Optimise execution speed

[malcolmstill]

Goal

• Speed up the interpreter. fib(39) takes around 26 seconds on my machine. This is in comparison to, say, lua on the same machine taking ~6 seconds.

Details

• The interpreter loop is currently a large switch statement. One problem is this switch statement ends up with a single jmp instruction which won't be great for the branch predictor
• Andrew is proposing adding support for labeled continue syntax introduce labeled continue syntax inside a switch expression ziglang/zig#8220, that would generate better machine code including per-instruction jmps
• In the same issue haberman suggests using a tail-call interpreter which I'm experimenting with in Tail call interpreter dispatch + unified stack #118 (which also unifies the stack...I have another branch that doesn't unify the branch and I'd like to complete both so I can compare them against one another). With the tail call interpreter I'm also experimenting with which parameters to include in the call as we'll get speed up if certain parameters can just remain in registers / avoid register spilling

References

• https://mstill.dev/posts/13/
• introduce labeled continue syntax inside a switch expression ziglang/zig#8220

[malcolmstill]

This is closed (at least initially) by #120