Trying to compile a new VM
Julian Fitzell
julian at beta4.com
Sat Jan 10 00:53:11 UTC 2004
Tim Rowledge wrote:
> Julian Fitzell <julian at beta4.com> wrote:
>
>
>>The other problem then turned out to be that the older VMMaker in Ian's
>>image seemed to produce a VM that wouldn't run the new images. So I
>>need to go back up to the newest VMMaker, but Ian has patches applied
>>againt the old one and I have no idea whether I still need to apply
>>those or not. And I also have no idea how new an image I need to be
>>using to build the new VM.
>
> I'd love to know what patches to VMMaker are involved. I don't recall
> any recent changes that should affect runnability and if there have
> been we need to find them soon.
Well, his configured image is at
http://www-sor.inria.fr/~piumarta/squeak/unix/devel/vmm.tar.bz2
There a VMMUnix-ikp for sure. There were also a bunch of other patches
in his working directory and I wasn't sure which of those might be
needed or might be incorporated into VMMaker (keep in mind I'm a VMMaker
newbie :) ).
They may all be included in the new VMMaker, but that was what I was
completely unable to determine.
> The only almost-recent change I can see affecting performance is the
> move to keeping (most of) the globals in a big array. IIRC Ian had
> changed to making use of that quite a while ago.
Yeah, I saw references to that in the mail archive. The code does
indeed have a foo array with globals in it. I managed to get a VM
compiled using the 3.6g-2 distribution code with ned's floatbug patch
and the truncate primitive fix I reposted earlier today. This VM still
seems to run at about half the speed of Ian's binary so I guess it must
be difference of gcc versions (though it seems depressing that the speed
dropped by 1/2 from 3.3.1 to 3.3.3).
This image at least runs the new images and doesn't give primitive
failed errors on the truncate prim so I think I'm going to just start
using it. Only so much time I can spend on these things with deadlines
looming...
>>And finally, this generated source compiles fine on Linux but appears
>>not to compile on Solaris (and I've compiled the 3.6-g-2 sources on the
>>same Solaris box just fine). I don't know if this is a problem with
>>Ned's branch or with the old VMMaker (I haven't yet tried a Solaris
>>compile of a ned-branch/new VMMaker source tree). The error I get is:
>>
>>gnu-interp.c: In function `primitiveResponse':
>>gnu-interp.c:16425: `primitiveIndex' undeclared (first use in this function)
>>gnu-interp.c:16425: (Each undeclared identifier is reported only once
>>gnu-interp.c:16425: for each function it appears in.)
>
> primitiveIndex is one of the globals that sh/w/ould be in the array
> 'foo' and so this suggests to me that soemthing is wrong with that
> mechanism here. Look at the interp.c file and see where primitiveIndex
> is declared and where it is used. In my latest interp.c it is declared
> within 'foo' and first used in addNewMethodToCache as
> foo->primitiveIndex.
Yeah, mine is defined and first used in the same places. When I looked
at the code earlier, I couldn't for the life of me see why the error was
coming up and I still can't. But the source I generated from the 3.6g-2
release instead of the CVS code seems to have compiled ok on Solaris now.
I just tried a diff between interp.c in two the trees and, except for
variable definition order differences, the only major changes I can pick
out (on a quick skim) are that the one generated from CVS with Ian's vmm
image (old VMMaker) includes emphemeron stuff and a bunch of references
to displayBits. The one generated from the 3.6g-2 release using the
newest VMMaker doesn't have either of those.
If anyone wants a copy of the tree or diffs or anything, I'm happy to
provide. But it's working (albeit slowly) now so I don't think I'll
monkey with it any more.
Julian
More information about the Squeak-dev
mailing list
|