On Mon, Apr 9, 2018 at 12:27 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
Alexander Korotkov <a.korotkov@postgrespro.ru> writes: > There is also patch for valgrind error in CheckIndexCompatible(). > It appears that this function didn't initialize ii_NumIndexAttrs and > ii_NumIndexKeyAttrs before calling ComputeIndexAttrs(). This > variables weren't used before, but now they're used to determine > whether index column is key. Also, this patch makes ComputeIndexAttrs() > initialize classOidP elements with InvalidOid for non-key attributes.
You didn't test this quite enough, because when I tried it I got a batch of warnings about memory chunk inconsistencies. The reason was that DefineIndex only allocated a classObjectId array large enough for the key columns, so that the change in ComputeIndexAttrs made it write past the end of that array.
Sorry, my bad.
But I think it's a good idea to make those output arrays all the same length, so I kept that change and changed DefineIndex.
Pushed; I've not checked to see if this makes valgrind happy, but the buildfarm will tell us.