- Templates scope state changes
- State-modifying function operate in place instead of returning a template
- Internal template representation contains actual owned nodes instead of a pointer to a syntax tree + an expression map
- No more wide calls
The previous paragraph layout algorithm had a couple of flaws:
- It always produced line break opportunities between runs although on
the textual level there might have been none.
- It didn't handle trailing spacing correctly in some cases.
- It wouldn't have been easily adaptable to Knuth-Plass style optimal
line breaking because it was fundamentally structured first-fit
run-by-run.
The new paragraph layout algorithm fixes these flaws. It proceeds
roughly in the following stages:
1. Collect all text in the paragraph.
2. Compute BiDi embedding levels.
3. Shape all runs, layout all children and store the resulting items in
a reusable (possibly even cacheable) `ParLayout`.
3. Iterate over all line breaks in the concatenated text.
4. Construct lightweight `LineLayout` objects for full lines instead of
runs. These mostly borrow from the `ParLayout` and only reshape the
first and last run if necessary. The design allows to use Harfbuzz's
UNSAFE_TO_BREAK mechanism to make reshaping more efficient. The size
of a `LineLayout` can be measured without building the line's frame.
5. Build only the selected line's frames and stack them.
This creates a smaller state machine helper type for softness coalescing, which does not own the resulting nodes. While this creates a bit more duplication in stack and par builder, it makes it a lot easier to integrate additional logic into the paragraph builder.
Furthermore:
- Line breaks are now "hard", that is, not coalesced with each other.
- Text nodes with equal style are now merged allowing for example `f{}i` to form a ligature.
