On Fri, Sep 30, 2016 at 7:10 AM, Bernd Schmidt <bschm...@redhat.com> wrote:
> On 09/30/2016 12:51 PM, Jakub Jelinek wrote:
>> On Fri, Sep 30, 2016 at 12:42:20PM +0200, Bernd Schmidt wrote:
>>> On 09/30/2016 11:45 AM, Jakub Jelinek wrote:
>>>
>>>> See Tom Tromey's explanation why accepting too much is bad (at least
>>>> unless we want multiple levels).
>>>
>>> And I still don't buy it. The case where someone writes "Don't fall
>>> through" is artificial to begin with, and also forgetting to put the break;
>>> there
>>> really isn't something for us to be concerned about.
Agreed.
>> It doesn't have to be exactly that, fall is a substring of 200+ english,
>> thr is a substring of 1000+ english words, if you allow arbitrary stuff in
>> between or before it or after it, it can say anything, completely
>> unrelated to falling through across case labels.
>
> True, but IMO irrelevant. We have to consider what is likely in real code.
> We're trying to catch likely problems without adding a prohibitive cost to
> enabling -Wextra. It's not helpful to punish people for writing code that is
> valid and documents intentional fallthrough, but does it in a style or a
> language we didn't expect.
>
> A complete solution would require working AI; maybe someone from Google can
> contribute one. Failing that, we need heuristics, and I still like Michael's
> suggestion of not printing a warning if we see _any_ comment, on the grounds
> that this would still catch the vast majority of actual errors, without the
> huge false positive rate we currently have.
I think this would lead to too many false negatives, as explanatory
comments before a case label are extremely common. I prefer the
*fall*thr{u,ough}* heuristic.
Jason
