usize/isize vector comparison within function

[travisstaloch]

on 30 Jan 2020 with zig version 0.5.0+7ebc624a1 pulled and built from master last night.

In the following code, two @Vector(4, usize) are compared with == in the test scope with no problem. However if the same comparison is made within a struct fn, the == comparison results in:

error: expected type '@Vector(4, u64)', found '@Vector(4, usize)'

It appears that the lhs @Vector(4, usize) is being converted to @Vector(4, u64) when it gets passed into a function. This error is not observed with u64 vector type.

//
// $ zig test repro.zig
// ./repro.zig:15:29: error: expected type '@Vector(4, u64)', found '@Vector(4, usize)'
//             const eqs = self.limbs == other.limbs; // <--- here
//                             ^
// ./repro.zig:15:36: note: referenced here
//             const eqs = self.limbs == other.limbs; // <--- here
//                                    ^
//

test "repro: vector == with usize/isize" {
    const S = struct {
        limbs: @Vector(4, usize), // this error also occurs with isize
        pub fn equals(self: @This(), other: @This()) bool {
            // error: expected type '@Vector(1, u64)', found '@Vector(1, usize)'
            const eqs = self.limbs == other.limbs; // <--- here
            return true;
        }
    };
    const s1 = S{ .limbs = @splat(4, @as(usize, 0)) };
    const s2 = S{ .limbs = @splat(4, @as(usize, 0)) };
    const eqs = s1.limbs == s2.limbs; // here == with usize vectors is ok
    const eq = s1.equals(s2);
}

[FireFox317]

Maybe it has something to do with the fact that @This() is called twice? A pattern that is common is put a definition in the struct: const Self = @This() and use Self instead.

[LemonBoy]

The code that lowers the comparison between two vectors uses @Vector(4, u64) as type for the intermediate computations since it's compatible (size-wise and signedness-wise) with both the LHS and the RHS. Too bad that you cannot (yet?) coerce a @Vector(4, usize) to a @Vector(4, u64) even though usize is equal to u64. The error message reflects this implementation detail.

[travisstaloch]

@FireFox317 I tried with const Self = @This() but removed it trying to make minimal repro.
@LemonBoy any ideas where to start on making the coercion happen? I might attempt myself.

[LemonBoy]

any ideas where to start on making the coercion happen? I might attempt myself.

You should check first with @andrewrk if that's supposed to work. Check out ir_analyze_bin_op_cmp_numeric to see where this happens and how dest_scalar_type is evaluated.

[travisstaloch]

Too bad that you cannot (yet?) coerce a @vector(4, usize) to a @vector(4, u64) even though usize is equal to u64

@andrewrk should this coerce?

Related to: #903 and #3641

[andrewrk]

Yes that should coerce. As long as each of the element values coerce and the length matches then the whole vector should.

A workaround is @bitCast.