This should be an easy 5-20% gain. We should benchmark the following changes to the compiler options:

• add -flto
• look into LTO related options like -fwhole-program and -fuse-linker-plugin
• do not use -fno-tree-vectorize
• choose a minimum instruction set/architecture we support and specify it with -march=X. I'd suggest sandy bridge (2011). We would not want to tune for sandy bridge, so also specify -mtune=generic

C++11 should also give us an easy gain on platforms that support it; will have to mod cmakelists.txt to detect if it is supported of course.