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