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/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" 25 #include <climits> 26 27 using namespace llvm; 28 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")); 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 // Temporary flag to be removed in 3.0 43 static cl::opt<bool> 44 NoSCEVUnroll("disable-unroll-scev", cl::init(false), cl::Hidden, 45 cl::desc("Use ScalarEvolution to analyze loop trip counts for unrolling")); 46 47 namespace { 48 class LoopUnroll : public LoopPass { 49 public: 50 static char ID; // Pass ID, replacement for typeid 51 LoopUnroll(int T = -1, int C = -1, int P = -1) : LoopPass(ID) { 52 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T); 53 CurrentCount = (C == -1) ? UnrollCount : unsigned(C); 54 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P; 55 56 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0); 57 58 initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); 59 } 60 61 /// A magic value for use with the Threshold parameter to indicate 62 /// that the loop unroll should be performed regardless of how much 63 /// code expansion would result. 64 static const unsigned NoThreshold = UINT_MAX; 65 66 // Threshold to use when optsize is specified (and there is no 67 // explicit -unroll-threshold). 68 static const unsigned OptSizeUnrollThreshold = 50; 69 70 unsigned CurrentCount; 71 unsigned CurrentThreshold; 72 bool CurrentAllowPartial; 73 bool UserThreshold; // CurrentThreshold is user-specified. 74 75 bool runOnLoop(Loop *L, LPPassManager &LPM); 76 77 /// This transformation requires natural loop information & requires that 78 /// loop preheaders be inserted into the CFG... 79 /// 80 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 81 AU.addRequired<LoopInfo>(); 82 AU.addPreserved<LoopInfo>(); 83 AU.addRequiredID(LoopSimplifyID); 84 AU.addPreservedID(LoopSimplifyID); 85 AU.addRequiredID(LCSSAID); 86 AU.addPreservedID(LCSSAID); 87 AU.addRequired<ScalarEvolution>(); 88 AU.addPreserved<ScalarEvolution>(); 89 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 90 // If loop unroll does not preserve dom info then LCSSA pass on next 91 // loop will receive invalid dom info. 92 // For now, recreate dom info, if loop is unrolled. 93 AU.addPreserved<DominatorTree>(); 94 } 95 }; 96 } 97 98 char LoopUnroll::ID = 0; 99 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 100 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 101 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 102 INITIALIZE_PASS_DEPENDENCY(LCSSA) 103 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 104 105 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial) { 106 return new LoopUnroll(Threshold, Count, AllowPartial); 107 } 108 109 /// ApproximateLoopSize - Approximate the size of the loop. 110 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls) { 111 CodeMetrics Metrics; 112 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 113 I != E; ++I) 114 Metrics.analyzeBasicBlock(*I); 115 NumCalls = Metrics.NumInlineCandidates; 116 117 unsigned LoopSize = Metrics.NumInsts; 118 119 // Don't allow an estimate of size zero. This would allows unrolling of loops 120 // with huge iteration counts, which is a compile time problem even if it's 121 // not a problem for code quality. 122 if (LoopSize == 0) LoopSize = 1; 123 124 return LoopSize; 125 } 126 127 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { 128 LoopInfo *LI = &getAnalysis<LoopInfo>(); 129 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>(); 130 131 BasicBlock *Header = L->getHeader(); 132 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 133 << "] Loop %" << Header->getName() << "\n"); 134 (void)Header; 135 136 // Determine the current unrolling threshold. While this is normally set 137 // from UnrollThreshold, it is overridden to a smaller value if the current 138 // function is marked as optimize-for-size, and the unroll threshold was 139 // not user specified. 140 unsigned Threshold = CurrentThreshold; 141 if (!UserThreshold && 142 Header->getParent()->hasFnAttr(Attribute::OptimizeForSize)) 143 Threshold = OptSizeUnrollThreshold; 144 145 // Find trip count and trip multiple if count is not available 146 unsigned TripCount = 0; 147 unsigned TripMultiple = 1; 148 if (!NoSCEVUnroll) { 149 // Find "latch trip count". UnrollLoop assumes that control cannot exit 150 // via the loop latch on any iteration prior to TripCount. The loop may exit 151 // early via an earlier branch. 152 BasicBlock *LatchBlock = L->getLoopLatch(); 153 if (LatchBlock) { 154 TripCount = SE->getSmallConstantTripCount(L, LatchBlock); 155 TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock); 156 } 157 } 158 else { 159 TripCount = L->getSmallConstantTripCount(); 160 if (TripCount == 0) 161 TripMultiple = L->getSmallConstantTripMultiple(); 162 } 163 // Automatically select an unroll count. 164 unsigned Count = CurrentCount; 165 if (Count == 0) { 166 // Conservative heuristic: if we know the trip count, see if we can 167 // completely unroll (subject to the threshold, checked below); otherwise 168 // try to find greatest modulo of the trip count which is still under 169 // threshold value. 170 if (TripCount == 0) 171 return false; 172 Count = TripCount; 173 } 174 175 // Enforce the threshold. 176 if (Threshold != NoThreshold) { 177 unsigned NumInlineCandidates; 178 unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates); 179 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 180 if (NumInlineCandidates != 0) { 181 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 182 return false; 183 } 184 uint64_t Size = (uint64_t)LoopSize*Count; 185 if (TripCount != 1 && Size > Threshold) { 186 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count 187 << " because size: " << Size << ">" << Threshold << "\n"); 188 if (!CurrentAllowPartial) { 189 DEBUG(dbgs() << " will not try to unroll partially because " 190 << "-unroll-allow-partial not given\n"); 191 return false; 192 } 193 // Reduce unroll count to be modulo of TripCount for partial unrolling 194 Count = Threshold / LoopSize; 195 while (Count != 0 && TripCount%Count != 0) { 196 Count--; 197 } 198 if (Count < 2) { 199 DEBUG(dbgs() << " could not unroll partially\n"); 200 return false; 201 } 202 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 203 } 204 } 205 206 // Unroll the loop. 207 if (!UnrollLoop(L, Count, TripCount, TripMultiple, LI, &LPM)) 208 return false; 209 210 return true; 211 } 212