order of magnitude type-unstable performance regression

[aplavin]

Taking an example from discource,

julia> using BenchmarkTools
julia> x_nan = [([1., 2., 3.],), ([1., 2., 3., NaN],)]
julia> x_missing = [([1., 2., 3.],), ([1., 2., 3., missing],)]
julia> @btime map(x -> sum.(x), $x_nan)
julia> @btime map(x -> sum.(x), $x_missing)

I get:

• 1.8:

  58.755 ns (1 allocation: 80 bytes)
  808.369 ns (8 allocations: 240 bytes)

• 1.9 and recent nightly:

  59.531 ns (1 allocation: 80 bytes)
  6.157 μs (16 allocations: 576 bytes)

So, the type-unstable performance suffers a lot, even for a very simple piece of code!

Unlike another 1.9 regression #48612, this one is present even on nightly.

[maleadt]

Bisected to #45062; maybe a dup of #49145?

[aplavin]

Same in 1.10.

[aplavin]

Same regression remains in 1.11 and nightly as well – about an order of magnitude slowdown wrt 1.8.

1.8:

  22.693 ns (1 allocation: 80 bytes)  # type stable
  262.298 ns (8 allocations: 240 bytes)  # type unstable

1.9:

  21.773 ns (1 allocation: 80 bytes)
  2.301 μs (16 allocations: 576 bytes)

1.10:

  20.520 ns (1 allocation: 80 bytes)
  2.449 μs (16 allocations: 576 bytes)

1.11:

  15.698 ns (2 allocations: 80 bytes)
  2.847 μs (22 allocations: 752 bytes)

nightly:

  15.907 ns (2 allocations: 80 bytes)
  2.819 μs (22 allocations: 752 bytes)