[Bug tree-optimization/31358] New: IV-OPTs produces a weird MEM_REF

[Bug tree-optimization/31358] New: IV-OPTs produces a weird MEM_REF

[pinskia at gcc dot gnu dot org]

For:
int *f(int *a, int b, int c)
{
  int i;
  a --;
  *a = 1;
  for (i = 0;i<b;i+=c)
    (a[i]) = 0;
  return a;
}
----------------CUT----------------
I get in the final_cleanup:
  MEM[index: ivtmp.35, offset: 4294967292] = 0;


And I noticed that ivtmp.29 is of type "long unsigned int" which seems wrong as
it is actually the pointer.  I think this in the end could also cause PR 28690
too.

Anyways the assembly in the ends up to be:
L4:
        add r9,r9,r5
        li r2,0
        subf r0,r5,r9
        stw r2,-4(r3)
        cmpw cr7,r4,r0
        add r3,r3,r11
        bgt+ cr7,L4


Which is bad, I will file a bug about not pulling out the "li r2, 0" in another
bug (that is a regression too).  But we have an add and a sub -4 which is just
wrong.  we should be able to use stwx instead which we did in 4.0.2.


-- 
           Summary: IV-OPTs produces a weird MEM_REF
           Product: gcc
           Version: 4.3.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: pinskia at gcc dot gnu dot org
GCC target triplet: powerpc-darwin


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358

[Bug tree-optimization/31358] IV-OPTs produces a weird MEM_REF

[rakdver at gcc dot gnu dot org]

------- Comment #1 from rakdver at gcc dot gnu dot org  2007-03-26 10:17 -------
(In reply to comment #0)
> For:
> int *f(int *a, int b, int c)
> {
>   int i;
>   a --;
>   *a = 1;
>   for (i = 0;i<b;i+=c)
>     (a[i]) = 0;
>   return a;
> }
> ----------------CUT----------------
> I get in the final_cleanup:
>   MEM[index: ivtmp.35, offset: 4294967292] = 0;
> 
> 
> And I noticed that ivtmp.29 is of type "long unsigned int" which seems wrong as
> it is actually the pointer.

no, this is OK (whether it is a good design choice is another thing, I will
think about that).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358

[Bug tree-optimization/31358] IV-OPTs produces a weird MEM_REF

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2007-03-26 17:35 -------
Even if this MEM is ok, we still get bad results as we have *(a+b-4) inside the
loop now.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358

[Bug tree-optimization/31358] IV-OPTs produces a weird MEM_REF

[pinskia at gcc dot gnu dot org]

------- Comment #3 from pinskia at gcc dot gnu dot org  2007-03-26 22:27 -------
3.2.3 produced:
.L6:
        slwi 0,9,2
        add 9,9,5
        cmpw 0,9,4
        stwx 11,3,0
        bgelr- 0
        b .L6

While 4.1.0 and above produces:
.L4:
        add 9,9,5
        stw 10,-4(3)
        add 3,3,11
        subf 0,5,9
        cmpw 7,4,0
        bgt 7,.L4


If we change "int *" to "char *" and set c to be 2, we get even worse code gen.
3.2.3:

.L9:
        stbx 11,3,9
        addi 9,9,2
        bdnz .L9
        blr


4.1.0:
.L4:
        add 9,3,11
        addi 11,11,2
        cmpw 7,4,11
        stb 0,-1(9)
        bgt 7,.L4

What happened to bdnz and we are now doing *(r9+r3-1) = r0 inside the loop?


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|                            |4.1.0 4.2.0 4.3.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358

[Bug tree-optimization/31358] IV-OPTs produces a weird MEM_REF

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2007-06-08 07:52 -------
(In reply to comment #1) 
> no, this is OK (whether it is a good design choice is another thing, I will
> think about that).

I actually have a fix for this issue (the MEM with just index: and offset:) but
it is going to wait until after the pointer_plus merge.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358

[Bug tree-optimization/31358] IV-OPTs produces a weird MEM_REF

[pinskia at gcc dot gnu dot org]

------- Comment #5 from pinskia at gcc dot gnu dot org  2007-08-06 05:58 -------
Yes this does cause some of PR 28690.

Anyways, mine.
A patch for this was posted at 
http://gcc.gnu.org/ml/gcc-patches/2007-08/msg00006.html

I need to update it for the comments still.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
OtherBugsDependingO|                            |28690
              nThis|                            |
         AssignedTo|unassigned at gcc dot gnu   |pinskia at gcc dot gnu dot
                   |dot org                     |org
             Status|UNCONFIRMED                 |ASSIGNED
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2007-08-06 05:58:46
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358

[Bug tree-optimization/31358] IV-OPTs produces a weird MEM_REF

[rguenth at gcc dot gnu dot org]

------- Comment #6 from rguenth at gcc dot gnu dot org  2008-03-11 09:37 -------
Subject: Bug 31358

Author: rguenth
Date: Tue Mar 11 09:36:51 2008
New Revision: 133102

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=133102
Log:
2008-03-11  Andrew Pinski  <andrew_pinski@playstation.sony.com>
        Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/31358
        * tree-ssa-loop-manip.c (create_iv): Call force_gimple_operand for
        the step with a NULL_TREE.
        * tree-ssa-loop-ivopts.c (find_bivs): Convert the step
        to sizetype if type is a pointer type.
        (add_candidate_1): Don't convert the base and step to
        the generic type if the orginal type is a pointer type.
        (add_iv_value_candidates): Use sizetype for the step
        if type is a pointer type.
        (cand_value_at): Likewise.
        * tree-ssa-address.c (add_to_parts): Use POINTER_PLUS_EXPR
        for pointer types.
        * tree-affine.c (tree_to_aff_combination <POINTER_PLUS_EXPR>):
        Don't convert the tem affine to the type.
        (add_elt_to_tree): Use sizetype for the step if a pointer.
        Use POINTER_PLUS_EXPR for pointers.
        (aff_combination_to_tree): Use sizetype for the step if a
        pointer.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-affine.c
    trunk/gcc/tree-ssa-loop-ivopts.c
    trunk/gcc/tree-ssa-loop-manip.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358

[Bug tree-optimization/31358] IV-OPTs produces a weird MEM_REF

[rguenth at gcc dot gnu dot org]

------- Comment #7 from rguenth at gcc dot gnu dot org  2008-03-11 09:46 -------
Fixed.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.4.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31358