Variable-length integer encodings.
Branch: master
Clone or download
Latest commit 6de14c7 Mar 5, 2018

README.md

Variable-Length Integers

This repository contains code for exploring the performance of different variable-length integer encoding schemes, specifically to determine the best encoding to use in the WebAssembly binary encoding. See the related WebAssembly design issue.

There are three primary interesting performance factors for a variable-length integer encoding:

  1. Compression ratio for the relevant distribution of integers.
  2. Decoding speed.
  3. Encoding speed.

We don't measure encoding speed here since it is less relevant for WebAssembly as long as it is reasonable.

Other desirable properties of a variable-length encoding:

  • Can the length of the whole encoded integer be determined from the first byte? This makes it faster to skip over numbers that are not needed, and it can improve instruction-level parallelism when decoding multiple consequtive numbers.
  • Patchability. Is it possible to reserve space for an unknown number, and write it back later? This requires a non-canonical encoding where a small number can be encoded with a larger length than strictly necessary.

Note that the C++ code in this repository is not production quality, and it can be tricked into executing undefined behavior by maliciously formed input.

LEB128

The Little Endian Base-128 encoding is used in DWARF debug information files and many other places. It is a simple byte-oriented encoding where each byte contains 7 bits of the integer being encoded, and the MSB is used as a tag bit which indicates if there are more bytes coming.

0xxxxxxx                     7 bits in 1 byte
1xxxxxxx 0xxxxxxx           14 bits in 2 bytes
1xxxxxxx 1xxxxxxx 0xxxxxxx  21 bits in 3 bytes
...

PrefixVarint

Brought up in WebAssembly/design#601, and probably invented independently many times, this encoding is very similar to LEB128, but it moves all the tag bits to the LSBs of the first byte:

xxxxxxx1  7 bits in 1 byte
xxxxxx10 14 bits in 2 bytes
xxxxx100 21 bits in 3 bytes
xxxx1000 28 bits in 4 bytes
xxx10000 35 bits in 5 bytes
xx100000 42 bits in 6 bytes
x1000000 49 bits in 7 bytes
10000000 56 bits in 8 bytes
00000000 64 bits in 9 bytes

This has advantages on modern CPUs with fast unaligned loads and count trailing zeros instructions. The compression ratio is the same as for LEB128, except for those 64-bit numbers that require 10 bytes to encode in LEB128.

Like UTF-8, the length of a PrefixVarint-encoded number can be determined from the first byte. (UTF-8 is not considered here since it only encodes 6 bits per byte due to design constraints that are not relevant to WebAssembly.)

leSQLite

The SQLite variable-length integer encoding is biased towards integer distributions with more small numbers. It can encode the integers 0-240 in one byte.

The encoding implemented here is modified for better performance with WebAssembly (little-endian SQLite). The first byte, B0 determines the encoding:

0-184   1 byte    value = B0
185-248 2 bytes   value = 185 + 256 * (B0 - 185) + B1
249-255 3-9 bytes value = (B0 - 249 + 2) little-endian bytes following B0.

This encoding packs more than 7 bits into 1 byte and a bit more than 14 bits into 2 bytes. This has a cost in encoding size since the 3-byte encoding only holds 16 bits. The 3+ byte encoded numbers are very fast to decode with an unaligned load instruction.

leSQLite2

A second variation of the SQLite-inspired encoding has a smoother bump between the 2-byte and 3-byte encodings. It divides the values of the first byte into 4 ranges:

  1. The value of a 1-byte encoding.
  2. The high 6 bits of a 2-byte encoding.
  3. The high 3 bits of a 3-byte encoding.
  4. The number of bytes in a 4-9 byte encoding.

The ranges are assigned like this:

B0 Values Formula
0-177 177 B0
178-241 2^14 178 + ((B0-178) << 8) + B[1]
242-249 2^19 16562 + ((B0-242) << 16) + B[1..2]
250-255 2^24..2^64 B0 - 250 + 3 little-endian bytes.

This variant is a bit slower to decode than the first one because there are more cases.