[1022] in Coldmud discussion meeting

root meeting help first previous next last

[COLD] Big Floats

daemon@ATHENA.MIT.EDU (Thu Jun 20 20:55:05 1996 )

Date: Thu, 20 Jun 1996 18:29:19 -0600 (MDT)
From: Brandon Gillespie <brandon@tombstone.sunrem.com>
To: coldstuff@cold.org

Right now the float type is (in general) a 32bit size (i.e. 'float').  
Some people would like greater precision, so I have been adding a simple 
'USE_BIG_FLOATS' define to the driver which would kick it up to 64 bits 
(double).  There is one major kink, and that is having a constant float 
in a method.  Right now when you have a constant number it is stored 
within the list of opcodes (which is a list of long numbers, likely 32 
bits).  If you have a big float number (64 bits) we have to do one of two 
things:

    1) reduce its precision to 32 bits
    2) split it across two opcodes, keeping full precision.

I personally prefer the first option for many reasons, primarily being 
that despite all of this talk about bits I am doing my best to remove 
bitsize dependancy from the driver.  The first option would simply mean 
we cast the float to whatever the opcode type is, and be done with it 
(casting it back to a float when needed), we do this right now without 
problem.  The second option would require a bit of coding, and would be 
adding bitwise dependancy (a definite nono), but everything would work as 
it is expected.

Keep in mind this is _ONLY_ with constant float values, such as:

    public method .foo() {
        return 0.5;
    };

Frankly, I doubt anybody will be needing 64bits of precision in a 
constant value, but.. who knows.

However!  If you are using big floats, and you feel you must have full
'big' precision in your constant floats, you do have the option of also
enabling big numbers, which then uses 64 bits for a long.  In general this
is NOT suggested, unless you are on a DEC Alpha (as I dont know of a
single other processor that completely handles 64 bit words--many do
internal 64 bit words).  But this would then fix the 32bit constant float
size problem (when using big floats). 

Comments on me using this approach?

-Brandon Gillespie-