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/CodeMetrics.h"
17 #include "llvm/Analysis/LoopPass.h"
18 #include "llvm/Analysis/ScalarEvolution.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/DiagnosticInfo.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/Metadata.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Transforms/Utils/UnrollLoop.h"
33 #define DEBUG_TYPE "loop-unroll"
35 static cl::opt<unsigned>
36 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
37 cl::desc("The cut-off point for automatic loop unrolling"));
39 static cl::opt<unsigned>
40 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
41 cl::desc("Use this unroll count for all loops including those with "
42 "unroll_count pragma values, for testing purposes"));
45 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
46 cl::desc("Allows loops to be partially unrolled until "
47 "-unroll-threshold loop size is reached."));
50 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden,
51 cl::desc("Unroll loops with run-time trip counts"));
53 static cl::opt<unsigned>
54 PragmaUnrollThreshold("pragma-unroll-threshold", cl::init(16 * 1024), cl::Hidden,
55 cl::desc("Unrolled size limit for loops with an unroll(enable) or "
56 "unroll_count pragma."));
59 class LoopUnroll : public LoopPass {
61 static char ID; // Pass ID, replacement for typeid
62 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) {
63 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
64 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
65 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
66 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
68 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
69 UserAllowPartial = (P != -1) ||
70 (UnrollAllowPartial.getNumOccurrences() > 0);
71 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
72 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0);
74 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
77 /// A magic value for use with the Threshold parameter to indicate
78 /// that the loop unroll should be performed regardless of how much
79 /// code expansion would result.
80 static const unsigned NoThreshold = UINT_MAX;
82 // Threshold to use when optsize is specified (and there is no
83 // explicit -unroll-threshold).
84 static const unsigned OptSizeUnrollThreshold = 50;
86 // Default unroll count for loops with run-time trip count if
87 // -unroll-count is not set
88 static const unsigned UnrollRuntimeCount = 8;
90 unsigned CurrentCount;
91 unsigned CurrentThreshold;
92 bool CurrentAllowPartial;
94 bool UserCount; // CurrentCount is user-specified.
95 bool UserThreshold; // CurrentThreshold is user-specified.
96 bool UserAllowPartial; // CurrentAllowPartial is user-specified.
97 bool UserRuntime; // CurrentRuntime is user-specified.
99 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
101 /// This transformation requires natural loop information & requires that
102 /// loop preheaders be inserted into the CFG...
104 void getAnalysisUsage(AnalysisUsage &AU) const override {
105 AU.addRequired<LoopInfo>();
106 AU.addPreserved<LoopInfo>();
107 AU.addRequiredID(LoopSimplifyID);
108 AU.addPreservedID(LoopSimplifyID);
109 AU.addRequiredID(LCSSAID);
110 AU.addPreservedID(LCSSAID);
111 AU.addRequired<ScalarEvolution>();
112 AU.addPreserved<ScalarEvolution>();
113 AU.addRequired<TargetTransformInfo>();
114 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
115 // If loop unroll does not preserve dom info then LCSSA pass on next
116 // loop will receive invalid dom info.
117 // For now, recreate dom info, if loop is unrolled.
118 AU.addPreserved<DominatorTreeWrapperPass>();
121 // Fill in the UnrollingPreferences parameter with values from the
122 // TargetTransformationInfo.
123 void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI,
124 TargetTransformInfo::UnrollingPreferences &UP) {
125 UP.Threshold = CurrentThreshold;
126 UP.OptSizeThreshold = OptSizeUnrollThreshold;
127 UP.PartialThreshold = CurrentThreshold;
128 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
129 UP.Count = CurrentCount;
130 UP.MaxCount = UINT_MAX;
131 UP.Partial = CurrentAllowPartial;
132 UP.Runtime = CurrentRuntime;
133 TTI.getUnrollingPreferences(L, UP);
136 // Select and return an unroll count based on parameters from
137 // user, unroll preferences, unroll pragmas, or a heuristic.
138 // SetExplicitly is set to true if the unroll count is is set by
139 // the user or a pragma rather than selected heuristically.
141 selectUnrollCount(const Loop *L, unsigned TripCount, bool HasEnablePragma,
142 unsigned PragmaCount,
143 const TargetTransformInfo::UnrollingPreferences &UP,
144 bool &SetExplicitly);
147 // Select threshold values used to limit unrolling based on a
148 // total unrolled size. Parameters Threshold and PartialThreshold
149 // are set to the maximum unrolled size for fully and partially
150 // unrolled loops respectively.
151 void selectThresholds(const Loop *L, bool HasPragma,
152 const TargetTransformInfo::UnrollingPreferences &UP,
153 unsigned &Threshold, unsigned &PartialThreshold) {
154 // Determine the current unrolling threshold. While this is
155 // normally set from UnrollThreshold, it is overridden to a
156 // smaller value if the current function is marked as
157 // optimize-for-size, and the unroll threshold was not user
159 Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
160 PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
161 if (!UserThreshold &&
162 L->getHeader()->getParent()->getAttributes().
163 hasAttribute(AttributeSet::FunctionIndex,
164 Attribute::OptimizeForSize)) {
165 Threshold = UP.OptSizeThreshold;
166 PartialThreshold = UP.PartialOptSizeThreshold;
169 // If the loop has an unrolling pragma, we want to be more
170 // aggressive with unrolling limits. Set thresholds to at
171 // least the PragmaTheshold value which is larger than the
173 if (Threshold != NoThreshold)
174 Threshold = std::max<unsigned>(Threshold, PragmaUnrollThreshold);
175 if (PartialThreshold != NoThreshold)
177 std::max<unsigned>(PartialThreshold, PragmaUnrollThreshold);
183 char LoopUnroll::ID = 0;
184 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
185 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
186 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
187 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
188 INITIALIZE_PASS_DEPENDENCY(LCSSA)
189 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
190 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
192 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial,
194 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime);
197 Pass *llvm::createSimpleLoopUnrollPass() {
198 return llvm::createLoopUnrollPass(-1, -1, 0, 0);
201 /// ApproximateLoopSize - Approximate the size of the loop.
202 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
203 bool &NotDuplicatable,
204 const TargetTransformInfo &TTI) {
206 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
208 Metrics.analyzeBasicBlock(*I, TTI);
209 NumCalls = Metrics.NumInlineCandidates;
210 NotDuplicatable = Metrics.notDuplicatable;
212 unsigned LoopSize = Metrics.NumInsts;
214 // Don't allow an estimate of size zero. This would allows unrolling of loops
215 // with huge iteration counts, which is a compile time problem even if it's
216 // not a problem for code quality.
217 if (LoopSize == 0) LoopSize = 1;
222 // Returns the value associated with the given metadata node name (for
223 // example, "llvm.loopunroll.count"). If no such named metadata node
224 // exists, then nullptr is returned.
225 static const ConstantInt *GetUnrollMetadataValue(const Loop *L,
227 MDNode *LoopID = L->getLoopID();
228 if (!LoopID) return nullptr;
230 // First operand should refer to the loop id itself.
231 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
232 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
234 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
235 const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
238 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
241 if (Name.equals(S->getString())) {
242 assert(MD->getNumOperands() == 2 &&
243 "Unroll hint metadata should have two operands.");
244 return cast<ConstantInt>(MD->getOperand(1));
250 // Returns true if the loop has an unroll(enable) pragma.
251 static bool HasUnrollEnablePragma(const Loop *L) {
252 const ConstantInt *EnableValue =
253 GetUnrollMetadataValue(L, "llvm.loopunroll.enable");
254 return (EnableValue && EnableValue->getZExtValue());
258 // Returns true if the loop has an unroll(disable) pragma.
259 static bool HasUnrollDisablePragma(const Loop *L) {
260 const ConstantInt *EnableValue =
261 GetUnrollMetadataValue(L, "llvm.loopunroll.enable");
262 return (EnableValue && !EnableValue->getZExtValue());
266 // If loop has an unroll_count pragma return the (necessarily
267 // positive) value from the pragma. Otherwise return 0.
268 static unsigned UnrollCountPragmaValue(const Loop *L) {
269 const ConstantInt *CountValue =
270 GetUnrollMetadataValue(L, "llvm.loopunroll.count");
272 unsigned Count = CountValue->getZExtValue();
273 assert(Count >= 1 && "Unroll count must be positive.");
279 unsigned LoopUnroll::selectUnrollCount(
280 const Loop *L, unsigned TripCount, bool HasEnablePragma,
281 unsigned PragmaCount, const TargetTransformInfo::UnrollingPreferences &UP,
282 bool &SetExplicitly) {
283 SetExplicitly = true;
285 // User-specified count (either as a command-line option or
286 // constructor parameter) has highest precedence.
287 unsigned Count = UserCount ? CurrentCount : 0;
289 // If there is no user-specified count, unroll pragmas have the next
290 // highest precendence.
294 } else if (HasEnablePragma) {
295 // unroll(enable) pragma without an unroll_count pragma
296 // indicates to unroll loop fully.
305 SetExplicitly = false;
307 // Runtime trip count.
308 Count = UnrollRuntimeCount;
310 // Conservative heuristic: if we know the trip count, see if we can
311 // completely unroll (subject to the threshold, checked below); otherwise
312 // try to find greatest modulo of the trip count which is still under
316 if (TripCount && Count > TripCount)
321 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
322 if (skipOptnoneFunction(L))
325 LoopInfo *LI = &getAnalysis<LoopInfo>();
326 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
327 const TargetTransformInfo &TTI = getAnalysis<TargetTransformInfo>();
329 BasicBlock *Header = L->getHeader();
330 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
331 << "] Loop %" << Header->getName() << "\n");
333 if (HasUnrollDisablePragma(L)) {
336 bool HasEnablePragma = HasUnrollEnablePragma(L);
337 unsigned PragmaCount = UnrollCountPragmaValue(L);
338 bool HasPragma = HasEnablePragma || PragmaCount > 0;
340 TargetTransformInfo::UnrollingPreferences UP;
341 getUnrollingPreferences(L, TTI, UP);
343 // Find trip count and trip multiple if count is not available
344 unsigned TripCount = 0;
345 unsigned TripMultiple = 1;
346 // Find "latch trip count". UnrollLoop assumes that control cannot exit
347 // via the loop latch on any iteration prior to TripCount. The loop may exit
348 // early via an earlier branch.
349 BasicBlock *LatchBlock = L->getLoopLatch();
351 TripCount = SE->getSmallConstantTripCount(L, LatchBlock);
352 TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock);
355 // Select an initial unroll count. This may be reduced later based
356 // on size thresholds.
357 bool CountSetExplicitly;
358 unsigned Count = selectUnrollCount(L, TripCount, HasEnablePragma, PragmaCount,
359 UP, CountSetExplicitly);
361 unsigned NumInlineCandidates;
362 bool notDuplicatable;
364 ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI);
365 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
366 uint64_t UnrolledSize = (uint64_t)LoopSize * Count;
367 if (notDuplicatable) {
368 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
369 << " instructions.\n");
372 if (NumInlineCandidates != 0) {
373 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
377 unsigned Threshold, PartialThreshold;
378 selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold);
380 // Given Count, TripCount and thresholds determine the type of
381 // unrolling which is to be performed.
382 enum { Full = 0, Partial = 1, Runtime = 2 };
384 if (TripCount && Count == TripCount) {
385 if (Threshold != NoThreshold && UnrolledSize > Threshold) {
386 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
387 << " because size: " << UnrolledSize << ">" << Threshold
393 } else if (TripCount && Count < TripCount) {
399 // Reduce count based on the type of unrolling and the threshold values.
400 unsigned OriginalCount = Count;
401 bool AllowRuntime = UserRuntime ? CurrentRuntime : UP.Runtime;
402 if (Unrolling == Partial) {
403 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
404 if (!AllowPartial && !CountSetExplicitly) {
405 DEBUG(dbgs() << " will not try to unroll partially because "
406 << "-unroll-allow-partial not given\n");
409 if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
410 // Reduce unroll count to be modulo of TripCount for partial unrolling.
411 Count = PartialThreshold / LoopSize;
412 while (Count != 0 && TripCount % Count != 0)
415 } else if (Unrolling == Runtime) {
416 if (!AllowRuntime && !CountSetExplicitly) {
417 DEBUG(dbgs() << " will not try to unroll loop with runtime trip count "
418 << "-unroll-runtime not given\n");
421 // Reduce unroll count to be the largest power-of-two factor of
422 // the original count which satisfies the threshold limit.
423 while (Count != 0 && UnrolledSize > PartialThreshold) {
425 UnrolledSize = LoopSize * Count;
427 if (Count > UP.MaxCount)
429 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
433 // Emit optimization remarks if we are unable to unroll the loop
434 // as directed by a pragma.
435 DebugLoc LoopLoc = L->getStartLoc();
436 Function *F = Header->getParent();
437 LLVMContext &Ctx = F->getContext();
438 if (HasEnablePragma && PragmaCount == 0) {
439 if (TripCount && Count != TripCount) {
440 emitOptimizationRemarkMissed(
441 Ctx, DEBUG_TYPE, *F, LoopLoc,
442 "Unable to fully unroll loop as directed by unroll(enable) pragma "
443 "because unrolled size is too large.");
444 } else if (!TripCount) {
445 emitOptimizationRemarkMissed(
446 Ctx, DEBUG_TYPE, *F, LoopLoc,
447 "Unable to fully unroll loop as directed by unroll(enable) pragma "
448 "because loop has a runtime trip count.");
450 } else if (PragmaCount > 0 && Count != OriginalCount) {
451 emitOptimizationRemarkMissed(
452 Ctx, DEBUG_TYPE, *F, LoopLoc,
453 "Unable to unroll loop the number of times directed by "
454 "unroll_count pragma because unrolled size is too large.");
458 if (Unrolling != Full && Count < 2) {
459 // Partial unrolling by 1 is a nop. For full unrolling, a factor
460 // of 1 makes sense because loop control can be eliminated.
465 if (!UnrollLoop(L, Count, TripCount, AllowRuntime, TripMultiple, LI, this, &LPM))