Bug #22551: YJIT: String#<< clears the receiver's coderange

• Status: Open
• Assignee: -
• Target version: -
• ruby -v: 4.0.7
• Backport: 3.3: UNKNOWN, 3.4: UNKNOWN, 4.0: UNKNOWN

[rdil (Reece Dunham)]

With YJIT, str << other leaves str with an unknown coderange when both strings have the same encoding. The interpreter keeps the coderange. Any later operation that needs it (length, [], valid_encoding?) then rescans the whole string, so code that appends and reads in a loop is much slower with YJIT than without.

# Run twice: `ruby repro.rb` and `ruby --yjit repro.rb`
require 'objspace'
require 'json'

def append(buf, chunk) = buf << chunk
def coderange(str) = JSON.parse(ObjectSpace.dump(str))['coderange'] || 'unknown'

chunk = 'é' * 100 # valid UTF-8, not ASCII
chunk.valid_encoding?

buf = +'é'
buf.valid_encoding?
1_000.times { append(+'é', chunk) } # let YJIT compile append
append(buf, chunk)
puts RUBY_DESCRIPTION
puts "coderange after <<: #{coderange(buf)}"

buf = +''
started = Process.clock_gettime(Process::CLOCK_MONOTONIC)
2_000.times do
  append(buf, chunk)
  buf.length
end
puts format('2,000 appends, each followed by #length: %.0f ms',
            (Process.clock_gettime(Process::CLOCK_MONOTONIC) - started) * 1000)

Results on arm64 macOS:

ruby	without YJIT	with --yjit
3.3.13	valid, 17 ms	unknown, 400 ms
3.4.9	valid, 17 ms	unknown, 477 ms
4.0.5	valid, 17 ms	unknown, 479 ms
4.0.7	valid, 17 ms	unknown, 504 ms