Optimization for short strings that fit within TValue.

[Tomarty]

This could be a compiler flag that introduces 1-2 internal types (or additional state for LUA_TSTRING if there are unused bits.) Strings of sizes 0-11 come to mind, but there could be an additional type for strings of size 12. This could vary based on the size of TValue at compile time.

Reasoning:

• This optimization is common in modern string implementations. It improves cache locality, etc.
• Blazingly fast to construct lots of small arbitrary strings.
• Fewer strings to manage on the heap, with memory savings with respect to string-heavy OOP code (Roblox Enums currently have 2291 unique strings, 1459 of which would be optimized by this.)
• Strings would become a practical scaffold for low level user types. For example a quaternion rotation could be packed into 12 bytes, and have buffer read/write support to boot.

Some challenges:

• Complexity increase.
• Breaks C compatibility? I base this assumption on lstring appending '\0' to all strings. This seems odd because lengths are known, and Lua permits '\0' within strings; I'm not sure what the implications are.

Has this been discussed before?

[HaroldCindy]

Has this been discussed before?

It was at least brought up in #512, but Luau has generally not implemented changes that increase the complexity of the VM without specific metrics showing that it would be a net improvement in a Roblox environment. I'm sure the Luau team is aware of the short string optimizations since they're pretty common in other Lua implementations, and they seem to consult those for practices that've worked well for others.

A number of features from Lua 5.1 have actually been removed if they increase maintenance burden without a convincing use-case within Roblox, like the emergency garbage collector.

I believe Roblox tries to mitigate some of the negative effects of all strings being on the heap by doing things like having fewer VMs (and by extension, heaps) so they can benefit from the VM-level string interning inherent in Lua's heap-allocated strings.

[vegorov-rbx]

We do not plan to include divergent string types in our implementation of Luau VM and will not accept PRs with that functionality.