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(1000), 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()->getAttributes().
217 hasAttribute(AttributeSet::FunctionIndex,
218 Attribute::OptimizeForSize)) {
219 Threshold = UP.OptSizeThreshold;
220 PartialThreshold = UP.PartialOptSizeThreshold;
223 // If the loop has an unrolling pragma, we want to be more
224 // aggressive with unrolling limits. Set thresholds to at
225 // least the PragmaTheshold value which is larger than the
227 if (Threshold != NoThreshold)
228 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold);
229 if (PartialThreshold != NoThreshold)
231 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold);
237 char LoopUnroll::ID = 0;
238 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
239 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
240 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
241 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
242 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
243 INITIALIZE_PASS_DEPENDENCY(LCSSA)
244 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
245 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
247 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial,
249 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime);
252 Pass *llvm::createSimpleLoopUnrollPass() {
253 return llvm::createLoopUnrollPass(-1, -1, 0, 0);
256 static bool isLoadFromConstantInitializer(Value *V) {
257 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
258 if (GV->isConstant() && GV->hasDefinitiveInitializer())
259 return GV->getInitializer();
263 struct FindConstantPointers {
264 bool LoadCanBeConstantFolded;
265 bool IndexIsConstant;
271 FindConstantPointers(const Loop *loop, ScalarEvolution &SE)
272 : LoadCanBeConstantFolded(true), IndexIsConstant(true), L(loop), SE(SE) {}
274 bool follow(const SCEV *S) {
275 if (const SCEVUnknown *SC = dyn_cast<SCEVUnknown>(S)) {
276 // We've reached the leaf node of SCEV, it's most probably just a
277 // variable. Now it's time to see if it corresponds to a global constant
278 // global (in which case we can eliminate the load), or not.
279 BaseAddress = SC->getValue();
280 LoadCanBeConstantFolded =
281 IndexIsConstant && isLoadFromConstantInitializer(BaseAddress);
284 if (isa<SCEVConstant>(S))
286 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
287 // If the current SCEV expression is AddRec, and its loop isn't the loop
288 // we are about to unroll, then we won't get a constant address after
289 // unrolling, and thus, won't be able to eliminate the load.
290 if (AR->getLoop() != L)
291 return IndexIsConstant = false;
292 // If the step isn't constant, we won't get constant addresses in unrolled
293 // version. Bail out.
294 if (const SCEVConstant *StepSE =
295 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
296 Step = StepSE->getValue()->getValue();
298 return IndexIsConstant = false;
300 return IndexIsConstant;
302 // If Result is true, continue traversal.
303 // Otherwise, we have found something that prevents us from (possible) load
305 return IndexIsConstant;
307 bool isDone() const { return !IndexIsConstant; }
310 // This class is used to get an estimate of the optimization effects that we
311 // could get from complete loop unrolling. It comes from the fact that some
312 // loads might be replaced with concrete constant values and that could trigger
313 // a chain of instruction simplifications.
315 // E.g. we might have:
316 // int a[] = {0, 1, 0};
318 // for (i = 0; i < 3; i ++)
320 // If we completely unroll the loop, we would get:
321 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
322 // Which then will be simplified to:
323 // v = b[0]* 0 + b[1]* 1 + b[2]* 0
326 class UnrollAnalyzer : public InstVisitor<UnrollAnalyzer, bool> {
327 typedef InstVisitor<UnrollAnalyzer, bool> Base;
328 friend class InstVisitor<UnrollAnalyzer, bool>;
333 const TargetTransformInfo &TTI;
335 DenseMap<Value *, Constant *> SimplifiedValues;
336 DenseMap<LoadInst *, Value *> LoadBaseAddresses;
337 SmallPtrSet<Instruction *, 32> CountedInstructions;
339 /// \brief Count the number of optimized instructions.
340 unsigned NumberOfOptimizedInstructions;
342 // Provide base case for our instruction visit.
343 bool visitInstruction(Instruction &I) { return false; };
344 // TODO: We should also visit ICmp, FCmp, GetElementPtr, Trunc, ZExt, SExt,
345 // FPTrunc, FPExt, FPToUI, FPToSI, UIToFP, SIToFP, BitCast, Select,
346 // ExtractElement, InsertElement, ShuffleVector, ExtractValue, InsertValue.
348 // Probaly it's worth to hoist the code for estimating the simplifications
349 // effects to a separate class, since we have a very similar code in
350 // InlineCost already.
351 bool visitBinaryOperator(BinaryOperator &I) {
352 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
353 if (!isa<Constant>(LHS))
354 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
356 if (!isa<Constant>(RHS))
357 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
359 Value *SimpleV = nullptr;
360 if (auto FI = dyn_cast<FPMathOperator>(&I))
362 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags());
364 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS);
366 if (SimpleV && CountedInstructions.insert(&I).second)
367 NumberOfOptimizedInstructions += TTI.getUserCost(&I);
369 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
370 SimplifiedValues[&I] = C;
376 Constant *computeLoadValue(LoadInst *LI, unsigned Iteration) {
379 Value *BaseAddr = LoadBaseAddresses[LI];
383 auto GV = dyn_cast<GlobalVariable>(BaseAddr);
387 ConstantDataSequential *CDS =
388 dyn_cast<ConstantDataSequential>(GV->getInitializer());
392 const SCEV *BaseAddrSE = SE.getSCEV(BaseAddr);
393 const SCEV *S = SE.getSCEV(LI->getPointerOperand());
394 const SCEV *OffSE = SE.getMinusSCEV(S, BaseAddrSE);
397 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OffSE);
401 if (const SCEVConstant *StepSE =
402 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
403 StepC = StepSE->getValue()->getValue();
407 if (const SCEVConstant *StartSE = dyn_cast<SCEVConstant>(AR->getStart()))
408 StartC = StartSE->getValue()->getValue();
412 unsigned ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
413 unsigned Start = StartC.getLimitedValue();
414 unsigned Step = StepC.getLimitedValue();
416 unsigned Index = (Start + Step * Iteration) / ElemSize;
417 if (Index >= CDS->getNumElements())
420 Constant *CV = CDS->getElementAsConstant(Index);
426 UnrollAnalyzer(const Loop *L, unsigned TripCount, ScalarEvolution &SE,
427 const TargetTransformInfo &TTI)
428 : L(L), TripCount(TripCount), SE(SE), TTI(TTI),
429 NumberOfOptimizedInstructions(0) {}
431 // Visit all loads the loop L, and for those that, after complete loop
432 // unrolling, would have a constant address and it will point to a known
433 // constant initializer, record its base address for future use. It is used
434 // when we estimate number of potentially simplified instructions.
435 void findConstFoldableLoads() {
436 for (auto BB : L->getBlocks()) {
437 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
438 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
441 Value *AddrOp = LI->getPointerOperand();
442 const SCEV *S = SE.getSCEV(AddrOp);
443 FindConstantPointers Visitor(L, SE);
444 SCEVTraversal<FindConstantPointers> T(Visitor);
446 if (Visitor.IndexIsConstant && Visitor.LoadCanBeConstantFolded) {
447 LoadBaseAddresses[LI] = Visitor.BaseAddress;
454 // Given a list of loads that could be constant-folded (LoadBaseAddresses),
455 // estimate number of optimized instructions after substituting the concrete
456 // values for the given Iteration.
457 // Fill in SimplifiedValues map for future use in DCE-estimation.
458 unsigned estimateNumberOfSimplifiedInstructions(unsigned Iteration) {
459 SmallSetVector<Instruction *, 8> Worklist;
460 SimplifiedValues.clear();
461 CountedInstructions.clear();
462 NumberOfOptimizedInstructions = 0;
464 // We start by adding all loads to the worklist.
465 for (auto &LoadDescr : LoadBaseAddresses) {
466 LoadInst *LI = LoadDescr.first;
467 SimplifiedValues[LI] = computeLoadValue(LI, Iteration);
468 if (CountedInstructions.insert(LI).second)
469 NumberOfOptimizedInstructions += TTI.getUserCost(LI);
471 for (User *U : LI->users()) {
472 Instruction *UI = dyn_cast<Instruction>(U);
479 // And then we try to simplify every user of every instruction from the
480 // worklist. If we do simplify a user, add it to the worklist to process
481 // its users as well.
482 while (!Worklist.empty()) {
483 Instruction *I = Worklist.pop_back_val();
488 for (User *U : I->users()) {
489 Instruction *UI = dyn_cast<Instruction>(U);
495 return NumberOfOptimizedInstructions;
498 // Given a list of potentially simplifed instructions, estimate number of
499 // instructions that would become dead if we do perform the simplification.
500 unsigned estimateNumberOfDeadInstructions() {
501 NumberOfOptimizedInstructions = 0;
503 // We keep a set vector for the worklist so that we don't wast space in the
504 // worklist queuing up the same instruction repeatedly. This can happen due
505 // to multiple operands being the same instruction or due to the same
506 // instruction being an operand of lots of things that end up dead or
508 SmallSetVector<Instruction *, 8> Worklist;
510 // The dead instructions are held in a separate set. This is used to
511 // prevent us from re-examining instructions and make sure we only count
512 // the benifit once. The worklist's internal set handles insertion
514 SmallPtrSet<Instruction *, 16> DeadInstructions;
516 // Lambda to enque operands onto the worklist.
517 auto EnqueueOperands = [&](Instruction &I) {
518 for (auto *Op : I.operand_values())
519 if (auto *OpI = dyn_cast<Instruction>(Op))
520 if (!OpI->use_empty())
521 Worklist.insert(OpI);
524 // Start by initializing worklist with simplified instructions.
525 for (auto &FoldedKeyValue : SimplifiedValues)
526 if (auto *FoldedInst = dyn_cast<Instruction>(FoldedKeyValue.first)) {
527 DeadInstructions.insert(FoldedInst);
529 // Add each instruction operand of this dead instruction to the
531 EnqueueOperands(*FoldedInst);
534 // If a definition of an insn is only used by simplified or dead
535 // instructions, it's also dead. Check defs of all instructions from the
537 while (!Worklist.empty()) {
538 Instruction *I = Worklist.pop_back_val();
541 if (DeadInstructions.count(I))
544 if (std::all_of(I->user_begin(), I->user_end(), [&](User *U) {
545 return DeadInstructions.count(cast<Instruction>(U));
547 NumberOfOptimizedInstructions += TTI.getUserCost(I);
548 DeadInstructions.insert(I);
552 return NumberOfOptimizedInstructions;
556 // Complete loop unrolling can make some loads constant, and we need to know if
557 // that would expose any further optimization opportunities.
558 // This routine estimates this optimization effect and returns the number of
559 // instructions, that potentially might be optimized away.
561 approximateNumberOfOptimizedInstructions(const Loop *L, ScalarEvolution &SE,
563 const TargetTransformInfo &TTI) {
564 if (!TripCount || !UnrollMaxIterationsCountToAnalyze)
567 UnrollAnalyzer UA(L, TripCount, SE, TTI);
568 UA.findConstFoldableLoads();
570 // Estimate number of instructions, that could be simplified if we replace a
571 // load with the corresponding constant. Since the same load will take
572 // different values on different iterations, we have to go through all loop's
573 // iterations here. To limit ourselves here, we check only first N
574 // iterations, and then scale the found number, if necessary.
575 unsigned IterationsNumberForEstimate =
576 std::min<unsigned>(UnrollMaxIterationsCountToAnalyze, TripCount);
577 unsigned NumberOfOptimizedInstructions = 0;
578 for (unsigned i = 0; i < IterationsNumberForEstimate; ++i) {
579 NumberOfOptimizedInstructions +=
580 UA.estimateNumberOfSimplifiedInstructions(i);
581 NumberOfOptimizedInstructions += UA.estimateNumberOfDeadInstructions();
583 NumberOfOptimizedInstructions *= TripCount / IterationsNumberForEstimate;
585 return NumberOfOptimizedInstructions;
588 /// ApproximateLoopSize - Approximate the size of the loop.
589 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
590 bool &NotDuplicatable,
591 const TargetTransformInfo &TTI,
592 AssumptionCache *AC) {
593 SmallPtrSet<const Value *, 32> EphValues;
594 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
597 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
599 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
600 NumCalls = Metrics.NumInlineCandidates;
601 NotDuplicatable = Metrics.notDuplicatable;
603 unsigned LoopSize = Metrics.NumInsts;
605 // Don't allow an estimate of size zero. This would allows unrolling of loops
606 // with huge iteration counts, which is a compile time problem even if it's
607 // not a problem for code quality. Also, the code using this size may assume
608 // that each loop has at least three instructions (likely a conditional
609 // branch, a comparison feeding that branch, and some kind of loop increment
610 // feeding that comparison instruction).
611 LoopSize = std::max(LoopSize, 3u);
616 // Returns the loop hint metadata node with the given name (for example,
617 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
619 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
620 if (MDNode *LoopID = L->getLoopID())
621 return GetUnrollMetadata(LoopID, Name);
625 // Returns true if the loop has an unroll(full) pragma.
626 static bool HasUnrollFullPragma(const Loop *L) {
627 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
630 // Returns true if the loop has an unroll(disable) pragma.
631 static bool HasUnrollDisablePragma(const Loop *L) {
632 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
635 // If loop has an unroll_count pragma return the (necessarily
636 // positive) value from the pragma. Otherwise return 0.
637 static unsigned UnrollCountPragmaValue(const Loop *L) {
638 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
640 assert(MD->getNumOperands() == 2 &&
641 "Unroll count hint metadata should have two operands.");
643 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
644 assert(Count >= 1 && "Unroll count must be positive.");
650 // Remove existing unroll metadata and add unroll disable metadata to
651 // indicate the loop has already been unrolled. This prevents a loop
652 // from being unrolled more than is directed by a pragma if the loop
653 // unrolling pass is run more than once (which it generally is).
654 static void SetLoopAlreadyUnrolled(Loop *L) {
655 MDNode *LoopID = L->getLoopID();
658 // First remove any existing loop unrolling metadata.
659 SmallVector<Metadata *, 4> MDs;
660 // Reserve first location for self reference to the LoopID metadata node.
661 MDs.push_back(nullptr);
662 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
663 bool IsUnrollMetadata = false;
664 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
666 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
667 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
669 if (!IsUnrollMetadata)
670 MDs.push_back(LoopID->getOperand(i));
673 // Add unroll(disable) metadata to disable future unrolling.
674 LLVMContext &Context = L->getHeader()->getContext();
675 SmallVector<Metadata *, 1> DisableOperands;
676 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
677 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
678 MDs.push_back(DisableNode);
680 MDNode *NewLoopID = MDNode::get(Context, MDs);
681 // Set operand 0 to refer to the loop id itself.
682 NewLoopID->replaceOperandWith(0, NewLoopID);
683 L->setLoopID(NewLoopID);
686 unsigned LoopUnroll::selectUnrollCount(
687 const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
688 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
689 bool &SetExplicitly) {
690 SetExplicitly = true;
692 // User-specified count (either as a command-line option or
693 // constructor parameter) has highest precedence.
694 unsigned Count = UserCount ? CurrentCount : 0;
696 // If there is no user-specified count, unroll pragmas have the next
697 // highest precendence.
701 } else if (PragmaFullUnroll) {
710 SetExplicitly = false;
712 // Runtime trip count.
713 Count = UnrollRuntimeCount;
715 // Conservative heuristic: if we know the trip count, see if we can
716 // completely unroll (subject to the threshold, checked below); otherwise
717 // try to find greatest modulo of the trip count which is still under
721 if (TripCount && Count > TripCount)
726 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
727 if (skipOptnoneFunction(L))
730 Function &F = *L->getHeader()->getParent();
732 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
733 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
734 const TargetTransformInfo &TTI =
735 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
736 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
738 BasicBlock *Header = L->getHeader();
739 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
740 << "] Loop %" << Header->getName() << "\n");
742 if (HasUnrollDisablePragma(L)) {
745 bool PragmaFullUnroll = HasUnrollFullPragma(L);
746 unsigned PragmaCount = UnrollCountPragmaValue(L);
747 bool HasPragma = PragmaFullUnroll || PragmaCount > 0;
749 TargetTransformInfo::UnrollingPreferences UP;
750 getUnrollingPreferences(L, TTI, UP);
752 // Find trip count and trip multiple if count is not available
753 unsigned TripCount = 0;
754 unsigned TripMultiple = 1;
755 // If there are multiple exiting blocks but one of them is the latch, use the
756 // latch for the trip count estimation. Otherwise insist on a single exiting
757 // block for the trip count estimation.
758 BasicBlock *ExitingBlock = L->getLoopLatch();
759 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
760 ExitingBlock = L->getExitingBlock();
762 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
763 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
766 // Select an initial unroll count. This may be reduced later based
767 // on size thresholds.
768 bool CountSetExplicitly;
769 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll,
770 PragmaCount, UP, CountSetExplicitly);
772 unsigned NumInlineCandidates;
773 bool notDuplicatable;
775 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
776 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
778 // When computing the unrolled size, note that the conditional branch on the
779 // backedge and the comparison feeding it are not replicated like the rest of
780 // the loop body (which is why 2 is subtracted).
781 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
782 if (notDuplicatable) {
783 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
784 << " instructions.\n");
787 if (NumInlineCandidates != 0) {
788 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
792 unsigned NumberOfOptimizedInstructions =
793 approximateNumberOfOptimizedInstructions(L, *SE, TripCount, TTI);
794 DEBUG(dbgs() << " Complete unrolling could save: "
795 << NumberOfOptimizedInstructions << "\n");
797 unsigned Threshold, PartialThreshold;
798 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold,
799 NumberOfOptimizedInstructions);
801 // Given Count, TripCount and thresholds determine the type of
802 // unrolling which is to be performed.
803 enum { Full = 0, Partial = 1, Runtime = 2 };
805 if (TripCount && Count == TripCount) {
806 if (Threshold != NoThreshold && UnrolledSize > Threshold) {
807 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
808 << " because size: " << UnrolledSize << ">" << Threshold
814 } else if (TripCount && Count < TripCount) {
820 // Reduce count based on the type of unrolling and the threshold values.
821 unsigned OriginalCount = Count;
822 bool AllowRuntime = UserRuntime ? CurrentRuntime : UP.Runtime;
823 if (Unrolling == Partial) {
824 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
825 if (!AllowPartial && !CountSetExplicitly) {
826 DEBUG(dbgs() << " will not try to unroll partially because "
827 << "-unroll-allow-partial not given\n");
830 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
831 // Reduce unroll count to be modulo of TripCount for partial unrolling.
832 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
833 while (Count != 0 && TripCount % Count != 0)
836 } else if (Unrolling == Runtime) {
837 if (!AllowRuntime && !CountSetExplicitly) {
838 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
839 << "-unroll-runtime not given\n");
842 // Reduce unroll count to be the largest power-of-two factor of
843 // the original count which satisfies the threshold limit.
844 while (Count != 0 && UnrolledSize > PartialThreshold) {
846 UnrolledSize = (LoopSize-2) * Count + 2;
848 if (Count > UP.MaxCount)
850 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
854 if (PragmaCount != 0)
855 // If loop has an unroll count pragma mark loop as unrolled to prevent
856 // unrolling beyond that requested by the pragma.
857 SetLoopAlreadyUnrolled(L);
859 // Emit optimization remarks if we are unable to unroll the loop
860 // as directed by a pragma.
861 DebugLoc LoopLoc = L->getStartLoc();
862 Function *F = Header->getParent();
863 LLVMContext &Ctx = F->getContext();
864 if (PragmaFullUnroll && PragmaCount == 0) {
865 if (TripCount && Count != TripCount) {
866 emitOptimizationRemarkMissed(
867 Ctx, DEBUG_TYPE, *F, LoopLoc,
868 "Unable to fully unroll loop as directed by unroll(full) pragma "
869 "because unrolled size is too large.");
870 } else if (!TripCount) {
871 emitOptimizationRemarkMissed(
872 Ctx, DEBUG_TYPE, *F, LoopLoc,
873 "Unable to fully unroll loop as directed by unroll(full) pragma "
874 "because loop has a runtime trip count.");
876 } else if (PragmaCount > 0 && Count != OriginalCount) {
877 emitOptimizationRemarkMissed(
878 Ctx, DEBUG_TYPE, *F, LoopLoc,
879 "Unable to unroll loop the number of times directed by "
880 "unroll_count pragma because unrolled size is too large.");
884 if (Unrolling != Full && Count < 2) {
885 // Partial unrolling by 1 is a nop. For full unrolling, a factor
886 // of 1 makes sense because loop control can be eliminated.
891 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, TripMultiple, LI, this,