Bump llvm version required for Power and backport bugfix.

[vchuravy]

This backports a fix for #18555. The patch does not apply cleanly to LLVM 3.8 so I bumped the required version for Power support. I will try and nominate this patch as a backport to 3.9.1.

cc: @ViralBShah

[tkelman]

add the checksums for llvm 3.9.0

[tkelman]

This will raise minimum cmake version to 3.4.3 on power right? I'm guessing cmake doesn't provide convenient binaries for power, did you build cmake from source or use a really new distro?

[vchuravy]

Yes I compiled CMake and CMake doen's provide binaries for power as far as I can tell.

The tests still fail:

    JULIA test/threads
     * threads              Warning: threaded loop executed in order
Test Failed
  Expression: x[] == m
   Evaluated: 0 == -5000
ERROR: LoadError: LoadError: There was an error during testing
 in record(::Base.Test.FallbackTestSet, ::Base.Test.Fail) at ./test.jl:397
 in do_test(::Base.Test.Returned, ::Expr) at ./test.jl:281
 in test_threaded_atomic_minmax(::Int16, ::Int16) at /home/valentin/julia/test/threads.jl:56

I put the llvm ir and code_native here https://gist.github.com/vchuravy/ba27d29870e49df5f640ff0b5aa635c3 if anybody is interested

[tkelman]

Or do any of the recent-cmake PPA's support power? I was planning on writing a contrib/download_cmake.sh script once we upgrade to LLVM 3.9 for the sake of Linux x86 and amd64 (it's easy enough to get from homebrew and cygwin elsewhere), but maybe with exotic architectures should we think about writing a makefile to automate building cmake ourselves?

[vchuravy]

I think we should probably deal with CMake like with any other dependency. If in doubt download it and build it ourselves :)

[tkelman]

Build-only dependencies have generally gotten a bit different treatment if they're easy enough to get binaries for good-enough versions, like gmake and gfortran.

[vchuravy]

added the checksums.

[tkelman]

lgtm. did 3.9 include D14260?

[tkelman]

@staticfloat @ViralBShah what version of cmake is on the power buildbot? It's running ubuntu 14.04 right, do we need to build a newer version to keep nightlies working there?

[vchuravy]

@tkelman I think it is in 3.9.0

[valentin@ibm-power2 llvm-svn]$ git log --grep=D14260
commit 1ca1fcaa5b4c75a65a202badfd5df8240a36ca0f
Author: Arch D. Robison <arch.robison@intel.com>
Date:   Mon Apr 25 22:22:39 2016 +0000

    Optimize store of "bitcast" from vector to aggregate.

[valentin@ibm-power2 llvm-svn]$ git branch --contains 1ca1fca
  master
* release_39

[staticfloat]

Yeah, we're going to need to build a new cmake on pretty much everything. (Except osx, because that gets cmake from Homebrew)

$ ansible julia_buildslaves -m shell -a 'export PATH=$PATH:/usr/local/bin && cmake --version 2>/dev/null | head -1' -i ./hosts

buildslave_osx10.10-x64 | SUCCESS | rc=0 >>
cmake version 3.6.1

buildslave_osx10.11-x64 | SUCCESS | rc=0 >>
cmake version 3.6.1

buildslave_osx10.9-x64 | SUCCESS | rc=0 >>
cmake version 3.6.1

buildslave_ubuntu14.04-x86 | SUCCESS | rc=0 >>
cmake version 2.8.12.2

buildslave_ubuntu14.04-x64 | SUCCESS | rc=0 >>
cmake version 2.8.12.2

buildslave_centos5.11-x86 | SUCCESS | rc=0 >>
cmake version 2.6-patch 4

buildslave_centos5.11-x64 | SUCCESS | rc=0 >>
cmake version 2.6-patch 4

buildslave_centos7.1-x64 | SUCCESS | rc=0 >>
cmake version 2.8.11

buildslave_centos6.7-x64 | SUCCESS | rc=0 >>
cmake version 2.8.12.2

buildslave_ubuntu14.04-armv7l | SUCCESS | rc=0 >>
cmake version 2.8.12.2

buildslave_centos7.2-ppc64le | SUCCESS | rc=0 >>
cmake version 2.8.11

buildslave_win6.2-x86 | SUCCESS | rc=0 >>
cmake version 3.3.2

buildslave_win6.2-x64 | SUCCESS | rc=0 >>
cmake version 3.3.2

I'll do this once the buildbot reconfig window opens, after 0.5.0 final is built.

EDIT: Issue created.

[tkelman]

I'm the cmake maintainer on cygwin and uploaded 3.6.2 there recently. When you next upgrade cygwin packages please be very careful to not upgrade the mingw gcc cross-compilers to 5.4 though, we need to keep those held back to 4.9 until we have a solution for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77333

[ViralBShah]

Would this address all threading issues on power at least as far as the tests go?

[vchuravy]

The last two commits work around an upstream issue and all tests pass now.

[ViralBShah]

Should we backport this for 0.5?

[ViralBShah]

Once CI completes, let's merge this. Marking as a backport candidate.

[ViralBShah]

I just realized that the old cmake will break the buildbot. Should we revert, or just fix the cmake? I am happy to help fix the cmake immediately for the build environment.

[ViralBShah]

I just built cmake for myself. If we build it in the buildbot account and add it to the PATH, will that suffice? If somone can DM me the password to the account, I will do it. Seems like it should be easy enough to fix up.

[staticfloat]

@ViralBShah which buildbot are you referring to? We were planning on upgrading cmake after all our 0.5.0 and 0.4.7 binaries are built and verified working.

[ViralBShah]

The power one.

[staticfloat]

Ah, we specifically upgraded cmake on the ppc64le buildbot, so no worries. :)

[ViralBShah]

Ok - great! Thanks.

[vchuravy]

@tkelman to backport this I would open a PR against release-0.5?

[tkelman]

See the steps at #17418. Does that need to be done now or can it wait until we run pkgeval for 0.5.1?

[vchuravy]

It can wait till 0.5.1 in my opinion. Ping me when you want to backport it
and I can prepare it.

On Wed, 21 Sep 2016 at 02:08 Tony Kelman notifications@github.com wrote:

See the steps at #17418 #17418.
Does that need to be done now or can it wait until we run pkgeval for 0.5.1?

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#18557 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAI3anoAVbP7qf3HI614gJs-jXNT2c7Sks5qsBMbgaJpZM4J_dOf
.