Re: [musl] [PATCH] vfprintf: support C2x %b and %B conversion specifiers

[Rich Felker <dalias@libc.org>]

On Mon, Sep 12, 2022 at 04:42:51PM +0200, Jₑₙₛ Gustedt wrote:
> Rich,
> 
> on Mon, 12 Sep 2022 09:59:04 -0400 you (Rich Felker <dalias@libc.org>)
> wrote:
> 
> > I don't see anything else immediately wrong. No rush to submit an
> > updated patch. Let's first figure out the timeline for C2x features,
> > and I'll try to figure out what's going on with the weird buf[] size
> > expression and clean it up separately so that your patch is just
> > adding the feature and not also fighting with historical cruft.
> 
> The specification of C23 has been finished in July, what follows now
> in WG14 is merely editorial on the standard text itself and should not
> have much impact on how things should be implemented. The current
> draft is here
> 
>       https://open-std.org/jtc1/sc22/wg14/www/docs/n3047.pdf
> 
> and any changes from there should not have impact on implementations,
> unless a majority of the national bodies imposes withdrawal of
> specific new features. The timeline is publication in end of 2023, so
> this leaves basically one year to have things ready when the new
> standard is published
> 
> Already for `printf` and friends there are several additions, not only
> the `b` formats but, e.g, also formats for the fixed-width types and
> for the new bit-precise types.
> 
> Otherwise, there are changes to the C library all over the place, but
> in general these are small and should be doable with limited work
> load, in particular for implementations that also implement
> POSIX. Those headers that have changes also are now supposed to have a
> feature macro that describes the version for which a header works,
> such that support for C23 can be added stepwise to C library
> implementations.
> 
> There is one big addition, though, for the decimal floating point
> types, but support here is optional. I don't know if there is already
> an open stand-alone implementation of that stuff, but I think that
> that should be the goal, instead of repeating the effort for every C
> library implementation that is out there.

What do these entail? It looks like there's a requirement for printf
to support them, so I don't see how you'd do that as a separate
library. It looks like __STDC_IEC_60559_DFP__ is optional though, so
maybe we could just decline to define it and leave the support
sporadic at the level the compiler supports, as an extension rather
than part of the standard-specified functionality?

Implementing any of this seems rather blocking since it depends on new
types at the compiler level for the compiler compiling libc. The only
way out from that is using asm thunks. For just the printf case, we
could do it with an arch-provided asm fragment that grabs the decimal
float from the va_list, which would be fairly low-cost to add/maintain
and allow the rest of the decimal float library code to be relegated
to a separate standalone library like you suggested. The only part of
printf that looks hard to implement for decimal floats is %a, and from
a quick glance it looks like the reverse direction in scanf/strto*
doesn't support hex for decimal floats?

Rich