Most proofs of Rice theorem seem to be based on the undecidability of the halting problem. They are "reduction-based".
Are there "direct" elementary proofs, perhaps based on diagonalization?
I think that the answer is "YES there are".
However most textbooks I know (such as [Odifreddi, Moret, Jones, Phillips] present reduction-based proofs.
Here is a proof based on the recursion theorem, rather than a reduction of an undecidable problem.
Rice's Theorem. Suppose that $P$ is any set of computable functions, which is not empty and not all computable functions. Then the set $\{ e\mid \varphi_e\in P\}$ is not decidable, where $\varphi_e$ is the function computed by program $e$.
In other words, there is no general procedure to determine from a program whether the function it computes has property $P$ or not.
Proof. Suppose that the set were decidable. Fix a computable function $f$ that is in $P$, and another computable function $g$ that is not in $P$. Now, for any program $e$, let $h(e)$ be the program that on input $n$ first determines whether $\varphi_e\in P$; if so, it outputs $g(n)$, and otherwise $f(n)$. So $\varphi_{h(e)}$ is either $g$ or $f$, depending on whether $\varphi_e\in P$ or not, respectively (but note that we are using the opposite function). In particular, we'll have $$\varphi_e\in P\quad\iff\quad\varphi_{h(e)}\notin P.$$ Meanwhile, by the recursion theorem, there is a program $e$ such that $\varphi_e=\varphi_{h(e)}$, which now gives an immediate contradiction, since $\varphi_e$ and $\varphi_{h(e)}$ are supposed to be opposite with respect to $P$. QED
Here is a proof in a more 'constructive' form that can be directly applied to any real-world programming language. Just to make clear, a property of programs must be a predicate that only depends on the halting and output behaviour, in other words two equivalent programs will have exactly the same properties. I will use $X \equiv Y$ to denote that two programs are equivalent, and so $P$ is a property iff $P$ is a predicate and $P(X)=P(Y)$ for any programs $X \equiv Y$.
For any property $P$ of programs such that $P(T)$ and $\neg P(F)$ for some programs $T$,$F$:
If $P$ is decidable:
Let $D$ be a program that decides $P$
Let $M$ be the following program on input $x$:
Create $N$ to be the following program on input $y$:
If $x$ is an invalid program: Return "" [may be unnecessary in some languages]
Return $x(x)(y)$ [using closures or some interpreter written in the language itself]
If D(N): Return F
Return T
Let $R$ be the program that $M(M)$ creates in the variable $N$
Clearly $R \equiv M(M)$
Also $M(M)$ halts because $D$ halts on all inputs
Thus $M(M) \equiv T \vee M(M) \equiv F$
$P(M(M)) \Leftrightarrow P(R) \Leftrightarrow D(R) = True \Leftrightarrow M(M) \equiv F \Leftrightarrow \neg P(M(M))$
Contradiction
Therefore $P$ is not decidable
