I wrote:
> However, see my response to Robert: why are we passing the original node
> to the transform function at all? It would be more useful and easier to
> work with to pass the function's fully-processed argument list, I believe.
After a bit of looking around, I realize that the current implementation
of transform functions is flat-out wrong, because whenever a transform
actually fires, it proceeds to throw away all the work that
eval_const_expressions has done on the input, and instead return some
lightly-modified version of the original node tree. Thus for example
in the regression database:
regression=# create function foo(x float8, y int) returns numeric as
regression-# 'select ($1 + $2)::numeric' language sql;
CREATE FUNCTION
regression=# select "numeric"(foo(y := 1, x := f1), -1) from float8_tbl;
ERROR: unrecognized node type: 310
since the adjustment of foo's named arguments is thrown away.
So this patch is going to need some work. I continue to not see any
particular reason why the transform function should need the original
node tree. I think what it *should* be getting is the OID of the
function (currently, it's impossible for one transform to serve more
than one function, which seems like it might be useful); the input
collation (currently, transforms are basically unusable for any
collation-sensitive function), and the pre-simplified argument list.
regards, tom lane