stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8262323: do not special case JVMCI in tiered compilation policy

Browse Source 
Reviewed-by: kvn, never
This commit is contained in:
Doug Simon 2021-03-09 10:57:03 +00:00
parent 3022baa953
commit 4f0a12ec87
2 changed files with 21 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
src/hotspot/share/compiler/compilationPolicy.cpp
View File

@ -1054,33 +1054,18 @@ CompLevel CompilationPolicy::common(const methodHandle& method, CompLevel cur_le
if (common<Predicate>(method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) { if (common<Predicate>(method, CompLevel_full_profile, disable_feedback) == CompLevel_full_optimization) {
next_level = CompLevel_full_optimization; next_level = CompLevel_full_optimization;
} else if (!CompilationModeFlag::disable_intermediate() && Predicate::apply(i, b, cur_level, method)) { } else if (!CompilationModeFlag::disable_intermediate() && Predicate::apply(i, b, cur_level, method)) {
#if INCLUDE_JVMCI // C1-generated fully profiled code is about 30% slower than the limited profile
if (EnableJVMCI && UseJVMCICompiler) { // code that has only invocation and backedge counters. The observation is that
// Since JVMCI takes a while to warm up, its queue inevitably backs up during // if C2 queue is large enough we can spend too much time in the fully profiled code
// early VM execution. As of 2014-06-13, JVMCI's inliner assumes that the root // while waiting for C2 to pick the method from the queue. To alleviate this problem
// compilation method and all potential inlinees have mature profiles (which // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long
// includes type profiling). If it sees immature profiles, JVMCI's inliner // we choose to compile a limited profiled version and then recompile with full profiling
// can perform pathologically bad (e.g., causing OutOfMemoryErrors due to // when the load on C2 goes down.
// exploring/inlining too many graphs). Since a rewrite of the inliner is if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) >
// in progress, we simply disable the dialing back heuristic for now and will Tier3DelayOn * compiler_count(CompLevel_full_optimization)) {
// revisit this decision once the new inliner is completed. next_level = CompLevel_limited_profile;
} else {
next_level = CompLevel_full_profile; next_level = CompLevel_full_profile;
} else
#endif
{
// C1-generated fully profiled code is about 30% slower than the limited profile
// code that has only invocation and backedge counters. The observation is that
// if C2 queue is large enough we can spend too much time in the fully profiled code
// while waiting for C2 to pick the method from the queue. To alleviate this problem
// we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long
// we choose to compile a limited profiled version and then recompile with full profiling
// when the load on C2 goes down.
if (!disable_feedback && CompileBroker::queue_size(CompLevel_full_optimization) >
Tier3DelayOn * compiler_count(CompLevel_full_optimization)) {
next_level = CompLevel_limited_profile;
} else {
next_level = CompLevel_full_profile;
}
} }
} }
break; break;

10
src/hotspot/share/compiler/compilerDefinitions.cpp
View File

@ -451,6 +451,16 @@ void CompilerConfig::set_jvmci_specific_flags() {
if (FLAG_IS_DEFAULT(NewSizeThreadIncrease)) { if (FLAG_IS_DEFAULT(NewSizeThreadIncrease)) {
FLAG_SET_DEFAULT(NewSizeThreadIncrease, MAX2(4*K, NewSizeThreadIncrease)); FLAG_SET_DEFAULT(NewSizeThreadIncrease, MAX2(4*K, NewSizeThreadIncrease));
} }
if (FLAG_IS_DEFAULT(Tier3DelayOn)) {
// This effectively prevents the compile broker scheduling tier 2
// (i.e., limited C1 profiling) compilations instead of tier 3
// (i.e., full C1 profiling) compilations when the tier 4 queue
// backs up (which is quite likely when using a non-AOT compiled JVMCI
// compiler). The observation based on jargraal is that the downside
// of skipping full profiling is much worse for performance than the
// queue backing up.
FLAG_SET_DEFAULT(Tier3DelayOn, 100000);
}
} // !UseJVMCINativeLibrary } // !UseJVMCINativeLibrary
} // UseJVMCICompiler } // UseJVMCICompiler
} }