Why are STL algorithms much faster with pointers than std::vector iterators?

Question

Why does std::nth_element() run so much faster, when it is given pointers instead of iterators? I would expect std::vector and STL algorithms to be quite optimized, but my measurements show execution time drops by 75% when I change iterators to pointers.

Using iterators, the following code (not including allocation of the vector) ran in 1200 milliseconds:

std::vector<uint16_t> data(/* 50 million values */);

const double alfa = 0.01;
const double beta = 0.95;

std::nth_element(data.begin(), data.begin() + int(data.size() * alfa), data.end());
const uint16_t x = *(data.begin() + int(data.size() * alfa));

std::nth_element(data.begin(), data.begin() + int(data.size() * beta), data.end());
const uint16_t y = *(data.begin() + int(data.size() * beta));

Using pointers, the following code (not including allocation of the vector) ran in 350 milliseconds:

std::vector<uint16_t> data(/* 50 million values */);

const double alfa = 0.01;
const double beta = 0.95;

std::nth_element(&data.front(), &data.front() + int(data.size() * alfa),
    &data.front() + data.size());
const uint16_t x = *(data.begin() + int(data.size() * alfa));

std::nth_element(&data.front(), &data.front() + int(data.size() * beta),
    &data.front() + data.size());
const uint16_t y = *(data.begin() + int(data.size() * beta));

I observed similar speed increase with std::sort() as well. The examples were compiled with Embarcadero C++ Builder XE8 version 22.0.19027.8951, Release build and "Generate fastest possible code" setting. These tests were ran during different executions so they should not affect each other.

Answer 1

我的猜测是编译器要么没有做好优化，要么你正在调试模式下构建，编译器使用特殊的，调试（慢）版本的STL容器。