The Intel Optimizaton Manual states that branches with relative offsets
greater than 2GB suffer a penalty. They cite a 6% improvement when this
is avoided. Our code doesn't rely heavily on dynamically linked
libraries, so I don't expect a similar win, but it's still better to do
it than not.
Eliminate long branches by asking the dynamic linker to restrict itself
to the lower 4GB of the address space. I saw that it maps libraries
at 1GB+ addresses, so this satisfies the limitation.
Fix is from the Intel Optimization Manual as well.
This change was ported from ScyllaDB Enterprise.
Closesscylladb/scylladb#22498
5c647408c7