/* Linear Loop transforms
- Copyright (C) 2003, 2004 Free Software Foundation, Inc.
+ Copyright (C) 2003, 2004, 2005, 2007 Free Software Foundation, Inc.
Contributed by Daniel Berlin <dberlin@dberlin.org>.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
+Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
-#include "errors.h"
#include "ggc.h"
#include "tree.h"
#include "target.h"
#include "rtl.h"
#include "basic-block.h"
#include "diagnostic.h"
+#include "obstack.h"
#include "tree-flow.h"
#include "tree-dump.h"
#include "timevar.h"
#include "tree-data-ref.h"
#include "tree-scalar-evolution.h"
#include "tree-pass.h"
-#include "varray.h"
#include "lambda.h"
/* Linear loop transforms include any composition of interchange,
transform matrix for locality purposes.
TODO: Completion of partial transforms. */
-/* Gather statistics for loop interchange. LOOP_NUMBER is a relative
- index in the considered loop nest. The first loop in the
- considered loop nest is FIRST_LOOP, and consequently the index of
- the considered loop is obtained by FIRST_LOOP + LOOP_NUMBER.
+/* Gather statistics for loop interchange. LOOP is the loop being
+ considered. The first loop in the considered loop nest is
+ FIRST_LOOP, and consequently, the index of the considered loop is
+ obtained by LOOP->DEPTH - FIRST_LOOP->DEPTH
Initializes:
- DEPENDENCE_STEPS the sum of all the data dependence distances
- carried by loop LOOP_NUMBER,
+ carried by loop LOOP,
- NB_DEPS_NOT_CARRIED_BY_LOOP the number of dependence relations
- for which the loop LOOP_NUMBER is not carrying any dependence,
+ for which the loop LOOP is not carrying any dependence,
- - ACCESS_STRIDES the sum of all the strides in LOOP_NUMBER.
+ - ACCESS_STRIDES the sum of all the strides in LOOP.
Example: for the following loop,
*/
static void
-gather_interchange_stats (varray_type dependence_relations,
- varray_type datarefs,
- unsigned int loop_number,
- unsigned int first_loop,
+gather_interchange_stats (VEC (ddr_p, heap) *dependence_relations,
+ VEC (data_reference_p, heap) *datarefs,
+ struct loop *loop,
+ struct loop *first_loop,
unsigned int *dependence_steps,
unsigned int *nb_deps_not_carried_by_loop,
- unsigned int *access_strides)
+ double_int *access_strides)
{
- unsigned int i;
+ unsigned int i, j;
+ struct data_dependence_relation *ddr;
+ struct data_reference *dr;
*dependence_steps = 0;
*nb_deps_not_carried_by_loop = 0;
- *access_strides = 0;
+ *access_strides = double_int_zero;
- for (i = 0; i < VARRAY_ACTIVE_SIZE (dependence_relations); i++)
+ for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
{
- int dist;
- struct data_dependence_relation *ddr =
- (struct data_dependence_relation *)
- VARRAY_GENERIC_PTR (dependence_relations, i);
-
/* If we don't know anything about this dependence, or the distance
vector is NULL, or there is no dependence, then there is no reuse of
data. */
-
- if (DDR_DIST_VECT (ddr) == NULL
- || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know
- || DDR_ARE_DEPENDENT (ddr) == chrec_known)
+ if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know
+ || DDR_ARE_DEPENDENT (ddr) == chrec_known
+ || DDR_NUM_DIST_VECTS (ddr) == 0)
continue;
-
-
- dist = DDR_DIST_VECT (ddr)[loop_number];
- if (dist == 0)
- (*nb_deps_not_carried_by_loop) += 1;
- else if (dist < 0)
- (*dependence_steps) += -dist;
- else
- (*dependence_steps) += dist;
+ for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++)
+ {
+ int dist = DDR_DIST_VECT (ddr, j)[loop_depth (loop) - loop_depth (first_loop)];
+
+ if (dist == 0)
+ (*nb_deps_not_carried_by_loop) += 1;
+
+ else if (dist < 0)
+ (*dependence_steps) += -dist;
+
+ else
+ (*dependence_steps) += dist;
+ }
}
/* Compute the access strides. */
- for (i = 0; i < VARRAY_ACTIVE_SIZE (datarefs); i++)
+ for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
{
unsigned int it;
- struct data_reference *dr = VARRAY_GENERIC_PTR (datarefs, i);
+ tree ref = DR_REF (dr);
tree stmt = DR_STMT (dr);
struct loop *stmt_loop = loop_containing_stmt (stmt);
- struct loop *inner_loop = current_loops->parray[first_loop + 1];
+ struct loop *inner_loop = first_loop->inner;
- if (!flow_loop_nested_p (inner_loop, stmt_loop)
- && inner_loop->num != stmt_loop->num)
+ if (inner_loop != stmt_loop
+ && !flow_loop_nested_p (inner_loop, stmt_loop))
continue;
- for (it = 0; it < DR_NUM_DIMENSIONS (dr); it++)
+ for (it = 0; it < DR_NUM_DIMENSIONS (dr);
+ it++, ref = TREE_OPERAND (ref, 0))
{
- tree chrec = DR_ACCESS_FN (dr, it);
- tree tstride = evolution_part_in_loop_num
- (chrec, first_loop + loop_number);
-
- if (tstride == NULL_TREE
- || TREE_CODE (tstride) != INTEGER_CST)
+ int num = am_vector_index_for_loop (DR_ACCESS_MATRIX (dr), loop->num);
+ int istride = AM_GET_ACCESS_MATRIX_ELEMENT (DR_ACCESS_MATRIX (dr), it, num);
+ tree array_size = TYPE_SIZE (TREE_TYPE (ref));
+ double_int dstride;
+
+ if (array_size == NULL_TREE
+ || TREE_CODE (array_size) != INTEGER_CST)
continue;
-
- (*access_strides) += int_cst_value (tstride);
+
+ dstride = double_int_mul (tree_to_double_int (array_size),
+ shwi_to_double_int (istride));
+ (*access_strides) = double_int_add (*access_strides, dstride);
}
}
}
static lambda_trans_matrix
try_interchange_loops (lambda_trans_matrix trans,
unsigned int depth,
- varray_type dependence_relations,
- varray_type datarefs,
- unsigned int first_loop)
+ VEC (ddr_p, heap) *dependence_relations,
+ VEC (data_reference_p, heap) *datarefs,
+ struct loop *first_loop)
{
- unsigned int loop_i, loop_j;
+ bool res;
+ struct loop *loop_i;
+ struct loop *loop_j;
unsigned int dependence_steps_i, dependence_steps_j;
- unsigned int access_strides_i, access_strides_j;
+ double_int access_strides_i, access_strides_j;
+ double_int small, large, nb_iter;
+ double_int l1_cache_size, l2_cache_size;
+ int cmp;
unsigned int nb_deps_not_carried_by_i, nb_deps_not_carried_by_j;
struct data_dependence_relation *ddr;
+ if (VEC_length (ddr_p, dependence_relations) == 0)
+ return trans;
+
/* When there is an unknown relation in the dependence_relations, we
know that it is no worth looking at this loop nest: give up. */
- ddr = (struct data_dependence_relation *)
- VARRAY_GENERIC_PTR (dependence_relations, 0);
+ ddr = VEC_index (ddr_p, dependence_relations, 0);
if (ddr == NULL || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
return trans;
-
+
+ l1_cache_size = uhwi_to_double_int (L1_CACHE_SIZE * 1024);
+ l2_cache_size = uhwi_to_double_int (L2_CACHE_SIZE * 1024);
+
/* LOOP_I is always the outer loop. */
- for (loop_j = 1; loop_j < depth; loop_j++)
- for (loop_i = 0; loop_i < loop_j; loop_i++)
+ for (loop_j = first_loop->inner;
+ loop_j;
+ loop_j = loop_j->inner)
+ for (loop_i = first_loop;
+ loop_depth (loop_i) < loop_depth (loop_j);
+ loop_i = loop_i->inner)
{
gather_interchange_stats (dependence_relations, datarefs,
loop_i, first_loop,
/* Heuristics for loop interchange profitability:
+ 0. Don't transform if the smallest stride is larger than
+ the L2 cache, or if the largest stride multiplied by the
+ number of iterations is smaller than the L1 cache.
+
1. (spatial locality) Inner loops should have smallest
dependence steps.
2. (spatial locality) Inner loops should contain more
dependence relations not carried by the loop.
- 3. (temporal locality) Inner loops should have smallest
+ 3. (temporal locality) Inner loops should have smallest
array access strides.
*/
+
+ cmp = double_int_ucmp (access_strides_i, access_strides_j);
+ small = cmp < 0 ? access_strides_i : access_strides_j;
+ large = cmp < 0 ? access_strides_j : access_strides_i;
+
+ if (double_int_ucmp (small, l2_cache_size) > 0)
+ continue;
+
+ res = cmp < 0 ?
+ estimated_loop_iterations (loop_j, false, &nb_iter):
+ estimated_loop_iterations (loop_i, false, &nb_iter);
+ large = double_int_mul (large, nb_iter);
+
+ if (res && double_int_ucmp (large, l1_cache_size) < 0)
+ continue;
+
if (dependence_steps_i < dependence_steps_j
|| nb_deps_not_carried_by_i > nb_deps_not_carried_by_j
- || access_strides_i < access_strides_j)
+ || cmp < 0)
{
- lambda_matrix_row_exchange (LTM_MATRIX (trans), loop_i, loop_j);
+ lambda_matrix_row_exchange (LTM_MATRIX (trans),
+ loop_depth (loop_i) - loop_depth (first_loop),
+ loop_depth (loop_j) - loop_depth (first_loop));
/* Validate the resulting matrix. When the transformation
is not valid, reverse to the previous transformation. */
if (!lambda_transform_legal_p (trans, depth, dependence_relations))
- lambda_matrix_row_exchange (LTM_MATRIX (trans), loop_i, loop_j);
+ lambda_matrix_row_exchange (LTM_MATRIX (trans),
+ loop_depth (loop_i) - loop_depth (first_loop),
+ loop_depth (loop_j) - loop_depth (first_loop));
}
}
return trans;
}
-/* Perform a set of linear transforms on LOOPS. */
+/* Return the number of nested loops in LOOP_NEST, or 0 if the loops
+ are not perfectly nested. */
+
+static unsigned int
+perfect_loop_nest_depth (struct loop *loop_nest)
+{
+ struct loop *temp;
+ unsigned int depth = 1;
+
+ /* If it's not a loop nest, we don't want it. We also don't handle
+ sibling loops properly, which are loops of the following form:
+
+ | for (i = 0; i < 50; i++)
+ | {
+ | for (j = 0; j < 50; j++)
+ | {
+ | ...
+ | }
+ | for (j = 0; j < 50; j++)
+ | {
+ | ...
+ | }
+ | }
+ */
+
+ if (!loop_nest->inner || !single_exit (loop_nest))
+ return 0;
+
+ for (temp = loop_nest->inner; temp; temp = temp->inner)
+ {
+ /* If we have a sibling loop or multiple exit edges, jump ship. */
+ if (temp->next || !single_exit (temp))
+ return 0;
+
+ depth++;
+ }
+
+ return depth;
+}
+
+/* Perform a set of linear transforms on loops. */
void
-linear_transform_loops (struct loops *loops)
+linear_transform_loops (void)
{
- unsigned int i;
-
- compute_immediate_uses (TDFA_USE_OPS | TDFA_USE_VOPS, NULL);
- for (i = 1; i < loops->num; i++)
+ bool modified = false;
+ loop_iterator li;
+ VEC(tree,heap) *oldivs = NULL;
+ VEC(tree,heap) *invariants = NULL;
+ VEC(tree,heap) *lambda_parameters = NULL;
+ VEC(tree,heap) *remove_ivs = VEC_alloc (tree, heap, 3);
+ struct loop *loop_nest;
+ tree oldiv_stmt;
+ unsigned i;
+
+ FOR_EACH_LOOP (li, loop_nest, 0)
{
unsigned int depth = 0;
- varray_type datarefs;
- varray_type dependence_relations;
- struct loop *loop_nest = loops->parray[i];
- struct loop *temp;
- VEC (tree) *oldivs = NULL;
- VEC (tree) *invariants = NULL;
+ VEC (ddr_p, heap) *dependence_relations;
+ VEC (data_reference_p, heap) *datarefs;
+
lambda_loopnest before, after;
lambda_trans_matrix trans;
- bool problem = false;
- bool need_perfect_nest = false;
- /* If it's not a loop nest, we don't want it.
- We also don't handle sibling loops properly,
- which are loops of the following form:
- for (i = 0; i < 50; i++)
- {
- for (j = 0; j < 50; j++)
- {
- ...
- }
- for (j = 0; j < 50; j++)
- {
- ...
- }
- } */
- if (!loop_nest->inner)
- continue;
- depth = 1;
- for (temp = loop_nest->inner; temp; temp = temp->inner)
- {
- flow_loop_scan (temp, LOOP_ALL);
- /* If we have a sibling loop or multiple exit edges, jump ship. */
- if (temp->next || temp->num_exits != 1)
- {
- problem = true;
- break;
- }
- depth ++;
- }
- if (problem)
+ struct obstack lambda_obstack;
+ gcc_obstack_init (&lambda_obstack);
+
+ depth = perfect_loop_nest_depth (loop_nest);
+ if (depth == 0)
continue;
- /* Analyze data references and dependence relations using scev. */
-
- VARRAY_GENERIC_PTR_INIT (datarefs, 10, "datarefs");
- VARRAY_GENERIC_PTR_INIT (dependence_relations, 10,
- "dependence_relations");
+ VEC_truncate (tree, oldivs, 0);
+ VEC_truncate (tree, invariants, 0);
+ VEC_truncate (tree, lambda_parameters, 0);
+
+ datarefs = VEC_alloc (data_reference_p, heap, 10);
+ dependence_relations = VEC_alloc (ddr_p, heap, 10 * 10);
+ if (!compute_data_dependences_for_loop (loop_nest, true, &datarefs,
+ &dependence_relations))
+ continue;
-
- compute_data_dependences_for_loop (depth, loop_nest,
- &datarefs, &dependence_relations);
+ lambda_collect_parameters (datarefs, &lambda_parameters);
+ if (!lambda_compute_access_matrices (datarefs, lambda_parameters,
+ loop_nest->num))
+ continue;
+
if (dump_file && (dump_flags & TDF_DETAILS))
- {
- unsigned int j;
- for (j = 0; j < VARRAY_ACTIVE_SIZE (dependence_relations); j++)
- {
- struct data_dependence_relation *ddr =
- (struct data_dependence_relation *)
- VARRAY_GENERIC_PTR (dependence_relations, j);
-
- if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
- {
- fprintf (dump_file, "DISTANCE_V (");
- print_lambda_vector (dump_file, DDR_DIST_VECT (ddr),
- DDR_SIZE_VECT (ddr));
- fprintf (dump_file, ")\n");
- fprintf (dump_file, "DIRECTION_V (");
- print_lambda_vector (dump_file, DDR_DIR_VECT (ddr),
- DDR_SIZE_VECT (ddr));
- fprintf (dump_file, ")\n");
- }
- }
- fprintf (dump_file, "\n\n");
- }
+ dump_ddrs (dump_file, dependence_relations);
+
/* Build the transformation matrix. */
trans = lambda_trans_matrix_new (depth, depth);
lambda_matrix_id (LTM_MATRIX (trans), depth);
-
trans = try_interchange_loops (trans, depth, dependence_relations,
- datarefs, loop_nest->num);
+ datarefs, loop_nest);
if (lambda_trans_matrix_id_p (trans))
{
if (dump_file)
fprintf (dump_file, "Won't transform loop. Optimal transform is the identity transform\n");
- continue;
+ goto free_and_continue;
}
/* Check whether the transformation is legal. */
{
if (dump_file)
fprintf (dump_file, "Can't transform loop, transform is illegal:\n");
- continue;
+ goto free_and_continue;
}
- if (!perfect_nest_p (loop_nest))
- need_perfect_nest = true;
- before = gcc_loopnest_to_lambda_loopnest (loops,
- loop_nest, &oldivs,
- &invariants,
- need_perfect_nest);
+
+ before = gcc_loopnest_to_lambda_loopnest (loop_nest, &oldivs,
+ &invariants, &lambda_obstack);
+
if (!before)
- continue;
-
+ goto free_and_continue;
+
if (dump_file)
{
fprintf (dump_file, "Before:\n");
print_lambda_loopnest (dump_file, before, 'i');
}
- after = lambda_loopnest_transform (before, trans);
+ after = lambda_loopnest_transform (before, trans, &lambda_obstack);
+
if (dump_file)
{
fprintf (dump_file, "After:\n");
print_lambda_loopnest (dump_file, after, 'u');
}
+
lambda_loopnest_to_gcc_loopnest (loop_nest, oldivs, invariants,
- after, trans);
+ &remove_ivs,
+ after, trans, &lambda_obstack);
+ modified = true;
+
if (dump_file)
fprintf (dump_file, "Successfully transformed loop.\n");
- oldivs = NULL;
- invariants = NULL;
+
+ free_and_continue:
+ obstack_free (&lambda_obstack, NULL);
free_dependence_relations (dependence_relations);
free_data_refs (datarefs);
}
- free_df ();
+
+ for (i = 0; VEC_iterate (tree, remove_ivs, i, oldiv_stmt); i++)
+ remove_iv (oldiv_stmt);
+
+ VEC_free (tree, heap, oldivs);
+ VEC_free (tree, heap, invariants);
+ VEC_free (tree, heap, remove_ivs);
scev_reset ();
- rewrite_into_loop_closed_ssa ();
-#ifdef ENABLE_CHECKING
- verify_loop_closed_ssa ();
-#endif
+
+ if (modified)
+ rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_full_phi);
}