[PATCH] D22392: [Sema] Fix an invalid nullability warning for binary conditional operators

[Akira Hatanaka via cfe-commits]

ahatanak added a comment.

I'm thinking about taking the following steps to compute the nullability of ternary operators.

The first step is to compute the "merged" nullability of the LHS and RHS.

For normal ternary operators (not the shorthand version):

- If both LHS and RHS have the same nullability (one of nonnull, nullable, null_unspecified, or none), pick that as the merged nullability.
- Otherwise, if either side is nullable, the merged nullability is nullable too.
- Otherwise, if either side is nonnull, pick the nullability of the other side. For example, if (LHS,RHS) == (nonnull, none), the merged nullability is none because we can't guarantee that the result is nonnull.
- Otherwise, the nullability is either null_unspecified or none. I think we can pick either one in this case.

For binary conditional operators like "p0 ?: p1":

- If the LHS (p0) is nonnull, the merged nullability is nonnull.
- Otherwise, the merged nullability is the nullability of the RHS (p1).

Once we have the merged nullability, the next step is to compare it with the result's nullability. If the nullability kinds differ, a new type that has the merged nullability is created.

Note that it looks like we might be able to simplify the steps described above if we take advantage of the fact that clang only warns when a nullable pointer is assigned/passed to a nonnull pointer (I don't think clang issues any warnings, for example, when a null_unspecified pointer is assigned to a nonnull pointer).


https://reviews.llvm.org/D22392