Hello,

Am Dienstag, den 02.12.2014, 17:47 -0500 schrieb Rich Felker:
> On Tue, Dec 02, 2014 at 11:00:12PM +0100, Jens Gustedt wrote:
> Then what about something like
> 
> #if __clang__ // && __clang_major__*100+__clang_minor__ <= xxxxx
> // use clang compound literal hack
> #elif __GNUC__*100+__GNUC_MINOR__ >= 407
> // use __builtin_complex
> #else
> // use _Complex_I with mult and add
> #endif
> 
> or alternatively:
> 
> #if __clang__ // && __clang_major__*100+__clang_minor__ <= xxxxx
> // use clang compound literal hack
> #else
> // use __builtin_complex
> #endif
> 
> I somewhat prefer the latter since the _Complex_I form is never really
> correct (INF/NAN issues) and it gives better fallback behavior
> assuming other compilers will implement __builtin_complex but might
> not identify themselves as "GNU C >= 4.7". It also produces an error
> on compilers that can't give the right behavior for INF/NAN rather
> than silently miscompiling the code.

I agree with your analysis, we should avoid using _Complex_I, here.

For the implementation I don't agree completely. Version number games
are not the right thing for clang, they don't support this, in a sense
in the same way that you don't want a macro that identifies musl.

They have their feature/builtin test framework, we should use
that. This then would kick in whenever they decide to implement that
builtin.

But still, I'd like to have a version with _Imaginary_I, see below.

> > > > > _Imaginary_I
> > > > > is not supported yet at all and would require further changes to this
> > > > > header to add anyway (this header is responsible for defining it), so
> > > > > conditioning on its definition is not meaningful.
> > > > 
> > > > I am not sure of that. It could also come directly from the compiler
> > > > just as it defines some __SOMETHING__ macros before any include
> > > 
> > > I don't think so. Consider:
> > > 
> > > #undef _Imaginary_I
> > > #include <complex.h>
> > > 
> > > Perhaps this is non-conforming for technical reasons,
> > 
> > it is non-conforming
> 
> Because of the _-prefixed namespace?

Yes, an implementation simply *mustn't* touch any identifier that
starts with underscore and capital letter. Period.

> Per my understanding, it's
> conforming to #undef macros provided by the standard headers in many
> cases.

> Eliminating a function-like macro to get at the underlying
> function is one such usage mentioned in the standard, but I think
> there are others too.

I don't know of many, and these are explicitly named. And all of this
is *not* the same thing as the case above. This concerns e.g "complex"
and "I" which are otherwise not reserved, and not e.g "_Complex_I"
that falls under the same rule as _Imaginary_I.

> > But if you want to avoid that branch, I understand. _Imaginary_I would
> > be *the* way to solve all of this easily, sigh.
> 
> Yes, I really wish GCC would support it, especially since Annex G
> _requires_ _Imaginary_I, and glibc falsely advertises Annex G support
> via the __STDC_IEC_559_COMPLEX__ macro despite not having
> _Imaginary_I. But the glibc folks seem to think this is a non-issue
> and I haven't seen any interest in fixing it on the GCC side.

argh, I wasn't aware of that, what a mess. yes _Imaginary and all that
comes with it is really an important feature of that Annex, e.g
functions returning _Imaginary values when known to have such values.

So I keep my claim for a first case version with _Imaginary_I up :)
we should encourage compiler implementors to provide this feature.

I looked a bit around and I found that
 - Keil supports imaginary
 - for Cray compilers 1.0i is pure imaginary and not complex, but they
   don't explain exactly what happens if you use this in arithmetic

Jens

-- 
:: INRIA Nancy Grand Est ::: AlGorille ::: ICube/ICPS :::
:: ::::::::::::::: office Strasbourg : +33 368854536   ::
:: :::::::::::::::::::::: gsm France : +33 651400183   ::
:: ::::::::::::::: gsm international : +49 15737185122 ::
:: http://icube-icps.unistra.fr/index.php/Jens_Gustedt ::