Enable lazy translation for user code by default

[eregon]

Enable lazy translation from the JRuby AST to Truffle AST for user code by default (--lazy-translation-user).
This is already the case for core library code, but not for user code.

This might need a few extra checks during parsing and some fixes for instrumentation.

This should help to improve the time to load code (part of application startup time).

[deepj]

Hm, bundle or rake commands seem to be more responsive with this change. Though, rails command seems to be still challenging

[eregon]

Implemented in d1230eb

[chrisseaton]

Sorry this doesn't make a statistically significant difference for me.

[eregon]

Thanks for checking.

On require "rubygems" it's noticeable with require profiling (~10ms vs ~22ms for translating), but that's a very small part of total require time (~485ms):

$ ruby --experimental-options --cpusampler --cpusampler.Mode=roots --metrics-profile-require=summary -e 'require "rubygems"' |& grep "metrics "
 metrics execute     |  484ms  99.4% |   0.0% ||  266ms  54.6% |   0.0% | (metrics)~1:0 
 metrics parsing     |   32ms   6.6% |   0.0% ||   32ms   6.6% |   0.0% | (metrics)~1:0 
 metrics require     |  485ms  99.6% |   0.0% ||   23ms   4.7% |   0.0% | (metrics)~1:0 
 metrics translating |   10ms   2.1% |   0.0% ||   10ms   2.1% |   0.0% | (metrics)~1:0 
 metrics searching   |    1ms   0.2% |   0.0% ||    1ms   0.2% |   0.0% | (metrics)~1:0 

vs (disabled)

$ ruby --experimental-options --lazy-translation-user=false --cpusampler --cpusampler.Mode=roots --metrics-profile-require=summary -e 'require "rubygems"' |& grep "metrics "
 metrics execute     |  490ms  99.4% |   0.0% ||  280ms  56.8% |   0.0% | (metrics)~1:0 
 metrics parsing     |   28ms   5.7% |   0.0% ||   28ms   5.7% |   0.0% | (metrics)~1:0 
 metrics translating |   22ms   4.5% |   0.0% ||   22ms   4.5% |   0.0% | (metrics)~1:0 
 metrics require     |  491ms  99.6% |   0.0% ||   17ms   3.4% |   0.0% | (metrics)~1:0 
 metrics searching   |    1ms   0.2% |   0.0% ||    1ms   0.2% |   0.0% | (metrics)~1:0 

[deepj]

I just noticed rake and bundle are a bit more responsive. What probably shows @eregon metrics. But I haven't seen any better responsivity if I use rails commands

[eregon]

BTW, a good part (~11%) of require "rubygems" seems to be spent in Array#flatten and Thread.detect_recursion:

ruby --experimental-options --cpusampler --cpusampler.SampleInternal --metrics-profile-require=detail -e 'require "rubygems"' 
----------------------------------------------------------------------------------------------------------------------------------------------------
Sampling Histogram. Recorded 573 samples with period 1ms.
  Self Time: Time spent on the top of the stack.
  Total Time: Time spent somewhere on the stack.
  Opt %: Percent of time spent in compiled and therefore non-interpreted code.
----------------------------------------------------------------------------------------------------------------------------------------------------
 Thread: Thread[main,5,main]
 Name                                        |Total Time     |  Opt % || Self Time     |  Opt % | Location             
----------------------------------------------------------------------------------------------------------------------------------------------------
 metrics parsing                             |   23ms   4.0% |   0.0% ||   23ms   4.0% |   0.0% | (metrics)~1:0 
 Thread.detect_recursion                     |   67ms  11.7% |   0.0% ||   22ms   3.8% |   0.0% | resource:/truffleruby/core/thread.rb~93-144:4227-5434 
 metrics translating                         |   14ms   2.4% |   0.0% ||   14ms   2.4% |   0.0% | (metrics)~1:0 
 Gem::Specification#initialize               |  186ms  32.5% |   0.0% ||   13ms   2.3% |   0.0% | ../../../.rubies/truffleruby-dev/lib/mri/rubygems/specification.rb~2059-2078:56439-56909 
 block in Array#recursively_flatten          |   34ms   5.9% |   0.0% ||   12ms   2.1% |   0.0% | resource:/truffleruby/core/array.rb~1329:31520-31568 
 Truffle::FeatureLoader.find_feature_or_file |   50ms   8.7% |   0.0% ||   10ms   1.7% |   0.0% | resource:/truffleruby/core/truffle/feature_loader.rb~72-120:2146-3557 
 Gem::Version#<=>                            |   20ms   3.5% |   0.0% ||    9ms   1.6% |   0.0% | ../../../.rubies/truffleruby-dev/lib/mri/rubygems/version.rb~342-367:11043-11660 
 Hash#key?                                   |   10ms   1.7% |   0.0% ||    9ms   1.6% |   0.0% | resource:/truffleruby/core/hash.rb~424-426:10788-10858 
 Truffle::Type.coerce_to_path                |   22ms   3.8% |   0.0% ||    8ms   1.4% |   0.0% | resource:/truffleruby/core/type.rb~495-511:14840-15334 
 block in Enumerable#uniq                    |   15ms   2.6% |   0.0% ||    8ms   1.4% |   0.0% | resource:/truffleruby/core/enumerable.rb~968:22566-22582 
 Gem::Requirement.parse                      |   34ms   5.9% |   0.0% ||    8ms   1.4% |   0.0% | ../../../.rubies/truffleruby-dev/lib/mri/rubygems/requirement.rb~114-126:3211-3511 
 block in Object#<top (required)>            |    9ms   1.6% |   0.0% ||    7ms   1.2% |   0.0% | resource:/truffleruby/core/regexp.rb~383:10286-10336 
 Truffle::FeatureLoader.feature_provided?    |   44ms   7.7% |   0.0% ||    7ms   1.2% |   0.0% | resource:/truffleruby/core/truffle/feature_loader.rb~134-173:3955-5407 
 Array#bottom_up_merge                       |   11ms   1.9% |   0.0% ||    7ms   1.2% |   0.0% | resource:/truffleruby/core/array.rb~1416-1435:33664-34149 
 String#include?                             |    9ms   1.6% |   0.0% ||    7ms   1.2% |   0.0% | resource:/truffleruby/core/string.rb~162-164:4978-5063 
 Gem::Requirement#initialize                 |   96ms  16.8% |   0.0% ||    7ms   1.2% |   0.0% | ../../../.rubies/truffleruby-dev/lib/mri/rubygems/requirement.rb~140-151:3915-4220 
 Array#recursively_flatten                   |   62ms  10.8% |   0.0% ||    6ms   1.0% |   0.0% | resource:/truffleruby/core/array.rb~1319-1352:31299-32049 
...

[chrisseaton]

detect_recursion is super common. Maybe we could do this better as a primitive?

[eregon]

Yes indeed, I was thinking we should optimize Thread.detect_recursion as well, and there are many ways to improve over the current implementation: https://github.com/oracle/truffleruby/projects/1#card-37918829