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 27 using namespace llvm; 28 29 static cl::opt<unsigned> 30 UnrollThreshold("unroll-threshold", cl::init(200), cl::Hidden, 31 cl::desc("The cut-off point for automatic loop unrolling")); 32 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")); 36 37 static cl::opt<bool> 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.")); 41 42 namespace { 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 // Threshold to use when optsize is specified (and there is no 54 // explicit -unroll-threshold). 55 static const unsigned OptSizeUnrollThreshold = 50; 56 57 unsigned CurrentThreshold; 58 59 bool runOnLoop(Loop *L, LPPassManager &LPM); 60 61 /// This transformation requires natural loop information & requires that 62 /// loop preheaders be inserted into the CFG... 63 /// 64 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 65 AU.addRequired<LoopInfo>(); 66 AU.addPreserved<LoopInfo>(); 67 AU.addRequiredID(LoopSimplifyID); 68 AU.addPreservedID(LoopSimplifyID); 69 AU.addRequiredID(LCSSAID); 70 AU.addPreservedID(LCSSAID); 71 AU.addPreserved<ScalarEvolution>(); 72 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 73 // If loop unroll does not preserve dom info then LCSSA pass on next 74 // loop will receive invalid dom info. 75 // For now, recreate dom info, if loop is unrolled. 76 AU.addPreserved<DominatorTree>(); 77 } 78 }; 79 } 80 81 char LoopUnroll::ID = 0; 82 INITIALIZE_PASS(LoopUnroll, "loop-unroll", "Unroll loops", false, false); 83 84 Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); } 85 86 /// ApproximateLoopSize - Approximate the size of the loop. 87 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls) { 88 CodeMetrics Metrics; 89 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 90 I != E; ++I) 91 Metrics.analyzeBasicBlock(*I); 92 NumCalls = Metrics.NumCalls; 93 return Metrics.NumInsts; 94 } 95 96 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { 97 98 LoopInfo *LI = &getAnalysis<LoopInfo>(); 99 100 BasicBlock *Header = L->getHeader(); 101 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 102 << "] Loop %" << Header->getName() << "\n"); 103 (void)Header; 104 105 // Determine the current unrolling threshold. While this is normally set 106 // from UnrollThreshold, it is overridden to a smaller value if the current 107 // function is marked as optimize-for-size, and the unroll threshold was 108 // not user specified. 109 CurrentThreshold = UnrollThreshold; 110 if (Header->getParent()->hasFnAttr(Attribute::OptimizeForSize) && 111 UnrollThreshold.getNumOccurrences() == 0) 112 CurrentThreshold = OptSizeUnrollThreshold; 113 114 // Find trip count 115 unsigned TripCount = L->getSmallConstantTripCount(); 116 unsigned Count = UnrollCount; 117 118 // Automatically select an unroll count. 119 if (Count == 0) { 120 // Conservative heuristic: if we know the trip count, see if we can 121 // completely unroll (subject to the threshold, checked below); otherwise 122 // try to find greatest modulo of the trip count which is still under 123 // threshold value. 124 if (TripCount == 0) 125 return false; 126 Count = TripCount; 127 } 128 129 // Enforce the threshold. 130 if (CurrentThreshold != NoThreshold) { 131 unsigned NumCalls; 132 unsigned LoopSize = ApproximateLoopSize(L, NumCalls); 133 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 134 if (NumCalls != 0) { 135 DEBUG(dbgs() << " Not unrolling loop with function calls.\n"); 136 return false; 137 } 138 uint64_t Size = (uint64_t)LoopSize*Count; 139 if (TripCount != 1 && Size > CurrentThreshold) { 140 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count 141 << " because size: " << Size << ">" << CurrentThreshold << "\n"); 142 if (!UnrollAllowPartial) { 143 DEBUG(dbgs() << " will not try to unroll partially because " 144 << "-unroll-allow-partial not given\n"); 145 return false; 146 } 147 // Reduce unroll count to be modulo of TripCount for partial unrolling 148 Count = CurrentThreshold / LoopSize; 149 while (Count != 0 && TripCount%Count != 0) { 150 Count--; 151 } 152 if (Count < 2) { 153 DEBUG(dbgs() << " could not unroll partially\n"); 154 return false; 155 } 156 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 157 } 158 } 159 160 // Unroll the loop. 161 Function *F = L->getHeader()->getParent(); 162 if (!UnrollLoop(L, Count, LI, &LPM)) 163 return false; 164 165 // FIXME: Reconstruct dom info, because it is not preserved properly. 166 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) 167 DT->runOnFunction(*F); 168 return true; 169 } 170