-Clink-dead-code doesn't generate dead non-generic items appearing within (instantiated) generic impl blocks

[eggyal]

I tried this code:

struct Foo<T>(T);

impl<T> Foo<T> {
    fn foo(&self) {}
    fn bar(&self) {}
}

fn main() {
    let foo = Foo( () );
    foo.foo();
}

When compiled with -Clink-dead-code, I expected to see Foo::<()>::bar in the resulting binary.

Instead, it is omitted.

Meta

rustc --version --verbose:

rustc 1.47.0-nightly (d7f945163 2020-07-19)
binary: rustc
commit-hash: d7f94516345a36ddfcd68cbdf1df835d356795c3
commit-date: 2020-07-19
host: x86_64-apple-darwin
release: 1.47.0-nightly
LLVM version: 10.0

[eggyal]

Thanks @jonas-schievink—but is this not contrary to what -Clink-dead-code should be doing?

[jonas-schievink]

-Clink-dead-code only links dead code, it doesn't force all possible code to be generated in the first place.

[eggyal]

@jonas-schievink according to @nikomatsakis in #46467 (comment):

Discussed in @rust-lang/compiler meeting. Consensus was that -Clink-dead-code should make compiler generate dead code, as it used to, since there isn't much value in just having the linker preserve dead code (but not having compiler generate it). We considered changing the name but decided that was more trouble than it's worth, and it'd be better to just adjust the -Chelp message to clarify.

Did I misunderstand?

[jonas-schievink]

Hmm, yeah, that looks right.

[eggyal]

Just to clarify a little further what I had expected to see here, as obviously it is not generally reasonable to generate dead generic items (after all, for what types should they be monomorphized?). However, it does seem reasonable to generate non-generic items that appear within generic impl blocks for which other items are being generated.