeelLast, LI, SE: &SE, DT, AC: &AC, PreserveLCSSA,
1319 VMap);
1320 simplifyLoopAfterUnroll(L, SimplifyIVs: true, LI, SE: &SE, DT: &DT, AC: &AC, TTI: &TTI, AA: nullptr);
1321 // If the loop was peeled, we already "used up" the profile information
1322 // we had, so we don't want to unroll or peel again.
1323 if (PP.PeelProfiledIterations)
1324 L->setLoopAlreadyUnrolled();
1325 return LoopUnrollResult::PartiallyUnrolled;
1326 }
1327
1328 // Do not attempt partial/runtime unrolling in FullLoopUnrolling
1329 if (OnlyFullUnroll && ((!TripCount && !MaxTripCount) ||
1330 UP.Count < TripCount || UP.Count < MaxTripCount)) {
1331 LLVM_DEBUG(
1332 dbgs() << "Not attempting partial/runtime unroll in FullLoopUnroll.\n");
1333 return LoopUnrollResult::Unmodified;
1334 }
1335
1336 // At this point, UP.Runtime indicates that run-time unrolling is allowed.
1337 // However, we only want to actually perform it if we don't know the trip
1338 // count and the unroll count doesn't divide the known trip multiple.
1339 // TODO: This decision should probably be pushed up into
1340 // computeUnrollCount().
1341 UP.Runtime &= TripCount == 0 && TripMultiple % UP.Count != 0;
1342
1343 // Save loop properties before it is transformed.
1344 MDNode *OrigLoopID = L->getLoopID();
1345
1346 // Unroll the loop.
1347 Loop *RemainderLoop = nullptr;
1348 UnrollLoopOptions ULO;
1349 ULO.Count = UP.Count;
1350 ULO.Force = UP.Force;
1351 ULO.AllowExpensiveTripCount = UP.AllowExpensiveTripCount;
1352 ULO.UnrollRemainder = UP.UnrollRemainder;
1353 ULO.Runtime = UP.Runtime;
1354 ULO.ForgetAllSCEV = ForgetAllSCEV;
1355 ULO.Heart = getLoopConvergenceHeart(TheLoop: L);
1356 ULO.SCEVExpansionBudget = UP.SCEVExpansionBudget;
1357 ULO.RuntimeUnrollMultiExit = UP.RuntimeUnrollMultiExit;
1358 ULO.AddAdditionalAccumulators = UP.AddAdditionalAccumulators;
1359 LoopUnrollResult UnrollResult = UnrollLoop(
1360 L, ULO, LI, SE: &SE, DT: &DT, AC: &AC, TTI: &TTI, ORE: &ORE, PreserveLCSSA, RemainderLoop: &RemainderLoop, AA);
1361 if (UnrollResult == LoopUnrollResult::Unmodified)
1362 return LoopUnrollResult::Unmodified;
1363
1364 if (RemainderLoop) {
1365 std::optional<MDNode *> RemainderLoopID =
1366 makeFollowupLoopID(OrigLoopID, FollowupAttrs: {LLVMLoopUnrollFollowupAll,
1367 LLVMLoopUnrollFollowupRemainder});
1368 if (RemainderLoopID)
1369 RemainderLoop->setLoopID(*RemainderLoopID);
1370 }
1371
1372 if (UnrollResult != LoopUnrollResult::FullyUnrolled) {
1373 std::optional<MDNode *> NewLoopID =
1374 makeFollowupLoopID(OrigLoopID, FollowupAttrs: {LLVMLoopUnrollFollowupAll,
1375 LLVMLoopUnrollFollowupUnrolled});
1376 if (NewLoopID) {
1377 L->setLoopID(*NewLoopID);
1378
1379 // Do not setLoopAlreadyUnrolled if loop attributes have been specified
1380 // explicitly.
1381 return UnrollResult;
1382 }
1383 }
1384
1385 // If loop has an unroll count pragma or unrolled by explicitly set count
1386 // mark loop as unrolled to prevent unrolling beyond that requested.
1387 if (UnrollResult != LoopUnrollResult::FullyUnrolled && IsCountSetExplicitly)
1388 L->setLoopAlreadyUnrolled();
1389
1390 return UnrollResult;
1391}
1392
1393namespace {
1394
1395class LoopUnroll : public LoopPass {
1396public:
1397 static char ID; // Pass ID, replacement for typeid
1398
1399 int OptLevel;
1400
1401 /// If false, use a cost model to determine whether unrolling of a loop is
1402 /// profitable. If true, only loops that explicitly request unrolling via
1403 /// metadata are considered. All other loops are skipped.
1404 bool OnlyWhenForced;
1405
1406 /// If false, when SCEV is invalidated, only forget everything in the
1407 /// top-most loop (call forgetTopMostLoop), of the loop being processed.
1408 /// Otherwise, forgetAllLoops and rebuild when needed next.
1409 bool ForgetAllSCEV;
1410
1411 std::optional<unsigned> ProvidedCount;
1412 std::optional<unsigned> ProvidedThreshold;
1413 std::optional<bool> ProvidedAllowPartial;
1414 std::optional<bool> ProvidedRuntime;
1415 std::optional<bool> ProvidedUpperBound;
1416 std::optional<bool> ProvidedAllowPeeling;
1417 std::optional<bool> ProvidedAllowProfileBasedPeeling;
1418 std::optional<unsigned> ProvidedFullUnrollMaxCount;
1419
1420 LoopUnroll(int OptLevel = 2, bool OnlyWhenForced = false,
1421 bool ForgetAllSCEV = false,
1422 std::optional<unsigned> Threshold = std::nullopt,
1423 std::optional<unsigned> Count = std::nullopt,
1424 std::optional<bool> AllowPartial = std::nullopt,
1425 std::optional<bool> Runtime = std::nullopt,
1426 std::optional<bool> UpperBound = std::nullopt,
1427 std::optional<bool> AllowPeeling = std::nullopt,
1428 std::optional<bool> AllowProfileBasedPeeling = std::nullopt,
1429 std::optional<unsigned> ProvidedFullUnrollMaxCount = std::nullopt)
1430 : LoopPass(ID), OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
1431 ForgetAllSCEV(ForgetAllSCEV), ProvidedCount(std::move(Count)),
1432 ProvidedThreshold(Threshold), ProvidedAllowPartial(AllowPartial),
1433 ProvidedRuntime(Runtime), ProvidedUpperBound(UpperBound),
1434 ProvidedAllowPeeling(AllowPeeling),
1435 ProvidedAllowProfileBasedPeeling(AllowProfileBasedPeeling),
1436 ProvidedFullUnrollMaxCount(ProvidedFullUnrollMaxCount) {
1437 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
1438 }
1439
1440 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1441 if (skipLoop(L))
1442 return false;
1443
1444 Function &F = *L->getHeader()->getParent();
1445
1446 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1447 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1448 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1449 const TargetTransformInfo &TTI =
1450 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1451 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1452 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
1453 // pass. Function analyses need to be preserved across loop transformations
1454 // but ORE cannot be preserved (see comment before the pass definition).
1455 OptimizationRemarkEmitter ORE(&F);
1456 bool PreserveLCSSA = mustPreserveAnalysisID(AID&: LCSSAID);
1457
1458 LoopUnrollResult Result = tryToUnrollLoop(
1459 L, DT, LI, SE, TTI, AC, ORE, BFI: nullptr, PSI: nullptr, PreserveLCSSA, OptLevel,
1460 /*OnlyFullUnroll*/ false, OnlyWhenForced, ForgetAllSCEV, ProvidedCount,
1461 ProvidedThreshold, ProvidedAllowPartial, ProvidedRuntime,
1462 ProvidedUpperBound, ProvidedAllowPeeling,
1463 ProvidedAllowProfileBasedPeeling, ProvidedFullUnrollMaxCount);
1464
1465 if (Result == LoopUnrollResult::FullyUnrolled)
1466 LPM.markLoopAsDeleted(L&: *L);
1467
1468 return Result != LoopUnrollResult::Unmodified;
1469 }
1470
1471 /// This transformation requires natural loop information & requires that
1472 /// loop preheaders be inserted into the CFG...
1473 void getAnalysisUsage(AnalysisUsage &AU) const override {
1474 AU.addRequired<AssumptionCacheTracker>();
1475 AU.addRequired<TargetTransformInfoWrapperPass>();
1476 // FIXME: Loop passes are required to preserve domtree, and for now we just
1477 // recreate dom info if anything gets unrolled.
1478 getLoopAnalysisUsage(AU);
1479 }
1480};
1481
1482} // end anonymous namespace
1483
1484char LoopUnroll::ID = 0;
1485
1486INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1487INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1488INITIALIZE_PASS_DEPENDENCY(LoopPass)
1489INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1490INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1491
1492Pass *llvm::createLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1493 bool ForgetAllSCEV, int Threshold, int Count,
1494 int AllowPartial, int Runtime, int UpperBound,
1495 int AllowPeeling) {
1496 // TODO: It would make more sense for this function to take the optionals
1497 // directly, but that's dangerous since it would silently break out of tree
1498 // callers.
1499 return new LoopUnroll(
1500 OptLevel, OnlyWhenForced, ForgetAllSCEV,
1501 Threshold == -1 ? std::nullopt : std::optional<unsigned>(Threshold),
1502 Count == -1 ? std::nullopt : std::optional<unsigned>(Count),
1503 AllowPartial == -1 ? std::nullopt : std::optional<bool>(AllowPartial),
1504 Runtime == -1 ? std::nullopt : std::optional<bool>(Runtime),
1505 UpperBound == -1 ? std::nullopt : std::optional<bool>(UpperBound),
1506 AllowPeeling == -1 ? std::nullopt : std::optional<bool>(AllowPeeling));
1507}
1508
1509PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
1510 LoopStandardAnalysisResults &AR,
1511 LPMUpdater &Updater) {
1512 // For the new PM, we can't use OptimizationRemarkEmitter as an analysis
1513 // pass. Function analyses need to be preserved across loop transformations
1514 // but ORE cannot be preserved (see comment before the pass definition).
1515 OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
1516
1517 // Keep track of the previous loop structure so we can identify new loops
1518 // created by unrolling.
1519 Loop *ParentL = L.getParentLoop();
1520 SmallPtrSet<Loop *, 4> OldLoops;
1521 if (ParentL)
1522 OldLoops.insert_range(R&: *ParentL);
1523 else
1524 OldLoops.insert_range(R&: AR.LI);
1525
1526 std::string LoopName = std::string(L.getName());
1527
1528 bool Changed =
1529 tryToUnrollLoop(L: &L, DT&: AR.DT, LI: &AR.LI, SE&: AR.SE, TTI: AR.TTI, AC&: AR.AC, ORE,
1530 /*BFI*/ nullptr, /*PSI*/ nullptr,
1531 /*PreserveLCSSA*/ true, OptLevel, /*OnlyFullUnroll*/ true,
1532 OnlyWhenForced, ForgetAllSCEV: ForgetSCEV, /*Count*/ ProvidedCount: std::nullopt,
1533 /*Threshold*/ ProvidedThreshold: std::nullopt, /*AllowPartial*/ ProvidedAllowPartial: false,
1534 /*Runtime*/ ProvidedRuntime: false, /*UpperBound*/ ProvidedUpperBound: false,
1535 /*AllowPeeling*/ ProvidedAllowPeeling: true,
1536 /*AllowProfileBasedPeeling*/ ProvidedAllowProfileBasedPeeling: false,
1537 /*FullUnrollMaxCount*/ ProvidedFullUnrollMaxCount: std::nullopt) !=
1538 LoopUnrollResult::Unmodified;
1539 if (!Changed)
1540 return PreservedAnalyses::all();
1541
1542 // The parent must not be damaged by unrolling!
1543#ifndef NDEBUG
1544 if (ParentL)
1545 ParentL->verifyLoop();
1546#endif
1547
1548 // Unrolling can do several things to introduce new loops into a loop nest:
1549 // - Full unrolling clones child loops within the current loop but then
1550 // removes the current loop making all of the children appear to be new
1551 // sibling loops.
1552 //
1553 // When a new loop appears as a sibling loop after fully unrolling,
1554 // its nesting structure has fundamentally changed and we want to revisit
1555 // it to reflect that.
1556 //
1557 // When unrolling has removed the current loop, we need to tell the
1558 // infrastructure that it is gone.
1559 //
1560 // Finally, we support a debugging/testing mode where we revisit child loops
1561 // as well. These are not expected to require further optimizations as either
1562 // they or the loop they were cloned from have been directly visited already.
1563 // But the debugging mode allows us to check this assumption.
1564 bool IsCurrentLoopValid = false;
1565 SmallVector<Loop *, 4> SibLoops;
1566 if (ParentL)
1567 SibLoops.append(in_start: ParentL->begin(), in_end: ParentL->end());
1568 else
1569 SibLoops.append(in_start: AR.LI.begin(), in_end: AR.LI.end());
1570 erase_if(C&: SibLoops, P: [&](Loop *SibLoop) {
1571 if (SibLoop == &L) {
1572 IsCurrentLoopValid = true;
1573 return true;
1574 }
1575
1576 // Otherwise erase the loop from the list if it was in the old loops.
1577 return OldLoops.contains(Ptr: SibLoop);
1578 });
1579 Updater.addSiblingLoops(NewSibLoops: SibLoops);
1580
1581 if (!IsCurrentLoopValid) {
1582 Updater.markLoopAsDeleted(L, Name: LoopName);
1583 } else {
1584 // We can only walk child loops if the current loop remained valid.
1585 if (UnrollRevisitChildLoops) {
1586 // Walk *all* of the child loops.
1587 SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end());
1588 Updater.addChildLoops(NewChildLoops: ChildLoops);
1589 }
1590 }
1591
1592 return getLoopPassPreservedAnalyses();
1593}
1594
1595PreservedAnalyses LoopUnrollPass::run(Function &F,
1596 FunctionAnalysisManager &AM) {
1597 auto &LI = AM.getResult<LoopAnalysis>(IR&: F);
1598 // There are no loops in the function. Return before computing other expensive
1599 // analyses.
1600 if (LI.empty())
1601 return PreservedAnalyses::all();
1602 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(IR&: F);
1603 auto &TTI = AM.getResult<TargetIRAnalysis>(IR&: F);
1604 auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F);
1605 auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F);
1606 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F);
1607 AAResults &AA = AM.getResult<AAManager>(IR&: F);
1608
1609 LoopAnalysisManager *LAM = nullptr;
1610 if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(IR&: F))
1611 LAM = &LAMProxy->getManager();
1612
1613 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F);
1614 ProfileSummaryInfo *PSI =
1615 MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
1616 auto *BFI = (PSI && PSI->hasProfileSummary()) ?
1617 &AM.getResult<BlockFrequencyAnalysis>(IR&: F) : nullptr;
1618
1619 bool Changed = false;
1620
1621 // The unroller requires loops to be in simplified form, and also needs LCSSA.
1622 // Since simplification may add new inner loops, it has to run before the
1623 // legality and profitability checks. This means running the loop unroller
1624 // will simplify all loops, regardless of whether anything end up being
1625 // unrolled.
1626 for (const auto &L : LI) {
1627 Changed |=
1628 simplifyLoop(L, DT: &DT, LI: &LI, SE: &SE, AC: &AC, MSSAU: nullptr, PreserveLCSSA: false /* PreserveLCSSA */);
1629 Changed |= formLCSSARecursively(L&: *L, DT, LI: &LI, SE: &SE);
1630 }
1631
1632 // Add the loop nests in the reverse order of LoopInfo. See method
1633 // declaration.
1634 SmallPriorityWorklist<Loop *, 4> Worklist;
1635 appendLoopsToWorklist(LI, Worklist);
1636
1637 while (!Worklist.empty()) {
1638 // Because the LoopInfo stores the loops in RPO, we walk the worklist
1639 // from back to front so that we work forward across the CFG, which
1640 // for unrolling is only needed to get optimization remarks emitted in
1641 // a forward order.
1642 Loop &L = *Worklist.pop_back_val();
1643#ifndef NDEBUG
1644 Loop *ParentL = L.getParentLoop();
1645#endif
1646
1647 // Check if the profile summary indicates that the profiled application
1648 // has a huge working set size, in which case we disable peeling to avoid
1649 // bloating it further.
1650 std::optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling;
1651 if (PSI && PSI->hasHugeWorkingSetSize())
1652 LocalAllowPeeling = false;
1653 std::string LoopName = std::string(L.getName());
1654 // The API here is quite complex to call and we allow to select some
1655 // flavors of unrolling during construction time (by setting UnrollOpts).
1656 LoopUnrollResult Result = tryToUnrollLoop(
1657 L: &L, DT, LI: &LI, SE, TTI, AC, ORE, BFI, PSI,
1658 /*PreserveLCSSA*/ true, OptLevel: UnrollOpts.OptLevel, /*OnlyFullUnroll*/ false,
1659 OnlyWhenForced: UnrollOpts.OnlyWhenForced, ForgetAllSCEV: UnrollOpts.ForgetSCEV,
1660 /*Count*/ ProvidedCount: std::nullopt,
1661 /*Threshold*/ ProvidedThreshold: std::nullopt, ProvidedAllowPartial: UnrollOpts.AllowPartial,
1662 ProvidedRuntime: UnrollOpts.AllowRuntime, ProvidedUpperBound: UnrollOpts.AllowUpperBound, ProvidedAllowPeeling: LocalAllowPeeling,
1663 ProvidedAllowProfileBasedPeeling: UnrollOpts.AllowProfileBasedPeeling, ProvidedFullUnrollMaxCount: UnrollOpts.FullUnrollMaxCount,
1664 AA: &AA);
1665 Changed |= Result != LoopUnrollResult::Unmodified;
1666
1667 // The parent must not be damaged by unrolling!
1668#ifndef NDEBUG
1669 if (Result != LoopUnrollResult::Unmodified && ParentL)
1670 ParentL->verifyLoop();
1671#endif
1672
1673 // Clear any cached analysis results for L if we removed it completely.
1674 if (LAM && Result == LoopUnrollResult::FullyUnrolled)
1675 LAM->clear(IR&: L, Name: LoopName);
1676 }
1677
1678 if (!Changed)
1679 return PreservedAnalyses::all();
1680
1681 return getLoopPassPreservedAnalyses();
1682}
1683
1684void LoopUnrollPass::printPipeline(
1685 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1686 static_cast<PassInfoMixin<LoopUnrollPass> *>(this)->printPipeline(
1687 OS, MapClassName2PassName);
1688 OS << '<';
1689 if (UnrollOpts.AllowPartial != std::nullopt)
1690 OS << (*UnrollOpts.AllowPartial ? "" : "no-") << "partial;";
1691 if (UnrollOpts.AllowPeeling != std::nullopt)
1692 OS << (*UnrollOpts.AllowPeeling ? "" : "no-") << "peeling;";
1693 if (UnrollOpts.AllowRuntime != std::nullopt)
1694 OS << (*UnrollOpts.AllowRuntime ? "" : "no-") << "runtime;";
1695 if (UnrollOpts.AllowUpperBound != std::nullopt)
1696 OS << (*UnrollOpts.AllowUpperBound ? "" : "no-") << "upperbound;";
1697 if (UnrollOpts.AllowProfileBasedPeeling != std::nullopt)
1698 OS << (*UnrollOpts.AllowProfileBasedPeeling ? "" : "no-")
1699 << "profile-peeling;";
1700 if (UnrollOpts.FullUnrollMaxCount != std::nullopt)
1701 OS << "full-unroll-max=" << UnrollOpts.FullUnrollMaxCount << ';';
1702 OS << 'O' << UnrollOpts.OptLevel;
1703 OS << '>';
1704}
1705