Re: [Qemu-devel] [PATCH v2 02/10] elfload: fix size of registers for N32

[Paolo Bonzini]

Il 04/04/2013 17:46, Peter Maydell ha scritto:
> On 4 April 2013 16:41, Paolo Bonzini <address@hidden> wrote:
>> Il 04/04/2013 17:32, Peter Maydell ha scritto:
>>> This is kind of ugly but it looks like the kernel is kind
>>> of ugly too (ie elf_greg_t as a type is not defined the
>>> same way necessarily for all targets and ABIs). At some
>>> point this type should probably live in a header file
>>> in linux-user/$arch/ but for now I guess it can pass.
>>>
>>> Reviewed-by: Peter Maydell <address@hidden>
>>
>> I'll add abi_reg to abitypes.h and move tswapreg there, so
>> that the signal handling structures could use it too.
> 
> I started off thinking that was the right thing, but then
> I realised that the signal handling structures vary in what
> they use (eg the MIPS ones tend to really use u32 and u64),
> so they aren't necessarily the same type as elf_greg_t).

Right, MIPS o32 uses uint64_t.  So I'll keep target_elf_greg_t as in v2.

> I think the conceptual ideal is that we should have
> abi types which match the kernel's user-facing types.
> This is a bit messed up here because the kernel doesn't
> actually expose elf_greg_t in most archs (and because
> MIPS has different types for "what the kernel thinks
> long is" vs "what userspace thinks long is", which is
> just plain confusing when looking at non-user-facing
> struct definitions).

Paolo