On Wed, Jun 3, 2026 at 12:31 AM Filip Kastl <[email protected]> wrote:
>
> Hi.
>
> Thanks for doing this!  This was on my TODO list for some time.  I didn't
> realize it was as simple as this.
>
>
> Just something I've noticed: When both lhs and rhs have alignment/range info,
> maybe_duplicate_ssa_info_at_copy doesn't do anything.  So we lose any
> additional info asociated with lhs.  I assume that this also happens in other
> copy propagation passes.  I wonder if we could perhaps intersect the infos in
> that case.

Yes, that is correct. In fact I looked into copyprop to see what was
done and maybe_duplicate_ssa_info_at_copy was the only thing used.
forwprop does that only too (see fwprop_set_lattice_val).

>
> But then I guess a situation like this could happen:
>
> unsigned x, y;
> x = getchar();
> if (x <= 4)
>     error();
> y = x;
> if (y <= 4)
>     <do something>
> <never use y again>
>
> GCC maybe computes range of y to be [0,4]?  Then if we copy propagate y = x
> and intersected the range information, we would get that range of x is [0,4]
> but that is not true and will lead to a miscompilation.

Right copying that would be bad news.
Anyways with more and more passes using the ranger, global range
information is going to be less useful.

Thanks,
Andrea

>
> Filip
>
> On Tue 2026-06-02 19:06:34, Andrew Pinski wrote:
> > Currently if the ssa info (range or aliasing) is different
> > between the src and dest of a copy, sccopy will reject it.
> > This fixes the FIXME and handles it in a similar way as copyprop
> > handles it.  This is needed to able to move sccopy earlier (before phiopt1)
> > and to act like a copyprop pass.
> >
> > Bootstrapped and tested on x86_64-linux-gnu with no regressions.
> >
> > gcc/ChangeLog:
> >
> >       * gimple-ssa-sccopy.cc (stmt_may_generate_copy): Remove restriction
> >       on the ssa info being different for the "copy".
> >       (scc_copy_prop::replace_scc_by_value): Call 
> > maybe_duplicate_ssa_info_at_copy
> >       when copyproping a ssa name.
> >
> > Signed-off-by: Andrew Pinski <[email protected]>
> > ---
> >  gcc/gimple-ssa-sccopy.cc | 18 ++----------------
> >  1 file changed, 2 insertions(+), 16 deletions(-)
> >
> > diff --git a/gcc/gimple-ssa-sccopy.cc b/gcc/gimple-ssa-sccopy.cc
> > index 0badfb11b93..4252e5295df 100644
> > --- a/gcc/gimple-ssa-sccopy.cc
> > +++ b/gcc/gimple-ssa-sccopy.cc
> > @@ -405,22 +405,6 @@ stmt_may_generate_copy (gimple *stmt)
> >    if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs))
> >      return false;
> >
> > -  /* It is possible that lhs has more alignment or value range 
> > information.  By
> > -     propagating we would lose this information.  So in the case that 
> > alignment
> > -     or value range information differs, we are conservative and do not
> > -     propagate.
> > -
> > -     FIXME: Propagate alignment and value range info the same way copy-prop
> > -     does.  */
> > -  if (POINTER_TYPE_P (TREE_TYPE (lhs))
> > -      && POINTER_TYPE_P (TREE_TYPE (rhs))
> > -      && SSA_NAME_PTR_INFO (lhs) != SSA_NAME_PTR_INFO (rhs))
> > -    return false;
> > -  if (!POINTER_TYPE_P (TREE_TYPE (lhs))
> > -      && !POINTER_TYPE_P (TREE_TYPE (rhs))
> > -      && SSA_NAME_RANGE_INFO (lhs) != SSA_NAME_RANGE_INFO (rhs))
> > -    return false;
> > -
> >    return true;  /* A statement of type _2 = _1;.  */
> >  }
> >
> > @@ -497,6 +481,8 @@ scc_copy_prop::replace_scc_by_value (vec<gimple *> scc, 
> > tree val)
> >
> >       }
> >        replace_uses_by (name, val);
> > +      if (TREE_CODE (val) == SSA_NAME)
> > +        maybe_duplicate_ssa_info_at_copy (name, val);
> >        bitmap_set_bit (dead_stmts, SSA_NAME_VERSION (name));
> >        didsomething = true;
> >      }
> > --
> > 2.43.0
> >

Reply via email to