[Bug rtl-optimization/10050] ifcvt is not smart enough

[Bug rtl-optimization/10050] ifcvt is not smart enough

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10050

--- Comment #10 from Andrew Pinski <pinskia at gcc dot gnu.org> 2012-02-07 21:05:24 UTC ---
Note this optimized I think is invalid with the C11/C++11 memory model.

[Bug middle-end/10050] ifcvt is not smart enough

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=10050

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|rtl-optimization            |middle-end

--- Comment #11 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note for func2 we get:

  if (flag.0_2 != 0)
    goto <bb 4>; [50.00%]
  else
    goto <bb 3>; [50.00%]

  <bb 3> [local count: 536870912]:
  pretmp_18 = *a_10(D);
  pretmp_20 = MEM[(int *)a_10(D) + 4B];
  goto <bb 5>; [100.00%]

  <bb 4> [local count: 536870913]:
  *a_10(D) = _9;
  MEM[(int *)a_10(D) + 4B] = _1;

  <bb 5> [local count: 1073741824]:
  # prephitmp_19 = PHI <pretmp_18(3), _9(4)>
  # prephitmp_21 = PHI <pretmp_20(3), _1(4)>

the conditional stores are correct if not supplying -fallow-store-data-races . 
Now GCC does not optimize this even with -fallow-store-data-races but maybe it
could.

[Bug rtl-optimization/10050] ifcvt is not smart enough

[ian at airs dot com]

------- Comment #9 from ian at airs dot com  2006-01-15 06:27 -------
Upon further reflection, I don't think this is amenable to a good fix using
conditional moves.  func2 doesn't read from a.  Using conditional moves will
require adding reads from a.  It's not obvious that that is a good tradeoff.

When I look at the code generated for i686 at -O2, the only problem I see is
that for some reason gcc doesn't recognize that a[0] and a[1] can not alias
each other, and it reloads a[0] after writing to a[1].


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10050

[Bug rtl-optimization/10050] ifcvt is not smart enough

[ian at airs dot com]

------- Comment #8 from ian at airs dot com  2006-01-15 05:28 -------
My patch doesn't help for func2, at least on x86, because at the RTL level we
see stores to memory, and my patch doesn't attempt to handle that because of
all the aliasing issues between the then-block and the else-block.

It could be done easily enough with a similar patch, one which checked for a
conditional which only had a then-block or only had an else-block.


-- 

ian at airs dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ian at airs dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10050

[Bug rtl-optimization/10050] ifcvt is not smart enough

[pinskia at gcc dot gnu dot org]

------- Comment #7 from pinskia at gcc dot gnu dot org  2006-01-15 03:53 -------
func3 is fixed on x86 but not func2.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10050

[Bug rtl-optimization/10050] ifcvt is not smart enough

[pinskia at gcc dot gnu dot org]

------- Comment #6 from pinskia at gcc dot gnu dot org  2006-01-15 03:15 -------
This might have been fixed by (I am going to check soon) :
2006-01-14  Ian Lance Taylor  <ian@airs.com>

        * ifcvt.c (noce_init_if_info): New static function, broken out of
        noce_process_if_block.
        (noce_process_if_block): Call noce_init_if_info.
        (check_cond_move_block): New static function.
        (cond_move_process_if_block): New static function.
        (process_if_block): Call cond_move_process_if_block.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10050

[Bug rtl-optimization/10050] ifcvt is not smart enough

[pinskia at gcc dot gnu dot org]

------- Comment #5 from pinskia at gcc dot gnu dot org  2005-12-07 04:14 -------
This is only an if conversion failure now.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|2005-09-07 17:41:48         |2005-12-07 04:14:51
               date|                            |
            Summary|Weak optimization on array  |ifcvt is not smart enough
                   |references                  |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10050