Fix bug in unPathNative

[hasufell]

When passing a path like "F:/foo/bar" on windows
(both '/' and '\' are valid path separators on windows) we'll get garbage output like so:

mkPathNative "F:/foo/bar"
"F:/foo/bar"
unPathNative "F:/foo/bar"
"F:foo\bar"

...effectively turning an absolute path into a relative path ("F:foo\bar" on windows is the directory "foo\bar" relative to the current working directory on drive F).

This is because Posix and Windows splitDirectories behave differently:

System.FilePath.Posix.splitDirectories $ "F:/foo/bar"
["F:","foo","bar"]
System.FilePath.Windows.splitDirectories $ "F:/foo/bar"
["F:/","foo","bar"]

When joining paths on windows, the filepath library does not assume a trailing path separator after the drive (here "F:"). This is because as described above, "F:foo" is a valid relative filepath.

[hasufell]

To trigger the bug on windows: cabal --remote-repo-cache F:/build/packages update.

Resulting in:

Exception fromAbsoluteFilePath: not an absolute path:
F:build\packages\hackage.haskell.org\rooC8E3.json
CallStack (from HasCallStack):
error, called at src\Hackage\Security\Util\Path.hs:279:31 in
hackage-security-0.6.2.6-inplace:Hackage.Security.Util.Path when using mirror
http://hackage.haskell.org/

[hasufell]

Discussion:

• please don't use these splitPath/joinPath tricks... they're not there to do proper normalization
• my patch drops the implicit semi-normalization (we just replace the path separators now)... that means there may be duplicate path separators... I think that's ok, but someone more familiar with the codebase would have to double check whether the code anywhere assumes that there are no duplicate path separators
• if we really want normalisation, then we need to normalise first and THEN replace the path separators with forward slashes

[hasufell]

@andreabedini ran into this bug while working on stable-haskell/ghc#135

[andreasabel]

Thanks for the PR!
Backslash path separators are indeed a pest to be stomped out.

[andreasabel]

Since FilePath = String, how likely is an interesting test generated here?

Maybe it would be good to add a few unit tests.

[hasufell]

Filepath has a windows path BNF that's used to generate interesting output, but that would be some copy pasta: https://github.com/haskell/filepath/blob/master/tests/filepath-equivalent-tests/Gen.hs#L38

[andreasabel]

Do you think this generator would be worth publishing in a package we could use here?
That would be the preferred solution; copy-paste tends to multiply the maintenance effort.

If you consider this all feature-creep, you could also equip the property with a comment stating that this is the intended mathematical property but it is likely not established by the test because the coverage is very poor with the vanilla generator for FilePath = String.

[andreasabel]

To not have this PR fizzle out, I merged it and pushed a disclaimer: 36a5fbf

My original intent was to add the disclaimer as an extra commit to this PR, but the PR wasn't set up in a way to allow contributions from the maintainers.