Good question. I go with -O3 because it does even more optimizations, but to be quite candid I really don't know what impact -fpredictive-commoning, -ftree-vectorize, and others have on my resulting machine code, if any.
I did a quick experiment, compiling one C file with -O2 -finline-functions and another with -O3, using the -S flag so I could see the assembler output.
The only difference I saw was this:
.comm free_list,8,8
Versus this:
.comm free_list,8,16
Who knows. I guess I'm really using -O3 because "3 is more than 2" -- in other words, "Ours goes to 11!" :)
predictive commoning is a loop optimization that commons cross-loop redundancies.
It is basically CSE "around loops". You can generalize it to subsume loop store motion and strength reduction, but most compilers (including GCC) don't bother.
For example, it transforms
for (int i = 0; i < 50; i++)
a[i+2] = a[i] + a[i+1]
into
p0 = a[0]
p1 = a[1]
for (int i = 0; i < 50; i++) {
a[i+2]=p2=p0+p1;
p0=p1;
p1=p2;
}
Eliminating a whole ton of loads and stores.
It's been a while since i looked at GCC's implementation, but it did pretty well in the past (whether you can do commoning depends on your ability to identify and group sequences, etc)
-ftree-vectorize does the obvious thing (turn on vectorization). How effective it is depends on a lot of factors.
I did a quick experiment, compiling one C file with -O2 -finline-functions and another with -O3, using the -S flag so I could see the assembler output.
The only difference I saw was this:
Versus this: Who knows. I guess I'm really using -O3 because "3 is more than 2" -- in other words, "Ours goes to 11!" :)