Optimize JSON::Pure::Generator by 2x-4x for simple options cases

[eregon]

* ruby --yjit benchmarks/bench.rb dump pure ruby 3.3.1 (2024-04-23 revision c56cd86388) +YJIT [x86_64-linux]
  Before: JSON.dump(obj)    604.604  (± 0.3%) i/s    (1.65 ms/i) -      3.060k in   5.061200s
  After:  JSON.dump(obj)      2.531k (± 0.4%) i/s  (395.14 μs/i) -     12.801k in   5.058326s
* ruby benchmarks/bench.rb dump pure
  truffleruby 24.1.0-dev-a8ebb51b, like ruby 3.2.2, Oracle GraalVM JVM [x86_64-linux]
  Before: JSON.dump(obj)      3.728k (± 9.4%) i/s  (268.26 μs/i) -     18.559k in   5.068915s
  After:  JSON.dump(obj)      7.835k (± 8.5%) i/s  (127.63 μs/i) -     39.004k in   5.031116s

Using the benchmark from #580

BTW, the C extension, with YJIT gives:

JSON::Ext::Generator
ruby 3.3.1 (2024-04-23 revision c56cd86388) +YJIT [x86_64-linux]
JSON.dump(obj)      7.402k (± 1.0%) i/s  (135.10 μs/i) -     37.485k in   5.064832s

[eregon]

@hsbt Could you review this as well? 🙏

---

cc @byroot I mentioned this to you at RubyKaigi.
Might also be interesting to @maximecb (this is a pure-Ruby JSON parser, this PR is optimizing the pure-Ruby JSON generator).

[byroot]

Some minor comments, but looks good.

[byroot]

This feels like this could be precomputed in configure so you'd only need to check a single ivar, no?

[eregon]

Mmh, I'm not sure what's the protocol, the State class has a bunch of public attr_accessor so it looks like configure is not necessarily called after such accessors would be used to write.
Or maybe JSON users are not supposed to call these accessors? What are they for then?

BTW it seems pretty inefficient that e.g. JSON.generate would create a new State every time (which runs the State constructor and configure every time) instead of using SAFE_STATE_PROTOTYPE.dup or so, but that's a separate thing so improve later.

[eregon]

Out of curiosity I did

diff --git a/lib/json/pure/generator.rb b/lib/json/pure/generator.rb
index 9f44ef5..d07c2de 100644
--- a/lib/json/pure/generator.rb
+++ b/lib/json/pure/generator.rb
@@ -268,6 +268,10 @@ module JSON
           else
             @max_nesting = 0
           end
+
+          @fast = (@indent.empty? and @space.empty? and @space_before.empty? and @object_nl.empty? and @array_nl.empty? and
+            !@ascii_only and !@script_safe and @max_nesting == 0 and !@strict)
+
           self
         end
         alias merge configure
@@ -277,6 +281,7 @@ module JSON
         def to_h
           result = {}
           instance_variables.each do |iv|
+            next if iv == :@fast
             iv = iv.to_s[1..-1]
             result[iv.to_sym] = self[iv]
           end
@@ -292,8 +297,10 @@ module JSON
         def generate(obj)
           if @indent.empty? and @space.empty? and @space_before.empty? and @object_nl.empty? and @array_nl.empty? and
               !@ascii_only and !@script_safe and @max_nesting == 0 and !@strict
+            raise "OOPS" if !@fast
             result = generate_json(obj, '')
           else
+            raise "OOPS" if @fast
             result = obj.to_json(self)
           end
           JSON.valid_utf8?(result) or raise GeneratorError,

To see if there were any inconsistency and the test suite didn't find any.
But still, I think we need to understand better how these accessors are used (or maybe remove them if they're not needed).

[byroot]

At this point we might consider bumping the required version to 2.4 or even 2.5

[eregon]

That would be nice, but this is not much overhead, so that could be cleaned up later when we bump the min Ruby version.

[eregon]

Rebased and CI should be green