Skip to content


Repository files navigation

Roadroller: Flattens Your JavaScript Demo

Roadroller is a heavyweight JavaScript packer for large demos. It was originally designed for js13kGames, but it remains usable for demos as small as 4KB. Depending on the input it can provide up to 15% additional compression compared to best ZIP/gzip recompressors. Try it online!

Roadroller is considered "heavyweight" unlike typical JS packers such as JSCrush or RegPack, because it is quite resource intensive and requires both a considerable amount of memory and a non-negligible run time. The default should work for most devices, but you can configure both aspects as you need.

Quick Start

In addition to the online demo, Roadroller is available as an NPM package:

$ npx roadroller input.js -o output.js

You can also use Roadroller as a library to integrate with your build pipeline.

import { Packer } from 'roadroller';

const inputs = [
        data: 'console.log("Hello, world!");',
        type: 'js',
        action: 'eval',

const options = {
    // see the Usage for available options.

const packer = new Packer(inputs, options);
await packer.optimize(); // takes less than 10 seconds by default

const { firstLine, secondLine } = packer.makeDecoder();
console.log(firstLine + secondLine);

Roadroller as a library or a CLI command requires Node.js 14 or later. Node.js 16 is strongly recommended because Roadroller is substantially faster in 16 than in 14.


By default Roadroller receives your JS code and returns a compressed JS code that should be further compressed with ZIP, gzip or PNG bootstrap (or more accurately, DEFLATE). Ideally your JS code should be already minified, probably using Terser or Closure Compiler; Roadroller only does a minimal whitespace and comment suppression.

The resulting code will look like this: (the newline is mostly for the explanation and can be removed)

,...']charCodeAtUinyxp',"for(;e<12345;c[e++]=p-128)/* omitted */;return o")([],[],12345678,/* omitted */))

The first line is a compressed data. It can contain control characters like (U+001C) that might not render in certain environments. Nevertheless you should make sure that they are all copied in verbatim.

The second line is a compressor tuned for this particular input. By default the decompressed data immediately goes through eval, but you can configure what to do with that.

The first line is very incompressible unlike the second line, so ideally you should compress two lines separately. This is best done by using ADVZIP from AdvanceCOMP or ECT. The first line and second line may form a single statement as above so they should not be separated; you can only put whitespace between them.

Input Configuration

Each input can be further configured by input type and action. In the CLI you put corresponding options before the file path.

Input type (CLI -t|--type TYPE, API type in the input object) determines the preprocessing step to improve the compression.

  • JavaScript (js) assumes a valid JS code. Automatically removes all redundant whitespace and comments and enables a separate modelling for embedded strings. This also works for JSON.
  • Text (text) assumes a human-readable Unicode text that can be encoded in UTF-8. This can also be used for JavaScript code that should not undergo preprocessing.

Input action (CLI -a|--action ACTION, API action in the input object) determines what to do with the decompressed data.

  • Evaluate (eval) evaluates the decompressed JavaScript code. If there are multiple inputs there should be exactly one JavaScript input with evaluate action, since subsequent inputs will be decompressed in that code. The resulting value is always a code string, which may include decoders for subsequent inputs.
  • Write to document (write) writes a decompressed string to document. Typically used with HTML.

Output Configuration

Number of contexts (CLI -S|--selectors xCOUNT) relates to the complexity of modelling. The larger number of contexts will compress better, but at the expense of linear increase in both the time and memory usage. The default is 12, which targets at most 1 second of latency permitted for typical 30 KB input.

Maximum memory usage (CLI -M|--max-memory MEGABYTES, API maxMemoryMB in the options object) configures the maximum memory to be used for decompression. Increasing or decreasing memory usage mostly affects the compression ratio and not the run time. The default is 150 MB and a larger value is not recommended for various reasons:

  • Any further gain for larger memory use is negligible for typical inputs less than 100 KB.

  • The compression may use more memory than the decompression: an one-shot compression may use up to 50% more memory, the optimizer will use 50% more on top of that.

  • It does take time to allocate and initialize a larger memory (~500 ms for 1 GB), so it is not a good choice for small inputs.

The actual memory usage can be as low as a half of the specified due to the internal architecture; -v will print the actual memory usage to stderr.

Allowing the decoder to pollute the global scope (CLI -D|--dirty, API allowFreeVars in the options object) is unsafe especially when the Roadroller output should coexist with other code or there are elements with single letter id attributes and turned off by default. But if you can control your environment (typical for demos), you can turn this on for a smaller decoder.

Optimize parameters (CLI -O|--optimize LEVEL, API Packer.optimize) searches for better modelling parameters. If parameters are already given the optimizer will try to improve upon that, and the optimizer prints best parameters at the end which can be reused for faster iteration. Parameters are solely related to the compression ratio so you can try this as many as you can afford. Each level does the following:

  • Level 0 does absolutely nothing and uses given parameters or default parameters if none. This is the default when any optimizable parameters are given.

  • Level 1 runs a quick search with about 30 sets of parameters and takes less than 10 seconds for typical 30 KB input. This is the default when no optimizable parameters are given, and intended for the typical build process.

  • Level 2 runs a thorough search with about 300 sets of parameters and takes about a minute or two. This is best useful for the release build and you would like to save best parameters for later uses.

  • Level ∞ is a special option only available in the CLI (-OO, with two capital Ohs) and runs increasingly slower optimizations in a run. Once the highest level is reached it runs that level forever. You need to explicitly terminate the search (e.g. CTRL-C), then it will proceed with the best parameters so far.

Advanced Configuration

Number of context bits (CLI -Zco|--context-bits BITS, API contextBits in the options object) sets the size of individual model as opposed to the total memory use (-M), which is a product of the number of context and the size of each model. This explicit option is most useful for the fair benchmarking, since some parameters like -Zpr or -Zmc affect the memory use and therefore this parameter.

Following parameters can be automatically optimized and normally you don't have to touch them unless you want to reproduce a particular set of parameters. As such, the default optimization (-O1) is disabled if any of these arguments are given in the CLI.

Chosen contexts (CLI -S|--selectors SELECTOR,SELECTOR,..., API sparseSelectors in the options object) determine which byte contexts are used for each model. Kth bit of the number (where K > 0) is set if the context contains the Kth-to-last byte: 5 = 101(2) for example would correspond to the context of the last byte and third-to-last byte, also called a sparse context (0,2). There is no particular limit for the number, but Roadroller only considers up to 9th order for the optimization process.

Precision (CLI -Zpr|--precision BITS, API precision in the options object) is the number of fractional bits used in the internal fixed point representation. This is shared between the entropy coder and context models and can't be decoupled. The default of 16 should be enough, you can also try to decrease it.

Learning rate (CLI -Zlr|--learning-rate RATE, API recipLearningRate in the options object) adjusts how fast would the context mixer adapt, where smaller is faster. The default is 500 which should be fine for long enough inputs. If your demo is smaller than 10 KB you can also try smaller numbers.

Model max count (CLI -Zmc|--model-max-count COUNT, API modelMaxCount in the options object) adjusts how fast would individual contexts adapt, where smaller is faster. The model adapts fastest when a particular context is first seen, but that process becomes slower as the context is seen multiple times. This parameter limits how slowest the adaptation process can be. The default of 5 is specifically tuned for JS code inputs.

Model base divisor (CLI -Zmd|--model-base-divisor DIVISOR, API modelRecipBaseCount in the options object) adjusts how fast should individual contexts adapt initially, where larger is faster. The optimal value typically ranges from 10 to 100 for JS code inputs.

Dynamic model flags (CLI -Zdy|--dynamic-models FLAGS, API dynamicModels in the options object) are used to enable or disable specific dynamic models, where each bit is turned on if the model is in use. There is currently one supported model:

  • The bit 0 (value 1) models quoted strings (', " or `) and works well for source codes. It assumes that every quotes are paired, so it can't be used in English texts with contractions (e.g. isn't) and turned off by default in non-JS inputs.

Number of abbreviations (CLI -Zab|--num-abbreviations NUM, API numAbbreviations in the options object) affects the preprocessing for JS code inputs. Common identifiers and reserved words can be abbreviated to single otherwise unused bytes during the preprocessing; this lessens the burden of context modelling which can only look at the limited number of past bytes. If this parameter is less than the number of allowable abbreviations some identifiers will be left as is, which can sometimes improve the compression.

Tips and Tricks

  • The current algorithm slightly prefers 7-bit and 8-bit inputs for the decoder simplicity. You can still use emojis and other tricks that stuff many bits into Unicode code points, but the compression ratio might be decreased. Keep in mind that Roadroller is already doing the hard work for you and you might not need to repeat that.

  • The compressed JS code doesn't do anything beyond computation and the final action, so you can do anything before or after that. The online demo for example inserts a sort of splash screen as a fallback.

  • Roadroller, while being super effective for many inputs, is not a panacea. Roadroller is weaker at exploiting the duplication at a distance than DEFLATE. Make sure to check ADVZIP or ECT out.

See also the wiki for more information.


Roadroller itself and resulting packed codes are ECMAScript 2015 (ES6) compatible and should run in every modern Web browser and JS implementation. Implementations are assumed to be reasonably fast but otherwise it can run in slower interpreters. MSIE is not directly supported but it works fine (slowly) after simple transpiling.

Roadroller and packed codes extensively use Math.exp and Math.log that are implementation-approximated, so there is a small but real possibility that they behave differently in different implementations. This is known to be a non-issue for browser JS engines as well as V8 and Node.js as they use the same math library (fdlibm) for those functions, but you have been warned.


Roadroller is mostly possible due to the progress in data compression algorithms as recent as 2010s:

  • Bytewise rANS coder, adapted from Fabien Giesen's public domain code.

  • Logistic context mixing, which is a type of neural network specifically designed for the data compression.

  • Sparse context models up to 9th order. Models are tuned for each input with simulated annealing. (You may have noticed that this entire architecture is similar to Crinkler, but Roadroller uses a faster and possibly better parameter search algorithm.)

The minimal JS code for this algorithm was initially adapted from a submission by Hasegawa Sayuri (public domain). The main difference is that Roadroller implements hashed contexts and thus order 3+ context models.


The Roadroller compressor proper is licensed under the MIT license. In addition to this, any decoder code produced by Roadroller, that is, everything in the second line is put in the public domain.