Stack traces short/missing on Darwin/ARM64

[mcabbott]

I'm seeing stack traces with only a few layers, or none, on an M1 mac, with yesterday's master. @staticfloat suggests this is related to the "Unwinding from JIT frames doesn't work" item of #36617

julia> sqrt(-1)
ERROR: DomainError with -1.0:
sqrt will only return a complex result if called with a complex argument. Try sqrt(Complex(x)).
Stacktrace:
 [1] throw_complex_domainerror(f::Symbol, x::Float64)
   @ Base.Math ./math.jl:33

julia> Matrix{Int}([1 2; 3 4.5])
ERROR: InexactError: Int64(4.5)

julia> versioninfo()
Julia Version 1.7.0-DEV.685
Commit f3c66b3f9d (2021-03-10 18:21 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin20.3.0)
  CPU: Apple M1
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.1 (ORCJIT, cyclone)
Environment:
  JULIA_NUM_THREADS = 4
  JULIA_EDITOR = "/Applications/Visual Studio Code.app/Contents/Resources/app/bin/code"

[Keno]

@staticfloat suggests this is related to the "Unwinding from JIT frames doesn't work" item of #36617

Yes, that is that item.

[Keno]

I looked into this a little bit. We do register some FDEs with the unwinder, but for some reason, they don't cover the ip range that is actually used by the JIT frame (on which the unwinder seems to agree with LLDB at least). Haven't yet figured out why that is, but maybe LLVM is just getting that wrong.

[hahnjo]

Hi @Keno, I think the unwinding problems are related to what I describe in https://bugs.llvm.org/show_bug.cgi?id=49692#c8: M1 uses pointer authentication, and libunwind has to strip the tag from the upper bits before doing anything with the instruction pointer. Apparently Apple's libunwind does that, but the version available from opensource.apple.com doesn't have the necessary code, nor LLVM upstream. The comment I linked has a hacky workaround that you may want to include in your own libunwind (at least it's my understanding that you distribute your own and don't use the system library?).

A question in that regard: Is Julia using exception unwinding as well? That's what the bug report is about and what we're interested in for the ROOT framework (developed at CERN). In the latest comment, I suggest that there may be a bug in Apple's libunwind so we're looking who else might be affected...

[vchuravy]

Can someone test this with #41421?

[mcabbott]

This seems to work here, thanks!

julia> sqrt(-1)
ERROR: DomainError with -1.0:
sqrt will only return a complex result if called with a complex argument. Try sqrt(Complex(x)).
Stacktrace:
 [1] throw_complex_domainerror(f::Symbol, x::Float64)
   @ Base.Math ./math.jl:33
 [2] sqrt
   @ ./math.jl:567 [inlined]
 [3] sqrt(x::Int64)
   @ Base.Math ./math.jl:1221
 [4] top-level scope
   @ REPL[5]:1

julia> using Plots

julia> plot(rand(20), c=:nope)
ERROR: Unknown color: nope
Stacktrace:
  [1] error(::String, ::String)
    @ Base ./error.jl:42
  [2] _parse_colorant(desc::String)
    @ Colors ~/.julia/packages/Colors/yDxFN/src/parse.jl:151
  [3] parse(#unused#::Type{RGBA{Float64}}, desc::String)
    @ Colors ~/.julia/packages/Colors/yDxFN/src/parse.jl:207
... etc

[mcabbott]

Edit: It's gone, maybe just some local problem.

This problem seems to have returned on: 294b0dfcd3 (0 days old master)