Awkward alignments for ZBD

[John R Harrison <johnh@xxxxxxxxxxxxxxxx>]

Hi Michel,

(I took the liberty of changing the subject line from "Possible Wording 
Contradiction in Signaling", which didn't seem to match the topic.)

| Just curious (before I investigate further):  aren't scaling or
| normalisation shifts that are not a multiple of 3 awkward for ZBD?

Yes, these do seem tricky when using the millennial digits directly.
But as you suggest later, one can always convert all the way to BCD.

This is why ZBD seems strictly better than DPD for software: if the
operation admits a fast direct implementation you can shortcut it,
while if all else fails you can do whatever you would anyway for DPD.

| Perhaps there are tricks that can be pulled off with 16-bit MMX
| fields, but without that, it seems that declets/millennia have to
| be processed serially, with carry propagation.

I did once play round a bit with "manual SIMD" algorithms, doing
reciprocal multiplications by shifts and adds, and masking to keep the
subfields isolated. My tentative conclusion was that it was too
hacky, and probably not worth the trouble. But perhaps there are
things I didn't consider.

John.