Hi, When I went and split the vectorizer cost model to support thunderx2t99, I did not go back and fix up thunderx to be correct after the split. This patch fixes the problem by changing the vector cost for fp to be the same as scalar fp cost which is good enough and brings us back to where we were with GCC 6.
Thanks, Andrew Pinski ChangeLog: * config/aarch64/aarch64.c (thunderx_vector_cost): Fix vec_fp_stmt_cost.
Index: config/aarch64/aarch64.c =================================================================== --- config/aarch64/aarch64.c (revision 250529) +++ config/aarch64/aarch64.c (working copy) @@ -390,7 +390,7 @@ static const struct cpu_vector_cost thun 3, /* scalar_load_cost */ 1, /* scalar_store_cost */ 4, /* vec_int_stmt_cost */ - 4, /* vec_fp_stmt_cost */ + 1, /* vec_fp_stmt_cost */ 4, /* vec_permute_cost */ 2, /* vec_to_scalar_cost */ 2, /* scalar_to_vec_cost */
