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