diff options
author | Richard Biener <rguenther@suse.de> | 2021-08-10 10:54:58 +0200 |
---|---|---|
committer | Richard Biener <rguenther@suse.de> | 2021-08-10 12:25:19 +0200 |
commit | 08aa0e3d4f781fd6a6e293bb06d280365a0bdc1d (patch) | |
tree | 9d0e2d205aa98850d8c33f85b869c212a2ce5d3c | |
parent | 7665af0b1a964b1baae3a59b22fcc420369c63cf (diff) |
tree-optimization/101809 - support emulated gather for double[int]
This adds emulated gather support for index vectors with more
elements than the data vector. The internal function gather
vectorization code doesn't currently handle this (but the builtin
decl code does). This allows vectorization of double data gather
with int indexes on 32bit platforms where there isn't an implicit
widening to 64bit present.
2021-08-10 Richard Biener <rguenther@suse.de>
PR tree-optimization/101809
* tree-vect-stmts.c (get_load_store_type): Allow emulated
gathers with offset vector nunits being a constant multiple
of the data vector nunits.
(vect_get_gather_scatter_ops): Use the appropriate nunits
for the offset vector defs.
(vectorizable_store): Adjust call to
vect_get_gather_scatter_ops.
(vectorizable_load): Likewise. Handle the case of less
offset vectors than data vectors.
-rw-r--r-- | gcc/tree-vect-stmts.c | 47 |
1 files changed, 30 insertions, 17 deletions
diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c index 5a5a4dab3f2..ab402b57fb4 100644 --- a/gcc/tree-vect-stmts.c +++ b/gcc/tree-vect-stmts.c @@ -2377,9 +2377,11 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, return false; } else if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant () - || !known_eq (TYPE_VECTOR_SUBPARTS (vectype), - TYPE_VECTOR_SUBPARTS - (gs_info->offset_vectype))) + || !TYPE_VECTOR_SUBPARTS + (gs_info->offset_vectype).is_constant () + || !constant_multiple_p (TYPE_VECTOR_SUBPARTS + (gs_info->offset_vectype), + TYPE_VECTOR_SUBPARTS (vectype))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, @@ -2928,11 +2930,10 @@ vect_build_gather_load_calls (vec_info *vinfo, stmt_vec_info stmt_info, containing loop. */ static void -vect_get_gather_scatter_ops (vec_info *vinfo, +vect_get_gather_scatter_ops (loop_vec_info loop_vinfo, class loop *loop, stmt_vec_info stmt_info, gather_scatter_info *gs_info, - tree *dataref_ptr, vec<tree> *vec_offset, - unsigned ncopies) + tree *dataref_ptr, vec<tree> *vec_offset) { gimple_seq stmts = NULL; *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); @@ -2943,8 +2944,10 @@ vect_get_gather_scatter_ops (vec_info *vinfo, new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); gcc_assert (!new_bb); } - vect_get_vec_defs_for_operand (vinfo, stmt_info, ncopies, gs_info->offset, - vec_offset, gs_info->offset_vectype); + unsigned ncopies = vect_get_num_copies (loop_vinfo, gs_info->offset_vectype); + vect_get_vec_defs_for_operand (loop_vinfo, stmt_info, ncopies, + gs_info->offset, vec_offset, + gs_info->offset_vectype); } /* Prepare to implement a grouped or strided load or store using @@ -8072,8 +8075,9 @@ vectorizable_store (vec_info *vinfo, } else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) { - vect_get_gather_scatter_ops (vinfo, loop, stmt_info, &gs_info, - &dataref_ptr, &vec_offsets, ncopies); + vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info, + &gs_info, &dataref_ptr, + &vec_offsets); vec_offset = vec_offsets[0]; } else @@ -9376,9 +9380,9 @@ vectorizable_load (vec_info *vinfo, } else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) { - vect_get_gather_scatter_ops (vinfo, loop, stmt_info, &gs_info, - &dataref_ptr, &vec_offsets, ncopies); - vec_offset = vec_offsets[0]; + vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info, + &gs_info, &dataref_ptr, + &vec_offsets); } else dataref_ptr @@ -9395,9 +9399,7 @@ vectorizable_load (vec_info *vinfo, if (dataref_offset) dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump); - else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) - vec_offset = vec_offsets[j]; - else + else if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)) dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, stmt_info, bump); if (mask) @@ -9490,6 +9492,7 @@ vectorizable_load (vec_info *vinfo, if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.ifn != IFN_LAST) { + vec_offset = vec_offsets[j]; tree zero = build_zero_cst (vectype); tree scale = size_int (gs_info.scale); gcall *call; @@ -9512,9 +9515,18 @@ vectorizable_load (vec_info *vinfo, gcc_assert (!final_mask); unsigned HOST_WIDE_INT const_nunits = nunits.to_constant (); + unsigned HOST_WIDE_INT const_offset_nunits + = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype) + .to_constant (); vec<constructor_elt, va_gc> *ctor_elts; vec_alloc (ctor_elts, const_nunits); gimple_seq stmts = NULL; + /* We support offset vectors with more elements + than the data vector for now. */ + unsigned HOST_WIDE_INT factor + = const_offset_nunits / const_nunits; + vec_offset = vec_offsets[j / factor]; + unsigned elt_offset = (j % factor) * const_nunits; tree idx_type = TREE_TYPE (TREE_TYPE (vec_offset)); tree scale = size_int (gs_info.scale); align @@ -9525,7 +9537,8 @@ vectorizable_load (vec_info *vinfo, { tree boff = size_binop (MULT_EXPR, TYPE_SIZE (idx_type), - bitsize_int (k)); + bitsize_int + (k + elt_offset)); tree idx = gimple_build (&stmts, BIT_FIELD_REF, idx_type, vec_offset, TYPE_SIZE (idx_type), |