Support LIKE with nondeterministic collations

This allows for example using LIKE with case-insensitive collations.
There was previously no internal implementation of this, so it was met
with a not-supported error.  This adds the internal implementation and
removes the error.  The implementation follows the specification of
the SQL standard for this.

Unlike with deterministic collations, the LIKE matching cannot go
character by character but has to go substring by substring.  For
example, if we are matching against LIKE 'foo%bar', we can't start by
looking for an 'f', then an 'o', but instead with have to find
something that matches 'foo'.  This is because the collation could
consider substrings of different lengths to be equal.  This is all
internal to MatchText() in like_match.c.

The changes in GenericMatchText() in like.c just pass through the
locale information to MatchText(), which was previously not needed.
This matches exactly Generic_Text_IC_like() below.

ILIKE is not affected.  (It's unclear whether ILIKE makes sense under
nondeterministic collations.)

This also updates match_pattern_prefix() in like_support.c to support
optimizing the case of an exact pattern with nondeterministic
collations.  This was already alluded to in the previous code.

(includes documentation examples from Daniel Vérité and test cases
from Paul A Jungwirth)

Reviewed-by: Jian He <jian.universality@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/700d2e86-bf75-4607-9cf2-f5b7802f6e88@eisentraut.org

Author: petere

doc/src/sgml/charset.sgml

@@ -1197,7 +1197,7 @@ CREATE COLLATION ignore_accents (provider = icu, locale = 'und-u-ks-level1-kc-tr
      to a performance penalty.  Note, in particular, that B-tree cannot use
      deduplication with indexes that use a nondeterministic collation.  Also,
      certain operations are not possible with nondeterministic collations,
-     such as pattern matching operations.  Therefore, they should be used
+     such as some pattern matching operations.  Therefore, they should be used
      only in cases where they are specifically wanted.
     </para>
 

doc/src/sgml/func.sgml

@@ -5414,9 +5414,10 @@ cast(-44 as bit(12))           <lineannotation>111111010100</lineannotation>
    </caution>
 
    <para>
-    The pattern matching operators of all three kinds do not support
-    nondeterministic collations.  If required, apply a different collation to
-    the expression to work around this limitation.
+    <function>SIMILAR TO</function> and <acronym>POSIX</acronym>-style regular
+    expressions do not support nondeterministic collations.  If required, use
+    <function>LIKE</function> or apply a different collation to the expression
+    to work around this limitation.
    </para>
 
   <sect2 id="functions-like">
@@ -5462,6 +5463,46 @@ cast(-44 as bit(12))           <lineannotation>111111010100</lineannotation>
 </programlisting>
    </para>
 
+   <para>
+    <function>LIKE</function> pattern matching supports nondeterministic
+    collations (see <xref linkend="collation-nondeterministic"/>), such as
+    case-insensitive collations or collations that, say, ignore punctuation.
+    So with a case-insensitive collation, one could have:
+<programlisting>
+'AbC' LIKE 'abc' COLLATE case_insensitive    <lineannotation>true</lineannotation>
+'AbC' LIKE 'a%' COLLATE case_insensitive     <lineannotation>true</lineannotation>
+</programlisting>
+    With collations that ignore certain characters or in general that consider
+    strings of different lengths equal, the semantics can become a bit more
+    complicated.  Consider these examples:
+<programlisting>
+'.foo.' LIKE 'foo' COLLATE ign_punct    <lineannotation>true</lineannotation>
+'.foo.' LIKE 'f_o' COLLATE ign_punct    <lineannotation>true</lineannotation>
+'.foo.' LIKE '_oo' COLLATE ign_punct    <lineannotation>false</lineannotation>
+</programlisting>
+    The way the matching works is that the pattern is partitioned into
+    sequences of wildcards and non-wildcard strings (wildcards being
+    <literal>_</literal> and <literal>%</literal>).  For example, the pattern
+    <literal>f_o</literal> is partitioned into <literal>f, _, o</literal>, the
+    pattern <literal>_oo</literal> is partitioned into <literal>_,
+    oo</literal>.  The input string matches the pattern if it can be
+    partitioned in such a way that the wildcards match one character or any
+    number of characters respectively and the non-wildcard partitions are
+    equal under the applicable collation.  So for example, <literal>'.foo.'
+    LIKE 'f_o' COLLATE ign_punct</literal> is true because one can partition
+    <literal>.foo.</literal> into <literal>.f, o, o.</literal>, and then
+    <literal>'.f' = 'f' COLLATE ign_punct</literal>, <literal>'o'</literal>
+    matches the <literal>_</literal> wildcard, and <literal>'o.' = 'o' COLLATE
+    ign_punct</literal>.  But <literal>'.foo.' LIKE '_oo' COLLATE
+    ign_punct</literal> is false because <literal>.foo.</literal> cannot be
+    partitioned in a way that the first character is any character and the
+    rest of the string compares equal to <literal>oo</literal>.  (Note that
+    the single-character wildcard always matches exactly one character,
+    independent of the collation.  So in this example, the
+    <literal>_</literal> would match <literal>.</literal>, but then the rest
+    of the input string won't match the rest of the pattern.)
+   </para>
+
    <para>
     <function>LIKE</function> pattern matching always covers the entire
     string.  Therefore, if it's desired to match a sequence anywhere within
@@ -5503,8 +5544,9 @@ cast(-44 as bit(12))           <lineannotation>111111010100</lineannotation>
 
    <para>
     The key word <token>ILIKE</token> can be used instead of
-    <token>LIKE</token> to make the match case-insensitive according
-    to the active locale.  This is not in the <acronym>SQL</acronym> standard but is a
+    <token>LIKE</token> to make the match case-insensitive according to the
+    active locale.  (But this does not support nondeterministic collations.)
+    This is not in the <acronym>SQL</acronym> standard but is a
     <productname>PostgreSQL</productname> extension.
    </para>
 

src/backend/utils/adt/like.c

@@ -147,22 +147,28 @@ SB_lower_char(unsigned char c, pg_locale_t locale)
 static inline int
 GenericMatchText(const char *s, int slen, const char *p, int plen, Oid collation)
 {
-	if (collation)
-	{
-		pg_locale_t locale = pg_newlocale_from_collation(collation);
+	pg_locale_t locale;
 
-		if (!locale->deterministic)
-			ereport(ERROR,
-					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-					 errmsg("nondeterministic collations are not supported for LIKE")));
+	if (!OidIsValid(collation))
+	{
+		/*
+		 * This typically means that the parser could not resolve a conflict
+		 * of implicit collations, so report it that way.
+		 */
+		ereport(ERROR,
+				(errcode(ERRCODE_INDETERMINATE_COLLATION),
+				 errmsg("could not determine which collation to use for LIKE"),
+				 errhint("Use the COLLATE clause to set the collation explicitly.")));
 	}
 
+	locale = pg_newlocale_from_collation(collation);
+
 	if (pg_database_encoding_max_length() == 1)
-		return SB_MatchText(s, slen, p, plen, 0);
+		return SB_MatchText(s, slen, p, plen, locale);
 	else if (GetDatabaseEncoding() == PG_UTF8)
-		return UTF8_MatchText(s, slen, p, plen, 0);
+		return UTF8_MatchText(s, slen, p, plen, locale);
 	else
-		return MB_MatchText(s, slen, p, plen, 0);
+		return MB_MatchText(s, slen, p, plen, locale);
 }
 
 static inline int

src/backend/utils/adt/like_match.c

@@ -157,7 +157,9 @@ MatchText(const char *t, int tlen, const char *p, int plen, pg_locale_t locale)
 			 * the first pattern byte to each text byte to avoid recursing
 			 * more than we have to.  This fact also guarantees that we don't
 			 * have to consider a match to the zero-length substring at the
-			 * end of the text.
+			 * end of the text.  With a nondeterministic collation, we can't
+			 * rely on the first bytes being equal, so we have to recurse in
+			 * any case.
 			 */
 			if (*p == '\\')
 			{
@@ -172,7 +174,7 @@ MatchText(const char *t, int tlen, const char *p, int plen, pg_locale_t locale)
 
 			while (tlen > 0)
 			{
-				if (GETCHAR(*t, locale) == firstpat)
+				if (GETCHAR(*t, locale) == firstpat || (locale && !locale->deterministic))
 				{
 					int			matched = MatchText(t, tlen, p, plen, locale);
 
@@ -196,6 +198,149 @@ MatchText(const char *t, int tlen, const char *p, int plen, pg_locale_t locale)
 			NextByte(p, plen);
 			continue;
 		}
+		else if (locale && !locale->deterministic)
+		{
+			/*
+			 * For nondeterministic locales, we find the next substring of the
+			 * pattern that does not contain wildcards and try to find a
+			 * matching substring in the text.  Crucially, we cannot do this
+			 * character by character, as in the normal case, but must do it
+			 * substring by substring, partitioned by the wildcard characters.
+			 * (This is per SQL standard.)
+			 */
+			const char *p1;
+			size_t		p1len;
+			const char *t1;
+			size_t		t1len;
+			bool		found_escape;
+			const char *subpat;
+			size_t		subpatlen;
+			char	   *buf = NULL;
+
+			/*
+			 * Determine next substring of pattern without wildcards.  p is
+			 * the start of the subpattern, p1 is one past the last byte. Also
+			 * track if we found an escape character.
+			 */
+			p1 = p;
+			p1len = plen;
+			found_escape = false;
+			while (p1len > 0)
+			{
+				if (*p1 == '\\')
+				{
+					found_escape = true;
+					NextByte(p1, p1len);
+					if (p1len == 0)
+						ereport(ERROR,
+								(errcode(ERRCODE_INVALID_ESCAPE_SEQUENCE),
+								 errmsg("LIKE pattern must not end with escape character")));
+				}
+				else if (*p1 == '_' || *p1 == '%')
+					break;
+				NextByte(p1, p1len);
+			}
+
+			/*
+			 * If we found an escape character, then make an unescaped copy of
+			 * the subpattern.
+			 */
+			if (found_escape)
+			{
+				char	   *b;
+
+				b = buf = palloc(p1 - p);
+				for (const char *c = p; c < p1; c++)
+				{
+					if (*c == '\\')
+						;
+					else
+						*(b++) = *c;
+				}
+
+				subpat = buf;
+				subpatlen = b - buf;
+			}
+			else
+			{
+				subpat = p;
+				subpatlen = p1 - p;
+			}
+
+			/*
+			 * Shortcut: If this is the end of the pattern, then the rest of
+			 * the text has to match the rest of the pattern.
+			 */
+			if (p1len == 0)
+			{
+				int			cmp;
+
+				cmp = pg_strncoll(subpat, subpatlen, t, tlen, locale);
+
+				if (buf)
+					pfree(buf);
+				if (cmp == 0)
+					return LIKE_TRUE;
+				else
+					return LIKE_FALSE;
+			}
+
+			/*
+			 * Now build a substring of the text and try to match it against
+			 * the subpattern.  t is the start of the text, t1 is one past the
+			 * last byte.  We start with a zero-length string.
+			 */
+			t1 = t;
+			t1len = tlen;
+			for (;;)
+			{
+				int			cmp;
+
+				CHECK_FOR_INTERRUPTS();
+
+				cmp = pg_strncoll(subpat, subpatlen, t, (t1 - t), locale);
+
+				/*
+				 * If we found a match, we have to test if the rest of pattern
+				 * can match against the rest of the string.  Otherwise we
+				 * have to continue here try matching with a longer substring.
+				 * (This is similar to the recursion for the '%' wildcard
+				 * above.)
+				 *
+				 * Note that we can't just wind forward p and t and continue
+				 * with the main loop.  This would fail for example with
+				 *
+				 * U&'\0061\0308bc' LIKE U&'\00E4_c' COLLATE ignore_accents
+				 *
+				 * You'd find that t=\0061 matches p=\00E4, but then the rest
+				 * won't match; but t=\0061\0308 also matches p=\00E4, and
+				 * then the rest will match.
+				 */
+				if (cmp == 0)
+				{
+					int			matched = MatchText(t1, t1len, p1, p1len, locale);
+
+					if (matched == LIKE_TRUE)
+					{
+						if (buf)
+							pfree(buf);
+						return matched;
+					}
+				}
+
+				/*
+				 * Didn't match.  If we used up the whole text, then the match
+				 * fails.  Otherwise, try again with a longer substring.
+				 */
+				if (t1len == 0)
+					return LIKE_FALSE;
+				else
+					NextChar(t1, t1len);
+			}
+			if (buf)
+				pfree(buf);
+			continue;
+		}
 		else if (GETCHAR(*p, locale) != GETCHAR(*t, locale))
 		{
 			/* non-wildcard pattern char fails to match text char */

src/backend/utils/adt/like_support.c

@@ -272,22 +272,6 @@ match_pattern_prefix(Node *leftop,
 		return NIL;
 	patt = (Const *) rightop;
 
-	/*
-	 * Not supported if the expression collation is nondeterministic.  The
-	 * optimized equality or prefix tests use bytewise comparisons, which is
-	 * not consistent with nondeterministic collations.  The actual
-	 * pattern-matching implementation functions will later error out that
-	 * pattern-matching is not supported with nondeterministic collations. (We
-	 * could also error out here, but by doing it later we get more precise
-	 * error messages.)  (It should be possible to support at least
-	 * Pattern_Prefix_Exact, but no point as long as the actual
-	 * pattern-matching implementations don't support it.)
-	 *
-	 * expr_coll is not set for a non-collation-aware data type such as bytea.
-	 */
-	if (expr_coll && !get_collation_isdeterministic(expr_coll))
-		return NIL;
-
 	/*
 	 * Try to extract a fixed prefix from the pattern.
 	 */
@@ -404,13 +388,26 @@ match_pattern_prefix(Node *leftop,
 	{
 		if (!op_in_opfamily(eqopr, opfamily))
 			return NIL;
+		if (indexcollation != expr_coll)
+			return NIL;
 		expr = make_opclause(eqopr, BOOLOID, false,
 							 (Expr *) leftop, (Expr *) prefix,
 							 InvalidOid, indexcollation);
 		result = list_make1(expr);
 		return result;
 	}
 
+	/*
+	 * Anything other than Pattern_Prefix_Exact is not supported if the
+	 * expression collation is nondeterministic.  The optimized equality or
+	 * prefix tests use bytewise comparisons, which is not consistent with
+	 * nondeterministic collations.
+	 *
+	 * expr_coll is not set for a non-collation-aware data type such as bytea.
+	 */
+	if (expr_coll && !get_collation_isdeterministic(expr_coll))
+		return NIL;
+
 	/*
 	 * Otherwise, we have a nonempty required prefix of the values.  Some
 	 * opclasses support prefix checks directly, otherwise we'll try to