Subject: C442 JGSI Review
Resent-Date: Thu, 30 Sep 1999 23:17:43 -0400
Resent-From: Geoffrey Fox <gcf@npac.syr.edu>
Resent-To: p_gcf@npac.syr.edu
Date: Fri, 17 Sep 1999 15:52:27 -0400 (EDT)
From: Bryan Carpenter <dbc@npac.syr.edu>
To: gcf@npac.syr.edu (Geoffrey Fox)
CC: dbc@npac.syr.edu (Bryan Carpenter)

The Java Memory Model is Fatally Flawed
William Pugh

a)Overall Recommendation

Probably a significant paper, but I hope the author can expand on some of
the points mentioned below.  A summary of the organization and conclusions
of the paper would be helpful (either in the introduction or conclusion).

b)Words suitable for authors

I suspect this paper raises important issues.  Perhaps some of the
points could be made in ways that would be clearer to a general audience.

The abstract gives a very brief summary of the highlights of the paper.

Section 1 reviews some basic properties of the Java  Memory Model as
it stands.  Section 2 explains some apparently unintended features
of the model, which make important compiler optimizations difficult.
The example of section 2.1 seems simple and could probably be made clearer
by adding a few more sentences (for example explaining the importance
of the aliasing).  Javasoft bug report #4242244 by Pugh contains a
simple concrete program based on the example.  It apparently detects
violation of the Java specification in several major optimizing JVMs.
Some words in the paper about this program would make the discussion
more concrete to me.

The example in Section 2.2 is more complicated.  Again I think it needs
more words.  I suppose the constraint that is specially worrisome is that
the read of `r.y' must precede the write of `p.x', even though there
is no data dependence between the relevant assignments.  But I can't
see where the text actually states this.  I wish the difficulties were
stated more explicitly, and specifics of the situation could be
compared with weaker models, such as coherence (mentioned earlier in
the section).  The author goes on to deduce that byte code reorderings
are illegal.  Again I would have liked more explanation.

Section 3 discusses weak memory models.  I wasn't sure whether the
problems with weak memory models also present with the JMM, or whether
the points in 3.1.1 actually arguments against weaker models?

Section 4 discusses unsafe idioms with the standard Java Memory Model.
I would have liked more explanation about the security issues connected
with Strings being atomic and immutable (which apparently they aren't,
really, with the current JMM).

In section 5 I would have liked to have known who are the "group of
people" discussing the safety guarantees.

c)Words for me if necessary

None.