I benchmarked libjpeg-turbo on iOS and Mac and it was significantly faster than the built-in stuff (typically around 2-3x), which includes the hardware decoder on iPhone.
Honest questions: is this a complete re-implementation of libjpeg? (It was mentioned that this was originally based off libjpeg/SIMD, a libjpeg v6b fork.) If not, why can't this be backported to upstream libjpeg?