1 /* 2 * Copyright (c) 2010, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "cds/aotLinkedClassBulkLoader.hpp" 26 #include "code/aotCodeCache.hpp" 27 #include "code/scopeDesc.hpp" 28 #include "compiler/compilationPolicy.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compilerDefinitions.inline.hpp" 31 #include "compiler/compilerOracle.hpp" 32 #include "compiler/recompilationPolicy.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "oops/method.inline.hpp" 35 #include "oops/methodData.hpp" 36 #include "oops/oop.inline.hpp" 37 #include "oops/trainingData.hpp" 38 #include "prims/jvmtiExport.hpp" 39 #include "runtime/arguments.hpp" 40 #include "runtime/deoptimization.hpp" 41 #include "runtime/frame.hpp" 42 #include "runtime/frame.inline.hpp" 43 #include "runtime/globals_extension.hpp" 44 #include "runtime/handles.inline.hpp" 45 #include "runtime/safepoint.hpp" 46 #include "runtime/safepointVerifiers.hpp" 47 #ifdef COMPILER1 48 #include "c1/c1_Compiler.hpp" 49 #endif 50 #ifdef COMPILER2 51 #include "opto/c2compiler.hpp" 52 #endif 53 #if INCLUDE_JVMCI 54 #include "jvmci/jvmci.hpp" 55 #endif 56 57 int64_t CompilationPolicy::_start_time = 0; 58 int CompilationPolicy::_c1_count = 0; 59 int CompilationPolicy::_c2_count = 0; 60 int CompilationPolicy::_c3_count = 0; 61 int CompilationPolicy::_ac_count = 0; 62 double CompilationPolicy::_increase_threshold_at_ratio = 0; 63 64 CompilationPolicy::TrainingReplayQueue CompilationPolicy::_training_replay_queue; 65 66 void compilationPolicy_init() { 67 CompilationPolicy::initialize(); 68 } 69 70 int CompilationPolicy::compiler_count(CompLevel comp_level) { 71 if (is_c1_compile(comp_level)) { 72 return c1_count(); 73 } else if (is_c2_compile(comp_level)) { 74 return c2_count(); 75 } 76 return 0; 77 } 78 79 // Returns true if m must be compiled before executing it 80 // This is intended to force compiles for methods (usually for 81 // debugging) that would otherwise be interpreted for some reason. 82 bool CompilationPolicy::must_be_compiled(const methodHandle& m, int comp_level) { 83 // Don't allow Xcomp to cause compiles in replay mode 84 if (ReplayCompiles) return false; 85 86 if (m->has_compiled_code()) return false; // already compiled 87 if (!can_be_compiled(m, comp_level)) return false; 88 89 return !UseInterpreter || // must compile all methods 90 (AlwaysCompileLoopMethods && m->has_loops() && CompileBroker::should_compile_new_jobs()); // eagerly compile loop methods 91 } 92 93 void CompilationPolicy::maybe_compile_early(const methodHandle& m, TRAPS) { 94 if (m->method_holder()->is_not_initialized()) { 95 // 'is_not_initialized' means not only '!is_initialized', but also that 96 // initialization has not been started yet ('!being_initialized') 97 // Do not force compilation of methods in uninitialized classes. 98 return; 99 } 100 if (!m->is_native() && MethodTrainingData::have_data()) { 101 MethodTrainingData* mtd = MethodTrainingData::find_fast(m); 102 if (mtd == nullptr) { 103 return; // there is no training data recorded for m 104 } 105 bool recompile = m->code_has_clinit_barriers(); 106 CompLevel cur_level = static_cast<CompLevel>(m->highest_comp_level()); 107 CompLevel next_level = trained_transition(m, cur_level, mtd, THREAD); 108 if ((next_level != cur_level || recompile) && can_be_compiled(m, next_level) && !CompileBroker::compilation_is_in_queue(m)) { 109 bool requires_online_compilation = false; 110 CompileTrainingData* ctd = mtd->last_toplevel_compile(next_level); 111 if (ctd != nullptr) { 112 requires_online_compilation = (ctd->init_deps_left() > 0); 113 } 114 if (requires_online_compilation && recompile) { 115 return; 116 } 117 if (PrintTieredEvents) { 118 print_event(FORCE_COMPILE, m(), m(), InvocationEntryBci, next_level); 119 } 120 CompileBroker::compile_method(m, InvocationEntryBci, next_level, methodHandle(), 0, requires_online_compilation, CompileTask::Reason_MustBeCompiled, THREAD); 121 if (HAS_PENDING_EXCEPTION) { 122 CLEAR_PENDING_EXCEPTION; 123 } 124 } 125 } 126 } 127 128 void CompilationPolicy::maybe_compile_early_after_init(const methodHandle& m, TRAPS) { 129 assert(m->method_holder()->is_initialized(), "Should be called after class initialization"); 130 maybe_compile_early(m, THREAD); 131 } 132 133 void CompilationPolicy::compile_if_required(const methodHandle& m, TRAPS) { 134 if (!THREAD->can_call_java() || THREAD->is_Compiler_thread()) { 135 // don't force compilation, resolve was on behalf of compiler 136 return; 137 } 138 if (m->method_holder()->is_not_initialized()) { 139 // 'is_not_initialized' means not only '!is_initialized', but also that 140 // initialization has not been started yet ('!being_initialized') 141 // Do not force compilation of methods in uninitialized classes. 142 // Note that doing this would throw an assert later, 143 // in CompileBroker::compile_method. 144 // We sometimes use the link resolver to do reflective lookups 145 // even before classes are initialized. 146 return; 147 } 148 149 if (must_be_compiled(m)) { 150 // This path is unusual, mostly used by the '-Xcomp' stress test mode. 151 CompLevel level = initial_compile_level(m); 152 if (PrintTieredEvents) { 153 print_event(FORCE_COMPILE, m(), m(), InvocationEntryBci, level); 154 } 155 CompileBroker::compile_method(m, InvocationEntryBci, level, methodHandle(), 0, false, CompileTask::Reason_MustBeCompiled, THREAD); 156 } 157 } 158 159 void CompilationPolicy::replay_training_at_init_impl(InstanceKlass* klass, TRAPS) { 160 if (!klass->has_init_deps_processed()) { 161 ResourceMark rm; 162 log_debug(training)("Replay training: %s", klass->external_name()); 163 164 KlassTrainingData* ktd = KlassTrainingData::find(klass); 165 if (ktd != nullptr) { 166 guarantee(ktd->has_holder(), ""); 167 ktd->notice_fully_initialized(); // sets klass->has_init_deps_processed bit 168 assert(klass->has_init_deps_processed(), ""); 169 170 if (AOTCompileEagerly) { 171 ktd->iterate_comp_deps([&](CompileTrainingData* ctd) { 172 if (ctd->init_deps_left() == 0) { 173 MethodTrainingData* mtd = ctd->method(); 174 if (mtd->has_holder()) { 175 const methodHandle mh(THREAD, const_cast<Method*>(mtd->holder())); 176 CompilationPolicy::maybe_compile_early(mh, THREAD); 177 } 178 } 179 }); 180 } 181 } 182 } 183 } 184 185 void CompilationPolicy::flush_replay_training_at_init(TRAPS) { 186 MonitorLocker locker(THREAD, TrainingReplayQueue_lock); 187 while (!_training_replay_queue.is_empty_unlocked()) { 188 locker.wait(); // let the replay training thread drain the queue 189 } 190 } 191 192 void CompilationPolicy::replay_training_at_init(InstanceKlass* klass, TRAPS) { 193 assert(klass->is_initialized(), ""); 194 if (TrainingData::have_data() && klass->is_shared()) { 195 _training_replay_queue.push(klass, TrainingReplayQueue_lock, THREAD); 196 } 197 } 198 199 // For TrainingReplayQueue 200 template<> 201 void CompilationPolicyUtils::Queue<InstanceKlass>::print_on(outputStream* st) { 202 int pos = 0; 203 for (QueueNode* cur = _head; cur != nullptr; cur = cur->next()) { 204 ResourceMark rm; 205 InstanceKlass* ik = cur->value(); 206 st->print_cr("%3d: " INTPTR_FORMAT " %s", ++pos, p2i(ik), ik->external_name()); 207 } 208 } 209 210 void CompilationPolicy::replay_training_at_init_loop(TRAPS) { 211 while (!CompileBroker::is_compilation_disabled_forever() || AOTVerifyTrainingData) { 212 InstanceKlass* ik = _training_replay_queue.pop(TrainingReplayQueue_lock, THREAD); 213 if (ik != nullptr) { 214 replay_training_at_init_impl(ik, THREAD); 215 } 216 } 217 } 218 219 static inline CompLevel adjust_level_for_compilability_query(CompLevel comp_level) { 220 if (comp_level == CompLevel_any) { 221 if (CompilerConfig::is_c1_only()) { 222 comp_level = CompLevel_simple; 223 } else if (CompilerConfig::is_c2_or_jvmci_compiler_only()) { 224 comp_level = CompLevel_full_optimization; 225 } 226 } 227 return comp_level; 228 } 229 230 // Returns true if m is allowed to be compiled 231 bool CompilationPolicy::can_be_compiled(const methodHandle& m, int comp_level) { 232 // allow any levels for WhiteBox 233 assert(WhiteBoxAPI || comp_level == CompLevel_any || is_compile(comp_level), "illegal compilation level %d", comp_level); 234 235 if (m->is_abstract()) return false; 236 if (DontCompileHugeMethods && m->code_size() > HugeMethodLimit) return false; 237 238 // Math intrinsics should never be compiled as this can lead to 239 // monotonicity problems because the interpreter will prefer the 240 // compiled code to the intrinsic version. This can't happen in 241 // production because the invocation counter can't be incremented 242 // but we shouldn't expose the system to this problem in testing 243 // modes. 244 if (!AbstractInterpreter::can_be_compiled(m)) { 245 return false; 246 } 247 comp_level = adjust_level_for_compilability_query((CompLevel) comp_level); 248 if (comp_level == CompLevel_any || is_compile(comp_level)) { 249 return !m->is_not_compilable(comp_level); 250 } 251 return false; 252 } 253 254 // Returns true if m is allowed to be osr compiled 255 bool CompilationPolicy::can_be_osr_compiled(const methodHandle& m, int comp_level) { 256 bool result = false; 257 comp_level = adjust_level_for_compilability_query((CompLevel) comp_level); 258 if (comp_level == CompLevel_any || is_compile(comp_level)) { 259 result = !m->is_not_osr_compilable(comp_level); 260 } 261 return (result && can_be_compiled(m, comp_level)); 262 } 263 264 bool CompilationPolicy::is_compilation_enabled() { 265 // NOTE: CompileBroker::should_compile_new_jobs() checks for UseCompiler 266 return CompileBroker::should_compile_new_jobs(); 267 } 268 269 CompileTask* CompilationPolicy::select_task_helper(CompileQueue* compile_queue) { 270 // Remove unloaded methods from the queue 271 for (CompileTask* task = compile_queue->first(); task != nullptr; ) { 272 CompileTask* next = task->next(); 273 if (task->is_unloaded()) { 274 compile_queue->remove_and_mark_stale(task); 275 } 276 task = next; 277 } 278 #if INCLUDE_JVMCI 279 if (UseJVMCICompiler && !BackgroundCompilation) { 280 /* 281 * In blocking compilation mode, the CompileBroker will make 282 * compilations submitted by a JVMCI compiler thread non-blocking. These 283 * compilations should be scheduled after all blocking compilations 284 * to service non-compiler related compilations sooner and reduce the 285 * chance of such compilations timing out. 286 */ 287 for (CompileTask* task = compile_queue->first(); task != nullptr; task = task->next()) { 288 if (task->is_blocking()) { 289 return task; 290 } 291 } 292 } 293 #endif 294 return compile_queue->first(); 295 } 296 297 // Simple methods are as good being compiled with C1 as C2. 298 // Determine if a given method is such a case. 299 bool CompilationPolicy::is_trivial(const methodHandle& method) { 300 if (method->is_accessor() || 301 method->is_constant_getter()) { 302 return true; 303 } 304 return false; 305 } 306 307 bool CompilationPolicy::force_comp_at_level_simple(const methodHandle& method) { 308 if (CompilationModeFlag::quick_internal()) { 309 #if INCLUDE_JVMCI 310 if (UseJVMCICompiler) { 311 AbstractCompiler* comp = CompileBroker::compiler(CompLevel_full_optimization); 312 if (comp != nullptr && comp->is_jvmci() && ((JVMCICompiler*) comp)->force_comp_at_level_simple(method)) { 313 return !AOTCodeCache::is_C3_on(); 314 } 315 } 316 #endif 317 } 318 return false; 319 } 320 321 CompLevel CompilationPolicy::comp_level(Method* method) { 322 nmethod *nm = method->code(); 323 if (nm != nullptr && nm->is_in_use()) { 324 return (CompLevel)nm->comp_level(); 325 } 326 return CompLevel_none; 327 } 328 329 // Call and loop predicates determine whether a transition to a higher 330 // compilation level should be performed (pointers to predicate functions 331 // are passed to common()). 332 // Tier?LoadFeedback is basically a coefficient that determines of 333 // how many methods per compiler thread can be in the queue before 334 // the threshold values double. 335 class LoopPredicate : AllStatic { 336 public: 337 static bool apply_scaled(const methodHandle& method, CompLevel cur_level, int i, int b, double scale) { 338 double threshold_scaling; 339 if (CompilerOracle::has_option_value(method, CompileCommandEnum::CompileThresholdScaling, threshold_scaling)) { 340 scale *= threshold_scaling; 341 } 342 switch(cur_level) { 343 case CompLevel_none: 344 case CompLevel_limited_profile: 345 return b >= Tier3BackEdgeThreshold * scale; 346 case CompLevel_full_profile: 347 return b >= Tier4BackEdgeThreshold * scale; 348 default: 349 return true; 350 } 351 } 352 353 static bool apply(const methodHandle& method, CompLevel cur_level, int i, int b) { 354 double k = 1; 355 switch(cur_level) { 356 case CompLevel_none: 357 // Fall through 358 case CompLevel_limited_profile: { 359 k = CompilationPolicy::threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 360 break; 361 } 362 case CompLevel_full_profile: { 363 k = CompilationPolicy::threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 364 break; 365 } 366 default: 367 return true; 368 } 369 return apply_scaled(method, cur_level, i, b, k); 370 } 371 }; 372 373 class CallPredicate : AllStatic { 374 public: 375 static bool apply_scaled(const methodHandle& method, CompLevel cur_level, int i, int b, double scale) { 376 double threshold_scaling; 377 if (CompilerOracle::has_option_value(method, CompileCommandEnum::CompileThresholdScaling, threshold_scaling)) { 378 scale *= threshold_scaling; 379 } 380 switch(cur_level) { 381 case CompLevel_none: 382 case CompLevel_limited_profile: 383 return (i >= Tier3InvocationThreshold * scale) || 384 (i >= Tier3MinInvocationThreshold * scale && i + b >= Tier3CompileThreshold * scale); 385 case CompLevel_full_profile: 386 return (i >= Tier4InvocationThreshold * scale) || 387 (i >= Tier4MinInvocationThreshold * scale && i + b >= Tier4CompileThreshold * scale); 388 default: 389 return true; 390 } 391 } 392 393 static bool apply(const methodHandle& method, CompLevel cur_level, int i, int b) { 394 double k = 1; 395 switch(cur_level) { 396 case CompLevel_none: 397 case CompLevel_limited_profile: { 398 k = CompilationPolicy::threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 399 break; 400 } 401 case CompLevel_full_profile: { 402 k = CompilationPolicy::threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 403 break; 404 } 405 default: 406 return true; 407 } 408 return apply_scaled(method, cur_level, i, b, k); 409 } 410 }; 411 412 double CompilationPolicy::threshold_scale(CompLevel level, int feedback_k) { 413 int comp_count = compiler_count(level); 414 if (comp_count > 0) { 415 double queue_size = CompileBroker::queue_size(level); 416 double k = (double)queue_size / ((double)feedback_k * (double)comp_count) + 1; 417 418 // Increase C1 compile threshold when the code cache is filled more 419 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 420 // The main intention is to keep enough free space for C2 compiled code 421 // to achieve peak performance if the code cache is under stress. 422 if (CompilerConfig::is_tiered() && !CompilationModeFlag::disable_intermediate() && is_c1_compile(level)) { 423 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(); 424 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 425 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 426 } 427 } 428 return k; 429 } 430 return 1; 431 } 432 433 void CompilationPolicy::print_counters(const char* prefix, Method* m) { 434 int invocation_count = m->invocation_count(); 435 int backedge_count = m->backedge_count(); 436 MethodData* mdh = m->method_data(); 437 int mdo_invocations = 0, mdo_backedges = 0; 438 int mdo_invocations_start = 0, mdo_backedges_start = 0; 439 if (mdh != nullptr) { 440 mdo_invocations = mdh->invocation_count(); 441 mdo_backedges = mdh->backedge_count(); 442 mdo_invocations_start = mdh->invocation_count_start(); 443 mdo_backedges_start = mdh->backedge_count_start(); 444 } 445 tty->print(" %stotal=%d,%d %smdo=%d(%d),%d(%d)", prefix, 446 invocation_count, backedge_count, prefix, 447 mdo_invocations, mdo_invocations_start, 448 mdo_backedges, mdo_backedges_start); 449 tty->print(" %smax levels=%d,%d", prefix, 450 m->highest_comp_level(), m->highest_osr_comp_level()); 451 } 452 453 void CompilationPolicy::print_training_data(const char* prefix, Method* method) { 454 methodHandle m(Thread::current(), method); 455 tty->print(" %smtd: ", prefix); 456 MethodTrainingData* mtd = MethodTrainingData::find(m); 457 if (mtd == nullptr) { 458 tty->print("null"); 459 } else { 460 MethodData* md = mtd->final_profile(); 461 tty->print("mdo="); 462 if (md == nullptr) { 463 tty->print("null"); 464 } else { 465 int mdo_invocations = md->invocation_count(); 466 int mdo_backedges = md->backedge_count(); 467 int mdo_invocations_start = md->invocation_count_start(); 468 int mdo_backedges_start = md->backedge_count_start(); 469 tty->print("%d(%d), %d(%d)", mdo_invocations, mdo_invocations_start, mdo_backedges, mdo_backedges_start); 470 } 471 CompileTrainingData* ctd = mtd->last_toplevel_compile(CompLevel_full_optimization); 472 tty->print(", deps="); 473 if (ctd == nullptr) { 474 tty->print("null"); 475 } else { 476 tty->print("%d", ctd->init_deps_left()); 477 } 478 } 479 } 480 481 // Print an event. 482 void CompilationPolicy::print_event(EventType type, Method* m, Method* im, int bci, CompLevel level) { 483 bool inlinee_event = m != im; 484 485 ttyLocker tty_lock; 486 tty->print("%lf: [", os::elapsedTime()); 487 488 switch(type) { 489 case CALL: 490 tty->print("call"); 491 break; 492 case LOOP: 493 tty->print("loop"); 494 break; 495 case COMPILE: 496 tty->print("compile"); 497 break; 498 case FORCE_COMPILE: 499 tty->print("force-compile"); 500 break; 501 case FORCE_RECOMPILE: 502 tty->print("force-recompile"); 503 break; 504 case REMOVE_FROM_QUEUE: 505 tty->print("remove-from-queue"); 506 break; 507 case UPDATE_IN_QUEUE: 508 tty->print("update-in-queue"); 509 break; 510 case REPROFILE: 511 tty->print("reprofile"); 512 break; 513 case MAKE_NOT_ENTRANT: 514 tty->print("make-not-entrant"); 515 break; 516 default: 517 tty->print("unknown"); 518 } 519 520 tty->print(" level=%d ", level); 521 522 ResourceMark rm; 523 char *method_name = m->name_and_sig_as_C_string(); 524 tty->print("[%s", method_name); 525 if (inlinee_event) { 526 char *inlinee_name = im->name_and_sig_as_C_string(); 527 tty->print(" [%s]] ", inlinee_name); 528 } 529 else tty->print("] "); 530 tty->print("@%d queues=%d,%d", bci, CompileBroker::queue_size(CompLevel_full_profile), 531 CompileBroker::queue_size(CompLevel_full_optimization)); 532 533 tty->print(" rate="); 534 if (m->prev_time() == 0) tty->print("n/a"); 535 else tty->print("%f", m->rate()); 536 537 RecompilationPolicy::print_load_average(); 538 539 tty->print(" k=%.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback), 540 threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback)); 541 542 if (type != COMPILE) { 543 print_counters("", m); 544 if (inlinee_event) { 545 print_counters("inlinee ", im); 546 } 547 tty->print(" compilable="); 548 bool need_comma = false; 549 if (!m->is_not_compilable(CompLevel_full_profile)) { 550 tty->print("c1"); 551 need_comma = true; 552 } 553 if (!m->is_not_osr_compilable(CompLevel_full_profile)) { 554 if (need_comma) tty->print(","); 555 tty->print("c1-osr"); 556 need_comma = true; 557 } 558 if (!m->is_not_compilable(CompLevel_full_optimization)) { 559 if (need_comma) tty->print(","); 560 tty->print("c2"); 561 need_comma = true; 562 } 563 if (!m->is_not_osr_compilable(CompLevel_full_optimization)) { 564 if (need_comma) tty->print(","); 565 tty->print("c2-osr"); 566 } 567 tty->print(" status="); 568 if (m->queued_for_compilation()) { 569 tty->print("in-queue"); 570 } else tty->print("idle"); 571 print_training_data("", m); 572 if (inlinee_event) { 573 print_training_data("inlinee ", im); 574 } 575 } 576 tty->print_cr("]"); 577 } 578 579 void CompilationPolicy::initialize() { 580 if (!CompilerConfig::is_interpreter_only()) { 581 if (AOTCodeCache::is_dumping_code()) { 582 // Assembly phase runs C1 and C2 compilation in separate phases, 583 // and can use all the CPU threads it can reach. Adjust the common 584 // options before policy starts overwriting them. There is a block 585 // at the very end that overrides final thread counts. 586 if (FLAG_IS_DEFAULT(UseDynamicNumberOfCompilerThreads)) { 587 FLAG_SET_ERGO(UseDynamicNumberOfCompilerThreads, false); 588 } 589 if (FLAG_IS_DEFAULT(CICompilerCount)) { 590 FLAG_SET_ERGO(CICompilerCount, MAX2(2, os::active_processor_count())); 591 } 592 } 593 int count = CICompilerCount; 594 bool c1_only = CompilerConfig::is_c1_only(); 595 bool c2_only = CompilerConfig::is_c2_or_jvmci_compiler_only(); 596 597 #ifdef _LP64 598 // Turn on ergonomic compiler count selection 599 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { 600 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); 601 } 602 if (CICompilerCountPerCPU) { 603 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n 604 int log_cpu = log2i(os::active_processor_count()); 605 int loglog_cpu = log2i(MAX2(log_cpu, 1)); 606 count = MAX2(log_cpu * loglog_cpu * 3 / 2, 2); 607 // Make sure there is enough space in the code cache to hold all the compiler buffers 608 size_t c1_size = 0; 609 #ifdef COMPILER1 610 c1_size = Compiler::code_buffer_size(); 611 #endif 612 size_t c2_size = 0; 613 #ifdef COMPILER2 614 c2_size = C2Compiler::initial_code_buffer_size(); 615 #endif 616 size_t buffer_size = c1_only ? c1_size : (c1_size/3 + 2*c2_size/3); 617 int max_count = (ReservedCodeCacheSize - (CodeCacheMinimumUseSpace DEBUG_ONLY(* 3))) / (int)buffer_size; 618 if (count > max_count) { 619 // Lower the compiler count such that all buffers fit into the code cache 620 count = MAX2(max_count, c1_only ? 1 : 2); 621 } 622 FLAG_SET_ERGO(CICompilerCount, count); 623 } 624 #else 625 // On 32-bit systems, the number of compiler threads is limited to 3. 626 // On these systems, the virtual address space available to the JVM 627 // is usually limited to 2-4 GB (the exact value depends on the platform). 628 // As the compilers (especially C2) can consume a large amount of 629 // memory, scaling the number of compiler threads with the number of 630 // available cores can result in the exhaustion of the address space 631 /// available to the VM and thus cause the VM to crash. 632 if (FLAG_IS_DEFAULT(CICompilerCount)) { 633 count = 3; 634 FLAG_SET_ERGO(CICompilerCount, count); 635 } 636 #endif 637 638 if (c1_only) { 639 // No C2 compiler thread required 640 set_c1_count(count); 641 } else if (c2_only) { 642 set_c2_count(count); 643 } else { 644 #if INCLUDE_JVMCI 645 if (UseJVMCICompiler && UseJVMCINativeLibrary) { 646 int libjvmci_count = MAX2((int) (count * JVMCINativeLibraryThreadFraction), 1); 647 int c1_count = MAX2(count - libjvmci_count, 1); 648 set_c2_count(libjvmci_count); 649 set_c1_count(c1_count); 650 } else if (AOTCodeCache::is_C3_on()) { 651 set_c1_count(MAX2(count / 3, 1)); 652 set_c2_count(MAX2(count - c1_count(), 1)); 653 set_c3_count(1); 654 } else 655 #endif 656 { 657 set_c1_count(MAX2(count / 3, 1)); 658 set_c2_count(MAX2(count - c1_count(), 1)); 659 } 660 } 661 if (AOTCodeCache::is_dumping_code()) { 662 set_c1_count(count); 663 set_c2_count(count); 664 count *= 2; // satisfy the assert below 665 } 666 if (AOTCodeCache::is_code_load_thread_on()) { 667 set_ac_count((c1_only || c2_only) ? 1 : 2); // At minimum we need 2 threads to load C1 and C2 cached code in parallel 668 } 669 assert(count == c1_count() + c2_count(), "inconsistent compiler thread count"); 670 set_increase_threshold_at_ratio(); 671 } 672 673 set_start_time(nanos_to_millis(os::javaTimeNanos())); 674 } 675 676 677 678 679 #ifdef ASSERT 680 bool CompilationPolicy::verify_level(CompLevel level) { 681 if (TieredCompilation && level > TieredStopAtLevel) { 682 return false; 683 } 684 // Check if there is a compiler to process the requested level 685 if (!CompilerConfig::is_c1_enabled() && is_c1_compile(level)) { 686 return false; 687 } 688 if (!CompilerConfig::is_c2_or_jvmci_compiler_enabled() && is_c2_compile(level)) { 689 return false; 690 } 691 692 // Interpreter level is always valid. 693 if (level == CompLevel_none) { 694 return true; 695 } 696 if (CompilationModeFlag::normal()) { 697 return true; 698 } else if (CompilationModeFlag::quick_only()) { 699 return level == CompLevel_simple; 700 } else if (CompilationModeFlag::high_only()) { 701 return level == CompLevel_full_optimization; 702 } else if (CompilationModeFlag::high_only_quick_internal()) { 703 return level == CompLevel_full_optimization || level == CompLevel_simple; 704 } 705 return false; 706 } 707 #endif 708 709 710 CompLevel CompilationPolicy::highest_compile_level() { 711 CompLevel level = CompLevel_none; 712 // Setup the maximum level available for the current compiler configuration. 713 if (!CompilerConfig::is_interpreter_only()) { 714 if (CompilerConfig::is_c2_or_jvmci_compiler_enabled()) { 715 level = CompLevel_full_optimization; 716 } else if (CompilerConfig::is_c1_enabled()) { 717 if (CompilerConfig::is_c1_simple_only()) { 718 level = CompLevel_simple; 719 } else { 720 level = CompLevel_full_profile; 721 } 722 } 723 } 724 // Clamp the maximum level with TieredStopAtLevel. 725 if (TieredCompilation) { 726 level = MIN2(level, (CompLevel) TieredStopAtLevel); 727 } 728 729 // Fix it up if after the clamping it has become invalid. 730 // Bring it monotonically down depending on the next available level for 731 // the compilation mode. 732 if (!CompilationModeFlag::normal()) { 733 // a) quick_only - levels 2,3,4 are invalid; levels -1,0,1 are valid; 734 // b) high_only - levels 1,2,3 are invalid; levels -1,0,4 are valid; 735 // c) high_only_quick_internal - levels 2,3 are invalid; levels -1,0,1,4 are valid. 736 if (CompilationModeFlag::quick_only()) { 737 if (level == CompLevel_limited_profile || level == CompLevel_full_profile || level == CompLevel_full_optimization) { 738 level = CompLevel_simple; 739 } 740 } else if (CompilationModeFlag::high_only()) { 741 if (level == CompLevel_simple || level == CompLevel_limited_profile || level == CompLevel_full_profile) { 742 level = CompLevel_none; 743 } 744 } else if (CompilationModeFlag::high_only_quick_internal()) { 745 if (level == CompLevel_limited_profile || level == CompLevel_full_profile) { 746 level = CompLevel_simple; 747 } 748 } 749 } 750 751 assert(verify_level(level), "Invalid highest compilation level: %d", level); 752 return level; 753 } 754 755 CompLevel CompilationPolicy::limit_level(CompLevel level) { 756 level = MIN2(level, highest_compile_level()); 757 assert(verify_level(level), "Invalid compilation level: %d", level); 758 return level; 759 } 760 761 CompLevel CompilationPolicy::initial_compile_level(const methodHandle& method) { 762 CompLevel level = CompLevel_any; 763 if (CompilationModeFlag::normal()) { 764 level = CompLevel_full_profile; 765 } else if (CompilationModeFlag::quick_only()) { 766 level = CompLevel_simple; 767 } else if (CompilationModeFlag::high_only()) { 768 level = CompLevel_full_optimization; 769 } else if (CompilationModeFlag::high_only_quick_internal()) { 770 if (force_comp_at_level_simple(method)) { 771 level = CompLevel_simple; 772 } else { 773 level = CompLevel_full_optimization; 774 } 775 } 776 assert(level != CompLevel_any, "Unhandled compilation mode"); 777 return limit_level(level); 778 } 779 780 // Set carry flags on the counters if necessary 781 void CompilationPolicy::handle_counter_overflow(const methodHandle& method) { 782 MethodCounters *mcs = method->method_counters(); 783 if (mcs != nullptr) { 784 mcs->invocation_counter()->set_carry_on_overflow(); 785 mcs->backedge_counter()->set_carry_on_overflow(); 786 } 787 MethodData* mdo = method->method_data(); 788 if (mdo != nullptr) { 789 mdo->invocation_counter()->set_carry_on_overflow(); 790 mdo->backedge_counter()->set_carry_on_overflow(); 791 } 792 } 793 794 // Called with the queue locked and with at least one element 795 CompileTask* CompilationPolicy::select_task(CompileQueue* compile_queue, JavaThread* THREAD) { 796 CompileTask *max_blocking_task = nullptr; 797 CompileTask *max_task = nullptr; 798 Method* max_method = nullptr; 799 800 int64_t t = nanos_to_millis(os::javaTimeNanos()); 801 // Iterate through the queue and find a method with a maximum rate. 802 for (CompileTask* task = compile_queue->first(); task != nullptr;) { 803 CompileTask* next_task = task->next(); 804 // If a method was unloaded or has been stale for some time, remove it from the queue. 805 // Blocking tasks and tasks submitted from whitebox API don't become stale 806 if (task->is_unloaded()) { 807 compile_queue->remove_and_mark_stale(task); 808 task = next_task; 809 continue; 810 } 811 if (task->is_aot()) { 812 // AOTCodeCache tasks are on separate queue, and they should load fast. There is no need to walk 813 // the rest of the queue, just take the task and go. 814 return task; 815 } 816 if (task->is_blocking() && task->compile_reason() == CompileTask::Reason_Whitebox) { 817 // CTW tasks, submitted as blocking Whitebox requests, do not participate in rate 818 // selection and/or any level adjustments. Just return them in order. 819 return task; 820 } 821 Method* method = task->method(); 822 methodHandle mh(THREAD, method); 823 if (task->can_become_stale() && is_stale(t, TieredCompileTaskTimeout, mh) && !is_old(mh)) { 824 if (PrintTieredEvents) { 825 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel) task->comp_level()); 826 } 827 method->clear_queued_for_compilation(); 828 method->set_pending_queue_processed(false); 829 compile_queue->remove_and_mark_stale(task); 830 task = next_task; 831 continue; 832 } 833 update_rate(t, mh); 834 if (max_task == nullptr || compare_methods(method, max_method) || compare_tasks(task, max_task)) { 835 // Select a method with the highest rate 836 max_task = task; 837 max_method = method; 838 } 839 840 if (task->is_blocking()) { 841 if (max_blocking_task == nullptr || compare_methods(method, max_blocking_task->method())) { 842 max_blocking_task = task; 843 } 844 } 845 846 task = next_task; 847 } 848 849 if (max_blocking_task != nullptr) { 850 // In blocking compilation mode, the CompileBroker will make 851 // compilations submitted by a JVMCI compiler thread non-blocking. These 852 // compilations should be scheduled after all blocking compilations 853 // to service non-compiler related compilations sooner and reduce the 854 // chance of such compilations timing out. 855 max_task = max_blocking_task; 856 max_method = max_task->method(); 857 } 858 859 methodHandle max_method_h(THREAD, max_method); 860 861 if (max_task != nullptr && max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile && 862 max_method != nullptr && is_method_profiled(max_method_h) && !Arguments::is_compiler_only()) { 863 max_task->set_comp_level(CompLevel_limited_profile); 864 865 if (CompileBroker::compilation_is_complete(max_method_h(), max_task->osr_bci(), CompLevel_limited_profile, 866 false /* requires_online_compilation */, 867 CompileTask::Reason_None)) { 868 if (PrintTieredEvents) { 869 print_event(REMOVE_FROM_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 870 } 871 compile_queue->remove_and_mark_stale(max_task); 872 max_method->clear_queued_for_compilation(); 873 return nullptr; 874 } 875 876 if (PrintTieredEvents) { 877 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 878 } 879 } 880 return max_task; 881 } 882 883 void CompilationPolicy::reprofile(ScopeDesc* trap_scope, bool is_osr) { 884 for (ScopeDesc* sd = trap_scope;; sd = sd->sender()) { 885 if (PrintTieredEvents) { 886 print_event(REPROFILE, sd->method(), sd->method(), InvocationEntryBci, CompLevel_none); 887 } 888 MethodData* mdo = sd->method()->method_data(); 889 if (mdo != nullptr) { 890 mdo->reset_start_counters(); 891 } 892 if (sd->is_top()) break; 893 } 894 } 895 896 nmethod* CompilationPolicy::event(const methodHandle& method, const methodHandle& inlinee, 897 int branch_bci, int bci, CompLevel comp_level, nmethod* nm, TRAPS) { 898 if (PrintTieredEvents) { 899 print_event(bci == InvocationEntryBci ? CALL : LOOP, method(), inlinee(), bci, comp_level); 900 } 901 902 #if INCLUDE_JVMCI 903 if (EnableJVMCI && UseJVMCICompiler && 904 comp_level == CompLevel_full_optimization CDS_ONLY(&& !AOTLinkedClassBulkLoader::class_preloading_finished())) { 905 return nullptr; 906 } 907 #endif 908 909 if (comp_level == CompLevel_none && 910 JvmtiExport::can_post_interpreter_events() && 911 THREAD->is_interp_only_mode()) { 912 return nullptr; 913 } 914 if (ReplayCompiles) { 915 // Don't trigger other compiles in testing mode 916 return nullptr; 917 } 918 919 handle_counter_overflow(method); 920 if (method() != inlinee()) { 921 handle_counter_overflow(inlinee); 922 } 923 924 if (bci == InvocationEntryBci) { 925 method_invocation_event(method, inlinee, comp_level, nm, THREAD); 926 } else { 927 // method == inlinee if the event originated in the main method 928 method_back_branch_event(method, inlinee, bci, comp_level, nm, THREAD); 929 // Check if event led to a higher level OSR compilation 930 CompLevel expected_comp_level = MIN2(CompLevel_full_optimization, static_cast<CompLevel>(comp_level + 1)); 931 if (!CompilationModeFlag::disable_intermediate() && inlinee->is_not_osr_compilable(expected_comp_level)) { 932 // It's not possible to reach the expected level so fall back to simple. 933 expected_comp_level = CompLevel_simple; 934 } 935 CompLevel max_osr_level = static_cast<CompLevel>(inlinee->highest_osr_comp_level()); 936 if (max_osr_level >= expected_comp_level) { // fast check to avoid locking in a typical scenario 937 nmethod* osr_nm = inlinee->lookup_osr_nmethod_for(bci, expected_comp_level, false); 938 assert(osr_nm == nullptr || osr_nm->comp_level() >= expected_comp_level, "lookup_osr_nmethod_for is broken"); 939 if (osr_nm != nullptr && osr_nm->comp_level() != comp_level) { 940 // Perform OSR with new nmethod 941 return osr_nm; 942 } 943 } 944 } 945 return nullptr; 946 } 947 948 // Check if the method can be compiled, change level if necessary 949 void CompilationPolicy::compile(const methodHandle& mh, int bci, CompLevel level, TRAPS) { 950 assert(verify_level(level), "Invalid compilation level requested: %d", level); 951 952 if (level == CompLevel_none) { 953 if (mh->has_compiled_code()) { 954 // Happens when we switch to interpreter to profile. 955 MutexLocker ml(Compile_lock); 956 NoSafepointVerifier nsv; 957 if (mh->has_compiled_code()) { 958 mh->code()->make_not_used(); 959 } 960 // Deoptimize immediately (we don't have to wait for a compile). 961 JavaThread* jt = THREAD; 962 RegisterMap map(jt, 963 RegisterMap::UpdateMap::skip, 964 RegisterMap::ProcessFrames::include, 965 RegisterMap::WalkContinuation::skip); 966 frame fr = jt->last_frame().sender(&map); 967 Deoptimization::deoptimize_frame(jt, fr.id()); 968 } 969 return; 970 } 971 972 if (!CompilationModeFlag::disable_intermediate()) { 973 // Check if the method can be compiled. If it cannot be compiled with C1, continue profiling 974 // in the interpreter and then compile with C2 (the transition function will request that, 975 // see common() ). If the method cannot be compiled with C2 but still can with C1, compile it with 976 // pure C1. 977 if ((bci == InvocationEntryBci && !can_be_compiled(mh, level))) { 978 if (level == CompLevel_full_optimization && can_be_compiled(mh, CompLevel_simple)) { 979 compile(mh, bci, CompLevel_simple, THREAD); 980 } 981 return; 982 } 983 if ((bci != InvocationEntryBci && !can_be_osr_compiled(mh, level))) { 984 if (level == CompLevel_full_optimization && can_be_osr_compiled(mh, CompLevel_simple)) { 985 nmethod* osr_nm = mh->lookup_osr_nmethod_for(bci, CompLevel_simple, false); 986 if (osr_nm != nullptr && osr_nm->comp_level() > CompLevel_simple) { 987 // Invalidate the existing OSR nmethod so that a compile at CompLevel_simple is permitted. 988 osr_nm->make_not_entrant("OSR invalidation for compiling with C1"); 989 } 990 compile(mh, bci, CompLevel_simple, THREAD); 991 } 992 return; 993 } 994 } 995 if (bci != InvocationEntryBci && mh->is_not_osr_compilable(level)) { 996 return; 997 } 998 if (!CompileBroker::compilation_is_in_queue(mh)) { 999 if (PrintTieredEvents) { 1000 print_event(COMPILE, mh(), mh(), bci, level); 1001 } 1002 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 1003 update_rate(nanos_to_millis(os::javaTimeNanos()), mh); 1004 bool requires_online_compilation = false; 1005 if (TrainingData::have_data()) { 1006 MethodTrainingData* mtd = MethodTrainingData::find_fast(mh); 1007 if (mtd != nullptr) { 1008 CompileTrainingData* ctd = mtd->last_toplevel_compile(level); 1009 if (ctd != nullptr) { 1010 requires_online_compilation = (ctd->init_deps_left() > 0); 1011 } 1012 } 1013 } 1014 CompileBroker::compile_method(mh, bci, level, mh, hot_count, requires_online_compilation, CompileTask::Reason_Tiered, THREAD); 1015 } 1016 } 1017 1018 // update_rate() is called from select_task() while holding a compile queue lock. 1019 void CompilationPolicy::update_rate(int64_t t, const methodHandle& method) { 1020 // Skip update if counters are absent. 1021 // Can't allocate them since we are holding compile queue lock. 1022 if (method->method_counters() == nullptr) return; 1023 1024 if (is_old(method)) { 1025 // We don't remove old methods from the queue, 1026 // so we can just zero the rate. 1027 method->set_rate(0); 1028 return; 1029 } 1030 1031 // We don't update the rate if we've just came out of a safepoint. 1032 // delta_s is the time since last safepoint in milliseconds. 1033 int64_t delta_s = t - SafepointTracing::end_of_last_safepoint_ms(); 1034 int64_t delta_t = t - (method->prev_time() != 0 ? method->prev_time() : start_time()); // milliseconds since the last measurement 1035 // How many events were there since the last time? 1036 int event_count = method->invocation_count() + method->backedge_count(); 1037 int delta_e = event_count - method->prev_event_count(); 1038 1039 // We should be running for at least 1ms. 1040 if (delta_s >= TieredRateUpdateMinTime) { 1041 // And we must've taken the previous point at least 1ms before. 1042 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 1043 method->set_prev_time(t); 1044 method->set_prev_event_count(event_count); 1045 method->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 1046 } else { 1047 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 1048 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 1049 method->set_rate(0); 1050 } 1051 } 1052 } 1053 } 1054 1055 // Check if this method has been stale for a given number of milliseconds. 1056 // See select_task(). 1057 bool CompilationPolicy::is_stale(int64_t t, int64_t timeout, const methodHandle& method) { 1058 int64_t delta_s = t - SafepointTracing::end_of_last_safepoint_ms(); 1059 int64_t delta_t = t - method->prev_time(); 1060 if (delta_t > timeout && delta_s > timeout) { 1061 int event_count = method->invocation_count() + method->backedge_count(); 1062 int delta_e = event_count - method->prev_event_count(); 1063 // Return true if there were no events. 1064 return delta_e == 0; 1065 } 1066 return false; 1067 } 1068 1069 // We don't remove old methods from the compile queue even if they have 1070 // very low activity. See select_task(). 1071 bool CompilationPolicy::is_old(const methodHandle& method) { 1072 int i = method->invocation_count(); 1073 int b = method->backedge_count(); 1074 double k = TieredOldPercentage / 100.0; 1075 1076 return CallPredicate::apply_scaled(method, CompLevel_none, i, b, k) || LoopPredicate::apply_scaled(method, CompLevel_none, i, b, k); 1077 } 1078 1079 double CompilationPolicy::weight(Method* method) { 1080 return (double)(method->rate() + 1) * (method->invocation_count() + 1) * (method->backedge_count() + 1); 1081 } 1082 1083 // Apply heuristics and return true if x should be compiled before y 1084 bool CompilationPolicy::compare_methods(Method* x, Method* y) { 1085 if (x->highest_comp_level() > y->highest_comp_level()) { 1086 // recompilation after deopt 1087 return true; 1088 } else 1089 if (x->highest_comp_level() == y->highest_comp_level()) { 1090 if (weight(x) > weight(y)) { 1091 return true; 1092 } 1093 } 1094 return false; 1095 } 1096 1097 bool CompilationPolicy::compare_tasks(CompileTask* x, CompileTask* y) { 1098 assert(!x->is_aot() && !y->is_aot(), "AOT code caching tasks are not expected here"); 1099 if (x->compile_reason() != y->compile_reason() && y->compile_reason() == CompileTask::Reason_MustBeCompiled) { 1100 return true; 1101 } 1102 return false; 1103 } 1104 1105 // Is method profiled enough? 1106 bool CompilationPolicy::is_method_profiled(const methodHandle& method) { 1107 MethodData* mdo = method->method_data(); 1108 if (mdo != nullptr) { 1109 int i = mdo->invocation_count_delta(); 1110 int b = mdo->backedge_count_delta(); 1111 return CallPredicate::apply_scaled(method, CompLevel_full_profile, i, b, 1); 1112 } 1113 return false; 1114 } 1115 1116 1117 // Determine is a method is mature. 1118 bool CompilationPolicy::is_mature(MethodData* mdo) { 1119 if (Arguments::is_compiler_only()) { 1120 // Always report profiles as immature with -Xcomp 1121 return false; 1122 } 1123 methodHandle mh(Thread::current(), mdo->method()); 1124 if (mdo != nullptr) { 1125 int i = mdo->invocation_count(); 1126 int b = mdo->backedge_count(); 1127 double k = ProfileMaturityPercentage / 100.0; 1128 return CallPredicate::apply_scaled(mh, CompLevel_full_profile, i, b, k) || LoopPredicate::apply_scaled(mh, CompLevel_full_profile, i, b, k); 1129 } 1130 return false; 1131 } 1132 1133 // If a method is old enough and is still in the interpreter we would want to 1134 // start profiling without waiting for the compiled method to arrive. 1135 // We also take the load on compilers into the account. 1136 bool CompilationPolicy::should_create_mdo(const methodHandle& method, CompLevel cur_level) { 1137 if (cur_level != CompLevel_none || force_comp_at_level_simple(method) || CompilationModeFlag::quick_only() || !ProfileInterpreter) { 1138 return false; 1139 } 1140 1141 if (TrainingData::have_data()) { 1142 MethodTrainingData* mtd = MethodTrainingData::find_fast(method); 1143 if (mtd != nullptr && mtd->saw_level(CompLevel_full_optimization)) { 1144 return true; 1145 } 1146 } 1147 1148 if (is_old(method)) { 1149 return true; 1150 } 1151 1152 int i = method->invocation_count(); 1153 int b = method->backedge_count(); 1154 double k = Tier0ProfilingStartPercentage / 100.0; 1155 1156 // If the top level compiler is not keeping up, delay profiling. 1157 if (CompileBroker::queue_size(CompLevel_full_optimization) <= Tier0Delay * compiler_count(CompLevel_full_optimization)) { 1158 return CallPredicate::apply_scaled(method, CompLevel_none, i, b, k) || LoopPredicate::apply_scaled(method, CompLevel_none, i, b, k); 1159 } 1160 return false; 1161 } 1162 1163 // Inlining control: if we're compiling a profiled method with C1 and the callee 1164 // is known to have OSRed in a C2 version, don't inline it. 1165 bool CompilationPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 1166 CompLevel comp_level = (CompLevel)env->comp_level(); 1167 if (comp_level == CompLevel_full_profile || 1168 comp_level == CompLevel_limited_profile) { 1169 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 1170 } 1171 return false; 1172 } 1173 1174 // Create MDO if necessary. 1175 void CompilationPolicy::create_mdo(const methodHandle& mh, JavaThread* THREAD) { 1176 if (mh->is_native() || 1177 mh->is_abstract() || 1178 mh->is_accessor() || 1179 mh->is_constant_getter()) { 1180 return; 1181 } 1182 if (mh->method_data() == nullptr) { 1183 Method::build_profiling_method_data(mh, CHECK_AND_CLEAR); 1184 } 1185 if (ProfileInterpreter && THREAD->has_last_Java_frame()) { 1186 MethodData* mdo = mh->method_data(); 1187 if (mdo != nullptr) { 1188 frame last_frame = THREAD->last_frame(); 1189 if (last_frame.is_interpreted_frame() && mh == last_frame.interpreter_frame_method()) { 1190 int bci = last_frame.interpreter_frame_bci(); 1191 address dp = mdo->bci_to_dp(bci); 1192 last_frame.interpreter_frame_set_mdp(dp); 1193 } 1194 } 1195 } 1196 } 1197 1198 CompLevel CompilationPolicy::trained_transition_from_none(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1199 precond(mtd != nullptr); 1200 precond(cur_level == CompLevel_none); 1201 1202 if (mtd->only_inlined() && !mtd->saw_level(CompLevel_full_optimization)) { 1203 return CompLevel_none; 1204 } 1205 1206 bool training_has_profile = (mtd->final_profile() != nullptr); 1207 if (mtd->saw_level(CompLevel_full_optimization) && !training_has_profile) { 1208 return CompLevel_full_profile; 1209 } 1210 1211 CompLevel highest_training_level = static_cast<CompLevel>(mtd->highest_top_level()); 1212 switch (highest_training_level) { 1213 case CompLevel_limited_profile: 1214 case CompLevel_full_profile: 1215 return CompLevel_limited_profile; 1216 case CompLevel_simple: 1217 return CompLevel_simple; 1218 case CompLevel_none: 1219 return CompLevel_none; 1220 default: 1221 break; 1222 } 1223 1224 // Now handle the case of level 4. 1225 assert(highest_training_level == CompLevel_full_optimization, "Unexpected compilation level: %d", highest_training_level); 1226 if (!training_has_profile) { 1227 // The method was a part of a level 4 compile, but don't have a stored profile, 1228 // we need to profile it. 1229 return CompLevel_full_profile; 1230 } 1231 const bool deopt = (static_cast<CompLevel>(method->highest_comp_level()) == CompLevel_full_optimization); 1232 // If we deopted, then we reprofile 1233 if (deopt && !is_method_profiled(method)) { 1234 return CompLevel_full_profile; 1235 } 1236 1237 CompileTrainingData* ctd = mtd->last_toplevel_compile(CompLevel_full_optimization); 1238 assert(ctd != nullptr, "Should have CTD for CompLevel_full_optimization"); 1239 // With SkipTier2IfPossible and all deps satisfied, go to level 4 immediately 1240 if (SkipTier2IfPossible && ctd->init_deps_left() == 0) { 1241 if (method->method_data() == nullptr) { 1242 create_mdo(method, THREAD); 1243 } 1244 return CompLevel_full_optimization; 1245 } 1246 1247 // Otherwise go to level 2 1248 return CompLevel_limited_profile; 1249 } 1250 1251 1252 CompLevel CompilationPolicy::trained_transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1253 precond(mtd != nullptr); 1254 precond(cur_level == CompLevel_limited_profile); 1255 1256 // One of the main reasons that we can get here is that we're waiting for the stored C2 code to become ready. 1257 1258 // But first, check if we have a saved profile 1259 bool training_has_profile = (mtd->final_profile() != nullptr); 1260 if (!training_has_profile) { 1261 return CompLevel_full_profile; 1262 } 1263 1264 1265 assert(training_has_profile, "Have to have a profile to be here"); 1266 // Check if the method is ready 1267 CompileTrainingData* ctd = mtd->last_toplevel_compile(CompLevel_full_optimization); 1268 if (ctd != nullptr && ctd->init_deps_left() == 0) { 1269 if (method->method_data() == nullptr) { 1270 create_mdo(method, THREAD); 1271 } 1272 return CompLevel_full_optimization; 1273 } 1274 1275 // Otherwise stay at the current level 1276 return CompLevel_limited_profile; 1277 } 1278 1279 1280 CompLevel CompilationPolicy::trained_transition_from_full_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1281 precond(mtd != nullptr); 1282 precond(cur_level == CompLevel_full_profile); 1283 1284 CompLevel highest_training_level = static_cast<CompLevel>(mtd->highest_top_level()); 1285 // We have method at the full profile level and we also know that it's possibly an important method. 1286 if (highest_training_level == CompLevel_full_optimization && !mtd->only_inlined()) { 1287 // Check if it is adequately profiled 1288 if (is_method_profiled(method)) { 1289 return CompLevel_full_optimization; 1290 } 1291 } 1292 1293 // Otherwise stay at the current level 1294 return CompLevel_full_profile; 1295 } 1296 1297 CompLevel CompilationPolicy::trained_transition(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1298 precond(MethodTrainingData::have_data()); 1299 1300 // If there is no training data recorded for this method, bail out. 1301 if (mtd == nullptr) { 1302 return cur_level; 1303 } 1304 1305 CompLevel next_level = cur_level; 1306 switch(cur_level) { 1307 default: break; 1308 case CompLevel_none: 1309 next_level = trained_transition_from_none(method, cur_level, mtd, THREAD); 1310 break; 1311 case CompLevel_limited_profile: 1312 next_level = trained_transition_from_limited_profile(method, cur_level, mtd, THREAD); 1313 break; 1314 case CompLevel_full_profile: 1315 next_level = trained_transition_from_full_profile(method, cur_level, mtd, THREAD); 1316 break; 1317 } 1318 1319 // We don't have any special strategies for the C2-only compilation modes, so just fix up the levels for now. 1320 if (CompilationModeFlag::high_only_quick_internal() && CompLevel_simple < next_level && next_level < CompLevel_full_optimization) { 1321 return CompLevel_none; 1322 } 1323 if (CompilationModeFlag::high_only() && next_level < CompLevel_full_optimization) { 1324 return CompLevel_none; 1325 } 1326 return (cur_level != next_level) ? limit_level(next_level) : cur_level; 1327 } 1328 1329 /* 1330 * Method states: 1331 * 0 - interpreter (CompLevel_none) 1332 * 1 - pure C1 (CompLevel_simple) 1333 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 1334 * 3 - C1 with full profiling (CompLevel_full_profile) 1335 * 4 - C2 or Graal (CompLevel_full_optimization) 1336 * 1337 * Common state transition patterns: 1338 * a. 0 -> 3 -> 4. 1339 * The most common path. But note that even in this straightforward case 1340 * profiling can start at level 0 and finish at level 3. 1341 * 1342 * b. 0 -> 2 -> 3 -> 4. 1343 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning 1344 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 1345 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 1346 * 1347 * c. 0 -> (3->2) -> 4. 1348 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 1349 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 1350 * of the method to 2 while the request is still in-queue, because it'll allow it to run much faster 1351 * without full profiling while c2 is compiling. 1352 * 1353 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 1354 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 1355 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 1356 * 1357 * e. 0 -> 4. 1358 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 1359 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 1360 * the compiled version already exists). 1361 * 1362 * Note that since state 0 can be reached from any other state via deoptimization different loops 1363 * are possible. 1364 * 1365 */ 1366 1367 // Common transition function. Given a predicate determines if a method should transition to another level. 1368 template<typename Predicate> 1369 CompLevel CompilationPolicy::common(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD, bool disable_feedback) { 1370 CompLevel next_level = cur_level; 1371 1372 if (force_comp_at_level_simple(method)) { 1373 next_level = CompLevel_simple; 1374 } else if (is_trivial(method) || method->is_native()) { 1375 // We do not care if there is profiling data for these methods, throw them to compiler. 1376 next_level = CompilationModeFlag::disable_intermediate() ? CompLevel_full_optimization : CompLevel_simple; 1377 } else if (MethodTrainingData::have_data()) { 1378 MethodTrainingData* mtd = MethodTrainingData::find_fast(method); 1379 if (mtd == nullptr) { 1380 // We haven't see compilations of this method in training. It's either very cold or the behavior changed. 1381 // Feed it to the standard TF with no profiling delay. 1382 next_level = standard_transition<Predicate>(method, cur_level, false /*delay_profiling*/, disable_feedback); 1383 } else { 1384 next_level = trained_transition(method, cur_level, mtd, THREAD); 1385 if (cur_level == next_level) { 1386 // trained_transtion() is going to return the same level if no startup/warmup optimizations apply. 1387 // In order to catch possible pathologies due to behavior change we feed the event to the regular 1388 // TF but with profiling delay. 1389 next_level = standard_transition<Predicate>(method, cur_level, true /*delay_profiling*/, disable_feedback); 1390 } 1391 } 1392 } else { 1393 next_level = standard_transition<Predicate>(method, cur_level, false /*delay_profiling*/, disable_feedback); 1394 } 1395 return (next_level != cur_level) ? limit_level(next_level) : next_level; 1396 } 1397 1398 1399 template<typename Predicate> 1400 CompLevel CompilationPolicy::standard_transition(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback) { 1401 CompLevel next_level = cur_level; 1402 switch(cur_level) { 1403 default: break; 1404 case CompLevel_none: 1405 next_level = transition_from_none<Predicate>(method, cur_level, delay_profiling, disable_feedback); 1406 break; 1407 case CompLevel_limited_profile: 1408 next_level = transition_from_limited_profile<Predicate>(method, cur_level, delay_profiling, disable_feedback); 1409 break; 1410 case CompLevel_full_profile: 1411 next_level = transition_from_full_profile<Predicate>(method, cur_level); 1412 break; 1413 } 1414 return next_level; 1415 } 1416 1417 template<typename Predicate> 1418 CompLevel CompilationPolicy::transition_from_none(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback) { 1419 precond(cur_level == CompLevel_none); 1420 CompLevel next_level = cur_level; 1421 int i = method->invocation_count(); 1422 int b = method->backedge_count(); 1423 double scale = delay_profiling ? Tier0ProfileDelayFactor : 1.0; 1424 // If we were at full profile level, would we switch to full opt? 1425 if (transition_from_full_profile<Predicate>(method, CompLevel_full_profile) == CompLevel_full_optimization) { 1426 next_level = CompLevel_full_optimization; 1427 } else if (!CompilationModeFlag::disable_intermediate() && Predicate::apply_scaled(method, cur_level, i, b, scale)) { 1428 // C1-generated fully profiled code is about 30% slower than the limited profile 1429 // code that has only invocation and backedge counters. The observation is that 1430 // if C2 queue is large enough we can spend too much time in the fully profiled code 1431 // while waiting for C2 to pick the method from the queue. To alleviate this problem 1432 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 1433 // we choose to compile a limited profiled version and then recompile with full profiling 1434 // when the load on C2 goes down. 1435 if (delay_profiling || (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > Tier3DelayOn * compiler_count(CompLevel_full_optimization))) { 1436 next_level = CompLevel_limited_profile; 1437 } else { 1438 next_level = CompLevel_full_profile; 1439 } 1440 } 1441 return next_level; 1442 } 1443 1444 template<typename Predicate> 1445 CompLevel CompilationPolicy::transition_from_full_profile(const methodHandle& method, CompLevel cur_level) { 1446 precond(cur_level == CompLevel_full_profile); 1447 CompLevel next_level = cur_level; 1448 MethodData* mdo = method->method_data(); 1449 if (mdo != nullptr) { 1450 if (mdo->would_profile() || CompilationModeFlag::disable_intermediate()) { 1451 int mdo_i = mdo->invocation_count_delta(); 1452 int mdo_b = mdo->backedge_count_delta(); 1453 if (Predicate::apply(method, cur_level, mdo_i, mdo_b)) { 1454 next_level = CompLevel_full_optimization; 1455 } 1456 } else { 1457 next_level = CompLevel_full_optimization; 1458 } 1459 } 1460 return next_level; 1461 } 1462 1463 template<typename Predicate> 1464 CompLevel CompilationPolicy::transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback) { 1465 precond(cur_level == CompLevel_limited_profile); 1466 CompLevel next_level = cur_level; 1467 int i = method->invocation_count(); 1468 int b = method->backedge_count(); 1469 double scale = delay_profiling ? Tier2ProfileDelayFactor : 1.0; 1470 MethodData* mdo = method->method_data(); 1471 if (mdo != nullptr) { 1472 if (mdo->would_profile()) { 1473 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 1474 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 1475 Predicate::apply_scaled(method, cur_level, i, b, scale))) { 1476 next_level = CompLevel_full_profile; 1477 } 1478 } else { 1479 next_level = CompLevel_full_optimization; 1480 } 1481 } else { 1482 // If there is no MDO we need to profile 1483 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 1484 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 1485 Predicate::apply_scaled(method, cur_level, i, b, scale))) { 1486 next_level = CompLevel_full_profile; 1487 } 1488 } 1489 if (next_level == CompLevel_full_profile && is_method_profiled(method)) { 1490 next_level = CompLevel_full_optimization; 1491 } 1492 return next_level; 1493 } 1494 1495 1496 // Determine if a method should be compiled with a normal entry point at a different level. 1497 CompLevel CompilationPolicy::call_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD) { 1498 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), common<LoopPredicate>(method, cur_level, THREAD, true)); 1499 CompLevel next_level = common<CallPredicate>(method, cur_level, THREAD, !TrainingData::have_data() && is_old(method)); 1500 1501 // If OSR method level is greater than the regular method level, the levels should be 1502 // equalized by raising the regular method level in order to avoid OSRs during each 1503 // invocation of the method. 1504 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 1505 MethodData* mdo = method->method_data(); 1506 guarantee(mdo != nullptr, "MDO should not be nullptr"); 1507 if (mdo->invocation_count() >= 1) { 1508 next_level = CompLevel_full_optimization; 1509 } 1510 } else { 1511 next_level = MAX2(osr_level, next_level); 1512 } 1513 #if INCLUDE_JVMCI 1514 if (EnableJVMCI && UseJVMCICompiler && 1515 next_level == CompLevel_full_optimization CDS_ONLY(&& !AOTLinkedClassBulkLoader::class_preloading_finished())) { 1516 next_level = cur_level; 1517 } 1518 #endif 1519 return next_level; 1520 } 1521 1522 // Determine if we should do an OSR compilation of a given method. 1523 CompLevel CompilationPolicy::loop_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD) { 1524 CompLevel next_level = common<LoopPredicate>(method, cur_level, THREAD, true); 1525 if (cur_level == CompLevel_none) { 1526 // If there is a live OSR method that means that we deopted to the interpreter 1527 // for the transition. 1528 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 1529 if (osr_level > CompLevel_none) { 1530 return osr_level; 1531 } 1532 } 1533 return next_level; 1534 } 1535 1536 // Handle the invocation event. 1537 void CompilationPolicy::method_invocation_event(const methodHandle& mh, const methodHandle& imh, 1538 CompLevel level, nmethod* nm, TRAPS) { 1539 if (should_create_mdo(mh, level)) { 1540 create_mdo(mh, THREAD); 1541 } 1542 CompLevel next_level = call_event(mh, level, THREAD); 1543 if (next_level != level) { 1544 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 1545 compile(mh, InvocationEntryBci, next_level, THREAD); 1546 } 1547 } 1548 } 1549 1550 // Handle the back branch event. Notice that we can compile the method 1551 // with a regular entry from here. 1552 void CompilationPolicy::method_back_branch_event(const methodHandle& mh, const methodHandle& imh, 1553 int bci, CompLevel level, nmethod* nm, TRAPS) { 1554 if (should_create_mdo(mh, level)) { 1555 create_mdo(mh, THREAD); 1556 } 1557 // Check if MDO should be created for the inlined method 1558 if (should_create_mdo(imh, level)) { 1559 create_mdo(imh, THREAD); 1560 } 1561 1562 if (is_compilation_enabled()) { 1563 CompLevel next_osr_level = loop_event(imh, level, THREAD); 1564 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 1565 // At the very least compile the OSR version 1566 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 1567 compile(imh, bci, next_osr_level, CHECK); 1568 } 1569 1570 // Use loop event as an opportunity to also check if there's been 1571 // enough calls. 1572 CompLevel cur_level, next_level; 1573 if (mh() != imh()) { // If there is an enclosing method 1574 { 1575 guarantee(nm != nullptr, "Should have nmethod here"); 1576 cur_level = comp_level(mh()); 1577 next_level = call_event(mh, cur_level, THREAD); 1578 1579 if (max_osr_level == CompLevel_full_optimization) { 1580 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 1581 bool make_not_entrant = false; 1582 if (nm->is_osr_method()) { 1583 // This is an osr method, just make it not entrant and recompile later if needed 1584 make_not_entrant = true; 1585 } else { 1586 if (next_level != CompLevel_full_optimization) { 1587 // next_level is not full opt, so we need to recompile the 1588 // enclosing method without the inlinee 1589 cur_level = CompLevel_none; 1590 make_not_entrant = true; 1591 } 1592 } 1593 if (make_not_entrant) { 1594 if (PrintTieredEvents) { 1595 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 1596 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 1597 } 1598 nm->make_not_entrant("OSR invalidation, back branch"); 1599 } 1600 } 1601 // Fix up next_level if necessary to avoid deopts 1602 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 1603 next_level = CompLevel_full_profile; 1604 } 1605 if (cur_level != next_level) { 1606 if (!CompileBroker::compilation_is_in_queue(mh)) { 1607 compile(mh, InvocationEntryBci, next_level, THREAD); 1608 } 1609 } 1610 } 1611 } else { 1612 cur_level = comp_level(mh()); 1613 next_level = call_event(mh, cur_level, THREAD); 1614 if (next_level != cur_level) { 1615 if (!CompileBroker::compilation_is_in_queue(mh)) { 1616 compile(mh, InvocationEntryBci, next_level, THREAD); 1617 } 1618 } 1619 } 1620 } 1621 } 1622