font family-specific fontconfig directives apply globally to all fonts

Per downstream report,
https://bugzilla.redhat.com/show_bug.cgi?id=705348

fedora is using fontconfig directives to enable freetype autohinter on a per-font basis. turns out qt doesn't handle this well, in that these end up applying globally to all displayed fonts.

In particular, ff one installs a /etc/fonts/conf.d/foo.conf that enables autohinting for a particular font family with a construct like:
<match target="font">
<test name="family" compare="eq">
<string>insert_real_font_family_here</string>
</test>
<edit name="autohint" mode="assign">
<bool>true</bool>
</edit>
</match>

The short version of the story is that Qt ends up erroneously applying this rule to (almost?) all fonts.

Per some comments in the downstream report too,

"
The way Pango is using fontconfig is completely correct and the way fontconfig
is designed to be used. In short, this is how to use fontconfig:

• Create a pattern specifying the characteristics you are looking for (ie.
  "sans bold 18"). Call this pat.

• Call FcConfigSubstitute (NULL, pat, FcMatchPattern). That's where
  target="pattern" configuration is applied.

• Call FcDefaultSubstitute (pat)

• Then either: { - Call FcFontSort (NULL, pat, ...). This will return a fontset, lets call it fs. - For any item in fs that you want to use (call that 'font'), you should: - Call FcRenderPrepare (NULL, pat, font). This will implicitly call FcConfigSubstituteWithPat (NULL, res, pat, FcMatchFont) on the pattern that it returns. That's where target="font" configuration is applied. - Use the resulting pattern to open the font and render it. }

  or

  { - Call FcFontMatch (NULL, pat, &result). This will return what we call a "font" pattern. It already has FcFontRenderPrepare (NULL, pat, font) called on it by FcFontMatch. }

Now! The way Qt is abusing fontconfig is:

• qt_fontSetForPattern calls FcFontSort (NULL, pat, ...) and gets the
  resulting fontset 'fs' and saves it.

• • • Note that it should have called FcConfigSubstitute (pattern,
      FcMatchPattern) and FcDefaultSubstitute (pattern) first, but it didn't.

• In QFontEngineMultiFT::loadEngine(), it gets font patterns out of fs, and
  calls it pattern. Note that this is a target="font" kind of pattern.

• Then it calls:

FcConfigSubstitute(0, pattern, FcMatchPattern);
FcDefaultSubstitute(pattern);

• • • This is completely wrong, since pattern is a font pattern, not a query
      pattern!

So, that's it. It's completely b0rked. Those two calls should be made before
FcFontSort(), and FcRenderPrepare called in place of those two.
"

and followup

"Ok, I digged this more. In fact, it was far far worse than I thought. The
code is stupid. It calls FcFontMatch() every other function...

I hacked up a patch together. It should improve things a lot. Can you test?
...."

which was accompanied by said patch.

I hope this can be used as a hint and good starting place for someone more familiar with the code to take this and run. Or at least, to help collaboratively find a better solution.

p.s. I submitted this prior as http://bugreports.qt.nokia.com/browse/QTBUG-19846 , which was abruptly closed due to, imo, an assumption that we were supplying a patch as-is for inclusion, rather than reporting a specific bug/problem. The basis for that assumption was largely my own fault for not including enough detail in the initial report. Hope this helps.