hi;
I am in process of finishing my own kernel for 8052 compatible CPUs.i thought i shall share a few points:
1.Decide the context of task: if you are keeping a,b,psw,dph/l,b,ret_add ls/ms only as part of register set then your task context overhead reduces but then you can only run 4 tasks at a time.also check out the compiler you want to use.if it is a compiler like Tasking which deafaults to a
certain register bank then you are better of
including that bank as part of context.
2.Check the stack usage and local variable options of the compiler and have option of user settable stack depth.
3.Context switch can be done in different ways:
a.You keep seperate stack for each task.this reduces the context switch time but limits no.of tasks
b.If you decide to use shared stack then it involves swapping data between iram and xram during switch.this is really slow.

let your application decide between the two.

4.The classical concepts of preemptible kernel put too much overhead on 8052.you are better off making kernel calls non-preemptible.at most you may consider restarting by rejecting stack load when in non critical region.
5.Think again before you try to implement priority inheritance.as per my knowledge keil,uc-osii(2.04),rtlinux(fsm labs)do not implement it.the algo appaers simple but is too time consuming especially when nested critical regions come in play.if you must attempt for non nested one.
Keep an open mind.Best of luck.
pranav