Making the RBS Parser Faster

RubyKaigi 2026, Hakodate, Hokkaido

Transcript

1. Making the RBS Parser Faster Soutaro Matsumoto (@soutaro)
 Shopify, Inc.
2. Soutaro Matsumoto • Senior Software Engineer at Shopify Ruby DX team • A Ruby core committer • RBS designer • Steep developer
3. Soutaro Matsumoto • Senior Software Engineer at Shopify Ruby DX team • A Ruby core committer • RBS designer • Steep developer
4. Soutaro Matsumoto • Senior Software Engineer at Shopify Ruby DX team • A Ruby core committer • RBS designer • Steep developer
5. RBS 4.0 & Steep 2.0
6. steep server singleton(T)[S] type Generics lower bounds: T > S ruby-rbs and ruby-rbs-sys crates steep check -e "1 + true" Several type narrowing updat RBS 4.0 & Steep 2.0 steep query
7. steep server Inline RBS declaration support 🎉 singleton(T)[S] type Generics lower bounds: T > S ruby-rbs and ruby-rbs-sys crates steep check -e "1 + true" Several type narrowing updat RBS 4.0 & Steep 2.0 steep query
8. Inline RBS Type Declaration Experimental
9. Inline RBS Type Declaration Experimental Type of attributes
10. Inline RBS Type Declaration Experimental Type of attributes Method type
11. Inline RBS Type Declaration • RBS and Steep directly support the feature, without using rbs-inline gem • Enable it by adding inline: true to the check calls in your Steep fi le Experimental
14. steep query • steep query lets you navigate your codebase from the command line • Output is in the same JSON format as LSP $ steep query hover signature_service:3:15 $ steep query definition Steep::Typing
15. steep query • steep query lets you navigate your codebase from the command line • Output is in the same JSON format as LSP $ steep query hover signature_service:3:15 $ steep query definition Steep::Typing
16. steep query • steep query lets you navigate your codebase from the command line • Output is in the same JSON format as LSP $ steep query hover signature_service:3:15 $ steep query definition Steep::Typing
17. steep query • steep query lets you navigate your codebase from the command line • Output is in the same JSON format as LSP $ steep query hover signature_service:3:15 $ steep query definition Steep::Typing
18. steep query • steep query lets you navigate your codebase from the command line • Output is in the same JSON format as LSP $ steep query hover signature_service:3:15 $ steep query definition Steep::Typing
19. steep query • steep query lets you navigate your codebase from the command line • Output is in the same JSON format as LSP $ steep query hover signature_service:3:15 $ steep query definition Steep::Typing
20. The RBS Parser • RBS parser translates source text into an AST • Using RBS always starts from parsing
21. The New RBS Parser
22. RBS::Parser in RBS gem The pure C parser The New Parser Architecture • The new parser constructs a pure C AST and translates it into Ruby objects The source code C Struct AST Ruby Object AST
25. 🤔
26. 🤔 We can ignore the issue because parsing accounts for only a small portion of total tool execution time
27. 🤔 We can ignore the issue because parsing accounts for only a small portion of total tool execution time We could parallelize parsing for better performance if needed.
28. 🤔 We can ignore the issue because parsing accounts for only a small portion of total tool execution time The new parser scans the input twice. It is inevitable. We could parallelize parsing for better performance if needed.
29. 🤔 We can ignore the issue because parsing accounts for only a small portion of total tool execution time The new parser scans the input twice. It is inevitable. We could parallelize parsing for better performance if needed.
30. 🤔 We can ignore the issue because parsing accounts for only a small portion of total tool execution time The new parser scans the input twice. It is inevitable. We could parallelize parsing for better performance if needed.
31. 🤔 We can ignore the issue because parsing accounts for only a small portion of total tool execution time The new parser scans the input twice. It is inevitable. We could parallelize parsing for better performance if needed.
32. 🤔 We can ignore the issue because parsing accounts for only a small portion of total tool execution time The new parser scans the input twice. It is inevitable. We could parallelize parsing for better performance if needed.
37. Small Inputs • Small inputs are more common with inline RBS declarations, as used in both RBS and Sorbet • We have many small RBS inputs in one fi le for annotations
38. Small Inputs • Small inputs are more common with inline RBS declarations, as used in both RBS and Sorbet • We have many small RBS inputs in one fi le for annotations
39. Start Pro fi ling • Best practice: pro fi le the code before you start • I started with some random optimizations (didn't work) • We use Instruments, a pro fi ler bundled in Xcode on macOS
40. Instruments
55. Looks like there is a chance to optimize the lexer
56. Looks like there is a chance to optimize the lexer
57. Lexer • Lexer -- lexical analyzer or tokenizer -- groups characters into tokens • Parser operates on tokens rather than individual characters c l a s s I n t e g e r < N u m e r i c Characters Tokens
58. Lexer • Lexer -- lexical analyzer or tokenizer -- groups characters into tokens • Parser operates on tokens rather than individual characters c l a s s I n t e g e r < N u m e r i c Characters Tokens kCLASS
59. Lexer • Lexer -- lexical analyzer or tokenizer -- groups characters into tokens • Parser operates on tokens rather than individual characters c l a s s I n t e g e r < N u m e r i c Characters Tokens kCLASS tUIDENT
60. Lexer • Lexer -- lexical analyzer or tokenizer -- groups characters into tokens • Parser operates on tokens rather than individual characters c l a s s I n t e g e r < N u m e r i c Characters Tokens kCLASS tUIDENT pLT
61. Lexer • Lexer -- lexical analyzer or tokenizer -- groups characters into tokens • Parser operates on tokens rather than individual characters c l a s s I n t e g e r < N u m e r i c Characters Tokens kCLASS tUIDENT pLT tUIDENT
62. re2c RBS uses re2c lexer generator, which supports Unicode Generated C code re2c De fi nition
63. re2c RBS uses re2c lexer generator, which supports Unicode Generated C code re2c De fi nition
64. re2c RBS uses re2c lexer generator, which supports Unicode Generated C code re2c De fi nition
65. re2c RBS uses re2c lexer generator, which supports Unicode Generated C code re2c De fi nition
66. re2c RBS uses re2c lexer generator, which supports Unicode Generated C code re2c De fi nition
67. re2c RBS uses re2c lexer generator, which supports Unicode Generated C code re2c De fi nition
68. re2c RBS uses re2c lexer generator, which supports Unicode Generated C code re2c De fi nition
69. t y p e t = " വ ؗ "
70. t y p e t = " വ ؗ "
71. t y p e t = " വ ؗ " t (116)
72. t y p e t = " വ ؗ " t (116) Moves 1 byte
73. t y p e t = " വ ؗ " t (116) Moves 1 byte
74. t y p e t = " വ ؗ " t (116) Moves 1 byte വ (20989)
75. t y p e t = " വ ؗ " t (116) Moves 1 byte Moves 3 bytes വ (20989)
76. 🤔 • rbs_peek calculates UTF-8 codepoint • rbs_skip calculates encoding dependent character width
77. Encoding of RBS Source Text • There was no speci fi cation of encoding of RBS source text • De fi ne an encoding spec 💪 • Follow Ruby's spec • It supports multi-byte encoding, but they must be ASCII compatible • UTF-8, SJIS, EUC-JP are fi ne, UTF-16 and UTF-32 are not
78. The Problem • Our lexer generator doesn't support encoding other than Unicode • Do we need to implement the conversion from supported encodings to Unicode? 😫
79. We don't need actual codepoints for parsing Only comments and string literal types allow non-ASCII characters
80. We don't need actual codepoints for parsing type language = "ϧϏʔ" type language = "ϥετ" Only comments and string literal types allow non-ASCII characters
81. We don't need actual codepoints for parsing type language = "ϧϏʔ" type language = "ϥετ" # ू߹Λදݱ͢ΔΫϥε class Set[T] end # ϧϧϧϧϧϧϧϧϧϧ class Set[T] end Only comments and string literal types allow non-ASCII characters
82. Skip Codepoint Calculation • If the next character is multi-byte, returning any Unicode codepoint works • If it's a single-byte character, it's an ASCII character • The actual comment/string content can be fetched from the bu ff er
83. Skip Codepoint Calculation • If the next character is multi-byte, returning any Unicode codepoint works • If it's a single-byte character, it's an ASCII character • The actual comment/string content can be fetched from the bu ff er
92. Benchmarking RBS(4.0.0.dev.4 in Gemfile-unicode) parsing with 86 files (2148952 bytes)... ✅ 30.108 i/s (33.254 ms/i) (±3.321%) Benchmarking RBS(4.0.0.dev.4 in Gemfile-base) parsing with 86 files (2148952 bytes)... ✅ 25.854 i/s (38.759 ms/i) (±3.868%) Lexer Updated • Correctly handles non-UTF-8 input • Add character byte count to the lexer struct, and use it in rbs_peek to skip second encoding->char_width call 16% improvement with core RBS benchmark
93. 😫 Far from 5x speedup
97. You should try a malloc-based allocator
98. You should try a malloc-based allocator Sure! Thanks!
99. You should try a malloc-based allocator Sure! Thanks! 🙊 ... Why though? The pro fi ler doesn't show anything related to the allocator problem. 🤔
100. You should try a malloc-based allocator Sure! Thanks! 🙊 ... Why though? The pro fi ler doesn't show anything related to the allocator problem. 🤔 It solved the problem! 😳
101. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete
102. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete Arena
103. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete Arena
104. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete Arena
105. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete Arena
106. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete Arena
107. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete Arena
108. Custom Allocator 1. Allocate large memory for arena at the start of parsing 2. Allocate memory from arena during parsing 3. Free the entire arena when parsing is complete
109. mmap (2) • mmap (2) is a system call that maps fi les directly into a process's address space • We use mmap simply to allocate a large contiguous block of memory for our custom allocator • munmap to return the memory to OS
110. malloc-Based Allocator • Use malloc to allocate memory and free to release the memory • malloc/free actually uses mmap/munmap internally, but the C library runtime usually reuse the memory blocks for better performance
111. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 Parsing #2
112. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 Parsing #2
113. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 Parsing #2
114. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2
115. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults
116. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults
117. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults
118. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults
119. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults C library runtime keeps the blocks
120. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults C library runtime keeps the blocks
121. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults C library runtime keeps the blocks
122. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults C library runtime keeps the blocks
123. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults C library runtime keeps the blocks
124. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults C library runtime keeps the blocks
125. Reusing Memory Blocks mmap-based Allocator malloc-based Allocator Parsing #1 ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ ⚡ Parsing #2 Page-faults No page-fault! C library runtime keeps the blocks
126. malloc Is Faster! • It triggers fewer page-faults than mmap-based allocator • Page-fault cost is relatively big with small fi les
127. Final Results core Steep Timee Big 300 Small 300 Baseline Lexer Allocator Head
128. Final Results 1.67x faster core Steep Timee Big 300 Small 300 Baseline Lexer Allocator Head
129. Final Results 1.67x faster core Steep Timee Big 300 Small 300 Baseline Lexer Allocator Head For smaller RBS fi les, baseline is slower then 3.9
130. Conclusion • The new RBS parser is up to 1.6x faster than RBS 3.9 • It correctly handles UTF-8 and other encoding inputs • Pro fi lers are useful, but they are not a silver bullet • For this parser optimization, the biggest improvement didn't come from pro fi ling 🤷 • RBS 4 and Steep 2 are out 📣