That is my point. You are sacrificing the accuracy of a double for the speed of a float.
You are already storing all these values in floats which is inherently inaccurate. Also if the dt becomes very large you have bigger problems than gravity anyway.