(Experimental) Add mrb2cs AOT compiler

[hadashiA]

This feature significantly improves execution speed and memory footprint by converting mrb bytecode into C# code in advance.

This feature allows for a mix of regular mrb bytecode execution and mrb2cs-generated C#, and it can also fallback to the interpreter for features that mrb2cs cannot convert to C#. In other words, while maintaining full compatibility, it becomes extremely fast as long as you put in the effort for AOT-compiled release builds.

Design principle

Why AOT?

• The main benefits are improved execution speed and, above all, the reduction of allocations through escape analysis and placing objects on the C# stack.
• Since ChibiRuby primarily targets AOT compilation environments such as Unity IL2CPP, we chose the pre-compilation method instead of a JIT-like approach that emits IL at runtime.

Why not base it on Ruby AST?

• By taking mrb bytecode as input, it becomes easy to mix the current interpreter and mrb2cs from the same input.
• It can be utilized seamlessly by using the interpreter in development environments and mrb2cs (with automatic fallback to the interpreter) in release builds.
• Since ChibiRuby and the mrb compiler are separate, a feature that takes AST as input is generally complex.

Optimization Mechanism

• Based on ZJIT, it works as follows:
• First, the mrb bytecode is converted into a High-Level IR representation. By adopting the SSA form, escape analysis is made easier.
• After applying optimizations to the IR, it is converted into C# code.

Performance

Benchmark	Metric	Interpreter	AOT	Ratio
ao_render (alloc-heavy, float, cross-object)	time/iter	3863 ms	503 ms	7.7× faster
	alloc/iter	710.0 MB	120.2 MB	5.9× less
mandelbrot (numeric tight loop)	time/iter	894 ms	18.2 ms	49× faster
	alloc/iter	950 B	911 B	~same (near-zero already)
optcarrot (NES emulator, complex ruby)	fps	16.4 fps	54.8 fps	3.3× faster