aarch64 support

@gvolcy what needs to be done:

bootstrap ghcup on aarch64 and add it to the release cycle
add the ghc URLs for aarch64
bootstrap cabal for aarch64 (there are no tarballs for it)

hey!!I want to know if i could run the Haskell node on the rock pi 4b for Cardano stakepool right now I'm running rust node. Let me the know if you have any hope me
Thank you

Afais, aarch64 support for GHC is semi-broken? ghc/ghc#15275 (comment 272630)

I've found this issue after running the tests on ghc 8.6.5 because like @gvolcy I want to run the cardano node on a Pi4.

I managed to compile GHC 8.6.5 on my Pi4 through a systemd-nspawn aarch64 container since I'm using Raspbian withe 64bit kernel and the 32bit user land (it was quite tedious). I also compiled the Cardano Haskell node too. It appears to be working (full sync on current test net with 18 hours uptime)

I ran the GHC normal test suite for 8.6.5 and here is the list of the failing tests (283):

TEST="ClosedFam1TH ClosedFam2TH ManyAlternatives ManyConstructors MultiLayerModules T10052 T10279 T10306 T10508_api T10547 T10596 T10598_TH T10603 T10620 T10638 T10697_decided_1 T10697_decided_2 T10697_decided_3 T10796a T10796b T10810 T10811 T10820 T10828 T10828a T10828b T10858 T10891 T10989 T11046 T11145 T11195 T11276 T11303b T11341 T11345 T11374 T11463 T11484 T11629 T11680 T11721_TH T11797 T11809 T12227 T12387 T12403 T12407 T12478_1 T12478_2 T12478_3 T12478_4 T12478_5 T12503 T12513 T12530 T12545 T12646 T12707 T12977 T12993 T13018 T13035 T13098 T13123 T13366 T13379 T13618 T13642 T13701 T13719 T13776 T13781 T13782 T13825-ghci T13837 T13856 T13885 T13887 T13968 T14052 T14060 T14204 T14298 T14646 T14681 T14683 T14697 T1476 T14817 T14843 T14875 T14885a T14885b T14885c T15164 T15243 T15324 T15331 T15365 T1541 T15481 T15502 T15518 T15550 T15572 T1835 T1969 T2182ghci T2222 T2674 T2700 T2713 T2817 T3016 T3064 T3100 T3177 T3177a T3319 T3395 T3424 T3920 T4029 T4124 T4128 T4135 T4188 T4364 T4436 T4801 T5030 T5037 T5217 T5290 T5313 T5321FD T5321Fun T5358 T5362 T5379 T5404 T5410 T5435_v_asm_b T5435_v_gcc T5452 T5508 T5631 T5642 T5837 T5882 T5883 T5968 T5971 T5976 T6005 T6005a T6018th T6106 T6114 T7241 T7477 T7484 T7667 T7667a T7681 T783 T7910 T8031 T8100 T8186 T8353 T8412 T8499 T8507 T8628 T8639_api T8759 T8761 T8807 T8884 T8932 T8953 T8954 T8987 T9064 T9066 T9081 T9199 T9233 T9262 T9293 T9630 T9692 T9738 T9872a T9872b T9872c T9872d T9961 TH_1tuple TH_Lift TH_Promoted1Tuple TH_PromotedList TH_PromotedTuple TH_RichKinds TH_RichKinds2 TH_Roles1 TH_Roles2 TH_Roles3 TH_Roles4 TH_StaticPointers TH_StringPrimL TH_TyInstWhere1 TH_TyInstWhere2 TH_class1 TH_dupdecl TH_emptycase TH_exn1 TH_exn2 TH_fail TH_finalizer TH_foreignCallingConventions TH_foreignInterruptible TH_linePragma TH_mkName TH_overlaps TH_overloadedlabels TH_pragma TH_recover TH_recover_warns TH_reifyDecl1 TH_reifyDecl2 TH_reifyInstances TH_reifyLocalDefs TH_reifyLocalDefs2 TH_reifyMkName TH_repE2 TH_repGuard TH_repGuardOutput TH_repPrim TH_repPrim2 TH_repPrimOutput TH_repPrimOutput2 TH_repUnboxedTuples TH_runIO TH_scopedTvs TH_sections TH_spliceD2 TH_spliceDecl1 TH_spliceDecl2 TH_spliceE1 TH_spliceE3 TH_spliceE4 TH_spliceE6 TH_spliceExpr1 TH_spliceGuard TH_spliceInst TH_tf1 TH_tf3 TH_tuple1 TH_unboxedSingleton TH_unresolvedInfix2 andy_cherry barton-mangler-bug break008 break009 break026 divbyzero dynCompileExpr galois_raytrace ghci.prog009 ghci.prog010 ghci004 ghci006 ghci025 ghci038 ghci057 ghci058 ghci062 joao-circular linker_unload linkwhole parsing001 plugin-recomp-flags plugin-recomp-pure print007 prog001 prog002 prog003 prog012 prog013 recomp011 recomp015 regalloc_unit_tests"

I don't know Haskell nor GHC so I am not sure if the failure of these tests is a problem for the Cardano node. Most of them involve something called RTS apparently... Note that I compiled GHC 8.6.5 using GHC 8.4.4 (Debian Buster's packaged GHC which is available for aarch64. The normal upgrade procedure doesn't work though since aarch64 builds are not available on the Haskell side).

Among the above tests many are performance related where too much memory seems to be allocated (in the range of +600% of the expected for some tests)

Quote from the GHC channel:

15:17 maerwald: cardano Node is not officially supported on aarch64. Running it with anything but +RTS -N1, will almost certainly deadlock at some point. armv7 is most likely broken due to Word size issues. Also cardano-node doesn’t compile for anything other than 8.4 afaik.

Running it with anything but +RTS -N1, will almost certainly deadlock at some point.

Indeed, I experienced a few segmentation faults... so I did some researches and back-ported a few patches unto my 8.6.5 install ( ghc/ghc!734 (closed) (with a bit from ghc/ghc!1128 (merged)) and ghc/ghc!1139 (closed) ) - I chose !734 over ghc/ghc!1128 (merged) because of ghc/ghc#18201 (closed) where it's mentioned that ghc 8.6.4 + !734 is being actively used on Aarch64 with great success but !734 alone introduced a problem in /rts/sm/MarkWeak.c so I used !1128 for this file alone.

Less tests are failing now..

Let's see if it's more stable and sorry for polluting this thread.

I think the main thing missing for this to move forward is the cabal aarch binaries. There were attempts at improving the cabal CI, but it seems those efforts are stale:

But... ghcup is not technically constrained to upstream cabal bindists. We can build and ship our own. I have not gotten around doing that.

added statushelp-needed label

There are cabal aarch bindists for the rc: https://oleg.fi/cabal-install-3.4.0.0-rc3/

changed milestone to %0.1.13

release candidate with aarch64 binaries is here https://downloads.haskell.org/~ghcup/0.1.13-rc2/

closed

aarch64 support

Child items ...

Activity