Introduces a few ad-hoc modifications to the DAFSA aimed to increase
performance while keeping the data size small.
- The 'first layer' of nodes is extracted out and replaced with a lookup
table. This turns the search for the first character from O(n) to O
(1), and doesn't increase the data size because all first characters
in the set of named character references have the
values 'a'-'z'/'A'-'Z', so a lookup array of exactly 52 elements can
be used. The lookup table stores the cumulative "number" fields that
would be calculated by a linear scan that matches a given node, thus
allowing the unique index to be built-up as normal with a O(1) search
instead of a linear scan.
- The 'second layer' of nodes is also extracted out and searches of the
second layer are done using a bit field of 52 bits (the set bits of
the bit field depend on the first character's value), where each set
bit corresponds to one of 'a'-'z'/'A'-'Z' (similar to the first
layer, the second layer can only contain ASCII alphabetic
characters). The bit field is then re-used (along with an offset) to
get the index into the array of second layer nodes. This technique
ultimately allows for storing the minimum number of nodes in the
second layer, and therefore only increasing the size of the data by
the size of the 'first to second layer link' info which is 52 * 8 =
416 bytes.
- After the second layer, the rest of the data is stored using a
mostly-normal DAFSA, but there are still a few differences:
- The "number" field is cumulative, in the same way that the
first/second layer store a cumulative "number" field. This cuts
down slightly on the amount of work done during the search of a
list of children, and we can get away with it because the
cumulative "number" fields of the remaining nodes in the DAFSA
(after the first and second layer nodes were extracted out) happens
to require few enough bits that we can store the cumulative version
while staying under our 32-bit budget.
- Instead of storing a 'last sibling' flag to denote the end of a
list of children, the length of each node's list of children is
stored. Again, this is mostly done just because there are enough
bits available to do so while keeping the DAFSA node within 32
bits.
- Note: Together, these modifications open up the possibility of
using a binary search instead of a linear search over the
children, but due to the consistently small lengths of the lists
of children in the remaining DAFSA, a linear search actually seems
to be the better option.
The new data size is 24,724 bytes, up from 24,412 bytes (+312, -104 from
the 52 first layer nodes going from 4-bytes to 2-bytes, and +416 from
the addition of the 'first to second layer link' data).
In terms of raw matching speed (outside the context of the tokenizer),
this provides about a 1.72x speedup.
In very named-character-reference-heavy tokenizer benchmarks, this
provides about a 1.05x speedup (the effect of named character reference
matching speed is diluted when benchmarking the tokenizer).
Additionally, fixes the size of the named character reference data when
targeting Windows.
There are two changes happening here: a correctness fix, and an
optimization. In theory they are unrelated, but the optimization
actually paves the way for the correctness fix.
Before this commit, the HTML tokenizer would attempt to look for named
character references by checking from after the `&` until the end of
m_decoded_input, which meant that it was unable to recognize things like
named character references that are inserted via `document.write` one
byte at a time. For example, if `∉` was written one-byte-at-a-time
with `document.write`, then the tokenizer would only check against `n`
since that's all that would exist at the time of the check and therefore
erroneously conclude that it was an invalid named character reference.
This commit modifies the approach taken for named character reference
matching by using a trie-like structure (specifically, a deterministic
acyclic finite state automaton or DAFSA), which allows for efficiently
matching one-character-at-a-time and therefore it is able to pick up
matching where it left off after each code point is consumed.
Note: Because it's possible for a partial match to not actually develop
into a full match (e.g. `¬indo` which could lead to `⋵̸`),
some backtracking is performed after-the-fact in order to only consume
the code points within the longest match found (e.g. `¬indo` would
backtrack back to `¬`).
With this new approach, `document.write` being called one-byte-at-a-time
is handled correctly, which allows for passing more WPT tests, with the
most directly relevant tests being
`/html/syntax/parsing/html5lib_entities01.html`
and
`/html/syntax/parsing/html5lib_entities02.html`
when run with `?run_type=write_single`. Additionally, the implementation
now better conforms to the language of the spec (and resolves a FIXME)
because exactly the matched characters are consumed and nothing more, so
SWITCH_TO is able to be used as the spec says instead of RECONSUME_IN.
The new approach is also an optimization:
- Instead of a linear search using `starts_with`, the usage of a DAFSA
means that it is always aware of which characters can lead to a match
at any given point, and will bail out whenever a match is no longer
possible.
- The DAFSA is able to take advantage of the note in the section
`13.5 Named character references` that says "This list is static and
will not be expanded or changed in the future." and tailor its Node
struct accordingly to tightly pack each node's data into 32-bits.
Together with the inherent DAFSA property of redundant node
deduplication, the amount of data stored for named character reference
matching is minimized.
In my testing:
- A benchmark tokenizing an arbitrary set of HTML test files was about
1.23x faster (2070ms to 1682ms).
- A benchmark tokenizing a file with tens of thousands of named
character references mixed in with truncated named character
references and arbitrary ASCII characters/ampersands runs about 8x
faster (758ms to 93ms).
- The size of `liblagom-web.so` was reduced by 94.96KiB.
Some technical details:
A DAFSA (deterministic acyclic finite state automaton) is essentially a
trie flattened into an array, but it also uses techniques to minimize
redundant nodes. This provides fast lookups while minimizing the
required data size, but normally does not allow for associating data
related to each word. However, by adding a count of the number of
possible words from each node, it becomes possible to also use it to
achieve minimal perfect hashing for the set of words (which allows going
from word -> unique index as well as unique index -> word). This allows
us to store a second array of data so that the DAFSA can be used as a
lookup for e.g. the associated code points.
For the Swift implementation, the new NamedCharacterReferenceMatcher
was used to satisfy the previous API and the tokenizer was left alone
otherwise. In the future, the Swift implementation should be updated to
use the same implementation for its NamedCharacterReference state as
the updated C++ implementation.
