gitlab-ci: build the Switch core's combined archive

[WizzardSK]

Follow-up to #2465: the Switch job builds, but the archive it hands to RetroArch is incomplete.

The libnx template runs cmake --build $BUILD_DIR --target ${CORENAME}_libretro, and on Switch that target is flycast's own objects and nothing else. The dependencies — xxhash, chdr, zip, elf and the resources — are folded in by the combined target that CMakeLists.txt defines for exactly this case:

add_custom_target(combined ALL
    COMMAND ${CMAKE_AR} -x $<TARGET_FILE:xxHash::xxhash>
    ...
    COMMAND ${CMAKE_AR} -rs flycast_libretro_libnx.a *.o

ALL means a plain cmake --build gets it; naming a single target does not. So the job takes the template's script with that one word changed.

Measured on a devkitA64 container, configuring exactly the way the template does:

target	members	size	chd.o / xxhash.o / zip_*.o
flycast_libretro	251	160,395,522	absent
combined	409	171,338,952	present

The run builds the first, checks the archive, then builds the second and checks again.

Without this, make -f Makefile.libnx in retroarch-precompiled links against 158 missing objects.

The alternative would be to move the combined commands into a POST_BUILD step on the library target, so that building flycast_libretro always produces the complete archive and no CI knows about it — happy to do that instead if you prefer it.

[WizzardSK]

One thing worth knowing: the merge commit carries [ci skip], so the buildbot skipped the pipeline for it (108292), same as for #2465 and #2466. That is fine for most CI-only changes, but it does mean this particular one - a change to the Switch job itself - has not been exercised yet; it will be on the next master build the buildbot triggers.

[flyinghead]

Sorry about that. I didn't want to trigger a build since this file isn't used here.

[WizzardSK]

Sorry, I looks like a misunderstanding. My intention was to move the switch build from this project https://github.com/libretro/flycast to your project because I believe this project is more mature and up to date than that old switch-only fork.

[flyinghead]

There is indeed some misunderstanding. .gitlab-ci.yml is the libretro build file, and it's only used by libretro for their own builds. This project doesn't build on gitlab and uses the build files in .github/workflows. We do build the libretro core for some platforms but only to verify that the build is working and for testing. We don't publish these builds.
You should probably talk to the libretro team to sort this out.

[WizzardSK]

That matches my understanding, and it is the reason the PR came here rather than to libretro: the buildbot builds this repository. The recipe in libretro-super reads flycast … https://github.com/flyinghead/flycast master, and git.libretro.com/libretro/flycast-upstream is a mirror of it, so .gitlab-ci.yml only has an effect where it lives now — libretro cannot fix it on their side without carrying a permanent divergence from your tree. Nothing in the PR touches .github/workflows or how you build and publish.

If you would rather not carry libretro's CI quirks in this repo at all, there is a cleaner version of this particular fix: move the combined commands into a POST_BUILD step on the library target, so cmake --build --target flycast_libretro produces a complete archive by itself and no CI file needs to know about it. I am happy to send that instead, and to take the question of which repository the Switch core should come from to the libretro side.