Improve Array#shift performance on v8 > 7.1 (Node 12+)

[ggrossetie]

Command

$ bundle exec rake bench:ips FILE=benchmark-ips/bm_array_shift.rb

Before

Node 10

=== Files: benchmark-ips/bm_array_shift.rb
bundle exec opal --dynamic-require ignore --missing-require ignore -rnodejs -ropal/platform -gbenchmark-ips -rbenchmark/ips -A benchmark-ips/bm_array_shift.rb
""
Warming up --------------------------------------
        shift no arg    39.327k i/100ms
Calculating -------------------------------------
        shift no arg    979.605k (± 1.6%) i/s -      4.916M in   5.019468s

Node 12

=== Files: benchmark-ips/bm_array_shift.rb
bundle exec opal --dynamic-require ignore --missing-require ignore -rnodejs -ropal/platform -gbenchmark-ips -rbenchmark/ips -A benchmark-ips/bm_array_shift.rb
""
Warming up --------------------------------------
        shift no arg     2.364k i/100ms
Calculating -------------------------------------
        shift no arg     24.879k (± 0.5%) i/s -    125.292k in   5.036277s

After

Node 10

=== Files: benchmark-ips/bm_array_shift.rb
bundle exec opal --dynamic-require ignore --missing-require ignore -ropal/platform -gbenchmark-ips -rbenchmark/ips -A benchmark-ips/bm_array_shift.rb
Warming up --------------------------------------
        shift no arg    26.398k i/100ms
Calculating -------------------------------------
        shift no arg    449.226k (± 4.4%) i/s -      2.244M in   5.005441s

Node 12

=== Files: benchmark-ips/bm_array_shift.rb
bundle exec opal --dynamic-require ignore --missing-require ignore -ropal/platform -gbenchmark-ips -rbenchmark/ips -A benchmark-ips/bm_array_shift.rb
Warming up --------------------------------------
        shift no arg    37.305k i/100ms
Calculating -------------------------------------
        shift no arg    611.658k (± 1.0%) i/s -      3.059M in   5.001669s

Summary

With this change, the code is now 2.18x times slower when using Node 10 but 24.58x times faster when using Node 12.

We cannot beat the highly optimized code produced by v8 pre 7.1 (Node 10) but the code is now "only" 1.6x times slower when using a recent version of v8/Node.

Please note that the performance on Node 14 is roughly equivalent to Node 12.

Further improvement

The Array#shift() (JavaScript) function is also used in String#tr and String#tr_s. It might be worth to replace this call by the optimized shiftNoArg function? In this case, where should I declare the shiftNoArg function? in runtime.js?

I also found an optimization for splice when the second argument is 1 (ie. list.splice(position, 1)). I found 8 occurrences in the code where the second argument is explicitly 1. When a variable is used as the second argument it might be worth to check if the value is actually 1 to use the optimized code (I think it's only worth if the array is large enough but that means that we need to add an additional check that will slow down the execution).

[ggrossetie]

@elia I found a more generic implementation:

function spliceGen (list, start, deleteCount) {
  if (deleteCount <= 0) {
    return
  }
  const length = list.length
  for (var i = start, k = i + deleteCount; k < length; i += 1, k += 1)
    list[i] = list[k]

  list.length = i
  // or:
  //for (;i < length; i++)
  //  list.pop()
}

Depending on the list length, calling list.pop() can be faster than updating the list length. The goal is to remove "trailing" elements.

[elia]

Thanks for catching and taking care of this!
I left one nit, let me know if you don't have time to update the PR and I'll merge it as it is 👍

[ggrossetie]

@elia I will keep the focus of this pull request to just Array#shift and open a new pull request to generalize the usage of spliceOne / spliceGen.

[ggrossetie]

@elia I've added a changelog entry and rebased/squashed my work. Let me know if there's anything left to do.