encoding_rs is a high-decode-performance, low-legacy-encode-footprint and high-correctness implementation of the WHATWG Encoding Standard written in Rust. In Firefox 56, encoding_rs replaced uconv as the character encoding library used in Firefox. This wasn’t an addition of a component but an actual replacement: uconv was removed when encoding_rs landed. This writeup covers the motivation and design of encoding_rs, as well as some benchmark results.
Additionally, encoding_rs contains a submodule called
encoding_rs::mem that’s meant for efficient encoding-related operations on UTF-16, UTF-8, and Latin1 in-memory strings—i.e., the kind of strings that are used in Gecko C++ code. This module is discussed separately after describing encoding_rs proper.
The C++ integration of encoding_rs is not covered here and is covered in another write-up instead.
Rust’s borrow checker is used with on-stack structs that get optimized away to enforce an “at most once” property that matches reads and writes to buffer space availability checks in legacy CJK converters. Legacy CJK converters are the most risky area in terms of memory-safety bugs in a C or C++ implementation.
Decode is very fast relative to other libraries with the exception of some single-byte encodings on ARMv7. Particular effort has gone into validating UTF-8 and converting UTF-8 to UTF-16 efficiently. ASCII runs are handled using SIMD when it makes sense. There is tension between making ASCII even faster vs. making transitions between ASCII and non-ASCII more expensive. This tension is the clearest when encoding from UTF-16, but it’s there when decoding, too.
By default, there is no encode-specific data other than 32 bits per single-byte encoding. This makes legacy CJK encode extremely slow by default relative to other libraries but still fast enough in for the browser use cases. That is, the amount of text one could reasonably submit at a time in a form submission encodes so fast even on a Raspberry Pi 3 (standing in for a low-end phone) that the user will not notice. Even with only 32 bits of encode-oriented data, multiple single-byte encoders are competitive with ICU though only the windows-1252 applied to ASCII or almost ASCII input is competitive with Windows system encoders. Faster CJK legacy encode is available as a compile-time option. But ideally, you should only be using UTF-8 for output anyway.
(If you just want to see the benchmarks and don’t have time for the discussion of the API and implementation internals, you can skip to the benchmarking section.)
encoding_rs::mem submodule, which is discussed after encoding_rs proper, encoding_rs implements the character encoding conversions defined in the Encoding Standard as well as the mapping from labels (i.e. strings in protocol text that identify encodings) to encodings.
Specifically, encoding_rs does the following:
u16) into a sequence of bytes in an Encoding Standard-defined character encoding as if the lone surrogates had been replaced with the REPLACEMENT CHARACTER before performing the encode. (Gecko’s UTF-16 is potentially invalid.)
The Web is UTF-8 these days and Rust uses UTF-8 as the in-RAM Unicode representation, so why is a character encoding conversion library even needed anymore? The answer is, of course, “for legacy reasons”.
While the HTML spec requires the use of UTF-8 and the Web is over 90% UTF-8 (according to W3Techs, whose methodology is questionable considering that they report e.g. ISO-8859-1 separately from windows-1252 and GB2312 separately from GBK even though the Web Platform makes no such distinctions, but Google hasn’t published their numbers since 2012), users still need to access the part of the Web that has not migrated to UTF-8 yet. That part does not consist only of ancient static pages, either. For example, in Japan there are still news sites that publish new content every day in Shift_JIS. Over here in Finland, I do my banking using a Web UI that is still encoded in ISO-8859-15.
Regardless of the implementation language, the character encoding library in Gecko was in need of a rewrite for three reasons:
The addition of Rust code in Firefox brought about the need to be able to convert to and from UTF-8 directly and in terms of binary size, it didn’t make sense to have distinct libraries for converting to and from UTF-16 and for converting to and from UTF-8. Instead, a unified library using the same lookup tables for both was needed. The old code wasn’t designed to yield both UTF-16-targeting and UTF-8-targeting machine code from the same source. The addition of an efficient capability to decode to UTF-8 or to encode from UTF-8 would have involved a level of change comparable to a rewrite.
The old library was crufty enough that it was easier to make correctness improvements by the means of a rewrite than by the means of incremental fixes.
In Firefox 43, I had already rewritten the Big5 decoder and encoder in C++, because a rewrite was easier than modifying the old code. In that particular case, the old code used the Private Use Area (PUA) of the Basic Multilingual Plane (BMP) for Hong Kong Supplementary Character Set (HKSCS) characters. However, after the old code was written, HKSCS characters had been assigned proper code points in Unicode, but many of the assignments are on the Supplementary Ideographic Plane (Plane 2). When a fundamental assumption, such as all the characters in an encoding mapping to the BMP, no longer holds, a rewrite is easier than an incremental change.
As another example (that showed up after the initial rewrite proposal but before the implementation got properly going), the ISO-2022-JP decoder had an XSS vulnerability that was difficult to fix without restructuring the existing code. I actually tried to write a patch for the old code and gave up.
In general, the code structure of the old multi-byte decoders differed from the spec text so much that it would have been harder to try to figure out if the code does what the spec requires than to write new code according to the spec.
The old code was written at a time when the exact set of behaviors that Web-exposed character encodings exhibit wasn’t fully understood. For this reason, the old code had generality that is no longer useful now that we know the full set of Web-exposed legacy encodings and can be confident that there will be no additional legacy encodings introduced with additional behaviors anymore.
As the most notable example, the old code assumed that the lower half of single-byte encodings might not be ASCII. By the time of planning encoding_rs, single-byte encodings whose lower half wasn’t ASCII had already been removed as part of previous Encoding Standard-compliance efforts. Some of the multi-byte encoding handling code also had configurability for the single-byte mode that allowed for non-ASCII single-byte mode. However, some multi-byte encodings had already been migrated off the generic two-byte encoding handling code years ago.
There had been generic two-byte encoding handling code, but it no longer made sense when only EUC-KR remained as an encoding exhibiting the generic characteristics. Big5 was able to decode to Plane 2, GBK had grown four-byte sequences as part of the evolution to GB18030, EUC-JP had grown support for three-byte sequences in order to support JIS X 0212 and Shift_JIS never had the EUC structure to begin with and had single-byte half-width katakana. Even EUC-KR itself had deviated from the original EUC structure by being extended to support all precomposed Hangul syllables (not just the ones in common use) in windows-949.
When a rewrite made sense in any case, it made sense to do the rewrite in Rust, because a rewrite of a clearly identifiable subsystem is exactly the kind of thing that is suitable for rewriting in Rust and the problem domain could use memory-safety. The old library was created in early 1999, but it still had a buffer overrun discovered in it in 2016 (in code added in the 2001 and 2002). This shows that the notion that code written in a memory-unsafe language becomes safe by being “battle-hardened” if it has been broadly deployed for an extended period of time is a myth. Memory-safety needs a systematic approach. Calendar time and broad deployment are not sufficient to turn unsafe code into safe code.
(The above-mentioned bug discovered in 2016 wasn’t the last uconv security bug to be fixed. In 2018, a memory-safety-relevant integer overflow bug was discovered in uconv after uconv had already been replaced with encoding_rs in non-ESR Firefox but uconv was still within security support in ESR. However, that bug was in the new Big5 code that I wrote for Firefox 43, so it can’t be held against the ancient uconv code. I had fixed the corresponding encoding_rs bug before encoding_rs landed in Firefox 56. The uconv bug was fixed in Firefox ESR 52.7.)
As noted above, a key requirement was the ability to decode to and from both UTF-16 and UTF-8, but ICU supports only decoding to and from UTF-16 and rust-encoding supports only decoding to and from UTF-8. Perhaps one might argue that pivoting via another UTF would be fast enough, but experience indicated that pivoting via another UTF posed at least a mental barrier: Even after the benefits of UTF-8 as an in-memory Unicode representation were known, Gecko subsystems had been written to use UTF-16 because that was what uconv decoded to.
A further problem with ICU is that it does not treat the Encoding Standard as its conformance target. Chrome patches ICU substantially for conformance. I didn’t want to maintain a similar patch set in the Gecko context and instead wanted a library that treats the Encoding Standard as its conformance target.
The invasiveness of the changes to rust-encoding that would have been needed to meet the API design, performance and UTF-16 targeting goals would have been large enough that it made sense to pursue them in a new project instead of trying to impose the requirements onto an existing project.
In addition to internal problems, uconv also had a couple of API design problems. First, the decoder API lacked the ability to signal the end of the stream. This meant that there was no way for the decoder to generate a REPLACEMENT CHARACTER when the input stream ended with an incomplete byte sequence. It was possible for the caller to determine from the status code if the last buffer passed to the decoder ended with an incomplete byte sequence, but then it was up to the caller to generate the REPLACEMENT CHARACTER in that situation even though the decoder was generally expected to provide this service. As a result, only one caller in the code base, the
TextDecoder implementation, did the right thing. Furthermore, even though the encoder side had an explicit way to signal the end of the stream, it was a separate method leading to more complexity for callers than just being able to say that a buffer is the last buffer.
Additionally, the API contract was unclear on whether it was supposed to fill buffers exactly potentially splitting a surrogate pair across buffer boundaries or whether it was supposed to guarantee output validity on a per-method call basis. In a situation where the input and output buffers were exhausted simultaneously, it was unspecified whether the converter should signal that the input was exhausted or that the output was exhausted. In cases where it wasn’t the responsibility of the converter to handle the replacement of malformed byte sequences when decoding or unmappable characters when encoding, the API left needlessly much responsibility to the caller to advance over the faulty input and to figure out what the faulty input was in the case where that mattered, i.e. when encoding and producing numeric character references for unmappable characters.
Character encoding conversion APIs tend to exhibit common problems, so the above uconv issues didn’t make uconv particularly flawed compared to other character encoding conversion APIs out there. In fact, to uconv’s credit at least in the form that it had evolved into by the time I got involved, given enough output space uconv always consumed all the input provided to it. This is very important from the perspective of API usability. It’s all too common for character encoding conversion APIs to backtrack if the input buffer ends with an incomplete byte sequence and to report the incomplete byte sequence at the end of the input buffer as not consumed. This leaves it to the caller to take those unconsumed bytes and to copy them to the start of the next buffer so that they can be completed by the bytes that follow. Even worse, sometimes this behavior isn’t documented and is up to the caller of the API to discover by experimentation. This behavior also imposes a, typically undocumented, minimum input buffer size, because the input buffer has to be large enough for at least one complete byte sequence to fit. If the input trickles in byte by byte, it’s up to the caller to arrange them into chunks large enough to contain a complete byte sequence.
Sometimes, the API design problem described in the previous paragraph is conditional on requesting error reporting. When I was writing the Validator.nu HTML Parser, I discovered that the
java.nio.charset character encoding conversion API was well-behaved when it was asked to handle errors on its own, but when the caller asked for the errors to be reported, the behavior undocumentedly changed to not consuming all the input offered even if there was enough output space. This was because the error reporting mechanism sought to designate the exact bytes in error by giving the caller the number of erroneous bytes corresponding to a single error. In order to make a single number make sense, the bytes always had to be counted backwards from the current position, which meant that the current position had to be placed such that it was at the end of the erroneous sequence and additionally the API sought to make it so that the entire erroneous sequence was in the buffer provided and not partially in a past already discarded buffer.
Finally, some APIs fail to provide a high-performance streaming mode where the caller is responsible for output buffer allocation. (This means two potential failures: First, failure to provide a streaming mode and, second, providing a streaming mode but converter seeks to control the output buffer allocation.)
In summary, in my experience, common character encoding conversion API design problems are the following:
All but the last item are specific to a streaming mode. Streaming is hard.
There are other API design considerations that would be unfair to label as “problems”, but that are still very relevant to designing a new API. These relate mainly to error handling and byte order mark (BOM) handling.
It is typical for character encoding conversion APIs to treat error handling as a mode that is set on a converter object as opposed to treating error handling as a different API entry point. API-wise it makes sense to have different entry points in order to have different return values for the two cases. Specifically, when the converter handles errors, the status of the conversion call cannot be that conversion stopped on an error for the caller to handle. Additionally, when the converter handles errors, it may make sense to provide a flag that indicates whether there where errors even though they were automatically handled.
Implementation-wise, experience suggests that baking error handling into each converter complicates code considerably and adds opportunities for bugs. Making the converter implementation always signal errors and having an optional wrapper that deals with those errors so that the application developer doesn’t need to leads to a much cleaner design. This design is a natural match for exposing different entry points: one entry point goes directly to the underlying converter and the other goes through the wrapper.
BOM sniffing is subtle enough that it is a bad idea to leave it to the application. It’s more robust to bake it into the conversion library. In particular, getting BOM sniffing right when bytes arrive one at a time is not trivial for applications to handle. Like replacement of errors, different BOM handling modes can be implemented as wrappers around the underlying converters.
Especially in languages that provide a notion of inheritance, interfaces or traits it is alluring for the API designer to seek to define an abstract conversion API that others can write more converters for. However, in the case of the Web, the set of encodings is closed and includes only those that are defined in the Encoding Standard. As far as the use cases in the Web context go, extensibility is not needed. On the contrary, especially in a code base that is also used in a non-Web context like Gecko is used in Thunderbird in the email context, it is a feature that we can be confident on the Web side that if we have a type that represents an encoding defined in the Encoding Standard it can’t exhibit behaviors from outside the Encoding Standard. By design, encoding_rs is not extensible, so an encoding_rs
Encoding does not represent any imaginable character encoding but instead represents a character encoding from the Encoding Standard. For example, we know from the type that we don’t accidentally have a UTF-7 decoder in Gecko code that has Web expectations even though Thunderbird contains a UTF-7 decoder in its codebase. (If you are interested in decoding email in Rust, there is a crate that wraps encoding_rs, adds UTF-7 decoding and maintains a type distinction between Web encodings and email encodings.)
Additionally, in the context of Rust and its Foreign Function Interface (FFI), it helps that references are references to plain structs and not trait objects. Whereas C++ puts a vtable pointer on the objects allowing pointers to polymorphic types to have the same size as C pointers, Rust’s type erasure puts the vtable pointer in the reference. A Rust reference to a struct has the same machine representation as a plain (non-null) C pointer. A Rust reference to a trait-typed thing is actually two pointers: one to the instance and another to the vtable appropriate for the concrete type of the instance. Since interoperability with C++ is a core design goal for encoding_rs, using the kind of types whose references are the same as C pointers avoids the problem of losing the vtable pointer when crossing the FFI boundary.
Conceptually a character encoding is a mapping from a stream of bytes onto a stream of Unicode scalar values and, in most cases, vice versa. Therefore, it would seem that the right abstraction for a converter is an iterator adaptor that consumes an iterator over bytes and yields Unicode scalar values (or vice versa).
There are two problems with modeling character encoding converters as iterator adaptors. First, it leaves optimization to the compiler, when manual optimizations across runs of code units are desirable. Specifically, it is a core goal for encoding_rs to make ASCII handling fast using SIMD, and the compiler does not have enough information about the data to know to produce ASCII-sequence-biased autovectorization. Second, Rust iterators are ill-suited for efficient and (from the C perspective) idiomatic exposure over the FFI.
The API style of unconv,
java.nio.charset, iconv, etc., of providing input and output buffers of several code units at a time to the converter is friendly both to SIMD and to FFI (Rust slices trivially decompose to pointer and length in C). While this isn’t 100% rustic like iterators, slices still aren’t unrustic.
This finally brings us to the actual API. There are three public structs:
Encoder. From the point of view of the application developer, these act like traits (or interfaces or superclasses to use concepts from other languages) even though they are structs. Instead of using language implementation-provided vtables for dynamic dispatch, they internally have an enum that wraps private structs that are conceptually like subclasses. The use of private enum for dispatch avoids vtable pointers in FFI, makes the hierarchy intentionally non-extensible (see above) and allows BOM sniffing to change what encoding a
Decoder is a decoder for.
There is one statically allocated instance of
Encoding for each encoding defined in the Encoding Standard. These instances have publicly visible names that allow application code to statically refer to a specific encoding (commonly, you want to do this with UTF-8, windows-1252, and the replacement encoding). To find an
Encoding instance dynamically at runtime based on a label obtained from protocol text, there is a static method
fn Encoding::for_label(label: &[u8]) -> &'static Encoding.
Encoding struct provides convenience methods for non-streaming conversions. These are “convenience” methods in the sense that they are implemented on top of
Encoder. An application that only uses non-streaming conversions only needs to deal with
Encoding and doesn’t need to use
Encoder at all.
Encoder provide streaming conversions and are allocated at runtime, because they encapsulate state related to the streaming conversion. On the
Encoder side, only ISO-2022-JP is actually stateful, so most of the discussion here will focus on
Internally, the encoding-specific structs wrapped by
Decoder are macroized to generate decode to UTF-8 and decode to UTF-16 from the same source code (likewise for
Encoder). Even though Rust applications are expected to use the UTF-8 case, I’m going to give examples using the UTF-16 case, because it doesn’t involve the distinction between
&[u8] which would distract from the more important issues.
The fundamental function that
Decoder provides is:
dst: &mut [u16],
) -> (DecoderResult, usize, usize)
This function wraps BOM sniffing around an underlying encoding-specific implementation that takes the same arguments and has the same return value. The
Decoder-provided wrapper first exposes the input to a BOM sniffing state machine and once the state machine gets out of the way delegates to the underlying implementation.
Decoder instances can’t be constructed by the application directly. Instead, they need to be obtained from factory functions on
Encoding. The factory functions come in three flavors for three different BOM sniffing modes: full BOM sniffing (the default), which may cause the
Decoder to morph into a decoder for a different encoding than initially (using enum for dispatch shows its usefulness here!), BOM removal (no morphing but the BOM for the encoding itself is skipped) and without BOM handling. The struct is the same in all cases, but the different factory methods initialize the state of the BOM sniffing state machine differently.
The method takes an input buffer (
src) and an output
dst) both of which are caller-allocated. The method then decodes bytes from
src into Unicode scalar values that are stored (as UTF-16) into
dst until one of the following three things happens:
The return value is a tuple of a status indicating which one
of the three reasons to return happened, how many input bytes were read and
how many output code units were written. The status is a
DecoderResult enumeration (possibilities
OutputFull corresponding to the three cases listed above).
The output written into
dst is guaranteed to be valid UTF-16,
and the output after each call is guaranteed to consist of
complete characters. (I.e. the code unit sequence for the last character is
guaranteed not to be split across output buffers.) This implies that the output buffer must be long enough for an astral character to fit (two UTF-16 code units) and the output buffer might not be fully filled. While it may seem wasteful not to fill the last slot of the output buffer in the common case, this design significantly simplifies the implementation while also simplifying callers by guaranteeing to the caller that it won’t have to deal with split surrogate pairs.
The boolean argument
last indicates that the end of the stream is reached
when all the bytes in
src have been consumed.
Decoder object can be used to incrementally decode a byte stream. During the processing of a single stream, the caller must call the method
zero or more times with
last set to
false and then call
least once with
last set to
true. If the
last set to
the processing of the stream has ended. Otherwise, the caller must call
decode_* again with
last set to
true (or treat a
Malformed result as
a fatal error).
Once the stream has ended, the
Decoder object must not be used anymore.
That is, you need to create another one to process another stream. Unlike with some other libraries that encourage callers to recycle converters that are expensive to create, encoding_rs guarantees that converters are extremely cheap to create. (More on this later.)
When the decoder returns
OutputFull or the decoder returns
the caller does not wish to treat it as a fatal error, the input buffer
src may not have been completely consumed. In that case, the caller must
pass the unconsumed contents of
src to the method again upon the next
Typically the application doesn’t wish to do its own error handling and just wants errors to be replaced with the REPLACEMENT CHARACTER. For this use case, there is another method that wraps the previous method and provides the replacement. The wrapper looks like this:
dst: &mut [u16],
) -> (CoderResult, usize, usize, bool)
Notably, the status enum is different, because the case of malformed sequences doesn’t need to be communicated to the application. Also, the return tuple includes a boolean flag to indicate whether there where errors.
Additionally, there is a method for querying the worst case output size even the current state of the decoder and the length of an input buffer. If the length of the output buffer is at least the worst case, the decoder guarantees that it won’t return
Initially, the plan was simply not to support applications that need to identify which input bytes were in error, because I thought that it wasn’t possible to do so without complicating the API for everyone else. However, very early into the implementation phase, I realized that it is possible to identify which bytes are in error without burdening applications that don’t care if the applications that want to know are responsible for remembering the last N bytes decoded where N is relatively small. It turns out that N is 6.
For a malformed sequence that corresponds to a single decode error (i.e. a single REPLACEMENT CHARACTER) a
DecoderResult::Malformed(u8, u8) is returned. The first wrapped integer indicates the length of the malformed byte sequence. The second wrapped integer indicates the number of bytes that were consumed after the malformed sequence. If the second integer is zero, the last byte that was consumed is the last byte of the malformed sequence. The malformed bytes may have been part of an earlier input buffer, which is why it is the responsibility of the application that wants to identify the bytes that were in error.
The first wrapped integer can have values 1, 2, 3 or 4. The second wrapped integer can have values 0, 1, 2 or 3. The worst-case sum of the two is 6, which happens with ISO-2022-JP.
When encoding to an encoding other than UTF-8 (the Encoding Standard does not support encoding into UTF-16LE or UTF-16BE, and there is one Unicode scalar value that cannot be encoded into gb18030), it is possible that the encoding cannot represent a character that is being encoded. In this case, instead of returning backward-looking indices
EncoderResult::Unmappable(char) wraps the Unicode scalar value that needs to be replaced with a numeric character reference when performing replacement. In the case of ISO-2022-JP, this Unicode scalar value can be the REPLACEMENT CHARACTER instead of a value actually occurring in the input if the input contains U+000E, U+000F, or U+001B.
This asymmetry between how errors are signaled in the decoder and encoder scenarios makes the signaling appropriate for each scenario instead of optimizing for consistency where consistency isn’t needed.
As noted earlier,
Encoding provides non-streaming convenience methods built on top of the streaming functionality. Instead of being simply wrappers for the streaming conversion, the non-streaming methods first try to check if the input is borrowable as output without conversion. For example, if the input is all ASCII and the encoding is ASCII-compatible, a
Cow borrowing the input is returned. Likewise, the input is borrowed when the encoding is UTF-8 and the input is valid or when the encoding is ISO-2022-JP and the input contains no escape sequences. Here’s an example of a non-streaming conversion method:
bytes: &'a [u8]
) -> (Cow<'a, str>, bool)
Cow is a Rust standard library type that wraps either an owned type or a corresponding borrowed type, so a heap allocation and copy can be avoided if the caller only needs a borrow. E.g.,
Cow<'a, str> wraps either a heap-allocated string or a pointer and a length designating a string view into memory owned by someone else. Lifetime
'a indicates that the lifetime of borrowed output depends on the lifetime of the input.)
Internally, there are five guiding design principles.
Even though in principle compile-time abstraction over UTF-8 and UTF-16 is a matter of monomorphizing over
u16, handling the two cases using generics would be more complicated than handling them using macros. That’s why it’s handled using macros. The conversion algorithms are written as blocks of code that are inputs to macros that expand to provide the skeleton conversion loop and fill in the encoding-specific blocks of code. In the skeleton in the decode case, one instantiation uses a
Utf8Destination struct and another uses a
Utf16Destination struct both of which provide the same API for writing into them. In the encode case, the source struct varies similarly.
The old code in uconv was relatively ad hoc in how it accessed the input and output buffers. It maybe did stuff, advanced some pointers, checked if the pointers reached the end of the buffer and maybe even backed off a bit in some places. It didn’t have an overarching pattern to how space availability was checked and matched to memory accesses so that no accesses could happen without a space check having happened first. For encoding_rs, I wanted to make sure that buffer access only goes forwards without backtracking more than the one byte that might get unread in error cases and that no read happens without checking that there is still data to be read and no write happens without checking that there is space in the output buffer.
Rust’s lifetimes can be used to enforce an “at most once” property. Immediately upon entering a conversion function, the input and output slices are wrapped in source and destination structs that maintain the current read or write position. I’ll use the write case as the example, but the read case works analogously. A decoder that only ever produces characters in the basic multilingual plane uses a BMP space checking method on the destination that takes the destination as a mutable reference (
&mut self). If the destination is a UTF-8 destination, the method checks that there is space for at least three additional bytes. If the destination is a UTF-16 destination, the method checks that there is space for at least one additional code unit. If there is enough space, the caller receives a BMP handle whose lifetime is tied to the the lifetime of the destination due to the handle containing the mutable reference to the destination. A mutable reference in Rust means exclusive access. Since a mutable reference to the destination is hidden inside the handle, no other method can be called on the destination until the handle goes out of scope. The handle provides a method for writing one BMP scalar value. That method takes the handle’s
self by value consuming the handle and preventing reuse.
The general concept is that at the top of the loop, the conversion loop checks availability of data at the source and obtains a read handle or returns from the conversion function with
InputEmpty and then checks availability of space at the destination and obtains a write handle or returns from the conversion function with
OutputFull. If neither check caused a return out of the conversion function, the conversion loop now hasn’t read or written either buffer but can be fully confident that it can successfully read from the input at most once and write a predetermined amount of units to the output at most once during the loop body. The handles go out of scope at the end of the loop body, and once the loop starts again, it’s time to check for input availability and out space availability again.
As an added twist, the read operation yields not only a byte of input but also an unread handle for unreading it, because in various error cases the spec calls for prepending input that was already read back to the input stream. In practice, all the cases in the spec can be handled by being able to unread at most one unit of input even though the spec text occasionally prepends more than one unit.
In practice, the ISO-2022-JP converters, which don’t need to be fast for Web use cases, use the above concept in its general form. For the ASCII-compatible encodings that are actually performance-relevant for Web use cases, there are a couple of elaborations.
First, the UTF-8 destination and the UTF-16 destination know how to copy ASCII from a byte source in an efficient way that handles more than one ASCII character per register (either a SIMD register or even an ALU register). So the main conversion loop starts with a call to a method that first tries to copy ASCII from the source to the destination and then returns a non-ASCII byte and write handle if there’s space left the destination. Once a non-ASCII byte is found, another loop is entered into that actually works with the handles.
Second, the loop that works with the handles doesn’t have a single scope per loop body for multi-byte encodings. Once we’re done copying ASCII, the non-ASCII byte that we found is always a lead byte of a multi-byte sequence unless there is an error—and we are optimizing for the case where there is neither an error nor a buffer boundary. Therefore, it makes sense to start another scope that does the handle obtaining space check choreography again in the hope that the next byte will be a valid trail byte given the lead byte that we just saw. Then there is a third innermost loop for reading the next byte after that so that if non-ASCII we can continue the middle loop as if this non-ASCII byte had come from the end of the initial ASCII fast path and if the byte is ASCII punctuation, we can spin in the innermost loop without trying to handle a longer ASCII run using SIMD, which would likely fail within CJK plain text. However, if we see non-punctuation ASCII, we can continue the outermost loop and go back to the ASCII fast path.
Not matching on a state variable indicating whether we’re expecting a lead or trail byte on a per-byte basis and instead using the program counter for the state distinguishing between lead and trail byte expectations is good for performance. However, it poses a new problem: What if the input buffer ends in the middle of a multi-byte sequence? Since we are using the program counter for state, the code for handling the trail byte in a two-byte encoding is only reachable by first executing the code for handling the lead byte, and since Rust doesn’t have
goto or a way to store continuations, after a buffer boundary we can’t just restore the local variables and jump directly to the trail byte handling. To deal with this, the macro structure that allows the reuse of code for decoding both to UTF-8 and to UTF-16 also duplicates the block for handling the trail byte such that the same block occurs between the function method entry and the conversion loop. If the previous buffer ended in the middle of a byte sequence, the next call to the conversion function handles the trail of that sequence before entering the actual conversion loop.
The UTF-8 decoder does not use the same structure as the other multi-byte decoders. Dealing with invalid byte sequences in the middle of the buffer or valid byte sequences that cross a buffer boundary is implemented naïvely from the spec in a way that is instantiated via macro from the same code both when converting to UTF-8 and when converting to UTF-16. However, once that outer tier of conversion gets to a state where it expects the next UTF-8 byte sequence, it calls into fast-track code that only deals with valid UTF-8 and returns back to the outer tier that’s capable of dealing with invalid UTF-8 or partial sequences when it discovers an incomplete sequence at the end of the buffer or an invalid sequence in the middle. This inner fast track is implemented separately for decoding UTF-8 to UTF-8 and for decoding UTF-8 to UTF-16.
The UTF-8 to UTF-16 case is close enough to one might expect from the above description of legacy multibyte encodings. At the top of the loop, there is the call to the ASCII fast path that zero-extends ASCII to UTF-16 Basic Latin multiple code units at a time and then byte sequences that start with a non-ASCII lead byte are handles as three cases: two-byte sequence, three-byte sequence or four-byte sequence. Lookup tables are used to check the validity of the combination of lead byte and second byte as explained below. The sequence is considered consumed only if it’s found to be valid. The corresponding UTF-8 code units are then written to the destination as normal
The UTF-8 to UTF-8 case is different. The input is read twice, but the writing is maximally efficient. First, a UTF-8 validation function is run on the input. This function only reads and doesn’t write and uses an ASCII validation fast path that checks more than one code unit at a time using SIMD or multiple code units per ALU word. The UTF-8 validation function is the UTF-8 to UTF-16 conversion function with all the writes removed. After the validation, the valid UTF-8 run is copied to the destination using
std::ptr::copy_nonoverlapping(), which is the Rust interface to LLVM
memcpy(). This way, the writing, which is generally less efficient than reading, can be done maximally efficiently instead of being done on a byte-by-byte basis for non-ASCII as would result from a read-once implementation. (Note that in the non-streaming case when the input is valid, both the second read and the writing are avoided. More on that later.)
It is not totally clear if this kind of double-reading is smart, since it is a pessimization for the 100% ASCII case. Intuitively, it should help the non-ASCII case, since even the non-ASCII parts can be written using SIMD. However, 100% ASCII UTF-8 to UTF-8 streaming case, which copies instead of borrowing, runs on Haswell at about two thirds of
memcpy() speed while the 100% ASCII windows-1252 to UTF-8 case (which writes the SIMD vectors right away without re-reading) runs at about
The hard parts of looping over potentially-invalid UTF-8 are:
encoding_rs combines the solution for the first two problems. Once it’s known that the lead byte is not ASCII, the lead byte is used as an index to a lookup table that yields a byte whose lower two bits are always zero and that has exactly one of the other six bits set to represent the following cases:
The second byte is used as an index to a lookup table yielding a byte whose low two bits are always zere, whose bit in the position corresponding to the lead being illegal is always one and whose other five bits are zero if the second byte is legal given the type of lead the bit position represents and one otherwise. When the bytes from the two lookup tables are ANDed together, the result is zero if the combination of lead byte and second byte is legal and non-zero otherwise.
When a trail byte is always known to have the normal range, as the third byte in a three-byte sequence is, we can check that the most significant bit is 1 and the second-most significant bit is zero. Note how the ANDing described in the above paragraph always leaves the two least-significant bits of the AND result as zeros. We shift the third byte of a three-byte sequence right by six and OR it with the AND result from the previous paragraph. Now the validity of the three-byte sequence can be decided in a single branch: If the result is 0x2, the sequence is valid. Otherwise, it’s invalid.
In the case of four-byte sequences, the number computed per above is extended to 16 bits and the two most-significant bits of the fourth byte are masked and shifted to bit positions 8 and 9. Now the validitiy of the four-byte sequence can be decidded in a single branch: If the result is 0x202, the sequence is valid. Otherwise, it’s invalid.
The fast path checks that there is at least 4 bytes of input on each iteration, so the bytes of any valid byte sequence for a single scalar value can be read without further bound checks. The code does use branches to decide whether to try to match the bytes as a two-byte, three-byte or four-byte sequence. I tried to handle the distinction between two-byte sequences and three-byte sequences branchlessly when converting UTF-8 to UTF-16. In this case, the mask applied to the lead byte is taken from a lookup table and mask is taken from a lookup table to zero out the bits of the third byte and the third shift amount (from 6 to 0) in the two-byte case. The result was slower than just having a branch to distinguish between two-byte sequences and three-byte sequences.
Now that there is branching to categorize the sequence length, it becomes of interest to avoid that branching. It’s also of interest to avoid going back to the SIMD ASCII fast path when the next lead is not ASCII. After a non-ASCII byte sequence, instead of looping back to the ASCII fast path, the next byte is read and checked. After a two-byte sequence, the next lead is checked for ASCIIness. If it’s not ASCII, the code loops back to the point where the SIMD ASCII path has just exited. I.e. there’s a non-ASCII byte as when exiting the ASCII SIMD fast path, but its non-ASCIIness was decided without SIMD. If the byte is an ASCII byte, it is processed and then the code loops back to the ASCII SIMD fast path.
Obviously, this is far from ideal. Avoiding immediate return to ASCII fast path after a two-byte character works within a non-Latin-script word but it doesn’t really help to let one ASCII character signal a return to SIMD when the one ASCII character is a single space between two non-Latin words. Unfortunately, trying to be smarter about avoiding too early looping back to the SIMD fast path would mean more branching, which itself has a cost.
In the two-byte case, if the next lead is non-ASCII, looping back to immediately after the exit from the ASCII fast path means that the next branch is anyway the branch to check if the lead is for a two-byte sequence, so this works out OK for words in non-Latin scripts in the two-byte-per-character part of the Basic Multilingual Plane. In the three-byte case, however, looping back to the point where the ASCII SIMD fast path ends would first run the check for a two-byte lead even though after a three-byte sequence the next lead is more likely to be for another three-byte sequnces. Therefore, after a three-byte sequence, the first check performed on the next lead is to see if it, too, is for a three-byte sequence in which case the code loops back to the start of the three-byte sequence processing code.
UTF-16LE and UTF-16BE are rare enough on the Web that a browser can well get away with a totally naïve and slow from-the-spec implementation. Indeed, that’s what landed in Firefox 56. However, when talking about encoding_rs, it was annoying to always have the figurative asterisk next to UTF-16LE and UTF-16BE to disclose slowness when the rest was fast. To get rid of the figurative asterisk, UTF-16LE and UTF-16BE decode is now optimized, too.
If you read The Unicode Standard, you might be left with the impression that the difference between UTF-16 as an in-memory Unicode representation and UTF-16 as an interchange format is byte order. This is not the full story. There are three additional concerns. First, there is a concern of memory alignment. In the case of UTF-16 as an in-memory Unicode representation, a buffer of UTF-16 code units is aligned to start at a memory address that is a multiple of the size of the code unit. That is, such a buffer always starts at an even address. When UTF-16 as an interchange format is read using a byte-oriented I/O interface, it may happen that a buffer starts at an odd address. Even on CPU architectures that don’t distinguish between aligned and unaligned 16-bit reads and writes on the ISA layer, merely reinterpreting a pointer to bytes starting at an odd address as a pointer pointing to 16-bit units and then accessing it as if was a normal buffer of 16-bit units is Undefined Behavior in C, C++, and Rust (as can in practice be revealed by autovectorization performed on the assumption of correct alignment). Second, there is the concern of buffers being an odd number of bytes in length, so the special logic is needed to handle the split UTF-16 code unit at the buffer boundary. Third, there is the concern of unpaired surrogates, so even when decoding to UTF-16, the input can’t be just be copied into right alignment, potentially with byte order swapping, without inspecting the data.
The structure of the UTF-16LE and UTF-16BE decoders is modeled on the structure of the UTF-8 decoders: There’s a naïve from-the-spec outer tier that deals with invalid and partial sequences and an inner fast path that only deals with valid sequences.
At the core of the fast path is a struct called
UnalignedU16Slice that wraps
*const u8, i.e. a pointer that can point to either an ever or an odd address, and a length in 16-bit units. It provides a way to make the unaligned slice one code unit shorter (to exclude a trailing high surrogate when needed), a way to take a tail subslice and ways to read a
u16 or, if SIMD is enabled,
u16x8 in a way that assumes the slice might not be aligned. It also provides a way to copy, potentially with endianness swapping, Basic Multilingual Plane code units to a plain aligned
&mut [u16] until the end of the buffer or surrogate code unit is reached. If SIMD is enabled, both the endianness swapping and the surrogate check are SIMD-accelerated.
When decoding to UTF-16, there’s a loop that first tries to use the above-mentioned Basic Multilingual Plane fast path and once a surrogate is found, handles the surrogates on a per-code-unit and returns back to the top of the loop if there was a valid pair.
When decoding to UTF-8, code copied and pasted from the UTF-16 to UTF-8 encoder is used. The difference is that instead of using
&[u16] as the source, the source is an
UnalignedU16Slice and, additionally, reads are followed with potential endian swapping. Additionally, unpaired surrogates are reported as errors in decode while UTF-16 to UTF-8 encode silently replaces unpaired surrogates with the REPLACEMENT CHARACTER. If SIMD is enabled, SIMD is used for the ASCII fast path. Both when decoding to UTF-8 and when decoding to UTF-16, endianness swapping is represented by a trait parameter, so the conversions are monomorphized into two copies: One that swaps endianness and one that doesn’t. This results in four conversion functions: Opposite-endian UTF-16 to UTF-8, same-endian UTF-16 to UTF-8, opposite-endian UTF-16 to UTF-16, same-endian UTF-16 to UTF-16. All these assume the worst for alignment. That is, code isn’t monomorphized for the aligned and unaligned cases. Unaligned access is fast on aarch64 and on the several most recent x86_64 microarchitectures, so optimizing performance of UTF-16LE and UTF-16BE in the aligned case for Core2 Duo-era x86_64 or for ARMv7 at the expense of binary size and source code complexity would be a bit too much considering that UTF-16LE and UTF-16BE performance doesn’t even really matter for Web use cases.
Unlike the other decoders, the x-user-defined decoder doesn’t have an optimized ASCII fast path. This is because the main remaining use case for x-user-defined it is loading binary data via
XMLHttpRequest in code written before proper binary data support via
meta tag, the windows-1252 decoder is used in place of the x-user-defined decoder.)
When decoding to UTF-8, the byte length of the output varies depending on content, so the operation is not suitable for SIMD. The loop simply works in a per-byte basis. However, when decoding to UTF-16 with SIMD enabled, each
u8x16 vector is zero-extended into two
u16x8 vectors. A mask computed by a lane-wise greater-than comparison to see which lanes were not in the ASCII range. The mask is used to retain the corresponding lanes from a vector of all lanes set to 0xF700 and the result is added to the original
(Nightly) Rust provides access to portable SIMD which closely maps to LLVM’s notion of portable SIMD. There are portable types, such as the
u16x8 types used by encoding_rs. These map to SSE registers on x86 & x86_64 and NEON registers on ARMv7 & aarch64, for example. The portable types provide lane-wise basic arithmetic, bitwise operations, and comparisons in a portable manner and with generally predictable performance characteristics. Additionally, there are portable shuffles where the shuffle pattern is constant at compile time. The performance characteristics of shuffles rely heavily on the quality of implementation of specific LLVM back ends, so with shuffles it’s a good idea to inspect the generated assembly.
The portable types can be zero-cost transmuted into vendor-specific types in order to perform operations using vendor-specific intrinsics. This means that SIMD code can generally be written in a portable way and specific operations can be made even faster using vendor specific operations. For example, checking if a
u8x16 contains only ASCII can be done very efficiently on SSE2 and aarch64, so the SIMD “is this
u8x16 ASCII?” operation in encoding_rs has vendor-specific specializations for SSE2 and aarch64. This is an amazing improvement over C. With C, an entire langer function / algorithm that uses SIMD ends up being written separately for each instruction set using vendor intrinsics for everything—even the basic operations that are supported by practically all vendors. It often happens that such vendor-specific code is written only for x86/x86_64 with ARMv7 or aarch64 left as a todo with POWER, etc., completely ignored.
Despite Rust making SIMD portable, performance tuning for specific architectures using conditional compilation to turn alternative implementations on or off is still needed. For example, because NEON on ARMv7 lacks an efficient “is this
u8x16 ASCII?” check, using NEON for processing the ASCII runs in UTF-8 validation turned out not to be an improvement over ALU-only code on ARMv7, even though using SIMD in UTF-8 validation makes sense on x86 and x86_64. On the other hand, the difference between using aligned or unaligned SIMD loads and stores is negligible on aarch64 (tested on ThunderX), so on that architecture encoding_rs uses unaligned loads and stores unconditionally. However, especially on Core2 Duo-era x86_64, the difference between using aligned access compared to using unaligned loads and stores with addresses that are actually aligned is very significant, so in the SSE2 case encoding_rs checks for alignment first and has four-way specializations for the four combinations of the source and destination being aligned or unaligned. As of June 2018, 20% of the Firefox x86/x86_64 release population was still on the kind of x86/x86_64 CPU where there’s a substantial performance disparity between aligned and unaligned SIMD loads and stores with actually aligned addresses.
Using SIMD for ASCII poses the problem that many non-Latin scripts use ASCII spaces and punctuation. If we return directly to the SIMD path upon seeing a single ASCII byte after a sequence of non-ASCII, we may end up processing a SIMD vector only to find that it’s not fully ASCII, because it just starts with an ASCII space or an ASCII punctuation character followed by an ASCII space and then non-ASCII follows again.
For non-Latin scripts that use ASCII spaces and punctuation, after non-ASCII it is useful to have a loop that keeps processing ASCII bytes using the ALU as long as the byte values are below the less-than sign. This way, ASCII spaces, punctuation and digits do not result unhelpful use of SIMD, but HTML markup results in a jump back to the SIMD path.
In the case of the legacy CJK encodings, it’s easy to decide whether to have such a punctuation loop are not: Korean benefits from one, so EUC-KR gets such a loop. Chinese and Japanese don’t benefit from such a loop, so the rest of the legacy CJK encodings don’t get one.
The decision is trickier for single-byte encodings and UTF-8. In the interest of code size, all the single byte encodings (other than x-user-defined) are handled with the same code. For the Latin encodings, it would be beneficial not to have a punctuation loop. For Cyrillic, Greek, Arabic and Hebrew, it is beneficial to have the punctuation loop. Decoding the Latin single-byte encodings is faster anyway, so the punctuation loop is therefore all single-byte encodings for the benefit of the ones that are non-Latin but use ASCII spaces and punctuation.
UTF-8 calls for a one-size-fits-all solution. By the same logic, one should expect to put a punctuation loop in the UTF-8 to UTF-16 decoder. Yet, there is no punctuation loop in the UTF-8 to UTF-16 decoder. I don’t recall the details, but a punctuation loop didn’t behave well. I didn’t investigate why exactly a punctuation loop didn’t behave well in this case, but the conversion loop is pretty delicate even without a punctuation loop, so maybe there was some bad interaction in the optimizer. Rust has been through LLVM major version updates since I experimented with this code, so it might be worthwhile to experiment again.
Character encoding conversion libraries typically reserve the right to perform expensive operations when a decoder or an encoder is instantiated. Expensive operations could include loading lookup tables from the file system, decompressing lookup tables or deriving encode-oriented lookup tables from decode-oriented lookup tables. This is problematic.
When the instantiation of a converter is potentially expensive, libraries end up recommending that callers hold onto converters and reset them between uses. Since encoding_rs builds BOM handling into the decoders, does so by varying the initial state of a state machine and BOM sniffing can change what encoding the decoder is for, being able to reset a decoder would require storing a second copy of the initial state in the decoder. More importantly, though, the usage patterns for character encoding converters tend to be such (at least in a Web browser) that there isn’t a natural way for callers to hold onto converters and creating some kind of cache for recycled converters create threading-problems and shouldn’t be the callers’ responsibility anyway. Even a thread-safe once-per-process heap-allocation on first use would be a problem. Firefox is both a multi-threaded and a multi-process application. E.g. generating a heap-allocated encode-optimized lookup table in a thread-safe way on first use would end up costing the footprint of the table in each process even if sharing between threads appeared simple enough.
To avoid these problems, encoding_rs guarantees that instantiating a converter is a very cheap operation: just a matter of loading some constants into a few machine words. No up-front computation on the data tables is performed during the converter instantiation. The data tables are Plain Old Data arranged in the layout that the conversion algorithms access. Of course, if the relevant part of the program binary hasn’t been paged in yet, accessing the data tables can result in the operating system paging them in.
The Encoding Standard gives the mapping tables for the legacy encodings as arrays indexed by what the spec calls the “pointer”. For single-byte encodings, the pointer is simply the unsigned byte value minus 0x80. That is, the lower half passes through as ASCII and the higher half is used for simple table lookup when decoding.
Conceptually, the encoder side is a linear search through the mapping table. A linear search may seem inefficient and, of course, it is. Still, the encode operation with the legacy encodings is actually rather rare in the Web Platform. It is exposed in only two places: in the error handling for the query string URLs occurring as attribute values in HTML and in HTML form submission. The former is error handling for the case where the query string hasn’t been properly percent escaped and, therefore, relatively rarely has to handle non-ASCII code points. The latter happens mainly in response to a user action that is followed by a network delay. An encoding library can get away with slowness in this case, since the slowness can get blamed on the network anyway. Furthermore, encoder speed that is shockingly slow percentage-wise compared to how fast it could be can still be fast in terms of human-perceivable timescales for the kind of input sizes that typically occur in the text fields of an HTML form.
The design of encoding_rs took place in the context of the CLDR parts of ICU having been accepted as part of desktop Firefox but having been blocked from inclusion in Firefox for Android for an extended period of time out of concern of the impact on apk size. I wanted to make sure that encoding_rs could replace uconv without getting blocked on size concerns on Android. Therefore, since there wasn’t a pressing need for the encoders for legacy encodings to be fast and there was a binary size concern (and the performance concern of instantiating an encoder excluding the option of spending time computing and encode-specific lookup table from decode-specific tables at the time of encoder instantiation), I made it a design principle that encoding_rs would have no encoder-specific data tables and instead the encoders would search the decode-oriented data tables even if it meant linear search.
As shipped in Firefox 56, the single-byte encoders in encoding_rs performed forward linear search across each quadrant of the lookup table for the single-byte encoding such that the fourth quadrant was search first and the first quadrant was searched last. This search order make the most sense for the single byte encodings considered collectively, since most encodings have lower-case letters in the fourth quadrant and the first quadrant is either effectively unused or contains rare punctuation.
In encoding_rs 0.8.11 (Firefox 65), though, as a companion change to compile-time options to speed up legacy CJK encode (discussed below), I relaxed the principle of not having any encode-specific data a little based on the observation that adding just 32 bits (not bytes!) of encoder-specific data per single-byte encoding could significantly accelerate the encoders for Latin1-like and non-Latin single-byte encodings while not making the performance of non-Latin1-like Latin encodings notably worse. Adding 8 bits for an offset in the lookup table to the start of a run on the consecutive code points, 8 bits for the length of the run and 16 bits for an offset to the start of the run in the Unicode code points, the common case (the code points to encode falling within the range) could be handled without a linear search. Unlike in the case of CJK legacy encode compile time options, the addition of 32 bits per single-byte encoding was small enough in added footprint that I thought it did not make sense to make it a compile-time option. Instead, the 32 bits per single-byte encoding are there unconditionally.
For multi-byte legacy encodings, the pointer is computed from two or more bytes. In that case, the computation forms a linear offset to the array when not all values of the (typically) two bytes are valid or the valid values for the two bytes aren’t contiguous. This is in contrast to some previous formulations where two bytes are interpreted as a 16-bit big endian integer and then that integer is considered to map to Unicode. Since not all values of the two bytes are in use, simply interpreting the two bytes as a 16-bit big endian integer would result in a needlessly sparse lookup table. (A sparse lookup table can have the benefit of being able to combine bits from the lead and trail byte without an actual multiplication instruction, which may have been important in the past. E.g. Big5 with a dense lookup table involves multiplying by 157, which compiles to an actual multiplication instruction.)
Still with the linearization math provided by the spec, the lookup tables provided by the spec are not fully dense. Since legacy encodings are not exercised by the modern most performance-sensitive sites and binary size on Android was a concern, I sought to make the lookup tables more compact potentially trading off a bit of performance. Visualizing the lookup table for EUC-KR (warning: the link points to a page that may be too large for phones with little RAM) reveals that the lookup table has two unused vertical bands as well as an unused lower left quadrant. The Japanese lookup tables (JIS X 0208 with vendor extensions and JIS X 0212) also have unused ranges. The gbk lookup table has no unused parts but in place of unused parts has areas filled with consecutive Private Use Area code points. More generally, the lookup tables have ranges of pointers that map to consecutive Unicode code points. As the most obvious examples, the Hiragana and Katakana characters occur in the lookup tables in the same order as they appear in Unicode, therefore, forming ranges of consecutive code points. The handling of such ranges can be performed by excluding them from the lookup table and instead writing a range check (and offset addition) in the decoder program code. (Aside: The visualizations were essential in order to gain understanding of the structure of the legacy CJK encodings. I developed the visualizations when working on encoding_rs and contributed them to the spec.)
Furthermore, the way EUC-KR and gbk have been extended from their original designs has a relationship with Unicode. The original smaller lookup table appears in the visualizations of the extended lookup tables on the lower right. In the case of EUC-KR, the original KS X 1001 lookup table contains the Hangul syllables in common use. In the case of gbk, the original GB2312 lookup table contains the most common (simplified) Hanzi ideographs. The extended lookup table for EUC-KR, at the top and on the left, contains in the Unicode order all the Hangul syllabes from the Hangul Syllables Unicode block that weren’t already included in the original KS X 1001 part on the lower right. Likewise, the extended lookup table for gbk, at the top and on the left, contains in the Unicode order all the ideographs from the CJK Unified Ideographs Unicode block that weren’t already included in the original GB2312 part on the lower right.
That is, after omitting the empty vertical bands in EUC-KR, in both EUC-KR and gbk the top part and the bottom left part form runs of consecutive code points such that the last code point in each run is less than the first code point in the next run. These are stored as tables (one for top and another for botton left) that contain the (linearized) pointer for the start of each such run and tables of equal length that contain the first code point of each run. When decoding, binary search with the linearized pointer can be performed to locate the start of the run that the pointer belongs to. The code point at the start of the run can be then obtained by reading the corresponding item from the table of the first code points of the runs. The correct code point within the range can be obtained by adding to the first code point the offset obtained by subtracting the pointer to the start of run from the pointer being searched. On the encoder side, linear search with the code point can be performed in the table starting the first code point of each range instead after it has been established that the Hangul Syllable code point (in the EUC-KR case) or the CJK Unified Ideograph code point (in the gbk case) wasn’t found in the lower right part of the lookup table. (This process could be even optimized further by arranging the tables in the Eytzinger order instead.)
Adding more program code in order to make the lookup tables smaller worked in most cases. Replacing ranges like Hiragana and Katakana with explicitly-programmed range checks and by compressing the top and bottom left parts of EUC-KR and gbk as described above resulted in an overall binary size reduction except for big5. In the case of big5, the added program code seemed to exceed the savings from a slightly smaller lookup table. That’s why the above techniques were not applied in released code to Big5 after all.
However, Big5 did provide the opportunity to separate the Unicode plane from the lower 16 bits instead of having to store 32-bit scalar values. The other lookup tables (excluding the non-gbk part of gb18030, which is totally different) only contain code points from the Basic Multilingual Plane, so the code points can be stored in 16 bits. The lookup table for Big5, however, contains code points from above the Basic Multilingual Plane. Still, the code points from above the Basic Multilingual Plane are not arbitrary code points. Instead, they are all from the Supplementary Ideographic Plane. Therefore, the main lookup table can contain the low 16 bits and then there is a bitmap that indicates whether the code point is on the Basic Multilingual Plane or on the Supplementary Ideographic Plane.
It is worth noting that while the attempts to make the tables smaller strictly add branching when decoding to UTF-16, in some cases when decoding to UTF-8 they merely move a branch to a different place. For example, when the code has a branch to handle e.g. Hiragana by offset mapping, it knows that a Hiragana character will be three bytes in UTF-8, so the branch to decide the UTF-8 sequence length based on scalar value is avoided. (There are separate methods for writing output that is known to be three bytes in UTF-8, output that is known to be two bytes in UTF-8, and output that might be either two or three bytes in UTF-8. In the UTF-16 case, all these methods do the same thing an output a single UTF-16 code unit.)
The effort to reduce the binary size was successful in the sense that the binary size of Firefox was reduced when encoding_rs replaced uconv, even though encoding_rs added new functionality to support decoding directly to UTF-8 and encoding directly from UTF-8.
In the case of Hangul syllables when encoding to EUC-KR even the original unextended KS X 1001 part of the mapping table is in the Unicode order due to KS X 1001 and Unicode agreeing on how the syllables should be sorted. This enables the use of binary search when encoding Hangul into EUC-KR without encode-specific lookup tables.
However, with the exception of gbk extension part that was not in original GB2312, the way the CJK Unified Ideographs have been laid out in the legacy standards has no obvious correspondence to Unicode order. As far as I’m aware, the options are doing a linear search over the decode-oriented data tables or introducing additional encode-oriented data tables. The relative performance difference between these two approaches is, obviously, dramatic.
Even though testing indicated that linear search over the decode-oriented data tables yielded acceptable human-perceived performance for the browser-relevant use cases even on phone-like hardware, I wanted to have a backup plan in case my determination of the human-perceived performance was wrong and users ended up complaining. Still, I tried to come up with a backup plan that would reach uconv performance (which already wasn’t as fast as as an implementation willing to spend memory on encode-specific tables could be) without having to add lookup tables as large as the obviously fast solution of having a table large enough to index by the offset to the CJK Unified Ideographs block would require.
Ideographs appear to be practically unused in modern Korean online writing, so accelerating Hanja to EUC-KR encode wasn’t important. On the other hand, GB2312, original Big5 (without the HKSCS parts) and JIS X 0208 all have the ideographs organized into two ranges: Level 1 and Level 2, where Level 1 contains the more frequently used ideographs. As the backup plan, I developed compile-time-optional encode acceleration of the Level 1 areas of these three mapping tables.
Since this was a mere backup plan, instead of researching better data structures for the problem, I went with the most obvious one: For each of the three legacy standards, an array of the level Level 1 Hanzi/Kanji sorted in the Unicode order and another array of the same length sorted in the corresponding order containing arrays of two bytes already encoded in the target encoding. In the case of JIS X 0208, there are three target encodings, so I used the most common one, Shift_JIS, for the bytes and added functions to transform the bytes to EUC-JP and ISO-2022-JP.
This solution was enough to make encode to the legacy CJK encodings many times faster than uconv. The backup plan, however, didn’t end up needing to ship in Firefox. Linear search seems to be fast enough, considering that users didn’t complain. Indeed, a linear search-based Big5 encoder had already been shipped in Firefox 43 without complaints from users. (However, this, in itself, wasn’t a sufficient data point on its own, since, anecdotally, it seems that the migration from Big5 to UTF-8 on the Web is further along than the migration from Shift_JIS and gbk.)
Even though impressive relative to uconv performance, accelerating Level 1 Hanzi/Kanji encode using binary search remained very slow relative to other encoding conversion libraries. In order to remove the perception that encoding_rs is very slow for some use cases, I implemented a compile-time option to use encode-only lookup tables that are large enough to index into directly by the offset into the Hangul Syllables or CJK Unified Ideographs Unicode blocks. With these options enabled, encoding_rs legacy CJK encoder performance is within an order of magnitude from ICU and kernel32.dll though still generally not exceeding their performance for plain text (that doesn’t have a lot of ASCII markup). Presumably, to match or exceed their performance, encoding_rs would need to use even larger lookup tables directly indexable by Basic Multilingual Plane code point and to have even fewer branches. It is worth noting, though, that while even larger lookup tables might win micro-benchmarks, they might have adverse effects on other code in real application workloads by causing more data to be evicted from caches during the encoding process.
In general, a library that seeks high encoder performance should probably take the advice given in the Unicode Standard and use an array of 256 pointers indexed by the high half of the Basic Multilingual Plane code point where each pointer either points to an array of 256 pre-encoded byte pairs indexed by the lower half of the Basic Multilingual plane code point or is a null pointer if all possible low bit combinations are unmapped.
Still, considering that a Web browser gets away with relatively slow legacy encoders, chances are that many other applications do, too. In general, applications should use UTF-8 for interchange and, therefore, not use the legacy encoders except where truly needed for backward compatibility. Chances are that most applications won’t need to use the compile-time options to enhance encoder performance and if they do, it’s probably more about getting the performance on a level where untrusted input can’t exercise excessively slow code paths rather than about maximal imaginable performance being essential. At this point, it doesn’t make sense to introduce compile options that would deviate more from the Firefox-relevant code structure for the sake of winning legacy encoder benchmarks.
One generally expects Rust code to be safe. Rust code that doesn’t use
unsafe is obviously safe. Rust code that uses
unsafe is safe only if
unsafe has been used correctly. Semi-alarmingly, encoding_rs uses
unsafe quite a bit.
unsafe isn’t used in random ways. Instead, it’s for certain things and only in certain source files. In particular, it is not used inside the source files that implement the logic for legacy CJK encodings, which in the C++ implementation would be the riskiest area in terms of memory-safety bugs. This is not to say that all the
unsafe is appropriate. Some of it would be avoidable right now, but the better way either didn’t exist or didn’t exist outside nightly Rust when I wrote the code, and some will likely become avoidable in the future.
Here’s an overview of the kinds of
unsafe in encoding_rs:
A couple of internal APIs use
char to signify Unicode scalar value. However, the scalar value gets computed in a way that first yields the value as
u32. Since the value is in the right range by construction, it is reinterpreted as
char without the cost of the range check. Some of this use of
unsafe could be avoided by using
u32 instead of
char internally in some places. It’s unclear if the current usage is worthwhile.
Since dealing with character encodings is in the core competence of encoding_rs, it would be silly to run the standard library’s UTF-8 validation on encoding_rs’s UTF-8 output. Instead, encoding_rs uses
unsafe to assert the validity of its UTF-8 output to the type system. It doesn’t make sense to try to get rid of this use of
unsafe. It’s fundamental to the crate.
Rust makes intrinsics categorically
unsafe even in cases where there isn’t actually anything that logically requires a given intrinsic to be
unsafe. This results in the use of
unsafe to call vendor-specific SIMD operations and to annotate
if conditions for branch prediction using
unlikely. This kind of
unsafe makes the code look harder to read and scarier than it actually is, but it is easy to convince oneself that this kind of
unsafe is not risky in terms of the overall safety of the crate.
When working with SIMD, it is necessary to convert between different lane configurations in a way that is just a type-system-level operation and on the machine level is nothing: the register is the same and the operations determine how the contents of the register are interpreted. As a consequence, reinterpreting a SIMD type of a given width in bits (always 128 bits in encoding_rs) as another SIMD type of the same width in bits should be OK if both types have integer lanes (i.e. all bit patterns are valid). I expect that in the future, Rust will gain safe wrappers for performing these reinterpretations. Such wrappers already exist behind a feature flag in the
The ASCII acceleration code reads and writes slices of
usize (if SIMD isn’t enabled) or
u16x8 (if SIMD is enabled). This is done by casting pointers and by dereferencing pointers. This, obviously, is not ideal in terms of confidence in the correctness of the code. Indeed, this kind of code in the
mem module had a bug that made to a crates.io release of encoding_rs, though I believe no one actually deployed that code to end users before the problem is remedied.
While, based in fuzzing, I believe this code to be to correct, potentially in the future it could be more obviously correct by using
_mut) on primitive slices (stabilized in Rust 1.30.0) and
from_slice_aligned and, possibly, their
_unchecked variants on SIMD types in the
packed_simd crate. However, some of these, notably
align_to, are themselves
unsafe, even though
align_to wouldn’t need to be
unsafe when both slice item types allow all bit patterns as their value space as primitive integers and integer-lane SIMD vectors do.
Especially with SIMD but also with UTF-16LE and UTF-16BE, unaligned memory access is done with
std::ptr::copy_nonoverlapping which LLVM optimizes the same way as C’s
In some cases, data is copied from a slice to another using
std::ptr::copy_non_overlapping even when
copy_from_slice on primitive slice would do and the bound check wouldn’t be too much of a performance problem. Removing remaining cases like this would not remove the
unsafe that they are in, because they are right next to setting the logical length of
Vec in a way that exposes uninitialized memory. Since the length is set right there anyway, it doesn’t make much sense to worry about passing the wrong length to
Perhaps the most frustrating use of
unsafe is to omit bound checks on slice access that the compiler logically should be able to omit from safe code but doesn’t. I hope that in the future, LLVM gets taught more about optimizing away unnecessary bound checks in the kind of IR that rustc emits. At present, it might be possible to write the code differently without
unsafe such that the resulting IR would match the kind of patterns that LLVM knows how to optimize. It is not a nice programming experience, though, to try different logically equivalent ways of expressing the code and seeing what kind of assembly comes out of the compiler.
Additionally, there are cases where an array of 128 items is accessed with a byte minus 128 after the bytes is known to have its highest bit set. This can’t be expected to be known to the optimizer in cases where the fact that the highest bit is set has been established using vendor-specific SIMD.
In the opening paragraph, I claimed high correctness. encoding_rs has been tested in various ways. There are small manually-written tests in the Rust source files for edge cases that seemed interesting. Additionally, every index item for every lookup table-based encoding is tested by generating the expectations from the data provided along via different code than the main implementation. In the context of Firefox, encoding_rs is tested using the test cases in Web Platform Tests (WPT). All encoding tests in WPT pass, except tests that test for the new
Additionally, encoding_rs is fuzzed using cargo-fuzz, which wraps LLVM’s coverage-guided libFuzzer for use on Rust code.
Let’s finally take a look at how encoding_rs performs compared to other libraries.
When decoding from UTF-8, the test case is the Wikipedia article for Mars, the planet, for the language in question in HTML.
Reasons for choosing Wikipedia were:
The topic Mars, the planet, was chosen, because it is the most-featured topic across the different-language Wikipedias and, indeed, had non-trivial articles in all the languages needed. Trying to choose a typical-length article for each language separately wasn’t feasible in the Wikidata data set.
The languages were chosen to represent the languages that have Web-relevant legacy encodings. In the case of windows-1252, multiple languages with different non-ASCII frequencies were used. The main shortcoming of this kind of selection is that UTF-8 is not tested with a (South Asian) language that would use three bytes per character in UTF-8 with ASCII spaces and would have more characters per typical word than Korean has.
When decoding from a non-UTF-8 encoding, the test case is synthetized from the UTF-8 test case by converting the Wikipedia article to the encoding in question and replacing unmappable characters with numeric character references (and in the case of Big5 removing a couple of characters that glibc couldn’t deal with).
When testing x-user-defined decode, the test case is a JPEG image, because loading binary data over XHR is the main performance-sensitive use case for x-user-defined.
Vietnamese windows-1258 workloads are excluded, because windows-1258 uses combining characters in an unusual way, so a naïve synthetization of windows-1258 test data from precomposed UTF-8 content would not have represented a real workload.
The encoder work loads use plain text extracts from the decoder test cases in order to simulate form submission (
textarea) workloads. That is, the encoder benchmarks do not test ASCII runs of HTML markup, because that scenario isn’t relevant to Web-exposed browser features.
The other Web-relevant case for the encoders is the parsing of URL query strings. In the absence of errors, the query strings are ASCII.
Obviously, uconv is benchmarked to understand performance relative to what Gecko had before. rust-encoding is benchmarked to understand performance relative to what was already available in the Rust ecosystem.
ICU and WebKit are benchmarked to understand performance relative to other browsers. WebKit uses its own character encoding converters for UTF-8, UTF-16LE, UTF-16BE, x-user-defined, replacement, and windows-1252 and uses ICU for the others. Chrome inherits this approach from WebKit but has changed the error handling for UTF-8 to be spec-compliant and carries substantial patches to ICU for Encoding Standard compliance. WebKit internals were easier to make available to the benchmark harness, so only WebKit is benchmarked.
WebKit’s windows-1252 is not benchmarked, because trying to use it segfaulted and it wasn’t worthwhile to debug the failure. WebKit on macOS is built with clang, of course, but hopefully building with GCC gives a general idea.
ICU is benchmarked as shipped in Ubuntu, but hopefully that’s close enough performance-wise to the copies of ICU used by Safari and Chrome.
kernel32.dll and glibc represent system APIs. I believe Core Foundation on Mac uses ICU internally, so in that sense ICU also represents a system API. I have no idea if the converters in kernel32.dll are performance-wise representative of what Edge and IE use. (kernel32.dll provides only a non-streaming API publicly while Edge and IE clearly need streaming converters.)
Bob Steagall’s UTF-8 to UTF-16 decoder is benchmarked, because an entire talk claiming excellent results was recently dedicated to it at CppCon and it indeed turned out to be exceptional in its speed for non-ASCII input.
Some the comparisons could be considered to compare things that aren’t commensurable. In particular:
Except for kernel32, the measurements exclude the initialization and destruction of the converter. This is to the advantage of uconv, ICU and glibc, which perform more work during converter initialization than encoding_rs does. kernel32 does not expose converter initialization is a distinct operation and it’s not clear if there is an initialization cost the first time a given converter is used or every time.
When converting to and from UTF-8, in the comparison with rust-encoding, rust-encoding targets
Vec<u8> while encoding_rs uses
Cows. In this case, instead of trying to make the comparison fair by making encoding_rs make a useless copy, the comparison demonstrates the benefits of conditionally copy-free Rust API design.
The WebKit API shows traces of Qt’s converter design. This includes always allocating a buffer on the heap for output. As a result, the WebKit numbers include the allocation and deallocation of the output buffer but those numbers are compared with encoding_rs numbers that don’t include buffer allocation and deallocation.
Since a reference libraries do not fully conform to the Encoding Standard, the work being performed isn’t exactly the same. Instead, the closest approximation of a given legacy encoding is used. Even the error handling can differ: WebKit’s UTF-16BE and UTF-16LE converters don’t check for unpaired surrogates and kernel32 shows unpolished behavior on errors.
Arguably, UTF-8 isn’t the native application-side Unicode representation of glibc. However, since e.g. glib (the infrastructure library used by GTK+) uses UTF-8 as its native application-side Unicode representation and wraps glibc for the conversions from external encodings, testing glibc’s performance to and from UTF-8 is relevant to how glibc is used even if arguably unfair.
When encoding from UTF-8, encoding_rs and rust-encoding assume the input is valid, but glibc does not.
The columns are grouped into decode results and into encode results. Those groups, in turn, are grouped into using UTF-16 as the internal Unicode representation and into using UTF-8 as the internal Unicode representation. Both cases are supported by encoding_rs but the libraries being compared with support one or the other. Then there is a column for each library whose performance is being compared with.
-O2) with GCC 7.3.0 on Ubuntu 18.04.
Each row names a language and an external encoding to convert from or to. The numbers are encoding_rs speed factors relative to the library named in the column. 2.0 means that encoding_rs is twice as fast as the reference library named in the column header. 0.5 means that the reference library named in the column header is twice as fast as encoding_rs. 0.00 means that encoding_rs is relatively very slow (still user-perceptibly fast enough for the form submission use case in a browser) and the non-zero decimals didn’t show up in the second decimal position.
encoding_rs and rust-encoding are built with Rust’s default optimization level
opt_level=3 even though encoding_rs in Firefox is built at
opt_level=2 for the time being. encoding_rs in Firefox is expected to switch to
opt_level=3 soon. For these benchmarks, at least on x86_64 Haswell, there is no practical difference between
opt_level=3 being applied to encoding_rs. However, previously there have been issues with
opt_level=2 that I would rather not have investigated, so I am really looking forward to using
opt_level=3 in the Gecko context. Also kewb is built at
-O3. The Rust version was 1.32.0-nightly (9fefb6766 2018-11-13).
In all cases, the default rustc optimization target for a given instruction set architecture was used. That is, e.g. the Haswell numbers mean running the code compiled for the generic x86_64 target on a Haswell chip and do not mean asking the compiler to optimize for Haswell specifically.
encoding_rs uses SSE2 explicitly. Since SSE2 is part of the x86_64 baseline instruction set, other software is eligible for SSE2 autovectorization or to enable explicit SSE2 parts if they have them. At least uconv had an explicit SSE2 code path for ASCII in the UTF-8 to UTF-16 decoder.
|Simplified Chinese, UTF-8||3.02||3.40||1.67||1.96||1.06||1.90||8.93||9.49||1.15||1.58||1.03||1.55||3575.00||75.42|
|Traditional Chinese, UTF-8||3.05||3.42||1.68||1.96||1.07||1.90||8.98||9.54||1.15||1.58||1.03||1.55||3600.25||74.89|
|Simplified Chinese, gb18030||3.68||3.64||5.04||6.40||4.73||0.23||0.01||0.02||0.01||0.01|
|Traditional Chinese, Big5||3.24||3.08||1.87||6.13||4.36||1.29||0.01||0.00||0.01||0.02|
|Simplified Chinese, UTF-16LE||13.52||6.33||8.38||7.60||5.59|
|Traditional Chinese, UTF-16LE||13.48||6.33||8.38||7.58||5.58|
|Simplified Chinese, UTF-16BE||11.27||5.29||7.00||6.85||5.17|
|Traditional Chinese, UTF-16BE||11.31||5.29||7.00||6.84||5.16|
The above table shows the results with the SIMD enabled for encoding_rs but without encode-specific data tables beyond 32 bits of encode-specefic data for each single-byte encoding).
With indexable lookup tables for the CJK Unified Ideographs and Hangul Syllables Unicode blocks, but otherwise retaining the same encoder structure, encoding_rs performs CJK legacy encode like this:
|Simplified Chinese, gb18030||24.53||0.74||2.42||0.97||0.80|
|Traditional Chinese, Big5||160.73||0.68||0.31||1.30||2.07|
Windows 10 is not available, kewb is not optimized for ARM, and browsers are excluded due to compilation problems. encoding_rs and rust-encoding are compiled with NEON enabled. Only encoding_rs uses NEON explicitly. Notably, NEON is less suited for feeding back into control flow than SSE2, so NEON is not used for validating ASCII, so the comparison with the Rust standard library ends up being an ALU vs. ALU comparison.
|Simplified Chinese, UTF-8||2.46||1.43||4.09||7.94||1.82||5748.35||238.95|
|Traditional Chinese, UTF-8||2.46||1.43||4.16||8.01||1.82||5872.95||171.07|
|Simplified Chinese, gb18030||1.74||4.08||4.02||0.01||0.01||0.02|
|Traditional Chinese, Big5||1.73||4.57||4.40||0.01||0.02||0.04|
|Simplified Chinese, UTF-16LE||4.61||3.32||5.88|
|Traditional Chinese, UTF-16LE||4.61||3.32||5.87|
|Simplified Chinese, UTF-16BE||2.83||2.85||4.84|
|Traditional Chinese, UTF-16BE||2.82||2.87||4.84|
I lack access to Windows 10 on aarch64, kewb is not optimized for aarch64, either, and browsers were excluded for compilation problems. As with x86_64, SIMD is part of the baseline compiler target instruction set on aarch64.
While I was not paying attention, ALU code for ASCII validation has gained speed relative to SIMD-based ASCII validation. I suspect this might be due to LLVM updates since LLVM 4. For this reason, I have moved aarch64 to use ALU code for ASCII validation pending more investigation of how to fix the SIMD code.
These numbers are from ThunderX, which is a server chip. Furthermore, this is the first-generation ThunderX, which is an in-order design. Benchmarking on phones does not make sense, because their clock speeds vary all over all the time due to thermal throttling, so benchmark results are not repeatable. Moreover, the thermal throttling may be rather fine-grained, so it is not feasible to identify throttling by looking at a clear 50% drop as is feasible e.g. with Raspberry Pi 3. The problem with ThunderX and Raspberry Pi 3 is that they use cores with in-order designs while high-end phones use more advanced out-of-order designs. It is quite frustrating that there is not good information about what non-phone computers with aarch64 chips might be able to hold a stable clock speed when running a compute benchmark for the purpose of testing small changes in implementation details. Stable clock speed is not a characteristic of ARM hardware and kernel combination that gets advertised or talked about on forums. (In the ARMv7+NEON case, I just happened to discover that a piece of hardware, Samsung Chromebook 2 with Crouton, suited my needs.)
|Simplified Chinese, UTF-8||1.98||1.19||4.67||6.83||1.27||5704.42||35.09|
|Traditional Chinese, UTF-8||1.97||1.18||4.62||6.77||1.28||5706.46||35.22|
|Simplified Chinese, gb18030||1.93||6.94||3.22||0.01||0.01||0.01|
|Traditional Chinese, Big5||1.92||5.65||3.65||0.01||0.01||0.02|
|Simplified Chinese, UTF-16LE||3.26||5.79||4.59|
|Traditional Chinese, UTF-16LE||3.26||5.78||4.58|
|Simplified Chinese, UTF-16BE||2.56||4.92||3.64|
|Traditional Chinese, UTF-16BE||2.56||4.93||3.63|
Rather expectedly, for ASCII on x86_64, SIMD is a lot faster than not using SIMD and encode to legacy encodings without encode-oriented data tables is relatively slow (but, again, still user-perceptibly fast enough even on low-end hardware for the form submission use case for legacy encoders in a Web browser). Also, the naïve the code structure that remains in the ISO-2022-JP decoder is slower than the kind of code structure that uses the program counter as part of the two-byte state tracking leading to more predictable branches.
Unlike the other libraries that convert to UTF-16 or UTF-8, glibc supports conversions from any encoding into any other by pivoting via UTF-32 on a per scalar value basis. This generality has a cost. I think the main take-away for application developers is that a standard library implementation covers a lot of functionality and not all those areas are optimized, so you should not assume that a library is fast at everything just because it is a core system library that has been around for a long time.
As noted earlier, in the “Apples to Oranges Comparisons”, when encoding from UTF-8, glibc treats the input as potentially invalid, but encoding_rs assumes validity, so when encoding from UTF-8 to UTF-8, the encoding_rs numbers are basically for
memcpy but glibc inspects everything.
In contrast, the Windows system converters have been seriously optimized for the encodings that are the default “ANSI code page” for some Windows locale. Notably, this benchmark tested gb18030 (not default system code page for any locale) and not GBK (the default for Simplified Chinese), and gb18030 looks relatively slower than the code pages that are the default in some locale configuration of Windows. EUC-JP, however, looks well optimized in kernel32 despite it not being the default for any locale.
On the decode side, kernel32 is faster than encoding_rs for single-byte encodings for non-Latin scripts that use ASCII punctuation and spaces. However, for Thai and Latin scripts, encoding_rs is faster than kernel32 for single-byte encodings. This shows the cost of ASCII-acceleration when bouncing back to ASCII only for one or two bytes at a time and shows the downside of trying to limit the code footprint of encoding_rs by using the same code for all single-byte encodings with only the lookup table as a configurable variable.
On the encode side, kernel32 isextremely fast relative to other implementations for the encodings that are the default “ANSI code page” for some Windows locale (and for EUC-JP). Windows is not Open Source, so I haven’t seen the code, but from the performance characteristics it looks like kernel32 has a lookup table that can be directly indexed by a 16-bit Basic Multilingual Plane code point and that yields a pair of bytes that can be copied directly to output. In microbenchmarks that don’t involve SIMD-acceleratable ASCII runs, it’s basically impossible to do better. It is hard to know what the cache effects of a maximally large lookup table are outside microbenchmarks, but the lookup table footprint just for CJK Unified Ideographs or just for Hangul Syllables is a large number of cache lines anyway.
Considering the use cases for the kernel32 converters, optimizing for extreme speed rather than small footprint makes sense. When pre-Unicode legacy apps are run on Windows, all calls to systems APIs that involve strings convert between the application-side “ANSI code page” and the system-side UTF-16. Typically, all apps run with the same legacy “ANSI code page”, so only the lookup table for one encoding needs to be actively accessed.
If the mission of the legacy encoders in encoding_rs was to provide maximally fast conversion to legacy encodings as opposed to providing correct conversion to legacy encodings with minimal footprint and just enough speed for the user not to complain about form submission, it would totally make sense to use tables directly indexably by 16-bit Basic Multilingual Plane code point.
Overall, performance-wise the rewrite was an improvement. (More about UTF-16 to UTF-8 encode below.) As far as I can tell, the EUC-KR results for uconv are not a benchmark environment glitch but the EUC-KR implementation in uconv was just remarkably inefficient. The Big5 results say nothing about the original design of uconv. The uconv Big5 implementation being compared with in the one I wrote for Firefox 43, and that implementation already did away with encode-oriented data tables.
In encoding_rs, the ISO-2022-JP decoder uses a state variable while uconv was a bit faster thanks to using the program counter for state.
As noted earlier in the “Apples to Oranges Comparisons” section, the numbers to and from UTF-8 show how much better borrowing is compared to copying when borrowing is possible. That is, encoding_rs borrows and rust-encoding copies.
ICU is an immensely useful and important library, but I am somewhat worried about the mentality that everyone should just standardize on ICU, and that no one can afford to rewrite ICU. In particular, I’m worried about the “just use ICU” approach entrenching UTF-16 as an in-memory representation of Unicode even more at a time when it’s increasingly clear that UTF-8 should be used not only as the interchange representation but also as the in-memory representation of Unicode. I hope the x86_64 and aarch64 results here encourage others to try to do better than ICU, (piece-wise, as the Rust ecosystem is doing) instead of just settling on ICU.
On ARMv7, encoding_rs performs worse than ICU for decoding non-windows-1252 single-byte encodings into UTF-16. This shows how encoding_rs’s design relies heavily on SIMD. ARMv7 has weaker SIMD functionality than x86, x86_64 or aarch64, so the split between ASCII and non-ASCII is a pessimization on ARMv7. In the x86_64 case the benefits of SSE2 for markup offset the downsides of the ASCII/non-ASCII handling split for natural language in the Wikipedia case. Fortunately, mobile browsing skews even more towards UTF-8 than the Web in general, migration from the affected encodings to UTF-8 is, anecdotally, even further along than migration to UTF-8 in general, and aarch64 is around the corner, so I think it isn’t worthwhile to invest effort or binary footprint into having a different design for ARMv7.
While encoding_rs is a lot faster than the other libraries when encoding ASCII or almost-ASCII from UTF-16 to UTF-8, encoding_rs does worse than uconv, kernel32 and ICU in cases where there is only short runs of ASCII, typically one ASCII space, mixed with non-ASCII. This is consistent for the Arabic, Greek, Hebrew and Russian but relative to kernel32this shows up also for Korean and for the Latin script—not just for Vietnamese (with which the effect also shows up relative to uconv), Turkish and Czech that whose non-ASCII frequency is obviously high but even for French.
This shows that the cost of swiching between the ASCII fast path and the non-ASCII mode is higher for UTF-16 input than for single-byte input, which makes sense, since checking whether a SIMD vector of 16-bit units is in the Basic Latin range requires more SSE2 operations that checking a vector of 8-bit units. Considering that the benefit of the ASCII fast path is so large in the ASCII case, I ended up keeping the ASCII fast path, despite it being a pessimization, though, fortunately, not a huge one, for many languages.
Arabic, Hebrew, Greek and Russian are all written in non-Latin scripts that use ASCII spaces and punctuation. Why does Arabic encode perform so much worse? The reason is that the trick of identifying a contiguous dominant range of code points that maps by offset is not as good a fit for windows-1256 as it is for windows-1251, windows-1252, windows-1253, and windows-1255. While there is a range of Arabic characters that is contiguous in both Unicode and in windows-1256, some characters are not in that range. In contrast, all Hebrew consonants (the test data is not vocalized) map by offset between Unicode and windows-1255. The Cyrillic letters needed for Russian are likewise mappable by offset between Unicode and windows-1251 as are Greek lower-case letters (and some upper case ones) in windows-1253. Of course, the bulk of windows-1252 maps by offset.
The approach of offsetting one range does not work at all for windows-1250.
Considering how for Web browser use cases even the relatively extremely slow speed of legacy CJK encode is fast enough, non-ASCII single-byte encode is fast enough for Web browser use cases even when the approach of offsetting a range does not work. The offset approach is just a very small-footprint tweak that is a nice bonus when it does work.
UTF-8 validation in the Rust standard library is very fast. It took quite a bit of effort to do better. (I hope that the code from encoding_rs gets upstreamed to the standard library eventually.) I managed to make encoding_rs faster than the standard library for input that’s not 100% ASCII first, but even when encoding_rs was faster than the standard library for English Wikipedia, the standard library was still faster for 100% ASCII. To make encoding_rs faster even in that case, it was necessary to introduce a two-tier approach even to the ASCII fast path. Assuming that the input is long enough to use SIMD at all, first the ASCII fast path processes 16 bytes as an unalinged SSE2 read. If that finds non-ASCII, the cost of having bounced to the SIMD path is still modest. If the first 16 bytes are ASCII, the fast path enters an ever faster path that uses aligned reads and unrolls the loop by two.
The data cache footprint of the UTF-8 validation function in the Rust standard library is 256 bytes or four cache lines. The data cache footprint of encoding_rs’s UTF-8 validation function is 384 bytes or six cache lines, so 50% more. Using a lookup table to speed up a function that in principle should be doing just simple bit manipulation is a bit questionable, because benchmarks show behavior where the cost of bringing the lookup table to the cache is amortized across the benchmark iterations and the application-context cost of having to evict something else is not visible. For long inputs containing non-ASCII, using a lookup table is clearly justified. The effects on handling short strings as part of a larger system are unclear. As we’ve learned from Spectre, we shouldn’t assume that the 100% ASCII case avoids bringing the lookup table into the data cache.
Interestingly, WebKit’s ASCII fast path is written as ALU code. I didn’t bother trying to locate the right disassembly, but if the performance is any indication, GCC must be unrolling and autovectrorizing WebKit’s ASCII fast path.
Bob Steagall’s UTF-8 to UTF-16 decoder that combines SSE2 with a Deterministic Finite Automaton (DFA) is remarkably fast. While encoding_rs is a bit faster for Latin script with very infrequent non-ASCII (the threshold is between German and Portuguese) and for writing that doesn’t use use ASCII spaces (Thai, Chinese, and Japanese), the DFA is faster for everything that involves more frequent transitions between ASCII and non-ASCII. I haven’t studied properly how the implementation manages the transitions between SSE2 and the DFA, but the result is awesome.
Compared to encoding_rs’s lookup table of 384 bytes or six cache lines, the DFA has a larger data cache footprint: the presentation slides say 896 bytes or 14 cache lines. As noted earlier, in the benchmarks the cost of bringing the tables into the cache are amortized across benchmark iterations and the cost of having to evict something else in a real-world application is not visible in a benchmark. Considering that
encoding_rs::mem (discussed below) reuses encoding_rs’s UTF-8 to UTF-16 decoder for potentially short strings, I’m reluctant to adopt the DFA design that could have adverse cache effects in an application context.
The above discussion has been about encoding_rs in its role for converting between external encodings and the application-internal Unicode representation(s). That kind of usage calls for a well-designed streaming API when incremental processing of HTML (and XML) is one of the use cases. However, if an application that has, for legacy reasons, multiple application-internal representations, converting between those generally calls less for streaming generality and more for API simplicity.
encoding_rs::mem provides efficient conversions between these four cases as well as functionality for checking if UTF-16 or UTF-8 only contains code points in the Latin1 range. Furthermore, the module also is able to check if text for sure does not contain any right-to-left text. While this check seems to be out of place in this module, it makes sense to combine this check with a Latin1ness check when creating DOM text nodes. Also, it makes sense to optimize the check using portable SIMD. (In Gecko, documents start their life as left-to-right-only. As long as they stay that way, the Unicode Bidirectional Algorithm can be optimized out in layout. However, whenever text is added to the document, it needs to be inspected to see if it might contain right-to-left characters. Once at least one such character is encountered, the document transitions into the bidi mode and the Unicode Bidirectional Algorithm is used in layout from then on.)
Notably, the use case of converting in-memory text is different from converting incrementally-parsed HTML or XML. Instead of providing streaming conversion,
encoding_rs::mem provides conversions in a non-streaming manner, which enables a simpler API. In most cases, the caller is supposed to allocate the target buffer according to the maximum possible length requirement. As an exception, conversions to UTF-8 can be performed in multiple steps in order to avoid excessive allocation, considering that the maximum possible length requirement when converting from UTF-16 to UTF-8 is three times the minimum possible case. The general assumption is that when converting from UTF-16 to UTF-8, first this buffer is sized according to the minimum possible case and rounded up to the allocator bucket and if the result doesn’t fit, then the maximum possible case is tried. When converting XPCOM strings, though, there’s an additional heuristic that looks at the first two cache lines of the UTF-16 buffer in order to make a guess whether the initial allocation should be larger than the minimum possible size.
Since Gecko uses an allocator with power-of-two buckets, is not worthwhile to compute the buffer size requirements precisely. Being a bit wrong still often ends up in the same allocator bucket. Indeed, the new code that makes guesses and occasionally reallocates is generally faster than the old code that tried to compute the buffer size requirements precisely and ended up doing UTF math twice in the process.
The code for
encoding_rs::mem looks rather unreviewable. It is that way due performance reasons. The messy look arises from SIMD with raw pointers, manual alignment handling and manual loop unrolling. To convince myself and others that the code does what it is supposed to do, I created another implementation of the same API in the simplest way possible using the Rust standard library facilities. Then I benchmarked the two to verify that my complicated code indeed was faster. Then I used cargo-fuzz to pass the same fuzzing input to both implementations and seeing that their output agrees (and that there a no panics or Address Sanitizer-reported problems).
This description of
encoding_rs::mem looks thematically quite different from the earlier discussion of encoding_rs proper. Indeed, as far as API usage goes,
encoding_rs::mem should be a separate crate. The only reason why it is a submodule is that the two share implementation details that don’t make sense to expose as a third crate with the public API. Users of encoding_rs that don’t need
encoding_rs::mem should simply ignore the submodule and let link-time optimization discard it.