1 //===- BlockFrequencyInfo.cpp - Block Frequency Analysis ------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // Loops should be simplified before this analysis. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/Analysis/BlockFrequencyInfo.h" 14 #include "llvm/ADT/APInt.h" 15 #include "llvm/ADT/iterator.h" 16 #include "llvm/Analysis/BlockFrequencyInfoImpl.h" 17 #include "llvm/Analysis/BranchProbabilityInfo.h" 18 #include "llvm/Analysis/LoopInfo.h" 19 #include "llvm/IR/CFG.h" 20 #include "llvm/IR/Function.h" 21 #include "llvm/IR/PassManager.h" 22 #include "llvm/InitializePasses.h" 23 #include "llvm/Pass.h" 24 #include "llvm/Support/CommandLine.h" 25 #include "llvm/Support/GraphWriter.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include <cassert> 28 #include <optional> 29 #include <string> 30 31 using namespace llvm; 32 33 #define DEBUG_TYPE "block-freq" 34 35 static cl::opt<GVDAGType> ViewBlockFreqPropagationDAG( 36 "view-block-freq-propagation-dags", cl::Hidden, 37 cl::desc("Pop up a window to show a dag displaying how block " 38 "frequencies propagation through the CFG."), 39 cl::values(clEnumValN(GVDT_None, "none", "do not display graphs."), 40 clEnumValN(GVDT_Fraction, "fraction", 41 "display a graph using the " 42 "fractional block frequency representation."), 43 clEnumValN(GVDT_Integer, "integer", 44 "display a graph using the raw " 45 "integer fractional block frequency representation."), 46 clEnumValN(GVDT_Count, "count", "display a graph using the real " 47 "profile count if available."))); 48 49 namespace llvm { 50 cl::opt<std::string> 51 ViewBlockFreqFuncName("view-bfi-func-name", cl::Hidden, 52 cl::desc("The option to specify " 53 "the name of the function " 54 "whose CFG will be displayed.")); 55 56 cl::opt<unsigned> 57 ViewHotFreqPercent("view-hot-freq-percent", cl::init(10), cl::Hidden, 58 cl::desc("An integer in percent used to specify " 59 "the hot blocks/edges to be displayed " 60 "in red: a block or edge whose frequency " 61 "is no less than the max frequency of the " 62 "function multiplied by this percent.")); 63 64 // Command line option to turn on CFG dot or text dump after profile annotation. 65 cl::opt<PGOViewCountsType> PGOViewCounts( 66 "pgo-view-counts", cl::Hidden, 67 cl::desc("A boolean option to show CFG dag or text with " 68 "block profile counts and branch probabilities " 69 "right after PGO profile annotation step. The " 70 "profile counts are computed using branch " 71 "probabilities from the runtime profile data and " 72 "block frequency propagation algorithm. To view " 73 "the raw counts from the profile, use option " 74 "-pgo-view-raw-counts instead. To limit graph " 75 "display to only one function, use filtering option " 76 "-view-bfi-func-name."), 77 cl::values(clEnumValN(PGOVCT_None, "none", "do not show."), 78 clEnumValN(PGOVCT_Graph, "graph", "show a graph."), 79 clEnumValN(PGOVCT_Text, "text", "show in text."))); 80 81 static cl::opt<bool> PrintBlockFreq( 82 "print-bfi", cl::init(false), cl::Hidden, 83 cl::desc("Print the block frequency info.")); 84 85 cl::opt<std::string> PrintBlockFreqFuncName( 86 "print-bfi-func-name", cl::Hidden, 87 cl::desc("The option to specify the name of the function " 88 "whose block frequency info is printed.")); 89 } // namespace llvm 90 91 namespace llvm { 92 93 static GVDAGType getGVDT() { 94 if (PGOViewCounts == PGOVCT_Graph) 95 return GVDT_Count; 96 return ViewBlockFreqPropagationDAG; 97 } 98 99 template <> 100 struct GraphTraits<BlockFrequencyInfo *> { 101 using NodeRef = const BasicBlock *; 102 using ChildIteratorType = const_succ_iterator; 103 using nodes_iterator = pointer_iterator<Function::const_iterator>; 104 105 static NodeRef getEntryNode(const BlockFrequencyInfo *G) { 106 return &G->getFunction()->front(); 107 } 108 109 static ChildIteratorType child_begin(const NodeRef N) { 110 return succ_begin(N); 111 } 112 113 static ChildIteratorType child_end(const NodeRef N) { return succ_end(N); } 114 115 static nodes_iterator nodes_begin(const BlockFrequencyInfo *G) { 116 return nodes_iterator(G->getFunction()->begin()); 117 } 118 119 static nodes_iterator nodes_end(const BlockFrequencyInfo *G) { 120 return nodes_iterator(G->getFunction()->end()); 121 } 122 }; 123 124 using BFIDOTGTraitsBase = 125 BFIDOTGraphTraitsBase<BlockFrequencyInfo, BranchProbabilityInfo>; 126 127 template <> 128 struct DOTGraphTraits<BlockFrequencyInfo *> : public BFIDOTGTraitsBase { 129 explicit DOTGraphTraits(bool isSimple = false) 130 : BFIDOTGTraitsBase(isSimple) {} 131 132 std::string getNodeLabel(const BasicBlock *Node, 133 const BlockFrequencyInfo *Graph) { 134 135 return BFIDOTGTraitsBase::getNodeLabel(Node, Graph, getGVDT()); 136 } 137 138 std::string getNodeAttributes(const BasicBlock *Node, 139 const BlockFrequencyInfo *Graph) { 140 return BFIDOTGTraitsBase::getNodeAttributes(Node, Graph, 141 ViewHotFreqPercent); 142 } 143 144 std::string getEdgeAttributes(const BasicBlock *Node, EdgeIter EI, 145 const BlockFrequencyInfo *BFI) { 146 return BFIDOTGTraitsBase::getEdgeAttributes(Node, EI, BFI, BFI->getBPI(), 147 ViewHotFreqPercent); 148 } 149 }; 150 151 } // end namespace llvm 152 153 BlockFrequencyInfo::BlockFrequencyInfo() = default; 154 155 BlockFrequencyInfo::BlockFrequencyInfo(const Function &F, 156 const BranchProbabilityInfo &BPI, 157 const LoopInfo &LI) { 158 calculate(F, BPI, LI); 159 } 160 161 BlockFrequencyInfo::BlockFrequencyInfo(BlockFrequencyInfo &&Arg) 162 : BFI(std::move(Arg.BFI)) {} 163 164 BlockFrequencyInfo &BlockFrequencyInfo::operator=(BlockFrequencyInfo &&RHS) { 165 releaseMemory(); 166 BFI = std::move(RHS.BFI); 167 return *this; 168 } 169 170 // Explicitly define the default constructor otherwise it would be implicitly 171 // defined at the first ODR-use which is the BFI member in the 172 // LazyBlockFrequencyInfo header. The dtor needs the BlockFrequencyInfoImpl 173 // template instantiated which is not available in the header. 174 BlockFrequencyInfo::~BlockFrequencyInfo() = default; 175 176 bool BlockFrequencyInfo::invalidate(Function &F, const PreservedAnalyses &PA, 177 FunctionAnalysisManager::Invalidator &) { 178 // Check whether the analysis, all analyses on functions, or the function's 179 // CFG have been preserved. 180 auto PAC = PA.getChecker<BlockFrequencyAnalysis>(); 181 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() || 182 PAC.preservedSet<CFGAnalyses>()); 183 } 184 185 void BlockFrequencyInfo::calculate(const Function &F, 186 const BranchProbabilityInfo &BPI, 187 const LoopInfo &LI) { 188 if (!BFI) 189 BFI.reset(new ImplType); 190 BFI->calculate(F, BPI, LI); 191 if (ViewBlockFreqPropagationDAG != GVDT_None && 192 (ViewBlockFreqFuncName.empty() || 193 F.getName().equals(ViewBlockFreqFuncName))) { 194 view(); 195 } 196 if (PrintBlockFreq && 197 (PrintBlockFreqFuncName.empty() || 198 F.getName().equals(PrintBlockFreqFuncName))) { 199 print(dbgs()); 200 } 201 } 202 203 BlockFrequency BlockFrequencyInfo::getBlockFreq(const BasicBlock *BB) const { 204 return BFI ? BFI->getBlockFreq(BB) : 0; 205 } 206 207 std::optional<uint64_t> 208 BlockFrequencyInfo::getBlockProfileCount(const BasicBlock *BB, 209 bool AllowSynthetic) const { 210 if (!BFI) 211 return std::nullopt; 212 213 return BFI->getBlockProfileCount(*getFunction(), BB, AllowSynthetic); 214 } 215 216 std::optional<uint64_t> 217 BlockFrequencyInfo::getProfileCountFromFreq(uint64_t Freq) const { 218 if (!BFI) 219 return std::nullopt; 220 return BFI->getProfileCountFromFreq(*getFunction(), Freq); 221 } 222 223 bool BlockFrequencyInfo::isIrrLoopHeader(const BasicBlock *BB) { 224 assert(BFI && "Expected analysis to be available"); 225 return BFI->isIrrLoopHeader(BB); 226 } 227 228 void BlockFrequencyInfo::setBlockFreq(const BasicBlock *BB, uint64_t Freq) { 229 assert(BFI && "Expected analysis to be available"); 230 BFI->setBlockFreq(BB, Freq); 231 } 232 233 void BlockFrequencyInfo::setBlockFreqAndScale( 234 const BasicBlock *ReferenceBB, uint64_t Freq, 235 SmallPtrSetImpl<BasicBlock *> &BlocksToScale) { 236 assert(BFI && "Expected analysis to be available"); 237 // Use 128 bits APInt to avoid overflow. 238 APInt NewFreq(128, Freq); 239 APInt OldFreq(128, BFI->getBlockFreq(ReferenceBB).getFrequency()); 240 APInt BBFreq(128, 0); 241 for (auto *BB : BlocksToScale) { 242 BBFreq = BFI->getBlockFreq(BB).getFrequency(); 243 // Multiply first by NewFreq and then divide by OldFreq 244 // to minimize loss of precision. 245 BBFreq *= NewFreq; 246 // udiv is an expensive operation in the general case. If this ends up being 247 // a hot spot, one of the options proposed in 248 // https://reviews.llvm.org/D28535#650071 could be used to avoid this. 249 BBFreq = BBFreq.udiv(OldFreq); 250 BFI->setBlockFreq(BB, BBFreq.getLimitedValue()); 251 } 252 BFI->setBlockFreq(ReferenceBB, Freq); 253 } 254 255 /// Pop up a ghostview window with the current block frequency propagation 256 /// rendered using dot. 257 void BlockFrequencyInfo::view(StringRef title) const { 258 ViewGraph(const_cast<BlockFrequencyInfo *>(this), title); 259 } 260 261 const Function *BlockFrequencyInfo::getFunction() const { 262 return BFI ? BFI->getFunction() : nullptr; 263 } 264 265 const BranchProbabilityInfo *BlockFrequencyInfo::getBPI() const { 266 return BFI ? &BFI->getBPI() : nullptr; 267 } 268 269 raw_ostream &BlockFrequencyInfo:: 270 printBlockFreq(raw_ostream &OS, const BlockFrequency Freq) const { 271 return BFI ? BFI->printBlockFreq(OS, Freq) : OS; 272 } 273 274 raw_ostream & 275 BlockFrequencyInfo::printBlockFreq(raw_ostream &OS, 276 const BasicBlock *BB) const { 277 return BFI ? BFI->printBlockFreq(OS, BB) : OS; 278 } 279 280 uint64_t BlockFrequencyInfo::getEntryFreq() const { 281 return BFI ? BFI->getEntryFreq() : 0; 282 } 283 284 void BlockFrequencyInfo::releaseMemory() { BFI.reset(); } 285 286 void BlockFrequencyInfo::print(raw_ostream &OS) const { 287 if (BFI) 288 BFI->print(OS); 289 } 290 291 void BlockFrequencyInfo::verifyMatch(BlockFrequencyInfo &Other) const { 292 if (BFI) 293 BFI->verifyMatch(*Other.BFI); 294 } 295 296 INITIALIZE_PASS_BEGIN(BlockFrequencyInfoWrapperPass, "block-freq", 297 "Block Frequency Analysis", true, true) 298 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass) 299 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 300 INITIALIZE_PASS_END(BlockFrequencyInfoWrapperPass, "block-freq", 301 "Block Frequency Analysis", true, true) 302 303 char BlockFrequencyInfoWrapperPass::ID = 0; 304 305 BlockFrequencyInfoWrapperPass::BlockFrequencyInfoWrapperPass() 306 : FunctionPass(ID) { 307 initializeBlockFrequencyInfoWrapperPassPass(*PassRegistry::getPassRegistry()); 308 } 309 310 BlockFrequencyInfoWrapperPass::~BlockFrequencyInfoWrapperPass() = default; 311 312 void BlockFrequencyInfoWrapperPass::print(raw_ostream &OS, 313 const Module *) const { 314 BFI.print(OS); 315 } 316 317 void BlockFrequencyInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 318 AU.addRequired<BranchProbabilityInfoWrapperPass>(); 319 AU.addRequired<LoopInfoWrapperPass>(); 320 AU.setPreservesAll(); 321 } 322 323 void BlockFrequencyInfoWrapperPass::releaseMemory() { BFI.releaseMemory(); } 324 325 bool BlockFrequencyInfoWrapperPass::runOnFunction(Function &F) { 326 BranchProbabilityInfo &BPI = 327 getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI(); 328 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 329 BFI.calculate(F, BPI, LI); 330 return false; 331 } 332 333 AnalysisKey BlockFrequencyAnalysis::Key; 334 BlockFrequencyInfo BlockFrequencyAnalysis::run(Function &F, 335 FunctionAnalysisManager &AM) { 336 auto &BP = AM.getResult<BranchProbabilityAnalysis>(F); 337 auto &LI = AM.getResult<LoopAnalysis>(F); 338 BlockFrequencyInfo BFI; 339 BFI.calculate(F, BP, LI); 340 return BFI; 341 } 342 343 PreservedAnalyses 344 BlockFrequencyPrinterPass::run(Function &F, FunctionAnalysisManager &AM) { 345 OS << "Printing analysis results of BFI for function " 346 << "'" << F.getName() << "':" 347 << "\n"; 348 AM.getResult<BlockFrequencyAnalysis>(F).print(OS); 349 return PreservedAnalyses::all(); 350 } 351