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/LoopInfo.h"
19 #include "llvm/Analysis/LoopPass.h"
20 #include "llvm/Support/Compiler.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(100), 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 VISIBILITY_HIDDEN LoopUnroll : public LoopPass {
45 static char ID; // Pass ID, replacement for typeid
46 LoopUnroll() : LoopPass(&ID) {}
48 /// A magic value for use with the Threshold parameter to indicate
49 /// that the loop unroll should be performed regardless of how much
50 /// code expansion would result.
51 static const unsigned NoThreshold = UINT_MAX;
53 bool runOnLoop(Loop *L, LPPassManager &LPM);
55 /// This transformation requires natural loop information & requires that
56 /// loop preheaders be inserted into the CFG...
58 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
59 AU.addRequiredID(LoopSimplifyID);
60 AU.addRequiredID(LCSSAID);
61 AU.addRequired<LoopInfo>();
62 AU.addPreservedID(LCSSAID);
63 AU.addPreserved<LoopInfo>();
64 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
65 // If loop unroll does not preserve dom info then LCSSA pass on next
66 // loop will receive invalid dom info.
67 // For now, recreate dom info, if loop is unrolled.
68 AU.addPreserved<DominatorTree>();
69 AU.addPreserved<DominanceFrontier>();
74 char LoopUnroll::ID = 0;
75 static RegisterPass<LoopUnroll> X("loop-unroll", "Unroll loops");
77 Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
79 /// ApproximateLoopSize - Approximate the size of the loop.
80 static unsigned ApproximateLoopSize(const Loop *L) {
82 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
85 Instruction *Term = BB->getTerminator();
86 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
87 if (isa<PHINode>(I) && BB == L->getHeader()) {
88 // Ignore PHI nodes in the header.
89 } else if (I->hasOneUse() && I->use_back() == Term) {
90 // Ignore instructions only used by the loop terminator.
91 } else if (isa<DbgInfoIntrinsic>(I)) {
92 // Ignore debug instructions
93 } else if (isa<GetElementPtrInst>(I) && I->hasOneUse()) {
94 // Ignore GEP as they generally are subsumed into a load or store.
95 } else if (isa<CallInst>(I)) {
96 // Estimate size overhead introduced by call instructions which
97 // is higher than other instructions. Here 3 and 10 are magic
98 // numbers that help one isolated test case from PR2067 without
99 // negatively impacting measured benchmarks.
100 Size += isa<IntrinsicInst>(I) ? 3 : 10;
105 // TODO: Ignore expressions derived from PHI and constants if inval of phi
106 // is a constant, or if operation is associative. This will get induction
114 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
115 assert(L->isLCSSAForm());
116 LoopInfo *LI = &getAnalysis<LoopInfo>();
118 BasicBlock *Header = L->getHeader();
119 DEBUG(errs() << "Loop Unroll: F[" << Header->getParent()->getName()
120 << "] Loop %" << Header->getName() << "\n");
124 unsigned TripCount = L->getSmallConstantTripCount();
125 unsigned Count = UnrollCount;
127 // Automatically select an unroll count.
129 // Conservative heuristic: if we know the trip count, see if we can
130 // completely unroll (subject to the threshold, checked below); otherwise
131 // try to find greatest modulo of the trip count which is still under
138 // Enforce the threshold.
139 if (UnrollThreshold != NoThreshold) {
140 unsigned LoopSize = ApproximateLoopSize(L);
141 DEBUG(errs() << " Loop Size = " << LoopSize << "\n");
142 uint64_t Size = (uint64_t)LoopSize*Count;
143 if (TripCount != 1 && Size > UnrollThreshold) {
144 DEBUG(errs() << " Too large to fully unroll with count: " << Count
145 << " because size: " << Size << ">" << UnrollThreshold << "\n");
146 if (!UnrollAllowPartial) {
147 DEBUG(errs() << " will not try to unroll partially because "
148 << "-unroll-allow-partial not given\n");
151 // Reduce unroll count to be modulo of TripCount for partial unrolling
152 Count = UnrollThreshold / LoopSize;
153 while (Count != 0 && TripCount%Count != 0) {
157 DEBUG(errs() << " could not unroll partially\n");
160 DEBUG(errs() << " partially unrolling with count: " << Count << "\n");
165 Function *F = L->getHeader()->getParent();
166 if (!UnrollLoop(L, Count, LI, &LPM))
169 // FIXME: Reconstruct dom info, because it is not preserved properly.
170 DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>();
172 DT->runOnFunction(*F);
173 DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>();
175 DF->runOnFunction(*F);