Let me toss in my two cents, as the (co)designer of a concurrent GC
algorithm ("Sapphire", now patented!).
In Sapphire, we actually maintain, at some stages of the concurrent GC, TWO
copies of non-volatile instance variables. We essentially assume mutator
threads (i.e., not the collector) access non-volatiles in a race-free
way. Specifically, the collector provides for the possibility of a race
between one mutator writing a variable and the collector trying to copy it
to a different place, but if there are TWO (or more) unsynchronized writes
to the non-volatile, it is entirely possible that threads may observe the
value change and then change back, with one or more writes not showing up
in the new copy. In particular, if Thread 1 writes 1 into a variable that
is initially 0, and Thread 2 writes 2, and this happens concurrently with
collection, then depending on which copy of the variable Thread 3 reads,
and when, it may see the value go from 0 to 1 to 2 and then back to 1. We
do guarantee no "out of thin air" and that if there are no more writes, the
value eventually stabilizes.
My question is: In the memory model being proposed, is this behavior legal?
(I am hoping so, since if it is not, my clever way of making my algorithm
much cheaper than it would be for C (say) may be undermined.)
I suppose I can frame this as a test case if pressed :-).
-- Eliot
==============================================================================
J. Eliot B. Moss, Associate Professor http://www.cs.umass.edu/~moss www
Director, Arch. and Lang. Impl. Lab. +1-413-545-4206 voice
Department of Computer Science +1-413-695-4226 cell
140 Governor's Drive, Room 372 +1-413-545-1249 fax
University of Massachusetts at Amherst moss@cs.umass.edu email
Amherst, MA 01003-9264 USA +1-413-545-3733 Priscilla Coe sec'y
==============================================================================
-------------------------------
JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel
This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:01:00 EDT