We use &vector[0] in several places to get the pointer to vector data, but std::vector[i] states invariant that i must be < vector.size(). If the vector is not empty - then &vector[0] works ok. However if the vector is empty - then the invariant is broken.

Normally it should not be a problem, since with empty vector we usually use (&vector[0], vector.size()) pair, and for empty vector its size is zero. In other words even if &vector[0] returns some not good pointer with zero size that data should never be accessed. But a build of libstdc++ could have asserts turned on and verify that i is actually < size. For example with default install on Fedora the following program with plain select()

from golang import select
select()

crashes with assert triggered on the breakage of that invariant

[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".
/usr/include/c++/10/bits/stl_vector.h:1042: std::vector<_Tp, _Alloc>::reference std::vector<_Tp, _Alloc>::operator[](std::vector<_Tp, _Alloc>::size_type) [with _Tp = golang::_selcase; _Alloc = std::allocator<golang::_selcase>; std::vector<_Tp, _Alloc>::reference = golang::_selcase&; std::vector<_Tp, _Alloc>::size_type = long unsigned int]: Assertion '__builtin_expect(__n < this->size(), true)' failed.

Program received signal SIGABRT, Aborted.
0x00007ffff7e22a25 in raise () from /lib64/libc.so.6
(gdb) bt
#0  0x00007ffff7e22a25 in raise () from /lib64/libc.so.6
#1  0x00007ffff7e0b895 in abort () from /lib64/libc.so.6
#2  0x00007fffea638d6d in std::__replacement_assert (__condition=0x7fffea63ea88 "__builtin_expect(__n < this->size(), true)",
    __function=0x7fffea63eab8 "std::vector<_Tp, _Alloc>::reference std::vector<_Tp, _Alloc>::operator[](std::vector<_Tp, _Alloc>::size_type) [with _Tp = golang::_selcase; _Alloc = std::allocator<golang::_selcase>; std::vector<_Tp, "..., __line=1042, __file=0x7fffea63ebe8 "/usr/include/c++/10/bits/stl_vector.h")
    at /usr/include/c++/10/x86_64-redhat-linux/bits/c++config.h:2552
#3  std::vector<golang::_selcase, std::allocator<golang::_selcase> >::operator[] (this=<synthetic pointer>, __n=0)
    at /usr/include/c++/10/bits/stl_vector.h:1042
#4  std::vector<golang::_selcase, std::allocator<golang::_selcase> >::operator[] (__n=0, this=<synthetic pointer>)
    at /usr/include/c++/10/bits/stl_vector.h:1040
#5  __pyx_pf_6golang_7_golang_6pyselect (__pyx_self=<optimized out>, __pyx_v_pycasev=<optimized out>) at golang/_golang.cpp:7930
#6  __pyx_pw_6golang_7_golang_7pyselect (__pyx_self=<optimized out>, __pyx_args=<optimized out>, __pyx_kwds=<optimized out>) at golang/_golang.cpp:7284
#7  0x00007ffff7bceeab in cfunction_call_varargs (func=<built-in function pyselect>, args=(), kwargs=0x0)
    at /usr/src/debug/python3-3.8.10-1.fc32.x86_64/Objects/call.c:743
...

-> Fix it all via replacing &vector[0] with vector.data() everywhere.

No need to add new test since select() is used by test_blockforever() and running that test on Fedora was crashing as well.

/cc @jerome