Dot calling in global scope has excessively long compilation time

[KristofferC]

f(x) = x

0.6:

julia> @time f.(rand(2))
  0.170154 seconds (53.89 k allocations: 2.928 MiB)

0.7:

julia> @time f.(rand(2))
  4.075771 seconds (869.73 k allocations: 51.770 MiB, 0.41% gc time)

However, note on 0.7:

julia> f(x) = x
f (generic function with 1 method)

julia> g(x) = f.(x)
g (generic function with 1 method)

julia> @time g(rand(2)):
  0.069514 seconds (73.85 k allocations: 4.065 MiB, 10.33% gc time)

Is this related to the "fast path" of . in global scope cc @Keno, @vtjnash?

[JeffBezanson]

Would be good to bisect this.

[mbauman]

I happen to have a bunch of random branches already checked out and built:

# master/28e79562f4 (5 days ago)
julia> @time f.(rand(2))
  3.305299 seconds (1.11 M allocations: 66.015 MiB, 3.21% gc time)

# master/bf45de8c33 (15 days ago)
julia> @time f.(rand(2))
  1.633016 seconds (1.06 M allocations: 62.159 MiB, 1.84% gc time)

# teh-jn/lazydotfuse/8147932909 (0 days ago)
julia> @time f.(rand(2))
  0.085282 seconds (92.90 k allocations: 5.784 MiB)

[KristofferC]

Ok, so the commit that took it from 1.6 to 3.3 seconds is bce5bbe @vtjnash, looking for the one that made it go from 0.1 to 1.6 now.

[KristofferC]

Ok, so:

0.1s -> 0.5s: 8b395aa
0.5s -> 0.8s: 8345b78
0.8s -> 1.3s: 4ed54db
1.3s -> 1.6s: ?
1.6s -> 3.3s: bce5bbe

[martinholters]

On freshly pulled master (890239f), I get (without any prior warm-up):

julia> @time identity.(rand(2));
  1.097517 seconds (192.40 k allocations: 11.320 MiB, 0.40% gc time)

julia> f(x) = x;

julia> @time f.(rand(2));
  0.063176 seconds (30.62 k allocations: 1.681 MiB)

The first one is still significantly slower than on 0.6, but that could potentially be a question of what is precompiled in the sysimg? The second one is pretty similar to 0.6.

[KristofferC]

Yeah, seems much better now. Perhaps #26785 or did that only touch disabled code?

[martinholters]

Is this a dup of #26714?
Going back to bce5bbe, I have

julia> @time identity.(rand(2));
  4.359986 seconds (1.13 M allocations: 66.379 MiB, 1.29% gc time)

julia> @time @code_warntype 1+1
# ...
  2.755632 seconds (1.20 M allocations: 66.892 MiB, 0.43% gc time)

or (is fresh session):

julia> @time @code_warntype 1+1;
# ...
  6.947450 seconds (2.12 M allocations: 121.545 MiB, 1.01% gc time)

julia> @time identity.(rand(2));
  0.164185 seconds (210.29 k allocations: 11.828 MiB, 2.02% gc time)

[KristofferC]

Yeah, looks like both are testing the same thing, probably invoking inference for the first time. Closing as dup of #26714