Re:Android Project Changeset 4f8b683: libc/memset.c
by chadaustin on Tuesday May 25, 2010 @11:18PM
(#1379414)
While we're jumping down the architecture pedant rabbit hole, a simple loop like that will be trivially predicted, so the branch will be basically free. In addition, hardware prefetchers do a much better job at predicting linear memory access than manual prefetch instructions. On Core 2, iirc, if you have 2 or 3 L2 misses at fixed offsets either direction from each other, the hardware will automatically begin prefetching memory so it's there when you need. The problem with manual prefetch instructions is there high latency. They're best for hinting to the processor that you're about to make an unpredictable load.
The shown example, that just writes to RAM, is probably just fast enough (it is harder to optimize read cases), although a bit of unrolling could reduce jump penalty impact (specially on non superscalar or in in-order superscalar CPUs -typical ARM included on handheld devices-).