The future implementation is NOT stable. I don't know when the bug was introduced, but it appears clearly now on the regression test as I have wrapped the test on a loop to increase the change of error. Sorry, can I clarify, are you saying that the futures implementation in present Boost.Thread is not reliable?
This is what the regression test says :(
I started to see this bug with the clang asan11 tester, but looping on the test shows that the problem is not limited to this particular configuration.
I've worked hard to find out the issue but while I have found some minor issues, I have not reached to mix all the problems that appear now on the regression test.
btw, if the implementation used to work at one point and it is a regression, it should be possible to find the first broken commit using git-bisect ...