Throw a different error for `@inbounds [][-1]` when `--check-bounds=yes`

[LilithHafner]

A bounds error of the form x[out_of_bounds_index] = 0 is very different from a bounds error of the form @inbounds x[out_of_bounds_index] = 0. The former throws a bounds error and the latter is UB.

When --check-bounds=yes, I would like @inbounds x[out_of_bounds_index] = 0 to throw an UnsafeBoundsError or somesuch.

This will enable testing that no unsafe @inbounds usage is hit in a test while still allowing bounds errors. For example this currently passes with --check-bounds=yes and is UB by default:

struct OffsetVector <: AbstractVector{Int}
    x::Vector{Int}
    offset::Int
end
Base.axes(x::OffsetVector) = (x.offset:x.offset+length(x.x),)
function Base.getindex(x::OffsetVector, i::Int)
    checkbounds(x, i)
    @inbounds x.x[i-x.offset]
end

x = OffsetVector([1,2,3], 5);
@test_throws BoundsError x[5]
@test_throws BoundsError x[9]

I'd like x[5] to throw an UnsafeBoundsError error and fail the test.

[oscardssmith]

This will be a pain to do, but it seems like a good idea.

[Seelengrab]

This seems difficult to do in the presence of @propagate_inbounds, no?

[jakobnissen]

Surely this is too breaking. Even if there is precedent for changing exception types, BoundsError may just be the most well-defined and recognizable error in all of Julia.

[LilithHafner]

This would only break usage that is UB in julia with default startup options, e.g.:

function f(x, i) 
    try
        @inbounds x[i]
    catch x
        if x isa BoundsError
            println("This is fine.")
            return zero(eltype(x))
        else
            rethrow()
        end
    end
end

How widespread do we think that pattern is?

[jakobnissen]

There is probably a lot of at-test-throw BoundsError in various Julia test suites.