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::_ac_count = 0; 61 double CompilationPolicy::_increase_threshold_at_ratio = 0; 62 63 CompilationPolicy::TrainingReplayQueue CompilationPolicy::_training_replay_queue; 64 65 void compilationPolicy_init() { 66 CompilationPolicy::initialize(); 67 } 68 69 int CompilationPolicy::compiler_count(CompLevel comp_level) { 70 if (is_c1_compile(comp_level)) { 71 return c1_count(); 72 } else if (is_c2_compile(comp_level)) { 73 return c2_count(); 74 } 75 return 0; 76 } 77 78 // Returns true if m must be compiled before executing it 79 // This is intended to force compiles for methods (usually for 80 // debugging) that would otherwise be interpreted for some reason. 81 bool CompilationPolicy::must_be_compiled(const methodHandle& m, int comp_level) { 82 // Don't allow Xcomp to cause compiles in replay mode 83 if (ReplayCompiles) return false; 84 85 if (m->has_compiled_code()) return false; // already compiled 86 if (!can_be_compiled(m, comp_level)) return false; 87 88 return !UseInterpreter || // must compile all methods 89 (AlwaysCompileLoopMethods && m->has_loops() && CompileBroker::should_compile_new_jobs()); // eagerly compile loop methods 90 } 91 92 void CompilationPolicy::maybe_compile_early(const methodHandle& m, TRAPS) { 93 if (m->method_holder()->is_not_initialized()) { 94 // 'is_not_initialized' means not only '!is_initialized', but also that 95 // initialization has not been started yet ('!being_initialized') 96 // Do not force compilation of methods in uninitialized classes. 97 return; 98 } 99 if (!m->is_native() && MethodTrainingData::have_data()) { 100 MethodTrainingData* mtd = MethodTrainingData::find_fast(m); 101 if (mtd == nullptr) { 102 return; // there is no training data recorded for m 103 } 104 // AOT Preload code with class init barriers is used, 105 // consider replacing it with normal (faster) AOT code 106 bool recompile = m->code_has_clinit_barriers(); 107 CompLevel cur_level = static_cast<CompLevel>(m->highest_comp_level()); 108 CompLevel next_level = trained_transition(m, cur_level, mtd, THREAD); 109 if ((next_level != cur_level || recompile) && can_be_compiled(m, next_level) && !CompileBroker::compilation_is_in_queue(m)) { 110 bool requires_online_compilation = false; 111 CompileTrainingData* ctd = mtd->last_toplevel_compile(next_level); 112 if (ctd != nullptr) { 113 // Can't load normal AOT code - not all dependancies are ready, 114 // request normal compilation 115 requires_online_compilation = (ctd->init_deps_left_acquire() > 0); 116 } 117 if (requires_online_compilation && recompile) { 118 // Wait when dependencies are ready to load normal AOT code 119 // if AOT Preload code is used now. 120 // 121 // FIXME. We may never (or it take long time) get all dependencies 122 // be ready to replace AOT Preload code. Consider using time and how many 123 // dependencies left to allow normal JIT compilation for replacement. 124 return; 125 } 126 if (PrintTieredEvents) { 127 print_event(FORCE_COMPILE, m(), m(), InvocationEntryBci, next_level); 128 } 129 CompileBroker::compile_method(m, InvocationEntryBci, next_level, 0, requires_online_compilation, CompileTask::Reason_MustBeCompiled, THREAD); 130 if (HAS_PENDING_EXCEPTION) { 131 CLEAR_PENDING_EXCEPTION; 132 } 133 } 134 } 135 } 136 137 void CompilationPolicy::compile_if_required(const methodHandle& m, TRAPS) { 138 if (!THREAD->can_call_java() || THREAD->is_Compiler_thread()) { 139 // don't force compilation, resolve was on behalf of compiler 140 return; 141 } 142 if (m->method_holder()->is_not_initialized()) { 143 // 'is_not_initialized' means not only '!is_initialized', but also that 144 // initialization has not been started yet ('!being_initialized') 145 // Do not force compilation of methods in uninitialized classes. 146 // Note that doing this would throw an assert later, 147 // in CompileBroker::compile_method. 148 // We sometimes use the link resolver to do reflective lookups 149 // even before classes are initialized. 150 return; 151 } 152 153 if (must_be_compiled(m)) { 154 // This path is unusual, mostly used by the '-Xcomp' stress test mode. 155 CompLevel level = initial_compile_level(m); 156 if (PrintTieredEvents) { 157 print_event(FORCE_COMPILE, m(), m(), InvocationEntryBci, level); 158 } 159 // Test AOT code too 160 bool requires_online_compilation = false; 161 if (TrainingData::have_data()) { 162 MethodTrainingData* mtd = MethodTrainingData::find_fast(m); 163 if (mtd != nullptr) { 164 CompileTrainingData* ctd = mtd->last_toplevel_compile(level); 165 if (ctd != nullptr) { 166 requires_online_compilation = (ctd->init_deps_left_acquire() > 0); 167 } 168 } 169 } 170 CompileBroker::compile_method(m, InvocationEntryBci, level, 0, requires_online_compilation, CompileTask::Reason_MustBeCompiled, THREAD); 171 } 172 } 173 174 void CompilationPolicy::replay_training_at_init_impl(InstanceKlass* klass, JavaThread* current) { 175 if (!klass->has_init_deps_processed()) { 176 ResourceMark rm; 177 log_debug(training)("Replay training: %s", klass->external_name()); 178 179 KlassTrainingData* ktd = KlassTrainingData::find(klass); 180 if (ktd != nullptr) { 181 guarantee(ktd->has_holder(), ""); 182 ktd->notice_fully_initialized(); // sets klass->has_init_deps_processed bit 183 assert(klass->has_init_deps_processed(), ""); 184 185 if (AOTCompileEagerly) { 186 ktd->iterate_comp_deps([&](CompileTrainingData* ctd) { 187 if (ctd->init_deps_left_acquire() == 0) { 188 MethodTrainingData* mtd = ctd->method(); 189 if (mtd->has_holder()) { 190 const methodHandle mh(current, const_cast<Method*>(mtd->holder())); 191 CompilationPolicy::maybe_compile_early(mh, current); 192 } 193 } 194 }); 195 } 196 } 197 } 198 } 199 200 void CompilationPolicy::replay_training_at_init(InstanceKlass* klass, JavaThread* current) { 201 assert(klass->is_initialized(), ""); 202 if (TrainingData::have_data() && klass->in_aot_cache()) { 203 _training_replay_queue.push(klass, TrainingReplayQueue_lock, current); 204 } 205 } 206 207 // For TrainingReplayQueue 208 template<> 209 void CompilationPolicyUtils::Queue<InstanceKlass>::print_on(outputStream* st) { 210 int pos = 0; 211 for (QueueNode* cur = _head; cur != nullptr; cur = cur->next()) { 212 ResourceMark rm; 213 InstanceKlass* ik = cur->value(); 214 st->print_cr("%3d: " INTPTR_FORMAT " %s", ++pos, p2i(ik), ik->external_name()); 215 } 216 } 217 218 void CompilationPolicy::replay_training_at_init_loop(JavaThread* current) { 219 while (!CompileBroker::is_compilation_disabled_forever()) { 220 InstanceKlass* ik = _training_replay_queue.pop(TrainingReplayQueue_lock, current); 221 if (ik != nullptr) { 222 replay_training_at_init_impl(ik, current); 223 } 224 } 225 } 226 227 static inline CompLevel adjust_level_for_compilability_query(CompLevel comp_level) { 228 if (comp_level == CompLevel_any) { 229 if (CompilerConfig::is_c1_only()) { 230 comp_level = CompLevel_simple; 231 } else if (CompilerConfig::is_c2_or_jvmci_compiler_only()) { 232 comp_level = CompLevel_full_optimization; 233 } 234 } 235 return comp_level; 236 } 237 238 // Returns true if m is allowed to be compiled 239 bool CompilationPolicy::can_be_compiled(const methodHandle& m, int comp_level) { 240 // allow any levels for WhiteBox 241 assert(WhiteBoxAPI || comp_level == CompLevel_any || is_compile(comp_level), "illegal compilation level %d", comp_level); 242 243 if (m->is_abstract()) return false; 244 if (DontCompileHugeMethods && m->code_size() > HugeMethodLimit) return false; 245 246 // Math intrinsics should never be compiled as this can lead to 247 // monotonicity problems because the interpreter will prefer the 248 // compiled code to the intrinsic version. This can't happen in 249 // production because the invocation counter can't be incremented 250 // but we shouldn't expose the system to this problem in testing 251 // modes. 252 if (!AbstractInterpreter::can_be_compiled(m)) { 253 return false; 254 } 255 comp_level = adjust_level_for_compilability_query((CompLevel) comp_level); 256 if (comp_level == CompLevel_any || is_compile(comp_level)) { 257 return !m->is_not_compilable(comp_level); 258 } 259 return false; 260 } 261 262 // Returns true if m is allowed to be osr compiled 263 bool CompilationPolicy::can_be_osr_compiled(const methodHandle& m, int comp_level) { 264 bool result = false; 265 comp_level = adjust_level_for_compilability_query((CompLevel) comp_level); 266 if (comp_level == CompLevel_any || is_compile(comp_level)) { 267 result = !m->is_not_osr_compilable(comp_level); 268 } 269 return (result && can_be_compiled(m, comp_level)); 270 } 271 272 bool CompilationPolicy::is_compilation_enabled() { 273 // NOTE: CompileBroker::should_compile_new_jobs() checks for UseCompiler 274 return CompileBroker::should_compile_new_jobs(); 275 } 276 277 CompileTask* CompilationPolicy::select_task_helper(CompileQueue* compile_queue) { 278 // Remove unloaded methods from the queue 279 for (CompileTask* task = compile_queue->first(); task != nullptr; ) { 280 CompileTask* next = task->next(); 281 if (task->is_unloaded()) { 282 compile_queue->remove_and_mark_stale(task); 283 } 284 task = next; 285 } 286 #if INCLUDE_JVMCI 287 if (UseJVMCICompiler && !BackgroundCompilation) { 288 /* 289 * In blocking compilation mode, the CompileBroker will make 290 * compilations submitted by a JVMCI compiler thread non-blocking. These 291 * compilations should be scheduled after all blocking compilations 292 * to service non-compiler related compilations sooner and reduce the 293 * chance of such compilations timing out. 294 */ 295 for (CompileTask* task = compile_queue->first(); task != nullptr; task = task->next()) { 296 if (task->is_blocking()) { 297 return task; 298 } 299 } 300 } 301 #endif 302 return compile_queue->first(); 303 } 304 305 // Simple methods are as good being compiled with C1 as C2. 306 // Determine if a given method is such a case. 307 bool CompilationPolicy::is_trivial(const methodHandle& method) { 308 if (method->is_accessor() || 309 method->is_constant_getter()) { 310 return true; 311 } 312 return false; 313 } 314 315 bool CompilationPolicy::force_comp_at_level_simple(const methodHandle& method) { 316 if (CompilationModeFlag::quick_internal()) { 317 #if INCLUDE_JVMCI 318 if (UseJVMCICompiler) { 319 AbstractCompiler* comp = CompileBroker::compiler(CompLevel_full_optimization); 320 if (comp != nullptr && comp->is_jvmci() && ((JVMCICompiler*) comp)->force_comp_at_level_simple(method)) { 321 return true; 322 } 323 } 324 #endif 325 } 326 return false; 327 } 328 329 CompLevel CompilationPolicy::comp_level(Method* method) { 330 nmethod *nm = method->code(); 331 if (nm != nullptr && nm->is_in_use()) { 332 return (CompLevel)nm->comp_level(); 333 } 334 return CompLevel_none; 335 } 336 337 // Call and loop predicates determine whether a transition to a higher 338 // compilation level should be performed (pointers to predicate functions 339 // are passed to common()). 340 // Tier?LoadFeedback is basically a coefficient that determines of 341 // how many methods per compiler thread can be in the queue before 342 // the threshold values double. 343 class LoopPredicate : AllStatic { 344 public: 345 static bool apply_scaled(const methodHandle& method, CompLevel cur_level, int i, int b, double scale) { 346 double threshold_scaling; 347 if (CompilerOracle::has_option_value(method, CompileCommandEnum::CompileThresholdScaling, threshold_scaling)) { 348 scale *= threshold_scaling; 349 } 350 switch(cur_level) { 351 case CompLevel_none: 352 case CompLevel_limited_profile: 353 return b >= Tier3BackEdgeThreshold * scale; 354 case CompLevel_full_profile: 355 return b >= Tier4BackEdgeThreshold * scale; 356 default: 357 return true; 358 } 359 } 360 361 static bool apply(const methodHandle& method, CompLevel cur_level, int i, int b) { 362 double k = 1; 363 switch(cur_level) { 364 case CompLevel_none: 365 // Fall through 366 case CompLevel_limited_profile: { 367 k = CompilationPolicy::threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 368 break; 369 } 370 case CompLevel_full_profile: { 371 k = CompilationPolicy::threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 372 break; 373 } 374 default: 375 return true; 376 } 377 return apply_scaled(method, cur_level, i, b, k); 378 } 379 }; 380 381 class CallPredicate : AllStatic { 382 public: 383 static bool apply_scaled(const methodHandle& method, CompLevel cur_level, int i, int b, double scale) { 384 double threshold_scaling; 385 if (CompilerOracle::has_option_value(method, CompileCommandEnum::CompileThresholdScaling, threshold_scaling)) { 386 scale *= threshold_scaling; 387 } 388 switch(cur_level) { 389 case CompLevel_none: 390 case CompLevel_limited_profile: 391 return (i >= Tier3InvocationThreshold * scale) || 392 (i >= Tier3MinInvocationThreshold * scale && i + b >= Tier3CompileThreshold * scale); 393 case CompLevel_full_profile: 394 return (i >= Tier4InvocationThreshold * scale) || 395 (i >= Tier4MinInvocationThreshold * scale && i + b >= Tier4CompileThreshold * scale); 396 default: 397 return true; 398 } 399 } 400 401 static bool apply(const methodHandle& method, CompLevel cur_level, int i, int b) { 402 double k = 1; 403 switch(cur_level) { 404 case CompLevel_none: 405 case CompLevel_limited_profile: { 406 k = CompilationPolicy::threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 407 break; 408 } 409 case CompLevel_full_profile: { 410 k = CompilationPolicy::threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 411 break; 412 } 413 default: 414 return true; 415 } 416 return apply_scaled(method, cur_level, i, b, k); 417 } 418 }; 419 420 double CompilationPolicy::threshold_scale(CompLevel level, int feedback_k) { 421 int comp_count = compiler_count(level); 422 if (comp_count > 0 && feedback_k > 0) { 423 double queue_size = CompileBroker::queue_size(level); 424 double k = (double)queue_size / ((double)feedback_k * (double)comp_count) + 1; 425 426 // Increase C1 compile threshold when the code cache is filled more 427 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 428 // The main intention is to keep enough free space for C2 compiled code 429 // to achieve peak performance if the code cache is under stress. 430 if (CompilerConfig::is_tiered() && !CompilationModeFlag::disable_intermediate() && is_c1_compile(level)) { 431 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(); 432 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 433 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 434 } 435 } 436 return k; 437 } 438 return 1; 439 } 440 441 void CompilationPolicy::print_counters_on(outputStream* st, const char* prefix, Method* m) { 442 int invocation_count = m->invocation_count(); 443 int backedge_count = m->backedge_count(); 444 MethodData* mdh = m->method_data(); 445 int mdo_invocations = 0, mdo_backedges = 0; 446 int mdo_invocations_start = 0, mdo_backedges_start = 0; 447 if (mdh != nullptr) { 448 mdo_invocations = mdh->invocation_count(); 449 mdo_backedges = mdh->backedge_count(); 450 mdo_invocations_start = mdh->invocation_count_start(); 451 mdo_backedges_start = mdh->backedge_count_start(); 452 } 453 st->print(" %stotal=%d,%d %smdo=%d(%d),%d(%d)", prefix, 454 invocation_count, backedge_count, prefix, 455 mdo_invocations, mdo_invocations_start, 456 mdo_backedges, mdo_backedges_start); 457 st->print(" %smax levels=%d,%d", prefix, m->highest_comp_level(), m->highest_osr_comp_level()); 458 } 459 460 void CompilationPolicy::print_training_data_on(outputStream* st, const char* prefix, Method* method) { 461 methodHandle m(Thread::current(), method); 462 st->print(" %smtd: ", prefix); 463 MethodTrainingData* mtd = MethodTrainingData::find(m); 464 if (mtd == nullptr) { 465 st->print("null"); 466 } else { 467 MethodData* md = mtd->final_profile(); 468 st->print("mdo="); 469 if (md == nullptr) { 470 st->print("null"); 471 } else { 472 int mdo_invocations = md->invocation_count(); 473 int mdo_backedges = md->backedge_count(); 474 int mdo_invocations_start = md->invocation_count_start(); 475 int mdo_backedges_start = md->backedge_count_start(); 476 st->print("%d(%d), %d(%d)", mdo_invocations, mdo_invocations_start, mdo_backedges, mdo_backedges_start); 477 } 478 CompileTrainingData* ctd = mtd->last_toplevel_compile(CompLevel_full_optimization); 479 st->print(", deps="); 480 if (ctd == nullptr) { 481 st->print("null"); 482 } else { 483 st->print("%d", ctd->init_deps_left_acquire()); 484 } 485 } 486 } 487 488 // Print an event. 489 void CompilationPolicy::print_event_on(outputStream *st, EventType type, Method* m, Method* im, int bci, CompLevel level) { 490 bool inlinee_event = m != im; 491 492 st->print("%lf: [", os::elapsedTime()); 493 494 switch(type) { 495 case CALL: 496 st->print("call"); 497 break; 498 case LOOP: 499 st->print("loop"); 500 break; 501 case COMPILE: 502 st->print("compile"); 503 break; 504 case FORCE_COMPILE: 505 st->print("force-compile"); 506 break; 507 case FORCE_RECOMPILE: 508 st->print("force-recompile"); 509 break; 510 case REMOVE_FROM_QUEUE: 511 st->print("remove-from-queue"); 512 break; 513 case UPDATE_IN_QUEUE: 514 st->print("update-in-queue"); 515 break; 516 case REPROFILE: 517 st->print("reprofile"); 518 break; 519 case MAKE_NOT_ENTRANT: 520 st->print("make-not-entrant"); 521 break; 522 default: 523 st->print("unknown"); 524 } 525 526 st->print(" level=%d ", level); 527 528 ResourceMark rm; 529 char *method_name = m->name_and_sig_as_C_string(); 530 st->print("[%s", method_name); 531 if (inlinee_event) { 532 char *inlinee_name = im->name_and_sig_as_C_string(); 533 st->print(" [%s]] ", inlinee_name); 534 } 535 else st->print("] "); 536 st->print("@%d queues=%d,%d", bci, CompileBroker::queue_size(CompLevel_full_profile), 537 CompileBroker::queue_size(CompLevel_full_optimization)); 538 539 st->print(" rate="); 540 if (m->prev_time() == 0) st->print("n/a"); 541 else st->print("%f", m->rate()); 542 543 RecompilationPolicy::print_load_average(st); 544 545 st->print(" k=%.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback), 546 threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback)); 547 548 if (type != COMPILE) { 549 print_counters_on(st, "", m); 550 if (inlinee_event) { 551 print_counters_on(st, "inlinee ", im); 552 } 553 st->print(" compilable="); 554 bool need_comma = false; 555 if (!m->is_not_compilable(CompLevel_full_profile)) { 556 st->print("c1"); 557 need_comma = true; 558 } 559 if (!m->is_not_osr_compilable(CompLevel_full_profile)) { 560 if (need_comma) st->print(","); 561 st->print("c1-osr"); 562 need_comma = true; 563 } 564 if (!m->is_not_compilable(CompLevel_full_optimization)) { 565 if (need_comma) st->print(","); 566 st->print("c2"); 567 need_comma = true; 568 } 569 if (!m->is_not_osr_compilable(CompLevel_full_optimization)) { 570 if (need_comma) st->print(","); 571 st->print("c2-osr"); 572 } 573 st->print(" status="); 574 if (m->queued_for_compilation()) { 575 st->print("in-queue"); 576 } else st->print("idle"); 577 578 print_training_data_on(st, "", m); 579 if (inlinee_event) { 580 print_training_data_on(st, "inlinee ", im); 581 } 582 } 583 st->print_cr("]"); 584 585 } 586 587 void CompilationPolicy::print_event(EventType type, Method* m, Method* im, int bci, CompLevel level) { 588 stringStream s; 589 print_event_on(&s, type, m, im, bci, level); 590 ResourceMark rm; 591 tty->print("%s", s.as_string()); 592 } 593 594 void CompilationPolicy::initialize() { 595 if (!CompilerConfig::is_interpreter_only()) { 596 int count = CICompilerCount; 597 bool c1_only = CompilerConfig::is_c1_only(); 598 bool c2_only = CompilerConfig::is_c2_or_jvmci_compiler_only(); 599 int min_count = (c1_only || c2_only) ? 1 : 2; 600 601 #ifdef _LP64 602 // Turn on ergonomic compiler count selection 603 if (AOTCodeCache::maybe_dumping_code()) { 604 // Assembly phase runs C1 and C2 compilation in separate phases, 605 // and can use all the CPU threads it can reach. Adjust the common 606 // options before policy starts overwriting them. 607 FLAG_SET_ERGO_IF_DEFAULT(UseDynamicNumberOfCompilerThreads, false); 608 FLAG_SET_ERGO_IF_DEFAULT(CICompilerCountPerCPU, false); 609 if (FLAG_IS_DEFAULT(CICompilerCount)) { 610 count = MAX2(count, os::active_processor_count()); 611 } 612 } 613 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { 614 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); 615 } 616 if (CICompilerCountPerCPU) { 617 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n 618 int log_cpu = log2i(os::active_processor_count()); 619 int loglog_cpu = log2i(MAX2(log_cpu, 1)); 620 count = MAX2(log_cpu * loglog_cpu * 3 / 2, min_count); 621 } 622 if (FLAG_IS_DEFAULT(CICompilerCount)) { 623 // Make sure there is enough space in the code cache to hold all the compiler buffers 624 size_t c1_size = 0; 625 #ifdef COMPILER1 626 c1_size = Compiler::code_buffer_size(); 627 #endif 628 size_t c2_size = 0; 629 #ifdef COMPILER2 630 c2_size = C2Compiler::initial_code_buffer_size(); 631 #endif 632 size_t buffer_size = 0; 633 if (c1_only) { 634 buffer_size = c1_size; 635 } else if (c2_only) { 636 buffer_size = c2_size; 637 } else { 638 buffer_size = c1_size / 3 + 2 * c2_size / 3; 639 } 640 size_t max_count = (NonNMethodCodeHeapSize - (CodeCacheMinimumUseSpace DEBUG_ONLY(* 3))) / buffer_size; 641 if ((size_t)count > max_count) { 642 // Lower the compiler count such that all buffers fit into the code cache 643 count = MAX2((int)max_count, min_count); 644 } 645 FLAG_SET_ERGO(CICompilerCount, count); 646 } 647 #else 648 // On 32-bit systems, the number of compiler threads is limited to 3. 649 // On these systems, the virtual address space available to the JVM 650 // is usually limited to 2-4 GB (the exact value depends on the platform). 651 // As the compilers (especially C2) can consume a large amount of 652 // memory, scaling the number of compiler threads with the number of 653 // available cores can result in the exhaustion of the address space 654 /// available to the VM and thus cause the VM to crash. 655 if (FLAG_IS_DEFAULT(CICompilerCount)) { 656 count = 3; 657 FLAG_SET_ERGO(CICompilerCount, count); 658 } 659 #endif // _LP64 660 661 if (c1_only) { 662 // No C2 compiler threads are needed 663 set_c1_count(count); 664 } else if (c2_only) { 665 // No C1 compiler threads are needed 666 set_c2_count(count); 667 } else { 668 #if INCLUDE_JVMCI 669 if (UseJVMCICompiler && UseJVMCINativeLibrary) { 670 int libjvmci_count = MAX2((int) (count * JVMCINativeLibraryThreadFraction), 1); 671 int c1_count = MAX2(count - libjvmci_count, 1); 672 set_c2_count(libjvmci_count); 673 set_c1_count(c1_count); 674 } else 675 #endif 676 { 677 set_c1_count(MAX2(count / 3, 1)); 678 set_c2_count(MAX2(count - c1_count(), 1)); 679 } 680 } 681 if (AOTCodeCache::is_code_load_thread_on()) { 682 set_ac_count((c1_only || c2_only) ? 1 : 2); // At minimum we need 2 threads to load C1 and C2 AOT code in parallel 683 } 684 assert(count == c1_count() + c2_count(), "inconsistent compiler thread count"); 685 set_increase_threshold_at_ratio(); 686 } else { 687 // Interpreter mode creates no compilers 688 FLAG_SET_ERGO(CICompilerCount, 0); 689 } 690 set_start_time(nanos_to_millis(os::javaTimeNanos())); 691 } 692 693 694 #ifdef ASSERT 695 bool CompilationPolicy::verify_level(CompLevel level) { 696 if (TieredCompilation && level > TieredStopAtLevel) { 697 return false; 698 } 699 // Check if there is a compiler to process the requested level 700 if (!CompilerConfig::is_c1_enabled() && is_c1_compile(level)) { 701 return false; 702 } 703 if (!CompilerConfig::is_c2_or_jvmci_compiler_enabled() && is_c2_compile(level)) { 704 return false; 705 } 706 707 // Interpreter level is always valid. 708 if (level == CompLevel_none) { 709 return true; 710 } 711 if (CompilationModeFlag::normal()) { 712 return true; 713 } else if (CompilationModeFlag::quick_only()) { 714 return level == CompLevel_simple; 715 } else if (CompilationModeFlag::high_only()) { 716 return level == CompLevel_full_optimization; 717 } else if (CompilationModeFlag::high_only_quick_internal()) { 718 return level == CompLevel_full_optimization || level == CompLevel_simple; 719 } 720 return false; 721 } 722 #endif 723 724 725 CompLevel CompilationPolicy::highest_compile_level() { 726 CompLevel level = CompLevel_none; 727 // Setup the maximum level available for the current compiler configuration. 728 if (!CompilerConfig::is_interpreter_only()) { 729 if (CompilerConfig::is_c2_or_jvmci_compiler_enabled()) { 730 level = CompLevel_full_optimization; 731 } else if (CompilerConfig::is_c1_enabled()) { 732 if (CompilerConfig::is_c1_simple_only()) { 733 level = CompLevel_simple; 734 } else { 735 level = CompLevel_full_profile; 736 } 737 } 738 } 739 // Clamp the maximum level with TieredStopAtLevel. 740 if (TieredCompilation) { 741 level = MIN2(level, (CompLevel) TieredStopAtLevel); 742 } 743 744 // Fix it up if after the clamping it has become invalid. 745 // Bring it monotonically down depending on the next available level for 746 // the compilation mode. 747 if (!CompilationModeFlag::normal()) { 748 // a) quick_only - levels 2,3,4 are invalid; levels -1,0,1 are valid; 749 // b) high_only - levels 1,2,3 are invalid; levels -1,0,4 are valid; 750 // c) high_only_quick_internal - levels 2,3 are invalid; levels -1,0,1,4 are valid. 751 if (CompilationModeFlag::quick_only()) { 752 if (level == CompLevel_limited_profile || level == CompLevel_full_profile || level == CompLevel_full_optimization) { 753 level = CompLevel_simple; 754 } 755 } else if (CompilationModeFlag::high_only()) { 756 if (level == CompLevel_simple || level == CompLevel_limited_profile || level == CompLevel_full_profile) { 757 level = CompLevel_none; 758 } 759 } else if (CompilationModeFlag::high_only_quick_internal()) { 760 if (level == CompLevel_limited_profile || level == CompLevel_full_profile) { 761 level = CompLevel_simple; 762 } 763 } 764 } 765 766 assert(verify_level(level), "Invalid highest compilation level: %d", level); 767 return level; 768 } 769 770 CompLevel CompilationPolicy::limit_level(CompLevel level) { 771 level = MIN2(level, highest_compile_level()); 772 assert(verify_level(level), "Invalid compilation level: %d", level); 773 return level; 774 } 775 776 CompLevel CompilationPolicy::initial_compile_level(const methodHandle& method) { 777 CompLevel level = CompLevel_any; 778 if (CompilationModeFlag::normal()) { 779 level = CompLevel_full_profile; 780 } else if (CompilationModeFlag::quick_only()) { 781 level = CompLevel_simple; 782 } else if (CompilationModeFlag::high_only()) { 783 level = CompLevel_full_optimization; 784 } else if (CompilationModeFlag::high_only_quick_internal()) { 785 if (force_comp_at_level_simple(method)) { 786 level = CompLevel_simple; 787 } else { 788 level = CompLevel_full_optimization; 789 } 790 } 791 assert(level != CompLevel_any, "Unhandled compilation mode"); 792 return limit_level(level); 793 } 794 795 // Set carry flags on the counters if necessary 796 void CompilationPolicy::handle_counter_overflow(const methodHandle& method) { 797 MethodCounters *mcs = method->method_counters(); 798 if (mcs != nullptr) { 799 mcs->invocation_counter()->set_carry_on_overflow(); 800 mcs->backedge_counter()->set_carry_on_overflow(); 801 } 802 MethodData* mdo = method->method_data(); 803 if (mdo != nullptr) { 804 mdo->invocation_counter()->set_carry_on_overflow(); 805 mdo->backedge_counter()->set_carry_on_overflow(); 806 } 807 } 808 809 // Called with the queue locked and with at least one element 810 CompileTask* CompilationPolicy::select_task(CompileQueue* compile_queue, JavaThread* THREAD) { 811 CompileTask *max_blocking_task = nullptr; 812 CompileTask *max_task = nullptr; 813 Method* max_method = nullptr; 814 815 int64_t t = nanos_to_millis(os::javaTimeNanos()); 816 // Iterate through the queue and find a method with a maximum rate. 817 for (CompileTask* task = compile_queue->first(); task != nullptr;) { 818 CompileTask* next_task = task->next(); 819 // If a method was unloaded or has been stale for some time, remove it from the queue. 820 // Blocking tasks and tasks submitted from whitebox API don't become stale 821 if (task->is_unloaded()) { 822 compile_queue->remove_and_mark_stale(task); 823 task = next_task; 824 continue; 825 } 826 if (task->is_aot_load()) { 827 // AOTCodeCache tasks are on separate queue, and they should load fast. There is no need to walk 828 // the rest of the queue, just take the task and go. 829 return task; 830 } 831 if (task->is_blocking() && task->compile_reason() == CompileTask::Reason_Whitebox) { 832 // CTW tasks, submitted as blocking Whitebox requests, do not participate in rate 833 // selection and/or any level adjustments. Just return them in order. 834 return task; 835 } 836 Method* method = task->method(); 837 methodHandle mh(THREAD, method); 838 if (task->can_become_stale() && is_stale(t, TieredCompileTaskTimeout, mh) && !is_old(mh)) { 839 if (PrintTieredEvents) { 840 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel) task->comp_level()); 841 } 842 method->clear_queued_for_compilation(); 843 method->set_pending_queue_processed(false); 844 compile_queue->remove_and_mark_stale(task); 845 task = next_task; 846 continue; 847 } 848 update_rate(t, mh); 849 if (max_task == nullptr || compare_methods(method, max_method) || compare_tasks(task, max_task)) { 850 // Select a method with the highest rate 851 max_task = task; 852 max_method = method; 853 } 854 855 if (task->is_blocking()) { 856 if (max_blocking_task == nullptr || compare_methods(method, max_blocking_task->method())) { 857 max_blocking_task = task; 858 } 859 } 860 861 task = next_task; 862 } 863 864 if (max_blocking_task != nullptr) { 865 // In blocking compilation mode, the CompileBroker will make 866 // compilations submitted by a JVMCI compiler thread non-blocking. These 867 // compilations should be scheduled after all blocking compilations 868 // to service non-compiler related compilations sooner and reduce the 869 // chance of such compilations timing out. 870 max_task = max_blocking_task; 871 max_method = max_task->method(); 872 } 873 874 methodHandle max_method_h(THREAD, max_method); 875 876 if (max_task != nullptr && max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile && 877 max_method != nullptr && is_method_profiled(max_method_h) && !Arguments::is_compiler_only()) { 878 max_task->set_comp_level(CompLevel_limited_profile); 879 880 if (CompileBroker::compilation_is_complete(max_method_h(), max_task->osr_bci(), CompLevel_limited_profile, 881 false /* requires_online_compilation */, 882 CompileTask::Reason_None)) { 883 if (PrintTieredEvents) { 884 print_event(REMOVE_FROM_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 885 } 886 compile_queue->remove_and_mark_stale(max_task); 887 max_method->clear_queued_for_compilation(); 888 return nullptr; 889 } 890 891 if (PrintTieredEvents) { 892 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 893 } 894 } 895 896 return max_task; 897 } 898 899 void CompilationPolicy::reprofile(ScopeDesc* trap_scope, bool is_osr) { 900 for (ScopeDesc* sd = trap_scope;; sd = sd->sender()) { 901 if (PrintTieredEvents) { 902 print_event(REPROFILE, sd->method(), sd->method(), InvocationEntryBci, CompLevel_none); 903 } 904 MethodData* mdo = sd->method()->method_data(); 905 if (mdo != nullptr) { 906 mdo->reset_start_counters(); 907 } 908 if (sd->is_top()) break; 909 } 910 } 911 912 nmethod* CompilationPolicy::event(const methodHandle& method, const methodHandle& inlinee, 913 int branch_bci, int bci, CompLevel comp_level, nmethod* nm, TRAPS) { 914 if (PrintTieredEvents) { 915 print_event(bci == InvocationEntryBci ? CALL : LOOP, method(), inlinee(), bci, comp_level); 916 } 917 918 if (comp_level == CompLevel_none && 919 JvmtiExport::can_post_interpreter_events() && 920 THREAD->is_interp_only_mode()) { 921 return nullptr; 922 } 923 if (ReplayCompiles) { 924 // Don't trigger other compiles in testing mode 925 return nullptr; 926 } 927 928 handle_counter_overflow(method); 929 if (method() != inlinee()) { 930 handle_counter_overflow(inlinee); 931 } 932 933 if (bci == InvocationEntryBci) { 934 method_invocation_event(method, inlinee, comp_level, nm, THREAD); 935 } else { 936 // method == inlinee if the event originated in the main method 937 method_back_branch_event(method, inlinee, bci, comp_level, nm, THREAD); 938 // Check if event led to a higher level OSR compilation 939 CompLevel expected_comp_level = MIN2(CompLevel_full_optimization, static_cast<CompLevel>(comp_level + 1)); 940 if (!CompilationModeFlag::disable_intermediate() && inlinee->is_not_osr_compilable(expected_comp_level)) { 941 // It's not possible to reach the expected level so fall back to simple. 942 expected_comp_level = CompLevel_simple; 943 } 944 CompLevel max_osr_level = static_cast<CompLevel>(inlinee->highest_osr_comp_level()); 945 if (max_osr_level >= expected_comp_level) { // fast check to avoid locking in a typical scenario 946 nmethod* osr_nm = inlinee->lookup_osr_nmethod_for(bci, expected_comp_level, false); 947 assert(osr_nm == nullptr || osr_nm->comp_level() >= expected_comp_level, "lookup_osr_nmethod_for is broken"); 948 if (osr_nm != nullptr && osr_nm->comp_level() != comp_level) { 949 // Perform OSR with new nmethod 950 return osr_nm; 951 } 952 } 953 } 954 return nullptr; 955 } 956 957 // Check if the method can be compiled, change level if necessary 958 void CompilationPolicy::compile(const methodHandle& mh, int bci, CompLevel level, TRAPS) { 959 assert(verify_level(level), "Invalid compilation level requested: %d", level); 960 961 if (level == CompLevel_none) { 962 if (mh->has_compiled_code()) { 963 // Happens when we switch to interpreter to profile. 964 MutexLocker ml(Compile_lock); 965 NoSafepointVerifier nsv; 966 if (mh->has_compiled_code()) { 967 mh->code()->make_not_used(); 968 } 969 // Deoptimize immediately (we don't have to wait for a compile). 970 JavaThread* jt = THREAD; 971 RegisterMap map(jt, 972 RegisterMap::UpdateMap::skip, 973 RegisterMap::ProcessFrames::include, 974 RegisterMap::WalkContinuation::skip); 975 frame fr = jt->last_frame().sender(&map); 976 Deoptimization::deoptimize_frame(jt, fr.id()); 977 } 978 return; 979 } 980 981 // Check if the method can be compiled. Additional logic for TieredCompilation: 982 // If it cannot be compiled with C1, continue profiling in the interpreter 983 // and then compile with C2 (the transition function will request that, 984 // see common() ). If the method cannot be compiled with C2 but still can with C1, compile it with 985 // pure C1. 986 if ((bci == InvocationEntryBci && !can_be_compiled(mh, level))) { 987 if (!CompilationModeFlag::disable_intermediate() && 988 level == CompLevel_full_optimization && can_be_compiled(mh, CompLevel_simple)) { 989 compile(mh, bci, CompLevel_simple, THREAD); 990 } 991 return; 992 } 993 if ((bci != InvocationEntryBci && !can_be_osr_compiled(mh, level))) { 994 if (!CompilationModeFlag::disable_intermediate() && 995 level == CompLevel_full_optimization && can_be_osr_compiled(mh, CompLevel_simple)) { 996 nmethod* osr_nm = mh->lookup_osr_nmethod_for(bci, CompLevel_simple, false); 997 if (osr_nm != nullptr && osr_nm->comp_level() > CompLevel_simple) { 998 // Invalidate the existing OSR nmethod so that a compile at CompLevel_simple is permitted. 999 osr_nm->make_not_entrant(nmethod::InvalidationReason::OSR_INVALIDATION_FOR_COMPILING_WITH_C1); 1000 } 1001 compile(mh, bci, CompLevel_simple, THREAD); 1002 } 1003 return; 1004 } 1005 if (bci != InvocationEntryBci && mh->is_not_osr_compilable(level)) { 1006 return; 1007 } 1008 if (!CompileBroker::compilation_is_in_queue(mh)) { 1009 if (PrintTieredEvents) { 1010 print_event(COMPILE, mh(), mh(), bci, level); 1011 } 1012 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 1013 update_rate(nanos_to_millis(os::javaTimeNanos()), mh); 1014 bool requires_online_compilation = false; 1015 if (TrainingData::have_data()) { 1016 MethodTrainingData* mtd = MethodTrainingData::find_fast(mh); 1017 if (mtd != nullptr) { 1018 CompileTrainingData* ctd = mtd->last_toplevel_compile(level); 1019 if (ctd != nullptr) { 1020 requires_online_compilation = (ctd->init_deps_left_acquire() > 0); 1021 } 1022 } 1023 } 1024 CompileBroker::compile_method(mh, bci, level, hot_count, requires_online_compilation, CompileTask::Reason_Tiered, THREAD); 1025 } 1026 } 1027 1028 // update_rate() is called from select_task() while holding a compile queue lock. 1029 void CompilationPolicy::update_rate(int64_t t, const methodHandle& method) { 1030 // Skip update if counters are absent. 1031 // Can't allocate them since we are holding compile queue lock. 1032 if (method->method_counters() == nullptr) return; 1033 1034 if (is_old(method)) { 1035 // We don't remove old methods from the queue, 1036 // so we can just zero the rate. 1037 method->set_rate(0); 1038 return; 1039 } 1040 1041 // We don't update the rate if we've just came out of a safepoint. 1042 // delta_s is the time since last safepoint in milliseconds. 1043 int64_t delta_s = t - SafepointTracing::end_of_last_safepoint_ms(); 1044 int64_t delta_t = t - (method->prev_time() != 0 ? method->prev_time() : start_time()); // milliseconds since the last measurement 1045 // How many events were there since the last time? 1046 int event_count = method->invocation_count() + method->backedge_count(); 1047 int delta_e = event_count - method->prev_event_count(); 1048 1049 // We should be running for at least 1ms. 1050 if (delta_s >= TieredRateUpdateMinTime) { 1051 // And we must've taken the previous point at least 1ms before. 1052 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 1053 method->set_prev_time(t); 1054 method->set_prev_event_count(event_count); 1055 method->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 1056 } else { 1057 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 1058 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 1059 method->set_rate(0); 1060 } 1061 } 1062 } 1063 } 1064 1065 // Check if this method has been stale for a given number of milliseconds. 1066 // See select_task(). 1067 bool CompilationPolicy::is_stale(int64_t t, int64_t timeout, const methodHandle& method) { 1068 int64_t delta_s = t - SafepointTracing::end_of_last_safepoint_ms(); 1069 int64_t delta_t = t - method->prev_time(); 1070 if (delta_t > timeout && delta_s > timeout) { 1071 int event_count = method->invocation_count() + method->backedge_count(); 1072 int delta_e = event_count - method->prev_event_count(); 1073 // Return true if there were no events. 1074 return delta_e == 0; 1075 } 1076 return false; 1077 } 1078 1079 // We don't remove old methods from the compile queue even if they have 1080 // very low activity. See select_task(). 1081 bool CompilationPolicy::is_old(const methodHandle& method) { 1082 int i = method->invocation_count(); 1083 int b = method->backedge_count(); 1084 double k = TieredOldPercentage / 100.0; 1085 1086 return CallPredicate::apply_scaled(method, CompLevel_none, i, b, k) || LoopPredicate::apply_scaled(method, CompLevel_none, i, b, k); 1087 } 1088 1089 double CompilationPolicy::weight(Method* method) { 1090 return (double)(method->rate() + 1) * (method->invocation_count() + 1) * (method->backedge_count() + 1); 1091 } 1092 1093 // Apply heuristics and return true if x should be compiled before y 1094 bool CompilationPolicy::compare_methods(Method* x, Method* y) { 1095 if (x->highest_comp_level() > y->highest_comp_level()) { 1096 // recompilation after deopt 1097 return true; 1098 } else 1099 if (x->highest_comp_level() == y->highest_comp_level()) { 1100 if (weight(x) > weight(y)) { 1101 return true; 1102 } 1103 } 1104 return false; 1105 } 1106 1107 bool CompilationPolicy::compare_tasks(CompileTask* x, CompileTask* y) { 1108 assert(!x->is_aot_load() && !y->is_aot_load(), "AOT code caching tasks are not expected here"); 1109 if (x->compile_reason() != y->compile_reason() && y->compile_reason() == CompileTask::Reason_MustBeCompiled) { 1110 return true; 1111 } 1112 return false; 1113 } 1114 1115 // Is method profiled enough? 1116 bool CompilationPolicy::is_method_profiled(const methodHandle& method) { 1117 MethodData* mdo = method->method_data(); 1118 if (mdo != nullptr) { 1119 int i = mdo->invocation_count_delta(); 1120 int b = mdo->backedge_count_delta(); 1121 return CallPredicate::apply_scaled(method, CompLevel_full_profile, i, b, 1); 1122 } 1123 return false; 1124 } 1125 1126 1127 // Determine is a method is mature. 1128 bool CompilationPolicy::is_mature(MethodData* mdo) { 1129 if (Arguments::is_compiler_only()) { 1130 // Always report profiles as immature with -Xcomp 1131 return false; 1132 } 1133 methodHandle mh(Thread::current(), mdo->method()); 1134 if (mdo != nullptr) { 1135 int i = mdo->invocation_count(); 1136 int b = mdo->backedge_count(); 1137 double k = ProfileMaturityPercentage / 100.0; 1138 return CallPredicate::apply_scaled(mh, CompLevel_full_profile, i, b, k) || LoopPredicate::apply_scaled(mh, CompLevel_full_profile, i, b, k); 1139 } 1140 return false; 1141 } 1142 1143 // If a method is old enough and is still in the interpreter we would want to 1144 // start profiling without waiting for the compiled method to arrive. 1145 // We also take the load on compilers into the account. 1146 bool CompilationPolicy::should_create_mdo(const methodHandle& method, CompLevel cur_level) { 1147 if (cur_level != CompLevel_none || force_comp_at_level_simple(method) || CompilationModeFlag::quick_only() || !ProfileInterpreter) { 1148 return false; 1149 } 1150 1151 if (TrainingData::have_data()) { 1152 MethodTrainingData* mtd = MethodTrainingData::find_fast(method); 1153 if (mtd != nullptr && mtd->saw_level(CompLevel_full_optimization)) { 1154 return true; 1155 } 1156 } 1157 1158 if (is_old(method)) { 1159 return true; 1160 } 1161 int i = method->invocation_count(); 1162 int b = method->backedge_count(); 1163 double k = Tier0ProfilingStartPercentage / 100.0; 1164 1165 // If the top level compiler is not keeping up, delay profiling. 1166 if (CompileBroker::queue_size(CompLevel_full_optimization) <= Tier0Delay * compiler_count(CompLevel_full_optimization)) { 1167 return CallPredicate::apply_scaled(method, CompLevel_none, i, b, k) || LoopPredicate::apply_scaled(method, CompLevel_none, i, b, k); 1168 } 1169 return false; 1170 } 1171 1172 // Inlining control: if we're compiling a profiled method with C1 and the callee 1173 // is known to have OSRed in a C2 version, don't inline it. 1174 bool CompilationPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 1175 CompLevel comp_level = (CompLevel)env->comp_level(); 1176 if (comp_level == CompLevel_full_profile || 1177 comp_level == CompLevel_limited_profile) { 1178 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 1179 } 1180 return false; 1181 } 1182 1183 // Create MDO if necessary. 1184 void CompilationPolicy::create_mdo(const methodHandle& mh, JavaThread* THREAD) { 1185 if (mh->is_native() || 1186 mh->is_abstract() || 1187 mh->is_accessor() || 1188 mh->is_constant_getter()) { 1189 return; 1190 } 1191 if (mh->method_data() == nullptr) { 1192 Method::build_profiling_method_data(mh, CHECK_AND_CLEAR); 1193 } 1194 if (ProfileInterpreter && THREAD->has_last_Java_frame()) { 1195 MethodData* mdo = mh->method_data(); 1196 if (mdo != nullptr) { 1197 frame last_frame = THREAD->last_frame(); 1198 if (last_frame.is_interpreted_frame() && mh == last_frame.interpreter_frame_method()) { 1199 int bci = last_frame.interpreter_frame_bci(); 1200 address dp = mdo->bci_to_dp(bci); 1201 last_frame.interpreter_frame_set_mdp(dp); 1202 } 1203 } 1204 } 1205 } 1206 1207 CompLevel CompilationPolicy::trained_transition_from_none(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1208 precond(mtd != nullptr); 1209 precond(cur_level == CompLevel_none); 1210 1211 if (mtd->only_inlined() && !mtd->saw_level(CompLevel_full_optimization)) { 1212 return CompLevel_none; 1213 } 1214 1215 bool training_has_profile = (mtd->final_profile() != nullptr); 1216 if (mtd->saw_level(CompLevel_full_optimization) && !training_has_profile) { 1217 return CompLevel_full_profile; 1218 } 1219 1220 CompLevel highest_training_level = static_cast<CompLevel>(mtd->highest_top_level()); 1221 switch (highest_training_level) { 1222 case CompLevel_limited_profile: 1223 case CompLevel_full_profile: 1224 return CompLevel_limited_profile; 1225 case CompLevel_simple: 1226 return CompLevel_simple; 1227 case CompLevel_none: 1228 return CompLevel_none; 1229 default: 1230 break; 1231 } 1232 1233 // Now handle the case of level 4. 1234 assert(highest_training_level == CompLevel_full_optimization, "Unexpected compilation level: %d", highest_training_level); 1235 if (!training_has_profile) { 1236 // The method was a part of a level 4 compile, but don't have a stored profile, 1237 // we need to profile it. 1238 return CompLevel_full_profile; 1239 } 1240 const bool deopt = (static_cast<CompLevel>(method->highest_comp_level()) == CompLevel_full_optimization); 1241 // If we deopted, then we reprofile 1242 if (deopt && !is_method_profiled(method)) { 1243 return CompLevel_full_profile; 1244 } 1245 1246 CompileTrainingData* ctd = mtd->last_toplevel_compile(CompLevel_full_optimization); 1247 assert(ctd != nullptr, "Should have CTD for CompLevel_full_optimization"); 1248 // With SkipTier2IfPossible and all deps satisfied, go to level 4 immediately 1249 if (SkipTier2IfPossible && ctd->init_deps_left_acquire() == 0) { 1250 if (method->method_data() == nullptr) { 1251 create_mdo(method, THREAD); 1252 } 1253 return CompLevel_full_optimization; 1254 } 1255 1256 // Otherwise go to level 2 1257 return CompLevel_limited_profile; 1258 } 1259 1260 1261 CompLevel CompilationPolicy::trained_transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1262 precond(mtd != nullptr); 1263 precond(cur_level == CompLevel_limited_profile); 1264 1265 // One of the main reasons that we can get here is that we're waiting for the stored C2 code to become ready. 1266 1267 // But first, check if we have a saved profile 1268 bool training_has_profile = (mtd->final_profile() != nullptr); 1269 if (!training_has_profile) { 1270 return CompLevel_full_profile; 1271 } 1272 1273 1274 assert(training_has_profile, "Have to have a profile to be here"); 1275 // Check if the method is ready 1276 CompileTrainingData* ctd = mtd->last_toplevel_compile(CompLevel_full_optimization); 1277 if (ctd != nullptr && ctd->init_deps_left_acquire() == 0) { 1278 if (method->method_data() == nullptr) { 1279 create_mdo(method, THREAD); 1280 } 1281 return CompLevel_full_optimization; 1282 } 1283 1284 // Otherwise stay at the current level 1285 return CompLevel_limited_profile; 1286 } 1287 1288 1289 CompLevel CompilationPolicy::trained_transition_from_full_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1290 precond(mtd != nullptr); 1291 precond(cur_level == CompLevel_full_profile); 1292 1293 CompLevel highest_training_level = static_cast<CompLevel>(mtd->highest_top_level()); 1294 // We have method at the full profile level and we also know that it's possibly an important method. 1295 if (highest_training_level == CompLevel_full_optimization && !mtd->only_inlined()) { 1296 // Check if it is adequately profiled 1297 if (is_method_profiled(method)) { 1298 return CompLevel_full_optimization; 1299 } 1300 } 1301 1302 // Otherwise stay at the current level 1303 return CompLevel_full_profile; 1304 } 1305 1306 CompLevel CompilationPolicy::trained_transition(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD) { 1307 precond(MethodTrainingData::have_data()); 1308 1309 // If there is no training data recorded for this method, bail out. 1310 if (mtd == nullptr) { 1311 return cur_level; 1312 } 1313 1314 CompLevel next_level = cur_level; 1315 switch(cur_level) { 1316 default: break; 1317 case CompLevel_none: 1318 next_level = trained_transition_from_none(method, cur_level, mtd, THREAD); 1319 break; 1320 case CompLevel_limited_profile: 1321 next_level = trained_transition_from_limited_profile(method, cur_level, mtd, THREAD); 1322 break; 1323 case CompLevel_full_profile: 1324 next_level = trained_transition_from_full_profile(method, cur_level, mtd, THREAD); 1325 break; 1326 } 1327 1328 // We don't have any special strategies for the C2-only compilation modes, so just fix up the levels for now. 1329 if (CompilationModeFlag::high_only_quick_internal() && CompLevel_simple < next_level && next_level < CompLevel_full_optimization) { 1330 return CompLevel_none; 1331 } 1332 if (CompilationModeFlag::high_only() && next_level < CompLevel_full_optimization) { 1333 return CompLevel_none; 1334 } 1335 return (cur_level != next_level) ? limit_level(next_level) : cur_level; 1336 } 1337 1338 /* 1339 * Method states: 1340 * 0 - interpreter (CompLevel_none) 1341 * 1 - pure C1 (CompLevel_simple) 1342 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 1343 * 3 - C1 with full profiling (CompLevel_full_profile) 1344 * 4 - C2 or Graal (CompLevel_full_optimization) 1345 * 1346 * Common state transition patterns: 1347 * a. 0 -> 3 -> 4. 1348 * The most common path. But note that even in this straightforward case 1349 * profiling can start at level 0 and finish at level 3. 1350 * 1351 * b. 0 -> 2 -> 3 -> 4. 1352 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning 1353 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 1354 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 1355 * 1356 * c. 0 -> (3->2) -> 4. 1357 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 1358 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 1359 * of the method to 2 while the request is still in-queue, because it'll allow it to run much faster 1360 * without full profiling while c2 is compiling. 1361 * 1362 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 1363 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 1364 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 1365 * 1366 * e. 0 -> 4. 1367 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 1368 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 1369 * the compiled version already exists). 1370 * 1371 * Note that since state 0 can be reached from any other state via deoptimization different loops 1372 * are possible. 1373 * 1374 */ 1375 1376 // Common transition function. Given a predicate determines if a method should transition to another level. 1377 template<typename Predicate> 1378 CompLevel CompilationPolicy::common(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD, bool disable_feedback) { 1379 CompLevel next_level = cur_level; 1380 1381 if (force_comp_at_level_simple(method)) { 1382 next_level = CompLevel_simple; 1383 } else if (is_trivial(method) || method->is_native()) { 1384 // We do not care if there is profiling data for these methods, throw them to compiler. 1385 next_level = CompilationModeFlag::disable_intermediate() ? CompLevel_full_optimization : CompLevel_simple; 1386 } else if (MethodTrainingData::have_data()) { 1387 MethodTrainingData* mtd = MethodTrainingData::find_fast(method); 1388 if (mtd == nullptr) { 1389 // We haven't see compilations of this method in training. It's either very cold or the behavior changed. 1390 // Feed it to the standard TF with no profiling delay. 1391 next_level = standard_transition<Predicate>(method, cur_level, false /*delay_profiling*/, disable_feedback); 1392 } else { 1393 next_level = trained_transition(method, cur_level, mtd, THREAD); 1394 if (cur_level == next_level) { 1395 // trained_transtion() is going to return the same level if no startup/warmup optimizations apply. 1396 // In order to catch possible pathologies due to behavior change we feed the event to the regular 1397 // TF but with profiling delay. 1398 next_level = standard_transition<Predicate>(method, cur_level, true /*delay_profiling*/, disable_feedback); 1399 } 1400 } 1401 } else { 1402 next_level = standard_transition<Predicate>(method, cur_level, false /*delay_profiling*/, disable_feedback); 1403 } 1404 return (next_level != cur_level) ? limit_level(next_level) : next_level; 1405 } 1406 1407 1408 template<typename Predicate> 1409 CompLevel CompilationPolicy::standard_transition(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback) { 1410 CompLevel next_level = cur_level; 1411 switch(cur_level) { 1412 default: break; 1413 case CompLevel_none: 1414 next_level = transition_from_none<Predicate>(method, cur_level, delay_profiling, disable_feedback); 1415 break; 1416 case CompLevel_limited_profile: 1417 next_level = transition_from_limited_profile<Predicate>(method, cur_level, delay_profiling, disable_feedback); 1418 break; 1419 case CompLevel_full_profile: 1420 next_level = transition_from_full_profile<Predicate>(method, cur_level); 1421 break; 1422 } 1423 return next_level; 1424 } 1425 1426 template<typename Predicate> static inline bool apply_predicate(const methodHandle& method, CompLevel cur_level, int i, int b, bool delay_profiling, double delay_profiling_scale) { 1427 if (delay_profiling) { 1428 return Predicate::apply_scaled(method, cur_level, i, b, delay_profiling_scale); 1429 } else { 1430 return Predicate::apply(method, cur_level, i, b); 1431 } 1432 } 1433 1434 template<typename Predicate> 1435 CompLevel CompilationPolicy::transition_from_none(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback) { 1436 precond(cur_level == CompLevel_none); 1437 CompLevel next_level = cur_level; 1438 int i = method->invocation_count(); 1439 int b = method->backedge_count(); 1440 // If we were at full profile level, would we switch to full opt? 1441 if (transition_from_full_profile<Predicate>(method, CompLevel_full_profile) == CompLevel_full_optimization) { 1442 next_level = CompLevel_full_optimization; 1443 } else if (!CompilationModeFlag::disable_intermediate() && apply_predicate<Predicate>(method, cur_level, i, b, delay_profiling, Tier0ProfileDelayFactor)) { 1444 // C1-generated fully profiled code is about 30% slower than the limited profile 1445 // code that has only invocation and backedge counters. The observation is that 1446 // if C2 queue is large enough we can spend too much time in the fully profiled code 1447 // while waiting for C2 to pick the method from the queue. To alleviate this problem 1448 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 1449 // we choose to compile a limited profiled version and then recompile with full profiling 1450 // when the load on C2 goes down. 1451 if (delay_profiling || (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > Tier3DelayOn * compiler_count(CompLevel_full_optimization))) { 1452 next_level = CompLevel_limited_profile; 1453 } else { 1454 next_level = CompLevel_full_profile; 1455 } 1456 } 1457 return next_level; 1458 } 1459 1460 template<typename Predicate> 1461 CompLevel CompilationPolicy::transition_from_full_profile(const methodHandle& method, CompLevel cur_level) { 1462 precond(cur_level == CompLevel_full_profile); 1463 CompLevel next_level = cur_level; 1464 MethodData* mdo = method->method_data(); 1465 if (mdo != nullptr) { 1466 if (mdo->would_profile() || CompilationModeFlag::disable_intermediate()) { 1467 int mdo_i = mdo->invocation_count_delta(); 1468 int mdo_b = mdo->backedge_count_delta(); 1469 if (Predicate::apply(method, cur_level, mdo_i, mdo_b)) { 1470 next_level = CompLevel_full_optimization; 1471 } 1472 } else { 1473 next_level = CompLevel_full_optimization; 1474 } 1475 } 1476 return next_level; 1477 } 1478 1479 template<typename Predicate> 1480 CompLevel CompilationPolicy::transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback) { 1481 precond(cur_level == CompLevel_limited_profile); 1482 CompLevel next_level = cur_level; 1483 int i = method->invocation_count(); 1484 int b = method->backedge_count(); 1485 MethodData* mdo = method->method_data(); 1486 if (mdo != nullptr) { 1487 if (mdo->would_profile()) { 1488 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 1489 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 1490 apply_predicate<Predicate>(method, cur_level, i, b, delay_profiling, Tier2ProfileDelayFactor))) { 1491 next_level = CompLevel_full_profile; 1492 } 1493 } else { 1494 next_level = CompLevel_full_optimization; 1495 } 1496 } else { 1497 // If there is no MDO we need to profile 1498 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 1499 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 1500 apply_predicate<Predicate>(method, cur_level, i, b, delay_profiling, Tier2ProfileDelayFactor))) { 1501 next_level = CompLevel_full_profile; 1502 } 1503 } 1504 if (next_level == CompLevel_full_profile && is_method_profiled(method)) { 1505 next_level = CompLevel_full_optimization; 1506 } 1507 return next_level; 1508 } 1509 1510 1511 // Determine if a method should be compiled with a normal entry point at a different level. 1512 CompLevel CompilationPolicy::call_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD) { 1513 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), common<LoopPredicate>(method, cur_level, THREAD, true)); 1514 CompLevel next_level = common<CallPredicate>(method, cur_level, THREAD, !TrainingData::have_data() && is_old(method)); 1515 1516 // If OSR method level is greater than the regular method level, the levels should be 1517 // equalized by raising the regular method level in order to avoid OSRs during each 1518 // invocation of the method. 1519 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 1520 MethodData* mdo = method->method_data(); 1521 guarantee(mdo != nullptr, "MDO should not be nullptr"); 1522 if (mdo->invocation_count() >= 1) { 1523 next_level = CompLevel_full_optimization; 1524 } 1525 } else { 1526 next_level = MAX2(osr_level, next_level); 1527 } 1528 1529 return next_level; 1530 } 1531 1532 // Determine if we should do an OSR compilation of a given method. 1533 CompLevel CompilationPolicy::loop_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD) { 1534 CompLevel next_level = common<LoopPredicate>(method, cur_level, THREAD, true); 1535 if (cur_level == CompLevel_none) { 1536 // If there is a live OSR method that means that we deopted to the interpreter 1537 // for the transition. 1538 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 1539 if (osr_level > CompLevel_none) { 1540 return osr_level; 1541 } 1542 } 1543 return next_level; 1544 } 1545 1546 // Handle the invocation event. 1547 void CompilationPolicy::method_invocation_event(const methodHandle& mh, const methodHandle& imh, 1548 CompLevel level, nmethod* nm, TRAPS) { 1549 if (should_create_mdo(mh, level)) { 1550 create_mdo(mh, THREAD); 1551 } 1552 CompLevel next_level = call_event(mh, level, THREAD); 1553 if (next_level != level) { 1554 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 1555 compile(mh, InvocationEntryBci, next_level, THREAD); 1556 } 1557 } 1558 } 1559 1560 // Handle the back branch event. Notice that we can compile the method 1561 // with a regular entry from here. 1562 void CompilationPolicy::method_back_branch_event(const methodHandle& mh, const methodHandle& imh, 1563 int bci, CompLevel level, nmethod* nm, TRAPS) { 1564 if (should_create_mdo(mh, level)) { 1565 create_mdo(mh, THREAD); 1566 } 1567 // Check if MDO should be created for the inlined method 1568 if (should_create_mdo(imh, level)) { 1569 create_mdo(imh, THREAD); 1570 } 1571 1572 if (is_compilation_enabled()) { 1573 CompLevel next_osr_level = loop_event(imh, level, THREAD); 1574 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 1575 // At the very least compile the OSR version 1576 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 1577 compile(imh, bci, next_osr_level, CHECK); 1578 } 1579 1580 // Use loop event as an opportunity to also check if there's been 1581 // enough calls. 1582 CompLevel cur_level, next_level; 1583 if (mh() != imh()) { // If there is an enclosing method 1584 { 1585 guarantee(nm != nullptr, "Should have nmethod here"); 1586 cur_level = comp_level(mh()); 1587 next_level = call_event(mh, cur_level, THREAD); 1588 1589 if (max_osr_level == CompLevel_full_optimization) { 1590 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 1591 bool make_not_entrant = false; 1592 if (nm->is_osr_method()) { 1593 // This is an osr method, just make it not entrant and recompile later if needed 1594 make_not_entrant = true; 1595 } else { 1596 if (next_level != CompLevel_full_optimization) { 1597 // next_level is not full opt, so we need to recompile the 1598 // enclosing method without the inlinee 1599 cur_level = CompLevel_none; 1600 make_not_entrant = true; 1601 } 1602 } 1603 if (make_not_entrant) { 1604 if (PrintTieredEvents) { 1605 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 1606 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 1607 } 1608 nm->make_not_entrant(nmethod::InvalidationReason::OSR_INVALIDATION_BACK_BRANCH); 1609 } 1610 } 1611 // Fix up next_level if necessary to avoid deopts 1612 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 1613 next_level = CompLevel_full_profile; 1614 } 1615 if (cur_level != next_level) { 1616 if (!CompileBroker::compilation_is_in_queue(mh)) { 1617 compile(mh, InvocationEntryBci, next_level, THREAD); 1618 } 1619 } 1620 } 1621 } else { 1622 cur_level = comp_level(mh()); 1623 next_level = call_event(mh, cur_level, THREAD); 1624 if (next_level != cur_level) { 1625 if (!CompileBroker::compilation_is_in_queue(mh)) { 1626 compile(mh, InvocationEntryBci, next_level, THREAD); 1627 } 1628 } 1629 } 1630 } 1631 } 1632