On 13/07/2020 12.52, Mandeep Sandhu wrote:
>> ...you *might* get similar benefits from a collection of `const string`
>> instead of `string`, but I'm not sure to what extent `string` can be
>> constexpr, especially if you're stuck on C++11.
>
> I don't know. I guess its best to benchmark and see how much
> performance difference I'd get.
>
>> In either case, I wonder if we'll get to (or are already at?) the point
>> that the compiler can more-or-less build the set at compile time such
>> that the run-time construction cost is near-zero...
>
> I wonder if const string objects can point to string literals kept in
> the .rodata section without having the need to allocate for it at
> runtime (or do they already do something similar currently?).

Possibly (but see Thiago's reply). I was thinking about that, but what I
really meant with the previous statement was that the *set*, not just
its values, would be constructed at compile time. IOW, the ctor for the
set is just 'hey, I have size X and my nodes are over there → RODATA'.
(Uh... this probably still needs pointer fix-ups, though, so I guess it
can't actually be in RODATA, but maybe in DATA. And you'd probably need
the elsewhere-mentioned constexpr allocator.)

-- 
Matthew