1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements a simple loop unroller. It works best when loops have
11 // been canonicalized by the -indvars pass, allowing it to determine the trip
12 // counts of loops easily.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/Analysis/AssumptionCache.h"
18 #include "llvm/Analysis/CodeMetrics.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LoopPass.h"
21 #include "llvm/Analysis/ScalarEvolution.h"
22 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DiagnosticInfo.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/InstVisitor.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Metadata.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Transforms/Utils/UnrollLoop.h"
38 #define DEBUG_TYPE "loop-unroll"
40 static cl::opt<unsigned>
41 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
42 cl::desc("The baseline cost threshold for loop unrolling"));
44 static cl::opt<unsigned> UnrollPercentDynamicCostSavedThreshold(
45 "unroll-percent-dynamic-cost-saved-threshold", cl::init(20), cl::Hidden,
46 cl::desc("The percentage of estimated dynamic cost which must be saved by "
47 "unrolling to allow unrolling up to the max threshold."));
49 static cl::opt<unsigned> UnrollDynamicCostSavingsDiscount(
50 "unroll-dynamic-cost-savings-discount", cl::init(2000), cl::Hidden,
51 cl::desc("This is the amount discounted from the total unroll cost when "
52 "the unrolled form has a high dynamic cost savings (triggered by "
53 "the '-unroll-perecent-dynamic-cost-saved-threshold' flag)."));
55 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
56 "unroll-max-iteration-count-to-analyze", cl::init(0), cl::Hidden,
57 cl::desc("Don't allow loop unrolling to simulate more than this number of"
58 "iterations when checking full unroll profitability"));
60 static cl::opt<unsigned>
61 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
62 cl::desc("Use this unroll count for all loops including those with "
63 "unroll_count pragma values, for testing purposes"));
66 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
67 cl::desc("Allows loops to be partially unrolled until "
68 "-unroll-threshold loop size is reached."));
71 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden,
72 cl::desc("Unroll loops with run-time trip counts"));
74 static cl::opt<unsigned>
75 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
76 cl::desc("Unrolled size limit for loops with an unroll(full) or "
77 "unroll_count pragma."));
80 class LoopUnroll : public LoopPass {
82 static char ID; // Pass ID, replacement for typeid
83 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) {
84 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
85 CurrentPercentDynamicCostSavedThreshold =
86 UnrollPercentDynamicCostSavedThreshold;
87 CurrentDynamicCostSavingsDiscount = UnrollDynamicCostSavingsDiscount;
88 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
89 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
90 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
92 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
93 UserPercentDynamicCostSavedThreshold =
94 (UnrollPercentDynamicCostSavedThreshold.getNumOccurrences() > 0);
95 UserDynamicCostSavingsDiscount =
96 (UnrollDynamicCostSavingsDiscount.getNumOccurrences() > 0);
97 UserAllowPartial = (P != -1) ||
98 (UnrollAllowPartial.getNumOccurrences() > 0);
99 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
100 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0);
102 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
105 /// A magic value for use with the Threshold parameter to indicate
106 /// that the loop unroll should be performed regardless of how much
107 /// code expansion would result.
108 static const unsigned NoThreshold = UINT_MAX;
110 // Threshold to use when optsize is specified (and there is no
111 // explicit -unroll-threshold).
112 static const unsigned OptSizeUnrollThreshold = 50;
114 // Default unroll count for loops with run-time trip count if
115 // -unroll-count is not set
116 static const unsigned UnrollRuntimeCount = 8;
118 unsigned CurrentCount;
119 unsigned CurrentThreshold;
120 unsigned CurrentPercentDynamicCostSavedThreshold;
121 unsigned CurrentDynamicCostSavingsDiscount;
122 bool CurrentAllowPartial;
125 // Flags for whether the 'current' settings are user-specified.
128 bool UserPercentDynamicCostSavedThreshold;
129 bool UserDynamicCostSavingsDiscount;
130 bool UserAllowPartial;
133 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
135 /// This transformation requires natural loop information & requires that
136 /// loop preheaders be inserted into the CFG...
138 void getAnalysisUsage(AnalysisUsage &AU) const override {
139 AU.addRequired<AssumptionCacheTracker>();
140 AU.addRequired<DominatorTreeWrapperPass>();
141 AU.addRequired<LoopInfoWrapperPass>();
142 AU.addPreserved<LoopInfoWrapperPass>();
143 AU.addRequiredID(LoopSimplifyID);
144 AU.addPreservedID(LoopSimplifyID);
145 AU.addRequiredID(LCSSAID);
146 AU.addPreservedID(LCSSAID);
147 AU.addRequired<ScalarEvolution>();
148 AU.addPreserved<ScalarEvolution>();
149 AU.addRequired<TargetTransformInfoWrapperPass>();
150 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
151 // If loop unroll does not preserve dom info then LCSSA pass on next
152 // loop will receive invalid dom info.
153 // For now, recreate dom info, if loop is unrolled.
154 AU.addPreserved<DominatorTreeWrapperPass>();
157 // Fill in the UnrollingPreferences parameter with values from the
158 // TargetTransformationInfo.
159 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI,
160 TargetTransformInfo::UnrollingPreferences &UP) {
161 UP.Threshold = CurrentThreshold;
162 UP.PercentDynamicCostSavedThreshold =
163 CurrentPercentDynamicCostSavedThreshold;
164 UP.DynamicCostSavingsDiscount = CurrentDynamicCostSavingsDiscount;
165 UP.OptSizeThreshold = OptSizeUnrollThreshold;
166 UP.PartialThreshold = CurrentThreshold;
167 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
168 UP.Count = CurrentCount;
169 UP.MaxCount = UINT_MAX;
170 UP.Partial = CurrentAllowPartial;
171 UP.Runtime = CurrentRuntime;
172 UP.AllowExpensiveTripCount = false;
173 TTI.getUnrollingPreferences(L, UP);
176 // Select and return an unroll count based on parameters from
177 // user, unroll preferences, unroll pragmas, or a heuristic.
178 // SetExplicitly is set to true if the unroll count is is set by
179 // the user or a pragma rather than selected heuristically.
181 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
182 unsigned PragmaCount,
183 const TargetTransformInfo::UnrollingPreferences &UP,
184 bool &SetExplicitly);
186 // Select threshold values used to limit unrolling based on a
187 // total unrolled size. Parameters Threshold and PartialThreshold
188 // are set to the maximum unrolled size for fully and partially
189 // unrolled loops respectively.
190 void selectThresholds(const Loop *L, bool HasPragma,
191 const TargetTransformInfo::UnrollingPreferences &UP,
192 unsigned &Threshold, unsigned &PartialThreshold,
193 unsigned &PercentDynamicCostSavedThreshold,
194 unsigned &DynamicCostSavingsDiscount) {
195 // Determine the current unrolling threshold. While this is
196 // normally set from UnrollThreshold, it is overridden to a
197 // smaller value if the current function is marked as
198 // optimize-for-size, and the unroll threshold was not user
200 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
201 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
202 PercentDynamicCostSavedThreshold =
203 UserPercentDynamicCostSavedThreshold
204 ? CurrentPercentDynamicCostSavedThreshold
205 : UP.PercentDynamicCostSavedThreshold;
206 DynamicCostSavingsDiscount = UserDynamicCostSavingsDiscount
207 ? CurrentDynamicCostSavingsDiscount
208 : UP.DynamicCostSavingsDiscount;
210 if (!UserThreshold &&
211 L->getHeader()->getParent()->hasFnAttribute(
212 Attribute::OptimizeForSize)) {
213 Threshold = UP.OptSizeThreshold;
214 PartialThreshold = UP.PartialOptSizeThreshold;
217 // If the loop has an unrolling pragma, we want to be more
218 // aggressive with unrolling limits. Set thresholds to at
219 // least the PragmaTheshold value which is larger than the
221 if (Threshold != NoThreshold)
222 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold);
223 if (PartialThreshold != NoThreshold)
225 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold);
228 bool canUnrollCompletely(Loop *L, unsigned Threshold,
229 unsigned PercentDynamicCostSavedThreshold,
230 unsigned DynamicCostSavingsDiscount,
231 uint64_t UnrolledCost, uint64_t RolledDynamicCost);
235 char LoopUnroll::ID = 0;
236 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
237 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
238 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
239 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
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);
256 // This class is used to get an estimate of the optimization effects that we
257 // could get from complete loop unrolling. It comes from the fact that some
258 // loads might be replaced with concrete constant values and that could trigger
259 // a chain of instruction simplifications.
261 // E.g. we might have:
262 // int a[] = {0, 1, 0};
264 // for (i = 0; i < 3; i ++)
266 // If we completely unroll the loop, we would get:
267 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
268 // Which then will be simplified to:
269 // v = b[0]* 0 + b[1]* 1 + b[2]* 0
272 class UnrolledInstAnalyzer : private InstVisitor<UnrolledInstAnalyzer, bool> {
273 typedef InstVisitor<UnrolledInstAnalyzer, bool> Base;
274 friend class InstVisitor<UnrolledInstAnalyzer, bool>;
275 struct SimplifiedAddress {
276 Value *Base = nullptr;
277 ConstantInt *Offset = nullptr;
281 UnrolledInstAnalyzer(unsigned Iteration,
282 DenseMap<Value *, Constant *> &SimplifiedValues,
283 const Loop *L, ScalarEvolution &SE)
284 : Iteration(Iteration), SimplifiedValues(SimplifiedValues), L(L), SE(SE) {
285 IterationNumber = SE.getConstant(APInt(64, Iteration));
288 // Allow access to the initial visit method.
292 /// \brief A cache of pointer bases and constant-folded offsets corresponding
293 /// to GEP (or derived from GEP) instructions.
295 /// In order to find the base pointer one needs to perform non-trivial
296 /// traversal of the corresponding SCEV expression, so it's good to have the
298 DenseMap<Value *, SimplifiedAddress> SimplifiedAddresses;
300 /// \brief Number of currently simulated iteration.
302 /// If an expression is ConstAddress+Constant, then the Constant is
303 /// Start + Iteration*Step, where Start and Step could be obtained from
307 /// \brief SCEV expression corresponding to number of currently simulated
309 const SCEV *IterationNumber;
311 /// \brief A Value->Constant map for keeping values that we managed to
312 /// constant-fold on the given iteration.
314 /// While we walk the loop instructions, we build up and maintain a mapping
315 /// of simplified values specific to this iteration. The idea is to propagate
316 /// any special information we have about loads that can be replaced with
317 /// constants after complete unrolling, and account for likely simplifications
319 DenseMap<Value *, Constant *> &SimplifiedValues;
324 /// \brief Try to simplify instruction \param I using its SCEV expression.
326 /// The idea is that some AddRec expressions become constants, which then
327 /// could trigger folding of other instructions. However, that only happens
328 /// for expressions whose start value is also constant, which isn't always the
329 /// case. In another common and important case the start value is just some
330 /// address (i.e. SCEVUnknown) - in this case we compute the offset and save
331 /// it along with the base address instead.
332 bool simplifyInstWithSCEV(Instruction *I) {
333 if (!SE.isSCEVable(I->getType()))
336 const SCEV *S = SE.getSCEV(I);
337 if (auto *SC = dyn_cast<SCEVConstant>(S)) {
338 SimplifiedValues[I] = SC->getValue();
342 auto *AR = dyn_cast<SCEVAddRecExpr>(S);
346 const SCEV *ValueAtIteration = AR->evaluateAtIteration(IterationNumber, SE);
347 // Check if the AddRec expression becomes a constant.
348 if (auto *SC = dyn_cast<SCEVConstant>(ValueAtIteration)) {
349 SimplifiedValues[I] = SC->getValue();
353 // Check if the offset from the base address becomes a constant.
354 auto *Base = dyn_cast<SCEVUnknown>(SE.getPointerBase(S));
358 dyn_cast<SCEVConstant>(SE.getMinusSCEV(ValueAtIteration, Base));
361 SimplifiedAddress Address;
362 Address.Base = Base->getValue();
363 Address.Offset = Offset->getValue();
364 SimplifiedAddresses[I] = Address;
368 /// Base case for the instruction visitor.
369 bool visitInstruction(Instruction &I) {
370 return simplifyInstWithSCEV(&I);
373 /// Try to simplify binary operator I.
375 /// TODO: Probaly it's worth to hoist the code for estimating the
376 /// simplifications effects to a separate class, since we have a very similar
377 /// code in InlineCost already.
378 bool visitBinaryOperator(BinaryOperator &I) {
379 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
380 if (!isa<Constant>(LHS))
381 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
383 if (!isa<Constant>(RHS))
384 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
387 Value *SimpleV = nullptr;
388 const DataLayout &DL = I.getModule()->getDataLayout();
389 if (auto FI = dyn_cast<FPMathOperator>(&I))
391 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags(), DL);
393 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL);
395 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV))
396 SimplifiedValues[&I] = C;
400 return Base::visitBinaryOperator(I);
403 /// Try to fold load I.
404 bool visitLoad(LoadInst &I) {
405 Value *AddrOp = I.getPointerOperand();
407 auto AddressIt = SimplifiedAddresses.find(AddrOp);
408 if (AddressIt == SimplifiedAddresses.end())
410 ConstantInt *SimplifiedAddrOp = AddressIt->second.Offset;
412 auto *GV = dyn_cast<GlobalVariable>(AddressIt->second.Base);
413 // We're only interested in loads that can be completely folded to a
415 if (!GV || !GV->hasInitializer())
418 ConstantDataSequential *CDS =
419 dyn_cast<ConstantDataSequential>(GV->getInitializer());
423 int ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
424 assert(SimplifiedAddrOp->getValue().getActiveBits() < 64 &&
425 "Unexpectedly large index value.");
426 int64_t Index = SimplifiedAddrOp->getSExtValue() / ElemSize;
427 if (Index >= CDS->getNumElements()) {
428 // FIXME: For now we conservatively ignore out of bound accesses, but
429 // we're allowed to perform the optimization in this case.
433 Constant *CV = CDS->getElementAsConstant(Index);
434 assert(CV && "Constant expected.");
435 SimplifiedValues[&I] = CV;
440 bool visitCastInst(CastInst &I) {
441 // Propagate constants through casts.
442 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
444 COp = SimplifiedValues.lookup(I.getOperand(0));
447 ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) {
448 SimplifiedValues[&I] = C;
452 return Base::visitCastInst(I);
455 bool visitCmpInst(CmpInst &I) {
456 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
458 // First try to handle simplified comparisons.
459 if (!isa<Constant>(LHS))
460 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
462 if (!isa<Constant>(RHS))
463 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
466 if (!isa<Constant>(LHS) && !isa<Constant>(RHS)) {
467 auto SimplifiedLHS = SimplifiedAddresses.find(LHS);
468 if (SimplifiedLHS != SimplifiedAddresses.end()) {
469 auto SimplifiedRHS = SimplifiedAddresses.find(RHS);
470 if (SimplifiedRHS != SimplifiedAddresses.end()) {
471 SimplifiedAddress &LHSAddr = SimplifiedLHS->second;
472 SimplifiedAddress &RHSAddr = SimplifiedRHS->second;
473 if (LHSAddr.Base == RHSAddr.Base) {
474 LHS = LHSAddr.Offset;
475 RHS = RHSAddr.Offset;
481 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
482 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
483 if (Constant *C = ConstantExpr::getCompare(I.getPredicate(), CLHS, CRHS)) {
484 SimplifiedValues[&I] = C;
490 return Base::visitCmpInst(I);
497 struct EstimatedUnrollCost {
498 /// \brief The estimated cost after unrolling.
499 unsigned UnrolledCost;
501 /// \brief The estimated dynamic cost of executing the instructions in the
503 unsigned RolledDynamicCost;
507 /// \brief Figure out if the loop is worth full unrolling.
509 /// Complete loop unrolling can make some loads constant, and we need to know
510 /// if that would expose any further optimization opportunities. This routine
511 /// estimates this optimization. It computes cost of unrolled loop
512 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
513 /// dynamic cost we mean that we won't count costs of blocks that are known not
514 /// to be executed (i.e. if we have a branch in the loop and we know that at the
515 /// given iteration its condition would be resolved to true, we won't add up the
516 /// cost of the 'false'-block).
517 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
518 /// the analysis failed (no benefits expected from the unrolling, or the loop is
519 /// too big to analyze), the returned value is None.
520 Optional<EstimatedUnrollCost>
521 analyzeLoopUnrollCost(const Loop *L, unsigned TripCount, DominatorTree &DT,
522 ScalarEvolution &SE, const TargetTransformInfo &TTI,
523 unsigned MaxUnrolledLoopSize) {
524 // We want to be able to scale offsets by the trip count and add more offsets
525 // to them without checking for overflows, and we already don't want to
526 // analyze *massive* trip counts, so we force the max to be reasonably small.
527 assert(UnrollMaxIterationsCountToAnalyze < (INT_MAX / 2) &&
528 "The unroll iterations max is too large!");
530 // Don't simulate loops with a big or unknown tripcount
531 if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
532 TripCount > UnrollMaxIterationsCountToAnalyze)
535 SmallSetVector<BasicBlock *, 16> BBWorklist;
536 DenseMap<Value *, Constant *> SimplifiedValues;
537 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues;
539 // The estimated cost of the unrolled form of the loop. We try to estimate
540 // this by simplifying as much as we can while computing the estimate.
541 unsigned UnrolledCost = 0;
542 // We also track the estimated dynamic (that is, actually executed) cost in
543 // the rolled form. This helps identify cases when the savings from unrolling
544 // aren't just exposing dead control flows, but actual reduced dynamic
545 // instructions due to the simplifications which we expect to occur after
547 unsigned RolledDynamicCost = 0;
549 // Ensure that we don't violate the loop structure invariants relied on by
551 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
552 assert(L->isLCSSAForm(DT) &&
553 "Must have loops in LCSSA form to track live-out values.");
555 DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
557 // Simulate execution of each iteration of the loop counting instructions,
558 // which would be simplified.
559 // Since the same load will take different values on different iterations,
560 // we literally have to go through all loop's iterations.
561 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
562 DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
564 // Prepare for the iteration by collecting any simplified entry or backedge
566 for (Instruction &I : *L->getHeader()) {
567 auto *PHI = dyn_cast<PHINode>(&I);
571 // The loop header PHI nodes must have exactly two input: one from the
572 // loop preheader and one from the loop latch.
574 PHI->getNumIncomingValues() == 2 &&
575 "Must have an incoming value only for the preheader and the latch.");
577 Value *V = PHI->getIncomingValueForBlock(
578 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
579 Constant *C = dyn_cast<Constant>(V);
580 if (Iteration != 0 && !C)
581 C = SimplifiedValues.lookup(V);
583 SimplifiedInputValues.push_back({PHI, C});
586 // Now clear and re-populate the map for the next iteration.
587 SimplifiedValues.clear();
588 while (!SimplifiedInputValues.empty())
589 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val());
591 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, L, SE);
594 BBWorklist.insert(L->getHeader());
595 // Note that we *must not* cache the size, this loop grows the worklist.
596 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
597 BasicBlock *BB = BBWorklist[Idx];
599 // Visit all instructions in the given basic block and try to simplify
600 // it. We don't change the actual IR, just count optimization
602 for (Instruction &I : *BB) {
603 unsigned InstCost = TTI.getUserCost(&I);
605 // Visit the instruction to analyze its loop cost after unrolling,
606 // and if the visitor returns false, include this instruction in the
608 if (!Analyzer.visit(I))
609 UnrolledCost += InstCost;
611 DEBUG(dbgs() << " " << I
612 << " would be simplified if loop is unrolled.\n");
616 // Also track this instructions expected cost when executing the rolled
618 RolledDynamicCost += InstCost;
620 // If unrolled body turns out to be too big, bail out.
621 if (UnrolledCost > MaxUnrolledLoopSize) {
622 DEBUG(dbgs() << " Exceeded threshold.. exiting.\n"
623 << " UnrolledCost: " << UnrolledCost
624 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
630 TerminatorInst *TI = BB->getTerminator();
632 // Add in the live successors by first checking whether we have terminator
633 // that may be simplified based on the values simplified by this call.
634 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
635 if (BI->isConditional()) {
636 if (Constant *SimpleCond =
637 SimplifiedValues.lookup(BI->getCondition())) {
638 BasicBlock *Succ = nullptr;
639 // Just take the first successor if condition is undef
640 if (isa<UndefValue>(SimpleCond))
641 Succ = BI->getSuccessor(0);
643 Succ = BI->getSuccessor(
644 cast<ConstantInt>(SimpleCond)->isZero() ? 1 : 0);
645 if (L->contains(Succ))
646 BBWorklist.insert(Succ);
650 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
651 if (Constant *SimpleCond =
652 SimplifiedValues.lookup(SI->getCondition())) {
653 BasicBlock *Succ = nullptr;
654 // Just take the first successor if condition is undef
655 if (isa<UndefValue>(SimpleCond))
656 Succ = SI->getSuccessor(0);
658 Succ = SI->findCaseValue(cast<ConstantInt>(SimpleCond))
660 if (L->contains(Succ))
661 BBWorklist.insert(Succ);
666 // Add BB's successors to the worklist.
667 for (BasicBlock *Succ : successors(BB))
668 if (L->contains(Succ))
669 BBWorklist.insert(Succ);
672 // If we found no optimization opportunities on the first iteration, we
673 // won't find them on later ones too.
674 if (UnrolledCost == RolledDynamicCost) {
675 DEBUG(dbgs() << " No opportunities found.. exiting.\n"
676 << " UnrolledCost: " << UnrolledCost << "\n");
680 DEBUG(dbgs() << "Analysis finished:\n"
681 << "UnrolledCost: " << UnrolledCost << ", "
682 << "RolledDynamicCost: " << RolledDynamicCost << "\n");
683 return {{UnrolledCost, RolledDynamicCost}};
686 /// ApproximateLoopSize - Approximate the size of the loop.
687 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
688 bool &NotDuplicatable,
689 const TargetTransformInfo &TTI,
690 AssumptionCache *AC) {
691 SmallPtrSet<const Value *, 32> EphValues;
692 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
695 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
697 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
698 NumCalls = Metrics.NumInlineCandidates;
699 NotDuplicatable = Metrics.notDuplicatable;
701 unsigned LoopSize = Metrics.NumInsts;
703 // Don't allow an estimate of size zero. This would allows unrolling of loops
704 // with huge iteration counts, which is a compile time problem even if it's
705 // not a problem for code quality. Also, the code using this size may assume
706 // that each loop has at least three instructions (likely a conditional
707 // branch, a comparison feeding that branch, and some kind of loop increment
708 // feeding that comparison instruction).
709 LoopSize = std::max(LoopSize, 3u);
714 // Returns the loop hint metadata node with the given name (for example,
715 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
717 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
718 if (MDNode *LoopID = L->getLoopID())
719 return GetUnrollMetadata(LoopID, Name);
723 // Returns true if the loop has an unroll(full) pragma.
724 static bool HasUnrollFullPragma(const Loop *L) {
725 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
728 // Returns true if the loop has an unroll(disable) pragma.
729 static bool HasUnrollDisablePragma(const Loop *L) {
730 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
733 // Returns true if the loop has an runtime unroll(disable) pragma.
734 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
735 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
738 // If loop has an unroll_count pragma return the (necessarily
739 // positive) value from the pragma. Otherwise return 0.
740 static unsigned UnrollCountPragmaValue(const Loop *L) {
741 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
743 assert(MD->getNumOperands() == 2 &&
744 "Unroll count hint metadata should have two operands.");
746 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
747 assert(Count >= 1 && "Unroll count must be positive.");
753 // Remove existing unroll metadata and add unroll disable metadata to
754 // indicate the loop has already been unrolled. This prevents a loop
755 // from being unrolled more than is directed by a pragma if the loop
756 // unrolling pass is run more than once (which it generally is).
757 static void SetLoopAlreadyUnrolled(Loop *L) {
758 MDNode *LoopID = L->getLoopID();
761 // First remove any existing loop unrolling metadata.
762 SmallVector<Metadata *, 4> MDs;
763 // Reserve first location for self reference to the LoopID metadata node.
764 MDs.push_back(nullptr);
765 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
766 bool IsUnrollMetadata = false;
767 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
769 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
770 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
772 if (!IsUnrollMetadata)
773 MDs.push_back(LoopID->getOperand(i));
776 // Add unroll(disable) metadata to disable future unrolling.
777 LLVMContext &Context = L->getHeader()->getContext();
778 SmallVector<Metadata *, 1> DisableOperands;
779 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
780 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
781 MDs.push_back(DisableNode);
783 MDNode *NewLoopID = MDNode::get(Context, MDs);
784 // Set operand 0 to refer to the loop id itself.
785 NewLoopID->replaceOperandWith(0, NewLoopID);
786 L->setLoopID(NewLoopID);
789 bool LoopUnroll::canUnrollCompletely(Loop *L, unsigned Threshold,
790 unsigned PercentDynamicCostSavedThreshold,
791 unsigned DynamicCostSavingsDiscount,
792 uint64_t UnrolledCost,
793 uint64_t RolledDynamicCost) {
795 if (Threshold == NoThreshold) {
796 DEBUG(dbgs() << " Can fully unroll, because no threshold is set.\n");
800 if (UnrolledCost <= Threshold) {
801 DEBUG(dbgs() << " Can fully unroll, because unrolled cost: "
802 << UnrolledCost << "<" << Threshold << "\n");
806 assert(UnrolledCost && "UnrolledCost can't be 0 at this point.");
807 assert(RolledDynamicCost >= UnrolledCost &&
808 "Cannot have a higher unrolled cost than a rolled cost!");
810 // Compute the percentage of the dynamic cost in the rolled form that is
811 // saved when unrolled. If unrolling dramatically reduces the estimated
812 // dynamic cost of the loop, we use a higher threshold to allow more
814 unsigned PercentDynamicCostSaved =
815 (uint64_t)(RolledDynamicCost - UnrolledCost) * 100ull / RolledDynamicCost;
817 if (PercentDynamicCostSaved >= PercentDynamicCostSavedThreshold &&
818 (int64_t)UnrolledCost - (int64_t)DynamicCostSavingsDiscount <=
819 (int64_t)Threshold) {
820 DEBUG(dbgs() << " Can fully unroll, because unrolling will reduce the "
821 "expected dynamic cost by " << PercentDynamicCostSaved
822 << "% (threshold: " << PercentDynamicCostSavedThreshold
824 << " and the unrolled cost (" << UnrolledCost
825 << ") is less than the max threshold ("
826 << DynamicCostSavingsDiscount << ").\n");
830 DEBUG(dbgs() << " Too large to fully unroll:\n");
831 DEBUG(dbgs() << " Threshold: " << Threshold << "\n");
832 DEBUG(dbgs() << " Max threshold: " << DynamicCostSavingsDiscount << "\n");
833 DEBUG(dbgs() << " Percent cost saved threshold: "
834 << PercentDynamicCostSavedThreshold << "%\n");
835 DEBUG(dbgs() << " Unrolled cost: " << UnrolledCost << "\n");
836 DEBUG(dbgs() << " Rolled dynamic cost: " << RolledDynamicCost << "\n");
837 DEBUG(dbgs() << " Percent cost saved: " << PercentDynamicCostSaved
842 unsigned LoopUnroll::selectUnrollCount(
843 const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
844 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
845 bool &SetExplicitly) {
846 SetExplicitly = true;
848 // User-specified count (either as a command-line option or
849 // constructor parameter) has highest precedence.
850 unsigned Count = UserCount ? CurrentCount : 0;
852 // If there is no user-specified count, unroll pragmas have the next
853 // highest precendence.
857 } else if (PragmaFullUnroll) {
866 SetExplicitly = false;
868 // Runtime trip count.
869 Count = UnrollRuntimeCount;
871 // Conservative heuristic: if we know the trip count, see if we can
872 // completely unroll (subject to the threshold, checked below); otherwise
873 // try to find greatest modulo of the trip count which is still under
877 if (TripCount && Count > TripCount)
882 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
883 if (skipOptnoneFunction(L))
886 Function &F = *L->getHeader()->getParent();
888 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
889 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
890 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
891 const TargetTransformInfo &TTI =
892 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
893 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
895 BasicBlock *Header = L->getHeader();
896 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
897 << "] Loop %" << Header->getName() << "\n");
899 if (HasUnrollDisablePragma(L)) {
902 bool PragmaFullUnroll = HasUnrollFullPragma(L);
903 unsigned PragmaCount = UnrollCountPragmaValue(L);
904 bool HasPragma = PragmaFullUnroll || PragmaCount > 0;
906 TargetTransformInfo::UnrollingPreferences UP;
907 getUnrollingPreferences(L, TTI, UP);
909 // Find trip count and trip multiple if count is not available
910 unsigned TripCount = 0;
911 unsigned TripMultiple = 1;
912 // If there are multiple exiting blocks but one of them is the latch, use the
913 // latch for the trip count estimation. Otherwise insist on a single exiting
914 // block for the trip count estimation.
915 BasicBlock *ExitingBlock = L->getLoopLatch();
916 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
917 ExitingBlock = L->getExitingBlock();
919 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
920 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
923 // Select an initial unroll count. This may be reduced later based
924 // on size thresholds.
925 bool CountSetExplicitly;
926 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll,
927 PragmaCount, UP, CountSetExplicitly);
929 unsigned NumInlineCandidates;
930 bool notDuplicatable;
932 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
933 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
935 // When computing the unrolled size, note that the conditional branch on the
936 // backedge and the comparison feeding it are not replicated like the rest of
937 // the loop body (which is why 2 is subtracted).
938 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
939 if (notDuplicatable) {
940 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
941 << " instructions.\n");
944 if (NumInlineCandidates != 0) {
945 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
949 unsigned Threshold, PartialThreshold;
950 unsigned PercentDynamicCostSavedThreshold;
951 unsigned DynamicCostSavingsDiscount;
952 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold,
953 PercentDynamicCostSavedThreshold,
954 DynamicCostSavingsDiscount);
956 // Given Count, TripCount and thresholds determine the type of
957 // unrolling which is to be performed.
958 enum { Full = 0, Partial = 1, Runtime = 2 };
960 if (TripCount && Count == TripCount) {
962 // If the loop is really small, we don't need to run an expensive analysis.
963 if (canUnrollCompletely(L, Threshold, 100, DynamicCostSavingsDiscount,
964 UnrolledSize, UnrolledSize)) {
967 // The loop isn't that small, but we still can fully unroll it if that
968 // helps to remove a significant number of instructions.
969 // To check that, run additional analysis on the loop.
970 if (Optional<EstimatedUnrollCost> Cost =
971 analyzeLoopUnrollCost(L, TripCount, DT, *SE, TTI,
972 Threshold + DynamicCostSavingsDiscount))
973 if (canUnrollCompletely(L, Threshold, PercentDynamicCostSavedThreshold,
974 DynamicCostSavingsDiscount, Cost->UnrolledCost,
975 Cost->RolledDynamicCost)) {
979 } else if (TripCount && Count < TripCount) {
985 // Reduce count based on the type of unrolling and the threshold values.
986 unsigned OriginalCount = Count;
988 (PragmaCount > 0) || (UserRuntime ? CurrentRuntime : UP.Runtime);
989 // Don't unroll a runtime trip count loop with unroll full pragma.
990 if (HasRuntimeUnrollDisablePragma(L) || PragmaFullUnroll) {
991 AllowRuntime = false;
993 if (Unrolling == Partial) {
994 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
995 if (!AllowPartial && !CountSetExplicitly) {
996 DEBUG(dbgs() << " will not try to unroll partially because "
997 << "-unroll-allow-partial not given\n");
1000 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
1001 // Reduce unroll count to be modulo of TripCount for partial unrolling.
1002 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
1003 while (Count != 0 && TripCount % Count != 0)
1006 } else if (Unrolling == Runtime) {
1007 if (!AllowRuntime && !CountSetExplicitly) {
1008 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
1009 << "-unroll-runtime not given\n");
1012 // Reduce unroll count to be the largest power-of-two factor of
1013 // the original count which satisfies the threshold limit.
1014 while (Count != 0 && UnrolledSize > PartialThreshold) {
1016 UnrolledSize = (LoopSize-2) * Count + 2;
1018 if (Count > UP.MaxCount)
1019 Count = UP.MaxCount;
1020 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
1024 if (PragmaCount != 0)
1025 // If loop has an unroll count pragma mark loop as unrolled to prevent
1026 // unrolling beyond that requested by the pragma.
1027 SetLoopAlreadyUnrolled(L);
1029 // Emit optimization remarks if we are unable to unroll the loop
1030 // as directed by a pragma.
1031 DebugLoc LoopLoc = L->getStartLoc();
1032 Function *F = Header->getParent();
1033 LLVMContext &Ctx = F->getContext();
1034 if (PragmaFullUnroll && PragmaCount == 0) {
1035 if (TripCount && Count != TripCount) {
1036 emitOptimizationRemarkMissed(
1037 Ctx, DEBUG_TYPE, *F, LoopLoc,
1038 "Unable to fully unroll loop as directed by unroll(full) pragma "
1039 "because unrolled size is too large.");
1040 } else if (!TripCount) {
1041 emitOptimizationRemarkMissed(
1042 Ctx, DEBUG_TYPE, *F, LoopLoc,
1043 "Unable to fully unroll loop as directed by unroll(full) pragma "
1044 "because loop has a runtime trip count.");
1046 } else if (PragmaCount > 0 && Count != OriginalCount) {
1047 emitOptimizationRemarkMissed(
1048 Ctx, DEBUG_TYPE, *F, LoopLoc,
1049 "Unable to unroll loop the number of times directed by "
1050 "unroll_count pragma because unrolled size is too large.");
1054 if (Unrolling != Full && Count < 2) {
1055 // Partial unrolling by 1 is a nop. For full unrolling, a factor
1056 // of 1 makes sense because loop control can be eliminated.
1061 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, UP.AllowExpensiveTripCount,
1062 TripMultiple, LI, this, &LPM, &AC))