The `.capacity()` of a `std::vector<>` starts off at 1 and doubles each time the vector's size will exceed the current capacity. This capacity increase requires a reallocation of storage, and possibly a copying of the entire array contents to a new location. If you are generating 1 trillion primes, you'll need 40 reallocations and will have done 1 trillion number copies over the course of the 40 reallocations. This is wasting time. If the operating system gets involved with virtual memory paging, you are going to suck up a lot more time.
This overhead can be eliminated by simply [`.reserve(size_type n)`](http://www.cplusplus.com/reference/vector/vector/reserve/)-ing the expected size of the vector ahead of time. With sufficient memory allocated to the vector upfront, no reallocations will occur, and no copies will occur.
----
for (unsigned long long i = 7;; i += 2)
{
for (int j = 0; i % primes[j] != 0; ++j)
{
if (sqrt(i) <= primes[j])
{
When `i` is a prime number or semi-prime number in the order of, say, one million, you are looping over all the prime numbers upto the square root of a million, to see if any of them can divide your current number.
How many different values will `sqrt(i)` evaluate to over those thousand iterations? Or phrased another way, how many times are you computing the same square root? You may want to move that `sqrt(i)` calculation out of the inner loop.
----
for (unsigned long long i = 7;; i += 2)
If you let this loop run over night, or even over a fort-night, will this loop ever end? No! `i` will overflow the `long long` and become negative, and slowing increment back towards positive numbers and repeat. Forever is not long enough. Use `i > 0` as the loop test condition.
----
A `long long` has at least 64 bits. A `double` has only a 52 bit mantissa. This means when you pass a large `long long` to `sqrt( )`, you will end up losing a few bits, which can make your `sqrt()` return slightly the wrong value. When you test `sqrt(i) <= primes[j]`, if `i` is greater than 2^52, and is a perfect square, you might return a value slightly less than the correct value and fail to test the last prime value, and erroneously declare the perfect square a prime number.
----
You are stuffing `long long` values into a `std:vector<int>` container. After a while, they ain't gonna fit.
----
You are using `long long` for your prime number candidates, which means you expect to find some prime numbers above 2^31.
The Prime Number Theorem tells us the density of prime numbers in that range to be around 1/21. Or, after testing numbers up to 21 billion, you should have found around 1 billion prime numbers.
Your prime number index `j` is declared as an `int`. An `int` is only guaranteed to have 16 bits. You would need at least a `long` to guarantee 32 bits. But if you hope to find prime numbers up to 2^52, you have to expect to find 2^45 primes, which even exceeds a `long`. Your `j` index should be a `long long` as well.
----
Finally, as mentioned in the comments, look at the Sieve of Eratosthenes.