What I've found wreaks the most havoc on build times is having too many -iquote and -isystem flags. Last time I profiled TensorFlow compiles using strace, about

What I've found wreaks the most havoc on build times is having too many -iquote and -isystem flags. Last time I profiled TensorFlow compiles using strace, about half the gcc wall time was spent inside all the stat() system calls those flags generate.