Re: compiler standards (and/or min gcc version) supported withinstalled headers ?

["Joseph S. Myers" <joseph at codesourcery dot com>]

On Wed, 2 Jan 2013, Andreas Jaeger wrote:

> With us moving to require Linux 2.6.16 as oldest release (which is from 2006),
> we should be able to require a newer GCC version as well.

The minimum kernel version helps simplify the code and make a default 
build more efficient by not including compatibility code; removing older 
compiler support from installed headers (where not shared with gnulib) 
doesn't give so much benefit.

I guess that in the cases where the condition is really for "compiler 
supports feature X", we could move such __GNUC__ or __GNUC_PREREQ 
conditions from random (not shared with gnulib) headers into sys/cdefs.h 
and use appropriate feature macros / conditionals in the other headers - 
and where the conditionals are on a GCC version 2.7 or earlier, they can 
certainly become just __GNUC__ conditionals.  And where it's of the form 
"GCC version X optimizes this automatically, use a complicated macro 
definition to optimize it for older versions" we could decide that inline 
optimizations for older GCC aren't that important and the headers should 
be simplified by eliminating them (making older GCC use the non-GCC 
paths).  Much the same may apply for not-fully-conforming variants for 
older GCC (e.g. in bits/huge_val.h for versions not supporting 
__builtin_huge_val).

In some cases conditionals on GCC should simply be removable.  sys/stat.h 
has a condition "#if defined __GNUC__ && __GNUC__ >= 2 && defined 
__USE_EXTERN_INLINES" - but __USE_EXTERN_INLINES is only defined for GCC, 
so no __GNUC__ conditions should be needed here.  (The aim generally 
should I think be to prefer feature conditionals over GCC ones in 
individual headers.)  malloc.h defines __MALLOC_P and __MALLOC_PMT 
conditional on __GNUC__, but they are only used inside malloc.c, so should 
simply be eliminated.  There are probably other such examples as well.  
Similarly to the elimination of old kernel support, it makes sense to 
start by cleaning up the *really old* compiler support (pre-2.7) and GCC 
conditionals in random files that aren't needed at all (or conditions that 
aren't needed if you assume support for long long, for example).

> Btw. GCC 4.1 was released around the same time as Linux 2.6.16 in early 2006.

I think more of when support *ended* for a Linux kernel version as being 
relevant here (July 2009 for 2.6.16, according to Wikipedia), i.e. we 
might decide a certain time after the end of support for 2.6.16 that it is 
no longer relevant to current glibc and move to a newer long-term-support 
2.6 version such as 2.6.27 as the minimum.  (I think we can move rather 
sooner to increase the minimum kernel headers version - currently 2.6.19.1 
- as opposed to the minimum kernel version used at runtime.)

-- 
Joseph S. Myers
joseph@codesourcery.com