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 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 90 namespace llvm { 91 92 static GVDAGType getGVDT() { 93 if (PGOViewCounts == PGOVCT_Graph) 94 return GVDT_Count; 95 return ViewBlockFreqPropagationDAG; 96 } 97 98 template <> 99 struct GraphTraits<BlockFrequencyInfo *> { 100 using NodeRef = const BasicBlock *; 101 using ChildIteratorType = const_succ_iterator; 102 using nodes_iterator = pointer_iterator<Function::const_iterator>; 103 104 static NodeRef getEntryNode(const BlockFrequencyInfo *G) { 105 return &G->getFunction()->front(); 106 } 107 108 static ChildIteratorType child_begin(const NodeRef N) { 109 return succ_begin(N); 110 } 111 112 static ChildIteratorType child_end(const NodeRef N) { return succ_end(N); } 113 114 static nodes_iterator nodes_begin(const BlockFrequencyInfo *G) { 115 return nodes_iterator(G->getFunction()->begin()); 116 } 117 118 static nodes_iterator nodes_end(const BlockFrequencyInfo *G) { 119 return nodes_iterator(G->getFunction()->end()); 120 } 121 }; 122 123 using BFIDOTGTraitsBase = 124 BFIDOTGraphTraitsBase<BlockFrequencyInfo, BranchProbabilityInfo>; 125 126 template <> 127 struct DOTGraphTraits<BlockFrequencyInfo *> : public BFIDOTGTraitsBase { 128 explicit DOTGraphTraits(bool isSimple = false) 129 : BFIDOTGTraitsBase(isSimple) {} 130 131 std::string getNodeLabel(const BasicBlock *Node, 132 const BlockFrequencyInfo *Graph) { 133 134 return BFIDOTGTraitsBase::getNodeLabel(Node, Graph, getGVDT()); 135 } 136 137 std::string getNodeAttributes(const BasicBlock *Node, 138 const BlockFrequencyInfo *Graph) { 139 return BFIDOTGTraitsBase::getNodeAttributes(Node, Graph, 140 ViewHotFreqPercent); 141 } 142 143 std::string getEdgeAttributes(const BasicBlock *Node, EdgeIter EI, 144 const BlockFrequencyInfo *BFI) { 145 return BFIDOTGTraitsBase::getEdgeAttributes(Node, EI, BFI, BFI->getBPI(), 146 ViewHotFreqPercent); 147 } 148 }; 149 150 } // end namespace llvm 151 152 BlockFrequencyInfo::BlockFrequencyInfo() = default; 153 154 BlockFrequencyInfo::BlockFrequencyInfo(const Function &F, 155 const BranchProbabilityInfo &BPI, 156 const LoopInfo &LI) { 157 calculate(F, BPI, LI); 158 } 159 160 BlockFrequencyInfo::BlockFrequencyInfo(BlockFrequencyInfo &&Arg) 161 : BFI(std::move(Arg.BFI)) {} 162 163 BlockFrequencyInfo &BlockFrequencyInfo::operator=(BlockFrequencyInfo &&RHS) { 164 releaseMemory(); 165 BFI = std::move(RHS.BFI); 166 return *this; 167 } 168 169 // Explicitly define the default constructor otherwise it would be implicitly 170 // defined at the first ODR-use which is the BFI member in the 171 // LazyBlockFrequencyInfo header. The dtor needs the BlockFrequencyInfoImpl 172 // template instantiated which is not available in the header. 173 BlockFrequencyInfo::~BlockFrequencyInfo() = default; 174 175 bool BlockFrequencyInfo::invalidate(Function &F, const PreservedAnalyses &PA, 176 FunctionAnalysisManager::Invalidator &) { 177 // Check whether the analysis, all analyses on functions, or the function's 178 // CFG have been preserved. 179 auto PAC = PA.getChecker<BlockFrequencyAnalysis>(); 180 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() || 181 PAC.preservedSet<CFGAnalyses>()); 182 } 183 184 void BlockFrequencyInfo::calculate(const Function &F, 185 const BranchProbabilityInfo &BPI, 186 const LoopInfo &LI) { 187 if (!BFI) 188 BFI.reset(new ImplType); 189 BFI->calculate(F, BPI, LI); 190 if (ViewBlockFreqPropagationDAG != GVDT_None && 191 (ViewBlockFreqFuncName.empty() || 192 F.getName().equals(ViewBlockFreqFuncName))) { 193 view(); 194 } 195 if (PrintBlockFreq && 196 (PrintBlockFreqFuncName.empty() || 197 F.getName().equals(PrintBlockFreqFuncName))) { 198 print(dbgs()); 199 } 200 } 201 202 BlockFrequency BlockFrequencyInfo::getBlockFreq(const BasicBlock *BB) const { 203 return BFI ? BFI->getBlockFreq(BB) : 0; 204 } 205 206 Optional<uint64_t> 207 BlockFrequencyInfo::getBlockProfileCount(const BasicBlock *BB, 208 bool AllowSynthetic) const { 209 if (!BFI) 210 return None; 211 212 return BFI->getBlockProfileCount(*getFunction(), BB, AllowSynthetic); 213 } 214 215 Optional<uint64_t> 216 BlockFrequencyInfo::getProfileCountFromFreq(uint64_t Freq) const { 217 if (!BFI) 218 return None; 219 return BFI->getProfileCountFromFreq(*getFunction(), Freq); 220 } 221 222 bool BlockFrequencyInfo::isIrrLoopHeader(const BasicBlock *BB) { 223 assert(BFI && "Expected analysis to be available"); 224 return BFI->isIrrLoopHeader(BB); 225 } 226 227 void BlockFrequencyInfo::setBlockFreq(const BasicBlock *BB, uint64_t Freq) { 228 assert(BFI && "Expected analysis to be available"); 229 BFI->setBlockFreq(BB, Freq); 230 } 231 232 void BlockFrequencyInfo::setBlockFreqAndScale( 233 const BasicBlock *ReferenceBB, uint64_t Freq, 234 SmallPtrSetImpl<BasicBlock *> &BlocksToScale) { 235 assert(BFI && "Expected analysis to be available"); 236 // Use 128 bits APInt to avoid overflow. 237 APInt NewFreq(128, Freq); 238 APInt OldFreq(128, BFI->getBlockFreq(ReferenceBB).getFrequency()); 239 APInt BBFreq(128, 0); 240 for (auto *BB : BlocksToScale) { 241 BBFreq = BFI->getBlockFreq(BB).getFrequency(); 242 // Multiply first by NewFreq and then divide by OldFreq 243 // to minimize loss of precision. 244 BBFreq *= NewFreq; 245 // udiv is an expensive operation in the general case. If this ends up being 246 // a hot spot, one of the options proposed in 247 // https://reviews.llvm.org/D28535#650071 could be used to avoid this. 248 BBFreq = BBFreq.udiv(OldFreq); 249 BFI->setBlockFreq(BB, BBFreq.getLimitedValue()); 250 } 251 BFI->setBlockFreq(ReferenceBB, Freq); 252 } 253 254 /// Pop up a ghostview window with the current block frequency propagation 255 /// rendered using dot. 256 void BlockFrequencyInfo::view(StringRef title) const { 257 ViewGraph(const_cast<BlockFrequencyInfo *>(this), title); 258 } 259 260 const Function *BlockFrequencyInfo::getFunction() const { 261 return BFI ? BFI->getFunction() : nullptr; 262 } 263 264 const BranchProbabilityInfo *BlockFrequencyInfo::getBPI() const { 265 return BFI ? &BFI->getBPI() : nullptr; 266 } 267 268 raw_ostream &BlockFrequencyInfo:: 269 printBlockFreq(raw_ostream &OS, const BlockFrequency Freq) const { 270 return BFI ? BFI->printBlockFreq(OS, Freq) : OS; 271 } 272 273 raw_ostream & 274 BlockFrequencyInfo::printBlockFreq(raw_ostream &OS, 275 const BasicBlock *BB) const { 276 return BFI ? BFI->printBlockFreq(OS, BB) : OS; 277 } 278 279 uint64_t BlockFrequencyInfo::getEntryFreq() const { 280 return BFI ? BFI->getEntryFreq() : 0; 281 } 282 283 void BlockFrequencyInfo::releaseMemory() { BFI.reset(); } 284 285 void BlockFrequencyInfo::print(raw_ostream &OS) const { 286 if (BFI) 287 BFI->print(OS); 288 } 289 290 void BlockFrequencyInfo::verifyMatch(BlockFrequencyInfo &Other) const { 291 if (BFI) 292 BFI->verifyMatch(*Other.BFI); 293 } 294 295 INITIALIZE_PASS_BEGIN(BlockFrequencyInfoWrapperPass, "block-freq", 296 "Block Frequency Analysis", true, true) 297 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass) 298 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 299 INITIALIZE_PASS_END(BlockFrequencyInfoWrapperPass, "block-freq", 300 "Block Frequency Analysis", true, true) 301 302 char BlockFrequencyInfoWrapperPass::ID = 0; 303 304 BlockFrequencyInfoWrapperPass::BlockFrequencyInfoWrapperPass() 305 : FunctionPass(ID) { 306 initializeBlockFrequencyInfoWrapperPassPass(*PassRegistry::getPassRegistry()); 307 } 308 309 BlockFrequencyInfoWrapperPass::~BlockFrequencyInfoWrapperPass() = default; 310 311 void BlockFrequencyInfoWrapperPass::print(raw_ostream &OS, 312 const Module *) const { 313 BFI.print(OS); 314 } 315 316 void BlockFrequencyInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 317 AU.addRequired<BranchProbabilityInfoWrapperPass>(); 318 AU.addRequired<LoopInfoWrapperPass>(); 319 AU.setPreservesAll(); 320 } 321 322 void BlockFrequencyInfoWrapperPass::releaseMemory() { BFI.releaseMemory(); } 323 324 bool BlockFrequencyInfoWrapperPass::runOnFunction(Function &F) { 325 BranchProbabilityInfo &BPI = 326 getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI(); 327 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 328 BFI.calculate(F, BPI, LI); 329 return false; 330 } 331 332 AnalysisKey BlockFrequencyAnalysis::Key; 333 BlockFrequencyInfo BlockFrequencyAnalysis::run(Function &F, 334 FunctionAnalysisManager &AM) { 335 BlockFrequencyInfo BFI; 336 BFI.calculate(F, AM.getResult<BranchProbabilityAnalysis>(F), 337 AM.getResult<LoopAnalysis>(F)); 338 return BFI; 339 } 340 341 PreservedAnalyses 342 BlockFrequencyPrinterPass::run(Function &F, FunctionAnalysisManager &AM) { 343 OS << "Printing analysis results of BFI for function " 344 << "'" << F.getName() << "':" 345 << "\n"; 346 AM.getResult<BlockFrequencyAnalysis>(F).print(OS); 347 return PreservedAnalyses::all(); 348 } 349