It depends on the hardware. There may be offloading for h.264, VP8, for different profiles and for encoding or decoding. But not all combinations are available on all machines. So software decoding may still be needed.

Now software decoders are often written with some hand-optimized assembly loops, but those optimized variants don't cover all possible target platforms and not every single loop is hand-optimized. So everything that's either a plain-C fallback or just hasn't been optimized can benefit from compiler improvements.