https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117111
--- Comment #2 from Kazumoto Kojima <kkojima at gcc dot gnu.org> ---
dbr_schedule is trying to fill the delay slot of
(jump_insn 17 16 42 (set (pc)
(if_then_else (eq (reg:SI 147 t)
(const_int 0 [0]))
(label_ref:SI 94)
(pc))) "fpone.c":7:11 232 {*cbranch_t}
(int_list:REG_BR_PROB 536870916 (nil))
-> 94)
by fill_slots_from_thread function. It takes the insn for the slot from the
thread
(insn 46 39 79 (set (reg:SI 8 r8 [orig:167 _1 ] [167])
(reg:SI 147 t)) "fpone.c":7:11 303 {movt}
(expr_list:REG_DEAD (reg:SI 147 t)
(nil)))
(insn 79 46 80 (set (reg/i:SI 0 r0)
(reg:SI 8 r8 [orig:167 _1 ] [167])) "fpone.c":8:1 191 {movsi_ie}
(expr_list:REG_DEAD (reg:SI 8 r8 [orig:167 _1 ] [167])
(nil)))
and makes a candidate insn
(insn 79 46 80 (set (reg/i:SI 0 r0)
(reg:SI 147 t)) "fpone.c":8:1 303 {movt}
(expr_list:REG_DEAD (reg:SI 8 r8 [orig:167 _1 ] [167])
(nil)))
for trial. fill_slots_from_thread calls try_split for this candidate first.
Here is a gdb backtrace where trial is the insn 79 above.
#0 try_split (pat=pat@entry=0x7ffff6f5ec48, trial=trial@entry=0x7ffff6f5d400,
last=last@entry=0) at /git/gcc/gcc/emit-rtl.cc:3932
#1 0x0000000000f5a748 in fill_slots_from_thread (insn=0x7ffff6e0cd38,
condition=0x7ffff6f5f0d8, thread_or_return=<optimized out>,
opposite_thread=<optimized out>, likely=false,
thread_if_true=<optimized out>, own_thread=true,
slots_to_fill=<optimized out>, pslots_filled=0x7fffffffd7ac,
delay_list=0x7fffffffd7d0) at /git/gcc/gcc/reorg.cc:2430
#2 0x0000000000f5d381 in fill_eager_delay_slots ()
at /git/gcc/gcc/reorg.cc:2843
#3 dbr_schedule (first=<optimized out>) at /git/gcc/gcc/reorg.cc:3705
and try_split applies the splitter
sh.md: 11147
;; This is not a peephole, but it's here because it's actually supposed
;; to be one. It tries to convert a sequence such as
;; movt r2 -> movt r2
;; movt r13 mov r2,r13
;; This gives the schduler a bit more freedom to hoist a following
;; comparison insn. Moreover, it the reg-reg mov insn is MT group which has
;; better chances for parallel execution.
;; We can do this with a peephole2 pattern, but then the cprop_hardreg
;; pass will revert the change. See also PR 64331.
;; Thus do it manually in one of the split passes after register allocation.
;; Sometimes the cprop_hardreg pass might also eliminate the reg-reg copy.
(define_split
[(set (match_operand:SI 0 "arith_reg_dest")
(match_operand:SI 1 "t_reg_operand"))]
...
and returns
(insn 113 46 80 (set (reg/i:SI 0 r0)
(reg:SI 8 r8 [orig:167 _1 ] [167])) "fpone.c":8:1 -1
(nil))
Thus fill_slots_from_thread fills the slot with it. This situation would be
unexpected with both fill_slots_from_thread and the above splitter.
BTW, old RA makes a bit worse code for the thread:
.L6:
movt r1
mov.l r1,@r15
mov.l @r15,r0
add #4,r15
lds.l @r15+,pr
and the splitter cloudn't be applied.