Re: AC_C_LONG_DOUBLE is wrong on IRIX 5.3

[Paul Eggert]

> From: Oliver Kiddle <address@hidden>
> Date: Thu, 15 Nov 2001 18:18:49 +0000
> 
> Patch is below.

Thanks.  I have some further remarks:

> Is the line break in the AC_DEFINE near the end ok?

I'd avoid it by shortening the comment; please see below.

> > int a[LDBL_MAX <= DBL_MAX ? -1 : 1];
> 
> Ah, ok, thanks. Again, you should take a look that the patch is what
> you had in mind. I assume that you didn't suggest the simpler
>     int a = LDBL_MAX <= DBL_MAX ? -1 : 1
> because it has some problem such as compilers doing it at runtime after
> converting LDBL_MAX to a double or similar.

No, the point was to have a compile-time error, not a run-time error.

> I changed the -1 to 0

Please undo that change, as that will cause the program to succeed with
some compilers, as some compilers allow zero-size arrays.

> and used sizeof(a) == 0

This shouldn't be needed; please see below.

> -If the C compiler supports the @code{long double} type, define
> +If the C compiler supports the @code{long double} type and this type
> +provides greater precision than the @code{double} type, define

Please reword to:

If the C compiler supports a @code{long double} type with
greater range or precision than the @code{double} type, define

> -[AC_CACHE_CHECK(for long double, ac_cv_c_long_double,
> +[AC_CACHE_CHECK(for useful long double, ac_cv_c_long_double,

Please change 'useful long double' to 'long double with greater range
or precision than double'.

>  AC_TRY_RUN(

It should use AC_TRY_COMPILE, not AC_TRY_RUN.  I realize that the old
version used AC_TRY_RUN, but let's fix that bug while we're at it.
AC_TRY_RUN makes cross-compiling problematic.

> +#if defined (LDBL_MAX) && defined (DBL_MAX)

This #if should be removed.  If those symbols are not present, the
code will fail to compile, and this should do the right thing.

> +  int a[LDBL_MAX <= DBL_MAX ? 0 : 1];

Two things.  First (as described before) the 0 should be -1.  Second,
as you noted, we should also use 'long double' if it merely has
greater greater precision.  So please reword as follows:

int a[DBL_MAX < LDBL_MAX || LDBL_EPSILON < DBL_EPSILON ? 1 : -1];

>    /* On Ultrix 4.3 cc, long double is 4 and double is 8.  */

This comment isn't needed any more.

> +  exit ((sizeof (a) == 0) ||
> +      (sizeof (long double) <= sizeof (double)));

There's no need to invoke exit, as it's a compile-time test now.

> +            [Define if the `long double' type works and provides greater
> +          precision than `double'.])

Reword to 'Define if long double has greater range or precision than double.'
Then it'll fit on a line.

Can you please test and resubmit a patch along those lines?
(In case you're worried, this process should converge eventually.  :-)