Faster, Easier, Simpler Vectors - David Stone - CppCon 2021 

thanks!

Cache locality? No need for additional indirection? Ability to memcpy contents?

compiler magic

I think he was referring to the concept more so than any particular implementation. The idea is to reserve a memory range with VirtualAlloc/mmap based on the max capacity and then growing the capacity extends the commit range in place rather than reallocating. Both allocation/deallocation and first-time page access incurs kernel mode transitions (syscalls and page faults, respectively), so it's not something you want for short lived or high churn vectors. But it can be useful if your application has a relatively small handful of big vectors where you need stability and/or performance. (Of course, if you just want stability and performance is not an issue, you can allocate and store the elements outside of the vector itself, as std::unordered_map has to do. The vector would still own those out-of-line elements, so the value semantics would be preserved.)

@@pervognsen_bitwise Do you know any stable vector implementation which has just 1 memory chunk and uses mmap or VirtualAlloc and os-specific functions?

I've seen this kind of thing called append_uninitialized(size_t n) which just does ensure_capacity(size + n); size += n. You'll find something akin to this in many non-std vector implementations. In Unreal Engine, for example, it's TArray::AddUninitialized.

Presenter here, and I agree, the name is horrible. There is work to add this to std::vector, and part of that will include coming up with a (hopefully) great name.

What do you mean? The allocator type is the second template type parameter. For example you can specify your own custom allocator as follows: vector v.

Look for some of the talks from John Lacos. He explains why it is good for types to implement the allocator feature, for enterprise scale software.

Mathematicians are more often than not terrible programmers.