man gcc

To whom it may concern:

gcc -O3 does NOT enable -funroll-loops, and has not done so since at least version 2.95.3 (released March 16, 2001). I don't know if it used to back in the olden days (like last millennium), but it doesn't now.

gcc 2.95.3:
-O3
Optimize yet more. `-O3' turns on all optimizations specified by `-O2' and also turns on the `inline-functions' option.

gcc 3.04
-O3
Optimize yet more. `-O3' turns on all optimizations specified by `-O2' and also turns on the `-finline-functions' and `-frename-registers' options.

gcc 3.1.1
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions and -frename-registers options.

gcc 3.2.3
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions and -frename-registers options.

gcc 3.3.6
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions and -frename-registers options.

gcc 3.4.6
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions, -fweb, -frename-registers and -funswitch-loops options.

gcc 4.0.4
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions, -funswitch-loops and -fgcse-after-reload options.

gcc 4.1.2
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions, -funswitch-loops and -fgcse-after-reload options.

gcc 4.4.3
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions, -funswitch-loops, -fpredictive-commoning, -fgcse-after-reload and -ftree-vectorize options.

gcc 4.5.0
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and also turns on the -finline-functions, -funswitch-loops, -fpredictive-commoning, -fgcse-after-reload and -ftree-vectorize options.

This has been a public service announcement regarding gcc.
Thank you for your time. You may return to your regularly scheduled blogging.

More optimizations...

As half a followup to my last post, I just tried to recompile all of the C/C++ code I've written on my system (and a few small projects I didn't) with -O3 -ftree-vectorizer-verbose=2 to see what it could actually vectorize.

The results were rather dismal. It optimized one loop that looked like this:

for(int x=0;x<mySize;x++) {
    t[x] = myItems[x];
}

And a bunch that were constant initialization like this:

for(int x=0;x<mySize;x++) {
    t[x] = 0;
}

There were also a couple more in my BitArray that had errors, but looked like I might be able to rewrite them to be vectorizable. Granted, you spend a lot of time in loops, so a few optimizations go a long way, but overall it looks like it has a hard time working with most code.