alt Hacker NewsIf your language is typed it's good to know at least a bit, so you can do the type inference properly; there are many ways to shoot yourself in the foot when it's ad-hoc.
Bidirectional type inference is a type inference style where you traverse the syntax tree once. Sometimes the type info flows top to bottom and sometimes it flows bottom up. If your type inference algorithm works by traversing the syntax tree, I suggest reading more about bidirectional type inference to get a better idea of how to best coreograph when the type info goes up and when it goes down.
Hindley-Milner type inference works by solving constraints. First you go through the code and figure out all the type constraints (e.g. a function call f(x) introduces the constraint that x must have the same type as the argument of f). Then you solve the system of equations, as if you'd solve a sudoku puzzle. This sort of "global type inference" can sometimes figure out the types even if you don't have any type annotations at all. The catch is that some type system features introduce constraints that are hard to solve. For example, in object oriented languages the constraints are inequalities (instanceof) instead of equalities. If you plan to go this route it's worth learning how to make the algorithm efficient and which type system features would be difficult to infer.
> Bidirectional type inference is a type inference style where you traverse the syntax tree once.
Yes, in my language I just build code directly from syntax tree in single pass (with a couple of minor exceptions). No complex machinery for type deduction is involved. So, now I assume it's called bidirectional type inference.
Personally I find what Rust does with possibility to avoid specifying types for variables isn't that great. It allows writing code which is hard to read, since no type information is present in source code. Also I suppose that it makes compilation slower, since solving all these equations isn't so easy and computationally-cheap.