class Solution {
public:
    void moveZeroes(vector<int>& nums) {
        vector<int>::size_type nonZero = 0;
        vector<int>::size_type next = 0;
        while (next < nums.size()) {
            if (nums[next] != 0) {
                // 使用 std::swap() 会带来 8ms 的性能损失
                // swap(nums[next], nums[nonZero]);
                auto tmp = nums[next];
                nums[next] = nums[nonZero];
                nums[nonZero] = tmp;
                //用nums[next]^=nums[nonZero]^=nums[next]^=nums[nonZero]; WA
                //用nonZero++;next++;会增加8ms左右
                ++nonZero;
            }
            //next++;
            ++next;    
        }
    }
};