1 /* 2 * Copyright (c) 2010, 2011, 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 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, methodOop m) { 77 if (is_old(m)) { 78 // We don't remove old methods from the queue, 79 // so we can just zero the rate. 80 m->set_rate(0); 81 return; 82 } 83 84 // We don't update the rate if we've just came out of a safepoint. 85 // delta_s is the time since last safepoint in milliseconds. 86 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 87 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement 88 // How many events were there since the last time? 89 int event_count = m->invocation_count() + m->backedge_count(); 90 int delta_e = event_count - m->prev_event_count(); 91 92 // We should be running for at least 1ms. 93 if (delta_s >= TieredRateUpdateMinTime) { 94 // And we must've taken the previous point at least 1ms before. 95 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 96 m->set_prev_time(t); 97 m->set_prev_event_count(event_count); 98 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 99 } else 100 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 101 // If nothing happened for 25ms, zero the rate. Don't modify prev values. 102 m->set_rate(0); 103 } 104 } 105 } 106 107 // Check if this method has been stale from a given number of milliseconds. 108 // See select_task(). 109 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, methodOop m) { 110 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 111 jlong delta_t = t - m->prev_time(); 112 if (delta_t > timeout && delta_s > timeout) { 113 int event_count = m->invocation_count() + m->backedge_count(); 114 int delta_e = event_count - m->prev_event_count(); 115 // Return true if there were no events. 116 return delta_e == 0; 117 } 118 return false; 119 } 120 121 // We don't remove old methods from the compile queue even if they have 122 // very low activity. See select_task(). 123 bool AdvancedThresholdPolicy::is_old(methodOop method) { 124 return method->invocation_count() > 50000 || method->backedge_count() > 500000; 125 } 126 127 double AdvancedThresholdPolicy::weight(methodOop method) { 128 return (method->rate() + 1) * ((method->invocation_count() + 1) * (method->backedge_count() + 1)); 129 } 130 131 // Apply heuristics and return true if x should be compiled before y 132 bool AdvancedThresholdPolicy::compare_methods(methodOop x, methodOop y) { 133 if (x->highest_comp_level() > y->highest_comp_level()) { 134 // recompilation after deopt 135 return true; 136 } else 137 if (x->highest_comp_level() == y->highest_comp_level()) { 138 if (weight(x) > weight(y)) { 139 return true; 140 } 141 } 142 return false; 143 } 144 145 // Is method profiled enough? 146 bool AdvancedThresholdPolicy::is_method_profiled(methodOop method) { 147 methodDataOop mdo = method->method_data(); 148 if (mdo != NULL) { 149 int i = mdo->invocation_count_delta(); 150 int b = mdo->backedge_count_delta(); 151 return call_predicate_helper<CompLevel_full_profile>(i, b, 1); 152 } 153 return false; 154 } 155 156 // Called with the queue locked and with at least one element 157 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { 158 CompileTask *max_task = NULL; 159 methodOop max_method; 160 jlong t = os::javaTimeMillis(); 161 // Iterate through the queue and find a method with a maximum rate. 162 for (CompileTask* task = compile_queue->first(); task != NULL;) { 163 CompileTask* next_task = task->next(); 164 methodOop method = (methodOop)JNIHandles::resolve(task->method_handle()); 165 methodDataOop mdo = method->method_data(); 166 update_rate(t, method); 167 if (max_task == NULL) { 168 max_task = task; 169 max_method = method; 170 } else { 171 // If a method has been stale for some time, remove it from the queue. 172 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { 173 if (PrintTieredEvents) { 174 print_event(KILL, method, method, task->osr_bci(), (CompLevel)task->comp_level()); 175 } 176 CompileTaskWrapper ctw(task); // Frees the task 177 compile_queue->remove(task); 178 method->clear_queued_for_compilation(); 179 task = next_task; 180 continue; 181 } 182 183 // Select a method with a higher rate 184 if (compare_methods(method, max_method)) { 185 max_task = task; 186 max_method = method; 187 } 188 } 189 task = next_task; 190 } 191 192 if (max_task->comp_level() == CompLevel_full_profile && is_method_profiled(max_method)) { 193 max_task->set_comp_level(CompLevel_limited_profile); 194 if (PrintTieredEvents) { 195 print_event(UPDATE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 196 } 197 } 198 199 return max_task; 200 } 201 202 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 203 double queue_size = CompileBroker::queue_size(level); 204 int comp_count = compiler_count(level); 205 double k = queue_size / (feedback_k * comp_count) + 1; 206 return k; 207 } 208 209 // Call and loop predicates determine whether a transition to a higher 210 // compilation level should be performed (pointers to predicate functions 211 // are passed to common()). 212 // Tier?LoadFeedback is basically a coefficient that determines of 213 // how many methods per compiler thread can be in the queue before 214 // the threshold values double. 215 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level) { 216 switch(cur_level) { 217 case CompLevel_none: 218 case CompLevel_limited_profile: { 219 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 220 return loop_predicate_helper<CompLevel_none>(i, b, k); 221 } 222 case CompLevel_full_profile: { 223 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 224 return loop_predicate_helper<CompLevel_full_profile>(i, b, k); 225 } 226 default: 227 return true; 228 } 229 } 230 231 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level) { 232 switch(cur_level) { 233 case CompLevel_none: 234 case CompLevel_limited_profile: { 235 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 236 return call_predicate_helper<CompLevel_none>(i, b, k); 237 } 238 case CompLevel_full_profile: { 239 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 240 return call_predicate_helper<CompLevel_full_profile>(i, b, k); 241 } 242 default: 243 return true; 244 } 245 } 246 247 // If a method is old enough and is still in the interpreter we would want to 248 // start profiling without waiting for the compiled method to arrive. 249 // We also take the load on compilers into the account. 250 bool AdvancedThresholdPolicy::should_create_mdo(methodOop method, CompLevel cur_level) { 251 if (cur_level == CompLevel_none && 252 CompileBroker::queue_size(CompLevel_full_optimization) <= 253 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 254 int i = method->invocation_count(); 255 int b = method->backedge_count(); 256 double k = Tier0ProfilingStartPercentage / 100.0; 257 return call_predicate_helper<CompLevel_none>(i, b, k) || loop_predicate_helper<CompLevel_none>(i, b, k); 258 } 259 return false; 260 } 261 262 // Create MDO if necessary. 263 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, TRAPS) { 264 if (mh->is_native() || mh->is_abstract() || mh->is_accessor()) return; 265 if (mh->method_data() == NULL) { 266 methodOopDesc::build_interpreter_method_data(mh, THREAD); 267 if (HAS_PENDING_EXCEPTION) { 268 CLEAR_PENDING_EXCEPTION; 269 } 270 } 271 } 272 273 274 /* 275 * Method states: 276 * 0 - interpreter (CompLevel_none) 277 * 1 - pure C1 (CompLevel_simple) 278 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 279 * 3 - C1 with full profiling (CompLevel_full_profile) 280 * 4 - C2 (CompLevel_full_optimization) 281 * 282 * Common state transition patterns: 283 * a. 0 -> 3 -> 4. 284 * The most common path. But note that even in this straightforward case 285 * profiling can start at level 0 and finish at level 3. 286 * 287 * b. 0 -> 2 -> 3 -> 4. 288 * This case occures when the load on C2 is deemed too high. So, instead of transitioning 289 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 290 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 291 * 292 * c. 0 -> (3->2) -> 4. 293 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 294 * to enable the profiling to fully occur at level 0. In this case we change the compilation level 295 * of the method to 2, because it'll allow it to run much faster without full profiling while c2 296 * is compiling. 297 * 298 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 299 * After a method was once compiled with C1 it can be identified as trivial and be compiled to 300 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 301 * 302 * e. 0 -> 4. 303 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 304 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 305 * the compiled version already exists). 306 * 307 * Note that since state 0 can be reached from any other state via deoptimization different loops 308 * are possible. 309 * 310 */ 311 312 // Common transition function. Given a predicate determines if a method should transition to another level. 313 CompLevel AdvancedThresholdPolicy::common(Predicate p, methodOop method, CompLevel cur_level) { 314 if (is_trivial(method)) return CompLevel_simple; 315 316 CompLevel next_level = cur_level; 317 int i = method->invocation_count(); 318 int b = method->backedge_count(); 319 320 switch(cur_level) { 321 case CompLevel_none: 322 // If we were at full profile level, would we switch to full opt? 323 if (common(p, method, CompLevel_full_profile) == CompLevel_full_optimization) { 324 next_level = CompLevel_full_optimization; 325 } else if ((this->*p)(i, b, cur_level)) { 326 // C1-generated fully profiled code is about 30% slower than the limited profile 327 // code that has only invocation and backedge counters. The observation is that 328 // if C2 queue is large enough we can spend too much time in the fully profiled code 329 // while waiting for C2 to pick the method from the queue. To alleviate this problem 330 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 331 // we choose to compile a limited profiled version and then recompile with full profiling 332 // when the load on C2 goes down. 333 if (CompileBroker::queue_size(CompLevel_full_optimization) > 334 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 335 next_level = CompLevel_limited_profile; 336 } else { 337 next_level = CompLevel_full_profile; 338 } 339 } 340 break; 341 case CompLevel_limited_profile: 342 if (is_method_profiled(method)) { 343 // Special case: we got here because this method was fully profiled in the interpreter. 344 next_level = CompLevel_full_optimization; 345 } else { 346 methodDataOop mdo = method->method_data(); 347 if (mdo != NULL) { 348 if (mdo->would_profile()) { 349 if (CompileBroker::queue_size(CompLevel_full_optimization) <= 350 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 351 (this->*p)(i, b, cur_level)) { 352 next_level = CompLevel_full_profile; 353 } 354 } else { 355 next_level = CompLevel_full_optimization; 356 } 357 } 358 } 359 break; 360 case CompLevel_full_profile: 361 { 362 methodDataOop mdo = method->method_data(); 363 if (mdo != NULL) { 364 if (mdo->would_profile()) { 365 int mdo_i = mdo->invocation_count_delta(); 366 int mdo_b = mdo->backedge_count_delta(); 367 if ((this->*p)(mdo_i, mdo_b, cur_level)) { 368 next_level = CompLevel_full_optimization; 369 } 370 } else { 371 next_level = CompLevel_full_optimization; 372 } 373 } 374 } 375 break; 376 } 377 return next_level; 378 } 379 380 // Determine if a method should be compiled with a normal entry point at a different level. 381 CompLevel AdvancedThresholdPolicy::call_event(methodOop method, CompLevel cur_level) { 382 CompLevel osr_level = (CompLevel) method->highest_osr_comp_level(); 383 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 384 385 // If OSR method level is greater than the regular method level, the levels should be 386 // equalized by raising the regular method level in order to avoid OSRs during each 387 // invocation of the method. 388 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 389 methodDataOop mdo = method->method_data(); 390 guarantee(mdo != NULL, "MDO should not be NULL"); 391 if (mdo->invocation_count() >= 1) { 392 next_level = CompLevel_full_optimization; 393 } 394 } else { 395 next_level = MAX2(osr_level, next_level); 396 } 397 398 return next_level; 399 } 400 401 // Determine if we should do an OSR compilation of a given method. 402 CompLevel AdvancedThresholdPolicy::loop_event(methodOop method, CompLevel cur_level) { 403 if (cur_level == CompLevel_none) { 404 // If there is a live OSR method that means that we deopted to the interpreter 405 // for the transition. 406 CompLevel osr_level = (CompLevel)method->highest_osr_comp_level(); 407 if (osr_level > CompLevel_none) { 408 return osr_level; 409 } 410 } 411 return common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level); 412 } 413 414 // Update the rate and submit compile 415 void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, TRAPS) { 416 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 417 update_rate(os::javaTimeMillis(), mh()); 418 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", THREAD); 419 } 420 421 422 // Handle the invocation event. 423 void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh, 424 CompLevel level, TRAPS) { 425 if (should_create_mdo(mh(), level)) { 426 create_mdo(mh, THREAD); 427 } 428 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, InvocationEntryBci)) { 429 CompLevel next_level = call_event(mh(), level); 430 if (next_level != level) { 431 compile(mh, InvocationEntryBci, next_level, THREAD); 432 } 433 } 434 } 435 436 // Handle the back branch event. Notice that we can compile the method 437 // with a regular entry from here. 438 void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh, 439 int bci, CompLevel level, TRAPS) { 440 if (should_create_mdo(mh(), level)) { 441 create_mdo(mh, THREAD); 442 } 443 444 // If the method is already compiling, quickly bail out. 445 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, bci)) { 446 // Use loop event as an opportinity to also check there's been 447 // enough calls. 448 CompLevel cur_level = comp_level(mh()); 449 CompLevel next_level = call_event(mh(), cur_level); 450 CompLevel next_osr_level = loop_event(mh(), level); 451 if (next_osr_level == CompLevel_limited_profile) { 452 next_osr_level = CompLevel_full_profile; // OSRs are supposed to be for very hot methods. 453 } 454 next_level = MAX2(next_level, 455 next_osr_level < CompLevel_full_optimization ? next_osr_level : cur_level); 456 bool is_compiling = false; 457 if (next_level != cur_level) { 458 compile(mh, InvocationEntryBci, next_level, THREAD); 459 is_compiling = true; 460 } 461 462 // Do the OSR version 463 if (!is_compiling && next_osr_level != level) { 464 compile(mh, bci, next_osr_level, THREAD); 465 } 466 } 467 } 468 469 #endif // TIERED