At 10:40 AM -0500 7/30/03, Sarita Adve wrote:
>With CnC, this is insufficient reasoning for test case 2, you have to
>consider the prohibited reads stuff for test case 2.
>In contrast, SC- (original email version plus about 10 more words) passes
>all the tests that Bill has so far. In other words, if you are happy with
>the guarantee that all SC executions past the discontinuity should have that
>write, then you should choose SC- because it is much simpler. (And possibly
>allows more optimizations.)
Oops. Although test case 2 had been given as a motivating example for
prohibited executions, we figured out that we actually don't need
prohibited writes for case 2 (we do need them for case 3).
In case 2, y=1 can be performed presciently as the first action in
the causal order, because it occurs in all non-prescient executions.
Once y=1 is performed presciently, all the remaining executions can
be performed non-presciently.
In case 3, you do need prohibited executions, because y=1 is not performed in all non-prescient executions.
------------------------------- JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel
This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:00:49 EDT