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 baseline cost threshold for loop unrolling"));
44 static cl::opt<unsigned> UnrollPercentDynamicCostSavedThreshold(
45 "unroll-percent-dynamic-cost-saved-threshold", cl::init(20), cl::Hidden,
46 cl::desc("The percentage of estimated dynamic cost which must be saved by "
47 "unrolling to allow unrolling up to the max threshold."));
49 static cl::opt<unsigned> UnrollDynamicCostSavingsDiscount(
50 "unroll-dynamic-cost-savings-discount", cl::init(2000), cl::Hidden,
51 cl::desc("This is the amount discounted from the total unroll cost when "
52 "the unrolled form has a high dynamic cost savings (triggered by "
53 "the '-unroll-perecent-dynamic-cost-saved-threshold' flag)."));
55 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
56 "unroll-max-iteration-count-to-analyze", cl::init(0), cl::Hidden,
57 cl::desc("Don't allow loop unrolling to simulate more than this number of"
58 "iterations when checking full unroll profitability"));
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 CurrentPercentDynamicCostSavedThreshold =
86 UnrollPercentDynamicCostSavedThreshold;
87 CurrentDynamicCostSavingsDiscount = UnrollDynamicCostSavingsDiscount;
88 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
89 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
90 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
92 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
93 UserPercentDynamicCostSavedThreshold =
94 (UnrollPercentDynamicCostSavedThreshold.getNumOccurrences() > 0);
95 UserDynamicCostSavingsDiscount =
96 (UnrollDynamicCostSavingsDiscount.getNumOccurrences() > 0);
97 UserAllowPartial = (P != -1) ||
98 (UnrollAllowPartial.getNumOccurrences() > 0);
99 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
100 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0);
102 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
105 /// A magic value for use with the Threshold parameter to indicate
106 /// that the loop unroll should be performed regardless of how much
107 /// code expansion would result.
108 static const unsigned NoThreshold = UINT_MAX;
110 // Threshold to use when optsize is specified (and there is no
111 // explicit -unroll-threshold).
112 static const unsigned OptSizeUnrollThreshold = 50;
114 // Default unroll count for loops with run-time trip count if
115 // -unroll-count is not set
116 static const unsigned UnrollRuntimeCount = 8;
118 unsigned CurrentCount;
119 unsigned CurrentThreshold;
120 unsigned CurrentPercentDynamicCostSavedThreshold;
121 unsigned CurrentDynamicCostSavingsDiscount;
122 bool CurrentAllowPartial;
125 // Flags for whether the 'current' settings are user-specified.
128 bool UserPercentDynamicCostSavedThreshold;
129 bool UserDynamicCostSavingsDiscount;
130 bool UserAllowPartial;
133 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
135 /// This transformation requires natural loop information & requires that
136 /// loop preheaders be inserted into the CFG...
138 void getAnalysisUsage(AnalysisUsage &AU) const override {
139 AU.addRequired<AssumptionCacheTracker>();
140 AU.addRequired<LoopInfoWrapperPass>();
141 AU.addPreserved<LoopInfoWrapperPass>();
142 AU.addRequiredID(LoopSimplifyID);
143 AU.addPreservedID(LoopSimplifyID);
144 AU.addRequiredID(LCSSAID);
145 AU.addPreservedID(LCSSAID);
146 AU.addRequired<ScalarEvolution>();
147 AU.addPreserved<ScalarEvolution>();
148 AU.addRequired<TargetTransformInfoWrapperPass>();
149 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
150 // If loop unroll does not preserve dom info then LCSSA pass on next
151 // loop will receive invalid dom info.
152 // For now, recreate dom info, if loop is unrolled.
153 AU.addPreserved<DominatorTreeWrapperPass>();
156 // Fill in the UnrollingPreferences parameter with values from the
157 // TargetTransformationInfo.
158 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI,
159 TargetTransformInfo::UnrollingPreferences &UP) {
160 UP.Threshold = CurrentThreshold;
161 UP.PercentDynamicCostSavedThreshold =
162 CurrentPercentDynamicCostSavedThreshold;
163 UP.DynamicCostSavingsDiscount = CurrentDynamicCostSavingsDiscount;
164 UP.OptSizeThreshold = OptSizeUnrollThreshold;
165 UP.PartialThreshold = CurrentThreshold;
166 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
167 UP.Count = CurrentCount;
168 UP.MaxCount = UINT_MAX;
169 UP.Partial = CurrentAllowPartial;
170 UP.Runtime = CurrentRuntime;
171 UP.AllowExpensiveTripCount = false;
172 TTI.getUnrollingPreferences(L, UP);
175 // Select and return an unroll count based on parameters from
176 // user, unroll preferences, unroll pragmas, or a heuristic.
177 // SetExplicitly is set to true if the unroll count is is set by
178 // the user or a pragma rather than selected heuristically.
180 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
181 unsigned PragmaCount,
182 const TargetTransformInfo::UnrollingPreferences &UP,
183 bool &SetExplicitly);
185 // Select threshold values used to limit unrolling based on a
186 // total unrolled size. Parameters Threshold and PartialThreshold
187 // are set to the maximum unrolled size for fully and partially
188 // unrolled loops respectively.
189 void selectThresholds(const Loop *L, bool HasPragma,
190 const TargetTransformInfo::UnrollingPreferences &UP,
191 unsigned &Threshold, unsigned &PartialThreshold,
192 unsigned &PercentDynamicCostSavedThreshold,
193 unsigned &DynamicCostSavingsDiscount) {
194 // Determine the current unrolling threshold. While this is
195 // normally set from UnrollThreshold, it is overridden to a
196 // smaller value if the current function is marked as
197 // optimize-for-size, and the unroll threshold was not user
199 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
200 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
201 PercentDynamicCostSavedThreshold =
202 UserPercentDynamicCostSavedThreshold
203 ? CurrentPercentDynamicCostSavedThreshold
204 : UP.PercentDynamicCostSavedThreshold;
205 DynamicCostSavingsDiscount = UserDynamicCostSavingsDiscount
206 ? CurrentDynamicCostSavingsDiscount
207 : UP.DynamicCostSavingsDiscount;
209 if (!UserThreshold &&
210 L->getHeader()->getParent()->hasFnAttribute(
211 Attribute::OptimizeForSize)) {
212 Threshold = UP.OptSizeThreshold;
213 PartialThreshold = UP.PartialOptSizeThreshold;
216 // If the loop has an unrolling pragma, we want to be more
217 // aggressive with unrolling limits. Set thresholds to at
218 // least the PragmaTheshold value which is larger than the
220 if (Threshold != NoThreshold)
221 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold);
222 if (PartialThreshold != NoThreshold)
224 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold);
227 bool canUnrollCompletely(Loop *L, unsigned Threshold,
228 unsigned PercentDynamicCostSavedThreshold,
229 unsigned DynamicCostSavingsDiscount,
230 unsigned UnrolledCost, unsigned RolledDynamicCost);
234 char LoopUnroll::ID = 0;
235 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
236 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
237 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
238 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
239 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
240 INITIALIZE_PASS_DEPENDENCY(LCSSA)
241 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
242 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
244 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial,
246 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime);
249 Pass *llvm::createSimpleLoopUnrollPass() {
250 return llvm::createLoopUnrollPass(-1, -1, 0, 0);
254 /// \brief SCEV expressions visitor used for finding expressions that would
255 /// become constants if the loop L is unrolled.
256 struct FindConstantPointers {
257 /// \brief Shows whether the expression is ConstAddress+Constant or not.
258 bool IndexIsConstant;
260 /// \brief Used for filtering out SCEV expressions with two or more AddRec
263 /// Used to filter out complicated SCEV expressions, having several AddRec
264 /// sub-expressions. We don't handle them, because unrolling one loop
265 /// would help to replace only one of these inductions with a constant, and
266 /// consequently, the expression would remain non-constant.
269 /// \brief If the SCEV expression becomes ConstAddress+Constant, this value
270 /// holds ConstAddress. Otherwise, it's nullptr.
273 /// \brief The loop, which we try to completely unroll.
278 FindConstantPointers(const Loop *L, ScalarEvolution &SE)
279 : IndexIsConstant(true), HaveSeenAR(false), BaseAddress(nullptr),
282 /// Examine the given expression S and figure out, if it can be a part of an
283 /// expression, that could become a constant after the loop is unrolled.
284 /// The routine sets IndexIsConstant and HaveSeenAR according to the analysis
286 /// \returns true if we need to examine subexpressions, and false otherwise.
287 bool follow(const SCEV *S) {
288 if (const SCEVUnknown *SC = dyn_cast<SCEVUnknown>(S)) {
289 // We've reached the leaf node of SCEV, it's most probably just a
291 // If it's the only one SCEV-subexpression, then it might be a base
292 // address of an index expression.
293 // If we've already recorded base address, then just give up on this SCEV
294 // - it's too complicated.
296 IndexIsConstant = false;
299 BaseAddress = SC->getValue();
302 if (isa<SCEVConstant>(S))
304 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
305 // If the current SCEV expression is AddRec, and its loop isn't the loop
306 // we are about to unroll, then we won't get a constant address after
307 // unrolling, and thus, won't be able to eliminate the load.
308 if (AR->getLoop() != L) {
309 IndexIsConstant = false;
312 // We don't handle multiple AddRecs here, so give up in this case.
314 IndexIsConstant = false;
320 // Continue traversal.
323 bool isDone() const { return !IndexIsConstant; }
325 } // End anonymous namespace.
328 /// \brief A cache of SCEV results used to optimize repeated queries to SCEV on
329 /// the same set of instructions.
331 /// The primary cost this saves is the cost of checking the validity of a SCEV
332 /// every time it is looked up. However, in some cases we can provide a reduced
333 /// and especially useful model for an instruction based upon SCEV that is
334 /// non-trivial to compute but more useful to clients.
337 /// \brief Struct to represent a GEP whose start and step are known fixed
338 /// offsets from a base address due to SCEV's analysis.
339 struct GEPDescriptor {
340 Value *BaseAddr = nullptr;
345 Optional<GEPDescriptor> getGEPDescriptor(GetElementPtrInst *GEP);
347 SCEVCache(const Loop &L, ScalarEvolution &SE) : L(L), SE(SE) {}
353 SmallDenseMap<GetElementPtrInst *, GEPDescriptor> GEPDescriptors;
355 } // End anonymous namespace.
357 /// \brief Get a simplified descriptor for a GEP instruction.
359 /// Where possible, this produces a simplified descriptor for a GEP instruction
360 /// using SCEV analysis of the containing loop. If this isn't possible, it
361 /// returns an empty optional.
363 /// The model is a base address, an initial offset, and a per-iteration step.
364 /// This fits very common patterns of GEPs inside loops and is something we can
365 /// use to simulate the behavior of a particular iteration of a loop.
367 /// This is a cached interface. The first call may do non-trivial work to
368 /// compute the result, but all subsequent calls will return a fast answer
369 /// based on a cached result. This includes caching negative results.
370 Optional<SCEVCache::GEPDescriptor>
371 SCEVCache::getGEPDescriptor(GetElementPtrInst *GEP) {
372 decltype(GEPDescriptors)::iterator It;
375 std::tie(It, Inserted) = GEPDescriptors.insert({GEP, {}});
378 if (!It->second.BaseAddr)
384 // We've inserted a new record into the cache, so compute the GEP descriptor
386 Value *V = cast<Value>(GEP);
387 if (!SE.isSCEVable(V->getType()))
389 const SCEV *S = SE.getSCEV(V);
391 // FIXME: It'd be nice if the worklist and set used by the
392 // SCEVTraversal could be re-used between loop iterations, but the
393 // interface doesn't support that. There is no way to clear the visited
394 // sets between uses.
395 FindConstantPointers Visitor(&L, SE);
396 SCEVTraversal<FindConstantPointers> T(Visitor);
398 // Try to find (BaseAddress+Step+Offset) tuple.
399 // If succeeded, save it to the cache - it might help in folding
402 if (!Visitor.IndexIsConstant || !Visitor.BaseAddress)
405 const SCEV *BaseAddrSE = SE.getSCEV(Visitor.BaseAddress);
406 if (BaseAddrSE->getType() != S->getType())
408 const SCEV *OffSE = SE.getMinusSCEV(S, BaseAddrSE);
409 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OffSE);
414 const SCEVConstant *StepSE =
415 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE));
416 const SCEVConstant *StartSE = dyn_cast<SCEVConstant>(AR->getStart());
417 if (!StepSE || !StartSE)
420 // Check and skip caching if doing so would require lots of bits to
422 APInt Start = StartSE->getValue()->getValue();
423 APInt Step = StepSE->getValue()->getValue();
424 if (Start.getActiveBits() > 32 || Step.getActiveBits() > 32)
427 // We found a cacheable SCEV model for the GEP.
428 It->second.BaseAddr = Visitor.BaseAddress;
429 It->second.Start = Start.getLimitedValue();
430 It->second.Step = Step.getLimitedValue();
435 // This class is used to get an estimate of the optimization effects that we
436 // could get from complete loop unrolling. It comes from the fact that some
437 // loads might be replaced with concrete constant values and that could trigger
438 // a chain of instruction simplifications.
440 // E.g. we might have:
441 // int a[] = {0, 1, 0};
443 // for (i = 0; i < 3; i ++)
445 // If we completely unroll the loop, we would get:
446 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
447 // Which then will be simplified to:
448 // v = b[0]* 0 + b[1]* 1 + b[2]* 0
451 class UnrolledInstAnalyzer : private InstVisitor<UnrolledInstAnalyzer, bool> {
452 typedef InstVisitor<UnrolledInstAnalyzer, bool> Base;
453 friend class InstVisitor<UnrolledInstAnalyzer, bool>;
456 UnrolledInstAnalyzer(unsigned Iteration,
457 DenseMap<Value *, Constant *> &SimplifiedValues,
459 : Iteration(Iteration), SimplifiedValues(SimplifiedValues), SC(SC) {}
461 // Allow access to the initial visit method.
465 /// \brief Number of currently simulated iteration.
467 /// If an expression is ConstAddress+Constant, then the Constant is
468 /// Start + Iteration*Step, where Start and Step could be obtained from
472 // While we walk the loop instructions, we we build up and maintain a mapping
473 // of simplified values specific to this iteration. The idea is to propagate
474 // any special information we have about loads that can be replaced with
475 // constants after complete unrolling, and account for likely simplifications
477 DenseMap<Value *, Constant *> &SimplifiedValues;
479 // We use a cache to wrap all our SCEV queries.
482 /// Base case for the instruction visitor.
483 bool visitInstruction(Instruction &I) { return false; };
485 /// TODO: Add visitors for other instruction types, e.g. ZExt, SExt.
487 /// Try to simplify binary operator I.
489 /// TODO: Probaly it's worth to hoist the code for estimating the
490 /// simplifications effects to a separate class, since we have a very similar
491 /// code in InlineCost already.
492 bool visitBinaryOperator(BinaryOperator &I) {
493 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
494 if (!isa<Constant>(LHS))
495 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
497 if (!isa<Constant>(RHS))
498 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
500 Value *SimpleV = nullptr;
501 const DataLayout &DL = I.getModule()->getDataLayout();
502 if (auto FI = dyn_cast<FPMathOperator>(&I))
504 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags(), DL);
506 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL);
508 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV))
509 SimplifiedValues[&I] = C;
514 /// Try to fold load I.
515 bool visitLoad(LoadInst &I) {
516 Value *AddrOp = I.getPointerOperand();
517 if (!isa<Constant>(AddrOp))
518 if (Constant *SimplifiedAddrOp = SimplifiedValues.lookup(AddrOp))
519 AddrOp = SimplifiedAddrOp;
521 auto *GEP = dyn_cast<GetElementPtrInst>(AddrOp);
524 auto OptionalGEPDesc = SC.getGEPDescriptor(GEP);
525 if (!OptionalGEPDesc)
528 auto GV = dyn_cast<GlobalVariable>(OptionalGEPDesc->BaseAddr);
529 // We're only interested in loads that can be completely folded to a
531 if (!GV || !GV->hasInitializer())
534 ConstantDataSequential *CDS =
535 dyn_cast<ConstantDataSequential>(GV->getInitializer());
539 // This calculation should never overflow because we bound Iteration quite
540 // low and both the start and step are 32-bit integers. We use signed
541 // integers so that UBSan will catch if a bug sneaks into the code.
542 int ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
543 int64_t Index = ((int64_t)OptionalGEPDesc->Start +
544 (int64_t)OptionalGEPDesc->Step * (int64_t)Iteration) /
546 if (Index >= CDS->getNumElements()) {
547 // FIXME: For now we conservatively ignore out of bound accesses, but
548 // we're allowed to perform the optimization in this case.
552 Constant *CV = CDS->getElementAsConstant(Index);
553 assert(CV && "Constant expected.");
554 SimplifiedValues[&I] = CV;
563 struct EstimatedUnrollCost {
564 /// \brief The estimated cost after unrolling.
565 unsigned UnrolledCost;
567 /// \brief The estimated dynamic cost of executing the instructions in the
569 unsigned RolledDynamicCost;
573 /// \brief Figure out if the loop is worth full unrolling.
575 /// Complete loop unrolling can make some loads constant, and we need to know
576 /// if that would expose any further optimization opportunities. This routine
577 /// estimates this optimization. It assigns computed number of instructions,
578 /// that potentially might be optimized away, to
579 /// NumberOfOptimizedInstructions, and total number of instructions to
580 /// UnrolledLoopSize (not counting blocks that won't be reached, if we were
581 /// able to compute the condition).
582 /// \returns false if we can't analyze the loop, or if we discovered that
583 /// unrolling won't give anything. Otherwise, returns true.
584 Optional<EstimatedUnrollCost>
585 analyzeLoopUnrollCost(const Loop *L, unsigned TripCount, ScalarEvolution &SE,
586 const TargetTransformInfo &TTI,
587 unsigned MaxUnrolledLoopSize) {
588 // We want to be able to scale offsets by the trip count and add more offsets
589 // to them without checking for overflows, and we already don't want to
590 // analyze *massive* trip counts, so we force the max to be reasonably small.
591 assert(UnrollMaxIterationsCountToAnalyze < (INT_MAX / 2) &&
592 "The unroll iterations max is too large!");
594 // Don't simulate loops with a big or unknown tripcount
595 if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
596 TripCount > UnrollMaxIterationsCountToAnalyze)
599 SmallSetVector<BasicBlock *, 16> BBWorklist;
600 DenseMap<Value *, Constant *> SimplifiedValues;
602 // Use a cache to access SCEV expressions so that we don't pay the cost on
603 // each iteration. This cache is lazily self-populating.
604 SCEVCache SC(*L, SE);
606 // The estimated cost of the unrolled form of the loop. We try to estimate
607 // this by simplifying as much as we can while computing the estimate.
608 unsigned UnrolledCost = 0;
609 // We also track the estimated dynamic (that is, actually executed) cost in
610 // the rolled form. This helps identify cases when the savings from unrolling
611 // aren't just exposing dead control flows, but actual reduced dynamic
612 // instructions due to the simplifications which we expect to occur after
614 unsigned RolledDynamicCost = 0;
616 // Simulate execution of each iteration of the loop counting instructions,
617 // which would be simplified.
618 // Since the same load will take different values on different iterations,
619 // we literally have to go through all loop's iterations.
620 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
621 SimplifiedValues.clear();
622 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SC);
625 BBWorklist.insert(L->getHeader());
626 // Note that we *must not* cache the size, this loop grows the worklist.
627 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
628 BasicBlock *BB = BBWorklist[Idx];
630 // Visit all instructions in the given basic block and try to simplify
631 // it. We don't change the actual IR, just count optimization
633 for (Instruction &I : *BB) {
634 unsigned InstCost = TTI.getUserCost(&I);
636 // Visit the instruction to analyze its loop cost after unrolling,
637 // and if the visitor returns false, include this instruction in the
639 if (!Analyzer.visit(I))
640 UnrolledCost += InstCost;
642 // Also track this instructions expected cost when executing the rolled
644 RolledDynamicCost += InstCost;
646 // If unrolled body turns out to be too big, bail out.
647 if (UnrolledCost > MaxUnrolledLoopSize)
651 // Add BB's successors to the worklist.
652 for (BasicBlock *Succ : successors(BB))
653 if (L->contains(Succ))
654 BBWorklist.insert(Succ);
657 // If we found no optimization opportunities on the first iteration, we
658 // won't find them on later ones too.
659 if (UnrolledCost == RolledDynamicCost)
662 return {{UnrolledCost, RolledDynamicCost}};
665 /// ApproximateLoopSize - Approximate the size of the loop.
666 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
667 bool &NotDuplicatable,
668 const TargetTransformInfo &TTI,
669 AssumptionCache *AC) {
670 SmallPtrSet<const Value *, 32> EphValues;
671 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
674 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
676 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
677 NumCalls = Metrics.NumInlineCandidates;
678 NotDuplicatable = Metrics.notDuplicatable;
680 unsigned LoopSize = Metrics.NumInsts;
682 // Don't allow an estimate of size zero. This would allows unrolling of loops
683 // with huge iteration counts, which is a compile time problem even if it's
684 // not a problem for code quality. Also, the code using this size may assume
685 // that each loop has at least three instructions (likely a conditional
686 // branch, a comparison feeding that branch, and some kind of loop increment
687 // feeding that comparison instruction).
688 LoopSize = std::max(LoopSize, 3u);
693 // Returns the loop hint metadata node with the given name (for example,
694 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
696 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
697 if (MDNode *LoopID = L->getLoopID())
698 return GetUnrollMetadata(LoopID, Name);
702 // Returns true if the loop has an unroll(full) pragma.
703 static bool HasUnrollFullPragma(const Loop *L) {
704 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
707 // Returns true if the loop has an unroll(disable) pragma.
708 static bool HasUnrollDisablePragma(const Loop *L) {
709 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
712 // Returns true if the loop has an runtime unroll(disable) pragma.
713 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
714 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
717 // If loop has an unroll_count pragma return the (necessarily
718 // positive) value from the pragma. Otherwise return 0.
719 static unsigned UnrollCountPragmaValue(const Loop *L) {
720 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
722 assert(MD->getNumOperands() == 2 &&
723 "Unroll count hint metadata should have two operands.");
725 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
726 assert(Count >= 1 && "Unroll count must be positive.");
732 // Remove existing unroll metadata and add unroll disable metadata to
733 // indicate the loop has already been unrolled. This prevents a loop
734 // from being unrolled more than is directed by a pragma if the loop
735 // unrolling pass is run more than once (which it generally is).
736 static void SetLoopAlreadyUnrolled(Loop *L) {
737 MDNode *LoopID = L->getLoopID();
740 // First remove any existing loop unrolling metadata.
741 SmallVector<Metadata *, 4> MDs;
742 // Reserve first location for self reference to the LoopID metadata node.
743 MDs.push_back(nullptr);
744 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
745 bool IsUnrollMetadata = false;
746 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
748 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
749 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
751 if (!IsUnrollMetadata)
752 MDs.push_back(LoopID->getOperand(i));
755 // Add unroll(disable) metadata to disable future unrolling.
756 LLVMContext &Context = L->getHeader()->getContext();
757 SmallVector<Metadata *, 1> DisableOperands;
758 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
759 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
760 MDs.push_back(DisableNode);
762 MDNode *NewLoopID = MDNode::get(Context, MDs);
763 // Set operand 0 to refer to the loop id itself.
764 NewLoopID->replaceOperandWith(0, NewLoopID);
765 L->setLoopID(NewLoopID);
768 bool LoopUnroll::canUnrollCompletely(Loop *L, unsigned Threshold,
769 unsigned PercentDynamicCostSavedThreshold,
770 unsigned DynamicCostSavingsDiscount,
771 unsigned UnrolledCost,
772 unsigned RolledDynamicCost) {
774 if (Threshold == NoThreshold) {
775 DEBUG(dbgs() << " Can fully unroll, because no threshold is set.\n");
779 if (UnrolledCost <= Threshold) {
780 DEBUG(dbgs() << " Can fully unroll, because unrolled cost: "
781 << UnrolledCost << "<" << Threshold << "\n");
785 assert(UnrolledCost && "UnrolledCost can't be 0 at this point.");
786 assert(RolledDynamicCost >= UnrolledCost &&
787 "Cannot have a higher unrolled cost than a rolled cost!");
789 // Compute the percentage of the dynamic cost in the rolled form that is
790 // saved when unrolled. If unrolling dramatically reduces the estimated
791 // dynamic cost of the loop, we use a higher threshold to allow more
793 unsigned PercentDynamicCostSaved =
794 (uint64_t)(RolledDynamicCost - UnrolledCost) * 100ull / RolledDynamicCost;
796 if (PercentDynamicCostSaved >= PercentDynamicCostSavedThreshold &&
797 (int64_t)UnrolledCost - (int64_t)DynamicCostSavingsDiscount <=
798 (int64_t)Threshold) {
799 DEBUG(dbgs() << " Can fully unroll, because unrolling will reduce the "
800 "expected dynamic cost by " << PercentDynamicCostSaved
801 << "% (threshold: " << PercentDynamicCostSavedThreshold
803 << " and the unrolled cost (" << UnrolledCost
804 << ") is less than the max threshold ("
805 << DynamicCostSavingsDiscount << ").\n");
809 DEBUG(dbgs() << " Too large to fully unroll:\n");
810 DEBUG(dbgs() << " Threshold: " << Threshold << "\n");
811 DEBUG(dbgs() << " Max threshold: " << DynamicCostSavingsDiscount << "\n");
812 DEBUG(dbgs() << " Percent cost saved threshold: "
813 << PercentDynamicCostSavedThreshold << "%\n");
814 DEBUG(dbgs() << " Unrolled cost: " << UnrolledCost << "\n");
815 DEBUG(dbgs() << " Rolled dynamic cost: " << RolledDynamicCost << "\n");
816 DEBUG(dbgs() << " Percent cost saved: " << PercentDynamicCostSaved
821 unsigned LoopUnroll::selectUnrollCount(
822 const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
823 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
824 bool &SetExplicitly) {
825 SetExplicitly = true;
827 // User-specified count (either as a command-line option or
828 // constructor parameter) has highest precedence.
829 unsigned Count = UserCount ? CurrentCount : 0;
831 // If there is no user-specified count, unroll pragmas have the next
832 // highest precendence.
836 } else if (PragmaFullUnroll) {
845 SetExplicitly = false;
847 // Runtime trip count.
848 Count = UnrollRuntimeCount;
850 // Conservative heuristic: if we know the trip count, see if we can
851 // completely unroll (subject to the threshold, checked below); otherwise
852 // try to find greatest modulo of the trip count which is still under
856 if (TripCount && Count > TripCount)
861 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
862 if (skipOptnoneFunction(L))
865 Function &F = *L->getHeader()->getParent();
867 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
868 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
869 const TargetTransformInfo &TTI =
870 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
871 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
873 BasicBlock *Header = L->getHeader();
874 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
875 << "] Loop %" << Header->getName() << "\n");
877 if (HasUnrollDisablePragma(L)) {
880 bool PragmaFullUnroll = HasUnrollFullPragma(L);
881 unsigned PragmaCount = UnrollCountPragmaValue(L);
882 bool HasPragma = PragmaFullUnroll || PragmaCount > 0;
884 TargetTransformInfo::UnrollingPreferences UP;
885 getUnrollingPreferences(L, TTI, UP);
887 // Find trip count and trip multiple if count is not available
888 unsigned TripCount = 0;
889 unsigned TripMultiple = 1;
890 // If there are multiple exiting blocks but one of them is the latch, use the
891 // latch for the trip count estimation. Otherwise insist on a single exiting
892 // block for the trip count estimation.
893 BasicBlock *ExitingBlock = L->getLoopLatch();
894 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
895 ExitingBlock = L->getExitingBlock();
897 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
898 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
901 // Select an initial unroll count. This may be reduced later based
902 // on size thresholds.
903 bool CountSetExplicitly;
904 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll,
905 PragmaCount, UP, CountSetExplicitly);
907 unsigned NumInlineCandidates;
908 bool notDuplicatable;
910 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
911 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
913 // When computing the unrolled size, note that the conditional branch on the
914 // backedge and the comparison feeding it are not replicated like the rest of
915 // the loop body (which is why 2 is subtracted).
916 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
917 if (notDuplicatable) {
918 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
919 << " instructions.\n");
922 if (NumInlineCandidates != 0) {
923 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
927 unsigned Threshold, PartialThreshold;
928 unsigned PercentDynamicCostSavedThreshold;
929 unsigned DynamicCostSavingsDiscount;
930 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold,
931 PercentDynamicCostSavedThreshold,
932 DynamicCostSavingsDiscount);
934 // Given Count, TripCount and thresholds determine the type of
935 // unrolling which is to be performed.
936 enum { Full = 0, Partial = 1, Runtime = 2 };
938 if (TripCount && Count == TripCount) {
940 // If the loop is really small, we don't need to run an expensive analysis.
941 if (canUnrollCompletely(L, Threshold, 100, DynamicCostSavingsDiscount,
942 UnrolledSize, UnrolledSize)) {
945 // The loop isn't that small, but we still can fully unroll it if that
946 // helps to remove a significant number of instructions.
947 // To check that, run additional analysis on the loop.
948 if (Optional<EstimatedUnrollCost> Cost = analyzeLoopUnrollCost(
949 L, TripCount, *SE, TTI, Threshold + DynamicCostSavingsDiscount))
950 if (canUnrollCompletely(L, Threshold, PercentDynamicCostSavedThreshold,
951 DynamicCostSavingsDiscount, Cost->UnrolledCost,
952 Cost->RolledDynamicCost)) {
956 } else if (TripCount && Count < TripCount) {
962 // Reduce count based on the type of unrolling and the threshold values.
963 unsigned OriginalCount = Count;
964 bool AllowRuntime = UserRuntime ? CurrentRuntime : UP.Runtime;
965 if (HasRuntimeUnrollDisablePragma(L)) {
966 AllowRuntime = false;
968 if (Unrolling == Partial) {
969 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
970 if (!AllowPartial && !CountSetExplicitly) {
971 DEBUG(dbgs() << " will not try to unroll partially because "
972 << "-unroll-allow-partial not given\n");
975 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
976 // Reduce unroll count to be modulo of TripCount for partial unrolling.
977 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
978 while (Count != 0 && TripCount % Count != 0)
981 } else if (Unrolling == Runtime) {
982 if (!AllowRuntime && !CountSetExplicitly) {
983 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
984 << "-unroll-runtime not given\n");
987 // Reduce unroll count to be the largest power-of-two factor of
988 // the original count which satisfies the threshold limit.
989 while (Count != 0 && UnrolledSize > PartialThreshold) {
991 UnrolledSize = (LoopSize-2) * Count + 2;
993 if (Count > UP.MaxCount)
995 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
999 if (PragmaCount != 0)
1000 // If loop has an unroll count pragma mark loop as unrolled to prevent
1001 // unrolling beyond that requested by the pragma.
1002 SetLoopAlreadyUnrolled(L);
1004 // Emit optimization remarks if we are unable to unroll the loop
1005 // as directed by a pragma.
1006 DebugLoc LoopLoc = L->getStartLoc();
1007 Function *F = Header->getParent();
1008 LLVMContext &Ctx = F->getContext();
1009 if (PragmaFullUnroll && PragmaCount == 0) {
1010 if (TripCount && Count != TripCount) {
1011 emitOptimizationRemarkMissed(
1012 Ctx, DEBUG_TYPE, *F, LoopLoc,
1013 "Unable to fully unroll loop as directed by unroll(full) pragma "
1014 "because unrolled size is too large.");
1015 } else if (!TripCount) {
1016 emitOptimizationRemarkMissed(
1017 Ctx, DEBUG_TYPE, *F, LoopLoc,
1018 "Unable to fully unroll loop as directed by unroll(full) pragma "
1019 "because loop has a runtime trip count.");
1021 } else if (PragmaCount > 0 && Count != OriginalCount) {
1022 emitOptimizationRemarkMissed(
1023 Ctx, DEBUG_TYPE, *F, LoopLoc,
1024 "Unable to unroll loop the number of times directed by "
1025 "unroll_count pragma because unrolled size is too large.");
1029 if (Unrolling != Full && Count < 2) {
1030 // Partial unrolling by 1 is a nop. For full unrolling, a factor
1031 // of 1 makes sense because loop control can be eliminated.
1036 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, UP.AllowExpensiveTripCount,
1037 TripMultiple, LI, this, &LPM, &AC))