Thread: GCC regex broken

I am facing some weird bugs with using regex on GCC.

Code:

#include <iostream>
#include <regex>

int main()
{
        std::regex pattern("[abc][123]", std::regex_constants::extended);
        std::cout << std::regex_match("a", pattern) << std::endl;
}

GCC produces "1". LLVM produces "0".

Am I misunderstanding something or is GCC's regex support horribly broken?

This is with GCC 4.8.3 on OSX.

Thanks

I can confirm phantom's idea about gcc 4.9.x being better.

I used 4.9.1 and my output was 0, matching your LLVM output.

Aah I see. Thanks for checking. I will upgrade to 4.9 then.

The reason I need to support GCC is because I want my code to run on Linux also (though I'm developing on OSX).

Very disappointed in GCC/libstdc++. They shipped something that is obviously broken with no warnings or errors, and just return wrong values. It took me half an hour of debugging to narrow it down to regex since I assumed GCC would be able to get something so simple right and there was no warnings or errors.

The least they could have done is add a "#warning" whenever someone included <regex>. Or maybe throw an exception whenever any regex stuff is used.

Originally Posted by phantomotap

O_o

To be fair, they also documented online that "regex" was incomplete which had about a dozen footnotes.

They should have thrown a firewall over "regex" (Requiring a `CXX11EXPERIMENTAL" macro or similar to suppress the warnings.) Given the state of C++11 support in general, I'm not really surprised specific warnings weren't offered.

Soma

Yeah I'm pretty disappointed by libstdc++'s C++11 support in general. C++11 isn't even that new. It has been standardized for 3 years already, and as C++0x for another few years.

I independently discovered a serious bug in std::this_thread::sleep_until a while ago, too. Basically if the end time has already passed when the function checks it due to scheduling, etc, it will sleep forever due to unsigned underflow. It has been more than a year already, and patches have been submitted. It's a small clean fix, and it's still not applied.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58038

I am now thinking about switching to LLVM on Linux as well. It seems to be of generally higher quality at least for x86-64 and C++11. For most projects it seems to produce slightly faster binaries than GCC as well.

It's true that they do warn about regex being broken in docs (now that I checked), but most people wouldn't have guessed "incomplete" means "not even the simplest examples work". If it's broken to that level, I would think at least a "#warning <regex> support is highly experimental and will probably not work" in <regex> is warranted. It would just be 1 line of code for them to add.

And by the way, my code does work perfectly with GCC 4.9. Thanks!