fix bug in hlist.{Intersection, Union}

[aryairani]

Fix bugs described in #562.

Currently Intersection is witnessed by three cases:
hnilIntersection says ∅ ∩ M = ∅, which is good.
hlistIntersection1 says (H :: T ∩ M) = T ∩ M
hlistIntersection2 says: if H ∈ M, (H :: T ∩ M) = H :: (T ∩ (M - R))

Unfortunately, (H :: T ∩ M) = T ∩ M is false if H ∉ M. This patch adds the H ∉ M condition to hlistIntersection1 using a FilterNot.Aux[M,H,M] witness.

The patch also fixes an equivalent bug in Union.

---

I had a couple additional questions:

1. Should hlistIntersection1 and hlistIntersection2 reside in the same trait now, eliminating the need for LowPriorityIntersection?
2. Why is it that Intersection extends DepFn1 but Union extends DepFn2? It seems like they should be the same, but I don't know what the right answer is.

[aryairani]

Fails mima check — this is tricky, the patch not bincompat with existing, but existing is buggy. :-\

[OlivierBlanvillain]

Hi @refried, thanks for fixing this, your changes are LGTM!

Regarding your 2nd question, I think A ∩ B only contains elements from A, while A U B could have elements from both A and B, which would explain why you don't need the same number of arguments to produce the output. Note that for A - B, the type of B is also sufficient, as the output will only contain values from A.

[alexarchambault]

About MiMa, it is possible keep binary compatibility. You just have to keep the former methods as is, possibly modifying a bit their implementation, but removing their implicit modifier. (All of this along side the new methods you added)

This way, code compiled against older versions of shapeless will still find the methods they're looking for at runtime, and code compiled against the newer version will only find the newest methods, which are the only ones flagged as implicit.

[milessabin]

+1 to @alexarchambault's advice ... I'll merge when MiMa passes ... thanks! :-)

[aryairani]

@alexarchambault @milessabin Ok, I can make that MiMa change, but please advise/confirm: the binary-compatible version will necessarily continue to be buggy, which is pretty scary to me. (The fix was to add the additional witness.) My instinct to add a @deprecation warning won't help if there's no recompilation.

Do I add @deprecation for now we delete the buggy method in the next minor version? Is there a process for remembering to have that happen? Thanks!

[alexarchambault]

@refried As the methods kept for binary compatibility don't have the implicit modifier anymore, they won't be found when building implicit instances of Intersection or Union. Only code compiled against older versions of shapeless should use them.

I don't mind about adding @deprecated. These methods are seldomly directly called anyway...

[aryairani]

@alexarchambault Sure, I understand that — I just wished there was a way to warn the person linking against the old version; but I guess there isn't, and they wouldn't discover a bincompat issue until runtime.

I pushed the update, so fingers crossed! 😄

[milessabin]

LGTM. Many thanks ... merging :-)