From Katz & Lindell's book, theorem 7.19:
Let f be a one way permutation with hard-core predicate hc.
Then the algorithm $G(s)=f(s)||hc(s)$ is a PRG with expansion factor $\ell(n)=n+1$.
But what about if we use a one-way function? I know that there could be $f(x)=f(x^{'})$ when $x\neq x^{'}$, we lost some entropy here, but what breaks the theorem actually?
The proof of Theorem 7.19 explicitly uses that $f(r)$ for a uniformly random bitstring $r$ is distributed like $r$ when $f$ is a permutation*. I don't have the energy to reprint the proof here to show exactly how the distribution of $f(r)$ is used, but it should be clear that it can be useful to know this about $f(r)$ when arguing that $G$ is a PRG.
If you want a counterexample when $f$ is not a permutation you can see this answer.
*This is the part of the proof that says
using the fact that $f$ is a permutation for the second equality
for those who have access to the book.
