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/Analysis/AssumptionCache.h"
17 #include "llvm/Analysis/CodeMetrics.h"
18 #include "llvm/Analysis/LoopPass.h"
19 #include "llvm/Analysis/ScalarEvolution.h"
20 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/DiagnosticInfo.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/IntrinsicInst.h"
26 #include "llvm/IR/Metadata.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/Transforms/Utils/UnrollLoop.h"
31 #include "llvm/IR/InstVisitor.h"
32 #include "llvm/Analysis/InstructionSimplify.h"
37 #define DEBUG_TYPE "loop-unroll"
39 static cl::opt<unsigned>
40 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
41 cl::desc("The cut-off point for automatic loop unrolling"));
43 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
44 "unroll-max-iteration-count-to-analyze", cl::init(1000), cl::Hidden,
45 cl::desc("Don't allow loop unrolling to simulate more than this number of"
46 "iterations when checking full unroll profitability"));
48 static cl::opt<unsigned> UnrollMinPercentOfOptimized(
49 "unroll-percent-of-optimized-for-complete-unroll", cl::init(20), cl::Hidden,
50 cl::desc("If complete unrolling could trigger further optimizations, and, "
51 "by that, remove the given percent of instructions, perform the "
52 "complete unroll even if it's beyond the threshold"));
54 static cl::opt<unsigned> UnrollAbsoluteThreshold(
55 "unroll-absolute-threshold", cl::init(2000), cl::Hidden,
56 cl::desc("Don't unroll if the unrolled size is bigger than this threshold,"
57 " even if we can remove big portion of instructions later."));
59 static cl::opt<unsigned>
60 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
61 cl::desc("Use this unroll count for all loops including those with "
62 "unroll_count pragma values, for testing purposes"));
65 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
66 cl::desc("Allows loops to be partially unrolled until "
67 "-unroll-threshold loop size is reached."));
70 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden,
71 cl::desc("Unroll loops with run-time trip counts"));
73 static cl::opt<unsigned>
74 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
75 cl::desc("Unrolled size limit for loops with an unroll(full) or "
76 "unroll_count pragma."));
79 class LoopUnroll : public LoopPass {
81 static char ID; // Pass ID, replacement for typeid
82 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) {
83 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
84 CurrentAbsoluteThreshold = UnrollAbsoluteThreshold;
85 CurrentMinPercentOfOptimized = UnrollMinPercentOfOptimized;
86 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
87 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
88 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
90 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
91 UserAbsoluteThreshold = (UnrollAbsoluteThreshold.getNumOccurrences() > 0);
92 UserPercentOfOptimized =
93 (UnrollMinPercentOfOptimized.getNumOccurrences() > 0);
94 UserAllowPartial = (P != -1) ||
95 (UnrollAllowPartial.getNumOccurrences() > 0);
96 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
97 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0);
99 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
102 /// A magic value for use with the Threshold parameter to indicate
103 /// that the loop unroll should be performed regardless of how much
104 /// code expansion would result.
105 static const unsigned NoThreshold = UINT_MAX;
107 // Threshold to use when optsize is specified (and there is no
108 // explicit -unroll-threshold).
109 static const unsigned OptSizeUnrollThreshold = 50;
111 // Default unroll count for loops with run-time trip count if
112 // -unroll-count is not set
113 static const unsigned UnrollRuntimeCount = 8;
115 unsigned CurrentCount;
116 unsigned CurrentThreshold;
117 unsigned CurrentAbsoluteThreshold;
118 unsigned CurrentMinPercentOfOptimized;
119 bool CurrentAllowPartial;
121 bool UserCount; // CurrentCount is user-specified.
122 bool UserThreshold; // CurrentThreshold is user-specified.
123 bool UserAbsoluteThreshold; // CurrentAbsoluteThreshold is
125 bool UserPercentOfOptimized; // CurrentMinPercentOfOptimized is
127 bool UserAllowPartial; // CurrentAllowPartial is user-specified.
128 bool UserRuntime; // CurrentRuntime is user-specified.
130 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
132 /// This transformation requires natural loop information & requires that
133 /// loop preheaders be inserted into the CFG...
135 void getAnalysisUsage(AnalysisUsage &AU) const override {
136 AU.addRequired<AssumptionCacheTracker>();
137 AU.addRequired<LoopInfoWrapperPass>();
138 AU.addPreserved<LoopInfoWrapperPass>();
139 AU.addRequiredID(LoopSimplifyID);
140 AU.addPreservedID(LoopSimplifyID);
141 AU.addRequiredID(LCSSAID);
142 AU.addPreservedID(LCSSAID);
143 AU.addRequired<ScalarEvolution>();
144 AU.addPreserved<ScalarEvolution>();
145 AU.addRequired<TargetTransformInfoWrapperPass>();
146 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
147 // If loop unroll does not preserve dom info then LCSSA pass on next
148 // loop will receive invalid dom info.
149 // For now, recreate dom info, if loop is unrolled.
150 AU.addPreserved<DominatorTreeWrapperPass>();
153 // Fill in the UnrollingPreferences parameter with values from the
154 // TargetTransformationInfo.
155 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI,
156 TargetTransformInfo::UnrollingPreferences &UP) {
157 UP.Threshold = CurrentThreshold;
158 UP.AbsoluteThreshold = CurrentAbsoluteThreshold;
159 UP.MinPercentOfOptimized = CurrentMinPercentOfOptimized;
160 UP.OptSizeThreshold = OptSizeUnrollThreshold;
161 UP.PartialThreshold = CurrentThreshold;
162 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
163 UP.Count = CurrentCount;
164 UP.MaxCount = UINT_MAX;
165 UP.Partial = CurrentAllowPartial;
166 UP.Runtime = CurrentRuntime;
167 TTI.getUnrollingPreferences(L, UP);
170 // Select and return an unroll count based on parameters from
171 // user, unroll preferences, unroll pragmas, or a heuristic.
172 // SetExplicitly is set to true if the unroll count is is set by
173 // the user or a pragma rather than selected heuristically.
175 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
176 unsigned PragmaCount,
177 const TargetTransformInfo::UnrollingPreferences &UP,
178 bool &SetExplicitly);
180 // Select threshold values used to limit unrolling based on a
181 // total unrolled size. Parameters Threshold and PartialThreshold
182 // are set to the maximum unrolled size for fully and partially
183 // unrolled loops respectively.
184 void selectThresholds(const Loop *L, bool HasPragma,
185 const TargetTransformInfo::UnrollingPreferences &UP,
186 unsigned &Threshold, unsigned &PartialThreshold,
187 unsigned NumberOfOptimizedInstructions) {
188 // Determine the current unrolling threshold. While this is
189 // normally set from UnrollThreshold, it is overridden to a
190 // smaller value if the current function is marked as
191 // optimize-for-size, and the unroll threshold was not user
193 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
195 // If we are allowed to completely unroll if we can remove M% of
196 // instructions, and we know that with complete unrolling we'll be able
197 // to kill N instructions, then we can afford to completely unroll loops
198 // with unrolled size up to N*100/M.
199 // Adjust the threshold according to that:
200 unsigned PercentOfOptimizedForCompleteUnroll =
201 UserPercentOfOptimized ? CurrentMinPercentOfOptimized
202 : UP.MinPercentOfOptimized;
203 unsigned AbsoluteThreshold = UserAbsoluteThreshold
204 ? CurrentAbsoluteThreshold
205 : UP.AbsoluteThreshold;
206 if (PercentOfOptimizedForCompleteUnroll)
207 Threshold = std::max<unsigned>(Threshold,
208 NumberOfOptimizedInstructions * 100 /
209 PercentOfOptimizedForCompleteUnroll);
210 // But don't allow unrolling loops bigger than absolute threshold.
211 Threshold = std::min<unsigned>(Threshold, AbsoluteThreshold);
213 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
214 if (!UserThreshold &&
215 L->getHeader()->getParent()->getAttributes().
216 hasAttribute(AttributeSet::FunctionIndex,
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;
333 unsigned NumberOfOptimizedInstructions;
335 DenseMap<Value *, Constant *> SimplifiedValues;
336 DenseMap<LoadInst *, Value *> LoadBaseAddresses;
337 SmallPtrSet<Instruction *, 32> CountedInsns;
339 // Provide base case for our instruction visit.
340 bool visitInstruction(Instruction &I) { return false; };
341 // TODO: We should also visit ICmp, FCmp, GetElementPtr, Trunc, ZExt, SExt,
342 // FPTrunc, FPExt, FPToUI, FPToSI, UIToFP, SIToFP, BitCast, Select,
343 // ExtractElement, InsertElement, ShuffleVector, ExtractValue, InsertValue.
345 // Probaly it's worth to hoist the code for estimating the simplifications
346 // effects to a separate class, since we have a very similar code in
347 // InlineCost already.
348 bool visitBinaryOperator(BinaryOperator &I) {
349 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
350 if (!isa<Constant>(LHS))
351 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
353 if (!isa<Constant>(RHS))
354 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
356 Value *SimpleV = nullptr;
357 if (auto FI = dyn_cast<FPMathOperator>(&I))
359 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags());
361 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS);
363 if (SimpleV && CountedInsns.insert(&I).second)
364 NumberOfOptimizedInstructions += TTI.getUserCost(&I);
366 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
367 SimplifiedValues[&I] = C;
373 Constant *computeLoadValue(LoadInst *LI, unsigned Iteration) {
376 Value *BaseAddr = LoadBaseAddresses[LI];
380 auto GV = dyn_cast<GlobalVariable>(BaseAddr);
384 ConstantDataSequential *CDS =
385 dyn_cast<ConstantDataSequential>(GV->getInitializer());
389 const SCEV *BaseAddrSE = SE.getSCEV(BaseAddr);
390 const SCEV *S = SE.getSCEV(LI->getPointerOperand());
391 const SCEV *OffSE = SE.getMinusSCEV(S, BaseAddrSE);
394 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OffSE);
398 if (const SCEVConstant *StepSE =
399 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
400 StepC = StepSE->getValue()->getValue();
404 if (const SCEVConstant *StartSE = dyn_cast<SCEVConstant>(AR->getStart()))
405 StartC = StartSE->getValue()->getValue();
409 unsigned ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
410 unsigned Start = StartC.getLimitedValue();
411 unsigned Step = StepC.getLimitedValue();
413 unsigned Index = (Start + Step * Iteration) / ElemSize;
414 if (Index >= CDS->getNumElements())
417 Constant *CV = CDS->getElementAsConstant(Index);
423 UnrollAnalyzer(const Loop *L, unsigned TripCount, ScalarEvolution &SE,
424 const TargetTransformInfo &TTI)
425 : L(L), TripCount(TripCount), SE(SE), TTI(TTI),
426 NumberOfOptimizedInstructions(0) {}
428 // Visit all loads the loop L, and for those that, after complete loop
429 // unrolling, would have a constant address and it will point to a known
430 // constant initializer, record its base address for future use. It is used
431 // when we estimate number of potentially simplified instructions.
432 void findConstFoldableLoads() {
433 for (auto BB : L->getBlocks()) {
434 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
435 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
438 Value *AddrOp = LI->getPointerOperand();
439 const SCEV *S = SE.getSCEV(AddrOp);
440 FindConstantPointers Visitor(L, SE);
441 SCEVTraversal<FindConstantPointers> T(Visitor);
443 if (Visitor.IndexIsConstant && Visitor.LoadCanBeConstantFolded) {
444 LoadBaseAddresses[LI] = Visitor.BaseAddress;
451 // Given a list of loads that could be constant-folded (LoadBaseAddresses),
452 // estimate number of optimized instructions after substituting the concrete
453 // values for the given Iteration.
454 // Fill in SimplifiedInsns map for future use in DCE-estimation.
455 unsigned estimateNumberOfSimplifiedInsns(unsigned Iteration) {
456 SmallVector<Instruction *, 8> Worklist;
457 SimplifiedValues.clear();
458 CountedInsns.clear();
460 NumberOfOptimizedInstructions = 0;
461 // We start by adding all loads to the worklist.
462 for (auto LoadDescr : LoadBaseAddresses) {
463 LoadInst *LI = LoadDescr.first;
464 SimplifiedValues[LI] = computeLoadValue(LI, Iteration);
465 if (CountedInsns.insert(LI).second)
466 NumberOfOptimizedInstructions += TTI.getUserCost(LI);
468 for (auto U : LI->users()) {
469 Instruction *UI = dyn_cast<Instruction>(U);
472 if (!L->contains(UI))
474 Worklist.push_back(UI);
478 // And then we try to simplify every user of every instruction from the
479 // worklist. If we do simplify a user, add it to the worklist to process
480 // its users as well.
481 while (!Worklist.empty()) {
482 Instruction *I = Worklist.pop_back_val();
485 for (auto U : I->users()) {
486 Instruction *UI = dyn_cast<Instruction>(U);
489 if (!L->contains(UI))
491 Worklist.push_back(UI);
494 return NumberOfOptimizedInstructions;
497 // Given a list of potentially simplifed instructions, estimate number of
498 // instructions that would become dead if we do perform the simplification.
499 unsigned estimateNumberOfDeadInsns() {
500 NumberOfOptimizedInstructions = 0;
501 SmallVector<Instruction *, 8> Worklist;
502 DenseMap<Instruction *, bool> DeadInstructions;
503 // Start by initializing worklist with simplified instructions.
504 for (auto Folded : SimplifiedValues) {
505 if (auto FoldedInsn = dyn_cast<Instruction>(Folded.first)) {
506 Worklist.push_back(FoldedInsn);
507 DeadInstructions[FoldedInsn] = true;
510 // If a definition of an insn is only used by simplified or dead
511 // instructions, it's also dead. Check defs of all instructions from the
513 while (!Worklist.empty()) {
514 Instruction *FoldedInsn = Worklist.pop_back_val();
515 for (Value *Op : FoldedInsn->operands()) {
516 if (auto I = dyn_cast<Instruction>(Op)) {
519 if (SimplifiedValues[I])
520 continue; // This insn has been counted already.
521 if (I->getNumUses() == 0)
523 bool AllUsersFolded = true;
524 for (auto U : I->users()) {
525 Instruction *UI = dyn_cast<Instruction>(U);
526 if (!SimplifiedValues[UI] && !DeadInstructions[UI]) {
527 AllUsersFolded = false;
531 if (AllUsersFolded) {
532 NumberOfOptimizedInstructions += TTI.getUserCost(I);
533 Worklist.push_back(I);
534 DeadInstructions[I] = true;
539 return NumberOfOptimizedInstructions;
543 // Complete loop unrolling can make some loads constant, and we need to know if
544 // that would expose any further optimization opportunities.
545 // This routine estimates this optimization effect and returns the number of
546 // instructions, that potentially might be optimized away.
548 approximateNumberOfOptimizedInstructions(const Loop *L, ScalarEvolution &SE,
550 const TargetTransformInfo &TTI) {
551 if (!TripCount || !UnrollMaxIterationsCountToAnalyze)
554 UnrollAnalyzer UA(L, TripCount, SE, TTI);
555 UA.findConstFoldableLoads();
557 // Estimate number of instructions, that could be simplified if we replace a
558 // load with the corresponding constant. Since the same load will take
559 // different values on different iterations, we have to go through all loop's
560 // iterations here. To limit ourselves here, we check only first N
561 // iterations, and then scale the found number, if necessary.
562 unsigned IterationsNumberForEstimate =
563 std::min<unsigned>(UnrollMaxIterationsCountToAnalyze, TripCount);
564 unsigned NumberOfOptimizedInstructions = 0;
565 for (unsigned i = 0; i < IterationsNumberForEstimate; ++i) {
566 NumberOfOptimizedInstructions += UA.estimateNumberOfSimplifiedInsns(i);
567 NumberOfOptimizedInstructions += UA.estimateNumberOfDeadInsns();
569 NumberOfOptimizedInstructions *= TripCount / IterationsNumberForEstimate;
571 return NumberOfOptimizedInstructions;
574 /// ApproximateLoopSize - Approximate the size of the loop.
575 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
576 bool &NotDuplicatable,
577 const TargetTransformInfo &TTI,
578 AssumptionCache *AC) {
579 SmallPtrSet<const Value *, 32> EphValues;
580 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
583 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
585 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
586 NumCalls = Metrics.NumInlineCandidates;
587 NotDuplicatable = Metrics.notDuplicatable;
589 unsigned LoopSize = Metrics.NumInsts;
591 // Don't allow an estimate of size zero. This would allows unrolling of loops
592 // with huge iteration counts, which is a compile time problem even if it's
593 // not a problem for code quality. Also, the code using this size may assume
594 // that each loop has at least three instructions (likely a conditional
595 // branch, a comparison feeding that branch, and some kind of loop increment
596 // feeding that comparison instruction).
597 LoopSize = std::max(LoopSize, 3u);
602 // Returns the loop hint metadata node with the given name (for example,
603 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
605 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
606 if (MDNode *LoopID = L->getLoopID())
607 return GetUnrollMetadata(LoopID, Name);
611 // Returns true if the loop has an unroll(full) pragma.
612 static bool HasUnrollFullPragma(const Loop *L) {
613 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
616 // Returns true if the loop has an unroll(disable) pragma.
617 static bool HasUnrollDisablePragma(const Loop *L) {
618 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
621 // If loop has an unroll_count pragma return the (necessarily
622 // positive) value from the pragma. Otherwise return 0.
623 static unsigned UnrollCountPragmaValue(const Loop *L) {
624 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
626 assert(MD->getNumOperands() == 2 &&
627 "Unroll count hint metadata should have two operands.");
629 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
630 assert(Count >= 1 && "Unroll count must be positive.");
636 // Remove existing unroll metadata and add unroll disable metadata to
637 // indicate the loop has already been unrolled. This prevents a loop
638 // from being unrolled more than is directed by a pragma if the loop
639 // unrolling pass is run more than once (which it generally is).
640 static void SetLoopAlreadyUnrolled(Loop *L) {
641 MDNode *LoopID = L->getLoopID();
644 // First remove any existing loop unrolling metadata.
645 SmallVector<Metadata *, 4> MDs;
646 // Reserve first location for self reference to the LoopID metadata node.
647 MDs.push_back(nullptr);
648 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
649 bool IsUnrollMetadata = false;
650 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
652 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
653 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
655 if (!IsUnrollMetadata)
656 MDs.push_back(LoopID->getOperand(i));
659 // Add unroll(disable) metadata to disable future unrolling.
660 LLVMContext &Context = L->getHeader()->getContext();
661 SmallVector<Metadata *, 1> DisableOperands;
662 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
663 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
664 MDs.push_back(DisableNode);
666 MDNode *NewLoopID = MDNode::get(Context, MDs);
667 // Set operand 0 to refer to the loop id itself.
668 NewLoopID->replaceOperandWith(0, NewLoopID);
669 L->setLoopID(NewLoopID);
672 unsigned LoopUnroll::selectUnrollCount(
673 const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
674 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
675 bool &SetExplicitly) {
676 SetExplicitly = true;
678 // User-specified count (either as a command-line option or
679 // constructor parameter) has highest precedence.
680 unsigned Count = UserCount ? CurrentCount : 0;
682 // If there is no user-specified count, unroll pragmas have the next
683 // highest precendence.
687 } else if (PragmaFullUnroll) {
696 SetExplicitly = false;
698 // Runtime trip count.
699 Count = UnrollRuntimeCount;
701 // Conservative heuristic: if we know the trip count, see if we can
702 // completely unroll (subject to the threshold, checked below); otherwise
703 // try to find greatest modulo of the trip count which is still under
707 if (TripCount && Count > TripCount)
712 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
713 if (skipOptnoneFunction(L))
716 Function &F = *L->getHeader()->getParent();
718 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
719 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
720 const TargetTransformInfo &TTI =
721 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
722 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
724 BasicBlock *Header = L->getHeader();
725 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
726 << "] Loop %" << Header->getName() << "\n");
728 if (HasUnrollDisablePragma(L)) {
731 bool PragmaFullUnroll = HasUnrollFullPragma(L);
732 unsigned PragmaCount = UnrollCountPragmaValue(L);
733 bool HasPragma = PragmaFullUnroll || PragmaCount > 0;
735 TargetTransformInfo::UnrollingPreferences UP;
736 getUnrollingPreferences(L, TTI, UP);
738 // Find trip count and trip multiple if count is not available
739 unsigned TripCount = 0;
740 unsigned TripMultiple = 1;
741 // If there are multiple exiting blocks but one of them is the latch, use the
742 // latch for the trip count estimation. Otherwise insist on a single exiting
743 // block for the trip count estimation.
744 BasicBlock *ExitingBlock = L->getLoopLatch();
745 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
746 ExitingBlock = L->getExitingBlock();
748 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
749 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
752 // Select an initial unroll count. This may be reduced later based
753 // on size thresholds.
754 bool CountSetExplicitly;
755 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll,
756 PragmaCount, UP, CountSetExplicitly);
758 unsigned NumInlineCandidates;
759 bool notDuplicatable;
761 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
762 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
764 // When computing the unrolled size, note that the conditional branch on the
765 // backedge and the comparison feeding it are not replicated like the rest of
766 // the loop body (which is why 2 is subtracted).
767 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
768 if (notDuplicatable) {
769 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
770 << " instructions.\n");
773 if (NumInlineCandidates != 0) {
774 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
778 unsigned NumberOfOptimizedInstructions =
779 approximateNumberOfOptimizedInstructions(L, *SE, TripCount, TTI);
780 DEBUG(dbgs() << " Complete unrolling could save: "
781 << NumberOfOptimizedInstructions << "\n");
783 unsigned Threshold, PartialThreshold;
784 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold,
785 NumberOfOptimizedInstructions);
787 // Given Count, TripCount and thresholds determine the type of
788 // unrolling which is to be performed.
789 enum { Full = 0, Partial = 1, Runtime = 2 };
791 if (TripCount && Count == TripCount) {
792 if (Threshold != NoThreshold && UnrolledSize > Threshold) {
793 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
794 << " because size: " << UnrolledSize << ">" << Threshold
800 } else if (TripCount && Count < TripCount) {
806 // Reduce count based on the type of unrolling and the threshold values.
807 unsigned OriginalCount = Count;
808 bool AllowRuntime = UserRuntime ? CurrentRuntime : UP.Runtime;
809 if (Unrolling == Partial) {
810 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
811 if (!AllowPartial && !CountSetExplicitly) {
812 DEBUG(dbgs() << " will not try to unroll partially because "
813 << "-unroll-allow-partial not given\n");
816 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
817 // Reduce unroll count to be modulo of TripCount for partial unrolling.
818 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
819 while (Count != 0 && TripCount % Count != 0)
822 } else if (Unrolling == Runtime) {
823 if (!AllowRuntime && !CountSetExplicitly) {
824 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
825 << "-unroll-runtime not given\n");
828 // Reduce unroll count to be the largest power-of-two factor of
829 // the original count which satisfies the threshold limit.
830 while (Count != 0 && UnrolledSize > PartialThreshold) {
832 UnrolledSize = (LoopSize-2) * Count + 2;
834 if (Count > UP.MaxCount)
836 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
840 if (PragmaCount != 0)
841 // If loop has an unroll count pragma mark loop as unrolled to prevent
842 // unrolling beyond that requested by the pragma.
843 SetLoopAlreadyUnrolled(L);
845 // Emit optimization remarks if we are unable to unroll the loop
846 // as directed by a pragma.
847 DebugLoc LoopLoc = L->getStartLoc();
848 Function *F = Header->getParent();
849 LLVMContext &Ctx = F->getContext();
850 if (PragmaFullUnroll && PragmaCount == 0) {
851 if (TripCount && Count != TripCount) {
852 emitOptimizationRemarkMissed(
853 Ctx, DEBUG_TYPE, *F, LoopLoc,
854 "Unable to fully unroll loop as directed by unroll(full) pragma "
855 "because unrolled size is too large.");
856 } else if (!TripCount) {
857 emitOptimizationRemarkMissed(
858 Ctx, DEBUG_TYPE, *F, LoopLoc,
859 "Unable to fully unroll loop as directed by unroll(full) pragma "
860 "because loop has a runtime trip count.");
862 } else if (PragmaCount > 0 && Count != OriginalCount) {
863 emitOptimizationRemarkMissed(
864 Ctx, DEBUG_TYPE, *F, LoopLoc,
865 "Unable to unroll loop the number of times directed by "
866 "unroll_count pragma because unrolled size is too large.");
870 if (Unrolling != Full && Count < 2) {
871 // Partial unrolling by 1 is a nop. For full unrolling, a factor
872 // of 1 makes sense because loop control can be eliminated.
877 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, TripMultiple, LI, this,