[libc] Switch to using the generic `<gpuintrin.h>` implementations (#121810)
[llvm-project.git] / llvm / tools / llvm-exegesis / llvm-exegesis.cpp
blobfa37e05956be8c66e53b3cdc3fe4ad8c072f3bf9
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 //===----------------------------------------------------------------------===//
14 #include "lib/Analysis.h"
15 #include "lib/BenchmarkResult.h"
16 #include "lib/BenchmarkRunner.h"
17 #include "lib/Clustering.h"
18 #include "lib/CodeTemplate.h"
19 #include "lib/Error.h"
20 #include "lib/LlvmState.h"
21 #include "lib/PerfHelper.h"
22 #include "lib/ProgressMeter.h"
23 #include "lib/ResultAggregator.h"
24 #include "lib/SnippetFile.h"
25 #include "lib/SnippetRepetitor.h"
26 #include "lib/Target.h"
27 #include "lib/TargetSelect.h"
28 #include "lib/ValidationEvent.h"
29 #include "llvm/ADT/StringExtras.h"
30 #include "llvm/ADT/Twine.h"
31 #include "llvm/MC/MCInstBuilder.h"
32 #include "llvm/MC/MCObjectFileInfo.h"
33 #include "llvm/MC/MCParser/MCAsmParser.h"
34 #include "llvm/MC/MCParser/MCTargetAsmParser.h"
35 #include "llvm/MC/MCRegisterInfo.h"
36 #include "llvm/MC/MCSubtargetInfo.h"
37 #include "llvm/MC/TargetRegistry.h"
38 #include "llvm/Object/ObjectFile.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/FileSystem.h"
41 #include "llvm/Support/Format.h"
42 #include "llvm/Support/InitLLVM.h"
43 #include "llvm/Support/Path.h"
44 #include "llvm/Support/SourceMgr.h"
45 #include "llvm/Support/TargetSelect.h"
46 #include "llvm/TargetParser/Host.h"
47 #include <algorithm>
48 #include <string>
50 namespace llvm {
51 namespace exegesis {
53 static cl::opt<int> OpcodeIndex(
54 "opcode-index",
55 cl::desc("opcode to measure, by index, or -1 to measure all opcodes"),
56 cl::cat(BenchmarkOptions), cl::init(0));
58 static cl::opt<std::string>
59 OpcodeNames("opcode-name",
60 cl::desc("comma-separated list of opcodes to measure, by name"),
61 cl::cat(BenchmarkOptions), cl::init(""));
63 static cl::opt<std::string> SnippetsFile("snippets-file",
64 cl::desc("code snippets to measure"),
65 cl::cat(BenchmarkOptions),
66 cl::init(""));
68 static cl::opt<std::string>
69 BenchmarkFile("benchmarks-file",
70 cl::desc("File to read (analysis mode) or write "
71 "(latency/uops/inverse_throughput modes) benchmark "
72 "results. “-” uses stdin/stdout."),
73 cl::cat(Options), cl::init(""));
75 static cl::opt<Benchmark::ModeE> BenchmarkMode(
76 "mode", cl::desc("the mode to run"), cl::cat(Options),
77 cl::values(clEnumValN(Benchmark::Latency, "latency", "Instruction Latency"),
78 clEnumValN(Benchmark::InverseThroughput, "inverse_throughput",
79 "Instruction Inverse Throughput"),
80 clEnumValN(Benchmark::Uops, "uops", "Uop Decomposition"),
81 // When not asking for a specific benchmark mode,
82 // we'll analyse the results.
83 clEnumValN(Benchmark::Unknown, "analysis", "Analysis")));
85 static cl::opt<Benchmark::ResultAggregationModeE> ResultAggMode(
86 "result-aggregation-mode", cl::desc("How to aggregate multi-values result"),
87 cl::cat(BenchmarkOptions),
88 cl::values(clEnumValN(Benchmark::Min, "min", "Keep min reading"),
89 clEnumValN(Benchmark::Max, "max", "Keep max reading"),
90 clEnumValN(Benchmark::Mean, "mean",
91 "Compute mean of all readings"),
92 clEnumValN(Benchmark::MinVariance, "min-variance",
93 "Keep readings set with min-variance")),
94 cl::init(Benchmark::Min));
96 static cl::opt<Benchmark::RepetitionModeE> RepetitionMode(
97 "repetition-mode", cl::desc("how to repeat the instruction snippet"),
98 cl::cat(BenchmarkOptions),
99 cl::values(
100 clEnumValN(Benchmark::Duplicate, "duplicate", "Duplicate the snippet"),
101 clEnumValN(Benchmark::Loop, "loop", "Loop over the snippet"),
102 clEnumValN(Benchmark::AggregateMin, "min",
103 "All of the above and take the minimum of measurements"),
104 clEnumValN(Benchmark::MiddleHalfDuplicate, "middle-half-duplicate",
105 "Middle half duplicate mode"),
106 clEnumValN(Benchmark::MiddleHalfLoop, "middle-half-loop",
107 "Middle half loop mode")),
108 cl::init(Benchmark::Duplicate));
110 static cl::opt<bool> BenchmarkMeasurementsPrintProgress(
111 "measurements-print-progress",
112 cl::desc("Produce progress indicator when performing measurements"),
113 cl::cat(BenchmarkOptions), cl::init(false));
115 static cl::opt<BenchmarkPhaseSelectorE> BenchmarkPhaseSelector(
116 "benchmark-phase",
117 cl::desc(
118 "it is possible to stop the benchmarking process after some phase"),
119 cl::cat(BenchmarkOptions),
120 cl::values(
121 clEnumValN(BenchmarkPhaseSelectorE::PrepareSnippet, "prepare-snippet",
122 "Only generate the minimal instruction sequence"),
123 clEnumValN(BenchmarkPhaseSelectorE::PrepareAndAssembleSnippet,
124 "prepare-and-assemble-snippet",
125 "Same as prepare-snippet, but also dumps an excerpt of the "
126 "sequence (hex encoded)"),
127 clEnumValN(BenchmarkPhaseSelectorE::AssembleMeasuredCode,
128 "assemble-measured-code",
129 "Same as prepare-and-assemble-snippet, but also creates the "
130 "full sequence "
131 "that can be dumped to a file using --dump-object-to-disk"),
132 clEnumValN(
133 BenchmarkPhaseSelectorE::Measure, "measure",
134 "Same as prepare-measured-code, but also runs the measurement "
135 "(default)")),
136 cl::init(BenchmarkPhaseSelectorE::Measure));
138 static cl::opt<bool>
139 UseDummyPerfCounters("use-dummy-perf-counters",
140 cl::desc("Do not read real performance counters, use "
141 "dummy values (for testing)"),
142 cl::cat(BenchmarkOptions), cl::init(false));
144 static cl::opt<unsigned>
145 MinInstructions("min-instructions",
146 cl::desc("The minimum number of instructions that should "
147 "be included in the snippet"),
148 cl::cat(BenchmarkOptions), cl::init(10000));
150 static cl::opt<unsigned>
151 LoopBodySize("loop-body-size",
152 cl::desc("when repeating the instruction snippet by looping "
153 "over it, duplicate the snippet until the loop body "
154 "contains at least this many instruction"),
155 cl::cat(BenchmarkOptions), cl::init(0));
157 static cl::opt<unsigned> MaxConfigsPerOpcode(
158 "max-configs-per-opcode",
159 cl::desc(
160 "allow to snippet generator to generate at most that many configs"),
161 cl::cat(BenchmarkOptions), cl::init(1));
163 static cl::opt<bool> IgnoreInvalidSchedClass(
164 "ignore-invalid-sched-class",
165 cl::desc("ignore instructions that do not define a sched class"),
166 cl::cat(BenchmarkOptions), cl::init(false));
168 static cl::opt<BenchmarkFilter> AnalysisSnippetFilter(
169 "analysis-filter", cl::desc("Filter the benchmarks before analysing them"),
170 cl::cat(BenchmarkOptions),
171 cl::values(
172 clEnumValN(BenchmarkFilter::All, "all",
173 "Keep all benchmarks (default)"),
174 clEnumValN(BenchmarkFilter::RegOnly, "reg-only",
175 "Keep only those benchmarks that do *NOT* involve memory"),
176 clEnumValN(BenchmarkFilter::WithMem, "mem-only",
177 "Keep only the benchmarks that *DO* involve memory")),
178 cl::init(BenchmarkFilter::All));
180 static cl::opt<BenchmarkClustering::ModeE> AnalysisClusteringAlgorithm(
181 "analysis-clustering", cl::desc("the clustering algorithm to use"),
182 cl::cat(AnalysisOptions),
183 cl::values(clEnumValN(BenchmarkClustering::Dbscan, "dbscan",
184 "use DBSCAN/OPTICS algorithm"),
185 clEnumValN(BenchmarkClustering::Naive, "naive",
186 "one cluster per opcode")),
187 cl::init(BenchmarkClustering::Dbscan));
189 static cl::opt<unsigned> AnalysisDbscanNumPoints(
190 "analysis-numpoints",
191 cl::desc("minimum number of points in an analysis cluster (dbscan only)"),
192 cl::cat(AnalysisOptions), cl::init(3));
194 static cl::opt<float> AnalysisClusteringEpsilon(
195 "analysis-clustering-epsilon",
196 cl::desc("epsilon for benchmark point clustering"),
197 cl::cat(AnalysisOptions), cl::init(0.1));
199 static cl::opt<float> AnalysisInconsistencyEpsilon(
200 "analysis-inconsistency-epsilon",
201 cl::desc("epsilon for detection of when the cluster is different from the "
202 "LLVM schedule profile values"),
203 cl::cat(AnalysisOptions), cl::init(0.1));
205 static cl::opt<std::string>
206 AnalysisClustersOutputFile("analysis-clusters-output-file", cl::desc(""),
207 cl::cat(AnalysisOptions), cl::init(""));
208 static cl::opt<std::string>
209 AnalysisInconsistenciesOutputFile("analysis-inconsistencies-output-file",
210 cl::desc(""), cl::cat(AnalysisOptions),
211 cl::init(""));
213 static cl::opt<bool> AnalysisDisplayUnstableOpcodes(
214 "analysis-display-unstable-clusters",
215 cl::desc("if there is more than one benchmark for an opcode, said "
216 "benchmarks may end up not being clustered into the same cluster "
217 "if the measured performance characteristics are different. by "
218 "default all such opcodes are filtered out. this flag will "
219 "instead show only such unstable opcodes"),
220 cl::cat(AnalysisOptions), cl::init(false));
222 static cl::opt<bool> AnalysisOverrideBenchmarksTripleAndCpu(
223 "analysis-override-benchmark-triple-and-cpu",
224 cl::desc("By default, we analyze the benchmarks for the triple/CPU they "
225 "were measured for, but if you want to analyze them for some "
226 "other combination (specified via -mtriple/-mcpu), you can "
227 "pass this flag."),
228 cl::cat(AnalysisOptions), cl::init(false));
230 static cl::opt<std::string>
231 TripleName("mtriple",
232 cl::desc("Target triple. See -version for available targets"),
233 cl::cat(Options));
235 static cl::opt<std::string>
236 MCPU("mcpu",
237 cl::desc("Target a specific cpu type (-mcpu=help for details)"),
238 cl::value_desc("cpu-name"), cl::cat(Options), cl::init("native"));
240 static cl::opt<std::string>
241 DumpObjectToDisk("dump-object-to-disk",
242 cl::desc("dumps the generated benchmark object to disk "
243 "and prints a message to access it"),
244 cl::ValueOptional, cl::cat(BenchmarkOptions));
246 static cl::opt<BenchmarkRunner::ExecutionModeE> ExecutionMode(
247 "execution-mode",
248 cl::desc("Selects the execution mode to use for running snippets"),
249 cl::cat(BenchmarkOptions),
250 cl::values(clEnumValN(BenchmarkRunner::ExecutionModeE::InProcess,
251 "inprocess",
252 "Executes the snippets within the same process"),
253 clEnumValN(BenchmarkRunner::ExecutionModeE::SubProcess,
254 "subprocess",
255 "Spawns a subprocess for each snippet execution, "
256 "allows for the use of memory annotations")),
257 cl::init(BenchmarkRunner::ExecutionModeE::InProcess));
259 static cl::opt<unsigned> BenchmarkRepeatCount(
260 "benchmark-repeat-count",
261 cl::desc("The number of times to repeat measurements on the benchmark k "
262 "before aggregating the results"),
263 cl::cat(BenchmarkOptions), cl::init(30));
265 static cl::list<ValidationEvent> ValidationCounters(
266 "validation-counter",
267 cl::desc(
268 "The name of a validation counter to run concurrently with the main "
269 "counter to validate benchmarking assumptions"),
270 cl::CommaSeparated, cl::cat(BenchmarkOptions), ValidationEventOptions());
272 static cl::opt<int> BenchmarkProcessCPU(
273 "benchmark-process-cpu",
274 cl::desc("The CPU number that the benchmarking process should executon on"),
275 cl::cat(BenchmarkOptions), cl::init(-1));
277 static cl::opt<std::string> MAttr(
278 "mattr", cl::desc("comma-separated list of target architecture features"),
279 cl::value_desc("+feature1,-feature2,..."), cl::cat(Options), cl::init(""));
281 static ExitOnError ExitOnErr("llvm-exegesis error: ");
283 // Helper function that logs the error(s) and exits.
284 template <typename... ArgTs> static void ExitWithError(ArgTs &&... Args) {
285 ExitOnErr(make_error<Failure>(std::forward<ArgTs>(Args)...));
288 // Check Err. If it's in a failure state log the file error(s) and exit.
289 static void ExitOnFileError(const Twine &FileName, Error Err) {
290 if (Err) {
291 ExitOnErr(createFileError(FileName, std::move(Err)));
295 // Check E. If it's in a success state then return the contained value.
296 // If it's in a failure state log the file error(s) and exit.
297 template <typename T>
298 T ExitOnFileError(const Twine &FileName, Expected<T> &&E) {
299 ExitOnFileError(FileName, E.takeError());
300 return std::move(*E);
303 static const char *getIgnoredOpcodeReasonOrNull(const LLVMState &State,
304 unsigned Opcode) {
305 const MCInstrDesc &InstrDesc = State.getIC().getInstr(Opcode).Description;
306 if (InstrDesc.isPseudo() || InstrDesc.usesCustomInsertionHook())
307 return "Unsupported opcode: isPseudo/usesCustomInserter";
308 if (InstrDesc.isBranch() || InstrDesc.isIndirectBranch())
309 return "Unsupported opcode: isBranch/isIndirectBranch";
310 if (InstrDesc.isCall() || InstrDesc.isReturn())
311 return "Unsupported opcode: isCall/isReturn";
312 return nullptr;
315 // Checks that only one of OpcodeNames, OpcodeIndex or SnippetsFile is provided,
316 // and returns the opcode indices or {} if snippets should be read from
317 // `SnippetsFile`.
318 static std::vector<unsigned> getOpcodesOrDie(const LLVMState &State) {
319 const size_t NumSetFlags = (OpcodeNames.empty() ? 0 : 1) +
320 (OpcodeIndex == 0 ? 0 : 1) +
321 (SnippetsFile.empty() ? 0 : 1);
322 const auto &ET = State.getExegesisTarget();
323 const auto AvailableFeatures = State.getSubtargetInfo().getFeatureBits();
325 if (NumSetFlags != 1) {
326 ExitOnErr.setBanner("llvm-exegesis: ");
327 ExitWithError("please provide one and only one of 'opcode-index', "
328 "'opcode-name' or 'snippets-file'");
330 if (!SnippetsFile.empty())
331 return {};
332 if (OpcodeIndex > 0)
333 return {static_cast<unsigned>(OpcodeIndex)};
334 if (OpcodeIndex < 0) {
335 std::vector<unsigned> Result;
336 unsigned NumOpcodes = State.getInstrInfo().getNumOpcodes();
337 Result.reserve(NumOpcodes);
338 for (unsigned I = 0, E = NumOpcodes; I < E; ++I) {
339 if (!ET.isOpcodeAvailable(I, AvailableFeatures))
340 continue;
341 Result.push_back(I);
343 return Result;
345 // Resolve opcode name -> opcode.
346 const auto ResolveName = [&State](StringRef OpcodeName) -> unsigned {
347 const auto &Map = State.getOpcodeNameToOpcodeIdxMapping();
348 auto I = Map.find(OpcodeName);
349 if (I != Map.end())
350 return I->getSecond();
351 return 0u;
354 SmallVector<StringRef, 2> Pieces;
355 StringRef(OpcodeNames.getValue())
356 .split(Pieces, ",", /* MaxSplit */ -1, /* KeepEmpty */ false);
357 std::vector<unsigned> Result;
358 Result.reserve(Pieces.size());
359 for (const StringRef &OpcodeName : Pieces) {
360 if (unsigned Opcode = ResolveName(OpcodeName))
361 Result.push_back(Opcode);
362 else
363 ExitWithError(Twine("unknown opcode ").concat(OpcodeName));
365 return Result;
368 // Generates code snippets for opcode `Opcode`.
369 static Expected<std::vector<BenchmarkCode>>
370 generateSnippets(const LLVMState &State, unsigned Opcode,
371 const BitVector &ForbiddenRegs) {
372 // Ignore instructions that we cannot run.
373 if (const char *Reason = getIgnoredOpcodeReasonOrNull(State, Opcode))
374 return make_error<Failure>(Reason);
376 const Instruction &Instr = State.getIC().getInstr(Opcode);
377 const std::vector<InstructionTemplate> InstructionVariants =
378 State.getExegesisTarget().generateInstructionVariants(
379 Instr, MaxConfigsPerOpcode);
381 SnippetGenerator::Options SnippetOptions;
382 SnippetOptions.MaxConfigsPerOpcode = MaxConfigsPerOpcode;
383 const std::unique_ptr<SnippetGenerator> Generator =
384 State.getExegesisTarget().createSnippetGenerator(BenchmarkMode, State,
385 SnippetOptions);
386 if (!Generator)
387 ExitWithError("cannot create snippet generator");
389 std::vector<BenchmarkCode> Benchmarks;
390 for (const InstructionTemplate &Variant : InstructionVariants) {
391 if (Benchmarks.size() >= MaxConfigsPerOpcode)
392 break;
393 if (auto Err = Generator->generateConfigurations(Variant, Benchmarks,
394 ForbiddenRegs))
395 return std::move(Err);
397 return Benchmarks;
400 static void runBenchmarkConfigurations(
401 const LLVMState &State, ArrayRef<BenchmarkCode> Configurations,
402 ArrayRef<std::unique_ptr<const SnippetRepetitor>> Repetitors,
403 const BenchmarkRunner &Runner) {
404 assert(!Configurations.empty() && "Don't have any configurations to run.");
405 std::optional<raw_fd_ostream> FileOstr;
406 if (BenchmarkFile != "-") {
407 int ResultFD = 0;
408 // Create output file or open existing file and truncate it, once.
409 ExitOnErr(errorCodeToError(openFileForWrite(BenchmarkFile, ResultFD,
410 sys::fs::CD_CreateAlways,
411 sys::fs::OF_TextWithCRLF)));
412 FileOstr.emplace(ResultFD, true /*shouldClose*/);
414 raw_ostream &Ostr = FileOstr ? *FileOstr : outs();
416 std::optional<ProgressMeter<>> Meter;
417 if (BenchmarkMeasurementsPrintProgress)
418 Meter.emplace(Configurations.size());
420 SmallVector<unsigned, 2> MinInstructionCounts = {MinInstructions};
421 if (RepetitionMode == Benchmark::MiddleHalfDuplicate ||
422 RepetitionMode == Benchmark::MiddleHalfLoop)
423 MinInstructionCounts.push_back(MinInstructions * 2);
425 for (const BenchmarkCode &Conf : Configurations) {
426 ProgressMeter<>::ProgressMeterStep MeterStep(Meter ? &*Meter : nullptr);
427 SmallVector<Benchmark, 2> AllResults;
429 for (const std::unique_ptr<const SnippetRepetitor> &Repetitor :
430 Repetitors) {
431 for (unsigned IterationRepetitions : MinInstructionCounts) {
432 auto RC = ExitOnErr(Runner.getRunnableConfiguration(
433 Conf, IterationRepetitions, LoopBodySize, *Repetitor));
434 std::optional<StringRef> DumpFile;
435 if (DumpObjectToDisk.getNumOccurrences())
436 DumpFile = DumpObjectToDisk;
437 const std::optional<int> BenchmarkCPU =
438 BenchmarkProcessCPU == -1
439 ? std::nullopt
440 : std::optional(BenchmarkProcessCPU.getValue());
441 auto [Err, BenchmarkResult] =
442 Runner.runConfiguration(std::move(RC), DumpFile, BenchmarkCPU);
443 if (Err) {
444 // Errors from executing the snippets are fine.
445 // All other errors are a framework issue and should fail.
446 if (!Err.isA<SnippetExecutionFailure>())
447 ExitOnErr(std::move(Err));
449 BenchmarkResult.Error = toString(std::move(Err));
451 AllResults.push_back(std::move(BenchmarkResult));
455 Benchmark &Result = AllResults.front();
457 // If any of our measurements failed, pretend they all have failed.
458 if (AllResults.size() > 1 &&
459 any_of(AllResults, [](const Benchmark &R) {
460 return R.Measurements.empty();
462 Result.Measurements.clear();
464 std::unique_ptr<ResultAggregator> ResultAgg =
465 ResultAggregator::CreateAggregator(RepetitionMode);
466 ResultAgg->AggregateResults(Result,
467 ArrayRef<Benchmark>(AllResults).drop_front());
469 // With dummy counters, measurements are rather meaningless,
470 // so drop them altogether.
471 if (UseDummyPerfCounters)
472 Result.Measurements.clear();
474 ExitOnFileError(BenchmarkFile, Result.writeYamlTo(State, Ostr));
478 void benchmarkMain() {
479 if (BenchmarkPhaseSelector == BenchmarkPhaseSelectorE::Measure &&
480 !UseDummyPerfCounters) {
481 #ifndef HAVE_LIBPFM
482 ExitWithError(
483 "benchmarking unavailable, LLVM was built without libpfm. You can "
484 "pass --benchmark-phase=... to skip the actual benchmarking or "
485 "--use-dummy-perf-counters to not query the kernel for real event "
486 "counts.");
487 #else
488 if (pfm::pfmInitialize())
489 ExitWithError("cannot initialize libpfm");
490 #endif
493 InitializeAllExegesisTargets();
494 #define LLVM_EXEGESIS(TargetName) \
495 LLVMInitialize##TargetName##AsmPrinter(); \
496 LLVMInitialize##TargetName##AsmParser();
497 #include "llvm/Config/TargetExegesis.def"
499 const LLVMState State = ExitOnErr(
500 LLVMState::Create(TripleName, MCPU, MAttr, UseDummyPerfCounters));
502 // Preliminary check to ensure features needed for requested
503 // benchmark mode are present on target CPU and/or OS.
504 if (BenchmarkPhaseSelector == BenchmarkPhaseSelectorE::Measure)
505 ExitOnErr(State.getExegesisTarget().checkFeatureSupport());
507 if (ExecutionMode == BenchmarkRunner::ExecutionModeE::SubProcess &&
508 UseDummyPerfCounters)
509 ExitWithError("Dummy perf counters are not supported in the subprocess "
510 "execution mode.");
512 const std::unique_ptr<BenchmarkRunner> Runner =
513 ExitOnErr(State.getExegesisTarget().createBenchmarkRunner(
514 BenchmarkMode, State, BenchmarkPhaseSelector, ExecutionMode,
515 BenchmarkRepeatCount, ValidationCounters, ResultAggMode));
516 if (!Runner) {
517 ExitWithError("cannot create benchmark runner");
520 const auto Opcodes = getOpcodesOrDie(State);
521 std::vector<BenchmarkCode> Configurations;
523 unsigned LoopRegister =
524 State.getExegesisTarget().getDefaultLoopCounterRegister(
525 State.getTargetMachine().getTargetTriple());
527 if (Opcodes.empty()) {
528 Configurations = ExitOnErr(readSnippets(State, SnippetsFile));
529 for (const auto &Configuration : Configurations) {
530 if (ExecutionMode != BenchmarkRunner::ExecutionModeE::SubProcess &&
531 (Configuration.Key.MemoryMappings.size() != 0 ||
532 Configuration.Key.MemoryValues.size() != 0 ||
533 Configuration.Key.SnippetAddress != 0))
534 ExitWithError("Memory and snippet address annotations are only "
535 "supported in subprocess "
536 "execution mode");
538 LoopRegister = Configurations[0].Key.LoopRegister;
541 SmallVector<std::unique_ptr<const SnippetRepetitor>, 2> Repetitors;
542 if (RepetitionMode != Benchmark::RepetitionModeE::AggregateMin)
543 Repetitors.emplace_back(
544 SnippetRepetitor::Create(RepetitionMode, State, LoopRegister));
545 else {
546 for (Benchmark::RepetitionModeE RepMode :
547 {Benchmark::RepetitionModeE::Duplicate,
548 Benchmark::RepetitionModeE::Loop})
549 Repetitors.emplace_back(
550 SnippetRepetitor::Create(RepMode, State, LoopRegister));
553 BitVector AllReservedRegs;
554 for (const std::unique_ptr<const SnippetRepetitor> &Repetitor : Repetitors)
555 AllReservedRegs |= Repetitor->getReservedRegs();
557 if (!Opcodes.empty()) {
558 for (const unsigned Opcode : Opcodes) {
559 // Ignore instructions without a sched class if
560 // -ignore-invalid-sched-class is passed.
561 if (IgnoreInvalidSchedClass &&
562 State.getInstrInfo().get(Opcode).getSchedClass() == 0) {
563 errs() << State.getInstrInfo().getName(Opcode)
564 << ": ignoring instruction without sched class\n";
565 continue;
568 auto ConfigsForInstr = generateSnippets(State, Opcode, AllReservedRegs);
569 if (!ConfigsForInstr) {
570 logAllUnhandledErrors(
571 ConfigsForInstr.takeError(), errs(),
572 Twine(State.getInstrInfo().getName(Opcode)).concat(": "));
573 continue;
575 std::move(ConfigsForInstr->begin(), ConfigsForInstr->end(),
576 std::back_inserter(Configurations));
580 if (MinInstructions == 0) {
581 ExitOnErr.setBanner("llvm-exegesis: ");
582 ExitWithError("--min-instructions must be greater than zero");
585 // Write to standard output if file is not set.
586 if (BenchmarkFile.empty())
587 BenchmarkFile = "-";
589 if (!Configurations.empty())
590 runBenchmarkConfigurations(State, Configurations, Repetitors, *Runner);
592 pfm::pfmTerminate();
595 // Prints the results of running analysis pass `Pass` to file `OutputFilename`
596 // if OutputFilename is non-empty.
597 template <typename Pass>
598 static void maybeRunAnalysis(const Analysis &Analyzer, const std::string &Name,
599 const std::string &OutputFilename) {
600 if (OutputFilename.empty())
601 return;
602 if (OutputFilename != "-") {
603 errs() << "Printing " << Name << " results to file '" << OutputFilename
604 << "'\n";
606 std::error_code ErrorCode;
607 raw_fd_ostream ClustersOS(OutputFilename, ErrorCode,
608 sys::fs::FA_Read | sys::fs::FA_Write);
609 if (ErrorCode)
610 ExitOnFileError(OutputFilename, errorCodeToError(ErrorCode));
611 if (auto Err = Analyzer.run<Pass>(ClustersOS))
612 ExitOnFileError(OutputFilename, std::move(Err));
615 static void filterPoints(MutableArrayRef<Benchmark> Points,
616 const MCInstrInfo &MCII) {
617 if (AnalysisSnippetFilter == BenchmarkFilter::All)
618 return;
620 bool WantPointsWithMemOps = AnalysisSnippetFilter == BenchmarkFilter::WithMem;
621 for (Benchmark &Point : Points) {
622 if (!Point.Error.empty())
623 continue;
624 if (WantPointsWithMemOps ==
625 any_of(Point.Key.Instructions, [&MCII](const MCInst &Inst) {
626 const MCInstrDesc &MCDesc = MCII.get(Inst.getOpcode());
627 return MCDesc.mayLoad() || MCDesc.mayStore();
629 continue;
630 Point.Error = "filtered out by user";
634 static void analysisMain() {
635 ExitOnErr.setBanner("llvm-exegesis: ");
636 if (BenchmarkFile.empty())
637 ExitWithError("--benchmarks-file must be set");
639 if (AnalysisClustersOutputFile.empty() &&
640 AnalysisInconsistenciesOutputFile.empty()) {
641 ExitWithError(
642 "for --mode=analysis: At least one of --analysis-clusters-output-file "
643 "and --analysis-inconsistencies-output-file must be specified");
646 InitializeAllExegesisTargets();
647 #define LLVM_EXEGESIS(TargetName) \
648 LLVMInitialize##TargetName##AsmPrinter(); \
649 LLVMInitialize##TargetName##Disassembler();
650 #include "llvm/Config/TargetExegesis.def"
652 auto MemoryBuffer = ExitOnFileError(
653 BenchmarkFile,
654 errorOrToExpected(MemoryBuffer::getFile(BenchmarkFile, /*IsText=*/true)));
656 const auto TriplesAndCpus = ExitOnFileError(
657 BenchmarkFile,
658 Benchmark::readTriplesAndCpusFromYamls(*MemoryBuffer));
659 if (TriplesAndCpus.empty()) {
660 errs() << "no benchmarks to analyze\n";
661 return;
663 if (TriplesAndCpus.size() > 1) {
664 ExitWithError("analysis file contains benchmarks from several CPUs. This "
665 "is unsupported.");
667 auto TripleAndCpu = *TriplesAndCpus.begin();
668 if (AnalysisOverrideBenchmarksTripleAndCpu) {
669 errs() << "overridding file CPU name (" << TripleAndCpu.CpuName
670 << ") with provided tripled (" << TripleName << ") and CPU name ("
671 << MCPU << ")\n";
672 TripleAndCpu.LLVMTriple = TripleName;
673 TripleAndCpu.CpuName = MCPU;
675 errs() << "using Triple '" << TripleAndCpu.LLVMTriple << "' and CPU '"
676 << TripleAndCpu.CpuName << "'\n";
678 // Read benchmarks.
679 const LLVMState State = ExitOnErr(
680 LLVMState::Create(TripleAndCpu.LLVMTriple, TripleAndCpu.CpuName));
681 std::vector<Benchmark> Points = ExitOnFileError(
682 BenchmarkFile, Benchmark::readYamls(State, *MemoryBuffer));
684 outs() << "Parsed " << Points.size() << " benchmark points\n";
685 if (Points.empty()) {
686 errs() << "no benchmarks to analyze\n";
687 return;
689 // FIXME: Merge points from several runs (latency and uops).
691 filterPoints(Points, State.getInstrInfo());
693 const auto Clustering = ExitOnErr(BenchmarkClustering::create(
694 Points, AnalysisClusteringAlgorithm, AnalysisDbscanNumPoints,
695 AnalysisClusteringEpsilon, &State.getSubtargetInfo(),
696 &State.getInstrInfo()));
698 const Analysis Analyzer(State, Clustering, AnalysisInconsistencyEpsilon,
699 AnalysisDisplayUnstableOpcodes);
701 maybeRunAnalysis<Analysis::PrintClusters>(Analyzer, "analysis clusters",
702 AnalysisClustersOutputFile);
703 maybeRunAnalysis<Analysis::PrintSchedClassInconsistencies>(
704 Analyzer, "sched class consistency analysis",
705 AnalysisInconsistenciesOutputFile);
708 } // namespace exegesis
709 } // namespace llvm
711 int main(int Argc, char **Argv) {
712 using namespace llvm;
714 InitLLVM X(Argc, Argv);
716 // Initialize targets so we can print them when flag --version is specified.
717 #define LLVM_EXEGESIS(TargetName) \
718 LLVMInitialize##TargetName##Target(); \
719 LLVMInitialize##TargetName##TargetInfo(); \
720 LLVMInitialize##TargetName##TargetMC();
721 #include "llvm/Config/TargetExegesis.def"
723 // Register the Target and CPU printer for --version.
724 cl::AddExtraVersionPrinter(sys::printDefaultTargetAndDetectedCPU);
726 // Enable printing of available targets when flag --version is specified.
727 cl::AddExtraVersionPrinter(TargetRegistry::printRegisteredTargetsForVersion);
729 cl::HideUnrelatedOptions({&exegesis::Options, &exegesis::BenchmarkOptions,
730 &exegesis::AnalysisOptions});
732 cl::ParseCommandLineOptions(Argc, Argv,
733 "llvm host machine instruction characteristics "
734 "measurment and analysis.\n");
736 exegesis::ExitOnErr.setExitCodeMapper([](const Error &Err) {
737 if (Err.isA<exegesis::ClusteringError>())
738 return EXIT_SUCCESS;
739 return EXIT_FAILURE;
742 if (exegesis::BenchmarkMode == exegesis::Benchmark::Unknown) {
743 exegesis::analysisMain();
744 } else {
745 exegesis::benchmarkMain();
747 return EXIT_SUCCESS;