https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118924
--- Comment #15 from rguenther at suse dot de <rguenther at suse dot de> --- On Mon, 3 Mar 2025, jamborm at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118924 > > --- Comment #14 from Martin Jambor <jamborm at gcc dot gnu.org> --- > So something like the following - which is completely untested, the > type test may be a wrong one, I'd like to think this through a little > more before actually proposing this, but any comments still welcome: There is refs_same_for_tbaa_p and I believe we have some other compatibility function I can't find right now done by Honza for path-based analysis which should be used by ICF and friends. > diff --git a/gcc/tree-sra.cc b/gcc/tree-sra.cc > index c26559edc66..88b350800ce 100644 > --- a/gcc/tree-sra.cc > +++ b/gcc/tree-sra.cc > @@ -979,6 +979,7 @@ create_access (tree expr, gimple *stmt, bool write) > access->type = TREE_TYPE (expr); > access->write = write; > access->grp_unscalarizable_region = unscalarizable_region; > + access->grp_same_access_path = true; > access->stmt = stmt; > access->reverse = reverse; > > @@ -1522,6 +1523,10 @@ build_accesses_from_assign (gimple *stmt) > racc = build_access_from_expr_1 (rhs, stmt, false); > lacc = build_access_from_expr_1 (lhs, stmt, true); > > + bool tbaa_hazard > + = (TYPE_MAIN_VARIANT (TREE_TYPE (lhs)) > + == TYPE_MAIN_VARIANT (TREE_TYPE (rhs))); > + I think you'd want to formulate this in a TBAA way instead. Since we constrain the base objects to be decls and the access path to contain no V_C_Es it should be reasonable to do tbaa_hazard = !alias_sets_conflict_p (get_alias_set (lhs), get_alias_set (rhs)); which is a simplified form of what refs_same_for_tbaa_p checks but which is not in any way flow-sensitive (refs_same_for_tbaa_p is). If we then have a similar check when we process the access tree of 'lhs' or 'rhs', that is, when we have two different typed accesses to [0, 8] of lhs for example, then I think it should be "complete". > if (lacc) > { > lacc->grp_assignment_write = 1; > @@ -1536,6 +1541,8 @@ build_accesses_from_assign (gimple *stmt) > bitmap_set_bit (cannot_scalarize_away_bitmap, > DECL_UID (lacc->base)); > } > + if (tbaa_hazard) > + lacc->grp_same_access_path = false; > } > > if (racc) > @@ -1555,6 +1562,8 @@ build_accesses_from_assign (gimple *stmt) > } > if (storage_order_barrier_p (lhs)) > racc->grp_unscalarizable_region = 1; > + if (tbaa_hazard) > + racc->grp_same_access_path = false; > } > > if (lacc && racc > @@ -2396,7 +2405,7 @@ sort_and_splice_var_accesses (tree var) > bool grp_partial_lhs = access->grp_partial_lhs; > bool first_scalar = is_gimple_reg_type (access->type); > bool unscalarizable_region = access->grp_unscalarizable_region; > - bool grp_same_access_path = true; > + bool grp_same_access_path = access->grp_same_access_path; > bool bf_non_full_precision > = (INTEGRAL_TYPE_P (access->type) > && TYPE_PRECISION (access->type) != access->size > @@ -2432,7 +2441,8 @@ sort_and_splice_var_accesses (tree var) > return NULL; > } > > - grp_same_access_path = path_comparable_for_same_access (access->expr); > + if (grp_same_access_path) > + grp_same_access_path = path_comparable_for_same_access (access->expr); > > j = i + 1; > while (j < access_count) > >
