CodeFest 2019. Ujjwal Sharma (Node.js) — V8 by example: A journey through the compilation pipeline

V8 is complicated. Things change way too fast and it’s really hard to keep track of what’s the fastest way of doing every specific action. But not anymore. Join me, a V8 contributor, on a journey through the compilation pipeline of V8 and understand how it all works under the hood. We’ll take the example of a popular JavaScript builtin method and find that what does and does not trigger de-optimization. By the end of the talk, you will have a fairly decent idea of how builtins are written inside the V8 compilation pipeline, and how to make sure you always take the fast path, no matter what.

Transcript

1. @ryzokuken V8 by Example A journey through the compilation pipeline 1
2. @ryzokuken Ujjwal Sharma (@ryzokuken) • Node.js – Core Collaborator • Contribute to the JS/Node.js ecosystem ◦ V8 ◦ TC39 ◦ libuv ◦ … • Student • Google Summer of Code • Speaker 2
3. @ryzokuken @ryzokuken What is V8? 3
4. @ryzokuken @ryzokuken 4
5. @ryzokuken @ryzokuken How does V8 work? 5
6. @ryzokuken @ryzokuken 6 Source
7. @ryzokuken @ryzokuken 7 Source Parser
8. @ryzokuken @ryzokuken 8 Source Parser
9. @ryzokuken @ryzokuken 9 Source IIFE?
10. @ryzokuken @ryzokuken 10 Source IIFE? Eager
11. @ryzokuken @ryzokuken 11 Source IIFE? Lazy Eager
12. @ryzokuken @ryzokuken 12 Source IIFE? * Eventually Eager Lazy
13. @ryzokuken @ryzokuken 13 Source IIFE? Eager Lazy AST + Scopes
14. @ryzokuken @ryzokuken 14 Source Parser AST + Scopes
15. @ryzokuken @ryzokuken 15 Source Parser AST + Scopes Ignition
16. @ryzokuken @ryzokuken 16 Source Parser AST + Scopes Ignition Bytecode
17. @ryzokuken @ryzokuken 17 Source Parser AST + Scopes Ignition Bytecode
18. @ryzokuken @ryzokuken 18 But interpreters are so slow! Idea: Let’s not be slow. Let’s use a JIT compiler.
19. @ryzokuken @ryzokuken 19
20. @ryzokuken @ryzokuken 20 Source Parser AST + Scopes Ignition Bytecode Turbofan + profiling data OPTIMIZE!
21. @ryzokuken @ryzokuken 21 Source Parser AST + Scopes Ignition Bytecode Turbofan Optimized Code DEOPTIMIZE!
22. @ryzokuken @ryzokuken 22 Source Parser AST + Scopes Ignition Bytecode Turbofan Optimized Code
23. @ryzokuken @ryzokuken How you feel about everything. Let’s take a couple steps back. 23
24. @ryzokuken @ryzokuken How does V8 work? 24
25. @ryzokuken @ryzokuken How does V8 work? a compiler 25
26. @ryzokuken @ryzokuken Source Code Parser Abstract Syntax Tree Assembler Machine Code (R) Compiler Assembly Program Linker Machine Code (T) 26
27. @ryzokuken @ryzokuken Is it simple enough yet? 27
28. @ryzokuken @ryzokuken Source Code Parser Abstract Syntax Tree Assembler Machine Code (R) Compiler Assembly Program Linker Machine Code (T) 28
29. @ryzokuken @ryzokuken Source Code Parser Abstract Syntax Tree Compiler Machine Code (T) 29
30. @ryzokuken @ryzokuken 30 BAD NEWS: V8 isn’t this simple
31. @ryzokuken @ryzokuken Because ECMAScript is slow by default, that’s why. 31
32. @ryzokuken @ryzokuken Idea: Let’s not be slow. BUT HOW? Speculative Optimization 32
33. @ryzokuken @ryzokuken But wait… What even is Speculative Optimization? 33
34. @ryzokuken @ryzokuken Speculative Optimization [spek-yuh-ley-tiv, -luh-tiv][op-tuh-muh-zey-shuh n] Noun. Guessing what is about to happen based on what has happened. 34
35. @ryzokuken @ryzokuken Speculative Optimization [spek-yuh-ley-tiv, -luh-tiv][op-tuh-muh-zey-shuh n] Noun. Guessing what is about to happen based on what has happened. Making assumptions about possible future inputs based on previous inputs. 35
36. @ryzokuken @ryzokuken Top three reasons why a JavaScript function might need optimization: 1. Dynamic Types 2. Dynamic Types 3. Dynamic Types Alright, but how does it help? 36
37. @ryzokuken @ryzokuken Source Code Parser Abstract Syntax Tree Compiler Machine Code (T) 37
38. @ryzokuken @ryzokuken Source Code Parser Abstract Syntax Tree 1. Baseline Interpreter 2. Optimizing Compiler Machine Code (T) 38
39. @ryzokuken @ryzokuken AST Baseline Interpreter Optimizing Compiler Bytecode Optimized Code optimize deoptimize + profiling data 39
40. @ryzokuken @ryzokuken Compiler Baseline Interpreter Optimizing Compiler 40
41. @ryzokuken @ryzokuken AST Baseline Interpreter Optimizing Compiler Bytecode Optimized Code optimize deoptimize 41 + profiling data
42. @ryzokuken @ryzokuken AST Bytecode Optimized Code optimize deoptimize + profiling data 42
43. @ryzokuken @ryzokuken 43 Okay, that’s great. But how does it even work? Let’s consider an example.
44. @ryzokuken @ryzokuken 44 function add(x, y) { return x + y; } console.log(add(1, 2));
45. @ryzokuken @ryzokuken 45 Not too intuitive? I’ve got you covered, fam. Let’s see how the Abstract Syntax Tree actually looks.
46. @ryzokuken @ryzokuken 46 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y"
47. @ryzokuken @ryzokuken 47 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC
48. @ryzokuken @ryzokuken 48 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC
49. @ryzokuken @ryzokuken 49 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add”
50. @ryzokuken @ryzokuken 50 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add” PARAMS
51. @ryzokuken @ryzokuken 51 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add” PARAMS VAR “x”
52. @ryzokuken @ryzokuken 52 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add” PARAMS VAR “x” VAR “y”
53. @ryzokuken @ryzokuken 53 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add” PARAMS RETURN VAR “x” VAR “y”
54. @ryzokuken @ryzokuken 54 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add” PARAMS RETURN VAR “x” VAR “y” ADD
55. @ryzokuken @ryzokuken 55 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add” PARAMS RETURN VAR “x” VAR “y” ADD PROXY “x”
56. @ryzokuken @ryzokuken 56 [generating bytecode for function: add] --- AST --- FUNC at 12 . KIND 0 . SUSPEND COUNT 0 . NAME "add" . PARAMS . . VAR (...) (mode = VAR) "x" . . VAR (...) (mode = VAR) "y" . RETURN at 22 . . ADD at 31 . . . VAR PROXY parameter[0] (...) (mode = VAR) "x" . . . VAR PROXY parameter[1] (...) (mode = VAR) "y" FUNC NAME “add” PARAMS RETURN VAR “x” VAR “y” ADD PROXY “x” PROXY “y”
57. @ryzokuken @ryzokuken 57 Scope Resolution FUNC NAME “add” PARAMS RETURN VAR “x” VAR “y” ADD PROXY “y” PROXY “x”
58. @ryzokuken @ryzokuken 58 Let’s step through the bytecode. Hope you like Assembly. Once this is done, we can generate bytecode.
59. @ryzokuken @ryzokuken 59 [generated bytecode for function: add] Parameter count 3 Frame size 0 12 E> 0x38d5f59df42 @ 0 : a5 StackCheck 22 S> 0x38d5f59df43 @ 1 : 25 02 Ldar a1 31 E> 0x38d5f59df45 @ 3 : 34 03 00 Add a0, [0] 35 S> 0x38d5f59df48 @ 6 : a9 Return Constant pool (size = 0) Handler Table (size = 0)
60. @ryzokuken @ryzokuken 60 [generated bytecode for function: add] Parameter count 3 Frame size 0 12 E> 0x38d5f59df42 @ 0 : a5 StackCheck 22 S> 0x38d5f59df43 @ 1 : 25 02 Ldar a1 31 E> 0x38d5f59df45 @ 3 : 34 03 00 Add a0, [0] 35 S> 0x38d5f59df48 @ 6 : a9 Return Constant pool (size = 0) Handler Table (size = 0)
61. @ryzokuken @ryzokuken 61 [generated bytecode for function: add] Parameter count 3 Frame size 0 12 E> 0x38d5f59df42 @ 0 : a5 StackCheck 22 S> 0x38d5f59df43 @ 1 : 25 02 Ldar a1 31 E> 0x38d5f59df45 @ 3 : 34 03 00 Add a0, [0] 35 S> 0x38d5f59df48 @ 6 : a9 Return Constant pool (size = 0) Handler Table (size = 0)
62. @ryzokuken @ryzokuken 62 [generated bytecode for function: add] Parameter count 3 Frame size 0 12 E> 0x38d5f59df42 @ 0 : a5 StackCheck 22 S> 0x38d5f59df43 @ 1 : 25 02 Ldar a1 31 E> 0x38d5f59df45 @ 3 : 34 03 00 Add a0, [0] 35 S> 0x38d5f59df48 @ 6 : a9 Return Constant pool (size = 0) Handler Table (size = 0)
63. @ryzokuken @ryzokuken 63 StackCheck Ldar a1 Add a0, [0] Return
64. @ryzokuken @ryzokuken 64 StackCheck Ldar a1 Add a0, [0] Return
65. @ryzokuken @ryzokuken 65 StackCheck Ldar a1 Add a0, [0] Return
66. @ryzokuken @ryzokuken 66 StackCheck Ldar a1 Add a0, [0] Return
67. @ryzokuken @ryzokuken 67 Two important things happen here. 1. Addition: It’s complicated. 2. Profiling: Feedback Vectors help.
68. @ryzokuken @ryzokuken 68 StackCheck Ldar a1 Add a0, [0] Return
69. @ryzokuken @ryzokuken 69 What about that Speculative Optimization thingie he was talking about earlier? Now that we finally have the baseline code running, let’s put that profiling data to good use. But wait...
70. @ryzokuken @ryzokuken 70 Oh, also, remember that time when I said addition in JavaScript was complicated?
71. @ryzokuken @ryzokuken 71
72. @ryzokuken @ryzokuken 72 Let me break it down for you.
73. @ryzokuken @ryzokuken 73 Number String Object
74. @ryzokuken @ryzokuken 74
75. @ryzokuken @ryzokuken 75 Remember Feedback Vectors? Let’s see how they really look to see how it all works. Awesome! But how do we make that assumption?
76. @ryzokuken @ryzokuken 76 function add(x, y) { return x + y; } console.log(add(1, 2));
77. @ryzokuken @ryzokuken 77 function add(x, y) { return x + y; } console.log(add(1, 2)); %DebugPrint(add);
78. @ryzokuken @ryzokuken 78 ... - feedback vector: 0x18bc7711df89: [FeedbackVector] in OldSpace - map: 0x18bc38c00bc1 <Map> - length: 1 - shared function info: 0x18bc7711dc59 <SharedFunctionInfo add> - optimized code/marker: OptimizationMarker::kNone - invocation count: 1 - profiler ticks: 0 - slot #0 BinaryOp BinaryOp:SignedSmall { [0]: 1 } ...
79. @ryzokuken @ryzokuken 79 function add(x, y) { return x + y; } console.log(add(1, 2)); %DebugPrint(add);
80. @ryzokuken @ryzokuken 80 function add(x, y) { return x + y; } console.log(add(1, 2)); console.log(add(1.1, 2.2)); %DebugPrint(add);
81. @ryzokuken @ryzokuken 81 ... - feedback vector: 0xcbd92d1dfe1: [FeedbackVector] in OldSpace - map: 0x0cbd4f880bc1 <Map> - length: 1 - shared function info: 0x0cbd92d1dc59 <SharedFunctionInfo add> - optimized code/marker: OptimizationMarker::kNone - invocation count: 2 - profiler ticks: 0 - slot #0 BinaryOp BinaryOp:Number { [0]: 7 } ...
82. @ryzokuken @ryzokuken 82
83. @ryzokuken @ryzokuken 83 Question: Is there a method to all this? Answer: Yes, it is called the Feedback Lattice.
84. @ryzokuken @ryzokuken 84 None SignedSmall Number BigInt String NumberOrOddball Any
85. @ryzokuken @ryzokuken 85 None SignedSmall Number BigInt String NumberOrOddball Any
86. @ryzokuken @ryzokuken 86 None SignedSmall Number BigInt String NumberOrOddball Any
87. @ryzokuken @ryzokuken 87 None SignedSmall Number BigInt String NumberOrOddball Any
88. @ryzokuken @ryzokuken 88 None SignedSmall Number BigInt String NumberOrOddball Any
89. @ryzokuken @ryzokuken 89 None SignedSmall Number BigInt String NumberOrOddball Any
90. @ryzokuken @ryzokuken 90 None SignedSmall Number BigInt String NumberOrOddball Any
91. @ryzokuken @ryzokuken 91 None SignedSmall Number BigInt String NumberOrOddball Any
92. @ryzokuken @ryzokuken 92 This data in the Feedback Vectors is used to finally optimize your code once a function is hot. When is a function “hot”? Let’s see what actually happens by explicitly triggering optimization.
93. @ryzokuken @ryzokuken 93 function add(x, y) { return x + y; } add(1, 2);
94. @ryzokuken @ryzokuken 94
95. @ryzokuken @ryzokuken 95 function add(x, y) { return x + y; } add(1, 2); // Warm up with SignedSmall feedback. %OptimizeFunctionOnNextCall(add); add(1, 2); // Optimize and run generated code.
96. @ryzokuken @ryzokuken 96 Question: What did we just do?
97. @ryzokuken @ryzokuken 97 Let’s see the optimized code generated by passing the --print-opt-code flag to d8.
98. @ryzokuken @ryzokuken 98 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18
99. @ryzokuken @ryzokuken 99 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Prologue
100. @ryzokuken @ryzokuken 100 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Check x is Smi
101. @ryzokuken @ryzokuken 101 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Check y is Smi
102. @ryzokuken @ryzokuken 102 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Smi → Word32 (x)
103. @ryzokuken @ryzokuken 103 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Smi → Word32 (y)
104. @ryzokuken @ryzokuken 104 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Add x and y Overflow Check
105. @ryzokuken @ryzokuken 105 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Result to Smi
106. @ryzokuken @ryzokuken 106 leaq rbx,[rip+0xfffffff9] movq rbx,[rcx-0x20] testb [rbx+0xf],0x1 jz 0x198104882dfb <+0x3b> movq r10,0x10efbfde0 (CompileLazyDeoptimizedCode) jmp r10 push rbp movq rbp,rsp push rsi push rdi cmpq rsp,[r13+0x11e8] (root (stack_limit)) jna 0x198104882e55 <+0x95> movq rdx,[rbp+0x18] testb rdx,0x1 jnz 0x198104882e7b <+0xbb> movq rcx,[rbp+0x10] testb rcx,0x1 jnz 0x198104882e87 <+0xc7> movq rdi,rcx shrq rdi, 32 movq r8,rdx shrq r8, 32 addl rdi,r8 jo 0x198104882e93 <+0xd3> shlq rdi, 32 movq rax,rdi movq rsp,rbp pop rbp ret 0x18 Epilogue
107. @ryzokuken @ryzokuken 107 Now, let’s try something fun.
108. @ryzokuken @ryzokuken 108 function add(x, y) { return x + y; } add(1, 2); // Warm up with SignedSmall feedback. %OptimizeFunctionOnNextCall(add); add(1, 2); // Optimize and run generated code.
109. @ryzokuken @ryzokuken 109 function add(x, y) { return x + y; } add(1, 2); // Warm up with SignedSmall feedback. %OptimizeFunctionOnNextCall(add); add(1, 2); // Optimize and run generated code. add(1.1, 2.2) // DEOPTIMIZE!
110. @ryzokuken @ryzokuken 110
111. @ryzokuken @ryzokuken 111 Note: I’m obviously kidding. Deoptimization is no laughing matter.
112. @ryzokuken @ryzokuken 112 Let’s see how the code is actually deoptimized by passing the --trace-deopt flag to d8.
113. @ryzokuken @ryzokuken 113 [deoptimizing (DEOPT eager): begin 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> (opt #0) @0, FP to SP delta: 24, caller sp: 0x7ffee7bee248] ;;; deoptimize at <add.js:2:12>, not a Smi reading input frame add => bytecode_offset=0, args=3, height=1, retval=0(#0); inputs: 0: 0x08d97929dbc1 ; [fp - 16] 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> 1: 0x08d9c6b81521 ; [fp + 32] 0x08d9c6b81521 <JSGlobal Object> 2: 0x08d97929da31 ; rdx 0x08d97929da31 <HeapNumber 1.1> 3: 0x08d97929da41 ; [fp + 16] 0x08d97929da41 <HeapNumber 2.2> 4: 0x08d979281749 ; [fp - 24] 0x08d979281749 <NativeContext[247]> 5: 0x08d92b080e19 ; (literal 2) 0x08d92b080e19 <Odd Oddball: optimized_out> translating interpreted frame add => bytecode_offset=0, height=8 0x7ffee7bee240: [top + 72] <- 0x08d9c6b81521 <JSGlobal Object> ; stack parameter (input #1) 0x7ffee7bee238: [top + 64] <- 0x08d97929da31 <HeapNumber 1.1> ; stack parameter (input #2) 0x7ffee7bee230: [top + 56] <- 0x08d97929da41 <HeapNumber 2.2> ; stack parameter (input #3) ------------------------- 0x7ffee7bee228: [top + 48] <- 0x00010d2881e8 ; caller's pc 0x7ffee7bee220: [top + 40] <- 0x7ffee7bee290 ; caller's fp 0x7ffee7bee218: [top + 32] <- 0x08d979281749 <NativeContext[247]> ; context (input #4) 0x7ffee7bee210: [top + 24] <- 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> ; function (input #0) 0x7ffee7bee208: [top + 16] <- 0x08d97929dca1 <BytecodeArray[7]> ; bytecode array 0x7ffee7bee200: [top + 8] <- 0x003900000000 <Smi 57> ; bytecode offset ------------------------- 0x7ffee7bee1f8: [top + 0] <- 0x08d92b080e19 <Odd Oddball: optimized_out> ; accumulator (input #5) [deoptimizing (eager): end 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> @0 => node=0, pc=0x00010d2886c0, caller sp=0x7ffee7bee248, took 1.163 ms]
114. @ryzokuken @ryzokuken Don’t worry, I got you. 114
115. @ryzokuken @ryzokuken 115 [deoptimizing (DEOPT eager): begin 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> (opt #0) @0, FP to SP delta: 24, caller sp: 0x7ffee7bee248] ;;; deoptimize at <add.js:2:12>, not a Smi reading input frame add => bytecode_offset=0, args=3, height=1, retval=0(#0); inputs: 0: 0x08d97929dbc1 ; [fp - 16] 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> 1: 0x08d9c6b81521 ; [fp + 32] 0x08d9c6b81521 <JSGlobal Object> 2: 0x08d97929da31 ; rdx 0x08d97929da31 <HeapNumber 1.1> 3: 0x08d97929da41 ; [fp + 16] 0x08d97929da41 <HeapNumber 2.2> 4: 0x08d979281749 ; [fp - 24] 0x08d979281749 <NativeContext[247]> 5: 0x08d92b080e19 ; (literal 2) 0x08d92b080e19 <Odd Oddball: optimized_out> translating interpreted frame add => bytecode_offset=0, height=8 0x7ffee7bee240: [top + 72] <- 0x08d9c6b81521 <JSGlobal Object> ; stack parameter (input #1) 0x7ffee7bee238: [top + 64] <- 0x08d97929da31 <HeapNumber 1.1> ; stack parameter (input #2) 0x7ffee7bee230: [top + 56] <- 0x08d97929da41 <HeapNumber 2.2> ; stack parameter (input #3) ------------------------- 0x7ffee7bee228: [top + 48] <- 0x00010d2881e8 ; caller's pc 0x7ffee7bee220: [top + 40] <- 0x7ffee7bee290 ; caller's fp 0x7ffee7bee218: [top + 32] <- 0x08d979281749 <NativeContext[247]> ; context (input #4) 0x7ffee7bee210: [top + 24] <- 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> ; function (input #0) 0x7ffee7bee208: [top + 16] <- 0x08d97929dca1 <BytecodeArray[7]> ; bytecode array 0x7ffee7bee200: [top + 8] <- 0x003900000000 <Smi 57> ; bytecode offset ------------------------- 0x7ffee7bee1f8: [top + 0] <- 0x08d92b080e19 <Odd Oddball: optimized_out> ; accumulator (input #5) [deoptimizing (eager): end 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> @0 => node=0, pc=0x00010d2886c0, caller sp=0x7ffee7bee248, took 1.163 ms]
116. @ryzokuken @ryzokuken 116 [deoptimizing (DEOPT eager): begin 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> (opt #0) @0, FP to SP delta: 24, caller sp: 0x7ffee7bee248] ;;; deoptimize at <add.js:2:12>, not a Smi reading input frame add => bytecode_offset=0, args=3, height=1, retval=0(#0); inputs: 0: 0x08d97929dbc1 ; [fp - 16] 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> 1: 0x08d9c6b81521 ; [fp + 32] 0x08d9c6b81521 <JSGlobal Object> 2: 0x08d97929da31 ; rdx 0x08d97929da31 <HeapNumber 1.1> 3: 0x08d97929da41 ; [fp + 16] 0x08d97929da41 <HeapNumber 2.2> 4: 0x08d979281749 ; [fp - 24] 0x08d979281749 <NativeContext[247]> 5: 0x08d92b080e19 ; (literal 2) 0x08d92b080e19 <Odd Oddball: optimized_out> translating interpreted frame add => bytecode_offset=0, height=8 0x7ffee7bee240: [top + 72] <- 0x08d9c6b81521 <JSGlobal Object> ; stack parameter (input #1) 0x7ffee7bee238: [top + 64] <- 0x08d97929da31 <HeapNumber 1.1> ; stack parameter (input #2) 0x7ffee7bee230: [top + 56] <- 0x08d97929da41 <HeapNumber 2.2> ; stack parameter (input #3) ------------------------- 0x7ffee7bee228: [top + 48] <- 0x00010d2881e8 ; caller's pc 0x7ffee7bee220: [top + 40] <- 0x7ffee7bee290 ; caller's fp 0x7ffee7bee218: [top + 32] <- 0x08d979281749 <NativeContext[247]> ; context (input #4) 0x7ffee7bee210: [top + 24] <- 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> ; function (input #0) 0x7ffee7bee208: [top + 16] <- 0x08d97929dca1 <BytecodeArray[7]> ; bytecode array 0x7ffee7bee200: [top + 8] <- 0x003900000000 <Smi 57> ; bytecode offset ------------------------- 0x7ffee7bee1f8: [top + 0] <- 0x08d92b080e19 <Odd Oddball: optimized_out> ; accumulator (input #5) [deoptimizing (eager): end 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> @0 => node=0, pc=0x00010d2886c0, caller sp=0x7ffee7bee248, took 1.163 ms]
117. @ryzokuken @ryzokuken 117 [deoptimizing (DEOPT eager): begin 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> (opt #0) @0, FP to SP delta: 24, caller sp: 0x7ffee7bee248] ;;; deoptimize at <add.js:2:12>, not a Smi reading input frame add => bytecode_offset=0, args=3, height=1, retval=0(#0); inputs: 0: 0x08d97929dbc1 ; [fp - 16] 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> 1: 0x08d9c6b81521 ; [fp + 32] 0x08d9c6b81521 <JSGlobal Object> 2: 0x08d97929da31 ; rdx 0x08d97929da31 <HeapNumber 1.1> 3: 0x08d97929da41 ; [fp + 16] 0x08d97929da41 <HeapNumber 2.2> 4: 0x08d979281749 ; [fp - 24] 0x08d979281749 <NativeContext[247]> 5: 0x08d92b080e19 ; (literal 2) 0x08d92b080e19 <Odd Oddball: optimized_out> translating interpreted frame add => bytecode_offset=0, height=8 0x7ffee7bee240: [top + 72] <- 0x08d9c6b81521 <JSGlobal Object> ; stack parameter (input #1) 0x7ffee7bee238: [top + 64] <- 0x08d97929da31 <HeapNumber 1.1> ; stack parameter (input #2) 0x7ffee7bee230: [top + 56] <- 0x08d97929da41 <HeapNumber 2.2> ; stack parameter (input #3) ------------------------- 0x7ffee7bee228: [top + 48] <- 0x00010d2881e8 ; caller's pc 0x7ffee7bee220: [top + 40] <- 0x7ffee7bee290 ; caller's fp 0x7ffee7bee218: [top + 32] <- 0x08d979281749 <NativeContext[247]> ; context (input #4) 0x7ffee7bee210: [top + 24] <- 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> ; function (input #0) 0x7ffee7bee208: [top + 16] <- 0x08d97929dca1 <BytecodeArray[7]> ; bytecode array 0x7ffee7bee200: [top + 8] <- 0x003900000000 <Smi 57> ; bytecode offset ------------------------- 0x7ffee7bee1f8: [top + 0] <- 0x08d92b080e19 <Odd Oddball: optimized_out> ; accumulator (input #5) [deoptimizing (eager): end 0x08d97929dbc1 <JSFunction add (sfi = 0x8d97929d999)> @0 => node=0, pc=0x00010d2886c0, caller sp=0x7ffee7bee248, took 1.163 ms]
118. @ryzokuken @ryzokuken 118 Pro Tip: If you don’t want your code to be slow, try to stick to certain types. It makes things easier.
119. @ryzokuken Special Thanks • Benedikt Meurer (@bmeurer) • Yang Guo (@hashseed) • Sathya Gunasekaran (@_gsathya) • Jakob Gruber (@schuay) • Sigurd Schneider (@sigurdschn) • ...and everyone else from the V8 team. • The organizers. 119
120. @ryzokuken @ryzokuken Спасибо! 120