Re: [gentoo-alt] Prefix bootstrap state

[Ruud Koolen]

On Thursday 06 February 2014 09:45:10 Fabian Groffen wrote:
> All,
>
> I started last FOSDEM to give the bootstrap process some desparately
> needed love again.

I'm with you, brother :)

> Sofar, I've got bootstraps succeed on OpenIndiana x86 and x64.  (yay)
>
> Currently, bootstrap is broken on:
> - Darwin: it hangs on bash, which DEMANDS readline, despite disabling
>   it, I think the best course of action here is to actually compile
>   readline before bash, but I don't recall why we didn't do that
>   (probably because it used to break).

readline in term demands ncurses, and this gets you into trouble with 
libraries I mentioned earlier: the libraries built by the host compiler may 
not be ABI-compatible with the libraries built by prefix's gcc later on. This 
means that when you later rebuild readline (or ncurses) during emerge -e 
system, you'll suddenly find yourself without a working bash.

> - Solaris 10: the old gcc compiler can't compile
>   baselayout-prefix/libeinfo.  I'm about to patch that piece of code
>   tonight so it will compile.

There are two problems here: libeinfo uses c99 which is not available 
everywhere (and in particular not on solaris 10) (this is easily patched by 
getting rid of the stdbool.h include); and then there's the linker error, for 
which you found a workaround but which I don't understand properly.

> - Fedora 17: Python 3 (whatelse) gives problems compiling modules.  The
>   oddest one _decimal, which breaks on some weird funcs, but maybe those
>   are arithmetic ones from the compiler that we may or may not have due
>   to the bootstrap setup.  Solution might be to block Python 3 during
>   bootstrap stages if it demands a newer compiler or something.

Regarding _decimal, is this bug 490774, perhaps? That one SHOULD be fixed.

For _crypt and _nis, the problem is that python incorrectly hardcodes some 
libc paths. I was planning to look into solving this properly when the RAP 
infrastructure is in place, as it probably needs somewhat different patches 
for RAP and mainline prefix.

>
> I intend on running the bootstraps again on OpenIndiana once I fixed the
> above issues.  Any help is welcome of course.  Especially now the
> readline case for Darwin just needs people to try and see if that
> results in problems for them.
>
> Over the last few days I've fixed quite some issues caused by implicit
> dependencies.  One reason I think Darwin currently breaks is because it
> doesn't come with readline (but libedit) while most other systems /do/
> have readline installed.  Similar reason I had a problem with bison on
> Fedora 17 because its configure demanded perl, and Fedora doesn't
> install that by default, while most other systems have it around...
>
> Fabian

I hit the same problem in freebsd recently, didn't look into it in detail yet.

Why do we need bison in stage1, exactly? If we can get rid of that, adding 
perl to the early stage3 list should fix this one.

-- Ruud