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/Dominators.h" 23 #include "llvm/IR/IntrinsicInst.h" 24 #include "llvm/Support/CommandLine.h" 25 #include "llvm/Support/Debug.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include "llvm/Transforms/Utils/UnrollLoop.h" 28 #include <climits> 29 30 using namespace llvm; 31 32 static cl::opt<unsigned> 33 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden, 34 cl::desc("The cut-off point for automatic loop unrolling")); 35 36 static cl::opt<unsigned> 37 UnrollCount("unroll-count", cl::init(0), cl::Hidden, 38 cl::desc("Use this unroll count for all loops, for testing purposes")); 39 40 static cl::opt<bool> 41 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden, 42 cl::desc("Allows loops to be partially unrolled until " 43 "-unroll-threshold loop size is reached.")); 44 45 static cl::opt<bool> 46 UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden, 47 cl::desc("Unroll loops with run-time trip counts")); 48 49 namespace { 50 class LoopUnroll : public LoopPass { 51 public: 52 static char ID; // Pass ID, replacement for typeid 53 LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) { 54 CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T); 55 CurrentCount = (C == -1) ? UnrollCount : unsigned(C); 56 CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P; 57 CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R; 58 59 UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0); 60 UserAllowPartial = (P != -1) || 61 (UnrollAllowPartial.getNumOccurrences() > 0); 62 UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0); 63 UserCount = (C != -1) || (UnrollCount.getNumOccurrences() > 0); 64 65 initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); 66 } 67 68 /// A magic value for use with the Threshold parameter to indicate 69 /// that the loop unroll should be performed regardless of how much 70 /// code expansion would result. 71 static const unsigned NoThreshold = UINT_MAX; 72 73 // Threshold to use when optsize is specified (and there is no 74 // explicit -unroll-threshold). 75 static const unsigned OptSizeUnrollThreshold = 50; 76 77 // Default unroll count for loops with run-time trip count if 78 // -unroll-count is not set 79 static const unsigned UnrollRuntimeCount = 8; 80 81 unsigned CurrentCount; 82 unsigned CurrentThreshold; 83 bool CurrentAllowPartial; 84 bool CurrentRuntime; 85 bool UserCount; // CurrentCount is user-specified. 86 bool UserThreshold; // CurrentThreshold is user-specified. 87 bool UserAllowPartial; // CurrentAllowPartial is user-specified. 88 bool UserRuntime; // CurrentRuntime is user-specified. 89 90 bool runOnLoop(Loop *L, LPPassManager &LPM); 91 92 /// This transformation requires natural loop information & requires that 93 /// loop preheaders be inserted into the CFG... 94 /// 95 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 96 AU.addRequired<LoopInfo>(); 97 AU.addPreserved<LoopInfo>(); 98 AU.addRequiredID(LoopSimplifyID); 99 AU.addPreservedID(LoopSimplifyID); 100 AU.addRequiredID(LCSSAID); 101 AU.addPreservedID(LCSSAID); 102 AU.addRequired<ScalarEvolution>(); 103 AU.addPreserved<ScalarEvolution>(); 104 AU.addRequired<TargetTransformInfo>(); 105 // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. 106 // If loop unroll does not preserve dom info then LCSSA pass on next 107 // loop will receive invalid dom info. 108 // For now, recreate dom info, if loop is unrolled. 109 AU.addPreserved<DominatorTreeWrapperPass>(); 110 } 111 }; 112 } 113 114 char LoopUnroll::ID = 0; 115 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 116 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 117 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 118 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 119 INITIALIZE_PASS_DEPENDENCY(LCSSA) 120 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 121 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) 122 123 Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial, 124 int Runtime) { 125 return new LoopUnroll(Threshold, Count, AllowPartial, Runtime); 126 } 127 128 /// ApproximateLoopSize - Approximate the size of the loop. 129 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls, 130 bool &NotDuplicatable, 131 const TargetTransformInfo &TTI) { 132 CodeMetrics Metrics; 133 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 134 I != E; ++I) 135 Metrics.analyzeBasicBlock(*I, TTI); 136 NumCalls = Metrics.NumInlineCandidates; 137 NotDuplicatable = Metrics.notDuplicatable; 138 139 unsigned LoopSize = Metrics.NumInsts; 140 141 // Don't allow an estimate of size zero. This would allows unrolling of loops 142 // with huge iteration counts, which is a compile time problem even if it's 143 // not a problem for code quality. 144 if (LoopSize == 0) LoopSize = 1; 145 146 return LoopSize; 147 } 148 149 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { 150 LoopInfo *LI = &getAnalysis<LoopInfo>(); 151 ScalarEvolution *SE = &getAnalysis<ScalarEvolution>(); 152 const TargetTransformInfo &TTI = getAnalysis<TargetTransformInfo>(); 153 154 BasicBlock *Header = L->getHeader(); 155 DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() 156 << "] Loop %" << Header->getName() << "\n"); 157 (void)Header; 158 159 TargetTransformInfo::UnrollingPreferences UP; 160 UP.Threshold = CurrentThreshold; 161 UP.OptSizeThreshold = OptSizeUnrollThreshold; 162 UP.Count = CurrentCount; 163 UP.Partial = CurrentAllowPartial; 164 UP.Runtime = CurrentRuntime; 165 TTI.getUnrollingPreferences(L, UP); 166 167 // Determine the current unrolling threshold. While this is normally set 168 // from UnrollThreshold, it is overridden to a smaller value if the current 169 // function is marked as optimize-for-size, and the unroll threshold was 170 // not user specified. 171 unsigned Threshold = UserThreshold ? CurrentThreshold : UP.Threshold; 172 if (!UserThreshold && 173 Header->getParent()->getAttributes(). 174 hasAttribute(AttributeSet::FunctionIndex, 175 Attribute::OptimizeForSize)) 176 Threshold = UP.OptSizeThreshold; 177 178 // Find trip count and trip multiple if count is not available 179 unsigned TripCount = 0; 180 unsigned TripMultiple = 1; 181 // Find "latch trip count". UnrollLoop assumes that control cannot exit 182 // via the loop latch on any iteration prior to TripCount. The loop may exit 183 // early via an earlier branch. 184 BasicBlock *LatchBlock = L->getLoopLatch(); 185 if (LatchBlock) { 186 TripCount = SE->getSmallConstantTripCount(L, LatchBlock); 187 TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock); 188 } 189 190 bool Runtime = UserRuntime ? CurrentRuntime : UP.Runtime; 191 192 // Use a default unroll-count if the user doesn't specify a value 193 // and the trip count is a run-time value. The default is different 194 // for run-time or compile-time trip count loops. 195 unsigned Count = UserCount ? CurrentCount : UP.Count; 196 if (Runtime && Count == 0 && TripCount == 0) 197 Count = UnrollRuntimeCount; 198 199 if (Count == 0) { 200 // Conservative heuristic: if we know the trip count, see if we can 201 // completely unroll (subject to the threshold, checked below); otherwise 202 // try to find greatest modulo of the trip count which is still under 203 // threshold value. 204 if (TripCount == 0) 205 return false; 206 Count = TripCount; 207 } 208 209 // Enforce the threshold. 210 if (Threshold != NoThreshold) { 211 unsigned NumInlineCandidates; 212 bool notDuplicatable; 213 unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates, 214 notDuplicatable, TTI); 215 DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); 216 if (notDuplicatable) { 217 DEBUG(dbgs() << " Not unrolling loop which contains non-duplicatable" 218 << " instructions.\n"); 219 return false; 220 } 221 if (NumInlineCandidates != 0) { 222 DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); 223 return false; 224 } 225 uint64_t Size = (uint64_t)LoopSize*Count; 226 if (TripCount != 1 && Size > Threshold) { 227 DEBUG(dbgs() << " Too large to fully unroll with count: " << Count 228 << " because size: " << Size << ">" << Threshold << "\n"); 229 bool AllowPartial = UserAllowPartial ? CurrentAllowPartial : UP.Partial; 230 if (!AllowPartial && !(Runtime && TripCount == 0)) { 231 DEBUG(dbgs() << " will not try to unroll partially because " 232 << "-unroll-allow-partial not given\n"); 233 return false; 234 } 235 if (TripCount) { 236 // Reduce unroll count to be modulo of TripCount for partial unrolling 237 Count = Threshold / LoopSize; 238 while (Count != 0 && TripCount%Count != 0) 239 Count--; 240 } 241 else if (Runtime) { 242 // Reduce unroll count to be a lower power-of-two value 243 while (Count != 0 && Size > Threshold) { 244 Count >>= 1; 245 Size = LoopSize*Count; 246 } 247 } 248 if (Count < 2) { 249 DEBUG(dbgs() << " could not unroll partially\n"); 250 return false; 251 } 252 DEBUG(dbgs() << " partially unrolling with count: " << Count << "\n"); 253 } 254 } 255 256 // Unroll the loop. 257 if (!UnrollLoop(L, Count, TripCount, Runtime, TripMultiple, LI, &LPM)) 258 return false; 259 260 return true; 261 } 262