Optimize set! DSL (Round 2)

[moberegger]

Follow up to #15, which made some changes to how blocks were passed around. Did some due diligence and actually found that doing it in _set didn't help. I put the results of some IPS benchmarks below... while they all show "difference falls within error", this new revision does run a bit faster. I had some previous runs showing that this new revision is faster without the error margin, but can't consistently reproduce that. Only included IPS because no differences showed up in memory allocation.

I think it's best to go with this new revision for a few reasons:

• It might be faster. At worst, no difference.
• It's closer to what was there before, so fewer deltas with upstream.
• It keeps things consistent with the rest of the source code, so fewer refactors will be necessary
• I have seen calls to ::Kernel.block_given? show up as hotspots in our Datadog profiles, so maybe best to save on that if we can.

Oddly enough, having &block in JbuilderTemplate#set! is consistently slower. I have no idea why it matters in one case but not the other.

---

json.set! :foo, :bar

ruby 3.4.4 (2025-05-14 revision a38531fd3f) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
              before   750.032k i/100ms
               after   721.349k i/100ms
Calculating -------------------------------------
              before      9.085M (± 2.2%) i/s  (110.07 ns/i) -     45.752M in   5.038314s
               after      9.163M (± 1.7%) i/s  (109.14 ns/i) -     46.166M in   5.039896s

Comparison:
               after:  9162918.1 i/s
              before:  9085243.6 i/s - same-ish: difference falls within error

---

# Where object = { bar: 123 }
json.set! :foo do
  json.extract! object, :bar
end

Warming up --------------------------------------
              before   164.804k i/100ms
               after   184.442k i/100ms
Calculating -------------------------------------
              before      1.755M (± 3.9%) i/s  (569.71 ns/i) -      8.899M in   5.080089s
               after      1.894M (± 7.1%) i/s  (527.90 ns/i) -      9.407M in   4.999829s

Comparison:
               after:  1894297.3 i/s
              before:  1755275.3 i/s - same-ish: difference falls within error

---

# Where array = [1, 2, 3]
json.set! :foo, array do |item|
  json.set! :bar, item
end

ruby 3.4.4 (2025-05-14 revision a38531fd3f) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
              before   104.803k i/100ms
               after   111.370k i/100ms
Calculating -------------------------------------
              before    967.022k (±18.6%) i/s    (1.03 μs/i) -      4.611M in   5.036618s
               after      1.115M (± 4.2%) i/s  (897.07 ns/i) -      5.568M in   5.005477s

Comparison:
               after:  1114745.4 i/s
              before:   967022.0 i/s - same-ish: difference falls within error

[mscrivo]

I'm having trouble reasoning about this change. Is it actually equivalent to what it was before? since the block yielding an element before?

[moberegger]

Yeah, it's the same. This is what was originally there. The block provided to set! just accepts on param, so this just passes the block instead of yielding to it inside another block.