On Fri, Aug 5, 2022 at 2:59 PM Andre Vieira (lists) via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> This isn't really a 'PATCH' yet, it's something I was working on but had
> to put on hold. Feel free to re-use any bits or trash all of it if you'd
> like.
@@ -10264,6 +10264,44 @@ expand_expr_real_2 (sepops ops, rtx target,
machine_mode tmode,
case VEC_PERM_EXPR:
{
+ if (TREE_CODE (treeop2) == VECTOR_CST
+ && targetm.vectorize.vla_constructor)
+ {
+ tree ctor0, ctor1;
+ if (TREE_CODE (treeop0) == SSA_NAME
+ && is_gimple_assign (SSA_NAME_DEF_STMT (treeop0)))
+ ctor0 = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (treeop0));
+ else
+ ctor0 = treeop0;
+ if (TREE_CODE (treeop1) == SSA_NAME
+ && is_gimple_assign (SSA_NAME_DEF_STMT (treeop1)))
+ ctor1 = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (treeop1));
just to say - you can't lookup things like this, you have to go through the TER
machinery, otherwise the expansions for the CTOR elements might be
clobbered already. That means to be fully effective doing this during RTL
expansion is likely limited.
