• downs (13/13) May 04 2008 If I remember correctly, the biggest problem with creating a concurrent ...

downs <default_357-line yahoo.de> writes:

If I remember correctly, the biggest problem with creating a concurrent garbage
collector for a language like D, is moving pointer information from <before the
mark scan> to <behind the mark scan> while the scan is running,
thus creating the impression that an object isn't referenced anymore.

Thus, I request the introduction of a safemode { } block (keyword name up for
debate). While inside a safemode { } block, a thread is considered paused from
the GC's point of view.

When all, or a sufficient number of, threads are in safemode, the GC can be run
concurrently without worrying about pointer movement.

Basically, safemode { } is mainly intended for areas of code where most of the
time is spent on non-pointer logic.

This is because every pointer/reference write access in safemode { } is subject
to a GC callback, to check if it's okay to execute that write.

If the address being written lies in an area that has already been scanned, the
thread is paused until garbage collection ends.

(Alternatively, mark the written data manually).

If a safemode { } block ends before the GC has finished, the thread is blocked
until it has.

Programs should be able to use GC calls to give an estimate for the remaining
time in a safemode block, to aid the GC in deciding whether to run or not.

All of this should allow us, for a certain class of programs where pointer
logic consumes a very small part of total runtime, to run garbage collection
almost or entirely in parallel.

Whaddya think?

 --downs