std::filesystem::create_directory and std::filesystem::create_directories have inconsistent error handling

[gulrak]

Bugzilla Link	41683
Version	8.0
OS	Linux
CC	@mclow

Extended Description

std::filesystem::create_directories and std::filesystem::create_directory behave differently in case there is a regular file with the name of the directory to create.
While implementing a std::filesystem compatible library myself I found this while running my tests against the libc++ implementation:

std::filesystem::create_directories(p) throws filesystem_error wir "File exists"
std::filesystem::create_directories(p,ec) sets the ec to a corresponding error and returns false

std::filesystem::create_directory(p) reports false
std::filesystem::create_directory(p,ec) reports false and no error in ec

So for this code:

#include <filesystem>
#include <iostream>

namespace fs = std::filesystem;

int main()
{
    std::error_code ec;
    std::cout << fs::create_directory("foo.txt", ec);
    std::cout << (ec ? " error" : " no error") << std::endl;
    ec.clear();
    std::cout << fs::create_directories("foo.txt", ec);
    std::cout << (ec ? " error" : " no error") << std::endl;
}

with an existing text file "foo.txt" the current result is:

0 no error
0 error

(see runnable example here https://wandbox.org/permlink/Wa0Ie40k24rBwGmY)

No matter which of these two behaviours is the correct one, they should behave the same, in regard to this situation.

I know there is LWG issue 2935 (https://cplusplus.github.io/LWG/issue2935) that requested a change to handling this situation as no error, but I'm with Nicolai Josuttis, claiming this is not a good idea (see WG21 P1164R1, approved in Kona and adopted with cplusplus/draft#2703).

So in the end, all four operations should handle the case as an error.

(As additional information: GCCs libstdc++ had this behaviour swapped but inconsistent too, and adopted the P1164R1 solution.)

Thanks for your work!

[mclow]

If you take a look at https://libcxx.llvm.org/cxx2a_status.html, you'll see that neither LWG issue 2935 nor P1164 have been implemented (yet)

[gulrak]

True, and I know that, but my concern is about the -std=c++17 behaviour (and I should have written that, sorry, wandbox resets the language to 2a when switching versions, und I accidently permalinked a 2a one) and I just wanted to give a most complete picture, after I was pointed to 2935 when reporting the similar problem to GCC before we agreed that they should handle the case in the same way.

So this is about, the different behaviour of the two, regarding this situation, and with the reference of the issues I wanted to collect arguments for one or the other solution, just a consistent one.