1 /* 2 * Copyright (c) 2010, 2019, 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 "runtime/advancedThresholdPolicy.hpp" 27 #include "runtime/simpleThresholdPolicy.inline.hpp" 28 29 #ifdef TIERED 30 // Print an event. 31 void AdvancedThresholdPolicy::print_specific(EventType type, methodHandle mh, methodHandle imh, 32 int bci, CompLevel level) { 33 tty->print(" rate="); 34 if (mh->prev_time() == 0) tty->print("n/a"); 35 else tty->print("%f", mh->rate()); 36 37 tty->print(" k=%.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback), 38 threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback)); 39 40 } 41 42 void AdvancedThresholdPolicy::initialize() { 43 // Turn on ergonomic compiler count selection 44 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { 45 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); 46 } 47 int count = CICompilerCount; 48 if (CICompilerCountPerCPU) { 49 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n 50 int log_cpu = log2_int(os::active_processor_count()); 51 int loglog_cpu = log2_int(MAX2(log_cpu, 1)); 52 count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2; 53 } 54 55 set_c1_count(MAX2(count / 3, 1)); 56 set_c2_count(MAX2(count - c1_count(), 1)); 57 FLAG_SET_ERGO(intx, CICompilerCount, c1_count() + c2_count()); 58 59 // Some inlining tuning 60 #if defined(X86) || defined(AARCH64) 61 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 62 FLAG_SET_DEFAULT(InlineSmallCode, 2000); 63 } 64 #endif 65 66 #ifdef SPARC 67 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 68 FLAG_SET_DEFAULT(InlineSmallCode, 2500); 69 } 70 #endif 71 72 set_increase_threshold_at_ratio(); 73 set_start_time(os::javaTimeMillis()); 74 } 75 76 // update_rate() is called from select_task() while holding a compile queue lock. 77 void AdvancedThresholdPolicy::update_rate(jlong t, Method* m) { 78 // Skip update if counters are absent. 79 // Can't allocate them since we are holding compile queue lock. 80 if (m->method_counters() == NULL) return; 81 82 if (is_old(m)) { 83 // We don't remove old methods from the queue, 84 // so we can just zero the rate. 85 m->set_rate(0); 86 return; 87 } 88 89 // We don't update the rate if we've just came out of a safepoint. 90 // delta_s is the time since last safepoint in milliseconds. 91 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 92 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement 93 // How many events were there since the last time? 94 int event_count = m->invocation_count() + m->backedge_count(); 95 int delta_e = event_count - m->prev_event_count(); 96 97 // We should be running for at least 1ms. 98 if (delta_s >= TieredRateUpdateMinTime) { 99 // And we must've taken the previous point at least 1ms before. 100 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 101 m->set_prev_time(t); 102 m->set_prev_event_count(event_count); 103 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 104 } else { 105 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 106 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 107 m->set_rate(0); 108 } 109 } 110 } 111 } 112 113 // Check if this method has been stale from a given number of milliseconds. 114 // See select_task(). 115 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, Method* m) { 116 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 117 jlong delta_t = t - m->prev_time(); 118 if (delta_t > timeout && delta_s > timeout) { 119 int event_count = m->invocation_count() + m->backedge_count(); 120 int delta_e = event_count - m->prev_event_count(); 121 // Return true if there were no events. 122 return delta_e == 0; 123 } 124 return false; 125 } 126 127 // We don't remove old methods from the compile queue even if they have 128 // very low activity. See select_task(). 129 bool AdvancedThresholdPolicy::is_old(Method* method) { 130 return method->invocation_count() > 50000 || method->backedge_count() > 500000; 131 } 132 133 double AdvancedThresholdPolicy::weight(Method* method) { 134 return (double)(method->rate() + 1) * 135 (method->invocation_count() + 1) * (method->backedge_count() + 1); 136 } 137 138 // Apply heuristics and return true if x should be compiled before y 139 bool AdvancedThresholdPolicy::compare_methods(Method* x, Method* y) { 140 if (x->highest_comp_level() > y->highest_comp_level()) { 141 // recompilation after deopt 142 return true; 143 } else 144 if (x->highest_comp_level() == y->highest_comp_level()) { 145 if (weight(x) > weight(y)) { 146 return true; 147 } 148 } 149 return false; 150 } 151 152 // Is method profiled enough? 153 bool AdvancedThresholdPolicy::is_method_profiled(Method* method) { 154 MethodData* mdo = method->method_data(); 155 if (mdo != NULL) { 156 int i = mdo->invocation_count_delta(); 157 int b = mdo->backedge_count_delta(); 158 return call_predicate_helper<CompLevel_full_profile>(i, b, 1); 159 } 160 return false; 161 } 162 163 // Called with the queue locked and with at least one element 164 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { 165 CompileTask *max_task = NULL; 166 Method* max_method = NULL; 167 jlong t = os::javaTimeMillis(); 168 // Iterate through the queue and find a method with a maximum rate. 169 for (CompileTask* task = compile_queue->first(); task != NULL;) { 170 CompileTask* next_task = task->next(); 171 Method* method = task->method(); 172 update_rate(t, method); 173 if (max_task == NULL) { 174 max_task = task; 175 max_method = method; 176 } else { 177 // If a method has been stale for some time, remove it from the queue. 178 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { 179 if (PrintTieredEvents) { 180 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel)task->comp_level()); 181 } 182 compile_queue->remove_and_mark_stale(task); 183 method->clear_queued_for_compilation(); 184 task = next_task; 185 continue; 186 } 187 188 // Select a method with a higher rate 189 if (compare_methods(method, max_method)) { 190 max_task = task; 191 max_method = method; 192 } 193 } 194 task = next_task; 195 } 196 197 if (max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile 198 && is_method_profiled(max_method)) { 199 200 if (CompileBroker::compilation_is_complete(max_method, max_task->osr_bci(), CompLevel_limited_profile)) { 201 if (PrintTieredEvents) { 202 print_event(REMOVE_FROM_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 203 } 204 compile_queue->remove_and_mark_stale(max_task); 205 max_method->clear_queued_for_compilation(); 206 return NULL; 207 } 208 209 max_task->set_comp_level(CompLevel_limited_profile); 210 if (PrintTieredEvents) { 211 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 212 } 213 } 214 215 return max_task; 216 } 217 218 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 219 double queue_size = CompileBroker::queue_size(level); 220 int comp_count = compiler_count(level); 221 double k = queue_size / (feedback_k * comp_count) + 1; 222 223 // Increase C1 compile threshold when the code cache is filled more 224 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 225 // The main intention is to keep enough free space for C2 compiled code 226 // to achieve peak performance if the code cache is under stress. 227 if ((TieredStopAtLevel == CompLevel_full_optimization) && (level != CompLevel_full_optimization)) { 228 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(); 229 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 230 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 231 } 232 } 233 return k; 234 } 235 236 // Call and loop predicates determine whether a transition to a higher 237 // compilation level should be performed (pointers to predicate functions 238 // are passed to common()). 239 // Tier?LoadFeedback is basically a coefficient that determines of 240 // how many methods per compiler thread can be in the queue before 241 // the threshold values double. 242 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level) { 243 switch(cur_level) { 244 case CompLevel_none: 245 case CompLevel_limited_profile: { 246 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 247 return loop_predicate_helper<CompLevel_none>(i, b, k); 248 } 249 case CompLevel_full_profile: { 250 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 251 return loop_predicate_helper<CompLevel_full_profile>(i, b, k); 252 } 253 default: 254 return true; 255 } 256 } 257 258 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level) { 259 switch(cur_level) { 260 case CompLevel_none: 261 case CompLevel_limited_profile: { 262 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 263 return call_predicate_helper<CompLevel_none>(i, b, k); 264 } 265 case CompLevel_full_profile: { 266 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 267 return call_predicate_helper<CompLevel_full_profile>(i, b, k); 268 } 269 default: 270 return true; 271 } 272 } 273 274 // If a method is old enough and is still in the interpreter we would want to 275 // start profiling without waiting for the compiled method to arrive. 276 // We also take the load on compilers into the account. 277 bool AdvancedThresholdPolicy::should_create_mdo(Method* method, CompLevel cur_level) { 278 if (cur_level == CompLevel_none && 279 CompileBroker::queue_size(CompLevel_full_optimization) <= 280 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 281 int i = method->invocation_count(); 282 int b = method->backedge_count(); 283 double k = Tier0ProfilingStartPercentage / 100.0; 284 return call_predicate_helper<CompLevel_none>(i, b, k) || loop_predicate_helper<CompLevel_none>(i, b, k); 285 } 286 return false; 287 } 288 289 // Inlining control: if we're compiling a profiled method with C1 and the callee 290 // is known to have OSRed in a C2 version, don't inline it. 291 bool AdvancedThresholdPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 292 CompLevel comp_level = (CompLevel)env->comp_level(); 293 if (comp_level == CompLevel_full_profile || 294 comp_level == CompLevel_limited_profile) { 295 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 296 } 297 return false; 298 } 299 300 // Create MDO if necessary. 301 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, JavaThread* THREAD) { 302 if (mh->is_native() || mh->is_abstract() || mh->is_accessor()) return; 303 if (mh->method_data() == NULL) { 304 Method::build_interpreter_method_data(mh, CHECK_AND_CLEAR); 305 } 306 } 307 308 309 /* 310 * Method states: 311 * 0 - interpreter (CompLevel_none) 312 * 1 - pure C1 (CompLevel_simple) 313 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 314 * 3 - C1 with full profiling (CompLevel_full_profile) 315 * 4 - C2 (CompLevel_full_optimization) 316 * 317 * Common state transition patterns: 318 * a. 0 -> 3 -> 4. 319 * The most common path. But note that even in this straightforward case 320 * profiling can start at level 0 and finish at level 3. 321 * 322 * b. 0 -> 2 -> 3 -> 4. 323 * This case occures when the load on C2 is deemed too high. So, instead of transitioning 324 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 325 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 326 * 327 * c. 0 -> (3->2) -> 4. 328 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 329 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 330 * of the method to 2 while the request is still in-queue, because it'll allow it to run much faster 331 * without full profiling while c2 is compiling. 332 * 333 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 334 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 335 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 336 * 337 * e. 0 -> 4. 338 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 339 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 340 * the compiled version already exists). 341 * 342 * Note that since state 0 can be reached from any other state via deoptimization different loops 343 * are possible. 344 * 345 */ 346 347 // Common transition function. Given a predicate determines if a method should transition to another level. 348 CompLevel AdvancedThresholdPolicy::common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback) { 349 CompLevel next_level = cur_level; 350 int i = method->invocation_count(); 351 int b = method->backedge_count(); 352 353 if (is_trivial(method)) { 354 next_level = CompLevel_simple; 355 } else { 356 switch(cur_level) { 357 case CompLevel_none: 358 // If we were at full profile level, would we switch to full opt? 359 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { 360 next_level = CompLevel_full_optimization; 361 } else if ((this->*p)(i, b, cur_level)) { 362 // C1-generated fully profiled code is about 30% slower than the limited profile 363 // code that has only invocation and backedge counters. The observation is that 364 // if C2 queue is large enough we can spend too much time in the fully profiled code 365 // while waiting for C2 to pick the method from the queue. To alleviate this problem 366 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 367 // we choose to compile a limited profiled version and then recompile with full profiling 368 // when the load on C2 goes down. 369 if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > 370 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 371 next_level = CompLevel_limited_profile; 372 } else { 373 next_level = CompLevel_full_profile; 374 } 375 } 376 break; 377 case CompLevel_limited_profile: 378 if (is_method_profiled(method)) { 379 // Special case: we got here because this method was fully profiled in the interpreter. 380 next_level = CompLevel_full_optimization; 381 } else { 382 MethodData* mdo = method->method_data(); 383 if (mdo != NULL) { 384 if (mdo->would_profile()) { 385 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 386 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 387 (this->*p)(i, b, cur_level))) { 388 next_level = CompLevel_full_profile; 389 } 390 } else { 391 next_level = CompLevel_full_optimization; 392 } 393 } 394 } 395 break; 396 case CompLevel_full_profile: 397 { 398 MethodData* mdo = method->method_data(); 399 if (mdo != NULL) { 400 if (mdo->would_profile()) { 401 int mdo_i = mdo->invocation_count_delta(); 402 int mdo_b = mdo->backedge_count_delta(); 403 if ((this->*p)(mdo_i, mdo_b, cur_level)) { 404 next_level = CompLevel_full_optimization; 405 } 406 } else { 407 next_level = CompLevel_full_optimization; 408 } 409 } 410 } 411 break; 412 } 413 } 414 return MIN2(next_level, (CompLevel)TieredStopAtLevel); 415 } 416 417 // Determine if a method should be compiled with a normal entry point at a different level. 418 CompLevel AdvancedThresholdPolicy::call_event(Method* method, CompLevel cur_level) { 419 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), 420 common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true)); 421 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 422 423 // If OSR method level is greater than the regular method level, the levels should be 424 // equalized by raising the regular method level in order to avoid OSRs during each 425 // invocation of the method. 426 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 427 MethodData* mdo = method->method_data(); 428 guarantee(mdo != NULL, "MDO should not be NULL"); 429 if (mdo->invocation_count() >= 1) { 430 next_level = CompLevel_full_optimization; 431 } 432 } else { 433 next_level = MAX2(osr_level, next_level); 434 } 435 return next_level; 436 } 437 438 // Determine if we should do an OSR compilation of a given method. 439 CompLevel AdvancedThresholdPolicy::loop_event(Method* method, CompLevel cur_level) { 440 CompLevel next_level = common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true); 441 if (cur_level == CompLevel_none) { 442 // If there is a live OSR method that means that we deopted to the interpreter 443 // for the transition. 444 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 445 if (osr_level > CompLevel_none) { 446 return osr_level; 447 } 448 } 449 return next_level; 450 } 451 452 // Update the rate and submit compile 453 void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, JavaThread* thread) { 454 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 455 update_rate(os::javaTimeMillis(), mh()); 456 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", thread); 457 } 458 459 // Handle the invocation event. 460 void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh, 461 CompLevel level, nmethod* nm, JavaThread* thread) { 462 if (should_create_mdo(mh(), level)) { 463 create_mdo(mh, thread); 464 } 465 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 466 CompLevel next_level = call_event(mh(), level); 467 if (next_level != level) { 468 compile(mh, InvocationEntryBci, next_level, thread); 469 } 470 } 471 } 472 473 // Handle the back branch event. Notice that we can compile the method 474 // with a regular entry from here. 475 void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh, 476 int bci, CompLevel level, nmethod* nm, JavaThread* thread) { 477 if (should_create_mdo(mh(), level)) { 478 create_mdo(mh, thread); 479 } 480 // Check if MDO should be created for the inlined method 481 if (should_create_mdo(imh(), level)) { 482 create_mdo(imh, thread); 483 } 484 485 if (is_compilation_enabled()) { 486 CompLevel next_osr_level = loop_event(imh(), level); 487 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 488 // At the very least compile the OSR version 489 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 490 compile(imh, bci, next_osr_level, thread); 491 } 492 493 // Use loop event as an opportunity to also check if there's been 494 // enough calls. 495 CompLevel cur_level, next_level; 496 if (mh() != imh()) { // If there is an enclosing method 497 guarantee(nm != NULL, "Should have nmethod here"); 498 cur_level = comp_level(mh()); 499 next_level = call_event(mh(), cur_level); 500 501 if (max_osr_level == CompLevel_full_optimization) { 502 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 503 bool make_not_entrant = false; 504 if (nm->is_osr_method()) { 505 // This is an osr method, just make it not entrant and recompile later if needed 506 make_not_entrant = true; 507 } else { 508 if (next_level != CompLevel_full_optimization) { 509 // next_level is not full opt, so we need to recompile the 510 // enclosing method without the inlinee 511 cur_level = CompLevel_none; 512 make_not_entrant = true; 513 } 514 } 515 if (make_not_entrant) { 516 if (PrintTieredEvents) { 517 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 518 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 519 } 520 nm->make_not_entrant(); 521 } 522 } 523 if (!CompileBroker::compilation_is_in_queue(mh)) { 524 // Fix up next_level if necessary to avoid deopts 525 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 526 next_level = CompLevel_full_profile; 527 } 528 if (cur_level != next_level) { 529 compile(mh, InvocationEntryBci, next_level, thread); 530 } 531 } 532 } else { 533 cur_level = comp_level(imh()); 534 next_level = call_event(imh(), cur_level); 535 if (!CompileBroker::compilation_is_in_queue(imh) && (next_level != cur_level)) { 536 compile(imh, InvocationEntryBci, next_level, thread); 537 } 538 } 539 } 540 } 541 542 #endif // TIERED