Re: GDB broken on MIPS targets with unmarked binaries

[Daniel Jacobowitz <drow at mvista dot com>]

On Wed, Jun 05, 2002 at 04:58:34PM -0400, Daniel Jacobowitz wrote:
> On Wed, Jun 05, 2002 at 04:44:21PM -0400, Andrew Cagney wrote:
> > >There's an interesting crash on startup in a mips-elf GDB nowadays... the
> > >code reads:
> > >
> > >    default:
> > >      tdep->mips_abi_string = "default";
> > >      tdep->mips_default_saved_regsize = MIPS_REGSIZE;
> > >      tdep->mips_default_stack_argsize = MIPS_REGSIZE;
> > >      tdep->mips_fp_register_double = (REGISTER_VIRTUAL_SIZE (FP0_REGNUM) 
> > >      == 8);
> > >      tdep->mips_last_arg_regnum = A0_REGNUM + 8 - 1;
> > >      tdep->mips_last_fp_arg_regnum = FPA0_REGNUM + 8 - 1;
> > >      tdep->mips_regs_have_home_p = 1;
> > >      tdep->gdb_target_is_mips64 = 0;
> > >      tdep->default_mask_address_p = 0;
> > >      set_gdbarch_long_bit (gdbarch, 32);
> > >      set_gdbarch_ptr_bit (gdbarch, 32);
> > >      set_gdbarch_long_long_bit (gdbarch, 64);
> > >      break;
> > >
> > >Something brought in recently, probably by the OS ABI changes if I had to
> > >guess, is causing the check of REGISTER_VIRTUAL_SIZE to fail.  I've no idea
> > >why.  Compiling a mips-elf (not mips-linux, which sets MIPS_DEFAULT_ABI) 
> > >and
> > >running gdb on an unmarked ELF binary reproduces this.
> > 
> > I'd suspect this:
> > 
> > 2002-05-08  Andrew Cagney  <ac131313@redhat.com>
> > 
> >         * gdbarch.sh (init_gdbarch_swap): Do not clear the swap section.
> >         (clear_gdbarch_swap): New function.
> >         (initialize_non_multiarch): Call.
> >         (gdbarch_update_p): Before calling init(), swap out and clear the
> >         existing architecture.
> >         * gdbarch.c: Regenerate.
> > 
> > The comment from arm-tdep.c should help explain why the above code is 
> > broken:
> > 
> >   /* We can't use NUM_REGS nor NUM_PSEUDO_REGS here, since that still
> >      references the old architecture vector, not the one we are
> >      building here.  */
> >   prologue_cache.saved_regs = (CORE_ADDR *)
> >     xcalloc (1, (sizeof (CORE_ADDR)
> >                  * (gdbarch_num_regs (gdbarch)
> >                     + gdbarch_num_pseudo_regs (gdbarch))));
> > 
> > As far as I can tell though, REGISTER_VIRTUAL_SIZE expands to 
> > REGISTER_VIRTUAL_TYPE which is still a macro?
> 
> The expansion appears to be:
> tdep->mips_fp_register_double = (((((38) >= 38 && (38) < 38 +32) ?
> 	builtin_type_float : ((38) == 32 ) ?
>         builtin_type_uint32 : ((38) >= 70 && (38) <= 89) ?
> 	builtin_type_uint32 :
>         builtin_type_int))->main_type->length == 8);
> 
> I can't see any reason for it to crash; I haven't reproduced this
> locally yet, although I'm working on it.  The above expansion came from
> the second person to mention this to me.

I believe you're right about the patch at fault, though I don't know
what to do about it.  This makes builtin_type_int == NULL during
evaluation of that expression, along with all the others.

-- 
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer