On Tue, 12 Jan 2016, Luis Machado wrote:
I also wonder whether the bfd arch detection couldn't be always
compiled in, at least for elf. Why does bfd fail to detect that this
is an bfd_arch_i386 file in the first place?
The mapping between `e_machine' and `bfd_architecture' is only provided
by individual BFD ELF target backends, via the ELF_MACHINE_CODE and
ELF_ARCH macros.
It seems bfd also falls back to the default, which is mips in this case.
p bfd_default_vector[0]
$3 = (const bfd_target *) 0x9beac0 <mips_elf32_trad_be_vec>
Regardless, I'd expect a suitable generic ELF BFD target to be selected,
which is what AFAICT `bfd_check_format' does. It is called by our
`core_open' function and has a `core_file_p' handler, which makes suitable
checks including `e_machine' in particular, except for generic ELF BFD
targets, which are special-cased (and always come last). So in the
absence of specific ELF target support in BFD I'd expect a compatible
generic ELF target to be chosen rather than the default BFD target, which
might be incompatible.
Sounds like we have a couple issues. The mips backend not handling weird
abi/isa combinations and GDB not preventing clearly incompatible core files
from proceeding further into processing in the target's backend?
I have given it some thought and came to a conclusion that we should at
least try being consistent. Which means I think we should not try to
handle files within the MIPS backend which would not be passed in the
first place in an `--enable-targets=all' configuration. Rather than
checking `e_machine' explicitly I'd be leaning towards using BFD to detect
such a situation though, perhaps by using a condition like
if (info.abfd != NULL
&& bfd_get_flavour (info.abfd) == bfd_target_elf_flavour
&& bfd_get_arch (info.abfd) != bfd_arch_mips)
return NULL;
(maybe with an additional error message) though ultimately I think it
would make sense to define different BFD architecture codes for file
formats which by definition carry no architecture information and for ones
that do but are not supported. Then for the formers we could continue
selecting the target using the current algorithm and for the latters we'd
just reject them as incompatible with the given backend -- all somewhere
in generic code so that individual target backends do not have to repeat
it all.