]>
Commit | Line | Data |
---|---|---|
fbdec14e | 1 | /* Loop interchange. |
a5544970 | 2 | Copyright (C) 2017-2019 Free Software Foundation, Inc. |
fbdec14e BC |
3 | Contributed by ARM Ltd. |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it | |
8 | under the terms of the GNU General Public License as published by the | |
9 | Free Software Foundation; either version 3, or (at your option) any | |
10 | later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
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/>. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "backend.h" | |
25 | #include "is-a.h" | |
26 | #include "tree.h" | |
27 | #include "gimple.h" | |
28 | #include "tree-pass.h" | |
29 | #include "ssa.h" | |
30 | #include "gimple-pretty-print.h" | |
31 | #include "fold-const.h" | |
32 | #include "gimplify.h" | |
33 | #include "gimple-iterator.h" | |
34 | #include "gimplify-me.h" | |
35 | #include "cfgloop.h" | |
36 | #include "params.h" | |
37 | #include "tree-ssa.h" | |
38 | #include "tree-scalar-evolution.h" | |
39 | #include "tree-ssa-loop-manip.h" | |
40 | #include "tree-ssa-loop-niter.h" | |
41 | #include "tree-ssa-loop-ivopts.h" | |
42 | #include "tree-ssa-dce.h" | |
43 | #include "tree-data-ref.h" | |
44 | #include "tree-vectorizer.h" | |
45 | ||
46 | /* This pass performs loop interchange: for example, the loop nest | |
47 | ||
48 | for (int j = 0; j < N; j++) | |
49 | for (int k = 0; k < N; k++) | |
50 | for (int i = 0; i < N; i++) | |
51 | c[i][j] = c[i][j] + a[i][k]*b[k][j]; | |
52 | ||
53 | is transformed to | |
54 | ||
55 | for (int i = 0; i < N; i++) | |
56 | for (int j = 0; j < N; j++) | |
57 | for (int k = 0; k < N; k++) | |
58 | c[i][j] = c[i][j] + a[i][k]*b[k][j]; | |
59 | ||
60 | This pass implements loop interchange in the following steps: | |
61 | ||
62 | 1) Find perfect loop nest for each innermost loop and compute data | |
63 | dependence relations for it. For above example, loop nest is | |
64 | <loop_j, loop_k, loop_i>. | |
65 | 2) From innermost to outermost loop, this pass tries to interchange | |
66 | each loop pair. For above case, it firstly tries to interchange | |
67 | <loop_k, loop_i> and loop nest becomes <loop_j, loop_i, loop_k>. | |
68 | Then it tries to interchange <loop_j, loop_i> and loop nest becomes | |
69 | <loop_i, loop_j, loop_k>. The overall effect is to move innermost | |
70 | loop to the outermost position. For loop pair <loop_i, loop_j> | |
71 | to be interchanged, we: | |
72 | 3) Check if data dependence relations are valid for loop interchange. | |
73 | 4) Check if both loops can be interchanged in terms of transformation. | |
74 | 5) Check if interchanging the two loops is profitable. | |
75 | 6) Interchange the two loops by mapping induction variables. | |
76 | ||
77 | This pass also handles reductions in loop nest. So far we only support | |
78 | simple reduction of inner loop and double reduction of the loop nest. */ | |
79 | ||
80 | /* Maximum number of stmts in each loop that should be interchanged. */ | |
81 | #define MAX_NUM_STMT (PARAM_VALUE (PARAM_LOOP_INTERCHANGE_MAX_NUM_STMTS)) | |
82 | /* Maximum number of data references in loop nest. */ | |
83 | #define MAX_DATAREFS (PARAM_VALUE (PARAM_LOOP_MAX_DATAREFS_FOR_DATADEPS)) | |
84 | ||
85 | /* Comparison ratio of access stride between inner/outer loops to be | |
86 | interchanged. This is the minimum stride ratio for loop interchange | |
87 | to be profitable. */ | |
88 | #define OUTER_STRIDE_RATIO (PARAM_VALUE (PARAM_LOOP_INTERCHANGE_STRIDE_RATIO)) | |
89 | /* The same as above, but we require higher ratio for interchanging the | |
90 | innermost two loops. */ | |
91 | #define INNER_STRIDE_RATIO ((OUTER_STRIDE_RATIO) + 1) | |
92 | ||
1eeeda47 BC |
93 | /* Comparison ratio of stmt cost between inner/outer loops. Loops won't |
94 | be interchanged if outer loop has too many stmts. */ | |
95 | #define STMT_COST_RATIO (3) | |
96 | ||
fbdec14e BC |
97 | /* Vector of strides that DR accesses in each level loop of a loop nest. */ |
98 | #define DR_ACCESS_STRIDE(dr) ((vec<tree> *) dr->aux) | |
99 | ||
100 | /* Structure recording loop induction variable. */ | |
101 | typedef struct induction | |
102 | { | |
103 | /* IV itself. */ | |
104 | tree var; | |
105 | /* IV's initializing value, which is the init arg of the IV PHI node. */ | |
106 | tree init_val; | |
107 | /* IV's initializing expr, which is (the expanded result of) init_val. */ | |
108 | tree init_expr; | |
109 | /* IV's step. */ | |
110 | tree step; | |
111 | } *induction_p; | |
112 | ||
113 | /* Enum type for loop reduction variable. */ | |
114 | enum reduction_type | |
115 | { | |
116 | UNKNOWN_RTYPE = 0, | |
117 | SIMPLE_RTYPE, | |
118 | DOUBLE_RTYPE | |
119 | }; | |
120 | ||
121 | /* Structure recording loop reduction variable. */ | |
122 | typedef struct reduction | |
123 | { | |
124 | /* Reduction itself. */ | |
125 | tree var; | |
126 | /* PHI node defining reduction variable. */ | |
127 | gphi *phi; | |
128 | /* Init and next variables of the reduction. */ | |
129 | tree init; | |
130 | tree next; | |
131 | /* Lcssa PHI node if reduction is used outside of its definition loop. */ | |
132 | gphi *lcssa_phi; | |
133 | /* Stmts defining init and next. */ | |
134 | gimple *producer; | |
135 | gimple *consumer; | |
136 | /* If init is loaded from memory, this is the loading memory reference. */ | |
137 | tree init_ref; | |
138 | /* If reduction is finally stored to memory, this is the stored memory | |
139 | reference. */ | |
140 | tree fini_ref; | |
141 | enum reduction_type type; | |
142 | } *reduction_p; | |
143 | ||
144 | ||
145 | /* Dump reduction RE. */ | |
146 | ||
147 | static void | |
148 | dump_reduction (reduction_p re) | |
149 | { | |
150 | if (re->type == SIMPLE_RTYPE) | |
151 | fprintf (dump_file, " Simple reduction: "); | |
152 | else if (re->type == DOUBLE_RTYPE) | |
153 | fprintf (dump_file, " Double reduction: "); | |
154 | else | |
155 | fprintf (dump_file, " Unknown reduction: "); | |
156 | ||
157 | print_gimple_stmt (dump_file, re->phi, 0); | |
158 | } | |
159 | ||
160 | /* Dump LOOP's induction IV. */ | |
161 | static void | |
162 | dump_induction (struct loop *loop, induction_p iv) | |
163 | { | |
164 | fprintf (dump_file, " Induction: "); | |
165 | print_generic_expr (dump_file, iv->var, TDF_SLIM); | |
166 | fprintf (dump_file, " = {"); | |
167 | print_generic_expr (dump_file, iv->init_expr, TDF_SLIM); | |
168 | fprintf (dump_file, ", "); | |
169 | print_generic_expr (dump_file, iv->step, TDF_SLIM); | |
170 | fprintf (dump_file, "}_%d\n", loop->num); | |
171 | } | |
172 | ||
173 | /* Loop candidate for interchange. */ | |
174 | ||
6c1dae73 | 175 | class loop_cand |
fbdec14e | 176 | { |
6c1dae73 | 177 | public: |
fbdec14e BC |
178 | loop_cand (struct loop *, struct loop *); |
179 | ~loop_cand (); | |
180 | ||
181 | reduction_p find_reduction_by_stmt (gimple *); | |
182 | void classify_simple_reduction (reduction_p); | |
183 | bool analyze_iloop_reduction_var (tree); | |
184 | bool analyze_oloop_reduction_var (loop_cand *, tree); | |
185 | bool analyze_induction_var (tree, tree); | |
186 | bool analyze_carried_vars (loop_cand *); | |
187 | bool analyze_lcssa_phis (void); | |
188 | bool can_interchange_p (loop_cand *); | |
fbdec14e BC |
189 | void undo_simple_reduction (reduction_p, bitmap); |
190 | ||
191 | /* The loop itself. */ | |
192 | struct loop *m_loop; | |
193 | /* The outer loop for interchange. It equals to loop if this loop cand | |
194 | itself represents the outer loop. */ | |
195 | struct loop *m_outer; | |
196 | /* Vector of induction variables in loop. */ | |
197 | vec<induction_p> m_inductions; | |
198 | /* Vector of reduction variables in loop. */ | |
199 | vec<reduction_p> m_reductions; | |
200 | /* Lcssa PHI nodes of this loop. */ | |
201 | vec<gphi *> m_lcssa_nodes; | |
202 | /* Single exit edge of this loop. */ | |
203 | edge m_exit; | |
204 | /* Basic blocks of this loop. */ | |
205 | basic_block *m_bbs; | |
1eeeda47 BC |
206 | /* Number of stmts of this loop. Inner loops' stmts are not included. */ |
207 | int m_num_stmts; | |
208 | /* Number of constant initialized simple reduction. */ | |
209 | int m_const_init_reduc; | |
fbdec14e BC |
210 | }; |
211 | ||
212 | /* Constructor. */ | |
213 | ||
214 | loop_cand::loop_cand (struct loop *loop, struct loop *outer) | |
1eeeda47 BC |
215 | : m_loop (loop), m_outer (outer), m_exit (single_exit (loop)), |
216 | m_bbs (get_loop_body (loop)), m_num_stmts (0), m_const_init_reduc (0) | |
fbdec14e BC |
217 | { |
218 | m_inductions.create (3); | |
219 | m_reductions.create (3); | |
220 | m_lcssa_nodes.create (3); | |
221 | } | |
222 | ||
223 | /* Destructor. */ | |
224 | ||
225 | loop_cand::~loop_cand () | |
226 | { | |
227 | induction_p iv; | |
228 | for (unsigned i = 0; m_inductions.iterate (i, &iv); ++i) | |
229 | free (iv); | |
230 | ||
231 | reduction_p re; | |
232 | for (unsigned i = 0; m_reductions.iterate (i, &re); ++i) | |
233 | free (re); | |
234 | ||
235 | m_inductions.release (); | |
236 | m_reductions.release (); | |
237 | m_lcssa_nodes.release (); | |
238 | free (m_bbs); | |
239 | } | |
240 | ||
241 | /* Return single use stmt of VAR in LOOP, otherwise return NULL. */ | |
242 | ||
243 | static gimple * | |
244 | single_use_in_loop (tree var, struct loop *loop) | |
245 | { | |
246 | gimple *stmt, *res = NULL; | |
247 | use_operand_p use_p; | |
248 | imm_use_iterator iterator; | |
249 | ||
250 | FOR_EACH_IMM_USE_FAST (use_p, iterator, var) | |
251 | { | |
252 | stmt = USE_STMT (use_p); | |
253 | if (is_gimple_debug (stmt)) | |
254 | continue; | |
255 | ||
256 | if (!flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
257 | continue; | |
258 | ||
259 | if (res) | |
260 | return NULL; | |
261 | ||
262 | res = stmt; | |
263 | } | |
264 | return res; | |
265 | } | |
266 | ||
267 | /* Return true if E is unsupported in loop interchange, i.e, E is a complex | |
268 | edge or part of irreducible loop. */ | |
269 | ||
270 | static inline bool | |
271 | unsupported_edge (edge e) | |
272 | { | |
273 | return (e->flags & (EDGE_COMPLEX | EDGE_IRREDUCIBLE_LOOP)); | |
274 | } | |
275 | ||
276 | /* Return the reduction if STMT is one of its lcssa PHI, producer or consumer | |
277 | stmt. */ | |
278 | ||
279 | reduction_p | |
280 | loop_cand::find_reduction_by_stmt (gimple *stmt) | |
281 | { | |
282 | gphi *phi = dyn_cast <gphi *> (stmt); | |
283 | reduction_p re; | |
284 | ||
285 | for (unsigned i = 0; m_reductions.iterate (i, &re); ++i) | |
286 | if ((phi != NULL && phi == re->lcssa_phi) | |
287 | || (stmt == re->producer || stmt == re->consumer)) | |
288 | return re; | |
289 | ||
290 | return NULL; | |
291 | } | |
292 | ||
fbdec14e BC |
293 | /* Return true if current loop_cand be interchanged. ILOOP is not NULL if |
294 | current loop_cand is outer loop in loop nest. */ | |
295 | ||
296 | bool | |
297 | loop_cand::can_interchange_p (loop_cand *iloop) | |
298 | { | |
299 | /* For now we only support at most one reduction. */ | |
300 | unsigned allowed_reduction_num = 1; | |
301 | ||
302 | /* Only support reduction if the loop nest to be interchanged is the | |
303 | innermostin two loops. */ | |
304 | if ((iloop == NULL && m_loop->inner != NULL) | |
305 | || (iloop != NULL && iloop->m_loop->inner != NULL)) | |
306 | allowed_reduction_num = 0; | |
307 | ||
308 | if (m_reductions.length () > allowed_reduction_num | |
309 | || (m_reductions.length () == 1 | |
310 | && m_reductions[0]->type == UNKNOWN_RTYPE)) | |
311 | return false; | |
312 | ||
313 | /* Only support lcssa PHI node which is for reduction. */ | |
314 | if (m_lcssa_nodes.length () > allowed_reduction_num) | |
315 | return false; | |
316 | ||
1eeeda47 BC |
317 | /* Check if basic block has any unsupported operation. Note basic blocks |
318 | of inner loops are not checked here. */ | |
fbdec14e BC |
319 | for (unsigned i = 0; i < m_loop->num_nodes; i++) |
320 | { | |
321 | basic_block bb = m_bbs[i]; | |
1eeeda47 BC |
322 | gphi_iterator psi; |
323 | gimple_stmt_iterator gsi; | |
fbdec14e BC |
324 | |
325 | /* Skip basic blocks of inner loops. */ | |
326 | if (bb->loop_father != m_loop) | |
1eeeda47 | 327 | continue; |
fbdec14e | 328 | |
1eeeda47 BC |
329 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
330 | { | |
331 | gimple *stmt = gsi_stmt (gsi); | |
332 | if (is_gimple_debug (stmt)) | |
333 | continue; | |
334 | ||
335 | if (gimple_has_side_effects (stmt)) | |
336 | return false; | |
337 | ||
338 | m_num_stmts++; | |
339 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
340 | { | |
341 | /* In basic block of outer loop, the call should be cheap since | |
342 | it will be moved to inner loop. */ | |
343 | if (iloop != NULL | |
344 | && !gimple_inexpensive_call_p (call)) | |
345 | return false; | |
346 | continue; | |
347 | } | |
348 | ||
349 | if (!iloop || !gimple_vuse (stmt)) | |
350 | continue; | |
fbdec14e | 351 | |
1eeeda47 BC |
352 | /* Support stmt accessing memory in outer loop only if it is for |
353 | inner loop's reduction. */ | |
354 | if (iloop->find_reduction_by_stmt (stmt)) | |
355 | continue; | |
356 | ||
357 | tree lhs; | |
358 | /* Support loop invariant memory reference if it's only used once by | |
359 | inner loop. */ | |
360 | /* ??? How's this checking for invariantness? */ | |
361 | if (gimple_assign_single_p (stmt) | |
362 | && (lhs = gimple_assign_lhs (stmt)) != NULL_TREE | |
363 | && TREE_CODE (lhs) == SSA_NAME | |
364 | && single_use_in_loop (lhs, iloop->m_loop)) | |
365 | continue; | |
366 | ||
367 | return false; | |
368 | } | |
fbdec14e | 369 | /* Check if loop has too many stmts. */ |
1eeeda47 | 370 | if (m_num_stmts > MAX_NUM_STMT) |
fbdec14e | 371 | return false; |
1eeeda47 BC |
372 | |
373 | /* Allow PHI nodes in any basic block of inner loop, PHI nodes in outer | |
374 | loop's header, or PHI nodes in dest bb of inner loop's exit edge. */ | |
375 | if (!iloop || bb == m_loop->header | |
376 | || bb == iloop->m_exit->dest) | |
377 | continue; | |
378 | ||
379 | /* Don't allow any other PHI nodes. */ | |
380 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
381 | if (!virtual_operand_p (PHI_RESULT (psi.phi ()))) | |
382 | return false; | |
fbdec14e BC |
383 | } |
384 | ||
385 | return true; | |
386 | } | |
387 | ||
388 | /* Programmers and optimizers (like loop store motion) may optimize code: | |
389 | ||
390 | for (int i = 0; i < N; i++) | |
391 | for (int j = 0; j < N; j++) | |
392 | a[i] += b[j][i] * c[j][i]; | |
393 | ||
394 | into reduction: | |
395 | ||
396 | for (int i = 0; i < N; i++) | |
397 | { | |
398 | // producer. Note sum can be intitialized to a constant. | |
399 | int sum = a[i]; | |
400 | for (int j = 0; j < N; j++) | |
401 | { | |
402 | sum += b[j][i] * c[j][i]; | |
403 | } | |
404 | // consumer. | |
405 | a[i] = sum; | |
406 | } | |
407 | ||
408 | The result code can't be interchanged without undoing the optimization. | |
409 | This function classifies this kind reduction and records information so | |
410 | that we can undo the store motion during interchange. */ | |
411 | ||
412 | void | |
413 | loop_cand::classify_simple_reduction (reduction_p re) | |
414 | { | |
415 | gimple *producer, *consumer; | |
416 | ||
417 | /* Check init variable of reduction and how it is initialized. */ | |
418 | if (TREE_CODE (re->init) == SSA_NAME) | |
419 | { | |
420 | producer = SSA_NAME_DEF_STMT (re->init); | |
421 | re->producer = producer; | |
422 | basic_block bb = gimple_bb (producer); | |
423 | if (!bb || bb->loop_father != m_outer) | |
424 | return; | |
425 | ||
426 | if (!gimple_assign_load_p (producer)) | |
427 | return; | |
428 | ||
429 | re->init_ref = gimple_assign_rhs1 (producer); | |
430 | } | |
1eeeda47 BC |
431 | else if (CONSTANT_CLASS_P (re->init)) |
432 | m_const_init_reduc++; | |
433 | else | |
fbdec14e BC |
434 | return; |
435 | ||
436 | /* Check how reduction variable is used. */ | |
437 | consumer = single_use_in_loop (PHI_RESULT (re->lcssa_phi), m_outer); | |
438 | if (!consumer | |
439 | || !gimple_store_p (consumer)) | |
440 | return; | |
441 | ||
442 | re->fini_ref = gimple_get_lhs (consumer); | |
443 | re->consumer = consumer; | |
444 | ||
445 | /* Simple reduction with constant initializer. */ | |
446 | if (!re->init_ref) | |
447 | { | |
448 | gcc_assert (CONSTANT_CLASS_P (re->init)); | |
449 | re->init_ref = unshare_expr (re->fini_ref); | |
450 | } | |
451 | ||
452 | /* Require memory references in producer and consumer are the same so | |
453 | that we can undo reduction during interchange. */ | |
454 | if (re->init_ref && !operand_equal_p (re->init_ref, re->fini_ref, 0)) | |
455 | return; | |
456 | ||
457 | re->type = SIMPLE_RTYPE; | |
458 | } | |
459 | ||
460 | /* Analyze reduction variable VAR for inner loop of the loop nest to be | |
461 | interchanged. Return true if analysis succeeds. */ | |
462 | ||
463 | bool | |
464 | loop_cand::analyze_iloop_reduction_var (tree var) | |
465 | { | |
466 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (var)); | |
467 | gphi *lcssa_phi = NULL, *use_phi; | |
468 | tree init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (m_loop)); | |
469 | tree next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (m_loop)); | |
470 | reduction_p re; | |
471 | gimple *stmt, *next_def, *single_use = NULL; | |
472 | use_operand_p use_p; | |
473 | imm_use_iterator iterator; | |
474 | ||
475 | if (TREE_CODE (next) != SSA_NAME) | |
476 | return false; | |
477 | ||
478 | next_def = SSA_NAME_DEF_STMT (next); | |
479 | basic_block bb = gimple_bb (next_def); | |
480 | if (!bb || !flow_bb_inside_loop_p (m_loop, bb)) | |
481 | return false; | |
482 | ||
483 | /* In restricted reduction, the var is (and must be) used in defining | |
484 | the updated var. The process can be depicted as below: | |
485 | ||
486 | var ;; = PHI<init, next> | |
487 | | | |
488 | | | |
489 | v | |
490 | +---------------------+ | |
491 | | reduction operators | <-- other operands | |
492 | +---------------------+ | |
493 | | | |
494 | | | |
495 | v | |
496 | next | |
497 | ||
498 | In terms loop interchange, we don't change how NEXT is computed based | |
499 | on VAR and OTHER OPERANDS. In case of double reduction in loop nest | |
500 | to be interchanged, we don't changed it at all. In the case of simple | |
501 | reduction in inner loop, we only make change how VAR/NEXT is loaded or | |
502 | stored. With these conditions, we can relax restrictions on reduction | |
503 | in a way that reduction operation is seen as black box. In general, | |
504 | we can ignore reassociation of reduction operator; we can handle fake | |
505 | reductions in which VAR is not even used to compute NEXT. */ | |
506 | if (! single_imm_use (var, &use_p, &single_use) | |
507 | || ! flow_bb_inside_loop_p (m_loop, gimple_bb (single_use))) | |
508 | return false; | |
509 | ||
510 | /* Check the reduction operation. We require a left-associative operation. | |
511 | For FP math we also need to be allowed to associate operations. */ | |
512 | if (gassign *ass = dyn_cast <gassign *> (single_use)) | |
513 | { | |
514 | enum tree_code code = gimple_assign_rhs_code (ass); | |
515 | if (! (associative_tree_code (code) | |
516 | || (code == MINUS_EXPR | |
517 | && use_p->use == gimple_assign_rhs1_ptr (ass))) | |
518 | || (FLOAT_TYPE_P (TREE_TYPE (var)) | |
519 | && ! flag_associative_math)) | |
520 | return false; | |
521 | } | |
522 | else | |
523 | return false; | |
524 | ||
525 | /* Handle and verify a series of stmts feeding the reduction op. */ | |
526 | if (single_use != next_def | |
4f5b9c80 | 527 | && !check_reduction_path (dump_user_location_t (), m_loop, phi, next, |
fbdec14e BC |
528 | gimple_assign_rhs_code (single_use))) |
529 | return false; | |
530 | ||
531 | /* Only support cases in which INIT is used in inner loop. */ | |
532 | if (TREE_CODE (init) == SSA_NAME) | |
533 | FOR_EACH_IMM_USE_FAST (use_p, iterator, init) | |
534 | { | |
535 | stmt = USE_STMT (use_p); | |
536 | if (is_gimple_debug (stmt)) | |
537 | continue; | |
538 | ||
539 | if (!flow_bb_inside_loop_p (m_loop, gimple_bb (stmt))) | |
540 | return false; | |
541 | } | |
542 | ||
543 | FOR_EACH_IMM_USE_FAST (use_p, iterator, next) | |
544 | { | |
545 | stmt = USE_STMT (use_p); | |
546 | if (is_gimple_debug (stmt)) | |
547 | continue; | |
548 | ||
549 | /* Or else it's used in PHI itself. */ | |
550 | use_phi = dyn_cast <gphi *> (stmt); | |
551 | if (use_phi == phi) | |
552 | continue; | |
553 | ||
554 | if (use_phi != NULL | |
555 | && lcssa_phi == NULL | |
556 | && gimple_bb (stmt) == m_exit->dest | |
557 | && PHI_ARG_DEF_FROM_EDGE (use_phi, m_exit) == next) | |
558 | lcssa_phi = use_phi; | |
559 | else | |
560 | return false; | |
561 | } | |
562 | if (!lcssa_phi) | |
563 | return false; | |
564 | ||
565 | re = XCNEW (struct reduction); | |
566 | re->var = var; | |
567 | re->init = init; | |
568 | re->next = next; | |
569 | re->phi = phi; | |
570 | re->lcssa_phi = lcssa_phi; | |
571 | ||
572 | classify_simple_reduction (re); | |
573 | ||
574 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
575 | dump_reduction (re); | |
576 | ||
577 | m_reductions.safe_push (re); | |
578 | return true; | |
579 | } | |
580 | ||
581 | /* Analyze reduction variable VAR for outer loop of the loop nest to be | |
582 | interchanged. ILOOP is not NULL and points to inner loop. For the | |
583 | moment, we only support double reduction for outer loop, like: | |
584 | ||
585 | for (int i = 0; i < n; i++) | |
586 | { | |
587 | int sum = 0; | |
588 | ||
589 | for (int j = 0; j < n; j++) // outer loop | |
590 | for (int k = 0; k < n; k++) // inner loop | |
591 | sum += a[i][k]*b[k][j]; | |
592 | ||
593 | s[i] = sum; | |
594 | } | |
595 | ||
596 | Note the innermost two loops are the loop nest to be interchanged. | |
597 | Return true if analysis succeeds. */ | |
598 | ||
599 | bool | |
600 | loop_cand::analyze_oloop_reduction_var (loop_cand *iloop, tree var) | |
601 | { | |
602 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (var)); | |
603 | gphi *lcssa_phi = NULL, *use_phi; | |
604 | tree init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (m_loop)); | |
605 | tree next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (m_loop)); | |
606 | reduction_p re; | |
607 | gimple *stmt, *next_def; | |
608 | use_operand_p use_p; | |
609 | imm_use_iterator iterator; | |
610 | ||
611 | if (TREE_CODE (next) != SSA_NAME) | |
612 | return false; | |
613 | ||
614 | next_def = SSA_NAME_DEF_STMT (next); | |
615 | basic_block bb = gimple_bb (next_def); | |
616 | if (!bb || !flow_bb_inside_loop_p (m_loop, bb)) | |
617 | return false; | |
618 | ||
619 | /* Find inner loop's simple reduction that uses var as initializer. */ | |
620 | reduction_p inner_re = NULL; | |
621 | for (unsigned i = 0; iloop->m_reductions.iterate (i, &inner_re); ++i) | |
622 | if (inner_re->init == var || operand_equal_p (inner_re->init, var, 0)) | |
623 | break; | |
624 | ||
625 | if (inner_re == NULL | |
626 | || inner_re->type != UNKNOWN_RTYPE | |
627 | || inner_re->producer != phi) | |
628 | return false; | |
629 | ||
630 | /* In case of double reduction, outer loop's reduction should be updated | |
631 | by inner loop's simple reduction. */ | |
632 | if (next_def != inner_re->lcssa_phi) | |
633 | return false; | |
634 | ||
635 | /* Outer loop's reduction should only be used to initialize inner loop's | |
636 | simple reduction. */ | |
637 | if (! single_imm_use (var, &use_p, &stmt) | |
638 | || stmt != inner_re->phi) | |
639 | return false; | |
640 | ||
641 | /* Check this reduction is correctly used outside of loop via lcssa phi. */ | |
642 | FOR_EACH_IMM_USE_FAST (use_p, iterator, next) | |
643 | { | |
644 | stmt = USE_STMT (use_p); | |
645 | if (is_gimple_debug (stmt)) | |
646 | continue; | |
647 | ||
648 | /* Or else it's used in PHI itself. */ | |
649 | use_phi = dyn_cast <gphi *> (stmt); | |
650 | if (use_phi == phi) | |
651 | continue; | |
652 | ||
653 | if (lcssa_phi == NULL | |
654 | && use_phi != NULL | |
655 | && gimple_bb (stmt) == m_exit->dest | |
656 | && PHI_ARG_DEF_FROM_EDGE (use_phi, m_exit) == next) | |
657 | lcssa_phi = use_phi; | |
658 | else | |
659 | return false; | |
660 | } | |
661 | if (!lcssa_phi) | |
662 | return false; | |
663 | ||
664 | re = XCNEW (struct reduction); | |
665 | re->var = var; | |
666 | re->init = init; | |
667 | re->next = next; | |
668 | re->phi = phi; | |
669 | re->lcssa_phi = lcssa_phi; | |
670 | re->type = DOUBLE_RTYPE; | |
671 | inner_re->type = DOUBLE_RTYPE; | |
672 | ||
673 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
674 | dump_reduction (re); | |
675 | ||
676 | m_reductions.safe_push (re); | |
677 | return true; | |
678 | } | |
679 | ||
680 | /* Return true if VAR is induction variable of current loop whose scev is | |
681 | specified by CHREC. */ | |
682 | ||
683 | bool | |
684 | loop_cand::analyze_induction_var (tree var, tree chrec) | |
685 | { | |
686 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (var)); | |
687 | tree init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (m_loop)); | |
688 | ||
689 | /* Var is loop invariant, though it's unlikely to happen. */ | |
690 | if (tree_does_not_contain_chrecs (chrec)) | |
691 | { | |
37aa6856 JJ |
692 | /* Punt on floating point invariants if honoring signed zeros, |
693 | representing that as + 0.0 would change the result if init | |
f359ba2f JJ |
694 | is -0.0. Similarly for SNaNs it can raise exception. */ |
695 | if (HONOR_SIGNED_ZEROS (chrec) || HONOR_SNANS (chrec)) | |
37aa6856 | 696 | return false; |
fbdec14e BC |
697 | struct induction *iv = XCNEW (struct induction); |
698 | iv->var = var; | |
699 | iv->init_val = init; | |
700 | iv->init_expr = chrec; | |
37aa6856 | 701 | iv->step = build_zero_cst (TREE_TYPE (chrec)); |
fbdec14e BC |
702 | m_inductions.safe_push (iv); |
703 | return true; | |
704 | } | |
705 | ||
706 | if (TREE_CODE (chrec) != POLYNOMIAL_CHREC | |
707 | || CHREC_VARIABLE (chrec) != (unsigned) m_loop->num | |
708 | || tree_contains_chrecs (CHREC_LEFT (chrec), NULL) | |
709 | || tree_contains_chrecs (CHREC_RIGHT (chrec), NULL)) | |
710 | return false; | |
711 | ||
712 | struct induction *iv = XCNEW (struct induction); | |
713 | iv->var = var; | |
714 | iv->init_val = init; | |
715 | iv->init_expr = CHREC_LEFT (chrec); | |
716 | iv->step = CHREC_RIGHT (chrec); | |
717 | ||
718 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
719 | dump_induction (m_loop, iv); | |
720 | ||
721 | m_inductions.safe_push (iv); | |
722 | return true; | |
723 | } | |
724 | ||
725 | /* Return true if all loop carried variables defined in loop header can | |
726 | be successfully analyzed. */ | |
727 | ||
728 | bool | |
729 | loop_cand::analyze_carried_vars (loop_cand *iloop) | |
730 | { | |
731 | edge e = loop_preheader_edge (m_outer); | |
732 | gphi_iterator gsi; | |
733 | ||
734 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
735 | fprintf (dump_file, "\nLoop(%d) carried vars:\n", m_loop->num); | |
736 | ||
737 | for (gsi = gsi_start_phis (m_loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
738 | { | |
739 | gphi *phi = gsi.phi (); | |
740 | ||
741 | tree var = PHI_RESULT (phi); | |
742 | if (virtual_operand_p (var)) | |
743 | continue; | |
744 | ||
745 | tree chrec = analyze_scalar_evolution (m_loop, var); | |
746 | chrec = instantiate_scev (e, m_loop, chrec); | |
747 | ||
748 | /* Analyze var as reduction variable. */ | |
749 | if (chrec_contains_undetermined (chrec) | |
750 | || chrec_contains_symbols_defined_in_loop (chrec, m_outer->num)) | |
751 | { | |
752 | if (iloop && !analyze_oloop_reduction_var (iloop, var)) | |
753 | return false; | |
754 | if (!iloop && !analyze_iloop_reduction_var (var)) | |
755 | return false; | |
756 | } | |
757 | /* Analyze var as induction variable. */ | |
758 | else if (!analyze_induction_var (var, chrec)) | |
759 | return false; | |
760 | } | |
761 | ||
762 | return true; | |
763 | } | |
764 | ||
765 | /* Return TRUE if loop closed PHI nodes can be analyzed successfully. */ | |
766 | ||
767 | bool | |
768 | loop_cand::analyze_lcssa_phis (void) | |
769 | { | |
770 | gphi_iterator gsi; | |
771 | for (gsi = gsi_start_phis (m_exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
772 | { | |
773 | gphi *phi = gsi.phi (); | |
774 | ||
775 | if (virtual_operand_p (PHI_RESULT (phi))) | |
776 | continue; | |
777 | ||
778 | /* TODO: We only support lcssa phi for reduction for now. */ | |
779 | if (!find_reduction_by_stmt (phi)) | |
780 | return false; | |
781 | } | |
782 | ||
783 | return true; | |
784 | } | |
785 | ||
786 | /* CONSUMER is a stmt in BB storing reduction result into memory object. | |
787 | When the reduction is intialized from constant value, we need to add | |
788 | a stmt loading from the memory object to target basic block in inner | |
789 | loop during undoing the reduction. Problem is that memory reference | |
790 | may use ssa variables not dominating the target basic block. This | |
791 | function finds all stmts on which CONSUMER depends in basic block BB, | |
792 | records and returns them via STMTS. */ | |
793 | ||
794 | static void | |
795 | find_deps_in_bb_for_stmt (gimple_seq *stmts, basic_block bb, gimple *consumer) | |
796 | { | |
797 | auto_vec<gimple *, 4> worklist; | |
798 | use_operand_p use_p; | |
799 | ssa_op_iter iter; | |
800 | gimple *stmt, *def_stmt; | |
801 | gimple_stmt_iterator gsi; | |
802 | ||
803 | /* First clear flag for stmts in bb. */ | |
804 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
805 | gimple_set_plf (gsi_stmt (gsi), GF_PLF_1, false); | |
806 | ||
807 | /* DFS search all depended stmts in bb and mark flag for these stmts. */ | |
808 | worklist.safe_push (consumer); | |
809 | while (!worklist.is_empty ()) | |
810 | { | |
811 | stmt = worklist.pop (); | |
812 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) | |
813 | { | |
814 | def_stmt = SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p)); | |
815 | ||
816 | if (is_a <gphi *> (def_stmt) | |
817 | || gimple_bb (def_stmt) != bb | |
818 | || gimple_plf (def_stmt, GF_PLF_1)) | |
819 | continue; | |
820 | ||
821 | worklist.safe_push (def_stmt); | |
822 | } | |
823 | gimple_set_plf (stmt, GF_PLF_1, true); | |
824 | } | |
65f4b875 | 825 | for (gsi = gsi_start_nondebug_bb (bb); |
fbdec14e BC |
826 | !gsi_end_p (gsi) && (stmt = gsi_stmt (gsi)) != consumer;) |
827 | { | |
828 | /* Move dep stmts to sequence STMTS. */ | |
829 | if (gimple_plf (stmt, GF_PLF_1)) | |
830 | { | |
831 | gsi_remove (&gsi, false); | |
832 | gimple_seq_add_stmt_without_update (stmts, stmt); | |
833 | } | |
834 | else | |
835 | gsi_next_nondebug (&gsi); | |
836 | } | |
837 | } | |
838 | ||
839 | /* User can write, optimizers can generate simple reduction RE for inner | |
840 | loop. In order to make interchange valid, we have to undo reduction by | |
841 | moving producer and consumer stmts into the inner loop. For example, | |
842 | below code: | |
843 | ||
844 | init = MEM_REF[idx]; //producer | |
845 | loop: | |
846 | var = phi<init, next> | |
847 | next = var op ... | |
848 | reduc_sum = phi<next> | |
849 | MEM_REF[idx] = reduc_sum //consumer | |
850 | ||
851 | is transformed into: | |
852 | ||
853 | loop: | |
854 | new_var = MEM_REF[idx]; //producer after moving | |
855 | next = new_var op ... | |
856 | MEM_REF[idx] = next; //consumer after moving | |
857 | ||
858 | Note if the reduction variable is initialized to constant, like: | |
859 | ||
860 | var = phi<0.0, next> | |
861 | ||
862 | we compute new_var as below: | |
863 | ||
864 | loop: | |
865 | tmp = MEM_REF[idx]; | |
866 | new_var = !first_iteration ? tmp : 0.0; | |
867 | ||
868 | so that the initial const is used in the first iteration of loop. Also | |
869 | record ssa variables for dead code elimination in DCE_SEEDS. */ | |
870 | ||
871 | void | |
872 | loop_cand::undo_simple_reduction (reduction_p re, bitmap dce_seeds) | |
873 | { | |
874 | gimple *stmt; | |
875 | gimple_stmt_iterator from, to = gsi_after_labels (m_loop->header); | |
876 | gimple_seq stmts = NULL; | |
877 | tree new_var; | |
878 | ||
879 | /* Prepare the initialization stmts and insert it to inner loop. */ | |
880 | if (re->producer != NULL) | |
881 | { | |
882 | gimple_set_vuse (re->producer, NULL_TREE); | |
883 | from = gsi_for_stmt (re->producer); | |
884 | gsi_remove (&from, false); | |
885 | gimple_seq_add_stmt_without_update (&stmts, re->producer); | |
886 | new_var = re->init; | |
887 | } | |
888 | else | |
889 | { | |
890 | /* Find all stmts on which expression "MEM_REF[idx]" depends. */ | |
891 | find_deps_in_bb_for_stmt (&stmts, gimple_bb (re->consumer), re->consumer); | |
892 | /* Because we generate new stmt loading from the MEM_REF to TMP. */ | |
893 | tree cond, tmp = copy_ssa_name (re->var); | |
894 | stmt = gimple_build_assign (tmp, re->init_ref); | |
895 | gimple_seq_add_stmt_without_update (&stmts, stmt); | |
896 | ||
897 | /* Init new_var to MEM_REF or CONST depending on if it is the first | |
898 | iteration. */ | |
899 | induction_p iv = m_inductions[0]; | |
900 | cond = fold_build2 (NE_EXPR, boolean_type_node, iv->var, iv->init_val); | |
901 | new_var = copy_ssa_name (re->var); | |
902 | stmt = gimple_build_assign (new_var, COND_EXPR, cond, tmp, re->init); | |
903 | gimple_seq_add_stmt_without_update (&stmts, stmt); | |
904 | } | |
905 | gsi_insert_seq_before (&to, stmts, GSI_SAME_STMT); | |
906 | ||
907 | /* Replace all uses of reduction var with new variable. */ | |
908 | use_operand_p use_p; | |
909 | imm_use_iterator iterator; | |
910 | FOR_EACH_IMM_USE_STMT (stmt, iterator, re->var) | |
911 | { | |
912 | FOR_EACH_IMM_USE_ON_STMT (use_p, iterator) | |
913 | SET_USE (use_p, new_var); | |
914 | ||
915 | update_stmt (stmt); | |
916 | } | |
917 | ||
918 | /* Move consumer stmt into inner loop, just after reduction next's def. */ | |
919 | unlink_stmt_vdef (re->consumer); | |
920 | release_ssa_name (gimple_vdef (re->consumer)); | |
921 | gimple_set_vdef (re->consumer, NULL_TREE); | |
922 | gimple_set_vuse (re->consumer, NULL_TREE); | |
923 | gimple_assign_set_rhs1 (re->consumer, re->next); | |
924 | from = gsi_for_stmt (re->consumer); | |
925 | to = gsi_for_stmt (SSA_NAME_DEF_STMT (re->next)); | |
926 | gsi_move_after (&from, &to); | |
927 | ||
928 | /* Mark the reduction variables for DCE. */ | |
929 | bitmap_set_bit (dce_seeds, SSA_NAME_VERSION (re->var)); | |
930 | bitmap_set_bit (dce_seeds, SSA_NAME_VERSION (PHI_RESULT (re->lcssa_phi))); | |
931 | } | |
932 | ||
933 | /* Free DATAREFS and its auxiliary memory. */ | |
934 | ||
935 | static void | |
936 | free_data_refs_with_aux (vec<data_reference_p> datarefs) | |
937 | { | |
938 | data_reference_p dr; | |
939 | for (unsigned i = 0; datarefs.iterate (i, &dr); ++i) | |
940 | if (dr->aux != NULL) | |
941 | { | |
942 | DR_ACCESS_STRIDE (dr)->release (); | |
46bb9d29 | 943 | delete (vec<tree> *) dr->aux; |
fbdec14e BC |
944 | } |
945 | ||
946 | free_data_refs (datarefs); | |
947 | } | |
948 | ||
949 | /* Class for loop interchange transformation. */ | |
950 | ||
951 | class tree_loop_interchange | |
952 | { | |
953 | public: | |
954 | tree_loop_interchange (vec<struct loop *> loop_nest) | |
955 | : m_loop_nest (loop_nest), m_niters_iv_var (NULL_TREE), | |
956 | m_dce_seeds (BITMAP_ALLOC (NULL)) { } | |
957 | ~tree_loop_interchange () { BITMAP_FREE (m_dce_seeds); } | |
958 | bool interchange (vec<data_reference_p>, vec<ddr_p>); | |
959 | ||
960 | private: | |
961 | void update_data_info (unsigned, unsigned, vec<data_reference_p>, vec<ddr_p>); | |
962 | bool valid_data_dependences (unsigned, unsigned, vec<ddr_p>); | |
963 | void interchange_loops (loop_cand &, loop_cand &); | |
964 | void map_inductions_to_loop (loop_cand &, loop_cand &); | |
965 | void move_code_to_inner_loop (struct loop *, struct loop *, basic_block *); | |
966 | ||
967 | /* The whole loop nest in which interchange is ongoing. */ | |
968 | vec<struct loop *> m_loop_nest; | |
969 | /* We create new IV which is only used in loop's exit condition check. | |
970 | In case of 3-level loop nest interchange, when we interchange the | |
971 | innermost two loops, new IV created in the middle level loop does | |
972 | not need to be preserved in interchanging the outermost two loops | |
973 | later. We record the IV so that it can be skipped. */ | |
974 | tree m_niters_iv_var; | |
975 | /* Bitmap of seed variables for dead code elimination after interchange. */ | |
976 | bitmap m_dce_seeds; | |
977 | }; | |
978 | ||
979 | /* Update data refs' access stride and dependence information after loop | |
980 | interchange. I_IDX/O_IDX gives indices of interchanged loops in loop | |
981 | nest. DATAREFS are data references. DDRS are data dependences. */ | |
982 | ||
983 | void | |
984 | tree_loop_interchange::update_data_info (unsigned i_idx, unsigned o_idx, | |
985 | vec<data_reference_p> datarefs, | |
986 | vec<ddr_p> ddrs) | |
987 | { | |
988 | struct data_reference *dr; | |
989 | struct data_dependence_relation *ddr; | |
990 | ||
991 | /* Update strides of data references. */ | |
992 | for (unsigned i = 0; datarefs.iterate (i, &dr); ++i) | |
993 | { | |
994 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); | |
995 | gcc_assert (stride->length () > i_idx); | |
996 | std::swap ((*stride)[i_idx], (*stride)[o_idx]); | |
997 | } | |
998 | /* Update data dependences. */ | |
999 | for (unsigned i = 0; ddrs.iterate (i, &ddr); ++i) | |
1000 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known) | |
1001 | { | |
1002 | for (unsigned j = 0; j < DDR_NUM_DIST_VECTS (ddr); ++j) | |
1003 | { | |
1004 | lambda_vector dist_vect = DDR_DIST_VECT (ddr, j); | |
1005 | std::swap (dist_vect[i_idx], dist_vect[o_idx]); | |
1006 | } | |
1007 | } | |
1008 | } | |
1009 | ||
1010 | /* Check data dependence relations, return TRUE if it's valid to interchange | |
1011 | two loops specified by I_IDX/O_IDX. Theoretically, interchanging the two | |
1012 | loops is valid only if dist vector, after interchanging, doesn't have '>' | |
1013 | as the leftmost non-'=' direction. Practically, this function have been | |
1014 | conservative here by not checking some valid cases. */ | |
1015 | ||
1016 | bool | |
1017 | tree_loop_interchange::valid_data_dependences (unsigned i_idx, unsigned o_idx, | |
1018 | vec<ddr_p> ddrs) | |
1019 | { | |
1020 | struct data_dependence_relation *ddr; | |
1021 | ||
1022 | for (unsigned i = 0; ddrs.iterate (i, &ddr); ++i) | |
1023 | { | |
1024 | /* Skip no-dependence case. */ | |
1025 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1026 | continue; | |
1027 | ||
1028 | for (unsigned j = 0; j < DDR_NUM_DIST_VECTS (ddr); ++j) | |
1029 | { | |
1030 | lambda_vector dist_vect = DDR_DIST_VECT (ddr, j); | |
1031 | unsigned level = dependence_level (dist_vect, m_loop_nest.length ()); | |
1032 | ||
1033 | /* If there is no carried dependence. */ | |
1034 | if (level == 0) | |
1035 | continue; | |
1036 | ||
1037 | level --; | |
1038 | ||
1039 | /* If dependence is not carried by any loop in between the two | |
1040 | loops [oloop, iloop] to interchange. */ | |
1041 | if (level < o_idx || level > i_idx) | |
1042 | continue; | |
1043 | ||
1044 | /* Be conservative, skip case if either direction at i_idx/o_idx | |
1045 | levels is not '=' or '<'. */ | |
1046 | if (dist_vect[i_idx] < 0 || dist_vect[o_idx] < 0) | |
1047 | return false; | |
1048 | } | |
1049 | } | |
1050 | ||
1051 | return true; | |
1052 | } | |
1053 | ||
1054 | /* Interchange two loops specified by ILOOP and OLOOP. */ | |
1055 | ||
1056 | void | |
1057 | tree_loop_interchange::interchange_loops (loop_cand &iloop, loop_cand &oloop) | |
1058 | { | |
1059 | reduction_p re; | |
1060 | gimple_stmt_iterator gsi; | |
1061 | tree i_niters, o_niters, var_after; | |
1062 | ||
1063 | /* Undo inner loop's simple reduction. */ | |
1064 | for (unsigned i = 0; iloop.m_reductions.iterate (i, &re); ++i) | |
1065 | if (re->type != DOUBLE_RTYPE) | |
1066 | { | |
1067 | if (re->producer) | |
1068 | reset_debug_uses (re->producer); | |
1069 | ||
1070 | iloop.undo_simple_reduction (re, m_dce_seeds); | |
1071 | } | |
1072 | ||
1073 | /* Only need to reset debug uses for double reduction. */ | |
1074 | for (unsigned i = 0; oloop.m_reductions.iterate (i, &re); ++i) | |
1075 | { | |
1076 | gcc_assert (re->type == DOUBLE_RTYPE); | |
1077 | reset_debug_uses (SSA_NAME_DEF_STMT (re->var)); | |
1078 | reset_debug_uses (SSA_NAME_DEF_STMT (re->next)); | |
1079 | } | |
1080 | ||
1081 | /* Prepare niters for both loops. */ | |
1082 | struct loop *loop_nest = m_loop_nest[0]; | |
1083 | edge instantiate_below = loop_preheader_edge (loop_nest); | |
1084 | gsi = gsi_last_bb (loop_preheader_edge (loop_nest)->src); | |
1085 | i_niters = number_of_latch_executions (iloop.m_loop); | |
1086 | i_niters = analyze_scalar_evolution (loop_outer (iloop.m_loop), i_niters); | |
1087 | i_niters = instantiate_scev (instantiate_below, loop_outer (iloop.m_loop), | |
1088 | i_niters); | |
1089 | i_niters = force_gimple_operand_gsi (&gsi, unshare_expr (i_niters), true, | |
1090 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
1091 | o_niters = number_of_latch_executions (oloop.m_loop); | |
1092 | if (oloop.m_loop != loop_nest) | |
1093 | { | |
1094 | o_niters = analyze_scalar_evolution (loop_outer (oloop.m_loop), o_niters); | |
1095 | o_niters = instantiate_scev (instantiate_below, loop_outer (oloop.m_loop), | |
1096 | o_niters); | |
1097 | } | |
1098 | o_niters = force_gimple_operand_gsi (&gsi, unshare_expr (o_niters), true, | |
1099 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
1100 | ||
1101 | /* Move src's code to tgt loop. This is necessary when src is the outer | |
1102 | loop and tgt is the inner loop. */ | |
1103 | move_code_to_inner_loop (oloop.m_loop, iloop.m_loop, oloop.m_bbs); | |
1104 | ||
1105 | /* Map outer loop's IV to inner loop, and vice versa. */ | |
1106 | map_inductions_to_loop (oloop, iloop); | |
1107 | map_inductions_to_loop (iloop, oloop); | |
1108 | ||
1109 | /* Create canonical IV for both loops. Note canonical IV for outer/inner | |
1110 | loop is actually from inner/outer loop. Also we record the new IV | |
1111 | created for the outer loop so that it can be skipped in later loop | |
1112 | interchange. */ | |
1113 | create_canonical_iv (oloop.m_loop, oloop.m_exit, | |
1114 | i_niters, &m_niters_iv_var, &var_after); | |
1115 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after)); | |
1116 | create_canonical_iv (iloop.m_loop, iloop.m_exit, | |
1117 | o_niters, NULL, &var_after); | |
1118 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after)); | |
1119 | ||
1120 | /* Scrap niters estimation of interchanged loops. */ | |
1121 | iloop.m_loop->any_upper_bound = false; | |
1122 | iloop.m_loop->any_likely_upper_bound = false; | |
1123 | free_numbers_of_iterations_estimates (iloop.m_loop); | |
1124 | oloop.m_loop->any_upper_bound = false; | |
1125 | oloop.m_loop->any_likely_upper_bound = false; | |
1126 | free_numbers_of_iterations_estimates (oloop.m_loop); | |
1127 | ||
4e090bcc BC |
1128 | /* Clear all cached scev information. This is expensive but shouldn't be |
1129 | a problem given we interchange in very limited times. */ | |
1130 | scev_reset_htab (); | |
1131 | ||
fbdec14e BC |
1132 | /* ??? The association between the loop data structure and the |
1133 | CFG changed, so what was loop N at the source level is now | |
1134 | loop M. We should think of retaining the association or breaking | |
1135 | it fully by creating a new loop instead of re-using the "wrong" one. */ | |
1136 | } | |
1137 | ||
1138 | /* Map induction variables of SRC loop to TGT loop. The function firstly | |
1139 | creates the same IV of SRC loop in TGT loop, then deletes the original | |
1140 | IV and re-initialize it using the newly created IV. For example, loop | |
1141 | nest: | |
1142 | ||
1143 | for (i = 0; i < N; i++) | |
1144 | for (j = 0; j < M; j++) | |
1145 | { | |
1146 | //use of i; | |
1147 | //use of j; | |
1148 | } | |
1149 | ||
1150 | will be transformed into: | |
1151 | ||
1152 | for (jj = 0; jj < M; jj++) | |
1153 | for (ii = 0; ii < N; ii++) | |
1154 | { | |
1155 | //use of ii; | |
1156 | //use of jj; | |
1157 | } | |
1158 | ||
1159 | after loop interchange. */ | |
1160 | ||
1161 | void | |
1162 | tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt) | |
1163 | { | |
1164 | induction_p iv; | |
1165 | edge e = tgt.m_exit; | |
1166 | gimple_stmt_iterator incr_pos = gsi_last_bb (e->src), gsi; | |
1167 | ||
1168 | /* Map source loop's IV to target loop. */ | |
1169 | for (unsigned i = 0; src.m_inductions.iterate (i, &iv); ++i) | |
1170 | { | |
1171 | gimple *use_stmt, *stmt = SSA_NAME_DEF_STMT (iv->var); | |
1172 | gcc_assert (is_a <gphi *> (stmt)); | |
1173 | ||
1174 | use_operand_p use_p; | |
1175 | /* Only map original IV to target loop. */ | |
1176 | if (m_niters_iv_var != iv->var) | |
1177 | { | |
1178 | /* Map the IV by creating the same one in target loop. */ | |
1179 | tree var_before, var_after; | |
1180 | tree base = unshare_expr (iv->init_expr); | |
1181 | tree step = unshare_expr (iv->step); | |
1182 | create_iv (base, step, SSA_NAME_VAR (iv->var), | |
1183 | tgt.m_loop, &incr_pos, false, &var_before, &var_after); | |
1184 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before)); | |
1185 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after)); | |
1186 | ||
1187 | /* Replace uses of the original IV var with newly created IV var. */ | |
1188 | imm_use_iterator imm_iter; | |
1189 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, iv->var) | |
1190 | { | |
1191 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
1192 | SET_USE (use_p, var_before); | |
1193 | ||
1194 | update_stmt (use_stmt); | |
1195 | } | |
1196 | } | |
1197 | ||
1198 | /* Mark all uses for DCE. */ | |
1199 | ssa_op_iter op_iter; | |
1200 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, op_iter, SSA_OP_USE) | |
1201 | { | |
1202 | tree use = USE_FROM_PTR (use_p); | |
1203 | if (TREE_CODE (use) == SSA_NAME | |
1204 | && ! SSA_NAME_IS_DEFAULT_DEF (use)) | |
1205 | bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (use)); | |
1206 | } | |
1207 | ||
1208 | /* Delete definition of the original IV in the source loop. */ | |
1209 | gsi = gsi_for_stmt (stmt); | |
1210 | remove_phi_node (&gsi, true); | |
1211 | } | |
1212 | } | |
1213 | ||
1214 | /* Move stmts of outer loop to inner loop. */ | |
1215 | ||
1216 | void | |
1217 | tree_loop_interchange::move_code_to_inner_loop (struct loop *outer, | |
1218 | struct loop *inner, | |
1219 | basic_block *outer_bbs) | |
1220 | { | |
1221 | basic_block oloop_exit_bb = single_exit (outer)->src; | |
1222 | gimple_stmt_iterator gsi, to; | |
1223 | ||
1224 | for (unsigned i = 0; i < outer->num_nodes; i++) | |
1225 | { | |
1226 | basic_block bb = outer_bbs[i]; | |
1227 | ||
1228 | /* Skip basic blocks of inner loop. */ | |
1229 | if (flow_bb_inside_loop_p (inner, bb)) | |
1230 | continue; | |
1231 | ||
1232 | /* Move code from header/latch to header/latch. */ | |
1233 | if (bb == outer->header) | |
1234 | to = gsi_after_labels (inner->header); | |
1235 | else if (bb == outer->latch) | |
1236 | to = gsi_after_labels (inner->latch); | |
1237 | else | |
1238 | /* Otherwise, simply move to exit->src. */ | |
1239 | to = gsi_last_bb (single_exit (inner)->src); | |
1240 | ||
1241 | for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);) | |
1242 | { | |
1243 | gimple *stmt = gsi_stmt (gsi); | |
1244 | ||
1245 | if (oloop_exit_bb == bb | |
1246 | && stmt == gsi_stmt (gsi_last_bb (oloop_exit_bb))) | |
1247 | { | |
1248 | gsi_next (&gsi); | |
1249 | continue; | |
1250 | } | |
1251 | ||
1252 | if (gimple_vuse (stmt)) | |
1253 | gimple_set_vuse (stmt, NULL_TREE); | |
1254 | if (gimple_vdef (stmt)) | |
1255 | { | |
1256 | unlink_stmt_vdef (stmt); | |
1257 | release_ssa_name (gimple_vdef (stmt)); | |
1258 | gimple_set_vdef (stmt, NULL_TREE); | |
1259 | } | |
1260 | ||
1261 | reset_debug_uses (stmt); | |
1262 | gsi_move_before (&gsi, &to); | |
1263 | } | |
1264 | } | |
1265 | } | |
1266 | ||
1267 | /* Given data reference DR in LOOP_NEST, the function computes DR's access | |
1268 | stride at each level of loop from innermost LOOP to outer. On success, | |
1269 | it saves access stride at each level loop in a vector which is pointed | |
1270 | by DR->aux. For example: | |
1271 | ||
1272 | int arr[100][100][100]; | |
1273 | for (i = 0; i < 100; i++) ;(DR->aux)strides[0] = 40000 | |
1274 | for (j = 100; j > 0; j--) ;(DR->aux)strides[1] = 400 | |
1275 | for (k = 0; k < 100; k++) ;(DR->aux)strides[2] = 4 | |
1276 | arr[i][j - 1][k] = 0; */ | |
1277 | ||
1278 | static void | |
1279 | compute_access_stride (struct loop *loop_nest, struct loop *loop, | |
1280 | data_reference_p dr) | |
1281 | { | |
1282 | vec<tree> *strides = new vec<tree> (); | |
1283 | basic_block bb = gimple_bb (DR_STMT (dr)); | |
1284 | ||
1285 | while (!flow_bb_inside_loop_p (loop, bb)) | |
1286 | { | |
1287 | strides->safe_push (build_int_cst (sizetype, 0)); | |
1288 | loop = loop_outer (loop); | |
1289 | } | |
1290 | gcc_assert (loop == bb->loop_father); | |
1291 | ||
1292 | tree ref = DR_REF (dr); | |
df0f6bbb JJ |
1293 | if (TREE_CODE (ref) == COMPONENT_REF |
1294 | && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))) | |
1295 | { | |
1296 | /* We can't take address of bitfields. If the bitfield is at constant | |
1297 | offset from the start of the struct, just use address of the | |
1298 | struct, for analysis of the strides that shouldn't matter. */ | |
1299 | if (!TREE_OPERAND (ref, 2) | |
1300 | || TREE_CODE (TREE_OPERAND (ref, 2)) == INTEGER_CST) | |
1301 | ref = TREE_OPERAND (ref, 0); | |
1302 | /* Otherwise, if we have a bit field representative, use that. */ | |
1303 | else if (DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (ref, 1)) | |
1304 | != NULL_TREE) | |
1305 | { | |
1306 | tree repr = DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (ref, 1)); | |
1307 | ref = build3 (COMPONENT_REF, TREE_TYPE (repr), TREE_OPERAND (ref, 0), | |
1308 | repr, TREE_OPERAND (ref, 2)); | |
1309 | } | |
1310 | /* Otherwise punt. */ | |
1311 | else | |
1312 | { | |
1313 | dr->aux = strides; | |
1314 | return; | |
1315 | } | |
1316 | } | |
fbdec14e BC |
1317 | tree scev_base = build_fold_addr_expr (ref); |
1318 | tree scev = analyze_scalar_evolution (loop, scev_base); | |
1319 | scev = instantiate_scev (loop_preheader_edge (loop_nest), loop, scev); | |
1320 | if (! chrec_contains_undetermined (scev)) | |
1321 | { | |
1322 | tree sl = scev; | |
1323 | struct loop *expected = loop; | |
1324 | while (TREE_CODE (sl) == POLYNOMIAL_CHREC) | |
1325 | { | |
1326 | struct loop *sl_loop = get_chrec_loop (sl); | |
1327 | while (sl_loop != expected) | |
1328 | { | |
1329 | strides->safe_push (size_int (0)); | |
1330 | expected = loop_outer (expected); | |
1331 | } | |
1332 | strides->safe_push (CHREC_RIGHT (sl)); | |
1333 | sl = CHREC_LEFT (sl); | |
1334 | expected = loop_outer (expected); | |
1335 | } | |
1336 | if (! tree_contains_chrecs (sl, NULL)) | |
1337 | while (expected != loop_outer (loop_nest)) | |
1338 | { | |
1339 | strides->safe_push (size_int (0)); | |
1340 | expected = loop_outer (expected); | |
1341 | } | |
1342 | } | |
1343 | ||
1344 | dr->aux = strides; | |
1345 | } | |
1346 | ||
1347 | /* Given loop nest LOOP_NEST with innermost LOOP, the function computes | |
1348 | access strides with respect to each level loop for all data refs in | |
1349 | DATAREFS from inner loop to outer loop. On success, it returns the | |
1350 | outermost loop that access strides can be computed successfully for | |
1351 | all data references. If access strides cannot be computed at least | |
1352 | for two levels of loop for any data reference, it returns NULL. */ | |
1353 | ||
1354 | static struct loop * | |
1355 | compute_access_strides (struct loop *loop_nest, struct loop *loop, | |
1356 | vec<data_reference_p> datarefs) | |
1357 | { | |
1358 | unsigned i, j, num_loops = (unsigned) -1; | |
1359 | data_reference_p dr; | |
1360 | vec<tree> *stride; | |
1361 | ||
1362 | for (i = 0; datarefs.iterate (i, &dr); ++i) | |
1363 | { | |
1364 | compute_access_stride (loop_nest, loop, dr); | |
1365 | stride = DR_ACCESS_STRIDE (dr); | |
1366 | if (stride->length () < num_loops) | |
1367 | { | |
1368 | num_loops = stride->length (); | |
1369 | if (num_loops < 2) | |
1370 | return NULL; | |
1371 | } | |
1372 | } | |
1373 | ||
1374 | for (i = 0; datarefs.iterate (i, &dr); ++i) | |
1375 | { | |
1376 | stride = DR_ACCESS_STRIDE (dr); | |
1377 | if (stride->length () > num_loops) | |
1378 | stride->truncate (num_loops); | |
1379 | ||
1380 | for (j = 0; j < (num_loops >> 1); ++j) | |
1381 | std::swap ((*stride)[j], (*stride)[num_loops - j - 1]); | |
1382 | } | |
1383 | ||
1384 | loop = superloop_at_depth (loop, loop_depth (loop) + 1 - num_loops); | |
1385 | gcc_assert (loop_nest == loop || flow_loop_nested_p (loop_nest, loop)); | |
1386 | return loop; | |
1387 | } | |
1388 | ||
1389 | /* Prune access strides for data references in DATAREFS by removing strides | |
1390 | of loops that isn't in current LOOP_NEST. */ | |
1391 | ||
1392 | static void | |
1393 | prune_access_strides_not_in_loop (struct loop *loop_nest, | |
1394 | struct loop *innermost, | |
1395 | vec<data_reference_p> datarefs) | |
1396 | { | |
1397 | data_reference_p dr; | |
1398 | unsigned num_loops = loop_depth (innermost) - loop_depth (loop_nest) + 1; | |
1399 | gcc_assert (num_loops > 1); | |
1400 | ||
1401 | /* Block remove strides of loops that is not in current loop nest. */ | |
1402 | for (unsigned i = 0; datarefs.iterate (i, &dr); ++i) | |
1403 | { | |
1404 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); | |
1405 | if (stride->length () > num_loops) | |
1406 | stride->block_remove (0, stride->length () - num_loops); | |
1407 | } | |
1408 | } | |
1409 | ||
1410 | /* Dump access strides for all DATAREFS. */ | |
1411 | ||
1412 | static void | |
1413 | dump_access_strides (vec<data_reference_p> datarefs) | |
1414 | { | |
1415 | data_reference_p dr; | |
1416 | fprintf (dump_file, "Access Strides for DRs:\n"); | |
1417 | for (unsigned i = 0; datarefs.iterate (i, &dr); ++i) | |
1418 | { | |
1419 | fprintf (dump_file, " "); | |
1420 | print_generic_expr (dump_file, DR_REF (dr), TDF_SLIM); | |
1421 | fprintf (dump_file, ":\t\t<"); | |
1422 | ||
1423 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); | |
1424 | unsigned num_loops = stride->length (); | |
1425 | for (unsigned j = 0; j < num_loops; ++j) | |
1426 | { | |
1427 | print_generic_expr (dump_file, (*stride)[j], TDF_SLIM); | |
1428 | fprintf (dump_file, "%s", (j < num_loops - 1) ? ",\t" : ">\n"); | |
1429 | } | |
1430 | } | |
1431 | } | |
1432 | ||
1433 | /* Return true if it's profitable to interchange two loops whose index | |
1434 | in whole loop nest vector are I_IDX/O_IDX respectively. The function | |
1435 | computes and compares three types information from all DATAREFS: | |
1436 | 1) Access stride for loop I_IDX and O_IDX. | |
1437 | 2) Number of invariant memory references with respect to I_IDX before | |
1438 | and after loop interchange. | |
1439 | 3) Flags indicating if all memory references access sequential memory | |
1440 | in ILOOP, before and after loop interchange. | |
1441 | If INNMOST_LOOP_P is true, the two loops for interchanging are the two | |
1442 | innermost loops in loop nest. This function also dumps information if | |
1443 | DUMP_INFO_P is true. */ | |
1444 | ||
1445 | static bool | |
1446 | should_interchange_loops (unsigned i_idx, unsigned o_idx, | |
1447 | vec<data_reference_p> datarefs, | |
1eeeda47 | 1448 | unsigned i_stmt_cost, unsigned o_stmt_cost, |
fbdec14e BC |
1449 | bool innermost_loops_p, bool dump_info_p = true) |
1450 | { | |
1451 | unsigned HOST_WIDE_INT ratio; | |
1452 | unsigned i, j, num_old_inv_drs = 0, num_new_inv_drs = 0; | |
1453 | struct data_reference *dr; | |
1454 | bool all_seq_dr_before_p = true, all_seq_dr_after_p = true; | |
1455 | widest_int iloop_strides = 0, oloop_strides = 0; | |
1456 | unsigned num_unresolved_drs = 0; | |
1457 | unsigned num_resolved_ok_drs = 0; | |
1458 | unsigned num_resolved_not_ok_drs = 0; | |
1459 | ||
1460 | if (dump_info_p && dump_file && (dump_flags & TDF_DETAILS)) | |
1461 | fprintf (dump_file, "\nData ref strides:\n\tmem_ref:\t\tiloop\toloop\n"); | |
1462 | ||
1463 | for (i = 0; datarefs.iterate (i, &dr); ++i) | |
1464 | { | |
1465 | vec<tree> *stride = DR_ACCESS_STRIDE (dr); | |
1466 | tree iloop_stride = (*stride)[i_idx], oloop_stride = (*stride)[o_idx]; | |
1467 | ||
1468 | bool subloop_stride_p = false; | |
1469 | /* Data ref can't be invariant or sequential access at current loop if | |
1470 | its address changes with respect to any subloops. */ | |
1471 | for (j = i_idx + 1; j < stride->length (); ++j) | |
1472 | if (!integer_zerop ((*stride)[j])) | |
1473 | { | |
1474 | subloop_stride_p = true; | |
1475 | break; | |
1476 | } | |
1477 | ||
1478 | if (integer_zerop (iloop_stride)) | |
1479 | { | |
1480 | if (!subloop_stride_p) | |
1481 | num_old_inv_drs++; | |
1482 | } | |
1483 | if (integer_zerop (oloop_stride)) | |
1484 | { | |
1485 | if (!subloop_stride_p) | |
1486 | num_new_inv_drs++; | |
1487 | } | |
1488 | ||
1489 | if (TREE_CODE (iloop_stride) == INTEGER_CST | |
1490 | && TREE_CODE (oloop_stride) == INTEGER_CST) | |
1491 | { | |
1492 | iloop_strides = wi::add (iloop_strides, wi::to_widest (iloop_stride)); | |
1493 | oloop_strides = wi::add (oloop_strides, wi::to_widest (oloop_stride)); | |
1494 | } | |
1495 | else if (multiple_of_p (TREE_TYPE (iloop_stride), | |
1496 | iloop_stride, oloop_stride)) | |
1497 | num_resolved_ok_drs++; | |
1498 | else if (multiple_of_p (TREE_TYPE (iloop_stride), | |
1499 | oloop_stride, iloop_stride)) | |
1500 | num_resolved_not_ok_drs++; | |
1501 | else | |
1502 | num_unresolved_drs++; | |
1503 | ||
1504 | /* Data ref can't be sequential access if its address changes in sub | |
1505 | loop. */ | |
1506 | if (subloop_stride_p) | |
1507 | { | |
1508 | all_seq_dr_before_p = false; | |
1509 | all_seq_dr_after_p = false; | |
1510 | continue; | |
1511 | } | |
1512 | /* Track if all data references are sequential accesses before/after loop | |
1513 | interchange. Note invariant is considered sequential here. */ | |
1514 | tree access_size = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); | |
1515 | if (all_seq_dr_before_p | |
1516 | && ! (integer_zerop (iloop_stride) | |
1517 | || operand_equal_p (access_size, iloop_stride, 0))) | |
1518 | all_seq_dr_before_p = false; | |
1519 | if (all_seq_dr_after_p | |
1520 | && ! (integer_zerop (oloop_stride) | |
1521 | || operand_equal_p (access_size, oloop_stride, 0))) | |
1522 | all_seq_dr_after_p = false; | |
1523 | } | |
1524 | ||
1525 | if (dump_info_p && dump_file && (dump_flags & TDF_DETAILS)) | |
1526 | { | |
1527 | fprintf (dump_file, "\toverall:\t\t"); | |
1528 | print_decu (iloop_strides, dump_file); | |
1529 | fprintf (dump_file, "\t"); | |
1530 | print_decu (oloop_strides, dump_file); | |
1531 | fprintf (dump_file, "\n"); | |
1532 | ||
1533 | fprintf (dump_file, "Invariant data ref: before(%d), after(%d)\n", | |
1534 | num_old_inv_drs, num_new_inv_drs); | |
1535 | fprintf (dump_file, "All consecutive stride: before(%s), after(%s)\n", | |
1536 | all_seq_dr_before_p ? "true" : "false", | |
1537 | all_seq_dr_after_p ? "true" : "false"); | |
1538 | fprintf (dump_file, "OK to interchage with variable strides: %d\n", | |
1539 | num_resolved_ok_drs); | |
1540 | fprintf (dump_file, "Not OK to interchage with variable strides: %d\n", | |
1541 | num_resolved_not_ok_drs); | |
1542 | fprintf (dump_file, "Variable strides we cannot decide: %d\n", | |
1543 | num_unresolved_drs); | |
1eeeda47 BC |
1544 | fprintf (dump_file, "Stmt cost of inner loop: %d\n", i_stmt_cost); |
1545 | fprintf (dump_file, "Stmt cost of outer loop: %d\n", o_stmt_cost); | |
fbdec14e BC |
1546 | } |
1547 | ||
1548 | if (num_unresolved_drs != 0 || num_resolved_not_ok_drs != 0) | |
1549 | return false; | |
1550 | ||
1eeeda47 BC |
1551 | /* Stmts of outer loop will be moved to inner loop. If there are two many |
1552 | such stmts, it could make inner loop costly. Here we compare stmt cost | |
1553 | between outer and inner loops. */ | |
1554 | if (i_stmt_cost && o_stmt_cost | |
1555 | && num_old_inv_drs + o_stmt_cost > num_new_inv_drs | |
1556 | && o_stmt_cost * STMT_COST_RATIO > i_stmt_cost) | |
1557 | return false; | |
1558 | ||
fbdec14e BC |
1559 | /* We use different stride comparison ratio for interchanging innermost |
1560 | two loops or not. The idea is to be conservative in interchange for | |
1561 | the innermost loops. */ | |
1562 | ratio = innermost_loops_p ? INNER_STRIDE_RATIO : OUTER_STRIDE_RATIO; | |
1563 | /* Do interchange if it gives better data locality behavior. */ | |
1564 | if (wi::gtu_p (iloop_strides, wi::mul (oloop_strides, ratio))) | |
1565 | return true; | |
1566 | if (wi::gtu_p (iloop_strides, oloop_strides)) | |
1567 | { | |
1568 | /* Or it creates more invariant memory references. */ | |
1569 | if ((!all_seq_dr_before_p || all_seq_dr_after_p) | |
1570 | && num_new_inv_drs > num_old_inv_drs) | |
1571 | return true; | |
1572 | /* Or it makes all memory references sequential. */ | |
1573 | if (num_new_inv_drs >= num_old_inv_drs | |
1574 | && !all_seq_dr_before_p && all_seq_dr_after_p) | |
1575 | return true; | |
1576 | } | |
1577 | ||
1578 | return false; | |
1579 | } | |
1580 | ||
1581 | /* Try to interchange inner loop of a loop nest to outer level. */ | |
1582 | ||
1583 | bool | |
1584 | tree_loop_interchange::interchange (vec<data_reference_p> datarefs, | |
1585 | vec<ddr_p> ddrs) | |
1586 | { | |
4f5b9c80 | 1587 | dump_user_location_t loc = find_loop_location (m_loop_nest[0]); |
fbdec14e BC |
1588 | bool changed_p = false; |
1589 | /* In each iteration we try to interchange I-th loop with (I+1)-th loop. | |
1590 | The overall effect is to push inner loop to outermost level in whole | |
1591 | loop nest. */ | |
1592 | for (unsigned i = m_loop_nest.length (); i > 1; --i) | |
1593 | { | |
1594 | unsigned i_idx = i - 1, o_idx = i - 2; | |
1595 | ||
1596 | /* Check validity for loop interchange. */ | |
1597 | if (!valid_data_dependences (i_idx, o_idx, ddrs)) | |
1598 | break; | |
1599 | ||
1600 | loop_cand iloop (m_loop_nest[i_idx], m_loop_nest[o_idx]); | |
1601 | loop_cand oloop (m_loop_nest[o_idx], m_loop_nest[o_idx]); | |
1602 | ||
1603 | /* Check if we can do transformation for loop interchange. */ | |
1604 | if (!iloop.analyze_carried_vars (NULL) | |
1605 | || !iloop.analyze_lcssa_phis () | |
1606 | || !oloop.analyze_carried_vars (&iloop) | |
1607 | || !oloop.analyze_lcssa_phis () | |
1608 | || !iloop.can_interchange_p (NULL) | |
1609 | || !oloop.can_interchange_p (&iloop)) | |
1610 | break; | |
1611 | ||
1eeeda47 BC |
1612 | /* Outer loop's stmts will be moved to inner loop during interchange. |
1613 | If there are many of them, it may make inner loop very costly. We | |
1614 | need to check number of outer loop's stmts in profit cost model of | |
1615 | interchange. */ | |
1616 | int stmt_cost = oloop.m_num_stmts; | |
1617 | /* Count out the exit checking stmt of outer loop. */ | |
1618 | stmt_cost --; | |
1619 | /* Count out IV's increasing stmt, IVOPTs takes care if it. */ | |
1620 | stmt_cost -= oloop.m_inductions.length (); | |
1621 | /* Count in the additional load and cond_expr stmts caused by inner | |
1622 | loop's constant initialized reduction. */ | |
1623 | stmt_cost += iloop.m_const_init_reduc * 2; | |
1624 | if (stmt_cost < 0) | |
1625 | stmt_cost = 0; | |
1626 | ||
fbdec14e BC |
1627 | /* Check profitability for loop interchange. */ |
1628 | if (should_interchange_loops (i_idx, o_idx, datarefs, | |
1eeeda47 BC |
1629 | (unsigned) iloop.m_num_stmts, |
1630 | (unsigned) stmt_cost, | |
fbdec14e BC |
1631 | iloop.m_loop->inner == NULL)) |
1632 | { | |
1633 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1634 | fprintf (dump_file, | |
1635 | "Loop_pair<outer:%d, inner:%d> is interchanged\n\n", | |
1636 | oloop.m_loop->num, iloop.m_loop->num); | |
1637 | ||
1638 | changed_p = true; | |
1639 | interchange_loops (iloop, oloop); | |
1640 | /* No need to update if there is no further loop interchange. */ | |
1641 | if (o_idx > 0) | |
1642 | update_data_info (i_idx, o_idx, datarefs, ddrs); | |
1643 | } | |
1644 | else | |
1645 | { | |
1646 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1647 | fprintf (dump_file, | |
1648 | "Loop_pair<outer:%d, inner:%d> is not interchanged\n\n", | |
1649 | oloop.m_loop->num, iloop.m_loop->num); | |
1650 | } | |
1651 | } | |
fbdec14e | 1652 | simple_dce_from_worklist (m_dce_seeds); |
da472c1b | 1653 | |
bbeeac91 | 1654 | if (changed_p && dump_enabled_p ()) |
da472c1b RB |
1655 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, |
1656 | "loops interchanged in loop nest\n"); | |
1657 | ||
fbdec14e BC |
1658 | return changed_p; |
1659 | } | |
1660 | ||
1661 | ||
1662 | /* Loop interchange pass. */ | |
1663 | ||
1664 | namespace { | |
1665 | ||
1666 | const pass_data pass_data_linterchange = | |
1667 | { | |
1668 | GIMPLE_PASS, /* type */ | |
1669 | "linterchange", /* name */ | |
1670 | OPTGROUP_LOOP, /* optinfo_flags */ | |
1671 | TV_LINTERCHANGE, /* tv_id */ | |
1672 | PROP_cfg, /* properties_required */ | |
1673 | 0, /* properties_provided */ | |
1674 | 0, /* properties_destroyed */ | |
1675 | 0, /* todo_flags_start */ | |
1676 | 0, /* todo_flags_finish */ | |
1677 | }; | |
1678 | ||
1679 | class pass_linterchange : public gimple_opt_pass | |
1680 | { | |
1681 | public: | |
1682 | pass_linterchange (gcc::context *ctxt) | |
1683 | : gimple_opt_pass (pass_data_linterchange, ctxt) | |
1684 | {} | |
1685 | ||
1686 | /* opt_pass methods: */ | |
1687 | opt_pass * clone () { return new pass_linterchange (m_ctxt); } | |
1688 | virtual bool gate (function *) { return flag_loop_interchange; } | |
1689 | virtual unsigned int execute (function *); | |
1690 | ||
1691 | }; // class pass_linterchange | |
1692 | ||
1693 | ||
1694 | /* Return true if LOOP has proper form for interchange. We check three | |
1695 | conditions in the function: | |
1696 | 1) In general, a loop can be interchanged only if it doesn't have | |
1697 | basic blocks other than header, exit and latch besides possible | |
1698 | inner loop nest. This basically restricts loop interchange to | |
1699 | below form loop nests: | |
1700 | ||
1701 | header<---+ | |
1702 | | | | |
1703 | v | | |
1704 | INNER_LOOP | | |
1705 | | | | |
1706 | v | | |
1707 | exit--->latch | |
1708 | ||
1709 | 2) Data reference in basic block that executes in different times | |
1710 | than loop head/exit is not allowed. | |
1711 | 3) Record the innermost outer loop that doesn't form rectangle loop | |
1712 | nest with LOOP. */ | |
1713 | ||
1714 | static bool | |
1715 | proper_loop_form_for_interchange (struct loop *loop, struct loop **min_outer) | |
1716 | { | |
1717 | edge e0, e1, exit; | |
1718 | ||
1719 | /* Don't interchange if loop has unsupported information for the moment. */ | |
1720 | if (loop->safelen > 0 | |
1721 | || loop->constraints != 0 | |
1722 | || loop->can_be_parallel | |
1723 | || loop->dont_vectorize | |
1724 | || loop->force_vectorize | |
1725 | || loop->in_oacc_kernels_region | |
1726 | || loop->orig_loop_num != 0 | |
1727 | || loop->simduid != NULL_TREE) | |
1728 | return false; | |
1729 | ||
1730 | /* Don't interchange if outer loop has basic block other than header, exit | |
1731 | and latch. */ | |
1732 | if (loop->inner != NULL | |
1733 | && loop->num_nodes != loop->inner->num_nodes + 3) | |
1734 | return false; | |
1735 | ||
1736 | if ((exit = single_dom_exit (loop)) == NULL) | |
1737 | return false; | |
1738 | ||
1739 | /* Check control flow on loop header/exit blocks. */ | |
1740 | if (loop->header == exit->src | |
1741 | && (EDGE_COUNT (loop->header->preds) != 2 | |
1742 | || EDGE_COUNT (loop->header->succs) != 2)) | |
1743 | return false; | |
1744 | else if (loop->header != exit->src | |
1745 | && (EDGE_COUNT (loop->header->preds) != 2 | |
1746 | || !single_succ_p (loop->header) | |
1747 | || unsupported_edge (single_succ_edge (loop->header)) | |
1748 | || EDGE_COUNT (exit->src->succs) != 2 | |
1749 | || !single_pred_p (exit->src) | |
1750 | || unsupported_edge (single_pred_edge (exit->src)))) | |
1751 | return false; | |
1752 | ||
1753 | e0 = EDGE_PRED (loop->header, 0); | |
1754 | e1 = EDGE_PRED (loop->header, 1); | |
1755 | if (unsupported_edge (e0) || unsupported_edge (e1) | |
1756 | || (e0->src != loop->latch && e1->src != loop->latch) | |
1757 | || (e0->src->loop_father == loop && e1->src->loop_father == loop)) | |
1758 | return false; | |
1759 | ||
1760 | e0 = EDGE_SUCC (exit->src, 0); | |
1761 | e1 = EDGE_SUCC (exit->src, 1); | |
1762 | if (unsupported_edge (e0) || unsupported_edge (e1) | |
1763 | || (e0->dest != loop->latch && e1->dest != loop->latch) | |
1764 | || (e0->dest->loop_father == loop && e1->dest->loop_father == loop)) | |
1765 | return false; | |
1766 | ||
1767 | /* Don't interchange if any reference is in basic block that doesn't | |
1768 | dominate exit block. */ | |
1769 | basic_block *bbs = get_loop_body (loop); | |
1770 | for (unsigned i = 0; i < loop->num_nodes; i++) | |
1771 | { | |
1772 | basic_block bb = bbs[i]; | |
1773 | ||
1774 | if (bb->loop_father != loop | |
1775 | || bb == loop->header || bb == exit->src | |
1776 | || dominated_by_p (CDI_DOMINATORS, exit->src, bb)) | |
1777 | continue; | |
1778 | ||
65f4b875 | 1779 | for (gimple_stmt_iterator gsi = gsi_start_nondebug_bb (bb); |
fbdec14e BC |
1780 | !gsi_end_p (gsi); gsi_next_nondebug (&gsi)) |
1781 | if (gimple_vuse (gsi_stmt (gsi))) | |
1782 | { | |
1783 | free (bbs); | |
1784 | return false; | |
1785 | } | |
1786 | } | |
1787 | free (bbs); | |
1788 | ||
1789 | tree niters = number_of_latch_executions (loop); | |
1790 | niters = analyze_scalar_evolution (loop_outer (loop), niters); | |
1791 | if (!niters || chrec_contains_undetermined (niters)) | |
1792 | return false; | |
1793 | ||
1794 | /* Record the innermost outer loop that doesn't form rectangle loop nest. */ | |
1795 | for (loop_p loop2 = loop_outer (loop); | |
1796 | loop2 && flow_loop_nested_p (*min_outer, loop2); | |
1797 | loop2 = loop_outer (loop2)) | |
1798 | { | |
1799 | niters = instantiate_scev (loop_preheader_edge (loop2), | |
1800 | loop_outer (loop), niters); | |
1801 | if (!evolution_function_is_invariant_p (niters, loop2->num)) | |
1802 | { | |
1803 | *min_outer = loop2; | |
1804 | break; | |
1805 | } | |
1806 | } | |
1807 | return true; | |
1808 | } | |
1809 | ||
1810 | /* Return true if any two adjacent loops in loop nest [INNERMOST, LOOP_NEST] | |
1811 | should be interchanged by looking into all DATAREFS. */ | |
1812 | ||
1813 | static bool | |
1814 | should_interchange_loop_nest (struct loop *loop_nest, struct loop *innermost, | |
1815 | vec<data_reference_p> datarefs) | |
1816 | { | |
1817 | unsigned idx = loop_depth (innermost) - loop_depth (loop_nest); | |
1818 | gcc_assert (idx > 0); | |
1819 | ||
1820 | /* Check if any two adjacent loops should be interchanged. */ | |
1821 | for (struct loop *loop = innermost; | |
1822 | loop != loop_nest; loop = loop_outer (loop), idx--) | |
1eeeda47 | 1823 | if (should_interchange_loops (idx, idx - 1, datarefs, 0, 0, |
fbdec14e BC |
1824 | loop == innermost, false)) |
1825 | return true; | |
1826 | ||
1827 | return false; | |
1828 | } | |
1829 | ||
1830 | /* Given loop nest LOOP_NEST and data references DATAREFS, compute data | |
1831 | dependences for loop interchange and store it in DDRS. Note we compute | |
1832 | dependences directly rather than call generic interface so that we can | |
1833 | return on unknown dependence instantly. */ | |
1834 | ||
1835 | static bool | |
1836 | tree_loop_interchange_compute_ddrs (vec<loop_p> loop_nest, | |
1837 | vec<data_reference_p> datarefs, | |
1838 | vec<ddr_p> *ddrs) | |
1839 | { | |
1840 | struct data_reference *a, *b; | |
1841 | struct loop *innermost = loop_nest.last (); | |
1842 | ||
1843 | for (unsigned i = 0; datarefs.iterate (i, &a); ++i) | |
1844 | { | |
1845 | bool a_outer_p = gimple_bb (DR_STMT (a))->loop_father != innermost; | |
1846 | for (unsigned j = i + 1; datarefs.iterate (j, &b); ++j) | |
1847 | if (DR_IS_WRITE (a) || DR_IS_WRITE (b)) | |
1848 | { | |
1849 | bool b_outer_p = gimple_bb (DR_STMT (b))->loop_father != innermost; | |
1850 | /* Don't support multiple write references in outer loop. */ | |
1851 | if (a_outer_p && b_outer_p && DR_IS_WRITE (a) && DR_IS_WRITE (b)) | |
1852 | return false; | |
1853 | ||
1854 | ddr_p ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
1855 | ddrs->safe_push (ddr); | |
1856 | compute_affine_dependence (ddr, loop_nest[0]); | |
1857 | ||
1858 | /* Give up if ddr is unknown dependence or classic direct vector | |
1859 | is not available. */ | |
1860 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1861 | || (DDR_ARE_DEPENDENT (ddr) == NULL_TREE | |
1862 | && DDR_NUM_DIR_VECTS (ddr) == 0)) | |
1863 | return false; | |
1864 | ||
1865 | /* If either data references is in outer loop of nest, we require | |
1866 | no dependence here because the data reference need to be moved | |
1867 | into inner loop during interchange. */ | |
1868 | if (a_outer_p && b_outer_p | |
1869 | && operand_equal_p (DR_REF (a), DR_REF (b), 0)) | |
1870 | continue; | |
1871 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known | |
1872 | && (a_outer_p || b_outer_p)) | |
1873 | return false; | |
1874 | } | |
1875 | } | |
1876 | ||
1877 | return true; | |
1878 | } | |
1879 | ||
1880 | /* Prune DATAREFS by removing any data reference not inside of LOOP. */ | |
1881 | ||
1882 | static inline void | |
1883 | prune_datarefs_not_in_loop (struct loop *loop, vec<data_reference_p> datarefs) | |
1884 | { | |
1885 | unsigned i, j; | |
1886 | struct data_reference *dr; | |
1887 | ||
1888 | for (i = 0, j = 0; datarefs.iterate (i, &dr); ++i) | |
1889 | { | |
1890 | if (flow_bb_inside_loop_p (loop, gimple_bb (DR_STMT (dr)))) | |
1891 | datarefs[j++] = dr; | |
1892 | else | |
1893 | { | |
1894 | if (dr->aux) | |
1895 | { | |
1896 | DR_ACCESS_STRIDE (dr)->release (); | |
46bb9d29 | 1897 | delete (vec<tree> *) dr->aux; |
fbdec14e BC |
1898 | } |
1899 | free_data_ref (dr); | |
1900 | } | |
1901 | } | |
1902 | datarefs.truncate (j); | |
1903 | } | |
1904 | ||
1905 | /* Find and store data references in DATAREFS for LOOP nest. If there's | |
1906 | difficult data reference in a basic block, we shrink the loop nest to | |
1907 | inner loop of that basic block's father loop. On success, return the | |
1908 | outer loop of the result loop nest. */ | |
1909 | ||
1910 | static struct loop * | |
1911 | prepare_data_references (struct loop *loop, vec<data_reference_p> *datarefs) | |
1912 | { | |
1913 | struct loop *loop_nest = loop; | |
1914 | vec<data_reference_p> *bb_refs; | |
1915 | basic_block bb, *bbs = get_loop_body_in_dom_order (loop); | |
1916 | ||
1917 | for (unsigned i = 0; i < loop->num_nodes; i++) | |
1918 | bbs[i]->aux = NULL; | |
1919 | ||
1920 | /* Find data references for all basic blocks. Shrink loop nest on difficult | |
1921 | data reference. */ | |
1922 | for (unsigned i = 0; loop_nest && i < loop->num_nodes; ++i) | |
1923 | { | |
1924 | bb = bbs[i]; | |
1925 | if (!flow_bb_inside_loop_p (loop_nest, bb)) | |
1926 | continue; | |
1927 | ||
1928 | bb_refs = new vec<data_reference_p> (); | |
1929 | if (find_data_references_in_bb (loop, bb, bb_refs) == chrec_dont_know) | |
1930 | { | |
1931 | loop_nest = bb->loop_father->inner; | |
1932 | if (loop_nest && !loop_nest->inner) | |
1933 | loop_nest = NULL; | |
1934 | ||
1935 | free_data_refs (*bb_refs); | |
1936 | delete bb_refs; | |
1937 | } | |
1938 | else if (bb_refs->is_empty ()) | |
1939 | delete bb_refs; | |
1940 | else | |
1941 | bb->aux = bb_refs; | |
1942 | } | |
1943 | ||
1944 | /* Collect all data references in loop nest. */ | |
1945 | for (unsigned i = 0; i < loop->num_nodes; i++) | |
1946 | { | |
1947 | bb = bbs[i]; | |
1948 | if (!bb->aux) | |
1949 | continue; | |
1950 | ||
1951 | bb_refs = (vec<data_reference_p> *) bb->aux; | |
1952 | if (loop_nest && flow_bb_inside_loop_p (loop_nest, bb)) | |
1953 | datarefs->safe_splice (*bb_refs); | |
1954 | else | |
1955 | free_data_refs (*bb_refs); | |
1956 | ||
1957 | delete bb_refs; | |
1958 | bb->aux = NULL; | |
1959 | } | |
1960 | free (bbs); | |
1961 | ||
1962 | return loop_nest; | |
1963 | } | |
1964 | ||
1965 | /* Given innermost LOOP, return true if perfect loop nest can be found and | |
1966 | data dependences can be computed. If succeed, record the perfect loop | |
1967 | nest in LOOP_NEST; record all data references in DATAREFS and record all | |
1968 | data dependence relations in DDRS. | |
1969 | ||
1970 | We do support a restricted form of imperfect loop nest, i.e, loop nest | |
1971 | with load/store in outer loop initializing/finalizing simple reduction | |
1972 | of the innermost loop. For such outer loop reference, we require that | |
1973 | it has no dependence with others sinve it will be moved to inner loop | |
1974 | in interchange. */ | |
1975 | ||
1976 | static bool | |
1977 | prepare_perfect_loop_nest (struct loop *loop, vec<loop_p> *loop_nest, | |
1978 | vec<data_reference_p> *datarefs, vec<ddr_p> *ddrs) | |
1979 | { | |
1980 | struct loop *start_loop = NULL, *innermost = loop; | |
1981 | struct loop *outermost = loops_for_fn (cfun)->tree_root; | |
1982 | ||
1983 | /* Find loop nest from the innermost loop. The outermost is the innermost | |
1984 | outer*/ | |
1985 | while (loop->num != 0 && loop->inner == start_loop | |
1986 | && flow_loop_nested_p (outermost, loop)) | |
1987 | { | |
1988 | if (!proper_loop_form_for_interchange (loop, &outermost)) | |
1989 | break; | |
1990 | ||
1991 | start_loop = loop; | |
1992 | /* If this loop has sibling loop, the father loop won't be in perfect | |
1993 | loop nest. */ | |
1994 | if (loop->next != NULL) | |
1995 | break; | |
1996 | ||
1997 | loop = loop_outer (loop); | |
1998 | } | |
1999 | if (!start_loop || !start_loop->inner) | |
2000 | return false; | |
2001 | ||
2002 | /* Prepare the data reference vector for the loop nest, pruning outer | |
2003 | loops we cannot handle. */ | |
2004 | start_loop = prepare_data_references (start_loop, datarefs); | |
2005 | if (!start_loop | |
2006 | /* Check if there is no data reference. */ | |
2007 | || datarefs->is_empty () | |
2008 | /* Check if there are too many of data references. */ | |
2009 | || (int) datarefs->length () > MAX_DATAREFS) | |
2010 | return false; | |
2011 | ||
2012 | /* Compute access strides for all data references, pruning outer | |
2013 | loops we cannot analyze refs in. */ | |
2014 | start_loop = compute_access_strides (start_loop, innermost, *datarefs); | |
2015 | if (!start_loop) | |
2016 | return false; | |
2017 | ||
2018 | /* Check if any interchange is profitable in the loop nest. */ | |
2019 | if (!should_interchange_loop_nest (start_loop, innermost, *datarefs)) | |
2020 | return false; | |
2021 | ||
2022 | /* Check if data dependences can be computed for loop nest starting from | |
2023 | start_loop. */ | |
2024 | loop = start_loop; | |
2025 | do { | |
2026 | loop_nest->truncate (0); | |
2027 | ||
2028 | if (loop != start_loop) | |
2029 | prune_datarefs_not_in_loop (start_loop, *datarefs); | |
2030 | ||
2031 | if (find_loop_nest (start_loop, loop_nest) | |
2032 | && tree_loop_interchange_compute_ddrs (*loop_nest, *datarefs, ddrs)) | |
2033 | { | |
2034 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2035 | fprintf (dump_file, | |
2036 | "\nConsider loop interchange for loop_nest<%d - %d>\n", | |
2037 | start_loop->num, innermost->num); | |
2038 | ||
2039 | if (loop != start_loop) | |
2040 | prune_access_strides_not_in_loop (start_loop, innermost, *datarefs); | |
2041 | ||
2042 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2043 | dump_access_strides (*datarefs); | |
2044 | ||
2045 | return true; | |
2046 | } | |
2047 | ||
2048 | free_dependence_relations (*ddrs); | |
2049 | *ddrs = vNULL; | |
2050 | /* Try to compute data dependences with the outermost loop stripped. */ | |
2051 | loop = start_loop; | |
2052 | start_loop = start_loop->inner; | |
2053 | } while (start_loop && start_loop->inner); | |
2054 | ||
2055 | return false; | |
2056 | } | |
2057 | ||
2058 | /* Main entry for loop interchange pass. */ | |
2059 | ||
2060 | unsigned int | |
2061 | pass_linterchange::execute (function *fun) | |
2062 | { | |
2063 | if (number_of_loops (fun) <= 2) | |
2064 | return 0; | |
2065 | ||
2066 | bool changed_p = false; | |
2067 | struct loop *loop; | |
2068 | FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) | |
2069 | { | |
2070 | vec<loop_p> loop_nest = vNULL; | |
2071 | vec<data_reference_p> datarefs = vNULL; | |
2072 | vec<ddr_p> ddrs = vNULL; | |
2073 | if (prepare_perfect_loop_nest (loop, &loop_nest, &datarefs, &ddrs)) | |
2074 | { | |
2075 | tree_loop_interchange loop_interchange (loop_nest); | |
2076 | changed_p |= loop_interchange.interchange (datarefs, ddrs); | |
2077 | } | |
2078 | free_dependence_relations (ddrs); | |
2079 | free_data_refs_with_aux (datarefs); | |
2080 | loop_nest.release (); | |
2081 | } | |
2082 | ||
fbdec14e BC |
2083 | return changed_p ? (TODO_update_ssa_only_virtuals) : 0; |
2084 | } | |
2085 | ||
2086 | } // anon namespace | |
2087 | ||
2088 | gimple_opt_pass * | |
2089 | make_pass_linterchange (gcc::context *ctxt) | |
2090 | { | |
2091 | return new pass_linterchange (ctxt); | |
2092 | } |