Fix handling of phrase operator removal while removing tsquery stopwords.

The distance of a removed phrase operator should propagate up to a
parent phrase operator if there is one, but this only worked correctly
in left-deep trees.  Throwing in a few parentheses confused it completely,
as indeed was illustrated by bizarre results in existing regression test
cases.

To fix, track unaccounted-for distances that should propagate to the left
and to the right of the current node, rather than trying to make it work
with only one returned distance.

Also make some adjustments to behave as well as we can for cases of
intermixed phrase and regular (AND/OR) operators.  I don't think it's
possible to be 100% correct for that without a rethinking of the tsquery
representation; for example, maybe we should just not drop stopword nodes
at all underneath phrase operators.  But this is better than it was,
and changing tsquery representation wouldn't be safely back-patchable.

While at it, I simplified the API of the clean_fakeval_intree function
a bit by getting rid of the "char *result" output parameter; that wasn't
doing anything that wasn't redundant with whether the result node is
NULL or not, and testing for NULL seems a lot clearer/safer.

This is part of a larger project to fix various infelicities in the
phrase-search implementation, but this part seems comittable on its own.

Back-patch to 9.6 where phrase operators were introduced.

Discussion: https://postgr.es/m/28215.1481999808@sss.pgh.pa.us
Discussion: https://postgr.es/m/26706.1482087250@sss.pgh.pa.us
This commit is contained in:
Tom Lane 2016-12-19 13:49:45 -05:00
parent dd728826c5
commit 2604438472
2 changed files with 119 additions and 82 deletions

View File

@ -207,45 +207,59 @@ clean_NOT(QueryItem *ptr, int *len)
}
#ifdef V_UNKNOWN /* exists in Windows headers */
#undef V_UNKNOWN
#endif
#ifdef V_FALSE /* exists in Solaris headers */
#undef V_FALSE
#endif
/*
* output values for result output parameter of clean_fakeval_intree
*/
#define V_UNKNOWN 0 /* the expression can't be evaluated
* statically */
#define V_TRUE 1 /* the expression is always true (not
* implemented) */
#define V_FALSE 2 /* the expression is always false (not
* implemented) */
#define V_STOP 3 /* the expression is a stop word */
/*
* Remove QI_VALSTOP (stopword nodes) from query tree.
* Remove QI_VALSTOP (stopword) nodes from query tree.
*
* Returns NULL if the query degenerates to nothing. Input must not be NULL.
*
* When we remove a phrase operator due to removing one or both of its
* arguments, we might need to adjust the distance of a parent phrase
* operator. For example, 'a' is a stopword, so:
* (b <-> a) <-> c should become b <2> c
* b <-> (a <-> c) should become b <2> c
* (b <-> (a <-> a)) <-> c should become b <3> c
* b <-> ((a <-> a) <-> c) should become b <3> c
* To handle that, we define two output parameters:
* ladd: amount to add to a phrase distance to the left of this node
* radd: amount to add to a phrase distance to the right of this node
* We need two outputs because we could need to bubble up adjustments to two
* different parent phrase operators. Consider
* w <-> (((a <-> x) <2> (y <3> a)) <-> z)
* After we've removed the two a's and are considering the <2> node (which is
* now just x <2> y), we have an ladd distance of 1 that needs to propagate
* up to the topmost (leftmost) <->, and an radd distance of 3 that needs to
* propagate to the rightmost <->, so that we'll end up with
* w <2> ((x <2> y) <4> z)
* Near the bottom of the tree, we may have subtrees consisting only of
* stopwords. The distances of any phrase operators within such a subtree are
* summed and propagated to both ladd and radd, since we don't know which side
* of the lowest surviving phrase operator we are in. The rule is that any
* subtree that degenerates to NULL must return equal values of ladd and radd,
* and the parent node dealing with it should incorporate only one of those.
*
* Currently, we only implement this adjustment for adjacent phrase operators.
* Thus for example 'x <-> ((a <-> y) | z)' will become 'x <-> (y | z)', which
* isn't ideal, but there is no way to represent the really desired semantics
* without some redesign of the tsquery structure. Certainly it would not be
* any better to convert that to 'x <2> (y | z)'. Since this is such a weird
* corner case, let it go for now. But we can fix it in cases where the
* intervening non-phrase operator also gets removed, for example
* '((x <-> a) | a) <-> y' will become 'x <2> y'.
*/
static NODE *
clean_fakeval_intree(NODE *node, char *result, int *adddistance)
clean_stopword_intree(NODE *node, int *ladd, int *radd)
{
char lresult = V_UNKNOWN,
rresult = V_UNKNOWN;
/* since this function recurses, it could be driven to stack overflow. */
check_stack_depth();
if (adddistance)
*adddistance = 0;
/* default output parameters indicate no change in parent distance */
*ladd = *radd = 0;
if (node->valnode->type == QI_VAL)
return node;
else if (node->valnode->type == QI_VALSTOP)
{
pfree(node);
*result = V_STOP;
return NULL;
}
@ -253,10 +267,10 @@ clean_fakeval_intree(NODE *node, char *result, int *adddistance)
if (node->valnode->qoperator.oper == OP_NOT)
{
node->right = clean_fakeval_intree(node->right, &rresult, NULL);
/* NOT doesn't change pattern width, so just report child distances */
node->right = clean_stopword_intree(node->right, ladd, radd);
if (!node->right)
{
*result = V_STOP;
freetree(node);
return NULL;
}
@ -264,67 +278,89 @@ clean_fakeval_intree(NODE *node, char *result, int *adddistance)
else
{
NODE *res = node;
bool isphrase;
int ndistance,
ldistance = 0,
rdistance = 0;
lladd,
lradd,
rladd,
rradd;
ndistance = (node->valnode->qoperator.oper == OP_PHRASE) ?
node->valnode->qoperator.distance :
0;
/* First, recurse */
node->left = clean_stopword_intree(node->left, &lladd, &lradd);
node->right = clean_stopword_intree(node->right, &rladd, &rradd);
node->left = clean_fakeval_intree(node->left,
&lresult,
ndistance ? &ldistance : NULL);
/* Check if current node is OP_PHRASE, get its distance */
isphrase = (node->valnode->qoperator.oper == OP_PHRASE);
ndistance = isphrase ? node->valnode->qoperator.distance : 0;
node->right = clean_fakeval_intree(node->right,
&rresult,
ndistance ? &rdistance : NULL);
/*
* ndistance, ldistance and rdistance are greater than zero if their
* corresponding nodes are OP_PHRASE
*/
if (lresult == V_STOP && rresult == V_STOP)
if (node->left == NULL && node->right == NULL)
{
if (adddistance && ndistance)
*adddistance = ldistance + ndistance + rdistance;
/*
* When we collapse out a phrase node entirely, propagate its own
* distance into both *ladd and *radd; it is the responsibility of
* the parent node to count it only once. Also, for a phrase
* node, distances coming from children are summed and propagated
* up to parent (we assume lladd == lradd and rladd == rradd, else
* rule was broken at a lower level). But if this isn't a phrase
* node, take the larger of the two child distances; that
* corresponds to what TS_execute will do in non-stopword cases.
*/
if (isphrase)
*ladd = *radd = lladd + ndistance + rladd;
else
*ladd = *radd = Max(lladd, rladd);
freetree(node);
*result = V_STOP;
return NULL;
}
else if (lresult == V_STOP)
else if (node->left == NULL)
{
/* Removing this operator and left subnode */
/* lladd and lradd are equal/redundant, don't count both */
if (isphrase)
{
/* operator's own distance must propagate to left */
*ladd = lladd + ndistance + rladd;
*radd = rradd;
}
else
{
/* at non-phrase op, just forget the left subnode entirely */
*ladd = rladd;
*radd = rradd;
}
res = node->right;
/*
* propagate distance from current node to the right upper
* subtree.
*/
if (adddistance && ndistance)
*adddistance = rdistance;
pfree(node);
}
else if (rresult == V_STOP)
else if (node->right == NULL)
{
/* Removing this operator and right subnode */
/* rladd and rradd are equal/redundant, don't count both */
if (isphrase)
{
/* operator's own distance must propagate to right */
*ladd = lladd;
*radd = lradd + ndistance + rradd;
}
else
{
/* at non-phrase op, just forget the right subnode entirely */
*ladd = lladd;
*radd = lradd;
}
res = node->left;
/*
* propagate distance from current node to the upper tree.
*/
if (adddistance && ndistance)
*adddistance = ndistance + ldistance;
pfree(node);
}
else if (ndistance)
else if (isphrase)
{
node->valnode->qoperator.distance += ldistance;
if (adddistance)
*adddistance = 0;
/* Absorb appropriate corrections at this level */
node->valnode->qoperator.distance += lradd + rladd;
/* Propagate up any unaccounted-for corrections */
*ladd = lladd;
*radd = rradd;
}
else if (adddistance)
else
{
*adddistance = 0;
/* We're keeping a non-phrase operator, so ladd/radd remain 0 */
}
return res;
@ -585,7 +621,8 @@ cleanup_fakeval_and_phrase(TSQuery in)
commonlen,
i;
NODE *root;
char result = V_UNKNOWN;
int ladd,
radd;
TSQuery out;
QueryItem *items;
char *operands;
@ -594,8 +631,8 @@ cleanup_fakeval_and_phrase(TSQuery in)
return in;
/* eliminate stop words */
root = clean_fakeval_intree(maketree(GETQUERY(in)), &result, NULL);
if (result != V_UNKNOWN)
root = clean_stopword_intree(maketree(GETQUERY(in)), &ladd, &radd);
if (root == NULL)
{
ereport(NOTICE,
(errmsg("text-search query contains only stop words or doesn't contain lexemes, ignored")));

View File

@ -594,7 +594,7 @@ SELECT to_tsquery('english', 'a <-> (1 <-> 2)');
SELECT to_tsquery('english', '1 <-> (a <-> 2)');
to_tsquery
-------------
'1' <-> '2'
'1' <2> '2'
(1 row)
SELECT to_tsquery('english', '1 <-> (2 <-> a)');
@ -630,7 +630,7 @@ SELECT to_tsquery('english', 'a <3> (1 <-> 2)');
SELECT to_tsquery('english', '1 <3> (a <-> 2)');
to_tsquery
-------------
'1' <3> '2'
'1' <4> '2'
(1 row)
SELECT to_tsquery('english', '1 <3> (2 <-> a)');
@ -666,7 +666,7 @@ SELECT to_tsquery('english', 'a <-> (1 <3> 2)');
SELECT to_tsquery('english', '1 <-> (a <3> 2)');
to_tsquery
-------------
'1' <-> '2'
'1' <4> '2'
(1 row)
SELECT to_tsquery('english', '1 <-> (2 <3> a)');
@ -684,7 +684,7 @@ SELECT to_tsquery('english', '((a <-> 1) <-> 2) <-> s');
SELECT to_tsquery('english', '(2 <-> (a <-> 1)) <-> s');
to_tsquery
-------------
'2' <-> '1'
'2' <2> '1'
(1 row)
SELECT to_tsquery('english', '((1 <-> a) <-> 2) <-> s');
@ -708,7 +708,7 @@ SELECT to_tsquery('english', 's <-> ((a <-> 1) <-> 2)');
SELECT to_tsquery('english', 's <-> (2 <-> (a <-> 1))');
to_tsquery
-------------
'2' <-> '1'
'2' <2> '1'
(1 row)
SELECT to_tsquery('english', 's <-> ((1 <-> a) <-> 2)');
@ -750,13 +750,13 @@ SELECT to_tsquery('english', '(s <-> (1 <-> a)) <-> 2');
SELECT to_tsquery('english', '2 <-> ((a <-> 1) <-> s)');
to_tsquery
-------------
'2' <-> '1'
'2' <2> '1'
(1 row)
SELECT to_tsquery('english', '2 <-> (s <-> (a <-> 1))');
to_tsquery
-------------
'2' <-> '1'
'2' <3> '1'
(1 row)
SELECT to_tsquery('english', '2 <-> ((1 <-> a) <-> s)');
@ -768,13 +768,13 @@ SELECT to_tsquery('english', '2 <-> ((1 <-> a) <-> s)');
SELECT to_tsquery('english', '2 <-> (s <-> (1 <-> a))');
to_tsquery
-------------
'2' <-> '1'
'2' <2> '1'
(1 row)
SELECT to_tsquery('english', 'foo <-> (a <-> (the <-> bar))');
to_tsquery
-----------------
'foo' <-> 'bar'
'foo' <3> 'bar'
(1 row)
SELECT to_tsquery('english', '((foo <-> a) <-> the) <-> bar');