Support s390x

[Kaiden42]

Hello, I'm referring to Issue #21474. I've been working in programming for a few years at one of the largest German insurance companies (R+V Versicherung) and I'm very interested in Julia. Since our company strategically relies on the mainframe alongside other platforms, I asked our management if we could provide community access to our s390x machines. The management is also very interested that we gain experience with Julia because there are many mathematical operations in the company that need to be processed. So I talked to our z/OS and zLinux groups and we have some questions for you:

• is there still any interest in a port for the s390x?
• do you need access to zLinux or zOS?
• what parameters do you need for a z guest roughly (mem, disk, IO bandwidth, #cpu, ...)?

We would be happy to support the community.

[cdluminate]

Let me add a hint: s390x architecture lacks libunwind support. After disabling libunwind, the buildlog suggests that some work is needed in LLVM. https://launchpad.net/~lumin0/+archive/ubuntu/ppa/+sourcepub/9482958/+listing-archive-extra https://launchpadlibrarian.net/393051331/buildlog_ubuntu-cosmic-s390x.julia_1.0.1-2~u1_BUILDING.txt.gz

[ViralBShah]

I think it would be best to approaching ibm to support this commercially. In the past ibm has given small sums of money for a ppc port as a one off but not at a sustainable way.

[ViralBShah]

Also isn't it big endian? That would be a considerable porting effort, quite likely.

[cdluminate]

Also isn't it big endian? That would be a considerable porting effort, quite likely.

s390x is Big Endian.

[ViralBShah]

I think at this time the best thing would be to go with one of the 1.0 supported architectures. If supporting s390x is essential, it would be best to start a joint dialogue with IBM.

[cfilleke]

• Julia full 1.0.5 (tarball download) compiles cleanly with make debug USE_SYSTEM_PCRE=1 USE_SYSTEM_LIBM=1 DISABLE_LIBUNWIND=1 CFLAGS=-DNDEBUG on a z/VM running RHEL 8.1
  • all prerequisite tools installed and dependencies resolved out of the official @redhat repositories.
• it bails on bootstrap so looking at the debug compile through the eyes of gdb
• compiles with same bootstrap error with and without debug flags

  $make debug  USE_SYSTEM_PCRE=1 USE_SYSTEM_LIBM=1 DISABLE_LIBUNWIND=1 CFLAGS=-DNDEBUG
  ...
  JULIA usr/lib/julia/corecompiler.ji
  LLVM ERROR: Cannot select: 0x6dee280: i64 = addrspacecast[0 -> 10] 0x6dee218
  0x6dee218: i64,ch = load<LD8[@"jl_global#16"](tbaa=<0x6d125b8>)(dereferenceable)(invariant)> 0x6ded100:1, 0x6ddee50, undef:i64
  0x6ddee50: i64,ch = load<LD8[GOT]> 0x6d79fe0, 0x6dee5c0, undef:i64
    0x6dee5c0: i64 = SystemZISD::PCREL_WRAPPER TargetGlobalAddress:i64<%jl_value_t** @"jl_global#16"> 0 [TF=1]
      0x6dee968: i64 = TargetGlobalAddress<%jl_value_t** @"jl_global#16"> 0 [TF=1]
    0x6ddecb0: i64 = undef
  0x6ddecb0: i64 = undef
  In function: japi1_top-level scope_0
  make[1]: *** [Makefile:188: /root/julia-1.0.5/usr/lib/julia/corecompiler.ji] Error 1
  make: *** [Makefile:75: julia-sysimg] Error 2

• Not too familiar with the code base, but it looks like the PCRE wrapper and ABI code would be the first place to start looking at architectural mods that would need to be made
• Any further hints as to where in the julia code a high concentration of little endian assumptions are made would be useful
• Looking at seeing if IBM's libunwind mods for Z can be used
• Julia on Z is essential imho @edelsohn

[alfasst]

Some update. Tried to build the latest version of Julia - v1.6.1 on an IBM Cloud Hyper Protect Virtual Server-iz. Got this error.

make[2]: *** No rule to make target 'trampolines/trampolines_s390x.S', needed by 'build/trampolines/trampolines_s390x.o'.  Stop.
/root/julia/deps/blastrampoline.mk:14: recipe for target 'scratch/blastrampoline-23de7a09bf354fe6f655c457bab5bf47fdd2486d/build-compiled' failed
make[1]: *** [scratch/blastrampoline-23de7a09bf354fe6f655c457bab5bf47fdd2486d/build-compiled] Error 2
Makefile:60: recipe for target 'julia-deps' failed
make: *** [julia-deps] Error 2

[StefanKarpinski]

That's a new feature where some simple assembly code is needed on each platform to allow dynamically swapping out BLAS libraries, which is a common request since many people want to use MKL on Intel. The trampoline code for s390x will need to be provided.