| 1 | // Licensed to the .NET Foundation under one or more agreements. |
| 2 | // The .NET Foundation licenses this file to you under the MIT license. |
| 3 | // See the LICENSE file in the project root for more information. |
| 4 | // =========================================================================== |
| 5 | // File: TieredCompilation.CPP |
| 6 | // |
| 7 | // =========================================================================== |
| 8 | |
| 9 | |
| 10 | |
| 11 | #include "common.h" |
| 12 | #include "excep.h" |
| 13 | #include "log.h" |
| 14 | #include "win32threadpool.h" |
| 15 | #include "threadsuspend.h" |
| 16 | #include "tieredcompilation.h" |
| 17 | |
| 18 | // TieredCompilationManager determines which methods should be recompiled and |
| 19 | // how they should be recompiled to best optimize the running code. It then |
| 20 | // handles logistics of getting new code created and installed. |
| 21 | // |
| 22 | // |
| 23 | // # Important entrypoints in this code: |
| 24 | // |
| 25 | // |
| 26 | // a) .ctor and Init(...) - called once during AppDomain initialization |
| 27 | // b) OnMethodCalled(...) - called when a method is being invoked. When a method |
| 28 | // has been called enough times this is currently the only |
| 29 | // trigger that initiates re-compilation. |
| 30 | // c) Shutdown() - called during AppDomain::Exit() to begin the process |
| 31 | // of stopping tiered compilation. After this point no more |
| 32 | // background optimization work will be initiated but in-progress |
| 33 | // work still needs to complete. |
| 34 | // d) ShutdownAllDomains() - Called from EEShutdownHelper to block until all async work is |
| 35 | // complete. We must do this before we shutdown the JIT. |
| 36 | // |
| 37 | // # Overall workflow |
| 38 | // |
| 39 | // Methods initially call into OnMethodCalled() and once the call count exceeds |
| 40 | // a fixed limit we queue work on to our internal list of methods needing to |
| 41 | // be recompiled (m_methodsToOptimize). If there is currently no thread |
| 42 | // servicing our queue asynchronously then we use the runtime threadpool |
| 43 | // QueueUserWorkItem to recruit one. During the callback for each threadpool work |
| 44 | // item we handle as many methods as possible in a fixed period of time, then |
| 45 | // queue another threadpool work item if m_methodsToOptimize hasn't been drained. |
| 46 | // |
| 47 | // The background thread enters at StaticOptimizeMethodsCallback(), enters the |
| 48 | // appdomain, and then begins calling OptimizeMethod on each method in the |
| 49 | // queue. For each method we jit it, then update the precode so that future |
| 50 | // entrypoint callers will run the new code. |
| 51 | // |
| 52 | // # Error handling |
| 53 | // |
| 54 | // The overall principle is don't swallow terminal failures that may have corrupted the |
| 55 | // process (AV for example), but otherwise for any transient issue or functional limitation |
| 56 | // that prevents us from optimizing log it for diagnostics and then back out gracefully, |
| 57 | // continuing to run the less optimal code. The feature should be constructed so that |
| 58 | // errors are limited to OS resource exhaustion or poorly behaved managed code |
| 59 | // (for example within an AssemblyResolve event or static constructor triggered by the JIT). |
| 60 | |
| 61 | #if defined(FEATURE_TIERED_COMPILATION) && !defined(DACCESS_COMPILE) |
| 62 | |
| 63 | // Called at AppDomain construction |
| 64 | TieredCompilationManager::TieredCompilationManager() : |
| 65 | m_lock(CrstTieredCompilation), |
| 66 | m_isAppDomainShuttingDown(FALSE), |
| 67 | m_countOptimizationThreadsRunning(0), |
| 68 | m_callCountOptimizationThreshhold(1), |
| 69 | m_optimizationQuantumMs(50), |
| 70 | m_methodsPendingCountingForTier1(nullptr), |
| 71 | m_tieringDelayTimerHandle(nullptr), |
| 72 | m_tier1CallCountingCandidateMethodRecentlyRecorded(false) |
| 73 | { |
| 74 | WRAPPER_NO_CONTRACT; |
| 75 | // On Unix, we can reach here before EEConfig is initialized, so defer config-based initialization to Init() |
| 76 | } |
| 77 | |
| 78 | // Called at AppDomain Init |
| 79 | void TieredCompilationManager::Init(ADID appDomainId) |
| 80 | { |
| 81 | CONTRACTL |
| 82 | { |
| 83 | GC_NOTRIGGER; |
| 84 | CAN_TAKE_LOCK; |
| 85 | MODE_PREEMPTIVE; |
| 86 | } |
| 87 | CONTRACTL_END; |
| 88 | |
| 89 | CrstHolder holder(&m_lock); |
| 90 | m_domainId = appDomainId; |
| 91 | m_callCountOptimizationThreshhold = g_pConfig->TieredCompilation_Tier1CallCountThreshold(); |
| 92 | } |
| 93 | |
| 94 | #endif // FEATURE_TIERED_COMPILATION && !DACCESS_COMPILE |
| 95 | |
| 96 | NativeCodeVersion::OptimizationTier TieredCompilationManager::GetInitialOptimizationTier(PTR_MethodDesc pMethodDesc) |
| 97 | { |
| 98 | WRAPPER_NO_CONTRACT; |
| 99 | _ASSERTE(pMethodDesc != NULL); |
| 100 | |
| 101 | #ifdef FEATURE_TIERED_COMPILATION |
| 102 | if (pMethodDesc->RequestedAggressiveOptimization()) |
| 103 | { |
| 104 | // Methods flagged with MethodImplOptions.AggressiveOptimization begin at tier 1, as a workaround to cold methods with |
| 105 | // hot loops performing poorly (https://github.com/dotnet/coreclr/issues/19751) |
| 106 | return NativeCodeVersion::OptimizationTier1; |
| 107 | } |
| 108 | #endif // FEATURE_TIERED_COMPILATION |
| 109 | |
| 110 | return NativeCodeVersion::OptimizationTier0; |
| 111 | } |
| 112 | |
| 113 | #if defined(FEATURE_TIERED_COMPILATION) && !defined(DACCESS_COMPILE) |
| 114 | |
| 115 | bool TieredCompilationManager::RequiresCallCounting(MethodDesc* pMethodDesc) |
| 116 | { |
| 117 | WRAPPER_NO_CONTRACT; |
| 118 | _ASSERTE(pMethodDesc != NULL); |
| 119 | _ASSERTE(pMethodDesc->IsEligibleForTieredCompilation()); |
| 120 | |
| 121 | return |
| 122 | g_pConfig->TieredCompilation_CallCounting() && |
| 123 | GetInitialOptimizationTier(pMethodDesc) == NativeCodeVersion::OptimizationTier0; |
| 124 | } |
| 125 | |
| 126 | // Called each time code in this AppDomain has been run. This is our sole entrypoint to begin |
| 127 | // tiered compilation for now. Returns TRUE if no more notifications are necessary, but |
| 128 | // more notifications may come anyways. |
| 129 | // |
| 130 | // currentCallCount is pre-incremented, that is to say the value is 1 on first call for a given |
| 131 | // method. |
| 132 | void TieredCompilationManager::OnMethodCalled( |
| 133 | MethodDesc* pMethodDesc, |
| 134 | DWORD currentCallCount, |
| 135 | BOOL* shouldStopCountingCallsRef, |
| 136 | BOOL* wasPromotedToTier1Ref) |
| 137 | { |
| 138 | WRAPPER_NO_CONTRACT; |
| 139 | _ASSERTE(pMethodDesc->IsEligibleForTieredCompilation()); |
| 140 | _ASSERTE(shouldStopCountingCallsRef != nullptr); |
| 141 | _ASSERTE(wasPromotedToTier1Ref != nullptr); |
| 142 | |
| 143 | *shouldStopCountingCallsRef = |
| 144 | // Stop call counting when the delay is in effect |
| 145 | IsTieringDelayActive() || |
| 146 | // Initiate the delay on tier 0 activity (when a new eligible method is called the first time) |
| 147 | (currentCallCount == 1 && g_pConfig->TieredCompilation_Tier1CallCountingDelayMs() != 0) || |
| 148 | // Stop call counting when ready for tier 1 promotion |
| 149 | currentCallCount >= m_callCountOptimizationThreshhold; |
| 150 | |
| 151 | *wasPromotedToTier1Ref = currentCallCount >= m_callCountOptimizationThreshhold; |
| 152 | |
| 153 | if (currentCallCount == m_callCountOptimizationThreshhold) |
| 154 | { |
| 155 | AsyncPromoteMethodToTier1(pMethodDesc); |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | void TieredCompilationManager::OnMethodCallCountingStoppedWithoutTier1Promotion(MethodDesc* pMethodDesc) |
| 160 | { |
| 161 | WRAPPER_NO_CONTRACT; |
| 162 | _ASSERTE(pMethodDesc != nullptr); |
| 163 | _ASSERTE(pMethodDesc->IsEligibleForTieredCompilation()); |
| 164 | |
| 165 | if (g_pConfig->TieredCompilation_Tier1CallCountingDelayMs() == 0) |
| 166 | { |
| 167 | return; |
| 168 | } |
| 169 | |
| 170 | while (true) |
| 171 | { |
| 172 | bool attemptedToInitiateDelay = false; |
| 173 | if (!IsTieringDelayActive()) |
| 174 | { |
| 175 | if (!TryInitiateTieringDelay()) |
| 176 | { |
| 177 | break; |
| 178 | } |
| 179 | attemptedToInitiateDelay = true; |
| 180 | } |
| 181 | |
| 182 | { |
| 183 | CrstHolder holder(&m_lock); |
| 184 | |
| 185 | SArray<MethodDesc*>* methodsPendingCountingForTier1 = m_methodsPendingCountingForTier1; |
| 186 | if (methodsPendingCountingForTier1 == nullptr) |
| 187 | { |
| 188 | // Timer tick callback race, try again |
| 189 | continue; |
| 190 | } |
| 191 | |
| 192 | // Record the method to resume counting later (see Tier1DelayTimerCallback) |
| 193 | bool success = false; |
| 194 | EX_TRY |
| 195 | { |
| 196 | methodsPendingCountingForTier1->Append(pMethodDesc); |
| 197 | success = true; |
| 198 | } |
| 199 | EX_CATCH |
| 200 | { |
| 201 | } |
| 202 | EX_END_CATCH(RethrowTerminalExceptions); |
| 203 | if (!success) |
| 204 | { |
| 205 | break; |
| 206 | } |
| 207 | |
| 208 | if (!attemptedToInitiateDelay) |
| 209 | { |
| 210 | // Delay call counting for currently recoded methods further |
| 211 | m_tier1CallCountingCandidateMethodRecentlyRecorded = true; |
| 212 | } |
| 213 | } |
| 214 | return; |
| 215 | } |
| 216 | |
| 217 | ResumeCountingCalls(pMethodDesc); |
| 218 | } |
| 219 | |
| 220 | void TieredCompilationManager::AsyncPromoteMethodToTier1(MethodDesc* pMethodDesc) |
| 221 | { |
| 222 | STANDARD_VM_CONTRACT; |
| 223 | |
| 224 | NativeCodeVersion t1NativeCodeVersion; |
| 225 | |
| 226 | // Add an inactive native code entry in the versioning table to track the tier1 |
| 227 | // compilation we are going to create. This entry binds the compilation to a |
| 228 | // particular version of the IL code regardless of any changes that may |
| 229 | // occur between now and when jitting completes. If the IL does change in that |
| 230 | // interval the new code entry won't be activated. |
| 231 | { |
| 232 | CodeVersionManager* pCodeVersionManager = pMethodDesc->GetCodeVersionManager(); |
| 233 | CodeVersionManager::TableLockHolder lock(pCodeVersionManager); |
| 234 | ILCodeVersion ilVersion = pCodeVersionManager->GetActiveILCodeVersion(pMethodDesc); |
| 235 | NativeCodeVersionCollection nativeVersions = ilVersion.GetNativeCodeVersions(pMethodDesc); |
| 236 | for (NativeCodeVersionIterator cur = nativeVersions.Begin(), end = nativeVersions.End(); cur != end; cur++) |
| 237 | { |
| 238 | if (cur->GetOptimizationTier() == NativeCodeVersion::OptimizationTier1) |
| 239 | { |
| 240 | // we've already promoted |
| 241 | LOG((LF_TIEREDCOMPILATION, LL_INFO100000, "TieredCompilationManager::AsyncPromoteMethodToTier1 Method=0x%pM (%s::%s) ignoring already promoted method\n" , |
| 242 | pMethodDesc, pMethodDesc->m_pszDebugClassName, pMethodDesc->m_pszDebugMethodName)); |
| 243 | return; |
| 244 | } |
| 245 | } |
| 246 | |
| 247 | HRESULT hr = S_OK; |
| 248 | if (FAILED(hr = ilVersion.AddNativeCodeVersion(pMethodDesc, NativeCodeVersion::OptimizationTier1, &t1NativeCodeVersion))) |
| 249 | { |
| 250 | // optimization didn't work for some reason (presumably OOM) |
| 251 | // just give up and continue on |
| 252 | STRESS_LOG2(LF_TIEREDCOMPILATION, LL_WARNING, "TieredCompilationManager::AsyncPromoteMethodToTier1: " |
| 253 | "AddNativeCodeVersion failed hr=0x%x, method=%pM\n" , |
| 254 | hr, pMethodDesc); |
| 255 | return; |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | // Insert the method into the optimization queue and trigger a thread to service |
| 260 | // the queue if needed. |
| 261 | // |
| 262 | // Note an error here could affect concurrent threads running this |
| 263 | // code. Those threads will observe m_countOptimizationThreadsRunning > 0 and return, |
| 264 | // then QueueUserWorkItem fails on this thread lowering the count and leaves them |
| 265 | // unserviced. Synchronous retries appear unlikely to offer any material improvement |
| 266 | // and complicating the code to narrow an already rare error case isn't desirable. |
| 267 | { |
| 268 | SListElem<NativeCodeVersion>* pMethodListItem = new (nothrow) SListElem<NativeCodeVersion>(t1NativeCodeVersion); |
| 269 | CrstHolder holder(&m_lock); |
| 270 | if (pMethodListItem != NULL) |
| 271 | { |
| 272 | m_methodsToOptimize.InsertTail(pMethodListItem); |
| 273 | } |
| 274 | |
| 275 | LOG((LF_TIEREDCOMPILATION, LL_INFO10000, "TieredCompilationManager::AsyncPromoteMethodToTier1 Method=0x%pM (%s::%s), code version id=0x%x queued\n" , |
| 276 | pMethodDesc, pMethodDesc->m_pszDebugClassName, pMethodDesc->m_pszDebugMethodName, |
| 277 | t1NativeCodeVersion.GetVersionId())); |
| 278 | |
| 279 | if (!IncrementWorkerThreadCountIfNeeded()) |
| 280 | { |
| 281 | return; |
| 282 | } |
| 283 | } |
| 284 | |
| 285 | if (!TryAsyncOptimizeMethods()) |
| 286 | { |
| 287 | CrstHolder holder(&m_lock); |
| 288 | DecrementWorkerThreadCount(); |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | void TieredCompilationManager::Shutdown() |
| 293 | { |
| 294 | STANDARD_VM_CONTRACT; |
| 295 | |
| 296 | CrstHolder holder(&m_lock); |
| 297 | m_isAppDomainShuttingDown = TRUE; |
| 298 | } |
| 299 | |
| 300 | bool TieredCompilationManager::IsTieringDelayActive() |
| 301 | { |
| 302 | LIMITED_METHOD_CONTRACT; |
| 303 | return m_methodsPendingCountingForTier1 != nullptr; |
| 304 | } |
| 305 | |
| 306 | bool TieredCompilationManager::TryInitiateTieringDelay() |
| 307 | { |
| 308 | WRAPPER_NO_CONTRACT; |
| 309 | _ASSERTE(g_pConfig->TieredCompilation()); |
| 310 | _ASSERTE(g_pConfig->TieredCompilation_Tier1CallCountingDelayMs() != 0); |
| 311 | |
| 312 | NewHolder<SArray<MethodDesc*>> methodsPendingCountingHolder = new(nothrow) SArray<MethodDesc*>(); |
| 313 | if (methodsPendingCountingHolder == nullptr) |
| 314 | { |
| 315 | return false; |
| 316 | } |
| 317 | |
| 318 | bool success = false; |
| 319 | EX_TRY |
| 320 | { |
| 321 | methodsPendingCountingHolder->Preallocate(64); |
| 322 | success = true; |
| 323 | } |
| 324 | EX_CATCH |
| 325 | { |
| 326 | } |
| 327 | EX_END_CATCH(RethrowTerminalExceptions); |
| 328 | if (!success) |
| 329 | { |
| 330 | return false; |
| 331 | } |
| 332 | |
| 333 | NewHolder<ThreadpoolMgr::TimerInfoContext> timerContextHolder = new(nothrow) ThreadpoolMgr::TimerInfoContext(); |
| 334 | if (timerContextHolder == nullptr) |
| 335 | { |
| 336 | return false; |
| 337 | } |
| 338 | timerContextHolder->AppDomainId = m_domainId; |
| 339 | timerContextHolder->TimerId = 0; |
| 340 | |
| 341 | { |
| 342 | CrstHolder holder(&m_lock); |
| 343 | |
| 344 | if (IsTieringDelayActive()) |
| 345 | { |
| 346 | return true; |
| 347 | } |
| 348 | |
| 349 | // The timer is created inside the lock to avoid some unnecessary additional complexity that would otherwise arise from |
| 350 | // there being a failure point after the timer is successfully created. For instance, if the timer is created outside |
| 351 | // the lock and then inside the lock it is found that another thread beat us to it, there would be two active timers |
| 352 | // that may tick before the extra timer is deleted, along with additional concurrency issues. |
| 353 | _ASSERTE(m_tieringDelayTimerHandle == nullptr); |
| 354 | success = false; |
| 355 | EX_TRY |
| 356 | { |
| 357 | if (ThreadpoolMgr::CreateTimerQueueTimer( |
| 358 | &m_tieringDelayTimerHandle, |
| 359 | TieringDelayTimerCallback, |
| 360 | timerContextHolder, |
| 361 | g_pConfig->TieredCompilation_Tier1CallCountingDelayMs(), |
| 362 | (DWORD)-1 /* Period, non-repeating */, |
| 363 | 0 /* flags */)) |
| 364 | { |
| 365 | success = true; |
| 366 | } |
| 367 | } |
| 368 | EX_CATCH |
| 369 | { |
| 370 | } |
| 371 | EX_END_CATCH(RethrowTerminalExceptions); |
| 372 | if (!success) |
| 373 | { |
| 374 | _ASSERTE(m_tieringDelayTimerHandle == nullptr); |
| 375 | return false; |
| 376 | } |
| 377 | |
| 378 | m_methodsPendingCountingForTier1 = methodsPendingCountingHolder.Extract(); |
| 379 | _ASSERTE(IsTieringDelayActive()); |
| 380 | } |
| 381 | |
| 382 | timerContextHolder.SuppressRelease(); // the timer context is automatically deleted by the timer infrastructure |
| 383 | return true; |
| 384 | } |
| 385 | |
| 386 | void WINAPI TieredCompilationManager::TieringDelayTimerCallback(PVOID parameter, BOOLEAN timerFired) |
| 387 | { |
| 388 | WRAPPER_NO_CONTRACT; |
| 389 | _ASSERTE(timerFired); |
| 390 | |
| 391 | ThreadpoolMgr::TimerInfoContext* timerContext = (ThreadpoolMgr::TimerInfoContext*)parameter; |
| 392 | EX_TRY |
| 393 | { |
| 394 | GCX_COOP(); |
| 395 | ManagedThreadBase::ThreadPool(timerContext->AppDomainId, TieringDelayTimerCallbackInAppDomain, nullptr); |
| 396 | } |
| 397 | EX_CATCH |
| 398 | { |
| 399 | STRESS_LOG1(LF_TIEREDCOMPILATION, LL_ERROR, "TieredCompilationManager::Tier1DelayTimerCallback: " |
| 400 | "Unhandled exception, hr=0x%x\n" , |
| 401 | GET_EXCEPTION()->GetHR()); |
| 402 | } |
| 403 | EX_END_CATCH(RethrowTerminalExceptions); |
| 404 | } |
| 405 | |
| 406 | void TieredCompilationManager::TieringDelayTimerCallbackInAppDomain(LPVOID parameter) |
| 407 | { |
| 408 | WRAPPER_NO_CONTRACT; |
| 409 | GetAppDomain()->GetTieredCompilationManager()->TieringDelayTimerCallbackWorker(); |
| 410 | } |
| 411 | |
| 412 | void TieredCompilationManager::TieringDelayTimerCallbackWorker() |
| 413 | { |
| 414 | WRAPPER_NO_CONTRACT; |
| 415 | _ASSERTE(GetAppDomain()->GetId() == m_domainId); |
| 416 | |
| 417 | HANDLE tieringDelayTimerHandle; |
| 418 | bool tier1CallCountingCandidateMethodRecentlyRecorded; |
| 419 | { |
| 420 | // It's possible for the timer to tick before it is recorded that the delay is in effect. This lock guarantees that the |
| 421 | // delay is in effect. |
| 422 | CrstHolder holder(&m_lock); |
| 423 | _ASSERTE(IsTieringDelayActive()); |
| 424 | |
| 425 | tieringDelayTimerHandle = m_tieringDelayTimerHandle; |
| 426 | _ASSERTE(tieringDelayTimerHandle != nullptr); |
| 427 | |
| 428 | tier1CallCountingCandidateMethodRecentlyRecorded = m_tier1CallCountingCandidateMethodRecentlyRecorded; |
| 429 | if (tier1CallCountingCandidateMethodRecentlyRecorded) |
| 430 | { |
| 431 | m_tier1CallCountingCandidateMethodRecentlyRecorded = false; |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | // Reschedule the timer if there has been recent tier 0 activity (when a new eligible method is called the first time) to |
| 436 | // further delay call counting |
| 437 | if (tier1CallCountingCandidateMethodRecentlyRecorded) |
| 438 | { |
| 439 | bool success = false; |
| 440 | EX_TRY |
| 441 | { |
| 442 | if (ThreadpoolMgr::ChangeTimerQueueTimer( |
| 443 | tieringDelayTimerHandle, |
| 444 | g_pConfig->TieredCompilation_Tier1CallCountingDelayMs(), |
| 445 | (DWORD)-1 /* Period, non-repeating */)) |
| 446 | { |
| 447 | success = true; |
| 448 | } |
| 449 | } |
| 450 | EX_CATCH |
| 451 | { |
| 452 | } |
| 453 | EX_END_CATCH(RethrowTerminalExceptions); |
| 454 | if (success) |
| 455 | { |
| 456 | return; |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | // Exchange information into locals inside the lock |
| 461 | SArray<MethodDesc*>* methodsPendingCountingForTier1; |
| 462 | bool optimizeMethods; |
| 463 | { |
| 464 | CrstHolder holder(&m_lock); |
| 465 | |
| 466 | methodsPendingCountingForTier1 = m_methodsPendingCountingForTier1; |
| 467 | _ASSERTE(methodsPendingCountingForTier1 != nullptr); |
| 468 | m_methodsPendingCountingForTier1 = nullptr; |
| 469 | |
| 470 | _ASSERTE(tieringDelayTimerHandle == m_tieringDelayTimerHandle); |
| 471 | m_tieringDelayTimerHandle = nullptr; |
| 472 | |
| 473 | _ASSERTE(!IsTieringDelayActive()); |
| 474 | optimizeMethods = IncrementWorkerThreadCountIfNeeded(); |
| 475 | } |
| 476 | |
| 477 | // Install call counters |
| 478 | MethodDesc** methods = methodsPendingCountingForTier1->GetElements(); |
| 479 | COUNT_T methodCount = methodsPendingCountingForTier1->GetCount(); |
| 480 | for (COUNT_T i = 0; i < methodCount; ++i) |
| 481 | { |
| 482 | ResumeCountingCalls(methods[i]); |
| 483 | } |
| 484 | delete methodsPendingCountingForTier1; |
| 485 | |
| 486 | ThreadpoolMgr::DeleteTimerQueueTimer(tieringDelayTimerHandle, nullptr); |
| 487 | |
| 488 | if (optimizeMethods) |
| 489 | { |
| 490 | OptimizeMethods(); |
| 491 | } |
| 492 | } |
| 493 | |
| 494 | void TieredCompilationManager::ResumeCountingCalls(MethodDesc* pMethodDesc) |
| 495 | { |
| 496 | WRAPPER_NO_CONTRACT; |
| 497 | _ASSERTE(pMethodDesc != nullptr); |
| 498 | _ASSERTE(pMethodDesc->IsVersionableWithPrecode()); |
| 499 | |
| 500 | pMethodDesc->GetPrecode()->ResetTargetInterlocked(); |
| 501 | } |
| 502 | |
| 503 | bool TieredCompilationManager::TryAsyncOptimizeMethods() |
| 504 | { |
| 505 | WRAPPER_NO_CONTRACT; |
| 506 | _ASSERTE(DebugGetWorkerThreadCount() != 0); |
| 507 | |
| 508 | // Terminal exceptions escape as exceptions, but all other errors should gracefully |
| 509 | // return to the caller. Non-terminal error conditions should be rare (ie OOM, |
| 510 | // OS failure to create thread) and we consider it reasonable for some methods |
| 511 | // to go unoptimized or have their optimization arbitrarily delayed under these |
| 512 | // circumstances. |
| 513 | bool success = false; |
| 514 | EX_TRY |
| 515 | { |
| 516 | if (ThreadpoolMgr::QueueUserWorkItem(StaticOptimizeMethodsCallback, this, QUEUE_ONLY, TRUE)) |
| 517 | { |
| 518 | success = true; |
| 519 | } |
| 520 | else |
| 521 | { |
| 522 | STRESS_LOG0(LF_TIEREDCOMPILATION, LL_WARNING, "TieredCompilationManager::OnMethodCalled: " |
| 523 | "ThreadpoolMgr::QueueUserWorkItem returned FALSE (no thread will run)\n" ); |
| 524 | } |
| 525 | } |
| 526 | EX_CATCH |
| 527 | { |
| 528 | STRESS_LOG1(LF_TIEREDCOMPILATION, LL_WARNING, "TieredCompilationManager::OnMethodCalled: " |
| 529 | "Exception queuing work item to threadpool, hr=0x%x\n" , |
| 530 | GET_EXCEPTION()->GetHR()); |
| 531 | } |
| 532 | EX_END_CATCH(RethrowTerminalExceptions); |
| 533 | return success; |
| 534 | } |
| 535 | |
| 536 | // This is the initial entrypoint for the background thread, called by |
| 537 | // the threadpool. |
| 538 | DWORD WINAPI TieredCompilationManager::StaticOptimizeMethodsCallback(void *args) |
| 539 | { |
| 540 | STANDARD_VM_CONTRACT; |
| 541 | |
| 542 | TieredCompilationManager * pTieredCompilationManager = (TieredCompilationManager *)args; |
| 543 | pTieredCompilationManager->OptimizeMethodsCallback(); |
| 544 | |
| 545 | return 0; |
| 546 | } |
| 547 | |
| 548 | void TieredCompilationManager::OptimizeMethodsCallback() |
| 549 | { |
| 550 | STANDARD_VM_CONTRACT; |
| 551 | _ASSERTE(DebugGetWorkerThreadCount() != 0); |
| 552 | |
| 553 | // This app domain shutdown check isn't required for correctness |
| 554 | // but it should reduce some unneeded exceptions trying |
| 555 | // to enter a closed AppDomain |
| 556 | { |
| 557 | CrstHolder holder(&m_lock); |
| 558 | if (m_isAppDomainShuttingDown) |
| 559 | { |
| 560 | DecrementWorkerThreadCount(); |
| 561 | return; |
| 562 | } |
| 563 | } |
| 564 | |
| 565 | EX_TRY |
| 566 | { |
| 567 | GCX_COOP(); |
| 568 | ENTER_DOMAIN_ID(m_domainId); |
| 569 | { |
| 570 | OptimizeMethods(); |
| 571 | } |
| 572 | END_DOMAIN_TRANSITION; |
| 573 | } |
| 574 | EX_CATCH |
| 575 | { |
| 576 | STRESS_LOG1(LF_TIEREDCOMPILATION, LL_ERROR, "TieredCompilationManager::OptimizeMethodsCallback: " |
| 577 | "Unhandled exception on domain transition, hr=0x%x\n" , |
| 578 | GET_EXCEPTION()->GetHR()); |
| 579 | } |
| 580 | EX_END_CATCH(RethrowTerminalExceptions); |
| 581 | } |
| 582 | |
| 583 | //This method will process one or more methods from optimization queue |
| 584 | // on a background thread. Each such method will be jitted with code |
| 585 | // optimizations enabled and then installed as the active implementation |
| 586 | // of the method entrypoint. |
| 587 | // |
| 588 | // We need to be carefuly not to work for too long in a single invocation |
| 589 | // of this method or we could starve the threadpool and force |
| 590 | // it to create unnecessary additional threads. |
| 591 | void TieredCompilationManager::OptimizeMethods() |
| 592 | { |
| 593 | WRAPPER_NO_CONTRACT; |
| 594 | _ASSERTE(DebugGetWorkerThreadCount() != 0); |
| 595 | _ASSERTE(GetAppDomain()->GetId() == m_domainId); |
| 596 | |
| 597 | ULONGLONG startTickCount = CLRGetTickCount64(); |
| 598 | NativeCodeVersion nativeCodeVersion; |
| 599 | EX_TRY |
| 600 | { |
| 601 | GCX_PREEMP(); |
| 602 | while (true) |
| 603 | { |
| 604 | { |
| 605 | CrstHolder holder(&m_lock); |
| 606 | |
| 607 | if (IsTieringDelayActive() || m_isAppDomainShuttingDown) |
| 608 | { |
| 609 | DecrementWorkerThreadCount(); |
| 610 | break; |
| 611 | } |
| 612 | |
| 613 | nativeCodeVersion = GetNextMethodToOptimize(); |
| 614 | if (nativeCodeVersion.IsNull()) |
| 615 | { |
| 616 | DecrementWorkerThreadCount(); |
| 617 | break; |
| 618 | } |
| 619 | } |
| 620 | OptimizeMethod(nativeCodeVersion); |
| 621 | |
| 622 | // If we have been running for too long return the thread to the threadpool and queue another event |
| 623 | // This gives the threadpool a chance to service other requests on this thread before returning to |
| 624 | // this work. |
| 625 | ULONGLONG currentTickCount = CLRGetTickCount64(); |
| 626 | if (currentTickCount >= startTickCount + m_optimizationQuantumMs) |
| 627 | { |
| 628 | if (!TryAsyncOptimizeMethods()) |
| 629 | { |
| 630 | CrstHolder holder(&m_lock); |
| 631 | DecrementWorkerThreadCount(); |
| 632 | } |
| 633 | break; |
| 634 | } |
| 635 | } |
| 636 | } |
| 637 | EX_CATCH |
| 638 | { |
| 639 | { |
| 640 | CrstHolder holder(&m_lock); |
| 641 | DecrementWorkerThreadCount(); |
| 642 | } |
| 643 | STRESS_LOG2(LF_TIEREDCOMPILATION, LL_ERROR, "TieredCompilationManager::OptimizeMethods: " |
| 644 | "Unhandled exception during method optimization, hr=0x%x, last method=%p\n" , |
| 645 | GET_EXCEPTION()->GetHR(), nativeCodeVersion.GetMethodDesc()); |
| 646 | } |
| 647 | EX_END_CATCH(RethrowTerminalExceptions); |
| 648 | } |
| 649 | |
| 650 | // Jit compiles and installs new optimized code for a method. |
| 651 | // Called on a background thread. |
| 652 | void TieredCompilationManager::OptimizeMethod(NativeCodeVersion nativeCodeVersion) |
| 653 | { |
| 654 | STANDARD_VM_CONTRACT; |
| 655 | |
| 656 | _ASSERTE(nativeCodeVersion.GetMethodDesc()->IsEligibleForTieredCompilation()); |
| 657 | if (CompileCodeVersion(nativeCodeVersion)) |
| 658 | { |
| 659 | ActivateCodeVersion(nativeCodeVersion); |
| 660 | } |
| 661 | } |
| 662 | |
| 663 | // Compiles new optimized code for a method. |
| 664 | // Called on a background thread. |
| 665 | BOOL TieredCompilationManager::CompileCodeVersion(NativeCodeVersion nativeCodeVersion) |
| 666 | { |
| 667 | STANDARD_VM_CONTRACT; |
| 668 | |
| 669 | PCODE pCode = NULL; |
| 670 | MethodDesc* pMethod = nativeCodeVersion.GetMethodDesc(); |
| 671 | EX_TRY |
| 672 | { |
| 673 | pCode = pMethod->PrepareCode(nativeCodeVersion); |
| 674 | LOG((LF_TIEREDCOMPILATION, LL_INFO10000, "TieredCompilationManager::CompileCodeVersion Method=0x%pM (%s::%s), code version id=0x%x, code ptr=0x%p\n" , |
| 675 | pMethod, pMethod->m_pszDebugClassName, pMethod->m_pszDebugMethodName, |
| 676 | nativeCodeVersion.GetVersionId(), |
| 677 | pCode)); |
| 678 | } |
| 679 | EX_CATCH |
| 680 | { |
| 681 | // Failing to jit should be rare but acceptable. We will leave whatever code already exists in place. |
| 682 | STRESS_LOG2(LF_TIEREDCOMPILATION, LL_INFO10, "TieredCompilationManager::CompileCodeVersion: Method %pM failed to jit, hr=0x%x\n" , |
| 683 | pMethod, GET_EXCEPTION()->GetHR()); |
| 684 | } |
| 685 | EX_END_CATCH(RethrowTerminalExceptions) |
| 686 | |
| 687 | return pCode != NULL; |
| 688 | } |
| 689 | |
| 690 | // Updates the MethodDesc and precode so that future invocations of a method will |
| 691 | // execute the native code pointed to by pCode. |
| 692 | // Called on a background thread. |
| 693 | void TieredCompilationManager::ActivateCodeVersion(NativeCodeVersion nativeCodeVersion) |
| 694 | { |
| 695 | STANDARD_VM_CONTRACT; |
| 696 | |
| 697 | MethodDesc* pMethod = nativeCodeVersion.GetMethodDesc(); |
| 698 | CodeVersionManager* pCodeVersionManager = pMethod->GetCodeVersionManager(); |
| 699 | |
| 700 | // If the ilParent version is active this will activate the native code version now. |
| 701 | // Otherwise if the ilParent version becomes active again in the future the native |
| 702 | // code version will activate then. |
| 703 | ILCodeVersion ilParent; |
| 704 | HRESULT hr = S_OK; |
| 705 | { |
| 706 | // As long as we are exclusively using precode publishing for tiered compilation |
| 707 | // methods this first attempt should succeed |
| 708 | CodeVersionManager::TableLockHolder lock(pCodeVersionManager); |
| 709 | ilParent = nativeCodeVersion.GetILCodeVersion(); |
| 710 | hr = ilParent.SetActiveNativeCodeVersion(nativeCodeVersion, FALSE); |
| 711 | LOG((LF_TIEREDCOMPILATION, LL_INFO10000, "TieredCompilationManager::ActivateCodeVersion Method=0x%pM (%s::%s), code version id=0x%x. SetActiveNativeCodeVersion ret=0x%x\n" , |
| 712 | pMethod, pMethod->m_pszDebugClassName, pMethod->m_pszDebugMethodName, |
| 713 | nativeCodeVersion.GetVersionId(), |
| 714 | hr)); |
| 715 | } |
| 716 | if (hr == CORPROF_E_RUNTIME_SUSPEND_REQUIRED) |
| 717 | { |
| 718 | // if we start using jump-stamp publishing for tiered compilation, the first attempt |
| 719 | // without the runtime suspended will fail and then this second attempt will |
| 720 | // succeed. |
| 721 | // Even though this works performance is likely to be quite bad. Realistically |
| 722 | // we are going to need batched updates to makes tiered-compilation + jump-stamp |
| 723 | // viable. This fallback path is just here as a proof-of-concept. |
| 724 | ThreadSuspend::SuspendEE(ThreadSuspend::SUSPEND_FOR_REJIT); |
| 725 | { |
| 726 | CodeVersionManager::TableLockHolder lock(pCodeVersionManager); |
| 727 | hr = ilParent.SetActiveNativeCodeVersion(nativeCodeVersion, TRUE); |
| 728 | LOG((LF_TIEREDCOMPILATION, LL_INFO10000, "TieredCompilationManager::ActivateCodeVersion Method=0x%pM (%s::%s), code version id=0x%x. [Suspended] SetActiveNativeCodeVersion ret=0x%x\n" , |
| 729 | pMethod, pMethod->m_pszDebugClassName, pMethod->m_pszDebugMethodName, |
| 730 | nativeCodeVersion.GetVersionId(), |
| 731 | hr)); |
| 732 | } |
| 733 | ThreadSuspend::RestartEE(FALSE, TRUE); |
| 734 | } |
| 735 | if (FAILED(hr)) |
| 736 | { |
| 737 | STRESS_LOG2(LF_TIEREDCOMPILATION, LL_INFO10, "TieredCompilationManager::ActivateCodeVersion: Method %pM failed to publish native code for native code version %d\n" , |
| 738 | pMethod, nativeCodeVersion.GetVersionId()); |
| 739 | } |
| 740 | } |
| 741 | |
| 742 | // Dequeues the next method in the optmization queue. |
| 743 | // This should be called with m_lock already held and runs |
| 744 | // on the background thread. |
| 745 | NativeCodeVersion TieredCompilationManager::GetNextMethodToOptimize() |
| 746 | { |
| 747 | STANDARD_VM_CONTRACT; |
| 748 | |
| 749 | SListElem<NativeCodeVersion>* pElem = m_methodsToOptimize.RemoveHead(); |
| 750 | if (pElem != NULL) |
| 751 | { |
| 752 | NativeCodeVersion nativeCodeVersion = pElem->GetValue(); |
| 753 | delete pElem; |
| 754 | return nativeCodeVersion; |
| 755 | } |
| 756 | return NativeCodeVersion(); |
| 757 | } |
| 758 | |
| 759 | bool TieredCompilationManager::IncrementWorkerThreadCountIfNeeded() |
| 760 | { |
| 761 | WRAPPER_NO_CONTRACT; |
| 762 | // m_lock should be held |
| 763 | |
| 764 | if (0 == m_countOptimizationThreadsRunning && |
| 765 | !m_isAppDomainShuttingDown && |
| 766 | !m_methodsToOptimize.IsEmpty() && |
| 767 | !IsTieringDelayActive()) |
| 768 | { |
| 769 | // Our current policy throttles at 1 thread, but in the future we |
| 770 | // could experiment with more parallelism. |
| 771 | m_countOptimizationThreadsRunning++; |
| 772 | return true; |
| 773 | } |
| 774 | return false; |
| 775 | } |
| 776 | |
| 777 | void TieredCompilationManager::DecrementWorkerThreadCount() |
| 778 | { |
| 779 | STANDARD_VM_CONTRACT; |
| 780 | // m_lock should be held |
| 781 | _ASSERTE(m_countOptimizationThreadsRunning != 0); |
| 782 | |
| 783 | m_countOptimizationThreadsRunning--; |
| 784 | } |
| 785 | |
| 786 | #ifdef _DEBUG |
| 787 | DWORD TieredCompilationManager::DebugGetWorkerThreadCount() |
| 788 | { |
| 789 | WRAPPER_NO_CONTRACT; |
| 790 | |
| 791 | CrstHolder holder(&m_lock); |
| 792 | return m_countOptimizationThreadsRunning; |
| 793 | } |
| 794 | #endif |
| 795 | |
| 796 | //static |
| 797 | CORJIT_FLAGS TieredCompilationManager::GetJitFlags(NativeCodeVersion nativeCodeVersion) |
| 798 | { |
| 799 | LIMITED_METHOD_CONTRACT; |
| 800 | |
| 801 | CORJIT_FLAGS flags; |
| 802 | if (!nativeCodeVersion.GetMethodDesc()->IsEligibleForTieredCompilation()) |
| 803 | { |
| 804 | #ifdef FEATURE_INTERPRETER |
| 805 | flags.Set(CORJIT_FLAGS::CORJIT_FLAG_MAKEFINALCODE); |
| 806 | #endif |
| 807 | return flags; |
| 808 | } |
| 809 | |
| 810 | if (nativeCodeVersion.GetOptimizationTier() == NativeCodeVersion::OptimizationTier0 && |
| 811 | !g_pConfig->TieredCompilation_OptimizeTier0()) |
| 812 | { |
| 813 | flags.Set(CORJIT_FLAGS::CORJIT_FLAG_TIER0); |
| 814 | } |
| 815 | else |
| 816 | { |
| 817 | flags.Set(CORJIT_FLAGS::CORJIT_FLAG_TIER1); |
| 818 | #ifdef FEATURE_INTERPRETER |
| 819 | flags.Set(CORJIT_FLAGS::CORJIT_FLAG_MAKEFINALCODE); |
| 820 | #endif |
| 821 | } |
| 822 | return flags; |
| 823 | } |
| 824 | |
| 825 | #endif // FEATURE_TIERED_COMPILATION && !DACCESS_COMPILE |
| 826 | |