Re: [patch] Remove invalid test on non-const empty strings

[chris jefferson <caj at cs dot york dot ac dot uk>]

Paolo Carlini wrote:

> Chris Jefferson wrote:
>
> > As we are testing empty strings, size()==0. Therefore we can only 
> > perform the operation "s[0]" where s is the empty string when s is 
> > const (when s[0] will return charT()). When s is non-const the 
> > operation is undefined. The test tries to access s[0] on an empty 
> > string s on both const and non-const strings. Therefore this patch 
> > fixes things up.
>
>
> I see your point, but I think the issue is somewhat debatable, and 
> much more general, actually. In this sense: according to 21.3.4, as 
> you point out above, "the behavior is undefined". Now, it happens, 
> that, *given our current implementation*, if things are not broken, we 
> mus return '\0' in this case too.
>
> Therefore, the general issue is the following: should our testsuite 
> only test what the standard mandates or, when appropriate, request, so 
> to speak, "more" from the code, for instance specific behaviors when 
> the behavior is strictly speaking undefined? I don't think we have a 
> very rigorous policy about this issue and, frankly, being an 
> implementor that cares about *regressions* and *unexpected* changes of 
> behavior, that is not being a seller of a "conformancy testsuite", I 
> think the latter interpretation has its own advantages, in some cases.
>
> What do other people think?
Good point!

Perhaps then a better idea would be to wrap tests like this in a #ifndef 
_GLIBCXX_DEBUG / #endif pair?