Update LLVM to 15.0.7

[vchuravy]

• Fix attributes and opaque pointer usage for LLVM15 #49528
• Support both Float16 ABIs depending on LLVM and platform #49527
• Use NewPM for ASAN/MSAN #49530

[gbaraldi]

The clang++ symlink is missing. Do you have any idea why @vchuravy ?

[vchuravy]

Because symlinks are lies? No not sure why that one would be missing, but I am pretty sure we don't require it as an output.

[gbaraldi]

It does seem some things were expecting it to be there 🤷

[vchuravy]

Can you for now just create it in place and then we can go back and look where in the process of building Clang.jl we lost it.

[gbaraldi]

I will see if it works at least locally, but it seems something very odd is going on with the tools.

[gbaraldi]

@vchuravy I think this might be ready for a pkgeval. I'm not sure what's up with ppc, and the other failures are pending me rebuilding LLVM but I'm kinda waiting to see if we don't find anything else that might also want to go into the rebuild.

The only test that failed is sockets which is kind of a known issue.

[vchuravy]

@nanosoldier runtests(configuration = (buildflags=["LLVM_ASSERTIONS=1", "FORCE_ASSERTIONS=1"],), vs_configuration = (buildflags = ["LLVM_ASSERTIONS=1", "FORCE_ASSERTIONS=1"],))

[vchuravy]

I will try to find some cycles for the PPC failure.

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected.
A full report can be found here.

[gbaraldi]

That's a surprisingly good result. I see most failures are from the LLVM world.

About the ppc thing I believe it's Float16 related.

[vchuravy]

PPC issue seem PWR8 specific:

error during bootstrap:
LoadError(at "sysimg.jl" line 3: LoadError(at "Base.jl" line 206: LoadError(at "float.jl" line 862: StackOverflowError())))                                                          
unknown function (ip: 0x7fff9a8900b8)
unknown function (ip: 0x7fff9a89001f)

*** This error might be fixed by running `make clean`. If the error persists, try `make cleanall`. ***                                                                               
make[1]: *** [/nobackup/users/vchuravy/dev/julia/sysimage.mk:67: /nobackup/users/vchuravy/dev/julia/jl-pwr8/usr/lib/julia/sys.ji] Error 1                                            
make: *** [/nobackup/users/vchuravy/dev/julia/Makefile:97: julia-sysimg-ji] Error 2
[vchuravy@satori-login-001 jl-pwr8]$ vim ../base/float.jl

Whereas PWR9 completes the build successfully.

[gbaraldi]

That is odd, do any of them support native Float16? Also StackOverflow is a funny error to have here, I do wonder what the backtrace is.

[vchuravy]

That is odd, do any of them support native Float16? Also StackOverflow is a funny error to have here, I do wonder what the backtrace is.

Experience tells me that this is likely a hw loop issue... The curse of PWR8. Trying to get see if I can get a backtrace and maybe that gives me a better indication.
Since I have a sysimage built for PPC9 I might be able to reproduce this in JIT mode.

[vchuravy]

AH nothing as nefarious as I thought.

Thread 1 "julia" received signal SIGSEGV, Segmentation fault.                                                                                                                         
0x00007ffff71b8e14 in julia__gnu_f2h_ieee (param=-128) at /nobackup/users/vchuravy/dev/julia/src/runtime_intrinsics.c:195                                                             
195         return float_to_half(param);                                                                                                                                              
(gdb)                                                                                                                                                                                 
(gdb) bt                                                                                                                                                                              
#0  0x00007ffff71b8e14 in julia__gnu_f2h_ieee (param=-128) at /nobackup/users/vchuravy/dev/julia/src/runtime_intrinsics.c:195                                                         
#1  0x00007fffe9ee00c0 in __gnu_f2h_ieee ()
...
#476 0x00007fffe9ee0020 in __gnu_h2f_ieee ()
#477 0x00007fffe9ee00cc in __gnu_f2h_ieee ()
...
#130712 0x00007fffea4e52c4 in julia_Float16_4059 () at float.jl:159                                                                                                                   
#130713 0x00007fffea4e5360 in jfptr_Float16_4060 ()                                                                                                                                   
#130714 0x00007ffff714fa78 in _jl_invoke (world=9137, mfunc=<optimized out>, nargs=1, args=0x7fffffffa738, F=0x7fffe87caff0) at /nobackup/users/vchuravy/dev/julia/src/gf.c:2720      
#130715 ijl_apply_generic (F=0x7fffe87caff0, args=0x7fffffffa738, nargs=<optimized out>) at /nobackup/users/vchuravy/dev/julia/src/gf.c:2921 

And it's reproducible on the PWR9 build with -C pwr8

So our interposer is calling itself xD

[gbaraldi]

@nanosoldier runbenchmarks()

[gbaraldi]

@nanosoldier runbenchmarks(ALL, vs = "master")

[gbaraldi]

The sanitizer passes are just gone from the old PM, so we have to turn the new one on here @pchintalapudi

[vchuravy]

@nanosoldier runtests(configuration = (buildflags=["LLVM_ASSERTIONS=1", "FORCE_ASSERTIONS=1"],), vs_configuration = (buildflags = ["LLVM_ASSERTIONS=1", "FORCE_ASSERTIONS=1"],))

[vchuravy]

@nanosoldier runbenchmarks(!"scalar", vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected.
A full report can be found here.

[vchuravy]

@nanosoldier runbenchmarks(!"scalar", vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[vchuravy]

@nanosoldier runbenchmarks(!"scalar", vs=":master")

[vchuravy]

We are expecting nanosoldier to come back with happy news and intend to merge this PR thereafter.

There is a second step that maybe @Keno or others could help with and that is to enable opaque-pointers,
but currently opqaue-pointers regresses performance to much.

Overall LLVM 15 is an important stepping stone for improved GPU support (primarily AMD), the migration the NewPM infrastructure.

This PR raises the minimum MacOS version to 10.14 and raises toolchain requirements as well.

With LLVM 15.x we will raise the version requirements of the toolchain used to build LLVM. The new requirements are as follows:

GCC >= 7.1
Clang >= 5.0
Apple Clang >= 9.3
Visual Studio 2019 >= 16.7

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[vchuravy]

Many thanks to @gbaraldi and @pchintalapudi for helping out massively with this PR. This turned out to be a much harder upgrade than usual.