[HarfBuzz] Slowdown if pre/post context is provided

[Grigori Goronzy]

Hi,

I changed some old shaping code to provide pre/post context as 
documented in [1]. I simply pass the full text and segment info for the 
relevant part of the text into an hb_buffer_t, instead of passing only 
the run to be shaped into the buffer. This works, but there is a major 
slowdown! I didn't get any precise numbers and didn't do any profiling 
yet, but shaping is 4-5x slower.

I can't imagine that's supposed to slow down the shaper so much when 
processing typical runs. Any idea?

Grigori

[1] 
http://cgit.freedesktop.org/harfbuzz/commit/?id=05207a79e0ae1769c5feaebe3fd99bdf9cfcf834