https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89795
--- Comment #5 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
So the combiner first eliminates a ZERO_EXTEND between 2 instructions as
redundant, which is OK in isolation, but IRA (combine_and_move_insns) later
combines again the same 2 instructions without using the
WORD_REGISTER_OPERATIONS semantics of the combiner.
This mini-combine pass in IRA isn't always run:
/* Don't move insns if live range shrinkage or register
pressure-sensitive scheduling were done because it will not
improve allocation but likely worsen insn scheduling. */
if (optimize
&& !flag_live_range_shrinkage
&& !(flag_sched_pressure && flag_schedule_insns))
combine_and_move_insns ();
which explains the need to -fno-sched-pressure on the ARM.
