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