Re: [fixed_string] Designing the contract

From: Alexander Grund (alexander.grund_at_[hidden])
Date: 2019-12-17 07:36:13

On 16.12.19 23:43, Gavin Lambert via Boost wrote:
> This is why, as I mentioned elsewhere in a related context, macros 
> should only ever select between different explicitly-named-differently 
> implementations (such as setting a default for that template policy 
> parameter, but not replacing the template policy parameter).
>
> Otherwise you get ODR violations and inconsistent ABI.  At least when 
> explicitly part of the signature, both implementations can co-exist in 
> the same binary with minimal issues as long as they're kept strictly 
> separate -- and if someone tries to "cross the streams" then they'll 
> get linker errors.
>
> Or, as I also said elsewhere, just pick one and stick to your guns. 
> Personally, I prefer throwing exceptions for that case; it's much 
> safer than allowing a buffer overflow.
+1 to the whole. Having macros to select default template params avoids 
the ODR and mixing issues. HOWEVER: What about you set 
BOOST_USE_THROW=1, then include libraryX then Boost and then use 
`static_string<char/*, default chosen by macro*/>`? libraryX might 
redefine BOOST_USE_THROW=0 and you get the wrong type.
So IMO either template policy (only) or "stick to your guns".
I'm also pro-exception on "overflow". Often enough (especially in hot 
loops where bounds are usually clear) the optimizer will remove all those.
At least, just std::abort.
> +1. Although throw_exception itself is not entirely immune to the 
> problems above, it is at least a customisation point that only the 
> application author is allowed to touch, which is better than a macro.
>
> [Although things get a bit dicier with shared objects and private 
> visibility.]
Oh yes... Damn shared objects. And with boost::throw_exception opens the 
door for ODR violations again...