[Bug tree-optimization/82697] [6/7/8 Regression] Wrong optimization with aliasing and "if"

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82697

Richard Biener <rguenth at gcc dot gnu.org> changed:

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

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
So cselim transforms

  *pi = 1;
  if (f)
    *pl = 2;

into

  *pi = 1;
  if (!f)
    tem = *pl;
  else
    tem = 2;
  *pl = tem;

that is of course only valid in case we can validly load from *pl which we
can't in this case.

Given there's no check for validity here there's nothing else than using
alias-set zero for the load ...

We also have to avoid changing the dynamic type of the object which means
the store _also_ has to use alias-set zero.

I have a fix.