On Thu, 13 Jul 2017, Jakub Jelinek wrote:

> Hi!
> 
> As mentioned in the PR, for aggregate copies we fail to verify there
> are no loads in between the PHI and the aggregate copy that could alias with 
> the
> lhs of the copy, which is needed, because we want to hoist the aggregate
> copy onto predecessor edges of the bb with the PHI.
> 
> The following patch implements that, bootstrapped/regtested on x86_64-linux
> and i686-linux, ok for trunk and after a while 7.x?
> 
> 2017-07-13  Jakub Jelinek  <ja...@redhat.com>
> 
>       PR tree-optimization/81365
>       * tree-ssa-phiprop.c (propagate_with_phi): When considering hoisting
>       aggregate moves onto bb predecessor edges, make sure there are no
>       loads that could alias the lhs in between the start of bb and the
>       loads from *phi.  If there are any debug stmts that could alias,
>       reset them.
> 
>       * g++.dg/torture/pr81365.C: New test.
> 
> --- gcc/tree-ssa-phiprop.c.jj 2017-05-22 10:50:11.000000000 +0200
> +++ gcc/tree-ssa-phiprop.c    2017-07-11 16:52:41.012340615 +0200
> @@ -327,7 +327,7 @@ propagate_with_phi (basic_block bb, gphi
>        if (!dominated_by_p (CDI_POST_DOMINATORS,
>                          bb, gimple_bb (use_stmt)))
>       continue;
> -         
> +
>        /* Check whether this is a load of *ptr.  */
>        if (!(is_gimple_assign (use_stmt)
>           && gimple_assign_rhs_code (use_stmt) == MEM_REF
> @@ -356,6 +356,9 @@ propagate_with_phi (basic_block bb, gphi
>           insert aggregate copies on the edges instead.  */
>        if (!is_gimple_reg_type (TREE_TYPE (TREE_TYPE (ptr))))
>       {
> +       if (!gimple_vdef (use_stmt))
> +         goto next;
> +
>         /* As we replicate the lhs on each incoming edge all
>            used SSA names have to be available there.  */
>         if (! for_each_index (gimple_assign_lhs_ptr (use_stmt),
> @@ -363,6 +366,51 @@ propagate_with_phi (basic_block bb, gphi
>                               get_immediate_dominator (CDI_DOMINATORS,
>                                                        gimple_bb (phi))))
>           goto next;
> +
> +       gimple *vuse_stmt;
> +       imm_use_iterator vui;
> +       use_operand_p vuse_p;
> +       bool debug_use_seen = false;
> +       /* In order to move the aggregate copies earlier, make sure
> +          there are no statements that could read from memory
> +          aliasing the lhs in between the start of bb and use_stmt.
> +          As we require use_stmt to have a VDEF above, loads after
> +          use_stmt will use a different virtual SSA_NAME.  */
> +       FOR_EACH_IMM_USE_FAST (vuse_p, vui, vuse)
> +         {
> +           vuse_stmt = USE_STMT (vuse_p);
> +           if (vuse_stmt == use_stmt)
> +             continue;
> +           if (!dominated_by_p (CDI_DOMINATORS,
> +                                gimple_bb (vuse_stmt), bb))
> +             continue;
> +           if (ref_maybe_used_by_stmt_p (vuse_stmt,
> +                                         gimple_assign_lhs (use_stmt)))
> +             {
> +               if (is_gimple_debug (vuse_stmt))

debug stmts do not have virtual operands so their handling is not
necessary here.

Ok with that change.

Thanks,
Richard.


> +                 debug_use_seen = true;
> +               else
> +                 goto next;
> +             }
> +         }
> +       /* Debug stmt uses should not prevent the transformation, but
> +          if we saw any, reset those debug stmts.  */
> +       if (debug_use_seen)
> +         FOR_EACH_IMM_USE_STMT (vuse_stmt, vui, vuse)
> +           {
> +             if (!is_gimple_debug (vuse_stmt))
> +               continue;
> +             if (!dominated_by_p (CDI_DOMINATORS,
> +                                  gimple_bb (vuse_stmt), bb))
> +               continue;
> +             if (ref_maybe_used_by_stmt_p (vuse_stmt,
> +                                           gimple_assign_lhs (use_stmt)))
> +               {
> +                 gimple_debug_bind_reset_value (vuse_stmt);
> +                 update_stmt (vuse_stmt);
> +               }
> +           }
> +
>         phiprop_insert_phi (bb, phi, use_stmt, phivn, n);
>  
>         /* Remove old stmt.  The phi is taken care of by DCE.  */
> --- gcc/testsuite/g++.dg/torture/pr81365.C.jj 2017-07-11 17:07:11.107130111 
> +0200
> +++ gcc/testsuite/g++.dg/torture/pr81365.C    2017-07-11 17:06:52.000000000 
> +0200
> @@ -0,0 +1,39 @@
> +// PR tree-optimization/81365
> +// { dg-do run }
> +
> +struct A { unsigned a; };
> +
> +struct B {
> +  B (const A *x)
> +  {
> +    __builtin_memcpy (b, x, 3 * sizeof (A));
> +    __builtin_memcpy (c, x + 3, sizeof (A));
> +    __builtin_memset (c + 1, 0, sizeof (A));
> +  }
> +  bool
> +  foo (unsigned x)
> +  {
> +    A *it = c;
> +    if (it->a == x || (++it)->a == x)
> +      {
> +     A t(b[0]);
> +     b[0] = *it;
> +     *it = t;
> +     return true;
> +      }
> +    return false;
> +  }
> +  A b[3];
> +  A c[2];
> +};
> +
> +int
> +main ()
> +{
> +  A x[] = { 4, 8, 12, 18 };
> +  B y(x);
> +  if (!y.foo (18))
> +    __builtin_abort ();
> +  if (!y.foo (4))
> +    __builtin_abort ();
> +}
> 
>       Jakub
> 
> 

-- 
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 
21284 (AG Nuernberg)

Reply via email to