Thread overview
Idea: Qualified jumps
Mar 14, 2022
Quirin Schroll
Mar 14, 2022
H. S. Teoh
Mar 15, 2022
Elronnd
March 14, 2022

Idea

Imagine you have a loop that does multiple things slightly differently. You think about refactoring those similar parts into a single function that takes a few parameters and could even have a nice, descriplive name. And you can’t — because there's a pesky return or break statement that just wouldn't work when put into another function. How would it even work anyways? Well, how would it? It couldn’t ever work in a global-scope function, that’s for sure, but it could — in principle — work in a nested function or a lambda under some reasonable conditions.
I've thought some time about the syntax, and parnet.return value or parent.break label (i.e. ⟨qualified-name⟩.⟨statement⟩ ⟨args⟩) felt the best, so I'm using. It is borrowed from object.BaseClass.method syntax.

Example

Show, don’t tell:

int parent(int[] xss)
{
    loop:
    foreach (xs; xss)
    {
        foreach (x; xs)
        {
            void child(int y)
            {
                if (isPrime(y))
                    parent.break loop;
                else if (y == 0xF00D)
                    parent.return 1234;
                else if (y > 1_000_000)
                    parent.goto veryBig;
            }
            child(x);
            higherOrderFunc(&child, x);
        }
    }

    return 0;

veryBig:
    writeln("wow!");
    return 1;
}

The function child must be declared in the loop. Declare it before and it does not see the label loop (you could object to this). Declare it after and its name is not avaliable inside the loop (this has always been the case).

The reasonable conditions are: If you ever take the address of that function, it must be assigned to a scope variable.
Otherwiese when the function is called, the stack frame it referred to might not exist anymore. It cannot be returned by the parent function. And it cannot be static.

The only case without ⟨args⟩ is return in a void context. break and continue must use the labelled form.

The Juicy Part

If we go further and deprecate lambdas that start with a brace (i.e. auto lam = {} lambdas, not auto lam = (){} lambdas), we could change their semtantics: Every return, break, continue, and goto statement would be implicitly a parent.return, parent.break etc. targeting the nearest control-flow structure.

Together with some proposal to omit parentheses when the only (or last) argument is such a special lambda, it would allow for custom control-flow statements.

int parent() {
  myControlFlowStmt(args) { lambda content; return 0; }
}

is lowered to

int parent() {
  myControlFlowStmt(args, () { lambda content; parent.return 0; });
}

The lowering for foreach does something similar, but there isn't really a way to utilize that functionality outside of it.

Corner Cases

If you have multiple nested functions, search without a leading dot is inside-out and outside-in with a leading dot:

void parent() // 1
{
    void parent() // 2
    {
        void parent() // 3
        {
            void child()
            {
                return; // exits child
                child.return; // (same) exits child
                parent.return; // exits 3
                parent.parent.return; // exits 2
                parent.parent.parent.return; // exits 1
                .parent.return; // exits 1
                .parent.parent.return; // exits 2
                .parent.parent.parent.return; // exits 3
                .child.return; // error (child is not at global scope)
                .parent.child.return; // error (child not directly under global .parent)
            }
        }
    }
}

If you name the parents differently, of course you don't need all these; parent2.return will suffice. Also note that nesting functions with the same name is legal in current D.

Who Is This For?

This looks like it asks for spaghetti code. But it's rather for being able to express something one cannot otherwise (easily) do. You would want that if you’re caught between code duplication and optimization.

March 14, 2022
On Mon, Mar 14, 2022 at 05:33:29PM +0000, Quirin Schroll via Digitalmars-d wrote:
> #### Idea
> Imagine you have a loop that does multiple things slightly
> differently. You think about refactoring those similar parts into a
> single function that takes a few parameters and could even have a
> nice, descriplive name. And you can’t — because there's a pesky
> `return` or `break` statement that just wouldn't work when put into
> another function. How would it even work anyways? Well, how would it?
> It couldn’t ever work in a global-scope function, that’s for sure, but
> it could — in principle — work in a nested function or a lambda under
> some reasonable conditions.
[...]

This is a similar problem to the one already solved by opApply: have the function return a value that indicates whether the containing loop should continue executing or not. For example, when iterating over an aggregate containing .opApply, the loop body is transformed such that `continue;` becomes `return 0;` and `break;` becomes `return 1;`. The loop implementation stops the loop if the delegate returns non-zero.

In a similar vein your refactored code could just replace a hoisted code block with:

	if (refactoredCode(...)) break;

Depending on the circumstances you may need to further differentiate between normal return and short-circuit `continue`, so you use a scheme where 0 = normal return, 1 = break, -1 = continue. So you'd do:

	L1: foreach (...) {
		...
		switch (auto ret = refactoredCode(...)) {
			case 1: break L1;
			case -1: continue;
		}
		...
	}

Kinda verbose, I'll grant you that, but I'm sure you can think of ways of making it more concise.


T

-- 
I think Debian's doing something wrong, `apt-get install pesticide', doesn't seem to remove the bugs on my system! -- Mike Dresser
March 15, 2022
On Monday, 14 March 2022 at 17:33:29 UTC, Quirin Schroll wrote:
> *snip*

Common lisp does this, if you are looking for precedent.