Kerning and ligatures

[Pip Cet]

This is a snapshot of my work on supporting kerning and ligatures; the
goal of this patch is to support those features for variable-pitch
fonts in English, but I'm not ignoring RTL languages, or scripts that
require shaping.

Like the first patch, this is meant to demonstrate the behavior I
would like to see, not the implementation. It's not meant for
immediate inclusion; perhaps it can serve as a starting point for
someone wishing to work on this further. That said, this patch does
not re-use the composite.c code. I tried to, several times, but what
it does is simply too different to allow useful code-sharing.

What's it all about is entering text like "affiliation" or "AVATAR"
while using a variable-pitch font. The proper way of displaying this,
for many OpenType fonts with variable pitch, involves ligatures and
kerning.

A ligature is an OpenType glyph that represents several characters;
for example, "ffi" in "affiliation".  Kerning is the determination of
each character's x advance based on its context; for example, in
"AVATAR", the distance between "A", "V", and the following "A" is
reduced.

None of this complicated logic is in the patch; instead, it uses the
HarfBuzz library.

This is somewhat unfortunate because the HarfBuzz API is fairly
limited: instead of requesting context incrementally from the calling
program, it expects to be provided with a "complete" context which it
then shapes. Doing that is probably too expensive, so we provide
"enough" context based on a heuristic. (This is complicated because
getting context is expensive, and too much of it can be harmful; the
current approach is to take a lot of context so the horrible
performance reminds me to do something about it :-) ).

This code supports M-to-N correspondences between code points and
glyphs points; it does so by first producing a single struct glyph
covering all (N) glyphs, then slicing it into one slice for each of
the (M) code points.

Important differences between this code and the existing composite.c code:

1. You can "enter" ligatures, such as by moving point to the second
"f" in "affiliation". This also applies to clusters of glyphs not
usually thought of as ligatures: For accented letters written as two
codepoints, such as á, this may be useful.  For שָׁ, it's irritating.

2. This code does not allow HarfBuzz to default to the language your
system is set to use. That prevents something as innocuous as a
screenshot from revealing your likely location. (The current language
default is "en").

3. Ligatures and kerning work with display strings or overlay strings,
rather than ending there or even revealing replaced buffer text as the
existing code does.

4. This code breaks lines after every glyph slice. That is suboptimal
for many languages.

Can someone help with the Telugu "hello"? I wonder whether there
should be a gap, and how much of one, between స్కా and రం in నమస్కారం?

PS: Sorry for the amount of recent noise.

0001-Snapshot-of-kerning-ligatures-work.-Do-not-merge.patch
Description: Text Data