Infinite loop when std::uncaught_exception is true

[tm604]

Been seeing some unusual behaviour with code that uses std::exception_ptr, finally tracked it down to a small-ish test case so thought it might be worth submitting. For anyone who stumbles across this while searching: this one has similar symptoms to #485 but I think the cause is different.

The example code in
https://gist.github.com/tm604/ade0a9354d894a0e71a4 uses std::exception_ptr to store an exception and rethrow it later.

This check:

        if( std::uncaught_exception() ) {
            m_unfinishedSections.push_back( UnfinishedSections( info, prevAssertions, _durationInSeconds ) );
            return;
        }

in https://github.com/philsquared/Catch/blob/1dd0d4c61aa7a5e1bf72fd3c1cca8ab6d7ad5460/include/internal/catch_runner_impl.hpp#L181 is returning true (not entirely sure why it's true, std::current_exception returns false which usually means stack unwinding is still in progress). This is causing an infinite loop because the section is never marked as completed.

Also, probably related: if I remove the REQUIRE/CHECK lines from that test file I get "free(): invalid pointer: 0x0000000001d07c68" instead of an infinite loop - ASAN's not too happy about the code there either (-fsanitize=address), have attached the output as generated when using the v1.2.1 single-header release.

So far I've just been removing that if() block entirely, not sure if there's a better fix - at least that allows the test to complete (and pass, in this case).

cheers,

Tom

[erhhh]

I wound up debugging separately for myself before finding this issue, and figured I should share.

The ASAN issue is likely because this

m_unfinishedSections.push_back( UnfinishedSections( info, prevAssertions, _durationInSeconds ) );

is happening inside this loop

for( std::vector<UnfinishedSections>::const_iterator it = m_unfinishedSections.begin(),
            itEnd = m_unfinishedSections.end();
        it != itEnd;
        ++it )

which is causing it and itEnd to be invalidated. Using std::list instead of std::vector should avoid that.

std::uncaught_exception being true when it shouldn't be is due to https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62258

[philsquared]

Thanks for all the triaging here (sorry it's taken a while to respond).
I've just committed a few changes in this area (on Develop) which should certainly address the vector issue (it should no longer re-enter while iterating) - I'm hoping it will also avoid the uncaught_exception bug (which I'm unable to reproduce, so can't be sure). Please let me know if this change addresses either or both issues for you (I'll be particularly pleased if the uncaught_exception issue goes away without having to resort to the hacks mentioned off that link).

[tm604]

Thanks guys - the uncaught_exception is definitely that GCC bug, switching to a patched version of the compiler makes that go away.

The latest version in develop no longer has the infinite loop issue, even without the compiler patch, so that's a good start. However, I'm seeing new test failures with that build - will try to put together a minimal test case for those. It may be related to the uncaught_exception bug again, so I'll try the patched compiler as well to see if that helps.

[philsquared]

Hi @tm604 - thanks for the report back.
Did you ever get to putting that "minimal test case" for the new failures?
Any clues what the failures are?

[philsquared]

Any objection to me closing this issue, now? (@tm604 if you're still seeing other issues please raise a new ticket if not related to this one).