[Bug tree-optimization/41244] New: "&data[i] - data" isn't converted to "i"

[Bug tree-optimization/41244] New: "&data[i] - data" isn't converted to "i"

[zsojka at seznam dot cz]

In attached code:
in find(), generated code computes offset using multiplication/division.
in set(), generated code computes &data[10] and compares &data[i] with that (to
verify find() fails to be optimised because of overflow rules)

tested gcc: 4.3.4, 4.4.1, 4.5.0-alpha20090827

command line:
gcc main.c -o main.o -c -O3
-f(no-)strict-overflow or -m32 doesn't have any effect on that behaviour

---------------------------------
static struct {
        int a, b, c;
        char pad[200]; /* padding so offset computation has to use expensive
div/mul */
} data[1000]; /* 1000 so loop isn't unrolled */

int find(int val)
{
        int i;
        for (i = 0; i < 1000; i++) {
                if (data[i].a == val) return &data[i] - data; /* does *212,
/212 */
        }
        return -1;
}

void set(int val)
{
        int i;
        for (i = 0; i < 1000; i++) {
                data[i].b = &data[i] - data < 10 || data[i].a == val; /*
compares with &data[10] */
        }
}
---------------------------------

Relevant generated ASM:
(eax is "i", sar+imul is converted division by 212)
---------------------------------
        cdqe
        imul    rax, rax, 212
        sar     rax, 2
        imul    eax, eax, -1944890851
        ret
---------------------------------


-- 
           Summary: "&data[i] - data" isn't converted to "i"
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: zsojka at seznam dot cz
  GCC host triplet: x86_64-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[zsojka at seznam dot cz]

------- Comment #1 from zsojka at seznam dot cz  2009-09-03 13:50 -------
Created an attachment (id=18475)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18475&action=view)
preprocessed source


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[rguenth at gcc dot gnu dot org]

------- Comment #2 from rguenth at gcc dot gnu dot org  2009-09-03 14:12 -------
Because the index arithmetic is done unsigned.

  return (int) ((unsigned int) i * 212) /[ex] 212;

We lost the information that i * 212 cannot overflow.

Simpler testcase:

extern int data[];

int find(int i)
{
  return &data[i] - data;
}

from which we enter c_fully_fold_internal with

((int) ((int *) &data + (unsigned int) ((unsigned int) i * 4)) - (int) &data)
/[ex] 4

addressed on the no-undefined-overflow branch.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
           Keywords|                            |missed-optimization
   Last reconfirmed|0000-00-00 00:00:00         |2009-09-03 14:12:37
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Here are some better testcases as the power of 2 case can be handled on the rtl
level as it is just a shift.

extern struct s{
 int a, b, c, d, e, f, g, h;
}  data[];

int find(int i, int j)
{
  struct s *a = &data[i];
  struct s *b = &data[j];
  return a - b;
}


extern struct {
 int a, b, c;
}  data1[];

int find1(int i, int j)
{
  return &data1[i] - data1;
}

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |pinskia at gcc dot gnu.org

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Mine.

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---

  _14 = _13 * 32;
  _15 = (long int) _14;
  _2 = _15 /[ex] 32;

I think this will work ....
(simplify
 (exact_div (nop_convert (mult @0 INTEGER_CST@1)) INTEGER_CST@2)
 (if (!TYPE_UNSIGNED (@0)
      && wi::to_wide (@1) == wi::to_wide (@2))
  (convert @0)))

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

--- Comment #9 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 55302
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55302&action=edit
Patch which I will be testing

This extends the pattern that already handles `(t * 2) / 2) -> t`.

The one thing which I need to double check and understand if :c can/should be
used here, or should use another with here.

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

--- Comment #10 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 55303
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55303&action=edit
Better patch

This is a better patch, operand_equal_p already does the integer cst check too.

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |NEW

--- Comment #11 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
You know what, both patches are wrong with respect to the overflow check. 
Anyways Someone else can fix this.

[Bug tree-optimization/41244] "&data[i] - data" isn't converted to "i"

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=41244

--- Comment #12 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So basically the biggest issue is we do the sign extend and then do the
multiply in an unsigned type; this causes all negative values causing a
wrapping which is not correct really but we don't know any better. I tried to
change pass_laddress::execute to do the multiply in a signed type but that
still is not able to optimize find in comment #6 to work and the IR is so much
more complex.