[goal] modifying private method in impl should not require users of public methods to recompile

[nikomatsakis]

One of the goals for the first alpha is that if you change the body of a private, inherent method, it should not cause the callers of public methods to recompile, nor to require type-checking.

Currently, however, it does. There are two reasons for this. Fixing either of these would fix the immediate problem described in this issue. Fixing both would be ideal.

• Separate fn signature from fn body in HIR #35078 -- method body and signature are not separated
  • this means that changing the body of a private method is considered to potentially change its signature
• incr. comp.: Methods should have their own dep-graph nodes #36349 -- methods do not get their own dep-node, just impls
  • this means that changing the body of a private method also contaminates the entire impl, which in turn contaminates the list of inherent methods

[nikomatsakis]

There are test cases checked in that reference this issue with FIXME(#37121) markers (or will be shortly).

[petrochenkov]

Does this goal apply to methods using impl Trait?

struct S;

impl S {
    pub fn public() -> impl Clone {
        S::private()
    }
    fn private() -> impl Clone {
        0u8
    }
}

[steveklabnik]

One of the goals for the first alpha

To be clear, this is about goals for incremental recompilation?

[nikomatsakis]

@petrochenkov

Does this goal apply to methods using impl Trait?

No, clearly not. I didn't mean this in a blanket way ("no possible change!"). But I meant "most of the time."

---

@steveklabnik

To be clear, this is about goals for incremental recompilation?

right.

[nikomatsakis]

@petrochenkov but good point, that'd make a nice test case :)