Re: RFC: C++ PATCH to adjust empty class parameter passing ABI

[Ramana Radhakrishnan <ramana dot gcc at googlemail dot com>]

On Thu, Apr 14, 2016 at 11:36 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Thu, Apr 14, 2016 at 11:25:07AM +0100, Ramana Radhakrishnan wrote:
>> > I see the test failing on aarch64-none-linux-gnu (native)
>> > with no output, just:
>> > spawn [open ...]
>> > FAIL: g++.dg/abi/empty13.C  -std=gnu++98 execution test
>> >
>> > And I see it passing on aarch64-none-elf.
>>


I don't understand why this passes on aarch64-none-elf btw.

>> IIRC on AArch64 we document that empty classes and structs take up a
>> slot as per the PCS .
>>
>> http://infocenter.arm.com/help/topic/com.arm.doc.ihi0059b/IHI0059B_cppabi64.pdf
>>
>> See section 3.1
>>
>> For the purposes of parameter passing in
>> [
>> AAPCS64
>> ], a parameter whose type is an empty class shall be treated as if its
>> type were an aggregate with a
>> single
>> member of type unsigned byte.
>
> One thing is passing such arguments in registers, another is how they are
> passed on the stack (e.g. if there are several other arguments that are
> passed in registers).


It would just be treated like a structure with an unsigned byte and
treated the same way for parameter passing on the stack - i.e. IIRC
take a slot of 8 bytes.

> Also, note that the (new) definition of C++ empty class I believe includes
> even empty classes that are normally very large, i.e. don't pretend they
> are just single byte, but can be e.g. 1024 bytes or longer.
> Do you want to still pass them in registers and/or on the stack as if
> it is 1 byte, 0 bytes or more bytes?

I don't know off-hand.

Is this a valid example for what you have in mind ?

struct baz
{
  char a[1024];
};
struct foo : baz
{
};

int bar (struct foo b, int x)

if so, that would get treated as though it is just a 1 byte quantity
for C++. as per current behaviour by both gcc and llvm for aarch64.

> How do you pass empty C structures?

I don't think we do anything special for empty C structures which IIRC
means they don't take a slot.


>
> E.g. looking at sparc-solaris2.12 cross, C for empty structures passes
> both in registers and on the stack as if there is one argument (and C++ did
> too for the empty classes that pretended to have size 1), so appart from the
> ICE in emit_move_insn there is no ABI change, only perhaps for the larger
> structs.
>
> Guess we need to figure out what do the various targets do for C empty
> structures and decide what we do want to do with C++ empty classes.
> Perhaps the C++ FE should just set some flag on the empty class parameters,
> and let the backends decide what to do with those.



Ramana

>
>         Jakub