Add support for the experimental MiniZinc black-box propagator interface

[Dekker1]

I'm hoping to sneak this in before the seemingly upcoming Gecode release. This adds support for the experimental Blackbox interface for MiniZinc (and solvers). The idea being that a user can delegate some evaluation/propagation to an external program/library.

The implementation for Chuffed and the MiniZinc compiler is also ready, the latter I hope to merge after the former and this merges such that CI will pass the required tests.

This work will be presented this year at ModRef.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[kalvdans]

Please document both the DLL interface functions and the external process protocol, so reviewers can review that the PR adheres to the specification, and so that people can write their own conforming black boxes.

[Dekker1]

Please document both the DLL interface functions and the external process protocol, so reviewers can review that the PR adheres to the specification, and so that people can write their own conforming black boxes.

The documentation for the ABI and executable streaming format will be part of the MiniZinc documentation.

[kalvdans]

The documentation for the ABI and executable streaming format will be part of the MiniZinc documentation.

Great! Do you have a link to it?

[Dekker1]

You can have a look at the current draft in my fork: https://github.com/Dekker1/libminizinc/blob/feat/blackbox/docs/en/blackbox.rst

[kalvdans]

You can have a look at the current draft in my fork: https://github.com/Dekker1/libminizinc/blob/feat/blackbox/docs/en/blackbox.rst

Thanks! Looks well documented. Perhaps mention whether fzn_blackbox function must be threadsafe, and there's a descrepancy who decides the number of output values; in the API it is the caller, in the streaming interface it is the black box.

[Dekker1]

Thanks for the feedback, I've updated the documentation with notes on these points. Regarding the seeming discrepancy, it should be noted that the fzn_blackbox constraint call always declares the amount of expected outputs, so if the stream contains a different amount of return values, this should generally be a runtime error.

[Dekker1]

I just pushed a small update to avoid an additional allocation for each propagate call (using int64_t directly) that I discovered working on the Chuffed implementation. This was a left-over from a previous version of the ABI where we used int. I'd updated the ABI connection, but not the internal run methods.