1 /* 2 * Copyright (c) 2010, 2013, 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_intptr(os::active_processor_count()); 51 int loglog_cpu = log2_intptr(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 #ifdef X86 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 JavaThread* THREAD = JavaThread::current(); 79 if (is_old(m)) { 80 // We don't remove old methods from the queue, 81 // so we can just zero the rate. 82 m->set_rate(0, THREAD); 83 return; 84 } 85 86 // We don't update the rate if we've just came out of a safepoint. 87 // delta_s is the time since last safepoint in milliseconds. 88 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 89 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement 90 // How many events were there since the last time? 91 int event_count = m->invocation_count() + m->backedge_count(); 92 int delta_e = event_count - m->prev_event_count(); 93 94 // We should be running for at least 1ms. 95 if (delta_s >= TieredRateUpdateMinTime) { 96 // And we must've taken the previous point at least 1ms before. 97 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 98 m->set_prev_time(t, THREAD); 99 m->set_prev_event_count(event_count, THREAD); 100 m->set_rate((float)delta_e / (float)delta_t, THREAD); // Rate is events per millisecond 101 } else 102 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 103 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 104 m->set_rate(0, THREAD); 105 } 106 } 107 } 108 109 // Check if this method has been stale from a given number of milliseconds. 110 // See select_task(). 111 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, Method* m) { 112 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 113 jlong delta_t = t - m->prev_time(); 114 if (delta_t > timeout && delta_s > timeout) { 115 int event_count = m->invocation_count() + m->backedge_count(); 116 int delta_e = event_count - m->prev_event_count(); 117 // Return true if there were no events. 118 return delta_e == 0; 119 } 120 return false; 121 } 122 123 // We don't remove old methods from the compile queue even if they have 124 // very low activity. See select_task(). 125 bool AdvancedThresholdPolicy::is_old(Method* method) { 126 return method->invocation_count() > 50000 || method->backedge_count() > 500000; 127 } 128 129 double AdvancedThresholdPolicy::weight(Method* method) { 130 return (method->rate() + 1) * ((method->invocation_count() + 1) * (method->backedge_count() + 1)); 131 } 132 133 // Apply heuristics and return true if x should be compiled before y 134 bool AdvancedThresholdPolicy::compare_methods(Method* x, Method* y) { 135 if (x->highest_comp_level() > y->highest_comp_level()) { 136 // recompilation after deopt 137 return true; 138 } else 139 if (x->highest_comp_level() == y->highest_comp_level()) { 140 if (weight(x) > weight(y)) { 141 return true; 142 } 143 } 144 return false; 145 } 146 147 // Is method profiled enough? 148 bool AdvancedThresholdPolicy::is_method_profiled(Method* method) { 149 MethodData* mdo = method->method_data(); 150 if (mdo != NULL) { 151 int i = mdo->invocation_count_delta(); 152 int b = mdo->backedge_count_delta(); 153 return call_predicate_helper<CompLevel_full_profile>(i, b, 1); 154 } 155 return false; 156 } 157 158 // Called with the queue locked and with at least one element 159 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { 160 CompileTask *max_task = NULL; 161 Method* max_method = NULL; 162 jlong t = os::javaTimeMillis(); 163 // Iterate through the queue and find a method with a maximum rate. 164 for (CompileTask* task = compile_queue->first(); task != NULL;) { 165 CompileTask* next_task = task->next(); 166 Method* method = task->method(); 167 MethodData* mdo = method->method_data(); 168 update_rate(t, method); 169 if (max_task == NULL) { 170 max_task = task; 171 max_method = method; 172 } else { 173 // If a method has been stale for some time, remove it from the queue. 174 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { 175 if (PrintTieredEvents) { 176 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel)task->comp_level()); 177 } 178 CompileTaskWrapper ctw(task); // Frees the task 179 compile_queue->remove(task); 180 method->clear_queued_for_compilation(); 181 task = next_task; 182 continue; 183 } 184 185 // Select a method with a higher rate 186 if (compare_methods(method, max_method)) { 187 max_task = task; 188 max_method = method; 189 } 190 } 191 task = next_task; 192 } 193 194 if (max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile 195 && is_method_profiled(max_method)) { 196 max_task->set_comp_level(CompLevel_limited_profile); 197 if (PrintTieredEvents) { 198 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 199 } 200 } 201 202 return max_task; 203 } 204 205 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 206 double queue_size = CompileBroker::queue_size(level); 207 int comp_count = compiler_count(level); 208 double k = queue_size / (feedback_k * comp_count) + 1; 209 210 // Increase C1 compile threshold when the code cache is filled more 211 // than specified by IncreaseFirstTierCompileThresholdAt percentage. 212 // The main intention is to keep enough free space for C2 compiled code 213 // to achieve peak performance if the code cache is under stress. 214 if ((TieredStopAtLevel == CompLevel_full_optimization) && (level != CompLevel_full_optimization)) { 215 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(); 216 if (current_reverse_free_ratio > _increase_threshold_at_ratio) { 217 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio); 218 } 219 } 220 return k; 221 } 222 223 // Call and loop predicates determine whether a transition to a higher 224 // compilation level should be performed (pointers to predicate functions 225 // are passed to common()). 226 // Tier?LoadFeedback is basically a coefficient that determines of 227 // how many methods per compiler thread can be in the queue before 228 // the threshold values double. 229 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level) { 230 switch(cur_level) { 231 case CompLevel_none: 232 case CompLevel_limited_profile: { 233 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 234 return loop_predicate_helper<CompLevel_none>(i, b, k); 235 } 236 case CompLevel_full_profile: { 237 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 238 return loop_predicate_helper<CompLevel_full_profile>(i, b, k); 239 } 240 default: 241 return true; 242 } 243 } 244 245 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level) { 246 switch(cur_level) { 247 case CompLevel_none: 248 case CompLevel_limited_profile: { 249 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 250 return call_predicate_helper<CompLevel_none>(i, b, k); 251 } 252 case CompLevel_full_profile: { 253 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 254 return call_predicate_helper<CompLevel_full_profile>(i, b, k); 255 } 256 default: 257 return true; 258 } 259 } 260 261 // If a method is old enough and is still in the interpreter we would want to 262 // start profiling without waiting for the compiled method to arrive. 263 // We also take the load on compilers into the account. 264 bool AdvancedThresholdPolicy::should_create_mdo(Method* method, CompLevel cur_level) { 265 if (cur_level == CompLevel_none && 266 CompileBroker::queue_size(CompLevel_full_optimization) <= 267 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 268 int i = method->invocation_count(); 269 int b = method->backedge_count(); 270 double k = Tier0ProfilingStartPercentage / 100.0; 271 return call_predicate_helper<CompLevel_none>(i, b, k) || loop_predicate_helper<CompLevel_none>(i, b, k); 272 } 273 return false; 274 } 275 276 // Inlining control: if we're compiling a profiled method with C1 and the callee 277 // is known to have OSRed in a C2 version, don't inline it. 278 bool AdvancedThresholdPolicy::should_not_inline(ciEnv* env, ciMethod* callee) { 279 CompLevel comp_level = (CompLevel)env->comp_level(); 280 if (comp_level == CompLevel_full_profile || 281 comp_level == CompLevel_limited_profile) { 282 return callee->highest_osr_comp_level() == CompLevel_full_optimization; 283 } 284 return false; 285 } 286 287 // Create MDO if necessary. 288 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, JavaThread* THREAD) { 289 if (mh->is_native() || mh->is_abstract() || mh->is_accessor()) return; 290 if (mh->method_data() == NULL) { 291 Method::build_interpreter_method_data(mh, CHECK_AND_CLEAR); 292 } 293 } 294 295 296 /* 297 * Method states: 298 * 0 - interpreter (CompLevel_none) 299 * 1 - pure C1 (CompLevel_simple) 300 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 301 * 3 - C1 with full profiling (CompLevel_full_profile) 302 * 4 - C2 (CompLevel_full_optimization) 303 * 304 * Common state transition patterns: 305 * a. 0 -> 3 -> 4. 306 * The most common path. But note that even in this straightforward case 307 * profiling can start at level 0 and finish at level 3. 308 * 309 * b. 0 -> 2 -> 3 -> 4. 310 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning 311 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 312 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 313 * 314 * c. 0 -> (3->2) -> 4. 315 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 316 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 317 * of the method to 2, because it'll allow it to run much faster without full profiling while c2 318 * is compiling. 319 * 320 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 321 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 322 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 323 * 324 * e. 0 -> 4. 325 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 326 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 327 * the compiled version already exists). 328 * 329 * Note that since state 0 can be reached from any other state via deoptimization different loops 330 * are possible. 331 * 332 */ 333 334 // Common transition function. Given a predicate determines if a method should transition to another level. 335 CompLevel AdvancedThresholdPolicy::common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback) { 336 CompLevel next_level = cur_level; 337 int i = method->invocation_count(); 338 int b = method->backedge_count(); 339 340 if (is_trivial(method)) { 341 next_level = CompLevel_simple; 342 } else { 343 switch(cur_level) { 344 case CompLevel_none: 345 // If we were at full profile level, would we switch to full opt? 346 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { 347 next_level = CompLevel_full_optimization; 348 } else if ((this->*p)(i, b, cur_level)) { 349 // C1-generated fully profiled code is about 30% slower than the limited profile 350 // code that has only invocation and backedge counters. The observation is that 351 // if C2 queue is large enough we can spend too much time in the fully profiled code 352 // while waiting for C2 to pick the method from the queue. To alleviate this problem 353 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 354 // we choose to compile a limited profiled version and then recompile with full profiling 355 // when the load on C2 goes down. 356 if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) > 357 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 358 next_level = CompLevel_limited_profile; 359 } else { 360 next_level = CompLevel_full_profile; 361 } 362 } 363 break; 364 case CompLevel_limited_profile: 365 if (is_method_profiled(method)) { 366 // Special case: we got here because this method was fully profiled in the interpreter. 367 next_level = CompLevel_full_optimization; 368 } else { 369 MethodData* mdo = method->method_data(); 370 if (mdo != NULL) { 371 if (mdo->would_profile()) { 372 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <= 373 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 374 (this->*p)(i, b, cur_level))) { 375 next_level = CompLevel_full_profile; 376 } 377 } else { 378 next_level = CompLevel_full_optimization; 379 } 380 } 381 } 382 break; 383 case CompLevel_full_profile: 384 { 385 MethodData* mdo = method->method_data(); 386 if (mdo != NULL) { 387 if (mdo->would_profile()) { 388 int mdo_i = mdo->invocation_count_delta(); 389 int mdo_b = mdo->backedge_count_delta(); 390 if ((this->*p)(mdo_i, mdo_b, cur_level)) { 391 next_level = CompLevel_full_optimization; 392 } 393 } else { 394 next_level = CompLevel_full_optimization; 395 } 396 } 397 } 398 break; 399 } 400 } 401 return MIN2(next_level, (CompLevel)TieredStopAtLevel); 402 } 403 404 // Determine if a method should be compiled with a normal entry point at a different level. 405 CompLevel AdvancedThresholdPolicy::call_event(Method* method, CompLevel cur_level) { 406 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(), 407 common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true)); 408 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 409 410 // If OSR method level is greater than the regular method level, the levels should be 411 // equalized by raising the regular method level in order to avoid OSRs during each 412 // invocation of the method. 413 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 414 MethodData* mdo = method->method_data(); 415 guarantee(mdo != NULL, "MDO should not be NULL"); 416 if (mdo->invocation_count() >= 1) { 417 next_level = CompLevel_full_optimization; 418 } 419 } else { 420 next_level = MAX2(osr_level, next_level); 421 } 422 return next_level; 423 } 424 425 // Determine if we should do an OSR compilation of a given method. 426 CompLevel AdvancedThresholdPolicy::loop_event(Method* method, CompLevel cur_level) { 427 CompLevel next_level = common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true); 428 if (cur_level == CompLevel_none) { 429 // If there is a live OSR method that means that we deopted to the interpreter 430 // for the transition. 431 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level); 432 if (osr_level > CompLevel_none) { 433 return osr_level; 434 } 435 } 436 return next_level; 437 } 438 439 // Update the rate and submit compile 440 void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, JavaThread* thread) { 441 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 442 update_rate(os::javaTimeMillis(), mh()); 443 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", thread); 444 } 445 446 // Handle the invocation event. 447 void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh, 448 CompLevel level, nmethod* nm, JavaThread* thread) { 449 if (should_create_mdo(mh(), level)) { 450 create_mdo(mh, thread); 451 } 452 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) { 453 CompLevel next_level = call_event(mh(), level); 454 if (next_level != level) { 455 compile(mh, InvocationEntryBci, next_level, thread); 456 } 457 } 458 } 459 460 // Handle the back branch event. Notice that we can compile the method 461 // with a regular entry from here. 462 void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh, 463 int bci, CompLevel level, nmethod* nm, JavaThread* thread) { 464 if (should_create_mdo(mh(), level)) { 465 create_mdo(mh, thread); 466 } 467 // Check if MDO should be created for the inlined method 468 if (should_create_mdo(imh(), level)) { 469 create_mdo(imh, thread); 470 } 471 472 if (is_compilation_enabled()) { 473 CompLevel next_osr_level = loop_event(imh(), level); 474 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level(); 475 // At the very least compile the OSR version 476 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) { 477 compile(imh, bci, next_osr_level, thread); 478 } 479 480 // Use loop event as an opportunity to also check if there's been 481 // enough calls. 482 CompLevel cur_level, next_level; 483 if (mh() != imh()) { // If there is an enclosing method 484 guarantee(nm != NULL, "Should have nmethod here"); 485 cur_level = comp_level(mh()); 486 next_level = call_event(mh(), cur_level); 487 488 if (max_osr_level == CompLevel_full_optimization) { 489 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts 490 bool make_not_entrant = false; 491 if (nm->is_osr_method()) { 492 // This is an osr method, just make it not entrant and recompile later if needed 493 make_not_entrant = true; 494 } else { 495 if (next_level != CompLevel_full_optimization) { 496 // next_level is not full opt, so we need to recompile the 497 // enclosing method without the inlinee 498 cur_level = CompLevel_none; 499 make_not_entrant = true; 500 } 501 } 502 if (make_not_entrant) { 503 if (PrintTieredEvents) { 504 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci; 505 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level); 506 } 507 nm->make_not_entrant(); 508 } 509 } 510 if (!CompileBroker::compilation_is_in_queue(mh)) { 511 // Fix up next_level if necessary to avoid deopts 512 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) { 513 next_level = CompLevel_full_profile; 514 } 515 if (cur_level != next_level) { 516 compile(mh, InvocationEntryBci, next_level, thread); 517 } 518 } 519 } else { 520 cur_level = comp_level(imh()); 521 next_level = call_event(imh(), cur_level); 522 if (!CompileBroker::compilation_is_in_queue(imh) && (next_level != cur_level)) { 523 compile(imh, InvocationEntryBci, next_level, thread); 524 } 525 } 526 } 527 } 528 529 #endif // TIERED