This enables -ftree-loop-distribute-patterns at -O[2s] and also
arranges cold loops to be still processed but for pattern
recognition to save code-size.
Bootstrap and regtest running on x86_64-unknown-linux-gnu.
Martin has done extensive compile-time testing on SPEC
identifying only a single regression I'll have a look into.
Richard.
2019-05-22 Richard Biener <rguent...@suse.de>
PR tree-optimization/88440
* opts.c (default_options_table): Enable -ftree-loop-distribute-patterns
at -O[2s]+.
* tree-loop-distribution.c (generate_memset_builtin): Fold the
generated call.
(generate_memcpy_builtin): Likewise.
(distribute_loop): Pass in whether to only distribute patterns.
(prepare_perfect_loop_nest): Also allow size optimization.
(pass_loop_distribution::execute): When optimizing a loop
nest for size allow pattern replacement.
* gcc.dg/tree-ssa/ldist-37.c: New testcase.
* gcc.dg/tree-ssa/ldist-38.c: Likewise.
Index: gcc/opts.c
===================================================================
--- gcc/opts.c (revision 271463)
+++ gcc/opts.c (working copy)
@@ -550,7 +550,7 @@ static const struct default_options defa
{ OPT_LEVELS_3_PLUS, OPT_fpredictive_commoning, NULL, 1 },
{ OPT_LEVELS_3_PLUS, OPT_fsplit_loops, NULL, 1 },
{ OPT_LEVELS_3_PLUS, OPT_fsplit_paths, NULL, 1 },
- { OPT_LEVELS_3_PLUS, OPT_ftree_loop_distribute_patterns, NULL, 1 },
+ { OPT_LEVELS_2_PLUS, OPT_ftree_loop_distribute_patterns, NULL, 1 },
{ OPT_LEVELS_3_PLUS, OPT_ftree_loop_distribution, NULL, 1 },
{ OPT_LEVELS_3_PLUS, OPT_ftree_loop_vectorize, NULL, 1 },
{ OPT_LEVELS_3_PLUS, OPT_ftree_partial_pre, NULL, 1 },
Index: gcc/tree-loop-distribution.c
===================================================================
--- gcc/tree-loop-distribution.c (revision 271463)
+++ gcc/tree-loop-distribution.c (working copy)
@@ -115,6 +115,7 @@ along with GCC; see the file COPYING3.
#include "params.h"
#include "tree-vectorizer.h"
#include "tree-eh.h"
+#include "gimple-fold.h"
#define MAX_DATAREFS_NUM \
@@ -1028,6 +1029,7 @@ generate_memset_builtin (struct loop *lo
fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET));
fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes);
gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING);
+ fold_stmt (&gsi);
if (dump_file && (dump_flags & TDF_DETAILS))
{
@@ -1071,6 +1073,7 @@ generate_memcpy_builtin (struct loop *lo
fn = build_fold_addr_expr (builtin_decl_implicit (kind));
fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes);
gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING);
+ fold_stmt (&gsi);
if (dump_file && (dump_flags & TDF_DETAILS))
{
@@ -2769,7 +2772,8 @@ finalize_partitions (struct loop *loop,
static int
distribute_loop (struct loop *loop, vec<gimple *> stmts,
- control_dependences *cd, int *nb_calls, bool *destroy_p)
+ control_dependences *cd, int *nb_calls, bool *destroy_p,
+ bool only_patterns_p)
{
ddrs_table = new hash_table<ddr_hasher> (389);
struct graph *rdg;
@@ -2843,7 +2847,7 @@ distribute_loop (struct loop *loop, vec<
/* If we are only distributing patterns but did not detect any,
simply bail out. */
- if (!flag_tree_loop_distribution
+ if (only_patterns_p
&& !any_builtin)
{
nbp = 0;
@@ -2855,7 +2859,7 @@ distribute_loop (struct loop *loop, vec<
a loop into pieces, separated by builtin calls. That is, we
only want no or a single loop body remaining. */
struct partition *into;
- if (!flag_tree_loop_distribution)
+ if (only_patterns_p)
{
for (i = 0; partitions.iterate (i, &into); ++i)
if (!partition_builtin_p (into))
@@ -3085,7 +3089,6 @@ prepare_perfect_loop_nest (struct loop *
&& loop_outer (outer)
&& outer->inner == loop && loop->next == NULL
&& single_exit (outer)
- && optimize_loop_for_speed_p (outer)
&& !chrec_contains_symbols_defined_in_loop (niters, outer->num)
&& (niters = number_of_latch_executions (outer)) != NULL_TREE
&& niters != chrec_dont_know)
@@ -3139,9 +3142,11 @@ pass_loop_distribution::execute (functio
walking to innermost loops. */
FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST)
{
- /* Don't distribute multiple exit edges loop, or cold loop. */
+ /* Don't distribute multiple exit edges loop, or cold loop when
+ not doing pattern detection. */
if (!single_exit (loop)
- || !optimize_loop_for_speed_p (loop))
+ || (!flag_tree_loop_distribute_patterns
+ && !optimize_loop_for_speed_p (loop)))
continue;
/* Don't distribute loop if niters is unknown. */
@@ -3169,9 +3174,10 @@ pass_loop_distribution::execute (functio
bool destroy_p;
int nb_generated_loops, nb_generated_calls;
- nb_generated_loops = distribute_loop (loop, work_list, cd,
- &nb_generated_calls,
- &destroy_p);
+ nb_generated_loops
+ = distribute_loop (loop, work_list, cd, &nb_generated_calls,
+ &destroy_p, (!optimize_loop_for_speed_p (loop)
+ || !flag_tree_loop_distribution));
if (destroy_p)
loops_to_be_destroyed.safe_push (loop);
Index: gcc/testsuite/gcc.dg/tree-ssa/ldist-37.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/ldist-37.c (nonexistent)
+++ gcc/testsuite/gcc.dg/tree-ssa/ldist-37.c (working copy)
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* { dg-options "-Os -fdump-tree-ldist-optimized" } */
+
+void foo(char* restrict dst, const char* buf)
+{
+ for (int i=0; i<8; ++i)
+ *dst++ = *buf++;
+}
+
+/* { dg-final { scan-tree-dump "split to 0 loops and 1 library calls"
"optimized" } } */
Index: gcc/testsuite/gcc.dg/tree-ssa/ldist-38.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/ldist-38.c (nonexistent)
+++ gcc/testsuite/gcc.dg/tree-ssa/ldist-38.c (working copy)
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-ldist-optimized" } */
+
+void foo(char* restrict dst, const char* buf)
+{
+ for (int i=0; i<8; ++i)
+ *dst++ = *buf++;
+}
+
+/* { dg-final { scan-tree-dump "split to 0 loops and 1 library calls" "ldist"
} } */
