Fallback from -O3 to -O2 -ftree-vectorize

[V10lator]

According to https://gcc.gnu.org/ml/gcc/2019-01/msg00019.html it might be a good idea to use "-O2 -ftree-vectorize" in case -O3 fails on modern CPUs. Only if -ftree-vectorize fails (which didn't happen for me while compiling ~600 packages) one should fallback to plain -O2.

With GCC 10 this might no longer be needed (if you look at the link you'll notice that GCC devs think about including -ftree-vectorize into -O2) but untill then I think it would be nice to have.

[ionenwks]

On that topic, I just found something curious with -O2 -ftree-vectorize on rxvt-unicode 9.22. If I build it with -O3, it's fine, with -O2 it's also fine, but if I build it with -O2 -ftree-vectorize it segfaults while creating the window. Adding -ftree-loop-distribute-patterns (which is normally enabled by -O3) fixes it.

Not sure if it's a gcc bug, something misleading the compiler in rxvt, or something more specific to my setup. Haven't found anything else affected but it may possibly be unsafe on its own.