To: J3                                                        09-179
From: John Reid
Subject:  Atomic subroutines
Date: 2009 April 03
References: J3/09-007r1

Discussion

This sentence, in para 2 of 13.1 specifies the effect of an atomic
subroutine:
"The effect of executing an atomic subroutine is as if the action on
the ATOM argument occurs instantaneously, and thus does not interfere
with other atomic actions that might occur asynchronously."

This sentence has an unintended effect.

Actions that occur instantaneously are ordered in time so the processor
is required to model this. For a single variable, this is easy enough.
The image holding the variable performs the actions in some order.
The image that makes the call waits for confirmation
that the action has occurred before doing anything else, which means
that the effect for it is as if the action were done instantaneously.

For variables on different images, this model does not work. Suppose
the value of a[1] is altered from 1 to 2 just before a[2] is altered
from 1 to 2. Suppose further that image 3 has better communication
with image 2 than image 1. This would mean that these statements,
executed by image 3
                call atomic_ref(x,a[1])
                call atomic_ref(y,a[2])
at about the same time might get 1 for x and 2 for y despite a[1]
being altered first. To make the model work requires synchronization,
which was not intended.

An edit is provided to correct this.


Edits to 09-007

[319:12-14] In 13.1, para 2, change the second sentence from
"The effect of executing an atomic subroutine is as if the action on
the ATOM argument occurs instantaneously, and thus does not interfere
with other atomic actions that might occur asynchronously."
to
"The effect of any two atomic subroutine calls for a particular
variable as the ATOM argument is as if one occurs before the other so
that they do not interfere with each other."