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/DenseMap.h"
17 #include "llvm/ADT/DenseMapInfo.h"
18 #include "llvm/ADT/DenseSet.h"
19 #include "llvm/ADT/None.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/Analysis/AssumptionCache.h"
27 #include "llvm/Analysis/CodeMetrics.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/LoopInfo.h"
30 #include "llvm/Analysis/LoopPass.h"
31 #include "llvm/Analysis/LoopUnrollAnalyzer.h"
32 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
33 #include "llvm/Analysis/ProfileSummaryInfo.h"
34 #include "llvm/Analysis/ScalarEvolution.h"
35 #include "llvm/Analysis/TargetTransformInfo.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CFG.h"
38 #include "llvm/IR/Constant.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DiagnosticInfo.h"
41 #include "llvm/IR/Dominators.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Metadata.h"
47 #include "llvm/IR/PassManager.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/Casting.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/Transforms/Scalar/LoopPassManager.h"
56 #include "llvm/Transforms/Utils.h"
57 #include "llvm/Transforms/Utils/LoopSimplify.h"
58 #include "llvm/Transforms/Utils/LoopUtils.h"
59 #include "llvm/Transforms/Utils/UnrollLoop.h"
70 #define DEBUG_TYPE "loop-unroll"
72 static cl::opt<unsigned>
73 UnrollThreshold("unroll-threshold", cl::Hidden,
74 cl::desc("The cost threshold for loop unrolling"));
76 static cl::opt<unsigned> UnrollPartialThreshold(
77 "unroll-partial-threshold", cl::Hidden,
78 cl::desc("The cost threshold for partial loop unrolling"));
80 static cl::opt<unsigned> UnrollMaxPercentThresholdBoost(
81 "unroll-max-percent-threshold-boost", cl::init(400), cl::Hidden,
82 cl::desc("The maximum 'boost' (represented as a percentage >= 100) applied "
83 "to the threshold when aggressively unrolling a loop due to the "
84 "dynamic cost savings. If completely unrolling a loop will reduce "
85 "the total runtime from X to Y, we boost the loop unroll "
86 "threshold to DefaultThreshold*std::min(MaxPercentThresholdBoost, "
87 "X/Y). This limit avoids excessive code bloat."));
89 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
90 "unroll-max-iteration-count-to-analyze", cl::init(10), cl::Hidden,
91 cl::desc("Don't allow loop unrolling to simulate more than this number of"
92 "iterations when checking full unroll profitability"));
94 static cl::opt<unsigned> UnrollCount(
95 "unroll-count", cl::Hidden,
96 cl::desc("Use this unroll count for all loops including those with "
97 "unroll_count pragma values, for testing purposes"));
99 static cl::opt<unsigned> UnrollMaxCount(
100 "unroll-max-count", cl::Hidden,
101 cl::desc("Set the max unroll count for partial and runtime unrolling, for"
102 "testing purposes"));
104 static cl::opt<unsigned> UnrollFullMaxCount(
105 "unroll-full-max-count", cl::Hidden,
107 "Set the max unroll count for full unrolling, for testing purposes"));
109 static cl::opt<unsigned> UnrollPeelCount(
110 "unroll-peel-count", cl::Hidden,
111 cl::desc("Set the unroll peeling count, for testing purposes"));
114 UnrollAllowPartial("unroll-allow-partial", cl::Hidden,
115 cl::desc("Allows loops to be partially unrolled until "
116 "-unroll-threshold loop size is reached."));
118 static cl::opt<bool> UnrollAllowRemainder(
119 "unroll-allow-remainder", cl::Hidden,
120 cl::desc("Allow generation of a loop remainder (extra iterations) "
121 "when unrolling a loop."));
124 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::Hidden,
125 cl::desc("Unroll loops with run-time trip counts"));
127 static cl::opt<unsigned> UnrollMaxUpperBound(
128 "unroll-max-upperbound", cl::init(8), cl::Hidden,
130 "The max of trip count upper bound that is considered in unrolling"));
132 static cl::opt<unsigned> PragmaUnrollThreshold(
133 "pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
134 cl::desc("Unrolled size limit for loops with an unroll(full) or "
135 "unroll_count pragma."));
137 static cl::opt<unsigned> FlatLoopTripCountThreshold(
138 "flat-loop-tripcount-threshold", cl::init(5), cl::Hidden,
139 cl::desc("If the runtime tripcount for the loop is lower than the "
140 "threshold, the loop is considered as flat and will be less "
141 "aggressively unrolled."));
144 UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden,
145 cl::desc("Allows loops to be peeled when the dynamic "
146 "trip count is known to be low."));
148 static cl::opt<bool> UnrollUnrollRemainder(
149 "unroll-remainder", cl::Hidden,
150 cl::desc("Allow the loop remainder to be unrolled."));
152 // This option isn't ever intended to be enabled, it serves to allow
153 // experiments to check the assumptions about when this kind of revisit is
155 static cl::opt<bool> UnrollRevisitChildLoops(
156 "unroll-revisit-child-loops", cl::Hidden,
157 cl::desc("Enqueue and re-visit child loops in the loop PM after unrolling. "
158 "This shouldn't typically be needed as child loops (or their "
159 "clones) were already visited."));
161 /// A magic value for use with the Threshold parameter to indicate
162 /// that the loop unroll should be performed regardless of how much
163 /// code expansion would result.
164 static const unsigned NoThreshold = std::numeric_limits<unsigned>::max();
166 /// Gather the various unrolling parameters based on the defaults, compiler
167 /// flags, TTI overrides and user specified parameters.
168 TargetTransformInfo::UnrollingPreferences llvm::gatherUnrollingPreferences(
169 Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, int OptLevel,
170 Optional<unsigned> UserThreshold, Optional<unsigned> UserCount,
171 Optional<bool> UserAllowPartial, Optional<bool> UserRuntime,
172 Optional<bool> UserUpperBound, Optional<bool> UserAllowPeeling) {
173 TargetTransformInfo::UnrollingPreferences UP;
175 // Set up the defaults
176 UP.Threshold = OptLevel > 2 ? 300 : 150;
177 UP.MaxPercentThresholdBoost = 400;
178 UP.OptSizeThreshold = 0;
179 UP.PartialThreshold = 150;
180 UP.PartialOptSizeThreshold = 0;
183 UP.DefaultUnrollRuntimeCount = 8;
184 UP.MaxCount = std::numeric_limits<unsigned>::max();
185 UP.FullUnrollMaxCount = std::numeric_limits<unsigned>::max();
189 UP.AllowRemainder = true;
190 UP.UnrollRemainder = false;
191 UP.AllowExpensiveTripCount = false;
193 UP.UpperBound = false;
194 UP.AllowPeeling = true;
195 UP.UnrollAndJam = false;
196 UP.UnrollAndJamInnerLoopThreshold = 60;
198 // Override with any target specific settings
199 TTI.getUnrollingPreferences(L, SE, UP);
201 // Apply size attributes
202 if (L->getHeader()->getParent()->optForSize()) {
203 UP.Threshold = UP.OptSizeThreshold;
204 UP.PartialThreshold = UP.PartialOptSizeThreshold;
207 // Apply any user values specified by cl::opt
208 if (UnrollThreshold.getNumOccurrences() > 0)
209 UP.Threshold = UnrollThreshold;
210 if (UnrollPartialThreshold.getNumOccurrences() > 0)
211 UP.PartialThreshold = UnrollPartialThreshold;
212 if (UnrollMaxPercentThresholdBoost.getNumOccurrences() > 0)
213 UP.MaxPercentThresholdBoost = UnrollMaxPercentThresholdBoost;
214 if (UnrollMaxCount.getNumOccurrences() > 0)
215 UP.MaxCount = UnrollMaxCount;
216 if (UnrollFullMaxCount.getNumOccurrences() > 0)
217 UP.FullUnrollMaxCount = UnrollFullMaxCount;
218 if (UnrollPeelCount.getNumOccurrences() > 0)
219 UP.PeelCount = UnrollPeelCount;
220 if (UnrollAllowPartial.getNumOccurrences() > 0)
221 UP.Partial = UnrollAllowPartial;
222 if (UnrollAllowRemainder.getNumOccurrences() > 0)
223 UP.AllowRemainder = UnrollAllowRemainder;
224 if (UnrollRuntime.getNumOccurrences() > 0)
225 UP.Runtime = UnrollRuntime;
226 if (UnrollMaxUpperBound == 0)
227 UP.UpperBound = false;
228 if (UnrollAllowPeeling.getNumOccurrences() > 0)
229 UP.AllowPeeling = UnrollAllowPeeling;
230 if (UnrollUnrollRemainder.getNumOccurrences() > 0)
231 UP.UnrollRemainder = UnrollUnrollRemainder;
233 // Apply user values provided by argument
234 if (UserThreshold.hasValue()) {
235 UP.Threshold = *UserThreshold;
236 UP.PartialThreshold = *UserThreshold;
238 if (UserCount.hasValue())
239 UP.Count = *UserCount;
240 if (UserAllowPartial.hasValue())
241 UP.Partial = *UserAllowPartial;
242 if (UserRuntime.hasValue())
243 UP.Runtime = *UserRuntime;
244 if (UserUpperBound.hasValue())
245 UP.UpperBound = *UserUpperBound;
246 if (UserAllowPeeling.hasValue())
247 UP.AllowPeeling = *UserAllowPeeling;
254 /// A struct to densely store the state of an instruction after unrolling at
257 /// This is designed to work like a tuple of <Instruction *, int> for the
258 /// purposes of hashing and lookup, but to be able to associate two boolean
259 /// states with each key.
260 struct UnrolledInstState {
264 unsigned IsCounted : 1;
267 /// Hashing and equality testing for a set of the instruction states.
268 struct UnrolledInstStateKeyInfo {
269 using PtrInfo = DenseMapInfo<Instruction *>;
270 using PairInfo = DenseMapInfo<std::pair<Instruction *, int>>;
272 static inline UnrolledInstState getEmptyKey() {
273 return {PtrInfo::getEmptyKey(), 0, 0, 0};
276 static inline UnrolledInstState getTombstoneKey() {
277 return {PtrInfo::getTombstoneKey(), 0, 0, 0};
280 static inline unsigned getHashValue(const UnrolledInstState &S) {
281 return PairInfo::getHashValue({S.I, S.Iteration});
284 static inline bool isEqual(const UnrolledInstState &LHS,
285 const UnrolledInstState &RHS) {
286 return PairInfo::isEqual({LHS.I, LHS.Iteration}, {RHS.I, RHS.Iteration});
290 struct EstimatedUnrollCost {
291 /// The estimated cost after unrolling.
292 unsigned UnrolledCost;
294 /// The estimated dynamic cost of executing the instructions in the
296 unsigned RolledDynamicCost;
299 } // end anonymous namespace
301 /// Figure out if the loop is worth full unrolling.
303 /// Complete loop unrolling can make some loads constant, and we need to know
304 /// if that would expose any further optimization opportunities. This routine
305 /// estimates this optimization. It computes cost of unrolled loop
306 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
307 /// dynamic cost we mean that we won't count costs of blocks that are known not
308 /// to be executed (i.e. if we have a branch in the loop and we know that at the
309 /// given iteration its condition would be resolved to true, we won't add up the
310 /// cost of the 'false'-block).
311 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
312 /// the analysis failed (no benefits expected from the unrolling, or the loop is
313 /// too big to analyze), the returned value is None.
314 static Optional<EstimatedUnrollCost> analyzeLoopUnrollCost(
315 const Loop *L, unsigned TripCount, DominatorTree &DT, ScalarEvolution &SE,
316 const SmallPtrSetImpl<const Value *> &EphValues,
317 const TargetTransformInfo &TTI, unsigned MaxUnrolledLoopSize) {
318 // We want to be able to scale offsets by the trip count and add more offsets
319 // to them without checking for overflows, and we already don't want to
320 // analyze *massive* trip counts, so we force the max to be reasonably small.
321 assert(UnrollMaxIterationsCountToAnalyze <
322 (unsigned)(std::numeric_limits<int>::max() / 2) &&
323 "The unroll iterations max is too large!");
325 // Only analyze inner loops. We can't properly estimate cost of nested loops
326 // and we won't visit inner loops again anyway.
330 // Don't simulate loops with a big or unknown tripcount
331 if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
332 TripCount > UnrollMaxIterationsCountToAnalyze)
335 SmallSetVector<BasicBlock *, 16> BBWorklist;
336 SmallSetVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitWorklist;
337 DenseMap<Value *, Constant *> SimplifiedValues;
338 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues;
340 // The estimated cost of the unrolled form of the loop. We try to estimate
341 // this by simplifying as much as we can while computing the estimate.
342 unsigned UnrolledCost = 0;
344 // We also track the estimated dynamic (that is, actually executed) cost in
345 // the rolled form. This helps identify cases when the savings from unrolling
346 // aren't just exposing dead control flows, but actual reduced dynamic
347 // instructions due to the simplifications which we expect to occur after
349 unsigned RolledDynamicCost = 0;
351 // We track the simplification of each instruction in each iteration. We use
352 // this to recursively merge costs into the unrolled cost on-demand so that
353 // we don't count the cost of any dead code. This is essentially a map from
354 // <instruction, int> to <bool, bool>, but stored as a densely packed struct.
355 DenseSet<UnrolledInstState, UnrolledInstStateKeyInfo> InstCostMap;
357 // A small worklist used to accumulate cost of instructions from each
358 // observable and reached root in the loop.
359 SmallVector<Instruction *, 16> CostWorklist;
361 // PHI-used worklist used between iterations while accumulating cost.
362 SmallVector<Instruction *, 4> PHIUsedList;
364 // Helper function to accumulate cost for instructions in the loop.
365 auto AddCostRecursively = [&](Instruction &RootI, int Iteration) {
366 assert(Iteration >= 0 && "Cannot have a negative iteration!");
367 assert(CostWorklist.empty() && "Must start with an empty cost list");
368 assert(PHIUsedList.empty() && "Must start with an empty phi used list");
369 CostWorklist.push_back(&RootI);
370 for (;; --Iteration) {
372 Instruction *I = CostWorklist.pop_back_val();
374 // InstCostMap only uses I and Iteration as a key, the other two values
375 // don't matter here.
376 auto CostIter = InstCostMap.find({I, Iteration, 0, 0});
377 if (CostIter == InstCostMap.end())
378 // If an input to a PHI node comes from a dead path through the loop
379 // we may have no cost data for it here. What that actually means is
382 auto &Cost = *CostIter;
384 // Already counted this instruction.
387 // Mark that we are counting the cost of this instruction now.
388 Cost.IsCounted = true;
390 // If this is a PHI node in the loop header, just add it to the PHI set.
391 if (auto *PhiI = dyn_cast<PHINode>(I))
392 if (PhiI->getParent() == L->getHeader()) {
393 assert(Cost.IsFree && "Loop PHIs shouldn't be evaluated as they "
394 "inherently simplify during unrolling.");
398 // Push the incoming value from the backedge into the PHI used list
399 // if it is an in-loop instruction. We'll use this to populate the
400 // cost worklist for the next iteration (as we count backwards).
401 if (auto *OpI = dyn_cast<Instruction>(
402 PhiI->getIncomingValueForBlock(L->getLoopLatch())))
403 if (L->contains(OpI))
404 PHIUsedList.push_back(OpI);
408 // First accumulate the cost of this instruction.
410 UnrolledCost += TTI.getUserCost(I);
411 LLVM_DEBUG(dbgs() << "Adding cost of instruction (iteration "
412 << Iteration << "): ");
413 LLVM_DEBUG(I->dump());
416 // We must count the cost of every operand which is not free,
417 // recursively. If we reach a loop PHI node, simply add it to the set
418 // to be considered on the next iteration (backwards!).
419 for (Value *Op : I->operands()) {
420 // Check whether this operand is free due to being a constant or
422 auto *OpI = dyn_cast<Instruction>(Op);
423 if (!OpI || !L->contains(OpI))
426 // Otherwise accumulate its cost.
427 CostWorklist.push_back(OpI);
429 } while (!CostWorklist.empty());
431 if (PHIUsedList.empty())
432 // We've exhausted the search.
435 assert(Iteration > 0 &&
436 "Cannot track PHI-used values past the first iteration!");
437 CostWorklist.append(PHIUsedList.begin(), PHIUsedList.end());
442 // Ensure that we don't violate the loop structure invariants relied on by
444 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
445 assert(L->isLCSSAForm(DT) &&
446 "Must have loops in LCSSA form to track live-out values.");
448 LLVM_DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
450 // Simulate execution of each iteration of the loop counting instructions,
451 // which would be simplified.
452 // Since the same load will take different values on different iterations,
453 // we literally have to go through all loop's iterations.
454 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
455 LLVM_DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
457 // Prepare for the iteration by collecting any simplified entry or backedge
459 for (Instruction &I : *L->getHeader()) {
460 auto *PHI = dyn_cast<PHINode>(&I);
464 // The loop header PHI nodes must have exactly two input: one from the
465 // loop preheader and one from the loop latch.
467 PHI->getNumIncomingValues() == 2 &&
468 "Must have an incoming value only for the preheader and the latch.");
470 Value *V = PHI->getIncomingValueForBlock(
471 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
472 Constant *C = dyn_cast<Constant>(V);
473 if (Iteration != 0 && !C)
474 C = SimplifiedValues.lookup(V);
476 SimplifiedInputValues.push_back({PHI, C});
479 // Now clear and re-populate the map for the next iteration.
480 SimplifiedValues.clear();
481 while (!SimplifiedInputValues.empty())
482 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val());
484 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE, L);
487 BBWorklist.insert(L->getHeader());
488 // Note that we *must not* cache the size, this loop grows the worklist.
489 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
490 BasicBlock *BB = BBWorklist[Idx];
492 // Visit all instructions in the given basic block and try to simplify
493 // it. We don't change the actual IR, just count optimization
495 for (Instruction &I : *BB) {
496 // These won't get into the final code - don't even try calculating the
498 if (isa<DbgInfoIntrinsic>(I) || EphValues.count(&I))
501 // Track this instruction's expected baseline cost when executing the
503 RolledDynamicCost += TTI.getUserCost(&I);
505 // Visit the instruction to analyze its loop cost after unrolling,
506 // and if the visitor returns true, mark the instruction as free after
507 // unrolling and continue.
508 bool IsFree = Analyzer.visit(I);
509 bool Inserted = InstCostMap.insert({&I, (int)Iteration,
511 /*IsCounted*/ false}).second;
513 assert(Inserted && "Cannot have a state for an unvisited instruction!");
518 // Can't properly model a cost of a call.
519 // FIXME: With a proper cost model we should be able to do it.
520 if (auto *CI = dyn_cast<CallInst>(&I)) {
521 const Function *Callee = CI->getCalledFunction();
522 if (!Callee || TTI.isLoweredToCall(Callee)) {
523 LLVM_DEBUG(dbgs() << "Can't analyze cost of loop with call\n");
528 // If the instruction might have a side-effect recursively account for
529 // the cost of it and all the instructions leading up to it.
530 if (I.mayHaveSideEffects())
531 AddCostRecursively(I, Iteration);
533 // If unrolled body turns out to be too big, bail out.
534 if (UnrolledCost > MaxUnrolledLoopSize) {
535 LLVM_DEBUG(dbgs() << " Exceeded threshold.. exiting.\n"
536 << " UnrolledCost: " << UnrolledCost
537 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
543 Instruction *TI = BB->getTerminator();
545 // Add in the live successors by first checking whether we have terminator
546 // that may be simplified based on the values simplified by this call.
547 BasicBlock *KnownSucc = nullptr;
548 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
549 if (BI->isConditional()) {
550 if (Constant *SimpleCond =
551 SimplifiedValues.lookup(BI->getCondition())) {
552 // Just take the first successor if condition is undef
553 if (isa<UndefValue>(SimpleCond))
554 KnownSucc = BI->getSuccessor(0);
555 else if (ConstantInt *SimpleCondVal =
556 dyn_cast<ConstantInt>(SimpleCond))
557 KnownSucc = BI->getSuccessor(SimpleCondVal->isZero() ? 1 : 0);
560 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
561 if (Constant *SimpleCond =
562 SimplifiedValues.lookup(SI->getCondition())) {
563 // Just take the first successor if condition is undef
564 if (isa<UndefValue>(SimpleCond))
565 KnownSucc = SI->getSuccessor(0);
566 else if (ConstantInt *SimpleCondVal =
567 dyn_cast<ConstantInt>(SimpleCond))
568 KnownSucc = SI->findCaseValue(SimpleCondVal)->getCaseSuccessor();
572 if (L->contains(KnownSucc))
573 BBWorklist.insert(KnownSucc);
575 ExitWorklist.insert({BB, KnownSucc});
579 // Add BB's successors to the worklist.
580 for (BasicBlock *Succ : successors(BB))
581 if (L->contains(Succ))
582 BBWorklist.insert(Succ);
584 ExitWorklist.insert({BB, Succ});
585 AddCostRecursively(*TI, Iteration);
588 // If we found no optimization opportunities on the first iteration, we
589 // won't find them on later ones too.
590 if (UnrolledCost == RolledDynamicCost) {
591 LLVM_DEBUG(dbgs() << " No opportunities found.. exiting.\n"
592 << " UnrolledCost: " << UnrolledCost << "\n");
597 while (!ExitWorklist.empty()) {
598 BasicBlock *ExitingBB, *ExitBB;
599 std::tie(ExitingBB, ExitBB) = ExitWorklist.pop_back_val();
601 for (Instruction &I : *ExitBB) {
602 auto *PN = dyn_cast<PHINode>(&I);
606 Value *Op = PN->getIncomingValueForBlock(ExitingBB);
607 if (auto *OpI = dyn_cast<Instruction>(Op))
608 if (L->contains(OpI))
609 AddCostRecursively(*OpI, TripCount - 1);
613 LLVM_DEBUG(dbgs() << "Analysis finished:\n"
614 << "UnrolledCost: " << UnrolledCost << ", "
615 << "RolledDynamicCost: " << RolledDynamicCost << "\n");
616 return {{UnrolledCost, RolledDynamicCost}};
619 /// ApproximateLoopSize - Approximate the size of the loop.
620 unsigned llvm::ApproximateLoopSize(
621 const Loop *L, unsigned &NumCalls, bool &NotDuplicatable, bool &Convergent,
622 const TargetTransformInfo &TTI,
623 const SmallPtrSetImpl<const Value *> &EphValues, unsigned BEInsns) {
625 for (BasicBlock *BB : L->blocks())
626 Metrics.analyzeBasicBlock(BB, TTI, EphValues);
627 NumCalls = Metrics.NumInlineCandidates;
628 NotDuplicatable = Metrics.notDuplicatable;
629 Convergent = Metrics.convergent;
631 unsigned LoopSize = Metrics.NumInsts;
633 // Don't allow an estimate of size zero. This would allows unrolling of loops
634 // with huge iteration counts, which is a compile time problem even if it's
635 // not a problem for code quality. Also, the code using this size may assume
636 // that each loop has at least three instructions (likely a conditional
637 // branch, a comparison feeding that branch, and some kind of loop increment
638 // feeding that comparison instruction).
639 LoopSize = std::max(LoopSize, BEInsns + 1);
644 // Returns the loop hint metadata node with the given name (for example,
645 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
647 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
648 if (MDNode *LoopID = L->getLoopID())
649 return GetUnrollMetadata(LoopID, Name);
653 // Returns true if the loop has an unroll(full) pragma.
654 static bool HasUnrollFullPragma(const Loop *L) {
655 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
658 // Returns true if the loop has an unroll(enable) pragma. This metadata is used
659 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
660 static bool HasUnrollEnablePragma(const Loop *L) {
661 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.enable");
664 // Returns true if the loop has an runtime unroll(disable) pragma.
665 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
666 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
669 // If loop has an unroll_count pragma return the (necessarily
670 // positive) value from the pragma. Otherwise return 0.
671 static unsigned UnrollCountPragmaValue(const Loop *L) {
672 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
674 assert(MD->getNumOperands() == 2 &&
675 "Unroll count hint metadata should have two operands.");
677 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
678 assert(Count >= 1 && "Unroll count must be positive.");
684 // Computes the boosting factor for complete unrolling.
685 // If fully unrolling the loop would save a lot of RolledDynamicCost, it would
686 // be beneficial to fully unroll the loop even if unrolledcost is large. We
687 // use (RolledDynamicCost / UnrolledCost) to model the unroll benefits to adjust
688 // the unroll threshold.
689 static unsigned getFullUnrollBoostingFactor(const EstimatedUnrollCost &Cost,
690 unsigned MaxPercentThresholdBoost) {
691 if (Cost.RolledDynamicCost >= std::numeric_limits<unsigned>::max() / 100)
693 else if (Cost.UnrolledCost != 0)
694 // The boosting factor is RolledDynamicCost / UnrolledCost
695 return std::min(100 * Cost.RolledDynamicCost / Cost.UnrolledCost,
696 MaxPercentThresholdBoost);
698 return MaxPercentThresholdBoost;
701 // Returns loop size estimation for unrolled loop.
702 static uint64_t getUnrolledLoopSize(
704 TargetTransformInfo::UnrollingPreferences &UP) {
705 assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
706 return (uint64_t)(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns;
709 // Returns true if unroll count was set explicitly.
710 // Calculates unroll count and writes it to UP.Count.
711 // Unless IgnoreUser is true, will also use metadata and command-line options
712 // that are specific to to the LoopUnroll pass (which, for instance, are
713 // irrelevant for the LoopUnrollAndJam pass).
714 // FIXME: This function is used by LoopUnroll and LoopUnrollAndJam, but consumes
715 // many LoopUnroll-specific options. The shared functionality should be
716 // refactored into it own function.
717 bool llvm::computeUnrollCount(
718 Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI,
719 ScalarEvolution &SE, const SmallPtrSetImpl<const Value *> &EphValues,
720 OptimizationRemarkEmitter *ORE, unsigned &TripCount, unsigned MaxTripCount,
721 unsigned &TripMultiple, unsigned LoopSize,
722 TargetTransformInfo::UnrollingPreferences &UP, bool &UseUpperBound) {
724 // Check for explicit Count.
725 // 1st priority is unroll count set by "unroll-count" option.
726 bool UserUnrollCount = UnrollCount.getNumOccurrences() > 0;
727 if (UserUnrollCount) {
728 UP.Count = UnrollCount;
729 UP.AllowExpensiveTripCount = true;
731 if (UP.AllowRemainder && getUnrolledLoopSize(LoopSize, UP) < UP.Threshold)
735 // 2nd priority is unroll count set by pragma.
736 unsigned PragmaCount = UnrollCountPragmaValue(L);
737 if (PragmaCount > 0) {
738 UP.Count = PragmaCount;
740 UP.AllowExpensiveTripCount = true;
742 if ((UP.AllowRemainder || (TripMultiple % PragmaCount == 0)) &&
743 getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
746 bool PragmaFullUnroll = HasUnrollFullPragma(L);
747 if (PragmaFullUnroll && TripCount != 0) {
748 UP.Count = TripCount;
749 if (getUnrolledLoopSize(LoopSize, UP) < PragmaUnrollThreshold)
753 bool PragmaEnableUnroll = HasUnrollEnablePragma(L);
754 bool ExplicitUnroll = PragmaCount > 0 || PragmaFullUnroll ||
755 PragmaEnableUnroll || UserUnrollCount;
757 if (ExplicitUnroll && TripCount != 0) {
758 // If the loop has an unrolling pragma, we want to be more aggressive with
759 // unrolling limits. Set thresholds to at least the PragmaUnrollThreshold
760 // value which is larger than the default limits.
761 UP.Threshold = std::max<unsigned>(UP.Threshold, PragmaUnrollThreshold);
762 UP.PartialThreshold =
763 std::max<unsigned>(UP.PartialThreshold, PragmaUnrollThreshold);
766 // 3rd priority is full unroll count.
767 // Full unroll makes sense only when TripCount or its upper bound could be
768 // statically calculated.
769 // Also we need to check if we exceed FullUnrollMaxCount.
770 // If using the upper bound to unroll, TripMultiple should be set to 1 because
771 // we do not know when loop may exit.
772 // MaxTripCount and ExactTripCount cannot both be non zero since we only
773 // compute the former when the latter is zero.
774 unsigned ExactTripCount = TripCount;
775 assert((ExactTripCount == 0 || MaxTripCount == 0) &&
776 "ExtractTripCount and MaxTripCount cannot both be non zero.");
777 unsigned FullUnrollTripCount = ExactTripCount ? ExactTripCount : MaxTripCount;
778 UP.Count = FullUnrollTripCount;
779 if (FullUnrollTripCount && FullUnrollTripCount <= UP.FullUnrollMaxCount) {
780 // When computing the unrolled size, note that BEInsns are not replicated
781 // like the rest of the loop body.
782 if (getUnrolledLoopSize(LoopSize, UP) < UP.Threshold) {
783 UseUpperBound = (MaxTripCount == FullUnrollTripCount);
784 TripCount = FullUnrollTripCount;
785 TripMultiple = UP.UpperBound ? 1 : TripMultiple;
786 return ExplicitUnroll;
788 // The loop isn't that small, but we still can fully unroll it if that
789 // helps to remove a significant number of instructions.
790 // To check that, run additional analysis on the loop.
791 if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
792 L, FullUnrollTripCount, DT, SE, EphValues, TTI,
793 UP.Threshold * UP.MaxPercentThresholdBoost / 100)) {
795 getFullUnrollBoostingFactor(*Cost, UP.MaxPercentThresholdBoost);
796 if (Cost->UnrolledCost < UP.Threshold * Boost / 100) {
797 UseUpperBound = (MaxTripCount == FullUnrollTripCount);
798 TripCount = FullUnrollTripCount;
799 TripMultiple = UP.UpperBound ? 1 : TripMultiple;
800 return ExplicitUnroll;
806 // 4th priority is loop peeling.
807 computePeelCount(L, LoopSize, UP, TripCount, SE);
811 return ExplicitUnroll;
814 // 5th priority is partial unrolling.
815 // Try partial unroll only when TripCount could be statically calculated.
817 UP.Partial |= ExplicitUnroll;
819 LLVM_DEBUG(dbgs() << " will not try to unroll partially because "
820 << "-unroll-allow-partial not given\n");
825 UP.Count = TripCount;
826 if (UP.PartialThreshold != NoThreshold) {
827 // Reduce unroll count to be modulo of TripCount for partial unrolling.
828 if (getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
830 (std::max(UP.PartialThreshold, UP.BEInsns + 1) - UP.BEInsns) /
831 (LoopSize - UP.BEInsns);
832 if (UP.Count > UP.MaxCount)
833 UP.Count = UP.MaxCount;
834 while (UP.Count != 0 && TripCount % UP.Count != 0)
836 if (UP.AllowRemainder && UP.Count <= 1) {
837 // If there is no Count that is modulo of TripCount, set Count to
838 // largest power-of-two factor that satisfies the threshold limit.
839 // As we'll create fixup loop, do the type of unrolling only if
840 // remainder loop is allowed.
841 UP.Count = UP.DefaultUnrollRuntimeCount;
842 while (UP.Count != 0 &&
843 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
847 if (PragmaEnableUnroll)
849 return OptimizationRemarkMissed(DEBUG_TYPE,
850 "UnrollAsDirectedTooLarge",
851 L->getStartLoc(), L->getHeader())
852 << "Unable to unroll loop as directed by unroll(enable) "
854 "because unrolled size is too large.";
859 UP.Count = TripCount;
861 if (UP.Count > UP.MaxCount)
862 UP.Count = UP.MaxCount;
863 if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
864 UP.Count != TripCount)
866 return OptimizationRemarkMissed(DEBUG_TYPE,
867 "FullUnrollAsDirectedTooLarge",
868 L->getStartLoc(), L->getHeader())
869 << "Unable to fully unroll loop as directed by unroll pragma "
871 "unrolled size is too large.";
873 return ExplicitUnroll;
875 assert(TripCount == 0 &&
876 "All cases when TripCount is constant should be covered here.");
877 if (PragmaFullUnroll)
879 return OptimizationRemarkMissed(
880 DEBUG_TYPE, "CantFullUnrollAsDirectedRuntimeTripCount",
881 L->getStartLoc(), L->getHeader())
882 << "Unable to fully unroll loop as directed by unroll(full) "
884 "because loop has a runtime trip count.";
887 // 6th priority is runtime unrolling.
888 // Don't unroll a runtime trip count loop when it is disabled.
889 if (HasRuntimeUnrollDisablePragma(L)) {
894 // Check if the runtime trip count is too small when profile is available.
895 if (L->getHeader()->getParent()->hasProfileData()) {
896 if (auto ProfileTripCount = getLoopEstimatedTripCount(L)) {
897 if (*ProfileTripCount < FlatLoopTripCountThreshold)
900 UP.AllowExpensiveTripCount = true;
904 // Reduce count based on the type of unrolling and the threshold values.
905 UP.Runtime |= PragmaEnableUnroll || PragmaCount > 0 || UserUnrollCount;
908 dbgs() << " will not try to unroll loop with runtime trip count "
909 << "-unroll-runtime not given\n");
914 UP.Count = UP.DefaultUnrollRuntimeCount;
916 // Reduce unroll count to be the largest power-of-two factor of
917 // the original count which satisfies the threshold limit.
918 while (UP.Count != 0 &&
919 getUnrolledLoopSize(LoopSize, UP) > UP.PartialThreshold)
923 unsigned OrigCount = UP.Count;
926 if (!UP.AllowRemainder && UP.Count != 0 && (TripMultiple % UP.Count) != 0) {
927 while (UP.Count != 0 && TripMultiple % UP.Count != 0)
930 dbgs() << "Remainder loop is restricted (that could architecture "
931 "specific or because the loop contains a convergent "
932 "instruction), so unroll count must divide the trip "
934 << TripMultiple << ". Reducing unroll count from " << OrigCount
935 << " to " << UP.Count << ".\n");
939 if (PragmaCount > 0 && !UP.AllowRemainder)
941 return OptimizationRemarkMissed(DEBUG_TYPE,
942 "DifferentUnrollCountFromDirected",
943 L->getStartLoc(), L->getHeader())
944 << "Unable to unroll loop the number of times directed by "
945 "unroll_count pragma because remainder loop is restricted "
946 "(that could architecture specific or because the loop "
947 "contains a convergent instruction) and so must have an "
949 "count that divides the loop trip multiple of "
950 << NV("TripMultiple", TripMultiple) << ". Unrolling instead "
951 << NV("UnrollCount", UP.Count) << " time(s).";
955 if (UP.Count > UP.MaxCount)
956 UP.Count = UP.MaxCount;
957 LLVM_DEBUG(dbgs() << " partially unrolling with count: " << UP.Count
961 return ExplicitUnroll;
964 static LoopUnrollResult tryToUnrollLoop(
965 Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE,
966 const TargetTransformInfo &TTI, AssumptionCache &AC,
967 OptimizationRemarkEmitter &ORE, bool PreserveLCSSA, int OptLevel,
968 bool OnlyWhenForced, Optional<unsigned> ProvidedCount,
969 Optional<unsigned> ProvidedThreshold, Optional<bool> ProvidedAllowPartial,
970 Optional<bool> ProvidedRuntime, Optional<bool> ProvidedUpperBound,
971 Optional<bool> ProvidedAllowPeeling) {
972 LLVM_DEBUG(dbgs() << "Loop Unroll: F["
973 << L->getHeader()->getParent()->getName() << "] Loop %"
974 << L->getHeader()->getName() << "\n");
975 TransformationMode TM = hasUnrollTransformation(L);
977 return LoopUnrollResult::Unmodified;
978 if (!L->isLoopSimplifyForm()) {
980 dbgs() << " Not unrolling loop which is not in loop-simplify form.\n");
981 return LoopUnrollResult::Unmodified;
984 // When automtatic unrolling is disabled, do not unroll unless overridden for
986 if (OnlyWhenForced && !(TM & TM_Enable))
987 return LoopUnrollResult::Unmodified;
989 unsigned NumInlineCandidates;
990 bool NotDuplicatable;
992 TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
993 L, SE, TTI, OptLevel, ProvidedThreshold, ProvidedCount,
994 ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound,
995 ProvidedAllowPeeling);
996 // Exit early if unrolling is disabled.
997 if (UP.Threshold == 0 && (!UP.Partial || UP.PartialThreshold == 0))
998 return LoopUnrollResult::Unmodified;
1000 SmallPtrSet<const Value *, 32> EphValues;
1001 CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
1004 ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent,
1005 TTI, EphValues, UP.BEInsns);
1006 LLVM_DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
1007 if (NotDuplicatable) {
1008 LLVM_DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
1009 << " instructions.\n");
1010 return LoopUnrollResult::Unmodified;
1012 if (NumInlineCandidates != 0) {
1013 LLVM_DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
1014 return LoopUnrollResult::Unmodified;
1017 // Find trip count and trip multiple if count is not available
1018 unsigned TripCount = 0;
1019 unsigned MaxTripCount = 0;
1020 unsigned TripMultiple = 1;
1021 // If there are multiple exiting blocks but one of them is the latch, use the
1022 // latch for the trip count estimation. Otherwise insist on a single exiting
1023 // block for the trip count estimation.
1024 BasicBlock *ExitingBlock = L->getLoopLatch();
1025 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
1026 ExitingBlock = L->getExitingBlock();
1028 TripCount = SE.getSmallConstantTripCount(L, ExitingBlock);
1029 TripMultiple = SE.getSmallConstantTripMultiple(L, ExitingBlock);
1032 // If the loop contains a convergent operation, the prelude we'd add
1033 // to do the first few instructions before we hit the unrolled loop
1034 // is unsafe -- it adds a control-flow dependency to the convergent
1035 // operation. Therefore restrict remainder loop (try unrollig without).
1037 // TODO: This is quite conservative. In practice, convergent_op()
1038 // is likely to be called unconditionally in the loop. In this
1039 // case, the program would be ill-formed (on most architectures)
1040 // unless n were the same on all threads in a thread group.
1041 // Assuming n is the same on all threads, any kind of unrolling is
1042 // safe. But currently llvm's notion of convergence isn't powerful
1043 // enough to express this.
1045 UP.AllowRemainder = false;
1047 // Try to find the trip count upper bound if we cannot find the exact trip
1049 bool MaxOrZero = false;
1051 MaxTripCount = SE.getSmallConstantMaxTripCount(L);
1052 MaxOrZero = SE.isBackedgeTakenCountMaxOrZero(L);
1053 // We can unroll by the upper bound amount if it's generally allowed or if
1054 // we know that the loop is executed either the upper bound or zero times.
1055 // (MaxOrZero unrolling keeps only the first loop test, so the number of
1056 // loop tests remains the same compared to the non-unrolled version, whereas
1057 // the generic upper bound unrolling keeps all but the last loop test so the
1058 // number of loop tests goes up which may end up being worse on targets with
1059 // constrained branch predictor resources so is controlled by an option.)
1060 // In addition we only unroll small upper bounds.
1061 if (!(UP.UpperBound || MaxOrZero) || MaxTripCount > UnrollMaxUpperBound) {
1066 // computeUnrollCount() decides whether it is beneficial to use upper bound to
1067 // fully unroll the loop.
1068 bool UseUpperBound = false;
1069 bool IsCountSetExplicitly = computeUnrollCount(
1070 L, TTI, DT, LI, SE, EphValues, &ORE, TripCount, MaxTripCount,
1071 TripMultiple, LoopSize, UP, UseUpperBound);
1073 return LoopUnrollResult::Unmodified;
1074 // Unroll factor (Count) must be less or equal to TripCount.
1075 if (TripCount && UP.Count > TripCount)
1076 UP.Count = TripCount;
1078 // Save loop properties before it is transformed.
1079 MDNode *OrigLoopID = L->getLoopID();
1082 Loop *RemainderLoop = nullptr;
1083 LoopUnrollResult UnrollResult = UnrollLoop(
1084 L, UP.Count, TripCount, UP.Force, UP.Runtime, UP.AllowExpensiveTripCount,
1085 UseUpperBound, MaxOrZero, TripMultiple, UP.PeelCount, UP.UnrollRemainder,
1086 LI, &SE, &DT, &AC, &ORE, PreserveLCSSA, &RemainderLoop);
1087 if (UnrollResult == LoopUnrollResult::Unmodified)
1088 return LoopUnrollResult::Unmodified;
1090 if (RemainderLoop) {
1091 Optional<MDNode *> RemainderLoopID =
1092 makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1093 LLVMLoopUnrollFollowupRemainder});
1094 if (RemainderLoopID.hasValue())
1095 RemainderLoop->setLoopID(RemainderLoopID.getValue());
1098 if (UnrollResult != LoopUnrollResult::FullyUnrolled) {
1099 Optional<MDNode *> NewLoopID =
1100 makeFollowupLoopID(OrigLoopID, {LLVMLoopUnrollFollowupAll,
1101 LLVMLoopUnrollFollowupUnrolled});
1102 if (NewLoopID.hasValue()) {
1103 L->setLoopID(NewLoopID.getValue());
1105 // Do not setLoopAlreadyUnrolled if loop attributes have been specified
1107 return UnrollResult;
1111 // If loop has an unroll count pragma or unrolled by explicitly set count
1112 // mark loop as unrolled to prevent unrolling beyond that requested.
1113 // If the loop was peeled, we already "used up" the profile information
1114 // we had, so we don't want to unroll or peel again.
1115 if (UnrollResult != LoopUnrollResult::FullyUnrolled &&
1116 (IsCountSetExplicitly || UP.PeelCount))
1117 L->setLoopAlreadyUnrolled();
1119 return UnrollResult;
1124 class LoopUnroll : public LoopPass {
1126 static char ID; // Pass ID, replacement for typeid
1130 /// If false, use a cost model to determine whether unrolling of a loop is
1131 /// profitable. If true, only loops that explicitly request unrolling via
1132 /// metadata are considered. All other loops are skipped.
1133 bool OnlyWhenForced;
1135 Optional<unsigned> ProvidedCount;
1136 Optional<unsigned> ProvidedThreshold;
1137 Optional<bool> ProvidedAllowPartial;
1138 Optional<bool> ProvidedRuntime;
1139 Optional<bool> ProvidedUpperBound;
1140 Optional<bool> ProvidedAllowPeeling;
1142 LoopUnroll(int OptLevel = 2, bool OnlyWhenForced = false,
1143 Optional<unsigned> Threshold = None,
1144 Optional<unsigned> Count = None,
1145 Optional<bool> AllowPartial = None, Optional<bool> Runtime = None,
1146 Optional<bool> UpperBound = None,
1147 Optional<bool> AllowPeeling = None)
1148 : LoopPass(ID), OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
1149 ProvidedCount(std::move(Count)), ProvidedThreshold(Threshold),
1150 ProvidedAllowPartial(AllowPartial), ProvidedRuntime(Runtime),
1151 ProvidedUpperBound(UpperBound), ProvidedAllowPeeling(AllowPeeling) {
1152 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
1155 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1159 Function &F = *L->getHeader()->getParent();
1161 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1162 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1163 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1164 const TargetTransformInfo &TTI =
1165 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1166 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1167 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
1168 // pass. Function analyses need to be preserved across loop transformations
1169 // but ORE cannot be preserved (see comment before the pass definition).
1170 OptimizationRemarkEmitter ORE(&F);
1171 bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1173 LoopUnrollResult Result = tryToUnrollLoop(
1174 L, DT, LI, SE, TTI, AC, ORE, PreserveLCSSA, OptLevel, OnlyWhenForced,
1175 ProvidedCount, ProvidedThreshold, ProvidedAllowPartial, ProvidedRuntime,
1176 ProvidedUpperBound, ProvidedAllowPeeling);
1178 if (Result == LoopUnrollResult::FullyUnrolled)
1179 LPM.markLoopAsDeleted(*L);
1181 return Result != LoopUnrollResult::Unmodified;
1184 /// This transformation requires natural loop information & requires that
1185 /// loop preheaders be inserted into the CFG...
1186 void getAnalysisUsage(AnalysisUsage &AU) const override {
1187 AU.addRequired<AssumptionCacheTracker>();
1188 AU.addRequired<TargetTransformInfoWrapperPass>();
1189 // FIXME: Loop passes are required to preserve domtree, and for now we just
1190 // recreate dom info if anything gets unrolled.
1191 getLoopAnalysisUsage(AU);
1195 } // end anonymous namespace
1197 char LoopUnroll::ID = 0;
1199 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1200 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1201 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1202 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1203 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
1205 Pass *llvm::createLoopUnrollPass(int OptLevel, bool OnlyWhenForced,
1206 int Threshold, int Count, int AllowPartial,
1207 int Runtime, int UpperBound,
1209 // TODO: It would make more sense for this function to take the optionals
1210 // directly, but that's dangerous since it would silently break out of tree
1212 return new LoopUnroll(
1213 OptLevel, OnlyWhenForced,
1214 Threshold == -1 ? None : Optional<unsigned>(Threshold),
1215 Count == -1 ? None : Optional<unsigned>(Count),
1216 AllowPartial == -1 ? None : Optional<bool>(AllowPartial),
1217 Runtime == -1 ? None : Optional<bool>(Runtime),
1218 UpperBound == -1 ? None : Optional<bool>(UpperBound),
1219 AllowPeeling == -1 ? None : Optional<bool>(AllowPeeling));
1222 Pass *llvm::createSimpleLoopUnrollPass(int OptLevel, bool OnlyWhenForced) {
1223 return createLoopUnrollPass(OptLevel, OnlyWhenForced, -1, -1, 0, 0, 0, 0);
1226 PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
1227 LoopStandardAnalysisResults &AR,
1228 LPMUpdater &Updater) {
1230 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
1231 Function *F = L.getHeader()->getParent();
1233 auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F);
1234 // FIXME: This should probably be optional rather than required.
1237 "LoopFullUnrollPass: OptimizationRemarkEmitterAnalysis not "
1238 "cached at a higher level");
1240 // Keep track of the previous loop structure so we can identify new loops
1241 // created by unrolling.
1242 Loop *ParentL = L.getParentLoop();
1243 SmallPtrSet<Loop *, 4> OldLoops;
1245 OldLoops.insert(ParentL->begin(), ParentL->end());
1247 OldLoops.insert(AR.LI.begin(), AR.LI.end());
1249 std::string LoopName = L.getName();
1252 tryToUnrollLoop(&L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, *ORE,
1253 /*PreserveLCSSA*/ true, OptLevel, OnlyWhenForced,
1255 /*Threshold*/ None, /*AllowPartial*/ false,
1256 /*Runtime*/ false, /*UpperBound*/ false,
1257 /*AllowPeeling*/ false) != LoopUnrollResult::Unmodified;
1259 return PreservedAnalyses::all();
1261 // The parent must not be damaged by unrolling!
1264 ParentL->verifyLoop();
1267 // Unrolling can do several things to introduce new loops into a loop nest:
1268 // - Full unrolling clones child loops within the current loop but then
1269 // removes the current loop making all of the children appear to be new
1272 // When a new loop appears as a sibling loop after fully unrolling,
1273 // its nesting structure has fundamentally changed and we want to revisit
1274 // it to reflect that.
1276 // When unrolling has removed the current loop, we need to tell the
1277 // infrastructure that it is gone.
1279 // Finally, we support a debugging/testing mode where we revisit child loops
1280 // as well. These are not expected to require further optimizations as either
1281 // they or the loop they were cloned from have been directly visited already.
1282 // But the debugging mode allows us to check this assumption.
1283 bool IsCurrentLoopValid = false;
1284 SmallVector<Loop *, 4> SibLoops;
1286 SibLoops.append(ParentL->begin(), ParentL->end());
1288 SibLoops.append(AR.LI.begin(), AR.LI.end());
1289 erase_if(SibLoops, [&](Loop *SibLoop) {
1290 if (SibLoop == &L) {
1291 IsCurrentLoopValid = true;
1295 // Otherwise erase the loop from the list if it was in the old loops.
1296 return OldLoops.count(SibLoop) != 0;
1298 Updater.addSiblingLoops(SibLoops);
1300 if (!IsCurrentLoopValid) {
1301 Updater.markLoopAsDeleted(L, LoopName);
1303 // We can only walk child loops if the current loop remained valid.
1304 if (UnrollRevisitChildLoops) {
1305 // Walk *all* of the child loops.
1306 SmallVector<Loop *, 4> ChildLoops(L.begin(), L.end());
1307 Updater.addChildLoops(ChildLoops);
1311 return getLoopPassPreservedAnalyses();
1314 template <typename RangeT>
1315 static SmallVector<Loop *, 8> appendLoopsToWorklist(RangeT &&Loops) {
1316 SmallVector<Loop *, 8> Worklist;
1317 // We use an internal worklist to build up the preorder traversal without
1319 SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist;
1321 for (Loop *RootL : Loops) {
1322 assert(PreOrderLoops.empty() && "Must start with an empty preorder walk.");
1323 assert(PreOrderWorklist.empty() &&
1324 "Must start with an empty preorder walk worklist.");
1325 PreOrderWorklist.push_back(RootL);
1327 Loop *L = PreOrderWorklist.pop_back_val();
1328 PreOrderWorklist.append(L->begin(), L->end());
1329 PreOrderLoops.push_back(L);
1330 } while (!PreOrderWorklist.empty());
1332 Worklist.append(PreOrderLoops.begin(), PreOrderLoops.end());
1333 PreOrderLoops.clear();
1338 PreservedAnalyses LoopUnrollPass::run(Function &F,
1339 FunctionAnalysisManager &AM) {
1340 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1341 auto &LI = AM.getResult<LoopAnalysis>(F);
1342 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1343 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1344 auto &AC = AM.getResult<AssumptionAnalysis>(F);
1345 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1347 LoopAnalysisManager *LAM = nullptr;
1348 if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(F))
1349 LAM = &LAMProxy->getManager();
1351 const ModuleAnalysisManager &MAM =
1352 AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
1353 ProfileSummaryInfo *PSI =
1354 MAM.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
1356 bool Changed = false;
1358 // The unroller requires loops to be in simplified form, and also needs LCSSA.
1359 // Since simplification may add new inner loops, it has to run before the
1360 // legality and profitability checks. This means running the loop unroller
1361 // will simplify all loops, regardless of whether anything end up being
1363 for (auto &L : LI) {
1364 Changed |= simplifyLoop(L, &DT, &LI, &SE, &AC, false /* PreserveLCSSA */);
1365 Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1368 SmallVector<Loop *, 8> Worklist = appendLoopsToWorklist(LI);
1370 while (!Worklist.empty()) {
1371 // Because the LoopInfo stores the loops in RPO, we walk the worklist
1372 // from back to front so that we work forward across the CFG, which
1373 // for unrolling is only needed to get optimization remarks emitted in
1375 Loop &L = *Worklist.pop_back_val();
1377 Loop *ParentL = L.getParentLoop();
1380 // Check if the profile summary indicates that the profiled application
1381 // has a huge working set size, in which case we disable peeling to avoid
1382 // bloating it further.
1383 Optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling;
1384 if (PSI && PSI->hasHugeWorkingSetSize())
1385 LocalAllowPeeling = false;
1386 std::string LoopName = L.getName();
1387 // The API here is quite complex to call and we allow to select some
1388 // flavors of unrolling during construction time (by setting UnrollOpts).
1389 LoopUnrollResult Result = tryToUnrollLoop(
1390 &L, DT, &LI, SE, TTI, AC, ORE,
1391 /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, UnrollOpts.OnlyWhenForced,
1393 /*Threshold*/ None, UnrollOpts.AllowPartial, UnrollOpts.AllowRuntime,
1394 UnrollOpts.AllowUpperBound, LocalAllowPeeling);
1395 Changed |= Result != LoopUnrollResult::Unmodified;
1397 // The parent must not be damaged by unrolling!
1399 if (Result != LoopUnrollResult::Unmodified && ParentL)
1400 ParentL->verifyLoop();
1403 // Clear any cached analysis results for L if we removed it completely.
1404 if (LAM && Result == LoopUnrollResult::FullyUnrolled)
1405 LAM->clear(L, LoopName);
1409 return PreservedAnalyses::all();
1411 return getLoopPassPreservedAnalyses();