When solving AtCoder Grand Contest 044 Problem B, I wrote some O(n^3) code with n<=500 like:

typedef vector<int> vi;
vector<vi> dirs({{0,1}, {0,-1}, {-1,0}, {1,0}});
...
for (auto dir : dirs){ // run O(n^3) times
...
}

and get TLE. On my pc it runs within 4900ms. During the contest I have no idea why.(dumb me) After the contest I looked at other people's solution and finally get accepted solution, just add an '&' after 'auto':

for (auto& dir : dirs){ // run O(n^3) times
...
}

480 ms on my pc! 10 times faster!

I did more tests:

for (int dd = 0; dd<4; ++dd){ // run O(n^3) times
    vi dir = dirs[dd];

4900 ms

for (int dd = 0; dd<4; ++dd){ // run O(n^3) times
    vi& dir = dirs[dd];

480 ms

vi dir;
for (int dd = 0; dd<4; ++dd){ // run O(n^3) times
    dir = dirs[dd];

1800 ms

Maybe vector is too big, I tried pair:

vector<pair<int, int>> dirs({{0,1}, {0,-1}, {-1,0}, {1,0}});
for (auto dir : dirs){ // run O(n^3) times

370 ms, no difference if I add ‘&’， maybe because pair of int is the same size as 64 bit pointers.

Final conclusion: Seems like

for (auto dir : dirs)

is equivalent to

for (int dd = 0; dd<4; ++dd){ // run O(n^3) times
    vi dir = dirs[dd];

that will construct a vector and assignment another vector to it every iteration. which cost much more time than simple data structures or reference.