https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120941
--- Comment #41 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to rguent...@suse.de from comment #40)
> On Wed, 30 Jul 2025, hjl.tools at gmail dot com wrote:
> > With my patch, we got
> >
> > basic_block bb = nearest_common_dominator_for_set (CDI_DOMINATORS, bbs);
> > /* For X86_CSE_VEC_DUP, don't place the vector set outside of the loop
> > to avoid extra spills. */
> > if (!load || load->kind != X86_CSE_VEC_DUP)
> > {
> > while (bb->loop_father->latch
> > != EXIT_BLOCK_PTR_FOR_FN (cfun))
> > bb = get_immediate_dominator (CDI_DOMINATORS,
> > bb->loop_father->header);
> > }
> >
> > When load->kind == X86_CSE_VEC_DUP, bb is the nearest common dominator which
> > may be inside the loop.
>
> That ensures this when there's a single reaching def for all of the uses
> in 'bbs'. So your patch then is also a correctness fix.
Before my patch, a single reaching def is hoisted outside of all loops, which
is still correct, but caused extra spill in some cases.
