We're currently hacking LLVM to install it to /usr/lib/llvm. I don't know how beneficial it is in the current (unversioned) form but one thing I know is that it causes issues. Most importantly, changing the install dir enforces a fair bit of hackery on our side to change the libdir and add RPATHs to compiled executables and llvm-config. In the future, more hacks may become necessary to support that. On the other hand, with that configuration we're not far from supporting installing multiple slots of LLVM in quite an easy way (bug 471330). We'd just have to add version to that directory (and since we're forcing it, it wouldn't be that hard), then hack the header-dir and llvm-config a bit. Yet another alternative (but I haven't looked at the possibility of doing that) is to install slots in custom prefixes. That is, make whole llvm land in /usr/lib/llvm-${SLOT}/{bin,include,lib32,lib64}. I think getting llvm-config to output proper flags for that would be relatively easy. Hackery with executables would be worse. When considering slotting, please think of how it was like with boost. I don't really want to introduce yet another semi-working mess, and waste my time improving it so that someone removes it the next day I finally achieved it. How should we proceed?
Switched to /usr/lib with the multilib conversion some time ago.
Oh wait, I moved clang and this is about llvm.
Well, I've started moving stuff to /usr/lib*.