@vitaut , do you have any idea why mingw fails and VS is fine? I don't have mingw environment at the moment

What I could do blindly is to catch std::runtime_error and use std::locale("") in that case.

Thanks for the PR. According to this thread:

The GCC's libstdc++ does not support locales other than the C locale on MinGW. Somebody has to implement it first.

So I changed it back to C locale but added a detection for empty lconv in 3cefa72. It would be even better to detect lconv::thousands_sep is there with SFINAE, but this simple solution should do for now. Does it work for you?

Hi, I have the same issue, but your fix (3cefa72) does not work. It's choosing the default thousands_sep:

In file included from fmt/posix.h:48:0,
                 from fmt/posix.cc:33:
fmt/format.h: In function 'fmt::StringRef fmt::internal::thousands_sep()':
fmt/format.h:915:60: error: 'struct lconv' has no member named 'thousands_sep'
 fmt::StringRef thousands_sep() { return std::localeconv()->thousands_sep; }

Anything I can do to help?

@Gachapen Thanks for testing it. Could you check what is sizeof(lconv) on Android? I guess the size check is not very reliable, maybe I should use a proper SFINAE detection for thousands_sep.

@vitaut Just checked, it is 1.

@Gachapen Thanks. Just realized that my "fix" is rubbish because the thousands_sep template is compiled anyway.

@vitaut Hehe, yes I see that now too :)

This is a bit too complex for me (I'd just #ifdef it). I'll happily test any fix you can come up with.

I will do the proper SFINAE version right now.

SFINAE with c++98 looks really messy... What do you think about this approach? dpantele@998bd2c

@dpantele Definitely looks better but I wonder why the tests are failing on mingw.

@vitaut I haven't pushed changes to the pull request branch, now all builds pass.

I can confirm that this now builds. Though is it safe to assume that the size should be less or equal to 1 when it has no members?

@dpantele Got it. Could you squash the commits and rebase to current master?

@Gachapen You are right, the size check is not very robust. A better option is to detect lconv::thousands_sep specifically using something like http://stackoverflow.com/a/1007175/471164, but it will require more work.

Ok, rebased and cleaned-up in 45a1509

Merged, thanks!