Re: Stack calibration

[Neil Jerram]

2008/9/28 Ludovic Courtès <address@hidden>:
>
> "Neil Jerram" <address@hidden> writes:
>
>> I've done this part a bit differently - see the libguile/Makefile.am
>> changes - because I couldn't see exactly how the recursive make
>> approach would work.  If you think recursive make would be
>> significantly better, can you describe or propose the detailed changes
>> that would be needed?
>
> Your proposition looks very good actually.  I suppose the generated
> makefile doesn't require recompilation of all `.lo' files to go from
> `libuguile' to `libguile', right?

If you mean does it actually compile them all again?: yes, I'm afraid
it does.  I think this is because the generated makefile thinks that
libuguile_la-eval.lo and libguile_la-eval.lo are separate objects.

If you mean does it need to?: no, it doesn't, because none of the
files apart from stackchk.c/stackchk-calibrated.c have actually
changed at all.

I currently don't know of a good solution for this.

It might work to define:

(i) a convenience library consisting of everything in libguile except
for stackchk.c/stackchk-calibrated.c

(ii) libuguile.la, consisting of the convenience library + stackchk.c

(iii) libguile.la, consisting of the convenience library +
stackchk-calibrated.c.

But I would be surprised if that didn't cause a regression on some
less mainstream platforms.

Do you have any suggestions?

> I'm not sure about cross-compilation (Dale Smith had also raised this
> issue on IRC some time ago).  IIUC, the user is expected to provide a
> `UGUILE_FOR_BUILD' at configure-time, which is then used to run
> `calibrate.scm'; however, `UGUILE_FOR_BUILD' runs on the host, not the
> target system, so the generated file will be erroneous, right?

Probably, yes.

> Thus, when cross-compiling, shouldn't we avoid stack calibration
> altogether and simply emit a warning a configure-time, for instance?

Well, ideally we should have a solution that works automatically in
all circumstances...

> At any rate, it's not a problem when cross-compiling with tools like
> Scratchbox, which actually "hide" the fact that we're cross-compiling
> and can run executables for the target system through an emulator.

Agreed.

Regards,
        Neil