RFC: Add `@__RELOCDIR__`, a relocatable version of `@__DIR__`

[fatteneder]

This adds the @__RELOCDIR__ macro which shall not hinder relocation of packages, unlike @__DIR__.

It works by splitting @__DIR__ into depot and subpath at macro expansion time. Resolution of @__RELOCDIR__ happens at runtime by trying to find a dir at depot/subpath.

If used in scripts or with julia -e <expr> the macro behaves like @__DIR__. If resolution of @__RELOCDIR__ fails the value of @__DIR__ is returned.

---

Fixes #54430

[fatteneder]

Open questions:

• Shall this just replace @__DIR__? Initially I thought no, because I am too paranoid that people use @__DIR__ to construct relative paths that refer outside a depot. On the other hand, introducing a new macro just misses out on easier migration in all the non-problematic cases.
• If it would replace current @__DIR__, would we need a mechanism to opt out of relocation? Given that kwargs for macros aren't a thing, the only option that would come to my mind would be @__DIR__(false) which looks ugly. Edit: Could also use @__DIR__(:relocatable) or so.

[giordano]

Initially I thought no, because I am too paranoid that people use @__DIR__ to construct relative paths that refer outside a depot.

I often use @__DIR__ in scripts, which have nothing to do with depots. And the script use case is also impossible to capture with PkgEval, so please absolutely don't do that.

[topolarity]

I'm not a big fan of the fact that something as simple as:

const data = joinpath(@__RELOCDIR__, "myfile.data")

is still non-relocatable.

The alternative path in my mind would be to have @__RELOCDIR__ return a path with a "virtual depot root" (e.g. /@depot/path/to/file) and then the low-library File I/O's in Julia would resolve this virtual root to it's "real" location. That's a bit similar to how Rust transparently translates paths on Windows to their "verbatim" form for transparent long-path support.

That might be too extreme to be worth hooking into low-level libraries, but it's the only way I can imagine getting good relocatability behavior "transparently"

[vchuravy]

If paths were more than just strings we could handle this elegantly. I can't find the issue for that just now.

We could perhaps handle the @depot handling during relocation. I.e. when we are writing an image we scan the ?beginning? of the string for @depot add it to an array of such strings and during deserialization we patch the result?

[topolarity]

We could perhaps handle the @depot handling during relocation. I.e. when we are writing an image we scan the ?beginning? of the string for @depot add it to an array of such strings and during deserialization we patch the result?

I think that's too risky - Scanning all the String's in your application and modifying some of them seems like a great way to mess up programs. What happens if it was already used as a key in a Dict, etc.?

We wouldn't be able to mark String immutable any more either, since it may be modified by the serializer

[vchuravy]

I don't disagree. But having the cost of prefix checking before operations in the file handling code seems suboptimal.

I kinda want it to return a DepotPath("..../...") object that we can handle correctly

[topolarity]

I don't disagree. But having the cost of prefix checking before operations in the file handling code seems suboptimal.

The scan for a prefix is probably pretty cheap (compared to the syscall), but the fact that we have multiple depot paths is definitely a huge pain.. that would require additional syscalls and may not even be possible to soundly emulate

So yeah, maybe the DepotPath style solution is the only way forward