#include "coretypes.h"
#include "dumpfile.h"
#include "tm.h"
-#include "input.h"
#include "alias.h"
#include "symtab.h"
#include "tree.h"
#include "tree-ssa-alias.h"
#include "internal-fn.h"
#include "gimple-expr.h"
-#include "is-a.h"
#include "gimple.h"
#include "gimple-iterator.h"
#include "gimple-ssa.h"
enum tree_code first_cond_code = ERROR_MARK;
tree lhs;
bool need_same_oprnds = false;
- tree vectype, scalar_type, first_op1 = NULL_TREE;
+ tree vectype = NULL_TREE, scalar_type, first_op1 = NULL_TREE;
optab optab;
int icode;
machine_mode optab_op2_mode;
machine_mode vec_mode;
- struct data_reference *first_dr;
HOST_WIDE_INT dummy;
- gimple first_load = NULL, prev_first_load = NULL, old_first_load = NULL;
+ gimple first_load = NULL, prev_first_load = NULL;
tree cond;
/* For every stmt in NODE find its def stmt/s. */
return false;
}
- old_first_load = first_load;
first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt));
if (prev_first_load)
{
}
else
prev_first_load = first_load;
-
- /* In some cases a group of loads is just the same load
- repeated N times. Only analyze its cost once. */
- if (first_load == stmt && old_first_load != first_load)
- {
- first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt));
- if (vect_supportable_dr_alignment (first_dr, false)
- == dr_unaligned_unsupported)
- {
- if (dump_enabled_p ())
- {
- dump_printf_loc (MSG_MISSED_OPTIMIZATION,
- vect_location,
- "Build SLP failed: unsupported "
- "unaligned load ");
- dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM,
- stmt, 0);
- dump_printf (MSG_MISSED_OPTIMIZATION, "\n");
- }
- /* Fatal mismatch. */
- matches[0] = false;
- return false;
- }
- }
}
} /* Grouped access. */
else
for (j = 0; j < group_size; ++j)
if (!matches[j])
{
- gimple tem = oprnds_info[0]->def_stmts[j];
- oprnds_info[0]->def_stmts[j] = oprnds_info[1]->def_stmts[j];
- oprnds_info[1]->def_stmts[j] = tem;
+ std::swap (oprnds_info[0]->def_stmts[j],
+ oprnds_info[1]->def_stmts[j]);
dump_printf (MSG_NOTE, "%d ", j);
}
dump_printf (MSG_NOTE, "\n");
}
+/* Attempt to reorder stmts in a reduction chain so that we don't
+ require any load permutation. Return true if that was possible,
+ otherwise return false. */
+
+static bool
+vect_attempt_slp_rearrange_stmts (slp_instance slp_instn)
+{
+ unsigned int group_size = SLP_INSTANCE_GROUP_SIZE (slp_instn);
+ unsigned int i, j;
+ sbitmap load_index;
+ unsigned int lidx;
+ slp_tree node, load;
+
+ /* Compare all the permutation sequences to the first one. We know
+ that at least one load is permuted. */
+ node = SLP_INSTANCE_LOADS (slp_instn)[0];
+ if (!node->load_permutation.exists ())
+ return false;
+ for (i = 1; SLP_INSTANCE_LOADS (slp_instn).iterate (i, &load); ++i)
+ {
+ if (!load->load_permutation.exists ())
+ return false;
+ FOR_EACH_VEC_ELT (load->load_permutation, j, lidx)
+ if (lidx != node->load_permutation[j])
+ return false;
+ }
+
+ /* Check that the loads in the first sequence are different and there
+ are no gaps between them. */
+ load_index = sbitmap_alloc (group_size);
+ bitmap_clear (load_index);
+ FOR_EACH_VEC_ELT (node->load_permutation, i, lidx)
+ {
+ if (bitmap_bit_p (load_index, lidx))
+ {
+ sbitmap_free (load_index);
+ return false;
+ }
+ bitmap_set_bit (load_index, lidx);
+ }
+ for (i = 0; i < group_size; i++)
+ if (!bitmap_bit_p (load_index, i))
+ {
+ sbitmap_free (load_index);
+ return false;
+ }
+ sbitmap_free (load_index);
+
+ /* This permutation is valid for reduction. Since the order of the
+ statements in the nodes is not important unless they are memory
+ accesses, we can rearrange the statements in all the nodes
+ according to the order of the loads. */
+ vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn), group_size,
+ node->load_permutation);
+
+ /* We are done, no actual permutations need to be generated. */
+ FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
+ SLP_TREE_LOAD_PERMUTATION (node).release ();
+ return true;
+}
+
/* Check if the required load permutations in the SLP instance
SLP_INSTN are supported. */
{
unsigned int group_size = SLP_INSTANCE_GROUP_SIZE (slp_instn);
unsigned int i, j, k, next;
- sbitmap load_index;
slp_tree node;
gimple stmt, load, next_load, first_load;
struct data_reference *dr;
stmt = SLP_TREE_SCALAR_STMTS (node)[0];
/* Reduction (there are no data-refs in the root).
- In reduction chain the order of the loads is important. */
+ In reduction chain the order of the loads is not important. */
if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt))
&& !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
{
- slp_tree load;
- unsigned int lidx;
-
- /* Compare all the permutation sequences to the first one. We know
- that at least one load is permuted. */
- node = SLP_INSTANCE_LOADS (slp_instn)[0];
- if (!node->load_permutation.exists ())
- return false;
- for (i = 1; SLP_INSTANCE_LOADS (slp_instn).iterate (i, &load); ++i)
- {
- if (!load->load_permutation.exists ())
- return false;
- FOR_EACH_VEC_ELT (load->load_permutation, j, lidx)
- if (lidx != node->load_permutation[j])
- return false;
- }
+ if (vect_attempt_slp_rearrange_stmts (slp_instn))
+ return true;
- /* Check that the loads in the first sequence are different and there
- are no gaps between them. */
- load_index = sbitmap_alloc (group_size);
- bitmap_clear (load_index);
- FOR_EACH_VEC_ELT (node->load_permutation, i, lidx)
- {
- if (bitmap_bit_p (load_index, lidx))
- {
- sbitmap_free (load_index);
- return false;
- }
- bitmap_set_bit (load_index, lidx);
- }
- for (i = 0; i < group_size; i++)
- if (!bitmap_bit_p (load_index, i))
- {
- sbitmap_free (load_index);
- return false;
- }
- sbitmap_free (load_index);
-
- /* This permutation is valid for reduction. Since the order of the
- statements in the nodes is not important unless they are memory
- accesses, we can rearrange the statements in all the nodes
- according to the order of the loads. */
- vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn), group_size,
- node->load_permutation);
-
- /* We are done, no actual permutations need to be generated. */
- FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
- SLP_TREE_LOAD_PERMUTATION (node).release ();
- return true;
+ /* Fallthru to general load permutation handling. */
}
/* In basic block vectorization we allow any subchain of an interleaving
|| VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (use_vinfo)))
&& !(gimple_code (use_stmt) == GIMPLE_PHI
&& STMT_VINFO_DEF_TYPE (use_vinfo) == vect_reduction_def))
- stype = hybrid;
+ {
+ if (dump_enabled_p ())
+ {
+ dump_printf_loc (MSG_NOTE, vect_location, "use of SLP "
+ "def in non-SLP stmt: ");
+ dump_gimple_stmt (MSG_NOTE, TDF_SLIM, use_stmt, 0);
+ }
+ stype = hybrid;
+ }
}
}
- if (stype == hybrid)
+ if (stype == hybrid
+ && !HYBRID_SLP_STMT (stmt_vinfo))
{
if (dump_enabled_p ())
{
the created stmts must be inserted. */
static inline void
-vect_create_mask_and_perm (gimple stmt, gimple next_scalar_stmt,
+vect_create_mask_and_perm (gimple stmt,
tree mask, int first_vec_indx, int second_vec_indx,
gimple_stmt_iterator *gsi, slp_tree node,
tree vectype, vec<tree> dr_chain,
{
tree perm_dest;
gimple perm_stmt = NULL;
- stmt_vec_info next_stmt_info;
int i, stride;
tree first_vec, second_vec, data_ref;
first_vec_indx += stride;
second_vec_indx += stride;
}
-
- /* Mark the scalar stmt as vectorized. */
- next_stmt_info = vinfo_for_stmt (next_scalar_stmt);
- STMT_VINFO_VEC_STMT (next_stmt_info) = perm_stmt;
}
gimple stmt = SLP_TREE_SCALAR_STMTS (node)[0];
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
tree mask_element_type = NULL_TREE, mask_type;
- int i, j, k, nunits, vec_index = 0, scalar_index;
+ int i, j, k, nunits, vec_index = 0;
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
- gimple next_scalar_stmt;
int group_size = SLP_INSTANCE_GROUP_SIZE (slp_node_instance);
int first_mask_element;
int index, unroll_factor, current_mask_element, ncopies;
bool needs_first_vector = false;
machine_mode mode;
+ if (!STMT_VINFO_GROUPED_ACCESS (stmt_info))
+ return false;
+
+ stmt_info = vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info));
+
mode = TYPE_MODE (vectype);
if (!can_vec_perm_p (mode, false, NULL))
/* The number of vector stmts to generate based only on SLP_NODE_INSTANCE
unrolling factor. */
- orig_vec_stmts_num = group_size *
- SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance) / nunits;
+ orig_vec_stmts_num
+ = (STMT_VINFO_GROUP_SIZE (stmt_info)
+ * SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance)
+ + nunits - 1) / nunits;
if (orig_vec_stmts_num == 1)
only_one_vec = true;
relatively to SLP_NODE_INSTANCE unrolling factor. */
ncopies = vf / SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance);
- if (!STMT_VINFO_GROUPED_ACCESS (stmt_info))
- return false;
-
- stmt_info = vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info));
-
/* Generate permutation masks for every NODE. Number of masks for each NODE
is equal to GROUP_SIZE.
E.g., we have a group of three nodes with three loads from the same
{c2,a3,b3,c3}. */
{
- scalar_index = 0;
index = 0;
vect_stmts_counter = 0;
vec_index = 0;
second_vec_index = vec_index;
}
- next_scalar_stmt
- = SLP_TREE_SCALAR_STMTS (node)[scalar_index++];
-
- vect_create_mask_and_perm (stmt, next_scalar_stmt,
+ vect_create_mask_and_perm (stmt,
mask_vec, first_vec_index, second_vec_index,
gsi, node, vectype, dr_chain,
ncopies, vect_stmts_counter++);