https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128483

--- Comment #7 from Reshma Roy <Reshma.Roy at amd dot com> ---
I have posted an initial draft of the fix in
https://gcc.gnu.org/pipermail/gcc-patches/2026-November/738932.html
As per the discussion, the patch is created such that it places broadcasts of a
nonzero constant with LCM, and hence the broadcast is no longer hoisted into
the hot candidate-loop header. The all-zeros, all-ones, and CONST_VECTOR cases
are unchanged, and the PR92080 and PR120941 tests pass with the patch.

Reply via email to