Re: [PATCH] Fix PR 52631 (VN does not use simplified expression for lookup)

[Richard Guenther <richard dot guenther at gmail dot com>]

On Tue, Jul 24, 2012 at 5:50 PM, Andrew Pinski
<andrew.pinski@caviumnetworks.com> wrote:
> Hi,
>   Before tuples was introduced, VN used to lookup the simplified
> expression to see if it was available already and use that instead of
> the non simplified one.  This patch adds the support back to VN to do
> exactly that.
>
> OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.

I think this should be done for all RHS and SSA name LHS, not only
for UNARY/BINARY/TERNARY - because even for SINGLE rhs we
can end up simplifying (for REALPART_EXPR for example which we
handle as nary, too).  I think we constrain try_to_simplify enough
so that

+		      /* First try to lookup the simplified expression. */
+		      if (simplified && valid_gimple_rhs_p (simplified))
+			{
+			  tree result = vn_nary_op_lookup (simplified, NULL);
+			  if (result)
+			    {
+			      changed = set_ssa_val_to (lhs, result);
+			      goto done;
+			    }
+			  changed = set_ssa_val_to (lhs, lhs);
+			  vn_nary_op_insert (simplified, lhs);
+			}
                  switch (get_gimple_rhs_class (code))
                    {
                    case GIMPLE_UNARY_RHS:
                    case GIMPLE_BINARY_RHS:
...

should work.  As you also insert the simplified variant I think we really
(finally) want to have a valid_nary_op routine rather than relying on
valid_gimple_rhs_p which is way too generic.

Thanks,
Richard.

> Thanks,
> Andrew Pinski
>
> ChangeLog:
>
>         * tree-ssa-sccvn.c (visit_use): Look up the simplified
>         expression before visting the original one.
>
>         * gcc.dg/tree-ssa/ssa-fre-9.c: Update the number of
>         eliminatations that happen.