Optimize the processing of text between nodes

[stof]

Instead of processing the text token one by one in the main loop, it is now processed in batch until the next special token (< and & which have special handling in the main loop and NUL characters which need to report a parse error).

https://blackfire.io/profiles/compare/8d7277d0-e2ed-40cf-b9b6-bffa6a523ae6/graph

There is a 51% improvement there

[goetas]

This looks promising :)

[stof]

php test/benchmark/run.php (current master, i.e. 182f34d)
Loading: 101.72620534897
Writing: 37.083342075348

php test/benchmark/run.php (this PR)
Loading: 69.69865322113
Writing: 37.433831691742

php test/benchmark/run_native.php (same benchmark using DOMDocument::loadHTML and DOMDocument::saveHTML instead)
Loading: 10.595810413361
Writing: 3.5749840736389

And for reference, here is the benchmark running on 2.4.0:
Loading: 127.82767772675
Writing: 37.827260494232

This is indeed quite promising (note that all my optimizations since 2.4.0 are focusing on the loading part only, that's why there is not much improvements on the writing side).

2.4.0 was 12x slower than the native parser
This PR reaches the level of 6.5x slower than native parser.

[goetas]

give the time to test this by my self tomorrow, but looks great!
Thanks a lot!

[stof]

And for the first time since I started this optimization work, the DOMTreeBuilder appears in the hot path defined by blackfire, instead of being entirely dominated by the Tokenizer 😄

[stof]

note that I still have a few ideas to keep going after that one (but not as big as that one)

[goetas]

Now that I see it, this looks so obvious :)

give the time to test this by my self tomorrow, but looks great!

impatience :)

[goetas]

Nice to see this benchmark:

v2.3.1

$ php test/benchmark/run.php 10
Loading: 230.20720481873

master

$ php test/benchmark/run.php 10
Loading: 66.839385032654

(php 7.2)

That is almost 4 time faster! (and my guess is that there is still room for improvements as example in the Tokenizer::attribute() function or moving the readUntilSequence into the scanner class)