1 //===-- llvm-mca.cpp - Machine Code Analyzer -------------------*- 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 // This utility is a simple driver that allows static performance analysis on
10 // machine code similarly to how IACA (Intel Architecture Code Analyzer) works.
11 //
12 // llvm-mca [options] <file-name>
13 // -march <type>
14 // -mcpu <cpu>
15 // -o <file>
16 //
17 // The target defaults to the host target.
18 // The cpu defaults to the 'native' host cpu.
19 // The output defaults to standard output.
20 //
21 //===----------------------------------------------------------------------===//
22
23 #include "CodeRegion.h"
24 #include "CodeRegionGenerator.h"
25 #include "PipelinePrinter.h"
26 #include "Views/BottleneckAnalysis.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/MCAsmBackend.h"
36 #include "llvm/MC/MCAsmInfo.h"
37 #include "llvm/MC/MCCodeEmitter.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCObjectFileInfo.h"
40 #include "llvm/MC/MCRegisterInfo.h"
41 #include "llvm/MC/MCSubtargetInfo.h"
42 #include "llvm/MC/MCTargetOptionsCommandFlags.h"
43 #include "llvm/MCA/CodeEmitter.h"
44 #include "llvm/MCA/Context.h"
45 #include "llvm/MCA/InstrBuilder.h"
46 #include "llvm/MCA/Pipeline.h"
47 #include "llvm/MCA/Stages/EntryStage.h"
48 #include "llvm/MCA/Stages/InstructionTables.h"
49 #include "llvm/MCA/Support.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/ErrorOr.h"
53 #include "llvm/Support/FileSystem.h"
54 #include "llvm/Support/Host.h"
55 #include "llvm/Support/InitLLVM.h"
56 #include "llvm/Support/MemoryBuffer.h"
57 #include "llvm/Support/SourceMgr.h"
58 #include "llvm/Support/TargetRegistry.h"
59 #include "llvm/Support/TargetSelect.h"
60 #include "llvm/Support/ToolOutputFile.h"
61 #include "llvm/Support/WithColor.h"
62
63 using namespace llvm;
64
65 static mc::RegisterMCTargetOptionsFlags MOF;
66
67 static cl::OptionCategory ToolOptions("Tool Options");
68 static cl::OptionCategory ViewOptions("View Options");
69
70 static cl::opt<std::string> InputFilename(cl::Positional,
71 cl::desc("<input file>"),
72 cl::cat(ToolOptions), cl::init("-"));
73
74 static cl::opt<std::string> OutputFilename("o", cl::desc("Output filename"),
75 cl::init("-"), cl::cat(ToolOptions),
76 cl::value_desc("filename"));
77
78 static cl::opt<std::string>
79 ArchName("march",
80 cl::desc("Target architecture. "
81 "See -version for available targets"),
82 cl::cat(ToolOptions));
83
84 static cl::opt<std::string>
85 TripleName("mtriple",
86 cl::desc("Target triple. See -version for available targets"),
87 cl::cat(ToolOptions));
88
89 static cl::opt<std::string>
90 MCPU("mcpu",
91 cl::desc("Target a specific cpu type (-mcpu=help for details)"),
92 cl::value_desc("cpu-name"), cl::cat(ToolOptions), cl::init("native"));
93
94 static cl::opt<std::string>
95 MATTR("mattr",
96 cl::desc("Additional target features."),
97 cl::cat(ToolOptions));
98
99 static cl::opt<bool>
100 PrintJson("json",
101 cl::desc("Print the output in json format"),
102 cl::cat(ToolOptions), cl::init(false));
103
104 static cl::opt<int>
105 OutputAsmVariant("output-asm-variant",
106 cl::desc("Syntax variant to use for output printing"),
107 cl::cat(ToolOptions), cl::init(-1));
108
109 static cl::opt<bool>
110 PrintImmHex("print-imm-hex", cl::cat(ToolOptions), cl::init(false),
111 cl::desc("Prefer hex format when printing immediate values"));
112
113 static cl::opt<unsigned> Iterations("iterations",
114 cl::desc("Number of iterations to run"),
115 cl::cat(ToolOptions), cl::init(0));
116
117 static cl::opt<unsigned>
118 DispatchWidth("dispatch", cl::desc("Override the processor dispatch width"),
119 cl::cat(ToolOptions), cl::init(0));
120
121 static cl::opt<unsigned>
122 RegisterFileSize("register-file-size",
123 cl::desc("Maximum number of physical registers which can "
124 "be used for register mappings"),
125 cl::cat(ToolOptions), cl::init(0));
126
127 static cl::opt<unsigned>
128 MicroOpQueue("micro-op-queue-size", cl::Hidden,
129 cl::desc("Number of entries in the micro-op queue"),
130 cl::cat(ToolOptions), cl::init(0));
131
132 static cl::opt<unsigned>
133 DecoderThroughput("decoder-throughput", cl::Hidden,
134 cl::desc("Maximum throughput from the decoders "
135 "(instructions per cycle)"),
136 cl::cat(ToolOptions), cl::init(0));
137
138 static cl::opt<bool>
139 PrintRegisterFileStats("register-file-stats",
140 cl::desc("Print register file statistics"),
141 cl::cat(ViewOptions), cl::init(false));
142
143 static cl::opt<bool> PrintDispatchStats("dispatch-stats",
144 cl::desc("Print dispatch statistics"),
145 cl::cat(ViewOptions), cl::init(false));
146
147 static cl::opt<bool>
148 PrintSummaryView("summary-view", cl::Hidden,
149 cl::desc("Print summary view (enabled by default)"),
150 cl::cat(ViewOptions), cl::init(true));
151
152 static cl::opt<bool> PrintSchedulerStats("scheduler-stats",
153 cl::desc("Print scheduler statistics"),
154 cl::cat(ViewOptions), cl::init(false));
155
156 static cl::opt<bool>
157 PrintRetireStats("retire-stats",
158 cl::desc("Print retire control unit statistics"),
159 cl::cat(ViewOptions), cl::init(false));
160
161 static cl::opt<bool> PrintResourcePressureView(
162 "resource-pressure",
163 cl::desc("Print the resource pressure view (enabled by default)"),
164 cl::cat(ViewOptions), cl::init(true));
165
166 static cl::opt<bool> PrintTimelineView("timeline",
167 cl::desc("Print the timeline view"),
168 cl::cat(ViewOptions), cl::init(false));
169
170 static cl::opt<unsigned> TimelineMaxIterations(
171 "timeline-max-iterations",
172 cl::desc("Maximum number of iterations to print in timeline view"),
173 cl::cat(ViewOptions), cl::init(0));
174
175 static cl::opt<unsigned> TimelineMaxCycles(
176 "timeline-max-cycles",
177 cl::desc(
178 "Maximum number of cycles in the timeline view. Defaults to 80 cycles"),
179 cl::cat(ViewOptions), cl::init(80));
180
181 static cl::opt<bool>
182 AssumeNoAlias("noalias",
183 cl::desc("If set, assume that loads and stores do not alias"),
184 cl::cat(ToolOptions), cl::init(true));
185
186 static cl::opt<unsigned> LoadQueueSize("lqueue",
187 cl::desc("Size of the load queue"),
188 cl::cat(ToolOptions), cl::init(0));
189
190 static cl::opt<unsigned> StoreQueueSize("squeue",
191 cl::desc("Size of the store queue"),
192 cl::cat(ToolOptions), cl::init(0));
193
194 static cl::opt<bool>
195 PrintInstructionTables("instruction-tables",
196 cl::desc("Print instruction tables"),
197 cl::cat(ToolOptions), cl::init(false));
198
199 static cl::opt<bool> PrintInstructionInfoView(
200 "instruction-info",
201 cl::desc("Print the instruction info view (enabled by default)"),
202 cl::cat(ViewOptions), cl::init(true));
203
204 static cl::opt<bool> EnableAllStats("all-stats",
205 cl::desc("Print all hardware statistics"),
206 cl::cat(ViewOptions), cl::init(false));
207
208 static cl::opt<bool>
209 EnableAllViews("all-views",
210 cl::desc("Print all views including hardware statistics"),
211 cl::cat(ViewOptions), cl::init(false));
212
213 static cl::opt<bool> EnableBottleneckAnalysis(
214 "bottleneck-analysis",
215 cl::desc("Enable bottleneck analysis (disabled by default)"),
216 cl::cat(ViewOptions), cl::init(false));
217
218 static cl::opt<bool> ShowEncoding(
219 "show-encoding",
220 cl::desc("Print encoding information in the instruction info view"),
221 cl::cat(ViewOptions), cl::init(false));
222
223 namespace {
224
getTarget(const char * ProgName)225 const Target *getTarget(const char *ProgName) {
226 if (TripleName.empty())
227 TripleName = Triple::normalize(sys::getDefaultTargetTriple());
228 Triple TheTriple(TripleName);
229
230 // Get the target specific parser.
231 std::string Error;
232 const Target *TheTarget =
233 TargetRegistry::lookupTarget(ArchName, TheTriple, Error);
234 if (!TheTarget) {
235 errs() << ProgName << ": " << Error;
236 return nullptr;
237 }
238
239 // Return the found target.
240 return TheTarget;
241 }
242
getOutputStream()243 ErrorOr<std::unique_ptr<ToolOutputFile>> getOutputStream() {
244 if (OutputFilename == "")
245 OutputFilename = "-";
246 std::error_code EC;
247 auto Out =
248 std::make_unique<ToolOutputFile>(OutputFilename, EC, sys::fs::OF_Text);
249 if (!EC)
250 return std::move(Out);
251 return EC;
252 }
253 } // end of anonymous namespace
254
processOptionImpl(cl::opt<bool> & O,const cl::opt<bool> & Default)255 static void processOptionImpl(cl::opt<bool> &O, const cl::opt<bool> &Default) {
256 if (!O.getNumOccurrences() || O.getPosition() < Default.getPosition())
257 O = Default.getValue();
258 }
259
processViewOptions()260 static void processViewOptions() {
261 if (!EnableAllViews.getNumOccurrences() &&
262 !EnableAllStats.getNumOccurrences())
263 return;
264
265 if (EnableAllViews.getNumOccurrences()) {
266 processOptionImpl(PrintSummaryView, EnableAllViews);
267 processOptionImpl(EnableBottleneckAnalysis, EnableAllViews);
268 processOptionImpl(PrintResourcePressureView, EnableAllViews);
269 processOptionImpl(PrintTimelineView, EnableAllViews);
270 processOptionImpl(PrintInstructionInfoView, EnableAllViews);
271 }
272
273 const cl::opt<bool> &Default =
274 EnableAllViews.getPosition() < EnableAllStats.getPosition()
275 ? EnableAllStats
276 : EnableAllViews;
277 processOptionImpl(PrintRegisterFileStats, Default);
278 processOptionImpl(PrintDispatchStats, Default);
279 processOptionImpl(PrintSchedulerStats, Default);
280 processOptionImpl(PrintRetireStats, Default);
281 }
282
283 // Returns true on success.
runPipeline(mca::Pipeline & P)284 static bool runPipeline(mca::Pipeline &P) {
285 // Handle pipeline errors here.
286 Expected<unsigned> Cycles = P.run();
287 if (!Cycles) {
288 WithColor::error() << toString(Cycles.takeError());
289 return false;
290 }
291 return true;
292 }
293
main(int argc,char ** argv)294 int main(int argc, char **argv) {
295 InitLLVM X(argc, argv);
296
297 // Initialize targets and assembly parsers.
298 InitializeAllTargetInfos();
299 InitializeAllTargetMCs();
300 InitializeAllAsmParsers();
301
302 // Enable printing of available targets when flag --version is specified.
303 cl::AddExtraVersionPrinter(TargetRegistry::printRegisteredTargetsForVersion);
304
305 cl::HideUnrelatedOptions({&ToolOptions, &ViewOptions});
306
307 // Parse flags and initialize target options.
308 cl::ParseCommandLineOptions(argc, argv,
309 "llvm machine code performance analyzer.\n");
310
311 // Get the target from the triple. If a triple is not specified, then select
312 // the default triple for the host. If the triple doesn't correspond to any
313 // registered target, then exit with an error message.
314 const char *ProgName = argv[0];
315 const Target *TheTarget = getTarget(ProgName);
316 if (!TheTarget)
317 return 1;
318
319 // GetTarget() may replaced TripleName with a default triple.
320 // For safety, reconstruct the Triple object.
321 Triple TheTriple(TripleName);
322
323 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferPtr =
324 MemoryBuffer::getFileOrSTDIN(InputFilename);
325 if (std::error_code EC = BufferPtr.getError()) {
326 WithColor::error() << InputFilename << ": " << EC.message() << '\n';
327 return 1;
328 }
329
330 // Apply overrides to llvm-mca specific options.
331 processViewOptions();
332
333 if (MCPU == "native")
334 MCPU = std::string(llvm::sys::getHostCPUName());
335
336 std::unique_ptr<MCSubtargetInfo> STI(
337 TheTarget->createMCSubtargetInfo(TripleName, MCPU, MATTR));
338 assert(STI && "Unable to create subtarget info!");
339 if (!STI->isCPUStringValid(MCPU))
340 return 1;
341
342 if (!PrintInstructionTables && !STI->getSchedModel().isOutOfOrder()) {
343 WithColor::error() << "please specify an out-of-order cpu. '" << MCPU
344 << "' is an in-order cpu.\n";
345 return 1;
346 }
347
348 if (!STI->getSchedModel().hasInstrSchedModel()) {
349 WithColor::error()
350 << "unable to find instruction-level scheduling information for"
351 << " target triple '" << TheTriple.normalize() << "' and cpu '" << MCPU
352 << "'.\n";
353
354 if (STI->getSchedModel().InstrItineraries)
355 WithColor::note()
356 << "cpu '" << MCPU << "' provides itineraries. However, "
357 << "instruction itineraries are currently unsupported.\n";
358 return 1;
359 }
360
361 std::unique_ptr<MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TripleName));
362 assert(MRI && "Unable to create target register info!");
363
364 MCTargetOptions MCOptions = mc::InitMCTargetOptionsFromFlags();
365 std::unique_ptr<MCAsmInfo> MAI(
366 TheTarget->createMCAsmInfo(*MRI, TripleName, MCOptions));
367 assert(MAI && "Unable to create target asm info!");
368
369 MCObjectFileInfo MOFI;
370 SourceMgr SrcMgr;
371
372 // Tell SrcMgr about this buffer, which is what the parser will pick up.
373 SrcMgr.AddNewSourceBuffer(std::move(*BufferPtr), SMLoc());
374
375 MCContext Ctx(MAI.get(), MRI.get(), &MOFI, &SrcMgr);
376
377 MOFI.InitMCObjectFileInfo(TheTriple, /* PIC= */ false, Ctx);
378
379 std::unique_ptr<buffer_ostream> BOS;
380
381 std::unique_ptr<MCInstrInfo> MCII(TheTarget->createMCInstrInfo());
382 assert(MCII && "Unable to create instruction info!");
383
384 std::unique_ptr<MCInstrAnalysis> MCIA(
385 TheTarget->createMCInstrAnalysis(MCII.get()));
386
387 // Parse the input and create CodeRegions that llvm-mca can analyze.
388 mca::AsmCodeRegionGenerator CRG(*TheTarget, SrcMgr, Ctx, *MAI, *STI, *MCII);
389 Expected<const mca::CodeRegions &> RegionsOrErr = CRG.parseCodeRegions();
390 if (!RegionsOrErr) {
391 if (auto Err =
392 handleErrors(RegionsOrErr.takeError(), [](const StringError &E) {
393 WithColor::error() << E.getMessage() << '\n';
394 })) {
395 // Default case.
396 WithColor::error() << toString(std::move(Err)) << '\n';
397 }
398 return 1;
399 }
400 const mca::CodeRegions &Regions = *RegionsOrErr;
401
402 // Early exit if errors were found by the code region parsing logic.
403 if (!Regions.isValid())
404 return 1;
405
406 if (Regions.empty()) {
407 WithColor::error() << "no assembly instructions found.\n";
408 return 1;
409 }
410
411 // Now initialize the output file.
412 auto OF = getOutputStream();
413 if (std::error_code EC = OF.getError()) {
414 WithColor::error() << EC.message() << '\n';
415 return 1;
416 }
417
418 unsigned AssemblerDialect = CRG.getAssemblerDialect();
419 if (OutputAsmVariant >= 0)
420 AssemblerDialect = static_cast<unsigned>(OutputAsmVariant);
421 std::unique_ptr<MCInstPrinter> IP(TheTarget->createMCInstPrinter(
422 Triple(TripleName), AssemblerDialect, *MAI, *MCII, *MRI));
423 if (!IP) {
424 WithColor::error()
425 << "unable to create instruction printer for target triple '"
426 << TheTriple.normalize() << "' with assembly variant "
427 << AssemblerDialect << ".\n";
428 return 1;
429 }
430
431 // Set the display preference for hex vs. decimal immediates.
432 IP->setPrintImmHex(PrintImmHex);
433
434 std::unique_ptr<ToolOutputFile> TOF = std::move(*OF);
435
436 const MCSchedModel &SM = STI->getSchedModel();
437
438 // Create an instruction builder.
439 mca::InstrBuilder IB(*STI, *MCII, *MRI, MCIA.get());
440
441 // Create a context to control ownership of the pipeline hardware.
442 mca::Context MCA(*MRI, *STI);
443
444 mca::PipelineOptions PO(MicroOpQueue, DecoderThroughput, DispatchWidth,
445 RegisterFileSize, LoadQueueSize, StoreQueueSize,
446 AssumeNoAlias, EnableBottleneckAnalysis);
447
448 // Number each region in the sequence.
449 unsigned RegionIdx = 0;
450
451 std::unique_ptr<MCCodeEmitter> MCE(
452 TheTarget->createMCCodeEmitter(*MCII, *MRI, Ctx));
453 assert(MCE && "Unable to create code emitter!");
454
455 std::unique_ptr<MCAsmBackend> MAB(TheTarget->createMCAsmBackend(
456 *STI, *MRI, mc::InitMCTargetOptionsFromFlags()));
457 assert(MAB && "Unable to create asm backend!");
458
459 for (const std::unique_ptr<mca::CodeRegion> &Region : Regions) {
460 // Skip empty code regions.
461 if (Region->empty())
462 continue;
463
464 // Don't print the header of this region if it is the default region, and
465 // it doesn't have an end location.
466 if (Region->startLoc().isValid() || Region->endLoc().isValid()) {
467 TOF->os() << "\n[" << RegionIdx++ << "] Code Region";
468 StringRef Desc = Region->getDescription();
469 if (!Desc.empty())
470 TOF->os() << " - " << Desc;
471 TOF->os() << "\n\n";
472 }
473
474 // Lower the MCInst sequence into an mca::Instruction sequence.
475 ArrayRef<MCInst> Insts = Region->getInstructions();
476 mca::CodeEmitter CE(*STI, *MAB, *MCE, Insts);
477 std::vector<std::unique_ptr<mca::Instruction>> LoweredSequence;
478 for (const MCInst &MCI : Insts) {
479 Expected<std::unique_ptr<mca::Instruction>> Inst =
480 IB.createInstruction(MCI);
481 if (!Inst) {
482 if (auto NewE = handleErrors(
483 Inst.takeError(),
484 [&IP, &STI](const mca::InstructionError<MCInst> &IE) {
485 std::string InstructionStr;
486 raw_string_ostream SS(InstructionStr);
487 WithColor::error() << IE.Message << '\n';
488 IP->printInst(&IE.Inst, 0, "", *STI, SS);
489 SS.flush();
490 WithColor::note()
491 << "instruction: " << InstructionStr << '\n';
492 })) {
493 // Default case.
494 WithColor::error() << toString(std::move(NewE));
495 }
496 return 1;
497 }
498
499 LoweredSequence.emplace_back(std::move(Inst.get()));
500 }
501
502 mca::SourceMgr S(LoweredSequence, PrintInstructionTables ? 1 : Iterations);
503
504 if (PrintInstructionTables) {
505 // Create a pipeline, stages, and a printer.
506 auto P = std::make_unique<mca::Pipeline>();
507 P->appendStage(std::make_unique<mca::EntryStage>(S));
508 P->appendStage(std::make_unique<mca::InstructionTables>(SM));
509 mca::PipelinePrinter Printer(*P, mca::View::OK_READABLE);
510
511 // Create the views for this pipeline, execute, and emit a report.
512 if (PrintInstructionInfoView) {
513 Printer.addView(std::make_unique<mca::InstructionInfoView>(
514 *STI, *MCII, CE, ShowEncoding, Insts, *IP));
515 }
516 Printer.addView(
517 std::make_unique<mca::ResourcePressureView>(*STI, *IP, Insts));
518
519 if (!runPipeline(*P))
520 return 1;
521
522 Printer.printReport(TOF->os());
523 continue;
524 }
525
526 // Create a basic pipeline simulating an out-of-order backend.
527 auto P = MCA.createDefaultPipeline(PO, S);
528 mca::PipelinePrinter Printer(*P, PrintJson ? mca::View::OK_JSON
529 : mca::View::OK_READABLE);
530
531 // When we output JSON, we add a view that contains the instructions
532 // and CPU resource information.
533 if (PrintJson)
534 Printer.addView(
535 std::make_unique<mca::InstructionView>(*STI, *IP, Insts, MCPU));
536
537 if (PrintSummaryView)
538 Printer.addView(
539 std::make_unique<mca::SummaryView>(SM, Insts, DispatchWidth));
540
541 if (EnableBottleneckAnalysis) {
542 Printer.addView(std::make_unique<mca::BottleneckAnalysis>(
543 *STI, *IP, Insts, S.getNumIterations()));
544 }
545
546 if (PrintInstructionInfoView)
547 Printer.addView(std::make_unique<mca::InstructionInfoView>(
548 *STI, *MCII, CE, ShowEncoding, Insts, *IP));
549
550 if (PrintDispatchStats)
551 Printer.addView(std::make_unique<mca::DispatchStatistics>());
552
553 if (PrintSchedulerStats)
554 Printer.addView(std::make_unique<mca::SchedulerStatistics>(*STI));
555
556 if (PrintRetireStats)
557 Printer.addView(std::make_unique<mca::RetireControlUnitStatistics>(SM));
558
559 if (PrintRegisterFileStats)
560 Printer.addView(std::make_unique<mca::RegisterFileStatistics>(*STI));
561
562 if (PrintResourcePressureView)
563 Printer.addView(
564 std::make_unique<mca::ResourcePressureView>(*STI, *IP, Insts));
565
566 if (PrintTimelineView) {
567 unsigned TimelineIterations =
568 TimelineMaxIterations ? TimelineMaxIterations : 10;
569 Printer.addView(std::make_unique<mca::TimelineView>(
570 *STI, *IP, Insts, std::min(TimelineIterations, S.getNumIterations()),
571 TimelineMaxCycles));
572 }
573
574 if (!runPipeline(*P))
575 return 1;
576
577 Printer.printReport(TOF->os());
578
579 // Clear the InstrBuilder internal state in preparation for another round.
580 IB.clear();
581 }
582
583 TOF->keep();
584 return 0;
585 }
586