significant regression in benchmarks

[Krastanov]

The whole point of ResumableFunctions is that it is much faster than Base.Channel tasks. That is still somewhat true, but the benchmarks (described in the README) have experienced significant loss of performance (with Int types).

These two should basically be the same, but currently there is a significant difference:

Direct: 
  23.115 ns (0 allocations: 0 bytes)
ResumableFunctions: 
  3.487 μs (249 allocations: 3.89 KiB)

After this is fixed, we should also add these benchmarks to the CI and we should add the check for zero allocations as an explicit test.

[gerlero]

Well, this is bad.

we should also add these benchmarks to the CI and we should add the check for zero allocations as an explicit test.

Agreed.

[pepijndevos]

I bisected the performance regression to 727ba4aab8826d9a6fff9f332ab1179dfefbb39c

[Krastanov]

@gerlero , any chance you can investigate this? It is probably some type instability in the modified iterate method that was introduced when yieldfrom was merged.

[gerlero]

@pepijndevos thanks; that's #57 then. @Krastanov I'll take a look as soon as I can.