On Mon, Oct 20, 2008 at 7:01 AM, Anssi Saari <anssi.saari at modeemi.fi> wrote: > On Sat, Oct 18, 2008 at 01:24:23PM -0700, Jason Garrett-Glaser wrote: > >> As far as I've experienced, there is zero measurable benefit from >> -march higher than pentium2 (except for SSE float math maybe), and >> GCC's autovectorization is patently worthless. > > Interesting. I actually tried building Firefox with -ftree-vectorize > recently, but all I got was a binary that segfaulted immediately. > > So is that why it's worthless or is it because even if it produced a > working binary, it wouldn't be any faster? It wouldn't be significantly faster, most likely. Last time I tried vectorization with -march=core2 it attempted to do an aligned load from an unaligned memory address when initializing an array and proceeded to crash, so I gave up on that. Dark Shikari
