https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108832

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |uros at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, this is on
(insn 263 304 17 2 (set (reg:DI 2 cx [98])
        (reg:DI 1 dx [98])) "pr108832.c":15:9 discrim 1 82 {*movdi_internal}
     (expr_list:REG_DEAD (reg:DI 1 dx [98])
        (nil)))
(insn 17 263 261 2 (parallel [
            (set (reg:CCZ 17 flags)
                (compare:CCZ (and:SI (reg:SI 2 cx [98])
                        (const_int -2 [0xfffffffffffffffe]))
                    (const_int 0 [0])))
            (set (reg:DI 2 cx)
                (zero_extend:DI (and:SI (reg:SI 2 cx [98])
                        (const_int -2 [0xfffffffffffffffe]))))
        ]) "pr108832.c":16:11 565 {*andsi_2_zext}
     (nil))
(insn 261 17 259 2 (set (reg:DI 0 ax)
        (const_int 1 [0x1])) 82 {*movdi_internal}
     (expr_list:REG_EQUIV (const_int 1 [0x1])
        (nil)))
(insn 259 261 4 2 (set (reg:DI 2 cx)
        (if_then_else:DI (ne (reg:CCZ 17 flags)
                (const_int 0 [0]))
            (reg:DI 2 cx)
            (reg:DI 0 ax))) 1304 {*movdicc_noc}
     (expr_list:REG_DEAD (reg:CCZ 17 flags)
        (expr_list:REG_DEAD (reg:DI 0 ax)
            (nil))))
on which
;; Eliminate a reg-reg mov by inverting the condition of a cmov (#1).
;; mov r0,r1; dec r0; mov r2,r3; cmov r0,r2 -> dec r1; mov r0,r3; cmov r0, r1
(define_peephole2
 [(set (match_operand:SWI248 0 "general_reg_operand")
       (match_operand:SWI248 1 "general_reg_operand"))
  (parallel [(set (reg FLAGS_REG) (match_operand 5))
             (set (match_dup 0) (match_operand:SWI248 6))])
  (set (match_operand:SWI248 2 "general_reg_operand")
       (match_operand:SWI248 3 "general_gr_operand"))
  (set (match_dup 0)
       (if_then_else:SWI248 (match_operator 4 "ix86_comparison_operator"
                             [(reg FLAGS_REG) (const_int 0)])
        (match_dup 0)
        (match_dup 2)))]
 "TARGET_CMOVE
  && REGNO (operands[2]) != REGNO (operands[0])
  && REGNO (operands[2]) != REGNO (operands[1])
  && peep2_reg_dead_p (1, operands[1])
  && peep2_reg_dead_p (4, operands[2])
  && !reg_overlap_mentioned_p (operands[0], operands[3])"
 [(parallel [(set (match_dup 7) (match_dup 8))
             (set (match_dup 1) (match_dup 9))])
  (set (match_dup 0) (match_dup 3))
  (set (match_dup 0) (if_then_else:SWI248 (match_dup 4)
                                          (match_dup 1)
                                          (match_dup 0)))]
{
  operands[7] = SET_DEST (XVECEXP (PATTERN (peep2_next_insn (1)), 0, 0));
  operands[8] = replace_rtx (operands[5], operands[0], operands[1], true);
  operands[9] = replace_rtx (operands[6], operands[0], operands[1], true);
})

applies.  replace_rtx has 2 modes, !all_regs in which it replaces just x ==
from with to
and all_regs, in which case it does:
  if (all_regs
      && REG_P (x)
      && REG_P (from)
      && REGNO (x) == REGNO (from))
    {
      gcc_assert (GET_MODE (x) == GET_MODE (from));
      return to;
    }
and so ICEs if we see the same REGNO as from in a different mode.
I think we actually don't need most of what replace_rtx is doing, we don't need
to simplify anything etc. because we are just changing one register to another
and can do it in place.
So, I think we need a different function for what the backend wants.
It can avoid all the simplify stuff because replace_rtx was destructive, so
could be implemented say using FOR_EACH_SUBRTX_PTR.  When seeing *loc == from,
it obviously
should set *loc = to, if it sees REG_P (*loc) && REGNO (*loc) == REGNO (from),
then
if the mode is the same, it can also just *loc = to, but if it is a different
mode,
I'd say for narrower mode it should *loc = gen_rtx_REG (GET_MODE (*loc), REGNO
(to));
and for wider mode (especially if say a multi-register reg) punt.
Not sure if such a case can occur though, but the punting would be hard if we
have made some changes already...

Reply via email to