The following fixes the same issue with scalar BB costing as I fixed
earlier this year with the loop scalar costing. We are currently
comparing apples and oranges on x86 where the add_stmt_cost hook
uses costs dependent on the stmt operation while the old hook
does not (cannot).
Fixed as follows.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
I am considering to backport this after some soaking/benchmarking.
Richard.
2018-05-09 Richard Biener <rguent...@suse.de>
* tree-vect-slp.c (vect_bb_slp_scalar_cost): Fill a cost
vector.
(vect_bb_vectorization_profitable_p): Adjust. Compute
actual scalar cost using the cost vector and the add_stmt_cost
machinery.
Index: gcc/tree-vect-slp.c
===================================================================
--- gcc/tree-vect-slp.c (revision 260072)
+++ gcc/tree-vect-slp.c (working copy)
@@ -2886,18 +2886,17 @@ vect_slp_analyze_operations (vec_info *v
and return it. Do not account defs that are marked in LIFE and
update LIFE according to uses of NODE. */
-static unsigned
+static void
vect_bb_slp_scalar_cost (basic_block bb,
- slp_tree node, vec<bool, va_heap> *life)
+ slp_tree node, vec<bool, va_heap> *life,
+ stmt_vector_for_cost *cost_vec)
{
- unsigned scalar_cost = 0;
unsigned i;
gimple *stmt;
slp_tree child;
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt)
{
- unsigned stmt_cost;
ssa_op_iter op_iter;
def_operand_p def_p;
stmt_vec_info stmt_info;
@@ -2933,17 +2932,17 @@ vect_bb_slp_scalar_cost (basic_block bb,
gimple_set_visited (stmt, true);
stmt_info = vinfo_for_stmt (stmt);
+ vect_cost_for_stmt kind;
if (STMT_VINFO_DATA_REF (stmt_info))
{
if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
- stmt_cost = vect_get_stmt_cost (scalar_load);
+ kind = scalar_load;
else
- stmt_cost = vect_get_stmt_cost (scalar_store);
+ kind = scalar_store;
}
else
- stmt_cost = vect_get_stmt_cost (scalar_stmt);
-
- scalar_cost += stmt_cost;
+ kind = scalar_stmt;
+ record_stmt_cost (cost_vec, 1, kind, stmt_info, 0, vect_body);
}
auto_vec<bool, 20> subtree_life;
@@ -2954,12 +2953,10 @@ vect_bb_slp_scalar_cost (basic_block bb,
/* Do not directly pass LIFE to the recursive call, copy it to
confine changes in the callee to the current child/subtree. */
subtree_life.safe_splice (*life);
- scalar_cost += vect_bb_slp_scalar_cost (bb, child, &subtree_life);
+ vect_bb_slp_scalar_cost (bb, child, &subtree_life, cost_vec);
subtree_life.truncate (0);
}
}
-
- return scalar_cost;
}
/* Check if vectorization of the basic block is profitable. */
@@ -2974,14 +2971,30 @@ vect_bb_vectorization_profitable_p (bb_v
unsigned int vec_prologue_cost = 0, vec_epilogue_cost = 0;
/* Calculate scalar cost. */
+ stmt_vector_for_cost scalar_costs;
+ scalar_costs.create (0);
FOR_EACH_VEC_ELT (slp_instances, i, instance)
{
auto_vec<bool, 20> life;
life.safe_grow_cleared (SLP_INSTANCE_GROUP_SIZE (instance));
- scalar_cost += vect_bb_slp_scalar_cost (BB_VINFO_BB (bb_vinfo),
- SLP_INSTANCE_TREE (instance),
- &life);
+ vect_bb_slp_scalar_cost (BB_VINFO_BB (bb_vinfo),
+ SLP_INSTANCE_TREE (instance),
+ &life, &scalar_costs);
+ }
+ void *target_cost_data = init_cost (NULL);
+ stmt_info_for_cost *si;
+ FOR_EACH_VEC_ELT (scalar_costs, i, si)
+ {
+ struct _stmt_vec_info *stmt_info
+ = si->stmt ? vinfo_for_stmt (si->stmt) : NULL;
+ (void) add_stmt_cost (target_cost_data, si->count,
+ si->kind, stmt_info, si->misalign,
+ vect_body);
}
+ scalar_costs.release ();
+ unsigned dummy;
+ finish_cost (target_cost_data, &dummy, &scalar_cost, &dummy);
+ destroy_cost_data (target_cost_data);
/* Unset visited flag. */
for (gimple_stmt_iterator gsi = bb_vinfo->region_begin;
