[Bug tree-optimization/90328] New: Wrong loop distribution with aliasing

[glisse at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90328

            Bug ID: 90328
           Summary: Wrong loop distribution with aliasing
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org
  Target Milestone: ---

I was investigating missed loop distribution opportunities probably related to
aliasing, and hit this case where gcc seems a bit too optimistic.

#include <new>
struct T { int* p; T(T const&t):p(t.p){} };
T* f(T* a,T* b){
    for(int i=0;i<1024;++i){
        a[i].~T();
        new(a+i)T(b[i]);
    }
    return a;
}

This gets optimized to a call to memcpy.
However, I believe it is legal to call f(x+1,x) where x is an array of size
1025, where it should replace every element with a copy of the first one.
This seems related to "new" and the use of a constructor. It is true that this
implies that a[i] and b[i] cannot alias, but not that the whole a and b cannot
alias.