Prevent writing out of bounds when final column is entirely NA

[TomAndrews]

If a column is entirely NA, then firstNonNACol will return the index
of the first entry of the next column. In the case where the final
column is entirely NA then this will result in writing off the end of
the allocated result.

Fix by subtracting one from the index in this case.

Hopefully fixes #234

[joshuaulrich]

Thanks for the report! I'll investigate.

In the future, please wait for feedback before opening a PR. The rationale for this is in the contributing guide. The change is small in this case, but changes in C code are less likely to be "trivial" than changes in R code.

[joshuaulrich]

Also, the unit test I added to your issue fails for your branch on "character" vectors. I suspect you simply missed that branch in the code. There should also be tests for calls where fromLast = TRUE.

It would be great if you add these tests to your PR and fix whatever else they uncover. Then please use git commit --amend to update your PR commit (instead of adding new commits). You'll also need to add --force to your push command.

[TomAndrews]

It turned out the problem was in firstNonNACol - I don't think the indexing was correct in the STRSXP case.

I've put my change in a different commit so it's easy for you to see but can squash and push if you prefer.

[joshuaulrich]

Yeah, it looks like you are correct. And I agree that it should be in a different commit in this case, since it's a different problem. I would only have squashed if the new commit only added another if (_first == nr + j*nr) _first--;.

[joshuaulrich]

I've merged your other PR, so go ahead and rebase this one when you have a chance. I can do it locally, but when I've done that in the past GitHub has marked the PR as "closed" instead of "merged"... and I'd prefer you get a "thumbs up".