[Info-vax] VMS Software Q1 '23 Update

[Arne Vajhøj]

On 1/29/2023 5:41 PM, Dave Froble wrote:
> On 1/29/2023 2:41 PM, John Dallman wrote:
>> In article <tr6dje$2s9rm$1 at dont-email.me>, arne at vajhoej.dk (Arne Vajhøj)
>> wrote:
>>> Didn't VAX to Alpha migration cause some issues with D floating?
>>>
>>> If I remember correct then Alpha did not fully support D.
>>
>> It didn't support D at all, as best I know. VAX had:

>> D format is double-precision, with 8 exponent bits, 1 sign bit, and 55
>> fraction bits, approximately plus or minus 2.9E-39 to plus or minus
>> 1.7E+38, with a precision of approximately 16 decimal digits. It's F
>> format with more precision, but no more range. This format was not
>> carried forwards into Alpha, and does not seem to have been much used.
>>
>> G format is also double-precision, with 11 exponent bits, 1 sign bit, and
>> 52 fraction bits, approximately plus or minus 5.6E-309 to plus or minus
>> 0.9E+308, with a precision of approximately 15 decimal digits. The
>> exponent has a bias of 1,024 (not 128). This format was carried forwards
>> into Alpha.

>       gfloat, sfloat, tfloat, and xfloat.  These  formats  correspond  
> to  the
>       SINGLE, DOUBLE, GFLOAT, SFLOAT, TFLOAT, and XFLOAT keywords.
> 
>       Keyword             Range                                  Precision

>       DOUBLE (64-bit)    .29 * 10^-38 to 1.7 * 10^38             16 digits
>       GFLOAT (64-bit)    .56 * 10^-308 to .90 * 10^308           15 digits

> Now, it has been my impression that that DOUBLE is D-float.
> 
> Walks like a D-float
> Quacks like a D-float
> Is it a D-float?

Good question.

It can certainly take D float as input and provide D float
as output.

The question is how.

The Alpha architecture manual says that it only
supports:
- load D from memory
- convert from D to G
- do G math
- convert back from G to D
- store D in memory
which sort of operate on D float but loose 3 bits precision.

The alternative is software emulation, which will be able to do VAX
compatible D math but at a speed cost.

So either hardware G math loosing 3 bits of precision (which is what
I assume when I asked the question above) or software full D math.

Arne