Make the code generator write-only to avoid exponential time generation

[loganfsmyth]

There's a bit of refactoring in here to get this all to work, but mostly it's the last two commits that are the meat of it:

1. Factor out all of the code responsible for mutating the output buffer into a discrete set of actions.
2. Re-implement those actions with a temporary queue so that queued changes can be reversed without requiring any reads from the output string itself.

Our old logic constantly referenced the output value, meaning every reference got more expensive as the output for larger, leading to an exponential increase in codegen times.

For example, using the example script from istanbuljs/babel-plugin-istanbul#5 (comment) which generates doubling AST array, generation time changes as shown:

Items: the size of the array being generated
Time: The time in ms to generate the code
Length: The number of characters in the output code

Items: 2 ,  time: 9 ,   length: 239
Items: 4 ,  time: 2 ,   length: 465
Items: 8 ,  time: 6 ,   length: 917
Items: 16 , time: 6 ,   length: 1840
Items: 32 , time: 15 ,  length: 3696
Items: 64 , time: 25 ,  length: 7408
Items: 128 ,    time: 93 ,  length: 14917
Items: 256 ,    time: 380 , length: 30149
Items: 512 ,    time: 1399 ,    length: 60613
Items: 1024 ,   time: 5301 ,    length: 121614
Items: 2048 ,   time: 20676 ,   length: 246542

to

Items: 2 ,  time: 7 ,   length: 239
Items: 4 ,  time: 5 ,   length: 465
Items: 8 ,  time: 5 ,   length: 917
Items: 16 , time: 6 ,   length: 1840
Items: 32 , time: 11 ,  length: 3696
Items: 64 , time: 3 ,   length: 7408
Items: 128 ,    time: 13 ,  length: 14917
Items: 256 ,    time: 18 ,  length: 30149
Items: 512 ,    time: 45 ,  length: 60613
Items: 1024 ,   time: 63 ,  length: 121614
Items: 2048 ,   time: 117 , length: 246542
Items: 4096 ,   time: 266 , length: 496398
Items: 8192 ,   time: 460 , length: 996110
Items: 16384 ,  time: 980 , length: 2014687
Items: 32768 ,  time: 2008 ,    length: 4062687
Items: 65536 ,  time: 3819 ,    length: 8158687
Items: 131072 , time: 7359 ,    length: 16443904

[codecov-io]

Current coverage is 87.79%

Merging #3565 into master will decrease coverage by 0.12%

@@             master      #3565   diff @@
==========================================
  Files           194        194          
  Lines          9640       9592    -48   
  Methods        1101       1099     -2   
  Messages          0          0          
  Branches       2204       2198     -6   
==========================================
- Hits           8475       8421    -54   
- Misses         1165       1171     +6   
  Partials          0          0          

Powered by Codecov. Last updated by c561312...65a677d

[danez]

A small micro optimization: Maybe putting force at the beginning, as the other checks seem more expensive?

[loganfsmyth]

I left it this way because the force param here is only true when printing VariableDeclaration nodes, so on the whole it's probably going to be false the vast majority of the time.

[danez]

👍

[hzoo]

I see endsWith has if (Array.isArray(str)) return str.some((s) => this.endsWith(s)); so this could be an array [" ", "\n"]?

Or we remove that (either way the name seems weird as str)

[loganfsmyth]

I'll ditch the array, we don't need it.

[hzoo]

👍

[jridgewell]

Should this be this._queue[0][0]? Right now, we're checking the last char of the item queued first (since we #unshift), not the item most recently queued.

[loganfsmyth]

Good call, it should be. Luckily, looks like in this case, the only place getLast is used is https://github.com/babel/babel/blob/master/packages/babel-generator/src/printer.js#L110, and the only cases that matter are non-queued tokens, so the queue case pretty much doesn't matter anyway.