Segfault with one(::Type) and Base.checked_add on Julia 1.0

[sostock]

The following code segfaults on Julia 1.0.4 (it works on 1.1.1 and 1.2.0-rc1):

struct HalfInt <: Real
    twice::Int8
    function HalfInt(x)
        i = Int8(x)
        twice = Base.checked_add(i, i)
        new(twice)
    end
end

convert(HalfInt, 1) # this works
one(HalfInt) # this segfaults, even though it should just call convert(HalfInt, 1)

The segfault also happens when Int8 is replaced by Int16, UInt8 or UInt16 (it works correctly for the other integer types). The segfault also doesn’t happen when Base.checked_add is replaced by +.

[KristofferC]

Would be good if someone could do a bisect and find the commit that fixed it. Then we might be able to backport it to the next 1.0 release.

[sostock]

I did a bisect, the relevant commit is 500d669:

500d669e84cad3ce8d1af90e79c9390396a464c7 is the first new commit
commit 500d669e84cad3ce8d1af90e79c9390396a464c7
Author: Jameson Nash <vtjnash@gmail.com>
Date:   Wed Oct 17 17:22:44 2018 -0400

    runtime: fix definition of checked-intrinsics
    
    Handle case where runtime_nbits is not a power of two,
    and be more careful to avoid C compiler undefined behavior
    and tricky promotion rules when defining these.

:040000 040000 fbc7f85db5b6626a19764fc0d3296f03f1c6016b 90610b2bc79f343ab771494088e2e48c38b0d521 M      src

This is part of #29721, which apparently cannot be backported so easily? At least that’s how I interpret the discussion there.

[vtjnash]

Closing as this was only a 1.0.x issue