[ft-devel] Progress towards binary backwards compatibility

[david turner]

Hello everyone,

I commited yesterday night a massive patch to the current CVS in order 
to respect
the layout of internal objects of FreeType 2.1.7. Note that the current 
code is not
ready for release yet, we need to change the build system to export the 
FT_BASE
functions again, and there are still some issues with the cache 
sub-system (more
details below).

I've been doing quite a lot of archeology and googling to see what 
not-to-bleeding
edge distributions are currently using, and 2.1.7 seems to be appropriate.

Note that we've made other subtle changes to the internal layouts 
between 2.1.7
and 2.1.10, but none that seems to have any impact (at least none I 
could have
found through googling mailing lists), so I've reverted all layouts to 
be compatible
with 2.1.7 anyway.

I've also discovered that Debian Woody uses 2.0.9, while Sarge uses 2.1.2.
There is no way we'll be able to support complete binary compatibility 
with these
versions, because the changes we've made to object layouts have been too
important. However, I'll later inspect the state of exported internal 
functions,
since most of them may be implementable.

I also discovered that we changed one *public* structure in 2.1.8, which is
extremely BAAAAAD. The culprit is the PS_PrivateRec defined in 
freetype/t1tables.h,
which gained an "expansion_factor" field towards the end of its structure.
Very fortunately, nobody seems to be interested by the fields that it 
has shifted
('language_group', 'password' and 'min_feature'), so this probably had 
zero impact
on production code.

Now, let's talk about the cache sub-system. As die-hard FreeType 
developers know,
this portion of the code is still dubbed "beta" or "experimental", and 
has changed quite
a lot, both in object layout and APIs, since its original version.

Unfortunately, I've found that some projects have been using it in 
production code,
this includes at least the following:

- FireFox (or even Mozilla), which contains a module used to perform 
client-side
  anti-aliased rendering through core X11 and FreeType 2. This code 
uses the old
  FTC_ImageCache types, including FTC_ImageDesc which disappeared in 2.1.8

- GNUStep, which uses it to cache font data. At least they acknowledge that
  our API is "in flux"

- NGL, which is a C++ OpenGL wrapper which uses the cache sub-system
 extensively, though in a rather stupid way (e.g. it sets "max_faces" 
to 65536,
 meaning that every opened FT_Face will be kept opened for the program's
 duration)

unfortunately, it is not possible to support all versions of the cache 
sub-system,
mainly because some function signatures have been changed rather 
heavily, without
any simple way to detect which kind of structures/parameters they're 
being passed
on.

I'll try some hack tonight that should at least save the day for 
FireFox, maybe
GNUStep, but certainly not NGL. I'll post my results tomorrow.

From now, and if no other problem is detected, we should consider the 
current layout
and internal APIs as gospel. The only acceptable things to do is to add 
new functions,
and eventually add fields at the end of internal structures, *only* if 
these are not
allocated in arrays.

More on this later,

- David


***********************************************************************************
Information contained in this email message is confidential and may be 
privileged, and is intended only for use of the individual or entity named 
above. If the reader of this message is not the intended recipient, or the 
employee or agent responsible to deliver it to the intended recipient, you are 
hereby notified that any dissemination, distribution or copying of this 
communication is strictly prohibited. If you have received this communication 
in error, please immediately notify the address@hidden and destroy the original 
message.
***********************************************************************************