https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122797
--- Comment #7 from Robin Dapp <rdapp at gcc dot gnu.org> ---
Also fails on riscv with e.g.
-O3 -march=rv64gcbv_zvbb_zvl512b -DQEMU -mtune=generic-ooo -mmax-vectorization
--param=riscv-autovec-mode=V4QI
Any specific options for x86 apart from -O3?
The issue is pretty simple/stupid and I believe is the whole reason why I added
the second parameter of
vect_load_perm_consecutive_p (slp_tree node, unsigned start_idx)
in the first place:
@@ -7855,15 +7867,7 @@ vect_optimize_slp_pass::remove_redundant_permutations ()
else
{
loop_vec_info loop_vinfo = as_a<loop_vec_info> (m_vinfo);
- stmt_vec_info load_info;
- bool this_load_permuted = false;
- unsigned j;
- FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
- if (SLP_TREE_LOAD_PERMUTATION (node)[j] != j)
- {
- this_load_permuted = true;
- break;
- }
+ bool this_load_permuted = !vect_load_perm_consecutive_p (node);
This is obviously wrong and what I wanted is that start_idx is not a start
index but a start value. It usually doesn't help to know that a permutation is
consecutive from somewhere in the middle. What is helpful, though, is to know
that it is consecutive starting with value x.
Testing a patch.
