regression in float to int performance on master

[JeffBezanson]

function g(x)
               a = similar(x,Int)
               for i=1:length(x)
                 a[i] = x[i]
               end
              a
              end

x=Float64.(rand(1:1000000,2000,2000));

On 0.5.0:

julia> @time for i=1:20; g(x); end
  0.284237 seconds (40 allocations: 610.353 MB, 26.00% gc time)

on master:

julia> @time for i=1:20; g(x); end
  0.921421 seconds (40 allocations: 610.353 MB, 3.41% gc time)

The difference in generated code is that convert is no longer inlined, and it uses branches instead of eager &.

[simonbyrne]

Could be #14763 related, will look into it.

[vchuravy]

We should add a set of benchmarks for this.

[stevengj]

@vchuravy, just pushed a PR to BaseBenchmarks.

[KristofferC]

I'm getting (using the fancy new benchmark macros from JuliaCI/BenchmarkTools.jl#37):

0.6

julia> @btime g(x)
  11.714 ms (2 allocations: 30.52 MiB)

0.5

julia> @btime g(x)
  11.308 ms (2 allocations: 30.52 MiB)

So seems fixed?

[tkelman]

Do we know what fixed this? Am considering backporting the original change, but it causes a performance regression in conversion and whatever fixed it was not marked for backporting.