Re: [fixincludes, v3] Enable full ISO C99 support for C++ on Solaris 10+

[Rainer Orth <ro at CeBiTec dot Uni-Bielefeld dot DE>]

Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE> writes:

> Jonathan Wakely <jwakely.gcc@gmail.com> writes:
>
>> On 19 May 2010 18:42, Jonathan Wakely wrote:
>>>
>>> The problem is I defined a "::wait" function which clashes with
>>> wait(2). ÂThe fix is to rename that function.
>>
>> Fixed on trunk. If you apply your changes to a branch then backporting
>> the testsuite fixes is pre-approved.
>
> Thanks.  Unfortunately, I've never received any comment on the libstdc++
> side of patch applicability to the 4.4 and 4.5 branches.
>
> Solaris 9/SPARC testing (with gas and TLS enabled to work around PR
> target/44132) revealed that the
> 26_numerics/headers/cmath/c99_classification_macros_c.c test XPASSes
> there, so I'll xfail it only on Solaris 10+.  Otherwise, C++ and
> libstdc++ test results are clean, so it seems the patch has no ill
> effects on earlier (pre-C99) Solaris versions.
>
> Solaris 8/SPARC and Solaris 8/9 x86 testing is still in progress.

I've finally gotten around to finishing this:

* As the testcase XPASSed on Solaris 8 and 9, I'm only XFAILing on
  Solaris 10 and up.

* Comparing libstdc++-v3 config.h before and after the patch, I noticed
  that there is one unexpected change:

 /* Define to 1 if you have the `isnan' function. */
-#define HAVE_ISNAN 1
+/* #undef HAVE_ISNAN */

  This happens because <math.h> includes <iso/math_c99.h> now, which has

/*
 * C99 7.12.3 classification macros
 */
#undef  isnan
#if __GNUC__ >= 4
#define isnan(x)        __builtin_isnan(x)
#else
#define isnan(x)        __extension__( \
                        { __typeof(x) __x_n = (x); \
                        __builtin_isunordered(__x_n, __x_n); })
#endif

  and compiling the configure testcode errors now:

% cat isnan.cpp
#include <math.h>

int
main (void)
{
  isnan (0);

  return 0;
}
% gcc -c -fno-builtin -D_GNU_SOURCE isnan.cpp
isnan.cpp: In function 'int main()':
isnan.cpp:6:3: error: non-floating-point argument in call to function '__builtin_isnan'

  which is correct, of course.  This seems to be a bug in linkage.m4
  (GLIBCXX_CHECK_MATH_DECL_1), which passes 0 to the function under
  test instead of 0.0.

Anyway, I've checked in the patch and will fix this up as a followup
once it has been agreed that passing 0.0 instead is the right way to
handle this.

	Rainer

-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University