Re: [bug-libunistring] Early Darwin and FAIL: test-float

[Jeffrey Walton]

On Sun, May 5, 2019 at 4:42 AM Bruno Haible <address@hidden> wrote:
>...
> >
> > Unistring 0.9.10 is failing one self test during 'make check'. The
> > other 500+ tests are OK.
> >
> > PASS: test-fdopen
> > PASS: test-fgetc
> > ../build-aux/test-driver: line 107: 85873 Abort trap              "$@"
> > > $log_file 2>&1
> > FAIL: test-float
>
> Thanks for the report.
>
> This platform (Mac OS X 10.5) is quite old [1], and already Mac OS X 10.11
> is end-of-life. This test-float failure indicates a problem only with
> 'long double's, which are hardly used in libunistring (only in the *printf
> routines). Therefore I won't spend time investigating it or providing a
> workaround (most likely: define LDBL_MAX differently).

The PowerMac long double characteristics look very similar to 32-bit
machines except for LDBL_MANT_DIG and sizeof. LDBL_MANT_DIG is twice
as large as expected. Other 32-bit machines use 53, and 64-bit
machines use 64. 106 looks unusual to me.

$ ./test.exe
sizeof(long double): 16
LDBL_MANT_DIG: 106
LDBL_MIN_EXP: -968
LDBL_MAX_EXP: 1024
LDBL_DIG: 31
LDBL_MIN_10_EXP: -291
LDBL_MAX_10_EXP: 308

That's causing this loop to fail:

   volatile long double m = LDBL_MAX;
   for (n = 0; n <= 2 * LDBL_MANT_DIG; n++)
      {
        volatile long double pow2_n = pow2l (n); /* 2^n */
        volatile long double x = m + (m / pow2_n);
        if (x > m)
          ASSERT (x + x == x);     <= Eventually fail here
        else
          ASSERT (!(x + x == x));
      }

What do you recommend?

Jeff