Re: felix regression on arm?
On Mon, 2007-10-01 at 12:05 +0200, Stefano Zacchiroli wrote:
> On Mon, Oct 01, 2007 at 08:02:50PM +1000, skaller wrote:
> > The indication is that flxg has failed. Unfortunately, no diagnostics
> > are visible here. The error code 0xb is a segfault I think, possibly
> > because the machine stack for ocaml is too small?
>
> I don't know and frankly do not have time to handle this by myself.
Of course this is not your job Stefano! However it IS almost
certainly a problem in Ocaml, so it is relevant to make sure:
I guess a stack overflow (but might be heap overflow).
> Please discuss it with the felix maintainer/uploader in Debian; he
> should have access to the arm machine to debug this.
CC'ing Mike Furr..
> If nobody has time to fix this we should either remove felix from
> testing or drop arm support for the felix package.
What happens if the problem is simply that Ocaml or Ocaml
generated code cannot run because the machine is too small?
What if the problem is that Debian-Ocaml-Maint or the
Ocaml development team has misconfigured the Arm arch,
by, for example, failing to provide an adequate stack
or heap?
This is I think relevant to all Ocamls, it tends to come up
with Felix because Felix is one of the few packages that
actually runs regression tests as part of the build.
That is .. it isn't a problem with Felix per se:
I do not think Felix can ensure programs are linked with
an adequate stack or heap .. how do other people handle this issue?
[Frankly I'd remove all programs and libraries that did NOT
come with test code unless it is the kind of software that
can't be tested automatically -- but that's another issue]
So actually I think this is probably a policy issue.
Felix DOES build, just fine, on the ARM: its just that
some test cases fail, and they're ones that work on
other arches. AFAIK, apart from Filename module, Felix
itself does not use ANY platform dependent code.
I know Dypgen in 1.1.4 does pack into integers and had an issue
with using 31 bits but forgetting the sign, should have used
30 bits. I mention that one because it is one of the few cases
where Ocaml behaviour can be platform dependent (the bug was
not manifest on a 64 bit arch .. all the developers with 32 bit
arches got a failure until I found the problem).
Anyhow the point is: this is very likely NOT to be a bug
in Felix. It's a bug in Ocaml, which is not my responsibility.
--
John Skaller <skaller at users dot sf dot net>
Felix, successor to C++: http://felix.sf.net
Reply to: