RFR: 8157698: PreservedMarksSet: don't launch workers if ParallelGCThreads==1 (extra small / simple)

[Thomas Schatzl]

Hi Tony,

On Di, 2016-05-24 at 12:11 -0400, Tony Printezis wrote:
> Just realized that restore_interval(WorkGang* workers) always calls
> run_task(), even if active worker num  == 1 (that was my bad). We
> generally call task.work(0) directly when active worker num == 1.
> 
> 
> http://cr.openjdk.java.net/~tonyp/8157698/webrev.1/
> 
> 
> Is this change reasonable?

  the change is okay.

However the RFE is based on a wrong assumption "HotSpot convention is to
call task.work(0) directly when ParallelGCThreads == 1." which is not
true any more.

This idiom has been removed (except for three places in CMS apparently)
quite some time ago (8u60 iirc?).

The reasons are that doing so does not have any noticeable advantage in
the general case. I.e. if you need to restore marks (which is supposed
to be uncommon), and you run the collector with a single thread (again,
unlikely) the actual work is typically much larger than this thread
switch.

Also, I would kind of prefer if code should not rely the existing value
of Workgang::active_workers(), as it then depends on random previous
phases to set it correctly. The reasons for the value it has right now
likely not apply to mark restoration at all.

So I object to push this change, unless it can be shown that it is
advantageous. Even then, I would like this change to recalculate the
number of active_workers according to some better policy.

Thanks,
  Thomas