[LLVM] Update to LLVM 12

[vchuravy]

No description provided.

[vchuravy]

@staticfloat do I need to regenerate the Artifacts.toml in stdlib/libLLVM_jll?

[staticfloat]

There shouldn't be an Artifacts.toml. There can be an StdlibArtifacts.toml, but that should not be checked in; it gets downloaded from GitHub at build time.

[vchuravy]

@yuyichao on AArch64 we are getting:

'-fmi' is not a recognized feature for this target (ignoring feature)
'-pa' is not a recognized feature for this target (ignoring feature)

It seems fmi was renamed in https://reviews.llvm.org/D94081 and pa in https://reviews.llvm.org/D94083

best things is probably to #ifdef on the LLVM version?

[nanosoldier]

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

[chriselrod]

All the failures I looked at originated from VectorizationBase.jl or SIMD.jl not removing experimental and v2 with LLVM 12:
llvm.experimental.vector.reduce.v2.fadd.v4f64
EDIT:
VectorizationBase.jl: JuliaSIMD/VectorizationBase.jl@441988f
SIMD.jl: eschnett/SIMD.jl#86

[nanosoldier]

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

[vchuravy]

All the failures I looked at originated from VectorizationBase.jl or SIMD.jl not removing experimental and v2 with LLVM 12:
llvm.experimental.vector.reduce.v2.fadd.v4f64
EDIT:
VectorizationBase.jl: JuliaSIMD/VectorizationBase.jl@441988f
SIMD.jl: eschnett/SIMD.jl#86

Thanks @chriselrod! Once the SIMD.jl one has been merged and tagged we can rerun PkgEval.

[nanosoldier]

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

[vtjnash]

Wow, that's truly horrendous. How did you manage to make even basic things like + and sin that much slower?

[vchuravy]

How did you manage

Hey now! How did they manage to make everything slower.

I fixed the 30000x regression, that seemed to be a missing pass and we will see what impact that has elsewhere.

[nanosoldier]

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

[vchuravy]

Ok, this is looking much better. Still some regression but they are all in the normal range. I will look at a few of the larger ones.

@staticfloat this is now blocked on getting python3 on the build it's.

[vchuravy]

@DilumAluthge any ideas why buildkite didn't get triggered automatically? There was a webhook triggered.

Manually: https://buildkite.com/julialang/julia/builds/200

[DilumAluthge]

Hmmm. I checked the settings and everything looked correct. Can you try pushing a new commit to this branch?

[vchuravy]

@nanosoldier runtests(["OpenQuantumBase", "LiveServer", "CBinding", "LLVM", "VectorizationBase"])

[vchuravy]

If these come back clean I am inclined to merge this tomorrow.

[nanosoldier]

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

[chriselrod]

Active lane mask fixed in VectorizationBase and tagged: JuliaRegistries/General#37133

Ugh, now actually fixed.

[nanosoldier]

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

[KristofferC]

The ["union", "array", ("perf_simplecopy", "Bool", 1)] and company might be relevant (4x regression).

[vchuravy]

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

[vchuravy]

@nanosoldier runbenchmarks(ALL, vs="@8ae50e0cea280e5cc7310a648c280e90c592ed04")

[vchuravy]

@nanosoldier runtests(ALL, vs = ":master")

[nanosoldier]

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

[nanosoldier]

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

[DilumAluthge]

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

@chriselrod The Tullio failures here seem to be coming from LoopVectorization and/or VectorizationBase? E.g.:

new arrays: Error During Test at /home/pkgeval/.julia/packages/Tullio/IHd6P/test/parsing.jl:4
  Got exception outside of a @test
  llvmcall only supports intrinsic calls
  Stacktrace:
    [1] macro expansion
      @ ~/.julia/packages/VectorizationBase/xil1r/src/llvm_intrin/intrin_funcs.jl:292 [inlined]
    [2] vsum
      @ ~/.julia/packages/VectorizationBase/xil1r/src/llvm_intrin/intrin_funcs.jl:292 [inlined]
    [3] vsum
      @ ~/.julia/packages/VectorizationBase/xil1r/src/llvm_intrin/intrin_funcs.jl:269 [inlined]
    [4] reduced_add
      @ ~/.julia/packages/VectorizationBase/xil1r/src/llvm_intrin/intrin_funcs.jl:319 [inlined]

[chriselrod]

@chriselrod The Tullio failures here seem to be coming from LoopVectorization and/or VectorizationBase? E.g.:

So, Tullio tests:

 [3d5dd08c] VectorizationBase v0.15.7

RegularizedLeastSquares tests:

  [3d5dd08c] + VectorizationBase v0.19.39

They should be on VectorizationBase 0.20.10.

ApproxManifoldProducts had:

  [3d5dd08c] VectorizationBase v0.20.10

Basic 3D multiply and cross dimension covariance test...: Error During Test at /home/pkgeval/.julia/packages/ApproxManifoldProducts/BhS8C/test/testBasicManiProduct.jl:26
  Got exception outside of a @test
  DomainError with -7.426431436168918e300:
  log will only return a complex result if called with a complex argument. Try log(Complex(x)).
  Stacktrace:
    [1] throw_complex_domainerror(f::Symbol, x::Float64)
      @ Base.Math ./math.jl:33

So maybe that's related, but I don't see how from a quick look. It depends on SLEEFPirates.jl, without actually using it, as far as I can tell.

BPFNative had:

  [929cbde3] + LLVM v3.7.1

empty: Error During Test at /home/pkgeval/.julia/packages/BPFnative/LpXdL/test/runtests.jl:27
  Got exception outside of a @test
  MethodError: no method matching LLVM.ConstantArray(::LLVM.IntegerType, ::Vector{UInt8})
  Closest candidates are:
    LLVM.ConstantArray(::LLVM.LLVMType) at /home/pkgeval/.julia/packages/LLVM/1GCWB/src/core/value/constant.jl:136
    LLVM.ConstantArray(::LLVM.LLVMType, !Matched::AbstractArray{T, N}) where {T<:LLVM.Constant, N} at /home/pkgeval/.julia/packages/LLVM/1GCWB/src/core/value/constant.jl:136
  Stacktrace:
    [1] finish_module!(job::GPUCompiler.CompilerJob{GPUCompiler.BPFCompilerTarget, BPFnative.BPFCompilerParams}, mod::LLVM.Module)
      @ GPUCompiler ~/.julia/packages/GPUCompiler/AdCnd/src/bpf.jl:54
    [2] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, kernel::LLVM.Function; strip::Bool, validate::Bool, format::LLVM.API.LLVMCodeGenFileType)
      @ GPUCompiler ~/.julia/packages/GPUCompiler/AdCnd/src/driver.jl:283
    [3] codegen(output::Symbol, job::GPUCompiler.CompilerJob; libraries::Bool, deferred_codegen::Bool, optimize::Bool, strip::Bool, validate::Bool, only_entry::Bool)
      @ GPUCompiler ~/.julia/packages/GPUCompiler/AdCnd/src/driver.jl:76
    [4] compile(target::Symbol, job::GPUCompiler.CompilerJob; libraries::Bool, deferred_codegen::Bool, optimize::Bool, strip::Bool, validate::Bool, only_entry::Bool)
      @ GPUCompiler ~/.julia/packages/GPUCompiler/AdCnd/src/driver.jl:39
    [5] bpffunction_compile(source::GPUCompiler.FunctionSpec{var"#kernel#1", Tuple{Int64}}; format::Symbol, license::String, prog_section::String, btf::Bool, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
      @ BPFnative ~/.julia/packages/BPFnative/LpXdL/src/compiler.jl:47
    [6] check_cache(cache::Dict{UInt64, Any}, compiler::Any, linker::Any, spec::GPUCompiler.FunctionSpec{var"#kernel#1", Tuple{Int64}}, prekey::UInt64; kwargs::Base.Pairs{Symbol, Any, Tuple{Symbol, Symbol, Symbol}, NamedTuple{(:format, :license, :btf), Tuple{Symbol, String, Bool}}})

[chriselrod]

@eschnett Can we get a new release for SIMD.jl? It crashes on the latest release:

JIT session error: Symbols not found: [ llvm.experimental.vector.reduce.v2.fadd.v4f64 ]
Failure value returned from cantFail wrapped call
Failed to materialize symbols: { (JuliaOJIT, { jfptr_sum_353, julia_sum_352 }) }
UNREACHABLE executed at /workspace/srcdir/usr/include/llvm/Support/Error.h:749!

but should be fine on master.

[vchuravy]

To look at:

ID	time ratio	memory ratio
["array", "accumulate", ("accumulate!", "Int")]	2.53 (5%)	1.00 (1%)
["array", "bool", "boolarray_true_load!"]	2.73 (5%)	1.00 (1%)
["array", "comprehension", ("collect", "Vector{Float64}")]	3.36 (5%)	1.00 (1%)
["array", "comprehension", ("comprehension_collect", "Vector{Float64}")]	3.04 (5%)	1.00 (1%)
["linalg", "arithmetic", ("cumsum!", "Int32", 1024)]	2.00 (45%)	1.00 (1%)

[nanosoldier]

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

[eschnett]

Released SIMD 3.3.1.

[vchuravy]

I will look at the remaining performance regressions over the next week.