Re: KISS4691, a potentially top-ranked RNG.

[Keith Thompson <kst-u@mib.org>]

glen herrmannsfeldt <gah@ugcs.caltech.edu> writes:
> In comp.lang.fortran Keith Thompson <kst-u@mib.org> wrote:
> (snip, I wrote)
>>> No.  You need CHAR_BIT to tell how many bits are in a char.
>>> It has been known to get to 64 on word addressed 64 bit machines.
>>> It must be at least 8, but can be more.
>  
>> Certainly CHAR_BIT *could* be 64 (or even more), but I've never heard
>> of a 64-bit system with CHAR_BIT==64.  Such an implementation would
>> have trouble dealing with octet-oriented data from other systems.
>> Some DSPs (digital signal processors) do have CHAR_BIT > 8.
>> There have been systems with, for example, 9-bit bytes, but they
>> had pretty much become obsolete by the time C was standardized.
>
> The story I remember was for a Cray machine, but I never used
> one to check on it.  I do know that there are some 64 bit
> word addressed machines, especially some from Cray.

I used to work on a Cray T90.  It had 64-bit words, and no direct
hardware access to smaller chunks of memory; to read an 8-bit byte,
you had to read the word containing it and then extract the bits
you wanted.  But the C compiler (C90, it never implemented C99)
had CHAR_BIT==8.  char was 8 bits; short, int, and long were all
64 bits.  A pointer to anything smaller than a 64-bit word stored
the byte offset in the high-order 3 bits (the address space was
much smaller than 64 bits, so the high-order bits weren't being
used for anything else).  This was all handled in software; the
compiler had to generate extra instructions to deal with byte data.

Making CHAR_BIT==64 would have made much more sense for the machine
itself in isolation, but it would have made interoperability with
other systems quite difficult.  It ran Unicos, Cray's BSDish Unix;
I'm not sure you can even implement Unix with bytes bigger than
8 bits.

It made it very slow for things like string processing, but since
it was primarily used for floating-point computations that wasn't
much of an issue.

Some older Cray vector systems ran other, non-Unix, operating systems
(COS, CTSS), but I never used any of those, and I don't know what
value their compilers used for CHAR_BIT -- if C was even supported.

-- 
Keith Thompson (The_Other_Keith) kst-u@mib.org  <http://www.ghoti.net/~kst>
Nokia
"We must do something.  This is something.  Therefore, we must do this."
    -- Antony Jay and Jonathan Lynn, "Yes Minister"