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"
29 static cl::opt<unsigned>
30 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
31 cl::desc("The cut-off point for automatic loop unrolling"));
33 static cl::opt<unsigned>
34 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
35 cl::desc("Use this unroll count for all loops, for testing purposes"));
38 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
39 cl::desc("Allows loops to be partially unrolled until "
40 "-unroll-threshold loop size is reached."));
43 class LoopUnroll : public LoopPass {
45 static char ID; // Pass ID, replacement for typeid
46 LoopUnroll() : LoopPass(ID) {
47 initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
50 /// A magic value for use with the Threshold parameter to indicate
51 /// that the loop unroll should be performed regardless of how much
52 /// code expansion would result.
53 static const unsigned NoThreshold = UINT_MAX;
55 // Threshold to use when optsize is specified (and there is no
56 // explicit -unroll-threshold).
57 static const unsigned OptSizeUnrollThreshold = 50;
59 unsigned CurrentThreshold;
61 bool runOnLoop(Loop *L, LPPassManager &LPM);
63 /// This transformation requires natural loop information & requires that
64 /// loop preheaders be inserted into the CFG...
66 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
67 AU.addRequired<LoopInfo>();
68 AU.addPreserved<LoopInfo>();
69 AU.addRequiredID(LoopSimplifyID);
70 AU.addPreservedID(LoopSimplifyID);
71 AU.addRequiredID(LCSSAID);
72 AU.addPreservedID(LCSSAID);
73 AU.addPreserved<ScalarEvolution>();
74 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
75 // If loop unroll does not preserve dom info then LCSSA pass on next
76 // loop will receive invalid dom info.
77 // For now, recreate dom info, if loop is unrolled.
78 AU.addPreserved<DominatorTree>();
83 char LoopUnroll::ID = 0;
84 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
85 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
86 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
87 INITIALIZE_PASS_DEPENDENCY(LCSSA)
88 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
90 Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
92 /// ApproximateLoopSize - Approximate the size of the loop.
93 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls) {
95 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
97 Metrics.analyzeBasicBlock(*I);
98 NumCalls = Metrics.NumInlineCandidates;
100 unsigned LoopSize = Metrics.NumInsts;
102 // If we can identify the induction variable, we know that it will become
103 // constant when we unroll the loop, so factor that into our loop size
105 // FIXME: We have to divide by InlineConstants::InstrCost because the
106 // measure returned by CountCodeReductionForConstant is not an instruction
107 // count, but rather a weight as defined by InlineConstants. It would
108 // probably be a good idea to standardize on a single weighting scheme by
109 // pushing more of the logic for weighting into CodeMetrics.
110 if (PHINode *IndVar = L->getCanonicalInductionVariable()) {
111 unsigned SizeDecrease = Metrics.CountCodeReductionForConstant(IndVar);
112 // NOTE: Because SizeDecrease is a fuzzy estimate, we don't want to allow
113 // it to totally negate the cost of unrolling a loop.
114 SizeDecrease = SizeDecrease > LoopSize / 2 ? LoopSize / 2 : SizeDecrease;
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>();
128 BasicBlock *Header = L->getHeader();
129 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
130 << "] Loop %" << Header->getName() << "\n");
133 // Determine the current unrolling threshold. While this is normally set
134 // from UnrollThreshold, it is overridden to a smaller value if the current
135 // function is marked as optimize-for-size, and the unroll threshold was
136 // not user specified.
137 CurrentThreshold = UnrollThreshold;
138 if (Header->getParent()->hasFnAttr(Attribute::OptimizeForSize) &&
139 UnrollThreshold.getNumOccurrences() == 0)
140 CurrentThreshold = OptSizeUnrollThreshold;
143 unsigned TripCount = L->getSmallConstantTripCount();
144 unsigned Count = UnrollCount;
146 // Automatically select an unroll count.
148 // Conservative heuristic: if we know the trip count, see if we can
149 // completely unroll (subject to the threshold, checked below); otherwise
150 // try to find greatest modulo of the trip count which is still under
157 // Enforce the threshold.
158 if (CurrentThreshold != NoThreshold) {
159 unsigned NumInlineCandidates;
160 unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates);
161 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n");
162 if (NumInlineCandidates != 0) {
163 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n");
166 uint64_t Size = (uint64_t)LoopSize*Count;
167 if (TripCount != 1 && Size > CurrentThreshold) {
168 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count
169 << " because size: " << Size << ">" << CurrentThreshold << "\n");
170 if (!UnrollAllowPartial) {
171 DEBUG(dbgs() << " will not try to unroll partially because "
172 << "-unroll-allow-partial not given\n");
175 // Reduce unroll count to be modulo of TripCount for partial unrolling
176 Count = CurrentThreshold / LoopSize;
177 while (Count != 0 && TripCount%Count != 0) {
181 DEBUG(dbgs() << " could not unroll partially\n");
184 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n");
189 Function *F = L->getHeader()->getParent();
190 if (!UnrollLoop(L, Count, LI, &LPM))
193 // FIXME: Reconstruct dom info, because it is not preserved properly.
194 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>())
195 DT->runOnFunction(*F);