Post: digging into struct initialization performance

[ysbaddaden]

Issue: the kramdown gfm doesn't support the > [!TIP] annotation for callouts. I'm not sure how to replace these?

[netlify]

✅ Deploy Preview for crystal-website ready!

Name	Link
🔨 Latest commit	3ef748d
🔍 Latest deploy log	https://app.netlify.com/sites/crystal-website/deploys/66e04389df644b0008c2c417
😎 Deploy Preview	https://deploy-preview-827--crystal-website.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify site configuration.

[beta-ziliani]

Nice! I left a few tips to improve it.

[HertzDevil]

IIRC this was first posted before ReferenceStorage was a thing, so I did my own benchmark:

class Digest
  module ClassMethods
    def digest(& : self ->) : Bytes
      context_buf = uninitialized ReferenceStorage(self)
      context = unsafe_construct(pointerof(context_buf))
      begin
        yield context
        context.final
      ensure
        context.finalize if context.responds_to?(:finalize)
      end
    end
  end
end

class Digest::Blake3 < ::Digest
  extend ::Digest::ClassMethods
end

struct Blake3Hasher
  # ditto
end

def blake3_hexstring(data)
  hasher = Blake3Hasher.new
  # ...
end

def blake3_hexstring_init(data)
  hasher = uninitialized Blake3Hasher
  hasher.init
  # ...
end

Benchmark.ips do |x|
  bytes = Random::Secure.random_bytes(32)
  x.report("SHA256") { Digest::SHA256.hexdigest(bytes) }
  x.report("Blake3 Digest") { Digest::Blake3.hexdigest(bytes) }
  x.report("Blake3 new") { blake3_hexstring(bytes) }
  x.report("Blake3 init") { blake3_hexstring_init(bytes) }
end

       SHA256   2.89M (346.37ns) (± 0.34%)   128B/op   2.54× slower
Blake3 Digest   6.41M (156.08ns) (± 0.48%)   128B/op   1.15× slower
   Blake3 new   3.48M (287.00ns) (± 0.30%)  80.0B/op   2.11× slower
  Blake3 init   7.34M (136.28ns) (± 0.26%)  80.0B/op        fastest

The remaining overhead most likely came from the mandatory memory zeroing for reference types.

Without redefining Digest::ClassMethods#digest:

       SHA256   2.44M (410.30ns) (± 0.48%)    224B/op   3.08× slower
Blake3 Digest   2.85M (351.27ns) (± 0.97%)  2.13kB/op   2.64× slower
   Blake3 new   3.53M (282.93ns) (± 0.43%)   80.0B/op   2.12× slower
  Blake3 init   7.50M (133.25ns) (± 0.40%)   80.0B/op        fastest

[ysbaddaden]

@HertzDevil Nice! I was thinking about this. A good demonstration of the experimental methods, and that RFC 4 could significantly improve performance!

[ysbaddaden]

@beta-ziliani I let you decide when to publish :)

[kostya]

I also have slow structs and fast tuples here crystal-lang/crystal#14225 (comment),
probably if you use tuples instead of struct it would be fast.