fun fact: unflake tests pass (on my laptop) in ~18 seconds when using #CppNix / upstream #Nix, but in just ~15 seconds when using #Lix

this unscientific benchmark shows a ~17% reduction in runtime, which is pretty cool

Had to switch back to base C++ #nix from #lix to package  #hledger's prebuilt release version, because lix still has an annoying builitins.fetchTarball bug¹ that prevents using tarballs that just contain a bunch of files without an extra single directory at the top level. (As you know, tarballs only EVER contain a top-level directory and never just a bunch of files, right?). In #cppnix it's already fixed.

¹https://git.lix.systems/lix-project/lix/issues/582
²https://github.com/NixOS/nix/pull/11195
³https://gitlab.com/nobodyinperson/nixconfig/-/commit/7198ae401c6b6dd875c7409bd009e6e59b9d8e0a#90cc0c517a4752361c635864aed25f106afccd7d_5_4

I'm still working on #Nix parsing. It's a nastier language than I realized; there are a couple strange constructions I've never seen before. Here's my favorite parser quirk of the day; in #CppNix, try:

> let x = 1; in let { inherit x; }

What do you think will happen? Why?

I finally think I've figured out a basic #Nix heap using the #Haskell/#GHC technique of flagging every heap object to detect infinite loops. Cache invalidation really is hard... Fortunately, Nix is pure, so I don't have to invalidate any heap objects directly. #RPython makes it easy to hack, although it's still a bit of a slog to debug segfaults.

I think I might have to actually look at #CppNix now, or at least reread Eelco's thesis. I need to understand the differences between evaluation and instantiation.