Re: [PATCH] Hashtable PR96088

["François Dumont" <frs.dumont@gmail.com>]

On 20/05/21 6:44 pm, Jonathan Wakely wrote:
> On 06/05/21 22:03 +0200, François Dumont via Libstdc++ wrote:
>> Hi
>>
>>     Considering your feedback on backtrace in debug mode is going to 
>> take me some time so here is another one.
>>
>>     Compared to latest submission I've added a _Hash_arg_t partial 
>> specialization for std::hash<>. It is not strictly necessary for the 
>> moment but when we will eventually remove its nested argument_type it 
>> will be. I also wonder if it is not easier to handle for the 
>> compiler, not sure about that thought.
>
> The std::hash specializations in libstdc++ define argument_type, but
> I'm already working on one that doesn't (forstd::stacktrace).
>
> And std::hash<acme::ProgramDefinedType> can be specialized by users,
> and is not required to provide argument_type.
>
> So it's already not valid to assume that std::hash<T>::argument_type
> exists.

Yes, I know that the plan is to get rid of argument_type. But as long as 
it is there we can still use it even if I didn't realize that you were 
already in the process of removing it so soon.


>
>> @@ -850,9 +852,56 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>>     iterator
>>     _M_emplace(const_iterator, false_type __uks, _Args&&... __args);
>>
>> +      template<typename _Kt, typename _Arg, typename _NodeGenerator>
>> +    std::pair<iterator, bool>
>> +    _M_insert_unique(_Kt&&, _Arg&&, const _NodeGenerator&);
>> +
>> +      // Detect nested argument_type.
>> +      template<typename _Kt, typename _Ht, typename = __void_t<>>
>> +    struct _Hash_arg_t
>> +    { typedef _Kt argument_type; };
>> +
>> +      // std::hash
>> +      template<typename _Kt, typename _Arg>
>> +    struct _Hash_arg_t<_Kt, std::hash<_Arg>>
>> +    { typedef _Arg argument_type; };
>> +
>> +      // Nested argument_type.
>> +      template<typename _Kt, typename _Ht>
>> +    struct _Hash_arg_t<_Kt, _Ht,
>> +              __void_t<typename _Ht::argument_type>>
>> +    { typedef typename _Ht::argument_type argument_type; };
>> +
>> +      // Function pointer.
>> +      template<typename _Kt, typename _Arg>
>> +    struct _Hash_arg_t<_Kt, std::size_t(*)(const _Arg&)>
>> +    { typedef _Arg argument_type; };
>> +
>> +      template<typename _Kt,
>> +           typename _ArgType
>> +         = typename _Hash_arg_t<_Kt, _Hash>::argument_type>
>> +    static typename conditional<
>> +      __is_nothrow_convertible<_Kt, _ArgType>::value, _Kt&&, 
>> key_type>::type
>
> Please use __conditional_t<...> here instead of
> typename conditional<...>::type.
>
> The purpose of the _Hash_arg_t type is to determine whether invoking
> the hash function with _Kt&& can throw, right?

No, the purpose of _Hash_arg_t is to find out what is the argument type 
of the _Hash functor. With this info I can check if invoking that 
functor is going to instantiate a temporary using a throwing operation. 
If so, I create a temporary at _Hashtable code level and move it to its 
final storage place when needed.

The tricky part is that _Hash can accept different argument types, for 
the moment I just do not create a temporary in this case.

>
> And if it can throw, you force a conversion early, and if it can't,
> you don't do the conversion.
>
> Can't you use __is_nothrow_invocable<_Hash&, _Kt> for that, instead of
> this fragile approach?
>
I think I already try but I'll check.

I fear that __is_nothrow_invocable<_Hash&, _Kt> tells if the chosen 
operator()(const _Arg&) is noexcept qualified. Not if the conversion 
from _Kt to _Arg is noexcept.