[HEADSUP] big-endian LP32 bulk builders

[Thomas Klausner <wiz%netbsd.org@localhost>]

Please make sure to restart builds after
https://mail-index.netbsd.org/pkgsrc-changes/2025/10/24/msg332623.html
from scratch, in particular packages using zlib have a problem. Details below.

Thanks,
 Thomas


On Sat, Nov 22, 2025 at 06:25:19PM +0100, Thomas Klausner wrote:
> On Sun, Oct 19, 2025 at 10:28:18PM +0100, Taylor R Campbell wrote:
> > It turns out NetBSD has been shipping with a bug in its zconf.h, part
> > of zlib, that causes certain inline macros -- and potentially function
> > prototypes -- to be wrong in a way that breaks the build of certain
> > packages using autoconf on LP32 platforms.
> > 
> > PR lib/59711: "#define HAVE_UNISTD_H 1" breaks 32-bit libz
> > 
> > Specifically, zconf.h has logic that roughly boils down to:
> > 
> > #ifndef z_off_t
> > #  ifdef HAVE_UNISTD_H
> > #    define z_off_t	off_t
> > #  else
> > #    define z_off_t	long
> > #  endif
> > #endif
> > 
> > and similarly for z_ptrdiff_t.
> > 
> > The system library is built with z_off_t=long (i.e., without
> > HAVE_UNISTD_H), but if you build a program with autoconf defining
> > HAVE_UNISTD_H, that program will be built with the wrong definition of
> > z_off_t in various functions and inline macros in zlib.h.
> > 
> > The wrong definition has different size on LP32 platforms.  On
> > little-endian LP32 platforms, as long as the structure in question is
> > padded, using off_t (64-bit) where you should use long (32-bit) may
> > not cause too much trouble because the placement of numbers below 2^32
> > coincides, but on big-endian LP32 the low bytes are in the wrong place
> > so things quickly go awry as in the PR.
> > 
> > Of course we'll fix this in NetBSD so that it ignores HAVE_UNISTD_H
> > and gives the same z_off_t with or without that macro.  But there will
> > remain a problem with binary packages built against the x.0 release
> > with the broken zconf.h.  I expect we can work around it in pkgsrc by
> > putting the following fragment in devel/zlib/builtin.mk (plus an
> > explanatory comment with references):
> > 
> > .    if ${OPSYS} == "NetBSD" && ${OPSYS_VERSION} < 110000
> > CPPFLAGS+=	-Dz_off_t=long -Dz_ptrdiff_t=long
> > CFLAGS+=	-Dz_off_t=long -Dz_ptrdiff_t=long
> > .    endif
> > 
> > But this will also require rebuilding any packages that were already
> > built against the builtin zlib -- those packages need to be revbumped.
> > 
> > Thoughts?
> 
> Since this probably means bumping >50% of pkgsrc, I'm slightly against
> the bump (not against the fix, of course).
> 
> I suggest instead that we loudly shout that bulk builds on the
> affected platforms (big-endian LP32, as I undestand you) should be
> restarted from scratch.
> 
> If that's fine with you, please commit the zlib diff (or ask me to)
> and then we can re-send your email with a "[HEADSUP] big-endian lp32
> bulk builders" subject.
>  Thomas