1 /* Linear Loop transforms
2 Copyright (C) 2003, 2004, 2005, 2007 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dberlin@dberlin.org>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
31 #include "basic-block.h"
32 #include "diagnostic.h"
34 #include "tree-flow.h"
35 #include "tree-dump.h"
40 #include "tree-chrec.h"
41 #include "tree-data-ref.h"
42 #include "tree-scalar-evolution.h"
43 #include "tree-pass.h"
46 /* Linear loop transforms include any composition of interchange,
47 scaling, skewing, and reversal. They are used to change the
48 iteration order of loop nests in order to optimize data locality of
49 traversals, or remove dependences that prevent
50 parallelization/vectorization/etc.
52 TODO: Determine reuse vectors/matrix and use it to determine optimal
53 transform matrix for locality purposes.
54 TODO: Completion of partial transforms. */
56 /* Gather statistics for loop interchange. LOOP is the loop being
57 considered. The first loop in the considered loop nest is
58 FIRST_LOOP, and consequently, the index of the considered loop is
59 obtained by LOOP->DEPTH - FIRST_LOOP->DEPTH
62 - DEPENDENCE_STEPS the sum of all the data dependence distances
65 - NB_DEPS_NOT_CARRIED_BY_LOOP the number of dependence relations
66 for which the loop LOOP is not carrying any dependence,
68 - ACCESS_STRIDES the sum of all the strides in LOOP.
70 Example: for the following loop,
72 | loop_1 runs 1335 times
73 | loop_2 runs 1335 times
74 | A[{{0, +, 1}_1, +, 1335}_2]
75 | B[{{0, +, 1}_1, +, 1335}_2]
80 gather_interchange_stats (in loop_1) will return
81 DEPENDENCE_STEPS = 3002
82 NB_DEPS_NOT_CARRIED_BY_LOOP = 5
83 ACCESS_STRIDES = 10694
85 gather_interchange_stats (in loop_2) will return
86 DEPENDENCE_STEPS = 3000
87 NB_DEPS_NOT_CARRIED_BY_LOOP = 7
92 gather_interchange_stats (VEC (ddr_p
, heap
) *dependence_relations
,
93 VEC (data_reference_p
, heap
) *datarefs
,
95 struct loop
*first_loop
,
96 unsigned int *dependence_steps
,
97 unsigned int *nb_deps_not_carried_by_loop
,
98 double_int
*access_strides
)
101 struct data_dependence_relation
*ddr
;
102 struct data_reference
*dr
;
104 *dependence_steps
= 0;
105 *nb_deps_not_carried_by_loop
= 0;
106 *access_strides
= double_int_zero
;
108 for (i
= 0; VEC_iterate (ddr_p
, dependence_relations
, i
, ddr
); i
++)
110 /* If we don't know anything about this dependence, or the distance
111 vector is NULL, or there is no dependence, then there is no reuse of
113 if (DDR_ARE_DEPENDENT (ddr
) == chrec_dont_know
114 || DDR_ARE_DEPENDENT (ddr
) == chrec_known
115 || DDR_NUM_DIST_VECTS (ddr
) == 0)
118 for (j
= 0; j
< DDR_NUM_DIST_VECTS (ddr
); j
++)
120 int dist
= DDR_DIST_VECT (ddr
, j
)[loop_depth (loop
) - loop_depth (first_loop
)];
123 (*nb_deps_not_carried_by_loop
) += 1;
126 (*dependence_steps
) += -dist
;
129 (*dependence_steps
) += dist
;
133 /* Compute the access strides. */
134 for (i
= 0; VEC_iterate (data_reference_p
, datarefs
, i
, dr
); i
++)
137 tree ref
= DR_REF (dr
);
138 tree stmt
= DR_STMT (dr
);
139 struct loop
*stmt_loop
= loop_containing_stmt (stmt
);
140 struct loop
*inner_loop
= first_loop
->inner
;
142 if (inner_loop
!= stmt_loop
143 && !flow_loop_nested_p (inner_loop
, stmt_loop
))
146 for (it
= 0; it
< DR_NUM_DIMENSIONS (dr
);
147 it
++, ref
= TREE_OPERAND (ref
, 0))
149 tree chrec
= DR_ACCESS_FN (dr
, it
);
150 tree tstride
= evolution_part_in_loop_num (chrec
, loop
->num
);
151 tree array_size
= TYPE_SIZE (TREE_TYPE (ref
));
154 if (tstride
== NULL_TREE
155 || array_size
== NULL_TREE
156 || TREE_CODE (tstride
) != INTEGER_CST
157 || TREE_CODE (array_size
) != INTEGER_CST
)
160 dstride
= double_int_mul (tree_to_double_int (array_size
),
161 tree_to_double_int (tstride
));
162 (*access_strides
) = double_int_add (*access_strides
, dstride
);
167 /* Attempt to apply interchange transformations to TRANS to maximize the
168 spatial and temporal locality of the loop.
169 Returns the new transform matrix. The smaller the reuse vector
170 distances in the inner loops, the fewer the cache misses.
171 FIRST_LOOP is the loop->num of the first loop in the analyzed loop
175 static lambda_trans_matrix
176 try_interchange_loops (lambda_trans_matrix trans
,
178 VEC (ddr_p
, heap
) *dependence_relations
,
179 VEC (data_reference_p
, heap
) *datarefs
,
180 struct loop
*first_loop
)
184 unsigned int dependence_steps_i
, dependence_steps_j
;
185 double_int access_strides_i
, access_strides_j
;
186 unsigned int nb_deps_not_carried_by_i
, nb_deps_not_carried_by_j
;
187 struct data_dependence_relation
*ddr
;
189 if (VEC_length (ddr_p
, dependence_relations
) == 0)
192 /* When there is an unknown relation in the dependence_relations, we
193 know that it is no worth looking at this loop nest: give up. */
194 ddr
= VEC_index (ddr_p
, dependence_relations
, 0);
195 if (ddr
== NULL
|| DDR_ARE_DEPENDENT (ddr
) == chrec_dont_know
)
198 /* LOOP_I is always the outer loop. */
199 for (loop_j
= first_loop
->inner
;
201 loop_j
= loop_j
->inner
)
202 for (loop_i
= first_loop
;
203 loop_depth (loop_i
) < loop_depth (loop_j
);
204 loop_i
= loop_i
->inner
)
206 gather_interchange_stats (dependence_relations
, datarefs
,
209 &nb_deps_not_carried_by_i
,
211 gather_interchange_stats (dependence_relations
, datarefs
,
214 &nb_deps_not_carried_by_j
,
217 /* Heuristics for loop interchange profitability:
219 1. (spatial locality) Inner loops should have smallest
222 2. (spatial locality) Inner loops should contain more
223 dependence relations not carried by the loop.
225 3. (temporal locality) Inner loops should have smallest
226 array access strides.
228 if (dependence_steps_i
< dependence_steps_j
229 || nb_deps_not_carried_by_i
> nb_deps_not_carried_by_j
230 || double_int_ucmp (access_strides_i
, access_strides_j
) < 0)
232 lambda_matrix_row_exchange (LTM_MATRIX (trans
),
233 loop_depth (loop_i
) - loop_depth (first_loop
),
234 loop_depth (loop_j
) - loop_depth (first_loop
));
235 /* Validate the resulting matrix. When the transformation
236 is not valid, reverse to the previous transformation. */
237 if (!lambda_transform_legal_p (trans
, depth
, dependence_relations
))
238 lambda_matrix_row_exchange (LTM_MATRIX (trans
),
239 loop_depth (loop_i
) - loop_depth (first_loop
),
240 loop_depth (loop_j
) - loop_depth (first_loop
));
247 /* Perform a set of linear transforms on loops. */
250 linear_transform_loops (void)
252 bool modified
= false;
254 VEC(tree
,heap
) *oldivs
= NULL
;
255 VEC(tree
,heap
) *invariants
= NULL
;
256 VEC(tree
,heap
) *remove_ivs
= VEC_alloc (tree
, heap
, 3);
257 struct loop
*loop_nest
;
261 FOR_EACH_LOOP (li
, loop_nest
, 0)
263 unsigned int depth
= 0;
264 VEC (ddr_p
, heap
) *dependence_relations
;
265 VEC (data_reference_p
, heap
) *datarefs
;
267 lambda_loopnest before
, after
;
268 lambda_trans_matrix trans
;
269 bool problem
= false;
270 struct obstack lambda_obstack
;
271 gcc_obstack_init (&lambda_obstack
);
273 /* If it's not a loop nest, we don't want it.
274 We also don't handle sibling loops properly,
275 which are loops of the following form:
276 for (i = 0; i < 50; i++)
278 for (j = 0; j < 50; j++)
282 for (j = 0; j < 50; j++)
287 if (!loop_nest
->inner
|| !single_exit (loop_nest
))
289 VEC_truncate (tree
, oldivs
, 0);
290 VEC_truncate (tree
, invariants
, 0);
292 for (temp
= loop_nest
->inner
; temp
; temp
= temp
->inner
)
294 /* If we have a sibling loop or multiple exit edges, jump ship. */
295 if (temp
->next
|| !single_exit (temp
))
305 /* Analyze data references and dependence relations using scev. */
307 datarefs
= VEC_alloc (data_reference_p
, heap
, 10);
308 dependence_relations
= VEC_alloc (ddr_p
, heap
, 10 * 10);
309 compute_data_dependences_for_loop (loop_nest
, true, &datarefs
,
310 &dependence_relations
);
312 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
313 dump_ddrs (dump_file
, dependence_relations
);
315 /* Build the transformation matrix. */
316 trans
= lambda_trans_matrix_new (depth
, depth
);
317 lambda_matrix_id (LTM_MATRIX (trans
), depth
);
318 trans
= try_interchange_loops (trans
, depth
, dependence_relations
,
319 datarefs
, loop_nest
);
321 if (lambda_trans_matrix_id_p (trans
))
324 fprintf (dump_file
, "Won't transform loop. Optimal transform is the identity transform\n");
325 goto free_and_continue
;
328 /* Check whether the transformation is legal. */
329 if (!lambda_transform_legal_p (trans
, depth
, dependence_relations
))
332 fprintf (dump_file
, "Can't transform loop, transform is illegal:\n");
333 goto free_and_continue
;
336 before
= gcc_loopnest_to_lambda_loopnest (loop_nest
, &oldivs
,
337 &invariants
, &lambda_obstack
);
340 goto free_and_continue
;
344 fprintf (dump_file
, "Before:\n");
345 print_lambda_loopnest (dump_file
, before
, 'i');
348 after
= lambda_loopnest_transform (before
, trans
, &lambda_obstack
);
352 fprintf (dump_file
, "After:\n");
353 print_lambda_loopnest (dump_file
, after
, 'u');
356 lambda_loopnest_to_gcc_loopnest (loop_nest
, oldivs
, invariants
,
358 after
, trans
, &lambda_obstack
);
362 fprintf (dump_file
, "Successfully transformed loop.\n");
365 obstack_free (&lambda_obstack
, NULL
);
366 free_dependence_relations (dependence_relations
);
367 free_data_refs (datarefs
);
370 for (i
= 0; VEC_iterate (tree
, remove_ivs
, i
, oldiv_stmt
); i
++)
371 remove_iv (oldiv_stmt
);
373 VEC_free (tree
, heap
, oldivs
);
374 VEC_free (tree
, heap
, invariants
);
375 VEC_free (tree
, heap
, remove_ivs
);
379 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa_full_phi
);