}
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "\n");
+ /* After swapping some operands we lost track whether an
+ operand has any pattern defs so be conservative here. */
+ if (oprnds_info[0]->any_pattern || oprnds_info[1]->any_pattern)
+ oprnds_info[0]->any_pattern = oprnds_info[1]->any_pattern = true;
/* And try again with scratch 'matches' ... */
bool *tem = XALLOCAVEC (bool, group_size);
if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts,
}
}
+/* Compute the set of scalar stmts participating in internal and external
+ nodes. */
+
+static void
+vect_slp_gather_vectorized_scalar_stmts (vec_info *vinfo, slp_tree node,
+ hash_set<slp_tree> &visited,
+ hash_set<stmt_vec_info> &vstmts,
+ hash_set<stmt_vec_info> &estmts)
+{
+ int i;
+ stmt_vec_info stmt_info;
+ slp_tree child;
+
+ if (visited.add (node))
+ return;
+
+ if (SLP_TREE_DEF_TYPE (node) == vect_internal_def)
+ {
+ FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
+ vstmts.add (stmt_info);
+
+ FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
+ if (child)
+ vect_slp_gather_vectorized_scalar_stmts (vinfo, child, visited,
+ vstmts, estmts);
+ }
+ else
+ for (tree def : SLP_TREE_SCALAR_OPS (node))
+ {
+ stmt_vec_info def_stmt = vinfo->lookup_def (def);
+ if (def_stmt)
+ estmts.add (def_stmt);
+ }
+}
+
+
/* Compute the scalar cost of the SLP node NODE and its children
and return it. Do not account defs that are marked in LIFE and
update LIFE according to uses of NODE. */
vect_bb_slp_scalar_cost (vec_info *vinfo,
slp_tree node, vec<bool, va_heap> *life,
stmt_vector_for_cost *cost_vec,
+ hash_set<stmt_vec_info> &vectorized_scalar_stmts,
hash_set<slp_tree> &visited)
{
unsigned i;
{
stmt_vec_info use_stmt_info = vinfo->lookup_stmt (use_stmt);
if (!use_stmt_info
- || !PURE_SLP_STMT
- (vect_stmt_to_vectorize (use_stmt_info)))
+ || !vectorized_scalar_stmts.contains (use_stmt_info))
{
(*life)[i] = true;
break;
subtree_life.safe_splice (*life);
}
vect_bb_slp_scalar_cost (vinfo, child, &subtree_life, cost_vec,
- visited);
+ vectorized_scalar_stmts, visited);
subtree_life.truncate (0);
}
}
static bool
vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo,
- vec<slp_instance> slp_instances)
+ vec<slp_instance> slp_instances,
+ loop_p orig_loop)
{
slp_instance instance;
int i;
SLP_INSTANCE_TREE (instance), visited);
}
+ /* Compute the set of scalar stmts we know will go away 'locally' when
+ vectorizing. This used to be tracked with just PURE_SLP_STMT but that's
+ not accurate for nodes promoted extern late or for scalar stmts that
+ are used both in extern defs and in vectorized defs. */
+ hash_set<stmt_vec_info> vectorized_scalar_stmts;
+ hash_set<stmt_vec_info> scalar_stmts_in_externs;
+ hash_set<slp_tree> visited;
+ FOR_EACH_VEC_ELT (slp_instances, i, instance)
+ {
+ vect_slp_gather_vectorized_scalar_stmts (bb_vinfo,
+ SLP_INSTANCE_TREE (instance),
+ visited,
+ vectorized_scalar_stmts,
+ scalar_stmts_in_externs);
+ for (stmt_vec_info rstmt : SLP_INSTANCE_ROOT_STMTS (instance))
+ vectorized_scalar_stmts.add (rstmt);
+ }
+ /* Scalar stmts used as defs in external nodes need to be preseved, so
+ remove them from vectorized_scalar_stmts. */
+ for (stmt_vec_info stmt : scalar_stmts_in_externs)
+ vectorized_scalar_stmts.remove (stmt);
+
/* Calculate scalar cost and sum the cost for the vector stmts
previously collected. */
stmt_vector_for_cost scalar_costs = vNULL;
stmt_vector_for_cost vector_costs = vNULL;
- hash_set<slp_tree> visited;
+ visited.empty ();
FOR_EACH_VEC_ELT (slp_instances, i, instance)
{
auto_vec<bool, 20> life;
SLP_INSTANCE_ROOT_STMTS (instance)[0], 0, vect_body);
vect_bb_slp_scalar_cost (bb_vinfo,
SLP_INSTANCE_TREE (instance),
- &life, &scalar_costs, visited);
+ &life, &scalar_costs, vectorized_scalar_stmts,
+ visited);
vector_costs.safe_splice (instance->cost_vec);
instance->cost_vec.release ();
}
- /* Unset visited flag. */
- stmt_info_for_cost *cost;
- FOR_EACH_VEC_ELT (scalar_costs, i, cost)
- gimple_set_visited (cost->stmt_info->stmt, false);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "Cost model analysis: \n");
li_scalar_costs (scalar_costs.length ());
auto_vec<std::pair<unsigned, stmt_info_for_cost *> >
li_vector_costs (vector_costs.length ());
+ stmt_info_for_cost *cost;
FOR_EACH_VEC_ELT (scalar_costs, i, cost)
{
unsigned l = gimple_bb (cost->stmt_info->stmt)->loop_father->num;
/* Now cost the portions individually. */
unsigned vi = 0;
unsigned si = 0;
+ bool profitable = true;
while (si < li_scalar_costs.length ()
&& vi < li_vector_costs.length ())
{
example). */
if (vec_outside_cost + vec_inside_cost > scalar_cost)
{
- scalar_costs.release ();
- vector_costs.release ();
- return false;
+ profitable = false;
+ break;
}
}
- if (vi < li_vector_costs.length ())
+ if (profitable && vi < li_vector_costs.length ())
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Excess vector cost for part in loop %d:\n",
li_vector_costs[vi].first);
- scalar_costs.release ();
- vector_costs.release ();
- return false;
+ profitable = false;
}
+ /* Unset visited flag. This is delayed when the subgraph is profitable
+ and we process the loop for remaining unvectorized if-converted code. */
+ if (!orig_loop || !profitable)
+ FOR_EACH_VEC_ELT (scalar_costs, i, cost)
+ gimple_set_visited (cost->stmt_info->stmt, false);
+
scalar_costs.release ();
vector_costs.release ();
- return true;
+
+ return profitable;
}
/* qsort comparator for lane defs. */
static bool
vect_slp_region (vec<basic_block> bbs, vec<data_reference_p> datarefs,
- vec<int> *dataref_groups, unsigned int n_stmts)
+ vec<int> *dataref_groups, unsigned int n_stmts,
+ loop_p orig_loop)
{
bb_vec_info bb_vinfo;
auto_vector_modes vector_modes;
bb_vinfo->shared->check_datarefs ();
- unsigned i;
- slp_instance instance;
- FOR_EACH_VEC_ELT (BB_VINFO_SLP_INSTANCES (bb_vinfo), i, instance)
+ auto_vec<slp_instance> profitable_subgraphs;
+ for (slp_instance instance : BB_VINFO_SLP_INSTANCES (bb_vinfo))
{
if (instance->subgraph_entries.is_empty ())
continue;
vect_location = instance->location ();
if (!unlimited_cost_model (NULL)
&& !vect_bb_vectorization_profitable_p
- (bb_vinfo, instance->subgraph_entries))
+ (bb_vinfo, instance->subgraph_entries, orig_loop))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
if (!dbg_cnt (vect_slp))
continue;
+ profitable_subgraphs.safe_push (instance);
+ }
+
+ /* When we're vectorizing an if-converted loop body with the
+ very-cheap cost model make sure we vectorized all if-converted
+ code. */
+ if (!profitable_subgraphs.is_empty ()
+ && orig_loop)
+ {
+ gcc_assert (bb_vinfo->bbs.length () == 1);
+ for (gimple_stmt_iterator gsi = gsi_start_bb (bb_vinfo->bbs[0]);
+ !gsi_end_p (gsi); gsi_next (&gsi))
+ {
+ /* The costing above left us with DCEable vectorized scalar
+ stmts having the visited flag set on profitable
+ subgraphs. Do the delayed clearing of the flag here. */
+ if (gimple_visited_p (gsi_stmt (gsi)))
+ {
+ gimple_set_visited (gsi_stmt (gsi), false);
+ continue;
+ }
+ if (flag_vect_cost_model != VECT_COST_MODEL_VERY_CHEAP)
+ continue;
+
+ if (gassign *ass = dyn_cast <gassign *> (gsi_stmt (gsi)))
+ if (gimple_assign_rhs_code (ass) == COND_EXPR)
+ {
+ if (!profitable_subgraphs.is_empty ()
+ && dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "not profitable because of "
+ "unprofitable if-converted scalar "
+ "code\n");
+ profitable_subgraphs.truncate (0);
+ }
+ }
+ }
+
+ /* Finally schedule the profitable subgraphs. */
+ for (slp_instance instance : profitable_subgraphs)
+ {
if (!vectorized && dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Basic block will be vectorized "
true if anything in the basic-block was vectorized. */
static bool
-vect_slp_bbs (const vec<basic_block> &bbs)
+vect_slp_bbs (const vec<basic_block> &bbs, loop_p orig_loop)
{
vec<data_reference_p> datarefs = vNULL;
auto_vec<int> dataref_groups;
&dataref_groups, current_group))
++current_group;
}
+ /* New BBs always start a new DR group. */
+ ++current_group;
}
- return vect_slp_region (bbs, datarefs, &dataref_groups, insns);
+ return vect_slp_region (bbs, datarefs, &dataref_groups, insns, orig_loop);
}
-/* Main entry for the BB vectorizer. Analyze and transform BB, returns
- true if anything in the basic-block was vectorized. */
+/* Special entry for the BB vectorizer. Analyze and transform a single
+ if-converted BB with ORIG_LOOPs body being the not if-converted
+ representation. Returns true if anything in the basic-block was
+ vectorized. */
bool
-vect_slp_bb (basic_block bb)
+vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop)
{
auto_vec<basic_block> bbs;
bbs.safe_push (bb);
- return vect_slp_bbs (bbs);
+ return vect_slp_bbs (bbs, orig_loop);
}
/* Main entry for the BB vectorizer. Analyze and transform BB, returns
if (split && !bbs.is_empty ())
{
- r |= vect_slp_bbs (bbs);
+ r |= vect_slp_bbs (bbs, NULL);
bbs.truncate (0);
bbs.quick_push (bb);
}
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"splitting region at control altering "
"definition %G", last);
- r |= vect_slp_bbs (bbs);
+ r |= vect_slp_bbs (bbs, NULL);
bbs.truncate (0);
}
}
if (!bbs.is_empty ())
- r |= vect_slp_bbs (bbs);
+ r |= vect_slp_bbs (bbs, NULL);
free (rpo);
projects / thirdparty / gcc.git / blobdiff