RFC: Add RelocPath type for userdefined relocatable paths

[fatteneder]

RelocPath(path) works by splitting path as depot/subpath based on DEPOT_PATH during construction.

When using RelocPath one explicitly opts into relocation, so it throws when path cannot be split.

A relocated path can be queried with String(r::RelocPath), which also throws on failure.

---

Fixes #54430
Alternative to #55146

[topolarity]

What's the idiomatic usage pattern for this look like?

[fatteneder]

The example where @__RELOCDIR__ fails #55146 (comment) would become now:

module MyPkg

const data = RelocPath(joinpath(@__DIR__, "myfile.data"))

function read_data()
    path = String(data)
    # load from path ...
end

end

[gbaraldi]

Is this how we get a Path type finally. I'd love that

[DilumAluthge]

It would be nice to get this into Julia 1.12.

[DilumAluthge]

Should we have a way of constructing a RelocPath by just providing the subpath?

[fatteneder]

Is this how we get a Path type finally. I'd love that

FWIW there is already https://juliahub.com/ui/Packages/General/FilePathsBase and https://juliahub.com/ui/Packages/General/FilePaths which provide such types.

Should we have a way of constructing a RelocPath by just providing the subpath?

I am not sure.

[DilumAluthge]

@vchuravy @gbaraldi and @topolarity Would you be able to review this PR?

As far as I can tell, this PR is an improvement compared to #55146, but I'm not particularly familiar with the relevant part of the codebase.

[DilumAluthge]

@staticfloat IIRC, you also worked on the depot-relocatibility code; it would be great if you could review this PR.

[DilumAluthge]

Is this how we get a Path type finally. I'd love that

The RelocPath constructor will error if the path isn't inside a depot, right? So IIUC this doesn't really cover the general "path type" request.

[vtjnash]

Thoughts about using OncePerProcess to make it so we only compute the result once:

Suggested change

	struct RelocPath
	subpath::String
	struct RelocPath
	getpath::OncePerProcess{String}
	end
	RelocPath(path::AbstractString) = let
	depot, _ = replace_depot_path_impl(path)
	if isnothing(depot)
	error("Failed to locate $(path) in any of DEPOT_PATH.")
	end
	subpath = String(replace(path, depot => ""; count=1))
	return RelocPath(OncePerProcess{String}() do
	for d in DEPOT_PATH
	if isdirpath(d)
	d = dirname(d)
	end
	local path = string(d, subpath)
	if ispath(path)
	return path
	end
	end
	error("Failed to relocate @depot$(subpath) in any of DEPOT_PATH.")
	end)
	end
	String(r::RelocPath) = r.getpath() # dynamically dispatches since getpath isn 't concrete :/

[DilumAluthge]

I don't love this. What if someone has modified Base.DEPOT_PATH during their Julia session?

[vtjnash]

This is pretty inaccurate, since it doesn't support @depot

Suggested change

	function show(io::IO, r::RelocPath)
	print(io, string("RelocPath(\"@depot", r.subpath, "\")"))
	end

[DilumAluthge]

How about we allow RelocPath to accept paths that are literally @depot/foo/bar?

[DilumAluthge]

Also, as far as the original show method goes, I think that the meaning of @depot/ is reasonably easy to understand here.

[fatteneder]

This is pretty inaccurate, since it doesn't support @depot

But it does behave like a @depot. (it behaves more like include_dependency than include to be precise)

What is inaccurate is that show assumes that nobody tinkered with DEPOT_PATH since construction.

[topolarity]

Should we just call this DepotPath? I think that might be a clearer name for users

[fatteneder]

Should we just call this DepotPath? I think that might be a clearer name for users

No strong feelings from my side.

[DilumAluthge]

I also don't feel strongly between RelocPath and DepotPath.

We could also do RelocatablePath if we wanted something a little more unambiguous.