[Bug tree-optimization/110932] [14 Regression] Dead Code Elimination Regression since r14-2230-g7e904d6c7f2

[law at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110932

Jeffrey A. Law  changed:

   What|Removed |Added

   Priority|P3  |P2
 CC||law at gcc dot gnu.org

[Bug tree-optimization/110932] [14 Regression] Dead Code Elimination Regression since r14-2230-g7e904d6c7f2

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110932

--- Comment #2 from Andrew Pinski  ---
(In reply to Andrew Pinski from comment #1)
> So the general rule is:
> (simplify
>  (eq:c @0 (convert (cmp @1 @2)))
>  (if (bitwise_equal_p (@0, @1))
>  (with {
> bool zeroalwaystrue = ...
> bool zeroalwaysfalse = ...
> bool onealwaystrue = ...
> bool onealwaysfalse = ...
>   }
>   (switch
>(if (zeroalwaystrue && onealwaystrue)
> (le @0:unsigned { one(unsigned); } )) // @0 == 0 | @0 == 1
>(if (zeroalwaysfalse && onealwaysfalse )
> { false_bool; }) // false
>(if (zeroalwaystrue && onealwaysfalse)
> (eq @0 { zero(TREE_TYPE(@0)); })) // @0 == 0
>(if (onealwaystrue && onealwaysfalse)
> (eq @0 { one(TREE_TYPE(@0)); })) // @0 == 1

If we want to do this only in VRP, we could just use match to match the
sequence and then handle the comparison there ...

[Bug tree-optimization/110932] [14 Regression] Dead Code Elimination Regression since r14-2230-g7e904d6c7f2

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110932

Andrew Pinski  changed:

   What|Removed |Added

 Status|UNCONFIRMED |NEW
 Ever confirmed|0   |1
   Last reconfirmed||2023-08-19

--- Comment #1 from Andrew Pinski  ---
  _41 = _43 <= 0;
  # RANGE [irange] unsigned short [0, 1] NONZERO 0x1
  _40 = (unsigned short) _41;
  # RANGE [irange] unsigned short [65534, +INF]
  _36 = ~_40;
  # RANGE [irange] short int [-2, -1]
  _19 = (short intD.17) _36;
  # RANGE [irange] unsigned int [4294967294, +INF]
  _17 = (unsigned intD.9) _19; // ~(unsigned)_41
   // rather _41 ? 4294967294 : 4294967295
  i.1_14 = (unsigned intD.9) pretmp_60;
  _13 = i.1_14 < _17;
  # RANGE [irange] int [0, 1] NONZERO 0x1
  _12 = (intD.6) _13;
  if (_12 == pretmp_60)


pretmp_60 == (((unsigned intD.9) pretmp_60) < _17)

Which is:
(pretmp_60  == 1 && ((unsigned intD.9) pretmp_60) < _17))
| (pretmp_60  == 0 && ((unsigned intD.9) pretmp_60) >= _17))
Though ((unsigned intD.9) pretmp_60) >= _17) is always false because _17's
range does not include 0.

So we are left with:
(pretmp_60  == 1 && 1 < _17)
But _17's range is always bigger than 0 so we are just left with:
pretmp_60 == 1


So the general rule is:
(simplify
 (eq:c @0 (convert (cmp @1 @2)))
 (if (bitwise_equal_p (@0, @1))
 (with {
bool zeroalwaystrue = ...
bool zeroalwaysfalse = ...
bool onealwaystrue = ...
bool onealwaysfalse = ...
  }
  (switch
   (if (zeroalwaystrue && onealwaystrue)
(le @0:unsigned { one(unsigned); } )) // @0 == 0 | @0 == 1
   (if (zeroalwaysfalse && onealwaysfalse )
{ false_bool; }) // false
   (if (zeroalwaystrue && onealwaysfalse)
(eq @0 { zero(TREE_TYPE(@0)); })) // @0 == 0
   (if (onealwaystrue && onealwaysfalse)
(eq @0 { one(TREE_TYPE(@0)); })) // @0 == 1

Later on instead of `if (_12 == _50)` we would have `if (_50 == 1)`
```
  if (_50 == 1)
goto ; [51.12%]
  else
goto ; [48.88%]
...
   [local count: 278223726]:
  if (_50 != 0)
goto ; [33.00%]
  else
goto ; [67.00%]
```
Which is an obvious case where we know _50 is 1 in bb 4 and therefore !=0.
etc.

[Bug tree-optimization/110932] [14 Regression] Dead Code Elimination Regression since r14-2230-g7e904d6c7f2

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110932

Richard Biener  changed:

   What|Removed |Added

   Target Milestone|--- |14.0
 CC||hubicka at gcc dot gnu.org
   Keywords||missed-optimization