3A/8.3

The example shows the use of a new value.
... but there is nothing about this in the "text".

Is it absolutely necessary to use a new value to get a real serializing behavior ? ... or not ?

Has someone an experience about it?

(someone, anyone ??? sorry, I don't know)

Most probably CPU does not have a caching mechanism for writes to any register. If you think about it, it is easier to just overwrite the register than to check if the two values are equal and then overwrite it.

I think the documentation is a bit ambiguous about this, but you can be certain that the value must not be "new" in the sense that this value already is in that register. Every write to a register overwrites the value in that register. New or not.

apparentlly, the serialisation here means that the execution of all instruction preceding a "Privileged serializing instruction" should be achieved before the instruction can be executed.

for example,

will need to complete every general purpose instructions before the lgdt [ebx] instruction. otherwise, it can crash probably the cpu.



when you enter a new value in cr, dr, gdtr, idtr, etc.. register.

this can be due (i suppose) to the fact that some initialisation should be made onthe structures and registers of this kind of extension.
because control reg, debug reg, etc, it is an extension of the processor.
a system part, where you don't execute code, but control the execution of code, the system controler circuitry.

then, a privileged instruction will induce serializing.
cpuid, iret and rsm don't induce serialization, but are privileged, because they depend on the system controler.