[Bug tree-optimization/112385] New: `(2 >> a) ^ (5 >> a)` is not optimized to `7 >> a` is 2 and 5 differ in signedness

[Bug tree-optimization/112385] New: `(2 >> a) ^ (5 >> a)` is not optimized to `7 >> a` is 2 and 5 differ in signedness

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112385

            Bug ID: 112385
           Summary: `(2 >> a) ^ (5 >> a)` is not optimized to `7 >> a` is
                    2 and 5 differ in signedness
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---

Take:
```
int f(int a)
{
  return (2 >> a) ^ (5u >> a);
}

```
This should be optimized to:
```
return (7 >> a);
```

But currently does not.

[Bug tree-optimization/112385] `(2 >> a) ^ (5 >> a)` is not optimized to `7 >> a` is 2 and 5 differ in signedness

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112385

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note a more complex case like:
```
int f(int a, unsigned b, int c)
{
  c &= 0xffff;
  return (c >> a) ^ (b >> a);
}
```
Should also be optimized to:
```
int f(int a, unsigned b, int c)
{
  c &= 0xffff;
  return (c^b) >> a;
}
```

That is if both are known to be non-negative we can do the combining.

[Bug tree-optimization/112385] `(2 >> a) ^ (5 >> a)` is not optimized to `7 >> a` is 2 and 5 differ in signedness

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112385

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
We can even do it if we know the unsigned value does not have the last bit set.
That is:
```
int f(int a, unsigned b, int c)
{
  b &= 0xffff;
  return (c >> a) ^ (b >> a);
}

```
Note clang(LLVM) does not handle the above.

The way LLVM handles the others is it changes ashift to lshift if the value is
known to be non-negative. But since in the gimple level we are dealing with
types we can't do that in a nice way; RTL level on the other hand yes that can
be done (though nonzero bits is limited there).

[Bug tree-optimization/112385] `(2 >> a) ^ (5 >> a)` is not optimized to `7 >> a` is 2 and 5 differ in signedness

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112385

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-11-12
             Status|UNCONFIRMED                 |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |pinskia at gcc dot gnu.org

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Mine, I will handle this.