Bug#274738: on mips{,el} atexit.o should be compiled with -mxgot
reassign 274738 binutils
severity 274738 important
thanks
At Tue, 05 Oct 2004 10:04:20 +0900,
GOTO Masanori wrote:
> At Mon, 4 Oct 2004 00:25:00 +0200,
> Thiemo Seufer wrote:
> > > current mainline libgcj fails to build on mips{,el}:
> > >
> > > /home/doko/gcc/gcc-snapshot-20041003/build/gcc/xgcc -shared-libgcc -B/home/doko/gcc/gcc-snapshot-20041003/build/gcc/ -nostdinc++ -L/home/doko/gcc/gcc-snapshot-20041003/build/mipsel-linux-gnu/libstdc++-v3/src -L/home/doko/gcc/gcc-snapshot-20041003/build/mipsel-linux-gnu/libstdc++-v3/src/.libs -B/usr/lib/gcc-snapshot/mipsel-linux-gnu/bin/ -B/usr/lib/gcc-snapshot/mipsel-linux-gnu/lib/ -isystem /usr/lib/gcc-snapshot/mipsel-linux-gnu/include -isystem /usr/lib/gcc-snapshot/mipsel-linux-gnu/sys-include -shared -nostdlib /usr/lib/crti.o /home/doko/gcc/gcc-snapshot-20041003/build/gcc/crtbeginS.o .libs/libgcj.la-2.o -Wl,--whole-archive ../libffi/.libs/libffi_convenience.a ../boehm-gc/.libs/libgcjgc_convenience.a ./libltdl/.libs/libltdlc.a -Wl,--no-whole-archive -L/home/doko/gcc/gcc-snapshot-20041003/build/mipsel-linux-gnu/libstdc++-v3/src -L/home/doko/gcc/gcc-snapshot-20041003/build/mipsel-linux-gnu/libstdc++-v3/src/.libs -L/home/doko/gcc/gcc-snapshot-20041003/build/mipsel-linux
-g
> nu/libjava ../libffi/.libs/libffi_convenience.a ../boehm-gc/.libs/libgcjgc_convenience.a -lpthread ./libltdl/.libs/libltdlc.a -ldl -lz -L/home/doko/gcc/gcc-snapshot-20041003/build/gcc -lgcc_s -lc -lgcc_s /home/doko/gcc/gcc-snapshot-20041003/build/gcc/crtendS.o /usr/lib/crtn.o -Wl,-soname -Wl,libgcj.so.6 -o .libs/libgcj.so.6.0.0
> > > /usr/lib/libc_nonshared.a(atexit.oS)(.text+0x38): In function `atexit':
> > > : relocation truncated to fit: R_MIPS_CALL16 __cxa_atexit@@GLIBC_2.2
> > > collect2: ld returned 1 exit status
> > > make[2]: *** [libgcj.la] Error 1
> > >
> > > can be fixed by compiling glibc's atexit with -mxgot
> >
> > This needs _two_ sets of those objects then, because a link employing
> > multigot will silently break over xgot objects.
>
> Does this problem come from glibc's compilation problem?
>
> > > (at least that's the way mozilla succeeded to build).
> >
> > Actually, no, mozilla uses the standard gcc/libc objects. It believe it
> > happens to work because those objects are linked in earlier, so their
> > GOT entry offset doesn't exceed the limit.
>
> Do you know what's the actual problem?
This is my memorandum, and the summary. If you find mistake, please
follow up it.
mips/mipsel has long-standing problem about GOT handling. The
GOT size is limited to 16KB and it's sometimes small. This
report is affected by this problem.
There're two ways to fix this problem; multiGOT and XGOT.
XGOT is originally developed in IRIX. It uses 32bit GOT, so
we don't need to worry about overflow of the table size.
Matthias suggested to use -mxgot. However, unfortunatelly, we
may lose binary compatibility (said by Thiemo). Thiemo said
it's nicer than multiGOT. OTOH, multiGOT is to have multiple
GOT - Daniel suggested it. He said binutils should be fixed
if it does not work. Note that multiGOT is firstly supported
by Alexandre Oliva in binutils 2.15 development.
In debian-mips, there were discussions about gcj compilation problem.
David Daney replied to Matthias that the current binutils + glibc
should work without any GOT problems:
http://lists.debian.org/debian-mips/2004/10/msg00008.html
I don't know which technically correct is. However, it's
long-standing problem, and not glibc problem. This bug should
actually be considered, but it's not sarge's target. I downgrade it
to important and reassign it to binutils. If you have another
opinion, let us know.
Regards,
-- gotom
Reply to: