CI: Remove Shippable & move AArch64 package generation to Travis

[kinke]

As the service is sadly no more.

[Geod24]

You probably want to migrate to .com in the next 10 days, too: https://blog.travis-ci.com/2021-05-07-orgshutdown
But then you'll be hit with the rate limiting which made so many people drop Travis.

[kinke]

Argh, thx for the heads-up. The announcement that they'd drop .org in the upcoming weeks has been there for months, so I started not believing it. ;) - Too bad, I was positively surprised about the recent performance gains (AArch64 job done in ~23 mins) and thought it would allow for a smooth Shippable replacement, incl. AArch64 Linux package generation. IIRC, I don't have sufficient rights to migrate our existing account (been playing with .com a while back with a personal account, to play with their AArch64 Graviton CPU).

[kinke]

Oh, according to https://docs.travis-ci.com/user/billing-overview/#partner-queue-solution, Travis' AArch64 offering is sponsored by ARM and won't cost any credits.

[kinke]

@dnadlinger @redstar: I cannot migrate the Travis ldc repo to .com, one of you guys will have to do it: https://travis-ci.com/organizations/ldc-developers/migrate. Edit: Oh and please migrate llvm-project too, I hope it's fast enough for prebuilding LLVM too.

[kinke]

TODO: add reggae as well, edit: and mimalloc

[kinke]

Alright, Travis and thus native AArch64 CI on Linux is back in business, and in better shape than ever, probably due to a combination of LLVM 12 and a more recent base image:

• lit-tests results aren't ignored any longer, just 2 tests had to be disabled (facilitated by dynamic-compile not working with LLVM 12).
• All gdb tests (except for one lit-test) now work and have been enabled.
• LTO integration tests re-enabled as well.
• Using default ld.gold for CI now, no more countless linker warnings requiring ld.bfd.
• Prebuilt package uses mimalloc and bundles reggae.
• CI performance has been really great so far (~25 mins), incl. single-core performance (multiple times higher than Shippable's). Overall, about twice as fast for the full LDC job here compared to Shippable; bottleneck being once more available RAM, not number of CPU cores (the lit-tests complete in ~10 seconds with 80 threads...).

[JohanEngelen]

This is great news, nice work!