Merge from mainline.

Pre-alloc splitting needs to be more careful to avoid inserting spills/restores
between call frame setup/restore points.  Unfortunately, this regresses 
code size a bit, but at least it's correct now!

llvm-svn: 64508