Re: C99 compliance for src/port/snprintf.c

Hi,

On 2018-08-25 13:08:18 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2018-08-16 11:41:30 -0400, Tom Lane wrote:
> >> Andres Freund <andres@anarazel.de> writes:
> >>> While I'd personally have no problem kicking gcc 3.4 to the curb, I'm
> >>> still confused what causes this error mode.  Kinda looks like
> >>> out-of-sync headers with gcc or something.
> 
> >> Yeah, this is *absolutely* unsurprising for a non-native gcc installation
> >> on an old platform.
> 
> > Sure, but that still requires the headers to behave differently between
> > C89 and C99 mode, as this worked before. But it turns out there's two
> > different math.h implementation headers, depending on c99 being enabled
> > (math_c99.h being the troublesome).  If I understand correctly the
> > problem is more that the system library headers are *newer* (and assume
> > a sun studio emulating/copying quite a bit of gcc) than the gcc that's
> > being used, and therefore gcc fails.
> 
> I have some more info on this issue, based on having successfully
> updated "gaur" using gcc 3.4.6 (which I picked because it was the last
> of the 3.x release series).  It seems very unlikely that there's much
> difference between 3.4.3 and 3.4.6 as far as external features go.
> What I find in the 3.4.6 documentation is
> 
>  -- Built-in Function: double __builtin_inf (void)
>      Similar to `__builtin_huge_val', except a warning is generated if
>      the target floating-point format does not support infinities.
>      This function is suitable for implementing the ISO C99 macro
>      `INFINITY'.
> 
> Note that the function is called "__builtin_inf", whereas what we see
> protosciurus choking on is "__builtin_infinity".  So I don't think this
> is a version skew issue at all.  I think that the system headers are
> written for the Solaris cc, and its name for the equivalent function is
> __builtin_infinity, whereas what gcc wants is __builtin_inf.  Likewise,
> the failures we see for __builtin_isinf and __builtin_isnan are because
> Solaris cc provides those but gcc does not.
> 
> If we wanted to keep protosciurus going without a compiler update, my
> thought would be to modify gcc's copy of math_c99.h to correct the
> function name underlying INFINITY, and change the definitions of isinf()
> and isnan() back to whatever was being used pre-C99.
> 
> It's possible that newer gcc releases have been tweaked so that they
> make appropriate corrections in this header file automatically, but
> that's not a sure thing.

I've pinged Dave about this, and he said:

On 2018-09-26 17:04:29 -0400, Dave Page wrote:
> Unfortunately, i think that whole machine is basically EOL now. It's
> already skipping OpenSSL for some builds, as being stuck on a very old
> version of the buildfarm client, as some of the modules used in newer
> versions just won't compile or work. I don't have any support contract, or
> access to newer versions of SunStudio, and the guys that used to package
> GCC for Solaris now charge to download their packages.
> 
> I could potentially build my own version of GCC, but I question whether
> it's really worth it, given the other problems.

He's now disabled building master on protosciurus and casteroides.  We
still have damselfly and rover_firefly so I don't feel too bad about
that.  I've pinged their owners to ask whether they could set up a sun
studio (or however that's called in their solaris descendants) version.

Greetings,

Andres Freund