2011/11/21 Paolo Carlini <paolo.carl...@oracle.com>: > On 11/21/2011 01:22 AM, Jason Merrill wrote: >> >> This bug fix required the following change to >> libstdc++-v3/include/std/tuple: >> >>> - template<typename... _UElements, typename = typename >>> - enable_if<__and_<integral_constant<bool, sizeof...(_UElements) >>> - == sizeof...(_Elements)>, >>> - __all_convertible<__conv_types<_UElements...>, >>> - __conv_types<_Elements...>> >>> - >::value>::type> >>> + template<typename... _UElements, >>> + typename = typename enable_if<sizeof...(_UElements) >>> + == >>> sizeof...(_Elements)>::type, >>> + typename = typename >>> + enable_if<__all_convertible<__conv_types<_UElements...>, >>> + __conv_types<_Elements...> >>> >::value >>> + >::type> >> >> __and_ doesn't have the shortcut semantics of &&, so checking for equal >> sizeof... didn't prevent us from trying the __all_convertible test. > > Ah!! So, now I'm puzzled: why are we using all those __and_ and __or_ in the > first place, if isn't because we want to avoid to avoid some instantiations? > Daniel can you remind me that?
AFAIK in my actual __all_convertible implementation, the second test was protected by lazy __and_ usage, so this could not happen. I assume that the defect occurred, when you adapted it to the existing structure. But of-course the more general __and_<is_something<_UElements, _Elements>...> expansion is preferred now. - Daniel
