performance regression on master for folds

[matthias314]

This may be related to #48240. For the code below I get

julia> @btime f($v, $w)
  34.609 μs (0 allocations: 0 bytes)   # 1.9.0-beta4
  69.175 μs (0 allocations: 0 bytes)   # 1.10.0-DEV.658 Commit 129d45169b2 (0 days old master)

Here is the code. The type UnsafeArray is a wrapper that turns bounds-checking off.

import Base: parent, getindex, axes, IndexStyle, eachindex, unsafe_getindex

struct UnsafeArray{T,N,P<:AbstractArray{T,N}} <: AbstractArray{T,N}
    parent::P
end

getindex(a::UnsafeArray, ii...) = unsafe_getindex(a.parent, ii...)
axes(a::UnsafeArray, d...) = axes(a.parent, d...)
IndexStyle(a::UnsafeArray) = IndexStyle(a.parent)
eachindex(is::IndexStyle, a::UnsafeArray) = eachindex(is, a.parent)
size(a::UnsafeArray, d...) = size(a.parent, d...)

f(v, w) = mapfoldl(splat(*), +, zip(v, w); init = zero(eltype(v)))

v = rand(Int8, 2^16);
w = UnsafeArray(view(v, :));

There is no regression for f(v,v), but that expression is involved in a mystery. If on master I time f(v,v) before f(v,w), then I get a different result:

julia> @btime f($v, $v); @btime f($v, $w)
  40.393 μs (0 allocations: 0 bytes)   # same as on 1.9.0-beta4
  44.629 μs (0 allocations: 0 bytes)   # this has suddenly dropped!

Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 4 × Intel(R) Core(TM) i3-10110U CPU @ 2.10GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, skylake)
  Threads: 1 on 4 virtual cores