Print this page
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/runtime/advancedThresholdPolicy.cpp
+++ new/src/share/vm/runtime/advancedThresholdPolicy.cpp
1 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 -*/
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 + */
5 24
6 25 #include "precompiled.hpp"
7 26 #include "runtime/advancedThresholdPolicy.hpp"
8 27 #include "runtime/simpleThresholdPolicy.inline.hpp"
9 28
10 29 #ifdef TIERED
11 30 // Print an event.
12 31 void AdvancedThresholdPolicy::print_specific(EventType type, methodHandle mh, methodHandle imh,
13 32 int bci, CompLevel level) {
14 33 tty->print(" rate: ");
15 34 if (mh->prev_time() == 0) tty->print("n/a");
16 35 else tty->print("%f", mh->rate());
17 36
18 37 tty->print(" k: %.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback),
19 38 threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback));
20 39
21 40 }
22 41
23 42 void AdvancedThresholdPolicy::initialize() {
24 43 // Turn on ergonomic compiler count selection
25 44 if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) {
26 45 FLAG_SET_DEFAULT(CICompilerCountPerCPU, true);
27 46 }
28 47 int count = CICompilerCount;
29 48 if (CICompilerCountPerCPU) {
30 49 // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n
31 50 int log_cpu = log2_intptr(os::active_processor_count());
32 51 int loglog_cpu = log2_intptr(MAX2(log_cpu, 1));
33 52 count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2;
34 53 }
35 54
36 55 set_c1_count(MAX2(count / 3, 1));
37 56 set_c2_count(MAX2(count - count / 3, 1));
38 57
39 58 // Some inlining tuning
40 59 #ifdef X86
41 60 if (FLAG_IS_DEFAULT(InlineSmallCode)) {
42 61 FLAG_SET_DEFAULT(InlineSmallCode, 2000);
43 62 }
44 63 #endif
45 64
46 65 #ifdef SPARC
47 66 if (FLAG_IS_DEFAULT(InlineSmallCode)) {
48 67 FLAG_SET_DEFAULT(InlineSmallCode, 2500);
49 68 }
50 69 #endif
51 70
52 71
53 72 set_start_time(os::javaTimeMillis());
54 73 }
55 74
56 75 // update_rate() is called from select_task() while holding a compile queue lock.
57 76 void AdvancedThresholdPolicy::update_rate(jlong t, methodOop m) {
58 77 if (is_old(m)) {
59 78 // We don't remove old methods from the queue,
60 79 // so we can just zero the rate.
61 80 m->set_rate(0);
62 81 return;
63 82 }
64 83
65 84 // We don't update the rate if we've just came out of a safepoint.
66 85 // delta_s is the time since last safepoint in milliseconds.
67 86 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint();
68 87 jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement
69 88 // How many events were there since the last time?
70 89 int event_count = m->invocation_count() + m->backedge_count();
71 90 int delta_e = event_count - m->prev_event_count();
72 91
73 92 // We should be running for at least 1ms.
74 93 if (delta_s >= TieredRateUpdateMinTime) {
75 94 // And we must've taken the previous point at least 1ms before.
76 95 if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) {
77 96 m->set_prev_time(t);
78 97 m->set_prev_event_count(event_count);
79 98 m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond
80 99 } else
81 100 if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) {
82 101 // If nothing happened for 25ms, zero the rate. Don't modify prev values.
83 102 m->set_rate(0);
84 103 }
85 104 }
86 105 }
87 106
88 107 // Check if this method has been stale from a given number of milliseconds.
89 108 // See select_task().
90 109 bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, methodOop m) {
91 110 jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint();
92 111 jlong delta_t = t - m->prev_time();
93 112 if (delta_t > timeout && delta_s > timeout) {
94 113 int event_count = m->invocation_count() + m->backedge_count();
95 114 int delta_e = event_count - m->prev_event_count();
96 115 // Return true if there were no events.
97 116 return delta_e == 0;
98 117 }
99 118 return false;
100 119 }
101 120
102 121 // We don't remove old methods from the compile queue even if they have
103 122 // very low activity. See select_task().
104 123 bool AdvancedThresholdPolicy::is_old(methodOop method) {
105 124 return method->invocation_count() > 50000 || method->backedge_count() > 500000;
106 125 }
107 126
108 127 double AdvancedThresholdPolicy::weight(methodOop method) {
109 128 return (method->rate() + 1) * ((method->invocation_count() + 1) * (method->backedge_count() + 1));
110 129 }
111 130
112 131 // Apply heuristics and return true if x should be compiled before y
113 132 bool AdvancedThresholdPolicy::compare_methods(methodOop x, methodOop y) {
114 133 if (x->highest_comp_level() > y->highest_comp_level()) {
115 134 // recompilation after deopt
116 135 return true;
117 136 } else
118 137 if (x->highest_comp_level() == y->highest_comp_level()) {
119 138 if (weight(x) > weight(y)) {
120 139 return true;
121 140 }
122 141 }
123 142 return false;
124 143 }
125 144
126 145 // Is method profiled enough?
127 146 bool AdvancedThresholdPolicy::is_method_profiled(methodOop method) {
128 147 methodDataOop mdo = method->method_data();
129 148 if (mdo != NULL) {
130 149 int i = mdo->invocation_count_delta();
131 150 int b = mdo->backedge_count_delta();
132 151 return call_predicate_helper<CompLevel_full_profile>(i, b, 1);
133 152 }
134 153 return false;
135 154 }
136 155
137 156 // Called with the queue locked and with at least one element
138 157 CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) {
139 158 CompileTask *max_task = NULL;
140 159 methodOop max_method;
141 160 jlong t = os::javaTimeMillis();
142 161 // Iterate through the queue and find a method with a maximum rate.
143 162 for (CompileTask* task = compile_queue->first(); task != NULL;) {
144 163 CompileTask* next_task = task->next();
145 164 methodOop method = (methodOop)JNIHandles::resolve(task->method_handle());
146 165 methodDataOop mdo = method->method_data();
147 166 update_rate(t, method);
148 167 if (max_task == NULL) {
149 168 max_task = task;
150 169 max_method = method;
151 170 } else {
152 171 // If a method has been stale for some time, remove it from the queue.
153 172 if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) {
154 173 if (PrintTieredEvents) {
155 174 print_event(KILL, method, method, task->osr_bci(), (CompLevel)task->comp_level());
156 175 }
157 176 CompileTaskWrapper ctw(task); // Frees the task
158 177 compile_queue->remove(task);
159 178 method->clear_queued_for_compilation();
160 179 task = next_task;
161 180 continue;
162 181 }
163 182
164 183 // Select a method with a higher rate
165 184 if (compare_methods(method, max_method)) {
166 185 max_task = task;
167 186 max_method = method;
168 187 }
169 188 }
170 189 task = next_task;
171 190 }
172 191
173 192 if (max_task->comp_level() == CompLevel_full_profile && is_method_profiled(max_method)) {
174 193 max_task->set_comp_level(CompLevel_limited_profile);
175 194 if (PrintTieredEvents) {
176 195 print_event(UPDATE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level());
177 196 }
178 197 }
179 198
180 199 return max_task;
181 200 }
182 201
183 202 double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) {
184 203 double queue_size = CompileBroker::queue_size(level);
185 204 int comp_count = compiler_count(level);
186 205 double k = queue_size / (feedback_k * comp_count) + 1;
187 206 return k;
188 207 }
189 208
190 209 // Call and loop predicates determine whether a transition to a higher
191 210 // compilation level should be performed (pointers to predicate functions
192 211 // are passed to common()).
193 212 // Tier?LoadFeedback is basically a coefficient that determines of
194 213 // how many methods per compiler thread can be in the queue before
195 214 // the threshold values double.
196 215 bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level) {
197 216 switch(cur_level) {
198 217 case CompLevel_none:
199 218 case CompLevel_limited_profile: {
200 219 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback);
201 220 return loop_predicate_helper<CompLevel_none>(i, b, k);
202 221 }
203 222 case CompLevel_full_profile: {
204 223 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback);
205 224 return loop_predicate_helper<CompLevel_full_profile>(i, b, k);
206 225 }
207 226 default:
208 227 return true;
209 228 }
210 229 }
211 230
212 231 bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level) {
213 232 switch(cur_level) {
214 233 case CompLevel_none:
215 234 case CompLevel_limited_profile: {
216 235 double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback);
217 236 return call_predicate_helper<CompLevel_none>(i, b, k);
218 237 }
219 238 case CompLevel_full_profile: {
220 239 double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback);
221 240 return call_predicate_helper<CompLevel_full_profile>(i, b, k);
222 241 }
223 242 default:
224 243 return true;
225 244 }
226 245 }
227 246
228 247 // If a method is old enough and is still in the interpreter we would want to
229 248 // start profiling without waiting for the compiled method to arrive.
230 249 // We also take the load on compilers into the account.
231 250 bool AdvancedThresholdPolicy::should_create_mdo(methodOop method, CompLevel cur_level) {
232 251 if (cur_level == CompLevel_none &&
233 252 CompileBroker::queue_size(CompLevel_full_optimization) <=
234 253 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) {
235 254 int i = method->invocation_count();
236 255 int b = method->backedge_count();
237 256 double k = Tier0ProfilingStartPercentage / 100.0;
238 257 return call_predicate_helper<CompLevel_none>(i, b, k) || loop_predicate_helper<CompLevel_none>(i, b, k);
239 258 }
240 259 return false;
241 260 }
242 261
243 262 // Create MDO if necessary.
244 263 void AdvancedThresholdPolicy::create_mdo(methodHandle mh, TRAPS) {
245 264 if (mh->is_native() || mh->is_abstract() || mh->is_accessor()) return;
246 265 if (mh->method_data() == NULL) {
247 266 methodOopDesc::build_interpreter_method_data(mh, THREAD);
248 267 if (HAS_PENDING_EXCEPTION) {
249 268 CLEAR_PENDING_EXCEPTION;
250 269 }
251 270 }
252 271 }
253 272
254 273
255 274 /*
256 275 * Method states:
257 276 * 0 - interpreter (CompLevel_none)
258 277 * 1 - pure C1 (CompLevel_simple)
259 278 * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile)
260 279 * 3 - C1 with full profiling (CompLevel_full_profile)
261 280 * 4 - C2 (CompLevel_full_optimization)
262 281 *
263 282 * Common state transition patterns:
264 283 * a. 0 -> 3 -> 4.
265 284 * The most common path. But note that even in this straightforward case
266 285 * profiling can start at level 0 and finish at level 3.
267 286 *
268 287 * b. 0 -> 2 -> 3 -> 4.
269 288 * This case occures when the load on C2 is deemed too high. So, instead of transitioning
270 289 * into state 3 directly and over-profiling while a method is in the C2 queue we transition to
271 290 * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs.
272 291 *
273 292 * c. 0 -> (3->2) -> 4.
274 293 * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough
275 294 * to enable the profiling to fully occur at level 0. In this case we change the compilation level
276 295 * of the method to 2, because it'll allow it to run much faster without full profiling while c2
277 296 * is compiling.
278 297 *
279 298 * d. 0 -> 3 -> 1 or 0 -> 2 -> 1.
280 299 * After a method was once compiled with C1 it can be identified as trivial and be compiled to
281 300 * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1.
282 301 *
283 302 * e. 0 -> 4.
284 303 * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter)
285 304 * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because
286 305 * the compiled version already exists).
287 306 *
288 307 * Note that since state 0 can be reached from any other state via deoptimization different loops
289 308 * are possible.
290 309 *
291 310 */
292 311
293 312 // Common transition function. Given a predicate determines if a method should transition to another level.
294 313 CompLevel AdvancedThresholdPolicy::common(Predicate p, methodOop method, CompLevel cur_level) {
295 314 if (is_trivial(method)) return CompLevel_simple;
296 315
297 316 CompLevel next_level = cur_level;
298 317 int i = method->invocation_count();
299 318 int b = method->backedge_count();
300 319
301 320 switch(cur_level) {
302 321 case CompLevel_none:
303 322 // If we were at full profile level, would we switch to full opt?
304 323 if (common(p, method, CompLevel_full_profile) == CompLevel_full_optimization) {
305 324 next_level = CompLevel_full_optimization;
306 325 } else if ((this->*p)(i, b, cur_level)) {
307 326 // C1-generated fully profiled code is about 30% slower than the limited profile
308 327 // code that has only invocation and backedge counters. The observation is that
309 328 // if C2 queue is large enough we can spend too much time in the fully profiled code
310 329 // while waiting for C2 to pick the method from the queue. To alleviate this problem
311 330 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long
312 331 // we choose to compile a limited profiled version and then recompile with full profiling
313 332 // when the load on C2 goes down.
314 333 if (CompileBroker::queue_size(CompLevel_full_optimization) >
315 334 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) {
316 335 next_level = CompLevel_limited_profile;
317 336 } else {
318 337 next_level = CompLevel_full_profile;
319 338 }
320 339 }
321 340 break;
322 341 case CompLevel_limited_profile:
323 342 if (is_method_profiled(method)) {
324 343 // Special case: we got here because this method was fully profiled in the interpreter.
325 344 next_level = CompLevel_full_optimization;
326 345 } else {
327 346 methodDataOop mdo = method->method_data();
328 347 if (mdo != NULL) {
329 348 if (mdo->would_profile()) {
330 349 if (CompileBroker::queue_size(CompLevel_full_optimization) <=
331 350 Tier3DelayOff * compiler_count(CompLevel_full_optimization) &&
332 351 (this->*p)(i, b, cur_level)) {
333 352 next_level = CompLevel_full_profile;
334 353 }
335 354 } else {
336 355 next_level = CompLevel_full_optimization;
337 356 }
338 357 }
339 358 }
340 359 break;
341 360 case CompLevel_full_profile:
342 361 {
343 362 methodDataOop mdo = method->method_data();
344 363 if (mdo != NULL) {
345 364 if (mdo->would_profile()) {
346 365 int mdo_i = mdo->invocation_count_delta();
347 366 int mdo_b = mdo->backedge_count_delta();
348 367 if ((this->*p)(mdo_i, mdo_b, cur_level)) {
349 368 next_level = CompLevel_full_optimization;
350 369 }
351 370 } else {
352 371 next_level = CompLevel_full_optimization;
353 372 }
354 373 }
355 374 }
356 375 break;
357 376 }
358 377 return next_level;
359 378 }
360 379
361 380 // Determine if a method should be compiled with a normal entry point at a different level.
362 381 CompLevel AdvancedThresholdPolicy::call_event(methodOop method, CompLevel cur_level) {
363 382 CompLevel osr_level = (CompLevel) method->highest_osr_comp_level();
364 383 CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level);
365 384
366 385 // If OSR method level is greater than the regular method level, the levels should be
367 386 // equalized by raising the regular method level in order to avoid OSRs during each
368 387 // invocation of the method.
369 388 if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) {
370 389 methodDataOop mdo = method->method_data();
371 390 guarantee(mdo != NULL, "MDO should not be NULL");
372 391 if (mdo->invocation_count() >= 1) {
373 392 next_level = CompLevel_full_optimization;
374 393 }
375 394 } else {
376 395 next_level = MAX2(osr_level, next_level);
377 396 }
378 397
379 398 return next_level;
380 399 }
381 400
382 401 // Determine if we should do an OSR compilation of a given method.
383 402 CompLevel AdvancedThresholdPolicy::loop_event(methodOop method, CompLevel cur_level) {
384 403 if (cur_level == CompLevel_none) {
385 404 // If there is a live OSR method that means that we deopted to the interpreter
386 405 // for the transition.
387 406 CompLevel osr_level = (CompLevel)method->highest_osr_comp_level();
388 407 if (osr_level > CompLevel_none) {
389 408 return osr_level;
390 409 }
391 410 }
392 411 return common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level);
393 412 }
394 413
395 414 // Update the rate and submit compile
396 415 void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, TRAPS) {
397 416 int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count();
398 417 update_rate(os::javaTimeMillis(), mh());
399 418 CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", THREAD);
400 419 }
401 420
402 421
403 422 // Handle the invocation event.
404 423 void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh,
405 424 CompLevel level, TRAPS) {
406 425 if (should_create_mdo(mh(), level)) {
407 426 create_mdo(mh, THREAD);
408 427 }
409 428 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, InvocationEntryBci)) {
410 429 CompLevel next_level = call_event(mh(), level);
411 430 if (next_level != level) {
412 431 compile(mh, InvocationEntryBci, next_level, THREAD);
413 432 }
414 433 }
415 434 }
416 435
417 436 // Handle the back branch event. Notice that we can compile the method
418 437 // with a regular entry from here.
419 438 void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh,
420 439 int bci, CompLevel level, TRAPS) {
421 440 if (should_create_mdo(mh(), level)) {
422 441 create_mdo(mh, THREAD);
423 442 }
424 443
425 444 // If the method is already compiling, quickly bail out.
426 445 if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, bci)) {
427 446 // Use loop event as an opportinity to also check there's been
428 447 // enough calls.
429 448 CompLevel cur_level = comp_level(mh());
430 449 CompLevel next_level = call_event(mh(), cur_level);
431 450 CompLevel next_osr_level = loop_event(mh(), level);
432 451 if (next_osr_level == CompLevel_limited_profile) {
433 452 next_osr_level = CompLevel_full_profile; // OSRs are supposed to be for very hot methods.
434 453 }
435 454 next_level = MAX2(next_level,
436 455 next_osr_level < CompLevel_full_optimization ? next_osr_level : cur_level);
437 456 bool is_compiling = false;
438 457 if (next_level != cur_level) {
439 458 compile(mh, InvocationEntryBci, next_level, THREAD);
440 459 is_compiling = true;
441 460 }
442 461
443 462 // Do the OSR version
444 463 if (!is_compiling && next_osr_level != level) {
445 464 compile(mh, bci, next_osr_level, THREAD);
446 465 }
447 466 }
448 467 }
449 468
450 469 #endif // TIERED
|
↓ open down ↓ |
436 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX