Remove idxfloor and shrink age, and maxprobe

[oscardssmith]

I believe this frees up 16 bytes for approximately free. idxfloor seems like a good idea, but it adds very little performance because with uniformly distributed hashes, on average, the idxfloor will be very low. The other change is to shrink age and maxprobe by 32 bits because there never should be 4 billion collisions next to each other, and age doesn't need to be 64 bits either.

[petvana]

The only major benefit of idxfloor is that we do not iterate over empty dict (#44411). We can add a fast path for count == 0 without idxfloor to get the same performance.

[oscardssmith]

that makes sense to me.

[simeonschaub]

I believe we have some basic Dict benchmarks in BaseBenchmarks, so would be good to verify with those before merging

[gbaraldi]

@nanosoldier runbenchmarks("!scalar", vs=":master")

[nanosoldier]

Your job failed.

[vtjnash]

maxprobe could be arbitrarily large, if the hash function is poorly chosen (it just degrades the dict into an array)

[oscardssmith]

if the hash function is poorly chosen everything else about our Dict's performance breaks anyway. I don't see why iteration should be special.

[petvana]

maxprobe could be arbitrarily large, if the hash function is poorly chosen (it just degrades the dict into an array)

For maxprobe to reach 32 bits, you first need to execute ~ 2^62 comparisons just to insert the elements that seems quite intractable for any non-quantum computer. Short example

julia> struct X; x::Int64; end

julia> Base.hash(x::X) = UInt64(1)

julia> for n = 5000:5000:20000; print("n=",n); 
         @time (d = Dict(); for i=1:n; d[X(i)] = i; end)
       end
n=5000  0.458749 seconds (12.52 M allocations: 213.679 MiB, 30.10% gc time, 0.72% compilation time)
n=10000  1.442519 seconds (50.03 M allocations: 785.996 MiB, 7.78% gc time)
n=15000  3.760717 seconds (112.53 M allocations: 1.699 GiB, 10.08% gc time)
n=20000  9.156206 seconds (200.07 M allocations: 3.070 GiB, 7.67% gc time)

[petvana]

I guess it can be moved to the previous line.

[petvana]

Now, the error from docs is that skip_deleted_floor! is missing in base/set.jl and base/weakkeydict.jl

[oscardssmith]

closing in favor of #59769