improve performance of setindex! on IdDict

[KristofferC]

The benchmark run of 1.3 vs 1.2 showed some performance regressions for IdDict. Profiling, it seems that the convert call now goes through jl_apply_generic:

              123 ./abstractdict.jl:479; push!
               62 ./abstractdict.jl:582; setindex!(::IdDict{Int64,Int64}, ::Any, ... # <--- convert call
                46 ...kristoffer/julia/src/gf.c:2296; jl_apply_generic
                 37 ...kristoffer/julia/src/gf.c:2247; jl_lookup_generic_
                  36 ...ia/src/./julia_internal.h:939; jl_typemap_assoc_exact
                   11 ...ffer/julia/src/typemap.c:850; jl_typemap_level_assoc_exact
                    3 ...ffer/julia/src/typemap.c:255; mtcache_hash_lookup
                     3 ...ffer/julia/src/typemap.c:186; jl_intref
                      3 ...fer/julia/src/./julia.h:802; jl_svecref
                    3 ...ffer/julia/src/typemap.c:269; mtcache_hash_lookup
                   21 ...ffer/julia/src/typemap.c:851; jl_typemap_level_assoc_exact
                    20 ...a/src/./julia_internal.h:935; jl_typemap_assoc_exact
                     4 ...ffer/julia/src/typemap.c:803; jl_typemap_entry_assoc_exact
                      2 ...fer/julia/src/typemap.c:123; sig_match_simple
                     6 ...ffer/julia/src/typemap.c:816; jl_typemap_entry_assoc_exact
                      2 ...fer/julia/src/typemap.c:136; sig_match_simple
                       1 ...fer/julia/src/./julia.h:1127; jl_is_type_type
                       1 ...fer/julia/src/./julia.h:1128; jl_is_type_type
                10 ...a/usr/lib/julia/sys.dylib:?; jfptr_convert_4167
                 7 ...ffer/julia/src/datatype.c:724; jl_box_int64
                  4 ...ia/src/./julia_internal.h:233; jl_gc_alloc_

Workaround this by only calling the convert when it is not known to be a no-op.

With

function perf_push!(c, elts)
    for e in elts
        push!(c, e)
    end
end

elts = [rand(Int) => rand(Int) for i in 1:100]

The relevant benchmarks are:

1.2

@btime perf_push!(IdDict{Int,Int}(), elts)
#  11.889 μs (206 allocations: 11.48 KiB)

@btime perf_push!(IdDict{Any,Any}(), elts)
#  12.150 μs (206 allocations: 11.48 KiB)

Master

@btime perf_push!(IdDict{Int,Int}(), elts)
#  20.218 μs (306 allocations: 13.05 KiB)

@btime perf_push!(IdDict{Any,Any}(), elts)
#  13.461 μs (206 allocations: 11.48 KiB)

PR

@btime perf_push!(IdDict{Int,Int}(), elts)
#  11.182 μs (205 allocations: 7.36 KiB)

@btime perf_push!(IdDict{Any,Any}(), elts)
#  11.121 μs (205 allocations: 7.36 KiB)

[KristofferC]

@nanosoldier runbenchmarks("collection", vs = ":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan

[KristofferC]

Performance looks good but would be good to know if the now dynamic call to convert is expected, cc @vtjnash, @JeffBezanson .