On Thu, 1 Feb 2024 at 17:22, Jonathan Wakely <jwak...@redhat.com> wrote:
>
> On Thu, 1 Feb 2024 at 16:34, Andreas Schwab <sch...@suse.de> wrote:
> >
> > On Feb 01 2024, Jonathan Wakely wrote:
> >
> > > This will result in an ABI change for targets that use 1-byte alignment
> > > for all integral types, e.g. cris-elf.
> >
> > Or 2-byte alignment as on m68k.
>
> Ah yes.
>
> In fact it's a change for everybody, because previously it was 15 bits
> of bit-fields, 1 bit padding, 2 x 16-bit short, 16 bits padding, then
> a char32_t, but now it's 15+17 bits of bit-fields, 2x 16-bit short,
> char32_t. So the shorts moved internally even if the size didn't
> change.
>
> I could make it 15+1 bit-fields, 2x16-bit short, 16 bit-field to avoid
> moving the shorts.
>
> template<typename _CharT>
> struct _Spec
> {
> _Align _M_align : 2;
> _Sign _M_sign : 2;
> unsigned _M_alt : 1;
> unsigned _M_localized : 1;
> unsigned _M_zero_fill : 1;
> _WidthPrec _M_width_kind : 2;
> _WidthPrec _M_prec_kind : 2;
> _Pres_type _M_type : 4;
> unsigned _M_reserved : 1;
> unsigned short _M_width;
> unsigned short _M_prec;
> unsigned _M_reserved2 : 16;
> char32_t _M_fill = ' ';
>
> That would also address the PCC_BITFIELD_TYPE_MATTERS point H-P made.
Oh, and I forgot to mention in the first email that I already changed
the ABI of that type in r14-6991-g37a4c5c23a270c so it's already
different from gcc-13 anyway.
I think I'd prefer to keep the reserved bits together, but a simpler
way to avoid 'unsigned long' making a difference for
PCC_BITFIELD_TYPE_MATTERS targets would be to use no more than 16 bits
but do:
unsigned _M_reserved : 1;
unsigned _M_reserved2 : 16;
