1 /* 2 * Copyright (c) 2010, 2015, 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 // Turn on ergonomic compiler count selection 46 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { 47 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); 48 } 49 int count = CICompilerCount; 50 if (CICompilerCountPerCPU) { 51 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n 52 int log_cpu = log2_intptr(os::active_processor_count()); 53 int loglog_cpu = log2_intptr(MAX2(log_cpu, 1)); 54 count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2; 55 } 56 57 set_c1_count(MAX2(count / 3, 1)); 58 set_c2_count(MAX2(count - c1_count(), 1)); 59 FLAG_SET_ERGO(intx, CICompilerCount, c1_count() + c2_count()); 60 61 // Some inlining tuning 62 #ifdef X86 63 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 64 FLAG_SET_DEFAULT(InlineSmallCode, 2000); 65 } 66 #endif 67 68 #if defined SPARC || defined AARCH64 69 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 70 FLAG_SET_DEFAULT(InlineSmallCode, 2500); 71 } 72 #endif 73 74 set_increase_threshold_at_ratio(); 75 set_start_time(os::javaTimeMillis()); 76 } 77 78 // update_rate() is called from select_task() while holding a compile queue lock. 79 void AdvancedThresholdPolicy::update_rate(jlong t, Method* m) { 80 // Skip update if counters are absent. 81 // Can't allocate them since we are holding compile queue lock. 82 if (m->method_counters() == NULL) return; 83 84 if (is_old(m)) { 85 // We don't remove old methods from the queue, 86 // so we can just zero the rate. 87 m->set_rate(0); 88 return; 89 } 90 91 // We don't update the rate if we've just came out of a safepoint. 92 // delta_s is the time since last safepoint in milliseconds. 93 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 94 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement 95 // How many events were there since the last time? 96 int event_count = m->invocation_count() + m->backedge_count(); 97 int delta_e = event_count - m->prev_event_count(); 98 99 // We should be running for at least 1ms. 100 if (delta_s >= TieredRateUpdateMinTime) { 101 // And we must've taken the previous point at least 1ms before. 102 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 103 m->set_prev_time(t); 104 m->set_prev_event_count(event_count); 105 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 106 } else { 107 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 108 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 109 m->set_rate(0); 110 } 111 } 112 } 113 } 114 115 // Check if this method has been stale from a given number of milliseconds. 116 // See select_task(). 117 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, Method* m) { 118 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 119 jlong delta_t = t - m->prev_time(); 120 if (delta_t > timeout && delta_s > timeout) { 121 int event_count = m->invocation_count() + m->backedge_count(); 122 int delta_e = event_count - m->prev_event_count(); 123 // Return true if there were no events. 124 return delta_e == 0; 125 } 126 return false; 127 } 128 129 // We don't remove old methods from the compile queue even if they have 130 // very low activity. See select_task(). 131 bool AdvancedThresholdPolicy::is_old(Method* method) { 132 return method->invocation_count() > 50000 || method->backedge_count() > 500000; 133 } 134 135 double AdvancedThresholdPolicy::weight(Method* method) { 136 return (method->rate() + 1) * ((method->invocation_count() + 1) * (method->backedge_count() + 1)); 137 } 138 139 // Apply heuristics and return true if x should be compiled before y 140 bool AdvancedThresholdPolicy::compare_methods(Method* x, Method* y) { 141 if (x->highest_comp_level() > y->highest_comp_level()) { 142 // recompilation after deopt 143 return true; 144 } else 145 if (x->highest_comp_level() == y->highest_comp_level()) { 146 if (weight(x) > weight(y)) { 147 return true; 148 } 149 } 150 return false; 151 } 152 153 // Is method profiled enough? 154 bool AdvancedThresholdPolicy::is_method_profiled(Method* method) { 155 MethodData* mdo = method->method_data(); 156 if (mdo != NULL) { 157 int i = mdo->invocation_count_delta(); 158 int b = mdo->backedge_count_delta(); 159 return call_predicate_helper<CompLevel_full_profile>(i, b, 1, method); 160 } 161 return false; 162 } 163 164 // Called with the queue locked and with at least one element 165 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { 166 #if INCLUDE_JVMCI 167 CompileTask *max_non_jvmci_task = NULL; 168 #endif 169 CompileTask *max_task = NULL; 170 Method* max_method = NULL; 171 jlong t = os::javaTimeMillis(); 172 // Iterate through the queue and find a method with a maximum rate. 173 for (CompileTask* task = compile_queue->first(); task != NULL;) { 174 CompileTask* next_task = task->next(); 175 Method* method = task->method(); 176 update_rate(t, method); 177 if (max_task == NULL) { 178 max_task = task; 179 max_method = method; 180 } else { 181 // If a method has been stale for some time, remove it from the queue. 182 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { 183 if (PrintTieredEvents) { 184 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel)task->comp_level()); 185 } 186 task->log_task_dequeued("stale"); 187 compile_queue->remove_and_mark_stale(task); 188 method->clear_queued_for_compilation(); 189 task = next_task; 190 continue; 191 } 192 193 // Select a method with a higher rate 194 if (compare_methods(method, max_method)) { 195 max_task = task; 196 max_method = method; 197 } 198 } 199 task = next_task; 200 } 201 202 #if INCLUDE_JVMCI 203 if (UseJVMCICompiler) { 204 if (max_non_jvmci_task != NULL) { 205 max_task = max_non_jvmci_task; 206 max_method = max_task->method(); 207 } 208 } 209 #endif 210 211 if (max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile 212 && is_method_profiled(max_method)) { 213 max_task->set_comp_level(CompLevel_limited_profile); 214 if (PrintTieredEvents) { 215 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 216 } 217 } 218 219 return max_task; 220 } 221 222 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 223 double queue_size = CompileBroker::queue_size(level); 224 int comp_count = compiler_count(level); 225 double k = queue_size / (feedback_k * comp_count) + 1; 226 227 // Increase C1 compile threshold when the code cache is filled more 228 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 229 // The main intention is to keep enough free space for C2 compiled code 230 // to achieve peak performance if the code cache is under stress. 231 if ((TieredStopAtLevel == CompLevel_full_optimization) && (level != CompLevel_full_optimization)) { 232 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(CodeCache::get_code_blob_type(level)); 233 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 234 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 235 } 236 } 237 return k; 238 } 239 240 // Call and loop predicates determine whether a transition to a higher 241 // compilation level should be performed (pointers to predicate functions 242 // are passed to common()). 243 // Tier?LoadFeedback is basically a coefficient that determines of 244 // how many methods per compiler thread can be in the queue before 245 // the threshold values double. 246 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level, Method* method) { 247 switch(cur_level) { 248 case CompLevel_none: 249 case CompLevel_limited_profile: { 250 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 251 return loop_predicate_helper<CompLevel_none>(i, b, k, method); 252 } 253 case CompLevel_full_profile: { 254 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 255 return loop_predicate_helper<CompLevel_full_profile>(i, b, k, method); 256 } 257 default: 258 return true; 259 } 260 } 261 262 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level, Method* method) { 263 switch(cur_level) { 264 case CompLevel_none: 265 case CompLevel_limited_profile: { 266 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 267 return call_predicate_helper<CompLevel_none>(i, b, k, method); 268 } 269 case CompLevel_full_profile: { 270 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 271 return call_predicate_helper<CompLevel_full_profile>(i, b, k, method); 272 } 273 default: 274 return true; 275 } 276 } 277 278 // If a method is old enough and is still in the interpreter we would want to 279 // start profiling without waiting for the compiled method to arrive. 280 // We also take the load on compilers into the account. 281 bool AdvancedThresholdPolicy::should_create_mdo(Method* method, CompLevel cur_level) { 282 if (cur_level == CompLevel_none && 283 CompileBroker::queue_size(CompLevel_full_optimization) <= 284 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 285 int i = method->invocation_count(); 286 int b = method->backedge_count(); 287 double k = Tier0ProfilingStartPercentage / 100.0; 288 return call_predicate_helper<CompLevel_none>(i, b, k, method) || loop_predicate_helper<CompLevel_none>(i, b, k, method); 289 } 290 return false; 291 } 292 293 // Inlining control: if we're compiling a profiled method with C1 and the callee 294 // is known to have OSRed in a C2 version, don't inline it. 295 bool AdvancedThresholdPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 296 CompLevel comp_level = (CompLevel)env->comp_level(); 297 if (comp_level == CompLevel_full_profile || 298 comp_level == CompLevel_limited_profile) { 299 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 300 } 301 return false; 302 } 303 304 // Create MDO if necessary. 305 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, JavaThread* THREAD) { 306 if (mh->is_native() || 307 mh->is_abstract() || 308 mh->is_accessor() || 309 mh->is_constant_getter()) { 310 return; 311 } 312 if (mh->method_data() == NULL) { 313 Method::build_interpreter_method_data(mh, CHECK_AND_CLEAR); 314 } 315 } 316 317 318 /* 319 * Method states: 320 * 0 - interpreter (CompLevel_none) 321 * 1 - pure C1 (CompLevel_simple) 322 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 323 * 3 - C1 with full profiling (CompLevel_full_profile) 324 * 4 - C2 (CompLevel_full_optimization) 325 * 326 * Common state transition patterns: 327 * a. 0 -> 3 -> 4. 328 * The most common path. But note that even in this straightforward case 329 * profiling can start at level 0 and finish at level 3. 330 * 331 * b. 0 -> 2 -> 3 -> 4. 332 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning 333 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 334 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 335 * 336 * c. 0 -> (3->2) -> 4. 337 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 338 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 339 * of the method to 2 while the request is still in-queue, because it'll allow it to run much faster 340 * without full profiling while c2 is compiling. 341 * 342 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 343 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 344 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 345 * 346 * e. 0 -> 4. 347 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 348 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 349 * the compiled version already exists). 350 * 351 * Note that since state 0 can be reached from any other state via deoptimization different loops 352 * are possible. 353 * 354 */ 355 356 // Common transition function. Given a predicate determines if a method should transition to another level. 357 CompLevel AdvancedThresholdPolicy::common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback) { 358 CompLevel next_level = cur_level; 359 int i = method->invocation_count(); 360 int b = method->backedge_count(); 361 362 if (is_trivial(method)) { 363 next_level = CompLevel_simple; 364 } else { 365 switch(cur_level) { 366 case CompLevel_none: 367 // If we were at full profile level, would we switch to full opt? 368 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { 369 next_level = CompLevel_full_optimization; 370 } else if ((this->*p)(i, b, cur_level, method)) { 371 #if INCLUDE_JVMCI 372 if (UseJVMCICompiler) { 373 // Since JVMCI takes a while to warm up, its queue inevitably backs up during 374 // early VM execution. 375 next_level = CompLevel_full_profile; 376 break; 377 } 378 #endif 379 // C1-generated fully profiled code is about 30% slower than the limited profile 380 // code that has only invocation and backedge counters. The observation is that 381 // if C2 queue is large enough we can spend too much time in the fully profiled code 382 // while waiting for C2 to pick the method from the queue. To alleviate this problem 383 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 384 // we choose to compile a limited profiled version and then recompile with full profiling 385 // when the load on C2 goes down. 386 if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > 387 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 388 next_level = CompLevel_limited_profile; 389 } else { 390 next_level = CompLevel_full_profile; 391 } 392 } 393 break; 394 case CompLevel_limited_profile: 395 if (is_method_profiled(method)) { 396 // Special case: we got here because this method was fully profiled in the interpreter. 397 next_level = CompLevel_full_optimization; 398 } else { 399 MethodData* mdo = method->method_data(); 400 if (mdo != NULL) { 401 if (mdo->would_profile()) { 402 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 403 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 404 (this->*p)(i, b, cur_level, method))) { 405 next_level = CompLevel_full_profile; 406 } 407 } else { 408 next_level = CompLevel_full_optimization; 409 } 410 } 411 } 412 break; 413 case CompLevel_full_profile: 414 { 415 MethodData* mdo = method->method_data(); 416 if (mdo != NULL) { 417 if (mdo->would_profile()) { 418 int mdo_i = mdo->invocation_count_delta(); 419 int mdo_b = mdo->backedge_count_delta(); 420 if ((this->*p)(mdo_i, mdo_b, cur_level, method)) { 421 next_level = CompLevel_full_optimization; 422 } 423 } else { 424 next_level = CompLevel_full_optimization; 425 } 426 } 427 } 428 break; 429 } 430 } 431 return MIN2(next_level, (CompLevel)TieredStopAtLevel); 432 } 433 434 // Determine if a method should be compiled with a normal entry point at a different level. 435 CompLevel AdvancedThresholdPolicy::call_event(Method* method, CompLevel cur_level) { 436 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), 437 common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true)); 438 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 439 440 // If OSR method level is greater than the regular method level, the levels should be 441 // equalized by raising the regular method level in order to avoid OSRs during each 442 // invocation of the method. 443 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 444 MethodData* mdo = method->method_data(); 445 guarantee(mdo != NULL, "MDO should not be NULL"); 446 if (mdo->invocation_count() >= 1) { 447 next_level = CompLevel_full_optimization; 448 } 449 } else { 450 next_level = MAX2(osr_level, next_level); 451 } 452 return next_level; 453 } 454 455 // Determine if we should do an OSR compilation of a given method. 456 CompLevel AdvancedThresholdPolicy::loop_event(Method* method, CompLevel cur_level) { 457 CompLevel next_level = common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true); 458 if (cur_level == CompLevel_none) { 459 // If there is a live OSR method that means that we deopted to the interpreter 460 // for the transition. 461 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 462 if (osr_level > CompLevel_none) { 463 return osr_level; 464 } 465 } 466 return next_level; 467 } 468 469 // Update the rate and submit compile 470 void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, JavaThread* thread) { 471 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 472 update_rate(os::javaTimeMillis(), mh()); 473 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", thread); 474 } 475 476 // Handle the invocation event. 477 void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh, 478 CompLevel level, nmethod* nm, JavaThread* thread) { 479 if (should_create_mdo(mh(), level)) { 480 create_mdo(mh, thread); 481 } 482 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 483 CompLevel next_level = call_event(mh(), level); 484 if (next_level != level) { 485 compile(mh, InvocationEntryBci, next_level, thread); 486 } 487 } 488 } 489 490 // Handle the back branch event. Notice that we can compile the method 491 // with a regular entry from here. 492 void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh, 493 int bci, CompLevel level, nmethod* nm, JavaThread* thread) { 494 if (should_create_mdo(mh(), level)) { 495 create_mdo(mh, thread); 496 } 497 // Check if MDO should be created for the inlined method 498 if (should_create_mdo(imh(), level)) { 499 create_mdo(imh, thread); 500 } 501 502 if (is_compilation_enabled()) { 503 CompLevel next_osr_level = loop_event(imh(), level); 504 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 505 // At the very least compile the OSR version 506 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 507 compile(imh, bci, next_osr_level, thread); 508 } 509 510 // Use loop event as an opportunity to also check if there's been 511 // enough calls. 512 CompLevel cur_level, next_level; 513 if (mh() != imh()) { // If there is an enclosing method 514 guarantee(nm != NULL, "Should have nmethod here"); 515 cur_level = comp_level(mh()); 516 next_level = call_event(mh(), cur_level); 517 518 if (max_osr_level == CompLevel_full_optimization) { 519 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 520 bool make_not_entrant = false; 521 if (nm->is_osr_method()) { 522 // This is an osr method, just make it not entrant and recompile later if needed 523 make_not_entrant = true; 524 } else { 525 if (next_level != CompLevel_full_optimization) { 526 // next_level is not full opt, so we need to recompile the 527 // enclosing method without the inlinee 528 cur_level = CompLevel_none; 529 make_not_entrant = true; 530 } 531 } 532 if (make_not_entrant) { 533 if (PrintTieredEvents) { 534 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 535 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 536 } 537 nm->make_not_entrant(); 538 } 539 } 540 if (!CompileBroker::compilation_is_in_queue(mh)) { 541 // Fix up next_level if necessary to avoid deopts 542 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 543 next_level = CompLevel_full_profile; 544 } 545 if (cur_level != next_level) { 546 compile(mh, InvocationEntryBci, next_level, thread); 547 } 548 } 549 } else { 550 cur_level = comp_level(imh()); 551 next_level = call_event(imh(), cur_level); 552 if (!CompileBroker::compilation_is_in_queue(imh) && (next_level != cur_level)) { 553 compile(imh, InvocationEntryBci, next_level, thread); 554 } 555 } 556 } 557 } 558 559 #endif // TIERED