Extend Infer ty for binary operators

[chenyukang]

Fixes #106138

[rustbot]

r? @compiler-errors

(rustbot has picked a reviewer for you, use r? to override)

[chenyukang]

This may a little bit aggressive, my intution is when the operand ty is a Infer var, neg and not may also return a Infer var type. Otherwise we need to check whether an Infer var type support the neg and not operation, which will introduce an error type.

With this change, validate.rs also need to change, codegen is OK. The validation for binary op is introduced by 9ac5e98, without test case, I'm not sure the reason for this.

[compiler-errors]

This inference strategy is incorrect. There's no guarantee that the input and output type of Neg or Not are related, and this causes an ICE in the compiler:

#[derive(Copy, Clone, Default)]
struct A;

struct B;

impl std::ops::Not for A {
    type Output = B;

    fn not(self) -> B { B }
}

fn main() {
    let x = Default::default();
    let y: A = !x;
}

[compiler-errors]

In this example x is constrained to be A because of the inference, but that would imply that y is type B, not A.

[chenyukang]

make sense 👍
this is a trivial difference here, binary op is more eager, so unary will infer type error.

pub fn bools(x: &Vec<bool>) {
    let binary = |i, a: &Vec<bool>| {
        a[i] && a[i+1] // ok
    };

    let unary = |i, a: &Vec<bool>| {
        !a[i] // cannot infer type
    };

    binary(0, x);
    unary(0, x);
}

Seems we can not find proper fix for this scenario?

[compiler-errors]

This validation should not be removed without at least a valid explanation for why it's not necessary anymore 😓

[compiler-errors]

cc @JakobDegen who added this validation

[chenyukang]

It's too late last night, this should be a Draft.
I just found more document for this validations, 6343691

[compiler-errors]

@rustbot author

[bors]

☔ The latest upstream changes (presumably #107267) made this pull request unmergeable. Please resolve the merge conflicts.