Feature proposal: opam install --justinstall

[AltGr]

There is currently still something that feels like a hole in our workflows, when working in an opam-aware project: these two are very good

• use opam install . --deps-only to have opam handle the local source dependencies
• use make to compile by hand

These can be enough, but sometimes, you need to also test the installation of your package by hand, and things get a bit worse:

• either run opam install ., but if you do it repeatedly, there will be no incremental compilation, so this can get quite long
• or run make install (possibly calling opam-installer if you use .install files), to use the results of your make. But then the resulting installation gets into your opam prefix, without being registered by opam, so this is a bit unclean (I admit I am using this a lot, though).

The option --reuse-build-dir and --inplace-build have somehow been attempts at helping here, but are not really satisfactory to that purpose: they are a bit heavy to use, bring some quirks and don't always behave as expected (e.g. --inplace-build will re-run ./configure with its own options from the opam file in your project dir, which may get annoying).

---

Enough for the summary of the issue, here is an idea to fill that gap:
let the user run the builds with their own tools, but let opam handle just the install part, from the source directory, and assuming the build was done properly. Could be something like:

opam install --justinstall foo

assuming foo is a locally-pinned package, or a directory containing one or more opam packages, run their installation instructions, including handling their .install and .config files if any, and register their installation into opam (then possibly recompiling dependents if any). Another plus is that uninstallation can then be cleanly handled by opam.

This could typically be used just after make, or directly from a Makefile install rule. Implementation-wise, care should be taken that the command is not allowed to modify any dependency of foo.

I'd like to have the input from the users here!

Another idea could be to just extend opam-installer (or a related command built-in opam) so that it registers the installation of the package within opam. It would certainly be simpler, but wouldn't handle installation outside of .install files, consistency with the package definition, or recompilation of dependents.

[Drup]

That sounds quite useful. I think it could also solve many of @gasche's woes regarding compiler development with opam.

[gasche]

The feature you propose sounds very nice. I have no strong opinion over whether it should be in opam install or opam-installer -- having it in the opam binary makes the setup slightly simpler, but well.

(I hadn't thought about the way opam-installer is limited by not being able to register packages within the opam database. Would it make sense for the (un)registration features to be available in an advanced root-administration tool provided in the opam library, that opam-installer or users could call to notify opam of such changes to their roots?)

[gasche]

P.S.: I would like to re-do my opam-compiler-conf performance tests under rc2, but the update approach that looked the nicest is stalled because, if I understand correctly,opam-devel 2~rc2 is not in opam yet.

[AltGr]

WHAT? Gaaah, I was sure that was merged, should have been done before the announcement !

[gasche]

(For the record, it has just been merged by @AltGr)

[AltGr]

yes, sorry about the delay...

[avsm]

If we do have this flag, it might be better called --install-only to match --deps-only.