bug#35721: 27.0.50; Strange Arabic shaping behavior

[Basil L. Contovounesios]

Eli Zaretskii <eliz@gnu.org> writes:

>> From: "Basil L. Contovounesios" <contovob@tcd.ie>
>> Date: Mon, 13 May 2019 23:09:06 +0100
>> 
>> I see the following on the master, harfbuzz, and emacs-26 branches
>> (precise versions follow my signature), but I'm not sure how much of
>> this is expected or due to e.g. my font.
>> 
>> 0. emacs -Q
>> 1. C-x 8 RET 0634 RET
>> 
>> The "tail" of the sheen is truncated by the fringe:
>
> After looking at the code and thinking about this, I think this is a
> feature (as strange as it may sound); see below for why I think so.
> And yes, it definitely depends on the font, in this case DejaVu Sans
> Mono.  I don't see this with any other fixed-pitch font I have.
>
> I'm not 100% sure I'm right here, so I CC Behdad and Kenichi in the
> hope that they will comment on this.  Behdad, I see a very similar
> issue with hb-view, when it renders this character using DejaVu Sans
> Mono, so it isn't just an Emacs issue (and seeing what hb-view
> produces actually made me think my opinion is correct about this).
>
>> 3. SPC
>> 
>> The "tail" of the sheen becomes visible, but falls outside of the box
>> cursor:
>
> Yes, this particular font's glyph for sheen has a negative value of
> left bearing.  Which AFAIU means it extends beyond the box dimensions
> to the left.

OK.

>> 4. C-x 8 RET 0643 RET
>> 
>> The kaf is correctly shaped in its initial form:
>> 
>> 5. C-SPC
>> 
>> The kaf changes to its isolated form:
>
> This is different problem, related to how we redraw portions of the
> buffer inside the region (more generally, those which have colors
> different from the default face).
>
> The problem is that we only pass to the shaping engine stretches of
> text that have the same face.  The basic reason for that is that a
> different face can use a different font, and we can only handle
> character composition for characters supported by the same font.
> Another fundamental reason is that the display engine processes text
> in chunks that have the same face.  So when the active region, or some
> other Emacs feature, paints portions of text in some non-default face,
> we redraw the display, and pass to the shaping engine only the portion
> that has that different face.  If that portion is a single character,
> you will see that it loses its correct shape and is rendered in its
> isolated form.  And if the colors change between two characters that
> need to be shaped together, the shaping will break.
>
> You can easily see this effect if you display HELLO, and then
> shift-select portions of the Arabic greeting (or any other script that
> is a heavy user of character compositions).
>
> To fix this, we need some mechanism that will pass larger chunks of
> text to the shaper in these cases, which will need some changes in how
> the display engine iterates through buffer/string text when it
> prepares them for display: we currently stop at every change of face.

Makes sense, thanks for explaining.

> Patches to fix this are most welcome.

I don't think I need to tell you not to hold your breath.
Maybe one day...

>> I occasionally see this happen even without typing anything, as if by a
>> timer, but I'm not sure how to reproduce it.  I think, without being
>> 100% certain, that it's only happened while using the 'arabic' input
>> method.
>
> Maybe, but given my description above, I'm not surprised, because it's
> enough that Emacs decides, for some reason, to redraw just that one
> character.

Your description above explains the case where the mark is activated in
the middle of a composition, but I don't think it explains why inserting
characters further down the buffer would affect compositions on previous
lines, as in steps 9 and 10 in the OP, where no face change is involved.

> Now to the original problem.  Let me turn the table and ask you: what
> did you expect to happen instead?  This is a fixed-pitch font, so how
> can Emacs display a character that extends to the left from its box,
> at the left-most window coordinate?  It has no choice but consider its
> extension be off-screen, as if the window was hscrolled.
>
> The "normal" case for this character is to be part of R2L text, which
> begins at the right window margin, and flows to the left.  In that
> case, the extension will overlap the character cell of the next (in
> the logical order) character.

Right.  I only noticed the truncation while writing up the rest of the
report, but it seemed relevant enough to mention.  Following your
explanation, the current truncating behaviour seems reasonable to me.

> So I think we have no bug here, we behave as expected.  If not, I'm
> sure Behdad and Kenichi will correct me.

Thanks,

-- 
Basil