contrib/llvm/tools/llvm-mca/llvm-mca.cpp

   1 //===-- llvm-mca.cpp - Machine Code Analyzer -------------------*- C++ -* -===//
   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 utility is a simple driver that allows static performance analysis on
  11 // machine code similarly to how IACA (Intel Architecture Code Analyzer) works.
  12 //
  13 //   llvm-mca [options] <file-name>
  14 //      -march <type>
  15 //      -mcpu <cpu>
  16 //      -o <file>
  17 //
  18 // The target defaults to the host target.
  19 // The cpu defaults to the 'native' host cpu.
  20 // The output defaults to standard output.
  21 //
  22 //===----------------------------------------------------------------------===//
  23
  24 #include "CodeRegion.h"
  25 #include "CodeRegionGenerator.h"
  26 #include "PipelinePrinter.h"
  27 #include "Views/DispatchStatistics.h"
  28 #include "Views/InstructionInfoView.h"
  29 #include "Views/RegisterFileStatistics.h"
  30 #include "Views/ResourcePressureView.h"
  31 #include "Views/RetireControlUnitStatistics.h"
  32 #include "Views/SchedulerStatistics.h"
  33 #include "Views/SummaryView.h"
  34 #include "Views/TimelineView.h"
  35 #include "llvm/MC/MCAsmInfo.h"
  36 #include "llvm/MC/MCContext.h"
  37 #include "llvm/MC/MCObjectFileInfo.h"
  38 #include "llvm/MC/MCRegisterInfo.h"
  39 #include "llvm/MCA/Context.h"
  40 #include "llvm/MCA/Pipeline.h"
  41 #include "llvm/MCA/Stages/EntryStage.h"
  42 #include "llvm/MCA/Stages/InstructionTables.h"
  43 #include "llvm/MCA/Support.h"
  44 #include "llvm/Support/CommandLine.h"
  45 #include "llvm/Support/ErrorHandling.h"
  46 #include "llvm/Support/ErrorOr.h"
  47 #include "llvm/Support/FileSystem.h"
  48 #include "llvm/Support/Host.h"
  49 #include "llvm/Support/InitLLVM.h"
  50 #include "llvm/Support/MemoryBuffer.h"
  51 #include "llvm/Support/SourceMgr.h"
  52 #include "llvm/Support/TargetRegistry.h"
  53 #include "llvm/Support/TargetSelect.h"
  54 #include "llvm/Support/ToolOutputFile.h"
  55 #include "llvm/Support/WithColor.h"
  56
  57 using namespace llvm;
  58
  59 static cl::OptionCategory ToolOptions("Tool Options");
  60 static cl::OptionCategory ViewOptions("View Options");
  61
  62 static cl::opt<std::string> InputFilename(cl::Positional,
  63                                           cl::desc("<input file>"),
  64                                           cl::cat(ToolOptions), cl::init("-"));
  65
  66 static cl::opt<std::string> OutputFilename("o", cl::desc("Output filename"),
  67                                            cl::init("-"), cl::cat(ToolOptions),
  68                                            cl::value_desc("filename"));
  69
  70 static cl::opt<std::string>
  71     ArchName("march", cl::desc("Target architecture. "
  72                                "See -version for available targets"),
  73              cl::cat(ToolOptions));
  74
  75 static cl::opt<std::string>
  76     TripleName("mtriple",
  77                cl::desc("Target triple. See -version for available targets"),
  78                cl::cat(ToolOptions));
  79
  80 static cl::opt<std::string>
  81     MCPU("mcpu",
  82          cl::desc("Target a specific cpu type (-mcpu=help for details)"),
  83          cl::value_desc("cpu-name"), cl::cat(ToolOptions), cl::init("native"));
  84
  85 static cl::opt<int>
  86     OutputAsmVariant("output-asm-variant",
  87                      cl::desc("Syntax variant to use for output printing"),
  88                      cl::cat(ToolOptions), cl::init(-1));
  89
  90 static cl::opt<unsigned> Iterations("iterations",
  91                                     cl::desc("Number of iterations to run"),
  92                                     cl::cat(ToolOptions), cl::init(0));
  93
  94 static cl::opt<unsigned>
  95     DispatchWidth("dispatch", cl::desc("Override the processor dispatch width"),
  96                   cl::cat(ToolOptions), cl::init(0));
  97
  98 static cl::opt<unsigned>
  99     RegisterFileSize("register-file-size",
 100                      cl::desc("Maximum number of physical registers which can "
 101                               "be used for register mappings"),
 102                      cl::cat(ToolOptions), cl::init(0));
 103
 104 static cl::opt<bool>
 105     PrintRegisterFileStats("register-file-stats",
 106                            cl::desc("Print register file statistics"),
 107                            cl::cat(ViewOptions), cl::init(false));
 108
 109 static cl::opt<bool> PrintDispatchStats("dispatch-stats",
 110                                         cl::desc("Print dispatch statistics"),
 111                                         cl::cat(ViewOptions), cl::init(false));
 112
 113 static cl::opt<bool>
 114     PrintSummaryView("summary-view", cl::Hidden,
 115                      cl::desc("Print summary view (enabled by default)"),
 116                      cl::cat(ViewOptions), cl::init(true));
 117
 118 static cl::opt<bool> PrintSchedulerStats("scheduler-stats",
 119                                          cl::desc("Print scheduler statistics"),
 120                                          cl::cat(ViewOptions), cl::init(false));
 121
 122 static cl::opt<bool>
 123     PrintRetireStats("retire-stats",
 124                      cl::desc("Print retire control unit statistics"),
 125                      cl::cat(ViewOptions), cl::init(false));
 126
 127 static cl::opt<bool> PrintResourcePressureView(
 128     "resource-pressure",
 129     cl::desc("Print the resource pressure view (enabled by default)"),
 130     cl::cat(ViewOptions), cl::init(true));
 131
 132 static cl::opt<bool> PrintTimelineView("timeline",
 133                                        cl::desc("Print the timeline view"),
 134                                        cl::cat(ViewOptions), cl::init(false));
 135
 136 static cl::opt<unsigned> TimelineMaxIterations(
 137     "timeline-max-iterations",
 138     cl::desc("Maximum number of iterations to print in timeline view"),
 139     cl::cat(ViewOptions), cl::init(0));
 140
 141 static cl::opt<unsigned> TimelineMaxCycles(
 142     "timeline-max-cycles",
 143     cl::desc(
 144         "Maximum number of cycles in the timeline view. Defaults to 80 cycles"),
 145     cl::cat(ViewOptions), cl::init(80));
 146
 147 static cl::opt<bool>
 148     AssumeNoAlias("noalias",
 149                   cl::desc("If set, assume that loads and stores do not alias"),
 150                   cl::cat(ToolOptions), cl::init(true));
 151
 152 static cl::opt<unsigned> LoadQueueSize("lqueue",
 153                                        cl::desc("Size of the load queue"),
 154                                        cl::cat(ToolOptions), cl::init(0));
 155
 156 static cl::opt<unsigned> StoreQueueSize("squeue",
 157                                         cl::desc("Size of the store queue"),
 158                                         cl::cat(ToolOptions), cl::init(0));
 159
 160 static cl::opt<bool>
 161     PrintInstructionTables("instruction-tables",
 162                            cl::desc("Print instruction tables"),
 163                            cl::cat(ToolOptions), cl::init(false));
 164
 165 static cl::opt<bool> PrintInstructionInfoView(
 166     "instruction-info",
 167     cl::desc("Print the instruction info view (enabled by default)"),
 168     cl::cat(ViewOptions), cl::init(true));
 169
 170 static cl::opt<bool> EnableAllStats("all-stats",
 171                                     cl::desc("Print all hardware statistics"),
 172                                     cl::cat(ViewOptions), cl::init(false));
 173
 174 static cl::opt<bool>
 175     EnableAllViews("all-views",
 176                    cl::desc("Print all views including hardware statistics"),
 177                    cl::cat(ViewOptions), cl::init(false));
 178
 179 namespace {
 180
 181 const Target *getTarget(const char *ProgName) {
 182   if (TripleName.empty())
 183     TripleName = Triple::normalize(sys::getDefaultTargetTriple());
 184   Triple TheTriple(TripleName);
 185
 186   // Get the target specific parser.
 187   std::string Error;
 188   const Target *TheTarget =
 189       TargetRegistry::lookupTarget(ArchName, TheTriple, Error);
 190   if (!TheTarget) {
 191     errs() << ProgName << ": " << Error;
 192     return nullptr;
 193   }
 194
 195   // Return the found target.
 196   return TheTarget;
 197 }
 198
 199 ErrorOr<std::unique_ptr<ToolOutputFile>> getOutputStream() {
 200   if (OutputFilename == "")
 201     OutputFilename = "-";
 202   std::error_code EC;
 203   auto Out =
 204       llvm::make_unique<ToolOutputFile>(OutputFilename, EC, sys::fs::F_None);
 205   if (!EC)
 206     return std::move(Out);
 207   return EC;
 208 }
 209 } // end of anonymous namespace
 210
 211 static void processOptionImpl(cl::opt<bool> &O, const cl::opt<bool> &Default) {
 212   if (!O.getNumOccurrences() || O.getPosition() < Default.getPosition())
 213     O = Default.getValue();
 214 }
 215
 216 static void processViewOptions() {
 217   if (!EnableAllViews.getNumOccurrences() &&
 218       !EnableAllStats.getNumOccurrences())
 219     return;
 220
 221   if (EnableAllViews.getNumOccurrences()) {
 222     processOptionImpl(PrintSummaryView, EnableAllViews);
 223     processOptionImpl(PrintResourcePressureView, EnableAllViews);
 224     processOptionImpl(PrintTimelineView, EnableAllViews);
 225     processOptionImpl(PrintInstructionInfoView, EnableAllViews);
 226   }
 227
 228   const cl::opt<bool> &Default =
 229       EnableAllViews.getPosition() < EnableAllStats.getPosition()
 230           ? EnableAllStats
 231           : EnableAllViews;
 232   processOptionImpl(PrintRegisterFileStats, Default);
 233   processOptionImpl(PrintDispatchStats, Default);
 234   processOptionImpl(PrintSchedulerStats, Default);
 235   processOptionImpl(PrintRetireStats, Default);
 236 }
 237
 238 // Returns true on success.
 239 static bool runPipeline(mca::Pipeline &P) {
 240   // Handle pipeline errors here.
 241   Expected<unsigned> Cycles = P.run();
 242   if (!Cycles) {
 243     WithColor::error() << toString(Cycles.takeError());
 244     return false;
 245   }
 246   return true;
 247 }
 248
 249 int main(int argc, char **argv) {
 250   InitLLVM X(argc, argv);
 251
 252   // Initialize targets and assembly parsers.
 253   InitializeAllTargetInfos();
 254   InitializeAllTargetMCs();
 255   InitializeAllAsmParsers();
 256
 257   // Enable printing of available targets when flag --version is specified.
 258   cl::AddExtraVersionPrinter(TargetRegistry::printRegisteredTargetsForVersion);
 259
 260   cl::HideUnrelatedOptions({&ToolOptions, &ViewOptions});
 261
 262   // Parse flags and initialize target options.
 263   cl::ParseCommandLineOptions(argc, argv,
 264                               "llvm machine code performance analyzer.\n");
 265
 266   // Get the target from the triple. If a triple is not specified, then select
 267   // the default triple for the host. If the triple doesn't correspond to any
 268   // registered target, then exit with an error message.
 269   const char *ProgName = argv[0];
 270   const Target *TheTarget = getTarget(ProgName);
 271   if (!TheTarget)
 272     return 1;
 273
 274   // GetTarget() may replaced TripleName with a default triple.
 275   // For safety, reconstruct the Triple object.
 276   Triple TheTriple(TripleName);
 277
 278   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferPtr =
 279       MemoryBuffer::getFileOrSTDIN(InputFilename);
 280   if (std::error_code EC = BufferPtr.getError()) {
 281     WithColor::error() << InputFilename << ": " << EC.message() << '\n';
 282     return 1;
 283   }
 284
 285   // Apply overrides to llvm-mca specific options.
 286   processViewOptions();
 287
 288   SourceMgr SrcMgr;
 289
 290   // Tell SrcMgr about this buffer, which is what the parser will pick up.
 291   SrcMgr.AddNewSourceBuffer(std::move(*BufferPtr), SMLoc());
 292
 293   std::unique_ptr<MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TripleName));
 294   assert(MRI && "Unable to create target register info!");
 295
 296   std::unique_ptr<MCAsmInfo> MAI(TheTarget->createMCAsmInfo(*MRI, TripleName));
 297   assert(MAI && "Unable to create target asm info!");
 298
 299   MCObjectFileInfo MOFI;
 300   MCContext Ctx(MAI.get(), MRI.get(), &MOFI, &SrcMgr);
 301   MOFI.InitMCObjectFileInfo(TheTriple, /* PIC= */ false, Ctx);
 302
 303   std::unique_ptr<buffer_ostream> BOS;
 304
 305   std::unique_ptr<MCInstrInfo> MCII(TheTarget->createMCInstrInfo());
 306
 307   std::unique_ptr<MCInstrAnalysis> MCIA(
 308       TheTarget->createMCInstrAnalysis(MCII.get()));
 309
 310   if (!MCPU.compare("native"))
 311     MCPU = llvm::sys::getHostCPUName();
 312
 313   std::unique_ptr<MCSubtargetInfo> STI(
 314       TheTarget->createMCSubtargetInfo(TripleName, MCPU, /* FeaturesStr */ ""));
 315   if (!STI->isCPUStringValid(MCPU))
 316     return 1;
 317
 318   if (!PrintInstructionTables && !STI->getSchedModel().isOutOfOrder()) {
 319     WithColor::error() << "please specify an out-of-order cpu. '" << MCPU
 320                        << "' is an in-order cpu.\n";
 321     return 1;
 322   }
 323
 324   if (!STI->getSchedModel().hasInstrSchedModel()) {
 325     WithColor::error()
 326         << "unable to find instruction-level scheduling information for"
 327         << " target triple '" << TheTriple.normalize() << "' and cpu '" << MCPU
 328         << "'.\n";
 329
 330     if (STI->getSchedModel().InstrItineraries)
 331       WithColor::note()
 332           << "cpu '" << MCPU << "' provides itineraries. However, "
 333           << "instruction itineraries are currently unsupported.\n";
 334     return 1;
 335   }
 336
 337   // Parse the input and create CodeRegions that llvm-mca can analyze.
 338   mca::AsmCodeRegionGenerator CRG(*TheTarget, SrcMgr, Ctx, *MAI, *STI, *MCII);
 339   Expected<const mca::CodeRegions &> RegionsOrErr = CRG.parseCodeRegions();
 340   if (!RegionsOrErr) {
 341     if (auto Err =
 342             handleErrors(RegionsOrErr.takeError(), [](const StringError &E) {
 343               WithColor::error() << E.getMessage() << '\n';
 344             })) {
 345       // Default case.
 346       WithColor::error() << toString(std::move(Err)) << '\n';
 347     }
 348     return 1;
 349   }
 350   const mca::CodeRegions &Regions = *RegionsOrErr;
 351   if (Regions.empty()) {
 352     WithColor::error() << "no assembly instructions found.\n";
 353     return 1;
 354   }
 355
 356   // Now initialize the output file.
 357   auto OF = getOutputStream();
 358   if (std::error_code EC = OF.getError()) {
 359     WithColor::error() << EC.message() << '\n';
 360     return 1;
 361   }
 362
 363   unsigned AssemblerDialect = CRG.getAssemblerDialect();
 364   if (OutputAsmVariant >= 0)
 365     AssemblerDialect = static_cast<unsigned>(OutputAsmVariant);
 366   std::unique_ptr<MCInstPrinter> IP(TheTarget->createMCInstPrinter(
 367       Triple(TripleName), AssemblerDialect, *MAI, *MCII, *MRI));
 368   if (!IP) {
 369     WithColor::error()
 370         << "unable to create instruction printer for target triple '"
 371         << TheTriple.normalize() << "' with assembly variant "
 372         << AssemblerDialect << ".\n";
 373     return 1;
 374   }
 375
 376   std::unique_ptr<ToolOutputFile> TOF = std::move(*OF);
 377
 378   const MCSchedModel &SM = STI->getSchedModel();
 379
 380   unsigned Width = SM.IssueWidth;
 381   if (DispatchWidth)
 382     Width = DispatchWidth;
 383
 384   // Create an instruction builder.
 385   mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get());
 386
 387   // Create a context to control ownership of the pipeline hardware.
 388   mca::Context MCA(*MRI, *STI);
 389
 390   mca::PipelineOptions PO(Width, RegisterFileSize, LoadQueueSize,
 391                           StoreQueueSize, AssumeNoAlias);
 392
 393   // Number each region in the sequence.
 394   unsigned RegionIdx = 0;
 395
 396   for (const std::unique_ptr<mca::CodeRegion> &Region : Regions) {
 397     // Skip empty code regions.
 398     if (Region->empty())
 399       continue;
 400
 401     // Don't print the header of this region if it is the default region, and
 402     // it doesn't have an end location.
 403     if (Region->startLoc().isValid() || Region->endLoc().isValid()) {
 404       TOF->os() << "\n[" << RegionIdx++ << "] Code Region";
 405       StringRef Desc = Region->getDescription();
 406       if (!Desc.empty())
 407         TOF->os() << " - " << Desc;
 408       TOF->os() << "\n\n";
 409     }
 410
 411     // Lower the MCInst sequence into an mca::Instruction sequence.
 412     ArrayRef<MCInst> Insts = Region->getInstructions();
 413     std::vector<std::unique_ptr<mca::Instruction>> LoweredSequence;
 414     for (const MCInst &MCI : Insts) {
 415       Expected<std::unique_ptr<mca::Instruction>> Inst =
 416           IB.createInstruction(MCI);
 417       if (!Inst) {
 418         if (auto NewE = handleErrors(
 419                 Inst.takeError(),
 420                 [&IP, &STI](const mca::InstructionError<MCInst> &IE) {
 421                   std::string InstructionStr;
 422                   raw_string_ostream SS(InstructionStr);
 423                   WithColor::error() << IE.Message << '\n';
 424                   IP->printInst(&IE.Inst, SS, "", *STI);
 425                   SS.flush();
 426                   WithColor::note() << "instruction: " << InstructionStr
 427                                     << '\n';
 428                 })) {
 429           // Default case.
 430           WithColor::error() << toString(std::move(NewE));
 431         }
 432         return 1;
 433       }
 434
 435       LoweredSequence.emplace_back(std::move(Inst.get()));
 436     }
 437
 438     mca::SourceMgr S(LoweredSequence, PrintInstructionTables ? 1 : Iterations);
 439
 440     if (PrintInstructionTables) {
 441       //  Create a pipeline, stages, and a printer.
 442       auto P = llvm::make_unique<mca::Pipeline>();
 443       P->appendStage(llvm::make_unique<mca::EntryStage>(S));
 444       P->appendStage(llvm::make_unique<mca::InstructionTables>(SM));
 445       mca::PipelinePrinter Printer(*P);
 446
 447       // Create the views for this pipeline, execute, and emit a report.
 448       if (PrintInstructionInfoView) {
 449         Printer.addView(llvm::make_unique<mca::InstructionInfoView>(
 450             *STI, *MCII, Insts, *IP));
 451       }
 452       Printer.addView(
 453           llvm::make_unique<mca::ResourcePressureView>(*STI, *IP, Insts));
 454
 455       if (!runPipeline(*P))
 456         return 1;
 457
 458       Printer.printReport(TOF->os());
 459       continue;
 460     }
 461
 462     // Create a basic pipeline simulating an out-of-order backend.
 463     auto P = MCA.createDefaultPipeline(PO, IB, S);
 464     mca::PipelinePrinter Printer(*P);
 465
 466     if (PrintSummaryView)
 467       Printer.addView(llvm::make_unique<mca::SummaryView>(SM, Insts, Width));
 468
 469     if (PrintInstructionInfoView)
 470       Printer.addView(
 471           llvm::make_unique<mca::InstructionInfoView>(*STI, *MCII, Insts, *IP));
 472
 473     if (PrintDispatchStats)
 474       Printer.addView(llvm::make_unique<mca::DispatchStatistics>());
 475
 476     if (PrintSchedulerStats)
 477       Printer.addView(llvm::make_unique<mca::SchedulerStatistics>(*STI));
 478
 479     if (PrintRetireStats)
 480       Printer.addView(llvm::make_unique<mca::RetireControlUnitStatistics>(SM));
 481
 482     if (PrintRegisterFileStats)
 483       Printer.addView(llvm::make_unique<mca::RegisterFileStatistics>(*STI));
 484
 485     if (PrintResourcePressureView)
 486       Printer.addView(
 487           llvm::make_unique<mca::ResourcePressureView>(*STI, *IP, Insts));
 488
 489     if (PrintTimelineView) {
 490       unsigned TimelineIterations =
 491           TimelineMaxIterations ? TimelineMaxIterations : 10;
 492       Printer.addView(llvm::make_unique<mca::TimelineView>(
 493           *STI, *IP, Insts, std::min(TimelineIterations, S.getNumIterations()),
 494           TimelineMaxCycles));
 495     }
 496
 497     if (!runPipeline(*P))
 498       return 1;
 499
 500     Printer.printReport(TOF->os());
 501
 502     // Clear the InstrBuilder internal state in preparation for another round.
 503     IB.clear();
 504   }
 505
 506   TOF->keep();
 507   return 0;
 508 }