Re: PATCH: New test for breakage from real.c rewrite on i386-unknown-freebsd

[Loren James Rittle <rittle at latour dot rsch dot comm dot mot dot com>]

> These values are only correct for 64-bit double, not 80-bit long double.
> The compiler is completely unprepared for you to do

> #define __glibcpp_long_double_bits __glibcpp_double_bits

> without doing

> #define LONG_DOUBLE_TYPE_SIZE 64

Well, I will not protest your statement but it is odd that it used to
"work"; I guess we just got lucky with the old real.c code configuration.

OK, based on reading the new documentation (sorry I didn't fish a bit
more before posting my observations), a new correct way to get this C ABI:

	`long double' using "64-bit double precision" IEEE mode
	(confirmed by close reading of system headers and other
	sources) with 96-bit size and alignments (but only god knows why)

would be something like this:

1. Create and document a new macro ala INTEL_EXTENDED_IEEE_FORMAT.

2. Use it in port's config file:

#define LONG_DOUBLE_TYPE_SIZE 96
#undef INTEL_EXTENDED_IEEE_FORMAT
#define FREEBSD_I386_LONG_DOUBLE_FORMAT

3. Add a default section to defaults.h so that modern macro usage works out.

4. Handle it in format_for_size() case IEEE_FLOAT_FORMAT, case 96

(Under the untested assumption that padding is placed at the end...)

        case 96:
	  if (FREEBSD_I386_LONG_DOUBLE_FORMAT)
	    return &ieee_double;
          else if (!INTEL_EXTENDED_IEEE_FORMAT)
            return &ieee_extended_motorola;
          else
            return &ieee_extended_intel_96;

Regards,
Loren