Re: AArch64 bfloat16 mangling

[Richard Sandiford <richard.sandiford@arm.com>]

Sorry for the slow response.

Jakub Jelinek <jakub@redhat.com> writes:
> Hi!
>
> On Mon, Jan 30, 2023 at 11:07:23PM +0000, Richard Sandiford wrote:
>> Jakub Jelinek <jakub@redhat.com> writes:
>> > https://gcc.gnu.org/pipermail/gcc-patches/2022-November/605965.html
>> >   - ABI - aarch64: Add bfloat16_t support for aarch64 (enabling it in GCC 14
>> >     will be harder)
>> 
>> Sorry for the delay on this.  There's still an ongoing debate about
>> whether to keep the current AArch64 mangling or switch to the new one.
>
> If it helps, I'll try to repeat the options I see:
> 1) don't do anything right now; problem is if it is done later (GCC 14+),
>    libstdc++ would need to conditionalize the std::bfloat16_t RTTI symbols,
>    have them in one symbol version for x86 and in another for aarch64
> 2) similarly to x86 __bf16 would be the underlying type for std::bfloat16_t
>    where the latter needs to act as usable extended floating point type with
>    all arithmetics, mangling is DF16b which is how std::bfloat16_t should
>    mangle according to the Itanitum ABI pull request; decltype (0.0bf16) is
>    __bf16; disadvantage is that existing code using __bf16 in argument
>    passing and templates changes mangling
> 3) keep __bf16 as is with its u6__bf16 mangling and use for std::bfloat16_t
>    a distinct type (the latter would be the bfloat16_type_node);
>    decltype (0.0bf16) would be that new type which would mangle DF16b and
>    would allow arithmetics/casts etc.  How exactly would the new type be
>    named is up to you (__bfloat16_t, __bfloat16, __std_bfloat16_t,
>    whatever else); in theory it could be created without a user accessible
>    name as well; libstdc++ only uses decltype (0.0bf16) to get at it
> 4) like 3), including keeping the mangling of __bf16 as u6__bf16, but
>    make also __bf16 a usable arithmetic type, not just a storage only type;
>    for C++ FE it would be simply another non-standard type like say
>    __float128 is on x86
> 5) like 2), but make the mangling of __bf16 depend on flag_abi_version;
>    flag_abi_version >= 18 (aka GCC 13+ ABI) mangles it as DF16b,
>    flag_abi_version < 18 mangles it as u6__bf16; the default for
>    -fabi-compat-version= is I think GCC 8 ABI compatibility, so GCC normally
>    emits mangling aliases, so say void foo (std::bfloat16_t) {} would
>    mangle as _Z3fooDF16b and for a few years there would be
>    an alias _Z3foou6__bf16 to it
>
> Of course, it is possible I've missed some options.
>
> 	Jakub

We decided to keep the current mangling of __bf16 and use it for
std::bfloat16_t too.  __bf16 will become a non-standard arithmetic type.
This will be an explicit diversion from the Itanium ABI.

I think that's equivalent to your (2) without the part about following
the Itanium ABI.

Thanks,
Richard