Hello-

PR61369 (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61369) points out
that std::discrete_distribution can return an event even if it has 0
probability, and proposes a simple fix. It seems that this fix was never
applied, because there was an expectation of redoing this code anyway to
use a more efficient algorithm (PR57925). Given that this new algorithm
has not been implemented so far, would it make sense to apply the simple
fix to address this issue? The attached patch does this.

One question about the patch, a slight annoyance is that only
std::lower_bound() is currently available in random.tcc, as this file
includes only bits/stl_algobase.h and not bits/stl_algo.h (via including
<vector>). Is there a preference between simply including stl_algo.h, or
moving upper_bound to stl_algobase.h, where lower_bound is? I noticed
that in C++20 mode, <vector> includes stl_algo.h already, so I figured
it would be fine to just include it in random.tcc unconditionally.

bootstrap + testing were done on x86-64 GNU/Linux, all tests the same
before + after plus 2 new passes from the new test. Thanks for taking a
look!

-Lewis