* [PATCH] Improve vect cost model for PR37150
@ 2016-11-04 13:22 Richard Biener
0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2016-11-04 13:22 UTC (permalink / raw)
To: gcc-patches
The following implements some easy improvements for the SLP cost model
for PR37150 which shows excess cost for dead loads and permutes
accounted when vectorizing a basic-block.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Still doesn't vectorize the testcase in that PR without
-fno-vect-cost-model though. Real improvements are only possible
with re-doing the vectorizer data structures.
Richard.
2016-11-04 Richard Biener <rguenther@suse.de>
PR tree-optimization/37150
* tree-vectorizer.h (vect_transform_slp_perm_load): Add n_perms
parameter.
* tree-vect-slp.c (vect_supported_load_permutation_p): Adjust.
(vect_analyze_slp_cost_1): Account for the real number of
permutations emitted and for dead loads.
(vect_transform_slp_perm_load): Add n_perms parameter counting
the number of emitted permutations.
* tree-vect-stmts.c (vectorizable_load): Adjust.
Index: gcc/tree-vectorizer.h
===================================================================
--- gcc/tree-vectorizer.h (revision 241791)
+++ gcc/tree-vectorizer.h (working copy)
@@ -1166,7 +1168,7 @@ extern int vect_get_known_peeling_cost (
extern void vect_free_slp_instance (slp_instance);
extern bool vect_transform_slp_perm_load (slp_tree, vec<tree> ,
gimple_stmt_iterator *, int,
- slp_instance, bool);
+ slp_instance, bool, unsigned *);
extern bool vect_slp_analyze_operations (vec<slp_instance> slp_instances,
void *);
extern bool vect_schedule_slp (vec_info *);
Index: gcc/tree-vect-slp.c
===================================================================
--- gcc/tree-vect-slp.c (revision 241791)
+++ gcc/tree-vect-slp.c (working copy)
@@ -1461,8 +1461,9 @@ vect_supported_load_permutation_p (slp_i
{
/* Verify the permutation can be generated. */
vec<tree> tem;
+ unsigned n_perms;
if (!vect_transform_slp_perm_load (node, tem, NULL,
- 1, slp_instn, true))
+ 1, slp_instn, true, &n_perms))
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION,
vect_location,
@@ -1475,11 +1476,13 @@ vect_supported_load_permutation_p (slp_i
}
/* For loop vectorization verify we can generate the permutation. */
+ unsigned n_perms;
FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
if (node->load_permutation.exists ()
&& !vect_transform_slp_perm_load
(node, vNULL, NULL,
- SLP_INSTANCE_UNROLLING_FACTOR (slp_instn), slp_instn, true))
+ SLP_INSTANCE_UNROLLING_FACTOR (slp_instn), slp_instn, true,
+ &n_perms))
return false;
return true;
@@ -1548,14 +1551,38 @@ vect_analyze_slp_cost_1 (slp_instance in
stmt = GROUP_FIRST_ELEMENT (stmt_info);
stmt_info = vinfo_for_stmt (stmt);
/* Record the cost for the permutation. */
- record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm,
+ unsigned n_perms;
+ vect_transform_slp_perm_load (node, vNULL, NULL,
+ ncopies_for_cost, instance, true,
+ &n_perms);
+ record_stmt_cost (body_cost_vec, n_perms, vec_perm,
stmt_info, 0, vect_body);
- /* And adjust the number of loads performed. */
unsigned nunits
= TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
- ncopies_for_cost
- = (GROUP_SIZE (stmt_info) - GROUP_GAP (stmt_info)
- + nunits - 1) / nunits;
+ /* And adjust the number of loads performed. This handles
+ redundancies as well as loads that are later dead. */
+ auto_sbitmap perm (GROUP_SIZE (stmt_info));
+ bitmap_clear (perm);
+ for (i = 0; i < SLP_TREE_LOAD_PERMUTATION (node).length (); ++i)
+ bitmap_set_bit (perm, SLP_TREE_LOAD_PERMUTATION (node)[i]);
+ ncopies_for_cost = 0;
+ bool load_seen = false;
+ for (i = 0; i < GROUP_SIZE (stmt_info); ++i)
+ {
+ if (i % nunits == 0)
+ {
+ if (load_seen)
+ ncopies_for_cost++;
+ load_seen = false;
+ }
+ if (bitmap_bit_p (perm, i))
+ load_seen = true;
+ }
+ if (load_seen)
+ ncopies_for_cost++;
+ gcc_assert (ncopies_for_cost
+ <= (GROUP_SIZE (stmt_info) - GROUP_GAP (stmt_info)
+ + nunits - 1) / nunits);
ncopies_for_cost *= SLP_INSTANCE_UNROLLING_FACTOR (instance);
}
/* Record the cost for the vector loads. */
@@ -3402,7 +3489,8 @@ vect_create_mask_and_perm (gimple *stmt,
bool
vect_transform_slp_perm_load (slp_tree node, vec<tree> dr_chain,
gimple_stmt_iterator *gsi, int vf,
- slp_instance slp_node_instance, bool analyze_only)
+ slp_instance slp_node_instance, bool analyze_only,
+ unsigned *n_perms)
{
gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0];
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
@@ -3457,6 +3545,7 @@ vect_transform_slp_perm_load (slp_tree n
int first_vec_index = -1;
int second_vec_index = -1;
bool noop_p = true;
+ *n_perms = 0;
for (int j = 0; j < unroll_factor; j++)
{
@@ -3513,6 +3602,9 @@ vect_transform_slp_perm_load (slp_tree n
return false;
}
+ if (! noop_p)
+ ++*n_perms;
+
if (!analyze_only)
{
tree mask_vec = NULL_TREE;
Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c (revision 241791)
+++ gcc/tree-vect-stmts.c (working copy)
@@ -6978,8 +7041,11 @@ vectorizable_load (gimple *stmt, gimple_
}
}
if (slp_perm)
- vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf,
- slp_node_instance, false);
+ {
+ unsigned n_perms;
+ vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf,
+ slp_node_instance, false, &n_perms);
+ }
return true;
}
@@ -7497,8 +7563,10 @@ vectorizable_load (gimple *stmt, gimple_
if (slp_perm)
{
+ unsigned n_perms;
if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf,
- slp_node_instance, false))
+ slp_node_instance, false,
+ &n_perms))
{
dr_chain.release ();
return false;
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2016-11-04 13:22 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-11-04 13:22 [PATCH] Improve vect cost model for PR37150 Richard Biener
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).