Frames: Design of Method Activation State about pysom

[smarr]

At least as an intermediate solution, the bytecode interpreter, needs a slightly adapted design to store the stack.

It is pretty much the same as the design for the AST-based interpreter, but we reserve space for the stack at the end of the Frame. The design of Inner remains unchanged.

Frame for Bytecode Interpreter

+-----------+
| Inner     | (Optional: for args/vars accessed from inner scopes, and non-local returns)
+-----------+
| Receiver  |
+-----------+
| Arg 1     |
| ...       |
| Arg n     |
+-----------+
| Local 1   |
| ...       |
| Local n   |  <-- initial StackPtr
+-----------+
| Stack     |
| ...       |
+-----------+
| StackPtr  |
+-----------+

To not have to pass around where to find the stack pointer, it's encoded in the last slot. The main reason here is not to have to change the common parts between AST and BC frames. Though, performance-wise it might not be optimal, because it's not a constant offset. The semantics of the stack pointer is that it points to the current element. For an empty stack, i.e., the initial StackPtr this means it points to the element before the first stack element.

In the long term, I hope the stack won't need to be part of the frame anymore. In TruffleSOM, it already works nicely to keep the stack simply in the interpreter loop and never escape it from there. The hope is that a compiler has an easier time to simply eliminate the array for the stack. For PySOM's bytecode interpreter, I'll have to see whether this works out and is worth it.

[smarr]

RPython Implementation Note

To make the RPython type system happy, everything stored in the Frame needs to be type erased (see rlib.rerased).
This is needed to be able to store a list together with the AbstractObject SOM objects, and the int for the StackPtr in the same list.

Since the Inner only stores AbstractObject instances, it doesn't need erasure.

For simplicity, the OnStack marker is simply a trueObject or falseObject, which should boil down to simply pointer comparisons.

[smarr]

Performance results for the AST interpreter changes:

https://rebench.stefan-marr.de/compare/RPySOM/3fd5865e788069ccf94775fc033242061ef1a2a2/41f23dc5073d8a8b21222e7c687739dee581cb6a#micro-steady-RPySOM-ast-jit