kFloat v1.0 - Improves execution time of some floating point
              words of ju:float.ffp
Jerry Kallaus  05/08/89
                  
JForth and assembly code which dramatically improves execution
time of FFP floating point arithmetic of ju:float.ffp in
JForth v2.0.

When this code is "included", it "includes" ju:float.ffp if
necessary and redefines several words.  Functionally, everything
should basically work as before.

The floating point package contained in ju:float.ffp provides
conditional operator words which check the condition code from
the last floating point operation and leave a flag on the stack.
These words are FLT FLE FEQ FNE FGE FGT FVC FVS.

Unfortunately, while the code for these words is fast, providing
them requires that the condition code resulting from every
floating point operation must be retrieved and saved in case one
of these words is used in following code.  This can be eliminated
in areas of the users code by setting the compile time flag
FP-STATUS? OFF.  When this flag is off, the condition checking
words listed above may NOT be used.  Note that
FP-STATUS? ON must be used prior to the use of any floating point
code for which the condition will be checked by one of
the words listed above.  Note also that there are several other
floating point relational and condition checking words for
which are NOT affected by any of this.

In any case, a faster library interface is provided for the
four arithmetic operators f+,f-,f* and f/.  Faster code
is also provided for the following words:
FLOAT, INT, FIX, F2*, F2/, DEG>RAD, and RAD>DEG.
The constant for RAD/DEG is added.
None of the transcendental words have been changed.


TIMING

Timing comparison of words of ju:float.ffp (old) and
words redefined by kFloat (new) with status saving
(w/ stat) and without status saving (wo/ stat).
Times are approximate in microseconds.
 
op     old   new   new
             w/    wo/
             stat  stat
 
f+     116    78    56
f-     116    78    58
f*     150   116    90
f/     186   148   120
float  134    42    42
int    122    24    22
fix    150    38    38
f2*     54    32     4
f2/     54    32     4



BUGS

None known, but would appreciate knowing of any that are found.

This package corrects the following  problems in both
the Amiga 1.3 mathtrans.library and JFORTH 2.0 ju:float.ffp.


Amiga FFP floating point math library.

_MathBase SPFix

Any argument with a low order byte of zero results in the
input returned as the result even though bits are set in the
high order three bytes, a large integer.
Should return zero since a ffp exponent of zero represents
an exponent of 2**(-64).



JForth ju:float.ffp floating point math utilities.

ju:float.ffp F2*

Given an argument of zero, F2* will increment the exponent part
to a one and return $ 00000001 as a result; an attempt to print
with F. will give the error message
"...warning fp too large to fix".
F2* should return zero for an argument of zero.


ju:float.ffp FLOAT

Has a problem with large negative integer arguments in the range
$ ffdf,ffff through $ 8000,0001.  For example
$ 8000,0001 results in $ 8020,00E0, should be $ 8000,00E0
$ FFDF,FF00 results in $ 8000,02D7, should be $ 8000,04D6
The comments in the code for FLOAT suggest that it needs to
do something special for arguments greater than 2**21 because
the library FLOAT cannot handle them.  Actually, the library
code seems to work fine, but gets messed up by this special code???
Maybe the library was fixed from an earlier version.


ju:float.ffp INT, and thus also FIX

When an overflow occurs, returns a value of zero!?
The library returns maximum positive or negative integer;
this seems more appropriate to me.