MIME-Version: 1.0
In-Reply-To: <504CD18B.1060207@iki.fi>
References: <504CD18B DOT 1060207 AT iki DOT fi>
Date: Sun, 9 Sep 2012 16:19:51 -0500
Message-ID: <CAA-ihx-HZWik70LhPd4d16wN4LifdO35_NkcPjziSzOfwRr9OQ@mail.gmail.com>
Subject: Re: Invalid tests for __STDC_VERSION__ in DJGPP header files
From: Rugxulo <rugxulo AT gmail DOT com>
To: djgpp-workers AT delorie DOT com
Content-Type: text/plain; charset=ISO-8859-1
Reply-To: djgpp-workers AT delorie DOT com
Errors-To: nobody AT delorie DOT com
Precedence: bulk

Hi,

On Sun, Sep 9, 2012 at 12:27 PM, Andris Pavenis <andris DOT pavenis AT iki DOT fi> wrote:
> -_STDC_VERSION__ is not defined in many cases any recent GCC version. It is
> only seems to be defined in several cases depending on the command line
> parameter -std and for C language only

Most C compilers seem to only define __STDC__ when "ANSI"
compatibility is turned on. GCC however must consider itself "close
enough" by default ("gnu89") that it defines it anyways. However,
anything beyond that is (usually) turned off.

Apparently GCC (from gcc.info) claims "-std=c89" and "-std=c90" and
"-ansi" are all the same thing. (Though I swear I halfway read
somewhere recently that __STDC_VERSION__ was first defined in ISO
1990. Dunno!)

By default, "gnu89" disables trigraphs (dunno what else), so it's
close but not quite strict, hence the obvious needs for other switches
(-ansi -pedantic).

> Some examples:
>
> [andris AT ap ~]$ i586-pc-msdosdjgpp-gcc -dD -E -std=c99 -x c /dev/null | grep
> STDC
> [andris AT ap ~]$ i586-pc-msdosdjgpp-gcc -dD -E -std=c90 -x c /dev/null | grep
> STDC
> [andris AT ap ~]$ i586-pc-msdosdjgpp-gcc -dD -E -std=c89 -x c /dev/null | grep
> STDC
> [andris AT ap ~]$ i586-pc-msdosdjgpp-gcc -dD -E -std=c9x -x c /dev/null | grep
> STDC
> [andris AT ap ~]$ i586-pc-msdosdjgpp-gcc -dD -E -std=c11 -x c /dev/null | grep
> STDC

You forgot "-std=iso9899:199409 (aka, C94 or C95 or AMD1 or NA1),
which first added all that wchar crud.  :-P

> As the result tests used in header files like
>
> #if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) \
>   || !defined(__STRICT_ANSI__)
>
> are invalid.

Dunno, it's hard to guess what's going on there. "#if BLAH" works only
if BLAH is non-zero, right? (Allegedly some old compilers defined
__STDC__ to zero to show lack of full compliance, ugh.) So a simple
"#define BLAH" should define it to equal 1, right?

> I see several possibilities
> 1) assume that the version is new enough when __STDC_VERSION__ is not
> defined
> #if !defined(__STDC_VERSION__) || __STDC_VERSION__ >= 199901L) \
>   || !defined(__STRICT_ANSI__)
> 2) remove tests for version 199001 or later at all (one may need tests for
> newer
>    versions though)

I assume you mean "199901L" (C99) here.

Dunno about removing anything, but if GCC was going to switch to
"gnu99" by default, they probably would've done it by now (as Clang
has from the beginning).

> 3) require __STDC_VERSION__ to be defined for very old GCC versions (as far
> as
>    I tested gcc-3.4.4 is considered not very old)

Old DJGPP 2.95.3 seems to say C94 by default (as does OpenWatcom,
IIRC), and that's with DJGPP 2.03p2 and no fancy switches. Even using
"-std=c9x" [sic] doesn't change that.

Seems 3.0.4 corrected it as default it doesn't define
__STDC_VERSION__. And it does indeed define it for "-std=c99" [sic].
Though if you use "-std=iso9899:199409", it will define it also (as
appropriate).

> Use of the broken test for __STDC_VERSION__ prevents libstdc++-v3 (GCC-4.8.0
> development versions from some last weeks) building OK due to actual
> contents of stdint.h being compiled out. Variant 1 is verified to fix the
> build (only tested cross-native build from Linux)

I don't know what versions are recommended or used or tested by
default. DJ and/or Eli will have to chime in. I'm pretty sure that
2.95.3 is way too old, but I would blindly guess that 3.4.4 may? be
acceptable??