#string #zero-copy #compact #type


Compact buffer/string type for zero-copy parsing

16 releases

0.4.1 Dec 7, 2018
0.4.0 Jun 29, 2017
0.3.1 Jun 29, 2017
0.2.3 Aug 26, 2016
0.1.3 Jul 27, 2015
Download history 8202/week @ 2018-12-20 7288/week @ 2018-12-27 8337/week @ 2019-01-03 7939/week @ 2019-01-10 8202/week @ 2019-01-17 8115/week @ 2019-01-24 8692/week @ 2019-01-31 10444/week @ 2019-02-07 8789/week @ 2019-02-14 9990/week @ 2019-02-21 8641/week @ 2019-02-28 9378/week @ 2019-03-07 8902/week @ 2019-03-14 8652/week @ 2019-03-21 8871/week @ 2019-03-28

36,119 downloads per month
Used in 129 crates (13 directly)




Warning: This library is at a very early stage of development, and it contains a substantial amount of unsafe code. Use at your own risk!

Build Status

API Documentation


Tendril is a compact string/buffer type, optimized for zero-copy parsing. Tendrils have the semantics of owned strings, but are sometimes views into shared buffers. When you mutate a tendril, an owned copy is made if necessary. Further mutations occur in-place until the string becomes shared, e.g. with clone() or subtendril().

Buffer sharing is accomplished through thread-local (non-atomic) reference counting, which has very low overhead. The Rust type system will prevent you at compile time from sending a tendril between threads. (See below for thoughts on relaxing this restriction.)

Whereas String allocates in the heap for any non-empty string, Tendril can store small strings (up to 8 bytes) in-line, without a heap allocation. Tendril is also smaller than String on 64-bit platforms — 16 bytes versus 24. Option<Tendril> is the same size as Tendril, thanks to NonZero.

The maximum length of a tendril is 4 GB. The library will panic if you attempt to go over the limit.

Formats and encoding

Tendril uses phantom types to track a buffer's format. This determines at compile time which operations are available on a given tendril. For example, Tendril<UTF8> and Tendril<Bytes> can be borrowed as &str and &[u8] respectively.

Tendril also integrates with rust-encoding and has preliminary support for WTF-8 buffers.

C interface

Tendril provides a C API, which allows Rust to efficiently exchange buffers with C or any other language.

#include "tendril.h"

int main() {
    tendril t = TENDRIL_INIT;
    tendril_sprintf(&t, "Hello, %d!\n", 2015);
    tendril_fwrite(&t, stdout);
    some_rust_library(t);  // transfer ownership
    return 0;

See the API documentation and the test program.

Plans for the future


html5ever will use Tendril as a zero-copy text representation. It would be good to preserve this all the way through to Servo's DOM. This would reduce memory consumption, and possibly speed up text shaping and painting. However, DOM text may conceivably be larger than 4 GB, and will anyway not be contiguous in memory around e.g. a character entity reference.

Solution: Build a rope on top of these strings and use that as Servo's representation of DOM text. We can perhaps do text shaping and/or painting in parallel for different chunks of a rope. html5ever can additionally use this rope type as a replacement for BufferQueue.

Because the underlying buffers are reference-counted, the bulk of this rope is already a persistent data structure. Consider what happens when appending two ropes to get a "new" rope. A vector-backed rope would copy a vector of small structs, one for each chunk, and would bump the corresponding refcounts. But it would not copy any of the string data.

If we want more sharing, then a 2-3 finger tree could be a good choice. We would probably stick with VecDeque for ropes under a certain size.

UTF-16 compatibility

SpiderMonkey expects text to be in UCS-2 format for the most part. The semantics of JavaScript strings are difficult to implement on UTF-8. This also applies to HTML parsing via document.write. Also, passing SpiderMonkey a string that isn't contiguous in memory will incur additional overhead and complexity, if not a full copy.

Solution: Use WTF-8 in parsing and in the DOM. Servo will convert to contiguous UTF-16 when necessary. The conversion can easily be parallelized, if we find a practical need to convert huge chunks of text all at once.

Source span information

Some html5ever API consumers want to know the originating location in the HTML source file(s) of each token or parse error. An example application would be a command-line HTML validator with diagnostic output similar to rustc's.

Solution: Accept some metadata along with each input string. The type of metadata is chosen by the API consumer; it defaults to (), which has size zero. For any non-inline string, we can provide the associated metadata as well as a byte offset.