1 //===- LoopUnroll.cpp - Loop unroller pass --------------------------------===//
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
7 //===----------------------------------------------------------------------===//
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 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar/LoopUnrollPass.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseMapInfo.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/StringRef.h"
25 #include "llvm/Analysis/AssumptionCache.h"
26 #include "llvm/Analysis/BlockFrequencyInfo.h"
27 #include "llvm/Analysis/CodeMetrics.h"
28 #include "llvm/Analysis/LazyBlockFrequencyInfo.h"
29 #include "llvm/Analysis/LoopAnalysisManager.h"
30 #include "llvm/Analysis/LoopInfo.h"
31 #include "llvm/Analysis/LoopPass.h"
32 #include "llvm/Analysis/LoopUnrollAnalyzer.h"
33 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
34 #include "llvm/Analysis/ProfileSummaryInfo.h"
35 #include "llvm/Analysis/ScalarEvolution.h"
36 #include "llvm/Analysis/TargetTransformInfo.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DiagnosticInfo.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Function.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Metadata.h"
48 #include "llvm/IR/PassManager.h"
49 #include "llvm/Pass.h"
50 #include "llvm/Support/Casting.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/ErrorHandling.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/Scalar.h"
56 #include "llvm/Transforms/Scalar/LoopPassManager.h"
57 #include "llvm/Transforms/Utils.h"
58 #include "llvm/Transforms/Utils/LoopSimplify.h"
59 #include "llvm/Transforms/Utils/LoopUtils.h"
60 #include "llvm/Transforms/Utils/SizeOpts.h"
61 #include "llvm/Transforms/Utils/UnrollLoop.h"
72 #define DEBUG_TYPE "loop-unroll"
74 cl::opt<bool> llvm::ForgetSCEVInLoopUnroll(
75 "forget-scev-loop-unroll", cl::init(false), cl::Hidden,
76 cl::desc("Forget everything in SCEV when doing LoopUnroll, instead of just"
77 " the current top-most loop. This is somtimes preferred to reduce"
80 static cl::opt<unsigned>
81 UnrollThreshold("unroll-threshold", cl::Hidden,
82 cl::desc("The cost threshold for loop unrolling"));
84 static cl::opt<unsigned> UnrollPartialThreshold(
85 "unroll-partial-threshold", cl::Hidden,
86 cl::desc("The cost threshold for partial loop unrolling"));
88 static cl::opt<unsigned> UnrollMaxPercentThresholdBoost(
89 "unroll-max-percent-threshold-boost", cl::init(400), cl::Hidden,
90 cl::desc("The maximum 'boost' (represented as a percentage >= 100) applied "
91 "to the threshold when aggressively unrolling a loop due to the "
92 "dynamic cost savings. If completely unrolling a loop will reduce "
93 "the total runtime from X to Y, we boost the loop unroll "
94 "threshold to DefaultThreshold*std::min(MaxPercentThresholdBoost, "
95 "X/Y). This limit avoids excessive code bloat."));
97 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
98 "unroll-max-iteration-count-to-analyze", cl::init(10), cl::Hidden,
99 cl::desc("Don't allow loop unrolling to simulate more than this number of"
100 "iterations when checking full unroll profitability"));
102 static cl::opt<unsigned> UnrollCount(
103 "unroll-count", cl::Hidden,
104 cl::desc("Use this unroll count for all loops including those with "
105 "unroll_count pragma values, for testing purposes"));
107 static cl::opt<unsigned> UnrollMaxCount(
108 "unroll-max-count", cl::Hidden,
109 cl::desc("Set the max unroll count for partial and runtime unrolling, for"
110 "testing purposes"));
112 static cl::opt<unsigned> UnrollFullMaxCount(
113 "unroll-full-max-count", cl::Hidden,
115 "Set the max unroll count for full unrolling, for testing purposes"));
117 static cl::opt<unsigned> UnrollPeelCount(
118 "unroll-peel-count", cl::Hidden,
119 cl::desc("Set the unroll peeling count, for testing purposes"));
122 UnrollAllowPartial("unroll-allow-partial", cl::Hidden,
123 cl::desc("Allows loops to be partially unrolled until "
124 "-unroll-threshold loop size is reached."));
126 static cl::opt<bool> UnrollAllowRemainder(
127 "unroll-allow-remainder", cl::Hidden,
128 cl::desc("Allow generation of a loop remainder (extra iterations) "
129 "when unrolling a loop."));
132 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::Hidden,
133 cl::desc("Unroll loops with run-time trip counts"));
135 static cl::opt<unsigned> UnrollMaxUpperBound(
136 "unroll-max-upperbound", cl::init(8), cl::Hidden,
138 "The max of trip count upper bound that is considered in unrolling"));
140 static cl::opt<unsigned> PragmaUnrollThreshold(
141 "pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
142 cl::desc("Unrolled size limit for loops with an unroll(full) or "
143 "unroll_count pragma."));
145 static cl::opt<unsigned> FlatLoopTripCountThreshold(
146 "flat-loop-tripcount-threshold", cl::init(5), cl::Hidden,
147 cl::desc("If the runtime tripcount for the loop is lower than the "
148 "threshold, the loop is considered as flat and will be less "
149 "aggressively unrolled."));
152 UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden,
153 cl::desc("Allows loops to be peeled when the dynamic "
154 "trip count is known to be low."));
156 static cl::opt<bool> UnrollUnrollRemainder(
157 "unroll-remainder", cl::Hidden,
158 cl::desc("Allow the loop remainder to be unrolled."));
160 // This option isn't ever intended to be enabled, it serves to allow
161 // experiments to check the assumptions about when this kind of revisit is
163 static cl::opt<bool> UnrollRevisitChildLoops(
164 "unroll-revisit-child-loops", cl::Hidden,
165 cl::desc("Enqueue and re-visit child loops in the loop PM after unrolling. "
166 "This shouldn't typically be needed as child loops (or their "
167 "clones) were already visited."));
169 /// A magic value for use with the Threshold parameter to indicate
170 /// that the loop unroll should be performed regardless of how much
171 /// code expansion would result.
172 static const unsigned NoThreshold = std::numeric_limits<unsigned>::max();
174 /// Gather the various unrolling parameters based on the defaults, compiler
175 /// flags, TTI overrides and user specified parameters.
176 TargetTransformInfo::UnrollingPreferences llvm::gatherUnrollingPreferences(
177 Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
178 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, int OptLevel,
179 Optional<unsigned> UserThreshold, Optional<unsigned> UserCount,
180 Optional<bool> UserAllowPartial, Optional<bool> UserRuntime,
181 Optional<bool> UserUpperBound, Optional<bool> UserAllowPeeling) {
182 TargetTransformInfo::UnrollingPreferences UP;
184 // Set up the defaults
185 UP.Threshold = OptLevel > 2 ? 300 : 150;
186 UP.MaxPercentThresholdBoost = 400;
187 UP.OptSizeThreshold = 0;
188 UP.PartialThreshold = 150;
189 UP.PartialOptSizeThreshold = 0;
192 UP.DefaultUnrollRuntimeCount = 8;
193 UP.MaxCount = std::numeric_limits<unsigned>::max();
194 UP.FullUnrollMaxCount = std::numeric_limits<unsigned>::max();
198 UP.AllowRemainder = true;
199 UP.UnrollRemainder = false;
200 UP.AllowExpensiveTripCount = false;
202 UP.UpperBound = false;
203 UP.AllowPeeling = true;
204 UP.UnrollAndJam = false;
205 UP.UnrollAndJamInnerLoopThreshold = 60;
207 // Override with any target specific settings
208 TTI.getUnrollingPreferences(L, SE, UP);
210 // Apply size attributes
211 bool OptForSize = L->getHeader()->getParent()->hasOptSize() ||
212 llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI);
214 UP.Threshold = UP.OptSizeThreshold;
215 UP.PartialThreshold = UP.PartialOptSizeThreshold;
216 UP.MaxPercentThresholdBoost = 100;
219 // Apply any user values specified by cl::opt
220 if (UnrollThreshold.getNumOccurrences() > 0)
221 UP.Threshold = UnrollThreshold;
222 if (UnrollPartialThreshold.getNumOccurrences() > 0)
223 UP.PartialThreshold = UnrollPartialThreshold;
224 if (UnrollMaxPercentThresholdBoost.getNumOccurrences() > 0)
225 UP.MaxPercentThresholdBoost = UnrollMaxPercentThresholdBoost;
226 if (UnrollMaxCount.getNumOccurrences() > 0)
227 UP.MaxCount = UnrollMaxCount;
228 if (UnrollFullMaxCount.getNumOccurrences() > 0)
229 UP.FullUnrollMaxCount = UnrollFullMaxCount;
230 if (UnrollPeelCount.getNumOccurrences() > 0)
231 UP.PeelCount = UnrollPeelCount;
232 if (UnrollAllowPartial.getNumOccurrences() > 0)
233 UP.Partial = UnrollAllowPartial;
234 if (UnrollAllowRemainder.getNumOccurrences() > 0)
235 UP.AllowRemainder = UnrollAllowRemainder;
236 if (UnrollRuntime.getNumOccurrences() > 0)
237 UP.Runtime = UnrollRuntime;
238 if (UnrollMaxUpperBound == 0)
239 UP.UpperBound = false;
240 if (UnrollAllowPeeling.getNumOccurrences() > 0)
241 UP.AllowPeeling = UnrollAllowPeeling;
242 if (UnrollUnrollRemainder.getNumOccurrences() > 0)
243 UP.UnrollRemainder = UnrollUnrollRemainder;
245 // Apply user values provided by argument
246 if (UserThreshold.hasValue()) {
247 UP.Threshold = *UserThreshold;
248 UP.PartialThreshold = *UserThreshold;
250 if (UserCount.hasValue())
251 UP.Count = *UserCount;
252 if (UserAllowPartial.hasValue())
253 UP.Partial = *UserAllowPartial;
254 if (UserRuntime.hasValue())
255 UP.Runtime = *UserRuntime;
256 if (UserUpperBound.hasValue())
257 UP.UpperBound = *UserUpperBound;
258 if (UserAllowPeeling.hasValue())
259 UP.AllowPeeling = *UserAllowPeeling;
266 /// A struct to densely store the state of an instruction after unrolling at
269 /// This is designed to work like a tuple of <Instruction *, int> for the
270 /// purposes of hashing and lookup, but to be able to associate two boolean
271 /// states with each key.
272 struct UnrolledInstState {
276 unsigned IsCounted : 1;
279 /// Hashing and equality testing for a set of the instruction states.
280 struct UnrolledInstStateKeyInfo {
281 using PtrInfo = DenseMapInfo<Instruction *>;
282 using PairInfo = DenseMapInfo<std::pair<Instruction *, int>>;
284 static inline UnrolledInstState getEmptyKey() {
285 return {PtrInfo::getEmptyKey(), 0, 0, 0};
288 static inline UnrolledInstState getTombstoneKey() {
289 return {PtrInfo::getTombstoneKey(), 0, 0, 0};
292 static inline unsigned getHashValue(const UnrolledInstState &S) {
293 return PairInfo::getHashValue({S.I, S.Iteration});
296 static inline bool isEqual(const UnrolledInstState &LHS,
297 const UnrolledInstState &RHS) {
298 return PairInfo::isEqual({LHS.I, LHS.Iteration}, {RHS.I, RHS.Iteration});
302 struct EstimatedUnrollCost {
303 /// The estimated cost after unrolling.
304 unsigned UnrolledCost;
306 /// The estimated dynamic cost of executing the instructions in the
308 unsigned RolledDynamicCost;
311 } // end anonymous namespace
313 /// Figure out if the loop is worth full unrolling.
315 /// Complete loop unrolling can make some loads constant, and we need to know
316 /// if that would expose any further optimization opportunities. This routine
317 /// estimates this optimization. It computes cost of unrolled loop
318 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
319 /// dynamic cost we mean that we won't count costs of blocks that are known not
320 /// to be executed (i.e. if we have a branch in the loop and we know that at the
321 /// given iteration its condition would be resolved to true, we won't add up the
322 /// cost of the 'false'-block).
323 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
324 /// the analysis failed (no benefits expected from the unrolling, or the loop is
325 /// too big to analyze), the returned value is None.
326 static Optional<EstimatedUnrollCost> analyzeLoopUnrollCost(
327 const Loop *L, unsigned TripCount, DominatorTree &DT, ScalarEvolution &SE,
328 const SmallPtrSetImpl<const Value *> &EphValues,
329 const TargetTransformInfo &TTI, unsigned MaxUnrolledLoopSize) {
330 // We want to be able to scale offsets by the trip count and add more offsets
331 // to them without checking for overflows, and we already don't want to
332 // analyze *massive* trip counts, so we force the max to be reasonably small.
333 assert(UnrollMaxIterationsCountToAnalyze <
334 (unsigned)(std::numeric_limits<int>::max() / 2) &&
335 "The unroll iterations max is too large!");
337 // Only analyze inner loops. We can't properly estimate cost of nested loops
338 // and we won't visit inner loops again anyway.
342 // Don't simulate loops with a big or unknown tripcount
343 if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
344 TripCount > UnrollMaxIterationsCountToAnalyze)
347 SmallSetVector<BasicBlock *, 16> BBWorklist;
348 SmallSetVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitWorklist;
349 DenseMap<Value *, Constant *> SimplifiedValues;
350 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues;
352 // The estimated cost of the unrolled form of the loop. We try to estimate
353 // this by simplifying as much as we can while computing the estimate.
354 unsigned UnrolledCost = 0;
356 // We also track the estimated dynamic (that is, actually executed) cost in
357 // the rolled form. This helps identify cases when the savings from unrolling
358 // aren't just exposing dead control flows, but actual reduced dynamic
359 // instructions due to the simplifications which we expect to occur after
361 unsigned RolledDynamicCost = 0;
363 // We track the simplification of each instruction in each iteration. We use
364 // this to recursively merge costs into the unrolled cost on-demand so that
365 // we don't count the cost of any dead code. This is essentially a map from
366 // <instruction, int> to <bool, bool>, but stored as a densely packed struct.
367 DenseSet<UnrolledInstState, UnrolledInstStateKeyInfo> InstCostMap;
369 // A small worklist used to accumulate cost of instructions from each
370 // observable and reached root in the loop.
371 SmallVector<Instruction *, 16> CostWorklist;
373 // PHI-used worklist used between iterations while accumulating cost.
374 SmallVector<Instruction *, 4> PHIUsedList;
376 // Helper function to accumulate cost for instructions in the loop.
377 auto AddCostRecursively = [&](Instruction &RootI, int Iteration) {
378 assert(Iteration >= 0 && "Cannot have a negative iteration!");
379 assert(CostWorklist.empty() && "Must start with an empty cost list");
380 assert(PHIUsedList.empty() && "Must start with an empty phi used list");
381 CostWorklist.push_back(&RootI);
382 for (;; --Iteration) {
384 Instruction *I = CostWorklist.pop_back_val();
386 // InstCostMap only uses I and Iteration as a key, the other two values
387 // don't matter here.
388 auto CostIter = InstCostMap.find({I, Iteration, 0, 0});
389 if (CostIter == InstCostMap.end())
390 // If an input to a PHI node comes from a dead path through the loop
391 // we may have no cost data for it here. What that actually means is
394 auto &Cost = *CostIter;
396 // Already counted this instruction.
399 // Mark that we are counting the cost of this instruction now.
400 Cost.IsCounted = true;
402 // If this is a PHI node in the loop header, just add it to the PHI set.
403 if (auto *PhiI = dyn_cast<PHINode>(I))
404 if (PhiI->getParent() == L->getHeader()) {
405 assert(Cost.IsFree && "Loop PHIs shouldn't be evaluated as they "
406 "inherently simplify during unrolling.");
410 // Push the incoming value from the backedge into the PHI used list
411 // if it is an in-loop instruction. We'll use this to populate the
412 // cost worklist for the next iteration (as we count backwards).
413 if (auto *OpI = dyn_cast<Instruction>(
414 PhiI->getIncomingValueForBlock(L->getLoopLatch())))
415 if (L->contains(OpI))
416 PHIUsedList.push_back(OpI);
420 // First accumulate the cost of this instruction.
422 UnrolledCost += TTI.getUserCost(I);
423 LLVM_DEBUG(dbgs() << "Adding cost of instruction (iteration "
424 << Iteration << "): ");
425 LLVM_DEBUG(I->dump());
428 // We must count the cost of every operand which is not free,
429 // recursively. If we reach a loop PHI node, simply add it to the set
430 // to be considered on the next iteration (backwards!).
431 for (Value *Op : I->operands()) {
432 // Check whether this operand is free due to being a constant or
434 auto *OpI = dyn_cast<Instruction>(Op);
435 if (!OpI || !L->contains(OpI))
438 // Otherwise accumulate its cost.
439 CostWorklist.push_back(OpI);
441 } while (!CostWorklist.empty());
443 if (PHIUsedList.empty())
444 // We've exhausted the search.
447 assert(Iteration > 0 &&
448 "Cannot track PHI-used values past the first iteration!");
449 CostWorklist.append(PHIUsedList.begin(), PHIUsedList.end());
454 // Ensure that we don't violate the loop structure invariants relied on by
456 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
457 assert(L->isLCSSAForm(DT) &&
458 "Must have loops in LCSSA form to track live-out values.");
460 LLVM_DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
462 // Simulate execution of each iteration of the loop counting instructions,
463 // which would be simplified.
464 // Since the same load will take different values on different iterations,
465 // we literally have to go through all loop's iterations.
466 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
467 LLVM_DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
469 // Prepare for the iteration by collecting any simplified entry or backedge
471 for (Instruction &I : *L->getHeader()) {
472 auto *PHI = dyn_cast<PHINode>(&I);
476 // The loop header PHI nodes must have exactly two input: one from the
477 // loop preheader and one from the loop latch.
479 PHI->getNumIncomingValues() == 2 &&
480 "Must have an incoming value only for the preheader and the latch.");
482 Value *V = PHI->getIncomingValueForBlock(
483 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
484 Constant *C = dyn_cast<Constant>(V);
485 if (Iteration != 0 && !C)
486 C = SimplifiedValues.lookup(V);
488 SimplifiedInputValues.push_back({PHI, C});
491 // Now clear and re-populate the map for the next iteration.
492 SimplifiedValues.clear();
493 while (!SimplifiedInputValues.empty())
494 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val());
496 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L);
499 BBWorklist.insert(L->getHeader());
500 // Note that we *must not* cache the size, this loop grows the worklist.
501 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
502 BasicBlock *BB = BBWorklist[Idx];
504 // Visit all instructions in the given basic block and try to simplify
505 // it. We don't change the actual IR, just count optimization
507 for (Instruction &I : *BB) {
508 // These won't get into the final code - don't even try calculating the
510 if (isa<DbgInfoIntrinsic>(I) || EphValues.count(&I))
513 // Track this instruction's expected baseline cost when executing the
515 RolledDynamicCost += TTI.getUserCost(&I);
517 // Visit the instruction to analyze its loop cost after unrolling,
518 // and if the visitor returns true, mark the instruction as free after
519 // unrolling and continue.
520 bool IsFree = Analyzer.visit(I);
521 bool Inserted = InstCostMap.insert({&I, (int)Iteration,
523 /*IsCounted*/ false}).second;
525 assert(Inserted && "Cannot have a state for an unvisited instruction!");
530 // Can't properly model a cost of a call.
531 // FIXME: With a proper cost model we should be able to do it.
532 if (auto *CI = dyn_cast<CallInst>(&I)) {
533 const Function *Callee = CI->getCalledFunction();
534 if (!Callee || TTI.isLoweredToCall(Callee)) {
535 LLVM_DEBUG(dbgs() << "Can't analyze cost of loop with call\n");
540 // If the instruction might have a side-effect recursively account for
541 // the cost of it and all the instructions leading up to it.
542 if (I.mayHaveSideEffects())
543 AddCostRecursively(I, Iteration);
545 // If unrolled body turns out to be too big, bail out.
546 if (UnrolledCost > MaxUnrolledLoopSize) {
547 LLVM_DEBUG(dbgs() << " Exceeded threshold.. exiting.\n"
548 << " UnrolledCost: " << UnrolledCost
549 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
555 Instruction *TI = BB->getTerminator();
557 // Add in the live successors by first checking whether we have terminator
558 // that may be simplified based on the values simplified by this call.
559 BasicBlock *KnownSucc = nullptr;
560 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
561 if (BI->isConditional()) {
562 if (Constant *SimpleCond =
563 SimplifiedValues.lookup(BI->getCondition())) {
564 // Just take the first successor if condition is undef
565 if (isa<UndefValue>(SimpleCond))
566 KnownSucc = BI->getSuccessor(0);
567 else if (ConstantInt *SimpleCondVal =
568 dyn_cast<ConstantInt>(SimpleCond))
569 KnownSucc = BI->getSuccessor(SimpleCondVal->isZero() ? 1 : 0);
572 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
573 if (Constant *SimpleCond =
574 SimplifiedValues.lookup(SI->getCondition())) {
575 // Just take the first successor if condition is undef
576 if (isa<UndefValue>(SimpleCond))
577 KnownSucc = SI->getSuccessor(0);
578 else if (ConstantInt *SimpleCondVal =
579 dyn_cast<ConstantInt>(SimpleCond))
580 KnownSucc = SI->findCaseValue(SimpleCondVal)->getCaseSuccessor();
584 if (L->contains(KnownSucc))
585 BBWorklist.insert(KnownSucc);
587 ExitWorklist.insert({BB, KnownSucc});
591 // Add BB's successors to the worklist.
592 for (BasicBlock *Succ : successors(BB))
593 if (L->contains(Succ))
594 BBWorklist.insert(Succ);
596 ExitWorklist.insert({BB, Succ});
597 AddCostRecursively(*TI, Iteration);
600 // If we found no optimization opportunities on the first iteration, we
601 // won't find them on later ones too.
602 if (UnrolledCost == RolledDynamicCost) {
603 LLVM_DEBUG(dbgs() << " No opportunities found.. exiting.\n"
604 << " UnrolledCost: " << UnrolledCost << "\n");
609 while (!ExitWorklist.empty()) {
610 BasicBlock *ExitingBB, *ExitBB;
611 std::tie(ExitingBB, ExitBB) = ExitWorklist.pop_back_val();
613 for (Instruction &I : *ExitBB) {
614 auto *PN = dyn_cast<PHINode>(&I);
618 Value *Op = PN->getIncomingValueForBlock(ExitingBB);
619 if (auto *OpI = dyn_cast<Instruction>(Op))
620 if (L->contains(OpI))
621 AddCostRecursively(*OpI, TripCount - 1);
625 LLVM_DEBUG(dbgs() << "Analysis finished:\n"
626 << "UnrolledCost: " << UnrolledCost << ", "
627 << "RolledDynamicCost: " << RolledDynamicCost << "\n");
628 return {{UnrolledCost, RolledDynamicCost}};
631 /// ApproximateLoopSize - Approximate the size of the loop.
632 unsigned llvm::ApproximateLoopSize(
633 const Loop *L, unsigned &NumCalls, bool &NotDuplicatable, bool &Convergent,
634 const TargetTransformInfo &TTI,
635 const SmallPtrSetImpl<const Value *> &EphValues, unsigned BEInsns) {
637 for (BasicBlock *BB : L->blocks())
638 Metrics.analyzeBasicBlock(BB, TTI, EphValues);
639 NumCalls = Metrics.NumInlineCandidates;
640 NotDuplicatable = Metrics.notDuplicatable;
641 Convergent = Metrics.convergent;
643 unsigned LoopSize = Metrics.NumInsts;
645 // Don't allow an estimate of size zero. This would allows unrolling of loops
646 // with huge iteration counts, which is a compile time problem even if it's
647 // not a problem for code quality. Also, the code using this size may assume
648 // that each loop has at least three instructions (likely a conditional
649 // branch, a comparison feeding that branch, and some kind of loop increment
650 // feeding that comparison instruction).
651 LoopSize = std::max(LoopSize, BEInsns + 1);
656 // Returns the loop hint metadata node with the given name (for example,
657 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
659 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
660 if (MDNode *LoopID = L->getLoopID())
661 return GetUnrollMetadata(LoopID, Name);
665 // Returns true if the loop has an unroll(full) pragma.
666 static bool HasUnrollFullPragma(const Loop *L) {
667 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
670 // Returns true if the loop has an unroll(enable) pragma. This metadata is used
671 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
672 static bool HasUnrollEnablePragma(const Loop *L) {
673 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.enable");
676 // Returns true if the loop has an runtime unroll(disable) pragma.
677 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
678 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
681 // If loop has an unroll_count pragma return the (necessarily
682 // positive) value from the pragma. Otherwise return 0.
683 static unsigned UnrollCountPragmaValue(const Loop *L) {
684 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
686 assert(MD->getNumOperands() == 2 &&
687 "Unroll count hint metadata should have two operands.");
689 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
690 assert(Count >= 1 && "Unroll count must be positive.");
696 // Computes the boosting factor for complete unrolling.
697 // If fully unrolling the loop would save a lot of RolledDynamicCost, it would
698 // be beneficial to fully unroll the loop even if unrolledcost is large. We
699 // use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust
700 // the unroll threshold.
701 static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost,
702 unsigned MaxPercentThresholdBoost) {
703 if (Cost.RolledDynamicCost >= std::numeric_limits<unsigned>::max() / 100)
705 else if (Cost.UnrolledCost != 0)
706 // The boosting factor is RolledDynamicCost / UnrolledCost
707 return std::min(100 * Cost.RolledDynamicCost / Cost.UnrolledCost,
708 MaxPercentThresholdBoost);
710 return MaxPercentThresholdBoost;
713 // Returns loop size estimation for unrolled loop.
714 static uint64_t getUnrolledLoopSize(
716 TargetTransformInfo::UnrollingPreferences &UP) {
717 assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
718 return (uint64_t)(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns;
721 // Returns true if unroll count was set explicitly.
722 // Calculates unroll count and writes it to UP.Count.
723 // Unless IgnoreUser is true, will also use metadata and command-line options
724 // that are specific to to the LoopUnroll pass (which, for instance, are
725 // irrelevant for the LoopUnrollAndJam pass).
726 // FIXME: This function is used by LoopUnroll and LoopUnrollAndJam, but consumes
727 // many LoopUnroll-specific options. The shared functionality should be
728 // refactored into it own function.
729 bool llvm::computeUnrollCount(
730 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI,
731 ScalarEvolution &SE, const SmallPtrSetImpl<const Value *> &EphValues,
732 OptimizationRemarkEmitter *ORE, unsigned &TripCount, unsigned MaxTripCount,
733 unsigned &TripMultiple, unsigned LoopSize,
734 TargetTransformInfo::UnrollingPreferences &UP, bool &UseUpperBound) {
736 // Check for explicit Count.
737 // 1st priority is unroll count set by "unroll-count" option.
738 bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0;
739 if (UserUnrollCount) {
740 UP.Count = UnrollCount;
741 UP.AllowExpensiveTripCount = true;
743 if (UP.AllowRemainder && getUnrolledLoopSize(LoopSize, UP) < UP.Threshold)
747 // 2nd priority is unroll count set by pragma.
748 unsigned PragmaCount = UnrollCountPragmaValue(L);
749 if (PragmaCount > 0) {
750 UP.Count = PragmaCount;
752 UP.AllowExpensiveTripCount = true;
754 if ((UP.AllowRemainder || (TripMultiple % PragmaCount == 0)) &&
755 getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
758 bool PragmaFullUnroll = HasUnrollFullPragma(L);
759 if (PragmaFullUnroll && TripCount != 0) {
760 UP.Count = TripCount;
761 if (getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
765 bool PragmaEnableUnroll = HasUnrollEnablePragma(L);
766 bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll ||
767 PragmaEnableUnroll || UserUnrollCount;
769 if (ExplicitUnroll && TripCount != 0) {
770 // If the loop has an unrolling pragma, we want to be more aggressive with
771 // unrolling limits. Set thresholds to at least the PragmaUnrollThreshold
772 // value which is larger than the default limits.
773 UP.Threshold = std::max<unsigned>(UP.Threshold, PragmaUnrollThreshold);
774 UP.PartialThreshold =
775 std::max<unsigned>(UP.PartialThreshold, PragmaUnrollThreshold);
778 // 3rd priority is full unroll count.
779 // Full unroll makes sense only when TripCount or its upper bound could be
780 // statically calculated.
781 // Also we need to check if we exceed FullUnrollMaxCount.
782 // If using the upper bound to unroll, TripMultiple should be set to 1 because
783 // we do not know when loop may exit.
784 // MaxTripCount and ExactTripCount cannot both be non zero since we only
785 // compute the former when the latter is zero.
786 unsigned ExactTripCount = TripCount;
787 assert((ExactTripCount == 0 || MaxTripCount == 0) &&
788 "ExtractTripCount and MaxTripCount cannot both be non zero.");
789 unsigned FullUnrollTripCount = ExactTripCount ? ExactTripCount : MaxTripCount;
790 UP.Count = FullUnrollTripCount;
791 if (FullUnrollTripCount && FullUnrollTripCount <= UP.FullUnrollMaxCount) {
792 // When computing the unrolled size, note that BEInsns are not replicated
793 // like the rest of the loop body.
794 if (getUnrolledLoopSize(LoopSize, UP) < UP.Threshold) {
795 UseUpperBound = (MaxTripCount == FullUnrollTripCount);
796 TripCount = FullUnrollTripCount;
797 TripMultiple = UP.UpperBound ? 1 : TripMultiple;
798 return ExplicitUnroll;
800 // The loop isn't that small, but we still can fully unroll it if that
801 // helps to remove a significant number of instructions.
802 // To check that, run additional analysis on the loop.
803 if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
804 L, FullUnrollTripCount, DT, SE, EphValues, TTI,
805 UP.Threshold * UP.MaxPercentThresholdBoost / 100)) {
807 getFullUnrollBoostingFactor(*Cost, UP.MaxPercentThresholdBoost);
808 if (Cost->UnrolledCost < UP.Threshold * Boost / 100) {
809 UseUpperBound = (MaxTripCount == FullUnrollTripCount);
810 TripCount = FullUnrollTripCount;
811 TripMultiple = UP.UpperBound ? 1 : TripMultiple;
812 return ExplicitUnroll;
818 // 4th priority is loop peeling.
819 computePeelCount(L, LoopSize, UP, TripCount, SE);
823 return ExplicitUnroll;
826 // 5th priority is partial unrolling.
827 // Try partial unroll only when TripCount could be statically calculated.
829 UP.Partial |= ExplicitUnroll;
831 LLVM_DEBUG(dbgs() << " will not try to unroll partially because "
832 << "-unroll-allow-partial not given\n");
837 UP.Count = TripCount;
838 if (UP.PartialThreshold != NoThreshold) {
839 // Reduce unroll count to be modulo of TripCount for partial unrolling.
840 if (getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
842 (std::max(UP.PartialThreshold, UP.BEInsns + 1) - UP.BEInsns) /
843 (LoopSize - UP.BEInsns);
844 if (UP.Count > UP.MaxCount)
845 UP.Count = UP.MaxCount;
846 while (UP.Count != 0 && TripCount % UP.Count != 0)
848 if (UP.AllowRemainder && UP.Count <= 1) {
849 // If there is no Count that is modulo of TripCount, set Count to
850 // largest power-of-two factor that satisfies the threshold limit.
851 // As we'll create fixup loop, do the type of unrolling only if
852 // remainder loop is allowed.
853 UP.Count = UP.DefaultUnrollRuntimeCount;
854 while (UP.Count != 0 &&
855 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
859 if (PragmaEnableUnroll)
861 return OptimizationRemarkMissed(DEBUG_TYPE,
862 "UnrollAsDirectedTooLarge",
863 L->getStartLoc(), L->getHeader())
864 << "Unable to unroll loop as directed by unroll(enable) "
866 "because unrolled size is too large.";
871 UP.Count = TripCount;
873 if (UP.Count > UP.MaxCount)
874 UP.Count = UP.MaxCount;
875 if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
876 UP.Count != TripCount)
878 return OptimizationRemarkMissed(DEBUG_TYPE,
879 "FullUnrollAsDirectedTooLarge",
880 L->getStartLoc(), L->getHeader())
881 << "Unable to fully unroll loop as directed by unroll pragma "
883 "unrolled size is too large.";
885 return ExplicitUnroll;
887 assert(TripCount == 0 &&
888 "All cases when TripCount is constant should be covered here.");
889 if (PragmaFullUnroll)
891 return OptimizationRemarkMissed(
892 DEBUG_TYPE, "CantFullUnrollAsDirectedRuntimeTripCount",
893 L->getStartLoc(), L->getHeader())
894 << "Unable to fully unroll loop as directed by unroll(full) "
896 "because loop has a runtime trip count.";
899 // 6th priority is runtime unrolling.
900 // Don't unroll a runtime trip count loop when it is disabled.
901 if (HasRuntimeUnrollDisablePragma(L)) {
906 // Check if the runtime trip count is too small when profile is available.
907 if (L->getHeader()->getParent()->hasProfileData()) {
908 if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) {
909 if (*ProfileTripCount < FlatLoopTripCountThreshold)
912 UP.AllowExpensiveTripCount = true;
916 // Reduce count based on the type of unrolling and the threshold values.
917 UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount;
920 dbgs() << " will not try to unroll loop with runtime trip count "
921 << "-unroll-runtime not given\n");
926 UP.Count = UP.DefaultUnrollRuntimeCount;
928 // Reduce unroll count to be the largest power-of-two factor of
929 // the original count which satisfies the threshold limit.
930 while (UP.Count != 0 &&
931 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
935 unsigned OrigCount = UP.Count;
938 if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) {
939 while (UP.Count != 0 && TripMultiple % UP.Count != 0)
942 dbgs() << "Remainder loop is restricted (that could architecture "
943 "specific or because the loop contains a convergent "
944 "instruction), so unroll count must divide the trip "
946 << TripMultiple << ". Reducing unroll count from " << OrigCount
947 << " to " << UP.Count << ".\n");
951 if (PragmaCount > 0 && !UP.AllowRemainder)
953 return OptimizationRemarkMissed(DEBUG_TYPE,
954 "DifferentUnrollCountFromDirected",
955 L->getStartLoc(), L->getHeader())
956 << "Unable to unroll loop the number of times directed by "
957 "unroll_count pragma because remainder loop is restricted "
958 "(that could architecture specific or because the loop "
959 "contains a convergent instruction) and so must have an "
961 "count that divides the loop trip multiple of "
962 << NV("TripMultiple", TripMultiple) << ". Unrolling instead "
963 << NV("UnrollCount", UP.Count) << " time(s).";
967 if (UP.Count > UP.MaxCount)
968 UP.Count = UP.MaxCount;
969 LLVM_DEBUG(dbgs() << " partially unrolling with count: " << UP.Count
973 return ExplicitUnroll;
976 static LoopUnrollResult tryToUnrollLoop(
977 Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE,
978 const TargetTransformInfo &TTI, AssumptionCache &AC,
979 OptimizationRemarkEmitter &ORE,
980 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
981 bool PreserveLCSSA, int OptLevel,
982 bool OnlyWhenForced, bool ForgetAllSCEV, Optional<unsigned> ProvidedCount,
983 Optional<unsigned> ProvidedThreshold, Optional<bool> ProvidedAllowPartial,
984 Optional<bool> ProvidedRuntime, Optional<bool> ProvidedUpperBound,
985 Optional<bool> ProvidedAllowPeeling) {
986 LLVM_DEBUG(dbgs() << "Loop Unroll: F["
987 << L->getHeader()->getParent()->getName() << "] Loop %"
988 << L->getHeader()->getName() << "\n");
989 TransformationMode TM = hasUnrollTransformation(L);
991 return LoopUnrollResult::Unmodified;
992 if (!L->isLoopSimplifyForm()) {
994 dbgs() << " Not unrolling loop which is not in loop-simplify form.\n");
995 return LoopUnrollResult::Unmodified;
998 // When automtatic unrolling is disabled, do not unroll unless overridden for
1000 if (OnlyWhenForced && !(TM & TM_Enable))
1001 return LoopUnrollResult::Unmodified;
1003 bool OptForSize = L->getHeader()->getParent()->hasOptSize();
1004 unsigned NumInlineCandidates;
1005 bool NotDuplicatable;
1007 TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
1008 L, SE, TTI, BFI, PSI, OptLevel, ProvidedThreshold, ProvidedCount,
1009 ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound,
1010 ProvidedAllowPeeling);
1012 // Exit early if unrolling is disabled. For OptForSize, we pick the loop size
1013 // as threshold later on.
1014 if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0) &&
1016 return LoopUnrollResult::Unmodified;
1018 SmallPtrSet<const Value *, 32> EphValues;
1019 CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
1022 ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent,
1023 TTI, EphValues, UP.BEInsns);
1024 LLVM_DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
1025 if (NotDuplicatable) {
1026 LLVM_DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
1027 << " instructions.\n");
1028 return LoopUnrollResult::Unmodified;
1031 // When optimizing for size, use LoopSize as threshold, to (fully) unroll
1032 // loops, if it does not increase code size.
1034 UP.Threshold = std::max(UP.Threshold, LoopSize);
1036 if (NumInlineCandidates != 0) {
1037 LLVM_DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
1038 return LoopUnrollResult::Unmodified;
1041 // Find trip count and trip multiple if count is not available
1042 unsigned TripCount = 0;
1043 unsigned MaxTripCount = 0;
1044 unsigned TripMultiple = 1;
1045 // If there are multiple exiting blocks but one of them is the latch, use the
1046 // latch for the trip count estimation. Otherwise insist on a single exiting
1047 // block for the trip count estimation.
1048 BasicBlock *ExitingBlock = L->getLoopLatch();
1049 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
1050 ExitingBlock = L->getExitingBlock();
1052 TripCount = SE.getSmallConstantTripCount(L, ExitingBlock);
1053 TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock);
1056 // If the loop contains a convergent operation, the prelude we'd add
1057 // to do the first few instructions before we hit the unrolled loop
1058 // is unsafe -- it adds a control-flow dependency to the convergent
1059 // operation. Therefore restrict remainder loop (try unrollig without).
1061 // TODO: This is quite conservative. In practice, convergent_op()
1062 // is likely to be called unconditionally in the loop. In this
1063 // case, the program would be ill-formed (on most architectures)
1064 // unless n were the same on all threads in a thread group.
1065 // Assuming n is the same on all threads, any kind of unrolling is
1066 // safe. But currently llvm's notion of convergence isn't powerful
1067 // enough to express this.
1069 UP.AllowRemainder = false;
1071 // Try to find the trip count upper bound if we cannot find the exact trip
1073 bool MaxOrZero = false;
1075 MaxTripCount = SE.getSmallConstantMaxTripCount(L);
1076 MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L);
1077 // We can unroll by the upper bound amount if it's generally allowed or if
1078 // we know that the loop is executed either the upper bound or zero times.
1079 // (MaxOrZero unrolling keeps only the first loop test, so the number of
1080 // loop tests remains the same compared to the non-unrolled version, whereas
1081 // the generic upper bound unrolling keeps all but the last loop test so the
1082 // number of loop tests goes up which may end up being worse on targets with
1083 // constrained branch predictor resources so is controlled by an option.)
1084 // In addition we only unroll small upper bounds.
1085 if (!(UP.UpperBound || MaxOrZero) || MaxTripCount > UnrollMaxUpperBound) {
1090 // computeUnrollCount() decides whether it is beneficial to use upper bound to
1091 // fully unroll the loop.
1092 bool UseUpperBound = false;
1093 bool IsCountSetExplicitly = computeUnrollCount(
1094 L, TTI, DT, LI, SE, EphValues, &ORE, TripCount, MaxTripCount,
1095 TripMultiple, LoopSize, UP, UseUpperBound);
1097 return LoopUnrollResult::Unmodified;
1098 // Unroll factor (Count) must be less or equal to TripCount.
1099 if (TripCount && UP.Count > TripCount)
1100 UP.Count = TripCount;
1102 // Save loop properties before it is transformed.
1103 MDNode *OrigLoopID = L->getLoopID();
1106 Loop *RemainderLoop = nullptr;
1107 LoopUnrollResult UnrollResult = UnrollLoop(
1109 {UP.Count, TripCount, UP.Force, UP.Runtime, UP.AllowExpensiveTripCount,
1110 UseUpperBound, MaxOrZero, TripMultiple, UP.PeelCount, UP.UnrollRemainder,
1112 LI, &SE, &DT, &AC, &ORE, PreserveLCSSA, &RemainderLoop);
1113 if (UnrollResult == LoopUnrollResult::Unmodified)
1114 return LoopUnrollResult::Unmodified;
1116 if (RemainderLoop) {
1117 Optional<MDNode *> RemainderLoopID =
1118 makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1119 LLVMLoopUnrollFollowupRemainder});
1120 if (RemainderLoopID.hasValue())
1121 RemainderLoop->setLoopID(RemainderLoopID.getValue());
1124 if (UnrollResult != LoopUnrollResult::FullyUnrolled) {
1125 Optional<MDNode *> NewLoopID =
1126 makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1127 LLVMLoopUnrollFollowupUnrolled});
1128 if (NewLoopID.hasValue()) {
1129 L->setLoopID(NewLoopID.getValue());
1131 // Do not setLoopAlreadyUnrolled if loop attributes have been specified
1133 return UnrollResult;
1137 // If loop has an unroll count pragma or unrolled by explicitly set count
1138 // mark loop as unrolled to prevent unrolling beyond that requested.
1139 // If the loop was peeled, we already "used up" the profile information
1140 // we had, so we don't want to unroll or peel again.
1141 if (UnrollResult != LoopUnrollResult::FullyUnrolled &&
1142 (IsCountSetExplicitly || UP.PeelCount))
1143 L->setLoopAlreadyUnrolled();
1145 return UnrollResult;
1150 class LoopUnroll : public LoopPass {
1152 static char ID; // Pass ID, replacement for typeid
1156 /// If false, use a cost model to determine whether unrolling of a loop is
1157 /// profitable. If true, only loops that explicitly request unrolling via
1158 /// metadata are considered. All other loops are skipped.
1159 bool OnlyWhenForced;
1161 /// If false, when SCEV is invalidated, only forget everything in the
1162 /// top-most loop (call forgetTopMostLoop), of the loop being processed.
1163 /// Otherwise, forgetAllLoops and rebuild when needed next.
1166 Optional<unsigned> ProvidedCount;
1167 Optional<unsigned> ProvidedThreshold;
1168 Optional<bool> ProvidedAllowPartial;
1169 Optional<bool> ProvidedRuntime;
1170 Optional<bool> ProvidedUpperBound;
1171 Optional<bool> ProvidedAllowPeeling;
1173 LoopUnroll(int OptLevel = 2, bool OnlyWhenForced = false,
1174 bool ForgetAllSCEV = false, Optional<unsigned> Threshold = None,
1175 Optional<unsigned> Count = None,
1176 Optional<bool> AllowPartial = None, Optional<bool> Runtime = None,
1177 Optional<bool> UpperBound = None,
1178 Optional<bool> AllowPeeling = None)
1179 : LoopPass(ID), OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
1180 ForgetAllSCEV(ForgetAllSCEV), ProvidedCount(std::move(Count)),
1181 ProvidedThreshold(Threshold), ProvidedAllowPartial(AllowPartial),
1182 ProvidedRuntime(Runtime), ProvidedUpperBound(UpperBound),
1183 ProvidedAllowPeeling(AllowPeeling) {
1184 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
1187 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1191 Function &F = *L->getHeader()->getParent();
1193 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1194 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1195 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1196 const TargetTransformInfo &TTI =
1197 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1198 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1199 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
1200 // pass. Function analyses need to be preserved across loop transformations
1201 // but ORE cannot be preserved (see comment before the pass definition).
1202 OptimizationRemarkEmitter ORE(&F);
1203 bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1205 LoopUnrollResult Result = tryToUnrollLoop(
1206 L, DT, LI, SE, TTI, AC, ORE, nullptr, nullptr,
1207 PreserveLCSSA, OptLevel, OnlyWhenForced,
1208 ForgetAllSCEV, ProvidedCount, ProvidedThreshold, ProvidedAllowPartial,
1209 ProvidedRuntime, ProvidedUpperBound, ProvidedAllowPeeling);
1211 if (Result == LoopUnrollResult::FullyUnrolled)
1212 LPM.markLoopAsDeleted(*L);
1214 return Result != LoopUnrollResult::Unmodified;
1217 /// This transformation requires natural loop information & requires that
1218 /// loop preheaders be inserted into the CFG...
1219 void getAnalysisUsage(AnalysisUsage &AU) const override {
1220 AU.addRequired<AssumptionCacheTracker>();
1221 AU.addRequired<TargetTransformInfoWrapperPass>();
1222 // FIXME: Loop passes are required to preserve domtree, and for now we just
1223 // recreate dom info if anything gets unrolled.
1224 getLoopAnalysisUsage(AU);
1228 } // end anonymous namespace
1230 char LoopUnroll::ID = 0;
1232 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1233 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1234 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1235 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1236 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1238 Pass *llvm::createLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1239 bool ForgetAllSCEV, int Threshold, int Count,
1240 int AllowPartial, int Runtime, int UpperBound,
1242 // TODO: It would make more sense for this function to take the optionals
1243 // directly, but that's dangerous since it would silently break out of tree
1245 return new LoopUnroll(
1246 OptLevel, OnlyWhenForced, ForgetAllSCEV,
1247 Threshold == -1 ? None : Optional<unsigned>(Threshold),
1248 Count == -1 ? None : Optional<unsigned>(Count),
1249 AllowPartial == -1 ? None : Optional<bool>(AllowPartial),
1250 Runtime == -1 ? None : Optional<bool>(Runtime),
1251 UpperBound == -1 ? None : Optional<bool>(UpperBound),
1252 AllowPeeling == -1 ? None : Optional<bool>(AllowPeeling));
1255 Pass *llvm::createSimpleLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1256 bool ForgetAllSCEV) {
1257 return createLoopUnrollPass(OptLevel, OnlyWhenForced, ForgetAllSCEV, -1, -1,
1261 PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
1262 LoopStandardAnalysisResults &AR,
1263 LPMUpdater &Updater) {
1265 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
1266 Function *F = L.getHeader()->getParent();
1268 auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F);
1269 // FIXME: This should probably be optional rather than required.
1272 "LoopFullUnrollPass: OptimizationRemarkEmitterAnalysis not "
1273 "cached at a higher level");
1275 // Keep track of the previous loop structure so we can identify new loops
1276 // created by unrolling.
1277 Loop *ParentL = L.getParentLoop();
1278 SmallPtrSet<Loop *, 4> OldLoops;
1280 OldLoops.insert(ParentL->begin(), ParentL->end());
1282 OldLoops.insert(AR.LI.begin(), AR.LI.end());
1284 std::string LoopName = L.getName();
1287 tryToUnrollLoop(&L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, *ORE,
1288 /*BFI*/ nullptr, /*PSI*/ nullptr,
1289 /*PreserveLCSSA*/ true, OptLevel, OnlyWhenForced,
1290 ForgetSCEV, /*Count*/ None,
1291 /*Threshold*/ None, /*AllowPartial*/ false,
1292 /*Runtime*/ false, /*UpperBound*/ false,
1293 /*AllowPeeling*/ false) != LoopUnrollResult::Unmodified;
1295 return PreservedAnalyses::all();
1297 // The parent must not be damaged by unrolling!
1300 ParentL->verifyLoop();
1303 // Unrolling can do several things to introduce new loops into a loop nest:
1304 // - Full unrolling clones child loops within the current loop but then
1305 // removes the current loop making all of the children appear to be new
1308 // When a new loop appears as a sibling loop after fully unrolling,
1309 // its nesting structure has fundamentally changed and we want to revisit
1310 // it to reflect that.
1312 // When unrolling has removed the current loop, we need to tell the
1313 // infrastructure that it is gone.
1315 // Finally, we support a debugging/testing mode where we revisit child loops
1316 // as well. These are not expected to require further optimizations as either
1317 // they or the loop they were cloned from have been directly visited already.
1318 // But the debugging mode allows us to check this assumption.
1319 bool IsCurrentLoopValid = false;
1320 SmallVector<Loop *, 4> SibLoops;
1322 SibLoops.append(ParentL->begin(), ParentL->end());
1324 SibLoops.append(AR.LI.begin(), AR.LI.end());
1325 erase_if(SibLoops, [&](Loop *SibLoop) {
1326 if (SibLoop == &L) {
1327 IsCurrentLoopValid = true;
1331 // Otherwise erase the loop from the list if it was in the old loops.
1332 return OldLoops.count(SibLoop) != 0;
1334 Updater.addSiblingLoops(SibLoops);
1336 if (!IsCurrentLoopValid) {
1337 Updater.markLoopAsDeleted(L, LoopName);
1339 // We can only walk child loops if the current loop remained valid.
1340 if (UnrollRevisitChildLoops) {
1341 // Walk *all* of the child loops.
1342 SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end());
1343 Updater.addChildLoops(ChildLoops);
1347 return getLoopPassPreservedAnalyses();
1350 template <typename RangeT>
1351 static SmallVector<Loop *, 8> appendLoopsToWorklist(RangeT &&Loops) {
1352 SmallVector<Loop *, 8> Worklist;
1353 // We use an internal worklist to build up the preorder traversal without
1355 SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist;
1357 for (Loop *RootL : Loops) {
1358 assert(PreOrderLoops.empty() && "Must start with an empty preorder walk.");
1359 assert(PreOrderWorklist.empty() &&
1360 "Must start with an empty preorder walk worklist.");
1361 PreOrderWorklist.push_back(RootL);
1363 Loop *L = PreOrderWorklist.pop_back_val();
1364 PreOrderWorklist.append(L->begin(), L->end());
1365 PreOrderLoops.push_back(L);
1366 } while (!PreOrderWorklist.empty());
1368 Worklist.append(PreOrderLoops.begin(), PreOrderLoops.end());
1369 PreOrderLoops.clear();
1374 PreservedAnalyses LoopUnrollPass::run(Function &F,
1375 FunctionAnalysisManager &AM) {
1376 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1377 auto &LI = AM.getResult<LoopAnalysis>(F);
1378 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1379 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1380 auto &AC = AM.getResult<AssumptionAnalysis>(F);
1381 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1383 LoopAnalysisManager *LAM = nullptr;
1384 if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(F))
1385 LAM = &LAMProxy->getManager();
1387 const ModuleAnalysisManager &MAM =
1388 AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
1389 ProfileSummaryInfo *PSI =
1390 MAM.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
1391 auto *BFI = (PSI && PSI->hasProfileSummary()) ?
1392 &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
1394 bool Changed = false;
1396 // The unroller requires loops to be in simplified form, and also needs LCSSA.
1397 // Since simplification may add new inner loops, it has to run before the
1398 // legality and profitability checks. This means running the loop unroller
1399 // will simplify all loops, regardless of whether anything end up being
1401 for (auto &L : LI) {
1403 simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */);
1404 Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1407 SmallVector<Loop *, 8> Worklist = appendLoopsToWorklist(LI);
1409 while (!Worklist.empty()) {
1410 // Because the LoopInfo stores the loops in RPO, we walk the worklist
1411 // from back to front so that we work forward across the CFG, which
1412 // for unrolling is only needed to get optimization remarks emitted in
1414 Loop &L = *Worklist.pop_back_val();
1416 Loop *ParentL = L.getParentLoop();
1419 // Check if the profile summary indicates that the profiled application
1420 // has a huge working set size, in which case we disable peeling to avoid
1421 // bloating it further.
1422 Optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling;
1423 if (PSI && PSI->hasHugeWorkingSetSize())
1424 LocalAllowPeeling = false;
1425 std::string LoopName = L.getName();
1426 // The API here is quite complex to call and we allow to select some
1427 // flavors of unrolling during construction time (by setting UnrollOpts).
1428 LoopUnrollResult Result = tryToUnrollLoop(
1429 &L, DT, &LI, SE, TTI, AC, ORE, BFI, PSI,
1430 /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, UnrollOpts.OnlyWhenForced,
1431 UnrollOpts.ForgetSCEV, /*Count*/ None,
1432 /*Threshold*/ None, UnrollOpts.AllowPartial, UnrollOpts.AllowRuntime,
1433 UnrollOpts.AllowUpperBound, LocalAllowPeeling);
1434 Changed |= Result != LoopUnrollResult::Unmodified;
1436 // The parent must not be damaged by unrolling!
1438 if (Result != LoopUnrollResult::Unmodified && ParentL)
1439 ParentL->verifyLoop();
1442 // Clear any cached analysis results for L if we removed it completely.
1443 if (LAM && Result == LoopUnrollResult::FullyUnrolled)
1444 LAM->clear(L, LoopName);
1448 return PreservedAnalyses::all();
1450 return getLoopPassPreservedAnalyses();