Mercurial > hg > graal-jvmci-8

comparison src/share/vm/runtime/advancedThresholdPolicy.cpp @ 16107:c7db9dc73694

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
disabled dialing C1 compilation back to partial profiling when Graal is the top tier compiler
author Doug Simon <doug.simon@oracle.com>
date Fri, 13 Jun 2014 18:42:33 +0200
parents 4ca6dc0799b6
children 52b4284cb496
comparison
equal deleted inserted replaced
16106:5d84706642ce 16107:c7db9dc73694
343 case CompLevel_none: 343 case CompLevel_none:
344 // If we were at full profile level, would we switch to full opt? 344 // If we were at full profile level, would we switch to full opt?
345 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { 345 if (common(p, method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) {
346 next_level = CompLevel_full_optimization; 346 next_level = CompLevel_full_optimization;
347 } else if ((this->*p)(i, b, cur_level)) { 347 } else if ((this->*p)(i, b, cur_level)) {
348 #ifdef COMPILERGRAAL
349 // Since Graal takes a while to warm up, its queue inevitably backs up during
350 // early VM execution. As of 2014-06-13, Graal's inliner assumes that the root
351 // compilation method and all potential inlinees have mature profiles (which
352 // includes type profiling). If it sees immature profiles, Graal's inliner
353 // can perform pathologically bad (e.g., causing OutOfMemoryErrors due to
354 // exploring/inlining too many graphs). Since a rewrite of the inliner is
355 // in progress, we simply disable the dialing back heuristic for now and will
356 // revisit this decision once the new inliner is completed.
357 next_level = CompLevel_full_profile;
358 #else
348 // C1-generated fully profiled code is about 30% slower than the limited profile 359 // C1-generated fully profiled code is about 30% slower than the limited profile
349 // code that has only invocation and backedge counters. The observation is that 360 // code that has only invocation and backedge counters. The observation is that
350 // if C2 queue is large enough we can spend too much time in the fully profiled code 361 // if C2 queue is large enough we can spend too much time in the fully profiled code
351 // while waiting for C2 to pick the method from the queue. To alleviate this problem 362 // while waiting for C2 to pick the method from the queue. To alleviate this problem
352 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 363 // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long
356 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 367 Tier3DelayOn * compiler_count(CompLevel_full_optimization)) {
357 next_level = CompLevel_limited_profile; 368 next_level = CompLevel_limited_profile;
358 } else { 369 } else {
359 next_level = CompLevel_full_profile; 370 next_level = CompLevel_full_profile;
360 } 371 }
372 #endif
361 } 373 }
362 break; 374 break;
363 case CompLevel_limited_profile: 375 case CompLevel_limited_profile:
364 if (is_method_profiled(method)) { 376 if (is_method_profiled(method)) {
365 // Special case: we got here because this method was fully profiled in the interpreter. 377 // Special case: we got here because this method was fully profiled in the interpreter.