JIT support for LUA_VECTOR_SIZE == 4

[HaroldCindy]

The codegen code has a check that bails if sizeof(TValue) != 16, is there any interest in code to make codegen work when LUA_VECTOR_SIZE == 4, which changes sizeof(TValue)? Are there any obvious areas that'd need to be changed to support it (other than the obvious vector op stuff?) I don't see any notes on why JIT isn't supported in this configuration. Does the JITted code rely on TValues existing on 16-byte boundaries?

[vegorov-rbx]

Luau team at Roblox is not interested in codegen support under LUA_VECTOR_SIZE == 4 configuration.

There are no obvious areas, one will likely need to create vastly different lowering, register allocation and manual assembly sequences for both x64 and arm64 platforms.
Main reason is that 24 byte TValue doesn't fit in 128 bit registers.

I would also say that it's very unlikely that we will accept a PR doing all this work because codegen is an active stages of development and we, as main maintainers of this repository are not likely to take the burden of maintaining that code going forward.

People are always free to maintain their own forks with added features.
Also, native code generation was designed in a way to work as a 'plugin' into VM, it is possible to write custom standalone jits.

[HaroldCindy]

Oh, duh, I hadn't considered how expanding sizeof(TValue) beyond 16 bytes would affect register allocation. In light of that, the current limit makes a lot of sense, and JIT for 24-byte TValue probably isn't worth looking into. Thanks for your insights here!