I wrote:
> I hypothesize that the reason we're not seeing equivalent failures
> on x86_64 is one of
> 1. x86_64 valgrind is stupider than aarch64, and fails to track that
> the contents of the SIMD registers are only partially defined.
> 2. x86_64 valgrind is smarter than aarch64, and is able to see
> that the "mask off invalid entries" step removes all the
> potentially-uninitialized bits.
Side note: it struck me that this could also be a valgrind version
skew issue. But the machine I'm seeing the failure on is running
valgrind-3.22.0-1.fc38.aarch64, which is the same upstream version
as valgrind-3.22.0-2.el8.x86_64, where I don't see it. So apparently
not. (There is a 3.23.0 out recently, but its release notes don't
mention anything that seems related.)
regards, tom lane