[WIP] BST: allocate instructions and operands adjacent to each other

[undingen]

transform BST to gapless bytecode
- all instructions of the CFG are copied into a single memory region.
- all instructions are packed and follow each other directly
- the CFGBlock just point into the bytecode block where the first instruction of the block is located

[undingen]

@kmod could you please do a initial review.
I think the first patches are fine but the last two (interned string change + the actual bytecode change) are little ugly.
The main problems for me are:

• the InternedString could in most cases just be a BoxedString* (we also already intern all of them)
• I don't like that I had to copy the CFG and CFGBlock classes and how I update the CFGBlock* inside the instructions. Probably I should remove the pointers and use indices instead.
• The whole copying of nodes is ugly and we should definitely use reader/writer methods instead of directly using the struct layout (this is helps with implementing the "set one bit in the opcode for 1byte vregs instead of 4byte"
• enforce that all blocks actual have a single terminator at the end because this is how we know when we reached the end of a CFGBlock

[kmod]

At this point InternedString is just a thin wrapper around an interned BoxedString, so it should definitely be possible to get rid of it. I'm actually pretty ok with not doing that just yet though: we need to do a larger change to properly use the fact that CPython's parser returns all these objects as Python objects (rather than doing the "PyObject -> serialized version -> getFooConstant" dance we currently do). So anyway the way you have it seems totally fine to me.

I'm a little bit more concerned about the bytecode-packing change because it feels like it's a bit of a detour away from the way we probably want to do it in the long run (constructing the bytecode directly in-place). How much work do you think it'd be to do that? I don't have a huge opinion about either getting it in as-is or trying to take some more time to do the construct-in-place.

[undingen]

I split all changes except the bytecode packing one into a new PR.
I agree with your concerns and I worked on a new way of this patch which directly inplace allocates the instruction gapless. Unfortunately I did not get it ready before my holiday :-( (it's trickier then I thought)
I'm a little worried that with directly emitting the instructions the use 1byte vregs instead of 4 when possible is harder to implement. With this copying approach it should be quite easy to add.

[undingen]

closing: this approach got replaced by the one in #1391