Fix ocamldep-postproc in case-insensitive, case-preserving filesystems.

[mfp]

On case-insensitive, case-preserving filesystems, omake finds invalid dependencies corresponding to paths that do not exist on disk since they have the wrong capitalization.

In particular, given module A that depends on B, and the corresponding source files a.ml and b.ml, if you have a target like OCamlPackage(runtime, a b), omake will invoke ocamlc as in ocamlc -pack -o runtime.cmo B.cmo a.cmo. Now, on a case-insensitive, case-preserving filesystem, b.cmo can be open(2)ed under both b.cmo and B.cmo paths, but ocamlc fails -- it must be doing readdir and sees B.cmo is missing.

This is caused by a bug in the ocamldep-postproc builtin. I verified this by overriding its definition in an affected codebase (extprot) under macOS to use OCamlScannerPostproc, which did work.

This PR fixes the builtin itself.

Credits

The work on this PR was performed on behalf of Ahrefs open-source work day:
https://twitter.com/javierwchavarri/status/1466774912148910088

[mfp]

This change was required for omake's bootstrapping to work.

[rr0gi]

readdir inside stat (wrapper) sounds expensive

[mfp]

It is :-(, but I'm not aware of any BSD/POSIX syscall/API to obtain the original capitalization -- I read even realpath(3) lies. And I don't know how to determine whether the filesystem is case-insensitive to restrict the above check to that case (without write ops that is). At least it's performed only when there's a successful (l)stat.

The only alternative I can think of is finding the mount point, probing the FS and caching this info, but it still feels a bit dirty.

[gerdstolpmann]

I'd rather say that this is a bug/problem in ocamlc. We really cannot afford to do a readdir on each stat.