[Bug libstdc++/50724] New: isnan broken by -ffinite-math-only in g++

["ejt at andrew dot cmu.edu" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50724

             Bug #: 50724
           Summary: isnan broken by -ffinite-math-only in g++
    Classification: Unclassified
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libstdc++
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: ejt@andrew.cmu.edu


Given the test code:

    static const float f_nan = 0.f/0.f;
    const int i_nan = *((int*)(&f_nan));
    printf("%g %d %x\n", f_nan, isnan(f_nan), i_nan);

I hope to see:
nan 1 7fc00000

but with the combination of -ffast-math and <cmath>, instead I get:
nan 0 7fc00000

Using just <math.h>, isnan will work. (hence I have filed against libstdc++) 
Not surprisingly, adding -fno-finite-math-only can also fix this.

(regarding cmath, both 'isnan' and 'std::isnan' give the same behavior; BTW I'm
a little surprised a unscoped 'isnan' compiles without 'using namespace
std;'... previous gcc (e.g. Apple 4.2.1) required the 'std'.  I see cmath
undef's the isnan macro from math.h, so where is the global isnan coming from?)

Tested with 4.4.3 and 4.5.2.

Ironically, a test with 4.1.2 shows the reverse situation: <math.h> didn't work
and <cmath> did work.

I realize the docs for -fno-finite-math-only basically warn "all bets are off"
for NaN handling, but as a quality of implementation issue, I'd like to see
consistent handling for floating-point classification functions.  A user may
not care about strict compliance for operations between non-finite values, but
if they are testing for 'bad' values (e.g. isnan), that's probably significant
-- otherwise the user wouldn't have added the classification calls to their
code in the first place.

Alternatively, a warning would be nice, like "isnan always evaluates to false"
when -ffinite-math-only is in effect.