We do want to allow LoadPRE to perform LICM-like transformations: we already consider PHI nodes to be negligible for
code size (making this transform code size neutral), and it allows us to hoist values out of loops, which is always
a good thing.

llvm-svn: 115205