Regress-test failure in current sources

With this afternoon's CVS sources, I got a failure in the 'polygon'
regression test while running the parallel-mode tests:

--
CREATE TABLE POLYGON_TBL(f1 polygon);
+ ERROR: heap_update: (am)invalid tid
INSERT INTO POLYGON_TBL(f1) VALUES ('(2.0,0.0),(2.0,4.0),(0.0,0.0)');

(plus many subsequent diffs arising from the lack of polygon_tbl).

This doesn't seem to be very repeatable --- I've only seen it once in
about a dozen tries. But there's a problem in there somewhere. I am
cycling the tests now in hopes of duplicating the failure.

There are only four heap_update() calls performed during this command,
all arising directly or indirectly from creation of the TOAST table:

#1 0xbe110 in UpdateStats (relid=10, reltuples=1000) at index.c:1742

#1 0xbe110 in UpdateStats (relid=1, reltuples=1000) at index.c:1742

#1 0xbdb70 in setRelhasindex (relid=145156, hasindex=1 '\001') at index.c:1487

#1 0xddd78 in AlterTableCreateToastTable (relationName=0x23710 "",
silent=-48 '') at command.c:1696

My suspicions are focused on the last one in AlterTableCreateToastTable,
but I can't prove it yet ...

Has anyone else seen something like this? Any ideas what's wrong?

regards, tom lane