https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107153
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
OK, so the main issue is that when we parallelize a loop without stores we
insert stores and thus would require a loop PHI for the virtual operand plus
LC SSA updates.
We eventually end up in gen_parallel_loop at
/* We assume that the loop usually iterates a lot. */
loop_version (loop, many_iterations_cond, NULL,
profile_probability::likely (),
profile_probability::unlikely (),
profile_probability::likely (),
profile_probability::unlikely (), true);
update_ssa (TODO_update_ssa_no_phi);
which performs the update but doesn't insert PHIs - the above assumed
that we'd have no SSA update pending before the call to loop_version.
separate_decls_in_region is what creates the stores (with no virtual
operands, thus triggering renaming and the need for a LC PHI).
I'm going to work around.
