On Mon, 17 Jan 2005 13:33:18 +0900, "Stephen J. Turnbull" said: -march=pentium4 seems to be the next biggest winner after -Os. I've never encountered code where -O3 helped much - maybe for numerical crunching all the loop unrolling helps, but for most real-life code, the code shrinkage caused by -Os wins due to L1 i-cache and L2 cache occupancy issues. I've not found any flags that help a measurable amount past -O2/-Os -march=pentium4. I'm sure there's 1 or 2 .c files in xemacs that would get 5 or 10% faster with hand-tweaking of the flags, but the seconds saved with said tweaking are outweighed by the hours spent identifying the tweaks. At that point, you're better off profiling and looking for algorithmic wins.