We may not classify a BB vectorization load as VMAT_ELEMENTWISE as that will ICE. Instead we build vectors from existing scalar loads. Make that explicit.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed. * tree-vect-stmts.cc (get_load_store_type): Explicitly fail when we end up with VMAT_ELEMENTWISE for BB vectorization. --- gcc/tree-vect-stmts.cc | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc index bdedb25efe9..cfc4f323a22 100644 --- a/gcc/tree-vect-stmts.cc +++ b/gcc/tree-vect-stmts.cc @@ -2492,6 +2492,11 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, "not falling back to elementwise accesses\n"); return false; } + + /* For BB vectorization build up the vector from existing scalar defs. */ + if (!loop_vinfo && *memory_access_type == VMAT_ELEMENTWISE) + return false; + return true; } -- 2.51.0