tg at gmplib.org
Tue May 3 22:43:44 CEST 2011
Marc Glisse <marc.glisse at inria.fr> writes:
> The latter still is not robust, since we don't make any promises about
> the existence of configfsf.guess. (They could provide their own
> config.guess, stored outside of GMP, though.)
Would it make sense to introduce some option to gmp to do this robustly?
Is there a compelling reason for that?
(kind of the reverse of --enable-fat, compile only for the lowest
common base of the category, but still with asm) Maybe not, but that
would be a way to avoid people doing this kind of hack.
I'll advise them to use --enable-fat.
Good idea. It will be great to have all x86 assembly code tested on
The libtoolize thing is not for darwin, it is for places where the
failures look like: "no rule to make target add_n.lo" and are because
recent autoconf is incompatible with ancient ltmain.sh.
Ah, they use a much newer autoconf than we have in ~gmp/usr/bin on
Hmm, actually, now I've applied that patch, the CXXFLAGS look correct
on Darwin. I have no idea why but I won't look into it more.
One should be humbly grateful when things mysteriously just work. :-)
Their Darwin build of the main branch (5.0 works) now fails with:
ld: in mpn/.libs/mod_34lsub1.o, in section __DATA,__data reloc 0:
X86_64_RELOC_SUBTRACTOR must have r_extern=1
which is much more like the usual bugs that get reported :-)
Yep, welcome to the Darwin linker... http://gmplib.org/macos.html
So this used to work, but magically broke? I bet it is the jumptab used
in mpn/x86_64/mod_34lsub1.asm that triggers some bug. But it has not
changed for about 5 months...
Perhaps subtracting .rodata and .text symbols isn't entirely kosher (see
L(tab) in the file). If I understand the pedagogical error message
right, they want some of these symbols to be exported.
Perhaps sticking this inside the file will help? :-)
More information about the gmp-devel