#include "tree.h"
#include "target.h"
#include "basic-block.h"
-#include "diagnostic.h"
+#include "tree-pretty-print.h"
+#include "gimple-pretty-print.h"
#include "tree-flow.h"
#include "tree-dump.h"
#include "cfgloop.h"
gimple stmt = VEC_index (gimple, stmts, 0);
enum vect_def_type first_stmt_dt0 = vect_uninitialized_def;
enum vect_def_type first_stmt_dt1 = vect_uninitialized_def;
- enum tree_code first_stmt_code = ERROR_MARK, rhs_code;
+ enum tree_code first_stmt_code = ERROR_MARK, rhs_code = ERROR_MARK;
tree first_stmt_def1_type = NULL_TREE, first_stmt_def0_type = NULL_TREE;
tree lhs;
bool stop_recursion = false, need_same_oprnds = false;
gimple first_load, prev_first_load = NULL;
/* For every stmt in NODE find its def stmt/s. */
- for (i = 0; VEC_iterate (gimple, stmts, i, stmt); i++)
+ FOR_EACH_VEC_ELT (gimple, stmts, i, stmt)
{
if (vect_print_dump_info (REPORT_SLP))
{
print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
}
+ /* Fail to vectorize statements marked as unvectorizable. */
+ if (!STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt)))
+ {
+ if (vect_print_dump_info (REPORT_SLP))
+ {
+ fprintf (vect_dump,
+ "Build SLP failed: unvectorizable statement ");
+ print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
+ }
+
+ return false;
+ }
+
lhs = gimple_get_lhs (stmt);
if (lhs == NULL_TREE)
{
optab_vector);
if (!optab
- || (optab->handlers[(int) vec_mode].insn_code
- == CODE_FOR_nothing))
+ || optab_handler (optab, vec_mode) == CODE_FOR_nothing)
{
/* No vector/vector shift, try for a vector/scalar shift. */
optab = optab_for_tree_code (rhs_code, vectype,
fprintf (vect_dump, "Build SLP failed: no optab.");
return false;
}
- icode = (int) optab->handlers[(int) vec_mode].insn_code;
+ icode = (int) optab_handler (optab, vec_mode);
if (icode == CODE_FOR_nothing)
{
if (vect_print_dump_info (REPORT_SLP))
&& (first_stmt_code != IMAGPART_EXPR
|| rhs_code != REALPART_EXPR)
&& (first_stmt_code != REALPART_EXPR
- || rhs_code != IMAGPART_EXPR))
+ || rhs_code != IMAGPART_EXPR)
+ && !(STMT_VINFO_STRIDED_ACCESS (vinfo_for_stmt (stmt))
+ && (first_stmt_code == ARRAY_REF
+ || first_stmt_code == INDIRECT_REF
+ || first_stmt_code == COMPONENT_REF
+ || first_stmt_code == MEM_REF)))
{
if (vect_print_dump_info (REPORT_SLP))
{
if (first_load == stmt)
{
first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt));
- if (vect_supportable_dr_alignment (first_dr)
+ if (vect_supportable_dr_alignment (first_dr, false)
== dr_unaligned_unsupported)
{
if (vect_print_dump_info (REPORT_SLP))
if (permutation)
{
VEC_safe_push (slp_tree, heap, *loads, *node);
- *inside_cost += TARG_VEC_PERMUTE_COST * group_size;
+ *inside_cost
+ += targetm.vectorize.builtin_vectorization_cost (vec_perm, NULL, 0)
+ * group_size;
+ }
+ else
+ {
+ /* We don't check here complex numbers chains, so we keep them in
+ LOADS for further check in vect_supported_load_permutation_p. */
+ if (rhs_code == REALPART_EXPR || rhs_code == IMAGPART_EXPR)
+ VEC_safe_push (slp_tree, heap, *loads, *node);
}
return true;
return;
fprintf (vect_dump, "node ");
- for (i = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt); i++)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt)
{
fprintf (vect_dump, "\n\tstmt %d ", i);
print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
if (!node)
return;
- for (i = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt); i++)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt)
if (j < 0 || i == j)
STMT_SLP_TYPE (vinfo_for_stmt (stmt)) = mark;
if (!node)
return;
- for (i = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt); i++)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt)
{
stmt_info = vinfo_for_stmt (stmt);
gcc_assert (!STMT_VINFO_RELEVANT (stmt_info)
for (i = 0; i < group_size; i++)
VEC_safe_push (gimple, heap, tmp_stmts, NULL);
- for (i = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt); i++)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt)
{
index = VEC_index (int, permutation, i);
VEC_replace (gimple, tmp_stmts, index, stmt);
int i = 0, j, prev = -1, next, k, number_of_groups;
bool supported, bad_permutation = false;
sbitmap load_index;
- slp_tree node;
- gimple stmt;
+ slp_tree node, other_complex_node;
+ gimple stmt, first = NULL, other_node_first;
+ unsigned complex_numbers = 0;
/* FORNOW: permutations are only supported in SLP. */
if (!slp_instn)
if (vect_print_dump_info (REPORT_SLP))
{
fprintf (vect_dump, "Load permutation ");
- for (i = 0; VEC_iterate (int, load_permutation, i, next); i++)
+ FOR_EACH_VEC_ELT (int, load_permutation, i, next)
fprintf (vect_dump, "%d ", next);
}
permutation). */
/* Check that all the load nodes are of the same size. */
- for (i = 0;
- VEC_iterate (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node);
- i++)
- if (VEC_length (gimple, SLP_TREE_SCALAR_STMTS (node))
- != (unsigned) group_size)
- return false;
-
+ FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node)
+ {
+ if (VEC_length (gimple, SLP_TREE_SCALAR_STMTS (node))
+ != (unsigned) group_size)
+ return false;
+
+ stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0);
+ if (is_gimple_assign (stmt)
+ && (gimple_assign_rhs_code (stmt) == REALPART_EXPR
+ || gimple_assign_rhs_code (stmt) == IMAGPART_EXPR))
+ complex_numbers++;
+ }
+
+ /* Complex operands can be swapped as following:
+ real_c = real_b + real_a;
+ imag_c = imag_a + imag_b;
+ i.e., we have {real_b, imag_a} and {real_a, imag_b} instead of
+ {real_a, imag_a} and {real_b, imag_b}. We check here that if interleaving
+ chains are mixed, they match the above pattern. */
+ if (complex_numbers)
+ {
+ FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_instn), i, node)
+ {
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), j, stmt)
+ {
+ if (j == 0)
+ first = stmt;
+ else
+ {
+ if (DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt)) != first)
+ {
+ if (complex_numbers != 2)
+ return false;
+
+ if (i == 0)
+ k = 1;
+ else
+ k = 0;
+
+ other_complex_node = VEC_index (slp_tree,
+ SLP_INSTANCE_LOADS (slp_instn), k);
+ other_node_first = VEC_index (gimple,
+ SLP_TREE_SCALAR_STMTS (other_complex_node), 0);
+
+ if (DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt))
+ != other_node_first)
+ return false;
+ }
+ }
+ }
+ }
+ }
+
+ /* We checked that this case ok, so there is no need to proceed with
+ permutation tests. */
+ if (complex_numbers == 2)
+ {
+ VEC_free (slp_tree, heap, SLP_INSTANCE_LOADS (slp_instn));
+ VEC_free (int, heap, SLP_INSTANCE_LOAD_PERMUTATION (slp_instn));
+ return true;
+ }
+
node = SLP_INSTANCE_TREE (slp_instn);
stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (node), 0);
/* LOAD_PERMUTATION is a list of indices of all the loads of the SLP
slp_tree load_node;
gimple first_load = NULL, load;
- for (i = 0;
- VEC_iterate (slp_tree, SLP_INSTANCE_LOADS (instance), i, load_node);
- i++)
- for (j = 0;
- VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (load_node), j, load);
- j++)
+ FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (instance), i, load_node)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (load_node), j, load)
first_load = get_earlier_stmt (load, first_load);
return first_load;
strided_stores = BB_VINFO_STRIDED_STORES (bb_vinfo);
/* Find SLP sequences starting from groups of strided stores. */
- for (i = 0; VEC_iterate (gimple, strided_stores, i, store); i++)
+ FOR_EACH_VEC_ELT (gimple, strided_stores, i, store)
if (vect_analyze_slp_instance (loop_vinfo, bb_vinfo, store))
ok = true;
}
/* Find SLP sequences starting from groups of reductions. */
- if (loop_vinfo && VEC_length (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo))
+ if (loop_vinfo && VEC_length (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo)) > 1
&& vect_analyze_slp_instance (loop_vinfo, bb_vinfo,
VEC_index (gimple, reductions, 0)))
ok = true;
if (vect_print_dump_info (REPORT_SLP))
fprintf (vect_dump, "=== vect_make_slp_decision ===");
- for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
+ FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
{
/* FORNOW: SLP if you can. */
if (unrolling_factor < SLP_INSTANCE_UNROLLING_FACTOR (instance))
if (!node)
return;
- for (i = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt); i++)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt)
if (PURE_SLP_STMT (vinfo_for_stmt (stmt))
&& TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME)
FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, gimple_op (stmt, 0))
if (vect_print_dump_info (REPORT_SLP))
fprintf (vect_dump, "=== vect_detect_hybrid_slp ===");
- for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
+ FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
vect_detect_hybrid_slp_stmts (SLP_INSTANCE_TREE (instance));
}
|| !vect_slp_analyze_node_operations (bb_vinfo, SLP_TREE_RIGHT (node)))
return false;
- for (i = 0; VEC_iterate (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt); i++)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_SCALAR_STMTS (node), i, stmt)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
gcc_assert (stmt_info);
}
-/* Cheick if the basic block can be vectorized. */
+/* Check if vectorization of the basic block is profitable. */
+
+static bool
+vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo)
+{
+ VEC (slp_instance, heap) *slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo);
+ slp_instance instance;
+ int i;
+ unsigned int vec_outside_cost = 0, vec_inside_cost = 0, scalar_cost = 0;
+ unsigned int stmt_cost;
+ gimple stmt;
+ gimple_stmt_iterator si;
+ basic_block bb = BB_VINFO_BB (bb_vinfo);
+ stmt_vec_info stmt_info = NULL;
+ tree dummy_type = NULL;
+ int dummy = 0;
+
+ /* Calculate vector costs. */
+ FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
+ {
+ vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance);
+ vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance);
+ }
+
+ /* Calculate scalar cost. */
+ for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
+ {
+ stmt = gsi_stmt (si);
+ stmt_info = vinfo_for_stmt (stmt);
+
+ if (!stmt_info || !STMT_VINFO_VECTORIZABLE (stmt_info)
+ || !PURE_SLP_STMT (stmt_info))
+ continue;
+
+ if (STMT_VINFO_DATA_REF (stmt_info))
+ {
+ if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
+ stmt_cost = targetm.vectorize.builtin_vectorization_cost
+ (scalar_load, dummy_type, dummy);
+ else
+ stmt_cost = targetm.vectorize.builtin_vectorization_cost
+ (scalar_store, dummy_type, dummy);
+ }
+ else
+ stmt_cost = targetm.vectorize.builtin_vectorization_cost
+ (scalar_stmt, dummy_type, dummy);
+
+ scalar_cost += stmt_cost;
+ }
+
+ if (vect_print_dump_info (REPORT_COST))
+ {
+ fprintf (vect_dump, "Cost model analysis: \n");
+ fprintf (vect_dump, " Vector inside of basic block cost: %d\n",
+ vec_inside_cost);
+ fprintf (vect_dump, " Vector outside of basic block cost: %d\n",
+ vec_outside_cost);
+ fprintf (vect_dump, " Scalar cost of basic block: %d", scalar_cost);
+ }
+
+ /* Vectorization is profitable if its cost is less than the cost of scalar
+ version. */
+ if (vec_outside_cost + vec_inside_cost >= scalar_cost)
+ return false;
+
+ return true;
+}
+
+/* Check if the basic block can be vectorized. */
bb_vec_info
vect_slp_analyze_bb (basic_block bb)
/* Mark all the statements that we want to vectorize as pure SLP and
relevant. */
- for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
+ FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
{
vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1);
vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance));
return NULL;
}
+ /* Cost model: check if the vectorization is worthwhile. */
+ if (flag_vect_cost_model
+ && !vect_bb_vectorization_profitable_p (bb_vinfo))
+ {
+ if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
+ fprintf (vect_dump, "not vectorized: vectorization is not "
+ "profitable.\n");
+
+ destroy_bb_vec_info (bb_vinfo);
+ return NULL;
+ }
+
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "Basic block will be vectorized using SLP\n");
if (vect_print_dump_info (REPORT_SLP))
fprintf (vect_dump, "=== vect_update_slp_costs_according_to_vf ===");
- for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
+ FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
/* We assume that costs are linear in ncopies. */
SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance) *= vf
/ SLP_INSTANCE_UNROLLING_FACTOR (instance);
break;
case MULT_EXPR:
- case BIT_AND_EXPR:
if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op)))
neutral_op = build_real (TREE_TYPE (op), dconst1);
else
break;
+ case BIT_AND_EXPR:
+ neutral_op = build_int_cst (TREE_TYPE (op), -1);
+ break;
+
default:
neutral_op = NULL;
}
gcc_assert (SLP_TREE_VEC_STMTS (slp_node));
- for (i = 0;
- VEC_iterate (gimple, SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt);
- i++)
+ FOR_EACH_VEC_ELT (gimple, SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt)
{
gcc_assert (vec_def_stmt);
vec_oprnd = gimple_get_lhs (vec_def_stmt);
stmt_vec_info next_stmt_info;
int i, stride;
tree first_vec, second_vec, data_ref;
- VEC (tree, heap) *params = NULL;
stride = SLP_TREE_NUMBER_OF_VEC_STMTS (node) / ncopies;
first_vec = VEC_index (tree, dr_chain, first_vec_indx);
second_vec = VEC_index (tree, dr_chain, second_vec_indx);
- /* Build argument list for the vectorized call. */
- VEC_free (tree, heap, params);
- params = VEC_alloc (tree, heap, 3);
- VEC_quick_push (tree, params, first_vec);
- VEC_quick_push (tree, params, second_vec);
- VEC_quick_push (tree, params, mask);
-
/* Generate the permute statement. */
- perm_stmt = gimple_build_call_vec (builtin_decl, params);
+ perm_stmt = gimple_build_call (builtin_decl,
+ 3, first_vec, second_vec, mask);
data_ref = make_ssa_name (perm_dest, perm_stmt);
gimple_call_set_lhs (perm_stmt, data_ref);
vect_finish_stmt_generation (stmt, perm_stmt, gsi);
we need the second and the third vectors: {b1,c1,a2,b2} and
{c2,a3,b3,c3}. */
- for (i = 0;
- VEC_iterate (slp_tree, SLP_INSTANCE_LOADS (slp_node_instance),
- i, node);
- i++)
+ FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (slp_node_instance), i, node)
{
scalar_index = 0;
index = 0;
all the nodes that participate in that permutation. */
if (SLP_INSTANCE_LOAD_PERMUTATION (instance))
{
- for (i = 0;
- VEC_iterate (slp_tree, SLP_INSTANCE_LOADS (instance), i, loads_node);
- i++)
+ FOR_EACH_VEC_ELT (slp_tree, SLP_INSTANCE_LOADS (instance), i, loads_node)
{
if (!SLP_TREE_VEC_STMTS (loads_node))
{
vf = 1;
}
- for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
+ FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
{
/* Schedule the tree of INSTANCE. */
is_store = vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance),
fprintf (vect_dump, "vectorizing stmts using SLP.");
}
- for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
+ FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
{
slp_tree root = SLP_INSTANCE_TREE (instance);
gimple store;
OSDN Git Service
