1 /* 2 * Copyright (c) 2010, 2014, 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 "runtime/advancedThresholdPolicy.hpp" 28 #include "runtime/simpleThresholdPolicy.inline.hpp" 29 30 #ifdef TIERED 31 // Print an event. 32 void AdvancedThresholdPolicy::print_specific(EventType type, methodHandle mh, methodHandle imh, 33 int bci, CompLevel level) { 34 tty->print(" rate="); 35 if (mh->prev_time() == 0) tty->print("n/a"); 36 else tty->print("%f", mh->rate()); 37 38 tty->print(" k=%.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback), 39 threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback)); 40 41 } 42 43 void AdvancedThresholdPolicy::initialize() { 44 // Turn on ergonomic compiler count selection 45 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { 46 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); 47 } 48 int count = CICompilerCount; 49 if (CICompilerCountPerCPU) { 50 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n 51 int log_cpu = log2_intptr(os::active_processor_count()); 52 int loglog_cpu = log2_intptr(MAX2(log_cpu, 1)); 53 count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2; 54 } 55 56 set_c1_count(MAX2(count / 3, 1)); 57 set_c2_count(MAX2(count - c1_count(), 1)); 58 FLAG_SET_ERGO(intx, CICompilerCount, c1_count() + c2_count()); 59 60 // Some inlining tuning 61 #ifdef X86 62 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 63 FLAG_SET_DEFAULT(InlineSmallCode, 2000); 64 } 65 #endif 66 67 #ifdef SPARC 68 if (FLAG_IS_DEFAULT(InlineSmallCode)) { 69 FLAG_SET_DEFAULT(InlineSmallCode, 2500); 70 } 71 #endif 72 73 set_increase_threshold_at_ratio(); 74 set_start_time(os::javaTimeMillis()); 75 } 76 77 // update_rate() is called from select_task() while holding a compile queue lock. 78 void AdvancedThresholdPolicy::update_rate(jlong t, Method* m) { 79 // Skip update if counters are absent. 80 // Can't allocate them since we are holding compile queue lock. 81 if (m->method_counters() == NULL) return; 82 83 if (is_old(m)) { 84 // We don't remove old methods from the queue, 85 // so we can just zero the rate. 86 m->set_rate(0); 87 return; 88 } 89 90 // We don't update the rate if we've just came out of a safepoint. 91 // delta_s is the time since last safepoint in milliseconds. 92 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 93 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement 94 // How many events were there since the last time? 95 int event_count = m->invocation_count() + m->backedge_count(); 96 int delta_e = event_count - m->prev_event_count(); 97 98 // We should be running for at least 1ms. 99 if (delta_s >= TieredRateUpdateMinTime) { 100 // And we must've taken the previous point at least 1ms before. 101 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 102 m->set_prev_time(t); 103 m->set_prev_event_count(event_count); 104 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 105 } else { 106 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 107 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 108 m->set_rate(0); 109 } 110 } 111 } 112 } 113 114 // Check if this method has been stale from a given number of milliseconds. 115 // See select_task(). 116 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, Method* m) { 117 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 118 jlong delta_t = t - m->prev_time(); 119 if (delta_t > timeout && delta_s > timeout) { 120 int event_count = m->invocation_count() + m->backedge_count(); 121 int delta_e = event_count - m->prev_event_count(); 122 // Return true if there were no events. 123 return delta_e == 0; 124 } 125 return false; 126 } 127 128 // We don't remove old methods from the compile queue even if they have 129 // very low activity. See select_task(). 130 bool AdvancedThresholdPolicy::is_old(Method* method) { 131 return method->invocation_count() > 50000 || method->backedge_count() > 500000; 132 } 133 134 double AdvancedThresholdPolicy::weight(Method* method) { 135 return (method->rate() + 1) * ((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 max_task->set_comp_level(CompLevel_limited_profile); 200 if (PrintTieredEvents) { 201 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 202 } 203 } 204 205 return max_task; 206 } 207 208 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 209 double queue_size = CompileBroker::queue_size(level); 210 int comp_count = compiler_count(level); 211 double k = queue_size / (feedback_k * comp_count) + 1; 212 213 // Increase C1 compile threshold when the code cache is filled more 214 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 215 // The main intention is to keep enough free space for C2 compiled code 216 // to achieve peak performance if the code cache is under stress. 217 if ((TieredStopAtLevel == CompLevel_full_optimization) && (level != CompLevel_full_optimization)) { 218 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(CodeCache::get_code_blob_type(level)); 219 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 220 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 221 } 222 } 223 return k; 224 } 225 226 // Call and loop predicates determine whether a transition to a higher 227 // compilation level should be performed (pointers to predicate functions 228 // are passed to common()). 229 // Tier?LoadFeedback is basically a coefficient that determines of 230 // how many methods per compiler thread can be in the queue before 231 // the threshold values double. 232 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level) { 233 switch(cur_level) { 234 case CompLevel_none: 235 case CompLevel_limited_profile: { 236 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 237 return loop_predicate_helper<CompLevel_none>(i, b, k); 238 } 239 case CompLevel_full_profile: { 240 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 241 return loop_predicate_helper<CompLevel_full_profile>(i, b, k); 242 } 243 default: 244 return true; 245 } 246 } 247 248 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level) { 249 switch(cur_level) { 250 case CompLevel_none: 251 case CompLevel_limited_profile: { 252 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 253 return call_predicate_helper<CompLevel_none>(i, b, k); 254 } 255 case CompLevel_full_profile: { 256 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 257 return call_predicate_helper<CompLevel_full_profile>(i, b, k); 258 } 259 default: 260 return true; 261 } 262 } 263 264 // If a method is old enough and is still in the interpreter we would want to 265 // start profiling without waiting for the compiled method to arrive. 266 // We also take the load on compilers into the account. 267 bool AdvancedThresholdPolicy::should_create_mdo(Method* method, CompLevel cur_level) { 268 if (cur_level == CompLevel_none && 269 CompileBroker::queue_size(CompLevel_full_optimization) <= 270 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 271 int i = method->invocation_count(); 272 int b = method->backedge_count(); 273 double k = Tier0ProfilingStartPercentage / 100.0; 274 return call_predicate_helper<CompLevel_none>(i, b, k) || loop_predicate_helper<CompLevel_none>(i, b, k); 275 } 276 return false; 277 } 278 279 // Inlining control: if we're compiling a profiled method with C1 and the callee 280 // is known to have OSRed in a C2 version, don't inline it. 281 bool AdvancedThresholdPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 282 CompLevel comp_level = (CompLevel)env->comp_level(); 283 if (comp_level == CompLevel_full_profile || 284 comp_level == CompLevel_limited_profile) { 285 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 286 } 287 return false; 288 } 289 290 // Create MDO if necessary. 291 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, JavaThread* THREAD) { 292 if (mh->is_native() || mh->is_abstract() || mh->is_accessor()) return; 293 if (mh->method_data() == NULL) { 294 Method::build_interpreter_method_data(mh, CHECK_AND_CLEAR); 295 } 296 } 297 298 299 /* 300 * Method states: 301 * 0 - interpreter (CompLevel_none) 302 * 1 - pure C1 (CompLevel_simple) 303 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 304 * 3 - C1 with full profiling (CompLevel_full_profile) 305 * 4 - C2 (CompLevel_full_optimization) 306 * 307 * Common state transition patterns: 308 * a. 0 -> 3 -> 4. 309 * The most common path. But note that even in this straightforward case 310 * profiling can start at level 0 and finish at level 3. 311 * 312 * b. 0 -> 2 -> 3 -> 4. 313 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning 314 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 315 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 316 * 317 * c. 0 -> (3->2) -> 4. 318 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 319 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 320 * of the method to 2, because it'll allow it to run much faster without full profiling while c2 321 * is compiling. 322 * 323 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 324 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 325 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 326 * 327 * e. 0 -> 4. 328 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 329 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 330 * the compiled version already exists). 331 * 332 * Note that since state 0 can be reached from any other state via deoptimization different loops 333 * are possible. 334 * 335 */ 336 337 // Common transition function. Given a predicate determines if a method should transition to another level. 338 CompLevel AdvancedThresholdPolicy::common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback) { 339 CompLevel next_level = cur_level; 340 int i = method->invocation_count(); 341 int b = method->backedge_count(); 342 343 if (is_trivial(method)) { 344 next_level = CompLevel_simple; 345 } else { 346 switch(cur_level) { 347 case CompLevel_none: 348 // If we were at full profile level, would we switch to full opt? 349 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { 350 next_level = CompLevel_full_optimization; 351 } else if ((this->*p)(i, b, cur_level)) { 352 // C1-generated fully profiled code is about 30% slower than the limited profile 353 // code that has only invocation and backedge counters. The observation is that 354 // if C2 queue is large enough we can spend too much time in the fully profiled code 355 // while waiting for C2 to pick the method from the queue. To alleviate this problem 356 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 357 // we choose to compile a limited profiled version and then recompile with full profiling 358 // when the load on C2 goes down. 359 if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > 360 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 361 next_level = CompLevel_limited_profile; 362 } else { 363 next_level = CompLevel_full_profile; 364 } 365 } 366 break; 367 case CompLevel_limited_profile: 368 if (is_method_profiled(method)) { 369 // Special case: we got here because this method was fully profiled in the interpreter. 370 next_level = CompLevel_full_optimization; 371 } else { 372 MethodData* mdo = method->method_data(); 373 if (mdo != NULL) { 374 if (mdo->would_profile()) { 375 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 376 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 377 (this->*p)(i, b, cur_level))) { 378 next_level = CompLevel_full_profile; 379 } 380 } else { 381 next_level = CompLevel_full_optimization; 382 } 383 } 384 } 385 break; 386 case CompLevel_full_profile: 387 { 388 MethodData* mdo = method->method_data(); 389 if (mdo != NULL) { 390 if (mdo->would_profile()) { 391 int mdo_i = mdo->invocation_count_delta(); 392 int mdo_b = mdo->backedge_count_delta(); 393 if ((this->*p)(mdo_i, mdo_b, cur_level)) { 394 next_level = CompLevel_full_optimization; 395 } 396 } else { 397 next_level = CompLevel_full_optimization; 398 } 399 } 400 } 401 break; 402 } 403 } 404 return MIN2(next_level, (CompLevel)TieredStopAtLevel); 405 } 406 407 // Determine if a method should be compiled with a normal entry point at a different level. 408 CompLevel AdvancedThresholdPolicy::call_event(Method* method, CompLevel cur_level) { 409 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), 410 common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true)); 411 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 412 413 // If OSR method level is greater than the regular method level, the levels should be 414 // equalized by raising the regular method level in order to avoid OSRs during each 415 // invocation of the method. 416 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 417 MethodData* mdo = method->method_data(); 418 guarantee(mdo != NULL, "MDO should not be NULL"); 419 if (mdo->invocation_count() >= 1) { 420 next_level = CompLevel_full_optimization; 421 } 422 } else { 423 next_level = MAX2(osr_level, next_level); 424 } 425 return next_level; 426 } 427 428 // Determine if we should do an OSR compilation of a given method. 429 CompLevel AdvancedThresholdPolicy::loop_event(Method* method, CompLevel cur_level) { 430 CompLevel next_level = common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true); 431 if (cur_level == CompLevel_none) { 432 // If there is a live OSR method that means that we deopted to the interpreter 433 // for the transition. 434 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 435 if (osr_level > CompLevel_none) { 436 return osr_level; 437 } 438 } 439 return next_level; 440 } 441 442 // Update the rate and submit compile 443 void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, JavaThread* thread) { 444 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 445 update_rate(os::javaTimeMillis(), mh()); 446 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", thread); 447 } 448 449 // Handle the invocation event. 450 void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh, 451 CompLevel level, nmethod* nm, JavaThread* thread) { 452 if (should_create_mdo(mh(), level)) { 453 create_mdo(mh, thread); 454 } 455 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 456 CompLevel next_level = call_event(mh(), level); 457 if (next_level != level) { 458 compile(mh, InvocationEntryBci, next_level, thread); 459 } 460 } 461 } 462 463 // Handle the back branch event. Notice that we can compile the method 464 // with a regular entry from here. 465 void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh, 466 int bci, CompLevel level, nmethod* nm, JavaThread* thread) { 467 if (should_create_mdo(mh(), level)) { 468 create_mdo(mh, thread); 469 } 470 // Check if MDO should be created for the inlined method 471 if (should_create_mdo(imh(), level)) { 472 create_mdo(imh, thread); 473 } 474 475 if (is_compilation_enabled()) { 476 CompLevel next_osr_level = loop_event(imh(), level); 477 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 478 // At the very least compile the OSR version 479 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 480 compile(imh, bci, next_osr_level, thread); 481 } 482 483 // Use loop event as an opportunity to also check if there's been 484 // enough calls. 485 CompLevel cur_level, next_level; 486 if (mh() != imh()) { // If there is an enclosing method 487 guarantee(nm != NULL, "Should have nmethod here"); 488 cur_level = comp_level(mh()); 489 next_level = call_event(mh(), cur_level); 490 491 if (max_osr_level == CompLevel_full_optimization) { 492 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 493 bool make_not_entrant = false; 494 if (nm->is_osr_method()) { 495 // This is an osr method, just make it not entrant and recompile later if needed 496 make_not_entrant = true; 497 } else { 498 if (next_level != CompLevel_full_optimization) { 499 // next_level is not full opt, so we need to recompile the 500 // enclosing method without the inlinee 501 cur_level = CompLevel_none; 502 make_not_entrant = true; 503 } 504 } 505 if (make_not_entrant) { 506 if (PrintTieredEvents) { 507 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 508 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 509 } 510 nm->make_not_entrant(); 511 } 512 } 513 if (!CompileBroker::compilation_is_in_queue(mh)) { 514 // Fix up next_level if necessary to avoid deopts 515 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 516 next_level = CompLevel_full_profile; 517 } 518 if (cur_level != next_level) { 519 compile(mh, InvocationEntryBci, next_level, thread); 520 } 521 } 522 } else { 523 cur_level = comp_level(imh()); 524 next_level = call_event(imh(), cur_level); 525 if (!CompileBroker::compilation_is_in_queue(imh) && (next_level != cur_level)) { 526 compile(imh, InvocationEntryBci, next_level, thread); 527 } 528 } 529 } 530 } 531 532 #endif // TIERED