Towards Ruby 4 JIT / RubyKaigi 2022

RubyKaigi 2022

Transcript

1. Towards Ruby 4 JIT @k0kubun
2. @k0kubun Maintain: MJIT, Haml, ERB Shopify team
3. GitHub Sponsors
4. Haml 6
5. Introduction to Ruby JIT
6. How does Ruby JIT work? Ruby
7. How does Ruby JIT work? 1 + 2 Ruby Abstract 
 Syntax 
 Tree
8. How does Ruby JIT work? 1 + 2 putobject 1 putobject 2 opt_plus leave Ruby Abstract 
 Syntax 
 Tree Instruction 
 Sequence 
 (Bytecode)
9. How does Ruby JIT work? 1 + 2 putobject 1 putobject 2 opt_plus leave Ruby Abstract 
 Syntax 
 Tree Instruction 
 Sequence 
 (Bytecode) Machine 
 Code
10. How does Ruby JIT work?
11. CRuby JIT 1: MJIT
12. CRuby JIT 2: YJIT
13. Current CRuby JITs speed.yjit.org
14. Current CRuby JITs speed.yjit.org
15. Current CRuby JITs • YJIT • Available since Ruby 3.1 • --jit or --yjit • MJIT • Available since Ruby 2.6 • --mjit
16. Current CRuby JITs • YJIT • Ruby 3.1: x86_64 only, no code GC, written in C • Ruby 3.2: arm64 support, (hopefully) code GC, written in Rust • MJIT • Ruby 3.1: Stable-ish, portable, native threads, written in C • Ruby 3.2: Experimental, fork + SIGCHLD, written in Ruby
17. MJIT in Ruby
20. mjit.rb: Secret "standard library" in Ruby 3.2 • mjit.rb • Even more powerful than TracePoint • You can monkey-patch CRuby JIT • No compatibility guarantee • Every module is private, so const_get is required
21. BYOJ: Bring Your Own JIT
22. BYOJ: Bring Your Own JIT • Load and pause MJIT with --mjit=pause • Define RubyVM::MJIT.compile • Use RubyVM::MJIT.const_get(:C) to hack RubyVM • Call RubyVM::MJIT.resume to start JIT With Ruby 3.2:
23. YJIT-style JIT • Monkey-patch RubyVM::MJIT.compile
24. MJIT-style JIT • Monkey-patch RubyVM::MJIT::Compiler.compile
25. MJIT-style JIT
26. Everyone is writing CRuby JIT
27. Benchmarking Ruby JIT
28. yjit-bench
29. yjit-bench • yjit-bench has three kinds of benchmarks: 1. Headlining Benchmarks 2. Other Benchmarks 3. Micro Benchmarks
30. 1. Headlining benchmarks • activerecord • hexapdf • liquid-render • mail • psych-load • railsbench ✉
31. 2. Other Benchmarks • binarytrees, fankuchredux, nbody • chunky_png • erubi, erubi_rails • lee • optcarrot • rubykon
32. 3. Micro Benchmarks • 30k_ifelse, 30k_methods • cfunc_itself, str_concat • fib • getivar, setivar • keyword_args • respond_to
34. Benchmark Your Own JIT • ./run_benchmarks.rb -e “/path/to/ruby --any-option” • Pass multiple -e options to compare different JITs
35. Towards Ruby 4 JIT
36. My wish on Ruby 4 JIT • I want Ruby 4 to be as fast as Java or JavaScript • Ruby 4's performance should be a reason to leave Python
38. More Concrete Examples
42. Ruby 4 Canary • true is mov-ed (immediate) • No opt_* VM instruction • Constant folding • Ruby / C method inlining
43. Ruby 4 Canary’ • Single branch instruction to access @one • Single register to access two • No heap allocation • No stack frame
46. Ruby 4 Canary 2 • 5000050000 is mov-ed (immediate) • Ruby -> C -> Ruby inlining
47. How can we get there?
48. Optimization Challenges 1. Constants 2. Variables 3. Method calls 4. Garbage collection
49. 1. Constants
50. 1. Constants
51. 1. Constants
52. 1. Constants
53. 1. Constants
54. 1. Constants
55. 1. Constants
56. 1. Constants
57. 2. Variables
58. 2. Variables
59. 2. Variables
60. 2. Variables
61. 2. Variables
62. 2. Variables
63. 2. Variables
64. 2. Variables
65. 2. Variables
66. 2. Variables
67. 2. Variables
68. 2. Variables 2021 2022 (tomorrow)
69. 3. Method calls
70. 3. Method calls
71. 3. Method calls
72. 3. Method calls
73. 3. Method calls
74. 3. Method calls
75. 3. Method calls • Code locality • Method inlining: C 㱻 Ruby • Pass arguments with native ABI • Deoptimization on redefinition or interruption (or TracePoint)
76. 4. Garbage collection
77. 4. Garbage collection
78. 4. Garbage collection
79. Next Steps • We still have a lot of rooms for improvements on yjit-bench • More cross-instruction optimizations • More method inlining over Ruby and C
80. Conclusion • Build your own JIT with Ruby 3.2 • Benchmark your JIT with yjit-bench