Re: [PATCH 2/3] relocate the entry point addess when used

[Pedro Alves <palves at redhat dot com>]

On 01/13/2014 08:46 PM, Tom Tromey wrote:
>>>>>> "Pedro" == Pedro Alves <palves@redhat.com> writes:
> 
> Tom> I think the existing code here is wrong.  It computes the entry point
> Tom> address directly from the BFD, not applying any runtime offsets.
> Tom> However, then objfile_relocate1 passes this address to find_pc_section
> Tom> -- which does use the offsets .  So, it seems to me that the current
> Tom> code can only find the correct address by luck.
> 
> Pedro> It's twisted, but I don't think it's luck.  You can convince yourself
> Pedro> it works by debugging a PIE, and trying a backtrace past main
> Pedro> ("set backtrace past-main" will then trigger the code to stop the
> Pedro> backtrace at the entry point), or doing "info files" (shows the entry).
> 
> Well, what I don't understand is that most addresses in dwarf2read.c are
> offset:
> 
>   baseaddr = ANOFFSET (objfile->section_offsets, SECT_OFF_TEXT (objfile));
> 
> ... however this is not done for the entry point, which comes directly
> from the BFD:
> 
>       objfile->per_bfd->ei.entry_point = bfd_get_start_address (objfile->obfd);
> 
> I suppose there is some other invariant ensuring that the entry point is
> only computed when all the objfile offsets are zero.  This part is not
> obvious to me.

I think it's just that the entry point is only really used for the main
executable, and if that is loaded at some relocation offset, we'll
always go through objfile_relocate (either PIE handling, or RSP
qOffsets handling for embedded systems) after reading in symbols,
while shared libraries are read in with the offsets already handy
upfront.  Even though init_entry_point_info seemingly computes the
entry point for shared libraries, that's really still for the
main executable:
 https://sourceware.org/ml/gdb-patches/2008-05/msg00112.html
and is probably indeed wrong for real shared libraries not
the main executable.

I do think your patch makes things much clearer and sturdier.

> 
> [...]
> Pedro> This is assuming osect ends up NULL after iterating over all.
> Pedro> It's violating the abstraction of the macro.  And, actually,
> Pedro> it's wrong, showing exactly why such assumptions are a bad idea:
> 
> Sorry about that, I'll fix it up locally.

Thanks.


-- 
Pedro Alves