This avoids necessarily having to allocate a virtual machine when dealing with function objects and calls which involve other units or contexts, we simply just maintain the call stack inside of a single virtual machine that might refer to instruction pointers in differing contexts.

Benchmarks indicate that relatively simple cross unit/context function have a 20% performance bump (bench external_functions) on my machine due to this, most of it would be due to avoiding additional allocations which are needed to set up individual stacks for separate context calls: