What is the relation between a programming language and the language of its input?

Question

I find some references say that

all the features of programming language fall within what can be captured by context-sensitive grammars. In fact, no programming language known to humankind anything that cann't.

and

Actually, most of programming language is context-free language.

Q1: Are the claims correct? I did not find some bibiographies about a precise description or even a proof.

Let $\Gamma$ be the alphabet of the instructions of a programming language. Given a specific procedure $P$, all of its control flows belong to $L_{P}$. $L_{P}$ describes how does it control its data.

At the same time, for every $x \in \Sigma^*$, $P$ computes it and gives an output $P(x)$. Assume that $P$ is used to slove a decision problem, i.e., $P(x) \in \{ 0,1 \}$. There is a language $L'_{P} = \{ x \in \Sigma* \mid P(x) = 1 \}$. If we regard $P$ as a machine, $L'_P$ describes the computational power of $P$.

It is obvious that $L_P$ and $L'_{P}$ are totally different.

Q2: What is the relation between $L_P$ and $L'_{P}$? Is there a conclusion such that $$L_P \in CFL \Leftrightarrow L'_P \in CFL$$ or $$L_P \in CSL \Leftrightarrow L'_P \in CSL$$

Answer 1

There is no interesting relation between a programming language (the set of words which are syntactically valid programs) and the set of words that are accepted by a program written in said programming language.

Note that it is possible to design a Turing powerful programming language which, as a formal language, is the set $1^* = \{\epsilon,1,11,111,\ldots\}$.

Still, program $1^n$ will accept a language $L_n$ which is completely unrelated to $1^*$. Indeed, by Turing completeness, for any r.e. language $A$ there is some $n$ such that $A=L_n$.

The crucial aspect to consider here is the semantics of the programming language, the way we map each program $1^n$ to its behavior, and ultimately to the set of accepted words $L_n$. Even if $1^*$ is a very simple language, the semantics can still make it Turing powerful, so that $L_n$ can be any r.e. language.