things I've changed from the PR:
- dropped legacy (i.e. non-flake) compat stuff, which turns out to account for most of the diff
- dropped `packages.garnet` since it doesn't work with `nix build .#garnet`
- back to using Rust-extended packages everywhere, which _seems_ fine
the rest is just re-inlining things and other refactors
note that changes we keep are essentially:
- not using Crane's extended packages everywhere
- bumping `nix-haskell` to avoid shell hook workaround
- various changes in how we call `nix-haskell`
- using `libCgarnet_rs` name, which Cabal expects
- adding proper non-dev-shell targets, so that e.g. `nix run` works
stuff we should still take a look at:
- seems a bit weird that `garnet-rs` arg to `project.nix` was always same (and therefore now we always pass the ) - might be a mistake, and we're supposed to be using local for local build?
- similar `libCgarnet_rs.a` logic appears in three places - see if we can deduplicate to some extent somehow
- I haven't even checked all steps are necessary
This reverts commit 1f1c0d959da699ce04f7951ecbcdb7976c8c0750.
This doesn't work well with multi-component builds. For example, it requires `"haskell.sessionLoading": "singleComponent"` in VSCode, which makes HLS work less reliably.
The script addition is a bit hacky, but there's no obvious straightforward arch/version-independent way to get most of the build path. And eventually, once issues with HLS etc. are sorted out we will revert to using Cabal hooks anyway.
Even though `Int` and `isize` should be the same in practice, we can't cleanly convert, as the type information isn't quite there. And anyway, strictly speaking per the report, `Int` is only guaranteed to hold 30 bits.
Note that this is essentially unchanged even if we specify `usize_is_size_t = true` for `cbindgen`.
