Fix dune path in nested shells

[gridbugs]

When starting a shell from an existing shell where the shell config has already been sourced, or when sourcing a shell config for a second time within the same shell session, any opam setup in the shell config will re-insert the bin directory from the current opam switch at the beginning of PATH.

Dune's setup logic doesn't re-insert directories already present in PATH as dune may share its bin directory with other programs reliant on the current PATH order. Thus opam's re-initialization will result in any dune executable in the current opam switch taking precedence over the binary distro.

To prevent this, the __dune_env function detects when it's called subsequent times in a shell session, and if it succeeded the first time in correctly setting up the environment for the dune binary distribution will remove any opam directories from PATH which preceed the first instance of location of the dune binary distro in PATH. Opam bin directories following the location of the binary distro in PATH, as well as all other directories in PATH are left unchanged.

[gridbugs]

Fixes #10

[gridbugs]

@Sudha247 this is now ready for review. Heads up that I replaced the awk script we wrote with a sed script. The reason is that the pattern the script needs to recognize is a path like "/home/user/.opam/bin", and the slashes in the path interfere with awk's regex syntax (/pattern/ { ... }). The benefit of sed is that the pattern delimiter character does not need to be slash, which makes it much easier to match patterns which are file paths.

The sed script is sed "1,\#^$__DUNE_ROOT/bin\$# { \#^$HOME/.opam#d; }". It uses # as a pattern delimiter. It applies logic on the range of lines from its input 1,\#^$__DUNE_ROOT/bin\$#, which refers to the range from line 1 to the first occurrence of the value of $__DUNE_ROOT/bin ($__DUNE_ROOT is expanded as a shell variable before sed sees it) which is the dune binary distro's bin directory. The range is inclusive at both ends. For lines within the range, sed runs \#^$HOME/.opam#d. This has 2 parts. \#^$HOME/.opam# is another pattern matching paths that look like opam switch bin dirs (again, $HOME is expanded as a shell variable before sed sees it so it just sees the variable's value). The d is the "delete" command which omits matching lines from output. Sed's default behaviour is to print each line, so lines outside the range, and lines inside the range which don't look like opam bin dirs are printed.

[Sudha247]

Looks good to me! One question below.

This is a very impressive script and I learnt a lot of new stuff about shell scripting by reviewing these PRs! Thanks for the work. However, I do wonder about the maintenance of the env scripts - we have the same logic coded to four different shells. Future changes need to be cognisant of the fact that the changes should occur in all of the env scripts. Would we benefit from fleshing this out in a meta script and importing it to the various configs? (this is not to block the PR)

[Sudha247]

Is __DUNE_SETUP_STATE not being exported here because it's already in success state, and we want to continue it being in the same state?

[gridbugs]

Yep exactly

[gridbugs]

However, I do wonder about the maintenance of the env scripts - we have the same logic coded to four different shells. Future changes need to be cognisant of the fact that the changes should occur in all of the env scripts. Would we benefit from fleshing this out in a meta script and importing it to the various configs? (this is not to block the PR)

It's tricky because these files are sourced in the user's shell config, so they have to be written in the user's shell language. If there was some common language that could be compiled to bash, zsh, fish, and posix shell but which also allowed us to add additional logic so the completion script could be loaded for bash and zsh but not other shells then that would let us centralize the logic. However I'm not aware of such a language. For the time being I think we should rely on the tests and code review process to ensure that changes are applied to each file.