On Sep 19, 2022, 23:14 +0800, Tom Lane <tgl@sss.pgh.pa.us>, wrote:
Very little of the planner bothers with freeing small allocations
like that.
I think so too, as said, not sure if it worths.
Can you demonstrate a case where this would actually
make a meaningful difference?
Offhand, an example may help a little:
create table t1(id int);
explain select max(id), min(id), sum(id), count(id), avg(id) from t1;
Modify codes to test:
@@ -139,6 +139,7 @@ preprocess_aggref(Aggref *aggref, PlannerInfo *root)
int16 transtypeLen;
Oid inputTypes[FUNC_MAX_ARGS];
int numArguments;
+ static size_t accumulate_list_size = 0;
Assert(aggref->agglevelsup == 0);
@@ -265,7 +266,7 @@ preprocess_aggref(Aggref *aggref, PlannerInfo *root)
aggserialfn, aggdeserialfn,
initValue, initValueIsNull,
same_input_transnos);
- list_free(same_input_transnos);
+ accumulate_list_size += sizeof(int) * list_length(same_input_transnos);
Gdb and print accumulate_list_size for each iteration:
SaveBytes = Sum results of accumulate_list_size: 32(4+4+8+8), as we have 5 aggs in sql.
If there were N sets of that aggs (more columns as id, with above aggs ), the bytes will be N*SaveBytes.
Seems we don’t have so many agg functions that could share the same trans function, Does it worth?
