New antisymmetrical debug check

[François Dumont]

On 22/10/2015 21:41, Daniel Krügler wrote:
> 2015-10-22 21:18 GMT+02:00 François Dumont <frs.dumont@gmail.com>:
>> Hi
>>
>>     I leveraged on predefined operators to inject a new debug check for
>> strict weak ordering operators. It checks that for a given functor f if
>> f(a, b) then !f(b, a). In terms of < operator: if (a < b) assert(!(b < a)).
>>
>>     I had to replace the generic functors like _Iter_comp_iter by more
>> explicit _Iter_less_iter where this check can take place. As algos can
>> be used with different iterator types those checks are only performed in
>> C++11 and after to make sure that b < a is a valid expression.
> [..]
>> Tested under Linux x86_64.
>>
>> Ok to commit ?
> I'm wondering: Why does the _M_reverse_check overload set require
> _GLIBCXX14_CONSTEXPR and not _GLIBCXX11_CONSTEXPR? None of them is
> written in a way that it could not be a valid C++11 constexpr
> function?

Because it only needs _GLIBCXX14_CONSTEXPR, this is to make them usable
for std::min, std::max, std::minmax, and some others.

>
> I believe that operator()(_Iterator1 __it1, _Iterator2 __it2) could
> easily be adapted to be also a valid C++11 constexpr function by
> replacing the if-directive by an conditional expression and please
> note that your __cplusplus tests are always C++11 compatible.
>
> On the other hand in line 311 of the patch you define a
> _GLIBCXX14_CONSTEXPR operator(), but that one seems to invoke
> _M_reverse_check functions where none of them is constexpr, so this
> operator() could never be constexpr. Note also that
> _Iter_less_val_comp on line 306 doesn't have a constexpr constructor,
> so there is no way to invoke the constexpr operator() member.
>
> It seems as if a similar problem exists for line 377 and the following
> _M_reverse_check overloads.
>
> Comparing to these two, why is in line 136 and line 187 the operator()
> template and it's _M_reverse_check not constexpr as well?

You are right, I haven't been careful enough with those. So I finally
consider where those operators are used and kept only the mandatory
ones. I kept _GLIBCXX14_CONSTEXPR also on the _M_reverse_check even if
_GLIBCXX_DEBUG mode is not compatible with constexpr. You won't be able
to build a constexpr when you activate it so those specification are not
necessary.

François

-------------- next part --------------
A non-text attachment was scrubbed...
Name: antisymmetrical_check.patch
Type: text/x-patch
Size: 35362 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/libstdc++/attachments/20151025/fa01aaf4/attachment.bin>