Improving memory safety

[thomasvergne]

Plume doesn't deal with memory management within the virtual machine. This is a real issue due to memory leaks, so to resolve this issue, we might implement an algorithm for memory management:

• Garbage collectors like mark and sweep: easy to implement, but I don't know anything about the performance impact it can have
• Reference counting like Perceus: less easy to implement but looks like it has better performance

Reference counting seems to be the solution for now, but it requires some modifications in Plume:

• a new interface drop which would describe two functions: ref, and deref
• a way to write directly bytecode within Plume code, like raw code, to implement dereferencing operations on object types without the need to allocate a completely new type object with just one field for reference count and another for the value

This solution needs also to resolve issues with primitive types such as int and float. These types do not need to be referenced or dereferenced as they're primitive values, raw values on the stack. This could be achieved either directly during type-checking or during the type-erasure step.

[thomasvergne]

This issue is not completely finished as the current algorithm presents some runtime errors. They are probably dealing with too early freeing.

[thomasvergne]

Plume has made great strides in memory safety. We've implemented a very light garbage collector in C with an API that can be used directly from your libraries. Under investigation for leaks, but on the right track!

Taken from an example program from Plume repository

[thomasvergne]

The previous version generated runtime errors, causing Plume to almost always crash when running a bytecode file. Some changes have been made, such as modifying some factors and default minimum values for the GC.