[Fontconfig] Application startup performance

[Nick Alcock]

On 6 Apr 2016, Akira TAGOH stated:

> Nick, just reverted the changes of f44bfad. so please test and figure
> out if qsort() thing is still matter.

Reverting that commit does indeed make all the file I/O go away: we
stat() and then open() each directory, then fstatfs() and close it. This
does do a network roundtrip for each dir not in the NFS attribute cache,
but only one per directory, and it's an unavoidable one if we're to
detect stale cache dirs at all.

The effect of the qsort() now appears to be in the noise to me (though
someone with a million fonts might disagree).

-- 
NULL && (void)