Clean out old msys2/mingw stuff in CI

[workingjubilee]

Recently we had increasing network failures due to redownloading mingw tools. The first few times this fell to the usual "retry it again", but that's really not a great solution as we were losing quite a few half-finished runs for essentially no reason. An attempt to fix it caused some struggle with rollups due to the first try at a patch not quite panning out. In other words, we could use some more attention to the windows-gnu CI setup.

Per @mati865 in #125546 (comment) we have shell scripts that we run in our CI that has somewhat arcane names, as it discusses old tools we don't actually use.

@ChrisDenton intends to follow up but given all the above-mentioned network failures I wanted to make sure there was an issue for this.

Chris thinks this was enabled by #121876 in part. Thanks @jieyouxu and @Oneirical!

[mati865]

Chris thinks this was enabled by #121876 in part. Thanks @jieyouxu and @Oneirical!

I don't think this was the case since there are still many tests that are not ported like for an example run-make-fulldeps. Maybe GHA now include make on Windows by the default (or by an accident)?

Guess we'll find out soon.

[jieyouxu]

I don't know if run-make (and rmake-fulldeps) tests depend on msys2, especially as mentioned not all tests have been ported. make is still definitely a dependency.

[ChrisDenton]

To be clear, my hypothesis is that at least some run-make tests assumed the existence of many gnu-ish tools so the switch to recipes (even if not complete) has greatly reduced the things that need to be assumed.

[ChrisDenton]

An alternative explanation is that the comment in the file was just always wrong ;)

[ChrisDenton]

Recently we had increasing network failures due to redownloading mingw tools

Note that this has been randomly failing for months now, basically since it was introduced. The only difference here was it failing a number of times in succession. Which might have just been a random coincidence or might not have been.

An attempt to fix it caused some struggle with rollups due to the first try at a patch not quite panning out. In other words, we could use some more attention to the windows-gnu CI setup.

The attempt failed because I forgot to test windows-msvc. The windows-gnu part worked fine.

[workingjubilee]

Note that this has been randomly failing for months now, basically since it was introduced.

oh lol I didn't notice.

[mati865]

To be clear, my hypothesis is that at least some run-make tests assumed the existence of many gnu-ish tools so the switch to recipes (even if not complete) has greatly reduced the things that need to be assumed.

They always required make provided by MSYS2 and still do (albeit something else seems to provide it now). In the past they also used to depend on toolchain specific tools like nm, objdump but those were provided by the toolchain and later migrated to LLVM tools that Rust always builds anyway.

So while porting of run-make tests to Rust is fantastic it doesn't impact removal of MSYS2. If you want to know what allowed the removal you can run which make.exe on CI.

[ChrisDenton]

I'm not at all doubting the make part is still needed.

[workingjubilee]

was just trying to throw out a kudos, didn't mean to start an argument. :ferrisCluelesser:

[jieyouxu]

I opened an issue #125585 to remind us that we should eventually document all assumed external dependecies for the test suites.