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