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.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/Analysis/AssumptionCache.h"
18 #include "llvm/Analysis/CodeMetrics.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LoopPass.h"
21 #include "llvm/Analysis/ScalarEvolution.h"
22 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DiagnosticInfo.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/InstVisitor.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Metadata.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Transforms/Utils/UnrollLoop.h"
38 #define DEBUG_TYPE "loop-unroll"
40 static cl::opt<unsigned>
41 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
42 cl::desc("The cut-off point for automatic loop unrolling"));
44 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
45 "unroll-max-iteration-count-to-analyze", cl::init(0), cl::Hidden,
46 cl::desc("Don't allow loop unrolling to simulate more than this number of"
47 "iterations when checking full unroll profitability"));
49 static cl::opt<unsigned> UnrollMinPercentOfOptimized(
50 "unroll-percent-of-optimized-for-complete-unroll", cl::init(20), cl::Hidden,
51 cl::desc("If complete unrolling could trigger further optimizations, and, "
52 "by that, remove the given percent of instructions, perform the "
53 "complete unroll even if it's beyond the threshold"));
55 static cl::opt<unsigned> UnrollAbsoluteThreshold(
56 "unroll-absolute-threshold", cl::init(2000), cl::Hidden,
57 cl::desc("Don't unroll if the unrolled size is bigger than this threshold,"
58 " even if we can remove big portion of instructions later."));
60 static cl::opt<unsigned>
61 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
62 cl::desc("Use this unroll count for all loops including those with "
63 "unroll_count pragma values, for testing purposes"));
66 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
67 cl::desc("Allows loops to be partially unrolled until "
68 "-unroll-threshold loop size is reached."));
71 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden,
72 cl::desc("Unroll loops with run-time trip counts"));
74 static cl::opt<unsigned>
75 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
76 cl::desc("Unrolled size limit for loops with an unroll(full) or "
77 "unroll_count pragma."));
80 class LoopUnroll : public LoopPass {
82 static char ID; // Pass ID, replacement for typeid
83 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) {
84 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
85 CurrentAbsoluteThreshold = UnrollAbsoluteThreshold;
86 CurrentMinPercentOfOptimized = UnrollMinPercentOfOptimized;
87 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
88 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
89 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
91 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
92 UserAbsoluteThreshold = (UnrollAbsoluteThreshold.getNumOccurrences() > 0);
93 UserPercentOfOptimized =
94 (UnrollMinPercentOfOptimized.getNumOccurrences() > 0);
95 UserAllowPartial = (P != -1) ||
96 (UnrollAllowPartial.getNumOccurrences() > 0);
97 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
98 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0);
100 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
103 /// A magic value for use with the Threshold parameter to indicate
104 /// that the loop unroll should be performed regardless of how much
105 /// code expansion would result.
106 static const unsigned NoThreshold = UINT_MAX;
108 // Threshold to use when optsize is specified (and there is no
109 // explicit -unroll-threshold).
110 static const unsigned OptSizeUnrollThreshold = 50;
112 // Default unroll count for loops with run-time trip count if
113 // -unroll-count is not set
114 static const unsigned UnrollRuntimeCount = 8;
116 unsigned CurrentCount;
117 unsigned CurrentThreshold;
118 unsigned CurrentAbsoluteThreshold;
119 unsigned CurrentMinPercentOfOptimized;
120 bool CurrentAllowPartial;
122 bool UserCount; // CurrentCount is user-specified.
123 bool UserThreshold; // CurrentThreshold is user-specified.
124 bool UserAbsoluteThreshold; // CurrentAbsoluteThreshold is
126 bool UserPercentOfOptimized; // CurrentMinPercentOfOptimized is
128 bool UserAllowPartial; // CurrentAllowPartial is user-specified.
129 bool UserRuntime; // CurrentRuntime is user-specified.
131 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
133 /// This transformation requires natural loop information & requires that
134 /// loop preheaders be inserted into the CFG...
136 void getAnalysisUsage(AnalysisUsage &AU) const override {
137 AU.addRequired<AssumptionCacheTracker>();
138 AU.addRequired<LoopInfoWrapperPass>();
139 AU.addPreserved<LoopInfoWrapperPass>();
140 AU.addRequiredID(LoopSimplifyID);
141 AU.addPreservedID(LoopSimplifyID);
142 AU.addRequiredID(LCSSAID);
143 AU.addPreservedID(LCSSAID);
144 AU.addRequired<ScalarEvolution>();
145 AU.addPreserved<ScalarEvolution>();
146 AU.addRequired<TargetTransformInfoWrapperPass>();
147 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
148 // If loop unroll does not preserve dom info then LCSSA pass on next
149 // loop will receive invalid dom info.
150 // For now, recreate dom info, if loop is unrolled.
151 AU.addPreserved<DominatorTreeWrapperPass>();
154 // Fill in the UnrollingPreferences parameter with values from the
155 // TargetTransformationInfo.
156 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI,
157 TargetTransformInfo::UnrollingPreferences &UP) {
158 UP.Threshold = CurrentThreshold;
159 UP.AbsoluteThreshold = CurrentAbsoluteThreshold;
160 UP.MinPercentOfOptimized = CurrentMinPercentOfOptimized;
161 UP.OptSizeThreshold = OptSizeUnrollThreshold;
162 UP.PartialThreshold = CurrentThreshold;
163 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
164 UP.Count = CurrentCount;
165 UP.MaxCount = UINT_MAX;
166 UP.Partial = CurrentAllowPartial;
167 UP.Runtime = CurrentRuntime;
168 UP.AllowExpensiveTripCount = false;
169 TTI.getUnrollingPreferences(L, UP);
172 // Select and return an unroll count based on parameters from
173 // user, unroll preferences, unroll pragmas, or a heuristic.
174 // SetExplicitly is set to true if the unroll count is is set by
175 // the user or a pragma rather than selected heuristically.
177 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
178 unsigned PragmaCount,
179 const TargetTransformInfo::UnrollingPreferences &UP,
180 bool &SetExplicitly);
182 // Select threshold values used to limit unrolling based on a
183 // total unrolled size. Parameters Threshold and PartialThreshold
184 // are set to the maximum unrolled size for fully and partially
185 // unrolled loops respectively.
186 void selectThresholds(const Loop *L, bool HasPragma,
187 const TargetTransformInfo::UnrollingPreferences &UP,
188 unsigned &Threshold, unsigned &PartialThreshold,
189 unsigned &AbsoluteThreshold,
190 unsigned &PercentOfOptimizedForCompleteUnroll) {
191 // Determine the current unrolling threshold. While this is
192 // normally set from UnrollThreshold, it is overridden to a
193 // smaller value if the current function is marked as
194 // optimize-for-size, and the unroll threshold was not user
196 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
197 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
198 AbsoluteThreshold = UserAbsoluteThreshold ? CurrentAbsoluteThreshold
199 : UP.AbsoluteThreshold;
200 PercentOfOptimizedForCompleteUnroll = UserPercentOfOptimized
201 ? CurrentMinPercentOfOptimized
202 : UP.MinPercentOfOptimized;
204 if (!UserThreshold &&
205 L->getHeader()->getParent()->hasFnAttribute(
206 Attribute::OptimizeForSize)) {
207 Threshold = UP.OptSizeThreshold;
208 PartialThreshold = UP.PartialOptSizeThreshold;
211 // If the loop has an unrolling pragma, we want to be more
212 // aggressive with unrolling limits. Set thresholds to at
213 // least the PragmaTheshold value which is larger than the
215 if (Threshold != NoThreshold)
216 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold);
217 if (PartialThreshold != NoThreshold)
219 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold);
222 bool canUnrollCompletely(Loop *L, unsigned Threshold,
223 unsigned AbsoluteThreshold, uint64_t UnrolledSize,
224 unsigned NumberOfOptimizedInstructions,
225 unsigned PercentOfOptimizedForCompleteUnroll);
229 char LoopUnroll::ID = 0;
230 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
231 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
232 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
233 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
234 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
235 INITIALIZE_PASS_DEPENDENCY(LCSSA)
236 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
237 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
239 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial,
241 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime);
244 Pass *llvm::createSimpleLoopUnrollPass() {
245 return llvm::createLoopUnrollPass(-1, -1, 0, 0);
249 /// \brief SCEV expressions visitor used for finding expressions that would
250 /// become constants if the loop L is unrolled.
251 struct FindConstantPointers {
252 /// \brief Shows whether the expression is ConstAddress+Constant or not.
253 bool IndexIsConstant;
255 /// \brief Used for filtering out SCEV expressions with two or more AddRec
258 /// Used to filter out complicated SCEV expressions, having several AddRec
259 /// sub-expressions. We don't handle them, because unrolling one loop
260 /// would help to replace only one of these inductions with a constant, and
261 /// consequently, the expression would remain non-constant.
264 /// \brief If the SCEV expression becomes ConstAddress+Constant, this value
265 /// holds ConstAddress. Otherwise, it's nullptr.
268 /// \brief The loop, which we try to completely unroll.
273 FindConstantPointers(const Loop *L, ScalarEvolution &SE)
274 : IndexIsConstant(true), HaveSeenAR(false), BaseAddress(nullptr),
277 /// Examine the given expression S and figure out, if it can be a part of an
278 /// expression, that could become a constant after the loop is unrolled.
279 /// The routine sets IndexIsConstant and HaveSeenAR according to the analysis
281 /// \returns true if we need to examine subexpressions, and false otherwise.
282 bool follow(const SCEV *S) {
283 if (const SCEVUnknown *SC = dyn_cast<SCEVUnknown>(S)) {
284 // We've reached the leaf node of SCEV, it's most probably just a
286 // If it's the only one SCEV-subexpression, then it might be a base
287 // address of an index expression.
288 // If we've already recorded base address, then just give up on this SCEV
289 // - it's too complicated.
291 IndexIsConstant = false;
294 BaseAddress = SC->getValue();
297 if (isa<SCEVConstant>(S))
299 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
300 // If the current SCEV expression is AddRec, and its loop isn't the loop
301 // we are about to unroll, then we won't get a constant address after
302 // unrolling, and thus, won't be able to eliminate the load.
303 if (AR->getLoop() != L) {
304 IndexIsConstant = false;
307 // We don't handle multiple AddRecs here, so give up in this case.
309 IndexIsConstant = false;
315 // Continue traversal.
318 bool isDone() const { return !IndexIsConstant; }
320 } // End anonymous namespace.
323 /// \brief A cache of SCEV results used to optimize repeated queries to SCEV on
324 /// the same set of instructions.
326 /// The primary cost this saves is the cost of checking the validity of a SCEV
327 /// every time it is looked up. However, in some cases we can provide a reduced
328 /// and especially useful model for an instruction based upon SCEV that is
329 /// non-trivial to compute but more useful to clients.
332 /// \brief Struct to represent a GEP whose start and step are known fixed
333 /// offsets from a base address due to SCEV's analysis.
334 struct GEPDescriptor {
335 Value *BaseAddr = nullptr;
340 Optional<GEPDescriptor> getGEPDescriptor(GetElementPtrInst *GEP);
342 SCEVCache(const Loop &L, ScalarEvolution &SE) : L(L), SE(SE) {}
348 SmallDenseMap<GetElementPtrInst *, GEPDescriptor> GEPDescriptors;
350 } // End anonymous namespace.
352 /// \brief Get a simplified descriptor for a GEP instruction.
354 /// Where possible, this produces a simplified descriptor for a GEP instruction
355 /// using SCEV analysis of the containing loop. If this isn't possible, it
356 /// returns an empty optional.
358 /// The model is a base address, an initial offset, and a per-iteration step.
359 /// This fits very common patterns of GEPs inside loops and is something we can
360 /// use to simulate the behavior of a particular iteration of a loop.
362 /// This is a cached interface. The first call may do non-trivial work to
363 /// compute the result, but all subsequent calls will return a fast answer
364 /// based on a cached result. This includes caching negative results.
365 Optional<SCEVCache::GEPDescriptor>
366 SCEVCache::getGEPDescriptor(GetElementPtrInst *GEP) {
367 decltype(GEPDescriptors)::iterator It;
370 std::tie(It, Inserted) = GEPDescriptors.insert({GEP, {}});
373 if (!It->second.BaseAddr)
379 // We've inserted a new record into the cache, so compute the GEP descriptor
381 Value *V = cast<Value>(GEP);
382 if (!SE.isSCEVable(V->getType()))
384 const SCEV *S = SE.getSCEV(V);
386 // FIXME: It'd be nice if the worklist and set used by the
387 // SCEVTraversal could be re-used between loop iterations, but the
388 // interface doesn't support that. There is no way to clear the visited
389 // sets between uses.
390 FindConstantPointers Visitor(&L, SE);
391 SCEVTraversal<FindConstantPointers> T(Visitor);
393 // Try to find (BaseAddress+Step+Offset) tuple.
394 // If succeeded, save it to the cache - it might help in folding
397 if (!Visitor.IndexIsConstant || !Visitor.BaseAddress)
400 const SCEV *BaseAddrSE = SE.getSCEV(Visitor.BaseAddress);
401 if (BaseAddrSE->getType() != S->getType())
403 const SCEV *OffSE = SE.getMinusSCEV(S, BaseAddrSE);
404 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OffSE);
409 const SCEVConstant *StepSE =
410 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE));
411 const SCEVConstant *StartSE = dyn_cast<SCEVConstant>(AR->getStart());
412 if (!StepSE || !StartSE)
415 // Check and skip caching if doing so would require lots of bits to
417 APInt Start = StartSE->getValue()->getValue();
418 APInt Step = StepSE->getValue()->getValue();
419 if (Start.getActiveBits() > 32 || Step.getActiveBits() > 32)
422 // We found a cacheable SCEV model for the GEP.
423 It->second.BaseAddr = Visitor.BaseAddress;
424 It->second.Start = Start.getLimitedValue();
425 It->second.Step = Step.getLimitedValue();
430 // This class is used to get an estimate of the optimization effects that we
431 // could get from complete loop unrolling. It comes from the fact that some
432 // loads might be replaced with concrete constant values and that could trigger
433 // a chain of instruction simplifications.
435 // E.g. we might have:
436 // int a[] = {0, 1, 0};
438 // for (i = 0; i < 3; i ++)
440 // If we completely unroll the loop, we would get:
441 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
442 // Which then will be simplified to:
443 // v = b[0]* 0 + b[1]* 1 + b[2]* 0
446 class UnrolledInstAnalyzer : private InstVisitor<UnrolledInstAnalyzer, bool> {
447 typedef InstVisitor<UnrolledInstAnalyzer, bool> Base;
448 friend class InstVisitor<UnrolledInstAnalyzer, bool>;
451 UnrolledInstAnalyzer(unsigned Iteration,
452 DenseMap<Value *, Constant *> &SimplifiedValues,
454 : Iteration(Iteration), SimplifiedValues(SimplifiedValues), SC(SC) {}
456 // Allow access to the initial visit method.
460 /// \brief Number of currently simulated iteration.
462 /// If an expression is ConstAddress+Constant, then the Constant is
463 /// Start + Iteration*Step, where Start and Step could be obtained from
467 // While we walk the loop instructions, we we build up and maintain a mapping
468 // of simplified values specific to this iteration. The idea is to propagate
469 // any special information we have about loads that can be replaced with
470 // constants after complete unrolling, and account for likely simplifications
472 DenseMap<Value *, Constant *> &SimplifiedValues;
474 // We use a cache to wrap all our SCEV queries.
477 /// Base case for the instruction visitor.
478 bool visitInstruction(Instruction &I) { return false; };
480 /// TODO: Add visitors for other instruction types, e.g. ZExt, SExt.
482 /// Try to simplify binary operator I.
484 /// TODO: Probaly it's worth to hoist the code for estimating the
485 /// simplifications effects to a separate class, since we have a very similar
486 /// code in InlineCost already.
487 bool visitBinaryOperator(BinaryOperator &I) {
488 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
489 if (!isa<Constant>(LHS))
490 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
492 if (!isa<Constant>(RHS))
493 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
495 Value *SimpleV = nullptr;
496 const DataLayout &DL = I.getModule()->getDataLayout();
497 if (auto FI = dyn_cast<FPMathOperator>(&I))
499 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags(), DL);
501 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL);
503 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV))
504 SimplifiedValues[&I] = C;
509 /// Try to fold load I.
510 bool visitLoad(LoadInst &I) {
511 Value *AddrOp = I.getPointerOperand();
512 if (!isa<Constant>(AddrOp))
513 if (Constant *SimplifiedAddrOp = SimplifiedValues.lookup(AddrOp))
514 AddrOp = SimplifiedAddrOp;
516 auto *GEP = dyn_cast<GetElementPtrInst>(AddrOp);
519 auto OptionalGEPDesc = SC.getGEPDescriptor(GEP);
520 if (!OptionalGEPDesc)
523 auto GV = dyn_cast<GlobalVariable>(OptionalGEPDesc->BaseAddr);
524 // We're only interested in loads that can be completely folded to a
526 if (!GV || !GV->hasInitializer())
529 ConstantDataSequential *CDS =
530 dyn_cast<ConstantDataSequential>(GV->getInitializer());
534 // This calculation should never overflow because we bound Iteration quite
535 // low and both the start and step are 32-bit integers. We use signed
536 // integers so that UBSan will catch if a bug sneaks into the code.
537 int ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
538 int64_t Index = ((int64_t)OptionalGEPDesc->Start +
539 (int64_t)OptionalGEPDesc->Step * (int64_t)Iteration) /
541 if (Index >= CDS->getNumElements()) {
542 // FIXME: For now we conservatively ignore out of bound accesses, but
543 // we're allowed to perform the optimization in this case.
547 Constant *CV = CDS->getElementAsConstant(Index);
548 assert(CV && "Constant expected.");
549 SimplifiedValues[&I] = CV;
558 struct EstimatedUnrollCost {
559 /// \brief Count the number of optimized instructions.
560 unsigned NumberOfOptimizedInstructions;
562 /// \brief Count the total number of instructions.
563 unsigned UnrolledLoopSize;
567 /// \brief Figure out if the loop is worth full unrolling.
569 /// Complete loop unrolling can make some loads constant, and we need to know
570 /// if that would expose any further optimization opportunities. This routine
571 /// estimates this optimization. It assigns computed number of instructions,
572 /// that potentially might be optimized away, to
573 /// NumberOfOptimizedInstructions, and total number of instructions to
574 /// UnrolledLoopSize (not counting blocks that won't be reached, if we were
575 /// able to compute the condition).
576 /// \returns false if we can't analyze the loop, or if we discovered that
577 /// unrolling won't give anything. Otherwise, returns true.
578 Optional<EstimatedUnrollCost>
579 analyzeLoopUnrollCost(const Loop *L, unsigned TripCount, ScalarEvolution &SE,
580 const TargetTransformInfo &TTI,
581 unsigned MaxUnrolledLoopSize) {
582 // We want to be able to scale offsets by the trip count and add more offsets
583 // to them without checking for overflows, and we already don't want to
584 // analyze *massive* trip counts, so we force the max to be reasonably small.
585 assert(UnrollMaxIterationsCountToAnalyze < (INT_MAX / 2) &&
586 "The unroll iterations max is too large!");
588 // Don't simulate loops with a big or unknown tripcount
589 if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
590 TripCount > UnrollMaxIterationsCountToAnalyze)
593 SmallSetVector<BasicBlock *, 16> BBWorklist;
594 DenseMap<Value *, Constant *> SimplifiedValues;
596 // Use a cache to access SCEV expressions so that we don't pay the cost on
597 // each iteration. This cache is lazily self-populating.
598 SCEVCache SC(*L, SE);
600 unsigned NumberOfOptimizedInstructions = 0;
601 unsigned UnrolledLoopSize = 0;
603 // Simulate execution of each iteration of the loop counting instructions,
604 // which would be simplified.
605 // Since the same load will take different values on different iterations,
606 // we literally have to go through all loop's iterations.
607 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
608 SimplifiedValues.clear();
609 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SC);
612 BBWorklist.insert(L->getHeader());
613 // Note that we *must not* cache the size, this loop grows the worklist.
614 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
615 BasicBlock *BB = BBWorklist[Idx];
617 // Visit all instructions in the given basic block and try to simplify
618 // it. We don't change the actual IR, just count optimization
620 for (Instruction &I : *BB) {
621 UnrolledLoopSize += TTI.getUserCost(&I);
623 // Visit the instruction to analyze its loop cost after unrolling,
624 // and if the visitor returns true, then we can optimize this
626 if (Analyzer.visit(I))
627 NumberOfOptimizedInstructions += TTI.getUserCost(&I);
629 // If unrolled body turns out to be too big, bail out.
630 if (UnrolledLoopSize - NumberOfOptimizedInstructions >
635 // Add BB's successors to the worklist.
636 for (BasicBlock *Succ : successors(BB))
637 if (L->contains(Succ))
638 BBWorklist.insert(Succ);
641 // If we found no optimization opportunities on the first iteration, we
642 // won't find them on later ones too.
643 if (!NumberOfOptimizedInstructions)
646 return {{NumberOfOptimizedInstructions, UnrolledLoopSize}};
649 /// ApproximateLoopSize - Approximate the size of the loop.
650 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
651 bool &NotDuplicatable,
652 const TargetTransformInfo &TTI,
653 AssumptionCache *AC) {
654 SmallPtrSet<const Value *, 32> EphValues;
655 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
658 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
660 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
661 NumCalls = Metrics.NumInlineCandidates;
662 NotDuplicatable = Metrics.notDuplicatable;
664 unsigned LoopSize = Metrics.NumInsts;
666 // Don't allow an estimate of size zero. This would allows unrolling of loops
667 // with huge iteration counts, which is a compile time problem even if it's
668 // not a problem for code quality. Also, the code using this size may assume
669 // that each loop has at least three instructions (likely a conditional
670 // branch, a comparison feeding that branch, and some kind of loop increment
671 // feeding that comparison instruction).
672 LoopSize = std::max(LoopSize, 3u);
677 // Returns the loop hint metadata node with the given name (for example,
678 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
680 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
681 if (MDNode *LoopID = L->getLoopID())
682 return GetUnrollMetadata(LoopID, Name);
686 // Returns true if the loop has an unroll(full) pragma.
687 static bool HasUnrollFullPragma(const Loop *L) {
688 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
691 // Returns true if the loop has an unroll(disable) pragma.
692 static bool HasUnrollDisablePragma(const Loop *L) {
693 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
696 // Returns true if the loop has an runtime unroll(disable) pragma.
697 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
698 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
701 // If loop has an unroll_count pragma return the (necessarily
702 // positive) value from the pragma. Otherwise return 0.
703 static unsigned UnrollCountPragmaValue(const Loop *L) {
704 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
706 assert(MD->getNumOperands() == 2 &&
707 "Unroll count hint metadata should have two operands.");
709 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
710 assert(Count >= 1 && "Unroll count must be positive.");
716 // Remove existing unroll metadata and add unroll disable metadata to
717 // indicate the loop has already been unrolled. This prevents a loop
718 // from being unrolled more than is directed by a pragma if the loop
719 // unrolling pass is run more than once (which it generally is).
720 static void SetLoopAlreadyUnrolled(Loop *L) {
721 MDNode *LoopID = L->getLoopID();
724 // First remove any existing loop unrolling metadata.
725 SmallVector<Metadata *, 4> MDs;
726 // Reserve first location for self reference to the LoopID metadata node.
727 MDs.push_back(nullptr);
728 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
729 bool IsUnrollMetadata = false;
730 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
732 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
733 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
735 if (!IsUnrollMetadata)
736 MDs.push_back(LoopID->getOperand(i));
739 // Add unroll(disable) metadata to disable future unrolling.
740 LLVMContext &Context = L->getHeader()->getContext();
741 SmallVector<Metadata *, 1> DisableOperands;
742 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
743 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
744 MDs.push_back(DisableNode);
746 MDNode *NewLoopID = MDNode::get(Context, MDs);
747 // Set operand 0 to refer to the loop id itself.
748 NewLoopID->replaceOperandWith(0, NewLoopID);
749 L->setLoopID(NewLoopID);
752 bool LoopUnroll::canUnrollCompletely(
753 Loop *L, unsigned Threshold, unsigned AbsoluteThreshold,
754 uint64_t UnrolledSize, unsigned NumberOfOptimizedInstructions,
755 unsigned PercentOfOptimizedForCompleteUnroll) {
757 if (Threshold == NoThreshold) {
758 DEBUG(dbgs() << " Can fully unroll, because no threshold is set.\n");
762 if (UnrolledSize <= Threshold) {
763 DEBUG(dbgs() << " Can fully unroll, because unrolled size: "
764 << UnrolledSize << "<" << Threshold << "\n");
768 assert(UnrolledSize && "UnrolledSize can't be 0 at this point.");
769 unsigned PercentOfOptimizedInstructions =
770 (uint64_t)NumberOfOptimizedInstructions * 100ull / UnrolledSize;
772 if (UnrolledSize <= AbsoluteThreshold &&
773 PercentOfOptimizedInstructions >= PercentOfOptimizedForCompleteUnroll) {
774 DEBUG(dbgs() << " Can fully unroll, because unrolling will help removing "
775 << PercentOfOptimizedInstructions
776 << "% instructions (threshold: "
777 << PercentOfOptimizedForCompleteUnroll << "%)\n");
778 DEBUG(dbgs() << " Unrolled size (" << UnrolledSize
779 << ") is less than the threshold (" << AbsoluteThreshold
784 DEBUG(dbgs() << " Too large to fully unroll:\n");
785 DEBUG(dbgs() << " Unrolled size: " << UnrolledSize << "\n");
786 DEBUG(dbgs() << " Estimated number of optimized instructions: "
787 << NumberOfOptimizedInstructions << "\n");
788 DEBUG(dbgs() << " Absolute threshold: " << AbsoluteThreshold << "\n");
789 DEBUG(dbgs() << " Minimum percent of removed instructions: "
790 << PercentOfOptimizedForCompleteUnroll << "\n");
791 DEBUG(dbgs() << " Threshold for small loops: " << Threshold << "\n");
795 unsigned LoopUnroll::selectUnrollCount(
796 const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
797 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
798 bool &SetExplicitly) {
799 SetExplicitly = true;
801 // User-specified count (either as a command-line option or
802 // constructor parameter) has highest precedence.
803 unsigned Count = UserCount ? CurrentCount : 0;
805 // If there is no user-specified count, unroll pragmas have the next
806 // highest precendence.
810 } else if (PragmaFullUnroll) {
819 SetExplicitly = false;
821 // Runtime trip count.
822 Count = UnrollRuntimeCount;
824 // Conservative heuristic: if we know the trip count, see if we can
825 // completely unroll (subject to the threshold, checked below); otherwise
826 // try to find greatest modulo of the trip count which is still under
830 if (TripCount && Count > TripCount)
835 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
836 if (skipOptnoneFunction(L))
839 Function &F = *L->getHeader()->getParent();
841 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
842 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
843 const TargetTransformInfo &TTI =
844 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
845 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
847 BasicBlock *Header = L->getHeader();
848 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
849 << "] Loop %" << Header->getName() << "\n");
851 if (HasUnrollDisablePragma(L)) {
854 bool PragmaFullUnroll = HasUnrollFullPragma(L);
855 unsigned PragmaCount = UnrollCountPragmaValue(L);
856 bool HasPragma = PragmaFullUnroll || PragmaCount > 0;
858 TargetTransformInfo::UnrollingPreferences UP;
859 getUnrollingPreferences(L, TTI, UP);
861 // Find trip count and trip multiple if count is not available
862 unsigned TripCount = 0;
863 unsigned TripMultiple = 1;
864 // If there are multiple exiting blocks but one of them is the latch, use the
865 // latch for the trip count estimation. Otherwise insist on a single exiting
866 // block for the trip count estimation.
867 BasicBlock *ExitingBlock = L->getLoopLatch();
868 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
869 ExitingBlock = L->getExitingBlock();
871 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
872 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
875 // Select an initial unroll count. This may be reduced later based
876 // on size thresholds.
877 bool CountSetExplicitly;
878 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll,
879 PragmaCount, UP, CountSetExplicitly);
881 unsigned NumInlineCandidates;
882 bool notDuplicatable;
884 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
885 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
887 // When computing the unrolled size, note that the conditional branch on the
888 // backedge and the comparison feeding it are not replicated like the rest of
889 // the loop body (which is why 2 is subtracted).
890 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
891 if (notDuplicatable) {
892 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
893 << " instructions.\n");
896 if (NumInlineCandidates != 0) {
897 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
901 unsigned Threshold, PartialThreshold;
902 unsigned AbsoluteThreshold, PercentOfOptimizedForCompleteUnroll;
903 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold,
904 AbsoluteThreshold, PercentOfOptimizedForCompleteUnroll);
906 // Given Count, TripCount and thresholds determine the type of
907 // unrolling which is to be performed.
908 enum { Full = 0, Partial = 1, Runtime = 2 };
910 if (TripCount && Count == TripCount) {
912 // If the loop is really small, we don't need to run an expensive analysis.
913 if (canUnrollCompletely(
914 L, Threshold, AbsoluteThreshold,
915 UnrolledSize, 0, 100)) {
918 // The loop isn't that small, but we still can fully unroll it if that
919 // helps to remove a significant number of instructions.
920 // To check that, run additional analysis on the loop.
921 if (Optional<EstimatedUnrollCost> Cost =
922 analyzeLoopUnrollCost(L, TripCount, *SE, TTI, AbsoluteThreshold))
923 if (canUnrollCompletely(L, Threshold, AbsoluteThreshold,
924 Cost->UnrolledLoopSize,
925 Cost->NumberOfOptimizedInstructions,
926 PercentOfOptimizedForCompleteUnroll)) {
930 } else if (TripCount && Count < TripCount) {
936 // Reduce count based on the type of unrolling and the threshold values.
937 unsigned OriginalCount = Count;
938 bool AllowRuntime = UserRuntime ? CurrentRuntime : UP.Runtime;
939 if (HasRuntimeUnrollDisablePragma(L)) {
940 AllowRuntime = false;
942 if (Unrolling == Partial) {
943 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
944 if (!AllowPartial && !CountSetExplicitly) {
945 DEBUG(dbgs() << " will not try to unroll partially because "
946 << "-unroll-allow-partial not given\n");
949 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
950 // Reduce unroll count to be modulo of TripCount for partial unrolling.
951 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
952 while (Count != 0 && TripCount % Count != 0)
955 } else if (Unrolling == Runtime) {
956 if (!AllowRuntime && !CountSetExplicitly) {
957 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
958 << "-unroll-runtime not given\n");
961 // Reduce unroll count to be the largest power-of-two factor of
962 // the original count which satisfies the threshold limit.
963 while (Count != 0 && UnrolledSize > PartialThreshold) {
965 UnrolledSize = (LoopSize-2) * Count + 2;
967 if (Count > UP.MaxCount)
969 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
973 if (PragmaCount != 0)
974 // If loop has an unroll count pragma mark loop as unrolled to prevent
975 // unrolling beyond that requested by the pragma.
976 SetLoopAlreadyUnrolled(L);
978 // Emit optimization remarks if we are unable to unroll the loop
979 // as directed by a pragma.
980 DebugLoc LoopLoc = L->getStartLoc();
981 Function *F = Header->getParent();
982 LLVMContext &Ctx = F->getContext();
983 if (PragmaFullUnroll && PragmaCount == 0) {
984 if (TripCount && Count != TripCount) {
985 emitOptimizationRemarkMissed(
986 Ctx, DEBUG_TYPE, *F, LoopLoc,
987 "Unable to fully unroll loop as directed by unroll(full) pragma "
988 "because unrolled size is too large.");
989 } else if (!TripCount) {
990 emitOptimizationRemarkMissed(
991 Ctx, DEBUG_TYPE, *F, LoopLoc,
992 "Unable to fully unroll loop as directed by unroll(full) pragma "
993 "because loop has a runtime trip count.");
995 } else if (PragmaCount > 0 && Count != OriginalCount) {
996 emitOptimizationRemarkMissed(
997 Ctx, DEBUG_TYPE, *F, LoopLoc,
998 "Unable to unroll loop the number of times directed by "
999 "unroll_count pragma because unrolled size is too large.");
1003 if (Unrolling != Full && Count < 2) {
1004 // Partial unrolling by 1 is a nop. For full unrolling, a factor
1005 // of 1 makes sense because loop control can be eliminated.
1010 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, UP.AllowExpensiveTripCount,
1011 TripMultiple, LI, this, &LPM, &AC))