Re: Fwd: cselib_record_set breaks due to auto_inc_dec

[Kenneth Zadeck <zadeck at naturalbridge dot com>]

if i had to guess what was going on here, and since i do not have the 
example or even know the platform, i am just guessing, i would say that 
at some earlier pass, the sub expression in the regnote and the second 
set in the parallel were the same expression and that someone changed 
the second set but did not delete the note or modify it.

and once they became different, it was just a case of "dead man walking".
On 12/19/2012 11:28 PM, David Edelsohn wrote:
> ---------- Forwarded message ----------
> From: Paulo Matos <pmatos@broadcom.com>
> Date: Wed, Dec 19, 2012 at 11:59 AM
> Subject: cselib_record_set breaks due to auto_inc_dec
> To: "gcc@gcc.gnu.org" <gcc@gcc.gnu.org>
>
>
> Hi,
>
> After pro and epilogue pass I have the following epilogue insn:
> (insn/f 20 19 21 2 (parallel [
>              (set (mem:DI (plus:SI (reg:SI 26 r26)
>                          (reg/f:SI 57 r57)) [0 S8 A8])
>                  (reg:DI 57 r57))
>              (set (reg/f:SI 57 r57)
>                  (plus:SI (reg:SI 26 r26)
>                      (reg/f:SI 57 r57)))
>          ]) bdra_main.i:2 -1
>       (expr_list:REG_DEAD (reg:SI 26 r26)
>          (expr_list:REG_FRAME_RELATED_EXPR (set (mem:DI (pre_modify:SI
> (reg/f:SI 57 r57)
>                          (plus:SI (reg/f:SI 57 r57)
>                              (const_int -8208 [0xffffffffffffdff0]))) [0 S8 A8])
>                  (reg:DI 57 r57))
>              (nil))))
>
> During cselib_record_sets, 2 sets are recognized (the ones in the
> parallel) and then a third one is added by cselib_record_autoinc_cb:
> (set (reg:SI 57) (plus (reg:SI 57) (const_int -8208)))
>
> After for_each_inc_dec calling cselib_record_autoinc_cb sets array is:
> sets[0]:
>    dest: (reg/f:SI 57 r57)
>    src: (plus:SI (reg:SI 26 r26)
>                      (reg/f:SI 57 r57))
> sets[1]:
>    dest: (mem:DI (plus:SI (reg:SI 26 r26)
>                          (reg/f:SI 57 r57)))
>    src: (reg:DI 57 r57)
> sets[2]:
>    dest: (reg/f:SI 57 r57)
>    src: (plus:SI (reg/f:SI 57 r57)
>                              (const_int -8208 [0xffffffffffffdff0]))))
>
> The last loop in cselib_record_sets is:
>    for (i = 0; i < n_sets; i++)
>      {
>        rtx dest = sets[i].dest;
>        if (REG_P (dest)
>            || (MEM_P (dest) && cselib_record_memory))
>          cselib_record_set (dest, sets[i].src_elt, sets[i].dest_addr_elt);
>      }
>
> This fails while trying to register the second set to reg:SI 57, in
>            /* The register should have been invalidated.  */
>            gcc_assert (REG_VALUES (dreg)->elt == 0);
>
> This seems to be a manifestation of PR32335, where the extra
> assignment to the same register (in this case r57) is added by
> auto_inc_dec analysis and not explicit in the parallel.
>
> Has anyone come across this before?
>
> Paulo Matos