1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements a simple loop unroller. It works best when loops have
11 // been canonicalized by the -indvars pass, allowing it to determine the trip
12 // counts of loops easily.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Scalar/LoopUnrollPass.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/Analysis/AssumptionCache.h"
18 #include "llvm/Analysis/CodeMetrics.h"
19 #include "llvm/Analysis/GlobalsModRef.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LoopPass.h"
22 #include "llvm/Analysis/LoopUnrollAnalyzer.h"
23 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/InstVisitor.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Transforms/Scalar.h"
35 #include "llvm/Transforms/Scalar/LoopPassManager.h"
36 #include "llvm/Transforms/Utils/LoopUtils.h"
37 #include "llvm/Transforms/Utils/UnrollLoop.h"
43 #define DEBUG_TYPE "loop-unroll"
45 static cl::opt<unsigned>
46 UnrollThreshold("unroll-threshold", cl::Hidden,
47 cl::desc("The cost threshold for loop unrolling"));
49 static cl::opt<unsigned> UnrollPartialThreshold(
50 "unroll-partial-threshold", cl::Hidden,
51 cl::desc("The cost threshold for partial loop unrolling"));
53 static cl::opt<unsigned> UnrollMaxPercentThresholdBoost(
54 "unroll-max-percent-threshold-boost", cl::init(400), cl::Hidden,
55 cl::desc("The maximum 'boost' (represented as a percentage >= 100) applied "
56 "to the threshold when aggressively unrolling a loop due to the "
57 "dynamic cost savings. If completely unrolling a loop will reduce "
58 "the total runtime from X to Y, we boost the loop unroll "
59 "threshold to DefaultThreshold*std::min(MaxPercentThresholdBoost, "
60 "X/Y). This limit avoids excessive code bloat."));
62 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
63 "unroll-max-iteration-count-to-analyze", cl::init(10), cl::Hidden,
64 cl::desc("Don't allow loop unrolling to simulate more than this number of"
65 "iterations when checking full unroll profitability"));
67 static cl::opt<unsigned> UnrollCount(
68 "unroll-count", cl::Hidden,
69 cl::desc("Use this unroll count for all loops including those with "
70 "unroll_count pragma values, for testing purposes"));
72 static cl::opt<unsigned> UnrollMaxCount(
73 "unroll-max-count", cl::Hidden,
74 cl::desc("Set the max unroll count for partial and runtime unrolling, for"
77 static cl::opt<unsigned> UnrollFullMaxCount(
78 "unroll-full-max-count", cl::Hidden,
80 "Set the max unroll count for full unrolling, for testing purposes"));
83 UnrollAllowPartial("unroll-allow-partial", cl::Hidden,
84 cl::desc("Allows loops to be partially unrolled until "
85 "-unroll-threshold loop size is reached."));
87 static cl::opt<bool> UnrollAllowRemainder(
88 "unroll-allow-remainder", cl::Hidden,
89 cl::desc("Allow generation of a loop remainder (extra iterations) "
90 "when unrolling a loop."));
93 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::Hidden,
94 cl::desc("Unroll loops with run-time trip counts"));
96 static cl::opt<unsigned> UnrollMaxUpperBound(
97 "unroll-max-upperbound", cl::init(8), cl::Hidden,
99 "The max of trip count upper bound that is considered in unrolling"));
101 static cl::opt<unsigned> PragmaUnrollThreshold(
102 "pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
103 cl::desc("Unrolled size limit for loops with an unroll(full) or "
104 "unroll_count pragma."));
106 static cl::opt<unsigned> FlatLoopTripCountThreshold(
107 "flat-loop-tripcount-threshold", cl::init(5), cl::Hidden,
108 cl::desc("If the runtime tripcount for the loop is lower than the "
109 "threshold, the loop is considered as flat and will be less "
110 "aggressively unrolled."));
113 UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden,
114 cl::desc("Allows loops to be peeled when the dynamic "
115 "trip count is known to be low."));
117 // This option isn't ever intended to be enabled, it serves to allow
118 // experiments to check the assumptions about when this kind of revisit is
120 static cl::opt<bool> UnrollRevisitChildLoops(
121 "unroll-revisit-child-loops", cl::Hidden,
122 cl::desc("Enqueue and re-visit child loops in the loop PM after unrolling. "
123 "This shouldn't typically be needed as child loops (or their "
124 "clones) were already visited."));
126 /// A magic value for use with the Threshold parameter to indicate
127 /// that the loop unroll should be performed regardless of how much
128 /// code expansion would result.
129 static const unsigned NoThreshold = UINT_MAX;
131 /// Gather the various unrolling parameters based on the defaults, compiler
132 /// flags, TTI overrides and user specified parameters.
133 static TargetTransformInfo::UnrollingPreferences gatherUnrollingPreferences(
134 Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, int OptLevel,
135 Optional<unsigned> UserThreshold, Optional<unsigned> UserCount,
136 Optional<bool> UserAllowPartial, Optional<bool> UserRuntime,
137 Optional<bool> UserUpperBound) {
138 TargetTransformInfo::UnrollingPreferences UP;
140 // Set up the defaults
141 UP.Threshold = OptLevel > 2 ? 300 : 150;
142 UP.MaxPercentThresholdBoost = 400;
143 UP.OptSizeThreshold = 0;
144 UP.PartialThreshold = 150;
145 UP.PartialOptSizeThreshold = 0;
148 UP.DefaultUnrollRuntimeCount = 8;
149 UP.MaxCount = UINT_MAX;
150 UP.FullUnrollMaxCount = UINT_MAX;
154 UP.AllowRemainder = true;
155 UP.AllowExpensiveTripCount = false;
157 UP.UpperBound = false;
158 UP.AllowPeeling = true;
160 // Override with any target specific settings
161 TTI.getUnrollingPreferences(L, SE, UP);
163 // Apply size attributes
164 if (L->getHeader()->getParent()->optForSize()) {
165 UP.Threshold = UP.OptSizeThreshold;
166 UP.PartialThreshold = UP.PartialOptSizeThreshold;
169 // Apply any user values specified by cl::opt
170 if (UnrollThreshold.getNumOccurrences() > 0)
171 UP.Threshold = UnrollThreshold;
172 if (UnrollPartialThreshold.getNumOccurrences() > 0)
173 UP.PartialThreshold = UnrollPartialThreshold;
174 if (UnrollMaxPercentThresholdBoost.getNumOccurrences() > 0)
175 UP.MaxPercentThresholdBoost = UnrollMaxPercentThresholdBoost;
176 if (UnrollMaxCount.getNumOccurrences() > 0)
177 UP.MaxCount = UnrollMaxCount;
178 if (UnrollFullMaxCount.getNumOccurrences() > 0)
179 UP.FullUnrollMaxCount = UnrollFullMaxCount;
180 if (UnrollAllowPartial.getNumOccurrences() > 0)
181 UP.Partial = UnrollAllowPartial;
182 if (UnrollAllowRemainder.getNumOccurrences() > 0)
183 UP.AllowRemainder = UnrollAllowRemainder;
184 if (UnrollRuntime.getNumOccurrences() > 0)
185 UP.Runtime = UnrollRuntime;
186 if (UnrollMaxUpperBound == 0)
187 UP.UpperBound = false;
188 if (UnrollAllowPeeling.getNumOccurrences() > 0)
189 UP.AllowPeeling = UnrollAllowPeeling;
191 // Apply user values provided by argument
192 if (UserThreshold.hasValue()) {
193 UP.Threshold = *UserThreshold;
194 UP.PartialThreshold = *UserThreshold;
196 if (UserCount.hasValue())
197 UP.Count = *UserCount;
198 if (UserAllowPartial.hasValue())
199 UP.Partial = *UserAllowPartial;
200 if (UserRuntime.hasValue())
201 UP.Runtime = *UserRuntime;
202 if (UserUpperBound.hasValue())
203 UP.UpperBound = *UserUpperBound;
209 /// A struct to densely store the state of an instruction after unrolling at
212 /// This is designed to work like a tuple of <Instruction *, int> for the
213 /// purposes of hashing and lookup, but to be able to associate two boolean
214 /// states with each key.
215 struct UnrolledInstState {
219 unsigned IsCounted : 1;
222 /// Hashing and equality testing for a set of the instruction states.
223 struct UnrolledInstStateKeyInfo {
224 typedef DenseMapInfo<Instruction *> PtrInfo;
225 typedef DenseMapInfo<std::pair<Instruction *, int>> PairInfo;
226 static inline UnrolledInstState getEmptyKey() {
227 return {PtrInfo::getEmptyKey(), 0, 0, 0};
229 static inline UnrolledInstState getTombstoneKey() {
230 return {PtrInfo::getTombstoneKey(), 0, 0, 0};
232 static inline unsigned getHashValue(const UnrolledInstState &S) {
233 return PairInfo::getHashValue({S.I, S.Iteration});
235 static inline bool isEqual(const UnrolledInstState &LHS,
236 const UnrolledInstState &RHS) {
237 return PairInfo::isEqual({LHS.I, LHS.Iteration}, {RHS.I, RHS.Iteration});
243 struct EstimatedUnrollCost {
244 /// \brief The estimated cost after unrolling.
245 unsigned UnrolledCost;
247 /// \brief The estimated dynamic cost of executing the instructions in the
249 unsigned RolledDynamicCost;
253 /// \brief Figure out if the loop is worth full unrolling.
255 /// Complete loop unrolling can make some loads constant, and we need to know
256 /// if that would expose any further optimization opportunities. This routine
257 /// estimates this optimization. It computes cost of unrolled loop
258 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
259 /// dynamic cost we mean that we won't count costs of blocks that are known not
260 /// to be executed (i.e. if we have a branch in the loop and we know that at the
261 /// given iteration its condition would be resolved to true, we won't add up the
262 /// cost of the 'false'-block).
263 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
264 /// the analysis failed (no benefits expected from the unrolling, or the loop is
265 /// too big to analyze), the returned value is None.
266 static Optional<EstimatedUnrollCost>
267 analyzeLoopUnrollCost(const Loop *L, unsigned TripCount, DominatorTree &DT,
268 ScalarEvolution &SE, const TargetTransformInfo &TTI,
269 unsigned MaxUnrolledLoopSize) {
270 // We want to be able to scale offsets by the trip count and add more offsets
271 // to them without checking for overflows, and we already don't want to
272 // analyze *massive* trip counts, so we force the max to be reasonably small.
273 assert(UnrollMaxIterationsCountToAnalyze < (INT_MAX / 2) &&
274 "The unroll iterations max is too large!");
276 // Only analyze inner loops. We can't properly estimate cost of nested loops
277 // and we won't visit inner loops again anyway.
281 // Don't simulate loops with a big or unknown tripcount
282 if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
283 TripCount > UnrollMaxIterationsCountToAnalyze)
286 SmallSetVector<BasicBlock *, 16> BBWorklist;
287 SmallSetVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitWorklist;
288 DenseMap<Value *, Constant *> SimplifiedValues;
289 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues;
291 // The estimated cost of the unrolled form of the loop. We try to estimate
292 // this by simplifying as much as we can while computing the estimate.
293 unsigned UnrolledCost = 0;
295 // We also track the estimated dynamic (that is, actually executed) cost in
296 // the rolled form. This helps identify cases when the savings from unrolling
297 // aren't just exposing dead control flows, but actual reduced dynamic
298 // instructions due to the simplifications which we expect to occur after
300 unsigned RolledDynamicCost = 0;
302 // We track the simplification of each instruction in each iteration. We use
303 // this to recursively merge costs into the unrolled cost on-demand so that
304 // we don't count the cost of any dead code. This is essentially a map from
305 // <instruction, int> to <bool, bool>, but stored as a densely packed struct.
306 DenseSet<UnrolledInstState, UnrolledInstStateKeyInfo> InstCostMap;
308 // A small worklist used to accumulate cost of instructions from each
309 // observable and reached root in the loop.
310 SmallVector<Instruction *, 16> CostWorklist;
312 // PHI-used worklist used between iterations while accumulating cost.
313 SmallVector<Instruction *, 4> PHIUsedList;
315 // Helper function to accumulate cost for instructions in the loop.
316 auto AddCostRecursively = [&](Instruction &RootI, int Iteration) {
317 assert(Iteration >= 0 && "Cannot have a negative iteration!");
318 assert(CostWorklist.empty() && "Must start with an empty cost list");
319 assert(PHIUsedList.empty() && "Must start with an empty phi used list");
320 CostWorklist.push_back(&RootI);
321 for (;; --Iteration) {
323 Instruction *I = CostWorklist.pop_back_val();
325 // InstCostMap only uses I and Iteration as a key, the other two values
326 // don't matter here.
327 auto CostIter = InstCostMap.find({I, Iteration, 0, 0});
328 if (CostIter == InstCostMap.end())
329 // If an input to a PHI node comes from a dead path through the loop
330 // we may have no cost data for it here. What that actually means is
333 auto &Cost = *CostIter;
335 // Already counted this instruction.
338 // Mark that we are counting the cost of this instruction now.
339 Cost.IsCounted = true;
341 // If this is a PHI node in the loop header, just add it to the PHI set.
342 if (auto *PhiI = dyn_cast<PHINode>(I))
343 if (PhiI->getParent() == L->getHeader()) {
344 assert(Cost.IsFree && "Loop PHIs shouldn't be evaluated as they "
345 "inherently simplify during unrolling.");
349 // Push the incoming value from the backedge into the PHI used list
350 // if it is an in-loop instruction. We'll use this to populate the
351 // cost worklist for the next iteration (as we count backwards).
352 if (auto *OpI = dyn_cast<Instruction>(
353 PhiI->getIncomingValueForBlock(L->getLoopLatch())))
354 if (L->contains(OpI))
355 PHIUsedList.push_back(OpI);
359 // First accumulate the cost of this instruction.
361 UnrolledCost += TTI.getUserCost(I);
362 DEBUG(dbgs() << "Adding cost of instruction (iteration " << Iteration
367 // We must count the cost of every operand which is not free,
368 // recursively. If we reach a loop PHI node, simply add it to the set
369 // to be considered on the next iteration (backwards!).
370 for (Value *Op : I->operands()) {
371 // Check whether this operand is free due to being a constant or
373 auto *OpI = dyn_cast<Instruction>(Op);
374 if (!OpI || !L->contains(OpI))
377 // Otherwise accumulate its cost.
378 CostWorklist.push_back(OpI);
380 } while (!CostWorklist.empty());
382 if (PHIUsedList.empty())
383 // We've exhausted the search.
386 assert(Iteration > 0 &&
387 "Cannot track PHI-used values past the first iteration!");
388 CostWorklist.append(PHIUsedList.begin(), PHIUsedList.end());
393 // Ensure that we don't violate the loop structure invariants relied on by
395 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
396 assert(L->isLCSSAForm(DT) &&
397 "Must have loops in LCSSA form to track live-out values.");
399 DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
401 // Simulate execution of each iteration of the loop counting instructions,
402 // which would be simplified.
403 // Since the same load will take different values on different iterations,
404 // we literally have to go through all loop's iterations.
405 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
406 DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
408 // Prepare for the iteration by collecting any simplified entry or backedge
410 for (Instruction &I : *L->getHeader()) {
411 auto *PHI = dyn_cast<PHINode>(&I);
415 // The loop header PHI nodes must have exactly two input: one from the
416 // loop preheader and one from the loop latch.
418 PHI->getNumIncomingValues() == 2 &&
419 "Must have an incoming value only for the preheader and the latch.");
421 Value *V = PHI->getIncomingValueForBlock(
422 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
423 Constant *C = dyn_cast<Constant>(V);
424 if (Iteration != 0 && !C)
425 C = SimplifiedValues.lookup(V);
427 SimplifiedInputValues.push_back({PHI, C});
430 // Now clear and re-populate the map for the next iteration.
431 SimplifiedValues.clear();
432 while (!SimplifiedInputValues.empty())
433 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val());
435 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L);
438 BBWorklist.insert(L->getHeader());
439 // Note that we *must not* cache the size, this loop grows the worklist.
440 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
441 BasicBlock *BB = BBWorklist[Idx];
443 // Visit all instructions in the given basic block and try to simplify
444 // it. We don't change the actual IR, just count optimization
446 for (Instruction &I : *BB) {
447 if (isa<DbgInfoIntrinsic>(I))
450 // Track this instruction's expected baseline cost when executing the
452 RolledDynamicCost += TTI.getUserCost(&I);
454 // Visit the instruction to analyze its loop cost after unrolling,
455 // and if the visitor returns true, mark the instruction as free after
456 // unrolling and continue.
457 bool IsFree = Analyzer.visit(I);
458 bool Inserted = InstCostMap.insert({&I, (int)Iteration,
460 /*IsCounted*/ false}).second;
462 assert(Inserted && "Cannot have a state for an unvisited instruction!");
467 // Can't properly model a cost of a call.
468 // FIXME: With a proper cost model we should be able to do it.
469 if(isa<CallInst>(&I))
472 // If the instruction might have a side-effect recursively account for
473 // the cost of it and all the instructions leading up to it.
474 if (I.mayHaveSideEffects())
475 AddCostRecursively(I, Iteration);
477 // If unrolled body turns out to be too big, bail out.
478 if (UnrolledCost > MaxUnrolledLoopSize) {
479 DEBUG(dbgs() << " Exceeded threshold.. exiting.\n"
480 << " UnrolledCost: " << UnrolledCost
481 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
487 TerminatorInst *TI = BB->getTerminator();
489 // Add in the live successors by first checking whether we have terminator
490 // that may be simplified based on the values simplified by this call.
491 BasicBlock *KnownSucc = nullptr;
492 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
493 if (BI->isConditional()) {
494 if (Constant *SimpleCond =
495 SimplifiedValues.lookup(BI->getCondition())) {
496 // Just take the first successor if condition is undef
497 if (isa<UndefValue>(SimpleCond))
498 KnownSucc = BI->getSuccessor(0);
499 else if (ConstantInt *SimpleCondVal =
500 dyn_cast<ConstantInt>(SimpleCond))
501 KnownSucc = BI->getSuccessor(SimpleCondVal->isZero() ? 1 : 0);
504 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
505 if (Constant *SimpleCond =
506 SimplifiedValues.lookup(SI->getCondition())) {
507 // Just take the first successor if condition is undef
508 if (isa<UndefValue>(SimpleCond))
509 KnownSucc = SI->getSuccessor(0);
510 else if (ConstantInt *SimpleCondVal =
511 dyn_cast<ConstantInt>(SimpleCond))
512 KnownSucc = SI->findCaseValue(SimpleCondVal)->getCaseSuccessor();
516 if (L->contains(KnownSucc))
517 BBWorklist.insert(KnownSucc);
519 ExitWorklist.insert({BB, KnownSucc});
523 // Add BB's successors to the worklist.
524 for (BasicBlock *Succ : successors(BB))
525 if (L->contains(Succ))
526 BBWorklist.insert(Succ);
528 ExitWorklist.insert({BB, Succ});
529 AddCostRecursively(*TI, Iteration);
532 // If we found no optimization opportunities on the first iteration, we
533 // won't find them on later ones too.
534 if (UnrolledCost == RolledDynamicCost) {
535 DEBUG(dbgs() << " No opportunities found.. exiting.\n"
536 << " UnrolledCost: " << UnrolledCost << "\n");
541 while (!ExitWorklist.empty()) {
542 BasicBlock *ExitingBB, *ExitBB;
543 std::tie(ExitingBB, ExitBB) = ExitWorklist.pop_back_val();
545 for (Instruction &I : *ExitBB) {
546 auto *PN = dyn_cast<PHINode>(&I);
550 Value *Op = PN->getIncomingValueForBlock(ExitingBB);
551 if (auto *OpI = dyn_cast<Instruction>(Op))
552 if (L->contains(OpI))
553 AddCostRecursively(*OpI, TripCount - 1);
557 DEBUG(dbgs() << "Analysis finished:\n"
558 << "UnrolledCost: " << UnrolledCost << ", "
559 << "RolledDynamicCost: " << RolledDynamicCost << "\n");
560 return {{UnrolledCost, RolledDynamicCost}};
563 /// ApproximateLoopSize - Approximate the size of the loop.
564 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
565 bool &NotDuplicatable, bool &Convergent,
566 const TargetTransformInfo &TTI,
567 AssumptionCache *AC, unsigned BEInsns) {
568 SmallPtrSet<const Value *, 32> EphValues;
569 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
572 for (BasicBlock *BB : L->blocks())
573 Metrics.analyzeBasicBlock(BB, TTI, EphValues);
574 NumCalls = Metrics.NumInlineCandidates;
575 NotDuplicatable = Metrics.notDuplicatable;
576 Convergent = Metrics.convergent;
578 unsigned LoopSize = Metrics.NumInsts;
580 // Don't allow an estimate of size zero. This would allows unrolling of loops
581 // with huge iteration counts, which is a compile time problem even if it's
582 // not a problem for code quality. Also, the code using this size may assume
583 // that each loop has at least three instructions (likely a conditional
584 // branch, a comparison feeding that branch, and some kind of loop increment
585 // feeding that comparison instruction).
586 LoopSize = std::max(LoopSize, BEInsns + 1);
591 // Returns the loop hint metadata node with the given name (for example,
592 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
594 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
595 if (MDNode *LoopID = L->getLoopID())
596 return GetUnrollMetadata(LoopID, Name);
600 // Returns true if the loop has an unroll(full) pragma.
601 static bool HasUnrollFullPragma(const Loop *L) {
602 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
605 // Returns true if the loop has an unroll(enable) pragma. This metadata is used
606 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
607 static bool HasUnrollEnablePragma(const Loop *L) {
608 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.enable");
611 // Returns true if the loop has an unroll(disable) pragma.
612 static bool HasUnrollDisablePragma(const Loop *L) {
613 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
616 // Returns true if the loop has an runtime unroll(disable) pragma.
617 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
618 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
621 // If loop has an unroll_count pragma return the (necessarily
622 // positive) value from the pragma. Otherwise return 0.
623 static unsigned UnrollCountPragmaValue(const Loop *L) {
624 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
626 assert(MD->getNumOperands() == 2 &&
627 "Unroll count hint metadata should have two operands.");
629 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
630 assert(Count >= 1 && "Unroll count must be positive.");
636 // Remove existing unroll metadata and add unroll disable metadata to
637 // indicate the loop has already been unrolled. This prevents a loop
638 // from being unrolled more than is directed by a pragma if the loop
639 // unrolling pass is run more than once (which it generally is).
640 static void SetLoopAlreadyUnrolled(Loop *L) {
641 MDNode *LoopID = L->getLoopID();
642 // First remove any existing loop unrolling metadata.
643 SmallVector<Metadata *, 4> MDs;
644 // Reserve first location for self reference to the LoopID metadata node.
645 MDs.push_back(nullptr);
648 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
649 bool IsUnrollMetadata = false;
650 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
652 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
653 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
655 if (!IsUnrollMetadata)
656 MDs.push_back(LoopID->getOperand(i));
660 // Add unroll(disable) metadata to disable future unrolling.
661 LLVMContext &Context = L->getHeader()->getContext();
662 SmallVector<Metadata *, 1> DisableOperands;
663 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
664 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
665 MDs.push_back(DisableNode);
667 MDNode *NewLoopID = MDNode::get(Context, MDs);
668 // Set operand 0 to refer to the loop id itself.
669 NewLoopID->replaceOperandWith(0, NewLoopID);
670 L->setLoopID(NewLoopID);
673 // Computes the boosting factor for complete unrolling.
674 // If fully unrolling the loop would save a lot of RolledDynamicCost, it would
675 // be beneficial to fully unroll the loop even if unrolledcost is large. We
676 // use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust
677 // the unroll threshold.
678 static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost,
679 unsigned MaxPercentThresholdBoost) {
680 if (Cost.RolledDynamicCost >= UINT_MAX / 100)
682 else if (Cost.UnrolledCost != 0)
683 // The boosting factor is RolledDynamicCost / UnrolledCost
684 return std::min(100 * Cost.RolledDynamicCost / Cost.UnrolledCost,
685 MaxPercentThresholdBoost);
687 return MaxPercentThresholdBoost;
690 // Returns loop size estimation for unrolled loop.
691 static uint64_t getUnrolledLoopSize(
693 TargetTransformInfo::UnrollingPreferences &UP) {
694 assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
695 return (uint64_t)(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns;
698 // Returns true if unroll count was set explicitly.
699 // Calculates unroll count and writes it to UP.Count.
700 static bool computeUnrollCount(
701 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI,
702 ScalarEvolution &SE, OptimizationRemarkEmitter *ORE, unsigned &TripCount,
703 unsigned MaxTripCount, unsigned &TripMultiple, unsigned LoopSize,
704 TargetTransformInfo::UnrollingPreferences &UP, bool &UseUpperBound) {
705 // Check for explicit Count.
706 // 1st priority is unroll count set by "unroll-count" option.
707 bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0;
708 if (UserUnrollCount) {
709 UP.Count = UnrollCount;
710 UP.AllowExpensiveTripCount = true;
712 if (UP.AllowRemainder && getUnrolledLoopSize(LoopSize, UP) < UP.Threshold)
716 // 2nd priority is unroll count set by pragma.
717 unsigned PragmaCount = UnrollCountPragmaValue(L);
718 if (PragmaCount > 0) {
719 UP.Count = PragmaCount;
721 UP.AllowExpensiveTripCount = true;
723 if (UP.AllowRemainder &&
724 getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
727 bool PragmaFullUnroll = HasUnrollFullPragma(L);
728 if (PragmaFullUnroll && TripCount != 0) {
729 UP.Count = TripCount;
730 if (getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
734 bool PragmaEnableUnroll = HasUnrollEnablePragma(L);
735 bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll ||
736 PragmaEnableUnroll || UserUnrollCount;
738 if (ExplicitUnroll && TripCount != 0) {
739 // If the loop has an unrolling pragma, we want to be more aggressive with
740 // unrolling limits. Set thresholds to at least the PragmaThreshold value
741 // which is larger than the default limits.
742 UP.Threshold = std::max<unsigned>(UP.Threshold, PragmaUnrollThreshold);
743 UP.PartialThreshold =
744 std::max<unsigned>(UP.PartialThreshold, PragmaUnrollThreshold);
747 // 3rd priority is full unroll count.
748 // Full unroll makes sense only when TripCount or its upper bound could be
749 // statically calculated.
750 // Also we need to check if we exceed FullUnrollMaxCount.
751 // If using the upper bound to unroll, TripMultiple should be set to 1 because
752 // we do not know when loop may exit.
753 // MaxTripCount and ExactTripCount cannot both be non zero since we only
754 // compute the former when the latter is zero.
755 unsigned ExactTripCount = TripCount;
756 assert((ExactTripCount == 0 || MaxTripCount == 0) &&
757 "ExtractTripCound and MaxTripCount cannot both be non zero.");
758 unsigned FullUnrollTripCount = ExactTripCount ? ExactTripCount : MaxTripCount;
759 UP.Count = FullUnrollTripCount;
760 if (FullUnrollTripCount && FullUnrollTripCount <= UP.FullUnrollMaxCount) {
761 // When computing the unrolled size, note that BEInsns are not replicated
762 // like the rest of the loop body.
763 if (getUnrolledLoopSize(LoopSize, UP) < UP.Threshold) {
764 UseUpperBound = (MaxTripCount == FullUnrollTripCount);
765 TripCount = FullUnrollTripCount;
766 TripMultiple = UP.UpperBound ? 1 : TripMultiple;
767 return ExplicitUnroll;
769 // The loop isn't that small, but we still can fully unroll it if that
770 // helps to remove a significant number of instructions.
771 // To check that, run additional analysis on the loop.
772 if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
773 L, FullUnrollTripCount, DT, SE, TTI,
774 UP.Threshold * UP.MaxPercentThresholdBoost / 100)) {
776 getFullUnrollBoostingFactor(*Cost, UP.MaxPercentThresholdBoost);
777 if (Cost->UnrolledCost < UP.Threshold * Boost / 100) {
778 UseUpperBound = (MaxTripCount == FullUnrollTripCount);
779 TripCount = FullUnrollTripCount;
780 TripMultiple = UP.UpperBound ? 1 : TripMultiple;
781 return ExplicitUnroll;
787 // 4th priority is loop peeling
788 computePeelCount(L, LoopSize, UP, TripCount);
792 return ExplicitUnroll;
795 // 5th priority is partial unrolling.
796 // Try partial unroll only when TripCount could be staticaly calculated.
798 UP.Partial |= ExplicitUnroll;
800 DEBUG(dbgs() << " will not try to unroll partially because "
801 << "-unroll-allow-partial not given\n");
806 UP.Count = TripCount;
807 if (UP.PartialThreshold != NoThreshold) {
808 // Reduce unroll count to be modulo of TripCount for partial unrolling.
809 if (getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
811 (std::max(UP.PartialThreshold, UP.BEInsns + 1) - UP.BEInsns) /
812 (LoopSize - UP.BEInsns);
813 if (UP.Count > UP.MaxCount)
814 UP.Count = UP.MaxCount;
815 while (UP.Count != 0 && TripCount % UP.Count != 0)
817 if (UP.AllowRemainder && UP.Count <= 1) {
818 // If there is no Count that is modulo of TripCount, set Count to
819 // largest power-of-two factor that satisfies the threshold limit.
820 // As we'll create fixup loop, do the type of unrolling only if
821 // remainder loop is allowed.
822 UP.Count = UP.DefaultUnrollRuntimeCount;
823 while (UP.Count != 0 &&
824 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
828 if (PragmaEnableUnroll)
830 OptimizationRemarkMissed(DEBUG_TYPE, "UnrollAsDirectedTooLarge",
831 L->getStartLoc(), L->getHeader())
832 << "Unable to unroll loop as directed by unroll(enable) pragma "
833 "because unrolled size is too large.");
837 UP.Count = TripCount;
839 if (UP.Count > UP.MaxCount)
840 UP.Count = UP.MaxCount;
841 if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
842 UP.Count != TripCount)
844 OptimizationRemarkMissed(DEBUG_TYPE, "FullUnrollAsDirectedTooLarge",
845 L->getStartLoc(), L->getHeader())
846 << "Unable to fully unroll loop as directed by unroll pragma because "
847 "unrolled size is too large.");
848 return ExplicitUnroll;
850 assert(TripCount == 0 &&
851 "All cases when TripCount is constant should be covered here.");
852 if (PragmaFullUnroll)
854 OptimizationRemarkMissed(DEBUG_TYPE,
855 "CantFullUnrollAsDirectedRuntimeTripCount",
856 L->getStartLoc(), L->getHeader())
857 << "Unable to fully unroll loop as directed by unroll(full) pragma "
858 "because loop has a runtime trip count.");
860 // 6th priority is runtime unrolling.
861 // Don't unroll a runtime trip count loop when it is disabled.
862 if (HasRuntimeUnrollDisablePragma(L)) {
867 // Check if the runtime trip count is too small when profile is available.
868 if (L->getHeader()->getParent()->getEntryCount()) {
869 if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) {
870 if (*ProfileTripCount < FlatLoopTripCountThreshold)
873 UP.AllowExpensiveTripCount = true;
877 // Reduce count based on the type of unrolling and the threshold values.
878 UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount;
880 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
881 << "-unroll-runtime not given\n");
886 UP.Count = UP.DefaultUnrollRuntimeCount;
888 // Reduce unroll count to be the largest power-of-two factor of
889 // the original count which satisfies the threshold limit.
890 while (UP.Count != 0 &&
891 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
895 unsigned OrigCount = UP.Count;
898 if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) {
899 while (UP.Count != 0 && TripMultiple % UP.Count != 0)
901 DEBUG(dbgs() << "Remainder loop is restricted (that could architecture "
902 "specific or because the loop contains a convergent "
903 "instruction), so unroll count must divide the trip "
905 << TripMultiple << ". Reducing unroll count from "
906 << OrigCount << " to " << UP.Count << ".\n");
908 if (PragmaCount > 0 && !UP.AllowRemainder)
910 OptimizationRemarkMissed(DEBUG_TYPE,
911 "DifferentUnrollCountFromDirected",
912 L->getStartLoc(), L->getHeader())
913 << "Unable to unroll loop the number of times directed by "
914 "unroll_count pragma because remainder loop is restricted "
915 "(that could architecture specific or because the loop "
916 "contains a convergent instruction) and so must have an unroll "
917 "count that divides the loop trip multiple of "
918 << NV("TripMultiple", TripMultiple) << ". Unrolling instead "
919 << NV("UnrollCount", UP.Count) << " time(s).");
922 if (UP.Count > UP.MaxCount)
923 UP.Count = UP.MaxCount;
924 DEBUG(dbgs() << " partially unrolling with count: " << UP.Count << "\n");
927 return ExplicitUnroll;
930 static bool tryToUnrollLoop(Loop *L, DominatorTree &DT, LoopInfo *LI,
931 ScalarEvolution &SE, const TargetTransformInfo &TTI,
932 AssumptionCache &AC, OptimizationRemarkEmitter &ORE,
933 bool PreserveLCSSA, int OptLevel,
934 Optional<unsigned> ProvidedCount,
935 Optional<unsigned> ProvidedThreshold,
936 Optional<bool> ProvidedAllowPartial,
937 Optional<bool> ProvidedRuntime,
938 Optional<bool> ProvidedUpperBound) {
939 DEBUG(dbgs() << "Loop Unroll: F[" << L->getHeader()->getParent()->getName()
940 << "] Loop %" << L->getHeader()->getName() << "\n");
941 if (HasUnrollDisablePragma(L))
943 if (!L->isLoopSimplifyForm()) {
945 dbgs() << " Not unrolling loop which is not in loop-simplify form.\n");
949 unsigned NumInlineCandidates;
950 bool NotDuplicatable;
952 TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
953 L, SE, TTI, OptLevel, ProvidedThreshold, ProvidedCount,
954 ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound);
955 // Exit early if unrolling is disabled.
956 if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0))
958 unsigned LoopSize = ApproximateLoopSize(
959 L, NumInlineCandidates, NotDuplicatable, Convergent, TTI, &AC, UP.BEInsns);
960 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
961 if (NotDuplicatable) {
962 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
963 << " instructions.\n");
966 if (NumInlineCandidates != 0) {
967 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
971 // Find trip count and trip multiple if count is not available
972 unsigned TripCount = 0;
973 unsigned MaxTripCount = 0;
974 unsigned TripMultiple = 1;
975 // If there are multiple exiting blocks but one of them is the latch, use the
976 // latch for the trip count estimation. Otherwise insist on a single exiting
977 // block for the trip count estimation.
978 BasicBlock *ExitingBlock = L->getLoopLatch();
979 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
980 ExitingBlock = L->getExitingBlock();
982 TripCount = SE.getSmallConstantTripCount(L, ExitingBlock);
983 TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock);
986 // If the loop contains a convergent operation, the prelude we'd add
987 // to do the first few instructions before we hit the unrolled loop
988 // is unsafe -- it adds a control-flow dependency to the convergent
989 // operation. Therefore restrict remainder loop (try unrollig without).
991 // TODO: This is quite conservative. In practice, convergent_op()
992 // is likely to be called unconditionally in the loop. In this
993 // case, the program would be ill-formed (on most architectures)
994 // unless n were the same on all threads in a thread group.
995 // Assuming n is the same on all threads, any kind of unrolling is
996 // safe. But currently llvm's notion of convergence isn't powerful
997 // enough to express this.
999 UP.AllowRemainder = false;
1001 // Try to find the trip count upper bound if we cannot find the exact trip
1003 bool MaxOrZero = false;
1005 MaxTripCount = SE.getSmallConstantMaxTripCount(L);
1006 MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L);
1007 // We can unroll by the upper bound amount if it's generally allowed or if
1008 // we know that the loop is executed either the upper bound or zero times.
1009 // (MaxOrZero unrolling keeps only the first loop test, so the number of
1010 // loop tests remains the same compared to the non-unrolled version, whereas
1011 // the generic upper bound unrolling keeps all but the last loop test so the
1012 // number of loop tests goes up which may end up being worse on targets with
1013 // constriained branch predictor resources so is controlled by an option.)
1014 // In addition we only unroll small upper bounds.
1015 if (!(UP.UpperBound || MaxOrZero) || MaxTripCount > UnrollMaxUpperBound) {
1020 // computeUnrollCount() decides whether it is beneficial to use upper bound to
1021 // fully unroll the loop.
1022 bool UseUpperBound = false;
1023 bool IsCountSetExplicitly =
1024 computeUnrollCount(L, TTI, DT, LI, SE, &ORE, TripCount, MaxTripCount,
1025 TripMultiple, LoopSize, UP, UseUpperBound);
1028 // Unroll factor (Count) must be less or equal to TripCount.
1029 if (TripCount && UP.Count > TripCount)
1030 UP.Count = TripCount;
1033 if (!UnrollLoop(L, UP.Count, TripCount, UP.Force, UP.Runtime,
1034 UP.AllowExpensiveTripCount, UseUpperBound, MaxOrZero,
1035 TripMultiple, UP.PeelCount, LI, &SE, &DT, &AC, &ORE,
1039 // If loop has an unroll count pragma or unrolled by explicitly set count
1040 // mark loop as unrolled to prevent unrolling beyond that requested.
1041 // If the loop was peeled, we already "used up" the profile information
1042 // we had, so we don't want to unroll or peel again.
1043 if (IsCountSetExplicitly || UP.PeelCount)
1044 SetLoopAlreadyUnrolled(L);
1050 class LoopUnroll : public LoopPass {
1052 static char ID; // Pass ID, replacement for typeid
1053 LoopUnroll(int OptLevel = 2, Optional<unsigned> Threshold = None,
1054 Optional<unsigned> Count = None,
1055 Optional<bool> AllowPartial = None, Optional<bool> Runtime = None,
1056 Optional<bool> UpperBound = None)
1057 : LoopPass(ID), OptLevel(OptLevel), ProvidedCount(std::move(Count)),
1058 ProvidedThreshold(Threshold), ProvidedAllowPartial(AllowPartial),
1059 ProvidedRuntime(Runtime), ProvidedUpperBound(UpperBound) {
1060 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
1064 Optional<unsigned> ProvidedCount;
1065 Optional<unsigned> ProvidedThreshold;
1066 Optional<bool> ProvidedAllowPartial;
1067 Optional<bool> ProvidedRuntime;
1068 Optional<bool> ProvidedUpperBound;
1070 bool runOnLoop(Loop *L, LPPassManager &) override {
1074 Function &F = *L->getHeader()->getParent();
1076 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1077 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1078 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1079 const TargetTransformInfo &TTI =
1080 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1081 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1082 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
1083 // pass. Function analyses need to be preserved across loop transformations
1084 // but ORE cannot be preserved (see comment before the pass definition).
1085 OptimizationRemarkEmitter ORE(&F);
1086 bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1088 return tryToUnrollLoop(L, DT, LI, SE, TTI, AC, ORE, PreserveLCSSA, OptLevel,
1089 ProvidedCount, ProvidedThreshold,
1090 ProvidedAllowPartial, ProvidedRuntime,
1091 ProvidedUpperBound);
1094 /// This transformation requires natural loop information & requires that
1095 /// loop preheaders be inserted into the CFG...
1097 void getAnalysisUsage(AnalysisUsage &AU) const override {
1098 AU.addRequired<AssumptionCacheTracker>();
1099 AU.addRequired<TargetTransformInfoWrapperPass>();
1100 // FIXME: Loop passes are required to preserve domtree, and for now we just
1101 // recreate dom info if anything gets unrolled.
1102 getLoopAnalysisUsage(AU);
1107 char LoopUnroll::ID = 0;
1108 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1109 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1110 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1111 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1112 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1114 Pass *llvm::createLoopUnrollPass(int OptLevel, int Threshold, int Count,
1115 int AllowPartial, int Runtime,
1117 // TODO: It would make more sense for this function to take the optionals
1118 // directly, but that's dangerous since it would silently break out of tree
1120 return new LoopUnroll(
1121 OptLevel, Threshold == -1 ? None : Optional<unsigned>(Threshold),
1122 Count == -1 ? None : Optional<unsigned>(Count),
1123 AllowPartial == -1 ? None : Optional<bool>(AllowPartial),
1124 Runtime == -1 ? None : Optional<bool>(Runtime),
1125 UpperBound == -1 ? None : Optional<bool>(UpperBound));
1128 Pass *llvm::createSimpleLoopUnrollPass(int OptLevel) {
1129 return llvm::createLoopUnrollPass(OptLevel, -1, -1, 0, 0, 0);
1132 PreservedAnalyses LoopUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
1133 LoopStandardAnalysisResults &AR,
1134 LPMUpdater &Updater) {
1136 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
1137 Function *F = L.getHeader()->getParent();
1139 auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F);
1140 // FIXME: This should probably be optional rather than required.
1142 report_fatal_error("LoopUnrollPass: OptimizationRemarkEmitterAnalysis not "
1143 "cached at a higher level");
1145 // Keep track of the previous loop structure so we can identify new loops
1146 // created by unrolling.
1147 Loop *ParentL = L.getParentLoop();
1148 SmallPtrSet<Loop *, 4> OldLoops;
1150 OldLoops.insert(ParentL->begin(), ParentL->end());
1152 OldLoops.insert(AR.LI.begin(), AR.LI.end());
1154 // The API here is quite complex to call, but there are only two interesting
1155 // states we support: partial and full (or "simple") unrolling. However, to
1156 // enable these things we actually pass "None" in for the optional to avoid
1157 // providing an explicit choice.
1158 Optional<bool> AllowPartialParam, RuntimeParam, UpperBoundParam;
1159 if (!AllowPartialUnrolling)
1160 AllowPartialParam = RuntimeParam = UpperBoundParam = false;
1161 bool Changed = tryToUnrollLoop(
1162 &L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, *ORE,
1163 /*PreserveLCSSA*/ true, OptLevel, /*Count*/ None,
1164 /*Threshold*/ None, AllowPartialParam, RuntimeParam, UpperBoundParam);
1166 return PreservedAnalyses::all();
1168 // The parent must not be damaged by unrolling!
1171 ParentL->verifyLoop();
1174 // Unrolling can do several things to introduce new loops into a loop nest:
1175 // - Partial unrolling clones child loops within the current loop. If it
1176 // uses a remainder, then it can also create any number of sibling loops.
1177 // - Full unrolling clones child loops within the current loop but then
1178 // removes the current loop making all of the children appear to be new
1180 // - Loop peeling can directly introduce new sibling loops by peeling one
1183 // When a new loop appears as a sibling loop, either from peeling an
1184 // iteration or fully unrolling, its nesting structure has fundamentally
1185 // changed and we want to revisit it to reflect that.
1187 // When unrolling has removed the current loop, we need to tell the
1188 // infrastructure that it is gone.
1190 // Finally, we support a debugging/testing mode where we revisit child loops
1191 // as well. These are not expected to require further optimizations as either
1192 // they or the loop they were cloned from have been directly visited already.
1193 // But the debugging mode allows us to check this assumption.
1194 bool IsCurrentLoopValid = false;
1195 SmallVector<Loop *, 4> SibLoops;
1197 SibLoops.append(ParentL->begin(), ParentL->end());
1199 SibLoops.append(AR.LI.begin(), AR.LI.end());
1200 erase_if(SibLoops, [&](Loop *SibLoop) {
1201 if (SibLoop == &L) {
1202 IsCurrentLoopValid = true;
1206 // Otherwise erase the loop from the list if it was in the old loops.
1207 return OldLoops.count(SibLoop) != 0;
1209 Updater.addSiblingLoops(SibLoops);
1211 if (!IsCurrentLoopValid) {
1212 Updater.markLoopAsDeleted(L);
1214 // We can only walk child loops if the current loop remained valid.
1215 if (UnrollRevisitChildLoops) {
1216 // Walk *all* of the child loops. This is a highly speculative mode
1217 // anyways so look for any simplifications that arose from partial
1218 // unrolling or peeling off of iterations.
1219 SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end());
1220 Updater.addChildLoops(ChildLoops);
1224 return getLoopPassPreservedAnalyses();