Generic Parse Method without Boxing

For better readability, you could use a generic dictionary with an anonymous function as follows:

var parserFuncs = new Dictionary<Type, Func<string, object>>() {
    { typeof(int), p => (int) int.Parse(p) },
    { typeof(bool), p => (bool) bool.Parse(p) },
    { typeof(long), p => (long) long.Parse(p) },
    { typeof(short), p => (short) short.Parse(p) },
    { typeof(DateTime), p => (DateTime) DateTime.Parse(p) }
    /* ...same for all the other primitive types */
};

return (T) parserFuncs[typeof(T)](value);

I'll add a little undocumented way:

public static T Convert<T>()
{
    if (typeof(T) == typeof(int))
    {
        int a = 5;
        T value = __refvalue(__makeref(a), T);
        return value;
    }
    else if (typeof(T) == typeof(long))
    {
        long a = 6;
        T value = __refvalue(__makeref(a), T);
        return value;
    }

    throw new NotImplementedException();
}

There is little documentation about them, but they work as of C# 4.0. Read for example here Hidden Features of C#? Remember that undocumented means unsupported, blah blah blah could not work in the future blah blah blah if you use them the devil will come for you blah blah blah :-)

public static T Parse<T>(this NameValueCollection col, string key)
{
  return (T)Convert.ChangeType(col[key], typeof(T));
}

I'm not entirely sure of ChangeType boxes or not (I guess reading the docs would tell me, but I'm pressed for time right now), but at least it gets rid of all that type-checking. The boxing overhead is not very high, though, so I wouldn't worry too much about it. If you're worried about run-time type consistency, I'd write the function as:

public static T Parse<T>(this NameValueCollection col, string key)
{
  T value;

  try
  {
    value = (T)Convert.ChangeType(col[key], typeof(T));
  }
  catch
  {
    value = default(T);
  }

  return value;
}

This way the function won't bomb if the value cannot be converted for whatever reason. That means, of course, that you'll have to check the returned value (which you'd have to do anyway since the user can edit the querystring).

I think you are over estimating the impact of the boxing/unboxing. The parse method will have a much bigger overhead (string parsing), dwarfing the boxing overhead. Also all the if statements will have a bigger impact. Reflection has the biggest impact of all.

I'd would not like to see this kind of code in production, as there is a cleaner way of doing it. The major problem I have with it is the large number of if statements you will need to cover all cases and the fact that someone could pass any old type to it.

What I would do is write a parse function for each type I want to parse (ie ParseInt()). It's clearer and it is well defined what the function will try to do. Also with short static methods, the compiler is more likely to inline them, saving a function call.

I think this is a bad application of generics, any particular reason for doing it this way?