Re: Status - compiling on hppa / ld sigsegv's - glibc 2.2.93
On Wed, Sep 11, 2002 at 10:06:15AM -0700, Randolph Chung wrote:
> > (gdb) set args --help
> > (gdb) r
> > Starting program: /mnt/flaire/src/glibc-upstream-cvs/libc-build/elf/ld.so.1 --help
> >
> > Program received signal SIGSEGV, Segmentation fault.
> > 0x410026a4 in _dl_start () at do-rel.h:73
> > 73 r = r + MIN (nrelative, relsize / sizeof (ElfW(Rel)));
> > (gdb) bt
> > #0 0x410026a4 in _dl_start () at do-rel.h:73
> > #1 0x4100209c in _start () at rtld.c:154
> > (gdb) disassemble
> > ...
> > 0x410026a4 <_dl_start+624>: fldd 0(sr0,r1),fr22
>
> Carlos,
>
> Some bits of info that might be useful for sorting this out...
>
> The relsize / sizeof(ElfW(Rel)) is translated by gcc into a series of
> floating point operations (as you disassembled above). Through some test
> programs, it seems like gcc's strategy for handling a code segment like
> that is:
>
> f.e. given sizeof(foo) == 12, and we want to compute bar / sizeof(foo)
> put 1/12 in the data section, load that from memory into a floating
> point reg, load bar into another floating point reg, and multiple them.
>
> r1 at this point in the code should be pointing to the constant in
> memory. since it is 0, we get a segfault (indeed if we replace the
> sizeof() with a constant, like 8, ld.so works).
>
> I'm guessing this means that our lp (r19) is not setup correctly when
> _dl_start is called, but i haven't dug enough into the code to figure
> out the details yet.
>
In _dl_start I'm not sure if r19 is expected to be correct or not - I
suspect that it may not be, and they didn't take into account the idea
that an integer divide would need to load from a constant pool... why
on earth is GCC generating an FP divide anyway?
--
Daniel Jacobowitz
MontaVista Software Debian GNU/Linux Developer
Reply to: