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/IntrinsicInst.h"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/Analysis/LoopPass.h"
19 #include "llvm/Analysis/CodeMetrics.h"
20 #include "llvm/Analysis/ScalarEvolution.h"
21 #include "llvm/Support/CommandLine.h"
22 #include "llvm/Support/Debug.h"
23 #include "llvm/Support/raw_ostream.h"
24 #include "llvm/Transforms/Utils/UnrollLoop.h"
25 #include "llvm/DataLayout.h"
30 static cl::opt<unsigned>
31 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
32 cl::desc("The cut-off point for automatic loop unrolling"));
34 static cl::opt<unsigned>
35 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
36 cl::desc("Use this unroll count for all loops, for testing purposes"));
39 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
40 cl::desc("Allows loops to be partially unrolled until "
41 "-unroll-threshold loop size is reached."));
44 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden,
45 cl::desc("Unroll loops with run-time trip counts"));
48 class LoopUnroll : public LoopPass {
50 static char ID; // Pass ID, replacement for typeid
51 LoopUnroll(int T = -1, int C = -1, int P = -1) : LoopPass(ID) {
52 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
53 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
54 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
56 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
58 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
61 /// A magic value for use with the Threshold parameter to indicate
62 /// that the loop unroll should be performed regardless of how much
63 /// code expansion would result.
64 static const unsigned NoThreshold = UINT_MAX;
66 // Threshold to use when optsize is specified (and there is no
67 // explicit -unroll-threshold).
68 static const unsigned OptSizeUnrollThreshold = 50;
70 // Default unroll count for loops with run-time trip count if
71 // -unroll-count is not set
72 static const unsigned UnrollRuntimeCount = 8;
74 unsigned CurrentCount;
75 unsigned CurrentThreshold;
76 bool CurrentAllowPartial;
77 bool UserThreshold; // CurrentThreshold is user-specified.
79 bool runOnLoop(Loop *L, LPPassManager &LPM);
81 /// This transformation requires natural loop information & requires that
82 /// loop preheaders be inserted into the CFG...
84 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addRequired<LoopInfo>();
86 AU.addPreserved<LoopInfo>();
87 AU.addRequiredID(LoopSimplifyID);
88 AU.addPreservedID(LoopSimplifyID);
89 AU.addRequiredID(LCSSAID);
90 AU.addPreservedID(LCSSAID);
91 AU.addRequired<ScalarEvolution>();
92 AU.addPreserved<ScalarEvolution>();
93 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
94 // If loop unroll does not preserve dom info then LCSSA pass on next
95 // loop will receive invalid dom info.
96 // For now, recreate dom info, if loop is unrolled.
97 AU.addPreserved<DominatorTree>();
102 char LoopUnroll::ID = 0;
103 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
104 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
105 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
106 INITIALIZE_PASS_DEPENDENCY(LCSSA)
107 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
108 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
110 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial) {
111 return new LoopUnroll(Threshold, Count, AllowPartial);
114 /// ApproximateLoopSize - Approximate the size of the loop.
115 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
116 const DataLayout *TD) {
118 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
120 Metrics.analyzeBasicBlock(*I, TD);
121 NumCalls = Metrics.NumInlineCandidates;
123 unsigned LoopSize = Metrics.NumInsts;
125 // Don't allow an estimate of size zero. This would allows unrolling of loops
126 // with huge iteration counts, which is a compile time problem even if it's
127 // not a problem for code quality.
128 if (LoopSize == 0) LoopSize = 1;
133 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
134 LoopInfo *LI = &getAnalysis<LoopInfo>();
135 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
137 BasicBlock *Header = L->getHeader();
138 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
139 << "] Loop %" << Header->getName() << "\n");
142 // Determine the current unrolling threshold. While this is normally set
143 // from UnrollThreshold, it is overridden to a smaller value if the current
144 // function is marked as optimize-for-size, and the unroll threshold was
145 // not user specified.
146 unsigned Threshold = CurrentThreshold;
147 if (!UserThreshold &&
148 Header->getParent()->getFnAttributes().
149 hasAttribute(Attributes::OptimizeForSize))
150 Threshold = OptSizeUnrollThreshold;
152 // Find trip count and trip multiple if count is not available
153 unsigned TripCount = 0;
154 unsigned TripMultiple = 1;
155 // Find "latch trip count". UnrollLoop assumes that control cannot exit
156 // via the loop latch on any iteration prior to TripCount. The loop may exit
157 // early via an earlier branch.
158 BasicBlock *LatchBlock = L->getLoopLatch();
160 TripCount = SE->getSmallConstantTripCount(L, LatchBlock);
161 TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock);
163 // Use a default unroll-count if the user doesn't specify a value
164 // and the trip count is a run-time value. The default is different
165 // for run-time or compile-time trip count loops.
166 unsigned Count = CurrentCount;
167 if (UnrollRuntime && CurrentCount == 0 && TripCount == 0)
168 Count = UnrollRuntimeCount;
171 // Conservative heuristic: if we know the trip count, see if we can
172 // completely unroll (subject to the threshold, checked below); otherwise
173 // try to find greatest modulo of the trip count which is still under
180 // Enforce the threshold.
181 if (Threshold != NoThreshold) {
182 const DataLayout *TD = getAnalysisIfAvailable<DataLayout>();
183 unsigned NumInlineCandidates;
184 unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates, TD);
185 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
186 if (NumInlineCandidates != 0) {
187 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
190 uint64_t Size = (uint64_t)LoopSize*Count;
191 if (TripCount != 1 && Size > Threshold) {
192 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
193 << " because size: " << Size << ">" << Threshold << "\n");
194 if (!CurrentAllowPartial && !(UnrollRuntime && TripCount == 0)) {
195 DEBUG(dbgs() << " will not try to unroll partially because "
196 << "-unroll-allow-partial not given\n");
200 // Reduce unroll count to be modulo of TripCount for partial unrolling
201 Count = Threshold / LoopSize;
202 while (Count != 0 && TripCount%Count != 0)
205 else if (UnrollRuntime) {
206 // Reduce unroll count to be a lower power-of-two value
207 while (Count != 0 && Size > Threshold) {
209 Size = LoopSize*Count;
213 DEBUG(dbgs() << " could not unroll partially\n");
216 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
221 if (!UnrollLoop(L, Count, TripCount, UnrollRuntime, TripMultiple, LI, &LPM))