Mercurial > hg > graal-compiler
comparison src/share/vm/runtime/advancedThresholdPolicy.cpp @ 2348:5d8f5a6dced7
7020403: Add AdvancedCompilationPolicy for tiered
Summary: This implements adaptive tiered compilation policy.
Reviewed-by: kvn, never
author | iveresov |
---|---|
date | Fri, 04 Mar 2011 15:14:16 -0800 |
parents | |
children | 97b64f73103b |
comparison
equal
deleted
inserted
replaced
2325:8c9c9ee30d71 | 2348:5d8f5a6dced7 |
---|---|
1 /* | |
2 * Copyright (c) 2010, 2011 Oracle and/or its affiliates. All rights reserved. | |
3 * ORACLE PROPRIETARY/CONFIDENTIAL. Use is subject to license terms. | |
4 */ | |
5 | |
6 #include "precompiled.hpp" | |
7 #include "runtime/advancedThresholdPolicy.hpp" | |
8 #include "runtime/simpleThresholdPolicy.inline.hpp" | |
9 | |
10 #ifdef TIERED | |
11 // Print an event. | |
12 void AdvancedThresholdPolicy::print_specific(EventType type, methodHandle mh, methodHandle imh, | |
13 int bci, CompLevel level) { | |
14 tty->print(" rate: "); | |
15 if (mh->prev_time() == 0) tty->print("n/a"); | |
16 else tty->print("%f", mh->rate()); | |
17 | |
18 tty->print(" k: %.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback), | |
19 threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback)); | |
20 | |
21 } | |
22 | |
23 void AdvancedThresholdPolicy::initialize() { | |
24 // Turn on ergonomic compiler count selection | |
25 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { | |
26 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); | |
27 } | |
28 int count = CICompilerCount; | |
29 if (CICompilerCountPerCPU) { | |
30 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n | |
31 int log_cpu = log2_intptr(os::active_processor_count()); | |
32 int loglog_cpu = log2_intptr(MAX2(log_cpu, 1)); | |
33 count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2; | |
34 } | |
35 | |
36 set_c1_count(MAX2(count / 3, 1)); | |
37 set_c2_count(MAX2(count - count / 3, 1)); | |
38 | |
39 // Some inlining tuning | |
40 #ifdef X86 | |
41 if (FLAG_IS_DEFAULT(InlineSmallCode)) { | |
42 FLAG_SET_DEFAULT(InlineSmallCode, 2000); | |
43 } | |
44 #endif | |
45 | |
46 #ifdef SPARC | |
47 if (FLAG_IS_DEFAULT(InlineSmallCode)) { | |
48 FLAG_SET_DEFAULT(InlineSmallCode, 2500); | |
49 } | |
50 #endif | |
51 | |
52 | |
53 set_start_time(os::javaTimeMillis()); | |
54 } | |
55 | |
56 // update_rate() is called from select_task() while holding a compile queue lock. | |
57 void AdvancedThresholdPolicy::update_rate(jlong t, methodOop m) { | |
58 if (is_old(m)) { | |
59 // We don't remove old methods from the queue, | |
60 // so we can just zero the rate. | |
61 m->set_rate(0); | |
62 return; | |
63 } | |
64 | |
65 // We don't update the rate if we've just came out of a safepoint. | |
66 // delta_s is the time since last safepoint in milliseconds. | |
67 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); | |
68 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement | |
69 // How many events were there since the last time? | |
70 int event_count = m->invocation_count() + m->backedge_count(); | |
71 int delta_e = event_count - m->prev_event_count(); | |
72 | |
73 // We should be running for at least 1ms. | |
74 if (delta_s >= TieredRateUpdateMinTime) { | |
75 // And we must've taken the previous point at least 1ms before. | |
76 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { | |
77 m->set_prev_time(t); | |
78 m->set_prev_event_count(event_count); | |
79 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond | |
80 } else | |
81 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { | |
82 // If nothing happened for 25ms, zero the rate. Don't modify prev values. | |
83 m->set_rate(0); | |
84 } | |
85 } | |
86 } | |
87 | |
88 // Check if this method has been stale from a given number of milliseconds. | |
89 // See select_task(). | |
90 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, methodOop m) { | |
91 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); | |
92 jlong delta_t = t - m->prev_time(); | |
93 if (delta_t > timeout && delta_s > timeout) { | |
94 int event_count = m->invocation_count() + m->backedge_count(); | |
95 int delta_e = event_count - m->prev_event_count(); | |
96 // Return true if there were no events. | |
97 return delta_e == 0; | |
98 } | |
99 return false; | |
100 } | |
101 | |
102 // We don't remove old methods from the compile queue even if they have | |
103 // very low activity. See select_task(). | |
104 bool AdvancedThresholdPolicy::is_old(methodOop method) { | |
105 return method->invocation_count() > 50000 || method->backedge_count() > 500000; | |
106 } | |
107 | |
108 double AdvancedThresholdPolicy::weight(methodOop method) { | |
109 return (method->rate() + 1) * ((method->invocation_count() + 1) * (method->backedge_count() + 1)); | |
110 } | |
111 | |
112 // Apply heuristics and return true if x should be compiled before y | |
113 bool AdvancedThresholdPolicy::compare_methods(methodOop x, methodOop y) { | |
114 if (x->highest_comp_level() > y->highest_comp_level()) { | |
115 // recompilation after deopt | |
116 return true; | |
117 } else | |
118 if (x->highest_comp_level() == y->highest_comp_level()) { | |
119 if (weight(x) > weight(y)) { | |
120 return true; | |
121 } | |
122 } | |
123 return false; | |
124 } | |
125 | |
126 // Is method profiled enough? | |
127 bool AdvancedThresholdPolicy::is_method_profiled(methodOop method) { | |
128 methodDataOop mdo = method->method_data(); | |
129 if (mdo != NULL) { | |
130 int i = mdo->invocation_count_delta(); | |
131 int b = mdo->backedge_count_delta(); | |
132 return call_predicate_helper<CompLevel_full_profile>(i, b, 1); | |
133 } | |
134 return false; | |
135 } | |
136 | |
137 // Called with the queue locked and with at least one element | |
138 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { | |
139 CompileTask *max_task = NULL; | |
140 methodOop max_method; | |
141 jlong t = os::javaTimeMillis(); | |
142 // Iterate through the queue and find a method with a maximum rate. | |
143 for (CompileTask* task = compile_queue->first(); task != NULL;) { | |
144 CompileTask* next_task = task->next(); | |
145 methodOop method = (methodOop)JNIHandles::resolve(task->method_handle()); | |
146 methodDataOop mdo = method->method_data(); | |
147 update_rate(t, method); | |
148 if (max_task == NULL) { | |
149 max_task = task; | |
150 max_method = method; | |
151 } else { | |
152 // If a method has been stale for some time, remove it from the queue. | |
153 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { | |
154 if (PrintTieredEvents) { | |
155 print_event(KILL, method, method, task->osr_bci(), (CompLevel)task->comp_level()); | |
156 } | |
157 CompileTaskWrapper ctw(task); // Frees the task | |
158 compile_queue->remove(task); | |
159 method->clear_queued_for_compilation(); | |
160 task = next_task; | |
161 continue; | |
162 } | |
163 | |
164 // Select a method with a higher rate | |
165 if (compare_methods(method, max_method)) { | |
166 max_task = task; | |
167 max_method = method; | |
168 } | |
169 } | |
170 task = next_task; | |
171 } | |
172 | |
173 if (max_task->comp_level() == CompLevel_full_profile && is_method_profiled(max_method)) { | |
174 max_task->set_comp_level(CompLevel_limited_profile); | |
175 if (PrintTieredEvents) { | |
176 print_event(UPDATE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); | |
177 } | |
178 } | |
179 | |
180 return max_task; | |
181 } | |
182 | |
183 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { | |
184 double queue_size = CompileBroker::queue_size(level); | |
185 int comp_count = compiler_count(level); | |
186 double k = queue_size / (feedback_k * comp_count) + 1; | |
187 return k; | |
188 } | |
189 | |
190 // Call and loop predicates determine whether a transition to a higher | |
191 // compilation level should be performed (pointers to predicate functions | |
192 // are passed to common()). | |
193 // Tier?LoadFeedback is basically a coefficient that determines of | |
194 // how many methods per compiler thread can be in the queue before | |
195 // the threshold values double. | |
196 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level) { | |
197 switch(cur_level) { | |
198 case CompLevel_none: | |
199 case CompLevel_limited_profile: { | |
200 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); | |
201 return loop_predicate_helper<CompLevel_none>(i, b, k); | |
202 } | |
203 case CompLevel_full_profile: { | |
204 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); | |
205 return loop_predicate_helper<CompLevel_full_profile>(i, b, k); | |
206 } | |
207 default: | |
208 return true; | |
209 } | |
210 } | |
211 | |
212 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level) { | |
213 switch(cur_level) { | |
214 case CompLevel_none: | |
215 case CompLevel_limited_profile: { | |
216 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); | |
217 return call_predicate_helper<CompLevel_none>(i, b, k); | |
218 } | |
219 case CompLevel_full_profile: { | |
220 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); | |
221 return call_predicate_helper<CompLevel_full_profile>(i, b, k); | |
222 } | |
223 default: | |
224 return true; | |
225 } | |
226 } | |
227 | |
228 // If a method is old enough and is still in the interpreter we would want to | |
229 // start profiling without waiting for the compiled method to arrive. | |
230 // We also take the load on compilers into the account. | |
231 bool AdvancedThresholdPolicy::should_create_mdo(methodOop method, CompLevel cur_level) { | |
232 if (cur_level == CompLevel_none && | |
233 CompileBroker::queue_size(CompLevel_full_optimization) <= | |
234 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { | |
235 int i = method->invocation_count(); | |
236 int b = method->backedge_count(); | |
237 double k = Tier0ProfilingStartPercentage / 100.0; | |
238 return call_predicate_helper<CompLevel_none>(i, b, k) || loop_predicate_helper<CompLevel_none>(i, b, k); | |
239 } | |
240 return false; | |
241 } | |
242 | |
243 // Create MDO if necessary. | |
244 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, TRAPS) { | |
245 if (mh->is_native() || mh->is_abstract() || mh->is_accessor()) return; | |
246 if (mh->method_data() == NULL) { | |
247 methodOopDesc::build_interpreter_method_data(mh, THREAD); | |
248 if (HAS_PENDING_EXCEPTION) { | |
249 CLEAR_PENDING_EXCEPTION; | |
250 } | |
251 } | |
252 } | |
253 | |
254 | |
255 /* | |
256 * Method states: | |
257 * 0 - interpreter (CompLevel_none) | |
258 * 1 - pure C1 (CompLevel_simple) | |
259 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) | |
260 * 3 - C1 with full profiling (CompLevel_full_profile) | |
261 * 4 - C2 (CompLevel_full_optimization) | |
262 * | |
263 * Common state transition patterns: | |
264 * a. 0 -> 3 -> 4. | |
265 * The most common path. But note that even in this straightforward case | |
266 * profiling can start at level 0 and finish at level 3. | |
267 * | |
268 * b. 0 -> 2 -> 3 -> 4. | |
269 * This case occures when the load on C2 is deemed too high. So, instead of transitioning | |
270 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to | |
271 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. | |
272 * | |
273 * c. 0 -> (3->2) -> 4. | |
274 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough | |
275 * to enable the profiling to fully occur at level 0. In this case we change the compilation level | |
276 * of the method to 2, because it'll allow it to run much faster without full profiling while c2 | |
277 * is compiling. | |
278 * | |
279 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. | |
280 * After a method was once compiled with C1 it can be identified as trivial and be compiled to | |
281 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. | |
282 * | |
283 * e. 0 -> 4. | |
284 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) | |
285 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because | |
286 * the compiled version already exists). | |
287 * | |
288 * Note that since state 0 can be reached from any other state via deoptimization different loops | |
289 * are possible. | |
290 * | |
291 */ | |
292 | |
293 // Common transition function. Given a predicate determines if a method should transition to another level. | |
294 CompLevel AdvancedThresholdPolicy::common(Predicate p, methodOop method, CompLevel cur_level) { | |
295 if (is_trivial(method)) return CompLevel_simple; | |
296 | |
297 CompLevel next_level = cur_level; | |
298 int i = method->invocation_count(); | |
299 int b = method->backedge_count(); | |
300 | |
301 switch(cur_level) { | |
302 case CompLevel_none: | |
303 // If we were at full profile level, would we switch to full opt? | |
304 if (common(p, method, CompLevel_full_profile) == CompLevel_full_optimization) { | |
305 next_level = CompLevel_full_optimization; | |
306 } else if ((this->*p)(i, b, cur_level)) { | |
307 // C1-generated fully profiled code is about 30% slower than the limited profile | |
308 // code that has only invocation and backedge counters. The observation is that | |
309 // if C2 queue is large enough we can spend too much time in the fully profiled code | |
310 // while waiting for C2 to pick the method from the queue. To alleviate this problem | |
311 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long | |
312 // we choose to compile a limited profiled version and then recompile with full profiling | |
313 // when the load on C2 goes down. | |
314 if (CompileBroker::queue_size(CompLevel_full_optimization) > | |
315 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { | |
316 next_level = CompLevel_limited_profile; | |
317 } else { | |
318 next_level = CompLevel_full_profile; | |
319 } | |
320 } | |
321 break; | |
322 case CompLevel_limited_profile: | |
323 if (is_method_profiled(method)) { | |
324 // Special case: we got here because this method was fully profiled in the interpreter. | |
325 next_level = CompLevel_full_optimization; | |
326 } else { | |
327 methodDataOop mdo = method->method_data(); | |
328 if (mdo != NULL) { | |
329 if (mdo->would_profile()) { | |
330 if (CompileBroker::queue_size(CompLevel_full_optimization) <= | |
331 Tier3DelayOff * compiler_count(CompLevel_full_optimization) && | |
332 (this->*p)(i, b, cur_level)) { | |
333 next_level = CompLevel_full_profile; | |
334 } | |
335 } else { | |
336 next_level = CompLevel_full_optimization; | |
337 } | |
338 } | |
339 } | |
340 break; | |
341 case CompLevel_full_profile: | |
342 { | |
343 methodDataOop mdo = method->method_data(); | |
344 if (mdo != NULL) { | |
345 if (mdo->would_profile()) { | |
346 int mdo_i = mdo->invocation_count_delta(); | |
347 int mdo_b = mdo->backedge_count_delta(); | |
348 if ((this->*p)(mdo_i, mdo_b, cur_level)) { | |
349 next_level = CompLevel_full_optimization; | |
350 } | |
351 } else { | |
352 next_level = CompLevel_full_optimization; | |
353 } | |
354 } | |
355 } | |
356 break; | |
357 } | |
358 return next_level; | |
359 } | |
360 | |
361 // Determine if a method should be compiled with a normal entry point at a different level. | |
362 CompLevel AdvancedThresholdPolicy::call_event(methodOop method, CompLevel cur_level) { | |
363 CompLevel osr_level = (CompLevel) method->highest_osr_comp_level(); | |
364 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); | |
365 | |
366 // If OSR method level is greater than the regular method level, the levels should be | |
367 // equalized by raising the regular method level in order to avoid OSRs during each | |
368 // invocation of the method. | |
369 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { | |
370 methodDataOop mdo = method->method_data(); | |
371 guarantee(mdo != NULL, "MDO should not be NULL"); | |
372 if (mdo->invocation_count() >= 1) { | |
373 next_level = CompLevel_full_optimization; | |
374 } | |
375 } else { | |
376 next_level = MAX2(osr_level, next_level); | |
377 } | |
378 | |
379 return next_level; | |
380 } | |
381 | |
382 // Determine if we should do an OSR compilation of a given method. | |
383 CompLevel AdvancedThresholdPolicy::loop_event(methodOop method, CompLevel cur_level) { | |
384 if (cur_level == CompLevel_none) { | |
385 // If there is a live OSR method that means that we deopted to the interpreter | |
386 // for the transition. | |
387 CompLevel osr_level = (CompLevel)method->highest_osr_comp_level(); | |
388 if (osr_level > CompLevel_none) { | |
389 return osr_level; | |
390 } | |
391 } | |
392 return common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level); | |
393 } | |
394 | |
395 // Update the rate and submit compile | |
396 void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, TRAPS) { | |
397 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); | |
398 update_rate(os::javaTimeMillis(), mh()); | |
399 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", THREAD); | |
400 } | |
401 | |
402 | |
403 // Handle the invocation event. | |
404 void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh, | |
405 CompLevel level, TRAPS) { | |
406 if (should_create_mdo(mh(), level)) { | |
407 create_mdo(mh, THREAD); | |
408 } | |
409 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, InvocationEntryBci)) { | |
410 CompLevel next_level = call_event(mh(), level); | |
411 if (next_level != level) { | |
412 compile(mh, InvocationEntryBci, next_level, THREAD); | |
413 } | |
414 } | |
415 } | |
416 | |
417 // Handle the back branch event. Notice that we can compile the method | |
418 // with a regular entry from here. | |
419 void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh, | |
420 int bci, CompLevel level, TRAPS) { | |
421 if (should_create_mdo(mh(), level)) { | |
422 create_mdo(mh, THREAD); | |
423 } | |
424 | |
425 // If the method is already compiling, quickly bail out. | |
426 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, bci)) { | |
427 // Use loop event as an opportinity to also check there's been | |
428 // enough calls. | |
429 CompLevel cur_level = comp_level(mh()); | |
430 CompLevel next_level = call_event(mh(), cur_level); | |
431 CompLevel next_osr_level = loop_event(mh(), level); | |
432 if (next_osr_level == CompLevel_limited_profile) { | |
433 next_osr_level = CompLevel_full_profile; // OSRs are supposed to be for very hot methods. | |
434 } | |
435 next_level = MAX2(next_level, | |
436 next_osr_level < CompLevel_full_optimization ? next_osr_level : cur_level); | |
437 bool is_compiling = false; | |
438 if (next_level != cur_level) { | |
439 compile(mh, InvocationEntryBci, next_level, THREAD); | |
440 is_compiling = true; | |
441 } | |
442 | |
443 // Do the OSR version | |
444 if (!is_compiling && next_osr_level != level) { | |
445 compile(mh, bci, next_osr_level, THREAD); | |
446 } | |
447 } | |
448 } | |
449 | |
450 #endif // TIERED |