C++ PATCH: PR 20599 (1/3)

[Douglas Gregor]

[re-sending to the list]

On Sep 24, 2006, at 2:22 PM, Mark Mitchell wrote:
> * Therefore, building on the suggestions in this thread, my  
> suggestion is that we:
>
> (a) add C++0x features only with a command-line option (off by  
> default, for now) so that users have to explicitly request the  
> features,
>
> (b) document that option in the manual as enabling experimental  
> features and warning people that C++0x is subject to change, and  
> that the GCC Gimplementation will track the actual standard,  
> without regard for backwards compatibility with previous GCC releases,
>
> (c) refrain from adding a feature until it is actually part of the  
> WP, so that "feature oscillation" is minimized.
>
> What do people think of that suggestion?

This is a reasonable policy.

One part I'm not clear about is whether you envision this to be a  
single flag (e.g., -std=c++0x, -std=gnu++0x) or whether it should be  
multiple flags (-experimental-rvalue-ref, -experimental-decltype,  
etc.). Personally, I have a strong preference for the single-flag  
approach, for two reasons:
   (1) It doesn't force the user to think about which features s/he  
is enabling to "try out a C++0x compiler." One flag enables all that  
GCC supports; the documentation and feature-test macros tell us what  
those features are.
   (2) All of these features have to work together anyway. If they  
don't work together in the language, we need to know early; if they  
don't work together in the compiler, they don't belong in an FSF  
release.

Since the policy we're agreeing on excludes variadic templates from  
consideration, I'd like to see this patch discussed on its own  
merits. Variadic templates are a huge win for libstdc++. They take  
the ugliest parts of libstdc++, the TR1 function objects and tuple  
implementations, and improve them greatly, both for libstdc++  
maintainers and for users. We can eliminate more than 50k of  
repetitive code from libstdc++, replacing a barely-maintainable mess  
of preprocessor metaprogramming with a simple, concise implementation  
that compiles in less than 1/10th of the time. Moreover, the variadic- 
template--enabled libstdc++ provides a better user experience. The  
current implementation limits the number of parameters to function/ 
bind/mem_fn/etc. to 10; try to use 11 parameters and you get a big,  
ugly error message. We can't even bump up the maximum number of  
parameters supported by the library (we've tried), because compile  
times become prohibitive.

GCC already contains other extensions meant to improve libstdc++,  
such as extern templates and strong using directives. The former  
helps improve compile times (among other benefits) and has now become  
part of the C++0x working draft. The latter helps improve the user  
experience for libstdc++'s debug mode. Varadic templates do both, and  
drastically simplify one of the worst parts of libstdc++.

   Cheers,
   Doug