So I shall setup the base for a stack frame setup, one that will be easy to implement and easy to use.


I can put the stack direction, alignment, and the stack pointers used in the layout. Basically anything related to the stack would be placed in the layout. This way, all of that information is together. I suppose initially I should have it so the code can handle PowerPC's stack frame layout, which has a number of requirements.


So hopefully with the creation of the stack frame layout, I can very much more easily and more simply write the stack frame that is needed. My hope is that the current invokeMethod() method is much smaller. It starts from line 218 and ends at 394 and as such is nearly 100 lines and I would guess that it would be about 20-40 lines more. However even with only 100 lines it is still quite complex.


Ok, so I had an idea. The code decoder stream needs flags. These flags will be stuff such as NO_CODE which disables code parsing and just states that only the layout of a class is to be parsed. Then another thought I had during my walk is that I really really like a fixed size stack where I can preallocate every variable as needed. What I can do with this new flagging system is have it where I can have a special flag which is used to go through all operations and then set in a bit field, all types which are assigned to a given location.


Then this way, I can determine how I want to allocate registers. When it comes to the stack, I will still need to select premade allocations although I do not have to worry about frame pointers or setting the stack pointer. Also the stack pointer can completely act as the base. So this means that the stack direction needs a top which is the opposite of base.


Actually I will do that regardless since it is difficult to forward JIT specific options currently.


I am considering making SquirrelJME a kind of hybrid Java/C program where the C program runs code, but has symbolic access to the byte code (which is compiled). That is, they can call each other and can execute each other. I would have to write my own C compiler, because the system one would be unable to handle access to Java and there would be a great difference in the target machine. I would want the C part to have the same size types for int and such regardless of the underlying system. Otherwise, the mass number of defines would not be maintainable.


A pure Java approach can work, but everything is a heap reference for the most part and there would be overhead of some calls for example. However, if the JIT is written partly in C then I would need a way to execute that C code in a way where it can be bridged with the JVM. Oracle's Java ME VM is written in C and is pretty much a pure C program. For hybrid programs, I would still need to recompile the byte code to native code. However, right now I output to namespaces which are part of objects.


However, switching to C would be quite the scope creep. I believe what I need is to actually splice apart the JIT. Basically there will be NO namespaces writen by the JIT. There would be a namespace target specific output (say to ELF or similar) and another which can generate machine code, similar to the native code generator.


So basically, the JIT for the class is going to splice the class handler. I would say it would be split into two parts: Container and NativeCode. Since Container is already is a class then Binary should be used instead.


Essentially the JIT on top of the class decoder will be split in where it handles things. There will be a basic output which writes to the container along with one that is attached to the native machine code generator. So the machine code generator would be for just generating machine code spit out by the JIT. Then I can perform a kind of optimization of sorts so that I can have two different binary containers: one that outputs to a native binary for usage for system execution (say an ELF with symbols) and another which is used by the JIT at runtime to execute JIT compiled code. One would be static while the other would be more dynamic. This would then make it where outputting to ELF or say PE at a later step is removed and is handled by the binary container code. This would essentially remove the link step along with needing a kind of cache for output. Basically it can write the ELF and such in a single go for the most part. Since it might not support linear output, I can use FileChannel to write ELFs and such and update it that so that it does not need to store anything in memory. So it would initialize it with unknowns and write the known information on close.


So basically this would be a splitting refactor that would remove the need to have the link stage to ELF and do that output directly, and potentially having symbols and other linkage information. This should actually make output generation be simpler and remove the potential need for namespaces. I just need a kind of descriptor for actually being able to use the namespaces at runtime. This could actually be separate and I could even support code and data completely alone.