Allow all characters in runfile source and target paths

[fmeum]

Adds support for spaces and newlines in source and target paths of runfiles symlinks to build-runfiles as well as to the Bash, C++, and Java runfiles libraries (the Python runfiles library has moved out of the repo).

If a symlink has spaces or newlines in its source path, it is prefixed with a space in the manifest and spaces, newlines, and backslashes in the source path are escaped with \s, \n, and \b respectively. This scheme has been chosen as it has the following properties:

1. There is no change to the format of manifest lines for entries whose source and target paths don't contain a space. This ensures compatibility with existing runfiles libraries.
2. There is even no change to the format of manifest lines for entries whose target path contains spaces but whose source path does not. These manifests previously failed in build-runfiles, but were usable on Windows and supported by the official runfiles libraries. This also ensures that the initialization snippet for the Bash runfiles library doesn't need to change, even when used on Unix with a space in the output base path.
3. The scheme is fully reversible and only depends on the source path, which gives runfiles libraries a choice between reversing the escaping when parsing the manifest (C++, Java) or applying the escaping when searching the manifest (Bash).

Fixes #4327

RELNOTES: Bazel now supports all characters in the rlocation and target paths of runfiles and can be run from workspaces with a space in their full path.

[fmeum]

@lberki I assigned you as reviewer for your general knowledge of runfiles, but feel free to redirect if needed.

[fmeum]

CC @meteorcloudy for Windows

[meteorcloudy]

Wow, this is great!

[meteorcloudy]

Is the runfiles path the only reason we didn't support space in workspace path? /cc @lberki

[fmeum]

I could see missing escaping in the test setup shell script as one other source of issues. When the rest of this PR and the general approach look good to you, I could run Bazel's test suite under a path with a space to shake out these issues.

[lberki]

Yeah, @fmeum 's approach is reasonable; spaces in workspace paths is the kind of thing that could break in unexpected places. Yes, we should properly quote everywhere, but that doesn't mean that we do if the functionality is untested.

[sgowroji]

Hi @meteorcloudy, Since i can see that this PR has been approved, Please let us know whether gTech team should proceed with importing it. Thanks!

[fmeum]

@sgowroji I would prefer to wait for a second opinion, for example by @lberki.

[matthewjh]

@fmeum: is the aim to get this into Bazel 8.0.0?

[lberki]

In principle, adding the functionality to encode spaces in runfiles manifests is fine. I'm not sure, though, how I feel about this particular one because it's tailored to spaces and only spaces. How about implementing a scheme that could be more easily extended?

Spitballing, something like this: if a line in the runfiles manifest starts with a space, the algorithm to parse the line is as follows: the separator is between the two paths is a space, end of the line is a newline, any other character is a file name. If a file name contains a space, a newline or a backslash, these need to be escaped as \s, \n or \b, respectively, any other backslash sequences are reserved for future extension.

This is easy to implement, provably supports everything is just as incompatible with the existing system and looks more conventional than the encoding proposed in this pull request.

@fmeum WDYT? It looks like if we go through the pain of adding syntax to runfiles manifests, we could as well make it future-proof. I could also be convinced to make them some subset of JSON so that parsing them is possible with standard libraries but also with random logic so that one doesn't need to add additional dependencies.

[fmeum]

This is easy to implement, provably supports everything is just as incompatible with the existing system and looks more conventional than the encoding proposed in this pull request.

It lacks one important property, but we can also fix that: With this scheme in its current form, the Bash runfiles setup snippet as well as third-party runfiles libraries would need to be modified to continue to support manifest entries in which the rlocationpath contains no spaces, but the target path does (common on Windows).

We can fix that by not escaping spaces in the target paths and declaring that the first space on a line that doesn't start with a space is the separator. We would then only add a space at the beginning of a line if either the rlocationpath contains a space or newline or if the target path contains a newline.

At that point it's similar to the scheme in this PR, it just uses escape characters instead of length encoding. I agree that escaping is more conventional and probably not much harder to implement in most languages (it does change the length of the strings vs. just taking a substring, but that seems okay).

@lberki What do you think, should I go ahead and implement the modified version of your scheme?

[fmeum]

I implemented the scheme without support for newlines for now. They could always be supported in a follow-up, which is much simpler with the new scheme (just introduce a new escape character everywhere).

[lberki]

The modified scheme looks reasonable. I'm travelling today so I can't review the pull request but will do so as soon as I can.

The main question in my mind is whether we do some sort of custom encoding (like this one) or go with a standard one. Of the standard ones, JSON seems to be the most palatable given the use case (simple list of key/value pairs intended to be parseable without any dependencies), but I'm not sure if it's worth the hassle and if it is, we are probably better off having an additional file called MANIFEST.json or something like that

[fmeum]

@Wyverald Done