Table of Content  Branch Up  

Early (very crazy) Ideas

If I managed to write a Linux Kernel that runs on my machine, then it would become a Linux Mini-Computer. Not only it will behave similar to 1970's UNIX machines but it also (and most importantly) will be able to run existing (Open Source) Linux software on top of the Kernel.

I know this is too big for me alone so much likely will never happen... just a crazy idea that came to my mind.

05/15/2009

A serial (RS232) port. I can't imagine a simplest way of connecting the machine to the outside world. This would be the way for dumping large amount of machine-code directly into memory.

05/16/2009

This is really crazy. Since modern operating systems such as Linux always try to distribute process among available processors, it occurred to me that I could build a machine using many little microprocessors each of them being capable of handling a user process.

By "little microprocessors" I mean cheap ones, such as the legendary Z80. They lack a "protected mode", indeed, but this were something to be handled by an "arbitror" circuit (anyways needed), not by the processor itself.

Hey, Craig... could you please prove this proposition for me using your so called "Object-Z notation"?

05/21/2009

I can imagine a computer with two CPUs. But not two identical CPUs but on the contrary, two CPUs that are very different each other.

One (lets calle it "CPU-0") is very low profile whereas the other ("CPU-1") is full featured. Each one has its own separate physical memory; they don't share memory and possibly no storage either; they are totally isolated from each other.

CPU-0 runs the Kernel and it does in "real mode": no virtual memory, no protected mode. As a matter of fact, this CPU does not implement those advanced technologies.

CPU-1 runs the user processes in protected mode; possible this is the only way it can operate: in cannot be "switched back to real mode". The OS Shell also runs in CPU-1 as an ordinary user process. Both Kernel and applications runs concurrently implying that the Kernel never stops: there is no switching between Kernel and User modes.

The only link between the two CPUs (apart from the power supply) is by the mean of hardware interrupt requests. Actually, all processes in CPU-1 runs as IRS routines. The Kernel switches processes by issuing hardware interrupts to CPU-1. Similarly, CPU-1 notifies the Kernel via hardware interrupts. Or maybe there is a "secret path" between the two. We may call it "Inter-Bus" or maybe: "Tunel".

Estrange architecture... Just another idea among many.

I asked my self how can possible be for a processor to execute every instruction in a single clock and, besides, what is the gain if it will have to fetch the code anyways. The answer is, of course, prefetch and pipelining. Then it occurred to me that I could "easily" implement a prefetch scheme in which code is unmistakable available to the CPU. This is the idea:

There are actually two processors, one into the other. The outer one is in charge of prefetching, nothing else. If my "Trans Matrix" idea holds, then this would be processor who dealed with the Matrix for making address translation and that stuff but this is not important now; the interesting part is how this processor will do prefetch and how is the prefetched storage area.

Well, the prefetch area is actually a stack. The inner CPU (the "real" one) will use its PC as a stack pointer. It simply pops code from the "stack" and executes it right away, no fetch needed. Moreover, the stack itself could be used as IR with no need for a transfer. I will forget about pipelines for now; with just having the code right in place all the time would be a tremendous improvement with respect to the old-fashioned approach I'm taking for the Heritage/1.

So the outer processor (or "preprocessor", as we may call it) needs to make sure that the code is being placed in the right order. So this gay is smart, it knows the instructions set or at least it knows branch instructions.

Knowing this, it will do jumps along its memory reading, following the program flow and making sure the "stack" content is not only prefetched but also "pre-sorted". You may be asking: What about conditional braches? You are right; prefetch prediction is too complicated so the "pre-processor" will simply stop prefetching as soon as it encounter a conditional branch.

Another question you may be asking is about data. Well, in reallity there are two stacks: one for code, the other for data. The computer is Havard inside, Von Neuman outside... something like that.

Well, that is the idea. I know it's not as easy I've said (I know by experience that finding things easy is not a good sign) but I found it cool and wanted to write it down.