Optimize ruby visitor

[Earlopain]

#3808 (comment)

Looking at things like node.compact_child_nodes.each maybe we should have node.each_child_node? It would avoid the Array allocation and since there is no point to yield nil there is no need to name it node.each_compact_child_node.

This also came up someplace else before but I don't remember where. I agree this makes sense. cc @eregon

---

compact_child_nodes allocates an array. We can skip that step by simply yielding the nodes.

Benchmark for visiting the rails codebase:

require "prism"
require "benchmark/ips"

files = Dir.glob("../rails/**/*.rb")
results = files.map { Prism.parse_file(it) }
visitor = Prism::Visitor.new

Benchmark.ips do |x|
  x.config(warmup: 3, time: 10)

  x.report do
    results.each do
      visitor.visit(it.value)
    end
  end
end

RubyVM::YJIT.enable

Benchmark.ips do |x|
  x.config(warmup: 3, time: 10)

  x.report do
    results.each do
      visitor.visit(it.value)
    end
  end
end

Before:

ruby 3.4.8 (2025-12-17 revision 995b59f666) +PRISM [x86_64-linux]
Warming up --------------------------------------
                         1.000 i/100ms
Calculating -------------------------------------
                          2.691 (± 0.0%) i/s  (371.55 ms/i) -     27.000 in  10.089422s
ruby 3.4.8 (2025-12-17 revision 995b59f666) +YJIT +PRISM [x86_64-linux]
Warming up --------------------------------------
                         1.000 i/100ms
Calculating -------------------------------------
                          7.278 (±13.7%) i/s  (137.39 ms/i) -     70.000 in  10.071568s

After:

ruby 3.4.8 (2025-12-17 revision 995b59f666) +PRISM [x86_64-linux]
Warming up --------------------------------------
                         1.000 i/100ms
Calculating -------------------------------------
                          3.429 (± 0.0%) i/s  (291.65 ms/i) -     35.000 in  10.208580s
ruby 3.4.8 (2025-12-17 revision 995b59f666) +YJIT +PRISM [x86_64-linux]
Warming up --------------------------------------
                         1.000 i/100ms
Calculating -------------------------------------
                         16.815 (± 0.0%) i/s   (59.47 ms/i) -    169.000 in  10.054668s

~21% faster on the interpreter, ~56% with YJIT

[Earlopain]

Unsure if this is correct. I'm also missing sorbet signatures which I don't know how to write for this

[kddnewton]

This looks right to me, and it would have failed to typecheck otherwise.

[Earlopain]

This can just be implemented as each_child_node.to_a but performance is not so great

[kddnewton]

Nope, you're right to keep this the same.