[autofit] Return correct linear advance width values.

This was quite a subtle bug which accidentally showed up with glyph
`afii10023' of arial.ttf (version 2.76).  This glyph is a composite;
the first component, `E', has an advance width of 1366 font units,
while the advance width of the composite itself (which looks like
uppercase `E' with dieresis) is 1367 font units.  I think this is
actually a bug in the font itself, because there is no reason that
this glyph has not the same width as uppercase `E' without the
dieresis.  Anyway, it helped identify this problem.

Using the TrueType hinter, the correct value (1367) of `afii10023'
was returned, but the autohinter mysteriously returned 1366.

Digging in the code showed that the autohinter recursively calls
FT_Load_Glyph to load the glyph, adding the FT_LOAD_NO_SCALE load
flag.  However, the `linearHoriAdvance' field is still returned as a
scaled value.  To avoid scaling twice, the old code in autofit reset
`linearHoriAdvance', using the `horiAdvance' field.  This seemed to
work since FT_LOAD_NO_SCALE was in use, but it failed actually,
because `horiAdvance' is defined as the distance of the first
subglyph's phantom points, which in turn are initialized using the
advance width of the first subglyph.  And as the given example
shows, these widths can differ.

* src/autofit/afloader.c (af_loader_load_g): Temporarily set
FT_LOAD_LINEAR_DESIGN while calling FT_Load_Glyph to get unscaled
values for the linear advance widths.

ChangeLog

@@ -1,3 +1,34 @@
+2012-03-11  Werner Lemberg  <wl@gnu.org>
+
+	[autofit] Return correct linear advance width values.
+
+	This was quite a subtle bug which accidentally showed up with glyph
+	`afii10023' of arial.ttf (version 2.76).  This glyph is a composite;
+	the first component, `E', has an advance width of 1366 font units,
+	while the advance width of the composite itself (which looks like
+	uppercase `E' with dieresis) is 1367 font units.  I think this is
+	actually a bug in the font itself, because there is no reason that
+	this glyph has not the same width as uppercase `E' without the
+	dieresis.  Anyway, it helped identify this problem.
+
+	Using the TrueType hinter, the correct value (1367) of `afii10023'
+	was returned, but the autohinter mysteriously returned 1366.
+
+	Digging in the code showed that the autohinter recursively calls
+	FT_Load_Glyph to load the glyph, adding the FT_LOAD_NO_SCALE load
+	flag.  However, the `linearHoriAdvance' field is still returned as a
+	scaled value.  To avoid scaling twice, the old code in autofit reset
+	`linearHoriAdvance', using the `horiAdvance' field.  This seemed to
+	work since FT_LOAD_NO_SCALE was in use, but it failed actually,
+	because `horiAdvance' is defined as the distance of the first
+	subglyph's phantom points, which in turn are initialized using the
+	advance width of the first subglyph.  And as the given example
+	shows, these widths can differ.
+
+	* src/autofit/afloader.c (af_loader_load_g): Temporarily set
+	FT_LOAD_LINEAR_DESIGN while calling FT_Load_Glyph to get unscaled
+	values for the linear advance widths.
+
 2012-03-10  Werner Lemberg  <wl@gnu.org>
 
 	[truetype] Fix SSW instruction.

src/autofit/afloader.c

@@ -104,9 +104,11 @@
     AF_GlyphHints     hints    = &loader->hints;
     FT_GlyphSlot      slot     = face->glyph;
     FT_Slot_Internal  internal = slot->internal;
+    FT_Int32          flags;
 
 
-    error = FT_Load_Glyph( face, glyph_index, load_flags );
+    flags = load_flags | FT_LOAD_LINEAR_DESIGN;
+    error = FT_Load_Glyph( face, glyph_index, flags );
     if ( error )
       goto Exit;
 
@@ -124,10 +126,6 @@
       FT_Vector_Transform( &loader->trans_delta, &inverse );
     }
 
-    /* set linear metrics */
-    slot->linearHoriAdvance = slot->metrics.horiAdvance;
-    slot->linearVertAdvance = slot->metrics.vertAdvance;
-
     switch ( slot->format )
     {
     case FT_GLYPH_FORMAT_OUTLINE: