Fix bogus handling of XQuery regex option flags.

The SQL spec defers to XQuery to define what the option flags are
for LIKE_REGEX patterns.  XQuery says that:
* 's' allows the dot character to match newlines, which by
  default it will not;
* 'm' allows ^ and $ to match at newlines, not only at the
  start/end of the whole string.
Thus, these are *not* inverses as they are for the similarly-named
POSIX options, and neither one corresponds to the POSIX 'n' option.
Fortunately, Spencer's library does expose these two behaviors as
separately twiddlable flags, so we just have to fix the mapping from
JSP flag bits to REG flag bits.  I also chose to rename the symbol
for 's' to DOTALL, to make it clearer that it's not the inverse
of MLINE.

Also, XQuery says that if the 'q' flag "is used together with the m, s,
or x flag, that flag has no effect".  I read this as saying that 'q'
overrides the other flags; whoever wrote our code seems to have read
it backwards.

Lastly, while XQuery's 'x' flag is related to what Spencer's code
does for REG_EXPANDED, it's not the same or a subset.  It seems best
to treat XQuery's 'x' as unimplemented for now.  Maybe later we can
expand our regex code to offer 'x'-style parsing as a separate option.

While at it, refactor the jsonpath code so that (a) there's only
one copy of the flag transformation logic not two, and (b) the
processing of flags is independent of the order in which the flags
are written.

We need some documentation updates to go with this, but I'll
tackle that separately.

Back-patch to v12 where this code originated.

Discussion: https://postgr.es/m/CAPpHfdvDci4iqNF9fhRkTqhe-5_8HmzeLt56drH%2B_Rv2rNRqfg@mail.gmail.com
Reference: https://www.w3.org/TR/2017/REC-xpath-functions-31-20170321/#flags

1 files changed, 50 insertions, 14 deletions

diff --git a/src/backend/utils/adt/jsonpath_gram.y b/src/backend/utils/adt/jsonpath_gram.y
index 91b4b2f5985..1725502ede2 100644
--- a/src/backend/utils/adt/jsonpath_gram.y
+++ b/src/backend/utils/adt/jsonpath_gram.y
@@ -481,42 +481,32 @@ makeItemLikeRegex(JsonPathParseItem *expr, JsonPathString *pattern,
 {
 	JsonPathParseItem  *v = makeItemType(jpiLikeRegex);
 	int					i;
-	int					cflags = REG_ADVANCED;
+	int					cflags;
 
 	v->value.like_regex.expr = expr;
 	v->value.like_regex.pattern = pattern->val;
 	v->value.like_regex.patternlen = pattern->len;
-	v->value.like_regex.flags = 0;
 
+	/* Parse the flags string, convert to bitmask.  Duplicate flags are OK. */
+	v->value.like_regex.flags = 0;
 	for (i = 0; flags && i < flags->len; i++)
 	{
 		switch (flags->val[i])
 		{
 			case 'i':
 				v->value.like_regex.flags |= JSP_REGEX_ICASE;
-				cflags |= REG_ICASE;
 				break;
 			case 's':
-				v->value.like_regex.flags &= ~JSP_REGEX_MLINE;
-				v->value.like_regex.flags |= JSP_REGEX_SLINE;
-				cflags |= REG_NEWLINE;
+				v->value.like_regex.flags |= JSP_REGEX_DOTALL;
 				break;
 			case 'm':
-				v->value.like_regex.flags &= ~JSP_REGEX_SLINE;
 				v->value.like_regex.flags |= JSP_REGEX_MLINE;
-				cflags &= ~REG_NEWLINE;
 				break;
 			case 'x':
 				v->value.like_regex.flags |= JSP_REGEX_WSPACE;
-				cflags |= REG_EXPANDED;
 				break;
 			case 'q':
 				v->value.like_regex.flags |= JSP_REGEX_QUOTE;
-				if (!(v->value.like_regex.flags & (JSP_REGEX_MLINE | JSP_REGEX_SLINE | JSP_REGEX_WSPACE)))
-				{
-					cflags &= ~REG_ADVANCED;
-					cflags |= REG_QUOTE;
-				}
 				break;
 			default:
 				ereport(ERROR,
@@ -528,6 +518,9 @@ makeItemLikeRegex(JsonPathParseItem *expr, JsonPathString *pattern,
 		}
 	}
 
+	/* Convert flags to what RE_compile_and_cache needs */
+	cflags = jspConvertRegexFlags(v->value.like_regex.flags);
+
 	/* check regex validity */
 	(void) RE_compile_and_cache(cstring_to_text_with_len(pattern->val,
 														 pattern->len),
@@ -537,6 +530,49 @@ makeItemLikeRegex(JsonPathParseItem *expr, JsonPathString *pattern,
 }
 
 /*
+ * Convert from XQuery regex flags to those recognized by our regex library.
+ */
+int
+jspConvertRegexFlags(uint32 xflags)
+{
+	/* By default, XQuery is very nearly the same as Spencer's AREs */
+	int			cflags = REG_ADVANCED;
+
+	/* Ignore-case means the same thing, too, modulo locale issues */
+	if (xflags & JSP_REGEX_ICASE)
+		cflags |= REG_ICASE;
+
+	/* Per XQuery spec, if 'q' is specified then 'm', 's', 'x' are ignored */
+	if (xflags & JSP_REGEX_QUOTE)
+	{
+		cflags &= ~REG_ADVANCED;
+		cflags |= REG_QUOTE;
+	}
+	else
+	{
+		/* Note that dotall mode is the default in POSIX */
+		if (!(xflags & JSP_REGEX_DOTALL))
+			cflags |= REG_NLSTOP;
+		if (xflags & JSP_REGEX_MLINE)
+			cflags |= REG_NLANCH;
+
+		/*
+		 * XQuery's 'x' mode is related to Spencer's expanded mode, but it's
+		 * not really enough alike to justify treating JSP_REGEX_WSPACE as
+		 * REG_EXPANDED.  For now we treat 'x' as unimplemented; perhaps in
+		 * future we'll modify the regex library to have an option for
+		 * XQuery-style ignore-whitespace mode.
+		 */
+		if (xflags & JSP_REGEX_WSPACE)
+			ereport(ERROR,
+					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+					 errmsg("XQuery \"x\" flag (expanded regular expressions) is not implemented")));
+	}
+
+	return cflags;
+}
+
+/*
  * jsonpath_scan.l is compiled as part of jsonpath_gram.y.  Currently, this is
  * unavoidable because jsonpath_gram does not create a .h file to export its
  * token symbols.  If these files ever grow large enough to be worth compiling