To: J3                                                 08-304
From: Nick Maclaren
Subject: Atomic memory operations
Date: 2008 November 20
References: J3/08-007r2 (WG5/N1723), 08-297r1


Discussion:

This paper proposes an edit to 08-297r1 to specify that the ordering
semantics of atomic operations are processor dependent.  The issues
are discussed at length in N1744, N1745 and N1748, and referred to
in other documents.

The purpose of this paragraph is to state in normative text that Fortran
does not require sequential consistency for atomic actions, and
therefore that programs that assume it are not portable.  The reason
that this should be normative is that it is a partial exception to the
rules of a "sequence, in time" that apply to ordinary actions.  Because
of existing differences between systems, a more precise specification is
not feasible.

It is also to encourage implementors to document whatever it is that
they do provide.

[313:13.1p2++] Append to the end of the new paragraph 2 (from
08-297r1, beginning "An <<atomic subroutine>...") of "13.1 Classes of
intrinsic procedures":

"The sequence of atomic actions within ordered segments is specified
in 2.4.5.  How sequences of atomic actions in unordered segments
interleave with each other is processor dependent."

[c.460 Annex A] Add:

"- how the sequences of atomic actions in unordered segments
interleave;"