Roberto Guida wrote: > -O3 -mtune=pentium4 -march=pentium4 -mfpmath=sse -malign-double > -m128bit-long-double -mmmx -msse -msse2 -maccumulate-outgoing-args > -mno-push-args -march=pentium4 already implies all of -mtune=pentium4 -mmmx -msse -msse2 so there's no point in specifying them. And several of those options are ABI-changing which means you'll get random unexplainable crashes (or worse) when trying to intermix code, which is why they say "Warning" in bold in the manual. If it were me I would use "-O3 -march=pentium4" and stop reading whatever gentoo nonsense says that it's a good idea to enable dangerous options as defaults. You can add -mfpmath=sse if you know that you won't ever encounter code that relies on the excess precision of the 387 to function correctly. Yes, such code could probably be classified as broken, but have you audited the source of everything that you intend to compile so that you can rule it out? This is the fallacy with enabling these paragraph-long settings of options as defaults! Not all options make sense for every program, and the ones that do make sense in most situations are already included in -Ox. > then maybe (i'm less sure here, i've to study more! :) ) these ones too: > > -ffast-math -fstrength-reduce -fexpensive-optimizations -funroll-loops > -frerun-loop-opt -falign-loops It would be a bad idea to use -ffast-math unconditionally without understanding its consequences and testing the code it generates for correct behavior. -funroll-loops is also something that is not necessarily always a good idea, and can make code slower, so again not a good choice for a default option. (If it made sense in all cases don't you think it would be enabled at -Ox?) The rest are probably just going to make compilation take longer for unmeasurable returns. You might consider -ftree-vectorize though. Brian
