🚀 go-pugleaf

No doubt it's much too late, and I don't know if anyone from Intel
will read this anyhow, but I can't resist explaining to whoever's
listening how they could make the segmentation actually usable (IMHO).

1) With reference to the instructions that, in one instruction, load
16 bits of segment into a segment register plus 32 bits of pointer
into one of the other registers: 48 bits is too long for a pointer.
Create variants of these instructions that split up a 32 bit word; my
guess is the best compromise would be 10 bits of segment and thus 22
bits of offset, giving 256 segments (remember, two of the segment bits
are privilege level) of up to 4Mb each.

2) Add a TLB for the last several segment descriptors accessed -- 8
is probably plenty.  The idea is to make it possible, usually, to do
one of these segment+offset loads without having to go to memory to
get the descriptor.

With this combination of features, the segmentation becomes cheap
enough and general enough to use liberally.

Comments?

-- Scott
Gyro@Reasoning.COM