naga incorrectly throws multiplication operation overflowed error in expressions involving abstract numerics

[matthew-wong1]

Description
Hi! I think I've found an issue with the way in which naga evaluates certain expressions involving abstract numerics. Naga throws a multiplication error overflowed for the following expression but Tint does not.

const num : i32 = -1 * i32(-2147483648)

Both tint and naga correctly throw that error for the following expression:

const num : i32 = -1 * -2147483648

Extra materials
I've spoken to some of the WGSL spec maintainers, who kindly provided the following explanation as for why the first expression should not throw an error:

"The first example is: -1 * i32(-2147483648)
That's type-checked as a multiply with abstract-int on the left, and i32 on the right.
There is no overload of the multiply operator with those two type parameters. There's no up-conversion from i32 to abstract int, and the only viable automatic conversion is to convert the LHS from abstract int to i32. This yields i32 * i32. It's two-s complement arithmetic and therefore wrapping occurs.

In the second example, -1 * -2147483648
It's a multiply of two abstract-ints. There's an overload for that, so the result produces another abstract int, of value 2147483648.
That value is a const-expression, and does not fit in i32, and that causes the error."

[ErichDonGubler]

We definitely don't have correct behavior yet with automatic casting of literals; we expect to be pretty close once #4400 is implemented, modulo bugs (which, admittedly, will likely be similar to this one).

I won't close this bug as a duplicate, unless @gfx-rs/naga has a strong opinion about it.

[jimblandy]

This actually is not related to type conversions. It's simply that the constant evaluator uses checked_mul for i32 multiplication, but WGSL says that multiplication should just wrap.

[sagudev]

The spec says:

Expressions on concrete integer types that overflow produce a result that is modulo 2**bitwidth

so I think we need to change all checked_ to wrapped_, at least for (i32, i32).

[matthew-wong1]

@sagudev I've now made that change in #6835 😄 I think the only instance I found for checked_mul for (i32, i32) was in constant_evaluator.rs

[sagudev]

I meant other ops like add, or sub.

[matthew-wong1]

Ah my bad, I misunderstood - I'll update my pull request. Thanks!