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/Transforms/Utils/UnrollLoop.h"
28 static cl::opt<unsigned>
29 UnrollThreshold("unroll-threshold", cl::init(100), cl::Hidden,
30 cl::desc("The cut-off point for automatic loop unrolling"));
32 static cl::opt<unsigned>
33 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
34 cl::desc("Use this unroll count for all loops, for testing purposes"));
37 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
38 cl::desc("Allows loops to be partially unrolled until "
39 "-unroll-threshold loop size is reached."));
42 class VISIBILITY_HIDDEN LoopUnroll : public LoopPass {
44 static char ID; // Pass ID, replacement for typeid
45 LoopUnroll() : LoopPass(&ID) {}
47 /// A magic value for use with the Threshold parameter to indicate
48 /// that the loop unroll should be performed regardless of how much
49 /// code expansion would result.
50 static const unsigned NoThreshold = UINT_MAX;
52 bool runOnLoop(Loop *L, LPPassManager &LPM);
54 /// This transformation requires natural loop information & requires that
55 /// loop preheaders be inserted into the CFG...
57 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
58 AU.addRequiredID(LoopSimplifyID);
59 AU.addRequiredID(LCSSAID);
60 AU.addRequired<LoopInfo>();
61 AU.addPreservedID(LCSSAID);
62 AU.addPreserved<LoopInfo>();
63 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
64 // If loop unroll does not preserve dom info then LCSSA pass on next
65 // loop will receive invalid dom info.
66 // For now, recreate dom info, if loop is unrolled.
67 AU.addPreserved<DominatorTree>();
68 AU.addPreserved<DominanceFrontier>();
73 char LoopUnroll::ID = 0;
74 static RegisterPass<LoopUnroll> X("loop-unroll", "Unroll loops");
76 Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
78 /// ApproximateLoopSize - Approximate the size of the loop.
79 static unsigned ApproximateLoopSize(const Loop *L) {
81 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
84 Instruction *Term = BB->getTerminator();
85 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
86 if (isa<PHINode>(I) && BB == L->getHeader()) {
87 // Ignore PHI nodes in the header.
88 } else if (I->hasOneUse() && I->use_back() == Term) {
89 // Ignore instructions only used by the loop terminator.
90 } else if (isa<DbgInfoIntrinsic>(I)) {
91 // Ignore debug instructions
92 } else if (isa<GetElementPtrInst>(I) && I->hasOneUse()) {
93 // Ignore GEP as they generally are subsumed into a load or store.
94 } else if (isa<CallInst>(I)) {
95 // Estimate size overhead introduced by call instructions which
96 // is higher than other instructions. Here 3 and 10 are magic
97 // numbers that help one isolated test case from PR2067 without
98 // negatively impacting measured benchmarks.
99 if (isa<IntrinsicInst>(I))
107 // TODO: Ignore expressions derived from PHI and constants if inval of phi
108 // is a constant, or if operation is associative. This will get induction
116 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
117 assert(L->isLCSSAForm());
118 LoopInfo *LI = &getAnalysis<LoopInfo>();
120 BasicBlock *Header = L->getHeader();
121 DOUT << "Loop Unroll: F[" << Header->getParent()->getName()
122 << "] Loop %" << Header->getName() << "\n";
125 unsigned TripCount = L->getSmallConstantTripCount();
126 unsigned Count = UnrollCount;
128 // Automatically select an unroll count.
130 // Conservative heuristic: if we know the trip count, see if we can
131 // completely unroll (subject to the threshold, checked below); otherwise
132 // try to find greatest modulo of the trip count which is still under
134 if (TripCount != 0) {
141 // Enforce the threshold.
142 if (UnrollThreshold != NoThreshold) {
143 unsigned LoopSize = ApproximateLoopSize(L);
144 DOUT << " Loop Size = " << LoopSize << "\n";
145 uint64_t Size = (uint64_t)LoopSize*Count;
146 if (TripCount != 1 && Size > UnrollThreshold) {
147 DOUT << " Too large to fully unroll with count: " << Count
148 << " because size: " << Size << ">" << UnrollThreshold << "\n";
149 if (UnrollAllowPartial) {
150 // Reduce unroll count to be modulo of TripCount for partial unrolling
151 Count = UnrollThreshold / LoopSize;
152 while (Count != 0 && TripCount%Count != 0) {
156 DOUT << " could not unroll partially\n";
159 DOUT << " partially unrolling with count: " << Count << "\n";
162 DOUT << " will not try to unroll partially because "
163 << "-unroll-allow-partial not given\n";
170 Function *F = L->getHeader()->getParent();
171 if (!UnrollLoop(L, Count, LI, &LPM))
174 // FIXME: Reconstruct dom info, because it is not preserved properly.
175 DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>();
177 DT->runOnFunction(*F);
178 DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>();
180 DF->runOnFunction(*F);