Method JIT Compiler for MRI

RubyElixirConf 2018
https://2018.rubyconf.tw/

Transcript

1. Method JIT Compiler for MRI RubyElixirConf Taiwan 2018 ~ Optimizations in Ruby 2.6.0 preview1, 2 ~ @k0kubun / Treasure Data Inc.
2. @k0kubun Treasure Data Inc. ERB maintainer, developing Ruby’s JIT
3. The history of MRI JIT
4. March 2017: RTL & MJIT
5. October 2017: YARV-MJIT 0QUDBSSPU
   XJUI
   -BO@.BTUFS GQT      Ruby 2.0 Ruby 2.5 YARV-MJIT RTL MJIT     https://github.com/k0kubun/yarv-mjit/tree/master-171211#optcarrot-benchmark
6. February 2018: Merge MJIT infrastructure
7. February 2018: Released in 2.6.0-preview1
8. How does it work?
9. Optionally enabled by "--jit" Tips: RUBYOPT="--jit" ruby … works too
10. New runtime dependency: gcc / clang
11. How Ruby’s method JIT works Methods Interpret
12. Methods Interpret Frequent calls ! How Ruby’s method JIT works
13. Methods Compile Machine code Interpret How Ruby’s method JIT works
14. Methods Machine code Interpret Call How Ruby’s method JIT works
15. Methods Machine code Interpret Call How Ruby’s method JIT works Compile
16. Methods Machine code Call How Ruby’s method JIT works Compile
17. Machine code Call How Ruby’s method JIT works
18. Latest Ruby’s performance benchmarks
19. Ruby 2.6.0-preview1 https://benchmark-driver.github.io/benchmarks/optcarrot/releases.html
20. Ruby trunk https://benchmark-driver.github.io/benchmarks/optcarrot/commits.html
21. Ruby trunk https://benchmark-driver.github.io/benchmarks/optcarrot/commits.html 2.6.0 Preview1 2.6.0 Preview2 ?
22. Micro benchmark: while 5.7x faster 2.6.0 Preview1 2.6.0 Preview2 ? https://benchmark-driver.github.io/benchmarks/mjit/commits.html
23. 0QUDBSSPU
    XJUI
    -BO@.BTUFS GQT      Ruby 2.0 trunk trunk+JIT RTL+JIT Ruby 3x3      But… we’re still far from Ruby 3x3 https://gist.github.com/k0kubun/7074ad434d0affd1bd98edaaa011ac1d 39fps to go
24. How to get there? Just inlining method doesn’t help if code is too complex We need more effort to exploit C compiler optimizations Let’s see what we’ve done so far
25. 2.6.0-Preview1 Optimizations
26. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code
27. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code ISeq Compile putself send :bar, cache: nil leave
28. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: nil leave ISeq Ruby VM Program Counter Interpret
29. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: nil leave ISeq Ruby VM Program Counter Call putself() { val = GET_SELF(); } C code for instruction
30. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: nil leave ISeq Ruby VM Program Counter Interpret
31. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: nil leave ISeq Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction
32. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: nil leave ISeq Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction Ruby method push C method call attr_reader attr_writer . . . Which type will be called?
33. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: Ruby leave ISeq Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction Store C function pointer w/ class timestamp Ruby method push C method call attr_reader attr_writer . . .
34. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: Ruby leave Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction Ruby method push Dispatch it by calling function pointer (compiler can't optimize)
35. 1. Basic inlining of Ruby method (r62197) def foo bar end Ruby code putself send :bar, cache: Ruby leave Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction Ruby method push Ruby method push In JIT, we can inline this operation by checking cache in ISeq ISeq
36. 1. Basic inlining of Ruby method (r62197) Using “method cache”, we can bypass method dispatch and inline the C function to push Ruby method frame If it's inlined, C compiler can apply various optimizations to Ruby method call, which is known as slow Optcarrot: 53.84fps -> 57.52fps
37. 2. Bypass Array/Hash check for #[] (r62398) optimized_#[](recv, key) { if recv.is_a?(Array) { fast_Array#[](recv, key); } else if recv.is_a?(Hash) { fast_Hash#[](recv, key); } else { dispatch(recv, #[], key); } }
38. 2. Bypass Array/Hash check for #[] (r62398) optimized_#[](recv, key) { if recv.is_a?(Array) { fast_Array#[](recv, key); } else if recv.is_a?(Hash) { fast_Hash#[](recv, key); } else { dispatch(recv, #[], key); } } array = [1,2,3] array[1]
39. 2. Bypass Array/Hash check for #[] (r62398) optimized_#[](recv, key) { if recv.is_a?(Array) { fast_Array#[](recv, key); } else if recv.is_a?(Hash) { fast_Hash#[](recv, key); } else { dispatch(recv, #[], key); } } hash = { foo: 1} hash[:foo]
40. 2. Bypass Array/Hash check for #[] (r62398) optimized_#[](recv, key) { if recv.is_a?(Array) { fast_Array#[](recv, key); } else if recv.is_a?(Hash) { fast_Hash#[](recv, key); } else { dispatch(recv, #[], key); } } def show params[:id] end ActionController::Parameters#[]
41. 2. Bypass Array/Hash check for #[] (r62398) optimized_#[](recv, key) { if recv.is_a?(Array) { fast_Array#[](recv, key); } else if recv.is_a?(Hash) { fast_Hash#[](recv, key); } else { dispatch(recv, #[], key); } } def show params[:id] end ActionController::Parameters#[] These checks are NOT needed for classes other than Array, Hash
42. 2. Bypass Array/Hash check for #[] (r62398) jit_#[](recv, key) { dispatch(recv, #[], key); } def show params[:id] end ActionController::Parameters#[]
43. 2. Bypass Array/Hash check for #[] (r62398) Ruby always optimizes #[] for Array/Hash, but it’s suboptimal for other classes JIT removes the guard for Array/Hash by seeing call cache, and also inlines pushing a method frame The same optimization can be applied to other methods later
44. 3. Inline Array#[] with Integer (r62388) optimized_#[](recv, key) { if recv.is_a?(Array) { fast_Array#[](recv, key); // extern } else if recv.is_a?(Hash) { fast_Hash#[](recv, key); } else { dispatch(recv, #[], key); } } It's not inlined and optimized well by compiler
45. 3. Inline Array#[] with Integer (r62388) optimized_#[](recv, key) { if recv.is_a?(Array) { if key.is_a?(Integer) { Array#[Integer](recv, key); // inline } else { fast_Array#[](recv, key); // extern } } else if recv.is_a?(Hash) { fast_Hash#[](recv, key); } else { dispatch(recv, #[], key); } } This special path is inlined and optimized well on JIT
46. 3. Inline Array#[] with Integer (r62388) Currently “JIT header“ has limited definitions of C functions in Ruby core I inlined a part of Array#[] definition, and then C compiler could optimize the code Optcarrot: 54.93fps -> 58.41fps
47. 2.6.0-Preview1 wrap up I mainly worked for portability, stability, maintainability Fix SEGV and deadlock, remove broken optimizations… Notable optimizations were only 3, so it wasn't fast yet
48. 2.6.0-Preview2 Optimizations
49. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code
50. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave
51. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack empty
52. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack 1
53. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack 1
54. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack 1 2
55. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack 1 2
56. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack 3
57. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack 3
58. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave Ruby VM Program Counter Stack Pointer VM stack 3 How to skip the stack pointer motion in JIT?
59. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave jit_three() { } JIT-ed code: before
60. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave jit_three() { *sp = 1; sp++; } JIT-ed code: before
61. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave jit_three() { *sp = 1; sp++; *sp = 2; sp++; } JIT-ed code: before
62. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave jit_three() { *sp = 1; sp++; *sp = 2; sp++; *(sp-2) = opt_plus( *(sp-2),*(sp-1)); sp--; } JIT-ed code: before
63. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave jit_three() { *sp = 1; sp++; *sp = 2; sp++; *(sp-2) = opt_plus( *(sp-2),*(sp-1)); sp--; return *(sp-1); } JIT-ed code: before
64. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave jit_three() { *sp = 1; sp++; *sp = 2; sp++; *(sp-2) = opt_plus( *(sp-2),*(sp-1)); sp--; return *(sp-1); } JIT-ed code: before jit_three() { VALUE stack[2]; stack[0] = 1; stack[1] = 2; stack[0] = opt_plus( stack[0], stack[1]); return stack[0]; } JIT-ed code: after
65. 1. Use C local variable for VM stack (r62655) def three 1 + 2 end Ruby code ISeq putobject 1 putobject 2 opt_plus leave jit_three() { *sp = 1; sp++; *sp = 2; sp++; *(sp-2) = opt_plus( *(sp-2),*(sp-1)); sp--; return *(sp-1); } JIT-ed code: before jit_three() { VALUE stack[2]; stack[0] = 1; stack[1] = 2; stack[0] = opt_plus( stack[0], stack[1]); return stack[0]; } JIT-ed code: after Array local variable This seems okay for just "1 + 2", but...
66. 1. Use C local variable for VM stack (r62655) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code
67. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C VM stack empty
68. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) VM stack empty
69. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) jit_three() stack[nil, nil] in jit_three() VM stack empty
70. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) jit_three() stack[1, nil] in jit_three() Push 1 to array local variable VM stack empty
71. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) jit_three() stack[1, nil] in jit_three() jit_err() VM stack empty
72. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) jit_three() stack[1, nil] in jit_three() jit_err() rb_raise() (call longjmp) VM stack empty
73. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) jit_three() stack[1, nil] in jit_three() jit_err() rb_raise() (call longjmp) longjmp purges JIT-ed frames VM stack empty
74. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) jit_three() stack[1, nil] in jit_three() jit_err() rb_raise() (call longjmp) VM stack empty 2
75. 1. Use C local variable for VM stack (r62655) def err # JIT-ed raise 'error' end def three # JIT-ed 1 + (err rescue 2) end Ruby code main() Call stack in C ruby_vm() (setjmp called) jit_three() stack[1, nil] in jit_three() jit_err() rb_raise() (call longjmp) VM stack empty 2 VM Stack doesn't have 2 values => SEGV 1 is expired
76. 1. Use C local variable for VM stack (r62655) When "catch table" (rescue, ensure, etc.) does not exist, we don't need to resurrect stack values on exception So we can use just C local variables to reproduce the stack of Ruby VM only when catch table does not exist Stack pointer is not moved and compiler can inline values Optcarrot: 57.13fps -> 62.14fps
77. 2. Bypass setjmp for yield (r62643) setjmp is slow If JIT-ed code is directly called from VM (no C function frames are created yet), we don’t need to call setjmp again Now yield is 1.3x faster than a non-JIT-ed case
78. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code
79. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code Ruby call stack #three Program Counter
80. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code Ruby call stack #three Program Counter
81. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code Ruby call stack #three Program Counter
82. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code Ruby call stack #three #err Program Counter Program Counter
83. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code Ruby call stack #three #err Program Counter Program Counter
84. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code Ruby call stack #three #err Program Counter Program Counter #raise Program Counter longjmp
85. 3. Skip moving program counter (r62678) def err raise 'error' end def three 1 + (err rescue 2) end Ruby code Ruby call stack #three #err Program Counter Program Counter #raise Program Counter Program counter is used to resurrect the position after longjmp
86. 3. Skip moving program counter (r62678) Same as the stack value's situation, we don't move the program counter only when catch table does not exist (rescue, ensure, etc.) Optcarrot: 64.92fps -> 68.08fps
87. 4. Force inlining arithmetic instructions (r62677) C compiler has a threshold of function size to be inlined Some Ruby's instructions (+, -, *, /, ...) are too large to be inlined by default, so I applied an "always inline" attribute In the future, we should reduce the size of code instead Optcarrot: 60.19fps -> 64.92fps
88. 5. Force inlining ivar instructions (r62693) Not only arithmetic instructions, but also instructions for instance variable are large too, so I force-inlined it Optcarrot: 67.04fps -> 68.20fps
89. 6. Disable stack consistency check (r63092) Ruby VM is always asserting the size of stack when returning from a method, and it's slow We can skip it on JIT because it's already checked by VM Optcarrot: 67.43fps -> 69.92fps
90. 7. Inline attr_reader method call (r63212) . def foo bar end Ruby code putself send :bar, cache: nil leave ISeq Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction Ruby method push C method call attr_reader attr_writer . .
91. 7. Inline attr_reader method call (r63212) def foo bar end Ruby code putself send :bar, cache: attr leave ISeq Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction Ruby method push C method call attr_reader attr_writer . . .
92. 7. Inline attr_reader method call (r63212) def foo bar end Ruby code putself send :bar, cache: attr leave ISeq Ruby VM Program Counter Call send(cache) { search_method(cache); CALL_METHOD(cache); } C code for instruction get_istance_variable() attr_reader
93. 7. Inline attr_reader method call (r63212) Using call cache in the same way as Ruby method, we can fully inline attr_reader without large compilation time The cost becomes the same as reference to normal instance variables Calling attr_reader is made 4x faster
94. 2.6.0-Preview2 (trunk) wrap up I've mainly worked on performance because it's useless if it's slow Generated code is much simplified and made fast by removing program counter and stack pointer motions But it still has some complexity and it blocks significant performance improvement by Ruby method inlining
95. Future of Ruby's JIT
96. 1. Deoptimization by longjmp We can generate aggressive code and cancel all JIT-ed calls by longjmp when something unexpected happens I’m going to remove guard for TracePoint and cancel it later It should also be used when all method caches are purged
97. 2. Instruction specialization for types Currently the same code is generated for both Hash#[] and Array#[] We need some instrumentation to detect the type which is passed to an optimized instruction Vladimir's RTL instruction achieves this by dynamic modification of instruction
98. 3. Multi-tier JIT Some other languages have multiple stages for JIT Depending on how frequently it's called, it may be better to balance compilation time and optimization level Vladimir is working on light JIT compilation Sometimes people deploy an application every 10 minutes
99. 4. Profile-guided JIT C compiler has a feature to profile compiled code and generate faster code using the profiling result Using the multi-tier JIT, we may be able to profile code in the first tier and generate faster code in the second tier
100. 5. Better JIT scheduler for Rails In Rails, an application becomes slower only during JIT compilation happens The possible cause might be the number of methods to be JIT-ed, compared to some other benchmarks Possibly we should reduce the number of methods to be JIT-ed or reduce frequency of JIT compilation
101. 6. Ruby / C method inlining I already succeeded to implement Ruby method inlining, but it increases compilation time I have ideas to implement C method inlining, but which method to be inlined should be solved first
102. 7. Exploit more C compiler optimizations Loop invariant motion Folding Ruby's constant Type check removal by type inference Reduce unnecessary memory accesses to VM registers
103. Conclusion 2.6.0-preview2 will be much faster than 2.6.0-preview1 (Still not ready for Rails) We still have so many things to be done for Ruby 3x3