So a processor executes instructions, and an instruction/opcode is made up of several cycles. Say for an "INCrement" instruction, there would be a cycle to fetch the opcode prefix, another to fetch the location to increment, another to read the value at that address, another to increment the value, another to write the value back to that address.
There can arise a case where one CPU is writing to memory right as another is reading that same address. And depending on whether you synchronize the two CPUs between opcodes or between cycles, the one reading can end up seeing a different value. This can, in certain cases, break emulated software if not done correctly.
It gets even hairier when you factor in that each cycle consists of several clock ticks. There is time required to hold a value on a bus before a read or write is acknowledged.
So as I continued to try and increase the precision, I found myself nesting state machines within state machines. Each CPU core had one for the instruction table to select the active instruction. And then each instruction had one for the active cycle. And then each cycle had one for the active clock. And then there was also an interrupt-driven DMA unit that could trigger within cycles that had to be accounted for. It was just too complex to collapse into one giant, 100,000 case state machine inside of a single function. Imagine trying to implement an x86 CPU inside one switch table inside one function.
So you would have to traverse through 3-4 state machines to do something as trivial as "increment an integer", and then return all the way back up the stack frame and switch to the other CPU. The code ended up being around 90% state machine maintenance to 10% actual simulation code. And it was painfully slow.
This led me to the idea of having a separate cooperative thread (coroutine+stack frame) for each CPU. Whenever one would read from or write to something the other CPU could see, it would make sure it was 'behind in time' to the other CPU. Otherwise, it would switch the stack frame and resume where the other thread left off. When that CPU stopped, it would resume right at our first CPU's pause. Very highly reciprocal.
The end result was a code reduction from around 350KB for a CPU core to around 35KB for the exact same code. And thanks to what I'll call "just in time switching", it was possible to run one CPU through hundreds of instructions when it wasn't communicating with the other, greatly reducing host CPU cache thrashing. I ended up with a huge speed bonus to boot.
The one problematic area ended up being serialization. Basically, this model moves the state machine into the host CPU's stack frames. But you can't write that out portably, so it becomes a real problem to capture an exact point in your program, write it out to disk, and then resume at that exact point in a future run. That took a long time to solve, and required some serious trickery involving "checkpoints" for serialization. So it's something to keep in mind if you want to use coroutines/cooperative threads and also want to serialize/unserialize things.
What was refreshing was that the core logic for switching between two threads is surprisingly simple. For x86, it is simply:
thread_switch(to = ecx; from = edx):
mov [edx],esp; mov esp,[ecx]; pop eax
mov [edx+ 4],ebp; mov [edx+ 8],esi; mov [edx+12],edi; mov [edx+16],ebx
mov ebp,[ecx+ 4]; mov esi,[ecx+ 8]; mov edi,[ecx+12]; mov ebx,[ecx+16]
jmp eax
Basically: save the old stack pointer and volatile registers, swap to the new stack pointer, restore the volatile registers and return. The function design is the program-equivalent of a palindrome.(push/pop tends to be slower on Athlon chips than indirect memory accesses; and getting the return address into eax instead of relying on ret allows x86 CPUs to start caching the new instructions quicker.)
So, like everyone else who has discovered this concept, I of course wrote my own library for it in C89, which can be downloaded here: http://byuu.org/programming/libco/
Instead of trying like other libraries to build in complicated schedulers and hundreds of functions, mine is just four functions taking zero to two arguments each. No data structures, no #defines. The idea was for the user to write their own scheduling system that works best for their intended use case, instead of trying to be one size fits all.