Ok, so all of this value shuffling in
__JITCodeStream__ is going to be a
complete mess. I would have to scan the entire tread to see which registers
are available. I would best use the
MultiSetDeque that I wrote for keeping
registers in check. It would just be for any kind of register, although kept
separate for float and int. There will be a constant initial pool of
registers. When a state is derived it will full the queue with that initial
register pool. Then for every register that is consumed it will remove those
from the pool. Then allocating of values will be handled by the tread code
with a link to the engine. I can also keep a mapping of what a slot is
aliased by too. That way I do not have to go through everything all the time.
This should make it much easier to code and remove all of the allocation
smashing code in the tread code.
So argument registers are considered temporary. I do wonder about my strategy for allocation. Using saved registers for local variables would probably be the most efficient choice. Then use temporaries for registers on the stack. Since local variables are preserved in the method calls. So I would need two pools of registers.
__removeSlot() will become
it will no longer be in the code stream handler.
I am going to need a test framework again like I did before. However the old framework was really complex and relied on services much. I suppose what I need is a general midlet that can be used for testing which contains all of the tests. I suppose it would be similar to before, but I will exploit packages and resources to locate tests rather than maintaining global lists. It should be a bit easier to maintain.
Since I would want the test system to run on existing Java ME systems (since I need to see how certain code runs) I would need a profile. However in some cases there might not be a profile to target for results. So what I would need in this case is a provider for a default profile set of expected values. So there is just a normal test which runs and outputs a result. If a profile is not being generated and no input profile was specified then the default profile is used to compare against. The testing system can be recursive also. I can use basic index files for tests which can recurse as needed. I would have the tests ordered so that they run the same. The same order would also mean that I would not need to load all the tests and sort them accordingly. The test indexes will just contain either directory references or class names. After they are read, they are sorted and executed.
Profile results should be diffable using the standard unified diff so that they may be compared by hand if needed.