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 "Context.h"
26 #include "DispatchStatistics.h"
27 #include "FetchStage.h"
28 #include "InstructionInfoView.h"
29 #include "InstructionTables.h"
30 #include "Pipeline.h"
31 #include "PipelinePrinter.h"
32 #include "RegisterFileStatistics.h"
33 #include "ResourcePressureView.h"
34 #include "RetireControlUnitStatistics.h"
35 #include "SchedulerStatistics.h"
36 #include "SummaryView.h"
37 #include "TimelineView.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCObjectFileInfo.h"
41 #include "llvm/MC/MCParser/MCTargetAsmParser.h"
42 #include "llvm/MC/MCRegisterInfo.h"
43 #include "llvm/MC/MCStreamer.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/ErrorOr.h"
46 #include "llvm/Support/FileSystem.h"
47 #include "llvm/Support/Host.h"
48 #include "llvm/Support/InitLLVM.h"
49 #include "llvm/Support/MemoryBuffer.h"
50 #include "llvm/Support/SourceMgr.h"
51 #include "llvm/Support/TargetRegistry.h"
52 #include "llvm/Support/TargetSelect.h"
53 #include "llvm/Support/ToolOutputFile.h"
54 #include "llvm/Support/WithColor.h"
55
56 using namespace llvm;
57
58 static cl::OptionCategory ToolOptions("Tool Options");
59 static cl::OptionCategory ViewOptions("View Options");
60
61 static cl::opt<std::string> InputFilename(cl::Positional,
62 cl::desc("<input file>"),
63 cl::cat(ToolOptions), cl::init("-"));
64
65 static cl::opt<std::string> OutputFilename("o", cl::desc("Output filename"),
66 cl::init("-"), cl::cat(ToolOptions),
67 cl::value_desc("filename"));
68
69 static cl::opt<std::string>
70 ArchName("march", cl::desc("Target arch to assemble for, "
71 "see -version for available targets"),
72 cl::cat(ToolOptions));
73
74 static cl::opt<std::string>
75 TripleName("mtriple", cl::desc("Target triple to assemble for, "
76 "see -version for available targets"),
77 cl::cat(ToolOptions));
78
79 static cl::opt<std::string>
80 MCPU("mcpu",
81 cl::desc("Target a specific cpu type (-mcpu=help for details)"),
82 cl::value_desc("cpu-name"), cl::cat(ToolOptions), cl::init("native"));
83
84 static cl::opt<int>
85 OutputAsmVariant("output-asm-variant",
86 cl::desc("Syntax variant to use for output printing"),
87 cl::cat(ToolOptions), cl::init(-1));
88
89 static cl::opt<unsigned> Iterations("iterations",
90 cl::desc("Number of iterations to run"),
91 cl::cat(ToolOptions), cl::init(0));
92
93 static cl::opt<unsigned>
94 DispatchWidth("dispatch", cl::desc("Override the processor dispatch width"),
95 cl::cat(ToolOptions), cl::init(0));
96
97 static cl::opt<unsigned>
98 RegisterFileSize("register-file-size",
99 cl::desc("Maximum number of physical registers which can "
100 "be used for register mappings"),
101 cl::cat(ToolOptions), cl::init(0));
102
103 static cl::opt<bool>
104 PrintRegisterFileStats("register-file-stats",
105 cl::desc("Print register file statistics"),
106 cl::cat(ViewOptions), cl::init(false));
107
108 static cl::opt<bool> PrintDispatchStats("dispatch-stats",
109 cl::desc("Print dispatch statistics"),
110 cl::cat(ViewOptions), cl::init(false));
111
112 static cl::opt<bool>
113 PrintSummaryView("summary-view", cl::Hidden,
114 cl::desc("Print summary view (enabled by default)"),
115 cl::cat(ViewOptions), cl::init(true));
116
117 static cl::opt<bool> PrintSchedulerStats("scheduler-stats",
118 cl::desc("Print scheduler statistics"),
119 cl::cat(ViewOptions), cl::init(false));
120
121 static cl::opt<bool>
122 PrintRetireStats("retire-stats",
123 cl::desc("Print retire control unit statistics"),
124 cl::cat(ViewOptions), cl::init(false));
125
126 static cl::opt<bool> PrintResourcePressureView(
127 "resource-pressure",
128 cl::desc("Print the resource pressure view (enabled by default)"),
129 cl::cat(ViewOptions), cl::init(true));
130
131 static cl::opt<bool> PrintTimelineView("timeline",
132 cl::desc("Print the timeline view"),
133 cl::cat(ViewOptions), cl::init(false));
134
135 static cl::opt<unsigned> TimelineMaxIterations(
136 "timeline-max-iterations",
137 cl::desc("Maximum number of iterations to print in timeline view"),
138 cl::cat(ViewOptions), cl::init(0));
139
140 static cl::opt<unsigned> TimelineMaxCycles(
141 "timeline-max-cycles",
142 cl::desc(
143 "Maximum number of cycles in the timeline view. Defaults to 80 cycles"),
144 cl::cat(ViewOptions), cl::init(80));
145
146 static cl::opt<bool>
147 AssumeNoAlias("noalias",
148 cl::desc("If set, assume that loads and stores do not alias"),
149 cl::cat(ToolOptions), cl::init(true));
150
151 static cl::opt<unsigned>
152 LoadQueueSize("lqueue",
153 cl::desc("Size of the load queue (unbound by default)"),
154 cl::cat(ToolOptions), cl::init(0));
155
156 static cl::opt<unsigned>
157 StoreQueueSize("squeue",
158 cl::desc("Size of the store queue (unbound by default)"),
159 cl::cat(ToolOptions), cl::init(0));
160
161 static cl::opt<bool>
162 PrintInstructionTables("instruction-tables",
163 cl::desc("Print instruction tables"),
164 cl::cat(ToolOptions), cl::init(false));
165
166 static cl::opt<bool> PrintInstructionInfoView(
167 "instruction-info",
168 cl::desc("Print the instruction info view (enabled by default)"),
169 cl::cat(ViewOptions), cl::init(true));
170
171 static cl::opt<bool> EnableAllStats("all-stats",
172 cl::desc("Print all hardware statistics"),
173 cl::cat(ViewOptions), cl::init(false));
174
175 static cl::opt<bool>
176 EnableAllViews("all-views",
177 cl::desc("Print all views including hardware statistics"),
178 cl::cat(ViewOptions), cl::init(false));
179
180 namespace {
181
getTarget(const char * ProgName)182 const Target *getTarget(const char *ProgName) {
183 TripleName = Triple::normalize(TripleName);
184 if (TripleName.empty())
185 TripleName = Triple::normalize(sys::getDefaultTargetTriple());
186 Triple TheTriple(TripleName);
187
188 // Get the target specific parser.
189 std::string Error;
190 const Target *TheTarget =
191 TargetRegistry::lookupTarget(ArchName, TheTriple, Error);
192 if (!TheTarget) {
193 errs() << ProgName << ": " << Error;
194 return nullptr;
195 }
196
197 // Return the found target.
198 return TheTarget;
199 }
200
201 // A comment consumer that parses strings.
202 // The only valid tokens are strings.
203 class MCACommentConsumer : public AsmCommentConsumer {
204 public:
205 mca::CodeRegions &Regions;
206
MCACommentConsumer(mca::CodeRegions & R)207 MCACommentConsumer(mca::CodeRegions &R) : Regions(R) {}
HandleComment(SMLoc Loc,StringRef CommentText)208 void HandleComment(SMLoc Loc, StringRef CommentText) override {
209 // Skip empty comments.
210 StringRef Comment(CommentText);
211 if (Comment.empty())
212 return;
213
214 // Skip spaces and tabs
215 unsigned Position = Comment.find_first_not_of(" \t");
216 if (Position >= Comment.size())
217 // we reached the end of the comment. Bail out.
218 return;
219
220 Comment = Comment.drop_front(Position);
221 if (Comment.consume_front("LLVM-MCA-END")) {
222 Regions.endRegion(Loc);
223 return;
224 }
225
226 // Now try to parse string LLVM-MCA-BEGIN
227 if (!Comment.consume_front("LLVM-MCA-BEGIN"))
228 return;
229
230 // Skip spaces and tabs
231 Position = Comment.find_first_not_of(" \t");
232 if (Position < Comment.size())
233 Comment = Comment.drop_front(Position);
234 // Use the rest of the string as a descriptor for this code snippet.
235 Regions.beginRegion(Comment, Loc);
236 }
237 };
238
AssembleInput(const char * ProgName,MCAsmParser & Parser,const Target * TheTarget,MCSubtargetInfo & STI,MCInstrInfo & MCII,MCTargetOptions & MCOptions)239 int AssembleInput(const char *ProgName, MCAsmParser &Parser,
240 const Target *TheTarget, MCSubtargetInfo &STI,
241 MCInstrInfo &MCII, MCTargetOptions &MCOptions) {
242 std::unique_ptr<MCTargetAsmParser> TAP(
243 TheTarget->createMCAsmParser(STI, Parser, MCII, MCOptions));
244
245 if (!TAP) {
246 WithColor::error() << "this target does not support assembly parsing.\n";
247 return 1;
248 }
249
250 Parser.setTargetParser(*TAP);
251 return Parser.Run(false);
252 }
253
getOutputStream()254 ErrorOr<std::unique_ptr<ToolOutputFile>> getOutputStream() {
255 if (OutputFilename == "")
256 OutputFilename = "-";
257 std::error_code EC;
258 auto Out =
259 llvm::make_unique<ToolOutputFile>(OutputFilename, EC, sys::fs::F_None);
260 if (!EC)
261 return std::move(Out);
262 return EC;
263 }
264
265 class MCStreamerWrapper final : public MCStreamer {
266 mca::CodeRegions &Regions;
267
268 public:
MCStreamerWrapper(MCContext & Context,mca::CodeRegions & R)269 MCStreamerWrapper(MCContext &Context, mca::CodeRegions &R)
270 : MCStreamer(Context), Regions(R) {}
271
272 // We only want to intercept the emission of new instructions.
EmitInstruction(const MCInst & Inst,const MCSubtargetInfo & STI,bool)273 virtual void EmitInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
274 bool /* unused */) override {
275 Regions.addInstruction(llvm::make_unique<const MCInst>(Inst));
276 }
277
EmitSymbolAttribute(MCSymbol * Symbol,MCSymbolAttr Attribute)278 bool EmitSymbolAttribute(MCSymbol *Symbol, MCSymbolAttr Attribute) override {
279 return true;
280 }
281
EmitCommonSymbol(MCSymbol * Symbol,uint64_t Size,unsigned ByteAlignment)282 void EmitCommonSymbol(MCSymbol *Symbol, uint64_t Size,
283 unsigned ByteAlignment) override {}
EmitZerofill(MCSection * Section,MCSymbol * Symbol=nullptr,uint64_t Size=0,unsigned ByteAlignment=0,SMLoc Loc=SMLoc ())284 void EmitZerofill(MCSection *Section, MCSymbol *Symbol = nullptr,
285 uint64_t Size = 0, unsigned ByteAlignment = 0,
286 SMLoc Loc = SMLoc()) override {}
EmitGPRel32Value(const MCExpr * Value)287 void EmitGPRel32Value(const MCExpr *Value) override {}
BeginCOFFSymbolDef(const MCSymbol * Symbol)288 void BeginCOFFSymbolDef(const MCSymbol *Symbol) override {}
EmitCOFFSymbolStorageClass(int StorageClass)289 void EmitCOFFSymbolStorageClass(int StorageClass) override {}
EmitCOFFSymbolType(int Type)290 void EmitCOFFSymbolType(int Type) override {}
EndCOFFSymbolDef()291 void EndCOFFSymbolDef() override {}
292
293 const std::vector<std::unique_ptr<const MCInst>> &
GetInstructionSequence(unsigned Index) const294 GetInstructionSequence(unsigned Index) const {
295 return Regions.getInstructionSequence(Index);
296 }
297 };
298 } // end of anonymous namespace
299
processOptionImpl(cl::opt<bool> & O,const cl::opt<bool> & Default)300 static void processOptionImpl(cl::opt<bool> &O, const cl::opt<bool> &Default) {
301 if (!O.getNumOccurrences() || O.getPosition() < Default.getPosition())
302 O = Default.getValue();
303 }
304
processViewOptions()305 static void processViewOptions() {
306 if (!EnableAllViews.getNumOccurrences() &&
307 !EnableAllStats.getNumOccurrences())
308 return;
309
310 if (EnableAllViews.getNumOccurrences()) {
311 processOptionImpl(PrintSummaryView, EnableAllViews);
312 processOptionImpl(PrintResourcePressureView, EnableAllViews);
313 processOptionImpl(PrintTimelineView, EnableAllViews);
314 processOptionImpl(PrintInstructionInfoView, EnableAllViews);
315 }
316
317 const cl::opt<bool> &Default =
318 EnableAllViews.getPosition() < EnableAllStats.getPosition()
319 ? EnableAllStats
320 : EnableAllViews;
321 processOptionImpl(PrintSummaryView, Default);
322 processOptionImpl(PrintRegisterFileStats, Default);
323 processOptionImpl(PrintDispatchStats, Default);
324 processOptionImpl(PrintSchedulerStats, Default);
325 processOptionImpl(PrintRetireStats, Default);
326 }
327
main(int argc,char ** argv)328 int main(int argc, char **argv) {
329 InitLLVM X(argc, argv);
330
331 // Initialize targets and assembly parsers.
332 llvm::InitializeAllTargetInfos();
333 llvm::InitializeAllTargetMCs();
334 llvm::InitializeAllAsmParsers();
335
336 // Enable printing of available targets when flag --version is specified.
337 cl::AddExtraVersionPrinter(TargetRegistry::printRegisteredTargetsForVersion);
338
339 cl::HideUnrelatedOptions({&ToolOptions, &ViewOptions});
340
341 // Parse flags and initialize target options.
342 cl::ParseCommandLineOptions(argc, argv,
343 "llvm machine code performance analyzer.\n");
344
345 MCTargetOptions MCOptions;
346 MCOptions.PreserveAsmComments = false;
347
348 // Get the target from the triple. If a triple is not specified, then select
349 // the default triple for the host. If the triple doesn't correspond to any
350 // registered target, then exit with an error message.
351 const char *ProgName = argv[0];
352 const Target *TheTarget = getTarget(ProgName);
353 if (!TheTarget)
354 return 1;
355
356 // GetTarget() may replaced TripleName with a default triple.
357 // For safety, reconstruct the Triple object.
358 Triple TheTriple(TripleName);
359
360 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferPtr =
361 MemoryBuffer::getFileOrSTDIN(InputFilename);
362 if (std::error_code EC = BufferPtr.getError()) {
363 WithColor::error() << InputFilename << ": " << EC.message() << '\n';
364 return 1;
365 }
366
367 // Apply overrides to llvm-mca specific options.
368 processViewOptions();
369
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 std::unique_ptr<MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TripleName));
376 assert(MRI && "Unable to create target register info!");
377
378 std::unique_ptr<MCAsmInfo> MAI(TheTarget->createMCAsmInfo(*MRI, TripleName));
379 assert(MAI && "Unable to create target asm info!");
380
381 MCObjectFileInfo MOFI;
382 MCContext Ctx(MAI.get(), MRI.get(), &MOFI, &SrcMgr);
383 MOFI.InitMCObjectFileInfo(TheTriple, /* PIC= */ false, Ctx);
384
385 std::unique_ptr<buffer_ostream> BOS;
386
387 mca::CodeRegions Regions(SrcMgr);
388 MCStreamerWrapper Str(Ctx, Regions);
389
390 std::unique_ptr<MCInstrInfo> MCII(TheTarget->createMCInstrInfo());
391
392 std::unique_ptr<MCInstrAnalysis> MCIA(
393 TheTarget->createMCInstrAnalysis(MCII.get()));
394
395 if (!MCPU.compare("native"))
396 MCPU = llvm::sys::getHostCPUName();
397
398 std::unique_ptr<MCSubtargetInfo> STI(
399 TheTarget->createMCSubtargetInfo(TripleName, MCPU, /* FeaturesStr */ ""));
400 if (!STI->isCPUStringValid(MCPU))
401 return 1;
402
403 if (!PrintInstructionTables && !STI->getSchedModel().isOutOfOrder()) {
404 WithColor::error() << "please specify an out-of-order cpu. '" << MCPU
405 << "' is an in-order cpu.\n";
406 return 1;
407 }
408
409 if (!STI->getSchedModel().hasInstrSchedModel()) {
410 WithColor::error()
411 << "unable to find instruction-level scheduling information for"
412 << " target triple '" << TheTriple.normalize() << "' and cpu '" << MCPU
413 << "'.\n";
414
415 if (STI->getSchedModel().InstrItineraries)
416 WithColor::note()
417 << "cpu '" << MCPU << "' provides itineraries. However, "
418 << "instruction itineraries are currently unsupported.\n";
419 return 1;
420 }
421
422 std::unique_ptr<MCAsmParser> P(createMCAsmParser(SrcMgr, Ctx, Str, *MAI));
423 MCAsmLexer &Lexer = P->getLexer();
424 MCACommentConsumer CC(Regions);
425 Lexer.setCommentConsumer(&CC);
426
427 if (AssembleInput(ProgName, *P, TheTarget, *STI, *MCII, MCOptions))
428 return 1;
429
430 if (Regions.empty()) {
431 WithColor::error() << "no assembly instructions found.\n";
432 return 1;
433 }
434
435 // Now initialize the output file.
436 auto OF = getOutputStream();
437 if (std::error_code EC = OF.getError()) {
438 WithColor::error() << EC.message() << '\n';
439 return 1;
440 }
441
442 unsigned AssemblerDialect = P->getAssemblerDialect();
443 if (OutputAsmVariant >= 0)
444 AssemblerDialect = static_cast<unsigned>(OutputAsmVariant);
445 std::unique_ptr<MCInstPrinter> IP(TheTarget->createMCInstPrinter(
446 Triple(TripleName), AssemblerDialect, *MAI, *MCII, *MRI));
447 if (!IP) {
448 WithColor::error()
449 << "unable to create instruction printer for target triple '"
450 << TheTriple.normalize() << "' with assembly variant "
451 << AssemblerDialect << ".\n";
452 return 1;
453 }
454
455 std::unique_ptr<llvm::ToolOutputFile> TOF = std::move(*OF);
456
457 const MCSchedModel &SM = STI->getSchedModel();
458
459 unsigned Width = SM.IssueWidth;
460 if (DispatchWidth)
461 Width = DispatchWidth;
462
463 // Create an instruction builder.
464 mca::InstrBuilder IB(*STI, *MCII, *MRI, *MCIA, *IP);
465
466 // Create a context to control ownership of the pipeline hardware.
467 mca::Context MCA(*MRI, *STI);
468
469 mca::PipelineOptions PO(Width, RegisterFileSize, LoadQueueSize,
470 StoreQueueSize, AssumeNoAlias);
471
472 // Number each region in the sequence.
473 unsigned RegionIdx = 0;
474 for (const std::unique_ptr<mca::CodeRegion> &Region : Regions) {
475 // Skip empty code regions.
476 if (Region->empty())
477 continue;
478
479 // Don't print the header of this region if it is the default region, and
480 // it doesn't have an end location.
481 if (Region->startLoc().isValid() || Region->endLoc().isValid()) {
482 TOF->os() << "\n[" << RegionIdx++ << "] Code Region";
483 StringRef Desc = Region->getDescription();
484 if (!Desc.empty())
485 TOF->os() << " - " << Desc;
486 TOF->os() << "\n\n";
487 }
488
489 mca::SourceMgr S(Region->getInstructions(),
490 PrintInstructionTables ? 1 : Iterations);
491
492 if (PrintInstructionTables) {
493 // Create a pipeline, stages, and a printer.
494 auto P = llvm::make_unique<mca::Pipeline>();
495 P->appendStage(llvm::make_unique<mca::FetchStage>(IB, S));
496 P->appendStage(llvm::make_unique<mca::InstructionTables>(SM, IB));
497 mca::PipelinePrinter Printer(*P);
498
499 // Create the views for this pipeline, execute, and emit a report.
500 if (PrintInstructionInfoView) {
501 Printer.addView(
502 llvm::make_unique<mca::InstructionInfoView>(*STI, *MCII, S, *IP));
503 }
504 Printer.addView(
505 llvm::make_unique<mca::ResourcePressureView>(*STI, *IP, S));
506 P->run();
507 Printer.printReport(TOF->os());
508 continue;
509 }
510
511 // Create a basic pipeline simulating an out-of-order backend.
512 auto P = MCA.createDefaultPipeline(PO, IB, S);
513 mca::PipelinePrinter Printer(*P);
514
515 if (PrintSummaryView)
516 Printer.addView(llvm::make_unique<mca::SummaryView>(SM, S, Width));
517
518 if (PrintInstructionInfoView)
519 Printer.addView(
520 llvm::make_unique<mca::InstructionInfoView>(*STI, *MCII, S, *IP));
521
522 if (PrintDispatchStats)
523 Printer.addView(llvm::make_unique<mca::DispatchStatistics>());
524
525 if (PrintSchedulerStats)
526 Printer.addView(llvm::make_unique<mca::SchedulerStatistics>(*STI));
527
528 if (PrintRetireStats)
529 Printer.addView(llvm::make_unique<mca::RetireControlUnitStatistics>());
530
531 if (PrintRegisterFileStats)
532 Printer.addView(llvm::make_unique<mca::RegisterFileStatistics>(*STI));
533
534 if (PrintResourcePressureView)
535 Printer.addView(
536 llvm::make_unique<mca::ResourcePressureView>(*STI, *IP, S));
537
538 if (PrintTimelineView) {
539 Printer.addView(llvm::make_unique<mca::TimelineView>(
540 *STI, *IP, S, TimelineMaxIterations, TimelineMaxCycles));
541 }
542
543 P->run();
544 Printer.printReport(TOF->os());
545
546 // Clear the InstrBuilder internal state in preparation for another round.
547 IB.clear();
548 }
549
550 TOF->keep();
551 return 0;
552 }
553