"Refreshing" functions by re-evaling their definition leads to better performance

[KristofferC]

julia> function printfd(n)
           open("/dev/null", "w") do io
               for i = 1:n
                   print(io, i, UInt32(i+1))
               end
           end
       end
printfd (generic function with 1 method)

julia> @btime printfd(1000)
  511.730 μs (5494 allocations: 226.78 KiB)

julia> @btime printfd(1000)
  512.749 μs (5494 allocations: 226.78 KiB)

julia> @eval Base begin
       show(io::IO, n::Signed) = (write(io, string(n)); nothing)
       print(io::IO, n::Unsigned) = print(io, string(n))
       end
print (generic function with 52 methods)

julia> @btime printfd(1000)
  222.553 μs (5005 allocations: 219.14 KiB)

julia> @btime printfd(1000)
  221.671 μs (5005 allocations: 219.14 KiB)

This is consistent. The benchmarks was run with a source build of julia.

[MikeInnes]

I've noticed this particularly with generated functions. In my case, it seemed like type inference was happening but no inlining/optimisation, for some reason.

[KristofferC]

Could this be due to where things are placed in the method table and how fast the dynamic dispatch calls find them?

[mbauman]

Shot in the dark: if there are any type instabilities it might be that inference is getting a "luckier" first compilation sequence after redefinition and ending up with a better inferred result than it "should".

[KristofferC]

I don't notice any difference in the inferred code so it seems this is something with the runtime...

[yuyichao]

Last time I hit this issue (right after 265 fix) it was due to the fix being a behavior fix and not a performance fix. There was no need to recompile if a dynamic dispatch can be Replaced with a static one after a new method is added

…

On Fri, Aug 24, 2018, 11:26 Matt Bauman ***@***.***> wrote: Shot in the dark: if there are any type instabilities it might be that inference is getting a "luckier" first compilation sequence after redefinition and ending up with a better inferred result than it "should". — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#28683 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAreKA5u19AsAJvKqYpum2_GXOhFSBd7ks5uUBsjgaJpZM4V_cF_> .

[KristofferC]

This specific case seems fixed now...

julia> using BenchmarkTools

julia> function printfd(n)
           open("/dev/null", "w") do io
               for i = 1:n
                   print(io, i, UInt32(i+1))
               end
           end
       end

printfd (generic function with 1 method)

julia> @btime printfd(1000)
  197.087 μs (5005 allocations: 219.14 KiB)

julia> @btime printfd(1000)
  196.564 μs (5005 allocations: 219.14 KiB)

julia> @eval Base begin
       show(io::IO, n::Signed) = (write(io, string(n)); nothing)
       print(io::IO, n::Unsigned) = print(io, string(n))
       end

print (generic function with 52 methods)

julia> @btime printfd(1000)
  205.369 μs (5005 allocations: 219.14 KiB)

julia> @btime printfd(1000)
  205.492 μs (5005 allocations: 219.14 KiB)