AC_C_RESTRICT

[Jeff Squyres]

This mail could probably be considered a followup to this thread  
introduced by Rolf Vandevaart from Sun back in Jan 2009:

  http://lists.gnu.org/archive/html/bug-autoconf/2009-01/msg00056.html

Short version:
--------------

Is there any chance that an AC_CXX_RESTRICT macro will ever come into  
existence, and/or the output from AC_C_RESTRICT will become safe for  
both C/C++?  Specifically, Open MPI uses the "restrict" keyword in a  
header file that is included in both C and C++ source code; we now  
have a problem if the support for "restrict" is different between the  
C and C++ compiler.

More details:
-------------

As Rolf noted in the earlier thread, the Sun compiler suite has  
different support for "restrict" in its C and C++ compiler (i.e., the  
C compiler supports some "restrict"-like keywords, but the C++  
compiler does not).  I have been told that "restrict" is not defined  
in the C++ language, so I can understand the Sun compiler making this  
interpretation.  We (the Open MPI project) have recently found another  
compiler that has different support for restrict in its C and C++  
compilers: the PGI Linux compiler suite.

It seems that AC_C_RESTRICT can sometimes find "restrict" for the PGI  
C compiler, but it can sometimes find "__restrict".  Both of these  
values work for the PGI C compiler, but only "__restrict" works for  
the PGI C++ compiler.  The reasons for AC's difference in answers  
appear to do with the PGI C compiler installation and the version of  
the GCC/glibc installed, distro/version, ...etc.  I didn't delve too  
deeply to figure out why this happened.  See the thread starting here  
for more details:

    http://www.open-mpi.org/community/lists/users/2009/03/8402.php

This causes a problem for us because Open MPI has a C header file that  
uses "restrict" and relies on AC_C_RESTRICT to do the Right Thing.   
Across all of our C source code, this works great.  When this header  
file is included in C++ source code, bad things *may* happen -- by  
what Rolf noted, and now if AC_C_RESTRICT chose "restrict" instead of  
"__restrict".

Rolf solved the problem by putting a hackaround in for the Sun  
compiler.  But now that there's a second compiler suite that falls  
into the same category as the Sun compiler, I'm wondering if Autoconf  
shouldn't try to address this issue...?  E.g.:

- should there be an AC_CXX_RESTRICT test?  E.g., perhaps the output  
of AC_C_RESTRICT and AC_CXX_RESTRICT tests would be protected with  
appropriate #if's that determine if you're compiling in C or C++.
- Or should there be a new AC_RESTRICT test that makes a solution for  
both the C and C++ compilers?
- Or ...?

(please reply all so that the others CC'ed on this post also see the  
followups; thanks!)

--
Jeff Squyres
Cisco Systems