On 19 August 2017 at 19:49, Avi Kivity wrote: > > > On 08/19/2017 08:49 PM, Jonathan Wakely wrote: >> >> On 19 August 2017 at 18:02, Avi Kivity wrote: >>> >>> I noticed that std::deque's move constructor (gcc 7.1) is not noexcept. >>> Is >>> this conformant to the standard? >> >> Yes it's conforming, but it sucks. > > > :( > >> >>> It's hard to imagine a reasonable deque >>> implementation that can fail a move. >> >> We have one. Our implementation is never "empty" it always has a table >> of pointers even if there are no elements. That means we need to >> allocate a new table for the moved-from deque, and allocating can >> throw. >> >> I don't think we can change this in an ABI compatible way. > > > Oh, I assumed no one would do this and wrote a wrapper that just adds the > noexcept, without even checking (a type that uses it requires noexcept > moves). > > I guess it's unique_ptr time for me. I proposed this ABI breaking change for the --enable-symvers=gnu-versioned-namespace config and nobody replied: https://gcc.gnu.org/ml/libstdc++/2016-10/msg00017.html This gets rid of the allocations in the move operations, allowing "really empty" deques. It also keeps a single deque node around for reuse, so that using a deque as a FIFO with a "one in, one out" sequence doesn't keep deallocating and reallocating. I'm going to commit this for GCC 8. Longer term I also want to provide some kind of adaptor so that e.g. __gnu_cxx::deque<T, A> would be an alias for std::deque<T, __better_deque<A>> and would enable the changes for the default --enable-symvers=gnu configuration.
