inference failure on nested `zip` iteration

[stakaz]

Hello, I was hitting the following problem: when using nested zip functions julia is somehow not able to generate fast code on the first run but on the second run the function becomes fast. An example is

using Statistics 
using BenchmarkTools 

f(X) = reweight((a for a ∈ X.A), 0.1)
g(X) = reweight((a * b for (a, b) ∈ zip(X.A, X.B)), 0.1)

@inline function reweight(A, ω)
    N = 0.0
    for (a,) ∈ zip(A)
        N += a
    end
    return N
end

N = 100
X = (A = rand(N), B = rand(N))

println("f")
@btime f($X)

println("g")
@btime g($X)
@code_warntype g(X)

Here, the first time @code_warntype is reporting some bad things but running the code again removes this and the function is much faster.

There is a discussion on discourse and I was suggested to submit an issue for this here.

julia> versioninfo()
Julia Version 1.1.0
Commit 80516ca202 (2019-01-21 21:24 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Pentium(R) CPU  N3530  @ 2.16GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, silvermont)

Please help me to find a solution for this as it really slows down my calculations considerable. Thanks.

[c42f]

I believe the root cause is that inference gives up on nested zip iterators:

julia> @code_warntype sum(x->x[1][1], zip(zip([1,2,3])))
Body::Any
1 ─ %1 = Base.add_sum::Core.Compiler.Const(Base.add_sum, false)
│   %2 = invoke Base.mapfoldl_impl(_2::Function, %1::Function, $(QuoteNode(NamedTuple()))::NamedTuple{(),Tuple{}}, _3::Base.Iterators.Zip{Tuple{Base.Iterators.Zip{Tuple{Array{Int64,1}}}}})::Any
└──      return %2