Method invalidation on loading Polynomials

[jishnub]

julia> using SnoopCompileCore

julia> invalidations = @snoopr using Polynomials;

julia> using SnoopCompile

julia> trees = invalidation_trees(invalidations);

julia> length(trees)
6

julia> trees[end].method
convert(::Type{T}, p::P) where {T, P<:(AbstractPolynomial{T})} in Polynomials at /home/jishnu/Dropbox/JuliaPackages/Polynomials.jl/src/common.jl:433

julia> trees[end-1].method
promote_rule(::Type{<:SparsePolynomial{var"#336#T", var"#338#X"}}, ::Type{var"#337#S"}) where {var"#336#T", var"#337#S"<:Number, var"#338#X"} in Polynomials at /home/jishnu/Dropbox/JuliaPackages/Polynomials.jl/src/abstract.jl:90

julia> trees[end-3].method
any(pred, p::AbstractPolynomial) in Polynomials at /home/jishnu/Dropbox/JuliaPackages/Polynomials.jl/src/common.jl:493

These three methods introduced here cause a lot of invalidation, leading to slow package load times. Maybe we can think about ways in which this may be remedied.

[jverzani]

Thanks, I haven't used this tool before. I get only length(trees) being 4, the first method shows up only. Is that machine dependent? (This is with your freshly merged PR.)

[jverzani]

I have a fix for the first, but I don't have the second two showing up when I follow your report. If you have time, could you check? (If the tests pass I'll merge it in and tag, so this would be easier for you.)

[jishnub]

Yes, I can confirm that the invalidation of convert doesn't show up anymore. The ones that remain are any and promote_rule

[jverzani]

I'm on v1.7.2. How about you? (I wasn't seeing the latter 2 and am guessing it is version dependent.)

[jishnub]

I'm on 1.7.3. It might be version-dependent

[jverzani]

THanks, good reason to upgrade.

[jverzani]

Okay, see them now. Will try and address.

[jverzani]

Oddly, in #412 I only addressed the issue with any, but I have no clue why the promote_rule one doesn't seem present anymore.

[jverzani]

I found a new invalidation under the 1.8 release candidate. If you have a chance, can you test on your end if the promote_rule is still present, and perhaps if there are others?