It seems that a newtype definition is just a data definition that obeys some restrictions (e.g., only one constructor), and that due to these restrictions the runtime system can handle newtypes more efficiently. And the handling of pattern matching for undefined values is slightly different.
But suppose Haskell would only knew data definitions, no newtypes: couldn't the compiler find out for itself whether a given data definition obeys these restrictions, and automatically treat it more efficiently?
I'm sure I'm missing out on something, there must be some deeper reason for this.
Both newtype and the single-constructor data introduce a single value constructor, but the value constructor introduced by newtype is strict and the value constructor introduced by data is lazy. So if you have
data D = D Int
newtype N = N Int
Then N undefined is equivalent to undefined and causes an error when evaluated. But D undefined is not equivalent to undefined, and it can be evaluated as long as you don't try to peek inside.
Couldn't the compiler handle this for itself.
No, not really—this is a case where as the programmer you get to decide whether the constructor is strict or lazy. To understand when and how to make constructors strict or lazy, you have to have a much better understanding of lazy evaluation than I do. I stick to the idea in the Report, namely that newtype is there for you to rename an existing type, like having several different incompatible kinds of measurements:
newtype Feet = Feet Double
newtype Cm = Cm Double
both behave exactly like Double at run time, but the compiler promises not to let you confuse them.