LoopUnrollPass.cpp revision 90c579de5a383cee278acc3f7e7b9d0a656e6a35
1//===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 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//===----------------------------------------------------------------------===// 14 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/InlineCost.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 <climits> 26 27using namespace llvm; 28 29static cl::opt<unsigned> 30UnrollThreshold("unroll-threshold", cl::init(200), cl::Hidden, 31 cl::desc("The cut-off point for automatic loop unrolling")); 32 33static cl::opt<unsigned> 34UnrollCount("unroll-count", cl::init(0), cl::Hidden, 35 cl::desc("Use this unroll count for all loops, for testing purposes")); 36 37static cl::opt<bool> 38UnrollAllowPartial("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.")); 41 42namespace { 43 class LoopUnroll : public LoopPass { 44 public: 45 static char ID; // Pass ID, replacement for typeid 46 LoopUnroll() : LoopPass(ID) {} 47 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; 52 53 bool runOnLoop(Loop *L, LPPassManager &LPM); 54 55 /// This transformation requires natural loop information & requires that 56 /// loop preheaders be inserted into the CFG... 57 /// 58 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 59 AU.addRequired<LoopInfo>(); 60 AU.addPreserved<LoopInfo>(); 61 AU.addRequiredID(LoopSimplifyID); 62 AU.addPreservedID(LoopSimplifyID); 63 AU.addRequiredID(LCSSAID); 64 AU.addPreservedID(LCSSAID); 65 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 66 // If loop unroll does not preserve dom info then LCSSA pass on next 67 // loop will receive invalid dom info. 68 // For now, recreate dom info, if loop is unrolled. 69 AU.addPreserved<DominatorTree>(); 70 AU.addPreserved<DominanceFrontier>(); 71 AU.addPreserved<ScalarEvolution>(); 72 } 73 }; 74} 75 76char LoopUnroll::ID = 0; 77INITIALIZE_PASS(LoopUnroll, "loop-unroll", "Unroll loops", false, false); 78 79Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); } 80 81/// ApproximateLoopSize - Approximate the size of the loop. 82static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls) { 83 CodeMetrics Metrics; 84 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 85 I != E; ++I) 86 Metrics.analyzeBasicBlock(*I); 87 NumCalls = Metrics.NumCalls; 88 return Metrics.NumInsts; 89} 90 91bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { 92 LoopInfo *LI = &getAnalysis<LoopInfo>(); 93 94 BasicBlock *Header = L->getHeader(); 95 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 96 << "] Loop %" << Header->getName() << "\n"); 97 (void)Header; 98 99 // Find trip count 100 unsigned TripCount = L->getSmallConstantTripCount(); 101 unsigned Count = UnrollCount; 102 103 // Automatically select an unroll count. 104 if (Count == 0) { 105 // Conservative heuristic: if we know the trip count, see if we can 106 // completely unroll (subject to the threshold, checked below); otherwise 107 // try to find greatest modulo of the trip count which is still under 108 // threshold value. 109 if (TripCount == 0) 110 return false; 111 Count = TripCount; 112 } 113 114 // Enforce the threshold. 115 if (UnrollThreshold != NoThreshold) { 116 unsigned NumCalls; 117 unsigned LoopSize = ApproximateLoopSize(L, NumCalls); 118 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 119 if (NumCalls != 0) { 120 DEBUG(dbgs() << " Not unrolling loop with function calls.\n"); 121 return false; 122 } 123 uint64_t Size = (uint64_t)LoopSize*Count; 124 if (TripCount != 1 && Size > UnrollThreshold) { 125 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count 126 << " because size: " << Size << ">" << UnrollThreshold << "\n"); 127 if (!UnrollAllowPartial) { 128 DEBUG(dbgs() << " will not try to unroll partially because " 129 << "-unroll-allow-partial not given\n"); 130 return false; 131 } 132 // Reduce unroll count to be modulo of TripCount for partial unrolling 133 Count = UnrollThreshold / LoopSize; 134 while (Count != 0 && TripCount%Count != 0) { 135 Count--; 136 } 137 if (Count < 2) { 138 DEBUG(dbgs() << " could not unroll partially\n"); 139 return false; 140 } 141 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 142 } 143 } 144 145 // Unroll the loop. 146 Function *F = L->getHeader()->getParent(); 147 if (!UnrollLoop(L, Count, LI, &LPM)) 148 return false; 149 150 // FIXME: Reconstruct dom info, because it is not preserved properly. 151 DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>(); 152 if (DT) { 153 DT->runOnFunction(*F); 154 DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>(); 155 if (DF) 156 DF->runOnFunction(*F); 157 } 158 return true; 159} 160