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 int count = CICompilerCount;
46 #ifdef _LP64
47 // Turn on ergonomic compiler count selection
48 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) {
49 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true);
50 }
51 if (CICompilerCountPerCPU) {
52 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n
53 int log_cpu = log2_intptr(os::active_processor_count());
54 int loglog_cpu = log2_intptr(MAX2(log_cpu, 1));
55 count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2;
56 }
57 #else
58 // On 32-bit systems, the number of compiler threads is limited to 3.
59 // On these systems, the virtual address space available to the JVM
60 // is usually limited to 2-4 GB (the exact value depends on the platform).
61 // As the compilers (especially C2) can consume a large amount of
62 // memory, scaling the number of compiler threads with the number of
63 // available cores can result in the exhaustion of the address space
64 /// available to the VM and thus cause the VM to crash.
65 if (FLAG_IS_DEFAULT(CICompilerCount)) {
66 count = 3;
67 }
68 #endif
69
70 set_c1_count(MAX2(count / 3, 1));
71 set_c2_count(MAX2(count - c1_count(), 1));
72 FLAG_SET_ERGO(intx, CICompilerCount, c1_count() + c2_count());
73
74 // Some inlining tuning
75 #ifdef X86
76 if (FLAG_IS_DEFAULT(InlineSmallCode)) {
77 FLAG_SET_DEFAULT(InlineSmallCode, 2000);
78 }
79 #endif
80
81 #if defined SPARC || defined AARCH64
82 if (FLAG_IS_DEFAULT(InlineSmallCode)) {
83 FLAG_SET_DEFAULT(InlineSmallCode, 2500);
84 }
85 #endif
86
87 set_increase_threshold_at_ratio();
88 set_start_time(os::javaTimeMillis());
89 }
90
91 // update_rate() is called from select_task() while holding a compile queue lock.
92 void AdvancedThresholdPolicy::update_rate(jlong t, Method* m) {
93 // Skip update if counters are absent.
94 // Can't allocate them since we are holding compile queue lock.
95 if (m->method_counters() == NULL) return;
96
97 if (is_old(m)) {
98 // We don't remove old methods from the queue,
99 // so we can just zero the rate.
100 m->set_rate(0);
101 return;
102 }
103
104 // We don't update the rate if we've just came out of a safepoint.
105 // delta_s is the time since last safepoint in milliseconds.
106 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint();
107 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement
108 // How many events were there since the last time?
109 int event_count = m->invocation_count() + m->backedge_count();
110 int delta_e = event_count - m->prev_event_count();
111
112 // We should be running for at least 1ms.
113 if (delta_s >= TieredRateUpdateMinTime) {
114 // And we must've taken the previous point at least 1ms before.
115 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) {
116 m->set_prev_time(t);
117 m->set_prev_event_count(event_count);
118 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond
119 } else {
120 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) {
121 // If nothing happened for 25ms, zero the rate. Don't modify prev values.
122 m->set_rate(0);
123 }
124 }
125 }
126 }
127
128 // Check if this method has been stale from a given number of milliseconds.
129 // See select_task().
130 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, Method* m) {
131 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint();
132 jlong delta_t = t - m->prev_time();
133 if (delta_t > timeout && delta_s > timeout) {
134 int event_count = m->invocation_count() + m->backedge_count();
135 int delta_e = event_count - m->prev_event_count();
136 // Return true if there were no events.
137 return delta_e == 0;
138 }
139 return false;
140 }
141
142 // We don't remove old methods from the compile queue even if they have
143 // very low activity. See select_task().
144 bool AdvancedThresholdPolicy::is_old(Method* method) {
145 return method->invocation_count() > 50000 || method->backedge_count() > 500000;
146 }
147
148 double AdvancedThresholdPolicy::weight(Method* method) {
149 return (double)(method->rate() + 1) *
150 (method->invocation_count() + 1) * (method->backedge_count() + 1);
151 }
152
153 // Apply heuristics and return true if x should be compiled before y
154 bool AdvancedThresholdPolicy::compare_methods(Method* x, Method* y) {
155 if (x->highest_comp_level() > y->highest_comp_level()) {
156 // recompilation after deopt
157 return true;
158 } else
159 if (x->highest_comp_level() == y->highest_comp_level()) {
160 if (weight(x) > weight(y)) {
161 return true;
162 }
163 }
164 return false;
165 }
166
167 // Is method profiled enough?
168 bool AdvancedThresholdPolicy::is_method_profiled(Method* method) {
169 MethodData* mdo = method->method_data();
170 if (mdo != NULL) {
171 int i = mdo->invocation_count_delta();
172 int b = mdo->backedge_count_delta();
173 return call_predicate_helper<CompLevel_full_profile>(i, b, 1, method);
174 }
175 return false;
176 }
177
178 // Called with the queue locked and with at least one element
179 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) {
180 #if INCLUDE_JVMCI
181 CompileTask *max_blocking_task = NULL;
182 #endif
183 CompileTask *max_task = NULL;
184 Method* max_method = NULL;
185 jlong t = os::javaTimeMillis();
186 // Iterate through the queue and find a method with a maximum rate.
187 for (CompileTask* task = compile_queue->first(); task != NULL;) {
188 CompileTask* next_task = task->next();
189 Method* method = task->method();
190 update_rate(t, method);
191 if (max_task == NULL) {
192 max_task = task;
193 max_method = method;
194 } else {
195 // If a method has been stale for some time, remove it from the queue.
196 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) {
197 if (PrintTieredEvents) {
198 print_event(REMOVE_FROM_QUEUE, method, method, task->osr_bci(), (CompLevel)task->comp_level());
199 }
200 task->log_task_dequeued("stale");
201 compile_queue->remove_and_mark_stale(task);
202 method->clear_queued_for_compilation();
203 task = next_task;
204 continue;
205 }
206
207 // Select a method with a higher rate
208 if (compare_methods(method, max_method)) {
209 max_task = task;
210 max_method = method;
211 }
212 }
213 #if INCLUDE_JVMCI
214 if (UseJVMCICompiler && task->is_blocking()) {
215 if (max_blocking_task == NULL || compare_methods(method, max_blocking_task->method())) {
216 max_blocking_task = task;
217 }
218 }
219 #endif
220 task = next_task;
221 }
222
223 #if INCLUDE_JVMCI
224 if (UseJVMCICompiler) {
225 if (max_blocking_task != NULL) {
226 // In blocking compilation mode, the CompileBroker will make
227 // compilations submitted by a JVMCI compiler thread non-blocking. These
228 // compilations should be scheduled after all blocking compilations
229 // to service non-compiler related compilations sooner and reduce the
230 // chance of such compilations timing out.
231 max_task = max_blocking_task;
232 max_method = max_task->method();
233 }
234 }
235 #endif
236
237 if (max_task->comp_level() == CompLevel_full_profile && TieredStopAtLevel > CompLevel_full_profile
238 && is_method_profiled(max_method)) {
239 max_task->set_comp_level(CompLevel_limited_profile);
240 if (PrintTieredEvents) {
241 print_event(UPDATE_IN_QUEUE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level());
242 }
243 }
244
245 return max_task;
246 }
247
248 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) {
249 double queue_size = CompileBroker::queue_size(level);
250 int comp_count = compiler_count(level);
251 double k = queue_size / (feedback_k * comp_count) + 1;
252
253 // Increase C1 compile threshold when the code cache is filled more
254 // than specified by IncreaseFirstTierCompileThresholdAt percentage.
255 // The main intention is to keep enough free space for C2 compiled code
256 // to achieve peak performance if the code cache is under stress.
257 if ((TieredStopAtLevel == CompLevel_full_optimization) && (level != CompLevel_full_optimization)) {
258 double current_reverse_free_ratio = CodeCache::reverse_free_ratio(CodeCache::get_code_blob_type(level));
259 if (current_reverse_free_ratio > _increase_threshold_at_ratio) {
260 k *= exp(current_reverse_free_ratio - _increase_threshold_at_ratio);
261 }
262 }
263 return k;
264 }
265
266 // Call and loop predicates determine whether a transition to a higher
267 // compilation level should be performed (pointers to predicate functions
268 // are passed to common()).
269 // Tier?LoadFeedback is basically a coefficient that determines of
270 // how many methods per compiler thread can be in the queue before
271 // the threshold values double.
272 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level, Method* method) {
273 switch(cur_level) {
274 case CompLevel_none:
275 case CompLevel_limited_profile: {
276 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback);
277 return loop_predicate_helper<CompLevel_none>(i, b, k, method);
278 }
279 case CompLevel_full_profile: {
280 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback);
281 return loop_predicate_helper<CompLevel_full_profile>(i, b, k, method);
282 }
283 default:
284 return true;
285 }
286 }
287
288 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level, Method* method) {
289 switch(cur_level) {
290 case CompLevel_none:
291 case CompLevel_limited_profile: {
292 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback);
293 return call_predicate_helper<CompLevel_none>(i, b, k, method);
294 }
295 case CompLevel_full_profile: {
296 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback);
297 return call_predicate_helper<CompLevel_full_profile>(i, b, k, method);
298 }
299 default:
300 return true;
301 }
302 }
303
304 // If a method is old enough and is still in the interpreter we would want to
305 // start profiling without waiting for the compiled method to arrive.
306 // We also take the load on compilers into the account.
307 bool AdvancedThresholdPolicy::should_create_mdo(Method* method, CompLevel cur_level) {
308 if (cur_level == CompLevel_none &&
309 CompileBroker::queue_size(CompLevel_full_optimization) <=
310 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) {
311 int i = method->invocation_count();
312 int b = method->backedge_count();
313 double k = Tier0ProfilingStartPercentage / 100.0;
314 return call_predicate_helper<CompLevel_none>(i, b, k, method) || loop_predicate_helper<CompLevel_none>(i, b, k, method);
315 }
316 return false;
317 }
318
319 // Inlining control: if we're compiling a profiled method with C1 and the callee
320 // is known to have OSRed in a C2 version, don't inline it.
321 bool AdvancedThresholdPolicy::should_not_inline(ciEnv* env, ciMethod* callee) {
322 CompLevel comp_level = (CompLevel)env->comp_level();
323 if (comp_level == CompLevel_full_profile ||
324 comp_level == CompLevel_limited_profile) {
325 return callee->highest_osr_comp_level() == CompLevel_full_optimization;
326 }
327 return false;
328 }
329
330 // Create MDO if necessary.
331 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, JavaThread* THREAD) {
332 if (mh->is_native() ||
333 mh->is_abstract() ||
334 mh->is_accessor() ||
335 mh->is_constant_getter()) {
336 return;
337 }
338 if (mh->method_data() == NULL) {
339 Method::build_interpreter_method_data(mh, CHECK_AND_CLEAR);
340 }
341 }
342
343
344 /*
345 * Method states:
346 * 0 - interpreter (CompLevel_none)
347 * 1 - pure C1 (CompLevel_simple)
348 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile)
349 * 3 - C1 with full profiling (CompLevel_full_profile)
350 * 4 - C2 (CompLevel_full_optimization)
351 *
352 * Common state transition patterns:
353 * a. 0 -> 3 -> 4.
354 * The most common path. But note that even in this straightforward case
355 * profiling can start at level 0 and finish at level 3.
356 *
357 * b. 0 -> 2 -> 3 -> 4.
358 * This case occurs when the load on C2 is deemed too high. So, instead of transitioning
359 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to
360 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs.
361 *
362 * c. 0 -> (3->2) -> 4.
363 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough
364 * to enable the profiling to fully occur at level 0. In this case we change the compilation level
365 * of the method to 2 while the request is still in-queue, because it'll allow it to run much faster
366 * without full profiling while c2 is compiling.
367 *
368 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1.
369 * After a method was once compiled with C1 it can be identified as trivial and be compiled to
370 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1.
371 *
372 * e. 0 -> 4.
373 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter)
374 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because
375 * the compiled version already exists).
376 *
377 * Note that since state 0 can be reached from any other state via deoptimization different loops
378 * are possible.
379 *
380 */
381
382 // Common transition function. Given a predicate determines if a method should transition to another level.
383 CompLevel AdvancedThresholdPolicy::common(Predicate p, Method* method, CompLevel cur_level, bool disable_feedback) {
384 CompLevel next_level = cur_level;
385 int i = method->invocation_count();
386 int b = method->backedge_count();
387
388 if (is_trivial(method)) {
389 next_level = CompLevel_simple;
390 } else {
391 switch(cur_level) {
392 case CompLevel_none:
393 // If we were at full profile level, would we switch to full opt?
394 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) {
395 next_level = CompLevel_full_optimization;
396 } else if ((this->*p)(i, b, cur_level, method)) {
397 #if INCLUDE_JVMCI
398 if (UseJVMCICompiler) {
399 // Since JVMCI takes a while to warm up, its queue inevitably backs up during
400 // early VM execution.
401 next_level = CompLevel_full_profile;
402 break;
403 }
404 #endif
405 // C1-generated fully profiled code is about 30% slower than the limited profile
406 // code that has only invocation and backedge counters. The observation is that
407 // if C2 queue is large enough we can spend too much time in the fully profiled code
408 // while waiting for C2 to pick the method from the queue. To alleviate this problem
409 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long
410 // we choose to compile a limited profiled version and then recompile with full profiling
411 // when the load on C2 goes down.
412 if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) >
413 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) {
414 next_level = CompLevel_limited_profile;
415 } else {
416 next_level = CompLevel_full_profile;
417 }
418 }
419 break;
420 case CompLevel_limited_profile:
421 if (is_method_profiled(method)) {
422 // Special case: we got here because this method was fully profiled in the interpreter.
423 next_level = CompLevel_full_optimization;
424 } else {
425 MethodData* mdo = method->method_data();
426 if (mdo != NULL) {
427 if (mdo->would_profile()) {
428 if (disable_feedback || (CompileBroker::queue_size(CompLevel_full_optimization) <=
429 Tier3DelayOff * compiler_count(CompLevel_full_optimization) &&
430 (this->*p)(i, b, cur_level, method))) {
431 next_level = CompLevel_full_profile;
432 }
433 } else {
434 next_level = CompLevel_full_optimization;
435 }
436 }
437 }
438 break;
439 case CompLevel_full_profile:
440 {
441 MethodData* mdo = method->method_data();
442 if (mdo != NULL) {
443 if (mdo->would_profile()) {
444 int mdo_i = mdo->invocation_count_delta();
445 int mdo_b = mdo->backedge_count_delta();
446 if ((this->*p)(mdo_i, mdo_b, cur_level, method)) {
447 next_level = CompLevel_full_optimization;
448 }
449 } else {
450 next_level = CompLevel_full_optimization;
451 }
452 }
453 }
454 break;
455 }
456 }
457 return MIN2(next_level, (CompLevel)TieredStopAtLevel);
458 }
459
460 // Determine if a method should be compiled with a normal entry point at a different level.
461 CompLevel AdvancedThresholdPolicy::call_event(Method* method, CompLevel cur_level) {
462 CompLevel osr_level = MIN2((CompLevel) method->highest_osr_comp_level(),
463 common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true));
464 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level);
465
466 // If OSR method level is greater than the regular method level, the levels should be
467 // equalized by raising the regular method level in order to avoid OSRs during each
468 // invocation of the method.
469 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) {
470 MethodData* mdo = method->method_data();
471 guarantee(mdo != NULL, "MDO should not be NULL");
472 if (mdo->invocation_count() >= 1) {
473 next_level = CompLevel_full_optimization;
474 }
475 } else {
476 next_level = MAX2(osr_level, next_level);
477 }
478 return next_level;
479 }
480
481 // Determine if we should do an OSR compilation of a given method.
482 CompLevel AdvancedThresholdPolicy::loop_event(Method* method, CompLevel cur_level) {
483 CompLevel next_level = common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level, true);
484 if (cur_level == CompLevel_none) {
485 // If there is a live OSR method that means that we deopted to the interpreter
486 // for the transition.
487 CompLevel osr_level = MIN2((CompLevel)method->highest_osr_comp_level(), next_level);
488 if (osr_level > CompLevel_none) {
489 return osr_level;
490 }
491 }
492 return next_level;
493 }
494
495 // Update the rate and submit compile
496 void AdvancedThresholdPolicy::submit_compile(const methodHandle& mh, int bci, CompLevel level, JavaThread* thread) {
497 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count();
498 update_rate(os::javaTimeMillis(), mh());
499 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", thread);
500 }
501
502 // Handle the invocation event.
503 void AdvancedThresholdPolicy::method_invocation_event(const methodHandle& mh, const methodHandle& imh,
504 CompLevel level, nmethod* nm, JavaThread* thread) {
505 if (should_create_mdo(mh(), level)) {
506 create_mdo(mh, thread);
507 }
508 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh)) {
509 CompLevel next_level = call_event(mh(), level);
510 if (next_level != level) {
511 compile(mh, InvocationEntryBci, next_level, thread);
512 }
513 }
514 }
515
516 // Handle the back branch event. Notice that we can compile the method
517 // with a regular entry from here.
518 void AdvancedThresholdPolicy::method_back_branch_event(const methodHandle& mh, const methodHandle& imh,
519 int bci, CompLevel level, nmethod* nm, JavaThread* thread) {
520 if (should_create_mdo(mh(), level)) {
521 create_mdo(mh, thread);
522 }
523 // Check if MDO should be created for the inlined method
524 if (should_create_mdo(imh(), level)) {
525 create_mdo(imh, thread);
526 }
527
528 if (is_compilation_enabled()) {
529 CompLevel next_osr_level = loop_event(imh(), level);
530 CompLevel max_osr_level = (CompLevel)imh->highest_osr_comp_level();
531 // At the very least compile the OSR version
532 if (!CompileBroker::compilation_is_in_queue(imh) && (next_osr_level != level)) {
533 compile(imh, bci, next_osr_level, thread);
534 }
535
536 // Use loop event as an opportunity to also check if there's been
537 // enough calls.
538 CompLevel cur_level, next_level;
539 if (mh() != imh()) { // If there is an enclosing method
540 guarantee(nm != NULL, "Should have nmethod here");
541 cur_level = comp_level(mh());
542 next_level = call_event(mh(), cur_level);
543
544 if (max_osr_level == CompLevel_full_optimization) {
545 // The inlinee OSRed to full opt, we need to modify the enclosing method to avoid deopts
546 bool make_not_entrant = false;
547 if (nm->is_osr_method()) {
548 // This is an osr method, just make it not entrant and recompile later if needed
549 make_not_entrant = true;
550 } else {
551 if (next_level != CompLevel_full_optimization) {
552 // next_level is not full opt, so we need to recompile the
553 // enclosing method without the inlinee
554 cur_level = CompLevel_none;
555 make_not_entrant = true;
556 }
557 }
558 if (make_not_entrant) {
559 if (PrintTieredEvents) {
560 int osr_bci = nm->is_osr_method() ? nm->osr_entry_bci() : InvocationEntryBci;
561 print_event(MAKE_NOT_ENTRANT, mh(), mh(), osr_bci, level);
562 }
563 nm->make_not_entrant();
564 }
565 }
566 if (!CompileBroker::compilation_is_in_queue(mh)) {
567 // Fix up next_level if necessary to avoid deopts
568 if (next_level == CompLevel_limited_profile && max_osr_level == CompLevel_full_profile) {
569 next_level = CompLevel_full_profile;
570 }
571 if (cur_level != next_level) {
572 compile(mh, InvocationEntryBci, next_level, thread);
573 }
574 }
575 } else {
576 cur_level = comp_level(imh());
577 next_level = call_event(imh(), cur_level);
578 if (!CompileBroker::compilation_is_in_queue(imh) && (next_level != cur_level)) {
579 compile(imh, InvocationEntryBci, next_level, thread);
580 }
581 }
582 }
583 }
584
585 #endif // TIERED