1 //===-- LoopSink.cpp - Loop Sink 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 does the inverse transformation of what LICM does. 11 // It traverses all of the instructions in the loop's preheader and sinks 12 // them to the loop body where frequency is lower than the loop's preheader. 13 // This pass is a reverse-transformation of LICM. It differs from the Sink 14 // pass in the following ways: 15 // 16 // * It only handles sinking of instructions from the loop's preheader to the 17 // loop's body 18 // * It uses alias set tracker to get more accurate alias info 19 // * It uses block frequency info to find the optimal sinking locations 20 // 21 // Overall algorithm: 22 // 23 // For I in Preheader: 24 // InsertBBs = BBs that uses I 25 // For BB in sorted(LoopBBs): 26 // DomBBs = BBs in InsertBBs that are dominated by BB 27 // if freq(DomBBs) > freq(BB) 28 // InsertBBs = UseBBs - DomBBs + BB 29 // For BB in InsertBBs: 30 // Insert I at BB's beginning 31 // 32 //===----------------------------------------------------------------------===// 33 34 #include "llvm/Transforms/Scalar/LoopSink.h" 35 #include "llvm/ADT/Statistic.h" 36 #include "llvm/Analysis/AliasAnalysis.h" 37 #include "llvm/Analysis/AliasSetTracker.h" 38 #include "llvm/Analysis/BasicAliasAnalysis.h" 39 #include "llvm/Analysis/BlockFrequencyInfo.h" 40 #include "llvm/Analysis/Loads.h" 41 #include "llvm/Analysis/LoopInfo.h" 42 #include "llvm/Analysis/LoopPass.h" 43 #include "llvm/Analysis/ScalarEvolution.h" 44 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 45 #include "llvm/Transforms/Utils/Local.h" 46 #include "llvm/IR/Dominators.h" 47 #include "llvm/IR/Instructions.h" 48 #include "llvm/IR/LLVMContext.h" 49 #include "llvm/IR/Metadata.h" 50 #include "llvm/Support/CommandLine.h" 51 #include "llvm/Transforms/Scalar.h" 52 #include "llvm/Transforms/Scalar/LoopPassManager.h" 53 #include "llvm/Transforms/Utils/LoopUtils.h" 54 using namespace llvm; 55 56 #define DEBUG_TYPE "loopsink" 57 58 STATISTIC(NumLoopSunk, "Number of instructions sunk into loop"); 59 STATISTIC(NumLoopSunkCloned, "Number of cloned instructions sunk into loop"); 60 61 static cl::opt<unsigned> SinkFrequencyPercentThreshold( 62 "sink-freq-percent-threshold", cl::Hidden, cl::init(90), 63 cl::desc("Do not sink instructions that require cloning unless they " 64 "execute less than this percent of the time.")); 65 66 static cl::opt<unsigned> MaxNumberOfUseBBsForSinking( 67 "max-uses-for-sinking", cl::Hidden, cl::init(30), 68 cl::desc("Do not sink instructions that have too many uses.")); 69 70 /// Return adjusted total frequency of \p BBs. 71 /// 72 /// * If there is only one BB, sinking instruction will not introduce code 73 /// size increase. Thus there is no need to adjust the frequency. 74 /// * If there are more than one BB, sinking would lead to code size increase. 75 /// In this case, we add some "tax" to the total frequency to make it harder 76 /// to sink. E.g. 77 /// Freq(Preheader) = 100 78 /// Freq(BBs) = sum(50, 49) = 99 79 /// Even if Freq(BBs) < Freq(Preheader), we will not sink from Preheade to 80 /// BBs as the difference is too small to justify the code size increase. 81 /// To model this, The adjusted Freq(BBs) will be: 82 /// AdjustedFreq(BBs) = 99 / SinkFrequencyPercentThreshold% 83 static BlockFrequency adjustedSumFreq(SmallPtrSetImpl<BasicBlock *> &BBs, 84 BlockFrequencyInfo &BFI) { 85 BlockFrequency T = 0; 86 for (BasicBlock *B : BBs) 87 T += BFI.getBlockFreq(B); 88 if (BBs.size() > 1) 89 T /= BranchProbability(SinkFrequencyPercentThreshold, 100); 90 return T; 91 } 92 93 /// Return a set of basic blocks to insert sinked instructions. 94 /// 95 /// The returned set of basic blocks (BBsToSinkInto) should satisfy: 96 /// 97 /// * Inside the loop \p L 98 /// * For each UseBB in \p UseBBs, there is at least one BB in BBsToSinkInto 99 /// that domintates the UseBB 100 /// * Has minimum total frequency that is no greater than preheader frequency 101 /// 102 /// The purpose of the function is to find the optimal sinking points to 103 /// minimize execution cost, which is defined as "sum of frequency of 104 /// BBsToSinkInto". 105 /// As a result, the returned BBsToSinkInto needs to have minimum total 106 /// frequency. 107 /// Additionally, if the total frequency of BBsToSinkInto exceeds preheader 108 /// frequency, the optimal solution is not sinking (return empty set). 109 /// 110 /// \p ColdLoopBBs is used to help find the optimal sinking locations. 111 /// It stores a list of BBs that is: 112 /// 113 /// * Inside the loop \p L 114 /// * Has a frequency no larger than the loop's preheader 115 /// * Sorted by BB frequency 116 /// 117 /// The complexity of the function is O(UseBBs.size() * ColdLoopBBs.size()). 118 /// To avoid expensive computation, we cap the maximum UseBBs.size() in its 119 /// caller. 120 static SmallPtrSet<BasicBlock *, 2> 121 findBBsToSinkInto(const Loop &L, const SmallPtrSetImpl<BasicBlock *> &UseBBs, 122 const SmallVectorImpl<BasicBlock *> &ColdLoopBBs, 123 DominatorTree &DT, BlockFrequencyInfo &BFI) { 124 SmallPtrSet<BasicBlock *, 2> BBsToSinkInto; 125 if (UseBBs.size() == 0) 126 return BBsToSinkInto; 127 128 BBsToSinkInto.insert(UseBBs.begin(), UseBBs.end()); 129 SmallPtrSet<BasicBlock *, 2> BBsDominatedByColdestBB; 130 131 // For every iteration: 132 // * Pick the ColdestBB from ColdLoopBBs 133 // * Find the set BBsDominatedByColdestBB that satisfy: 134 // - BBsDominatedByColdestBB is a subset of BBsToSinkInto 135 // - Every BB in BBsDominatedByColdestBB is dominated by ColdestBB 136 // * If Freq(ColdestBB) < Freq(BBsDominatedByColdestBB), remove 137 // BBsDominatedByColdestBB from BBsToSinkInto, add ColdestBB to 138 // BBsToSinkInto 139 for (BasicBlock *ColdestBB : ColdLoopBBs) { 140 BBsDominatedByColdestBB.clear(); 141 for (BasicBlock *SinkedBB : BBsToSinkInto) 142 if (DT.dominates(ColdestBB, SinkedBB)) 143 BBsDominatedByColdestBB.insert(SinkedBB); 144 if (BBsDominatedByColdestBB.size() == 0) 145 continue; 146 if (adjustedSumFreq(BBsDominatedByColdestBB, BFI) > 147 BFI.getBlockFreq(ColdestBB)) { 148 for (BasicBlock *DominatedBB : BBsDominatedByColdestBB) { 149 BBsToSinkInto.erase(DominatedBB); 150 } 151 BBsToSinkInto.insert(ColdestBB); 152 } 153 } 154 155 // If the total frequency of BBsToSinkInto is larger than preheader frequency, 156 // do not sink. 157 if (adjustedSumFreq(BBsToSinkInto, BFI) > 158 BFI.getBlockFreq(L.getLoopPreheader())) 159 BBsToSinkInto.clear(); 160 return BBsToSinkInto; 161 } 162 163 // Sinks \p I from the loop \p L's preheader to its uses. Returns true if 164 // sinking is successful. 165 // \p LoopBlockNumber is used to sort the insertion blocks to ensure 166 // determinism. 167 static bool sinkInstruction(Loop &L, Instruction &I, 168 const SmallVectorImpl<BasicBlock *> &ColdLoopBBs, 169 const SmallDenseMap<BasicBlock *, int, 16> &LoopBlockNumber, 170 LoopInfo &LI, DominatorTree &DT, 171 BlockFrequencyInfo &BFI) { 172 // Compute the set of blocks in loop L which contain a use of I. 173 SmallPtrSet<BasicBlock *, 2> BBs; 174 for (auto &U : I.uses()) { 175 Instruction *UI = cast<Instruction>(U.getUser()); 176 // We cannot sink I to PHI-uses. 177 if (dyn_cast<PHINode>(UI)) 178 return false; 179 // We cannot sink I if it has uses outside of the loop. 180 if (!L.contains(LI.getLoopFor(UI->getParent()))) 181 return false; 182 BBs.insert(UI->getParent()); 183 } 184 185 // findBBsToSinkInto is O(BBs.size() * ColdLoopBBs.size()). We cap the max 186 // BBs.size() to avoid expensive computation. 187 // FIXME: Handle code size growth for min_size and opt_size. 188 if (BBs.size() > MaxNumberOfUseBBsForSinking) 189 return false; 190 191 // Find the set of BBs that we should insert a copy of I. 192 SmallPtrSet<BasicBlock *, 2> BBsToSinkInto = 193 findBBsToSinkInto(L, BBs, ColdLoopBBs, DT, BFI); 194 if (BBsToSinkInto.empty()) 195 return false; 196 197 // Copy the final BBs into a vector and sort them using the total ordering 198 // of the loop block numbers as iterating the set doesn't give a useful 199 // order. No need to stable sort as the block numbers are a total ordering. 200 SmallVector<BasicBlock *, 2> SortedBBsToSinkInto; 201 SortedBBsToSinkInto.insert(SortedBBsToSinkInto.begin(), BBsToSinkInto.begin(), 202 BBsToSinkInto.end()); 203 llvm::sort(SortedBBsToSinkInto.begin(), SortedBBsToSinkInto.end(), 204 [&](BasicBlock *A, BasicBlock *B) { 205 return LoopBlockNumber.find(A)->second < 206 LoopBlockNumber.find(B)->second; 207 }); 208 209 BasicBlock *MoveBB = *SortedBBsToSinkInto.begin(); 210 // FIXME: Optimize the efficiency for cloned value replacement. The current 211 // implementation is O(SortedBBsToSinkInto.size() * I.num_uses()). 212 for (BasicBlock *N : makeArrayRef(SortedBBsToSinkInto).drop_front(1)) { 213 assert(LoopBlockNumber.find(N)->second > 214 LoopBlockNumber.find(MoveBB)->second && 215 "BBs not sorted!"); 216 // Clone I and replace its uses. 217 Instruction *IC = I.clone(); 218 IC->setName(I.getName()); 219 IC->insertBefore(&*N->getFirstInsertionPt()); 220 // Replaces uses of I with IC in N 221 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;) { 222 Use &U = *UI++; 223 auto *I = cast<Instruction>(U.getUser()); 224 if (I->getParent() == N) 225 U.set(IC); 226 } 227 // Replaces uses of I with IC in blocks dominated by N 228 replaceDominatedUsesWith(&I, IC, DT, N); 229 LLVM_DEBUG(dbgs() << "Sinking a clone of " << I << " To: " << N->getName() 230 << '\n'); 231 NumLoopSunkCloned++; 232 } 233 LLVM_DEBUG(dbgs() << "Sinking " << I << " To: " << MoveBB->getName() << '\n'); 234 NumLoopSunk++; 235 I.moveBefore(&*MoveBB->getFirstInsertionPt()); 236 237 return true; 238 } 239 240 /// Sinks instructions from loop's preheader to the loop body if the 241 /// sum frequency of inserted copy is smaller than preheader's frequency. 242 static bool sinkLoopInvariantInstructions(Loop &L, AAResults &AA, LoopInfo &LI, 243 DominatorTree &DT, 244 BlockFrequencyInfo &BFI, 245 ScalarEvolution *SE) { 246 BasicBlock *Preheader = L.getLoopPreheader(); 247 if (!Preheader) 248 return false; 249 250 // Enable LoopSink only when runtime profile is available. 251 // With static profile, the sinking decision may be sub-optimal. 252 if (!Preheader->getParent()->hasProfileData()) 253 return false; 254 255 const BlockFrequency PreheaderFreq = BFI.getBlockFreq(Preheader); 256 // If there are no basic blocks with lower frequency than the preheader then 257 // we can avoid the detailed analysis as we will never find profitable sinking 258 // opportunities. 259 if (all_of(L.blocks(), [&](const BasicBlock *BB) { 260 return BFI.getBlockFreq(BB) > PreheaderFreq; 261 })) 262 return false; 263 264 bool Changed = false; 265 AliasSetTracker CurAST(AA); 266 267 // Compute alias set. 268 for (BasicBlock *BB : L.blocks()) 269 CurAST.add(*BB); 270 271 // Sort loop's basic blocks by frequency 272 SmallVector<BasicBlock *, 10> ColdLoopBBs; 273 SmallDenseMap<BasicBlock *, int, 16> LoopBlockNumber; 274 int i = 0; 275 for (BasicBlock *B : L.blocks()) 276 if (BFI.getBlockFreq(B) < BFI.getBlockFreq(L.getLoopPreheader())) { 277 ColdLoopBBs.push_back(B); 278 LoopBlockNumber[B] = ++i; 279 } 280 std::stable_sort(ColdLoopBBs.begin(), ColdLoopBBs.end(), 281 [&](BasicBlock *A, BasicBlock *B) { 282 return BFI.getBlockFreq(A) < BFI.getBlockFreq(B); 283 }); 284 285 // Traverse preheader's instructions in reverse order becaue if A depends 286 // on B (A appears after B), A needs to be sinked first before B can be 287 // sinked. 288 for (auto II = Preheader->rbegin(), E = Preheader->rend(); II != E;) { 289 Instruction *I = &*II++; 290 // No need to check for instruction's operands are loop invariant. 291 assert(L.hasLoopInvariantOperands(I) && 292 "Insts in a loop's preheader should have loop invariant operands!"); 293 if (!canSinkOrHoistInst(*I, &AA, &DT, &L, &CurAST, nullptr)) 294 continue; 295 if (sinkInstruction(L, *I, ColdLoopBBs, LoopBlockNumber, LI, DT, BFI)) 296 Changed = true; 297 } 298 299 if (Changed && SE) 300 SE->forgetLoopDispositions(&L); 301 return Changed; 302 } 303 304 PreservedAnalyses LoopSinkPass::run(Function &F, FunctionAnalysisManager &FAM) { 305 LoopInfo &LI = FAM.getResult<LoopAnalysis>(F); 306 // Nothing to do if there are no loops. 307 if (LI.empty()) 308 return PreservedAnalyses::all(); 309 310 AAResults &AA = FAM.getResult<AAManager>(F); 311 DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F); 312 BlockFrequencyInfo &BFI = FAM.getResult<BlockFrequencyAnalysis>(F); 313 314 // We want to do a postorder walk over the loops. Since loops are a tree this 315 // is equivalent to a reversed preorder walk and preorder is easy to compute 316 // without recursion. Since we reverse the preorder, we will visit siblings 317 // in reverse program order. This isn't expected to matter at all but is more 318 // consistent with sinking algorithms which generally work bottom-up. 319 SmallVector<Loop *, 4> PreorderLoops = LI.getLoopsInPreorder(); 320 321 bool Changed = false; 322 do { 323 Loop &L = *PreorderLoops.pop_back_val(); 324 325 // Note that we don't pass SCEV here because it is only used to invalidate 326 // loops in SCEV and we don't preserve (or request) SCEV at all making that 327 // unnecessary. 328 Changed |= sinkLoopInvariantInstructions(L, AA, LI, DT, BFI, 329 /*ScalarEvolution*/ nullptr); 330 } while (!PreorderLoops.empty()); 331 332 if (!Changed) 333 return PreservedAnalyses::all(); 334 335 PreservedAnalyses PA; 336 PA.preserveSet<CFGAnalyses>(); 337 return PA; 338 } 339 340 namespace { 341 struct LegacyLoopSinkPass : public LoopPass { 342 static char ID; 343 LegacyLoopSinkPass() : LoopPass(ID) { 344 initializeLegacyLoopSinkPassPass(*PassRegistry::getPassRegistry()); 345 } 346 347 bool runOnLoop(Loop *L, LPPassManager &LPM) override { 348 if (skipLoop(L)) 349 return false; 350 351 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); 352 return sinkLoopInvariantInstructions( 353 *L, getAnalysis<AAResultsWrapperPass>().getAAResults(), 354 getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), 355 getAnalysis<DominatorTreeWrapperPass>().getDomTree(), 356 getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(), 357 SE ? &SE->getSE() : nullptr); 358 } 359 360 void getAnalysisUsage(AnalysisUsage &AU) const override { 361 AU.setPreservesCFG(); 362 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 363 getLoopAnalysisUsage(AU); 364 } 365 }; 366 } 367 368 char LegacyLoopSinkPass::ID = 0; 369 INITIALIZE_PASS_BEGIN(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false, 370 false) 371 INITIALIZE_PASS_DEPENDENCY(LoopPass) 372 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 373 INITIALIZE_PASS_END(LegacyLoopSinkPass, "loop-sink", "Loop Sink", false, false) 374 375 Pass *llvm::createLoopSinkPass() { return new LegacyLoopSinkPass(); } 376