Idea for avoiding GC for lambdas

[H. S. Teoh]

On Tue, Jun 22, 2021 at 03:56:03PM -0400, Steven Schveighoffer via Digitalmars-d wrote:
[...]
> This again, is a misunderstanding of my proposal. I'm not talking
> about altering delegates in any way.
> 
> There are no delegates involved in `filter`. It is, instead, an inner
> struct with a hidden context pointer. Could just as well be a struct
> with context carried with it instead of allocated on the heap.
[...]

Ohhh I see what you mean now. So essentially what you're saying is,
whenever there's an alias parameter bound to a lambda, the compiler
should, in effect, insert additional parameters to the template
function that captures the context the lambda is bound to?

IOW, this:

	auto filter(alias fun, R)(R range) {
		struct Voldemort {
			...
			if (fun(range.front)) ...
			...
		}
		return Voldemort(range);
	}

	int x = 2, y = 3;
	auto r = [1, 2, 3].filter!(e => e == x || e == y);

gets lowered into something like this:

	struct __lambda_context {
		int x, y;
	}
	bool __lambda_fun(__lambda_context ctxt, int e) {
		return e == ctxt.x || e == ctxt.y;
	}
	auto __filter_instance(R)(__lambda_context ctxt, R range) {
		struct Voldemort {
			__lambda_context secretCtxt;
			...
			if (__lambda_fun(secretCtxt, range.front)) ...
			...
		}
		return Voldemort(ctxt, range);
	}

	int x = 2, y = 3;
	auto r = __filter_instance(__lambda_context(x, y), [1, 2, 3]);

Am I right?

__lambda_context may not necessarily carry a copy of the captured
variables; possibly it could use pointers instead. Though care would
have to be taken not to leave dangling pointers to out-of-scope local
variables -- probably you'd have to make the references `scope` unless
you copy them by value into the Voldemort struct.


T

-- 
Не дорог подарок, дорога любовь.