RFR: 8253600: G1: Fully support pinned regions for full gc [v4]

[Albert Mingkun Yang]

On Tue, 10 Nov 2020 09:55:25 GMT, Thomas Schatzl <tschatzl at openjdk.org> wrote:

>> src/hotspot/share/gc/g1/g1FullCollector.hpp line 76:
>> 
>>> 74:   ReferenceProcessorSubjectToDiscoveryMutator _is_subject_mutator;
>>> 75: 
>>> 76:   G1FullGCHeapRegionAttr _region_attr_table;
>> 
>> I don't really see the point of this auxiliary data structure. Why can't we just query the underlying region for its type, pinned, open/close archive?
>
> The only reason is performance:
> 
> Consider memory access with that auxiliary data structure: it is a single dereference/memory load on a very dense data structure.
> 
> Directly querying the HeapRegion* is two dependent dereferences (first getting the HeapRegion*, then to the HeapRegion, then accessing the member) on a fairly large data structure (HeapRegion).
> 
> Since these loads are done for every reference (millions of times) the second is much slower, at best filling the cache with useless data as (roughly) a HeapRegion is around a cache line in size (iirc).
> 
> Further, with the current encoding of the values in region attribute table, the closed-or-pinned check can be done with a single check instead of some disjunction (ie. region->is_pinned() || region->is_closed(), although it is true that closed regions are always pinned), so saving even more code (and branch locations).

I see; thank you for the explanation. Please add some comments around this code.

-------------

PR: https://git.openjdk.java.net/jdk/pull/824