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/LoopPass.h"
20 #include "llvm/Analysis/ScalarEvolution.h"
21 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DiagnosticInfo.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/Metadata.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Transforms/Utils/UnrollLoop.h"
32 #include "llvm/IR/InstVisitor.h"
33 #include "llvm/Analysis/InstructionSimplify.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 TTI.getUnrollingPreferences(L, UP);
171 // Select and return an unroll count based on parameters from
172 // user, unroll preferences, unroll pragmas, or a heuristic.
173 // SetExplicitly is set to true if the unroll count is is set by
174 // the user or a pragma rather than selected heuristically.
176 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
177 unsigned PragmaCount,
178 const TargetTransformInfo::UnrollingPreferences &UP,
179 bool &SetExplicitly);
181 // Select threshold values used to limit unrolling based on a
182 // total unrolled size. Parameters Threshold and PartialThreshold
183 // are set to the maximum unrolled size for fully and partially
184 // unrolled loops respectively.
185 void selectThresholds(const Loop *L, bool HasPragma,
186 const TargetTransformInfo::UnrollingPreferences &UP,
187 unsigned &Threshold, unsigned &PartialThreshold,
188 unsigned NumberOfOptimizedInstructions) {
189 // Determine the current unrolling threshold. While this is
190 // normally set from UnrollThreshold, it is overridden to a
191 // smaller value if the current function is marked as
192 // optimize-for-size, and the unroll threshold was not user
194 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
196 // If we are allowed to completely unroll if we can remove M% of
197 // instructions, and we know that with complete unrolling we'll be able
198 // to kill N instructions, then we can afford to completely unroll loops
199 // with unrolled size up to N*100/M.
200 // Adjust the threshold according to that:
201 unsigned PercentOfOptimizedForCompleteUnroll =
202 UserPercentOfOptimized ? CurrentMinPercentOfOptimized
203 : UP.MinPercentOfOptimized;
204 unsigned AbsoluteThreshold = UserAbsoluteThreshold
205 ? CurrentAbsoluteThreshold
206 : UP.AbsoluteThreshold;
207 if (PercentOfOptimizedForCompleteUnroll)
208 Threshold = std::max<unsigned>(Threshold,
209 NumberOfOptimizedInstructions * 100 /
210 PercentOfOptimizedForCompleteUnroll);
211 // But don't allow unrolling loops bigger than absolute threshold.
212 Threshold = std::min<unsigned>(Threshold, AbsoluteThreshold);
214 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
215 if (!UserThreshold &&
216 L->getHeader()->getParent()->hasFnAttribute(
217 Attribute::OptimizeForSize)) {
218 Threshold = UP.OptSizeThreshold;
219 PartialThreshold = UP.PartialOptSizeThreshold;
222 // If the loop has an unrolling pragma, we want to be more
223 // aggressive with unrolling limits. Set thresholds to at
224 // least the PragmaTheshold value which is larger than the
226 if (Threshold != NoThreshold)
227 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold);
228 if (PartialThreshold != NoThreshold)
230 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold);
236 char LoopUnroll::ID = 0;
237 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
238 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
239 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
240 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
241 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
242 INITIALIZE_PASS_DEPENDENCY(LCSSA)
243 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
244 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
246 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial,
248 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime);
251 Pass *llvm::createSimpleLoopUnrollPass() {
252 return llvm::createLoopUnrollPass(-1, -1, 0, 0);
255 static bool isLoadFromConstantInitializer(Value *V) {
256 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
257 if (GV->isConstant() && GV->hasDefinitiveInitializer())
258 return GV->getInitializer();
262 struct FindConstantPointers {
263 bool LoadCanBeConstantFolded;
264 bool IndexIsConstant;
270 FindConstantPointers(const Loop *loop, ScalarEvolution &SE)
271 : LoadCanBeConstantFolded(true), IndexIsConstant(true), L(loop), SE(SE) {}
273 bool follow(const SCEV *S) {
274 if (const SCEVUnknown *SC = dyn_cast<SCEVUnknown>(S)) {
275 // We've reached the leaf node of SCEV, it's most probably just a
276 // variable. Now it's time to see if it corresponds to a global constant
277 // global (in which case we can eliminate the load), or not.
278 BaseAddress = SC->getValue();
279 LoadCanBeConstantFolded =
280 IndexIsConstant && isLoadFromConstantInitializer(BaseAddress);
283 if (isa<SCEVConstant>(S))
285 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
286 // If the current SCEV expression is AddRec, and its loop isn't the loop
287 // we are about to unroll, then we won't get a constant address after
288 // unrolling, and thus, won't be able to eliminate the load.
289 if (AR->getLoop() != L)
290 return IndexIsConstant = false;
291 // If the step isn't constant, we won't get constant addresses in unrolled
292 // version. Bail out.
293 if (const SCEVConstant *StepSE =
294 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
295 Step = StepSE->getValue()->getValue();
297 return IndexIsConstant = false;
299 return IndexIsConstant;
301 // If Result is true, continue traversal.
302 // Otherwise, we have found something that prevents us from (possible) load
304 return IndexIsConstant;
306 bool isDone() const { return !IndexIsConstant; }
309 // This class is used to get an estimate of the optimization effects that we
310 // could get from complete loop unrolling. It comes from the fact that some
311 // loads might be replaced with concrete constant values and that could trigger
312 // a chain of instruction simplifications.
314 // E.g. we might have:
315 // int a[] = {0, 1, 0};
317 // for (i = 0; i < 3; i ++)
319 // If we completely unroll the loop, we would get:
320 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
321 // Which then will be simplified to:
322 // v = b[0]* 0 + b[1]* 1 + b[2]* 0
325 class UnrollAnalyzer : public InstVisitor<UnrollAnalyzer, bool> {
326 typedef InstVisitor<UnrollAnalyzer, bool> Base;
327 friend class InstVisitor<UnrollAnalyzer, bool>;
332 const TargetTransformInfo &TTI;
334 DenseMap<Value *, Constant *> SimplifiedValues;
335 DenseMap<LoadInst *, Value *> LoadBaseAddresses;
336 SmallPtrSet<Instruction *, 32> CountedInstructions;
338 /// \brief Count the number of optimized instructions.
339 unsigned NumberOfOptimizedInstructions;
341 // Provide base case for our instruction visit.
342 bool visitInstruction(Instruction &I) { return false; };
343 // TODO: We should also visit ICmp, FCmp, GetElementPtr, Trunc, ZExt, SExt,
344 // FPTrunc, FPExt, FPToUI, FPToSI, UIToFP, SIToFP, BitCast, Select,
345 // ExtractElement, InsertElement, ShuffleVector, ExtractValue, InsertValue.
347 // Probaly it's worth to hoist the code for estimating the simplifications
348 // effects to a separate class, since we have a very similar code in
349 // InlineCost already.
350 bool visitBinaryOperator(BinaryOperator &I) {
351 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
352 if (!isa<Constant>(LHS))
353 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
355 if (!isa<Constant>(RHS))
356 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
358 Value *SimpleV = nullptr;
359 if (auto FI = dyn_cast<FPMathOperator>(&I))
361 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags());
363 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS);
365 if (SimpleV && CountedInstructions.insert(&I).second)
366 NumberOfOptimizedInstructions += TTI.getUserCost(&I);
368 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
369 SimplifiedValues[&I] = C;
375 Constant *computeLoadValue(LoadInst *LI, unsigned Iteration) {
378 Value *BaseAddr = LoadBaseAddresses[LI];
382 auto GV = dyn_cast<GlobalVariable>(BaseAddr);
386 ConstantDataSequential *CDS =
387 dyn_cast<ConstantDataSequential>(GV->getInitializer());
391 const SCEV *BaseAddrSE = SE.getSCEV(BaseAddr);
392 const SCEV *S = SE.getSCEV(LI->getPointerOperand());
393 const SCEV *OffSE = SE.getMinusSCEV(S, BaseAddrSE);
396 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OffSE);
400 if (const SCEVConstant *StepSE =
401 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
402 StepC = StepSE->getValue()->getValue();
406 if (const SCEVConstant *StartSE = dyn_cast<SCEVConstant>(AR->getStart()))
407 StartC = StartSE->getValue()->getValue();
411 unsigned ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
412 unsigned Start = StartC.getLimitedValue();
413 unsigned Step = StepC.getLimitedValue();
415 unsigned Index = (Start + Step * Iteration) / ElemSize;
416 if (Index >= CDS->getNumElements())
419 Constant *CV = CDS->getElementAsConstant(Index);
425 UnrollAnalyzer(const Loop *L, unsigned TripCount, ScalarEvolution &SE,
426 const TargetTransformInfo &TTI)
427 : L(L), TripCount(TripCount), SE(SE), TTI(TTI),
428 NumberOfOptimizedInstructions(0) {}
430 // Visit all loads the loop L, and for those that, after complete loop
431 // unrolling, would have a constant address and it will point to a known
432 // constant initializer, record its base address for future use. It is used
433 // when we estimate number of potentially simplified instructions.
434 void findConstFoldableLoads() {
435 for (auto BB : L->getBlocks()) {
436 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
437 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
440 Value *AddrOp = LI->getPointerOperand();
441 const SCEV *S = SE.getSCEV(AddrOp);
442 FindConstantPointers Visitor(L, SE);
443 SCEVTraversal<FindConstantPointers> T(Visitor);
445 if (Visitor.IndexIsConstant && Visitor.LoadCanBeConstantFolded) {
446 LoadBaseAddresses[LI] = Visitor.BaseAddress;
453 // Given a list of loads that could be constant-folded (LoadBaseAddresses),
454 // estimate number of optimized instructions after substituting the concrete
455 // values for the given Iteration. Also track how many instructions become
456 // dead through this process.
457 unsigned estimateNumberOfOptimizedInstructions(unsigned Iteration) {
458 // We keep a set vector for the worklist so that we don't wast space in the
459 // worklist queuing up the same instruction repeatedly. This can happen due
460 // to multiple operands being the same instruction or due to the same
461 // instruction being an operand of lots of things that end up dead or
463 SmallSetVector<Instruction *, 8> Worklist;
465 // Clear the simplified values and counts for this iteration.
466 SimplifiedValues.clear();
467 CountedInstructions.clear();
468 NumberOfOptimizedInstructions = 0;
470 // We start by adding all loads to the worklist.
471 for (auto &LoadDescr : LoadBaseAddresses) {
472 LoadInst *LI = LoadDescr.first;
473 SimplifiedValues[LI] = computeLoadValue(LI, Iteration);
474 if (CountedInstructions.insert(LI).second)
475 NumberOfOptimizedInstructions += TTI.getUserCost(LI);
477 for (User *U : LI->users())
478 Worklist.insert(cast<Instruction>(U));
481 // And then we try to simplify every user of every instruction from the
482 // worklist. If we do simplify a user, add it to the worklist to process
483 // its users as well.
484 while (!Worklist.empty()) {
485 Instruction *I = Worklist.pop_back_val();
490 for (User *U : I->users())
491 Worklist.insert(cast<Instruction>(U));
494 // Now that we know the potentially simplifed instructions, estimate number
495 // of instructions that would become dead if we do perform the
498 // The dead instructions are held in a separate set. This is used to
499 // prevent us from re-examining instructions and make sure we only count
500 // the benifit once. The worklist's internal set handles insertion
502 SmallPtrSet<Instruction *, 16> DeadInstructions;
504 // Lambda to enque operands onto the worklist.
505 auto EnqueueOperands = [&](Instruction &I) {
506 for (auto *Op : I.operand_values())
507 if (auto *OpI = dyn_cast<Instruction>(Op))
508 if (!OpI->use_empty())
509 Worklist.insert(OpI);
512 // Start by initializing worklist with simplified instructions.
513 for (auto &FoldedKeyValue : SimplifiedValues)
514 if (auto *FoldedInst = dyn_cast<Instruction>(FoldedKeyValue.first)) {
515 DeadInstructions.insert(FoldedInst);
517 // Add each instruction operand of this dead instruction to the
519 EnqueueOperands(*FoldedInst);
522 // If a definition of an insn is only used by simplified or dead
523 // instructions, it's also dead. Check defs of all instructions from the
525 while (!Worklist.empty()) {
526 Instruction *I = Worklist.pop_back_val();
529 if (DeadInstructions.count(I))
532 if (std::all_of(I->user_begin(), I->user_end(), [&](User *U) {
533 return DeadInstructions.count(cast<Instruction>(U));
535 NumberOfOptimizedInstructions += TTI.getUserCost(I);
536 DeadInstructions.insert(I);
540 return NumberOfOptimizedInstructions;
544 // Complete loop unrolling can make some loads constant, and we need to know if
545 // that would expose any further optimization opportunities.
546 // This routine estimates this optimization effect and returns the number of
547 // instructions, that potentially might be optimized away.
549 approximateNumberOfOptimizedInstructions(const Loop *L, ScalarEvolution &SE,
551 const TargetTransformInfo &TTI) {
552 if (!TripCount || !UnrollMaxIterationsCountToAnalyze)
555 UnrollAnalyzer UA(L, TripCount, SE, TTI);
556 UA.findConstFoldableLoads();
558 // Estimate number of instructions, that could be simplified if we replace a
559 // load with the corresponding constant. Since the same load will take
560 // different values on different iterations, we have to go through all loop's
561 // iterations here. To limit ourselves here, we check only first N
562 // iterations, and then scale the found number, if necessary.
563 unsigned IterationsNumberForEstimate =
564 std::min<unsigned>(UnrollMaxIterationsCountToAnalyze, TripCount);
565 unsigned NumberOfOptimizedInstructions = 0;
566 for (unsigned i = 0; i < IterationsNumberForEstimate; ++i)
567 NumberOfOptimizedInstructions +=
568 UA.estimateNumberOfOptimizedInstructions(i);
570 NumberOfOptimizedInstructions *= TripCount / IterationsNumberForEstimate;
572 return NumberOfOptimizedInstructions;
575 /// ApproximateLoopSize - Approximate the size of the loop.
576 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
577 bool &NotDuplicatable,
578 const TargetTransformInfo &TTI,
579 AssumptionCache *AC) {
580 SmallPtrSet<const Value *, 32> EphValues;
581 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
584 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
586 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
587 NumCalls = Metrics.NumInlineCandidates;
588 NotDuplicatable = Metrics.notDuplicatable;
590 unsigned LoopSize = Metrics.NumInsts;
592 // Don't allow an estimate of size zero. This would allows unrolling of loops
593 // with huge iteration counts, which is a compile time problem even if it's
594 // not a problem for code quality. Also, the code using this size may assume
595 // that each loop has at least three instructions (likely a conditional
596 // branch, a comparison feeding that branch, and some kind of loop increment
597 // feeding that comparison instruction).
598 LoopSize = std::max(LoopSize, 3u);
603 // Returns the loop hint metadata node with the given name (for example,
604 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
606 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
607 if (MDNode *LoopID = L->getLoopID())
608 return GetUnrollMetadata(LoopID, Name);
612 // Returns true if the loop has an unroll(full) pragma.
613 static bool HasUnrollFullPragma(const Loop *L) {
614 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
617 // Returns true if the loop has an unroll(disable) pragma.
618 static bool HasUnrollDisablePragma(const Loop *L) {
619 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
622 // If loop has an unroll_count pragma return the (necessarily
623 // positive) value from the pragma. Otherwise return 0.
624 static unsigned UnrollCountPragmaValue(const Loop *L) {
625 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
627 assert(MD->getNumOperands() == 2 &&
628 "Unroll count hint metadata should have two operands.");
630 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
631 assert(Count >= 1 && "Unroll count must be positive.");
637 // Remove existing unroll metadata and add unroll disable metadata to
638 // indicate the loop has already been unrolled. This prevents a loop
639 // from being unrolled more than is directed by a pragma if the loop
640 // unrolling pass is run more than once (which it generally is).
641 static void SetLoopAlreadyUnrolled(Loop *L) {
642 MDNode *LoopID = L->getLoopID();
645 // First remove any existing loop unrolling metadata.
646 SmallVector<Metadata *, 4> MDs;
647 // Reserve first location for self reference to the LoopID metadata node.
648 MDs.push_back(nullptr);
649 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
650 bool IsUnrollMetadata = false;
651 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
653 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
654 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
656 if (!IsUnrollMetadata)
657 MDs.push_back(LoopID->getOperand(i));
660 // Add unroll(disable) metadata to disable future unrolling.
661 LLVMContext &Context = L->getHeader()->getContext();
662 SmallVector<Metadata *, 1> DisableOperands;
663 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
664 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
665 MDs.push_back(DisableNode);
667 MDNode *NewLoopID = MDNode::get(Context, MDs);
668 // Set operand 0 to refer to the loop id itself.
669 NewLoopID->replaceOperandWith(0, NewLoopID);
670 L->setLoopID(NewLoopID);
673 unsigned LoopUnroll::selectUnrollCount(
674 const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
675 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
676 bool &SetExplicitly) {
677 SetExplicitly = true;
679 // User-specified count (either as a command-line option or
680 // constructor parameter) has highest precedence.
681 unsigned Count = UserCount ? CurrentCount : 0;
683 // If there is no user-specified count, unroll pragmas have the next
684 // highest precendence.
688 } else if (PragmaFullUnroll) {
697 SetExplicitly = false;
699 // Runtime trip count.
700 Count = UnrollRuntimeCount;
702 // Conservative heuristic: if we know the trip count, see if we can
703 // completely unroll (subject to the threshold, checked below); otherwise
704 // try to find greatest modulo of the trip count which is still under
708 if (TripCount && Count > TripCount)
713 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
714 if (skipOptnoneFunction(L))
717 Function &F = *L->getHeader()->getParent();
719 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
720 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
721 const TargetTransformInfo &TTI =
722 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
723 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
725 BasicBlock *Header = L->getHeader();
726 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
727 << "] Loop %" << Header->getName() << "\n");
729 if (HasUnrollDisablePragma(L)) {
732 bool PragmaFullUnroll = HasUnrollFullPragma(L);
733 unsigned PragmaCount = UnrollCountPragmaValue(L);
734 bool HasPragma = PragmaFullUnroll || PragmaCount > 0;
736 TargetTransformInfo::UnrollingPreferences UP;
737 getUnrollingPreferences(L, TTI, UP);
739 // Find trip count and trip multiple if count is not available
740 unsigned TripCount = 0;
741 unsigned TripMultiple = 1;
742 // If there are multiple exiting blocks but one of them is the latch, use the
743 // latch for the trip count estimation. Otherwise insist on a single exiting
744 // block for the trip count estimation.
745 BasicBlock *ExitingBlock = L->getLoopLatch();
746 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
747 ExitingBlock = L->getExitingBlock();
749 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
750 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
753 // Select an initial unroll count. This may be reduced later based
754 // on size thresholds.
755 bool CountSetExplicitly;
756 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll,
757 PragmaCount, UP, CountSetExplicitly);
759 unsigned NumInlineCandidates;
760 bool notDuplicatable;
762 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
763 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
765 // When computing the unrolled size, note that the conditional branch on the
766 // backedge and the comparison feeding it are not replicated like the rest of
767 // the loop body (which is why 2 is subtracted).
768 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
769 if (notDuplicatable) {
770 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
771 << " instructions.\n");
774 if (NumInlineCandidates != 0) {
775 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
779 unsigned NumberOfOptimizedInstructions =
780 approximateNumberOfOptimizedInstructions(L, *SE, TripCount, TTI);
781 DEBUG(dbgs() << " Complete unrolling could save: "
782 << NumberOfOptimizedInstructions << "\n");
784 unsigned Threshold, PartialThreshold;
785 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold,
786 NumberOfOptimizedInstructions);
788 // Given Count, TripCount and thresholds determine the type of
789 // unrolling which is to be performed.
790 enum { Full = 0, Partial = 1, Runtime = 2 };
792 if (TripCount && Count == TripCount) {
793 if (Threshold != NoThreshold && UnrolledSize > Threshold) {
794 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
795 << " because size: " << UnrolledSize << ">" << Threshold
801 } else if (TripCount && Count < TripCount) {
807 // Reduce count based on the type of unrolling and the threshold values.
808 unsigned OriginalCount = Count;
809 bool AllowRuntime = UserRuntime ? CurrentRuntime : UP.Runtime;
810 if (Unrolling == Partial) {
811 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
812 if (!AllowPartial && !CountSetExplicitly) {
813 DEBUG(dbgs() << " will not try to unroll partially because "
814 << "-unroll-allow-partial not given\n");
817 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
818 // Reduce unroll count to be modulo of TripCount for partial unrolling.
819 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
820 while (Count != 0 && TripCount % Count != 0)
823 } else if (Unrolling == Runtime) {
824 if (!AllowRuntime && !CountSetExplicitly) {
825 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
826 << "-unroll-runtime not given\n");
829 // Reduce unroll count to be the largest power-of-two factor of
830 // the original count which satisfies the threshold limit.
831 while (Count != 0 && UnrolledSize > PartialThreshold) {
833 UnrolledSize = (LoopSize-2) * Count + 2;
835 if (Count > UP.MaxCount)
837 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
841 if (PragmaCount != 0)
842 // If loop has an unroll count pragma mark loop as unrolled to prevent
843 // unrolling beyond that requested by the pragma.
844 SetLoopAlreadyUnrolled(L);
846 // Emit optimization remarks if we are unable to unroll the loop
847 // as directed by a pragma.
848 DebugLoc LoopLoc = L->getStartLoc();
849 Function *F = Header->getParent();
850 LLVMContext &Ctx = F->getContext();
851 if (PragmaFullUnroll && PragmaCount == 0) {
852 if (TripCount && Count != TripCount) {
853 emitOptimizationRemarkMissed(
854 Ctx, DEBUG_TYPE, *F, LoopLoc,
855 "Unable to fully unroll loop as directed by unroll(full) pragma "
856 "because unrolled size is too large.");
857 } else if (!TripCount) {
858 emitOptimizationRemarkMissed(
859 Ctx, DEBUG_TYPE, *F, LoopLoc,
860 "Unable to fully unroll loop as directed by unroll(full) pragma "
861 "because loop has a runtime trip count.");
863 } else if (PragmaCount > 0 && Count != OriginalCount) {
864 emitOptimizationRemarkMissed(
865 Ctx, DEBUG_TYPE, *F, LoopLoc,
866 "Unable to unroll loop the number of times directed by "
867 "unroll_count pragma because unrolled size is too large.");
871 if (Unrolling != Full && Count < 2) {
872 // Partial unrolling by 1 is a nop. For full unrolling, a factor
873 // of 1 makes sense because loop control can be eliminated.
878 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, TripMultiple, LI, this,