Consider reframing internal compiler typing logic to be "locally-based constraint solving" instead of "globally-based"

[emilyaherbert]

Previously the compiler has relied on a global TypeEngine in order to complete type inference. Now that there is a PR to remove this (#3353), we can begin to think about introducing an internal change to the how the TypeEngine that will make understanding and contributing to the type system much easier and will reduce the scope of new changes significantly.

Instead of approaching type constraint solving "globally", we should change it to be more "local" (I'm making these phrases up as a go so I'll try my best to explain). Essentially what I'm trying to say is that exposing all types across all scopes of the program is adding unnecessary complexity. For example, when doing type inference in a function body, the types that you need are 1) from the scope surrounding the function body and 2) from the scope of the function body itself. You don't need to drag around the types from inside the declaration in the previous declaration block, which is what the global TypeEngine does.

I think that this issue manifests in the fact that local constraint solving is not fully solved right now. For example, #3324 demonstrates that there is an issue resolving types with match expressions (although there is a fix in #3348).

Notes:

• https://rustc-dev-guide.rust-lang.org/ty.html#question-why-not-substitute-inside-the-adtdef
• https://rustc-dev-guide.rust-lang.org/generics.html
• https://doc.rust-lang.org/nightly/nightly-rustc/rustc_middle/ty/subst/type.SubstsRef.html