https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118795
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rsandifo at gcc dot gnu.org
--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
With
(simplify
(vec_perm @0 @1 VECTOR_CST@2)
...
we have
/* Generate a canonical form of the selector. */
if (!ins && sel.encoding () != builder)
{
/* Some targets are deficient and fail to expand a single
argument permutation while still allowing an equivalent
2-argument version. */
tree oldop2 = op2;
if (sel.ninputs () == 2
|| can_vec_perm_const_p (result_mode, op_mode, sel, false))
op2 = vec_perm_indices_to_tree (TREE_TYPE (op2), sel);
else
{
vec_perm_indices sel2 (builder, 2, nelts);
if (can_vec_perm_const_p (result_mode, op_mode, sel2,
false))
op2 = vec_perm_indices_to_tree (TREE_TYPE (op2), sel2);
else
/* Not directly supported with either encoding,
so use the preferred form. */
op2 = vec_perm_indices_to_tree (TREE_TYPE (op2), sel);
where the last case can generate a not supported permute and the ninputs == 2
guard is botched. I think sth like the following was intended, but it
doesn't fix the testcase.
diff --git a/gcc/match.pd b/gcc/match.pd
index ad966766376..f5caa472411 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -11043,15 +11043,14 @@ and,
argument permutation while still allowing an equivalent
2-argument version. */
tree oldop2 = op2;
- if (sel.ninputs () == 2
- || can_vec_perm_const_p (result_mode, op_mode, sel, false))
+ if (can_vec_perm_const_p (result_mode, op_mode, sel, false))
op2 = vec_perm_indices_to_tree (TREE_TYPE (op2), sel);
- else
+ else if (sel.ninputs () == 1)
{
vec_perm_indices sel2 (builder, 2, nelts);
if (can_vec_perm_const_p (result_mode, op_mode, sel2,
false))
op2 = vec_perm_indices_to_tree (TREE_TYPE (op2), sel2);
- else
+ else if (optimize_vectors_before_lowering_p ())
/* Not directly supported with either encoding,
so use the preferred form. */
op2 = vec_perm_indices_to_tree (TREE_TYPE (op2), sel);
I'll note that RTL expansion does
/* Always specify two input vectors here and leave the target to handle
cases in which the inputs are equal. Not all backends can cope with
the single-input representation when testing for a double-input
target instruction. */
vec_perm_indices indices (sel, 2, GET_MODE_NUNITS (mode));
so the above target support checks might be "wrong" in this sense if it
matters at all which the above suggests.
