1 /* 2 * Copyright (c) 2010, 2016, 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 "precompiled.hpp" 26 #include "code/codeCache.hpp" 27 #include "compiler/compileTask.hpp" 28 #include "runtime/advancedThresholdPolicy.hpp" 29 #include "runtime/simpleThresholdPolicy.inline.hpp" 30 31 #ifdef TIERED 32 // Print an event. 33 void AdvancedThresholdPolicy::print_specific(EventType type, methodHandle mh, methodHandle imh, 34 int bci, CompLevel level) { 35 tty->print(" rate="); 36 if (mh->prev_time() == 0) tty->print("n/a"); 37 else tty->print("%f", mh->rate()); 38 39 tty->print(" k=%.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback), 40 threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback)); 41 42 } 43 44 void AdvancedThresholdPolicy::initialize() { 45 int count = CICompilerCount; 46 #ifdef _LP64 47 // Turn on ergonomic compiler count selection 48 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { 49 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); 50 } 51 if (CICompilerCountPerCPU) { 52 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n 53 int log_cpu = log2_intptr(os::active_processor_count()); 54 int loglog_cpu = log2_intptr(MAX2(log_cpu, 1)); 55 count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2; 56 } 57 #else 58 // On 32-bit systems, the number of compiler threads is limited to 3. 59 // On these systems, the virtual address space available to the JVM 60 // is usually limited to 2-4 GB (the exact value depends on the platform). 61 // As the compilers (especially C2) can consume a large amount of 62 // memory, scaling the number of compiler threads with the number of 63 // available cores can result in the exhaustion of the address space 64 /// available to the VM and thus cause the VM to crash. 65 if (FLAG_IS_DEFAULT(CICompilerCount)) { 66 count = 3; 67 } 68 #endif 69 70 set_c1_count(MAX2(count / 3, 1)); 71 set_c2_count(MAX2(count - c1_count(), 1)); 72 FLAG_SET_ERGO(intx, CICompilerCount, c1_count() + c2_count()); 73 74 // Some inlining tuning 75 #ifdef X86 76 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 77 FLAG_SET_DEFAULT(InlineSmallCode, 2000); 78 } 79 #endif 80 81 #if defined SPARC || defined AARCH64 82 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 83 FLAG_SET_DEFAULT(InlineSmallCode, 2500); 84 } 85 #endif 86 87 set_increase_threshold_at_ratio(); 88 set_start_time(os::javaTimeMillis()); 89 } 90 91 // update_rate() is called from select_task() while holding a compile queue lock. 92 void AdvancedThresholdPolicy::update_rate(jlong t, Method* m) { 93 // Skip update if counters are absent. 94 // Can't allocate them since we are holding compile queue lock. 95 if (m->method_counters() == NULL) return; 96 97 if (is_old(m)) { 98 // We don't remove old methods from the queue, 99 // so we can just zero the rate. 100 m->set_rate(0); 101 return; 102 } 103 104 // We don't update the rate if we've just came out of a safepoint. 105 // delta_s is the time since last safepoint in milliseconds. 106 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 107 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement 108 // How many events were there since the last time? 109 int event_count = m->invocation_count() + m->backedge_count(); 110 int delta_e = event_count - m->prev_event_count(); 111 112 // We should be running for at least 1ms. 113 if (delta_s >= TieredRateUpdateMinTime) { 114 // And we must've taken the previous point at least 1ms before. 115 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 116 m->set_prev_time(t); 117 m->set_prev_event_count(event_count); 118 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 119 } else { 120 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 121 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 122 m->set_rate(0); 123 } 124 } 125 } 126 } 127 128 // Check if this method has been stale from a given number of milliseconds. 129 // See select_task(). 130 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, Method* m) { 131 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 132 jlong delta_t = t - m->prev_time(); 133 if (delta_t > timeout && delta_s > timeout) { 134 int event_count = m->invocation_count() + m->backedge_count(); 135 int delta_e = event_count - m->prev_event_count(); 136 // Return true if there were no events. 137 return delta_e == 0; 138 } 139 return false; 140 } 141 142 // We don't remove old methods from the compile queue even if they have 143 // very low activity. See select_task(). 144 bool AdvancedThresholdPolicy::is_old(Method* method) { 145 return method->invocation_count() > 50000 || method->backedge_count() > 500000; 146 } 147 148 double AdvancedThresholdPolicy::weight(Method* method) { 149 return (double)(method->rate() + 1) * 150 (method->invocation_count() + 1) * (method->backedge_count() + 1); 151 } 152 153 // Apply heuristics and return true if x should be compiled before y 154 bool AdvancedThresholdPolicy::compare_methods(Method* x, Method* y) { 155 if (x->highest_comp_level() > y->highest_comp_level()) { 156 // recompilation after deopt 157 return true; 158 } else 159 if (x->highest_comp_level() == y->highest_comp_level()) { 160 if (weight(x) > weight(y)) { 161 return true; 162 } 163 } 164 return false; 165 } 166 167 // Is method profiled enough? 168 bool AdvancedThresholdPolicy::is_method_profiled(Method* method) { 169 MethodData* mdo = method->method_data(); 170 if (mdo != NULL) { 171 int i = mdo->invocation_count_delta(); 172 int b = mdo->backedge_count_delta(); 173 return call_predicate_helper<CompLevel_full_profile>(i, b, 1, method); 174 } 175 return false; 176 } 177 178 // Called with the queue locked and with at least one element 179 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { 180 CompileTask *max_blocking_task = NULL; 181 CompileTask *max_task = NULL; 182 Method* max_method = NULL; 183 jlong t = os::javaTimeMillis(); 184 // Iterate through the queue and find a method with a maximum rate. 185 for (CompileTask* task = compile_queue->first(); task != NULL;) { 186 CompileTask* next_task = task->next(); 187 Method* method = task->method(); 188 update_rate(t, method); 189 if (max_task == NULL) { 190 max_task = task; 191 max_method = method; 192 } else { 193 // If a method has been stale for some time, remove it from the queue. 194 // Blocking tasks don't become stale 195 if (!task->is_blocking() && is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { 196 if (PrintTieredEvents) { 197 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel)task->comp_level()); 198 } 199 task->log_task_dequeued("stale"); 200 compile_queue->remove_and_mark_stale(task); 201 method->clear_queued_for_compilation(); 202 task = next_task; 203 continue; 204 } 205 206 // Select a method with a higher rate 207 if (compare_methods(method, max_method)) { 208 max_task = task; 209 max_method = method; 210 } 211 } 212 213 if (task->is_blocking()) { 214 if (max_blocking_task == NULL || compare_methods(method, max_blocking_task->method())) { 215 max_blocking_task = task; 216 } 217 } 218 219 task = next_task; 220 } 221 222 if (max_blocking_task != NULL) { 223 // In blocking compilation mode, the CompileBroker will make 224 // compilations submitted by a JVMCI compiler thread non-blocking. These 225 // compilations should be scheduled after all blocking compilations 226 // to service non-compiler related compilations sooner and reduce the 227 // chance of such compilations timing out. 228 max_task = max_blocking_task; 229 max_method = max_task->method(); 230 } 231 232 if (max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile 233 && is_method_profiled(max_method)) { 234 max_task->set_comp_level(CompLevel_limited_profile); 235 if (PrintTieredEvents) { 236 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 237 } 238 } 239 240 return max_task; 241 } 242 243 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 244 double queue_size = CompileBroker::queue_size(level); 245 int comp_count = compiler_count(level); 246 double k = queue_size / (feedback_k * comp_count) + 1; 247 248 // Increase C1 compile threshold when the code cache is filled more 249 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 250 // The main intention is to keep enough free space for C2 compiled code 251 // to achieve peak performance if the code cache is under stress. 252 if ((TieredStopAtLevel == CompLevel_full_optimization) && (level != CompLevel_full_optimization)) { 253 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(CodeCache::get_code_blob_type(level)); 254 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 255 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 256 } 257 } 258 return k; 259 } 260 261 // Call and loop predicates determine whether a transition to a higher 262 // compilation level should be performed (pointers to predicate functions 263 // are passed to common()). 264 // Tier?LoadFeedback is basically a coefficient that determines of 265 // how many methods per compiler thread can be in the queue before 266 // the threshold values double. 267 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level, Method* method) { 268 switch(cur_level) { 269 case CompLevel_none: 270 case CompLevel_limited_profile: { 271 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 272 return loop_predicate_helper<CompLevel_none>(i, b, k, method); 273 } 274 case CompLevel_full_profile: { 275 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 276 return loop_predicate_helper<CompLevel_full_profile>(i, b, k, method); 277 } 278 default: 279 return true; 280 } 281 } 282 283 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level, Method* method) { 284 switch(cur_level) { 285 case CompLevel_none: 286 case CompLevel_limited_profile: { 287 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 288 return call_predicate_helper<CompLevel_none>(i, b, k, method); 289 } 290 case CompLevel_full_profile: { 291 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 292 return call_predicate_helper<CompLevel_full_profile>(i, b, k, method); 293 } 294 default: 295 return true; 296 } 297 } 298 299 // If a method is old enough and is still in the interpreter we would want to 300 // start profiling without waiting for the compiled method to arrive. 301 // We also take the load on compilers into the account. 302 bool AdvancedThresholdPolicy::should_create_mdo(Method* method, CompLevel cur_level) { 303 if (cur_level == CompLevel_none && 304 CompileBroker::queue_size(CompLevel_full_optimization) <= 305 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 306 int i = method->invocation_count(); 307 int b = method->backedge_count(); 308 double k = Tier0ProfilingStartPercentage / 100.0; 309 return call_predicate_helper<CompLevel_none>(i, b, k, method) || loop_predicate_helper<CompLevel_none>(i, b, k, method); 310 } 311 return false; 312 } 313 314 // Inlining control: if we're compiling a profiled method with C1 and the callee 315 // is known to have OSRed in a C2 version, don't inline it. 316 bool AdvancedThresholdPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 317 CompLevel comp_level = (CompLevel)env->comp_level(); 318 if (comp_level == CompLevel_full_profile || 319 comp_level == CompLevel_limited_profile) { 320 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 321 } 322 return false; 323 } 324 325 // Create MDO if necessary. 326 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, JavaThread* THREAD) { 327 if (mh->is_native() || 328 mh->is_abstract() || 329 mh->is_accessor() || 330 mh->is_constant_getter()) { 331 return; 332 } 333 if (mh->method_data() == NULL) { 334 Method::build_interpreter_method_data(mh, CHECK_AND_CLEAR); 335 } 336 } 337 338 339 /* 340 * Method states: 341 * 0 - interpreter (CompLevel_none) 342 * 1 - pure C1 (CompLevel_simple) 343 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 344 * 3 - C1 with full profiling (CompLevel_full_profile) 345 * 4 - C2 (CompLevel_full_optimization) 346 * 347 * Common state transition patterns: 348 * a. 0 -> 3 -> 4. 349 * The most common path. But note that even in this straightforward case 350 * profiling can start at level 0 and finish at level 3. 351 * 352 * b. 0 -> 2 -> 3 -> 4. 353 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning 354 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 355 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 356 * 357 * c. 0 -> (3->2) -> 4. 358 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 359 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 360 * of the method to 2 while the request is still in-queue, because it'll allow it to run much faster 361 * without full profiling while c2 is compiling. 362 * 363 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 364 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 365 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 366 * 367 * e. 0 -> 4. 368 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 369 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 370 * the compiled version already exists). 371 * 372 * Note that since state 0 can be reached from any other state via deoptimization different loops 373 * are possible. 374 * 375 */ 376 377 // Common transition function. Given a predicate determines if a method should transition to another level. 378 CompLevel AdvancedThresholdPolicy::common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback) { 379 CompLevel next_level = cur_level; 380 int i = method->invocation_count(); 381 int b = method->backedge_count(); 382 383 if (is_trivial(method)) { 384 next_level = CompLevel_simple; 385 } else { 386 switch(cur_level) { 387 case CompLevel_none: 388 // If we were at full profile level, would we switch to full opt? 389 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { 390 next_level = CompLevel_full_optimization; 391 } else if ((this->*p)(i, b, cur_level, method)) { 392 #if INCLUDE_JVMCI 393 if (UseJVMCICompiler) { 394 // Since JVMCI takes a while to warm up, its queue inevitably backs up during 395 // early VM execution. 396 next_level = CompLevel_full_profile; 397 break; 398 } 399 #endif 400 // C1-generated fully profiled code is about 30% slower than the limited profile 401 // code that has only invocation and backedge counters. The observation is that 402 // if C2 queue is large enough we can spend too much time in the fully profiled code 403 // while waiting for C2 to pick the method from the queue. To alleviate this problem 404 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 405 // we choose to compile a limited profiled version and then recompile with full profiling 406 // when the load on C2 goes down. 407 if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > 408 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 409 next_level = CompLevel_limited_profile; 410 } else { 411 next_level = CompLevel_full_profile; 412 } 413 } 414 break; 415 case CompLevel_limited_profile: 416 if (is_method_profiled(method)) { 417 // Special case: we got here because this method was fully profiled in the interpreter. 418 next_level = CompLevel_full_optimization; 419 } else { 420 MethodData* mdo = method->method_data(); 421 if (mdo != NULL) { 422 if (mdo->would_profile()) { 423 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 424 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 425 (this->*p)(i, b, cur_level, method))) { 426 next_level = CompLevel_full_profile; 427 } 428 } else { 429 next_level = CompLevel_full_optimization; 430 } 431 } 432 } 433 break; 434 case CompLevel_full_profile: 435 { 436 MethodData* mdo = method->method_data(); 437 if (mdo != NULL) { 438 if (mdo->would_profile()) { 439 int mdo_i = mdo->invocation_count_delta(); 440 int mdo_b = mdo->backedge_count_delta(); 441 if ((this->*p)(mdo_i, mdo_b, cur_level, method)) { 442 next_level = CompLevel_full_optimization; 443 } 444 } else { 445 next_level = CompLevel_full_optimization; 446 } 447 } 448 } 449 break; 450 } 451 } 452 return MIN2(next_level, (CompLevel)TieredStopAtLevel); 453 } 454 455 // Determine if a method should be compiled with a normal entry point at a different level. 456 CompLevel AdvancedThresholdPolicy::call_event(Method* method, CompLevel cur_level) { 457 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), 458 common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true)); 459 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 460 461 // If OSR method level is greater than the regular method level, the levels should be 462 // equalized by raising the regular method level in order to avoid OSRs during each 463 // invocation of the method. 464 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 465 MethodData* mdo = method->method_data(); 466 guarantee(mdo != NULL, "MDO should not be NULL"); 467 if (mdo->invocation_count() >= 1) { 468 next_level = CompLevel_full_optimization; 469 } 470 } else { 471 next_level = MAX2(osr_level, next_level); 472 } 473 return next_level; 474 } 475 476 // Determine if we should do an OSR compilation of a given method. 477 CompLevel AdvancedThresholdPolicy::loop_event(Method* method, CompLevel cur_level) { 478 CompLevel next_level = common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true); 479 if (cur_level == CompLevel_none) { 480 // If there is a live OSR method that means that we deopted to the interpreter 481 // for the transition. 482 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 483 if (osr_level > CompLevel_none) { 484 return osr_level; 485 } 486 } 487 return next_level; 488 } 489 490 // Update the rate and submit compile 491 void AdvancedThresholdPolicy::submit_compile(const methodHandle& mh, int bci, CompLevel level, JavaThread* thread) { 492 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 493 update_rate(os::javaTimeMillis(), mh()); 494 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", thread); 495 } 496 497 // Handle the invocation event. 498 void AdvancedThresholdPolicy::method_invocation_event(const methodHandle& mh, const methodHandle& imh, 499 CompLevel level, nmethod* nm, JavaThread* thread) { 500 if (should_create_mdo(mh(), level)) { 501 create_mdo(mh, thread); 502 } 503 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 504 CompLevel next_level = call_event(mh(), level); 505 if (next_level != level) { 506 compile(mh, InvocationEntryBci, next_level, thread); 507 } 508 } 509 } 510 511 // Handle the back branch event. Notice that we can compile the method 512 // with a regular entry from here. 513 void AdvancedThresholdPolicy::method_back_branch_event(const methodHandle& mh, const methodHandle& imh, 514 int bci, CompLevel level, nmethod* nm, JavaThread* thread) { 515 if (should_create_mdo(mh(), level)) { 516 create_mdo(mh, thread); 517 } 518 // Check if MDO should be created for the inlined method 519 if (should_create_mdo(imh(), level)) { 520 create_mdo(imh, thread); 521 } 522 523 if (is_compilation_enabled()) { 524 CompLevel next_osr_level = loop_event(imh(), level); 525 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 526 // At the very least compile the OSR version 527 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 528 compile(imh, bci, next_osr_level, thread); 529 } 530 531 // Use loop event as an opportunity to also check if there's been 532 // enough calls. 533 CompLevel cur_level, next_level; 534 if (mh() != imh()) { // If there is an enclosing method 535 guarantee(nm != NULL, "Should have nmethod here"); 536 cur_level = comp_level(mh()); 537 next_level = call_event(mh(), cur_level); 538 539 if (max_osr_level == CompLevel_full_optimization) { 540 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 541 bool make_not_entrant = false; 542 if (nm->is_osr_method()) { 543 // This is an osr method, just make it not entrant and recompile later if needed 544 make_not_entrant = true; 545 } else { 546 if (next_level != CompLevel_full_optimization) { 547 // next_level is not full opt, so we need to recompile the 548 // enclosing method without the inlinee 549 cur_level = CompLevel_none; 550 make_not_entrant = true; 551 } 552 } 553 if (make_not_entrant) { 554 if (PrintTieredEvents) { 555 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 556 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 557 } 558 nm->make_not_entrant(); 559 } 560 } 561 if (!CompileBroker::compilation_is_in_queue(mh)) { 562 // Fix up next_level if necessary to avoid deopts 563 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 564 next_level = CompLevel_full_profile; 565 } 566 if (cur_level != next_level) { 567 compile(mh, InvocationEntryBci, next_level, thread); 568 } 569 } 570 } else { 571 cur_level = comp_level(imh()); 572 next_level = call_event(imh(), cur_level); 573 if (!CompileBroker::compilation_is_in_queue(imh) && (next_level != cur_level)) { 574 compile(imh, InvocationEntryBci, next_level, thread); 575 } 576 } 577 } 578 } 579 580 #endif // TIERED