don't preserve names on IR instructions. This matches llvm-gcc's behavior and

speeds up the compiler by ~8% at -emit-llvm -O0.

llvm-svn: 58977