"Tom Linden" <tom@[EMAIL PROTECTED]
> wrote in message
news:ops3onzoiizgicya@[EMAIL PROTECTED]
> On Fri, 20 Jan 2006 15:08:17 GMT, robin <robin_v@[EMAIL PROTECTED]
> wrote:
>
> >> I don't believe that is true, I believe most floating point
> >> representations that
> >> have used a binary exponent have suppressed the leading one to obtain
> >> one more
> >> bit of accuracy. But with a radix 16 exponent you can't. of course
do
> >> that.
> >>
> >> Not sure how far this goes back in time, but i bet it is to the 50's
> >> anyway.
> > No. The leading bit wasn't suppressed, even wh
>
> Well, it was on a number of machines that I have worked on and since a
> normalized
> float always has a leading 1 for the characteristic (Mantissa) so no
test
> needed
A test is always needed because the hidden '1' must not be inserted
for a zero value when the hidden '1' is expanded.
>and small amount of additional logic needed for the ac***ulator.
Any logic of that type was not "trivial:" as it was all hardwired.
> Thus if the floating point number had a characteristic of n bits
> the ac***ulator would need a minimum of n+1 bits
n+2 bits, to allow for the expansion of the hidden '1', plus carry
out of the most-significant position (two non-zero mantissas
will always generate a carry (1+1=2))
> + possibly guard bits
Machines of that era didn't have guard digits.
> > The reason was that it was more expensive (if in hardware),
> > requiring a test and generation of the bit. In a serial machine,
> > that wasted two machine cycles.
>
> I don't see that, following an operation the result would need to be
> normalized anyway.
The test is required _beforehand_ when unpacking the
FPN for an arithmetic operation.
> > In software, all it gained was loss of time and loss of
> > fast memory (always in short supply).
|
