Re: Standardize roff (was: *roff `\~` support)

[Alejandro Colomar]

Hi,

On 8/14/22 16:49, DJ Chase wrote:
> On Sun Aug 14, 2022 at 9:56 AM EDT, Ingo Schwarze wrote:
> > Hi,
> >
> > DJ Chase wrote on Sat, Aug 13, 2022 at 05:27:34PM +0000:
> >
> > > Have we ever considered a de jure *roff standard?
> >
> > No, i think that would be pure madness given the amount of working
> > time available in any of the roff projects.
> >
> > […]
>
> This is very sad to hear.
>
> > > It could also lead to more users & use cases because existing
> > > users could count on systems supporting certain features, so
> > > they could use *roff in more situations, which would lead to
> > > more exposure.
> >
> > You appear to massively overrate the importance end-users
> > typically attribute to standardization.
>
> That’s probably because *I* massively overrate the importance of
> standardization (I mean I literally carry a standards binder with me).
> Still, though, it’s rather annoying that end users — especially
> programmers — don’t value standards as much.

(Official) standardization isn't necessarily a good thing.  With C, it 
was originally good, in the times of ISO C89.  Now, it's doing more 
damage to the language and current implementations than any good (it's 
still doing some good, but a lot of bad).

The best that a standardization process can do is limit itself to 
describe _only_ features already existing in the language, being a kind 
of arbiter that decides on which behavior is best for a given feature, 
so that all implementations follow the best existing one.  Where 
different implementations might have good reasons to do it differently, 
the standard should describe the behavior as implementation-specific. 
And of course, a standard should only standardize features that are 
expected to be good for every implementation, with optional features 
either not being standardized, or being marked optional by the standard 
(like Annex K was; although that one was broken, so it was later removed 
for good).

But that shouldn't be necessary if implementors had some decency and 
didn't implement features so that they are completely incompatible with 
those of other systems.  I.e., if an existing system has 'foo(int a);', 
you don't provide 'foo(int *b);'; you go for 'foo2(int *b);' or 'bar(int 
*b);'.  There's plenty of cases where this has happened, and in some 
cases it might be due to an accident, but in some other cases, it's just 
due to incompetence.  See an example that bit me a month ago: 
<https://github.com/nginx/unit/issues/737>.

And the bad things that standardization can do are several:

By reserving the power to centrally decide the future of a language, 
they take power from implementations, which now can't add some features, 
by fear that they might contradict a future standard.  This is very sad, 
because while the implementations are guided by usefulness and 
worthiness, and try to come up with the best feature for them (and by 
natural selection, implementations are then used or not used, depending 
on their quality), standards have a large part of bureaucracy, and that 
doesn't provide the best features.

A few examples of that are: a %b printf specifier for binary was 
rejected by glibc on the terms of something like "the feature is good, 
and the implementation seems correct, but %b is reserved by the 
standard, so we don't want to possibly conflict with a future standard"; 
luckily, the standard defined that, and the feature was added a few 
years later.  One example that is much more necessary is a way to get 
the size of an array, which currently is impossible in portable C (at 
least not in a way that safely rejects to compile on non-arrays).  I 
also proposed an addition to glibc, and the reasons to reject it were of 
the same kind, and arguing that the standard was discussing about adding 
such a feature; guess what? the standard hasn't added such a feature for 
C23, and we still have no portable way to do it (and the unportable ways 
are more cumbersome than what one would expect).  I hope C3x adds 
_Lengthof(arr), but who knows.

<https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2529.pdf>
<https://stackoverflow.com/questions/37538/how-do-i-determine-the-size-of-my-array-in-c/57537491#57537491>


And then we have another problem of standardization committees: their 
priorities are so broken, that they prefer inventing a completely new 
feature for C, with nothing even remotely resemblant to it within the 
existing language (I'm talking about nullptr and nullptr_t), rather than 
standardizing an existing good feature such as POSIX's NULL  ((void *) 
0).  So now we have 0, NULL, and nullptr for referring to a null pointer 
constant in C.  And none of them is perfect.  0 needs to be casted when 
passed to variadic functions, and has readability issues.  NULL is 
perfect within the POSIX world, but if you go out of POSIX, it's as bad 
as 0.  nullptr, apart from being incomprehensible, it is unsafe; okay 
it's not unsafe by itself, and if it were the only way to refer to a 
null pointer constant, it would be great, but it's not, and even the 
committee recognizes that it will never be.

<https://discourse.llvm.org/t/iso-c3x-proposal-nonnull-qualifier/59269/48>

Many existing projects that use NULL (especially POSIX projects), are 
not going to change their whole codebase to use nullptr.  nullptr_t adds 
some features that add safety against null pointer constants based on 
the type of the constant (by means of _Generic); but that means that one 
can easily bypass those features by using NULL or 0, which means that 
it's not really safe, and it might give a sense of safety that it has 
not.  So, without extending my rant about nullptr much more, it's just a 
feature broken from day 0, invented by the ISO C committee.

Maybe one of the worst problems of the committee (WG14) is that many of 
its members are also members of the WG21, and as such, they may have 
incompatible priorities.

I don't see standardization as good as it may seem at first glance.

And of course following the standard should come with a pinch of salt: 
one should follow the standard, when the standard isn't broken.

But then, the standard isn't better than any other implementation.  So, 
as a programmer, I think programs should target their expected systems, 
and not more (unless it's easy).  If a program is to be run on Linux, 
then target GNU C.  If you can add some partial support for ISO C 
without interfering in your way significantly, then okay, go for it; but 
complete ISO C support is unthinkable; a program conforming to ISO C is 
useless, or unnecessarily complex, or even unsafe.  I implement things 
thinking on my system first, then if it's easy, I can support other FOSS 
Unix systems, if it's easy, but only if it is.  Commercial systems are 
automatically out of support.  I'm not spending a single minute of my 
time to be nice to those systems when their not nice to me.

I think it's better to let natural selection to work out its way.  If a 
feature is good, other implementations will pick it, and maybe even 
improve it.  If a feature is not good (or it's not needed by other 
systems), it will not be portable.

Cheers,

Alex


--
Alejandro Colomar
<http://www.alejandro-colomar.es/>

OpenPGP_signature
Description: OpenPGP digital signature