Drop base64-js dependency, and optimise Buffer.write

[chjj]

Most string write functions (with the exception of hexWrite) were allocating arrays before copying them onto the buffer. This resulted in significant performance overhead.

This PR modifies said functions to write directly to the buffer. In addition, it drops the dependency on base64-js and adds support for the base64url encoding.

The numbers are as follows:

$ node perf/write-base64.js
OldBuffer.byteLength(8192, "base64") x 24,009 ops/sec ±0.55% (93 runs sampled)
NewBuffer.byteLength(8192, "base64") x 9,548,372 ops/sec ±0.18% (95 runs sampled)

OldBuffer#write(8192, "base64") x 19,789 ops/sec ±0.49% (95 runs sampled)
NewBuffer#write(8192, "base64") x 22,646 ops/sec ±0.04% (95 runs sampled)

OldBuffer#toString(8192, "base64") x 6,686 ops/sec ±0.39% (87 runs sampled)
NewBuffer#toString(8192, "base64") x 10,735 ops/sec ±0.35% (90 runs sampled)

$ node perf/write-binary.js
OldBuffer#write(8192, "binary") x 13,662 ops/sec ±1.37% (84 runs sampled)
NewBuffer#write(8192, "binary") x 46,397 ops/sec ±0.64% (91 runs sampled)

$ node perf/write-ucs2.js
OldBuffer#write(8192, "ucs2") x 3,537 ops/sec ±0.70% (86 runs sampled)
NewBuffer#write(8192, "ucs2") x 40,422 ops/sec ±1.14% (93 runs sampled)

$ node perf/write-utf8.js
OldBuffer.byteLength(2898, "utf8") x 15,304 ops/sec ±0.87% (85 runs sampled)
NewBuffer.byteLength(2898, "utf8") x 58,197 ops/sec ±0.05% (86 runs sampled)

OldBuffer#write(2898, "utf8") x 13,565 ops/sec ±0.82% (85 runs sampled)
NewBuffer#write(2898, "utf8") x 32,716 ops/sec ±0.25% (93 runs sampled)

There are some massive speedups here, particularly with asciiWrite, ucs2Write, and utf8Write.

I was disappointed to see base64 only has a marginal speedup, however, there is a significant speedup for base64Slice and byteLength('base64'). Likewise, byteLength('utf8') also received a significant speedup.

[chjj]

Added a new commit to further optimize base64Write. This logic is basically how the node.js c++ code works internally. It requires no string preprocessing (str.replace, str.split, str.indexOf, etc) which gives it an extra perf boost.

New Numbers:

OldBuffer#write(8192, "base64") x 19,789 ops/sec ±0.49% (95 runs sampled)
NewBuffer#write(8192, "base64") x 24,542 ops/sec ±0.02% (97 runs sampled)

[dcousens]

@chjj can you break this into two pull requests? The addition of all the new benchmarks (easy merge), and then an isolated pull request for the actual behavioural changes?

[chjj]

Done. Force pushed so this is now the isolated PR.

edit: I can rebase on master and force push again once you merge #352 if need be.

[chjj]

New base64 implementation using atob/btoa pushed.

[chjj]

After thinking about it, all of this atob/btoa inconsistency is really bugging me. I'm considering switching back to the custom base64 impl. Although harder to review, it provides consistent behavior that we know will work in everything and has predictable performance.

[chjj]

@dcousens, I think I will switch this PR back to my custom base64 implementation.

The more I think about it, the more I don't like atob/btoa for the following reasons:

1. It's hard to benchmark due to the node.js issue. The numbers may vary significantly in different browsers as well.
2. It screws up the tests if we want to shim in atob/btoa for node.js.
3. The node.js shim also causes issues because node.js has a history of breaking their base64 parser.
4. With regards to number 3, same could be true of atob in any javascript environment: we don't know whether that environment has some kind of bug/inconsistency with their base64 parser. We also don't know the performance properties of that environment's base64 decoder.
5. With atob, buf.write('base64') is basically guaranteed to be slower on a lot of base64url strings or a string with invalid characters because it has to run base64clean and retry parsing.
6. Using btoa is just straight up slower than our custom implementation in any environment (probably).
7. There are environments you could bundle for that do not have atob/btoa. An example of this is quickjs. If I want to write some javascript code and compile it to a binary that runs everywhere, quickjs combined with cosmopolitan is the answer. By using atob/btoa, we break Buffer in that environment.
8. It goes against the whole spirit of this PR which is to write bytes directly to the buffer during decoding.

Number 7 is one I recently thought of. I really really like the idea of compiling bundles for quickjs (I'm thinking of adding a quickjs target to my bundler to convert the quickjs api to the node.js api, but that's a different story).

The benefits of a custom implementation are as follows:

1. Predictable performance.
2. Predictable behavior.
3. Faster toString.
4. Supports any javascript environment, including things like quickjs and the d8 shell.
5. Less fussing over all of the stuff above.

The downsides:

1. Slower buf.write(str, 'base64') in environments like chrome (but again, chrome would only be significantly faster if the string is well-formed enough such that atob can parse it).
2. Larger code size, harder to review.

[dcousens]

I understand your plight for each of the points, but I don't know if this library should be a reference for this kind of functionality. This library should be using whatever underlying native functions exist, as that is what users expect.

The primary reason this library exists is to support cross-compatibility of the native API's, not to be an improved implementation of their shared functionality.

With atob, buf.write('base64') is basically guaranteed to be slower on a lot of base64url strings or a string with invalid characters because it has to run base64clean and retry parsing.

I don't understand this, we should know that the input string is base64url no?

This pull request is difficult to quickly review in the limited time I have, when you have touched and moved the utf8Write and hexWrite and ucs2Write and... can we reduce that diff?

[chjj]

This library should be using whatever underlying native functions exist, as that is what users expect.

Fair enough.

I don't understand this, we should know that the input string is base64url no?

We can use that as a hint, but the Buffer.from('base64') in node has supported the base64url character set for a long time. It's not always guaranteed that a user would pass in base64 versus base64url.

This pull request is difficult to quickly review in the limited time I have, when you have touched and moved the utf8Write and hexWrite and ucs2Write and... can we reduce that diff?

Yeah, I know I've been bugging you a lot lately. Sorry about that. Thank you for your attentiveness this past week.

I find it's best to look at the diff of each individual commit, but I see your point. I'll try to get it looking better.