Skip to content

Instantly share code, notes, and snippets.

@aszx87410
Last active December 8, 2023 21:59
  • Star 3 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
Star You must be signed in to star a gist
Save aszx87410/97f694127c9a455c5b3a8a108505a042 to your computer and use it in GitHub Desktop.
let vs var by investigating the bytecode generated by Node.js

It's the reply to the question raised by @getify on his twitter:

here's a variation on the question... will JS engines exhibit much performance difference between these two loops?

for (var i = 0; i < 100000000; i++) {
   // do some stuff, but not closure
}

for (let i = 0; i < 100000000; i++) {
   // do the same stuff (no closure)
}

Maybe every JS engine has their own implementation, but luckily in V8 we can generate the bytecode to see what's going on under the hood.

The code I used for test:

function let_for(){
  for (let i = 0; i < 10; i++) {
    console.log(i)
  }
}

function var_for() {
  for (var i = 0; i < 10; i++) {
    console.log(i)
  }
}

function let_timeout() {
  for (let i = 0; i < 10; i++) {
    setTimeout(function let_timeout_inner() {
      console.log(i)
    })
  }
}

function var_timeout() {
  for (var i = 0; i < 10; i++) {
    setTimeout(function var_timeout_inner() {
      console.log(i)
    })
  }
}

let_for()
var_for()
let_timeout()
var_timeout()

// node --print-bytecode a.js > out.txt

Environment: OSX El Capitan + Node.js v12.2.0

First, we can see the generated bytecode for var_for and let_for:

var_for

[generated bytecode for function: var_for]
Parameter count 1
Register count 3
Frame size 24
   95 E> 0x1e9bbe2f0972 @    0 : a5                StackCheck 
  115 S> 0x1e9bbe2f0973 @    1 : 0b                LdaZero 
         0x1e9bbe2f0974 @    2 : 26 fb             Star r0
  120 S> 0x1e9bbe2f0976 @    4 : 0c 0a             LdaSmi [10]
  120 E> 0x1e9bbe2f0978 @    6 : 69 fb 00          TestLessThan r0, [0]
         0x1e9bbe2f097b @    9 : 99 1c             JumpIfFalse [28] (0x1e9bbe2f0997 @ 37)
  102 E> 0x1e9bbe2f097d @   11 : a5                StackCheck 
  137 S> 0x1e9bbe2f097e @   12 : 13 00 01          LdaGlobal [0], [1]
         0x1e9bbe2f0981 @   15 : 26 f9             Star r2
  145 E> 0x1e9bbe2f0983 @   17 : 28 f9 01 03       LdaNamedProperty r2, [1], [3]
         0x1e9bbe2f0987 @   21 : 26 fa             Star r1
  145 E> 0x1e9bbe2f0989 @   23 : 59 fa f9 fb 05    CallProperty1 r1, r2, r0, [5]
  127 S> 0x1e9bbe2f098e @   28 : 25 fb             Ldar r0
         0x1e9bbe2f0990 @   30 : 4c 07             Inc [7]
         0x1e9bbe2f0992 @   32 : 26 fb             Star r0
         0x1e9bbe2f0994 @   34 : 8a 1e 00          JumpLoop [30], [0] (0x1e9bbe2f0976 @ 4)
         0x1e9bbe2f0997 @   37 : 0d                LdaUndefined 
  156 S> 0x1e9bbe2f0998 @   38 : a9                Return 
Constant pool (size = 2)
Handler Table (size = 0)

let_for

[generated bytecode for function: let_for]
Parameter count 1
Register count 3
Frame size 24
   16 E> 0x1e9bbe2d800a @    0 : a5                StackCheck 
   35 S> 0x1e9bbe2d800b @    1 : 0b                LdaZero 
         0x1e9bbe2d800c @    2 : 26 fb             Star r0
   40 S> 0x1e9bbe2d800e @    4 : 0c 0a             LdaSmi [10]
   40 E> 0x1e9bbe2d8010 @    6 : 69 fb 00          TestLessThan r0, [0]
         0x1e9bbe2d8013 @    9 : 99 1c             JumpIfFalse [28] (0x1e9bbe2d802f @ 37)
   22 E> 0x1e9bbe2d8015 @   11 : a5                StackCheck 
   57 S> 0x1e9bbe2d8016 @   12 : 13 00 01          LdaGlobal [0], [1]
         0x1e9bbe2d8019 @   15 : 26 f9             Star r2
   65 E> 0x1e9bbe2d801b @   17 : 28 f9 01 03       LdaNamedProperty r2, [1], [3]
         0x1e9bbe2d801f @   21 : 26 fa             Star r1
   65 E> 0x1e9bbe2d8021 @   23 : 59 fa f9 fb 05    CallProperty1 r1, r2, r0, [5]
   47 S> 0x1e9bbe2d8026 @   28 : 25 fb             Ldar r0
         0x1e9bbe2d8028 @   30 : 4c 07             Inc [7]
         0x1e9bbe2d802a @   32 : 26 fb             Star r0
         0x1e9bbe2d802c @   34 : 8a 1e 00          JumpLoop [30], [0] (0x1e9bbe2d800e @ 4)
         0x1e9bbe2d802f @   37 : 0d                LdaUndefined 
   76 S> 0x1e9bbe2d8030 @   38 : a9                Return 
Constant pool (size = 2)
Handler Table (size = 0)

It's exactly the same, and there is only one variable which stored in r0. So V8 won't create new variable in each iteration in this case.

How about the setTimeout? The case we really need different variable in each iteration.

We can start from the bytecode for var_timeout_inner and let_timeout_inner

var_timeout_inner

[generated bytecode for function: var_timeout_inner]
Parameter count 1
Register count 3
Frame size 24
  397 E> 0x2725c3d772e2 @    0 : a5                StackCheck 
  408 S> 0x2725c3d772e3 @    1 : 13 00 00          LdaGlobal [0], [0]
         0x2725c3d772e6 @    4 : 26 fa             Star r1
  416 E> 0x2725c3d772e8 @    6 : 28 fa 01 02       LdaNamedProperty r1, [1], [2]
         0x2725c3d772ec @   10 : 26 fb             Star r0
         0x2725c3d772ee @   12 : 1a 04             LdaCurrentContextSlot [4]
         0x2725c3d772f0 @   14 : 26 f9             Star r2
  416 E> 0x2725c3d772f2 @   16 : 59 fb fa f9 04    CallProperty1 r0, r1, r2, [4]
         0x2725c3d772f7 @   21 : 0d                LdaUndefined 
  427 S> 0x2725c3d772f8 @   22 : a9                Return 
Constant pool (size = 2)
Handler Table (size = 0)

let_timeout_inner

[generated bytecode for function: let_timeout_inner]
Parameter count 1
Register count 3
Frame size 24
  258 E> 0x2725c3d76972 @    0 : a5                StackCheck 
  269 S> 0x2725c3d76973 @    1 : 13 00 00          LdaGlobal [0], [0]
         0x2725c3d76976 @    4 : 26 fa             Star r1
  277 E> 0x2725c3d76978 @    6 : 28 fa 01 02       LdaNamedProperty r1, [1], [2]
         0x2725c3d7697c @   10 : 26 fb             Star r0
         0x2725c3d7697e @   12 : 1a 04             LdaCurrentContextSlot [4]
  281 E> 0x2725c3d76980 @   14 : aa 02             ThrowReferenceErrorIfHole [2]
         0x2725c3d76982 @   16 : 26 f9             Star r2
  277 E> 0x2725c3d76984 @   18 : 59 fb fa f9 04    CallProperty1 r0, r1, r2, [4]
         0x2725c3d76989 @   23 : 0d                LdaUndefined 
  288 S> 0x2725c3d7698a @   24 : a9                Return 
Constant pool (size = 3)
Handler Table (size = 0)

The only difference is this line ThrowReferenceErrorIfHole in let_timeout_inner. It's actually the implementation of TDZ, so it doesn't matter.

Finally, here is the main dish: var_timeout and let_timeout

var_timeout

[generated bytecode for function: var_timeout]
Parameter count 1
Register count 3
Frame size 24
         0x2725c3d7316a @    0 : 84 00 01          CreateFunctionContext [0], [1]
         0x2725c3d7316d @    3 : 16 fb             PushContext r0
  318 E> 0x2725c3d7316f @    5 : a5                StackCheck 
  338 S> 0x2725c3d73170 @    6 : 0b                LdaZero 
  338 E> 0x2725c3d73171 @    7 : 1d 04             StaCurrentContextSlot [4]
  343 S> 0x2725c3d73173 @    9 : 1a 04             LdaCurrentContextSlot [4]
         0x2725c3d73175 @   11 : 26 fa             Star r1
         0x2725c3d73177 @   13 : 0c 0a             LdaSmi [10]
  343 E> 0x2725c3d73179 @   15 : 69 fa 00          TestLessThan r1, [0]
         0x2725c3d7317c @   18 : 99 1b             JumpIfFalse [27] (0x2725c3d73197 @ 45)
  325 E> 0x2725c3d7317e @   20 : a5                StackCheck 
  360 S> 0x2725c3d7317f @   21 : 13 01 01          LdaGlobal [1], [1]
         0x2725c3d73182 @   24 : 26 fa             Star r1
         0x2725c3d73184 @   26 : 81 02 03 02       CreateClosure [2], [3], #2
         0x2725c3d73188 @   30 : 26 f9             Star r2
  360 E> 0x2725c3d7318a @   32 : 5d fa f9 04       CallUndefinedReceiver1 r1, r2, [4]
  350 S> 0x2725c3d7318e @   36 : 1a 04             LdaCurrentContextSlot [4]
         0x2725c3d73190 @   38 : 4c 06             Inc [6]
  350 E> 0x2725c3d73192 @   40 : 1d 04             StaCurrentContextSlot [4]
         0x2725c3d73194 @   42 : 8a 21 00          JumpLoop [33], [0] (0x2725c3d73173 @ 9)
         0x2725c3d73197 @   45 : 0d                LdaUndefined 
  434 S> 0x2725c3d73198 @   46 : a9                Return 
Constant pool (size = 3)
Handler Table (size = 0)

It seems V8 will create a FunctionContext and then create a variable and store it in the position [4] of CurrentContextSlot.

let_timeout

[generated bytecode for function: let_timeout]
Parameter count 1
Register count 7
Frame size 56
  179 E> 0x2725c3d70daa @    0 : a5                StackCheck 
  199 S> 0x2725c3d70dab @    1 : 0b                LdaZero 
         0x2725c3d70dac @    2 : 26 f8             Star r3
         0x2725c3d70dae @    4 : 26 fb             Star r0
         0x2725c3d70db0 @    6 : 0c 01             LdaSmi [1]
         0x2725c3d70db2 @    8 : 26 fa             Star r1
  293 E> 0x2725c3d70db4 @   10 : a5                StackCheck 
         0x2725c3d70db5 @   11 : 82 00             CreateBlockContext [0]
         0x2725c3d70db7 @   13 : 16 f7             PushContext r4
         0x2725c3d70db9 @   15 : 0f                LdaTheHole 
         0x2725c3d70dba @   16 : 1d 04             StaCurrentContextSlot [4]
         0x2725c3d70dbc @   18 : 25 fb             Ldar r0
         0x2725c3d70dbe @   20 : 1d 04             StaCurrentContextSlot [4]
         0x2725c3d70dc0 @   22 : 0c 01             LdaSmi [1]
         0x2725c3d70dc2 @   24 : 67 fa 00          TestEqual r1, [0]
         0x2725c3d70dc5 @   27 : 99 07             JumpIfFalse [7] (0x2725c3d70dcc @ 34)
         0x2725c3d70dc7 @   29 : 0b                LdaZero 
         0x2725c3d70dc8 @   30 : 26 fa             Star r1
         0x2725c3d70dca @   32 : 8b 08             Jump [8] (0x2725c3d70dd2 @ 40)
  211 S> 0x2725c3d70dcc @   34 : 1a 04             LdaCurrentContextSlot [4]
         0x2725c3d70dce @   36 : 4c 01             Inc [1]
  211 E> 0x2725c3d70dd0 @   38 : 1d 04             StaCurrentContextSlot [4]
         0x2725c3d70dd2 @   40 : 0c 01             LdaSmi [1]
         0x2725c3d70dd4 @   42 : 26 f9             Star r2
  204 S> 0x2725c3d70dd6 @   44 : 1a 04             LdaCurrentContextSlot [4]
         0x2725c3d70dd8 @   46 : 26 f6             Star r5
         0x2725c3d70dda @   48 : 0c 0a             LdaSmi [10]
  204 E> 0x2725c3d70ddc @   50 : 69 f6 02          TestLessThan r5, [2]
         0x2725c3d70ddf @   53 : 99 04             JumpIfFalse [4] (0x2725c3d70de3 @ 57)
         0x2725c3d70de1 @   55 : 8b 06             Jump [6] (0x2725c3d70de7 @ 61)
         0x2725c3d70de3 @   57 : 17 f7             PopContext r4
         0x2725c3d70de5 @   59 : 8b 33             Jump [51] (0x2725c3d70e18 @ 110)
         0x2725c3d70de7 @   61 : 0c 01             LdaSmi [1]
         0x2725c3d70de9 @   63 : 67 f9 03          TestEqual r2, [3]
         0x2725c3d70dec @   66 : 99 1c             JumpIfFalse [28] (0x2725c3d70e08 @ 94)
  186 E> 0x2725c3d70dee @   68 : a5                StackCheck 
  221 S> 0x2725c3d70def @   69 : 13 01 04          LdaGlobal [1], [4]
         0x2725c3d70df2 @   72 : 26 f6             Star r5
         0x2725c3d70df4 @   74 : 81 02 06 02       CreateClosure [2], [6], #2
         0x2725c3d70df8 @   78 : 26 f5             Star r6
  221 E> 0x2725c3d70dfa @   80 : 5d f6 f5 07       CallUndefinedReceiver1 r5, r6, [7]
         0x2725c3d70dfe @   84 : 0b                LdaZero 
         0x2725c3d70dff @   85 : 26 f9             Star r2
         0x2725c3d70e01 @   87 : 1a 04             LdaCurrentContextSlot [4]
         0x2725c3d70e03 @   89 : 26 fb             Star r0
         0x2725c3d70e05 @   91 : 8a 1e 01          JumpLoop [30], [1] (0x2725c3d70de7 @ 61)
         0x2725c3d70e08 @   94 : 0c 01             LdaSmi [1]
  293 E> 0x2725c3d70e0a @   96 : 67 f9 09          TestEqual r2, [9]
         0x2725c3d70e0d @   99 : 99 06             JumpIfFalse [6] (0x2725c3d70e13 @ 105)
         0x2725c3d70e0f @  101 : 17 f7             PopContext r4
         0x2725c3d70e11 @  103 : 8b 07             Jump [7] (0x2725c3d70e18 @ 110)
         0x2725c3d70e13 @  105 : 17 f7             PopContext r4
         0x2725c3d70e15 @  107 : 8a 61 00          JumpLoop [97], [0] (0x2725c3d70db4 @ 10)
         0x2725c3d70e18 @  110 : 0d                LdaUndefined 
  295 S> 0x2725c3d70e19 @  111 : a9                Return 
Constant pool (size = 3)
Handler Table (size = 0)

Compare to the var_timeout, it's obviously longer, so I tried to translate it into more readable format(and remove some code I think it's not necessary):

r1 = 1

loop:
r4.push(new BlockContext())
CurrentContextSlot = 0
if (r1 === 1) {
  r1 = 0
} else {
  CurrentContextSlot++
}
r2 = 1
r5 = CurrentContextSlot
if (!(r5 < 10)) { // end loop
  PopContext r4
  goto done
}

loop2:
if (r2 === 1) {
  setTimeout()
  r2 = 0
  r0 = CurrentContextSlot
  goto loop2
}

if (r2 === 1) {
  PopContext r4
  goto done
}
PopContext r4
goto loop

done:
return undefined

The code above is not accurate, but you can still see that it creates new BlockContext in each iteration.

Here is my conclusion: V8 treats let and var exactly the same if there is no closure in the for loop. But please noted that it's still guess, I am not the V8 engineer 😂

@paulosince
Copy link

Regarding the difference found for closure, why do you think there’s this difference? And which one would be better? (Your opinion)

@aszx87410
Copy link
Author

Regarding the difference found for closure, why do you think there’s this difference? And which one would be better? (Your opinion)

For closure, let must create a new block context in each iteration, because spec said so. And for the no-closure scenario, we can see this as a part of V8's optimization to remove unnecessary cost. let is preferable in most cases I think.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment