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 (double)(method->rate() + 1) * 137 (method->invocation_count() + 1) * (method->backedge_count() + 1); 138 } 139 140 // Apply heuristics and return true if x should be compiled before y 141 bool AdvancedThresholdPolicy::compare_methods(Method* x, Method* y) { 142 if (x->highest_comp_level() > y->highest_comp_level()) { 143 // recompilation after deopt 144 return true; 145 } else 146 if (x->highest_comp_level() == y->highest_comp_level()) { 147 if (weight(x) > weight(y)) { 148 return true; 149 } 150 } 151 return false; 152 } 153 154 // Is method profiled enough? 155 bool AdvancedThresholdPolicy::is_method_profiled(Method* method) { 156 MethodData* mdo = method->method_data(); 157 if (mdo != NULL) { 158 int i = mdo->invocation_count_delta(); 159 int b = mdo->backedge_count_delta(); 160 return call_predicate_helper<CompLevel_full_profile>(i, b, 1, method); 161 } 162 return false; 163 } 164 165 // Called with the queue locked and with at least one element 166 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { 167 #if INCLUDE_JVMCI 168 CompileTask *max_non_jvmci_task = NULL; 169 #endif 170 CompileTask *max_task = NULL; 171 Method* max_method = NULL; 172 jlong t = os::javaTimeMillis(); 173 // Iterate through the queue and find a method with a maximum rate. 174 for (CompileTask* task = compile_queue->first(); task != NULL;) { 175 CompileTask* next_task = task->next(); 176 Method* method = task->method(); 177 update_rate(t, method); 178 if (max_task == NULL) { 179 max_task = task; 180 max_method = method; 181 } else { 182 // If a method has been stale for some time, remove it from the queue. 183 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { 184 if (PrintTieredEvents) { 185 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel)task->comp_level()); 186 } 187 task->log_task_dequeued("stale"); 188 compile_queue->remove_and_mark_stale(task); 189 method->clear_queued_for_compilation(); 190 task = next_task; 191 continue; 192 } 193 194 // Select a method with a higher rate 195 if (compare_methods(method, max_method)) { 196 max_task = task; 197 max_method = method; 198 } 199 } 200 task = next_task; 201 } 202 203 #if INCLUDE_JVMCI 204 if (UseJVMCICompiler) { 205 if (max_non_jvmci_task != NULL) { 206 max_task = max_non_jvmci_task; 207 max_method = max_task->method(); 208 } 209 } 210 #endif 211 212 if (max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile 213 && is_method_profiled(max_method)) { 214 max_task->set_comp_level(CompLevel_limited_profile); 215 if (PrintTieredEvents) { 216 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 217 } 218 } 219 220 return max_task; 221 } 222 223 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 224 double queue_size = CompileBroker::queue_size(level); 225 int comp_count = compiler_count(level); 226 double k = queue_size / (feedback_k * comp_count) + 1; 227 228 // Increase C1 compile threshold when the code cache is filled more 229 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 230 // The main intention is to keep enough free space for C2 compiled code 231 // to achieve peak performance if the code cache is under stress. 232 if ((TieredStopAtLevel == CompLevel_full_optimization) && (level != CompLevel_full_optimization)) { 233 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(CodeCache::get_code_blob_type(level)); 234 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 235 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 236 } 237 } 238 return k; 239 } 240 241 // Call and loop predicates determine whether a transition to a higher 242 // compilation level should be performed (pointers to predicate functions 243 // are passed to common()). 244 // Tier?LoadFeedback is basically a coefficient that determines of 245 // how many methods per compiler thread can be in the queue before 246 // the threshold values double. 247 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level, Method* method) { 248 switch(cur_level) { 249 case CompLevel_none: 250 case CompLevel_limited_profile: { 251 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 252 return loop_predicate_helper<CompLevel_none>(i, b, k, method); 253 } 254 case CompLevel_full_profile: { 255 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 256 return loop_predicate_helper<CompLevel_full_profile>(i, b, k, method); 257 } 258 default: 259 return true; 260 } 261 } 262 263 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level, Method* method) { 264 switch(cur_level) { 265 case CompLevel_none: 266 case CompLevel_limited_profile: { 267 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 268 return call_predicate_helper<CompLevel_none>(i, b, k, method); 269 } 270 case CompLevel_full_profile: { 271 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 272 return call_predicate_helper<CompLevel_full_profile>(i, b, k, method); 273 } 274 default: 275 return true; 276 } 277 } 278 279 // If a method is old enough and is still in the interpreter we would want to 280 // start profiling without waiting for the compiled method to arrive. 281 // We also take the load on compilers into the account. 282 bool AdvancedThresholdPolicy::should_create_mdo(Method* method, CompLevel cur_level) { 283 if (cur_level == CompLevel_none && 284 CompileBroker::queue_size(CompLevel_full_optimization) <= 285 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 286 int i = method->invocation_count(); 287 int b = method->backedge_count(); 288 double k = Tier0ProfilingStartPercentage / 100.0; 289 return call_predicate_helper<CompLevel_none>(i, b, k, method) || loop_predicate_helper<CompLevel_none>(i, b, k, method); 290 } 291 return false; 292 } 293 294 // Inlining control: if we're compiling a profiled method with C1 and the callee 295 // is known to have OSRed in a C2 version, don't inline it. 296 bool AdvancedThresholdPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 297 CompLevel comp_level = (CompLevel)env->comp_level(); 298 if (comp_level == CompLevel_full_profile || 299 comp_level == CompLevel_limited_profile) { 300 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 301 } 302 return false; 303 } 304 305 // Create MDO if necessary. 306 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, JavaThread* THREAD) { 307 if (mh->is_native() || 308 mh->is_abstract() || 309 mh->is_accessor() || 310 mh->is_constant_getter()) { 311 return; 312 } 313 if (mh->method_data() == NULL) { 314 Method::build_interpreter_method_data(mh, CHECK_AND_CLEAR); 315 } 316 } 317 318 319 /* 320 * Method states: 321 * 0 - interpreter (CompLevel_none) 322 * 1 - pure C1 (CompLevel_simple) 323 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 324 * 3 - C1 with full profiling (CompLevel_full_profile) 325 * 4 - C2 (CompLevel_full_optimization) 326 * 327 * Common state transition patterns: 328 * a. 0 -> 3 -> 4. 329 * The most common path. But note that even in this straightforward case 330 * profiling can start at level 0 and finish at level 3. 331 * 332 * b. 0 -> 2 -> 3 -> 4. 333 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning 334 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 335 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 336 * 337 * c. 0 -> (3->2) -> 4. 338 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 339 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 340 * of the method to 2 while the request is still in-queue, because it'll allow it to run much faster 341 * without full profiling while c2 is compiling. 342 * 343 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 344 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 345 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 346 * 347 * e. 0 -> 4. 348 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 349 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 350 * the compiled version already exists). 351 * 352 * Note that since state 0 can be reached from any other state via deoptimization different loops 353 * are possible. 354 * 355 */ 356 357 // Common transition function. Given a predicate determines if a method should transition to another level. 358 CompLevel AdvancedThresholdPolicy::common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback) { 359 CompLevel next_level = cur_level; 360 int i = method->invocation_count(); 361 int b = method->backedge_count(); 362 363 if (is_trivial(method)) { 364 next_level = CompLevel_simple; 365 } else { 366 switch(cur_level) { 367 case CompLevel_none: 368 // If we were at full profile level, would we switch to full opt? 369 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { 370 next_level = CompLevel_full_optimization; 371 } else if ((this->*p)(i, b, cur_level, method)) { 372 #if INCLUDE_JVMCI 373 if (UseJVMCICompiler) { 374 // Since JVMCI takes a while to warm up, its queue inevitably backs up during 375 // early VM execution. 376 next_level = CompLevel_full_profile; 377 break; 378 } 379 #endif 380 // C1-generated fully profiled code is about 30% slower than the limited profile 381 // code that has only invocation and backedge counters. The observation is that 382 // if C2 queue is large enough we can spend too much time in the fully profiled code 383 // while waiting for C2 to pick the method from the queue. To alleviate this problem 384 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 385 // we choose to compile a limited profiled version and then recompile with full profiling 386 // when the load on C2 goes down. 387 if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > 388 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 389 next_level = CompLevel_limited_profile; 390 } else { 391 next_level = CompLevel_full_profile; 392 } 393 } 394 break; 395 case CompLevel_limited_profile: 396 if (is_method_profiled(method)) { 397 // Special case: we got here because this method was fully profiled in the interpreter. 398 next_level = CompLevel_full_optimization; 399 } else { 400 MethodData* mdo = method->method_data(); 401 if (mdo != NULL) { 402 if (mdo->would_profile()) { 403 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 404 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 405 (this->*p)(i, b, cur_level, method))) { 406 next_level = CompLevel_full_profile; 407 } 408 } else { 409 next_level = CompLevel_full_optimization; 410 } 411 } 412 } 413 break; 414 case CompLevel_full_profile: 415 { 416 MethodData* mdo = method->method_data(); 417 if (mdo != NULL) { 418 if (mdo->would_profile()) { 419 int mdo_i = mdo->invocation_count_delta(); 420 int mdo_b = mdo->backedge_count_delta(); 421 if ((this->*p)(mdo_i, mdo_b, cur_level, method)) { 422 next_level = CompLevel_full_optimization; 423 } 424 } else { 425 next_level = CompLevel_full_optimization; 426 } 427 } 428 } 429 break; 430 } 431 } 432 return MIN2(next_level, (CompLevel)TieredStopAtLevel); 433 } 434 435 // Determine if a method should be compiled with a normal entry point at a different level. 436 CompLevel AdvancedThresholdPolicy::call_event(Method* method, CompLevel cur_level) { 437 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), 438 common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true)); 439 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 440 441 // If OSR method level is greater than the regular method level, the levels should be 442 // equalized by raising the regular method level in order to avoid OSRs during each 443 // invocation of the method. 444 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 445 MethodData* mdo = method->method_data(); 446 guarantee(mdo != NULL, "MDO should not be NULL"); 447 if (mdo->invocation_count() >= 1) { 448 next_level = CompLevel_full_optimization; 449 } 450 } else { 451 next_level = MAX2(osr_level, next_level); 452 } 453 return next_level; 454 } 455 456 // Determine if we should do an OSR compilation of a given method. 457 CompLevel AdvancedThresholdPolicy::loop_event(Method* method, CompLevel cur_level) { 458 CompLevel next_level = common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true); 459 if (cur_level == CompLevel_none) { 460 // If there is a live OSR method that means that we deopted to the interpreter 461 // for the transition. 462 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 463 if (osr_level > CompLevel_none) { 464 return osr_level; 465 } 466 } 467 return next_level; 468 } 469 470 // Update the rate and submit compile 471 void AdvancedThresholdPolicy::submit_compile(const methodHandle& mh, int bci, CompLevel level, JavaThread* thread) { 472 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 473 update_rate(os::javaTimeMillis(), mh()); 474 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", thread); 475 } 476 477 // Handle the invocation event. 478 void AdvancedThresholdPolicy::method_invocation_event(const methodHandle& mh, const methodHandle& imh, 479 CompLevel level, nmethod* nm, JavaThread* thread) { 480 if (should_create_mdo(mh(), level)) { 481 create_mdo(mh, thread); 482 } 483 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 484 CompLevel next_level = call_event(mh(), level); 485 if (next_level != level) { 486 compile(mh, InvocationEntryBci, next_level, thread); 487 } 488 } 489 } 490 491 // Handle the back branch event. Notice that we can compile the method 492 // with a regular entry from here. 493 void AdvancedThresholdPolicy::method_back_branch_event(const methodHandle& mh, const methodHandle& imh, 494 int bci, CompLevel level, nmethod* nm, JavaThread* thread) { 495 if (should_create_mdo(mh(), level)) { 496 create_mdo(mh, thread); 497 } 498 // Check if MDO should be created for the inlined method 499 if (should_create_mdo(imh(), level)) { 500 create_mdo(imh, thread); 501 } 502 503 if (is_compilation_enabled()) { 504 CompLevel next_osr_level = loop_event(imh(), level); 505 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 506 // At the very least compile the OSR version 507 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 508 compile(imh, bci, next_osr_level, thread); 509 } 510 511 // Use loop event as an opportunity to also check if there's been 512 // enough calls. 513 CompLevel cur_level, next_level; 514 if (mh() != imh()) { // If there is an enclosing method 515 guarantee(nm != NULL, "Should have nmethod here"); 516 cur_level = comp_level(mh()); 517 next_level = call_event(mh(), cur_level); 518 519 if (max_osr_level == CompLevel_full_optimization) { 520 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 521 bool make_not_entrant = false; 522 if (nm->is_osr_method()) { 523 // This is an osr method, just make it not entrant and recompile later if needed 524 make_not_entrant = true; 525 } else { 526 if (next_level != CompLevel_full_optimization) { 527 // next_level is not full opt, so we need to recompile the 528 // enclosing method without the inlinee 529 cur_level = CompLevel_none; 530 make_not_entrant = true; 531 } 532 } 533 if (make_not_entrant) { 534 if (PrintTieredEvents) { 535 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 536 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 537 } 538 nm->make_not_entrant(); 539 } 540 } 541 if (!CompileBroker::compilation_is_in_queue(mh)) { 542 // Fix up next_level if necessary to avoid deopts 543 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 544 next_level = CompLevel_full_profile; 545 } 546 if (cur_level != next_level) { 547 compile(mh, InvocationEntryBci, next_level, thread); 548 } 549 } 550 } else { 551 cur_level = comp_level(imh()); 552 next_level = call_event(imh(), cur_level); 553 if (!CompileBroker::compilation_is_in_queue(imh) && (next_level != cur_level)) { 554 compile(imh, InvocationEntryBci, next_level, thread); 555 } 556 } 557 } 558 } 559 560 #endif // TIERED