Re: [PATCH] c++: Add target hook for emit_support_tinfos [PR108883]

[Jason Merrill <jason@redhat.com>]

On 2/27/23 18:51, Jakub Jelinek wrote:
> On Mon, Feb 27, 2023 at 06:26:04PM -0500, Jason Merrill wrote:
>>> The following patch instead adds a target hook which allows the backend
>>> to temporarily tweak registered types such that emit_support_tinfos
>>> emits whatever is needed.
>>
>> Why handle these types differently from the DFP handling at the end of
>> emit_support_tinfos?
> 
> One thing is that the fallback_* nodes look like a waste to me,
> the tinfo decls are mangled right away, and the fallback_* nodes need to be
> walked by GC, so I think we could get away even for the decimal tinfos
> to just do:
>        dfloat32_type_node = make_node (REAL_TYPE);
>        emit_support_tinfo_1 (dfloat32_type_node);
>        dfloat32_type_node = NULL_TREE;
> etc. and drop the fallback stuff.

I think you're right.

> If we wanted to do fallback_* even for the _Float*/decltype(0.0bf16)
> nodes, which are at least sometimes mangled in target hooks it would
> make stuff harder because fallback_* is C++ FE private.
> 
> And then there is a question whether we want to emit rtti for
> _Float{16,32,64,128}, _Float{32,64,128}x and decltype(0.0bf16) regardless
> of whether the target supports them at all or not.
> Emitting them always would have an advantage, if say bfloat16_t support
> isn't added for aarch64 for GCC 13 (it is still pending review), we wouldn't
> need to deal with symbol versioning for it in GCC 14 or later.
> On the other side, on some arches some types are very unlikely to be
> supported.  And e.g. _Float128x isn't supported on any arch right now.

A good point.  Incidentally, it seems problematic for embedded users 
that all the fundamental type_infos are emitted in the same .o, making 
it hard to link in only the ones you care about.  And new floating-point 
variants add to that problem.  So perhaps until that is addressed, it's 
better to avoid adding a bunch more on targets that don't support them.

Hmm.

> Though, if we can get rid of the fallback_* stuff and we wanted to emit
> all _Float{16,32,64,128}, _Float{32,64,128}x and decltype(0.0bf16) tinfos
> on all arches (or say for now all but _Float128x), we could do it simply
> by splitting the fundamentals array in emit_support_tinfos into
> one without fallback and one with fallback, put say
>      &dfloat32_type_node, &dfloat64_type_node, &dfloat128_type_node,
>      &bfloat16_type_node, &float16_type_node, &float32_type_node,
>      &float64_type_node, &float128_type_node, &float32x_type_node,
>      &float64x_type_node, &float128x_type_node, 0
> into the latter and simply handle the NULL case with a temporary fallback,
> like:
>    tree fallback = NULL_TREE;
>    for (ix = 0; fundamentals_with_fallback[ix]; ix++)
>      if (*fundamentals_with_fallback[ix])
>        emit_support_tinfo_1 (*fundamentals_with_fallback[ix]);
>      else
>        {
> 	if (fallback == NULL_TREE)
> 	  fallback = make_node (REAL_TYPE);
> 	*fundamentals_with_fallback[ix] = fallback;
> 	emit_support_tinfo_1 (fallback);
> 	*fundamentals_with_fallback[ix] = NULL_TREE;
>        }
> 
> 	Jakub
>