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/GlobalsModRef.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/CodeMetrics.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LoopPass.h"
22 #include "llvm/Analysis/ScalarEvolution.h"
23 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DiagnosticInfo.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/InstVisitor.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Transforms/Utils/UnrollLoop.h"
39 #define DEBUG_TYPE "loop-unroll"
41 static cl::opt<unsigned>
42 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
43 cl::desc("The baseline cost threshold for loop unrolling"));
45 static cl::opt<unsigned> UnrollPercentDynamicCostSavedThreshold(
46 "unroll-percent-dynamic-cost-saved-threshold", cl::init(20), cl::Hidden,
47 cl::desc("The percentage of estimated dynamic cost which must be saved by "
48 "unrolling to allow unrolling up to the max threshold."));
50 static cl::opt<unsigned> UnrollDynamicCostSavingsDiscount(
51 "unroll-dynamic-cost-savings-discount", cl::init(2000), cl::Hidden,
52 cl::desc("This is the amount discounted from the total unroll cost when "
53 "the unrolled form has a high dynamic cost savings (triggered by "
54 "the '-unroll-perecent-dynamic-cost-saved-threshold' flag)."));
56 static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
57 "unroll-max-iteration-count-to-analyze", cl::init(0), cl::Hidden,
58 cl::desc("Don't allow loop unrolling to simulate more than this number of"
59 "iterations when checking full unroll profitability"));
61 static cl::opt<unsigned>
62 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
63 cl::desc("Use this unroll count for all loops including those with "
64 "unroll_count pragma values, for testing purposes"));
67 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
68 cl::desc("Allows loops to be partially unrolled until "
69 "-unroll-threshold loop size is reached."));
72 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden,
73 cl::desc("Unroll loops with run-time trip counts"));
75 static cl::opt<unsigned>
76 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
77 cl::desc("Unrolled size limit for loops with an unroll(full) or "
78 "unroll_count pragma."));
81 class LoopUnroll : public LoopPass {
83 static char ID; // Pass ID, replacement for typeid
84 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) {
85 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
86 CurrentPercentDynamicCostSavedThreshold =
87 UnrollPercentDynamicCostSavedThreshold;
88 CurrentDynamicCostSavingsDiscount = UnrollDynamicCostSavingsDiscount;
89 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
90 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
91 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
93 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
94 UserPercentDynamicCostSavedThreshold =
95 (UnrollPercentDynamicCostSavedThreshold.getNumOccurrences() > 0);
96 UserDynamicCostSavingsDiscount =
97 (UnrollDynamicCostSavingsDiscount.getNumOccurrences() > 0);
98 UserAllowPartial = (P != -1) ||
99 (UnrollAllowPartial.getNumOccurrences() > 0);
100 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
101 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0);
103 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
106 /// A magic value for use with the Threshold parameter to indicate
107 /// that the loop unroll should be performed regardless of how much
108 /// code expansion would result.
109 static const unsigned NoThreshold = UINT_MAX;
111 // Threshold to use when optsize is specified (and there is no
112 // explicit -unroll-threshold).
113 static const unsigned OptSizeUnrollThreshold = 50;
115 // Default unroll count for loops with run-time trip count if
116 // -unroll-count is not set
117 static const unsigned UnrollRuntimeCount = 8;
119 unsigned CurrentCount;
120 unsigned CurrentThreshold;
121 unsigned CurrentPercentDynamicCostSavedThreshold;
122 unsigned CurrentDynamicCostSavingsDiscount;
123 bool CurrentAllowPartial;
126 // Flags for whether the 'current' settings are user-specified.
129 bool UserPercentDynamicCostSavedThreshold;
130 bool UserDynamicCostSavingsDiscount;
131 bool UserAllowPartial;
134 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
136 /// This transformation requires natural loop information & requires that
137 /// loop preheaders be inserted into the CFG...
139 void getAnalysisUsage(AnalysisUsage &AU) const override {
140 AU.addRequired<AssumptionCacheTracker>();
141 AU.addRequired<DominatorTreeWrapperPass>();
142 AU.addRequired<LoopInfoWrapperPass>();
143 AU.addPreserved<LoopInfoWrapperPass>();
144 AU.addRequiredID(LoopSimplifyID);
145 AU.addPreservedID(LoopSimplifyID);
146 AU.addRequiredID(LCSSAID);
147 AU.addPreservedID(LCSSAID);
148 AU.addRequired<ScalarEvolutionWrapperPass>();
149 AU.addPreserved<ScalarEvolutionWrapperPass>();
150 AU.addRequired<TargetTransformInfoWrapperPass>();
151 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
152 // If loop unroll does not preserve dom info then LCSSA pass on next
153 // loop will receive invalid dom info.
154 // For now, recreate dom info, if loop is unrolled.
155 AU.addPreserved<DominatorTreeWrapperPass>();
156 AU.addPreserved<GlobalsAAWrapperPass>();
159 // Fill in the UnrollingPreferences parameter with values from the
160 // TargetTransformationInfo.
161 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI,
162 TargetTransformInfo::UnrollingPreferences &UP) {
163 UP.Threshold = CurrentThreshold;
164 UP.PercentDynamicCostSavedThreshold =
165 CurrentPercentDynamicCostSavedThreshold;
166 UP.DynamicCostSavingsDiscount = CurrentDynamicCostSavingsDiscount;
167 UP.OptSizeThreshold = OptSizeUnrollThreshold;
168 UP.PartialThreshold = CurrentThreshold;
169 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
170 UP.Count = CurrentCount;
171 UP.MaxCount = UINT_MAX;
172 UP.Partial = CurrentAllowPartial;
173 UP.Runtime = CurrentRuntime;
174 UP.AllowExpensiveTripCount = false;
175 TTI.getUnrollingPreferences(L, UP);
178 // Select and return an unroll count based on parameters from
179 // user, unroll preferences, unroll pragmas, or a heuristic.
180 // SetExplicitly is set to true if the unroll count is is set by
181 // the user or a pragma rather than selected heuristically.
183 selectUnrollCount(const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
184 unsigned PragmaCount,
185 const TargetTransformInfo::UnrollingPreferences &UP,
186 bool &SetExplicitly);
188 // Select threshold values used to limit unrolling based on a
189 // total unrolled size. Parameters Threshold and PartialThreshold
190 // are set to the maximum unrolled size for fully and partially
191 // unrolled loops respectively.
192 void selectThresholds(const Loop *L, bool UsePragmaThreshold,
193 const TargetTransformInfo::UnrollingPreferences &UP,
194 unsigned &Threshold, unsigned &PartialThreshold,
195 unsigned &PercentDynamicCostSavedThreshold,
196 unsigned &DynamicCostSavingsDiscount) {
197 // Determine the current unrolling threshold. While this is
198 // normally set from UnrollThreshold, it is overridden to a
199 // smaller value if the current function is marked as
200 // optimize-for-size, and the unroll threshold was not user
202 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
203 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
204 PercentDynamicCostSavedThreshold =
205 UserPercentDynamicCostSavedThreshold
206 ? CurrentPercentDynamicCostSavedThreshold
207 : UP.PercentDynamicCostSavedThreshold;
208 DynamicCostSavingsDiscount = UserDynamicCostSavingsDiscount
209 ? CurrentDynamicCostSavingsDiscount
210 : UP.DynamicCostSavingsDiscount;
212 if (!UserThreshold &&
213 // FIXME: Use Function::optForSize().
214 L->getHeader()->getParent()->hasFnAttribute(
215 Attribute::OptimizeForSize)) {
216 Threshold = UP.OptSizeThreshold;
217 PartialThreshold = UP.PartialOptSizeThreshold;
219 if (UsePragmaThreshold) {
220 // If the loop has an unrolling pragma, we want to be more
221 // aggressive with unrolling limits. Set thresholds to at
222 // least the PragmaTheshold value which is larger than the
224 if (Threshold != NoThreshold)
225 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold);
226 if (PartialThreshold != NoThreshold)
228 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold);
231 bool canUnrollCompletely(Loop *L, unsigned Threshold,
232 unsigned PercentDynamicCostSavedThreshold,
233 unsigned DynamicCostSavingsDiscount,
234 uint64_t UnrolledCost, uint64_t RolledDynamicCost);
238 char LoopUnroll::ID = 0;
239 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
240 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
241 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
242 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
243 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
244 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
245 INITIALIZE_PASS_DEPENDENCY(LCSSA)
246 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
247 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
249 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial,
251 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime);
254 Pass *llvm::createSimpleLoopUnrollPass() {
255 return llvm::createLoopUnrollPass(-1, -1, 0, 0);
259 // This class is used to get an estimate of the optimization effects that we
260 // could get from complete loop unrolling. It comes from the fact that some
261 // loads might be replaced with concrete constant values and that could trigger
262 // a chain of instruction simplifications.
264 // E.g. we might have:
265 // int a[] = {0, 1, 0};
267 // for (i = 0; i < 3; i ++)
269 // If we completely unroll the loop, we would get:
270 // v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
271 // Which then will be simplified to:
272 // v = b[0]* 0 + b[1]* 1 + b[2]* 0
275 class UnrolledInstAnalyzer : private InstVisitor<UnrolledInstAnalyzer, bool> {
276 typedef InstVisitor<UnrolledInstAnalyzer, bool> Base;
277 friend class InstVisitor<UnrolledInstAnalyzer, bool>;
278 struct SimplifiedAddress {
279 Value *Base = nullptr;
280 ConstantInt *Offset = nullptr;
284 UnrolledInstAnalyzer(unsigned Iteration,
285 DenseMap<Value *, Constant *> &SimplifiedValues,
287 : SimplifiedValues(SimplifiedValues), SE(SE) {
288 IterationNumber = SE.getConstant(APInt(64, Iteration));
291 // Allow access to the initial visit method.
295 /// \brief A cache of pointer bases and constant-folded offsets corresponding
296 /// to GEP (or derived from GEP) instructions.
298 /// In order to find the base pointer one needs to perform non-trivial
299 /// traversal of the corresponding SCEV expression, so it's good to have the
301 DenseMap<Value *, SimplifiedAddress> SimplifiedAddresses;
303 /// \brief SCEV expression corresponding to number of currently simulated
305 const SCEV *IterationNumber;
307 /// \brief A Value->Constant map for keeping values that we managed to
308 /// constant-fold on the given iteration.
310 /// While we walk the loop instructions, we build up and maintain a mapping
311 /// of simplified values specific to this iteration. The idea is to propagate
312 /// any special information we have about loads that can be replaced with
313 /// constants after complete unrolling, and account for likely simplifications
315 DenseMap<Value *, Constant *> &SimplifiedValues;
319 /// \brief Try to simplify instruction \param I using its SCEV expression.
321 /// The idea is that some AddRec expressions become constants, which then
322 /// could trigger folding of other instructions. However, that only happens
323 /// for expressions whose start value is also constant, which isn't always the
324 /// case. In another common and important case the start value is just some
325 /// address (i.e. SCEVUnknown) - in this case we compute the offset and save
326 /// it along with the base address instead.
327 bool simplifyInstWithSCEV(Instruction *I) {
328 if (!SE.isSCEVable(I->getType()))
331 const SCEV *S = SE.getSCEV(I);
332 if (auto *SC = dyn_cast<SCEVConstant>(S)) {
333 SimplifiedValues[I] = SC->getValue();
337 auto *AR = dyn_cast<SCEVAddRecExpr>(S);
341 const SCEV *ValueAtIteration = AR->evaluateAtIteration(IterationNumber, SE);
342 // Check if the AddRec expression becomes a constant.
343 if (auto *SC = dyn_cast<SCEVConstant>(ValueAtIteration)) {
344 SimplifiedValues[I] = SC->getValue();
348 // Check if the offset from the base address becomes a constant.
349 auto *Base = dyn_cast<SCEVUnknown>(SE.getPointerBase(S));
353 dyn_cast<SCEVConstant>(SE.getMinusSCEV(ValueAtIteration, Base));
356 SimplifiedAddress Address;
357 Address.Base = Base->getValue();
358 Address.Offset = Offset->getValue();
359 SimplifiedAddresses[I] = Address;
363 /// Base case for the instruction visitor.
364 bool visitInstruction(Instruction &I) {
365 return simplifyInstWithSCEV(&I);
368 /// Try to simplify binary operator I.
370 /// TODO: Probably it's worth to hoist the code for estimating the
371 /// simplifications effects to a separate class, since we have a very similar
372 /// code in InlineCost already.
373 bool visitBinaryOperator(BinaryOperator &I) {
374 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
375 if (!isa<Constant>(LHS))
376 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
378 if (!isa<Constant>(RHS))
379 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
382 Value *SimpleV = nullptr;
383 const DataLayout &DL = I.getModule()->getDataLayout();
384 if (auto FI = dyn_cast<FPMathOperator>(&I))
386 SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags(), DL);
388 SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL);
390 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV))
391 SimplifiedValues[&I] = C;
395 return Base::visitBinaryOperator(I);
398 /// Try to fold load I.
399 bool visitLoad(LoadInst &I) {
400 Value *AddrOp = I.getPointerOperand();
402 auto AddressIt = SimplifiedAddresses.find(AddrOp);
403 if (AddressIt == SimplifiedAddresses.end())
405 ConstantInt *SimplifiedAddrOp = AddressIt->second.Offset;
407 auto *GV = dyn_cast<GlobalVariable>(AddressIt->second.Base);
408 // We're only interested in loads that can be completely folded to a
410 if (!GV || !GV->hasDefinitiveInitializer() || !GV->isConstant())
413 ConstantDataSequential *CDS =
414 dyn_cast<ConstantDataSequential>(GV->getInitializer());
418 // We might have a vector load from an array. FIXME: for now we just bail
419 // out in this case, but we should be able to resolve and simplify such
421 if(!CDS->isElementTypeCompatible(I.getType()))
424 int ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
425 assert(SimplifiedAddrOp->getValue().getActiveBits() < 64 &&
426 "Unexpectedly large index value.");
427 int64_t Index = SimplifiedAddrOp->getSExtValue() / ElemSize;
428 if (Index >= CDS->getNumElements()) {
429 // FIXME: For now we conservatively ignore out of bound accesses, but
430 // we're allowed to perform the optimization in this case.
434 Constant *CV = CDS->getElementAsConstant(Index);
435 assert(CV && "Constant expected.");
436 SimplifiedValues[&I] = CV;
441 bool visitCastInst(CastInst &I) {
442 // Propagate constants through casts.
443 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
445 COp = SimplifiedValues.lookup(I.getOperand(0));
448 ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) {
449 SimplifiedValues[&I] = C;
453 return Base::visitCastInst(I);
456 bool visitCmpInst(CmpInst &I) {
457 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
459 // First try to handle simplified comparisons.
460 if (!isa<Constant>(LHS))
461 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
463 if (!isa<Constant>(RHS))
464 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
467 if (!isa<Constant>(LHS) && !isa<Constant>(RHS)) {
468 auto SimplifiedLHS = SimplifiedAddresses.find(LHS);
469 if (SimplifiedLHS != SimplifiedAddresses.end()) {
470 auto SimplifiedRHS = SimplifiedAddresses.find(RHS);
471 if (SimplifiedRHS != SimplifiedAddresses.end()) {
472 SimplifiedAddress &LHSAddr = SimplifiedLHS->second;
473 SimplifiedAddress &RHSAddr = SimplifiedRHS->second;
474 if (LHSAddr.Base == RHSAddr.Base) {
475 LHS = LHSAddr.Offset;
476 RHS = RHSAddr.Offset;
482 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
483 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
484 if (Constant *C = ConstantExpr::getCompare(I.getPredicate(), CLHS, CRHS)) {
485 SimplifiedValues[&I] = C;
491 return Base::visitCmpInst(I);
498 struct EstimatedUnrollCost {
499 /// \brief The estimated cost after unrolling.
502 /// \brief The estimated dynamic cost of executing the instructions in the
504 int RolledDynamicCost;
508 /// \brief Figure out if the loop is worth full unrolling.
510 /// Complete loop unrolling can make some loads constant, and we need to know
511 /// if that would expose any further optimization opportunities. This routine
512 /// estimates this optimization. It computes cost of unrolled loop
513 /// (UnrolledCost) and dynamic cost of the original loop (RolledDynamicCost). By
514 /// dynamic cost we mean that we won't count costs of blocks that are known not
515 /// to be executed (i.e. if we have a branch in the loop and we know that at the
516 /// given iteration its condition would be resolved to true, we won't add up the
517 /// cost of the 'false'-block).
518 /// \returns Optional value, holding the RolledDynamicCost and UnrolledCost. If
519 /// the analysis failed (no benefits expected from the unrolling, or the loop is
520 /// too big to analyze), the returned value is None.
521 static Optional<EstimatedUnrollCost>
522 analyzeLoopUnrollCost(const Loop *L, unsigned TripCount, DominatorTree &DT,
523 ScalarEvolution &SE, const TargetTransformInfo &TTI,
524 int MaxUnrolledLoopSize) {
525 // We want to be able to scale offsets by the trip count and add more offsets
526 // to them without checking for overflows, and we already don't want to
527 // analyze *massive* trip counts, so we force the max to be reasonably small.
528 assert(UnrollMaxIterationsCountToAnalyze < (INT_MAX / 2) &&
529 "The unroll iterations max is too large!");
531 // Don't simulate loops with a big or unknown tripcount
532 if (!UnrollMaxIterationsCountToAnalyze || !TripCount ||
533 TripCount > UnrollMaxIterationsCountToAnalyze)
536 SmallSetVector<BasicBlock *, 16> BBWorklist;
537 DenseMap<Value *, Constant *> SimplifiedValues;
538 SmallVector<std::pair<Value *, Constant *>, 4> SimplifiedInputValues;
540 // The estimated cost of the unrolled form of the loop. We try to estimate
541 // this by simplifying as much as we can while computing the estimate.
542 int UnrolledCost = 0;
543 // We also track the estimated dynamic (that is, actually executed) cost in
544 // the rolled form. This helps identify cases when the savings from unrolling
545 // aren't just exposing dead control flows, but actual reduced dynamic
546 // instructions due to the simplifications which we expect to occur after
548 int RolledDynamicCost = 0;
550 // Ensure that we don't violate the loop structure invariants relied on by
552 assert(L->isLoopSimplifyForm() && "Must put loop into normal form first.");
553 assert(L->isLCSSAForm(DT) &&
554 "Must have loops in LCSSA form to track live-out values.");
556 DEBUG(dbgs() << "Starting LoopUnroll profitability analysis...\n");
558 // Simulate execution of each iteration of the loop counting instructions,
559 // which would be simplified.
560 // Since the same load will take different values on different iterations,
561 // we literally have to go through all loop's iterations.
562 for (unsigned Iteration = 0; Iteration < TripCount; ++Iteration) {
563 DEBUG(dbgs() << " Analyzing iteration " << Iteration << "\n");
565 // Prepare for the iteration by collecting any simplified entry or backedge
567 for (Instruction &I : *L->getHeader()) {
568 auto *PHI = dyn_cast<PHINode>(&I);
572 // The loop header PHI nodes must have exactly two input: one from the
573 // loop preheader and one from the loop latch.
575 PHI->getNumIncomingValues() == 2 &&
576 "Must have an incoming value only for the preheader and the latch.");
578 Value *V = PHI->getIncomingValueForBlock(
579 Iteration == 0 ? L->getLoopPreheader() : L->getLoopLatch());
580 Constant *C = dyn_cast<Constant>(V);
581 if (Iteration != 0 && !C)
582 C = SimplifiedValues.lookup(V);
584 SimplifiedInputValues.push_back({PHI, C});
587 // Now clear and re-populate the map for the next iteration.
588 SimplifiedValues.clear();
589 while (!SimplifiedInputValues.empty())
590 SimplifiedValues.insert(SimplifiedInputValues.pop_back_val());
592 UnrolledInstAnalyzer Analyzer(Iteration, SimplifiedValues, SE);
595 BBWorklist.insert(L->getHeader());
596 // Note that we *must not* cache the size, this loop grows the worklist.
597 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
598 BasicBlock *BB = BBWorklist[Idx];
600 // Visit all instructions in the given basic block and try to simplify
601 // it. We don't change the actual IR, just count optimization
603 for (Instruction &I : *BB) {
604 int InstCost = TTI.getUserCost(&I);
606 // Visit the instruction to analyze its loop cost after unrolling,
607 // and if the visitor returns false, include this instruction in the
609 if (!Analyzer.visit(I))
610 UnrolledCost += InstCost;
612 DEBUG(dbgs() << " " << I
613 << " would be simplified if loop is unrolled.\n");
617 // Also track this instructions expected cost when executing the rolled
619 RolledDynamicCost += InstCost;
621 // If unrolled body turns out to be too big, bail out.
622 if (UnrolledCost > MaxUnrolledLoopSize) {
623 DEBUG(dbgs() << " Exceeded threshold.. exiting.\n"
624 << " UnrolledCost: " << UnrolledCost
625 << ", MaxUnrolledLoopSize: " << MaxUnrolledLoopSize
631 TerminatorInst *TI = BB->getTerminator();
633 // Add in the live successors by first checking whether we have terminator
634 // that may be simplified based on the values simplified by this call.
635 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
636 if (BI->isConditional()) {
637 if (Constant *SimpleCond =
638 SimplifiedValues.lookup(BI->getCondition())) {
639 BasicBlock *Succ = nullptr;
640 // Just take the first successor if condition is undef
641 if (isa<UndefValue>(SimpleCond))
642 Succ = BI->getSuccessor(0);
644 Succ = BI->getSuccessor(
645 cast<ConstantInt>(SimpleCond)->isZero() ? 1 : 0);
646 if (L->contains(Succ))
647 BBWorklist.insert(Succ);
651 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
652 if (Constant *SimpleCond =
653 SimplifiedValues.lookup(SI->getCondition())) {
654 BasicBlock *Succ = nullptr;
655 // Just take the first successor if condition is undef
656 if (isa<UndefValue>(SimpleCond))
657 Succ = SI->getSuccessor(0);
659 Succ = SI->findCaseValue(cast<ConstantInt>(SimpleCond))
661 if (L->contains(Succ))
662 BBWorklist.insert(Succ);
667 // Add BB's successors to the worklist.
668 for (BasicBlock *Succ : successors(BB))
669 if (L->contains(Succ))
670 BBWorklist.insert(Succ);
673 // If we found no optimization opportunities on the first iteration, we
674 // won't find them on later ones too.
675 if (UnrolledCost == RolledDynamicCost) {
676 DEBUG(dbgs() << " No opportunities found.. exiting.\n"
677 << " UnrolledCost: " << UnrolledCost << "\n");
681 DEBUG(dbgs() << "Analysis finished:\n"
682 << "UnrolledCost: " << UnrolledCost << ", "
683 << "RolledDynamicCost: " << RolledDynamicCost << "\n");
684 return {{UnrolledCost, RolledDynamicCost}};
687 /// ApproximateLoopSize - Approximate the size of the loop.
688 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
689 bool &NotDuplicatable,
690 const TargetTransformInfo &TTI,
691 AssumptionCache *AC) {
692 SmallPtrSet<const Value *, 32> EphValues;
693 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
696 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
698 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
699 NumCalls = Metrics.NumInlineCandidates;
700 NotDuplicatable = Metrics.notDuplicatable;
702 unsigned LoopSize = Metrics.NumInsts;
704 // Don't allow an estimate of size zero. This would allows unrolling of loops
705 // with huge iteration counts, which is a compile time problem even if it's
706 // not a problem for code quality. Also, the code using this size may assume
707 // that each loop has at least three instructions (likely a conditional
708 // branch, a comparison feeding that branch, and some kind of loop increment
709 // feeding that comparison instruction).
710 LoopSize = std::max(LoopSize, 3u);
715 // Returns the loop hint metadata node with the given name (for example,
716 // "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
718 static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
719 if (MDNode *LoopID = L->getLoopID())
720 return GetUnrollMetadata(LoopID, Name);
724 // Returns true if the loop has an unroll(full) pragma.
725 static bool HasUnrollFullPragma(const Loop *L) {
726 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
729 // Returns true if the loop has an unroll(enable) pragma. This metadata is used
730 // for both "#pragma unroll" and "#pragma clang loop unroll(enable)" directives.
731 static bool HasUnrollEnablePragma(const Loop *L) {
732 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.enable");
735 // Returns true if the loop has an unroll(disable) pragma.
736 static bool HasUnrollDisablePragma(const Loop *L) {
737 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
740 // Returns true if the loop has an runtime unroll(disable) pragma.
741 static bool HasRuntimeUnrollDisablePragma(const Loop *L) {
742 return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.runtime.disable");
745 // If loop has an unroll_count pragma return the (necessarily
746 // positive) value from the pragma. Otherwise return 0.
747 static unsigned UnrollCountPragmaValue(const Loop *L) {
748 MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
750 assert(MD->getNumOperands() == 2 &&
751 "Unroll count hint metadata should have two operands.");
753 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
754 assert(Count >= 1 && "Unroll count must be positive.");
760 // Remove existing unroll metadata and add unroll disable metadata to
761 // indicate the loop has already been unrolled. This prevents a loop
762 // from being unrolled more than is directed by a pragma if the loop
763 // unrolling pass is run more than once (which it generally is).
764 static void SetLoopAlreadyUnrolled(Loop *L) {
765 MDNode *LoopID = L->getLoopID();
768 // First remove any existing loop unrolling metadata.
769 SmallVector<Metadata *, 4> MDs;
770 // Reserve first location for self reference to the LoopID metadata node.
771 MDs.push_back(nullptr);
772 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
773 bool IsUnrollMetadata = false;
774 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
776 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
777 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
779 if (!IsUnrollMetadata)
780 MDs.push_back(LoopID->getOperand(i));
783 // Add unroll(disable) metadata to disable future unrolling.
784 LLVMContext &Context = L->getHeader()->getContext();
785 SmallVector<Metadata *, 1> DisableOperands;
786 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
787 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
788 MDs.push_back(DisableNode);
790 MDNode *NewLoopID = MDNode::get(Context, MDs);
791 // Set operand 0 to refer to the loop id itself.
792 NewLoopID->replaceOperandWith(0, NewLoopID);
793 L->setLoopID(NewLoopID);
796 bool LoopUnroll::canUnrollCompletely(Loop *L, unsigned Threshold,
797 unsigned PercentDynamicCostSavedThreshold,
798 unsigned DynamicCostSavingsDiscount,
799 uint64_t UnrolledCost,
800 uint64_t RolledDynamicCost) {
802 if (Threshold == NoThreshold) {
803 DEBUG(dbgs() << " Can fully unroll, because no threshold is set.\n");
807 if (UnrolledCost <= Threshold) {
808 DEBUG(dbgs() << " Can fully unroll, because unrolled cost: "
809 << UnrolledCost << "<" << Threshold << "\n");
813 assert(UnrolledCost && "UnrolledCost can't be 0 at this point.");
814 assert(RolledDynamicCost >= UnrolledCost &&
815 "Cannot have a higher unrolled cost than a rolled cost!");
817 // Compute the percentage of the dynamic cost in the rolled form that is
818 // saved when unrolled. If unrolling dramatically reduces the estimated
819 // dynamic cost of the loop, we use a higher threshold to allow more
821 unsigned PercentDynamicCostSaved =
822 (uint64_t)(RolledDynamicCost - UnrolledCost) * 100ull / RolledDynamicCost;
824 if (PercentDynamicCostSaved >= PercentDynamicCostSavedThreshold &&
825 (int64_t)UnrolledCost - (int64_t)DynamicCostSavingsDiscount <=
826 (int64_t)Threshold) {
827 DEBUG(dbgs() << " Can fully unroll, because unrolling will reduce the "
828 "expected dynamic cost by " << PercentDynamicCostSaved
829 << "% (threshold: " << PercentDynamicCostSavedThreshold
831 << " and the unrolled cost (" << UnrolledCost
832 << ") is less than the max threshold ("
833 << DynamicCostSavingsDiscount << ").\n");
837 DEBUG(dbgs() << " Too large to fully unroll:\n");
838 DEBUG(dbgs() << " Threshold: " << Threshold << "\n");
839 DEBUG(dbgs() << " Max threshold: " << DynamicCostSavingsDiscount << "\n");
840 DEBUG(dbgs() << " Percent cost saved threshold: "
841 << PercentDynamicCostSavedThreshold << "%\n");
842 DEBUG(dbgs() << " Unrolled cost: " << UnrolledCost << "\n");
843 DEBUG(dbgs() << " Rolled dynamic cost: " << RolledDynamicCost << "\n");
844 DEBUG(dbgs() << " Percent cost saved: " << PercentDynamicCostSaved
849 unsigned LoopUnroll::selectUnrollCount(
850 const Loop *L, unsigned TripCount, bool PragmaFullUnroll,
851 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
852 bool &SetExplicitly) {
853 SetExplicitly = true;
855 // User-specified count (either as a command-line option or
856 // constructor parameter) has highest precedence.
857 unsigned Count = UserCount ? CurrentCount : 0;
859 // If there is no user-specified count, unroll pragmas have the next
860 // highest precedence.
864 } else if (PragmaFullUnroll) {
873 SetExplicitly = false;
875 // Runtime trip count.
876 Count = UnrollRuntimeCount;
878 // Conservative heuristic: if we know the trip count, see if we can
879 // completely unroll (subject to the threshold, checked below); otherwise
880 // try to find greatest modulo of the trip count which is still under
884 if (TripCount && Count > TripCount)
889 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
890 if (skipOptnoneFunction(L))
893 Function &F = *L->getHeader()->getParent();
895 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
896 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
897 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
898 const TargetTransformInfo &TTI =
899 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
900 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
902 BasicBlock *Header = L->getHeader();
903 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
904 << "] Loop %" << Header->getName() << "\n");
906 if (HasUnrollDisablePragma(L)) {
909 bool PragmaFullUnroll = HasUnrollFullPragma(L);
910 bool PragmaEnableUnroll = HasUnrollEnablePragma(L);
911 unsigned PragmaCount = UnrollCountPragmaValue(L);
912 bool HasPragma = PragmaFullUnroll || PragmaEnableUnroll || PragmaCount > 0;
914 TargetTransformInfo::UnrollingPreferences UP;
915 getUnrollingPreferences(L, TTI, UP);
917 // Find trip count and trip multiple if count is not available
918 unsigned TripCount = 0;
919 unsigned TripMultiple = 1;
920 // If there are multiple exiting blocks but one of them is the latch, use the
921 // latch for the trip count estimation. Otherwise insist on a single exiting
922 // block for the trip count estimation.
923 BasicBlock *ExitingBlock = L->getLoopLatch();
924 if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
925 ExitingBlock = L->getExitingBlock();
927 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
928 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
931 // Select an initial unroll count. This may be reduced later based
932 // on size thresholds.
933 bool CountSetExplicitly;
934 unsigned Count = selectUnrollCount(L, TripCount, PragmaFullUnroll,
935 PragmaCount, UP, CountSetExplicitly);
937 unsigned NumInlineCandidates;
938 bool notDuplicatable;
940 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
941 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
943 // When computing the unrolled size, note that the conditional branch on the
944 // backedge and the comparison feeding it are not replicated like the rest of
945 // the loop body (which is why 2 is subtracted).
946 uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
947 if (notDuplicatable) {
948 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
949 << " instructions.\n");
952 if (NumInlineCandidates != 0) {
953 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
957 unsigned Threshold, PartialThreshold;
958 unsigned PercentDynamicCostSavedThreshold;
959 unsigned DynamicCostSavingsDiscount;
960 // Only use the high pragma threshold when we have a target unroll factor such
961 // as with "#pragma unroll N" or a pragma indicating full unrolling and the
962 // trip count is known. Otherwise we rely on the standard threshold to
963 // heuristically select a reasonable unroll count.
964 bool UsePragmaThreshold =
966 ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount != 0);
968 selectThresholds(L, UsePragmaThreshold, UP, Threshold, PartialThreshold,
969 PercentDynamicCostSavedThreshold,
970 DynamicCostSavingsDiscount);
972 // Given Count, TripCount and thresholds determine the type of
973 // unrolling which is to be performed.
974 enum { Full = 0, Partial = 1, Runtime = 2 };
976 if (TripCount && Count == TripCount) {
978 // If the loop is really small, we don't need to run an expensive analysis.
979 if (canUnrollCompletely(L, Threshold, 100, DynamicCostSavingsDiscount,
980 UnrolledSize, UnrolledSize)) {
983 // The loop isn't that small, but we still can fully unroll it if that
984 // helps to remove a significant number of instructions.
985 // To check that, run additional analysis on the loop.
986 if (Optional<EstimatedUnrollCost> Cost =
987 analyzeLoopUnrollCost(L, TripCount, DT, *SE, TTI,
988 Threshold + DynamicCostSavingsDiscount))
989 if (canUnrollCompletely(L, Threshold, PercentDynamicCostSavedThreshold,
990 DynamicCostSavingsDiscount, Cost->UnrolledCost,
991 Cost->RolledDynamicCost)) {
995 } else if (TripCount && Count < TripCount) {
1001 // Reduce count based on the type of unrolling and the threshold values.
1002 unsigned OriginalCount = Count;
1003 bool AllowRuntime = PragmaEnableUnroll || (PragmaCount > 0) ||
1004 (UserRuntime ? CurrentRuntime : UP.Runtime);
1005 // Don't unroll a runtime trip count loop with unroll full pragma.
1006 if (HasRuntimeUnrollDisablePragma(L) || PragmaFullUnroll) {
1007 AllowRuntime = false;
1009 if (Unrolling == Partial) {
1010 bool AllowPartial = PragmaEnableUnroll ||
1011 (UserAllowPartial ? CurrentAllowPartial : UP.Partial);
1012 if (!AllowPartial && !CountSetExplicitly) {
1013 DEBUG(dbgs() << " will not try to unroll partially because "
1014 << "-unroll-allow-partial not given\n");
1017 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
1018 // Reduce unroll count to be modulo of TripCount for partial unrolling.
1019 Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
1020 while (Count != 0 && TripCount % Count != 0)
1023 } else if (Unrolling == Runtime) {
1024 if (!AllowRuntime && !CountSetExplicitly) {
1025 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
1026 << "-unroll-runtime not given\n");
1029 // Reduce unroll count to be the largest power-of-two factor of
1030 // the original count which satisfies the threshold limit.
1031 while (Count != 0 && UnrolledSize > PartialThreshold) {
1033 UnrolledSize = (LoopSize-2) * Count + 2;
1035 if (Count > UP.MaxCount)
1036 Count = UP.MaxCount;
1037 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
1041 if (PragmaCount != 0)
1042 // If loop has an unroll count pragma mark loop as unrolled to prevent
1043 // unrolling beyond that requested by the pragma.
1044 SetLoopAlreadyUnrolled(L);
1046 // Emit optimization remarks if we are unable to unroll the loop
1047 // as directed by a pragma.
1048 DebugLoc LoopLoc = L->getStartLoc();
1049 Function *F = Header->getParent();
1050 LLVMContext &Ctx = F->getContext();
1051 if ((PragmaCount > 0) && Count != OriginalCount) {
1052 emitOptimizationRemarkMissed(
1053 Ctx, DEBUG_TYPE, *F, LoopLoc,
1054 "Unable to unroll loop the number of times directed by "
1055 "unroll_count pragma because unrolled size is too large.");
1056 } else if (PragmaFullUnroll && !TripCount) {
1057 emitOptimizationRemarkMissed(
1058 Ctx, DEBUG_TYPE, *F, LoopLoc,
1059 "Unable to fully unroll loop as directed by unroll(full) pragma "
1060 "because loop has a runtime trip count.");
1061 } else if (PragmaEnableUnroll && Count != TripCount && Count < 2) {
1062 emitOptimizationRemarkMissed(
1063 Ctx, DEBUG_TYPE, *F, LoopLoc,
1064 "Unable to unroll loop as directed by unroll(enable) pragma because "
1065 "unrolled size is too large.");
1066 } else if ((PragmaFullUnroll || PragmaEnableUnroll) && TripCount &&
1067 Count != TripCount) {
1068 emitOptimizationRemarkMissed(
1069 Ctx, DEBUG_TYPE, *F, LoopLoc,
1070 "Unable to fully unroll loop as directed by unroll pragma because "
1071 "unrolled size is too large.");
1075 if (Unrolling != Full && Count < 2) {
1076 // Partial unrolling by 1 is a nop. For full unrolling, a factor
1077 // of 1 makes sense because loop control can be eliminated.
1082 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, UP.AllowExpensiveTripCount,
1083 TripMultiple, LI, this, &LPM, &AC))