Hi Kyrill,
On Thu, Oct 31, 2024 at 02:05:00PM +0000, Kyrylo Tkachov wrote:
> Hi Yury,
>
> > On 31 Oct 2024, at 14:23, Yury Khrustalev <yury.khrusta...@arm.com> wrote:
> >
> > From: Szabolcs Nagy <szabolcs.n...@arm.com>
> >
> > Add new builtins for GCS:
> >
> > void *__builtin_aarch64_gcspr (void)
> > uint64_t __builtin_aarch64_gcspopm (void)
> > void *__builtin_aarch64_gcsss (void *)
> >
> > + ftype = build_function_type_list (ptr_type_node, NULL);
> > + aarch64_builtin_decls[AARCH64_BUILTIN_GCSPR]
> > + = aarch64_general_add_builtin ("__builtin_aarch64_gcspr", ftype,
> > + AARCH64_BUILTIN_GCSPR);
>
> Are these builtins intended to eventually implement the ACLE intrinsics like
> __gcspr ()?
Yes, I will shortly send updated patch series that will include patch for
__gcs* () ACLE
intrinsics.
> I see that in the spec the type signature of that is const void * __gcspr
> (void).
> So I think you want to use const_ptr_type_node rather than just ptr_type_node
> in the ftype.
Yes, the spec [1] indeed suggests `const void *` for GCS pointers in the
corresponding ACLE
intrinsics, however current implementation uses `void *` intentionally as
explained in the
commit message above: GCS pointer can be writable in some cases via the gcsss
operation or
via the GCSSTR instruction.
I will comment in [1] to reflect this.
[1] https://github.com/ARM-software/acle/pull/260
> > + = aarch64_general_add_builtin ("__builtin_aarch64_gcsss", ftype,
> > + AARCH64_BUILTIN_GCSSS);
>
> Same here.
As above.
> Looks ok to me otherwise.
> Thanks,
> Kyrill
>
> > +}
> > +
