Bug: 192001186

Clone this repo:
  1. be1d5d7 Make base64 available to product and vendor am: 38dd59bc03 am: 38e79c7c96 by Matthew Maurer · 7 months ago build-tools-release main master
  2. 38e79c7 Make base64 available to product and vendor am: 38dd59bc03 by Matthew Maurer · 7 months ago android-u-beta-1-gpl
  3. 38dd59b Make base64 available to product and vendor by Matthew Maurer · 7 months ago
  4. fd1c730 Upgrade base64 to 0.21.0 am: bb574e5283 am: 7f650b709f by Jeff Vander Stoep · 8 months ago
  5. 7f650b7 Upgrade base64 to 0.21.0 am: bb574e5283 by Jeff Vander Stoep · 8 months ago


Docs CircleCI codecov unsafe forbidden

Made with CLion. Thanks to JetBrains for supporting open source!

It's base64. What more could anyone want?

This library‘s goals are to be correct and fast. It’s thoroughly tested and widely used. It exposes functionality at multiple levels of abstraction so you can choose the level of convenience vs performance that you want, e.g. decode_engine_slice decodes into an existing &mut [u8] and is pretty fast (2.6GiB/s for a 3 KiB input), whereas decode_engine allocates a new Vec<u8> and returns it, which might be more convenient in some cases, but is slower (although still fast enough for almost any purpose) at 2.1 GiB/s.

See the docs for all the details.


I need to decode base64 with whitespace/null bytes/other random things interspersed in it. What should I do?

Remove non-base64 characters from your input before decoding.

If you have a Vec of base64, retain can be used to strip out whatever you need removed.

If you have a Read (e.g. reading a file or network socket), there are various approaches.

  • Use iter_read together with Read's bytes() to filter out unwanted bytes.
  • Implement Read with a read() impl that delegates to your actual Read, and then drops any bytes you don't want.

I need to line-wrap base64, e.g. for MIME/PEM.

line-wrap does just that.

I want canonical base64 encoding/decoding.

First, don‘t do this. You should no more expect Base64 to be canonical than you should expect compression algorithms to produce canonical output across all usage in the wild (hint: they don’t). However, people are drawn to their own destruction like moths to a flame, so here we are.

There are two opportunities for non-canonical encoding (and thus, detection of the same during decoding): the final bits of the last encoded token in two or three token suffixes, and the = token used to inflate the suffix to a full four tokens.

The trailing bits issue is unavoidable: with 6 bits available in each encoded token, 1 input byte takes 2 tokens, with the second one having some bits unused. Same for two input bytes: 16 bits, but 3 tokens have 18 bits. Unless we decide to stop shipping whole bytes around, we're stuck with those extra bits that a sneaky or buggy encoder might set to 1 instead of 0.

The = pad bytes, on the other hand, are entirely a self-own by the Base64 standard. They do not affect decoding other than to provide an opportunity to say “that padding is incorrect”. Exabytes of storage and transfer have no doubt been wasted on pointless = bytes. Somehow we all seem to be quite comfortable with, say, hex-encoded data just stopping when it's done rather than requiring a confirmation that the author of the encoder could count to four. Anyway, there are two ways to make pad bytes predictable: require canonical padding to the next multiple of four bytes as per the RFC, or, if you control all producers and consumers, save a few bytes by requiring no padding (especially applicable to the url-safe alphabet).

All Engine implementations must at a minimum support treating non-canonical padding of both types as an error, and optionally may allow other behaviors.

Rust version compatibility

The minimum supported Rust version is 1.57.0.


Contributions are very welcome. However, because this library is used widely, and in security-sensitive contexts, all PRs will be carefully scrutinized. Beyond that, this sort of low level library simply needs to be 100% correct. Nobody wants to chase bugs in encoding of any sort.

All this means that it takes me a fair amount of time to review each PR, so it might take quite a while to carve out the free time to give each PR the attention it deserves. I will get to everyone eventually!


Benchmarks are in benches/. Running them requires nightly rust, but rustup makes it easy:

rustup run nightly cargo bench


This crate supports no_std. By default the crate targets std via the std feature. You can deactivate the default-features to target core instead. In that case you lose out on all the functionality revolving around std::io, std::error::Error, and heap allocations. There is an additional alloc feature that you can activate to bring back the support for heap allocations.


On Linux, you can use perf for profiling. Then compile the benchmarks with rustup nightly run cargo bench --no-run.

Run the benchmark binary with perf (shown here filtering to one particular benchmark, which will make the results easier to read). perf is only available to the root user on most systems as it fiddles with event counters in your CPU, so use sudo. We need to run the actual benchmark binary, hence the path into target. You can see the actual full path with rustup run nightly cargo bench -v; it will print out the commands it runs. If you use the exact path that bench outputs, make sure you get the one that's for the benchmarks, not the tests. You may also want to cargo clean so you have only one benchmarks- binary (they tend to accumulate).

sudo perf record target/release/deps/benchmarks-* --bench decode_10mib_reuse

Then analyze the results, again with perf:

sudo perf annotate -l

You'll see a bunch of interleaved rust source and assembly like this. The section with lib.rs:327 is telling us that 4.02% of samples saw the movzbl aka bit shift as the active instruction. However, this percentage is not as exact as it seems due to a phenomenon called skid. Basically, a consequence of how fancy modern CPUs are is that this sort of instruction profiling is inherently inaccurate, especially in branch-heavy code.

 lib.rs:322    0.70 :     10698:       mov    %rdi,%rax
    2.82 :        1069b:       shr    $0x38,%rax
         :                  if morsel == decode_tables::INVALID_VALUE {
         :                      bad_byte_index = input_index;
         :                      break;
         :                  };
         :                  accum = (morsel as u64) << 58;
 lib.rs:327    4.02 :     1069f:       movzbl (%r9,%rax,1),%r15d
         :              // fast loop of 8 bytes at a time
         :              while input_index < length_of_full_chunks {
         :                  let mut accum: u64;
         :                  let input_chunk = BigEndian::read_u64(&input_bytes[input_index..(input_index + 8)]);
         :                  morsel = decode_table[(input_chunk >> 56) as usize];
 lib.rs:322    3.68 :     106a4:       cmp    $0xff,%r15
         :                  if morsel == decode_tables::INVALID_VALUE {
    0.00 :        106ab:       je     1090e <base64::decode_config_buf::hbf68a45fefa299c1+0x46e>


This uses cargo-fuzz. See fuzz/fuzzers for the available fuzzing scripts. To run, use an invocation like these:

cargo +nightly fuzz run roundtrip
cargo +nightly fuzz run roundtrip_no_pad
cargo +nightly fuzz run roundtrip_random_config -- -max_len=10240
cargo +nightly fuzz run decode_random


This project is dual-licensed under MIT and Apache 2.0.