Hi,
replace_text() in varlena.c builds the result in a StringInfo buffer,
and finishes by copying it into a freshly allocated varlena structure
with cstring_to_text_with_len(), in the same memory context.
It looks like that copy step could be avoided by preprending the
varlena header to the StringInfo to begin with, and return the buffer
as a text*, as in the attached patch.
On large strings, the time saved can be significant. For instance
I'm seeing a ~20% decrease in total execution time on a test with
lengths in the 2-3 MB range, like this:
select sum(length(
replace(repeat('abcdefghijklmnopqrstuvwxyz', i*10), 'abc', 'ABC')
))
from generate_series(10000,12000) as i;
Also, at a glance, there are a few other functions with similar
StringInfo-to-varlena copies that seem avoidable:
concat_internal(), text_format(), replace_text_regexp().
Are there reasons not to do this? Otherwise, should it be considered
in in a more principled way, such as adding to the StringInfo API
functions like void InitStringInfoForVarlena() and
text *StringInfoAsVarlena()?
Best regards,
--
Daniel Vérité
PostgreSQL-powered mailer: http://www.manitou-mail.org
Twitter: @DanielVerite
diff --git a/src/backend/utils/adt/varlena.c b/src/backend/utils/adt/varlena.c
index 693ccc5..3df54ed 100644
--- a/src/backend/utils/adt/varlena.c
+++ b/src/backend/utils/adt/varlena.c
@@ -4136,6 +4136,10 @@ replace_text(PG_FUNCTION_ARGS)
initStringInfo(&str);
+ /* allocate a varlena header at the start of the stringinfo */
+ enlargeStringInfo(&str, VARHDRSZ);
+ str.len += VARHDRSZ;
+
do
{
CHECK_FOR_INTERRUPTS();
@@ -4160,8 +4164,8 @@ replace_text(PG_FUNCTION_ARGS)
text_position_cleanup(&state);
- ret_text = cstring_to_text_with_len(str.data, str.len);
- pfree(str.data);
+ ret_text = (text*) str.data;
+ SET_VARSIZE(ret_text, str.len); /* VARHDRSZ is already included in
str.len */
PG_RETURN_TEXT_P(ret_text);
}
