[Bug tree-optimization/57361] New: Remove self memory assignment

[Bug tree-optimization/57361] New: Remove self memory assignment

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57361

            Bug ID: 57361
           Summary: Remove self memory assignment
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org

Hello,

as noticed in PR 57303, tree level optimizations don't remove self memory
assignments:

typedef struct A{long l,m,n,o,p;} A;
void f(A*x){
  *x=*x;
}

compiled with -O3 still has the assignment in the .optimized dump.

I suggested removing statements where rhs is tcc_reference and
operand_equal_p(lhs,rhs), but Richard warns: "The obvious place would be dead
store elimination.  But beware that the above, if not _literally_ being *p = *p
can be changing the effective type of the memory location and thus might be not
dead in terms of type-based aliasing rules (which basically means that we need
to do sth more clever than just removing the store)."

[Bug tree-optimization/57361] Remove self memory assignment

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57361

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-05-22
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Assignment expansion makes this optimization:

  /* Optimize away no-op moves without side-effects.  */
  if (operand_equal_p (to, from, 0))
    return;

The operand_equal_p implementation looks safe at the moment (even in terms
of type-based aliasing rules).

I'd have done a more aggressive optimization in DSE, doing

  if (stmt_kills_ref_p (stmt, gimple_assign_rhs1 (stmt)))
    ... must be a self-assignment ...

which would definitely need extra care for self-assignments that change
the effective type of a memory location.  Though I have a hard time
creating a testcase that would exhibit sth like

  MEM[(struct X *)q_1] = MEM[(struct Y *)q_1];

which would change the dynamic type of *q_1 to X from Y.

[Bug tree-optimization/57361] Remove self memory assignment

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57361

--- Comment #2 from Marc Glisse <glisse at gcc dot gnu.org> ---
I'll let Richard decide if this should stay open for the more aggressive
optimization or not.

r200034 | glisse | 2013-06-12 19:56:30 +0200 (Wed, 12 Jun 2013) | 10 lines

2013-06-12  Marc Glisse  <marc.glisse@inria.fr>

        PR tree-optimization/57361
gcc/
        * tree-ssa-dse.c (dse_possible_dead_store_p): Handle self-assignment.

gcc/testsuite/
        * gcc.dg/tree-ssa/pr57361.c: New file.