findall has strange behavior for impure predicates on Bool Arrays

[LilithHafner]

julia> f(x) = x || rand(Bool);

julia> a = [false, false, true];

julia> Random.seed!(10);

julia> findall(f, a)
2-element Vector{Int64}:
 1
 2
   # Where is three?
julia> findall(f, a)
3-element Vector{Int64}:
 2
 3 # There it is!
 4 # Oh, hello four, I didn't expect to see you here.

julia> versioninfo() # 25 days ago
Julia Version 1.9.0-DEV.1053
Commit 9e22e567f29 (2022-07-26 14:21 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin21.5.0)
  CPU: 4 × Intel(R) Core(TM) i5-8210Y CPU @ 1.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.5 (ORCJIT, skylake)
  Threads: 1 on 2 virtual cores
Environment:
  LD_LIBRARY_PATH = /usr/local/lib
  JULIA_PKG_PRECOMPILE_AUTO = 0

Not returning 3 is bad. Returing 4 could lead to segfaults and is definitely a bug.

We have two implementations of findall(f::Function, a::AbstractArray)

1. Precompute f.(a) and then get indices from the BitArray
   

   julia/base/array.jl

   Line 2328 in aac466f

   findall(testf::Function, A::AbstractArray) = findall(testf.(A))

2. Count count(f, a), preallocate an index array, and iterate through a second time recomputing f(a)
   

   julia/base/array.jl

   Lines 2375 to 2393 in aac466f

   	function _findall(f::Function, A::AbstractArray{Bool})
   	n = count(f, A)
   	I = Vector{eltype(keys(A))}(undef, n)
   	isempty(I) && return I
   	_findall(f, I, A)
   	end
   	function _findall(f::Function, I::Vector, A::AbstractArray{Bool})
   	cnt = 1
   	len = length(I)
   	for (k, v) in pairs(A)
   	@inbounds I[cnt] = k
   	cnt += f(v)
   	cnt > len && return I
   	end
   	# In case of impure f, this line could potentially be hit. In that case,
   	# we can't assume I is the correct length.
   	resize!(I, cnt - 1)
   	end

For simple f, 2 is about 1.5x faster according to my rough benchmarks. If the runtime of f dominates, then 1 should be 2x faster. If f is impure then 1 behaves how one would expect and 2 can have bizarre consequences.

Ideally, we dispatch to 2 for simple pure f and 1 otherwise. Our current heuristic is to dispatch to 2 for a::AbstractArray{Bool} and 1 otherwise. This is a bad heuristic. It would be cool to dispatch based on effect analysis, but if that is not an option, my preference is to use f === identity as the heuristic (even though this is a performance hit in some cases).

I suspect this was introduced by #42202 (cc @jakobnissen) which fixed #42187.

[adienes]

potential candidate to tag for 1.10?

[LilithHafner]

Oops! This is broken on 1.9 and not on 1.8, so we probably should have put it on the 1.9 milestone and held back the release until it was fixed. Thanks for the bump.