Re: [HACKERS] Identifying no-op length coercions
On Tue, Jun 21, 2011 at 5:50 PM, Alexey Klyukin wrote: > On Jun 21, 2011, at 9:58 PM, Noah Misch wrote: >> A pg_regress test needs stable output, so we would do it roughly like this: >> >> CREATE TEMP TABLE relstorage AS SELECT 0::regclass AS oldnode; >> ... >> UPDATE relstorage SET oldnode = >> (SELECT relfilenode FROM pg_class WHERE oid = >> 'test'::regclass); >> ALTER TABLE test ALTER name TYPE varchar(65535); >> SELECT oldnode <> relfilenode AS rewritten >> FROM pg_class, relstorage WHERE oid = 'test'::regclass; >> >> I originally rejected that as too ugly to read. Perhaps not. > > Yes, your example is more appropriate. I think you can make it more > straightforward by getting rid of the temp table: > > CREATE TABLE test(oldnode oid, name varchar(5)); > > INSERT INTO test(oldnode) SELECT relfilenode FROM pg_class WHERE > oid='test'::regclass; > > ALTER TABLE test ALTER name TYPE varchar(10); > > SELECT oldnode <> relfilenode AS rewritten FROM pg_class, test WHERE > oid='test'::regclass; Without wishing to foreclose the possibility of adding a suitable regression test, I've committed the main patch. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Jun 21, 2011, at 9:58 PM, Noah Misch wrote: > > A pg_regress test needs stable output, so we would do it roughly like this: > > CREATE TEMP TABLE relstorage AS SELECT 0::regclass AS oldnode; > ... > UPDATE relstorage SET oldnode = > (SELECT relfilenode FROM pg_class WHERE oid = 'test'::regclass); > ALTER TABLE test ALTER name TYPE varchar(65535); > SELECT oldnode <> relfilenode AS rewritten > FROM pg_class, relstorage WHERE oid = 'test'::regclass; > > I originally rejected that as too ugly to read. Perhaps not. Yes, your example is more appropriate. I think you can make it more straightforward by getting rid of the temp table: CREATE TABLE test(oldnode oid, name varchar(5)); INSERT INTO test(oldnode) SELECT relfilenode FROM pg_class WHERE oid='test'::regclass; ALTER TABLE test ALTER name TYPE varchar(10); SELECT oldnode <> relfilenode AS rewritten FROM pg_class, test WHERE oid='test'::regclass; > >> The only nitpick code-wise is these lines in varchar_transform: >> >> +int32 old_max = exprTypmod(source) - VARHDRSZ; >> +int32 new_max = new_typmod - VARHDRSZ; >> >> I have a hard time understanding why VARHDRSZ is subtracted here, so I'd >> assume that's a bug. > > We track the varchar typmod internally as (max length) + VARHDRSZ. Oh, right, haven't thought that this is a varchar specific thing. Thank you, Alexey. -- Command Prompt, Inc. http://www.CommandPrompt.com PostgreSQL Replication, Consulting, Custom Development, 24x7 support -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Tue, Jun 21, 2011 at 06:31:44PM +0300, Alexey Klyukin wrote: > Here is my review of this patch. Thanks! > The patch applies cleanly to the HEAD, produces no additional warnings. It > doesn't include additional regression tests. One can include a test, using the > commands like the ones included below. > > Changes to the documentation are limited to the a description of a new > pg_class attribute. It would probably make sense to document all the > exceptions to the table's rewrite on ALTER TABLE documentation page, although > it could wait for more such exceptions. I like the current level of detail in the ALTER TABLE page. Having EXPLAIN ALTER TABLE would help here, but that's a bigger project than this one. > postgres=# alter table test alter column name type varchar(10); > ALTER TABLE > postgres=# select relfilenode from pg_class where oid='test'::regclass; > relfilenode > - >66308 > (1 row) > postgres=# alter table test alter column name type varchar(65535); > ALTER TABLE > postgres=# select relfilenode from pg_class where oid='test'::regclass; > relfilenode > - >66308 > (1 row) A pg_regress test needs stable output, so we would do it roughly like this: CREATE TEMP TABLE relstorage AS SELECT 0::regclass AS oldnode; ... UPDATE relstorage SET oldnode = (SELECT relfilenode FROM pg_class WHERE oid = 'test'::regclass); ALTER TABLE test ALTER name TYPE varchar(65535); SELECT oldnode <> relfilenode AS rewritten FROM pg_class, relstorage WHERE oid = 'test'::regclass; I originally rejected that as too ugly to read. Perhaps not. > The only nitpick code-wise is these lines in varchar_transform: > > + int32 old_max = exprTypmod(source) - VARHDRSZ; > + int32 new_max = new_typmod - VARHDRSZ; > > I have a hard time understanding why VARHDRSZ is subtracted here, so I'd > assume that's a bug. We track the varchar typmod internally as (max length) + VARHDRSZ. -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Hi, On Jun 19, 2011, at 2:10 PM, Noah Misch wrote: > On Sat, Jun 18, 2011 at 11:32:20PM -0400, Robert Haas wrote: >> On Sat, Jun 18, 2011 at 11:12 PM, Robert Haas wrote: >>> On Sat, Jun 18, 2011 at 11:06 PM, Noah Misch wrote: On Sat, Jun 18, 2011 at 10:57:13PM -0400, Robert Haas wrote: > On Sat, Jun 11, 2011 at 5:13 PM, Noah Misch wrote: >> Sounds good. ?Updated patch attached, incorporating your ideas. ?Before >> applying >> it, run this command to update the uninvolved pg_proc.h DATA entries: >> ?perl -pi -e 's/PGUID(\s+\d+){4}/$& 0/' src/include/catalog/pg_proc.h > > This doesn't quite apply any more. ?I think the pgindent run broke it > slightly. Hmm, I just get two one-line offsets when applying it to current master. ?Note that you need to run the perl invocation before applying the patch. ?Could you provide full output of your `patch' invocation, along with any reject files? >>> >>> Ah, crap. ?You're right. ?I didn't follow your directions for how to >>> apply the patch. ?Sorry. >> >> I think you need to update the comment in simplify_function() to say >> that we have three strategies, rather than two. I think it would also >> be appropriate to add a longish comment just before the test that >> calls protransform, explaining what the charter of that function is >> and why the mechanism exists. > > Good idea. See attached. > >> Documentation issues aside, I see very little not to like about this. > > Great! Thanks for reviewing. > Here is my review of this patch. The patch applies cleanly to the HEAD, produces no additional warnings. It doesn't include additional regression tests. One can include a test, using the commands like the ones included below. Changes to the documentation are limited to the a description of a new pg_class attribute. It would probably make sense to document all the exceptions to the table's rewrite on ALTER TABLE documentation page, although it could wait for more such exceptions. The feature works as intended and I haven't noticed any crashes or assertion failures, i.e. postgres=# create table test(id integer, name varchar); CREATE TABLE postgres=# insert into test values(1, 'test'); INSERT 0 1 postgres=# select relfilenode from pg_class where oid='test'::regclass; relfilenode - 66302 (1 row) postgres=# alter table test alter column name type varchar(10); ALTER TABLE postgres=# select relfilenode from pg_class where oid='test'::regclass; relfilenode - 66308 (1 row) postgres=# alter table test alter column name type varchar(65535); ALTER TABLE postgres=# select relfilenode from pg_class where oid='test'::regclass; relfilenode - 66308 (1 row) The code looks good and takes into account all the previous suggestions. The only nitpick code-wise is these lines in varchar_transform: + int32 old_max = exprTypmod(source) - VARHDRSZ; + int32 new_max = new_typmod - VARHDRSZ; I have a hard time understanding why VARHDRSZ is subtracted here, so I'd assume that's a bug. Other than that, I haven't noticed any issues w/ this patch. Sincerely, Alexey. -- Command Prompt, Inc. http://www.CommandPrompt.com PostgreSQL Replication, Consulting, Custom Development, 24x7 support -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Sat, Jun 18, 2011 at 11:32:20PM -0400, Robert Haas wrote: > On Sat, Jun 18, 2011 at 11:12 PM, Robert Haas wrote: > > On Sat, Jun 18, 2011 at 11:06 PM, Noah Misch wrote: > >> On Sat, Jun 18, 2011 at 10:57:13PM -0400, Robert Haas wrote: > >>> On Sat, Jun 11, 2011 at 5:13 PM, Noah Misch wrote: > >>> > Sounds good. ?Updated patch attached, incorporating your ideas. ?Before > >>> > applying > >>> > it, run this command to update the uninvolved pg_proc.h DATA entries: > >>> > ?perl -pi -e 's/PGUID(\s+\d+){4}/$& 0/' src/include/catalog/pg_proc.h > >>> > >>> This doesn't quite apply any more. ?I think the pgindent run broke it > >>> slightly. > >> > >> Hmm, I just get two one-line offsets when applying it to current master. > >> ?Note > >> that you need to run the perl invocation before applying the patch. ?Could > >> you > >> provide full output of your `patch' invocation, along with any reject > >> files? > > > > Ah, crap. ?You're right. ?I didn't follow your directions for how to > > apply the patch. ?Sorry. > > I think you need to update the comment in simplify_function() to say > that we have three strategies, rather than two. I think it would also > be appropriate to add a longish comment just before the test that > calls protransform, explaining what the charter of that function is > and why the mechanism exists. Good idea. See attached. > Documentation issues aside, I see very little not to like about this. Great! Thanks for reviewing. diff --git a/doc/src/sgml/catalogs.sgml b/doc/src/sgml/catalogs.sgml index 24d7d98..7a380ce 100644 *** a/doc/src/sgml/catalogs.sgml --- b/doc/src/sgml/catalogs.sgml *** *** 4338,4343 --- 4338,4350 + protransform + regproc + pg_proc.oid + Calls to function can be simplified by this other function + + + proisagg bool diff --git a/src/backend/catalog/index 6250b07..92be0a7 100644 *** a/src/backend/catalog/pg_proc.c --- b/src/backend/catalog/pg_proc.c *** *** 304,309 ProcedureCreate(const char *procedureName, --- 304,310 values[Anum_pg_proc_procost - 1] = Float4GetDatum(procost); values[Anum_pg_proc_prorows - 1] = Float4GetDatum(prorows); values[Anum_pg_proc_provariadic - 1] = ObjectIdGetDatum(variadicType); + values[Anum_pg_proc_protransform - 1] = ObjectIdGetDatum(InvalidOid); values[Anum_pg_proc_proisagg - 1] = BoolGetDatum(isAgg); values[Anum_pg_proc_proiswindow - 1] = BoolGetDatum(isWindowFunc); values[Anum_pg_proc_prosecdef - 1] = BoolGetDatum(security_definer); diff --git a/src/backend/commands/taindex 912f45c..4dffd64 100644 *** a/src/backend/commands/tablecmds.c --- b/src/backend/commands/tablecmds.c *** *** 56,61 --- 56,62 #include "nodes/nodeFuncs.h" #include "nodes/parsenodes.h" #include "optimizer/clauses.h" + #include "optimizer/planner.h" #include "parser/parse_clause.h" #include "parser/parse_coerce.h" #include "parser/parse_collate.h" *** *** 3495,3501 ATRewriteTable(AlteredTableInfo *tab, Oid OIDNewHeap, LOCKMODE lockmode) { NewColumnValue *ex = lfirst(l); ! ex->exprstate = ExecPrepareExpr((Expr *) ex->expr, estate); } notnull_attrs = NIL; --- 3496,3503 { NewColumnValue *ex = lfirst(l); ! /* expr already planned */ ! ex->exprstate = ExecInitExpr((Expr *) ex->expr, NULL); } notnull_attrs = NIL; *** *** 4398,4404 ATExecAddColumn(List **wqueue, AlteredTableInfo *tab, Relation rel, newval = (NewColumnValue *) palloc0(sizeof(NewColumnValue)); newval->attnum = attribute.attnum; ! newval->expr = defval; tab->newvals = lappend(tab->newvals, newval); tab->rewrite = true; --- 4400,4406 newval = (NewColumnValue *) palloc0(sizeof(NewColumnValue)); newval->attnum = attribute.attnum; ! newval->expr = expression_planner(defval); tab->newvals = lappend(tab->newvals, newval); tab->rewrite = true; *** *** 6707,6712 ATPrepAlterColumnType(List **wqueue, --- 6709,6717 /* Fix collations after all else */ assign_expr_collations(pstate, transform); + /* Plan the expr now so we can accurately assess the need to rewrite. */ + transform = (Node *) expression_planner((Expr *) transform); + /* * Add a work queue item to make ATRewriteTable update the column * contents. diff --git a/src/backend/optimizer/utilindex 2914c39..be0935d 100644 *** a/src/backend/optimizer/util/clauses.c --- b/src/backend/optimize
Re: [HACKERS] Identifying no-op length coercions
On Sat, Jun 18, 2011 at 11:12 PM, Robert Haas wrote: > On Sat, Jun 18, 2011 at 11:06 PM, Noah Misch wrote: >> On Sat, Jun 18, 2011 at 10:57:13PM -0400, Robert Haas wrote: >>> On Sat, Jun 11, 2011 at 5:13 PM, Noah Misch wrote: >>> > Sounds good. ?Updated patch attached, incorporating your ideas. ?Before >>> > applying >>> > it, run this command to update the uninvolved pg_proc.h DATA entries: >>> > ?perl -pi -e 's/PGUID(\s+\d+){4}/$& 0/' src/include/catalog/pg_proc.h >>> >>> This doesn't quite apply any more. I think the pgindent run broke it >>> slightly. >> >> Hmm, I just get two one-line offsets when applying it to current master. >> Note >> that you need to run the perl invocation before applying the patch. Could >> you >> provide full output of your `patch' invocation, along with any reject files? > > Ah, crap. You're right. I didn't follow your directions for how to > apply the patch. Sorry. I think you need to update the comment in simplify_function() to say that we have three strategies, rather than two. I think it would also be appropriate to add a longish comment just before the test that calls protransform, explaining what the charter of that function is and why the mechanism exists. Documentation issues aside, I see very little not to like about this. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Sat, Jun 18, 2011 at 11:06 PM, Noah Misch wrote: > On Sat, Jun 18, 2011 at 10:57:13PM -0400, Robert Haas wrote: >> On Sat, Jun 11, 2011 at 5:13 PM, Noah Misch wrote: >> > Sounds good. ?Updated patch attached, incorporating your ideas. ?Before >> > applying >> > it, run this command to update the uninvolved pg_proc.h DATA entries: >> > ?perl -pi -e 's/PGUID(\s+\d+){4}/$& 0/' src/include/catalog/pg_proc.h >> >> This doesn't quite apply any more. I think the pgindent run broke it >> slightly. > > Hmm, I just get two one-line offsets when applying it to current master. Note > that you need to run the perl invocation before applying the patch. Could you > provide full output of your `patch' invocation, along with any reject files? Ah, crap. You're right. I didn't follow your directions for how to apply the patch. Sorry. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Sat, Jun 18, 2011 at 10:57:13PM -0400, Robert Haas wrote: > On Sat, Jun 11, 2011 at 5:13 PM, Noah Misch wrote: > > Sounds good. ?Updated patch attached, incorporating your ideas. ?Before > > applying > > it, run this command to update the uninvolved pg_proc.h DATA entries: > > ?perl -pi -e 's/PGUID(\s+\d+){4}/$& 0/' src/include/catalog/pg_proc.h > > This doesn't quite apply any more. I think the pgindent run broke it > slightly. Hmm, I just get two one-line offsets when applying it to current master. Note that you need to run the perl invocation before applying the patch. Could you provide full output of your `patch' invocation, along with any reject files? Thanks, nm -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Sat, Jun 11, 2011 at 5:13 PM, Noah Misch wrote: > Sounds good. Updated patch attached, incorporating your ideas. Before > applying > it, run this command to update the uninvolved pg_proc.h DATA entries: > perl -pi -e 's/PGUID(\s+\d+){4}/$& 0/' src/include/catalog/pg_proc.h This doesn't quite apply any more. I think the pgindent run broke it slightly. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Sat, Jun 11, 2011 at 03:03:18PM -0400, Tom Lane wrote: > Noah Misch writes: > > Good points. I'm thinking, then, add an Expr argument to > > simplify_function() > > and have the CoerceViaIO branch of eval_const_expressions_mutator() pass > > NULL > > for both its simplify_function() calls. If simplify_function() gets a NULL > > Expr > > for a function that has a protransform, synthesize a FuncExpr based on its > > other > > arguments before calling the transform function. Seem reasonable? Granted, > > that would be dead code until someone applies a transform function to a > > type I/O > > function, which could easily never happen. Perhaps just ignore the > > transform > > function when we started with a CoerceViaIO node? > > Until we actually have a use-case for simplifying I/O functions like this, > I can't see going out of our way for it ... Sounds good. Updated patch attached, incorporating your ideas. Before applying it, run this command to update the uninvolved pg_proc.h DATA entries: perl -pi -e 's/PGUID(\s+\d+){4}/$& 0/' src/include/catalog/pg_proc.h Thanks, nm diff --git a/doc/src/sgml/catalogs.sgml b/doc/src/sgml/catalogs.sgml index 8504555..5d1a447 100644 *** a/doc/src/sgml/catalogs.sgml --- b/doc/src/sgml/catalogs.sgml *** *** 4337,4342 --- 4337,4349 + protransform + regproc + pg_proc.oid + Calls to function can be simplified by this other function + + + proisagg bool diff --git a/src/backend/catalog/index 6250b07..92be0a7 100644 *** a/src/backend/catalog/pg_proc.c --- b/src/backend/catalog/pg_proc.c *** *** 304,309 ProcedureCreate(const char *procedureName, --- 304,310 values[Anum_pg_proc_procost - 1] = Float4GetDatum(procost); values[Anum_pg_proc_prorows - 1] = Float4GetDatum(prorows); values[Anum_pg_proc_provariadic - 1] = ObjectIdGetDatum(variadicType); + values[Anum_pg_proc_protransform - 1] = ObjectIdGetDatum(InvalidOid); values[Anum_pg_proc_proisagg - 1] = BoolGetDatum(isAgg); values[Anum_pg_proc_proiswindow - 1] = BoolGetDatum(isWindowFunc); values[Anum_pg_proc_prosecdef - 1] = BoolGetDatum(security_definer); diff --git a/src/backend/commands/taindex 2c9f855..563a1b2 100644 *** a/src/backend/commands/tablecmds.c --- b/src/backend/commands/tablecmds.c *** *** 56,61 --- 56,62 #include "nodes/nodeFuncs.h" #include "nodes/parsenodes.h" #include "optimizer/clauses.h" + #include "optimizer/planner.h" #include "parser/parse_clause.h" #include "parser/parse_coerce.h" #include "parser/parse_collate.h" *** *** 3495,3501 ATRewriteTable(AlteredTableInfo *tab, Oid OIDNewHeap, LOCKMODE lockmode) { NewColumnValue *ex = lfirst(l); ! ex->exprstate = ExecPrepareExpr((Expr *) ex->expr, estate); } notnull_attrs = NIL; --- 3496,3503 { NewColumnValue *ex = lfirst(l); ! /* expr already planned */ ! ex->exprstate = ExecInitExpr((Expr *) ex->expr, NULL); } notnull_attrs = NIL; *** *** 4398,4404 ATExecAddColumn(List **wqueue, AlteredTableInfo *tab, Relation rel, newval = (NewColumnValue *) palloc0(sizeof(NewColumnValue)); newval->attnum = attribute.attnum; ! newval->expr = defval; tab->newvals = lappend(tab->newvals, newval); tab->rewrite = true; --- 4400,4406 newval = (NewColumnValue *) palloc0(sizeof(NewColumnValue)); newval->attnum = attribute.attnum; ! newval->expr = expression_planner(defval); tab->newvals = lappend(tab->newvals, newval); tab->rewrite = true; *** *** 6706,6711 ATPrepAlterColumnType(List **wqueue, --- 6708,6716 /* Fix collations after all else */ assign_expr_collations(pstate, transform); + /* Plan the expr now so we can accurately assess the need to rewrite. */ + transform = (Node *) expression_planner((Expr *) transform); + /* * Add a work queue item to make ATRewriteTable update the column * contents. diff --git a/src/backend/optimizer/utilindex 2914c39..e9fb750 100644 *** a/src/backend/optimizer/util/clauses.c --- b/src/backend/optimizer/util/clauses.c *** *** 106,114 static List *simplify_and_arguments(List *args, eval_const_expressions_context *context, bool *haveNull, bool *forceFalse); static Node *simplify_boolean_equality(Oid opno, List *args); ! static Expr *simplify_function(Oid funcid, !
Re: [HACKERS] Identifying no-op length coercions
Noah Misch writes: > Good points. I'm thinking, then, add an Expr argument to simplify_function() > and have the CoerceViaIO branch of eval_const_expressions_mutator() pass NULL > for both its simplify_function() calls. If simplify_function() gets a NULL > Expr > for a function that has a protransform, synthesize a FuncExpr based on its > other > arguments before calling the transform function. Seem reasonable? Granted, > that would be dead code until someone applies a transform function to a type > I/O > function, which could easily never happen. Perhaps just ignore the transform > function when we started with a CoerceViaIO node? Until we actually have a use-case for simplifying I/O functions like this, I can't see going out of our way for it ... regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Sat, Jun 11, 2011 at 02:11:55PM -0400, Tom Lane wrote: > Noah Misch writes: > > We originally discussed having the transform function take and return Expr > > nodes. It turns out that simplify_function() does not have the Expr, > > probably > > because the particular choice of node type among the many that can convey a > > function call does not matter to it. The same should be true of a transform > > function -- it should do the same thing whether the subject call arrives > > from > > a FuncExpr or an OpExpr. Therefore, I changed the transform function > > signature to "Expr *protransform(List *args)". > > That seems like the wrong thing. For one thing, it makes it quite > impossible to share one transform support function among multiple > functions/operators, since it won't know which one the argument list > is for. More generally, I foresee the transform needing to know > most of the other information that might be in the FuncExpr or OpExpr. > It certainly would need access to the collation, and it would probably > like to be able to copy the parse location to whatever it outputs, > and so forth and so on. So I really think passing the node to the > function is the most future-proof way to do it. If that means you > need to rejigger things a bit in clauses.c, so be it. Good points. I'm thinking, then, add an Expr argument to simplify_function() and have the CoerceViaIO branch of eval_const_expressions_mutator() pass NULL for both its simplify_function() calls. If simplify_function() gets a NULL Expr for a function that has a protransform, synthesize a FuncExpr based on its other arguments before calling the transform function. Seem reasonable? Granted, that would be dead code until someone applies a transform function to a type I/O function, which could easily never happen. Perhaps just ignore the transform function when we started with a CoerceViaIO node? > > The large pg_proc.h diff mostly just adds protransform = 0 to every > > function. > > Due to the resulting patch size, I've compressed it. There are > > new/otherwise > > changed DATA lines for OIDs 669 and 3097 only. > > The chances that this part will still apply cleanly by the time anyone > gets around to committing it are nil. I'd suggest you break it into two > separate patches, one that modifies the existing lines to add the > protransform = 0 column and then one to apply the remaining deltas in > the file. I envision the eventual committer doing the first step > on-the-fly (perhaps with an emacs macro, at least that's the way I > usually do it) and then wanting a patchable diff for the rest. Or if > you wanted to be really helpful, you could provide a throwaway perl > script to do the modifications of the existing lines ... That would be better; I'll do it for the next version. > I haven't read the actual patch, these are just some quick reactions > to your description. Thanks, nm -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Noah Misch writes: > We originally discussed having the transform function take and return Expr > nodes. It turns out that simplify_function() does not have the Expr, probably > because the particular choice of node type among the many that can convey a > function call does not matter to it. The same should be true of a transform > function -- it should do the same thing whether the subject call arrives from > a FuncExpr or an OpExpr. Therefore, I changed the transform function > signature to "Expr *protransform(List *args)". That seems like the wrong thing. For one thing, it makes it quite impossible to share one transform support function among multiple functions/operators, since it won't know which one the argument list is for. More generally, I foresee the transform needing to know most of the other information that might be in the FuncExpr or OpExpr. It certainly would need access to the collation, and it would probably like to be able to copy the parse location to whatever it outputs, and so forth and so on. So I really think passing the node to the function is the most future-proof way to do it. If that means you need to rejigger things a bit in clauses.c, so be it. > The large pg_proc.h diff mostly just adds protransform = 0 to every function. > Due to the resulting patch size, I've compressed it. There are new/otherwise > changed DATA lines for OIDs 669 and 3097 only. The chances that this part will still apply cleanly by the time anyone gets around to committing it are nil. I'd suggest you break it into two separate patches, one that modifies the existing lines to add the protransform = 0 column and then one to apply the remaining deltas in the file. I envision the eventual committer doing the first step on-the-fly (perhaps with an emacs macro, at least that's the way I usually do it) and then wanting a patchable diff for the rest. Or if you wanted to be really helpful, you could provide a throwaway perl script to do the modifications of the existing lines ... I haven't read the actual patch, these are just some quick reactions to your description. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Fri, Jun 03, 2011 at 10:43:17AM -0500, Jim Nasby wrote: > On Jun 3, 2011, at 10:11 AM, Alexey Klyukin wrote: > >> Is your interest in cheap varchar(N)->varchar(N+M) conversions > >> specifically, or > >> in some broader application of this facility? > > > > Exactly varchar conversions. > > Why limit it to varchar? Shouldn't we be able to do this for any varlena? The > only challenge I see is numeric; we'd need to ensure that both size and > precision are not decreasing. I've implemented support for varchar, varbit, numeric, time, timetz, timestamp, timestamptz, and interval. However, I'll probably submit only varchar in the initial infrastructure patch and the rest in followup patches in a later CF. For numeric, we store the display scale in every datum, so any change to it rewrites the table. You'll be able to cheaply change numeric(7,2) to numeric(9,2) but not to numeric(9,3). -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Fri, Jun 3, 2011 at 11:43 AM, Jim Nasby wrote: > On Jun 3, 2011, at 10:11 AM, Alexey Klyukin wrote: >>> Is your interest in cheap varchar(N)->varchar(N+M) conversions >>> specifically, or >>> in some broader application of this facility? >> >> Exactly varchar conversions. > > Why limit it to varchar? Shouldn't we be able to do this for any varlena? The > only challenge I see is numeric; we'd need to ensure that both size and > precision are not decreasing. More than that: you should also be able to make it work for things like xml -> text. Indeed, I believe Noah has plans to do just that. Which, quite frankly, will be awesome. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Jim Nasby wrote: > The only challenge I see is numeric; we'd need to ensure that both > size and precision are not decreasing. To be picky, wouldn't that need to be "neither (precision - scale) nor scale is decreasing"? -Kevin -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Jun 3, 2011, at 10:11 AM, Alexey Klyukin wrote: >> Is your interest in cheap varchar(N)->varchar(N+M) conversions specifically, >> or >> in some broader application of this facility? > > Exactly varchar conversions. Why limit it to varchar? Shouldn't we be able to do this for any varlena? The only challenge I see is numeric; we'd need to ensure that both size and precision are not decreasing. -- Jim C. Nasby, Database Architect j...@nasby.net 512.569.9461 (cell) http://jim.nasby.net -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Hi, On Jun 2, 2011, at 10:22 PM, Noah Misch wrote: > Hi Alexey, > ... > Is your interest in cheap varchar(N)->varchar(N+M) conversions specifically, > or > in some broader application of this facility? Exactly varchar conversions. > > Thanks for volunteering to review; that will be a big help. Actually, I could > especially use some feedback now on a related design and implementation: > > http://archives.postgresql.org/message-id/20110524104029.gb18...@tornado.gateway.2wire.net > Note that the third and fifth sentences of that description are incorrect. > The > rest stands without them. Even just some feedback on the mundane issue noted > in > the last paragraph would help. Ok, I'll review it. Thank you, Alexey. -- Command Prompt, Inc. http://www.CommandPrompt.com PostgreSQL Replication, Consulting, Custom Development, 24x7 support -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Hi Alexey, On Thu, Jun 02, 2011 at 05:08:51PM +0300, Alexey Klyukin wrote: > Looks like this thread has silently died out. Is there an agreement on the > syntax and implementation part? We (CMD) have a customer, who is interested in > pushing this through, so, if we have a patch, I'd be happy to assist in > reviewing it. I think we have a consensus on the implementation. We didn't totally lock down the syntax. Tom and I seem happy to have no SQL exposure at all, so that's what I'm planning to submit. However, we were pretty close to a syntax consensus in the event that it becomes desirable to do otherwise. Is your interest in cheap varchar(N)->varchar(N+M) conversions specifically, or in some broader application of this facility? Thanks for volunteering to review; that will be a big help. Actually, I could especially use some feedback now on a related design and implementation: http://archives.postgresql.org/message-id/20110524104029.gb18...@tornado.gateway.2wire.net Note that the third and fifth sentences of that description are incorrect. The rest stands without them. Even just some feedback on the mundane issue noted in the last paragraph would help. Thanks, nm -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On May 24, 2011, at 12:15 AM, Noah Misch wrote: > On Mon, May 23, 2011 at 03:01:40PM -0400, Tom Lane wrote: >> Noah Misch writes: >>> Good deal. Given that conclusion, the other policy decision I anticipate >>> affecting this particular patch is the choice of syntax. Presumably, it >>> will be >>> a new common_func_opt_item. When I last looked at the keywords list and >>> tried >>> to come up with something, these were the best I could do: >> >>> CREATE FUNCTION ... PARSER MAPPING helperfunc(args) >>> CREATE FUNCTION ... PLANS CONVERSION helperfunc(args) >> >> We could go with your previous idea of not bothering to expose this in >> the SQL syntax. Given that the helper function is going to have a >> signature along the lines of "(internal, internal) -> internal", it's >> going to be difficult for anyone to use it for non-builtin functions >> anyhow. >> >> But if you really don't like that, what about > > That would be just fine with me. We can always expose it later. > >> >> TRANSFORM helperfunctionname >> >> Although TRANSFORM isn't currently a keyword for us, it is a >> non-reserved keyword in SQL:2008, and it seems possible that we might >> someday think about implementing the associated features. > > I was thinking of that word too, along the lines of "PLAN TRANSFORM FUNCTION > helperfunctionname". Then again, that wrongly sounds somewhat like it's > transforming planner nodes. Perhaps plain TRANSFORM or TRANSFORM FUNCTION > would > be the way to go. Looks like this thread has silently died out. Is there an agreement on the syntax and implementation part? We (CMD) have a customer, who is interested in pushing this through, so, if we have a patch, I'd be happy to assist in reviewing it. -- Alexey Klyukin The PostgreSQL Company - Command Prompt, Inc. -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 03:01:40PM -0400, Tom Lane wrote: > Noah Misch writes: > > Good deal. Given that conclusion, the other policy decision I anticipate > > affecting this particular patch is the choice of syntax. Presumably, it > > will be > > a new common_func_opt_item. When I last looked at the keywords list and > > tried > > to come up with something, these were the best I could do: > > > CREATE FUNCTION ... PARSER MAPPING helperfunc(args) > > CREATE FUNCTION ... PLANS CONVERSION helperfunc(args) > > We could go with your previous idea of not bothering to expose this in > the SQL syntax. Given that the helper function is going to have a > signature along the lines of "(internal, internal) -> internal", it's > going to be difficult for anyone to use it for non-builtin functions > anyhow. > > But if you really don't like that, what about That would be just fine with me. We can always expose it later. > > TRANSFORM helperfunctionname > > Although TRANSFORM isn't currently a keyword for us, it is a > non-reserved keyword in SQL:2008, and it seems possible that we might > someday think about implementing the associated features. I was thinking of that word too, along the lines of "PLAN TRANSFORM FUNCTION helperfunctionname". Then again, that wrongly sounds somewhat like it's transforming planner nodes. Perhaps plain TRANSFORM or TRANSFORM FUNCTION would be the way to go. nm -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 02:53:01PM -0400, Robert Haas wrote: > On Mon, May 23, 2011 at 2:46 PM, Noah Misch wrote: > > On Mon, May 23, 2011 at 02:11:39PM -0400, Robert Haas wrote: > >> On Mon, May 23, 2011 at 1:53 PM, Tom Lane wrote: > >> > Maybe. ?But casts would be the least of our concerns if we were trying > >> > to change the column type. ?Changing typmod doesn't affect the set of > >> > operations that could be applied to a column, whereas changing type > >> > surely does. > >> > >> OK, this is the crucial point I was missing. ?Sorry for being a bit > >> fuzzy-headed about this. > >> > >> My mental model of our type system, or of what a type system ought to > >> do, just doesn't match the type system we've got. > >> > >> So let's do it the way you proposed. > > > > Good deal. ?Given that conclusion, the other policy decision I anticipate > > affecting this particular patch is the choice of syntax. ?Presumably, it > > will be > > a new common_func_opt_item. ?When I last looked at the keywords list and > > tried > > to come up with something, these were the best I could do: > > > > ?CREATE FUNCTION ... PARSER MAPPING helperfunc(args) > > ?CREATE FUNCTION ... PLANS CONVERSION helperfunc(args) > > > > Both feel forced, to put it generously. ?Any better ideas? ?Worth adding a > > keyword to get something decent? > > Do you have something specific in mind? I had not. Having thought about it some, maybe "PLAN TRANSFORM FUNCTION". Do we have a name for the pass over the tree rooted at eval_const_expressions()? > Just to throw out another few possibilities, how about INLINE FUNCTION > or ANALYZE FUNCTION? INLINE FUNCTION evokes, for me, having a different version of a function that gets substituted when it's inlined. ANALYZE FUNCTION seems reasonable, though. I don't think any name we'd pick will make a significant number of readers understand it without reading the full documentation. nm -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Greg Stark writes: > This was the thing that concerned me. If anyone wants to add this > feature for a new data type they're going to have to understand and > tie their code to all this internal parser node stuff. That means > their code will be much more closely tied to a specific version, will > have to be written in C, and will require much more in-depth > understanding of Postgres internal data structures. > By comparison the boolean cast predicate could be written in any > language and only required the data type implementor to understand > their data type. It seems much more likely to actually get used and be > used correctly. I don't think I believe that interesting length-conversion functions are going to be written in anything but C anyway, so I don't find that that argument holds much water. Generally, the low-level functions for a new datatype have to be in C. As for the other point, if all you want to do is examine the expression's typmod, the API for that is pretty stable (exprTypmod()). regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 12:01 PM, Tom Lane wrote: > Given that the helper function is going to have a > signature along the lines of "(internal, internal) -> internal", it's > going to be difficult for anyone to use it for non-builtin functions > anyhow. I hate to go around in circles on this but I didn't see the original discussion. This was the thing that concerned me. If anyone wants to add this feature for a new data type they're going to have to understand and tie their code to all this internal parser node stuff. That means their code will be much more closely tied to a specific version, will have to be written in C, and will require much more in-depth understanding of Postgres internal data structures. By comparison the boolean cast predicate could be written in any language and only required the data type implementor to understand their data type. It seems much more likely to actually get used and be used correctly. -- greg -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Noah Misch writes: > Good deal. Given that conclusion, the other policy decision I anticipate > affecting this particular patch is the choice of syntax. Presumably, it will > be > a new common_func_opt_item. When I last looked at the keywords list and tried > to come up with something, these were the best I could do: > CREATE FUNCTION ... PARSER MAPPING helperfunc(args) > CREATE FUNCTION ... PLANS CONVERSION helperfunc(args) We could go with your previous idea of not bothering to expose this in the SQL syntax. Given that the helper function is going to have a signature along the lines of "(internal, internal) -> internal", it's going to be difficult for anyone to use it for non-builtin functions anyhow. But if you really don't like that, what about TRANSFORM helperfunctionname Although TRANSFORM isn't currently a keyword for us, it is a non-reserved keyword in SQL:2008, and it seems possible that we might someday think about implementing the associated features. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 2:46 PM, Noah Misch wrote: > On Mon, May 23, 2011 at 02:11:39PM -0400, Robert Haas wrote: >> On Mon, May 23, 2011 at 1:53 PM, Tom Lane wrote: >> > Maybe. ?But casts would be the least of our concerns if we were trying >> > to change the column type. ?Changing typmod doesn't affect the set of >> > operations that could be applied to a column, whereas changing type >> > surely does. >> >> OK, this is the crucial point I was missing. Sorry for being a bit >> fuzzy-headed about this. >> >> My mental model of our type system, or of what a type system ought to >> do, just doesn't match the type system we've got. >> >> So let's do it the way you proposed. > > Good deal. Given that conclusion, the other policy decision I anticipate > affecting this particular patch is the choice of syntax. Presumably, it will > be > a new common_func_opt_item. When I last looked at the keywords list and tried > to come up with something, these were the best I could do: > > CREATE FUNCTION ... PARSER MAPPING helperfunc(args) > CREATE FUNCTION ... PLANS CONVERSION helperfunc(args) > > Both feel forced, to put it generously. Any better ideas? Worth adding a > keyword to get something decent? Do you have something specific in mind? Just to throw out another few possibilities, how about INLINE FUNCTION or ANALYZE FUNCTION? -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 02:11:39PM -0400, Robert Haas wrote: > On Mon, May 23, 2011 at 1:53 PM, Tom Lane wrote: > > Maybe. ?But casts would be the least of our concerns if we were trying > > to change the column type. ?Changing typmod doesn't affect the set of > > operations that could be applied to a column, whereas changing type > > surely does. > > OK, this is the crucial point I was missing. Sorry for being a bit > fuzzy-headed about this. > > My mental model of our type system, or of what a type system ought to > do, just doesn't match the type system we've got. > > So let's do it the way you proposed. Good deal. Given that conclusion, the other policy decision I anticipate affecting this particular patch is the choice of syntax. Presumably, it will be a new common_func_opt_item. When I last looked at the keywords list and tried to come up with something, these were the best I could do: CREATE FUNCTION ... PARSER MAPPING helperfunc(args) CREATE FUNCTION ... PLANS CONVERSION helperfunc(args) Both feel forced, to put it generously. Any better ideas? Worth adding a keyword to get something decent? Thanks, nm -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 1:53 PM, Tom Lane wrote: > Maybe. But casts would be the least of our concerns if we were trying > to change the column type. Changing typmod doesn't affect the set of > operations that could be applied to a column, whereas changing type > surely does. OK, this is the crucial point I was missing. Sorry for being a bit fuzzy-headed about this. My mental model of our type system, or of what a type system ought to do, just doesn't match the type system we've got. So let's do it the way you proposed. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Robert Haas writes: > On Mon, May 23, 2011 at 1:21 PM, Tom Lane wrote: >> ... But I remain of the opinion that that's the wrong place >> to put it. > So you said here: > http://archives.postgresql.org/pgsql-hackers/2011-01/msg02575.php > http://archives.postgresql.org/pgsql-hackers/2011-01/msg02585.php > The trouble is, I still can't see why type OIDs and typemods should be > handled differently. Taking your example again: > CREATE TABLE base (f1 varchar(4)); > CREATE VIEW vv AS SELECT f1::varchar(8) FROM base; > ALTER TABLE base ALTER COLUMN f1 TYPE varchar(16); > Your claim on the thread is that we want to someday allow this case. > But what if the last statement were instead: > ALTER TABLE base ALTER COLUMN f1 TYPE integer; > Should it also be our goal to handle that case? Maybe. But casts would be the least of our concerns if we were trying to change the column type. Changing typmod doesn't affect the set of operations that could be applied to a column, whereas changing type surely does. In any case, the fact that the current implementation can't readily support that is a poor excuse for building entirely new features that also can't support it, when said features could easily be designed without the restriction. But more generally, I don't believe that you've made any positive case whatever for doing it from pg_cast. It's not faster, it's not more flexible, so why should we choose that approach? regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 1:21 PM, Tom Lane wrote: > Robert Haas writes: >> On Mon, May 23, 2011 at 12:42 PM, Noah Misch wrote: >>> There were two proposals on the table: >>> >>> 1. Attach a "f(from_typmod, to_typmod, is_explicit) RETURNS boolean" >>> function >>> to the pg_cast; call it in find_coercion_pathway() >>> 2. Attach a "f(FuncExpr) RETURNS Expr" (actually internal/internal) function >>> to the pg_proc; call it in simplify_function() >>> >>> I tried and failed to write a summary of the respective arguments that could >>> legitimately substitute for (re-)reading the original thread, so I haven't >>> included one. I myself find the advantages of #2 mildly more compelling. > >> The main reason I preferred #1 is that it would only get invoked in >> the case of casts, whereas #2 would get invoked for all function >> calls. For us to pay that overhead, there has to be some use case, >> and I didn't find the examples that were offered very compelling. > > Well, as I pointed out in > http://archives.postgresql.org/pgsql-hackers/2011-01/msg02570.php > a hook function attached to pg_proc entries would cost nothing > measurable when not used. You could possibly make the same claim > for attaching the hook to pg_cast entries, if you cause the optimization > to occur during initial cast lookup rather than expression > simplification. But I remain of the opinion that that's the wrong place > to put it. So you said here: http://archives.postgresql.org/pgsql-hackers/2011-01/msg02575.php http://archives.postgresql.org/pgsql-hackers/2011-01/msg02585.php The trouble is, I still can't see why type OIDs and typemods should be handled differently. Taking your example again: CREATE TABLE base (f1 varchar(4)); CREATE VIEW vv AS SELECT f1::varchar(8) FROM base; ALTER TABLE base ALTER COLUMN f1 TYPE varchar(16); Your claim on the thread is that we want to someday allow this case. But what if the last statement were instead: ALTER TABLE base ALTER COLUMN f1 TYPE integer; Should it also be our goal to handle that case? If not, why are they different? -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
Robert Haas writes: > On Mon, May 23, 2011 at 12:42 PM, Noah Misch wrote: >> There were two proposals on the table: >> >> 1. Attach a "f(from_typmod, to_typmod, is_explicit) RETURNS boolean" function >> to the pg_cast; call it in find_coercion_pathway() >> 2. Attach a "f(FuncExpr) RETURNS Expr" (actually internal/internal) function >> to the pg_proc; call it in simplify_function() >> >> I tried and failed to write a summary of the respective arguments that could >> legitimately substitute for (re-)reading the original thread, so I haven't >> included one. I myself find the advantages of #2 mildly more compelling. > The main reason I preferred #1 is that it would only get invoked in > the case of casts, whereas #2 would get invoked for all function > calls. For us to pay that overhead, there has to be some use case, > and I didn't find the examples that were offered very compelling. Well, as I pointed out in http://archives.postgresql.org/pgsql-hackers/2011-01/msg02570.php a hook function attached to pg_proc entries would cost nothing measurable when not used. You could possibly make the same claim for attaching the hook to pg_cast entries, if you cause the optimization to occur during initial cast lookup rather than expression simplification. But I remain of the opinion that that's the wrong place to put it. > How > many CPU cycles are we willing to spend trying to simplify x + 0 to > just x, or x * 0 to just 0? I'm not sure that's worthwhile either, but it was an example of a possible future use-case rather than something that anybody was proposing doing right now. Even though I tend to agree that it wouldn't be worth looking for such cases with simple numeric datatypes, it's not that hard to believe that someone might want the ability for complicated datatypes with expensive operations. In the short term, the only actual cost we'd incur is that we'd be bloating pg_proc instead of pg_cast with an extra column, and there's more rows in pg_proc. But pg_proc rows are already pretty darn wide. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Identifying no-op length coercions
On Mon, May 23, 2011 at 12:42 PM, Noah Misch wrote: > I'd like to revive the discussion that arose during the last CommitFest over > the proper design for identifying no-op length coercions like varchar(4) -> > varchar(8). Here is the root of the original debate: > > http://archives.postgresql.org/message-id/20110109220353.gd5...@tornado.leadboat.com > > There were two proposals on the table: > > 1. Attach a "f(from_typmod, to_typmod, is_explicit) RETURNS boolean" function > to the pg_cast; call it in find_coercion_pathway() > 2. Attach a "f(FuncExpr) RETURNS Expr" (actually internal/internal) function > to the pg_proc; call it in simplify_function() > > I tried and failed to write a summary of the respective arguments that could > legitimately substitute for (re-)reading the original thread, so I haven't > included one. I myself find the advantages of #2 mildly more compelling. The main reason I preferred #1 is that it would only get invoked in the case of casts, whereas #2 would get invoked for all function calls. For us to pay that overhead, there has to be some use case, and I didn't find the examples that were offered very compelling. How many CPU cycles are we willing to spend trying to simplify x + 0 to just x, or x * 0 to just 0? I might be missing something here, but those strikes me as textbook examples of cases where it's not worth slowing down the whole system to fix a query that the user could have easily written in some less-pathological way to begin with. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers