Re: KISS4691, a potentially top-ranked RNG.

["robin" <robin51@dodo.mapson.com.au>]

<nmm1@cam.ac.uk> wrote in message news:iplmk8$2s6$1@gosset.csi.cam.ac.uk...

| As those of us with significant experience of portability will
| remember, building fixed sizes into a program is a cardinal
| error.  Sooner or later, you will want to run that code on a
| system with different sizes, and it won't work.

That's because FORTRAN was designed around a word machine,
where float precisions offered were either single and double precision.
That could cause different actions when moving from a 32-bit word machine
to 60/64-bit word machine,

Later languages, including PL/I, were designed around
byte machines.  The way in which float precisions are specified
lends itself to portability.  Thus, asking for say, 12 decimal digits
got you about that number of digits, whether the machine offered
32-bit or 60-bit words.  Some machines actually offered decimal
(BCD) floats in hardware,* and asking for 12 decimals got you about
that number of digits.

|  The vast
| majority of clean programs, even in C90, needed NO source
| changes to go from 32 to 64 bits; I have written code that was
| moved to systems with 16, 24, 32, 36, 48, 60, 64 and perhaps
| other sizes, and so have many other people.

Indeed.  (Including doing portable character handling before F77.)

| I accept that Marsaglia's generator is a valid case for using
| fixed-width types, but that is likely to make it very inefficient
| indeed on a system where they are not natural.

I very much doubt that any new machine is going to offer less than 32 bit
integer words ; for some years now the trend has been to machines of
64-bit integers.  Marsaglia's (integer) algorithms will run on such machines,
some possibly with minor alterations, but will not, even if modified,
will not run "inefficiently".

|  Masking is a
| far more portably efficient approach.

_______________
* some still do, in an updated verion of decimal.