feat: small-string optimization

[tomberek]

If a string can fit into a Value, without a context, then pack it in. This avoids extra allocation and clean up for common strings.

Motivation

Speed up Nix's string performance, reduce memory usage.

• memory: 98%
• cpu: ~same
  • I expected a bigger improvement as this avoids nearly 1M allocations. Very likely my C++-fu could be better.

Context

Looking at various approaches: small-strings, symbol table usage, storing strings as a list of pointers for efficient concat without allocation.

It is easy to violate GC assumptions when doing things like this, so I expect to go through a few iterations and plenty of testing.

ping: @pennae

Priorities and Process

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

[pennae]

this works and does reduce memory use, but it makes some hot paths in the code much more expensive and increases the number of instructions executed overall. it may be more productive to reduce the size of Value to a single tagged pointer, reusing most bits of what is now the type field to store non-gc pointers and the type tag. that too would be very intrusive though and itself not likely to speed things up all that much :(

[Mic92]

What does memory: 98% mean? Do you save 2% of memory? And in what benchmark?

[nixos-discourse]

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/2024-03-04-nix-team-meeting-minute-130/40830/1