Let say I have couple of thousands moving objects in my scene and i want to use frustum-sphere test to cull invisible ones.
Then in my code i have something like this:

struct BoundingSphere
{
   float x, y, z, radius;
};

struct ObjectsDetailPool
{
    BoundingSphere bs[MAX_OBJECTS];
    uint64_t visible[MAX_OBJECTS]; // 64 possible views
    DrawableThing dt[MAX_OBJECTS]; // mesh data
};

struct Object
{
    BoundingSphere* pbs;
    uint64_t* pvisible;
    DrawableThing* pdt;
};

vector<Object> vobjs;

... //here cull for all views

// draw
if(vobjs.empty())
   return;

...//set shader...
for(auto& obj : vobjs)
{
    if(obj.pvisible ...) // this is bad?
        obj.draw();
}

...// repeat draw for each view

But lot of texts say this is bad because of branch mispredictions and what not, usually there is a lot of criticism but none of the solutions shown.

I dont understand how should i construct/manage this list then?
Should i partition the vector? So i swap-back invisible objects and store iterator where visible ends, but then again i have a branch because i need to check if it is visible.
Create a new vector and push_back only visible objects? But then again i have a branch because i need to check if it is visible.