[Bug tree-optimization/62631] gcc.dg/tree-ssa/ivopts-lt-2.c FAILs

["ro at CeBiTec dot Uni-Bielefeld.DE" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62631

--- Comment #9 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> ---
> --- Comment #8 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
>> I think that's easiest for Eric to say.
>
> Not really, I guess you want to debug the function and replay the computation
> since the cost is synthetized and doesn't come directly from the back-end.

I've found what's going on:

* In expmed.c (init_expmed_one_mode), l.194

  set_shiftadd_cost (speed, mode, m, set_src_cost (all->shift_add, speed));

  with all->shift_add something like

(plus:QI (mult:QI (reg:QI 109 [0])
        (const_int 8 [0x8]))
    (reg:QI 109 [0]))

* For the mult part, rtx_code calls sparc_rtx_cost, which has

    case MULT:
      if (float_mode_p)
    *total = sparc_costs->float_mul;
      else if (! TARGET_HARD_MUL)
    *total = COSTS_N_INSNS (25);

  On SPARCv9/-m64, TARGET_HARD_MUL is false, so we get the 25*4 = 100 part,
  unlike v8, which explains why the test only fails for 64-bit.

    Rainer