Milion rzeczy, które musi zrobić YARV zanim wykona Twój kod

Transcript

1. 1.

   Historya prawdziwa opowiedziana 19 grudnia 2012, w której Jan Stępień

   wraz z gronem śmiałków spod chorągwi Czerwonego Kryształu pogrążą się w niezbadanych czeluściach Maszyn Wirtualnych a na dnie samym staną oko w oko z monstrum przez człowieczy rozum nieogarnionym, co YARV się zowie.
2. 2.

   Maszyny wirtualne czynią życie prostszym Abstrakcja nad pamięcią Dynamiczny wpływ

   na działanie Abstrakcja nad sprzętem Nie ma nic za darmo
3. 3.

   Lisp, SmallTalk, Self… To były czasy!

4. 4.

   So, Ruby was a Lisp originally, in theory. Let’s call

   it MatzLisp from now on. ;-) Matz Re: Ruby’s lisp features 13 lutego 2006
5. 5.

   I have always been more interested in designing the language

   than implementing it. So Ruby interpreter is always slower than it should be. Matz The Ruby VM: Episode I 2007
6. 6.

   Matz’s Ruby Interpreter

7. 7.

   Yet another Ruby VM 笹田耕一 (Koichi Sasada)

8. 8.

   None
9. 9.

   parse.c Kod → abstract syntax tree

10. 10.

    None
11. 11.

    irb(main):000:0> require ’ripper’ irb(main):001:0> pp Ripper.sexp ”Math.sqrt 3 + 4”

    [:program, [[:command_call, [:var_ref, [:@const, ”Math”, [1, 0]]], :”.”, [:@ident, ”sqrt”, [1, 5]], [:args_add_block, [[:binary, [:@int, ”3”, [1, 10]], :+, [:@int, ”4”, [1, 14]]]], false]]]]
12. 12.

    compile.c abstract syntax tree → bytecode

13. 13.

    RubyVM::InstructionSequence (znany również jako iseq)

14. 14.

    irb(main):000:0> iseq = RubyVM::InstructionSequence.compile ”x + 6” irb(main):001:0> puts iseq.disasm

    == disasm: <RubyVM::InstructionSequence> == 0000 trace 1 ( 1) 0002 putself 0003 send :x, 0, nil, 24, <ic:0> 0009 putobject 6 0011 opt_plus <ic:2> 0013 leave
15. 15.

    irb(main):002:0> pp iseq.to_a [”YARVInstructionSequence/SimpleDataFormat”, 1, 2, 1, {arg_size: 0, local_size:

    1, stack_max: 2}, ”<compiled>”, ”<compiled>”, nil, 1, :top, [], 0, [], [1, [:trace, 1], [:putself], [:send, :a, 0, nil, 24, 0], [:putself], [:send, :b, 0, nil, 24, 1], [:opt_plus, 3], [:leave]]]
16. 16.

    #if SUPPORT_JOKE ... CONST_ID(goto_id, ”__goto__”); CONST_ID(label_id, ”__label__”); if (nd_type(node) ==

    NODE_FCALL && (mid == goto_id || mid == label_id)) { ... #endif
17. 17.

    $ cd ruby $ sed -i vm_opts.h -e ’s/define SUPPORT_JOKE

    *0/define SUPPORT_JOKE 1/’ $ make
18. 18.

    __label__ :start puts ”haters gonna hate” __goto__ :start

19. 19.

    haters gonna hate haters gonna hate haters gonna ha gonna

    hate haters gonna hate haters gonna hate hate hate haters gonna hate haters gonna hate haters go haters gonna hate haters gonna hate haters gonna ha gonna hate haters gonna hate haters gonna hate hate hate haters gonna hate haters gonna hate haters go haters gonna hate haters gonna hate haters gonna ha gonna hate haters gonna hate haters gonna hate hate hate haters gonna hate haters gonna hate haters go haters gonna hate haters gonna hate haters gonna ha gonna hate haters gonna hate haters gonna hate hate hate haters gonna hate haters gonna hate haters go haters gonna hate haters gonna hate haters gonna ha
20. 20.

    insns.def Definicje instrukcji maszyny wirtualnej

21. 21.

    /** @c put @e put new array. @j 新しい配列をスタック上の num

    個の値で初期化して生成しプッシュする。 */ DEFINE_INSN newarray (rb_num_t num) (...) (VALUE val) // inc += 1 - num; { val = rb_ary_new4((long)num, STACK_ADDR_FROM_TOP(num)); POPN(num); }
22. 22.

    /** @c joke @e The Answer to Life, the Universe,

    and Everything @j 人生、宇宙、すべての答え。 */ DEFINE_INSN answer () () (VALUE ret) { ret = INT2FIX(42); }
23. 23.

    irb(main):000:0> RubyVM::INSTRUCTION_NAMES => [”nop”, ”getlocal”, ”setlocal”, ”getspecial”, ”setspecial”, ”getdynamic”, ”setdynamic”,

    ”getinstancevariable”, ”setinstancevariable”, ”getclassvariable”, ”setclassvariable”, ”getconstant”, ”setconstant”, ...
24. 24.

    YARVinstructions — How Ruby 1.9 Executes Your Ruby Script http://yarvinstructions.heroku.com

25. 25.

    $ cd ruby $ make V=1 ... ruby ./tool/insns2vm.rb ￿srcdir=”.”

    insns.inc ruby ./tool/insns2vm.rb ￿srcdir=”.” insns_info.inc ruby ./tool/insns2vm.rb ￿srcdir=”.” optinsn.inc ruby ./tool/insns2vm.rb ￿srcdir=”.” optunifs.inc ruby ./tool/insns2vm.rb ￿srcdir=”.” opt_sc.inc ruby ./tool/insns2vm.rb ￿srcdir=”.” vmtc.inc ruby ./tool/insns2vm.rb ￿srcdir=”.” vm.inc ...
26. 26.

    vm.inc To tu siedzi gros maszyny wirtualnej

27. 27.

    Przykład: getinstancevariable label_insn_getinstancevariable: id = stack_pop(); ic = stack_pop(); val

    = vm_get_ivar(get_self(), id, ic); stack_push(val); instruction_pointer += 1; goto instruction_labels[instruction_pointer];
28. 28.

    vm_exec_core Funkcja, która nie przeszłaby żadnego code review

29. 29.

    Chwila na zebranie myśli

30. 30.

    vm_call_cfunc 7 (0.4%) of 1285 (80.9%) rb_ary_each 0 (0.0%) of

    981 (61.8%) 1656 rb_f_catch 0 (0.0%) of 898 (56.5%) 883 rb_vm_invoke_proc 1 (0.1%) of 810 (51.0%) 872 t_run_machine_without_threads 0 (0.0%) of 724 (45.6%) 706 rb_require_safe 0 (0.0%) of 679 (42.8%) 1109 rb_obj_tap 0 (0.0%) of 538 (33.9%) 522 rb_ary_collect 0 (0.0%) of 416 (26.2%) 819 rb_push_glob 0 (0.0%) of 208 (13.1%) 208 rb_class_new_instance 0 (0.0%) of 195 (12.3%) 230 rb_ensure 0 (0.0%) of 184 (11.6%) 35 enum_all 0 (0.0%) of 115 (7.2%) 115 yield_under 0 (0.0%) of 102 (6.4%) 134 enum_any 0 (0.0%) of 90 (5.7%) 90 rb_yield 1 (0.1%) of 1216 (76.6%) 1891 rb_catch_obj 0 (0.0%) of 915 (57.6%) 898 vm_exec 6 (0.4%) of 1285 (80.9%) 910 EventMachine_t Run 0 (0.0%) of 737 (46.4%) 724 rb_load_internal 0 (0.0%) of 689 (43.4%) 1309 search_required 0 (0.0%) of 95 (6.0%) 95 542 818 push_glob 0 (0.0%) of 207 (13.0%) 207 vm_call0 5 (0.3%) of 1160 (73.0%) 214 34 rb_block_call 0 (0.0%) of 305 (19.2%) 115 vm_yield_with_cref 0 (0.0%) of 103 (6.5%) 137 90 vm_call_method 35 (2.2%) of 1285 (80.9%) 8049 35 6 vm_exec_core 73 (4.6%) of 1285 (80.9%) 8628 8101 73 rb_funcall 3 (0.2%) of 933 (58.8%) 3 rb_autoload_load 0 (0.0%) of 145 (9.1%) 3 vm_get_ev_const.isra.6 1 (0.1%) of 141 (8.9%) 285 rb_const_get_0 0 (0.0%) of 125 (7.9%) 4 2 1112 280 114 171 183 vm_yield 1 (0.1%) of 1235 (77.8%) 4414 vm_yield_with_cfunc 1 (0.1%) of 263 (16.6%) 273 39 all_iter_i 1 (0.1%) of 115 (7.2%) 115 any_iter_i 0 (0.0%) of 88 (5.5%) 88 3479 228 35 1307 5 catch_i 0 (0.0%) of 926 (58.3%) 926 914 event_callback_wrapper 0 (0.0%) of 835 (52.6%) event_callback_wrapper (inline) 0 (0.0%) of 835 (52.6%) 835 835 ConnectionDescriptor _DispatchInboundData 0 (0.0%) of 808 (50.9%) 808 EventMachine_t _RunSelectOnce 0 (0.0%) of 787 (49.6%) ConnectionDescriptor Read 0 (0.0%) of 779 (49.1%) 772 779 EventMachine_t _RunOnce 0 (0.0%) of 765 (48.2%) 759 737 rb_iseq_eval 0 (0.0%) of 697 (43.9%) 1286 1269 rb_feature_p 2 (0.1%) of 124 (7.8%) 79 rb_iseq_eval_main 0 (0.0%) of 396 (24.9%) 396 ruby_exec_internal 0 (0.0%) of 382 (24.1%) 382 ruby_exec_node 0 (0.0%) of 369 (23.2%) 369 ruby_run_node 0 (0.0%) of 358 (22.5%) 352 main 0 (0.0%) of 342 (21.5%) 342 __libc_start_main 0 (0.0%) of 330 (20.8%) 330 _start 0 (0.0%) of 306 (19.3%) 306 rb_iterate 0 (0.0%) of 305 (19.2%) 314 313 114 88 ruby_brace_glob0 0 (0.0%) of 207 (13.0%) 207 ruby_brace_expand.constprop.1 7 (0.4%) of 193 (12.2%) 193 ruby_glob0 0 (0.0%) of 182 (11.5%) 14 366 168 glob_helper 6 (0.4%) of 111 (7.0%) 111 125 2 rb_get_expanded_load_path 0 (0.0%) of 106 (6.7%) 105 139 eval_string_with_cref 1 (0.1%) of 98 (6.2%) 67 1 st_foreach 56 (3.5%) of 339 (21.3%) 56 mark_entry 2 (0.1%) of 241 (15.2%) 743 mark_method_entry_i 8 (0.5%) of 233 (14.7%) 986 mark_const_entry_i 3 (0.2%) of 165 (10.4%) 323 mark_keyvalue 0 (0.0%) of 138 (8.7%) 233 gc_mark_children 53 (3.3%) of 284 (17.9%) 729 proc_mark 0 (0.0%) of 165 (10.4%) 6 gc_mark 127 (8.0%) of 128 (8.1%) 15 8 96 iseq_mark 32 (2.0%) of 233 (14.7%) 739 117 12 3 319 212 2 7 3059 mark_tbl 3 (0.2%) of 256 (16.1%) 711 mark_m_tbl 0 (0.0%) of 234 (14.7%) 982 332 86 mark_const_tbl 0 (0.0%) of 163 (10.3%) 325 mark_hash 0 (0.0%) of 137 (8.6%) 230 60 751 3 993 1050 62 22 21 env_mark 0 (0.0%) of 158 (9.9%) 199 328 234 117 vm_xmalloc 1 (0.1%) of 223 (14.0%) vm_malloc_prepare 3 (0.2%) of 186 (11.7%) 173 2 garbage_collect 0 (0.0%) of 187 (11.8%) 183 gc_marks 0 (0.0%) of 110 (6.9%) 91 gc_sweep 0 (0.0%) of 96 (6.0%) 96 31 slot_sweep 53 (3.3%) of 99 (6.2%) 93 mark_locations_array 4 (0.3%) of 158 (9.9%) 255 203 60 4 str_new 2 (0.1%) of 115 (7.2%) 103 2 53 rb_usascii_str_new 1 (0.1%) of 94 (5.9%) 93
31. 31.

    glob_helper 6 (0.4%) of 111 (7.0%) 125 st_foreach 56 (3.5%)

    of 339 (21.3%) 56 mark_entry 2 (0.1%) of 241 (15.2%) 743 mark_method_entry_i 8 (0.5%) of 233 (14.7%) 986 mark_const_entry_i 3 (0.2%) of 165 (10.4%) 323 mark_keyvalue 0 (0.0%) of 138 (8.7%) 233 gc_mark_children 53 (3.3%) of 284 (17.9%) 729 proc_mark 0 (0.0%) of 165 (10.4%) 6 15 8 96 iseq_mark 32 (2.0%) of 233 (14.7%) 739 117 12 3 319 212 2 3059 mark_tbl 3 (0.2%) of 256 (16.1%) 711 982 332 86 mark_const_tbl 0 (0.0%) of 163 (10.3%) 325 mark_hash 0 (0.0%) of 137 (8.6%) 230 60 751 3 9 1050 62 22 21 env_mark 0 (0.0%) of 158 (9.9%) 199 328 234 gc_marks 0 (0.0%) of 110 (6.9%) 91 31 255 203 60
32. 32.

    872 1109 vm_exec 6 (0.4%) of 1285 (80.9%) 910 rb_block_call

    0 (0.0%) of 305 (19.2%) 115 vm_yield_with_cref 0 (0.0%) of 103 (6.5%) 137 90 vm_call_method 35 (2.2%) of 1285 (80.9%) 8049 35 6 vm_exec_core 73 (4.6%) of 1285 (80.9%) 8628 8101 73 3 3 vm_get_ev_const.isra.6 1 (0.1%) of 141 (8.9%) 285 rb_const_get_0 0 (0.0%) of 125 (7.9%) 4 114 171 183 14 1307 ventMachine_t RunSelectOnce 0 (0.0%) of 787 (49.6%) onDescriptor Read 0 (0.0%) of 779 (49.1%) 772 9 759 1286 rb_iseq_eval_main 0 (0.0%) of 396 (24.9%) 396 382 314 139 eval_string_with_cref 1 (0.1%) of 98 (6.2%) 67 1 mar
33. 33.

    Wywołanie metody w pięciu prostych krokach

34. 34.

    1. Przygotuj argumenty i blok 2. Przygotuj stos jeśli metoda

    zwraca wiele wartości 3. Ustal klasę odbiorcy 4. Znajdź wołaną metodę 5. Wywołaj metodę
35. 35.

    Cache lekarstwem na wszystkie zmartwienia

36. 36.

    Liczba klas obiektów, które pojawiają się w miejscach wywołań metod

37. 37.

    Inline cache Zapisuj znalezioną metodę w miejscach wywołania

38. 38.

    1. Przygotuj argumenty i blok 2. Przygotuj stos jeśli metoda

    zwraca wiele wartości 3. Ustal klasę odbiorcy 4. Zajrzyj do inline cache 5. Zajrzyj do globalnego cache 6. Znajdź wołaną metodę normalnie 7. Wywołaj metodę
39. 39.

    Specjalizacja instrukcji Przyspieszanie standardowo wykonywanych operacji

40. 40.

    send Instrukcja maszyny wirtualnej, która służy do wywoływania metod

41. 41.

    a + b → opt_plus

42. 42.

    a[b] → opt_aref

43. 43.

    YARV: Yet Another RubyVM Innovating the Ruby Interpreter Koichi Sasada

    Graduate School of Technology, Tokyo University of Agriculture and Technology 2-24-16 Nakacho, Koganei-shi, Tokyo, Japan. sasada@namikilab.tuat.ac.jp ABSTRACT Ruby - an Object-Oriented scripting language - is used world- wide because of its ease of use. However, the current in- terpreter is slow. To solve this problem, some virtual ma- chines were developed, but none with adequate performance or functionality. To fill this gap, I have developed a Ruby interpreter called YARV (Yet Another Ruby VM). YARV is based on a stack machine architecture and features op- timizations for high speed execution of Ruby programs. In this poster, I introduce the Ruby programming language, discuss certain characteristics of Ruby from the aspect of a Ruby interpreter implementer, and explain methods of implementation and optimization. Benchmark results are given at the end. Categories and Subject Descriptors D.3 [PROGRAMMING LANGUAGES]: Processors— Interpreters General Terms Languages • Normal OO features (class, method call, etc.) • Advanced OO features (all values are objects, Min-in, Singleton method, etc.) • Dynamic-typing, re-definable behavior, dynamic eval- uation • Operator overloading • Exception handling • Closure and method invocation with a block • Garbage collection support • Dynamic module loading • Many useful libraries • Highly portable However, the current Ruby intepreter (old-ruby) is slow. This is because it works by traversing abstract syntax tree and evaluating each node. To solve this problem, I have de-
44. 44.

    O czym się nawet nie zająknąłem Garbage collector MRI 2.0

    Inne maszyny wirtualne
45. 45.

    Podsumowanie! 1. Poczynając od MRI 1.9 Ruby jest wykonywany na

    prawdziwej maszynie wirtualnej, 2. Kod → abstract syntax tree → bytecode, 3. 80 instrukcji YARV jest zdefiniowanych w insns.def, 4. Optymalizacje: cache metod, inline cache, specjalizacja instrukcji.
46. 46.

    jan@stepien.cc hp://stepien.cc/~jan @janstepien Serdecznie dziękuję

47. 47.

    Prezentację przygotowano przy pomocy pakietu L A TEX. © 2012

    Jan Stępień. Część praw zastrzeżono.