ssize has not been declared in std

[ned14]

quickcpplib/include/quickcpplib/gsl-lite/include/gsl/gsl-lite.hpp:984:12: error: ‘ssize’ has not been declared in ‘std’
  984 | using std::ssize;
      |            ^~~~~

Cause:

#define gsl_HAVE_STD_SSIZE              ( gsl_COMPILER_GNUC_VERSION >= 1000 )

However ssize() is a C++ 20 only feature, not a GCC 10 feature. So this is the wrong check to do.

[mbeutel]

Thank you. std::ssize() support was added to libstdc++ only for GCC 10, so the version check is necessary, but we additionally need to check the language version, i.e. #if __cplusplus > 201703L. (GCC sets __cplusplus=201709L for tentative C++20 support.)

std::ssize() has since appeared in libc++ and in the MSVC STL as well, so this check will need an update anyway.

Edit: There is the __cpp_lib_ssize feature test macro but libstdc++ GCC 10 doesn't define it yet: https://gcc.godbolt.org/z/_JSaYy

[ned14]

Me personally, I would write a local ssize() implementation which prefers a std::ssize(), uses that if available, otherwise falls back to an emulation. You can see an example of such a technique for std::launder() at:

https://github.com/ned14/quickcpplib/blob/master/include/quickcpplib/start_lifetime_as.hpp#L33

[mbeutel]

That technique is prone to ADL hijacking:

struct Evil : std::array<int, 1>{ };
constexpr int ssize(Evil) { return 2; }
static_assert(gsl::ssize(Evil{ }) == 1);

Also, unlike std::launder() which requires compiler magic, gsl::ssize() has a straightforward portable implementation; what would be the benefit of conditionally delegating to std::ssize()?

[ned14]

I don't think that technique is prone to ADL hijacking. For something like:

namespace std_or_emulation
{
  using namespace std;
  template <class C>
  constexpr auto ssize(const C& c, ...)  // inferior overload 
    -> std::common_type_t<std::ptrdiff_t,
                          std::make_signed_t<decltype(c.size())>> 
  {
      using R = std::common_type_t<std::ptrdiff_t,
                                   std::make_signed_t<decltype(c.size())>>;
      return static_cast<R>(c.size());
  }
}
template<class T> constexpr auto ssize(const T&v) { return std_or_emulation::ssize(v); }

Here it would select the best available overload in namespace std_or_emulation. You deliberately make your emulation an inferior overload choice valid if and only if namespace std doesn't provide a better one.

ADL never comes into play here, because all lookup is namespace-qualified.

[mbeutel]

Right. But please note that the example you referenced is prone to hijacking because you don't use a qualified name there.

Nevertheless, what is the benefit of delegating to std::ssize()?

[ned14]

Nevertheless, what is the benefit of delegating to std::ssize()?

To get early notification of when your standard library implementator has done something weird, mainly. Otherwise, none.

[dskachan]

Hello everyone.
I have setup:
Ubuntu 20.04, gcc 10.2.0; cmake 3.18.0; conan 1.27.1. what else... ccache 3.7.7

I have a project, which uses -std=gnu++17 flag. so, 17th standard. someday we will move to 2a, but now not about it.
The project uses gsl-lite/0.36.0 over conan.

upon compilation, I'm getting mentioned above error:

/root/.conan/data/gsl-lite/0.36.0/_/_/package/5ab84d6acfe1f23c4fae0ab88f26e3a396351ac9/include/gsl/gsl-lite.hpp:974:12: error: ‘ssize’ has not been declared in ‘std’
  974 | using std::ssize;

@mbeutel mentioned that std::ssize_t is for gcc 10+; and standard 2a+. so, it should not include std::ssize_t, right?
what am I doing wrong?

cmake generated me such a line:

ccache /usr/local/bin/c++ -DGSL_THROW_ON_CONTRACT_VIOLATION -D_GLIBCXX_USE_CXX11_ABI=1  -isystem /root/.conan/data/gsl-lite/0.36.0/_/_/package/5ab84d6acfe1f23c4fae0ab88f26e3a396351ac9/include ####### MANY OTHER INCLUDES ##### -fdiagnostics-color=always -g -fPIC -Wall -Wextra -Wshadow -Winit-self -Wfloat-equal -Wtrigraphs -Wconversion -Wcast-align -Wlogical-op -Wunsafe-loop-optimizations -Wno-unused-parameter -Wdouble-promotion -Wpointer-arith -Wformat=2 -pthread -std=gnu++17 -MD -MT src/CMakeFiles/project.dir/Progress.cpp.o -MF src/CMakeFiles/project.dir/Progress.cpp.o.d -o src/CMakeFiles/project.dir/project.cpp.o -c ../src/Progress.cpp

[mbeutel]

The project uses gsl-lite/0.36.0 over conan.
[...]
what am I doing wrong?

The fix for this issue was committed in 913e86d which found its way into version 0.37, released May 13, 2020.

The latest version of gsl-lite in Conan Center seems to be 0.36.0. I guess one could open an issue in their repo to have the package updated.

[dskachan]

Sorry for off-topic
is tag 0.37.0 intentionally without "v" in the name?

[mbeutel]

is tag 0.37.0 intentionally without "v" in the name?

No, that was unintentional. Thanks for bringing this up; I just pushed an additional "v0.37.0" tag.