[webkit-dev] MathML layout refactor proposal

[Alejandro Garcia Castro]

Hi,

In the last months we have been working on refactoring the MathML
layout code to make it more maintainable, because in the current point
fixing issues or adding new features has been too complex.

We have reached the point where we think we have to discuss whether
this work makes sense and whether it is interesting for the WebKit
community.  If so, we will start pushing the initial patches from our
branch to upstream.

The idea behind the refactor is to remove FlexBox dependency and
create its own layout MathML methods. The main reasons to do this are:

   - Reduce code complexity: Adapting FlexBox layout (which is already
     complex) in order to create the MahtML layout made the code too
     complex, a big technical debt that made the improvement of MathML
     more difficult every day.

   - Avoid just another FlexBox dependent code: When we had to add
     general layout alignment support we had a lot of problems trying
     to solve some the MathML issues because it was not clear how it
     worked with the FlexBox.

   - Improve performance: We do not need all the features FlexBox
     layout adds for most of the MathML blocks but we are executing
     all that code. We also can simplify the render tree structures
     that were created to make the FlexBox layout work.

   - Make easier to improve the MathML implementation: Using
     independent renderer classes gives more flexibility to get exact
     positioning and spacing required to get high-quality math
     rendering based on TeX rules and the OpenType MATH table.  (cf
     http://www.mathml-association.org/MathMLinHTML5/)

We have a working prototype that basically passes the current MathML
tests and removes the FlexBox dependency in:

https://github.com/alexgcastro/webkit/tree/MathMLLayout

We have basically one commit per MathML renderer that we had to
replace. This is still initial code and we need more work and add more
tests to make sure we are improving the situation.

We want to do it incrementally with 2 steps:
1. Remove FlexBox dependency but do not break the tests, this means
   keep the RenderTree structure. This is basically done, and we just
   need review and try to push them.
2. Refactor the RenderTree structure, removing the anonymous nodes
   created to make FlexBox work and all the code that it was
   required. We already have the Fractions and Underover implemented
   and the code is much clearer. This will break the tests relying on
   a PNG image or a render tree reference. Hence we will do it also
   per renderer and rebasing the tests after each commit.

The main con of the change that the code could be bigger in some parts
of the renderers, but more direct and simple, so it should be actually
good regarding maintenance.

If you have any question, proposal or comment just send it, it would
be great to hear some more feedback, and check if we should proceed
with this effort.

Greetings,

Alex