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/Transforms/Scalar.h" 17 #include "llvm/Analysis/CodeMetrics.h" 18 #include "llvm/Analysis/LoopPass.h" 19 #include "llvm/Analysis/ScalarEvolution.h" 20 #include "llvm/Analysis/TargetTransformInfo.h" 21 #include "llvm/IR/DataLayout.h" 22 #include "llvm/IR/IntrinsicInst.h" 23 #include "llvm/Support/CommandLine.h" 24 #include "llvm/Support/Debug.h" 25 #include "llvm/Support/raw_ostream.h" 26 #include "llvm/Transforms/Utils/UnrollLoop.h" 27 #include <climits> 28 29 using namespace llvm; 30 31 static cl::opt<unsigned> 32 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden, 33 cl::desc("The cut-off point for automatic loop unrolling")); 34 35 static cl::opt<unsigned> 36 UnrollCount("unroll-count", cl::init(0), cl::Hidden, 37 cl::desc("Use this unroll count for all loops, for testing purposes")); 38 39 static cl::opt<bool> 40 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden, 41 cl::desc("Allows loops to be partially unrolled until " 42 "-unroll-threshold loop size is reached.")); 43 44 static cl::opt<bool> 45 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden, 46 cl::desc("Unroll loops with run-time trip counts")); 47 48 namespace { 49 class LoopUnroll : public LoopPass { 50 public: 51 static char ID; // Pass ID, replacement for typeid 52 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) { 53 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T); 54 CurrentCount = (C == -1) ? UnrollCount : unsigned(C); 55 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P; 56 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R; 57 58 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0); 59 UserAllowPartial = (P != -1) || 60 (UnrollAllowPartial.getNumOccurrences() > 0); 61 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0); 62 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0); 63 64 initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); 65 } 66 67 /// A magic value for use with the Threshold parameter to indicate 68 /// that the loop unroll should be performed regardless of how much 69 /// code expansion would result. 70 static const unsigned NoThreshold = UINT_MAX; 71 72 // Threshold to use when optsize is specified (and there is no 73 // explicit -unroll-threshold). 74 static const unsigned OptSizeUnrollThreshold = 50; 75 76 // Default unroll count for loops with run-time trip count if 77 // -unroll-count is not set 78 static const unsigned UnrollRuntimeCount = 8; 79 80 unsigned CurrentCount; 81 unsigned CurrentThreshold; 82 bool CurrentAllowPartial; 83 bool CurrentRuntime; 84 bool UserCount; // CurrentCount is user-specified. 85 bool UserThreshold; // CurrentThreshold is user-specified. 86 bool UserAllowPartial; // CurrentAllowPartial is user-specified. 87 bool UserRuntime; // CurrentRuntime is user-specified. 88 89 bool runOnLoop(Loop *L, LPPassManager &LPM); 90 91 /// This transformation requires natural loop information & requires that 92 /// loop preheaders be inserted into the CFG... 93 /// 94 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 95 AU.addRequired<LoopInfo>(); 96 AU.addPreserved<LoopInfo>(); 97 AU.addRequiredID(LoopSimplifyID); 98 AU.addPreservedID(LoopSimplifyID); 99 AU.addRequiredID(LCSSAID); 100 AU.addPreservedID(LCSSAID); 101 AU.addRequired<ScalarEvolution>(); 102 AU.addPreserved<ScalarEvolution>(); 103 AU.addRequired<TargetTransformInfo>(); 104 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 105 // If loop unroll does not preserve dom info then LCSSA pass on next 106 // loop will receive invalid dom info. 107 // For now, recreate dom info, if loop is unrolled. 108 AU.addPreserved<DominatorTree>(); 109 } 110 }; 111 } 112 113 char LoopUnroll::ID = 0; 114 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 115 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 116 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 117 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 118 INITIALIZE_PASS_DEPENDENCY(LCSSA) 119 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 120 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 121 122 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial, 123 int Runtime) { 124 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime); 125 } 126 127 /// ApproximateLoopSize - Approximate the size of the loop. 128 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls, 129 bool &NotDuplicatable, 130 const TargetTransformInfo &TTI) { 131 CodeMetrics Metrics; 132 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 133 I != E; ++I) 134 Metrics.analyzeBasicBlock(*I, TTI); 135 NumCalls = Metrics.NumInlineCandidates; 136 NotDuplicatable = Metrics.notDuplicatable; 137 138 unsigned LoopSize = Metrics.NumInsts; 139 140 // Don't allow an estimate of size zero. This would allows unrolling of loops 141 // with huge iteration counts, which is a compile time problem even if it's 142 // not a problem for code quality. 143 if (LoopSize == 0) LoopSize = 1; 144 145 return LoopSize; 146 } 147 148 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { 149 LoopInfo *LI = &getAnalysis<LoopInfo>(); 150 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>(); 151 const TargetTransformInfo &TTI = getAnalysis<TargetTransformInfo>(); 152 153 BasicBlock *Header = L->getHeader(); 154 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 155 << "] Loop %" << Header->getName() << "\n"); 156 (void)Header; 157 158 TargetTransformInfo::UnrollingPreferences UP; 159 UP.Threshold = CurrentThreshold; 160 UP.OptSizeThreshold = OptSizeUnrollThreshold; 161 UP.Count = CurrentCount; 162 UP.Partial = CurrentAllowPartial; 163 UP.Runtime = CurrentRuntime; 164 TTI.getUnrollingPreferences(L, UP); 165 166 // Determine the current unrolling threshold. While this is normally set 167 // from UnrollThreshold, it is overridden to a smaller value if the current 168 // function is marked as optimize-for-size, and the unroll threshold was 169 // not user specified. 170 unsigned Threshold = UserThreshold ? CurrentThreshold : UP.Threshold; 171 if (!UserThreshold && 172 Header->getParent()->getAttributes(). 173 hasAttribute(AttributeSet::FunctionIndex, 174 Attribute::OptimizeForSize)) 175 Threshold = UP.OptSizeThreshold; 176 177 // Find trip count and trip multiple if count is not available 178 unsigned TripCount = 0; 179 unsigned TripMultiple = 1; 180 // Find "latch trip count". UnrollLoop assumes that control cannot exit 181 // via the loop latch on any iteration prior to TripCount. The loop may exit 182 // early via an earlier branch. 183 BasicBlock *LatchBlock = L->getLoopLatch(); 184 if (LatchBlock) { 185 TripCount = SE->getSmallConstantTripCount(L, LatchBlock); 186 TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock); 187 } 188 189 bool Runtime = UserRuntime ? CurrentRuntime : UP.Runtime; 190 191 // Use a default unroll-count if the user doesn't specify a value 192 // and the trip count is a run-time value. The default is different 193 // for run-time or compile-time trip count loops. 194 unsigned Count = UserCount ? CurrentCount : UP.Count; 195 if (Runtime && Count == 0 && TripCount == 0) 196 Count = UnrollRuntimeCount; 197 198 if (Count == 0) { 199 // Conservative heuristic: if we know the trip count, see if we can 200 // completely unroll (subject to the threshold, checked below); otherwise 201 // try to find greatest modulo of the trip count which is still under 202 // threshold value. 203 if (TripCount == 0) 204 return false; 205 Count = TripCount; 206 } 207 208 // Enforce the threshold. 209 if (Threshold != NoThreshold) { 210 unsigned NumInlineCandidates; 211 bool notDuplicatable; 212 unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates, 213 notDuplicatable, TTI); 214 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 215 if (notDuplicatable) { 216 DEBUG(dbgs() << " Not unrolling loop which contains non duplicatable" 217 << " instructions.\n"); 218 return false; 219 } 220 if (NumInlineCandidates != 0) { 221 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 222 return false; 223 } 224 uint64_t Size = (uint64_t)LoopSize*Count; 225 if (TripCount != 1 && Size > Threshold) { 226 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count 227 << " because size: " << Size << ">" << Threshold << "\n"); 228 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial; 229 if (!AllowPartial && !(Runtime && TripCount == 0)) { 230 DEBUG(dbgs() << " will not try to unroll partially because " 231 << "-unroll-allow-partial not given\n"); 232 return false; 233 } 234 if (TripCount) { 235 // Reduce unroll count to be modulo of TripCount for partial unrolling 236 Count = Threshold / LoopSize; 237 while (Count != 0 && TripCount%Count != 0) 238 Count--; 239 } 240 else if (Runtime) { 241 // Reduce unroll count to be a lower power-of-two value 242 while (Count != 0 && Size > Threshold) { 243 Count >>= 1; 244 Size = LoopSize*Count; 245 } 246 } 247 if (Count < 2) { 248 DEBUG(dbgs() << " could not unroll partially\n"); 249 return false; 250 } 251 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 252 } 253 } 254 255 // Unroll the loop. 256 if (!UnrollLoop(L, Count, TripCount, Runtime, TripMultiple, LI, &LPM)) 257 return false; 258 259 return true; 260 } 261