tools/llvm-exegesis/llvm-exegesis.cpp

   1 //===-- llvm-exegesis.cpp ---------------------------------------*- C++ -*-===//
   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 /// \file
  10 /// Measures execution properties (latencies/uops) of an instruction.
  11 ///
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "lib/Analysis.h"
  15 #include "lib/BenchmarkResult.h"
  16 #include "lib/BenchmarkRunner.h"
  17 #include "lib/Clustering.h"
  18 #include "lib/Error.h"
  19 #include "lib/LlvmState.h"
  20 #include "lib/PerfHelper.h"
  21 #include "lib/SnippetFile.h"
  22 #include "lib/SnippetRepetitor.h"
  23 #include "lib/Target.h"
  24 #include "lib/TargetSelect.h"
  25 #include "llvm/ADT/StringExtras.h"
  26 #include "llvm/ADT/Twine.h"
  27 #include "llvm/MC/MCInstBuilder.h"
  28 #include "llvm/MC/MCObjectFileInfo.h"
  29 #include "llvm/MC/MCParser/MCAsmParser.h"
  30 #include "llvm/MC/MCParser/MCTargetAsmParser.h"
  31 #include "llvm/MC/MCRegisterInfo.h"
  32 #include "llvm/MC/MCSubtargetInfo.h"
  33 #include "llvm/Object/ObjectFile.h"
  34 #include "llvm/Support/CommandLine.h"
  35 #include "llvm/Support/Format.h"
  36 #include "llvm/Support/Path.h"
  37 #include "llvm/Support/SourceMgr.h"
  38 #include "llvm/Support/TargetRegistry.h"
  39 #include "llvm/Support/TargetSelect.h"
  40 #include <algorithm>
  41 #include <string>
  42
  43 namespace llvm {
  44 namespace exegesis {
  45
  46 static cl::OptionCategory Options("llvm-exegesis options");
  47 static cl::OptionCategory BenchmarkOptions("llvm-exegesis benchmark options");
  48 static cl::OptionCategory AnalysisOptions("llvm-exegesis analysis options");
  49
  50 static cl::opt<int> OpcodeIndex("opcode-index",
  51                                 cl::desc("opcode to measure, by index"),
  52                                 cl::cat(BenchmarkOptions), cl::init(0));
  53
  54 static cl::opt<std::string>
  55     OpcodeNames("opcode-name",
  56                 cl::desc("comma-separated list of opcodes to measure, by name"),
  57                 cl::cat(BenchmarkOptions), cl::init(""));
  58
  59 static cl::opt<std::string> SnippetsFile("snippets-file",
  60                                          cl::desc("code snippets to measure"),
  61                                          cl::cat(BenchmarkOptions),
  62                                          cl::init(""));
  63
  64 static cl::opt<std::string>
  65     BenchmarkFile("benchmarks-file",
  66                   cl::desc("File to read (analysis mode) or write "
  67                            "(latency/uops/inverse_throughput modes) benchmark "
  68                            "results. “-” uses stdin/stdout."),
  69                   cl::cat(Options), cl::init(""));
  70
  71 static cl::opt<exegesis::InstructionBenchmark::ModeE> BenchmarkMode(
  72     "mode", cl::desc("the mode to run"), cl::cat(Options),
  73     cl::values(clEnumValN(exegesis::InstructionBenchmark::Latency, "latency",
  74                           "Instruction Latency"),
  75                clEnumValN(exegesis::InstructionBenchmark::InverseThroughput,
  76                           "inverse_throughput",
  77                           "Instruction Inverse Throughput"),
  78                clEnumValN(exegesis::InstructionBenchmark::Uops, "uops",
  79                           "Uop Decomposition"),
  80                // When not asking for a specific benchmark mode,
  81                // we'll analyse the results.
  82                clEnumValN(exegesis::InstructionBenchmark::Unknown, "analysis",
  83                           "Analysis")));
  84
  85 static cl::opt<exegesis::InstructionBenchmark::RepetitionModeE> RepetitionMode(
  86     "repetition-mode", cl::desc("how to repeat the instruction snippet"),
  87     cl::cat(BenchmarkOptions),
  88     cl::values(clEnumValN(exegesis::InstructionBenchmark::Duplicate,
  89                           "duplicate", "Duplicate the snippet"),
  90                clEnumValN(exegesis::InstructionBenchmark::Loop, "loop",
  91                           "Loop over the snippet")));
  92
  93 static cl::opt<unsigned>
  94     NumRepetitions("num-repetitions",
  95                    cl::desc("number of time to repeat the asm snippet"),
  96                    cl::cat(BenchmarkOptions), cl::init(10000));
  97
  98 static cl::opt<unsigned> MaxConfigsPerOpcode(
  99     "max-configs-per-opcode",
 100     cl::desc(
 101         "allow to snippet generator to generate at most that many configs"),
 102     cl::cat(BenchmarkOptions), cl::init(1));
 103
 104 static cl::opt<bool> IgnoreInvalidSchedClass(
 105     "ignore-invalid-sched-class",
 106     cl::desc("ignore instructions that do not define a sched class"),
 107     cl::cat(BenchmarkOptions), cl::init(false));
 108
 109 static cl::opt<exegesis::InstructionBenchmarkClustering::ModeE>
 110     AnalysisClusteringAlgorithm(
 111         "analysis-clustering", cl::desc("the clustering algorithm to use"),
 112         cl::cat(AnalysisOptions),
 113         cl::values(clEnumValN(exegesis::InstructionBenchmarkClustering::Dbscan,
 114                               "dbscan", "use DBSCAN/OPTICS algorithm"),
 115                    clEnumValN(exegesis::InstructionBenchmarkClustering::Naive,
 116                               "naive", "one cluster per opcode")),
 117         cl::init(exegesis::InstructionBenchmarkClustering::Dbscan));
 118
 119 static cl::opt<unsigned> AnalysisDbscanNumPoints(
 120     "analysis-numpoints",
 121     cl::desc("minimum number of points in an analysis cluster (dbscan only)"),
 122     cl::cat(AnalysisOptions), cl::init(3));
 123
 124 static cl::opt<float> AnalysisClusteringEpsilon(
 125     "analysis-clustering-epsilon",
 126     cl::desc("epsilon for benchmark point clustering"),
 127     cl::cat(AnalysisOptions), cl::init(0.1));
 128
 129 static cl::opt<float> AnalysisInconsistencyEpsilon(
 130     "analysis-inconsistency-epsilon",
 131     cl::desc("epsilon for detection of when the cluster is different from the "
 132              "LLVM schedule profile values"),
 133     cl::cat(AnalysisOptions), cl::init(0.1));
 134
 135 static cl::opt<std::string>
 136     AnalysisClustersOutputFile("analysis-clusters-output-file", cl::desc(""),
 137                                cl::cat(AnalysisOptions), cl::init(""));
 138 static cl::opt<std::string>
 139     AnalysisInconsistenciesOutputFile("analysis-inconsistencies-output-file",
 140                                       cl::desc(""), cl::cat(AnalysisOptions),
 141                                       cl::init(""));
 142
 143 static cl::opt<bool> AnalysisDisplayUnstableOpcodes(
 144     "analysis-display-unstable-clusters",
 145     cl::desc("if there is more than one benchmark for an opcode, said "
 146              "benchmarks may end up not being clustered into the same cluster "
 147              "if the measured performance characteristics are different. by "
 148              "default all such opcodes are filtered out. this flag will "
 149              "instead show only such unstable opcodes"),
 150     cl::cat(AnalysisOptions), cl::init(false));
 151
 152 static cl::opt<std::string> CpuName(
 153     "mcpu",
 154     cl::desc("cpu name to use for pfm counters, leave empty to autodetect"),
 155     cl::cat(Options), cl::init(""));
 156
 157 static cl::opt<bool>
 158     DumpObjectToDisk("dump-object-to-disk",
 159                      cl::desc("dumps the generated benchmark object to disk "
 160                               "and prints a message to access it"),
 161                      cl::cat(BenchmarkOptions), cl::init(true));
 162
 163 static ExitOnError ExitOnErr;
 164
 165 // Checks that only one of OpcodeNames, OpcodeIndex or SnippetsFile is provided,
 166 // and returns the opcode indices or {} if snippets should be read from
 167 // `SnippetsFile`.
 168 static std::vector<unsigned> getOpcodesOrDie(const MCInstrInfo &MCInstrInfo) {
 169   const size_t NumSetFlags = (OpcodeNames.empty() ? 0 : 1) +
 170                              (OpcodeIndex == 0 ? 0 : 1) +
 171                              (SnippetsFile.empty() ? 0 : 1);
 172   if (NumSetFlags != 1)
 173     report_fatal_error(
 174         "please provide one and only one of 'opcode-index', 'opcode-name' or "
 175         "'snippets-file'");
 176   if (!SnippetsFile.empty())
 177     return {};
 178   if (OpcodeIndex > 0)
 179     return {static_cast<unsigned>(OpcodeIndex)};
 180   if (OpcodeIndex < 0) {
 181     std::vector<unsigned> Result;
 182     for (unsigned I = 1, E = MCInstrInfo.getNumOpcodes(); I < E; ++I)
 183       Result.push_back(I);
 184     return Result;
 185   }
 186   // Resolve opcode name -> opcode.
 187   const auto ResolveName = [&MCInstrInfo](StringRef OpcodeName) -> unsigned {
 188     for (unsigned I = 1, E = MCInstrInfo.getNumOpcodes(); I < E; ++I)
 189       if (MCInstrInfo.getName(I) == OpcodeName)
 190         return I;
 191     return 0u;
 192   };
 193   SmallVector<StringRef, 2> Pieces;
 194   StringRef(OpcodeNames.getValue())
 195       .split(Pieces, ",", /* MaxSplit */ -1, /* KeepEmpty */ false);
 196   std::vector<unsigned> Result;
 197   for (const StringRef OpcodeName : Pieces) {
 198     if (unsigned Opcode = ResolveName(OpcodeName))
 199       Result.push_back(Opcode);
 200     else
 201       report_fatal_error(Twine("unknown opcode ").concat(OpcodeName));
 202   }
 203   return Result;
 204 }
 205
 206 // Generates code snippets for opcode `Opcode`.
 207 static Expected<std::vector<BenchmarkCode>>
 208 generateSnippets(const LLVMState &State, unsigned Opcode,
 209                  const BitVector &ForbiddenRegs) {
 210   const Instruction &Instr = State.getIC().getInstr(Opcode);
 211   const MCInstrDesc &InstrDesc = *Instr.Description;
 212   // Ignore instructions that we cannot run.
 213   if (InstrDesc.isPseudo())
 214     return make_error<Failure>("Unsupported opcode: isPseudo");
 215   if (InstrDesc.isBranch() || InstrDesc.isIndirectBranch())
 216     return make_error<Failure>("Unsupported opcode: isBranch/isIndirectBranch");
 217   if (InstrDesc.isCall() || InstrDesc.isReturn())
 218     return make_error<Failure>("Unsupported opcode: isCall/isReturn");
 219
 220   SnippetGenerator::Options Options;
 221   Options.MaxConfigsPerOpcode = MaxConfigsPerOpcode;
 222   const std::unique_ptr<SnippetGenerator> Generator =
 223       State.getExegesisTarget().createSnippetGenerator(BenchmarkMode, State,
 224                                                        Options);
 225   if (!Generator)
 226     report_fatal_error("cannot create snippet generator");
 227   return Generator->generateConfigurations(Instr, ForbiddenRegs);
 228 }
 229
 230 void benchmarkMain() {
 231 #ifndef HAVE_LIBPFM
 232   report_fatal_error(
 233       "benchmarking unavailable, LLVM was built without libpfm.");
 234 #endif
 235
 236   if (exegesis::pfm::pfmInitialize())
 237     report_fatal_error("cannot initialize libpfm");
 238
 239   InitializeNativeTarget();
 240   InitializeNativeTargetAsmPrinter();
 241   InitializeNativeTargetAsmParser();
 242   InitializeNativeExegesisTarget();
 243
 244   const LLVMState State(CpuName);
 245   const auto Opcodes = getOpcodesOrDie(State.getInstrInfo());
 246
 247   const auto Repetitor = SnippetRepetitor::Create(RepetitionMode, State);
 248
 249   std::vector<BenchmarkCode> Configurations;
 250   if (!Opcodes.empty()) {
 251     for (const unsigned Opcode : Opcodes) {
 252       // Ignore instructions without a sched class if
 253       // -ignore-invalid-sched-class is passed.
 254       if (IgnoreInvalidSchedClass &&
 255           State.getInstrInfo().get(Opcode).getSchedClass() == 0) {
 256         errs() << State.getInstrInfo().getName(Opcode)
 257                << ": ignoring instruction without sched class\n";
 258         continue;
 259       }
 260       auto ConfigsForInstr =
 261           generateSnippets(State, Opcode, Repetitor->getReservedRegs());
 262       if (!ConfigsForInstr) {
 263         logAllUnhandledErrors(
 264             ConfigsForInstr.takeError(), errs(),
 265             Twine(State.getInstrInfo().getName(Opcode)).concat(": "));
 266         continue;
 267       }
 268       std::move(ConfigsForInstr->begin(), ConfigsForInstr->end(),
 269                 std::back_inserter(Configurations));
 270     }
 271   } else {
 272     Configurations = ExitOnErr(readSnippets(State, SnippetsFile));
 273   }
 274
 275   const std::unique_ptr<BenchmarkRunner> Runner =
 276       State.getExegesisTarget().createBenchmarkRunner(BenchmarkMode, State);
 277   if (!Runner) {
 278     report_fatal_error("cannot create benchmark runner");
 279   }
 280
 281   if (NumRepetitions == 0)
 282     report_fatal_error("--num-repetitions must be greater than zero");
 283
 284   // Write to standard output if file is not set.
 285   if (BenchmarkFile.empty())
 286     BenchmarkFile = "-";
 287
 288   for (const BenchmarkCode &Conf : Configurations) {
 289     InstructionBenchmark Result = Runner->runConfiguration(
 290         Conf, NumRepetitions, *Repetitor, DumpObjectToDisk);
 291     ExitOnErr(Result.writeYaml(State, BenchmarkFile));
 292   }
 293   exegesis::pfm::pfmTerminate();
 294 }
 295
 296 // Prints the results of running analysis pass `Pass` to file `OutputFilename`
 297 // if OutputFilename is non-empty.
 298 template <typename Pass>
 299 static void maybeRunAnalysis(const Analysis &Analyzer, const std::string &Name,
 300                              const std::string &OutputFilename) {
 301   if (OutputFilename.empty())
 302     return;
 303   if (OutputFilename != "-") {
 304     errs() << "Printing " << Name << " results to file '" << OutputFilename
 305            << "'\n";
 306   }
 307   std::error_code ErrorCode;
 308   raw_fd_ostream ClustersOS(OutputFilename, ErrorCode,
 309                             sys::fs::FA_Read | sys::fs::FA_Write);
 310   if (ErrorCode)
 311     report_fatal_error("cannot open out file: " + OutputFilename);
 312   if (auto Err = Analyzer.run<Pass>(ClustersOS))
 313     report_fatal_error(std::move(Err));
 314 }
 315
 316 static void analysisMain() {
 317   if (BenchmarkFile.empty())
 318     report_fatal_error("--benchmarks-file must be set.");
 319
 320   if (AnalysisClustersOutputFile.empty() &&
 321       AnalysisInconsistenciesOutputFile.empty()) {
 322     report_fatal_error(
 323         "At least one of --analysis-clusters-output-file and "
 324         "--analysis-inconsistencies-output-file must be specified.");
 325   }
 326
 327   InitializeNativeTarget();
 328   InitializeNativeTargetAsmPrinter();
 329   InitializeNativeTargetDisassembler();
 330   // Read benchmarks.
 331   const LLVMState State("");
 332   const std::vector<InstructionBenchmark> Points =
 333       ExitOnErr(InstructionBenchmark::readYamls(State, BenchmarkFile));
 334   outs() << "Parsed " << Points.size() << " benchmark points\n";
 335   if (Points.empty()) {
 336     errs() << "no benchmarks to analyze\n";
 337     return;
 338   }
 339   // FIXME: Check that all points have the same triple/cpu.
 340   // FIXME: Merge points from several runs (latency and uops).
 341
 342   std::string Error;
 343   const auto *TheTarget =
 344       TargetRegistry::lookupTarget(Points[0].LLVMTriple, Error);
 345   if (!TheTarget) {
 346     errs() << "unknown target '" << Points[0].LLVMTriple << "'\n";
 347     return;
 348   }
 349
 350   std::unique_ptr<MCInstrInfo> InstrInfo(TheTarget->createMCInstrInfo());
 351
 352   const auto Clustering = ExitOnErr(InstructionBenchmarkClustering::create(
 353       Points, AnalysisClusteringAlgorithm, AnalysisDbscanNumPoints,
 354       AnalysisClusteringEpsilon, InstrInfo->getNumOpcodes()));
 355
 356   const Analysis Analyzer(*TheTarget, std::move(InstrInfo), Clustering,
 357                           AnalysisInconsistencyEpsilon,
 358                           AnalysisDisplayUnstableOpcodes);
 359
 360   maybeRunAnalysis<Analysis::PrintClusters>(Analyzer, "analysis clusters",
 361                                             AnalysisClustersOutputFile);
 362   maybeRunAnalysis<Analysis::PrintSchedClassInconsistencies>(
 363       Analyzer, "sched class consistency analysis",
 364       AnalysisInconsistenciesOutputFile);
 365 }
 366
 367 } // namespace exegesis
 368 } // namespace llvm
 369
 370 int main(int Argc, char **Argv) {
 371   using namespace llvm;
 372   cl::ParseCommandLineOptions(Argc, Argv, "");
 373
 374   exegesis::ExitOnErr.setExitCodeMapper([](const Error &Err) {
 375     if (Err.isA<StringError>())
 376       return EXIT_SUCCESS;
 377     return EXIT_FAILURE;
 378   });
 379
 380   if (exegesis::BenchmarkMode == exegesis::InstructionBenchmark::Unknown) {
 381     exegesis::analysisMain();
 382   } else {
 383     exegesis::benchmarkMain();
 384   }
 385   return EXIT_SUCCESS;
 386 }