I think this patch is necessary:
diff --git a/src/interfaces/ecpg/preproc/pgc.l b/src/interfaces/ecpg/preproc/pgc.l
index 07fee80a9c..3529b2ea86 100644
--- a/src/interfaces/ecpg/preproc/pgc.l
+++ b/src/interfaces/ecpg/preproc/pgc.l
@@ -753,7 +753,7 @@ cppline {space}*#([^i][A-Za-z]*|{if}|{ifdef}|{ifndef}|{import})((\/\*[^*/]*\*+
}
<xui>{dquote} {
BEGIN(state_before_str_start);
- if (literallen == 2) /* "U&" */
+ if (literallen == 0)
mmerror(PARSE_ERROR, ET_ERROR, "zero-length delimited identifier");
/* The backend will truncate the identifier here. We do not as it does not change the result. */
base_yylval.str = psprintf("U&\"%s\"", literalbuf);
The old code doesn't make sense. The literallen is the length of the
data in literalbuf, which clearly doesn't include the "U&" as the
comment suggests.
A test case is to preprocess a file like this (ecpg test.pgc):
exec sql select u&"
which currently does *not* give the above error, but it should.