Re: varattno remapping

On 12/24/2013 03:21 PM, Abbas Butt wrote:

> Could you please explain a little bit more how would you solve the posed
> problem using map_variable_attnos?
> 
> I was recently working on a similar problem and used the following algo
> to solve it.
> 
> I had to find to which column of the base table does a column in
> the select statement of the view query belong.
> To relate a target list entry in the select query of the view to
> an actual column in base table.

Sounds similar. My problem is simplified by the constraint that the view
must be a simple view (only one base relation) and must only contain
simple column-references to single the base relation. No  expressions,
no joins. (These are the rules for simply updatable views).

I'm new to the planner and rewriter, so don't take what I do as any kind
of example beyond "this seems to work".

I generate a varattno mapping as follows:

/** Scan the passed view target list, whose members must consist solely* of Var nodes with a varno equal to the passed targetvarno.** A mapping is built from the resno (i.e. tlist index) of the view* tlist to the corresponding attribute number of the base relation* the varattno points to.** Must not be called with a targetlist containing non-Var entries.*/
static void
gen_view_base_attr_map(List *viewtlist, AttrNumber * attnomap, int
targetvarno)
{   ListCell    *lc;   TargetEntry *te;   Var         *tev;   int         l_viewtlist = list_length(viewtlist);
   foreach(lc, viewtlist)   {       te = (TargetEntry*) lfirst(lc);       /* Could relax this in future and map only the var entries,        * ignoring everything else, but currently pointless since we        * are only interested in simple views. */       Assert(IsA(te->expr, Var));       tev = (Var*) te->expr;       Assert(tev->varno == targetvarno);       Assert(te->resno - 1 < l_viewtlist);       attnomap[te->resno - 1] = tev->varattno;   }
}


producing a forward mapping of view attno to base relation attno.



I then apply the varattno remapping, in this case to the returning list
of a DML query acting on a view, with something like:
   varattno_map = palloc( list_length(viewquery->targetList) *                          sizeof(AttrNumber) );
   gen_view_base_attr_map(viewquery->targetList,                          varattno_map, rtr->rtindex);
   parsetree->returningList = map_variable_attnos(           (Node*) parsetree->returningList,           old_result_rt_index, 0,           varattno_map, list_length(viewquery->targetList),           &found_whole_row_var           );
   if (found_whole_row_var)   {       /* TODO: Extend map_variable_attnos API to          pass a mutator to handle whole-row vars. */       elog(ERROR, "RETURNING list contains a whole-row variable, "                   "which is not currently supported for updatable "                   "views");   }
   ChangeVarNodes((Node*) parsetree->returningList,                  old_result_rt_index,                  new_result_rt_index,                  0);


I'd prefer to be doing the map_variable_attnos and ChangeVarNodes work
in a single pass, but it looks like a clumsy and verbose process to
write a new walker. So I'm going to leave it as an "opportunity for
future optimisation" for now ;-)




(As it happens that "found_whole_row_var" is a real pain; I'm going to
have to deal with a TODO item in map_variable_attnos to provide a
callback that replaces a whole-row Var with an expansion of it into a
row-expression).


-- Craig Ringer                   http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services