On Tue, May 7, 2013 at 1:00 PM, John Spicer <jhs@edg.com> wrote:
I'm not personally fond of that direction.
At least for EDG, and I suspect others, the set of feature supported is going to depend on command-line options, which would mean that we'd have to define one set of macros (or something) to allow the <features> header to define another set. It also makes it more likely that you could have a <features> file that doesn't actually match what the compiler does.
Since our names are reserved identifier, an implementation can predefine them and provide a <features> which does nothing.
This allows feature test macros to be provided (a) with no compiler changes, (b) retroactively for an implementation which was released prior to our recommendations (provide a <features> which checks the compiler and version, and use -D__cpp_lib_header_features), and (c) with (essentially) zero cost for programs which do not use the macros. This also happens to be standardizing existing practice ;-)
That said, I agree with Clark that quantifying the cost in (c) would be useful. I have informally heard that this cost is measurable (especially for things like configure scripts which build a large number of tiny programs), but I've never seen numbers.
I don't have a problem with requiring an include of <features> in order to access feature macros.
I would hope, however, that compiler vendors would either recognize that header and then predefine the appropriate macros or have the file contain something like
#pragma define_feature_macros
What I would not like to see, which is actually what existing practice is, is a file that said something like:
#if defined(__GNUC__) && __GNUC__ >= <something> && __GNUC_MINOR__ >= <something>
#define __cpp_has_xxx <something>
#define __cpp_has_xxx <something>
#elif defined(_MSC_VER) && ...
#define __cpp_has_xxx <something>
#define __cpp_has_xxx <something>
... etc ...
So, I agree that this has the advantage of providing a bridge that can be used with implementations released prior to specifying the feature macros, and also has the advantage of having zero cost unless they are used.
But I think to most effectively address the issue, compiler support is required.
To quote an example that Clark provided:
#ifndef __USE_RVALUE_REFERENCES
#if (__GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ >= 3) || \
_MSC_VER >= 1600
#if __EDG_VERSION__ > 0
#define __USE_RVALUE_REFERENCES (__EDG_VERSION__ >= 410)
#else
#define __USE_RVALUE_REFERENCES 1
#endif
#elif __clang__
#define __USE_RVALUE_REFERENCES __has_feature(cxx_rvalue_references)
#else
#define __USE_RVALUE_REFERENCES 0
#endif
#endif
My hope is that our recommendations will eliminate the need for heuristics and rely on a definitive answer from the compiler.
John.
John.
On May 7, 2013, at 3:17 PM, Richard Smith <
richard@metafoo.co.uk> wrote:
> Hi,
>
> Has any thought been given to putting the feature-test macros into an implementation-supplied header, instead of predefining them? This would allow us to remove the cost associated with predefining these macros, for translation units which don't need them. Instead, we could supply a single predefined macro indicating whether the header is available, and user code would write something like:
>
> #ifdef __cpp_lib_header_features
> #include <features>
> #endif
>
> #ifdef __cpp_relaxed_constexpr
> constexpr
> #endif
> size_t strlen(const char *p) { /* ... */ }
>
> ... and so on.
> Features mailing list
> Features@isocpp.open-std.org
> http://www.open-std.org/mailman/listinfo/features
_______________________________________________
Features mailing list
Features@isocpp.open-std.org
http://www.open-std.org/mailman/listinfo/features