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/Target/TargetData.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 class LoopUnroll : public LoopPass {
46 static char ID; // Pass ID, replacement for typeid
47 LoopUnroll(int T = -1, int C = -1, int P = -1) : LoopPass(ID) {
48 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
49 CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
50 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
52 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
54 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
57 /// A magic value for use with the Threshold parameter to indicate
58 /// that the loop unroll should be performed regardless of how much
59 /// code expansion would result.
60 static const unsigned NoThreshold = UINT_MAX;
62 // Threshold to use when optsize is specified (and there is no
63 // explicit -unroll-threshold).
64 static const unsigned OptSizeUnrollThreshold = 50;
66 unsigned CurrentCount;
67 unsigned CurrentThreshold;
68 bool CurrentAllowPartial;
69 bool UserThreshold; // CurrentThreshold is user-specified.
71 bool runOnLoop(Loop *L, LPPassManager &LPM);
73 /// This transformation requires natural loop information & requires that
74 /// loop preheaders be inserted into the CFG...
76 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
77 AU.addRequired<LoopInfo>();
78 AU.addPreserved<LoopInfo>();
79 AU.addRequiredID(LoopSimplifyID);
80 AU.addPreservedID(LoopSimplifyID);
81 AU.addRequiredID(LCSSAID);
82 AU.addPreservedID(LCSSAID);
83 AU.addRequired<ScalarEvolution>();
84 AU.addPreserved<ScalarEvolution>();
85 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
86 // If loop unroll does not preserve dom info then LCSSA pass on next
87 // loop will receive invalid dom info.
88 // For now, recreate dom info, if loop is unrolled.
89 AU.addPreserved<DominatorTree>();
94 char LoopUnroll::ID = 0;
95 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
96 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
97 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
98 INITIALIZE_PASS_DEPENDENCY(LCSSA)
99 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
100 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
102 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial) {
103 return new LoopUnroll(Threshold, Count, AllowPartial);
106 /// ApproximateLoopSize - Approximate the size of the loop.
107 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
108 const TargetData *TD) {
110 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
112 Metrics.analyzeBasicBlock(*I, TD);
113 NumCalls = Metrics.NumInlineCandidates;
115 unsigned LoopSize = Metrics.NumInsts;
117 // Don't allow an estimate of size zero. This would allows unrolling of loops
118 // with huge iteration counts, which is a compile time problem even if it's
119 // not a problem for code quality.
120 if (LoopSize == 0) LoopSize = 1;
125 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
126 LoopInfo *LI = &getAnalysis<LoopInfo>();
127 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
129 BasicBlock *Header = L->getHeader();
130 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
131 << "] Loop %" << Header->getName() << "\n");
134 // Determine the current unrolling threshold. While this is normally set
135 // from UnrollThreshold, it is overridden to a smaller value if the current
136 // function is marked as optimize-for-size, and the unroll threshold was
137 // not user specified.
138 unsigned Threshold = CurrentThreshold;
139 if (!UserThreshold &&
140 Header->getParent()->hasFnAttr(Attribute::OptimizeForSize))
141 Threshold = OptSizeUnrollThreshold;
143 // Find trip count and trip multiple if count is not available
144 unsigned TripCount = 0;
145 unsigned TripMultiple = 1;
146 // Find "latch trip count". UnrollLoop assumes that control cannot exit
147 // via the loop latch on any iteration prior to TripCount. The loop may exit
148 // early via an earlier branch.
149 BasicBlock *LatchBlock = L->getLoopLatch();
151 TripCount = SE->getSmallConstantTripCount(L, LatchBlock);
152 TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock);
154 // Automatically select an unroll count.
155 unsigned Count = CurrentCount;
157 // Conservative heuristic: if we know the trip count, see if we can
158 // completely unroll (subject to the threshold, checked below); otherwise
159 // try to find greatest modulo of the trip count which is still under
166 // Enforce the threshold.
167 if (Threshold != NoThreshold) {
168 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
169 unsigned NumInlineCandidates;
170 unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates, TD);
171 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
172 if (NumInlineCandidates != 0) {
173 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
176 uint64_t Size = (uint64_t)LoopSize*Count;
177 if (TripCount != 1 && Size > Threshold) {
178 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
179 << " because size: " << Size << ">" << Threshold << "\n");
180 if (!CurrentAllowPartial) {
181 DEBUG(dbgs() << " will not try to unroll partially because "
182 << "-unroll-allow-partial not given\n");
185 // Reduce unroll count to be modulo of TripCount for partial unrolling
186 Count = Threshold / LoopSize;
187 while (Count != 0 && TripCount%Count != 0) {
191 DEBUG(dbgs() << " could not unroll partially\n");
194 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
199 if (!UnrollLoop(L, Count, TripCount, TripMultiple, LI, &LPM))