[Bug tree-optimization/79119] absolute value of a pointer difference can be assumed to be less than or equal to PTRDIFF_MAX

[Bug tree-optimization/79119] absolute value of a pointer difference can be assumed to be less than or equal to PTRDIFF_MAX

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79119

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So a couple of things need to happen really:
optimized now:
  if (p_10 < q_12)
    goto <bb 3>; [50.00%]
  else
    goto <bb 4>; [50.00%]

  <bb 3> [local count: 536870913]:
  _3 = _22 - _21;
  _4 = _3 /[ex] 4;
  iftmp.0_14 = (long unsigned int) _4;
  goto <bb 5>; [100.00%]

  <bb 4> [local count: 536870913]:
  _5 = _21 - _22;
  _6 = _5 /[ex] 4;
  iftmp.0_13 = (long unsigned int) _6;

  <bb 5> [local count: 1073741824]:
  # iftmp.0_7 = PHI <iftmp.0_14(3), iftmp.0_13(4)>

I have a patch which is able to move the cast out of the PHI (will be posting
it today or tomorrow).
The next thing is we need to move the /[ext] 4 out of the PHI, I have not
extended PHIOPT to do that just yet (just extended it to handle unary
operations though).
And then we have to see the abs but with _22/_21 being unrelated to p_10/q_12
except with a cast, it might be hard. I am looking into a patch for the case
where if _22/_21 was in the conditional.

[Bug tree-optimization/79119] absolute value of a pointer difference can be assumed to be less than or equal to PTRDIFF_MAX

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79119

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Hmm:
  # RANGE [irange] long unsigned int [0,
2305843009213693951][16140901064495857664, +INF]
  iftmp.0_7 = (long unsigned intD.10) _19;
  if (iftmp.0_7 > 2305843009213693951)


So that is _19 < 0 If I read this correctly.
The only reason why it was not converted is because of the /4 that was before.
So we have a hole in the range.

[Bug tree-optimization/79119] absolute value of a pointer difference can be assumed to be less than or equal to PTRDIFF_MAX

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79119

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=94274

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #3)
> So a couple of things need to happen really:
> optimized now:
>   if (p_10 < q_12)
>     goto <bb 3>; [50.00%]
>   else
>     goto <bb 4>; [50.00%]
> 
>   <bb 3> [local count: 536870913]:
>   _3 = _22 - _21;
>   _4 = _3 /[ex] 4;
>   iftmp.0_14 = (long unsigned int) _4;
>   goto <bb 5>; [100.00%]
> 
>   <bb 4> [local count: 536870913]:
>   _5 = _21 - _22;
>   _6 = _5 /[ex] 4;
>   iftmp.0_13 = (long unsigned int) _6;
> 
>   <bb 5> [local count: 1073741824]:
>   # iftmp.0_7 = PHI <iftmp.0_14(3), iftmp.0_13(4)>

Most of that is PR 94274 .