https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110204
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #5)
> Yeah, the issue is that PRE figures out a new value here but we've already
> done value-numbering so we can't alter the "old" solution here. So what we
> do
> is add a '0' with value '_42' (instead of value '0'). This "second order"
> value numbering leaves more opportunities on the plate. It was IMHO
> cleaner to insert a
>
> pretmp_163 = 0;
>
> than substituting '0' everywhere but then leaving around the third order
> optimizations in case we had _42 + 1 that would then simplify to sth = 1 ...
>
> Previously we'd have inserted a degenerate PHI, now we get these kind of
> copies. "Now" is for a long time so this isn't new for PRE at least.
>
> We could "hack" this by doing
>
> diff --git a/gcc/tree-ssa-sccvn.cc b/gcc/tree-ssa-sccvn.cc
> index 11061a374a2..effb4f4de73 100644
> --- a/gcc/tree-ssa-sccvn.cc
> +++ b/gcc/tree-ssa-sccvn.cc
> @@ -6615,6 +6615,13 @@ eliminate_dom_walker::eliminate_push_avail
> (basic_block, tree op)
> if (avail[SSA_NAME_VERSION (valnum)])
> pushop = avail[SSA_NAME_VERSION (valnum)];
> avail_stack.safe_push (pushop);
> + if (gassign *ass = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (op)))
> + if (gimple_assign_rhs_class (ass) == GIMPLE_SINGLE_RHS)
> + {
> + tree rhs1 = gimple_assign_rhs1 (ass);
> + if (CONSTANT_CLASS_P (rhs1) || TREE_CODE (rhs1) == SSA_NAME)
> + op = rhs1;
> + }
> avail[SSA_NAME_VERSION (valnum)] = op;
> }
> }
Ah, that breaks the avail stack handling. We can do the trick in
eliminate_avail though where it's more expensive or make the avail
stack more complicated.
