1 //===-- Analysis.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 #include "Analysis.h" 10 #include "BenchmarkResult.h" 11 #include "llvm/ADT/STLExtras.h" 12 #include "llvm/MC/MCAsmInfo.h" 13 #include "llvm/Support/FormatVariadic.h" 14 #include <limits> 15 #include <unordered_set> 16 #include <vector> 17 18 namespace llvm { 19 namespace exegesis { 20 21 static const char kCsvSep = ','; 22 23 static unsigned resolveSchedClassId(const llvm::MCSubtargetInfo &STI, 24 unsigned SchedClassId, 25 const llvm::MCInst &MCI) { 26 const auto &SM = STI.getSchedModel(); 27 while (SchedClassId && SM.getSchedClassDesc(SchedClassId)->isVariant()) 28 SchedClassId = 29 STI.resolveVariantSchedClass(SchedClassId, &MCI, SM.getProcessorID()); 30 return SchedClassId; 31 } 32 33 namespace { 34 35 enum EscapeTag { kEscapeCsv, kEscapeHtml, kEscapeHtmlString }; 36 37 template <EscapeTag Tag> 38 void writeEscaped(llvm::raw_ostream &OS, const llvm::StringRef S); 39 40 template <> 41 void writeEscaped<kEscapeCsv>(llvm::raw_ostream &OS, const llvm::StringRef S) { 42 if (std::find(S.begin(), S.end(), kCsvSep) == S.end()) { 43 OS << S; 44 } else { 45 // Needs escaping. 46 OS << '"'; 47 for (const char C : S) { 48 if (C == '"') 49 OS << "\"\""; 50 else 51 OS << C; 52 } 53 OS << '"'; 54 } 55 } 56 57 template <> 58 void writeEscaped<kEscapeHtml>(llvm::raw_ostream &OS, const llvm::StringRef S) { 59 for (const char C : S) { 60 if (C == '<') 61 OS << "<"; 62 else if (C == '>') 63 OS << ">"; 64 else if (C == '&') 65 OS << "&"; 66 else 67 OS << C; 68 } 69 } 70 71 template <> 72 void writeEscaped<kEscapeHtmlString>(llvm::raw_ostream &OS, 73 const llvm::StringRef S) { 74 for (const char C : S) { 75 if (C == '"') 76 OS << "\\\""; 77 else 78 OS << C; 79 } 80 } 81 82 } // namespace 83 84 template <EscapeTag Tag> 85 static void 86 writeClusterId(llvm::raw_ostream &OS, 87 const InstructionBenchmarkClustering::ClusterId &CID) { 88 if (CID.isNoise()) 89 writeEscaped<Tag>(OS, "[noise]"); 90 else if (CID.isError()) 91 writeEscaped<Tag>(OS, "[error]"); 92 else 93 OS << CID.getId(); 94 } 95 96 template <EscapeTag Tag> 97 static void writeMeasurementValue(llvm::raw_ostream &OS, const double Value) { 98 // Given Value, if we wanted to serialize it to a string, 99 // how many base-10 digits will we need to store, max? 100 static constexpr auto MaxDigitCount = 101 std::numeric_limits<decltype(Value)>::max_digits10; 102 // Also, we will need a decimal separator. 103 static constexpr auto DecimalSeparatorLen = 1; // '.' e.g. 104 // So how long of a string will the serialization produce, max? 105 static constexpr auto SerializationLen = MaxDigitCount + DecimalSeparatorLen; 106 107 // WARNING: when changing the format, also adjust the small-size estimate ^. 108 static constexpr StringLiteral SimpleFloatFormat = StringLiteral("{0:F}"); 109 110 writeEscaped<Tag>( 111 OS, 112 llvm::formatv(SimpleFloatFormat.data(), Value).sstr<SerializationLen>()); 113 } 114 115 template <typename EscapeTag, EscapeTag Tag> 116 void Analysis::writeSnippet(llvm::raw_ostream &OS, 117 llvm::ArrayRef<uint8_t> Bytes, 118 const char *Separator) const { 119 llvm::SmallVector<std::string, 3> Lines; 120 // Parse the asm snippet and print it. 121 while (!Bytes.empty()) { 122 llvm::MCInst MI; 123 uint64_t MISize = 0; 124 if (!Disasm_->getInstruction(MI, MISize, Bytes, 0, llvm::nulls(), 125 llvm::nulls())) { 126 writeEscaped<Tag>(OS, llvm::join(Lines, Separator)); 127 writeEscaped<Tag>(OS, Separator); 128 writeEscaped<Tag>(OS, "[error decoding asm snippet]"); 129 return; 130 } 131 llvm::SmallString<128> InstPrinterStr; // FIXME: magic number. 132 llvm::raw_svector_ostream OSS(InstPrinterStr); 133 InstPrinter_->printInst(&MI, OSS, "", *SubtargetInfo_); 134 Bytes = Bytes.drop_front(MISize); 135 Lines.emplace_back(llvm::StringRef(InstPrinterStr).trim()); 136 } 137 writeEscaped<Tag>(OS, llvm::join(Lines, Separator)); 138 } 139 140 // Prints a row representing an instruction, along with scheduling info and 141 // point coordinates (measurements). 142 void Analysis::printInstructionRowCsv(const size_t PointId, 143 llvm::raw_ostream &OS) const { 144 const InstructionBenchmark &Point = Clustering_.getPoints()[PointId]; 145 writeClusterId<kEscapeCsv>(OS, Clustering_.getClusterIdForPoint(PointId)); 146 OS << kCsvSep; 147 writeSnippet<EscapeTag, kEscapeCsv>(OS, Point.AssembledSnippet, "; "); 148 OS << kCsvSep; 149 writeEscaped<kEscapeCsv>(OS, Point.Key.Config); 150 OS << kCsvSep; 151 assert(!Point.Key.Instructions.empty()); 152 const llvm::MCInst &MCI = Point.keyInstruction(); 153 const unsigned SchedClassId = resolveSchedClassId( 154 *SubtargetInfo_, InstrInfo_->get(MCI.getOpcode()).getSchedClass(), MCI); 155 156 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 157 const llvm::MCSchedClassDesc *const SCDesc = 158 SubtargetInfo_->getSchedModel().getSchedClassDesc(SchedClassId); 159 writeEscaped<kEscapeCsv>(OS, SCDesc->Name); 160 #else 161 OS << SchedClassId; 162 #endif 163 for (const auto &Measurement : Point.Measurements) { 164 OS << kCsvSep; 165 writeMeasurementValue<kEscapeCsv>(OS, Measurement.PerInstructionValue); 166 } 167 OS << "\n"; 168 } 169 170 Analysis::Analysis(const llvm::Target &Target, 171 std::unique_ptr<llvm::MCInstrInfo> InstrInfo, 172 const InstructionBenchmarkClustering &Clustering, 173 bool AnalysisDisplayUnstableOpcodes) 174 : Clustering_(Clustering), InstrInfo_(std::move(InstrInfo)), 175 AnalysisDisplayUnstableOpcodes_(AnalysisDisplayUnstableOpcodes) { 176 if (Clustering.getPoints().empty()) 177 return; 178 179 const InstructionBenchmark &FirstPoint = Clustering.getPoints().front(); 180 RegInfo_.reset(Target.createMCRegInfo(FirstPoint.LLVMTriple)); 181 AsmInfo_.reset(Target.createMCAsmInfo(*RegInfo_, FirstPoint.LLVMTriple)); 182 SubtargetInfo_.reset(Target.createMCSubtargetInfo(FirstPoint.LLVMTriple, 183 FirstPoint.CpuName, "")); 184 InstPrinter_.reset(Target.createMCInstPrinter( 185 llvm::Triple(FirstPoint.LLVMTriple), 0 /*default variant*/, *AsmInfo_, 186 *InstrInfo_, *RegInfo_)); 187 188 Context_ = llvm::make_unique<llvm::MCContext>(AsmInfo_.get(), RegInfo_.get(), 189 &ObjectFileInfo_); 190 Disasm_.reset(Target.createMCDisassembler(*SubtargetInfo_, *Context_)); 191 assert(Disasm_ && "cannot create MCDisassembler. missing call to " 192 "InitializeXXXTargetDisassembler ?"); 193 } 194 195 template <> 196 llvm::Error 197 Analysis::run<Analysis::PrintClusters>(llvm::raw_ostream &OS) const { 198 if (Clustering_.getPoints().empty()) 199 return llvm::Error::success(); 200 201 // Write the header. 202 OS << "cluster_id" << kCsvSep << "opcode_name" << kCsvSep << "config" 203 << kCsvSep << "sched_class"; 204 for (const auto &Measurement : Clustering_.getPoints().front().Measurements) { 205 OS << kCsvSep; 206 writeEscaped<kEscapeCsv>(OS, Measurement.Key); 207 } 208 OS << "\n"; 209 210 // Write the points. 211 const auto &Clusters = Clustering_.getValidClusters(); 212 for (size_t I = 0, E = Clusters.size(); I < E; ++I) { 213 for (const size_t PointId : Clusters[I].PointIndices) { 214 printInstructionRowCsv(PointId, OS); 215 } 216 OS << "\n\n"; 217 } 218 return llvm::Error::success(); 219 } 220 221 Analysis::ResolvedSchedClassAndPoints::ResolvedSchedClassAndPoints( 222 ResolvedSchedClass &&RSC) 223 : RSC(std::move(RSC)) {} 224 225 std::vector<Analysis::ResolvedSchedClassAndPoints> 226 Analysis::makePointsPerSchedClass() const { 227 std::vector<ResolvedSchedClassAndPoints> Entries; 228 // Maps SchedClassIds to index in result. 229 std::unordered_map<unsigned, size_t> SchedClassIdToIndex; 230 const auto &Points = Clustering_.getPoints(); 231 for (size_t PointId = 0, E = Points.size(); PointId < E; ++PointId) { 232 const InstructionBenchmark &Point = Points[PointId]; 233 if (!Point.Error.empty()) 234 continue; 235 assert(!Point.Key.Instructions.empty()); 236 // FIXME: we should be using the tuple of classes for instructions in the 237 // snippet as key. 238 const llvm::MCInst &MCI = Point.keyInstruction(); 239 unsigned SchedClassId = InstrInfo_->get(MCI.getOpcode()).getSchedClass(); 240 const bool WasVariant = SchedClassId && SubtargetInfo_->getSchedModel() 241 .getSchedClassDesc(SchedClassId) 242 ->isVariant(); 243 SchedClassId = resolveSchedClassId(*SubtargetInfo_, SchedClassId, MCI); 244 const auto IndexIt = SchedClassIdToIndex.find(SchedClassId); 245 if (IndexIt == SchedClassIdToIndex.end()) { 246 // Create a new entry. 247 SchedClassIdToIndex.emplace(SchedClassId, Entries.size()); 248 ResolvedSchedClassAndPoints Entry( 249 ResolvedSchedClass(*SubtargetInfo_, SchedClassId, WasVariant)); 250 Entry.PointIds.push_back(PointId); 251 Entries.push_back(std::move(Entry)); 252 } else { 253 // Append to the existing entry. 254 Entries[IndexIt->second].PointIds.push_back(PointId); 255 } 256 } 257 return Entries; 258 } 259 260 // Uops repeat the same opcode over again. Just show this opcode and show the 261 // whole snippet only on hover. 262 static void writeUopsSnippetHtml(llvm::raw_ostream &OS, 263 const std::vector<llvm::MCInst> &Instructions, 264 const llvm::MCInstrInfo &InstrInfo) { 265 if (Instructions.empty()) 266 return; 267 writeEscaped<kEscapeHtml>(OS, InstrInfo.getName(Instructions[0].getOpcode())); 268 if (Instructions.size() > 1) 269 OS << " (x" << Instructions.size() << ")"; 270 } 271 272 // Latency tries to find a serial path. Just show the opcode path and show the 273 // whole snippet only on hover. 274 static void 275 writeLatencySnippetHtml(llvm::raw_ostream &OS, 276 const std::vector<llvm::MCInst> &Instructions, 277 const llvm::MCInstrInfo &InstrInfo) { 278 bool First = true; 279 for (const llvm::MCInst &Instr : Instructions) { 280 if (First) 281 First = false; 282 else 283 OS << " → "; 284 writeEscaped<kEscapeHtml>(OS, InstrInfo.getName(Instr.getOpcode())); 285 } 286 } 287 288 void Analysis::printSchedClassClustersHtml( 289 const std::vector<SchedClassCluster> &Clusters, 290 const ResolvedSchedClass &RSC, llvm::raw_ostream &OS) const { 291 const auto &Points = Clustering_.getPoints(); 292 OS << "<table class=\"sched-class-clusters\">"; 293 OS << "<tr><th>ClusterId</th><th>Opcode/Config</th>"; 294 assert(!Clusters.empty()); 295 for (const auto &Measurement : 296 Points[Clusters[0].getPointIds()[0]].Measurements) { 297 OS << "<th>"; 298 writeEscaped<kEscapeHtml>(OS, Measurement.Key); 299 OS << "</th>"; 300 } 301 OS << "</tr>"; 302 for (const SchedClassCluster &Cluster : Clusters) { 303 OS << "<tr class=\"" 304 << (Cluster.measurementsMatch(*SubtargetInfo_, RSC, Clustering_) 305 ? "good-cluster" 306 : "bad-cluster") 307 << "\"><td>"; 308 writeClusterId<kEscapeHtml>(OS, Cluster.id()); 309 OS << "</td><td><ul>"; 310 for (const size_t PointId : Cluster.getPointIds()) { 311 const auto &Point = Points[PointId]; 312 OS << "<li><span class=\"mono\" title=\""; 313 writeSnippet<EscapeTag, kEscapeHtmlString>(OS, Point.AssembledSnippet, 314 "\n"); 315 OS << "\">"; 316 switch (Point.Mode) { 317 case InstructionBenchmark::Latency: 318 writeLatencySnippetHtml(OS, Point.Key.Instructions, *InstrInfo_); 319 break; 320 case InstructionBenchmark::Uops: 321 case InstructionBenchmark::InverseThroughput: 322 writeUopsSnippetHtml(OS, Point.Key.Instructions, *InstrInfo_); 323 break; 324 default: 325 llvm_unreachable("invalid mode"); 326 } 327 OS << "</span> <span class=\"mono\">"; 328 writeEscaped<kEscapeHtml>(OS, Point.Key.Config); 329 OS << "</span></li>"; 330 } 331 OS << "</ul></td>"; 332 for (const auto &Stats : Cluster.getRepresentative()) { 333 OS << "<td class=\"measurement\">"; 334 writeMeasurementValue<kEscapeHtml>(OS, Stats.avg()); 335 OS << "<br><span class=\"minmax\">["; 336 writeMeasurementValue<kEscapeHtml>(OS, Stats.min()); 337 OS << ";"; 338 writeMeasurementValue<kEscapeHtml>(OS, Stats.max()); 339 OS << "]</span></td>"; 340 } 341 OS << "</tr>"; 342 } 343 OS << "</table>"; 344 } 345 346 // Return the non-redundant list of WriteProcRes used by the given sched class. 347 // The scheduling model for LLVM is such that each instruction has a certain 348 // number of uops which consume resources which are described by WriteProcRes 349 // entries. Each entry describe how many cycles are spent on a specific ProcRes 350 // kind. 351 // For example, an instruction might have 3 uOps, one dispatching on P0 352 // (ProcResIdx=1) and two on P06 (ProcResIdx = 7). 353 // Note that LLVM additionally denormalizes resource consumption to include 354 // usage of super resources by subresources. So in practice if there exists a 355 // P016 (ProcResIdx=10), then the cycles consumed by P0 are also consumed by 356 // P06 (ProcResIdx = 7) and P016 (ProcResIdx = 10), and the resources consumed 357 // by P06 are also consumed by P016. In the figure below, parenthesized cycles 358 // denote implied usage of superresources by subresources: 359 // P0 P06 P016 360 // uOp1 1 (1) (1) 361 // uOp2 1 (1) 362 // uOp3 1 (1) 363 // ============================= 364 // 1 3 3 365 // Eventually we end up with three entries for the WriteProcRes of the 366 // instruction: 367 // {ProcResIdx=1, Cycles=1} // P0 368 // {ProcResIdx=7, Cycles=3} // P06 369 // {ProcResIdx=10, Cycles=3} // P016 370 // 371 // Note that in this case, P016 does not contribute any cycles, so it would 372 // be removed by this function. 373 // FIXME: Move this to MCSubtargetInfo and use it in llvm-mca. 374 static llvm::SmallVector<llvm::MCWriteProcResEntry, 8> 375 getNonRedundantWriteProcRes(const llvm::MCSchedClassDesc &SCDesc, 376 const llvm::MCSubtargetInfo &STI) { 377 llvm::SmallVector<llvm::MCWriteProcResEntry, 8> Result; 378 const auto &SM = STI.getSchedModel(); 379 const unsigned NumProcRes = SM.getNumProcResourceKinds(); 380 381 // This assumes that the ProcResDescs are sorted in topological order, which 382 // is guaranteed by the tablegen backend. 383 llvm::SmallVector<float, 32> ProcResUnitUsage(NumProcRes); 384 for (const auto *WPR = STI.getWriteProcResBegin(&SCDesc), 385 *const WPREnd = STI.getWriteProcResEnd(&SCDesc); 386 WPR != WPREnd; ++WPR) { 387 const llvm::MCProcResourceDesc *const ProcResDesc = 388 SM.getProcResource(WPR->ProcResourceIdx); 389 if (ProcResDesc->SubUnitsIdxBegin == nullptr) { 390 // This is a ProcResUnit. 391 Result.push_back({WPR->ProcResourceIdx, WPR->Cycles}); 392 ProcResUnitUsage[WPR->ProcResourceIdx] += WPR->Cycles; 393 } else { 394 // This is a ProcResGroup. First see if it contributes any cycles or if 395 // it has cycles just from subunits. 396 float RemainingCycles = WPR->Cycles; 397 for (const auto *SubResIdx = ProcResDesc->SubUnitsIdxBegin; 398 SubResIdx != ProcResDesc->SubUnitsIdxBegin + ProcResDesc->NumUnits; 399 ++SubResIdx) { 400 RemainingCycles -= ProcResUnitUsage[*SubResIdx]; 401 } 402 if (RemainingCycles < 0.01f) { 403 // The ProcResGroup contributes no cycles of its own. 404 continue; 405 } 406 // The ProcResGroup contributes `RemainingCycles` cycles of its own. 407 Result.push_back({WPR->ProcResourceIdx, 408 static_cast<uint16_t>(std::round(RemainingCycles))}); 409 // Spread the remaining cycles over all subunits. 410 for (const auto *SubResIdx = ProcResDesc->SubUnitsIdxBegin; 411 SubResIdx != ProcResDesc->SubUnitsIdxBegin + ProcResDesc->NumUnits; 412 ++SubResIdx) { 413 ProcResUnitUsage[*SubResIdx] += RemainingCycles / ProcResDesc->NumUnits; 414 } 415 } 416 } 417 return Result; 418 } 419 420 Analysis::ResolvedSchedClass::ResolvedSchedClass( 421 const llvm::MCSubtargetInfo &STI, unsigned ResolvedSchedClassId, 422 bool WasVariant) 423 : SchedClassId(ResolvedSchedClassId), SCDesc(STI.getSchedModel().getSchedClassDesc(ResolvedSchedClassId)), 424 WasVariant(WasVariant), 425 NonRedundantWriteProcRes(getNonRedundantWriteProcRes(*SCDesc, STI)), 426 IdealizedProcResPressure(computeIdealizedProcResPressure( 427 STI.getSchedModel(), NonRedundantWriteProcRes)) { 428 assert((SCDesc == nullptr || !SCDesc->isVariant()) && 429 "ResolvedSchedClass should never be variant"); 430 } 431 432 void Analysis::SchedClassCluster::addPoint( 433 size_t PointId, const InstructionBenchmarkClustering &Clustering) { 434 PointIds.push_back(PointId); 435 const auto &Point = Clustering.getPoints()[PointId]; 436 if (ClusterId.isUndef()) { 437 ClusterId = Clustering.getClusterIdForPoint(PointId); 438 Representative.resize(Point.Measurements.size()); 439 } 440 for (size_t I = 0, E = Point.Measurements.size(); I < E; ++I) { 441 Representative[I].push(Point.Measurements[I]); 442 } 443 assert(ClusterId == Clustering.getClusterIdForPoint(PointId)); 444 } 445 446 // Returns a ProxResIdx by id or name. 447 static unsigned findProcResIdx(const llvm::MCSubtargetInfo &STI, 448 const llvm::StringRef NameOrId) { 449 // Interpret the key as an ProcResIdx. 450 unsigned ProcResIdx = 0; 451 if (llvm::to_integer(NameOrId, ProcResIdx, 10)) 452 return ProcResIdx; 453 // Interpret the key as a ProcRes name. 454 const auto &SchedModel = STI.getSchedModel(); 455 for (int I = 0, E = SchedModel.getNumProcResourceKinds(); I < E; ++I) { 456 if (NameOrId == SchedModel.getProcResource(I)->Name) 457 return I; 458 } 459 return 0; 460 } 461 462 bool Analysis::SchedClassCluster::measurementsMatch( 463 const llvm::MCSubtargetInfo &STI, const ResolvedSchedClass &RSC, 464 const InstructionBenchmarkClustering &Clustering) const { 465 const size_t NumMeasurements = Representative.size(); 466 std::vector<BenchmarkMeasure> ClusterCenterPoint(NumMeasurements); 467 std::vector<BenchmarkMeasure> SchedClassPoint(NumMeasurements); 468 // Latency case. 469 assert(!Clustering.getPoints().empty()); 470 const InstructionBenchmark::ModeE Mode = Clustering.getPoints()[0].Mode; 471 if (Mode == InstructionBenchmark::Latency) { 472 if (NumMeasurements != 1) { 473 llvm::errs() 474 << "invalid number of measurements in latency mode: expected 1, got " 475 << NumMeasurements << "\n"; 476 return false; 477 } 478 // Find the latency. 479 SchedClassPoint[0].PerInstructionValue = 0.0; 480 for (unsigned I = 0; I < RSC.SCDesc->NumWriteLatencyEntries; ++I) { 481 const llvm::MCWriteLatencyEntry *const WLE = 482 STI.getWriteLatencyEntry(RSC.SCDesc, I); 483 SchedClassPoint[0].PerInstructionValue = 484 std::max<double>(SchedClassPoint[0].PerInstructionValue, WLE->Cycles); 485 } 486 ClusterCenterPoint[0].PerInstructionValue = Representative[0].avg(); 487 } else if (Mode == InstructionBenchmark::Uops) { 488 for (int I = 0, E = Representative.size(); I < E; ++I) { 489 const auto Key = Representative[I].key(); 490 uint16_t ProcResIdx = findProcResIdx(STI, Key); 491 if (ProcResIdx > 0) { 492 // Find the pressure on ProcResIdx `Key`. 493 const auto ProcResPressureIt = 494 std::find_if(RSC.IdealizedProcResPressure.begin(), 495 RSC.IdealizedProcResPressure.end(), 496 [ProcResIdx](const std::pair<uint16_t, float> &WPR) { 497 return WPR.first == ProcResIdx; 498 }); 499 SchedClassPoint[I].PerInstructionValue = 500 ProcResPressureIt == RSC.IdealizedProcResPressure.end() 501 ? 0.0 502 : ProcResPressureIt->second; 503 } else if (Key == "NumMicroOps") { 504 SchedClassPoint[I].PerInstructionValue = RSC.SCDesc->NumMicroOps; 505 } else { 506 llvm::errs() << "expected `key` to be either a ProcResIdx or a ProcRes " 507 "name, got " 508 << Key << "\n"; 509 return false; 510 } 511 ClusterCenterPoint[I].PerInstructionValue = Representative[I].avg(); 512 } 513 } else if (Mode == InstructionBenchmark::InverseThroughput) { 514 for (int I = 0, E = Representative.size(); I < E; ++I) { 515 SchedClassPoint[I].PerInstructionValue = 516 MCSchedModel::getReciprocalThroughput(STI, *RSC.SCDesc); 517 ClusterCenterPoint[I].PerInstructionValue = Representative[I].min(); 518 } 519 } else { 520 llvm_unreachable("unimplemented measurement matching mode"); 521 return false; 522 } 523 return Clustering.isNeighbour(ClusterCenterPoint, SchedClassPoint); 524 } 525 526 void Analysis::printSchedClassDescHtml(const ResolvedSchedClass &RSC, 527 llvm::raw_ostream &OS) const { 528 OS << "<table class=\"sched-class-desc\">"; 529 OS << "<tr><th>Valid</th><th>Variant</th><th>NumMicroOps</th><th>Latency</" 530 "th><th>RThroughput</th><th>WriteProcRes</th><th title=\"This is the " 531 "idealized unit resource (port) pressure assuming ideal " 532 "distribution\">Idealized Resource Pressure</th></tr>"; 533 if (RSC.SCDesc->isValid()) { 534 const auto &SM = SubtargetInfo_->getSchedModel(); 535 OS << "<tr><td>✔</td>"; 536 OS << "<td>" << (RSC.WasVariant ? "✔" : "✕") << "</td>"; 537 OS << "<td>" << RSC.SCDesc->NumMicroOps << "</td>"; 538 // Latencies. 539 OS << "<td><ul>"; 540 for (int I = 0, E = RSC.SCDesc->NumWriteLatencyEntries; I < E; ++I) { 541 const auto *const Entry = 542 SubtargetInfo_->getWriteLatencyEntry(RSC.SCDesc, I); 543 OS << "<li>" << Entry->Cycles; 544 if (RSC.SCDesc->NumWriteLatencyEntries > 1) { 545 // Dismabiguate if more than 1 latency. 546 OS << " (WriteResourceID " << Entry->WriteResourceID << ")"; 547 } 548 OS << "</li>"; 549 } 550 OS << "</ul></td>"; 551 // inverse throughput. 552 OS << "<td>"; 553 writeMeasurementValue<kEscapeHtml>( 554 OS, 555 MCSchedModel::getReciprocalThroughput(*SubtargetInfo_, *RSC.SCDesc)); 556 OS << "</td>"; 557 // WriteProcRes. 558 OS << "<td><ul>"; 559 for (const auto &WPR : RSC.NonRedundantWriteProcRes) { 560 OS << "<li><span class=\"mono\">"; 561 writeEscaped<kEscapeHtml>(OS, 562 SM.getProcResource(WPR.ProcResourceIdx)->Name); 563 OS << "</span>: " << WPR.Cycles << "</li>"; 564 } 565 OS << "</ul></td>"; 566 // Idealized port pressure. 567 OS << "<td><ul>"; 568 for (const auto &Pressure : RSC.IdealizedProcResPressure) { 569 OS << "<li><span class=\"mono\">"; 570 writeEscaped<kEscapeHtml>(OS, SubtargetInfo_->getSchedModel() 571 .getProcResource(Pressure.first) 572 ->Name); 573 OS << "</span>: "; 574 writeMeasurementValue<kEscapeHtml>(OS, Pressure.second); 575 OS << "</li>"; 576 } 577 OS << "</ul></td>"; 578 OS << "</tr>"; 579 } else { 580 OS << "<tr><td>✕</td><td></td><td></td></tr>"; 581 } 582 OS << "</table>"; 583 } 584 585 static constexpr const char kHtmlHead[] = R"( 586 <head> 587 <title>llvm-exegesis Analysis Results</title> 588 <style> 589 body { 590 font-family: sans-serif 591 } 592 span.sched-class-name { 593 font-weight: bold; 594 font-family: monospace; 595 } 596 span.opcode { 597 font-family: monospace; 598 } 599 span.config { 600 font-family: monospace; 601 } 602 div.inconsistency { 603 margin-top: 50px; 604 } 605 table { 606 margin-left: 50px; 607 border-collapse: collapse; 608 } 609 table, table tr,td,th { 610 border: 1px solid #444; 611 } 612 table ul { 613 padding-left: 0px; 614 margin: 0px; 615 list-style-type: none; 616 } 617 table.sched-class-clusters td { 618 padding-left: 10px; 619 padding-right: 10px; 620 padding-top: 10px; 621 padding-bottom: 10px; 622 } 623 table.sched-class-desc td { 624 padding-left: 10px; 625 padding-right: 10px; 626 padding-top: 2px; 627 padding-bottom: 2px; 628 } 629 span.mono { 630 font-family: monospace; 631 } 632 td.measurement { 633 text-align: center; 634 } 635 tr.good-cluster td.measurement { 636 color: #292 637 } 638 tr.bad-cluster td.measurement { 639 color: #922 640 } 641 tr.good-cluster td.measurement span.minmax { 642 color: #888; 643 } 644 tr.bad-cluster td.measurement span.minmax { 645 color: #888; 646 } 647 </style> 648 </head> 649 )"; 650 651 template <> 652 llvm::Error Analysis::run<Analysis::PrintSchedClassInconsistencies>( 653 llvm::raw_ostream &OS) const { 654 const auto &FirstPoint = Clustering_.getPoints()[0]; 655 // Print the header. 656 OS << "<!DOCTYPE html><html>" << kHtmlHead << "<body>"; 657 OS << "<h1><span class=\"mono\">llvm-exegesis</span> Analysis Results</h1>"; 658 OS << "<h3>Triple: <span class=\"mono\">"; 659 writeEscaped<kEscapeHtml>(OS, FirstPoint.LLVMTriple); 660 OS << "</span></h3><h3>Cpu: <span class=\"mono\">"; 661 writeEscaped<kEscapeHtml>(OS, FirstPoint.CpuName); 662 OS << "</span></h3>"; 663 664 for (const auto &RSCAndPoints : makePointsPerSchedClass()) { 665 if (!RSCAndPoints.RSC.SCDesc) 666 continue; 667 // Bucket sched class points into sched class clusters. 668 std::vector<SchedClassCluster> SchedClassClusters; 669 for (const size_t PointId : RSCAndPoints.PointIds) { 670 const auto &ClusterId = Clustering_.getClusterIdForPoint(PointId); 671 if (!ClusterId.isValid()) 672 continue; // Ignore noise and errors. FIXME: take noise into account ? 673 if (ClusterId.isUnstable() ^ AnalysisDisplayUnstableOpcodes_) 674 continue; // Either display stable or unstable clusters only. 675 auto SchedClassClusterIt = 676 std::find_if(SchedClassClusters.begin(), SchedClassClusters.end(), 677 [ClusterId](const SchedClassCluster &C) { 678 return C.id() == ClusterId; 679 }); 680 if (SchedClassClusterIt == SchedClassClusters.end()) { 681 SchedClassClusters.emplace_back(); 682 SchedClassClusterIt = std::prev(SchedClassClusters.end()); 683 } 684 SchedClassClusterIt->addPoint(PointId, Clustering_); 685 } 686 687 // Print any scheduling class that has at least one cluster that does not 688 // match the checked-in data. 689 if (llvm::all_of(SchedClassClusters, 690 [this, &RSCAndPoints](const SchedClassCluster &C) { 691 return C.measurementsMatch( 692 *SubtargetInfo_, RSCAndPoints.RSC, Clustering_); 693 })) 694 continue; // Nothing weird. 695 696 OS << "<div class=\"inconsistency\"><p>Sched Class <span " 697 "class=\"sched-class-name\">"; 698 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 699 writeEscaped<kEscapeHtml>(OS, RSCAndPoints.RSC.SCDesc->Name); 700 #else 701 OS << RSCAndPoints.RSC.SchedClassId; 702 #endif 703 OS << "</span> contains instructions whose performance characteristics do" 704 " not match that of LLVM:</p>"; 705 printSchedClassClustersHtml(SchedClassClusters, RSCAndPoints.RSC, OS); 706 OS << "<p>llvm SchedModel data:</p>"; 707 printSchedClassDescHtml(RSCAndPoints.RSC, OS); 708 OS << "</div>"; 709 } 710 711 OS << "</body></html>"; 712 return llvm::Error::success(); 713 } 714 715 // Distributes a pressure budget as evenly as possible on the provided subunits 716 // given the already existing port pressure distribution. 717 // 718 // The algorithm is as follows: while there is remaining pressure to 719 // distribute, find the subunits with minimal pressure, and distribute 720 // remaining pressure equally up to the pressure of the unit with 721 // second-to-minimal pressure. 722 // For example, let's assume we want to distribute 2*P1256 723 // (Subunits = [P1,P2,P5,P6]), and the starting DensePressure is: 724 // DensePressure = P0 P1 P2 P3 P4 P5 P6 P7 725 // 0.1 0.3 0.2 0.0 0.0 0.5 0.5 0.5 726 // RemainingPressure = 2.0 727 // We sort the subunits by pressure: 728 // Subunits = [(P2,p=0.2), (P1,p=0.3), (P5,p=0.5), (P6, p=0.5)] 729 // We'll first start by the subunits with minimal pressure, which are at 730 // the beginning of the sorted array. In this example there is one (P2). 731 // The subunit with second-to-minimal pressure is the next one in the 732 // array (P1). So we distribute 0.1 pressure to P2, and remove 0.1 cycles 733 // from the budget. 734 // Subunits = [(P2,p=0.3), (P1,p=0.3), (P5,p=0.5), (P5,p=0.5)] 735 // RemainingPressure = 1.9 736 // We repeat this process: distribute 0.2 pressure on each of the minimal 737 // P2 and P1, decrease budget by 2*0.2: 738 // Subunits = [(P2,p=0.5), (P1,p=0.5), (P5,p=0.5), (P5,p=0.5)] 739 // RemainingPressure = 1.5 740 // There are no second-to-minimal subunits so we just share the remaining 741 // budget (1.5 cycles) equally: 742 // Subunits = [(P2,p=0.875), (P1,p=0.875), (P5,p=0.875), (P5,p=0.875)] 743 // RemainingPressure = 0.0 744 // We stop as there is no remaining budget to distribute. 745 void distributePressure(float RemainingPressure, 746 llvm::SmallVector<uint16_t, 32> Subunits, 747 llvm::SmallVector<float, 32> &DensePressure) { 748 // Find the number of subunits with minimal pressure (they are at the 749 // front). 750 llvm::sort(Subunits, [&DensePressure](const uint16_t A, const uint16_t B) { 751 return DensePressure[A] < DensePressure[B]; 752 }); 753 const auto getPressureForSubunit = [&DensePressure, 754 &Subunits](size_t I) -> float & { 755 return DensePressure[Subunits[I]]; 756 }; 757 size_t NumMinimalSU = 1; 758 while (NumMinimalSU < Subunits.size() && 759 getPressureForSubunit(NumMinimalSU) == getPressureForSubunit(0)) { 760 ++NumMinimalSU; 761 } 762 while (RemainingPressure > 0.0f) { 763 if (NumMinimalSU == Subunits.size()) { 764 // All units are minimal, just distribute evenly and be done. 765 for (size_t I = 0; I < NumMinimalSU; ++I) { 766 getPressureForSubunit(I) += RemainingPressure / NumMinimalSU; 767 } 768 return; 769 } 770 // Distribute the remaining pressure equally. 771 const float MinimalPressure = getPressureForSubunit(NumMinimalSU - 1); 772 const float SecondToMinimalPressure = getPressureForSubunit(NumMinimalSU); 773 assert(MinimalPressure < SecondToMinimalPressure); 774 const float Increment = SecondToMinimalPressure - MinimalPressure; 775 if (RemainingPressure <= NumMinimalSU * Increment) { 776 // There is not enough remaining pressure. 777 for (size_t I = 0; I < NumMinimalSU; ++I) { 778 getPressureForSubunit(I) += RemainingPressure / NumMinimalSU; 779 } 780 return; 781 } 782 // Bump all minimal pressure subunits to `SecondToMinimalPressure`. 783 for (size_t I = 0; I < NumMinimalSU; ++I) { 784 getPressureForSubunit(I) = SecondToMinimalPressure; 785 RemainingPressure -= SecondToMinimalPressure; 786 } 787 while (NumMinimalSU < Subunits.size() && 788 getPressureForSubunit(NumMinimalSU) == SecondToMinimalPressure) { 789 ++NumMinimalSU; 790 } 791 } 792 } 793 794 std::vector<std::pair<uint16_t, float>> computeIdealizedProcResPressure( 795 const llvm::MCSchedModel &SM, 796 llvm::SmallVector<llvm::MCWriteProcResEntry, 8> WPRS) { 797 // DensePressure[I] is the port pressure for Proc Resource I. 798 llvm::SmallVector<float, 32> DensePressure(SM.getNumProcResourceKinds()); 799 llvm::sort(WPRS, [](const llvm::MCWriteProcResEntry &A, 800 const llvm::MCWriteProcResEntry &B) { 801 return A.ProcResourceIdx < B.ProcResourceIdx; 802 }); 803 for (const llvm::MCWriteProcResEntry &WPR : WPRS) { 804 // Get units for the entry. 805 const llvm::MCProcResourceDesc *const ProcResDesc = 806 SM.getProcResource(WPR.ProcResourceIdx); 807 if (ProcResDesc->SubUnitsIdxBegin == nullptr) { 808 // This is a ProcResUnit. 809 DensePressure[WPR.ProcResourceIdx] += WPR.Cycles; 810 } else { 811 // This is a ProcResGroup. 812 llvm::SmallVector<uint16_t, 32> Subunits(ProcResDesc->SubUnitsIdxBegin, 813 ProcResDesc->SubUnitsIdxBegin + 814 ProcResDesc->NumUnits); 815 distributePressure(WPR.Cycles, Subunits, DensePressure); 816 } 817 } 818 // Turn dense pressure into sparse pressure by removing zero entries. 819 std::vector<std::pair<uint16_t, float>> Pressure; 820 for (unsigned I = 0, E = SM.getNumProcResourceKinds(); I < E; ++I) { 821 if (DensePressure[I] > 0.0f) 822 Pressure.emplace_back(I, DensePressure[I]); 823 } 824 return Pressure; 825 } 826 827 } // namespace exegesis 828 } // namespace llvm 829