cargo test takes too long if each crate has a dedicated salsa db

[xiyuzhai]

I use salsa-2022 to create my own programming language https://github.com/xiyuzhai-husky-lang/husky, and I construct a salsa db struct for almost every crate. Now that there are too many crates now, the cargo test time has increased dramatically. My last github push takes 40min to pass the checks.

For example, this is the db for testing in some late stage crate. The macro expands to a thousand lines of code.

#[salsa::db(
    CowordJar,
    husky_vfs::db::VfsJar,
    EntityPathJar,
    husky_token_data::db::TokenDataJar,
    TokenJar,
    AstJar,
    EntitySynTreeJar,
    TomlTokenJar,
    TomlAstJar,
    ManifestAstJar,
    CorgiConfigJar,
    CorgiConfigAstJar,
    // linkage
    husky_linkage::db::LinkageJar,
    ManifestJar,
    SynExprJar,
    SynDefnJar,
    SynDeclJar,
    TermPreludeJar,
    DeclarativeTermJar,
    DeclarativeSignatureJar,
    husky_declarative_ty::db::DeclarativeTypeJar,
    EtherealTermJar,
    EtherealSignatureJar,
    FluffyTermJar,
    SemaExprJar,
    husky_hir_ty::db::HirTypeJar,
    husky_hir_eager_expr::db::HirEagerExprJar,
    husky_hir_lazy_expr::db::HirLazyExprJar,
    husky_hir_expr::db::HirExprJar,
    husky_hir_decl::db::HirDeclJar,
    husky_hir_defn::db::HirDefnJar,
    RustTranspilationJar
)]
#[derive(Default)]
pub(crate) struct DB {
    storage: salsa::Storage<Self>,
}

I suggest that there is a dedicated salsa db for testing with no generics so that it needs only to be compiled once instead of being compiled for each individual crate. The Jars type could be something like

type Jars = HashMap<TypeId, Option<Box<dyn ...>>>

so that it's the same for different crates.

It can implement HasJar<SomeJar> by looking up the hashmap using the type id of SomeJar. In theory, this DB is the final object, it implements HasJar<Jar> for any Jar. It would be inefficient in release, but would save tons of compilation time for testing.