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 #define DEBUG_TYPE "loop-unroll"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/Analysis/CodeMetrics.h"
18 #include "llvm/Analysis/LoopPass.h"
19 #include "llvm/Analysis/ScalarEvolution.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include "llvm/Transforms/Utils/UnrollLoop.h"
32 static cl::opt<unsigned>
33 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
34 cl::desc("The cut-off point for automatic loop unrolling"));
36 static cl::opt<unsigned>
37 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
38 cl::desc("Use this unroll count for all loops, for testing purposes"));
41 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
42 cl::desc("Allows loops to be partially unrolled until "
43 "-unroll-threshold loop size is reached."));
46 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden,
47 cl::desc("Unroll loops with run-time trip counts"));
50 class LoopUnroll : public LoopPass {
52 static char ID; // Pass ID, replacement for typeid
53 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) {
54 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
55 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
56 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
57 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
59 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
60 UserAllowPartial = (P != -1) ||
61 (UnrollAllowPartial.getNumOccurrences() > 0);
62 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
63 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0);
65 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
68 /// A magic value for use with the Threshold parameter to indicate
69 /// that the loop unroll should be performed regardless of how much
70 /// code expansion would result.
71 static const unsigned NoThreshold = UINT_MAX;
73 // Threshold to use when optsize is specified (and there is no
74 // explicit -unroll-threshold).
75 static const unsigned OptSizeUnrollThreshold = 50;
77 // Default unroll count for loops with run-time trip count if
78 // -unroll-count is not set
79 static const unsigned UnrollRuntimeCount = 8;
81 unsigned CurrentCount;
82 unsigned CurrentThreshold;
83 bool CurrentAllowPartial;
85 bool UserCount; // CurrentCount is user-specified.
86 bool UserThreshold; // CurrentThreshold is user-specified.
87 bool UserAllowPartial; // CurrentAllowPartial is user-specified.
88 bool UserRuntime; // CurrentRuntime is user-specified.
90 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
92 /// This transformation requires natural loop information & requires that
93 /// loop preheaders be inserted into the CFG...
95 void getAnalysisUsage(AnalysisUsage &AU) const override {
96 AU.addRequired<LoopInfo>();
97 AU.addPreserved<LoopInfo>();
98 AU.addRequiredID(LoopSimplifyID);
99 AU.addPreservedID(LoopSimplifyID);
100 AU.addRequiredID(LCSSAID);
101 AU.addPreservedID(LCSSAID);
102 AU.addRequired<ScalarEvolution>();
103 AU.addPreserved<ScalarEvolution>();
104 AU.addRequired<TargetTransformInfo>();
105 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
106 // If loop unroll does not preserve dom info then LCSSA pass on next
107 // loop will receive invalid dom info.
108 // For now, recreate dom info, if loop is unrolled.
109 AU.addPreserved<DominatorTreeWrapperPass>();
114 char LoopUnroll::ID = 0;
115 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
116 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
117 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
118 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
119 INITIALIZE_PASS_DEPENDENCY(LCSSA)
120 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
121 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
123 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial,
125 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime);
128 Pass *llvm::createSimpleLoopUnrollPass() {
129 return llvm::createLoopUnrollPass(-1, -1, 0, 0);
132 /// ApproximateLoopSize - Approximate the size of the loop.
133 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
134 bool &NotDuplicatable,
135 const TargetTransformInfo &TTI) {
137 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
139 Metrics.analyzeBasicBlock(*I, TTI);
140 NumCalls = Metrics.NumInlineCandidates;
141 NotDuplicatable = Metrics.notDuplicatable;
143 unsigned LoopSize = Metrics.NumInsts;
145 // Don't allow an estimate of size zero. This would allows unrolling of loops
146 // with huge iteration counts, which is a compile time problem even if it's
147 // not a problem for code quality.
148 if (LoopSize == 0) LoopSize = 1;
153 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
154 if (skipOptnoneFunction(L))
157 LoopInfo *LI = &getAnalysis<LoopInfo>();
158 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
159 const TargetTransformInfo &TTI = getAnalysis<TargetTransformInfo>();
161 BasicBlock *Header = L->getHeader();
162 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
163 << "] Loop %" << Header->getName() << "\n");
166 TargetTransformInfo::UnrollingPreferences UP;
167 UP.Threshold = CurrentThreshold;
168 UP.OptSizeThreshold = OptSizeUnrollThreshold;
169 UP.PartialThreshold = CurrentThreshold;
170 UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
171 UP.Count = CurrentCount;
172 UP.MaxCount = UINT_MAX;
173 UP.Partial = CurrentAllowPartial;
174 UP.Runtime = CurrentRuntime;
175 TTI.getUnrollingPreferences(L, UP);
177 // Determine the current unrolling threshold. While this is normally set
178 // from UnrollThreshold, it is overridden to a smaller value if the current
179 // function is marked as optimize-for-size, and the unroll threshold was
180 // not user specified.
181 unsigned Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
182 unsigned PartialThreshold =
183 UserThreshold ? CurrentThreshold : UP.PartialThreshold;
184 if (!UserThreshold &&
185 Header->getParent()->getAttributes().
186 hasAttribute(AttributeSet::FunctionIndex,
187 Attribute::OptimizeForSize)) {
188 Threshold = UP.OptSizeThreshold;
189 PartialThreshold = UP.PartialOptSizeThreshold;
192 // Find trip count and trip multiple if count is not available
193 unsigned TripCount = 0;
194 unsigned TripMultiple = 1;
195 // Find "latch trip count". UnrollLoop assumes that control cannot exit
196 // via the loop latch on any iteration prior to TripCount. The loop may exit
197 // early via an earlier branch.
198 BasicBlock *LatchBlock = L->getLoopLatch();
200 TripCount = SE->getSmallConstantTripCount(L, LatchBlock);
201 TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock);
204 bool Runtime = UserRuntime ? CurrentRuntime : UP.Runtime;
206 // Use a default unroll-count if the user doesn't specify a value
207 // and the trip count is a run-time value. The default is different
208 // for run-time or compile-time trip count loops.
209 unsigned Count = UserCount ? CurrentCount : UP.Count;
210 if (Runtime && Count == 0 && TripCount == 0)
211 Count = UnrollRuntimeCount;
214 // Conservative heuristic: if we know the trip count, see if we can
215 // completely unroll (subject to the threshold, checked below); otherwise
216 // try to find greatest modulo of the trip count which is still under
223 // Enforce the threshold.
224 if (Threshold != NoThreshold && PartialThreshold != NoThreshold) {
225 unsigned NumInlineCandidates;
226 bool notDuplicatable;
227 unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates,
228 notDuplicatable, TTI);
229 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
230 if (notDuplicatable) {
231 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable"
232 << " instructions.\n");
235 if (NumInlineCandidates != 0) {
236 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
239 uint64_t Size = (uint64_t)LoopSize*Count;
240 if (TripCount != 1 && Size > Threshold) {
241 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
242 << " because size: " << Size << ">" << Threshold << "\n");
243 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial;
244 if (!AllowPartial && !(Runtime && TripCount == 0)) {
245 DEBUG(dbgs() << " will not try to unroll partially because "
246 << "-unroll-allow-partial not given\n");
250 // Reduce unroll count to be modulo of TripCount for partial unrolling
251 Count = PartialThreshold / LoopSize;
252 while (Count != 0 && TripCount%Count != 0)
256 // Reduce unroll count to be a lower power-of-two value
257 while (Count != 0 && Size > PartialThreshold) {
259 Size = LoopSize*Count;
262 if (Count > UP.MaxCount)
265 DEBUG(dbgs() << " could not unroll partially\n");
268 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
273 if (!UnrollLoop(L, Count, TripCount, Runtime, TripMultiple, LI, this, &LPM))