>>>>> "Gerd" == Gerd Moellmann <[EMAIL PROTECTED]> writes:
Gerd> Raymond Toy <[EMAIL PROTECTED]> writes:
>> I believe thread switching is slow because it has to copy the control
>> (?) stack on each switch.
Gerd> Maybe you mean the "binding stack", which holds bindings of special
Gerd> variables?
Actually it seems to be both (and an alien stack). There's the
control-stack-fork vop and friends that copy the control stack.
multi-proc has stuff about the binding stack and alien stack.
>> The obvious solution is adjusting the stack pointer to point to the
>> new control stack, with no copying. Do not know if this would work
>> or not.
Gerd> Well, I guess one would have to frob SYMBOL-VALUE to make this work.
Gerd> One could store symbol values for each thread in thread-local storage,
Gerd> and let SYMBOL-VALUE retrieve it from there.
That would make symbol-value much slower wouldn't it?
Gerd> On x86, I think its common to use one of the segment registers FS or
Gerd> GS to implement per-thread data. A thread switch changes the value of
Gerd> the segment register, and instructions with a segment prefix can be
Gerd> used to access thread-local data.
Gerd> I hope that's approximately right; it must have been >10 years since I
Gerd> last used that :).
I think that's how SBCL implements it's threads on Linux. :-)
Ray
