Looking at the revised version of initGISTstate, I am thinking that
I missed a step in updating the main indexing code for the new
pg_opclass definition. It seems to me that if opckeytype is not 0,
that datatype ought to be used to declare the index column type from
the beginning. Then initGISTstate wouldn't need to go through all
these pushups to develop a correct tuple descriptor for the index.
Any objections?
regards, tom lane