ext-encoding

This extension provides high-performance raw data encoding & decoding utilities for PHP.

It was designed to supersede pocketmine/binaryutils and the painfully slow PHP functions pack() and unpack().

Real-world performance tests

• pocketmine/nbt was tested with release 0.2.1, and showed 1.5x read and 2x write performance with some basic synthetic tests.

API

A recent IDE stub can usually be found in our custom stubs repository.

Warning

The API design is not yet finalized. Everything is still subject to change prior to the 1.0.0 release.

Note

Although ext-encoding was built as a replacement to pocketmine/binaryutils, it is not a drop-in replacement. Its API is completely different and incompatible.

The new API has been designed with the lessons learned from pocketmine/binaryutils in mind. Most notably:

• Readers and writers have fully separated APIs - no more accidentally writing while intending to read or vice versa
• Endian-reversible types are implemented in LE:: and BE:: static methods, which avoids accidentally using the wrong byte order
• All integer-accepting and returning functions explicitly state whether they work with Signed or Unsigned integers

FAQs

Why are BinaryStream and generally pocketmine/binaryutils so slow?

• VarInt encode/decode
• pack() and unpack()
• Linear buffer reallocations
• Array-of-type

VarInt encode/decode

VarInts are heavily used by the Bedrock protocol. This format is borrowed from protobuf.

There's no fast way to implement them in pure PHP. They require repeated calls to chr() and ord() in a loop, as well as needing workarounds for PHP's lack of logical rightshift.

Compared to BinaryStream, this extension's VarInt:: functions offer a performance improvement of 5-10x (depending on the size of the value and other conditions, YMMV) with both signed and unsigned varints.

This will significantly improve performance in PocketMine-MP when integrated. For example, chunk encoding will become significantly faster, and encoding & decoding of almost all packets will benefit too.

pack() and unpack()

PHP's pack() and unpack() functions are abysmally slow. Parsing the formatting code argument takes over 90% of the time spent in these functions. This overhead can be easily avoided when the types of data used are known in advance.

This extension implements specialized functions for writing big and little endian byte/short/int/long/float/double. Depending on the type and other factors, these functions typically show a 3-4x performance improvement compared to BinaryStream.

Linear buffer reallocations

BinaryStream and similar PHP-land byte-buffer implementations often use strings and use the .= concatenation operator. This is problematic, because the entire string will be reallocated every time something is appended to it. While this isn't a big issue for small buffers, the performance of writing to large buffers progressively degrades.

ByteBufferWriter uses exponential scaling (factor of 2) to minimize buffer reallocations at the cost of potentially wasting some memory. This means that the internal buffer size is doubled when the buffer runs out of space.

Array-of-type

All the above problems contribute to this one, in addition to:

• Extra function call overhead
• Dealing with PHP HashTable structures is generally slow (a problem not solved by this extension currently)

VarInt::readSignedIntArray(), for example, was found to be over 50 times faster in simple tests than a loop calling BinaryStream::getVarInt() when dealing with an array of 10k elements. The most obvious cases where this will benefit PocketMine-MP are in LevelChunkPacket encoding, and plugins using ClientboundMapItemDataPacket could also benefit from it.

In the future it'll probably make sense to add PHP wrappers for native array-of-type (e.g. IntArray, LongArray etc) so that we can avoid the performance and memory usage penalties of dealing with large primitive arrays at runtime.

Why are there SO MANY functions? Why not just accept something like bool $signed, ByteOrder $order parameters?

Runtime parameters would mean that these hot encoding paths would need to branch to decide how to encode everything. Branching is slow, so we want to avoid that.

Internally, we only have a handful of functions (defined in Serializers.h), which use C++ templates to inject type, signedness, and byte order arguments. The compiler expands these templates into optimised branchless native functions for each (type, signed, byte order) combination.

In addition, parsing arguments in PHP is slow, and since PHP doesn't have anything akin to C++ templates (or generics more generally), the only option to get compile-time knowledge of byte order and signedness is to bake them into the function name. There is a function for every combination of (type, signed, byte order).

The downside of this is that we can't use .stub.php files to generate arginfo, so the IDE stubs have to be generated from the extension using extension-stub-generator. Also, you'll probably need eye bleach after seeing the macros that generate the function matrix.

However, considering how critical binary data handling is to performance in PocketMine-MP, this is a trade absolutely worth making.

Why static methods instead of ByteBuffer(Reader|Writer) instance methods?

Two reasons:

• As described above, the static read/write methods can't be generated using .stub.php files. If we put the generated functions in ByteBufferReader/ByteBufferWriter, we'd be unable to use a .stub.php file to define the rest of its non-generated API.
• I've made too many mistakes with byte order due to IDE auto complete. With this API design, byte order is decided by the very first character you type, so auto complete can't trip you up (and you have to import BE or LE).

Why fully specify Signed or Unsigned in every function name? Why not just have e.g. readInt() and readUint()?

This library's first users will be people moving from BinaryStream, where the API is infamous for being inconsistent about signedness when not specified (pmmp/BinaryUtils#15). For example, getShort() is unsigned, and getInt() is signed.

I felt that it was better to be verbose to force developers to think about whether to use a signed or an unsigned type when migrating old code.