You're wrong. See e.g. https://llvm.org/pubs/2002-12-LattnerMSThesis.pdf, the original thesis proposing LLVM.

LLVM was always trying to pitch itself as a whole-lifetime optimization system. In practice, this meant focusing on LTO and PGO compilation modes. The idea of runtime or idle-time optimization were always kind of a next-todo step rather than anything actually implemented.