> On Jun 15, 2017, at 22:22, Eric Fiselier <e...@efcs.ca> wrote: > > > >> On Thu, Jun 15, 2017 at 11:00 PM, Duncan P. N. Exon Smith >> <dexonsm...@apple.com> wrote: >> Your suggestion is essentially to replace experimental/string_view with >> something like: >> >> namespace std { inline namespace __1 { namespace experimental { >> template <class CharT> >> using basic_string_view = _VSTD::basic_string_view; >> }}} >> >> That breaks: >> 1. User compiles 1.cpp with older toolchain. 1.cpp implements >> foo(std::experimental::string_view). >> 2. User compiles 2.cpp with newer toolchain. 2.cpp calls >> foo(std::experimental::string_view). >> 3. User links 1.o with 2.o. >> >> I'm not sure if this matters. > > It can't matter. <experimental/foo> are allowed to break both their API and > ABI as needed. > > Also I was suggesting > > namespace std { namespace experimental { > using std::basic_string_view; > using std::string_view; > }} > > This approach will break code that expects experimental::string_view and > std::string_view are different types: > Example: > > void foo(std::string_view); > void foo(std::experimental::string_view); > foo(std::string_view{}); // ambiguous > >> >>> On Jun 15, 2017, at 21:55, Eric Fiselier <e...@efcs.ca> wrote: >>> >>> I would also want to do serious performance analysis on this patch. Does >>> removing the string_view overloads cause less optimal overloads to be >>> chosen? Perhaps allocating ones? >>> That would be really unfortunate, and I'm not sure that's in the best >>> interest of our users at large. >> >> Less optimal compared to what? C++17 code? > > Not sure yet, I'm trying to figure out what types the `const Tp&` overloads > are attempting to soak up. Is it only string_view?
The type trait restricts it to things convertible to string_view that are not const char *. > > >> >>> /Eric >>> >>>> On Thu, Jun 15, 2017 at 10:51 PM, Duncan P. N. Exon Smith >>>> <dexonsm...@apple.com> wrote: >>>> >>>>> On Jun 15, 2017, at 19:42, Eric Fiselier <e...@efcs.ca> wrote: >>>>> >>>>> >>>>> >>>>> On Thu, Jun 15, 2017 at 8:38 PM, Duncan P. N. Exon Smith >>>>> <dexonsm...@apple.com> wrote: >>>>>> I just started working on a patch to add #if guards, and the first >>>>>> interesting thing I found was the basic_string constructor: >>>>>> >>>>>>> template <class _CharT, class _Traits, class _Allocator> >>>>>>> template <class _Tp> >>>>>>> basic_string<_CharT, _Traits, _Allocator>::basic_string( >>>>>>> const _Tp& __t, size_type __pos, size_type __n, const >>>>>>> allocator_type& __a, >>>>>>> typename >>>>>>> enable_if<__can_be_converted_to_string_view<_CharT, _Traits, >>>>>>> _Tp>::value, void>::type *) >>>>>>> : __r_(__second_tag(), __a) >>>>>>> { >>>>>>> __self_view __sv = __self_view(__t).substr(__pos, __n); >>>>>>> __init(__sv.data(), __sv.size()); >>>>>>> #if _LIBCPP_DEBUG_LEVEL >= 2 >>>>>>> __get_db()->__insert_c(this); >>>>>>> #endif >>>>>>> } >>>>>> >>>>> >>>>> That constructor was added in C++17, so removing it along with >>>>> string_view should be OK. >>>>> Assuming we don't use it to implement older constructors using a single >>>>> template. >>>>> >>>>> >>>>>> I suppose the decision was made so that std::string could take advantage >>>>>> of it. >>>>>> >>>>>> Is it a conforming extension? >>>>> >>>>> No, because it can change the meaning of otherwise well defined code, as >>>>> you pointed out initially. >>>> >>>> Let me know if this patch is along the right lines. If so, I'll finish it >>>> up and put it on phab. >>>> >>>> experimental/filesystem/path.cpp doesn't compile, since >>>> experimental/filesystem uses things like operator+=(string, string_view) >>>> extensively. But I'd like an early opinion on the approach before I dig >>>> in. >>>> >>>> In string, the only function that needed to be rewritten was >>>> string::compare(size, size, string, size, size). I'm nervous that >>>> filesystem will be a bigger job. >>>> >>>> >>>> >>>>> >>>>>> >>>>>>> On Jun 15, 2017, at 18:35, Eric Fiselier <e...@efcs.ca> wrote: >>>>>>> >>>>>>> It *shouldn't* include <string_view>, that's a given. >>>>>>> >>>>>>> IIRC, and Marshall would know better, I believe it was untenable to >>>>>>> maintain a version of <string> that didn't depend on <string_view> >>>>>>> after making >>>>>>> the changes required for C++17. >>>>>>> >>>>>>> However inspecting <string> now it does seem possible that the >>>>>>> entanglement >>>>>>> is avoidable.Though it's also likely I'm just not seeing the whole >>>>>>> picture. >>>>>>> >>>>>>> /Eric >>>>>>> >>>>>>> On Thu, Jun 15, 2017 at 6:42 PM, Duncan P. N. Exon Smith >>>>>>> <dexonsm...@apple.com>wrote: >>>>>>>> >>>>>>>> > On Jul 20, 2016, at 22:31, Marshall Clow via cfe-commits >>>>>>>> > <cfe-commits@lists.llvm.org> wrote: >>>>>>>> > >>>>>>>> > Modified: libcxx/trunk/include/string >>>>>>>> > URL: >>>>>>>> > http://llvm.org/viewvc/llvm-project/libcxx/trunk/include/string?rev=276238&r1=276237&r2=276238&view=diff >>>>>>>> > ============================================================================== >>>>>>>> > >>>>>>>> > @@ -435,6 +461,7 @@ basic_string<char32_t> operator "" s( co >>>>>>>> > */ >>>>>>>> > >>>>>>>> > #include <__config> >>>>>>>> > +#include <string_view> >>>>>>>> >>>>>>>> This breaks the following, valid, C++14 code: >>>>>>>> >>>>>>>> #include <string> >>>>>>>> #include <experimental/string_view> >>>>>>>> using namespace std; >>>>>>>> using std::experimental::string_view; >>>>>>>> void f() { string_view sv; } >>>>>>>> >>>>>>>> Should <string> #include <string_view> even when we're not in C++17 >>>>>>>> mode? Why? >>>>>>>> >>>>>>>> > #include <iosfwd> >>>>>>>> > #include <cstring> >>>>>>>> > #include <cstdio> // For EOF. >>>> >>>> >>> >> >
_______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits