[HarfBuzz] Sinhala split matra

[Harshula]

Hi Behdad,

On Fri, 2012-11-16 at 17:19 +1100, Harshula wrote:
> On Thu, 2012-11-15 at 00:12 -0800, Behdad Esfahbod wrote:

<snip>

> > we decided that we want to support both
> > categories of fonts: those that work fine with Uniscribe, and those that used
> > to work with old HarfBuzz / Pango.  I don't see how that can be a limitation
> > to a user.
> 
> Really? Would both those groups of fonts work at the same time in, for
> example, a Libre Office (using HarfBuzz) document?
> 
> > As such, I have no interest in arguing about how that decision was
> > made, and I don't think it's relevant to any user of HarfBuzz.  If you have a
> > font that is not addressed correctly with HarfBuzz as is, let us know and we
> > will try to accommodate that.
> 
> Did you notice that fonts made for Uniscribe now have a subtle rendering
> error without the environment variable? Try out කෝ with a font made for
> Uniscribe. For example, try:
> 
> http://www.icta.lk/attachments/1090_winnie.ttf
> http://www.icta.lk/attachments/1090_WARNA.ttf
> 
> If you can accommodate both groups of fonts at the same time by default,
> that would be great.

Thanks for undoing Commit 0736915b8ed789a209205fec762997af3a8af89c
([Indic] Decompose Sinhala split matras the way old HarfBuzz / Pango
did)! I did a quick test of Harfbuzz with the following Commit:

------------------------------------------------------
commit 43b653150081a2f9dc6b7481229ac4cd952575dc
Author: Behdad Esfahbod <behdad at behdad.org>
Date:   Fri Nov 16 13:12:35 2012 -0800

    [Indic] Another try to unbreak Sinhala split matras
    
    Just read the comments...

diff --git a/src/hb-ot-shape-complex-indic.cc b/src/hb-ot-shape-complex-indic.cc
index b185824..d924d1a 100644
--- a/src/hb-ot-shape-complex-indic.cc
+++ b/src/hb-ot-shape-complex-indic.cc
@@ -1317,15 +1317,42 @@ decompose_indic (const hb_ot_shape_normalize_context_t *c,
 #endif
   }
 
-  if (indic_options ().uniscribe_bug_compatible)
-  switch (ab)
+  if ((ab == 0x0DDA || hb_in_range<hb_codepoint_t> (ab, 0x0DDC, 0x0DDE)))
   {
-    /* These Sinhala ones have Unicode decompositions, but Uniscribe
-     * decomposes them "Khmer-style". */
-    case 0x0DDA  : *a = 0x0DD9; *b= 0x0DDA; return true;
-    case 0x0DDC  : *a = 0x0DD9; *b= 0x0DDC; return true;
-    case 0x0DDD  : *a = 0x0DD9; *b= 0x0DDD; return true;
-    case 0x0DDE  : *a = 0x0DD9; *b= 0x0DDE; return true;
+    /*
+     * Sinhala split matras...  Let the fun begin.
+     *
+     * These four characters have Unicode decompositions.  However, Uniscribe
+     * decomposes them "Khmer-style", that is, it uses the character itself to
+     * get the second half.  The first half of all four decompositions is always
+     * U+0DD9.
+     *
+     * Now, there are buggy fonts, namely, the widely used lklug.ttf, that are
+     * broken with Uniscribe.  But we need to support them.  As such, we only
+     * do the Uniscribe-style decomposition if the character is transformed into
+     * its "sec.half" form by the 'pstf' feature.  Otherwise, we fall back to
+     * Unicode decomposition.
+     *
+     * Note that we can't unconditionally use Unicode decomposition.  That would
+     * break some other fonts, that are designed to work with Uniscribe, and
+     * don't have positioning features for the Unicode-style decomposition.
+     *
+     * Argh...
+     */
+
+    const indic_shape_plan_t *indic_plan = (const indic_shape_plan_t *) c->plan->data;
+
+    hb_codepoint_t glyph;
+
+    if (indic_options ().uniscribe_bug_compatible ||
+       (c->font->get_glyph (ab, 0, &glyph) &&
+        indic_plan->pstf.would_substitute (&glyph, 1, true, c->font->face)))
+    {
+      /* Ok, safe to use Uniscribe-style decomposition. */
+      *a = 0x0DD9;
+      *b = ab;
+      return true;
+    }
   }
 
   return c->unicode->decompose (ab, a, b);
------------------------------------------------------

Now both groups of fonts appear to render correctly by default. That is
a fantastic outcome for Sinhala script users. BTW, the lklug.ttf font
has been deprecated/unmaintained for some time since FreeSerif's Sinhala
support improved.

Thanks again,
#