]>
Commit | Line | Data |
---|---|---|
c6bb733d | 1 | /* Conversion of SESE regions to Polyhedra. |
3aea1f79 | 2 | Copyright (C) 2009-2014 Free Software Foundation, Inc. |
c6bb733d | 3 | Contributed by Sebastian Pop <sebastian.pop@amd.com>. |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify | |
8 | it under the terms of the GNU General Public License as published by | |
9 | the Free Software Foundation; either version 3, or (at your option) | |
10 | any later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, | |
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | GNU General Public License 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" | |
87e20041 | 22 | |
429cca51 | 23 | #ifdef HAVE_isl |
87e20041 | 24 | #include <isl/set.h> |
25 | #include <isl/map.h> | |
26 | #include <isl/union_map.h> | |
27 | #include <isl/constraint.h> | |
28 | #include <isl/aff.h> | |
ea48ac01 | 29 | #include <isl/val.h> |
30 | /* For C++ linkage of C functions. | |
31 | Missing from isl/val_gmp.h in isl 0.12 versions. | |
32 | Appearing in isl/val_gmp.h in isl 0.13. | |
33 | To be removed when passing to isl 0.13. */ | |
34 | #if defined(__cplusplus) | |
35 | extern "C" { | |
36 | #endif | |
37 | #include <isl/val_gmp.h> | |
38 | #if defined(__cplusplus) | |
39 | } | |
40 | #endif | |
429cca51 | 41 | #ifdef HAVE_cloog |
87e20041 | 42 | #include <cloog/cloog.h> |
43 | #include <cloog/cloog.h> | |
44 | #include <cloog/isl/domain.h> | |
45 | #endif | |
429cca51 | 46 | #endif |
87e20041 | 47 | |
c6bb733d | 48 | #include "system.h" |
49 | #include "coretypes.h" | |
41a8aa41 | 50 | #include "tree.h" |
bc61cadb | 51 | #include "basic-block.h" |
52 | #include "tree-ssa-alias.h" | |
53 | #include "internal-fn.h" | |
54 | #include "gimple-expr.h" | |
55 | #include "is-a.h" | |
073c1fd5 | 56 | #include "gimple.h" |
dcf1a1ec | 57 | #include "gimple-iterator.h" |
e795d6e1 | 58 | #include "gimplify.h" |
59 | #include "gimplify-me.h" | |
073c1fd5 | 60 | #include "gimple-ssa.h" |
61 | #include "tree-cfg.h" | |
62 | #include "tree-phinodes.h" | |
63 | #include "ssa-iterators.h" | |
9ed99284 | 64 | #include "stringpool.h" |
073c1fd5 | 65 | #include "tree-ssanames.h" |
05d9c18a | 66 | #include "tree-ssa-loop-manip.h" |
67 | #include "tree-ssa-loop-niter.h" | |
073c1fd5 | 68 | #include "tree-ssa-loop.h" |
69 | #include "tree-into-ssa.h" | |
f0a18dd2 | 70 | #include "tree-pass.h" |
c6bb733d | 71 | #include "cfgloop.h" |
72 | #include "tree-chrec.h" | |
73 | #include "tree-data-ref.h" | |
74 | #include "tree-scalar-evolution.h" | |
c6bb733d | 75 | #include "domwalk.h" |
c6bb733d | 76 | #include "sese.h" |
f8ba3083 | 77 | #include "tree-ssa-propagate.h" |
c6bb733d | 78 | |
429cca51 | 79 | #ifdef HAVE_isl |
3ebca59c | 80 | #include "expr.h" |
c6bb733d | 81 | #include "graphite-poly.h" |
c6bb733d | 82 | #include "graphite-sese-to-poly.h" |
83 | ||
87e20041 | 84 | |
85 | /* Assigns to RES the value of the INTEGER_CST T. */ | |
86 | ||
87 | static inline void | |
88 | tree_int_to_gmp (tree t, mpz_t res) | |
89 | { | |
ab2c1de8 | 90 | wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t))); |
87e20041 | 91 | } |
92 | ||
9d828157 | 93 | /* Returns the index of the PHI argument defined in the outermost |
94 | loop. */ | |
c6bb733d | 95 | |
96 | static size_t | |
9d828157 | 97 | phi_arg_in_outermost_loop (gimple phi) |
c6bb733d | 98 | { |
99 | loop_p loop = gimple_bb (phi)->loop_father; | |
9d828157 | 100 | size_t i, res = 0; |
c6bb733d | 101 | |
102 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
103 | if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src)) | |
9d828157 | 104 | { |
105 | loop = gimple_phi_arg_edge (phi, i)->src->loop_father; | |
106 | res = i; | |
107 | } | |
c6bb733d | 108 | |
9d828157 | 109 | return res; |
c6bb733d | 110 | } |
111 | ||
112 | /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position | |
113 | PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */ | |
114 | ||
115 | static void | |
116 | remove_simple_copy_phi (gimple_stmt_iterator *psi) | |
117 | { | |
118 | gimple phi = gsi_stmt (*psi); | |
119 | tree res = gimple_phi_result (phi); | |
9d828157 | 120 | size_t entry = phi_arg_in_outermost_loop (phi); |
c6bb733d | 121 | tree init = gimple_phi_arg_def (phi, entry); |
122 | gimple stmt = gimple_build_assign (res, init); | |
123 | edge e = gimple_phi_arg_edge (phi, entry); | |
124 | ||
125 | remove_phi_node (psi, false); | |
126 | gsi_insert_on_edge_immediate (e, stmt); | |
c6bb733d | 127 | } |
128 | ||
129 | /* Removes an invariant phi node at position PSI by inserting on the | |
130 | loop ENTRY edge the assignment RES = INIT. */ | |
131 | ||
132 | static void | |
133 | remove_invariant_phi (sese region, gimple_stmt_iterator *psi) | |
134 | { | |
135 | gimple phi = gsi_stmt (*psi); | |
136 | loop_p loop = loop_containing_stmt (phi); | |
137 | tree res = gimple_phi_result (phi); | |
138 | tree scev = scalar_evolution_in_region (region, loop, res); | |
9d828157 | 139 | size_t entry = phi_arg_in_outermost_loop (phi); |
c6bb733d | 140 | edge e = gimple_phi_arg_edge (phi, entry); |
141 | tree var; | |
142 | gimple stmt; | |
e3a19533 | 143 | gimple_seq stmts = NULL; |
c6bb733d | 144 | |
145 | if (tree_contains_chrecs (scev, NULL)) | |
146 | scev = gimple_phi_arg_def (phi, entry); | |
147 | ||
148 | var = force_gimple_operand (scev, &stmts, true, NULL_TREE); | |
149 | stmt = gimple_build_assign (res, var); | |
150 | remove_phi_node (psi, false); | |
151 | ||
e3a19533 | 152 | gimple_seq_add_stmt (&stmts, stmt); |
c6bb733d | 153 | gsi_insert_seq_on_edge (e, stmts); |
154 | gsi_commit_edge_inserts (); | |
155 | SSA_NAME_DEF_STMT (res) = stmt; | |
156 | } | |
157 | ||
158 | /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */ | |
159 | ||
160 | static inline bool | |
161 | simple_copy_phi_p (gimple phi) | |
162 | { | |
163 | tree res; | |
164 | ||
165 | if (gimple_phi_num_args (phi) != 2) | |
166 | return false; | |
167 | ||
168 | res = gimple_phi_result (phi); | |
169 | return (res == gimple_phi_arg_def (phi, 0) | |
170 | || res == gimple_phi_arg_def (phi, 1)); | |
171 | } | |
172 | ||
173 | /* Returns true when the phi node at position PSI is a reduction phi | |
174 | node in REGION. Otherwise moves the pointer PSI to the next phi to | |
175 | be considered. */ | |
176 | ||
177 | static bool | |
178 | reduction_phi_p (sese region, gimple_stmt_iterator *psi) | |
179 | { | |
180 | loop_p loop; | |
c6bb733d | 181 | gimple phi = gsi_stmt (*psi); |
182 | tree res = gimple_phi_result (phi); | |
183 | ||
c6bb733d | 184 | loop = loop_containing_stmt (phi); |
185 | ||
186 | if (simple_copy_phi_p (phi)) | |
187 | { | |
5d92ac74 | 188 | /* PRE introduces phi nodes like these, for an example, |
c6bb733d | 189 | see id-5.f in the fortran graphite testsuite: |
190 | ||
191 | # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)> | |
192 | */ | |
193 | remove_simple_copy_phi (psi); | |
194 | return false; | |
195 | } | |
196 | ||
15822b8e | 197 | if (scev_analyzable_p (res, region)) |
c6bb733d | 198 | { |
15822b8e | 199 | tree scev = scalar_evolution_in_region (region, loop, res); |
200 | ||
201 | if (evolution_function_is_invariant_p (scev, loop->num)) | |
30105622 | 202 | remove_invariant_phi (region, psi); |
203 | else | |
204 | gsi_next (psi); | |
205 | ||
c6bb733d | 206 | return false; |
207 | } | |
208 | ||
c6bb733d | 209 | /* All the other cases are considered reductions. */ |
210 | return true; | |
211 | } | |
212 | ||
c6bb733d | 213 | /* Store the GRAPHITE representation of BB. */ |
214 | ||
215 | static gimple_bb_p | |
f1f41a6c | 216 | new_gimple_bb (basic_block bb, vec<data_reference_p> drs) |
c6bb733d | 217 | { |
218 | struct gimple_bb *gbb; | |
219 | ||
220 | gbb = XNEW (struct gimple_bb); | |
221 | bb->aux = gbb; | |
222 | GBB_BB (gbb) = bb; | |
223 | GBB_DATA_REFS (gbb) = drs; | |
f1f41a6c | 224 | GBB_CONDITIONS (gbb).create (0); |
225 | GBB_CONDITION_CASES (gbb).create (0); | |
c6bb733d | 226 | |
227 | return gbb; | |
228 | } | |
229 | ||
ae11f03b | 230 | static void |
f1f41a6c | 231 | free_data_refs_aux (vec<data_reference_p> datarefs) |
ae11f03b | 232 | { |
233 | unsigned int i; | |
234 | struct data_reference *dr; | |
1a95516e | 235 | |
f1f41a6c | 236 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
1a95516e | 237 | if (dr->aux) |
ae11f03b | 238 | { |
e9a3f95f | 239 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); |
1a95516e | 240 | |
dd045aee | 241 | free (bap->alias_set); |
1a95516e | 242 | |
e9a3f95f | 243 | free (bap); |
ae11f03b | 244 | dr->aux = NULL; |
245 | } | |
246 | } | |
c6bb733d | 247 | /* Frees GBB. */ |
248 | ||
249 | static void | |
250 | free_gimple_bb (struct gimple_bb *gbb) | |
251 | { | |
ae11f03b | 252 | free_data_refs_aux (GBB_DATA_REFS (gbb)); |
c6bb733d | 253 | free_data_refs (GBB_DATA_REFS (gbb)); |
254 | ||
f1f41a6c | 255 | GBB_CONDITIONS (gbb).release (); |
256 | GBB_CONDITION_CASES (gbb).release (); | |
c6bb733d | 257 | GBB_BB (gbb)->aux = 0; |
258 | XDELETE (gbb); | |
259 | } | |
260 | ||
261 | /* Deletes all gimple bbs in SCOP. */ | |
262 | ||
263 | static void | |
264 | remove_gbbs_in_scop (scop_p scop) | |
265 | { | |
266 | int i; | |
267 | poly_bb_p pbb; | |
268 | ||
f1f41a6c | 269 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
c6bb733d | 270 | free_gimple_bb (PBB_BLACK_BOX (pbb)); |
271 | } | |
272 | ||
273 | /* Deletes all scops in SCOPS. */ | |
274 | ||
275 | void | |
f1f41a6c | 276 | free_scops (vec<scop_p> scops) |
c6bb733d | 277 | { |
278 | int i; | |
279 | scop_p scop; | |
280 | ||
f1f41a6c | 281 | FOR_EACH_VEC_ELT (scops, i, scop) |
c6bb733d | 282 | { |
283 | remove_gbbs_in_scop (scop); | |
284 | free_sese (SCOP_REGION (scop)); | |
285 | free_scop (scop); | |
286 | } | |
287 | ||
f1f41a6c | 288 | scops.release (); |
c6bb733d | 289 | } |
290 | ||
221a697e | 291 | /* Same as outermost_loop_in_sese, returns the outermost loop |
292 | containing BB in REGION, but makes sure that the returned loop | |
293 | belongs to the REGION, and so this returns the first loop in the | |
294 | REGION when the loop containing BB does not belong to REGION. */ | |
295 | ||
296 | static loop_p | |
297 | outermost_loop_in_sese_1 (sese region, basic_block bb) | |
298 | { | |
299 | loop_p nest = outermost_loop_in_sese (region, bb); | |
300 | ||
301 | if (loop_in_sese_p (nest, region)) | |
302 | return nest; | |
303 | ||
304 | /* When the basic block BB does not belong to a loop in the region, | |
305 | return the first loop in the region. */ | |
306 | nest = nest->inner; | |
307 | while (nest) | |
308 | if (loop_in_sese_p (nest, region)) | |
309 | break; | |
310 | else | |
311 | nest = nest->next; | |
312 | ||
313 | gcc_assert (nest); | |
314 | return nest; | |
315 | } | |
316 | ||
c6bb733d | 317 | /* Generates a polyhedral black box only if the bb contains interesting |
318 | information. */ | |
319 | ||
8c6b3774 | 320 | static gimple_bb_p |
321 | try_generate_gimple_bb (scop_p scop, basic_block bb) | |
c6bb733d | 322 | { |
f1f41a6c | 323 | vec<data_reference_p> drs; |
324 | drs.create (5); | |
221a697e | 325 | sese region = SCOP_REGION (scop); |
326 | loop_p nest = outermost_loop_in_sese_1 (region, bb); | |
c6bb733d | 327 | gimple_stmt_iterator gsi; |
328 | ||
329 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3f6c0a40 | 330 | { |
331 | gimple stmt = gsi_stmt (gsi); | |
221a697e | 332 | loop_p loop; |
333 | ||
334 | if (is_gimple_debug (stmt)) | |
335 | continue; | |
336 | ||
337 | loop = loop_containing_stmt (stmt); | |
338 | if (!loop_in_sese_p (loop, region)) | |
339 | loop = nest; | |
340 | ||
341 | graphite_find_data_references_in_stmt (nest, loop, stmt, &drs); | |
3f6c0a40 | 342 | } |
c6bb733d | 343 | |
8c6b3774 | 344 | return new_gimple_bb (bb, drs); |
c6bb733d | 345 | } |
346 | ||
347 | /* Returns true if all predecessors of BB, that are not dominated by BB, are | |
348 | marked in MAP. The predecessors dominated by BB are loop latches and will | |
349 | be handled after BB. */ | |
350 | ||
351 | static bool | |
352 | all_non_dominated_preds_marked_p (basic_block bb, sbitmap map) | |
353 | { | |
354 | edge e; | |
355 | edge_iterator ei; | |
356 | ||
357 | FOR_EACH_EDGE (e, ei, bb->preds) | |
08b7917c | 358 | if (!bitmap_bit_p (map, e->src->index) |
c6bb733d | 359 | && !dominated_by_p (CDI_DOMINATORS, e->src, bb)) |
360 | return false; | |
361 | ||
362 | return true; | |
363 | } | |
364 | ||
365 | /* Compare the depth of two basic_block's P1 and P2. */ | |
366 | ||
367 | static int | |
368 | compare_bb_depths (const void *p1, const void *p2) | |
369 | { | |
370 | const_basic_block const bb1 = *(const_basic_block const*)p1; | |
371 | const_basic_block const bb2 = *(const_basic_block const*)p2; | |
372 | int d1 = loop_depth (bb1->loop_father); | |
373 | int d2 = loop_depth (bb2->loop_father); | |
374 | ||
375 | if (d1 < d2) | |
376 | return 1; | |
377 | ||
378 | if (d1 > d2) | |
379 | return -1; | |
380 | ||
381 | return 0; | |
382 | } | |
383 | ||
384 | /* Sort the basic blocks from DOM such that the first are the ones at | |
385 | a deepest loop level. */ | |
386 | ||
387 | static void | |
f1f41a6c | 388 | graphite_sort_dominated_info (vec<basic_block> dom) |
c6bb733d | 389 | { |
f1f41a6c | 390 | dom.qsort (compare_bb_depths); |
c6bb733d | 391 | } |
392 | ||
393 | /* Recursive helper function for build_scops_bbs. */ | |
394 | ||
395 | static void | |
8c6b3774 | 396 | build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb) |
c6bb733d | 397 | { |
398 | sese region = SCOP_REGION (scop); | |
f1f41a6c | 399 | vec<basic_block> dom; |
8c6b3774 | 400 | poly_bb_p pbb; |
c6bb733d | 401 | |
08b7917c | 402 | if (bitmap_bit_p (visited, bb->index) |
c6bb733d | 403 | || !bb_in_sese_p (bb, region)) |
404 | return; | |
405 | ||
8c6b3774 | 406 | pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb)); |
f1f41a6c | 407 | SCOP_BBS (scop).safe_push (pbb); |
08b7917c | 408 | bitmap_set_bit (visited, bb->index); |
c6bb733d | 409 | |
410 | dom = get_dominated_by (CDI_DOMINATORS, bb); | |
411 | ||
f1f41a6c | 412 | if (!dom.exists ()) |
c6bb733d | 413 | return; |
414 | ||
415 | graphite_sort_dominated_info (dom); | |
416 | ||
f1f41a6c | 417 | while (!dom.is_empty ()) |
c6bb733d | 418 | { |
419 | int i; | |
420 | basic_block dom_bb; | |
421 | ||
f1f41a6c | 422 | FOR_EACH_VEC_ELT (dom, i, dom_bb) |
c6bb733d | 423 | if (all_non_dominated_preds_marked_p (dom_bb, visited)) |
424 | { | |
8c6b3774 | 425 | build_scop_bbs_1 (scop, visited, dom_bb); |
f1f41a6c | 426 | dom.unordered_remove (i); |
c6bb733d | 427 | break; |
428 | } | |
429 | } | |
430 | ||
f1f41a6c | 431 | dom.release (); |
c6bb733d | 432 | } |
433 | ||
434 | /* Gather the basic blocks belonging to the SCOP. */ | |
435 | ||
8c6b3774 | 436 | static void |
437 | build_scop_bbs (scop_p scop) | |
c6bb733d | 438 | { |
fe672ac0 | 439 | sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
c6bb733d | 440 | sese region = SCOP_REGION (scop); |
441 | ||
53c5d9d4 | 442 | bitmap_clear (visited); |
8c6b3774 | 443 | build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region)); |
c6bb733d | 444 | sbitmap_free (visited); |
445 | } | |
446 | ||
87e20041 | 447 | /* Return an ISL identifier for the polyhedral basic block PBB. */ |
448 | ||
449 | static isl_id * | |
450 | isl_id_for_pbb (scop_p s, poly_bb_p pbb) | |
451 | { | |
452 | char name[50]; | |
453 | snprintf (name, sizeof (name), "S_%d", pbb_index (pbb)); | |
454 | return isl_id_alloc (s->ctx, name, pbb); | |
455 | } | |
456 | ||
c6bb733d | 457 | /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron. |
458 | We generate SCATTERING_DIMENSIONS scattering dimensions. | |
459 | ||
460 | CLooG 0.15.0 and previous versions require, that all | |
461 | scattering functions of one CloogProgram have the same number of | |
462 | scattering dimensions, therefore we allow to specify it. This | |
463 | should be removed in future versions of CLooG. | |
464 | ||
465 | The scattering polyhedron consists of these dimensions: scattering, | |
466 | loop_iterators, parameters. | |
467 | ||
468 | Example: | |
469 | ||
470 | | scattering_dimensions = 5 | |
471 | | used_scattering_dimensions = 3 | |
472 | | nb_iterators = 1 | |
473 | | scop_nb_params = 2 | |
474 | | | |
475 | | Schedule: | |
476 | | i | |
477 | | 4 5 | |
478 | | | |
479 | | Scattering polyhedron: | |
480 | | | |
481 | | scattering: {s1, s2, s3, s4, s5} | |
482 | | loop_iterators: {i} | |
483 | | parameters: {p1, p2} | |
484 | | | |
485 | | s1 s2 s3 s4 s5 i p1 p2 1 | |
486 | | 1 0 0 0 0 0 0 0 -4 = 0 | |
487 | | 0 1 0 0 0 -1 0 0 0 = 0 | |
488 | | 0 0 1 0 0 0 0 0 -5 = 0 */ | |
489 | ||
490 | static void | |
87e20041 | 491 | build_pbb_scattering_polyhedrons (isl_aff *static_sched, |
c6bb733d | 492 | poly_bb_p pbb, int scattering_dimensions) |
493 | { | |
494 | int i; | |
c6bb733d | 495 | int nb_iterators = pbb_dim_iter_domain (pbb); |
496 | int used_scattering_dimensions = nb_iterators * 2 + 1; | |
ea48ac01 | 497 | isl_val *val; |
87e20041 | 498 | isl_space *dc, *dm; |
c6bb733d | 499 | |
500 | gcc_assert (scattering_dimensions >= used_scattering_dimensions); | |
501 | ||
87e20041 | 502 | dc = isl_set_get_space (pbb->domain); |
503 | dm = isl_space_add_dims (isl_space_from_domain (dc), | |
504 | isl_dim_out, scattering_dimensions); | |
505 | pbb->schedule = isl_map_universe (dm); | |
c6bb733d | 506 | |
507 | for (i = 0; i < scattering_dimensions; i++) | |
508 | { | |
c6bb733d | 509 | /* Textual order inside this loop. */ |
510 | if ((i % 2) == 0) | |
511 | { | |
87e20041 | 512 | isl_constraint *c = isl_equality_alloc |
513 | (isl_local_space_from_space (isl_map_get_space (pbb->schedule))); | |
514 | ||
ea48ac01 | 515 | val = isl_aff_get_coefficient_val (static_sched, isl_dim_in, i / 2); |
87e20041 | 516 | |
ea48ac01 | 517 | val = isl_val_neg (val); |
518 | c = isl_constraint_set_constant_val (c, val); | |
87e20041 | 519 | c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1); |
520 | pbb->schedule = isl_map_add_constraint (pbb->schedule, c); | |
c6bb733d | 521 | } |
522 | ||
523 | /* Iterations of this loop. */ | |
524 | else /* if ((i % 2) == 1) */ | |
525 | { | |
526 | int loop = (i - 1) / 2; | |
87e20041 | 527 | pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop, |
528 | isl_dim_out, i); | |
c6bb733d | 529 | } |
c6bb733d | 530 | } |
531 | ||
87e20041 | 532 | pbb->transformed = isl_map_copy (pbb->schedule); |
c6bb733d | 533 | } |
534 | ||
535 | /* Build for BB the static schedule. | |
536 | ||
537 | The static schedule is a Dewey numbering of the abstract syntax | |
538 | tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification | |
539 | ||
540 | The following example informally defines the static schedule: | |
541 | ||
542 | A | |
543 | for (i: ...) | |
544 | { | |
545 | for (j: ...) | |
546 | { | |
547 | B | |
548 | C | |
549 | } | |
550 | ||
551 | for (k: ...) | |
552 | { | |
553 | D | |
554 | E | |
555 | } | |
556 | } | |
557 | F | |
558 | ||
559 | Static schedules for A to F: | |
560 | ||
561 | DEPTH | |
562 | 0 1 2 | |
563 | A 0 | |
564 | B 1 0 0 | |
565 | C 1 0 1 | |
566 | D 1 1 0 | |
567 | E 1 1 1 | |
568 | F 2 | |
569 | */ | |
570 | ||
571 | static void | |
572 | build_scop_scattering (scop_p scop) | |
573 | { | |
574 | int i; | |
575 | poly_bb_p pbb; | |
576 | gimple_bb_p previous_gbb = NULL; | |
87e20041 | 577 | isl_space *dc = isl_set_get_space (scop->context); |
578 | isl_aff *static_sched; | |
c6bb733d | 579 | |
41f75a99 | 580 | dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun)); |
87e20041 | 581 | static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc)); |
c6bb733d | 582 | |
583 | /* We have to start schedules at 0 on the first component and | |
584 | because we cannot compare_prefix_loops against a previous loop, | |
585 | prefix will be equal to zero, and that index will be | |
586 | incremented before copying. */ | |
87e20041 | 587 | static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1); |
c6bb733d | 588 | |
f1f41a6c | 589 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
c6bb733d | 590 | { |
591 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); | |
c6bb733d | 592 | int prefix; |
593 | int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1; | |
594 | ||
595 | if (previous_gbb) | |
596 | prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb); | |
597 | else | |
598 | prefix = 0; | |
599 | ||
600 | previous_gbb = gbb; | |
c6bb733d | 601 | |
87e20041 | 602 | static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, |
603 | prefix, 1); | |
604 | build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims); | |
605 | } | |
606 | ||
607 | isl_aff_free (static_sched); | |
608 | } | |
609 | ||
610 | static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space); | |
611 | ||
612 | /* Extract an affine expression from the chain of recurrence E. */ | |
c6bb733d | 613 | |
87e20041 | 614 | static isl_pw_aff * |
615 | extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space) | |
616 | { | |
617 | isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space)); | |
618 | isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space)); | |
619 | isl_local_space *ls = isl_local_space_from_space (space); | |
41f75a99 | 620 | unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1; |
87e20041 | 621 | isl_aff *loop = isl_aff_set_coefficient_si |
622 | (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1); | |
623 | isl_pw_aff *l = isl_pw_aff_from_aff (loop); | |
624 | ||
625 | /* Before multiplying, make sure that the result is affine. */ | |
626 | gcc_assert (isl_pw_aff_is_cst (rhs) | |
627 | || isl_pw_aff_is_cst (l)); | |
628 | ||
629 | return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l)); | |
630 | } | |
631 | ||
632 | /* Extract an affine expression from the mult_expr E. */ | |
633 | ||
634 | static isl_pw_aff * | |
635 | extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space) | |
636 | { | |
637 | isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0), | |
638 | isl_space_copy (space)); | |
639 | isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
c6bb733d | 640 | |
87e20041 | 641 | if (!isl_pw_aff_is_cst (lhs) |
642 | && !isl_pw_aff_is_cst (rhs)) | |
643 | { | |
644 | isl_pw_aff_free (lhs); | |
645 | isl_pw_aff_free (rhs); | |
646 | return NULL; | |
c6bb733d | 647 | } |
648 | ||
87e20041 | 649 | return isl_pw_aff_mul (lhs, rhs); |
c6bb733d | 650 | } |
651 | ||
87e20041 | 652 | /* Return an ISL identifier from the name of the ssa_name E. */ |
c6bb733d | 653 | |
87e20041 | 654 | static isl_id * |
655 | isl_id_for_ssa_name (scop_p s, tree e) | |
c6bb733d | 656 | { |
87e20041 | 657 | const char *name = get_name (e); |
658 | isl_id *id; | |
659 | ||
660 | if (name) | |
661 | id = isl_id_alloc (s->ctx, name, e); | |
662 | else | |
663 | { | |
664 | char name1[50]; | |
665 | snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e)); | |
666 | id = isl_id_alloc (s->ctx, name1, e); | |
667 | } | |
c6bb733d | 668 | |
87e20041 | 669 | return id; |
670 | } | |
c6bb733d | 671 | |
87e20041 | 672 | /* Return an ISL identifier for the data reference DR. */ |
c6bb733d | 673 | |
87e20041 | 674 | static isl_id * |
675 | isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED) | |
676 | { | |
677 | /* Data references all get the same isl_id. They need to be comparable | |
678 | and are distinguished through the first dimension, which contains the | |
679 | alias set number. */ | |
680 | return isl_id_alloc (s->ctx, "", 0); | |
c6bb733d | 681 | } |
682 | ||
87e20041 | 683 | /* Extract an affine expression from the ssa_name E. */ |
c6bb733d | 684 | |
87e20041 | 685 | static isl_pw_aff * |
686 | extract_affine_name (scop_p s, tree e, __isl_take isl_space *space) | |
c6bb733d | 687 | { |
87e20041 | 688 | isl_aff *aff; |
689 | isl_set *dom; | |
690 | isl_id *id; | |
691 | int dimension; | |
692 | ||
693 | id = isl_id_for_ssa_name (s, e); | |
694 | dimension = isl_space_find_dim_by_id (space, isl_dim_param, id); | |
9af5ce0c | 695 | isl_id_free (id); |
87e20041 | 696 | dom = isl_set_universe (isl_space_copy (space)); |
697 | aff = isl_aff_zero_on_domain (isl_local_space_from_space (space)); | |
698 | aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1); | |
699 | return isl_pw_aff_alloc (dom, aff); | |
700 | } | |
c6bb733d | 701 | |
87e20041 | 702 | /* Extract an affine expression from the gmp constant G. */ |
c6bb733d | 703 | |
87e20041 | 704 | static isl_pw_aff * |
705 | extract_affine_gmp (mpz_t g, __isl_take isl_space *space) | |
706 | { | |
707 | isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space)); | |
708 | isl_aff *aff = isl_aff_zero_on_domain (ls); | |
709 | isl_set *dom = isl_set_universe (space); | |
ea48ac01 | 710 | isl_val *v; |
711 | isl_ctx *ct; | |
c6bb733d | 712 | |
ea48ac01 | 713 | ct = isl_aff_get_ctx (aff); |
714 | v = isl_val_int_from_gmp (ct, g); | |
715 | aff = isl_aff_add_constant_val (aff, v); | |
c6bb733d | 716 | |
87e20041 | 717 | return isl_pw_aff_alloc (dom, aff); |
c6bb733d | 718 | } |
719 | ||
87e20041 | 720 | /* Extract an affine expression from the integer_cst E. */ |
c6bb733d | 721 | |
87e20041 | 722 | static isl_pw_aff * |
723 | extract_affine_int (tree e, __isl_take isl_space *space) | |
724 | { | |
725 | isl_pw_aff *res; | |
726 | mpz_t g; | |
727 | ||
728 | mpz_init (g); | |
729 | tree_int_to_gmp (e, g); | |
730 | res = extract_affine_gmp (g, space); | |
731 | mpz_clear (g); | |
732 | ||
733 | return res; | |
734 | } | |
735 | ||
736 | /* Compute pwaff mod 2^width. */ | |
737 | ||
ea48ac01 | 738 | extern isl_ctx *the_isl_ctx; |
739 | ||
87e20041 | 740 | static isl_pw_aff * |
741 | wrap (isl_pw_aff *pwaff, unsigned width) | |
c6bb733d | 742 | { |
ea48ac01 | 743 | isl_val *mod; |
c6bb733d | 744 | |
ea48ac01 | 745 | mod = isl_val_int_from_ui(the_isl_ctx, width); |
746 | mod = isl_val_2exp (mod); | |
747 | pwaff = isl_pw_aff_mod_val (pwaff, mod); | |
87e20041 | 748 | |
749 | return pwaff; | |
c6bb733d | 750 | } |
751 | ||
c6bb733d | 752 | /* When parameter NAME is in REGION, returns its index in SESE_PARAMS. |
753 | Otherwise returns -1. */ | |
754 | ||
755 | static inline int | |
756 | parameter_index_in_region_1 (tree name, sese region) | |
757 | { | |
758 | int i; | |
759 | tree p; | |
760 | ||
761 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
762 | ||
f1f41a6c | 763 | FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p) |
c6bb733d | 764 | if (p == name) |
765 | return i; | |
766 | ||
767 | return -1; | |
768 | } | |
769 | ||
770 | /* When the parameter NAME is in REGION, returns its index in | |
771 | SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS | |
772 | and returns the index of NAME. */ | |
773 | ||
774 | static int | |
775 | parameter_index_in_region (tree name, sese region) | |
776 | { | |
777 | int i; | |
778 | ||
779 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
780 | ||
781 | i = parameter_index_in_region_1 (name, region); | |
782 | if (i != -1) | |
783 | return i; | |
784 | ||
785 | gcc_assert (SESE_ADD_PARAMS (region)); | |
786 | ||
f1f41a6c | 787 | i = SESE_PARAMS (region).length (); |
788 | SESE_PARAMS (region).safe_push (name); | |
c6bb733d | 789 | return i; |
790 | } | |
791 | ||
87e20041 | 792 | /* Extract an affine expression from the tree E in the scop S. */ |
c6bb733d | 793 | |
87e20041 | 794 | static isl_pw_aff * |
795 | extract_affine (scop_p s, tree e, __isl_take isl_space *space) | |
c6bb733d | 796 | { |
87e20041 | 797 | isl_pw_aff *lhs, *rhs, *res; |
798 | tree type; | |
799 | ||
800 | if (e == chrec_dont_know) { | |
801 | isl_space_free (space); | |
802 | return NULL; | |
803 | } | |
c6bb733d | 804 | |
805 | switch (TREE_CODE (e)) | |
806 | { | |
807 | case POLYNOMIAL_CHREC: | |
87e20041 | 808 | res = extract_affine_chrec (s, e, space); |
c6bb733d | 809 | break; |
810 | ||
811 | case MULT_EXPR: | |
87e20041 | 812 | res = extract_affine_mul (s, e, space); |
c6bb733d | 813 | break; |
814 | ||
815 | case PLUS_EXPR: | |
816 | case POINTER_PLUS_EXPR: | |
87e20041 | 817 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); |
818 | rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
819 | res = isl_pw_aff_add (lhs, rhs); | |
c6bb733d | 820 | break; |
821 | ||
822 | case MINUS_EXPR: | |
87e20041 | 823 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); |
824 | rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
825 | res = isl_pw_aff_sub (lhs, rhs); | |
826 | break; | |
c6bb733d | 827 | |
828 | case NEGATE_EXPR: | |
87e20041 | 829 | case BIT_NOT_EXPR: |
830 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); | |
831 | rhs = extract_affine (s, integer_minus_one_node, space); | |
832 | res = isl_pw_aff_mul (lhs, rhs); | |
833 | break; | |
c6bb733d | 834 | |
87e20041 | 835 | case SSA_NAME: |
836 | gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s))); | |
837 | res = extract_affine_name (s, e, space); | |
838 | break; | |
c6bb733d | 839 | |
87e20041 | 840 | case INTEGER_CST: |
841 | res = extract_affine_int (e, space); | |
842 | /* No need to wrap a single integer. */ | |
843 | return res; | |
c6bb733d | 844 | |
87e20041 | 845 | CASE_CONVERT: |
846 | case NON_LVALUE_EXPR: | |
847 | res = extract_affine (s, TREE_OPERAND (e, 0), space); | |
848 | break; | |
c6bb733d | 849 | |
87e20041 | 850 | default: |
851 | gcc_unreachable (); | |
852 | break; | |
853 | } | |
c6bb733d | 854 | |
87e20041 | 855 | type = TREE_TYPE (e); |
856 | if (TYPE_UNSIGNED (type)) | |
857 | res = wrap (res, TYPE_PRECISION (type)); | |
c6bb733d | 858 | |
87e20041 | 859 | return res; |
860 | } | |
c6bb733d | 861 | |
87e20041 | 862 | /* In the context of sese S, scan the expression E and translate it to |
863 | a linear expression C. When parsing a symbolic multiplication, K | |
864 | represents the constant multiplier of an expression containing | |
865 | parameters. */ | |
c6bb733d | 866 | |
87e20041 | 867 | static void |
868 | scan_tree_for_params (sese s, tree e) | |
869 | { | |
870 | if (e == chrec_dont_know) | |
871 | return; | |
c6bb733d | 872 | |
87e20041 | 873 | switch (TREE_CODE (e)) |
874 | { | |
875 | case POLYNOMIAL_CHREC: | |
876 | scan_tree_for_params (s, CHREC_LEFT (e)); | |
877 | break; | |
c6bb733d | 878 | |
87e20041 | 879 | case MULT_EXPR: |
880 | if (chrec_contains_symbols (TREE_OPERAND (e, 0))) | |
881 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); | |
882 | else | |
883 | scan_tree_for_params (s, TREE_OPERAND (e, 1)); | |
884 | break; | |
c6bb733d | 885 | |
87e20041 | 886 | case PLUS_EXPR: |
887 | case POINTER_PLUS_EXPR: | |
888 | case MINUS_EXPR: | |
889 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); | |
890 | scan_tree_for_params (s, TREE_OPERAND (e, 1)); | |
c6bb733d | 891 | break; |
892 | ||
87e20041 | 893 | case NEGATE_EXPR: |
894 | case BIT_NOT_EXPR: | |
c6bb733d | 895 | CASE_CONVERT: |
896 | case NON_LVALUE_EXPR: | |
87e20041 | 897 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); |
c6bb733d | 898 | break; |
899 | ||
87e20041 | 900 | case SSA_NAME: |
901 | parameter_index_in_region (e, s); | |
902 | break; | |
903 | ||
904 | case INTEGER_CST: | |
a1b6cd6a | 905 | case ADDR_EXPR: |
906 | break; | |
907 | ||
c6bb733d | 908 | default: |
909 | gcc_unreachable (); | |
910 | break; | |
911 | } | |
912 | } | |
913 | ||
c6bb733d | 914 | /* Find parameters with respect to REGION in BB. We are looking in memory |
915 | access functions, conditions and loop bounds. */ | |
916 | ||
917 | static void | |
918 | find_params_in_bb (sese region, gimple_bb_p gbb) | |
919 | { | |
920 | int i; | |
db899978 | 921 | unsigned j; |
c6bb733d | 922 | data_reference_p dr; |
923 | gimple stmt; | |
924 | loop_p loop = GBB_BB (gbb)->loop_father; | |
c6bb733d | 925 | |
db899978 | 926 | /* Find parameters in the access functions of data references. */ |
f1f41a6c | 927 | FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr) |
db899978 | 928 | for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++) |
87e20041 | 929 | scan_tree_for_params (region, DR_ACCESS_FN (dr, j)); |
c6bb733d | 930 | |
931 | /* Find parameters in conditional statements. */ | |
f1f41a6c | 932 | FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt) |
c6bb733d | 933 | { |
c6bb733d | 934 | tree lhs = scalar_evolution_in_region (region, loop, |
935 | gimple_cond_lhs (stmt)); | |
936 | tree rhs = scalar_evolution_in_region (region, loop, | |
937 | gimple_cond_rhs (stmt)); | |
938 | ||
87e20041 | 939 | scan_tree_for_params (region, lhs); |
940 | scan_tree_for_params (region, rhs); | |
c6bb733d | 941 | } |
942 | } | |
943 | ||
944 | /* Record the parameters used in the SCOP. A variable is a parameter | |
945 | in a scop if it does not vary during the execution of that scop. */ | |
946 | ||
947 | static void | |
948 | find_scop_parameters (scop_p scop) | |
949 | { | |
950 | poly_bb_p pbb; | |
951 | unsigned i; | |
952 | sese region = SCOP_REGION (scop); | |
953 | struct loop *loop; | |
87e20041 | 954 | int nbp; |
c6bb733d | 955 | |
956 | /* Find the parameters used in the loop bounds. */ | |
f1f41a6c | 957 | FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop) |
c6bb733d | 958 | { |
959 | tree nb_iters = number_of_latch_executions (loop); | |
960 | ||
961 | if (!chrec_contains_symbols (nb_iters)) | |
962 | continue; | |
963 | ||
964 | nb_iters = scalar_evolution_in_region (region, loop, nb_iters); | |
87e20041 | 965 | scan_tree_for_params (region, nb_iters); |
c6bb733d | 966 | } |
967 | ||
c6bb733d | 968 | /* Find the parameters used in data accesses. */ |
f1f41a6c | 969 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
c6bb733d | 970 | find_params_in_bb (region, PBB_BLACK_BOX (pbb)); |
971 | ||
87e20041 | 972 | nbp = sese_nb_params (region); |
973 | scop_set_nb_params (scop, nbp); | |
c6bb733d | 974 | SESE_ADD_PARAMS (region) = false; |
acb3969f | 975 | |
5d92ac74 | 976 | { |
87e20041 | 977 | tree e; |
978 | isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0); | |
5d92ac74 | 979 | |
f1f41a6c | 980 | FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e) |
87e20041 | 981 | space = isl_space_set_dim_id (space, isl_dim_param, i, |
982 | isl_id_for_ssa_name (scop, e)); | |
5d92ac74 | 983 | |
87e20041 | 984 | scop->context = isl_set_universe (space); |
5d92ac74 | 985 | } |
5d92ac74 | 986 | } |
987 | ||
c6bb733d | 988 | /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives |
989 | the constraints for the surrounding loops. */ | |
990 | ||
991 | static void | |
992 | build_loop_iteration_domains (scop_p scop, struct loop *loop, | |
87e20041 | 993 | int nb, |
994 | isl_set *outer, isl_set **doms) | |
c6bb733d | 995 | { |
c6bb733d | 996 | tree nb_iters = number_of_latch_executions (loop); |
c6bb733d | 997 | sese region = SCOP_REGION (scop); |
998 | ||
87e20041 | 999 | isl_set *inner = isl_set_copy (outer); |
1000 | isl_space *space; | |
1001 | isl_constraint *c; | |
1002 | int pos = isl_set_dim (outer, isl_dim_set); | |
ea48ac01 | 1003 | isl_val *v; |
87e20041 | 1004 | mpz_t g; |
1005 | ||
1006 | mpz_init (g); | |
87e20041 | 1007 | |
1008 | inner = isl_set_add_dims (inner, isl_dim_set, 1); | |
1009 | space = isl_set_get_space (inner); | |
c6bb733d | 1010 | |
1011 | /* 0 <= loop_i */ | |
87e20041 | 1012 | c = isl_inequality_alloc |
1013 | (isl_local_space_from_space (isl_space_copy (space))); | |
1014 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1); | |
1015 | inner = isl_set_add_constraint (inner, c); | |
c6bb733d | 1016 | |
87e20041 | 1017 | /* loop_i <= cst_nb_iters */ |
c6bb733d | 1018 | if (TREE_CODE (nb_iters) == INTEGER_CST) |
1019 | { | |
87e20041 | 1020 | c = isl_inequality_alloc |
9af5ce0c | 1021 | (isl_local_space_from_space (isl_space_copy (space))); |
87e20041 | 1022 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1); |
1023 | tree_int_to_gmp (nb_iters, g); | |
ea48ac01 | 1024 | v = isl_val_int_from_gmp (the_isl_ctx, g); |
1025 | c = isl_constraint_set_constant_val (c, v); | |
87e20041 | 1026 | inner = isl_set_add_constraint (inner, c); |
c6bb733d | 1027 | } |
87e20041 | 1028 | |
1029 | /* loop_i <= expr_nb_iters */ | |
c6bb733d | 1030 | else if (!chrec_contains_undetermined (nb_iters)) |
1031 | { | |
5de9d3ed | 1032 | widest_int nit; |
87e20041 | 1033 | isl_pw_aff *aff; |
1034 | isl_set *valid; | |
1035 | isl_local_space *ls; | |
1036 | isl_aff *al; | |
1037 | isl_set *le; | |
c6bb733d | 1038 | |
c6bb733d | 1039 | nb_iters = scalar_evolution_in_region (region, loop, nb_iters); |
87e20041 | 1040 | |
1041 | aff = extract_affine (scop, nb_iters, isl_set_get_space (inner)); | |
1042 | valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff)); | |
1043 | valid = isl_set_project_out (valid, isl_dim_set, 0, | |
1044 | isl_set_dim (valid, isl_dim_set)); | |
1045 | scop->context = isl_set_intersect (scop->context, valid); | |
1046 | ||
1047 | ls = isl_local_space_from_space (isl_space_copy (space)); | |
1048 | al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls), | |
1049 | isl_dim_in, pos, 1); | |
1050 | le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al), | |
1051 | isl_pw_aff_copy (aff)); | |
1052 | inner = isl_set_intersect (inner, le); | |
c6bb733d | 1053 | |
fee017b3 | 1054 | if (max_stmt_executions (loop, &nit)) |
87e20041 | 1055 | { |
1056 | /* Insert in the context the constraints from the | |
1057 | estimation of the number of iterations NIT and the | |
1058 | symbolic number of iterations (involving parameter | |
1059 | names) NB_ITERS. First, build the affine expression | |
1060 | "NIT - NB_ITERS" and then say that it is positive, | |
1061 | i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */ | |
1062 | isl_pw_aff *approx; | |
1063 | mpz_t g; | |
1064 | isl_set *x; | |
1065 | isl_constraint *c; | |
1066 | ||
1067 | mpz_init (g); | |
796b6678 | 1068 | wi::to_mpz (nit, g, SIGNED); |
87e20041 | 1069 | mpz_sub_ui (g, g, 1); |
1070 | approx = extract_affine_gmp (g, isl_set_get_space (inner)); | |
1071 | x = isl_pw_aff_ge_set (approx, aff); | |
1072 | x = isl_set_project_out (x, isl_dim_set, 0, | |
1073 | isl_set_dim (x, isl_dim_set)); | |
1074 | scop->context = isl_set_intersect (scop->context, x); | |
1075 | ||
1076 | c = isl_inequality_alloc | |
1077 | (isl_local_space_from_space (isl_space_copy (space))); | |
1078 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1); | |
ea48ac01 | 1079 | v = isl_val_int_from_gmp (the_isl_ctx, g); |
87e20041 | 1080 | mpz_clear (g); |
ea48ac01 | 1081 | c = isl_constraint_set_constant_val (c, v); |
87e20041 | 1082 | inner = isl_set_add_constraint (inner, c); |
1083 | } | |
3fb5ff55 | 1084 | else |
1085 | isl_pw_aff_free (aff); | |
c6bb733d | 1086 | } |
1087 | else | |
1088 | gcc_unreachable (); | |
1089 | ||
1090 | if (loop->inner && loop_in_sese_p (loop->inner, region)) | |
87e20041 | 1091 | build_loop_iteration_domains (scop, loop->inner, nb + 1, |
1092 | isl_set_copy (inner), doms); | |
c6bb733d | 1093 | |
1094 | if (nb != 0 | |
1095 | && loop->next | |
1096 | && loop_in_sese_p (loop->next, region)) | |
87e20041 | 1097 | build_loop_iteration_domains (scop, loop->next, nb, |
1098 | isl_set_copy (outer), doms); | |
c6bb733d | 1099 | |
87e20041 | 1100 | doms[loop->num] = inner; |
c6bb733d | 1101 | |
87e20041 | 1102 | isl_set_free (outer); |
1103 | isl_space_free (space); | |
87e20041 | 1104 | mpz_clear (g); |
c6bb733d | 1105 | } |
1106 | ||
1107 | /* Returns a linear expression for tree T evaluated in PBB. */ | |
1108 | ||
87e20041 | 1109 | static isl_pw_aff * |
1110 | create_pw_aff_from_tree (poly_bb_p pbb, tree t) | |
c6bb733d | 1111 | { |
87e20041 | 1112 | scop_p scop = PBB_SCOP (pbb); |
c6bb733d | 1113 | |
87e20041 | 1114 | t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t); |
c6bb733d | 1115 | gcc_assert (!automatically_generated_chrec_p (t)); |
1116 | ||
87e20041 | 1117 | return extract_affine (scop, t, isl_set_get_space (pbb->domain)); |
c6bb733d | 1118 | } |
1119 | ||
87e20041 | 1120 | /* Add conditional statement STMT to pbb. CODE is used as the comparison |
1121 | operator. This allows us to invert the condition or to handle | |
1122 | inequalities. */ | |
c6bb733d | 1123 | |
1124 | static void | |
87e20041 | 1125 | add_condition_to_pbb (poly_bb_p pbb, gimple stmt, enum tree_code code) |
c6bb733d | 1126 | { |
87e20041 | 1127 | isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt)); |
1128 | isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt)); | |
1129 | isl_set *cond; | |
c6bb733d | 1130 | |
87e20041 | 1131 | switch (code) |
c6bb733d | 1132 | { |
87e20041 | 1133 | case LT_EXPR: |
1134 | cond = isl_pw_aff_lt_set (lhs, rhs); | |
1135 | break; | |
c6bb733d | 1136 | |
87e20041 | 1137 | case GT_EXPR: |
1138 | cond = isl_pw_aff_gt_set (lhs, rhs); | |
1139 | break; | |
c6bb733d | 1140 | |
87e20041 | 1141 | case LE_EXPR: |
1142 | cond = isl_pw_aff_le_set (lhs, rhs); | |
1143 | break; | |
c6bb733d | 1144 | |
87e20041 | 1145 | case GE_EXPR: |
1146 | cond = isl_pw_aff_ge_set (lhs, rhs); | |
1147 | break; | |
c6bb733d | 1148 | |
87e20041 | 1149 | case EQ_EXPR: |
1150 | cond = isl_pw_aff_eq_set (lhs, rhs); | |
1151 | break; | |
c6bb733d | 1152 | |
87e20041 | 1153 | case NE_EXPR: |
1154 | cond = isl_pw_aff_ne_set (lhs, rhs); | |
1155 | break; | |
c6bb733d | 1156 | |
87e20041 | 1157 | default: |
9af5ce0c | 1158 | isl_pw_aff_free (lhs); |
1159 | isl_pw_aff_free (rhs); | |
87e20041 | 1160 | return; |
c6bb733d | 1161 | } |
87e20041 | 1162 | |
1163 | cond = isl_set_coalesce (cond); | |
1164 | cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain)); | |
1165 | pbb->domain = isl_set_intersect (pbb->domain, cond); | |
c6bb733d | 1166 | } |
1167 | ||
1168 | /* Add conditions to the domain of PBB. */ | |
1169 | ||
1170 | static void | |
1171 | add_conditions_to_domain (poly_bb_p pbb) | |
1172 | { | |
1173 | unsigned int i; | |
1174 | gimple stmt; | |
1175 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); | |
c6bb733d | 1176 | |
f1f41a6c | 1177 | if (GBB_CONDITIONS (gbb).is_empty ()) |
c6bb733d | 1178 | return; |
1179 | ||
f1f41a6c | 1180 | FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt) |
c6bb733d | 1181 | switch (gimple_code (stmt)) |
1182 | { | |
1183 | case GIMPLE_COND: | |
1184 | { | |
1185 | enum tree_code code = gimple_cond_code (stmt); | |
1186 | ||
1187 | /* The conditions for ELSE-branches are inverted. */ | |
f1f41a6c | 1188 | if (!GBB_CONDITION_CASES (gbb)[i]) |
c6bb733d | 1189 | code = invert_tree_comparison (code, false); |
1190 | ||
1191 | add_condition_to_pbb (pbb, stmt, code); | |
1192 | break; | |
1193 | } | |
1194 | ||
1195 | case GIMPLE_SWITCH: | |
87e20041 | 1196 | /* Switch statements are not supported right now - fall through. */ |
c6bb733d | 1197 | |
1198 | default: | |
1199 | gcc_unreachable (); | |
1200 | break; | |
1201 | } | |
1202 | } | |
1203 | ||
8c6b3774 | 1204 | /* Traverses all the GBBs of the SCOP and add their constraints to the |
1205 | iteration domains. */ | |
1206 | ||
1207 | static void | |
1208 | add_conditions_to_constraints (scop_p scop) | |
1209 | { | |
1210 | int i; | |
1211 | poly_bb_p pbb; | |
1212 | ||
f1f41a6c | 1213 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
8c6b3774 | 1214 | add_conditions_to_domain (pbb); |
1215 | } | |
1216 | ||
dff64cac | 1217 | /* Returns a COND_EXPR statement when BB has a single predecessor, the |
1218 | edge between BB and its predecessor is not a loop exit edge, and | |
1219 | the last statement of the single predecessor is a COND_EXPR. */ | |
c6bb733d | 1220 | |
1221 | static gimple | |
dff64cac | 1222 | single_pred_cond_non_loop_exit (basic_block bb) |
c6bb733d | 1223 | { |
1224 | if (single_pred_p (bb)) | |
1225 | { | |
1226 | edge e = single_pred_edge (bb); | |
1227 | basic_block pred = e->src; | |
dff64cac | 1228 | gimple stmt; |
1229 | ||
1230 | if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father)) | |
1231 | return NULL; | |
1232 | ||
1233 | stmt = last_stmt (pred); | |
c6bb733d | 1234 | |
1235 | if (stmt && gimple_code (stmt) == GIMPLE_COND) | |
1236 | return stmt; | |
1237 | } | |
dff64cac | 1238 | |
c6bb733d | 1239 | return NULL; |
1240 | } | |
1241 | ||
54c91640 | 1242 | class sese_dom_walker : public dom_walker |
1243 | { | |
1244 | public: | |
1245 | sese_dom_walker (cdi_direction, sese); | |
54c91640 | 1246 | |
1247 | virtual void before_dom_children (basic_block); | |
1248 | virtual void after_dom_children (basic_block); | |
1249 | ||
1250 | private: | |
4997014d | 1251 | auto_vec<gimple, 3> m_conditions, m_cases; |
ae84f584 | 1252 | sese m_region; |
54c91640 | 1253 | }; |
1254 | ||
1255 | sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region) | |
ae84f584 | 1256 | : dom_walker (direction), m_region (region) |
54c91640 | 1257 | { |
54c91640 | 1258 | } |
1259 | ||
c6bb733d | 1260 | /* Call-back for dom_walk executed before visiting the dominated |
1261 | blocks. */ | |
1262 | ||
54c91640 | 1263 | void |
1264 | sese_dom_walker::before_dom_children (basic_block bb) | |
c6bb733d | 1265 | { |
d3746d81 | 1266 | gimple_bb_p gbb; |
1267 | gimple stmt; | |
c6bb733d | 1268 | |
ae84f584 | 1269 | if (!bb_in_sese_p (bb, m_region)) |
c6bb733d | 1270 | return; |
1271 | ||
dff64cac | 1272 | stmt = single_pred_cond_non_loop_exit (bb); |
d3746d81 | 1273 | |
c6bb733d | 1274 | if (stmt) |
1275 | { | |
1276 | edge e = single_pred_edge (bb); | |
1277 | ||
ae84f584 | 1278 | m_conditions.safe_push (stmt); |
c6bb733d | 1279 | |
1280 | if (e->flags & EDGE_TRUE_VALUE) | |
ae84f584 | 1281 | m_cases.safe_push (stmt); |
c6bb733d | 1282 | else |
ae84f584 | 1283 | m_cases.safe_push (NULL); |
c6bb733d | 1284 | } |
1285 | ||
d3746d81 | 1286 | gbb = gbb_from_bb (bb); |
1287 | ||
c6bb733d | 1288 | if (gbb) |
1289 | { | |
ae84f584 | 1290 | GBB_CONDITIONS (gbb) = m_conditions.copy (); |
1291 | GBB_CONDITION_CASES (gbb) = m_cases.copy (); | |
c6bb733d | 1292 | } |
1293 | } | |
1294 | ||
1295 | /* Call-back for dom_walk executed after visiting the dominated | |
1296 | blocks. */ | |
1297 | ||
54c91640 | 1298 | void |
1299 | sese_dom_walker::after_dom_children (basic_block bb) | |
c6bb733d | 1300 | { |
ae84f584 | 1301 | if (!bb_in_sese_p (bb, m_region)) |
c6bb733d | 1302 | return; |
1303 | ||
dff64cac | 1304 | if (single_pred_cond_non_loop_exit (bb)) |
c6bb733d | 1305 | { |
ae84f584 | 1306 | m_conditions.pop (); |
1307 | m_cases.pop (); | |
c6bb733d | 1308 | } |
1309 | } | |
1310 | ||
c6bb733d | 1311 | /* Add constraints on the possible values of parameter P from the type |
1312 | of P. */ | |
1313 | ||
1314 | static void | |
87e20041 | 1315 | add_param_constraints (scop_p scop, graphite_dim_t p) |
c6bb733d | 1316 | { |
f1f41a6c | 1317 | tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p]; |
c6bb733d | 1318 | tree type = TREE_TYPE (parameter); |
88a62e9b | 1319 | tree lb = NULL_TREE; |
1320 | tree ub = NULL_TREE; | |
c6bb733d | 1321 | |
8424df5f | 1322 | if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type)) |
1323 | lb = lower_bound_in_type (type, type); | |
1324 | else | |
1325 | lb = TYPE_MIN_VALUE (type); | |
1326 | ||
1327 | if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type)) | |
1328 | ub = upper_bound_in_type (type, type); | |
1329 | else | |
1330 | ub = TYPE_MAX_VALUE (type); | |
c6bb733d | 1331 | |
1332 | if (lb) | |
1333 | { | |
87e20041 | 1334 | isl_space *space = isl_set_get_space (scop->context); |
1335 | isl_constraint *c; | |
1336 | mpz_t g; | |
ea48ac01 | 1337 | isl_val *v; |
87e20041 | 1338 | |
1339 | c = isl_inequality_alloc (isl_local_space_from_space (space)); | |
1340 | mpz_init (g); | |
87e20041 | 1341 | tree_int_to_gmp (lb, g); |
ea48ac01 | 1342 | v = isl_val_int_from_gmp (the_isl_ctx, g); |
1343 | v = isl_val_neg (v); | |
87e20041 | 1344 | mpz_clear (g); |
ea48ac01 | 1345 | c = isl_constraint_set_constant_val (c, v); |
87e20041 | 1346 | c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1); |
1347 | ||
1348 | scop->context = isl_set_add_constraint (scop->context, c); | |
c6bb733d | 1349 | } |
1350 | ||
1351 | if (ub) | |
1352 | { | |
87e20041 | 1353 | isl_space *space = isl_set_get_space (scop->context); |
1354 | isl_constraint *c; | |
1355 | mpz_t g; | |
ea48ac01 | 1356 | isl_val *v; |
87e20041 | 1357 | |
1358 | c = isl_inequality_alloc (isl_local_space_from_space (space)); | |
1359 | ||
1360 | mpz_init (g); | |
87e20041 | 1361 | tree_int_to_gmp (ub, g); |
ea48ac01 | 1362 | v = isl_val_int_from_gmp (the_isl_ctx, g); |
87e20041 | 1363 | mpz_clear (g); |
ea48ac01 | 1364 | c = isl_constraint_set_constant_val (c, v); |
87e20041 | 1365 | c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1); |
1366 | ||
1367 | scop->context = isl_set_add_constraint (scop->context, c); | |
c6bb733d | 1368 | } |
1369 | } | |
1370 | ||
1371 | /* Build the context of the SCOP. The context usually contains extra | |
1372 | constraints that are added to the iteration domains that constrain | |
1373 | some parameters. */ | |
1374 | ||
1375 | static void | |
1376 | build_scop_context (scop_p scop) | |
1377 | { | |
c6bb733d | 1378 | graphite_dim_t p, n = scop_nb_params (scop); |
1379 | ||
c6bb733d | 1380 | for (p = 0; p < n; p++) |
87e20041 | 1381 | add_param_constraints (scop, p); |
c6bb733d | 1382 | } |
1383 | ||
1384 | /* Build the iteration domains: the loops belonging to the current | |
1385 | SCOP, and that vary for the execution of the current basic block. | |
1386 | Returns false if there is no loop in SCOP. */ | |
1387 | ||
1388 | static void | |
1389 | build_scop_iteration_domain (scop_p scop) | |
1390 | { | |
1391 | struct loop *loop; | |
1392 | sese region = SCOP_REGION (scop); | |
1393 | int i; | |
c6bb733d | 1394 | poly_bb_p pbb; |
41f75a99 | 1395 | int nb_loops = number_of_loops (cfun); |
87e20041 | 1396 | isl_set **doms = XCNEWVEC (isl_set *, nb_loops); |
c6bb733d | 1397 | |
f1f41a6c | 1398 | FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop) |
c6bb733d | 1399 | if (!loop_in_sese_p (loop_outer (loop), region)) |
87e20041 | 1400 | build_loop_iteration_domains (scop, loop, 0, |
1401 | isl_set_copy (scop->context), doms); | |
c6bb733d | 1402 | |
f1f41a6c | 1403 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
87e20041 | 1404 | { |
1405 | loop = pbb_loop (pbb); | |
1406 | ||
1407 | if (doms[loop->num]) | |
1408 | pbb->domain = isl_set_copy (doms[loop->num]); | |
1409 | else | |
1410 | pbb->domain = isl_set_copy (scop->context); | |
1411 | ||
1412 | pbb->domain = isl_set_set_tuple_id (pbb->domain, | |
1413 | isl_id_for_pbb (scop, pbb)); | |
1414 | } | |
c6bb733d | 1415 | |
628eaf60 | 1416 | for (i = 0; i < nb_loops; i++) |
87e20041 | 1417 | if (doms[i]) |
1418 | isl_set_free (doms[i]); | |
c6bb733d | 1419 | |
87e20041 | 1420 | free (doms); |
c6bb733d | 1421 | } |
1422 | ||
1423 | /* Add a constrain to the ACCESSES polyhedron for the alias set of | |
1424 | data reference DR. ACCESSP_NB_DIMS is the dimension of the | |
1425 | ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration | |
1426 | domain. */ | |
1427 | ||
87e20041 | 1428 | static isl_map * |
1429 | pdr_add_alias_set (isl_map *acc, data_reference_p dr) | |
c6bb733d | 1430 | { |
87e20041 | 1431 | isl_constraint *c; |
c6bb733d | 1432 | int alias_set_num = 0; |
e9a3f95f | 1433 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); |
c6bb733d | 1434 | |
1a95516e | 1435 | if (bap && bap->alias_set) |
e9a3f95f | 1436 | alias_set_num = *(bap->alias_set); |
c6bb733d | 1437 | |
87e20041 | 1438 | c = isl_equality_alloc |
1439 | (isl_local_space_from_space (isl_map_get_space (acc))); | |
1440 | c = isl_constraint_set_constant_si (c, -alias_set_num); | |
1441 | c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1); | |
1442 | ||
1443 | return isl_map_add_constraint (acc, c); | |
1444 | } | |
1445 | ||
1446 | /* Assign the affine expression INDEX to the output dimension POS of | |
1447 | MAP and return the result. */ | |
1448 | ||
1449 | static isl_map * | |
1450 | set_index (isl_map *map, int pos, isl_pw_aff *index) | |
1451 | { | |
1452 | isl_map *index_map; | |
1453 | int len = isl_map_dim (map, isl_dim_out); | |
1454 | isl_id *id; | |
1455 | ||
1456 | index_map = isl_map_from_pw_aff (index); | |
1457 | index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos); | |
1458 | index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1); | |
c6bb733d | 1459 | |
87e20041 | 1460 | id = isl_map_get_tuple_id (map, isl_dim_out); |
1461 | index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id); | |
1462 | id = isl_map_get_tuple_id (map, isl_dim_in); | |
1463 | index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id); | |
c6bb733d | 1464 | |
87e20041 | 1465 | return isl_map_intersect (map, index_map); |
c6bb733d | 1466 | } |
1467 | ||
1468 | /* Add to ACCESSES polyhedron equalities defining the access functions | |
1469 | to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES | |
1470 | polyhedron, DOM_NB_DIMS is the dimension of the iteration domain. | |
1471 | PBB is the poly_bb_p that contains the data reference DR. */ | |
1472 | ||
87e20041 | 1473 | static isl_map * |
1474 | pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb) | |
c6bb733d | 1475 | { |
1476 | int i, nb_subscripts = DR_NUM_DIMENSIONS (dr); | |
c6bb733d | 1477 | scop_p scop = PBB_SCOP (pbb); |
c6bb733d | 1478 | |
1479 | for (i = 0; i < nb_subscripts; i++) | |
1480 | { | |
87e20041 | 1481 | isl_pw_aff *aff; |
c6bb733d | 1482 | tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i); |
1483 | ||
87e20041 | 1484 | aff = extract_affine (scop, afn, |
1485 | isl_space_domain (isl_map_get_space (acc))); | |
1486 | acc = set_index (acc, i + 1, aff); | |
c6bb733d | 1487 | } |
1488 | ||
87e20041 | 1489 | return acc; |
c6bb733d | 1490 | } |
1491 | ||
1492 | /* Add constrains representing the size of the accessed data to the | |
19b42529 | 1493 | ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the |
1494 | ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration | |
c6bb733d | 1495 | domain. */ |
1496 | ||
87e20041 | 1497 | static isl_set * |
1498 | pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr) | |
c6bb733d | 1499 | { |
1500 | tree ref = DR_REF (dr); | |
1501 | int i, nb_subscripts = DR_NUM_DIMENSIONS (dr); | |
c6bb733d | 1502 | |
249d544d | 1503 | for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0)) |
c6bb733d | 1504 | { |
249d544d | 1505 | tree low, high; |
c6bb733d | 1506 | |
249d544d | 1507 | if (TREE_CODE (ref) != ARRAY_REF) |
c6bb733d | 1508 | break; |
1509 | ||
249d544d | 1510 | low = array_ref_low_bound (ref); |
249d544d | 1511 | high = array_ref_up_bound (ref); |
1512 | ||
87e20041 | 1513 | /* XXX The PPL code dealt separately with |
1514 | subscript - low >= 0 and high - subscript >= 0 in case one of | |
1515 | the two bounds isn't known. Do the same here? */ | |
1516 | ||
e913b5cd | 1517 | if (tree_fits_shwi_p (low) |
87e20041 | 1518 | && high |
e913b5cd | 1519 | && tree_fits_shwi_p (high) |
3354e72e | 1520 | /* 1-element arrays at end of structures may extend over |
1521 | their declared size. */ | |
1522 | && !(array_at_struct_end_p (ref) | |
1523 | && operand_equal_p (low, high, 0))) | |
249d544d | 1524 | { |
87e20041 | 1525 | isl_id *id; |
1526 | isl_aff *aff; | |
1527 | isl_set *univ, *lbs, *ubs; | |
1528 | isl_pw_aff *index; | |
1529 | isl_space *space; | |
1530 | isl_set *valid; | |
1531 | isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent)); | |
1532 | isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent)); | |
1533 | ||
1534 | /* high >= 0 */ | |
1535 | valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub)); | |
1536 | valid = isl_set_project_out (valid, isl_dim_set, 0, | |
1537 | isl_set_dim (valid, isl_dim_set)); | |
1538 | scop->context = isl_set_intersect (scop->context, valid); | |
1539 | ||
1540 | space = isl_set_get_space (extent); | |
1541 | aff = isl_aff_zero_on_domain (isl_local_space_from_space (space)); | |
1542 | aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1); | |
1543 | univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff))); | |
1544 | index = isl_pw_aff_alloc (univ, aff); | |
1545 | ||
1546 | id = isl_set_get_tuple_id (extent); | |
1547 | lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id)); | |
1548 | ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id); | |
1549 | ||
1550 | /* low <= sub_i <= high */ | |
1551 | lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb); | |
1552 | ubs = isl_pw_aff_le_set (index, ub); | |
1553 | extent = isl_set_intersect (extent, lbs); | |
1554 | extent = isl_set_intersect (extent, ubs); | |
249d544d | 1555 | } |
c6bb733d | 1556 | } |
87e20041 | 1557 | |
1558 | return extent; | |
c6bb733d | 1559 | } |
1560 | ||
1561 | /* Build data accesses for DR in PBB. */ | |
1562 | ||
1563 | static void | |
1564 | build_poly_dr (data_reference_p dr, poly_bb_p pbb) | |
1565 | { | |
ae11f03b | 1566 | int dr_base_object_set; |
87e20041 | 1567 | isl_map *acc; |
1568 | isl_set *extent; | |
1569 | scop_p scop = PBB_SCOP (pbb); | |
c6bb733d | 1570 | |
87e20041 | 1571 | { |
1572 | isl_space *dc = isl_set_get_space (pbb->domain); | |
1573 | int nb_out = 1 + DR_NUM_DIMENSIONS (dr); | |
1574 | isl_space *space = isl_space_add_dims (isl_space_from_domain (dc), | |
1575 | isl_dim_out, nb_out); | |
c6bb733d | 1576 | |
87e20041 | 1577 | acc = isl_map_universe (space); |
1578 | acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr)); | |
1579 | } | |
c6bb733d | 1580 | |
87e20041 | 1581 | acc = pdr_add_alias_set (acc, dr); |
1582 | acc = pdr_add_memory_accesses (acc, dr, pbb); | |
c6bb733d | 1583 | |
87e20041 | 1584 | { |
1585 | isl_id *id = isl_id_for_dr (scop, dr); | |
1586 | int nb = 1 + DR_NUM_DIMENSIONS (dr); | |
1587 | isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb); | |
1588 | int alias_set_num = 0; | |
1589 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); | |
1590 | ||
1591 | if (bap && bap->alias_set) | |
1592 | alias_set_num = *(bap->alias_set); | |
1593 | ||
1594 | space = isl_space_set_tuple_id (space, isl_dim_set, id); | |
1595 | extent = isl_set_nat_universe (space); | |
1596 | extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num); | |
1597 | extent = pdr_add_data_dimensions (extent, scop, dr); | |
1598 | } | |
c6bb733d | 1599 | |
3ebac17a | 1600 | gcc_assert (dr->aux); |
1601 | dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set; | |
ae11f03b | 1602 | |
87e20041 | 1603 | new_poly_dr (pbb, dr_base_object_set, |
3ebac17a | 1604 | DR_IS_READ (dr) ? PDR_READ : PDR_WRITE, |
87e20041 | 1605 | dr, DR_NUM_DIMENSIONS (dr), acc, extent); |
ae11f03b | 1606 | } |
c6bb733d | 1607 | |
96233858 | 1608 | /* Write to FILE the alias graph of data references in DIMACS format. */ |
8dee37ec | 1609 | |
1610 | static inline bool | |
1611 | write_alias_graph_to_ascii_dimacs (FILE *file, char *comment, | |
f1f41a6c | 1612 | vec<data_reference_p> drs) |
8dee37ec | 1613 | { |
f1f41a6c | 1614 | int num_vertex = drs.length (); |
8dee37ec | 1615 | int edge_num = 0; |
1616 | data_reference_p dr1, dr2; | |
1617 | int i, j; | |
1618 | ||
1619 | if (num_vertex == 0) | |
1620 | return true; | |
1621 | ||
f1f41a6c | 1622 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1623 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
5fc88ffd | 1624 | if (dr_may_alias_p (dr1, dr2, true)) |
8dee37ec | 1625 | edge_num++; |
1626 | ||
1627 | fprintf (file, "$\n"); | |
1628 | ||
1629 | if (comment) | |
1630 | fprintf (file, "c %s\n", comment); | |
1631 | ||
1632 | fprintf (file, "p edge %d %d\n", num_vertex, edge_num); | |
1633 | ||
f1f41a6c | 1634 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1635 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
5fc88ffd | 1636 | if (dr_may_alias_p (dr1, dr2, true)) |
8dee37ec | 1637 | fprintf (file, "e %d %d\n", i + 1, j + 1); |
1638 | ||
1639 | return true; | |
1640 | } | |
1641 | ||
96233858 | 1642 | /* Write to FILE the alias graph of data references in DOT format. */ |
1643 | ||
1644 | static inline bool | |
1645 | write_alias_graph_to_ascii_dot (FILE *file, char *comment, | |
f1f41a6c | 1646 | vec<data_reference_p> drs) |
96233858 | 1647 | { |
f1f41a6c | 1648 | int num_vertex = drs.length (); |
96233858 | 1649 | data_reference_p dr1, dr2; |
1650 | int i, j; | |
1651 | ||
1652 | if (num_vertex == 0) | |
1653 | return true; | |
1654 | ||
1655 | fprintf (file, "$\n"); | |
1656 | ||
1657 | if (comment) | |
1658 | fprintf (file, "c %s\n", comment); | |
1659 | ||
1660 | /* First print all the vertices. */ | |
f1f41a6c | 1661 | FOR_EACH_VEC_ELT (drs, i, dr1) |
96233858 | 1662 | fprintf (file, "n%d;\n", i); |
1663 | ||
f1f41a6c | 1664 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1665 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
5fc88ffd | 1666 | if (dr_may_alias_p (dr1, dr2, true)) |
96233858 | 1667 | fprintf (file, "n%d n%d\n", i, j); |
1668 | ||
1669 | return true; | |
1670 | } | |
1671 | ||
1672 | /* Write to FILE the alias graph of data references in ECC format. */ | |
1673 | ||
1674 | static inline bool | |
1675 | write_alias_graph_to_ascii_ecc (FILE *file, char *comment, | |
f1f41a6c | 1676 | vec<data_reference_p> drs) |
96233858 | 1677 | { |
f1f41a6c | 1678 | int num_vertex = drs.length (); |
96233858 | 1679 | data_reference_p dr1, dr2; |
1680 | int i, j; | |
1681 | ||
1682 | if (num_vertex == 0) | |
1683 | return true; | |
1684 | ||
1685 | fprintf (file, "$\n"); | |
1686 | ||
1687 | if (comment) | |
1688 | fprintf (file, "c %s\n", comment); | |
1689 | ||
f1f41a6c | 1690 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1691 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
5fc88ffd | 1692 | if (dr_may_alias_p (dr1, dr2, true)) |
96233858 | 1693 | fprintf (file, "%d %d\n", i, j); |
1694 | ||
1695 | return true; | |
1696 | } | |
1697 | ||
e9a3f95f | 1698 | /* Check if DR1 and DR2 are in the same object set. */ |
1699 | ||
1700 | static bool | |
1701 | dr_same_base_object_p (const struct data_reference *dr1, | |
1702 | const struct data_reference *dr2) | |
1703 | { | |
1704 | return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0); | |
1705 | } | |
96233858 | 1706 | |
1707 | /* Uses DFS component number as representative of alias-sets. Also tests for | |
1708 | optimality by verifying if every connected component is a clique. Returns | |
1709 | true (1) if the above test is true, and false (0) otherwise. */ | |
1710 | ||
1711 | static int | |
f1f41a6c | 1712 | build_alias_set_optimal_p (vec<data_reference_p> drs) |
c6bb733d | 1713 | { |
f1f41a6c | 1714 | int num_vertices = drs.length (); |
96233858 | 1715 | struct graph *g = new_graph (num_vertices); |
c6bb733d | 1716 | data_reference_p dr1, dr2; |
1717 | int i, j; | |
96233858 | 1718 | int num_connected_components; |
1719 | int v_indx1, v_indx2, num_vertices_in_component; | |
1720 | int *all_vertices; | |
1721 | int *vertices; | |
1722 | struct graph_edge *e; | |
f289f81b | 1723 | int this_component_is_clique; |
1724 | int all_components_are_cliques = 1; | |
c6bb733d | 1725 | |
f1f41a6c | 1726 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1727 | for (j = i+1; drs.iterate (j, &dr2); j++) | |
5fc88ffd | 1728 | if (dr_may_alias_p (dr1, dr2, true)) |
c6bb733d | 1729 | { |
1730 | add_edge (g, i, j); | |
1731 | add_edge (g, j, i); | |
1732 | } | |
1733 | ||
96233858 | 1734 | all_vertices = XNEWVEC (int, num_vertices); |
1735 | vertices = XNEWVEC (int, num_vertices); | |
1736 | for (i = 0; i < num_vertices; i++) | |
1737 | all_vertices[i] = i; | |
1738 | ||
e9a3f95f | 1739 | num_connected_components = graphds_dfs (g, all_vertices, num_vertices, |
1740 | NULL, true, NULL); | |
1741 | for (i = 0; i < g->n_vertices; i++) | |
1742 | { | |
f1f41a6c | 1743 | data_reference_p dr = drs[i]; |
e9a3f95f | 1744 | base_alias_pair *bap; |
1a95516e | 1745 | |
3ebac17a | 1746 | gcc_assert (dr->aux); |
1747 | bap = (base_alias_pair *)(dr->aux); | |
1a95516e | 1748 | |
e9a3f95f | 1749 | bap->alias_set = XNEW (int); |
1750 | *(bap->alias_set) = g->vertices[i].component + 1; | |
1751 | } | |
1752 | ||
96233858 | 1753 | /* Verify if the DFS numbering results in optimal solution. */ |
1754 | for (i = 0; i < num_connected_components; i++) | |
1755 | { | |
1756 | num_vertices_in_component = 0; | |
1757 | /* Get all vertices whose DFS component number is the same as i. */ | |
1758 | for (j = 0; j < num_vertices; j++) | |
1759 | if (g->vertices[j].component == i) | |
1760 | vertices[num_vertices_in_component++] = j; | |
1761 | ||
1762 | /* Now test if the vertices in 'vertices' form a clique, by testing | |
1763 | for edges among each pair. */ | |
1764 | this_component_is_clique = 1; | |
1765 | for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++) | |
1766 | { | |
1767 | for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++) | |
1768 | { | |
1769 | /* Check if the two vertices are connected by iterating | |
1770 | through all the edges which have one of these are source. */ | |
1771 | e = g->vertices[vertices[v_indx2]].pred; | |
1772 | while (e) | |
1773 | { | |
1774 | if (e->src == vertices[v_indx1]) | |
1775 | break; | |
1776 | e = e->pred_next; | |
1777 | } | |
1778 | if (!e) | |
1779 | { | |
1780 | this_component_is_clique = 0; | |
1781 | break; | |
1782 | } | |
1783 | } | |
1784 | if (!this_component_is_clique) | |
1785 | all_components_are_cliques = 0; | |
1786 | } | |
1787 | } | |
c6bb733d | 1788 | |
96233858 | 1789 | free (all_vertices); |
1790 | free (vertices); | |
c6bb733d | 1791 | free_graph (g); |
96233858 | 1792 | return all_components_are_cliques; |
c6bb733d | 1793 | } |
1794 | ||
8c6b3774 | 1795 | /* Group each data reference in DRS with its base object set num. */ |
ae11f03b | 1796 | |
1797 | static void | |
f1f41a6c | 1798 | build_base_obj_set_for_drs (vec<data_reference_p> drs) |
ae11f03b | 1799 | { |
f1f41a6c | 1800 | int num_vertex = drs.length (); |
e9a3f95f | 1801 | struct graph *g = new_graph (num_vertex); |
1802 | data_reference_p dr1, dr2; | |
1803 | int i, j; | |
e9a3f95f | 1804 | int *queue; |
1805 | ||
f1f41a6c | 1806 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1807 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
e9a3f95f | 1808 | if (dr_same_base_object_p (dr1, dr2)) |
1809 | { | |
1810 | add_edge (g, i, j); | |
1811 | add_edge (g, j, i); | |
1812 | } | |
1813 | ||
1814 | queue = XNEWVEC (int, num_vertex); | |
1815 | for (i = 0; i < num_vertex; i++) | |
1816 | queue[i] = i; | |
1817 | ||
1a95516e | 1818 | graphds_dfs (g, queue, num_vertex, NULL, true, NULL); |
e9a3f95f | 1819 | |
1820 | for (i = 0; i < g->n_vertices; i++) | |
1821 | { | |
f1f41a6c | 1822 | data_reference_p dr = drs[i]; |
e9a3f95f | 1823 | base_alias_pair *bap; |
1a95516e | 1824 | |
3ebac17a | 1825 | gcc_assert (dr->aux); |
1826 | bap = (base_alias_pair *)(dr->aux); | |
1a95516e | 1827 | |
e9a3f95f | 1828 | bap->base_obj_set = g->vertices[i].component + 1; |
1829 | } | |
1830 | ||
1831 | free (queue); | |
1832 | free_graph (g); | |
ae11f03b | 1833 | } |
1834 | ||
c6bb733d | 1835 | /* Build the data references for PBB. */ |
1836 | ||
1837 | static void | |
1838 | build_pbb_drs (poly_bb_p pbb) | |
1839 | { | |
1840 | int j; | |
1841 | data_reference_p dr; | |
f1f41a6c | 1842 | vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb)); |
c6bb733d | 1843 | |
f1f41a6c | 1844 | FOR_EACH_VEC_ELT (gbb_drs, j, dr) |
c6bb733d | 1845 | build_poly_dr (dr, pbb); |
1846 | } | |
1847 | ||
aa16918e | 1848 | /* Dump to file the alias graphs for the data references in DRS. */ |
1849 | ||
1850 | static void | |
f1f41a6c | 1851 | dump_alias_graphs (vec<data_reference_p> drs) |
aa16918e | 1852 | { |
1853 | char comment[100]; | |
1854 | FILE *file_dimacs, *file_ecc, *file_dot; | |
1855 | ||
1856 | file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab"); | |
1857 | if (file_dimacs) | |
1858 | { | |
1859 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1860 | current_function_name ()); | |
1861 | write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs); | |
1862 | fclose (file_dimacs); | |
1863 | } | |
1864 | ||
1865 | file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab"); | |
1866 | if (file_ecc) | |
1867 | { | |
1868 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1869 | current_function_name ()); | |
1870 | write_alias_graph_to_ascii_ecc (file_ecc, comment, drs); | |
1871 | fclose (file_ecc); | |
1872 | } | |
1873 | ||
1874 | file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab"); | |
1875 | if (file_dot) | |
1876 | { | |
1877 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1878 | current_function_name ()); | |
1879 | write_alias_graph_to_ascii_dot (file_dot, comment, drs); | |
1880 | fclose (file_dot); | |
1881 | } | |
1882 | } | |
1883 | ||
c6bb733d | 1884 | /* Build data references in SCOP. */ |
1885 | ||
1886 | static void | |
1887 | build_scop_drs (scop_p scop) | |
1888 | { | |
da6ec3dd | 1889 | int i, j; |
c6bb733d | 1890 | poly_bb_p pbb; |
da6ec3dd | 1891 | data_reference_p dr; |
4997014d | 1892 | auto_vec<data_reference_p, 3> drs; |
da6ec3dd | 1893 | |
8c6b3774 | 1894 | /* Remove all the PBBs that do not have data references: these basic |
1895 | blocks are not handled in the polyhedral representation. */ | |
f1f41a6c | 1896 | for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++) |
1897 | if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ()) | |
1688f172 | 1898 | { |
bf0d0e76 | 1899 | free_gimple_bb (PBB_BLACK_BOX (pbb)); |
479a6d79 | 1900 | free_poly_bb (pbb); |
f1f41a6c | 1901 | SCOP_BBS (scop).ordered_remove (i); |
1688f172 | 1902 | i--; |
1903 | } | |
8c6b3774 | 1904 | |
f1f41a6c | 1905 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
1906 | for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++) | |
1907 | drs.safe_push (dr); | |
da6ec3dd | 1908 | |
f1f41a6c | 1909 | FOR_EACH_VEC_ELT (drs, i, dr) |
e9a3f95f | 1910 | dr->aux = XNEW (base_alias_pair); |
1911 | ||
1912 | if (!build_alias_set_optimal_p (drs)) | |
1913 | { | |
1914 | /* TODO: Add support when building alias set is not optimal. */ | |
1915 | ; | |
1916 | } | |
1917 | ||
9c48819a | 1918 | build_base_obj_set_for_drs (drs); |
ae11f03b | 1919 | |
8dee37ec | 1920 | /* When debugging, enable the following code. This cannot be used |
1921 | in production compilers. */ | |
aa16918e | 1922 | if (0) |
1923 | dump_alias_graphs (drs); | |
8dee37ec | 1924 | |
f1f41a6c | 1925 | drs.release (); |
c6bb733d | 1926 | |
f1f41a6c | 1927 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
c6bb733d | 1928 | build_pbb_drs (pbb); |
1929 | } | |
1930 | ||
30f4f4a6 | 1931 | /* Return a gsi at the position of the phi node STMT. */ |
1932 | ||
1933 | static gimple_stmt_iterator | |
1934 | gsi_for_phi_node (gimple stmt) | |
1935 | { | |
1936 | gimple_stmt_iterator psi; | |
1937 | basic_block bb = gimple_bb (stmt); | |
1938 | ||
1939 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1940 | if (stmt == gsi_stmt (psi)) | |
1941 | return psi; | |
1942 | ||
1943 | gcc_unreachable (); | |
1944 | return psi; | |
1945 | } | |
1946 | ||
1688f172 | 1947 | /* Analyze all the data references of STMTS and add them to the |
1948 | GBB_DATA_REFS vector of BB. */ | |
1949 | ||
1950 | static void | |
f1f41a6c | 1951 | analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts) |
1688f172 | 1952 | { |
1953 | loop_p nest; | |
1688f172 | 1954 | gimple_bb_p gbb; |
1955 | gimple stmt; | |
1956 | int i; | |
221a697e | 1957 | sese region = SCOP_REGION (scop); |
1688f172 | 1958 | |
221a697e | 1959 | if (!bb_in_sese_p (bb, region)) |
1688f172 | 1960 | return; |
1961 | ||
221a697e | 1962 | nest = outermost_loop_in_sese_1 (region, bb); |
1688f172 | 1963 | gbb = gbb_from_bb (bb); |
1964 | ||
f1f41a6c | 1965 | FOR_EACH_VEC_ELT (stmts, i, stmt) |
221a697e | 1966 | { |
1967 | loop_p loop; | |
1968 | ||
1969 | if (is_gimple_debug (stmt)) | |
1970 | continue; | |
1971 | ||
1972 | loop = loop_containing_stmt (stmt); | |
1973 | if (!loop_in_sese_p (loop, region)) | |
1974 | loop = nest; | |
1975 | ||
1976 | graphite_find_data_references_in_stmt (nest, loop, stmt, | |
1688f172 | 1977 | &GBB_DATA_REFS (gbb)); |
221a697e | 1978 | } |
1688f172 | 1979 | } |
1980 | ||
1981 | /* Insert STMT at the end of the STMTS sequence and then insert the | |
1982 | statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts | |
1983 | on STMTS. */ | |
1984 | ||
1985 | static void | |
1986 | insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts, | |
1987 | gimple_stmt_iterator insert_gsi) | |
1988 | { | |
1989 | gimple_stmt_iterator gsi; | |
4997014d | 1990 | auto_vec<gimple, 3> x; |
1688f172 | 1991 | |
e3a19533 | 1992 | gimple_seq_add_stmt (&stmts, stmt); |
1688f172 | 1993 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
f1f41a6c | 1994 | x.safe_push (gsi_stmt (gsi)); |
1688f172 | 1995 | |
1996 | gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT); | |
1997 | analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x); | |
1688f172 | 1998 | } |
1999 | ||
8c6b3774 | 2000 | /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */ |
c6bb733d | 2001 | |
2002 | static void | |
1688f172 | 2003 | insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt) |
c6bb733d | 2004 | { |
c6bb733d | 2005 | gimple_seq stmts; |
52123e43 | 2006 | gimple_stmt_iterator gsi; |
8c6b3774 | 2007 | tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE); |
11db727d | 2008 | gimple stmt = gimple_build_assign (unshare_expr (res), var); |
4997014d | 2009 | auto_vec<gimple, 3> x; |
c6bb733d | 2010 | |
e3a19533 | 2011 | gimple_seq_add_stmt (&stmts, stmt); |
1688f172 | 2012 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
f1f41a6c | 2013 | x.safe_push (gsi_stmt (gsi)); |
52123e43 | 2014 | |
39a34dd8 | 2015 | if (gimple_code (after_stmt) == GIMPLE_PHI) |
52123e43 | 2016 | { |
39a34dd8 | 2017 | gsi = gsi_after_labels (gimple_bb (after_stmt)); |
52123e43 | 2018 | gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT); |
2019 | } | |
2020 | else | |
2021 | { | |
39a34dd8 | 2022 | gsi = gsi_for_stmt (after_stmt); |
52123e43 | 2023 | gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); |
2024 | } | |
1688f172 | 2025 | |
2026 | analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x); | |
c6bb733d | 2027 | } |
2028 | ||
8c6b3774 | 2029 | /* Creates a poly_bb_p for basic_block BB from the existing PBB. */ |
2030 | ||
2031 | static void | |
2032 | new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb) | |
2033 | { | |
f1f41a6c | 2034 | vec<data_reference_p> drs; |
2035 | drs.create (3); | |
8c6b3774 | 2036 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); |
2037 | gimple_bb_p gbb1 = new_gimple_bb (bb, drs); | |
2038 | poly_bb_p pbb1 = new_poly_bb (scop, gbb1); | |
f1f41a6c | 2039 | int index, n = SCOP_BBS (scop).length (); |
8c6b3774 | 2040 | |
2041 | /* The INDEX of PBB in SCOP_BBS. */ | |
2042 | for (index = 0; index < n; index++) | |
f1f41a6c | 2043 | if (SCOP_BBS (scop)[index] == pbb) |
8c6b3774 | 2044 | break; |
2045 | ||
87e20041 | 2046 | pbb1->domain = isl_set_copy (pbb->domain); |
24bc4589 | 2047 | pbb1->domain = isl_set_set_tuple_id (pbb1->domain, |
2048 | isl_id_for_pbb (scop, pbb1)); | |
93f9c161 | 2049 | |
8c6b3774 | 2050 | GBB_PBB (gbb1) = pbb1; |
f1f41a6c | 2051 | GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy (); |
2052 | GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy (); | |
2053 | SCOP_BBS (scop).safe_insert (index + 1, pbb1); | |
8c6b3774 | 2054 | } |
2055 | ||
c6bb733d | 2056 | /* Insert on edge E the assignment "RES := EXPR". */ |
2057 | ||
2058 | static void | |
8c6b3774 | 2059 | insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr) |
c6bb733d | 2060 | { |
2061 | gimple_stmt_iterator gsi; | |
e3a19533 | 2062 | gimple_seq stmts = NULL; |
c6bb733d | 2063 | tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE); |
11db727d | 2064 | gimple stmt = gimple_build_assign (unshare_expr (res), var); |
8c6b3774 | 2065 | basic_block bb; |
4997014d | 2066 | auto_vec<gimple, 3> x; |
c6bb733d | 2067 | |
e3a19533 | 2068 | gimple_seq_add_stmt (&stmts, stmt); |
1688f172 | 2069 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
f1f41a6c | 2070 | x.safe_push (gsi_stmt (gsi)); |
1688f172 | 2071 | |
c6bb733d | 2072 | gsi_insert_seq_on_edge (e, stmts); |
2073 | gsi_commit_edge_inserts (); | |
8c6b3774 | 2074 | bb = gimple_bb (stmt); |
2075 | ||
2076 | if (!bb_in_sese_p (bb, SCOP_REGION (scop))) | |
2077 | return; | |
2078 | ||
2079 | if (!gbb_from_bb (bb)) | |
2080 | new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb); | |
1688f172 | 2081 | |
2082 | analyze_drs_in_stmts (scop, bb, x); | |
c6bb733d | 2083 | } |
2084 | ||
2085 | /* Creates a zero dimension array of the same type as VAR. */ | |
2086 | ||
2087 | static tree | |
03c1327b | 2088 | create_zero_dim_array (tree var, const char *base_name) |
c6bb733d | 2089 | { |
2090 | tree index_type = build_index_type (integer_zero_node); | |
2091 | tree elt_type = TREE_TYPE (var); | |
2092 | tree array_type = build_array_type (elt_type, index_type); | |
03c1327b | 2093 | tree base = create_tmp_var (array_type, base_name); |
c6bb733d | 2094 | |
c6bb733d | 2095 | return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE, |
2096 | NULL_TREE); | |
2097 | } | |
2098 | ||
2099 | /* Returns true when PHI is a loop close phi node. */ | |
2100 | ||
2101 | static bool | |
2102 | scalar_close_phi_node_p (gimple phi) | |
2103 | { | |
30f4f4a6 | 2104 | if (gimple_code (phi) != GIMPLE_PHI |
7c782c9b | 2105 | || virtual_operand_p (gimple_phi_result (phi))) |
c6bb733d | 2106 | return false; |
2107 | ||
03ce78db | 2108 | /* Note that loop close phi nodes should have a single argument |
2109 | because we translated the representation into a canonical form | |
2110 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
c6bb733d | 2111 | return (gimple_phi_num_args (phi) == 1); |
2112 | } | |
2113 | ||
5d21c24a | 2114 | /* For a definition DEF in REGION, propagates the expression EXPR in |
2115 | all the uses of DEF outside REGION. */ | |
2116 | ||
2117 | static void | |
2118 | propagate_expr_outside_region (tree def, tree expr, sese region) | |
2119 | { | |
2120 | imm_use_iterator imm_iter; | |
2121 | gimple use_stmt; | |
2122 | gimple_seq stmts; | |
2123 | bool replaced_once = false; | |
2124 | ||
ed455480 | 2125 | gcc_assert (TREE_CODE (def) == SSA_NAME); |
5d21c24a | 2126 | |
2127 | expr = force_gimple_operand (unshare_expr (expr), &stmts, true, | |
2128 | NULL_TREE); | |
2129 | ||
2130 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2131 | if (!is_gimple_debug (use_stmt) | |
2132 | && !bb_in_sese_p (gimple_bb (use_stmt), region)) | |
2133 | { | |
2134 | ssa_op_iter iter; | |
2135 | use_operand_p use_p; | |
2136 | ||
2137 | FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES) | |
2138 | if (operand_equal_p (def, USE_FROM_PTR (use_p), 0) | |
2139 | && (replaced_once = true)) | |
2140 | replace_exp (use_p, expr); | |
2141 | ||
2142 | update_stmt (use_stmt); | |
2143 | } | |
2144 | ||
2145 | if (replaced_once) | |
2146 | { | |
2147 | gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts); | |
2148 | gsi_commit_edge_inserts (); | |
2149 | } | |
2150 | } | |
2151 | ||
c6bb733d | 2152 | /* Rewrite out of SSA the reduction phi node at PSI by creating a zero |
2153 | dimension array for it. */ | |
2154 | ||
2155 | static void | |
8c6b3774 | 2156 | rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi) |
c6bb733d | 2157 | { |
8c6b3774 | 2158 | sese region = SCOP_REGION (scop); |
c6bb733d | 2159 | gimple phi = gsi_stmt (*psi); |
2160 | tree res = gimple_phi_result (phi); | |
55c89f69 | 2161 | basic_block bb = gimple_bb (phi); |
2162 | gimple_stmt_iterator gsi = gsi_after_labels (bb); | |
c6bb733d | 2163 | tree arg = gimple_phi_arg_def (phi, 0); |
55c89f69 | 2164 | gimple stmt; |
c6bb733d | 2165 | |
03ce78db | 2166 | /* Note that loop close phi nodes should have a single argument |
2167 | because we translated the representation into a canonical form | |
2168 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
2169 | gcc_assert (gimple_phi_num_args (phi) == 1); | |
2170 | ||
55c89f69 | 2171 | /* The phi node can be a non close phi node, when its argument is |
51ec8951 | 2172 | invariant, or a default definition. */ |
55c89f69 | 2173 | if (is_gimple_min_invariant (arg) |
51ec8951 | 2174 | || SSA_NAME_IS_DEFAULT_DEF (arg)) |
ed455480 | 2175 | { |
2176 | propagate_expr_outside_region (res, arg, region); | |
2177 | gsi_next (psi); | |
2178 | return; | |
2179 | } | |
5d21c24a | 2180 | |
68a6a8ba | 2181 | else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father) |
2182 | { | |
2183 | propagate_expr_outside_region (res, arg, region); | |
2184 | stmt = gimple_build_assign (res, arg); | |
2185 | remove_phi_node (psi, false); | |
2186 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
68a6a8ba | 2187 | return; |
2188 | } | |
2189 | ||
5d21c24a | 2190 | /* If res is scev analyzable and is not a scalar value, it is safe |
2191 | to ignore the close phi node: it will be code generated in the | |
2192 | out of Graphite pass. */ | |
2193 | else if (scev_analyzable_p (res, region)) | |
2194 | { | |
2195 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res)); | |
2196 | tree scev; | |
2197 | ||
2198 | if (!loop_in_sese_p (loop, region)) | |
2199 | { | |
2200 | loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg)); | |
2201 | scev = scalar_evolution_in_region (region, loop, arg); | |
2202 | scev = compute_overall_effect_of_inner_loop (loop, scev); | |
2203 | } | |
2204 | else | |
ed455480 | 2205 | scev = scalar_evolution_in_region (region, loop, res); |
5d21c24a | 2206 | |
2207 | if (tree_does_not_contain_chrecs (scev)) | |
2208 | propagate_expr_outside_region (res, scev, region); | |
2209 | ||
2210 | gsi_next (psi); | |
2211 | return; | |
2212 | } | |
5184a05f | 2213 | else |
55c89f69 | 2214 | { |
874117c8 | 2215 | tree zero_dim_array = create_zero_dim_array (res, "Close_Phi"); |
55c89f69 | 2216 | |
11db727d | 2217 | stmt = gimple_build_assign (res, unshare_expr (zero_dim_array)); |
55c89f69 | 2218 | |
49253930 | 2219 | if (TREE_CODE (arg) == SSA_NAME) |
1688f172 | 2220 | insert_out_of_ssa_copy (scop, zero_dim_array, arg, |
8c6b3774 | 2221 | SSA_NAME_DEF_STMT (arg)); |
55c89f69 | 2222 | else |
8c6b3774 | 2223 | insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb), |
55c89f69 | 2224 | zero_dim_array, arg); |
2225 | } | |
c6bb733d | 2226 | |
2227 | remove_phi_node (psi, false); | |
c6bb733d | 2228 | SSA_NAME_DEF_STMT (res) = stmt; |
1688f172 | 2229 | |
2230 | insert_stmts (scop, stmt, NULL, gsi_after_labels (bb)); | |
c6bb733d | 2231 | } |
2232 | ||
2233 | /* Rewrite out of SSA the reduction phi node at PSI by creating a zero | |
2234 | dimension array for it. */ | |
2235 | ||
2236 | static void | |
8c6b3774 | 2237 | rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi) |
c6bb733d | 2238 | { |
2239 | size_t i; | |
2240 | gimple phi = gsi_stmt (*psi); | |
2241 | basic_block bb = gimple_bb (phi); | |
2242 | tree res = gimple_phi_result (phi); | |
874117c8 | 2243 | tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa"); |
c6bb733d | 2244 | gimple stmt; |
c6bb733d | 2245 | |
2246 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2247 | { | |
2248 | tree arg = gimple_phi_arg_def (phi, i); | |
097e870c | 2249 | edge e = gimple_phi_arg_edge (phi, i); |
c6bb733d | 2250 | |
097e870c | 2251 | /* Avoid the insertion of code in the loop latch to please the |
2252 | pattern matching of the vectorizer. */ | |
a865acd7 | 2253 | if (TREE_CODE (arg) == SSA_NAME |
f5f9fa15 | 2254 | && !SSA_NAME_IS_DEFAULT_DEF (arg) |
a865acd7 | 2255 | && e->src == bb->loop_father->latch) |
1688f172 | 2256 | insert_out_of_ssa_copy (scop, zero_dim_array, arg, |
8c6b3774 | 2257 | SSA_NAME_DEF_STMT (arg)); |
c6bb733d | 2258 | else |
8c6b3774 | 2259 | insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg); |
c6bb733d | 2260 | } |
2261 | ||
11db727d | 2262 | stmt = gimple_build_assign (res, unshare_expr (zero_dim_array)); |
c6bb733d | 2263 | remove_phi_node (psi, false); |
11db727d | 2264 | insert_stmts (scop, stmt, NULL, gsi_after_labels (bb)); |
c6bb733d | 2265 | } |
2266 | ||
b8046a12 | 2267 | /* Rewrite the degenerate phi node at position PSI from the degenerate |
2268 | form "x = phi (y, y, ..., y)" to "x = y". */ | |
2269 | ||
2270 | static void | |
2271 | rewrite_degenerate_phi (gimple_stmt_iterator *psi) | |
2272 | { | |
2273 | tree rhs; | |
2274 | gimple stmt; | |
2275 | gimple_stmt_iterator gsi; | |
2276 | gimple phi = gsi_stmt (*psi); | |
2277 | tree res = gimple_phi_result (phi); | |
2278 | basic_block bb; | |
2279 | ||
b8046a12 | 2280 | bb = gimple_bb (phi); |
2281 | rhs = degenerate_phi_result (phi); | |
2282 | gcc_assert (rhs); | |
2283 | ||
2284 | stmt = gimple_build_assign (res, rhs); | |
2285 | remove_phi_node (psi, false); | |
b8046a12 | 2286 | |
2287 | gsi = gsi_after_labels (bb); | |
2288 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
2289 | } | |
2290 | ||
42ad5d61 | 2291 | /* Rewrite out of SSA all the reduction phi nodes of SCOP. */ |
2292 | ||
8c6b3774 | 2293 | static void |
42ad5d61 | 2294 | rewrite_reductions_out_of_ssa (scop_p scop) |
2295 | { | |
2296 | basic_block bb; | |
2297 | gimple_stmt_iterator psi; | |
2298 | sese region = SCOP_REGION (scop); | |
2299 | ||
fc00614f | 2300 | FOR_EACH_BB_FN (bb, cfun) |
42ad5d61 | 2301 | if (bb_in_sese_p (bb, region)) |
2302 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi);) | |
2303 | { | |
b8046a12 | 2304 | gimple phi = gsi_stmt (psi); |
2305 | ||
7c782c9b | 2306 | if (virtual_operand_p (gimple_phi_result (phi))) |
7ba04629 | 2307 | { |
2308 | gsi_next (&psi); | |
2309 | continue; | |
2310 | } | |
2311 | ||
b8046a12 | 2312 | if (gimple_phi_num_args (phi) > 1 |
2313 | && degenerate_phi_result (phi)) | |
2314 | rewrite_degenerate_phi (&psi); | |
2315 | ||
2316 | else if (scalar_close_phi_node_p (phi)) | |
8c6b3774 | 2317 | rewrite_close_phi_out_of_ssa (scop, &psi); |
b8046a12 | 2318 | |
42ad5d61 | 2319 | else if (reduction_phi_p (region, &psi)) |
8c6b3774 | 2320 | rewrite_phi_out_of_ssa (scop, &psi); |
42ad5d61 | 2321 | } |
2322 | ||
2323 | update_ssa (TODO_update_ssa); | |
2324 | #ifdef ENABLE_CHECKING | |
2325 | verify_loop_closed_ssa (true); | |
2326 | #endif | |
2327 | } | |
2328 | ||
5061777c | 2329 | /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory |
2330 | read from ZERO_DIM_ARRAY. */ | |
2331 | ||
2332 | static void | |
1688f172 | 2333 | rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array, |
8c6b3774 | 2334 | tree def, gimple use_stmt) |
5061777c | 2335 | { |
874117c8 | 2336 | gimple name_stmt; |
2337 | tree name; | |
5061777c | 2338 | ssa_op_iter iter; |
2339 | use_operand_p use_p; | |
5061777c | 2340 | |
dd27e8aa | 2341 | gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI); |
5061777c | 2342 | |
874117c8 | 2343 | name = copy_ssa_name (def, NULL); |
2344 | name_stmt = gimple_build_assign (name, zero_dim_array); | |
2345 | ||
dd27e8aa | 2346 | gimple_assign_set_lhs (name_stmt, name); |
1688f172 | 2347 | insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt)); |
5061777c | 2348 | |
dd27e8aa | 2349 | FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES) |
2350 | if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)) | |
2351 | replace_exp (use_p, name); | |
5061777c | 2352 | |
2353 | update_stmt (use_stmt); | |
2354 | } | |
2355 | ||
3ac4f821 | 2356 | /* For every definition DEF in the SCOP that is used outside the scop, |
2357 | insert a closing-scop definition in the basic block just after this | |
2358 | SCOP. */ | |
2359 | ||
2360 | static void | |
2361 | handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt) | |
2362 | { | |
2363 | tree var = create_tmp_reg (TREE_TYPE (def), NULL); | |
2364 | tree new_name = make_ssa_name (var, stmt); | |
2365 | bool needs_copy = false; | |
2366 | use_operand_p use_p; | |
2367 | imm_use_iterator imm_iter; | |
2368 | gimple use_stmt; | |
2369 | sese region = SCOP_REGION (scop); | |
2370 | ||
2371 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2372 | { | |
2373 | if (!bb_in_sese_p (gimple_bb (use_stmt), region)) | |
2374 | { | |
2375 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
2376 | { | |
2377 | SET_USE (use_p, new_name); | |
2378 | } | |
2379 | update_stmt (use_stmt); | |
2380 | needs_copy = true; | |
2381 | } | |
2382 | } | |
2383 | ||
2384 | /* Insert in the empty BB just after the scop a use of DEF such | |
2385 | that the rewrite of cross_bb_scalar_dependences won't insert | |
2386 | arrays everywhere else. */ | |
2387 | if (needs_copy) | |
2388 | { | |
2389 | gimple assign = gimple_build_assign (new_name, def); | |
2390 | gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest); | |
2391 | ||
3ac4f821 | 2392 | update_stmt (assign); |
2393 | gsi_insert_before (&psi, assign, GSI_SAME_STMT); | |
2394 | } | |
2395 | } | |
2396 | ||
42ad5d61 | 2397 | /* Rewrite the scalar dependences crossing the boundary of the BB |
c9a67530 | 2398 | containing STMT with an array. Return true when something has been |
2399 | changed. */ | |
42ad5d61 | 2400 | |
c9a67530 | 2401 | static bool |
3ac4f821 | 2402 | rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi) |
42ad5d61 | 2403 | { |
3ac4f821 | 2404 | sese region = SCOP_REGION (scop); |
42ad5d61 | 2405 | gimple stmt = gsi_stmt (*gsi); |
2406 | imm_use_iterator imm_iter; | |
2407 | tree def; | |
2408 | basic_block def_bb; | |
2409 | tree zero_dim_array = NULL_TREE; | |
2410 | gimple use_stmt; | |
c9a67530 | 2411 | bool res = false; |
42ad5d61 | 2412 | |
b19b9f62 | 2413 | switch (gimple_code (stmt)) |
2414 | { | |
2415 | case GIMPLE_ASSIGN: | |
2416 | def = gimple_assign_lhs (stmt); | |
2417 | break; | |
2418 | ||
2419 | case GIMPLE_CALL: | |
2420 | def = gimple_call_lhs (stmt); | |
2421 | break; | |
2422 | ||
2423 | default: | |
c9a67530 | 2424 | return false; |
b19b9f62 | 2425 | } |
42ad5d61 | 2426 | |
a4a4b66a | 2427 | if (!def |
2428 | || !is_gimple_reg (def)) | |
c9a67530 | 2429 | return false; |
42ad5d61 | 2430 | |
5d21c24a | 2431 | if (scev_analyzable_p (def, region)) |
2432 | { | |
2433 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def)); | |
2434 | tree scev = scalar_evolution_in_region (region, loop, def); | |
2435 | ||
c9a67530 | 2436 | if (tree_contains_chrecs (scev, NULL)) |
2437 | return false; | |
5d21c24a | 2438 | |
c9a67530 | 2439 | propagate_expr_outside_region (def, scev, region); |
2440 | return true; | |
5d21c24a | 2441 | } |
2442 | ||
42ad5d61 | 2443 | def_bb = gimple_bb (stmt); |
2444 | ||
3ac4f821 | 2445 | handle_scalar_deps_crossing_scop_limits (scop, def, stmt); |
2446 | ||
42ad5d61 | 2447 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) |
c9a67530 | 2448 | if (gimple_code (use_stmt) == GIMPLE_PHI |
2449 | && (res = true)) | |
5061777c | 2450 | { |
ed455480 | 2451 | gimple_stmt_iterator psi = gsi_for_stmt (use_stmt); |
42ad5d61 | 2452 | |
ed455480 | 2453 | if (scalar_close_phi_node_p (gsi_stmt (psi))) |
8c6b3774 | 2454 | rewrite_close_phi_out_of_ssa (scop, &psi); |
ed455480 | 2455 | else |
8c6b3774 | 2456 | rewrite_phi_out_of_ssa (scop, &psi); |
ed455480 | 2457 | } |
2458 | ||
2459 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2460 | if (gimple_code (use_stmt) != GIMPLE_PHI | |
2461 | && def_bb != gimple_bb (use_stmt) | |
c9a67530 | 2462 | && !is_gimple_debug (use_stmt) |
2463 | && (res = true)) | |
ed455480 | 2464 | { |
5061777c | 2465 | if (!zero_dim_array) |
2466 | { | |
03c1327b | 2467 | zero_dim_array = create_zero_dim_array |
874117c8 | 2468 | (def, "Cross_BB_scalar_dependence"); |
1688f172 | 2469 | insert_out_of_ssa_copy (scop, zero_dim_array, def, |
39a34dd8 | 2470 | SSA_NAME_DEF_STMT (def)); |
5061777c | 2471 | gsi_next (gsi); |
2472 | } | |
2473 | ||
2bdae241 | 2474 | rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array), |
8c6b3774 | 2475 | def, use_stmt); |
5061777c | 2476 | } |
c9a67530 | 2477 | |
2478 | return res; | |
5061777c | 2479 | } |
2480 | ||
bf8b5699 | 2481 | /* Rewrite out of SSA all the reduction phi nodes of SCOP. */ |
2482 | ||
8c6b3774 | 2483 | static void |
bf8b5699 | 2484 | rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop) |
2485 | { | |
2486 | basic_block bb; | |
2487 | gimple_stmt_iterator psi; | |
2488 | sese region = SCOP_REGION (scop); | |
c9a67530 | 2489 | bool changed = false; |
5061777c | 2490 | |
3ac4f821 | 2491 | /* Create an extra empty BB after the scop. */ |
7133bc6f | 2492 | split_edge (SESE_EXIT (region)); |
3ac4f821 | 2493 | |
fc00614f | 2494 | FOR_EACH_BB_FN (bb, cfun) |
5061777c | 2495 | if (bb_in_sese_p (bb, region)) |
2496 | for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
3ac4f821 | 2497 | changed |= rewrite_cross_bb_scalar_deps (scop, &psi); |
5061777c | 2498 | |
c9a67530 | 2499 | if (changed) |
2500 | { | |
2501 | scev_reset_htab (); | |
2502 | update_ssa (TODO_update_ssa); | |
5061777c | 2503 | #ifdef ENABLE_CHECKING |
c9a67530 | 2504 | verify_loop_closed_ssa (true); |
5061777c | 2505 | #endif |
c9a67530 | 2506 | } |
c6bb733d | 2507 | } |
2508 | ||
2509 | /* Returns the number of pbbs that are in loops contained in SCOP. */ | |
2510 | ||
2511 | static int | |
2512 | nb_pbbs_in_loops (scop_p scop) | |
2513 | { | |
2514 | int i; | |
2515 | poly_bb_p pbb; | |
2516 | int res = 0; | |
2517 | ||
f1f41a6c | 2518 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
c6bb733d | 2519 | if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop))) |
2520 | res++; | |
2521 | ||
2522 | return res; | |
2523 | } | |
2524 | ||
f007fe97 | 2525 | /* Return the number of data references in BB that write in |
2526 | memory. */ | |
2527 | ||
2528 | static int | |
2529 | nb_data_writes_in_bb (basic_block bb) | |
2530 | { | |
2531 | int res = 0; | |
2532 | gimple_stmt_iterator gsi; | |
2533 | ||
2534 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2535 | if (gimple_vdef (gsi_stmt (gsi))) | |
2536 | res++; | |
2537 | ||
2538 | return res; | |
2539 | } | |
2540 | ||
8c6b3774 | 2541 | /* Splits at STMT the basic block BB represented as PBB in the |
2542 | polyhedral form. */ | |
2543 | ||
2544 | static edge | |
2545 | split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt) | |
2546 | { | |
2547 | edge e1 = split_block (bb, stmt); | |
2548 | new_pbb_from_pbb (scop, pbb, e1->dest); | |
2549 | return e1; | |
2550 | } | |
2551 | ||
2552 | /* Splits STMT out of its current BB. This is done for reduction | |
2553 | statements for which we want to ignore data dependences. */ | |
30f4f4a6 | 2554 | |
2555 | static basic_block | |
8c6b3774 | 2556 | split_reduction_stmt (scop_p scop, gimple stmt) |
30f4f4a6 | 2557 | { |
30f4f4a6 | 2558 | basic_block bb = gimple_bb (stmt); |
8c6b3774 | 2559 | poly_bb_p pbb = pbb_from_bb (bb); |
1688f172 | 2560 | gimple_bb_p gbb = gbb_from_bb (bb); |
8c6b3774 | 2561 | edge e1; |
1688f172 | 2562 | int i; |
2563 | data_reference_p dr; | |
30f4f4a6 | 2564 | |
f007fe97 | 2565 | /* Do not split basic blocks with no writes to memory: the reduction |
2566 | will be the only write to memory. */ | |
fc2372d9 | 2567 | if (nb_data_writes_in_bb (bb) == 0 |
2568 | /* Or if we have already marked BB as a reduction. */ | |
2569 | || PBB_IS_REDUCTION (pbb_from_bb (bb))) | |
f007fe97 | 2570 | return bb; |
2571 | ||
8c6b3774 | 2572 | e1 = split_pbb (scop, pbb, bb, stmt); |
30f4f4a6 | 2573 | |
8c6b3774 | 2574 | /* Split once more only when the reduction stmt is not the only one |
2575 | left in the original BB. */ | |
2576 | if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb))) | |
2577 | { | |
2578 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
2579 | gsi_prev (&gsi); | |
2580 | e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi)); | |
2581 | } | |
30f4f4a6 | 2582 | |
1688f172 | 2583 | /* A part of the data references will end in a different basic block |
2584 | after the split: move the DRs from the original GBB to the newly | |
2585 | created GBB1. */ | |
f1f41a6c | 2586 | FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr) |
1688f172 | 2587 | { |
2588 | basic_block bb1 = gimple_bb (DR_STMT (dr)); | |
2589 | ||
2590 | if (bb1 != bb) | |
2591 | { | |
2592 | gimple_bb_p gbb1 = gbb_from_bb (bb1); | |
f1f41a6c | 2593 | GBB_DATA_REFS (gbb1).safe_push (dr); |
2594 | GBB_DATA_REFS (gbb).ordered_remove (i); | |
1688f172 | 2595 | i--; |
2596 | } | |
2597 | } | |
2598 | ||
8c6b3774 | 2599 | return e1->dest; |
30f4f4a6 | 2600 | } |
2601 | ||
2602 | /* Return true when stmt is a reduction operation. */ | |
2603 | ||
2604 | static inline bool | |
2605 | is_reduction_operation_p (gimple stmt) | |
2606 | { | |
5daf19b1 | 2607 | enum tree_code code; |
2608 | ||
2609 | gcc_assert (is_gimple_assign (stmt)); | |
2610 | code = gimple_assign_rhs_code (stmt); | |
2611 | ||
30f4f4a6 | 2612 | return flag_associative_math |
5daf19b1 | 2613 | && commutative_tree_code (code) |
2614 | && associative_tree_code (code); | |
30f4f4a6 | 2615 | } |
2616 | ||
2617 | /* Returns true when PHI contains an argument ARG. */ | |
2618 | ||
2619 | static bool | |
2620 | phi_contains_arg (gimple phi, tree arg) | |
2621 | { | |
2622 | size_t i; | |
2623 | ||
2624 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2625 | if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0)) | |
2626 | return true; | |
2627 | ||
2628 | return false; | |
2629 | } | |
2630 | ||
2631 | /* Return a loop phi node that corresponds to a reduction containing LHS. */ | |
2632 | ||
2633 | static gimple | |
2634 | follow_ssa_with_commutative_ops (tree arg, tree lhs) | |
2635 | { | |
2636 | gimple stmt; | |
2637 | ||
2638 | if (TREE_CODE (arg) != SSA_NAME) | |
2639 | return NULL; | |
2640 | ||
2641 | stmt = SSA_NAME_DEF_STMT (arg); | |
2642 | ||
7f60ea7e | 2643 | if (gimple_code (stmt) == GIMPLE_NOP |
2644 | || gimple_code (stmt) == GIMPLE_CALL) | |
36f22aa0 | 2645 | return NULL; |
2646 | ||
30f4f4a6 | 2647 | if (gimple_code (stmt) == GIMPLE_PHI) |
2648 | { | |
2649 | if (phi_contains_arg (stmt, lhs)) | |
2650 | return stmt; | |
2651 | return NULL; | |
2652 | } | |
2653 | ||
5daf19b1 | 2654 | if (!is_gimple_assign (stmt)) |
2655 | return NULL; | |
2656 | ||
30f4f4a6 | 2657 | if (gimple_num_ops (stmt) == 2) |
2658 | return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs); | |
2659 | ||
2660 | if (is_reduction_operation_p (stmt)) | |
2661 | { | |
2662 | gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs); | |
2663 | ||
2664 | return res ? res : | |
2665 | follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs); | |
2666 | } | |
2667 | ||
2668 | return NULL; | |
2669 | } | |
2670 | ||
2671 | /* Detect commutative and associative scalar reductions starting at | |
5184a05f | 2672 | the STMT. Return the phi node of the reduction cycle, or NULL. */ |
30f4f4a6 | 2673 | |
2674 | static gimple | |
2675 | detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg, | |
f1f41a6c | 2676 | vec<gimple> *in, |
2677 | vec<gimple> *out) | |
30f4f4a6 | 2678 | { |
2679 | gimple phi = follow_ssa_with_commutative_ops (arg, lhs); | |
2680 | ||
5184a05f | 2681 | if (!phi) |
2682 | return NULL; | |
30f4f4a6 | 2683 | |
f1f41a6c | 2684 | in->safe_push (stmt); |
2685 | out->safe_push (stmt); | |
5184a05f | 2686 | return phi; |
30f4f4a6 | 2687 | } |
2688 | ||
2689 | /* Detect commutative and associative scalar reductions starting at | |
5d2603f9 | 2690 | STMT. Return the phi node of the reduction cycle, or NULL. */ |
30f4f4a6 | 2691 | |
2692 | static gimple | |
f1f41a6c | 2693 | detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in, |
2694 | vec<gimple> *out) | |
30f4f4a6 | 2695 | { |
2696 | tree lhs = gimple_assign_lhs (stmt); | |
2697 | ||
2698 | if (gimple_num_ops (stmt) == 2) | |
2699 | return detect_commutative_reduction_arg (lhs, stmt, | |
2700 | gimple_assign_rhs1 (stmt), | |
2701 | in, out); | |
2702 | ||
2703 | if (is_reduction_operation_p (stmt)) | |
2704 | { | |
2705 | gimple res = detect_commutative_reduction_arg (lhs, stmt, | |
2706 | gimple_assign_rhs1 (stmt), | |
2707 | in, out); | |
2708 | return res ? res | |
2709 | : detect_commutative_reduction_arg (lhs, stmt, | |
2710 | gimple_assign_rhs2 (stmt), | |
2711 | in, out); | |
2712 | } | |
2713 | ||
2714 | return NULL; | |
2715 | } | |
2716 | ||
2717 | /* Return a loop phi node that corresponds to a reduction containing LHS. */ | |
2718 | ||
2719 | static gimple | |
2720 | follow_inital_value_to_phi (tree arg, tree lhs) | |
2721 | { | |
2722 | gimple stmt; | |
2723 | ||
2724 | if (!arg || TREE_CODE (arg) != SSA_NAME) | |
2725 | return NULL; | |
2726 | ||
2727 | stmt = SSA_NAME_DEF_STMT (arg); | |
2728 | ||
2729 | if (gimple_code (stmt) == GIMPLE_PHI | |
2730 | && phi_contains_arg (stmt, lhs)) | |
2731 | return stmt; | |
2732 | ||
2733 | return NULL; | |
2734 | } | |
2735 | ||
2736 | ||
9d75589a | 2737 | /* Return the argument of the loop PHI that is the initial value coming |
30f4f4a6 | 2738 | from outside the loop. */ |
2739 | ||
2740 | static edge | |
2741 | edge_initial_value_for_loop_phi (gimple phi) | |
2742 | { | |
2743 | size_t i; | |
2744 | ||
2745 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2746 | { | |
2747 | edge e = gimple_phi_arg_edge (phi, i); | |
2748 | ||
2749 | if (loop_depth (e->src->loop_father) | |
2750 | < loop_depth (e->dest->loop_father)) | |
2751 | return e; | |
2752 | } | |
2753 | ||
2754 | return NULL; | |
2755 | } | |
2756 | ||
9d75589a | 2757 | /* Return the argument of the loop PHI that is the initial value coming |
30f4f4a6 | 2758 | from outside the loop. */ |
2759 | ||
2760 | static tree | |
2761 | initial_value_for_loop_phi (gimple phi) | |
2762 | { | |
2763 | size_t i; | |
2764 | ||
2765 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2766 | { | |
2767 | edge e = gimple_phi_arg_edge (phi, i); | |
2768 | ||
2769 | if (loop_depth (e->src->loop_father) | |
2770 | < loop_depth (e->dest->loop_father)) | |
2771 | return gimple_phi_arg_def (phi, i); | |
2772 | } | |
2773 | ||
2774 | return NULL_TREE; | |
2775 | } | |
2776 | ||
2528d7cf | 2777 | /* Returns true when DEF is used outside the reduction cycle of |
2778 | LOOP_PHI. */ | |
2779 | ||
2780 | static bool | |
2781 | used_outside_reduction (tree def, gimple loop_phi) | |
2782 | { | |
2783 | use_operand_p use_p; | |
2784 | imm_use_iterator imm_iter; | |
2785 | loop_p loop = loop_containing_stmt (loop_phi); | |
2786 | ||
2787 | /* In LOOP, DEF should be used only in LOOP_PHI. */ | |
2788 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2789 | { | |
2790 | gimple stmt = USE_STMT (use_p); | |
2791 | ||
2792 | if (stmt != loop_phi | |
2793 | && !is_gimple_debug (stmt) | |
2794 | && flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
2795 | return true; | |
2796 | } | |
2797 | ||
2798 | return false; | |
2799 | } | |
2800 | ||
93514494 | 2801 | /* Detect commutative and associative scalar reductions belonging to |
2802 | the SCOP starting at the loop closed phi node STMT. Return the phi | |
2803 | node of the reduction cycle, or NULL. */ | |
30f4f4a6 | 2804 | |
2805 | static gimple | |
f1f41a6c | 2806 | detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in, |
2807 | vec<gimple> *out) | |
30f4f4a6 | 2808 | { |
2809 | if (scalar_close_phi_node_p (stmt)) | |
2810 | { | |
2528d7cf | 2811 | gimple def, loop_phi, phi, close_phi = stmt; |
2812 | tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0); | |
5184a05f | 2813 | |
2814 | if (TREE_CODE (arg) != SSA_NAME) | |
2815 | return NULL; | |
2816 | ||
03ce78db | 2817 | /* Note that loop close phi nodes should have a single argument |
2818 | because we translated the representation into a canonical form | |
2819 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
2528d7cf | 2820 | gcc_assert (gimple_phi_num_args (close_phi) == 1); |
03ce78db | 2821 | |
5184a05f | 2822 | def = SSA_NAME_DEF_STMT (arg); |
2528d7cf | 2823 | if (!stmt_in_sese_p (def, SCOP_REGION (scop)) |
2824 | || !(loop_phi = detect_commutative_reduction (scop, def, in, out))) | |
93514494 | 2825 | return NULL; |
2826 | ||
2528d7cf | 2827 | lhs = gimple_phi_result (close_phi); |
2828 | init = initial_value_for_loop_phi (loop_phi); | |
2829 | phi = follow_inital_value_to_phi (init, lhs); | |
30f4f4a6 | 2830 | |
2528d7cf | 2831 | if (phi && (used_outside_reduction (lhs, phi) |
2832 | || !has_single_use (gimple_phi_result (phi)))) | |
30f4f4a6 | 2833 | return NULL; |
2528d7cf | 2834 | |
f1f41a6c | 2835 | in->safe_push (loop_phi); |
2836 | out->safe_push (close_phi); | |
2528d7cf | 2837 | return phi; |
30f4f4a6 | 2838 | } |
2839 | ||
2840 | if (gimple_code (stmt) == GIMPLE_ASSIGN) | |
2841 | return detect_commutative_reduction_assign (stmt, in, out); | |
2842 | ||
2843 | return NULL; | |
2844 | } | |
2845 | ||
2846 | /* Translate the scalar reduction statement STMT to an array RED | |
2847 | knowing that its recursive phi node is LOOP_PHI. */ | |
2848 | ||
2849 | static void | |
1688f172 | 2850 | translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red, |
2851 | gimple stmt, gimple loop_phi) | |
30f4f4a6 | 2852 | { |
30f4f4a6 | 2853 | tree res = gimple_phi_result (loop_phi); |
53b5bc41 | 2854 | gimple assign = gimple_build_assign (res, unshare_expr (red)); |
1688f172 | 2855 | gimple_stmt_iterator gsi; |
30f4f4a6 | 2856 | |
1688f172 | 2857 | insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi))); |
30f4f4a6 | 2858 | |
53b5bc41 | 2859 | assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt)); |
1688f172 | 2860 | gsi = gsi_for_stmt (stmt); |
2861 | gsi_next (&gsi); | |
2862 | insert_stmts (scop, assign, NULL, gsi); | |
30f4f4a6 | 2863 | } |
2864 | ||
eae8f2a1 | 2865 | /* Removes the PHI node and resets all the debug stmts that are using |
2866 | the PHI_RESULT. */ | |
2867 | ||
2868 | static void | |
2869 | remove_phi (gimple phi) | |
2870 | { | |
2871 | imm_use_iterator imm_iter; | |
2872 | tree def; | |
2873 | use_operand_p use_p; | |
2874 | gimple_stmt_iterator gsi; | |
4997014d | 2875 | auto_vec<gimple, 3> update; |
eae8f2a1 | 2876 | unsigned int i; |
2877 | gimple stmt; | |
2878 | ||
2879 | def = PHI_RESULT (phi); | |
2880 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2881 | { | |
2882 | stmt = USE_STMT (use_p); | |
2883 | ||
2884 | if (is_gimple_debug (stmt)) | |
2885 | { | |
2886 | gimple_debug_bind_reset_value (stmt); | |
f1f41a6c | 2887 | update.safe_push (stmt); |
eae8f2a1 | 2888 | } |
2889 | } | |
2890 | ||
f1f41a6c | 2891 | FOR_EACH_VEC_ELT (update, i, stmt) |
eae8f2a1 | 2892 | update_stmt (stmt); |
2893 | ||
eae8f2a1 | 2894 | gsi = gsi_for_phi_node (phi); |
2895 | remove_phi_node (&gsi, false); | |
2896 | } | |
2897 | ||
fa6ed0e9 | 2898 | /* Helper function for for_each_index. For each INDEX of the data |
2899 | reference REF, returns true when its indices are valid in the loop | |
2900 | nest LOOP passed in as DATA. */ | |
2901 | ||
2902 | static bool | |
2903 | dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data) | |
2904 | { | |
2905 | loop_p loop; | |
2906 | basic_block header, def_bb; | |
2907 | gimple stmt; | |
2908 | ||
2909 | if (TREE_CODE (*index) != SSA_NAME) | |
2910 | return true; | |
2911 | ||
2912 | loop = *((loop_p *) data); | |
2913 | header = loop->header; | |
2914 | stmt = SSA_NAME_DEF_STMT (*index); | |
2915 | ||
2916 | if (!stmt) | |
2917 | return true; | |
2918 | ||
2919 | def_bb = gimple_bb (stmt); | |
2920 | ||
2921 | if (!def_bb) | |
2922 | return true; | |
2923 | ||
2924 | return dominated_by_p (CDI_DOMINATORS, header, def_bb); | |
2925 | } | |
2926 | ||
53b5bc41 | 2927 | /* When the result of a CLOSE_PHI is written to a memory location, |
2928 | return a pointer to that memory reference, otherwise return | |
2929 | NULL_TREE. */ | |
2930 | ||
2931 | static tree | |
2932 | close_phi_written_to_memory (gimple close_phi) | |
2933 | { | |
2934 | imm_use_iterator imm_iter; | |
53b5bc41 | 2935 | use_operand_p use_p; |
2936 | gimple stmt; | |
fa6ed0e9 | 2937 | tree res, def = gimple_phi_result (close_phi); |
53b5bc41 | 2938 | |
2939 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2940 | if ((stmt = USE_STMT (use_p)) | |
2941 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
fa6ed0e9 | 2942 | && (res = gimple_assign_lhs (stmt))) |
2943 | { | |
2944 | switch (TREE_CODE (res)) | |
2945 | { | |
2946 | case VAR_DECL: | |
2947 | case PARM_DECL: | |
2948 | case RESULT_DECL: | |
2949 | return res; | |
2950 | ||
2951 | case ARRAY_REF: | |
2952 | case MEM_REF: | |
2953 | { | |
2954 | tree arg = gimple_phi_arg_def (close_phi, 0); | |
2955 | loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg)); | |
2956 | ||
2957 | /* FIXME: this restriction is for id-{24,25}.f and | |
2958 | could be handled by duplicating the computation of | |
2959 | array indices before the loop of the close_phi. */ | |
2960 | if (for_each_index (&res, dr_indices_valid_in_loop, &nest)) | |
2961 | return res; | |
2962 | } | |
2963 | /* Fallthru. */ | |
53b5bc41 | 2964 | |
fa6ed0e9 | 2965 | default: |
2966 | continue; | |
2967 | } | |
2968 | } | |
53b5bc41 | 2969 | return NULL_TREE; |
2970 | } | |
2971 | ||
30f4f4a6 | 2972 | /* Rewrite out of SSA the reduction described by the loop phi nodes |
2973 | IN, and the close phi nodes OUT. IN and OUT are structured by loop | |
2974 | levels like this: | |
2975 | ||
2976 | IN: stmt, loop_n, ..., loop_0 | |
2977 | OUT: stmt, close_n, ..., close_0 | |
2978 | ||
2979 | the first element is the reduction statement, and the next elements | |
2980 | are the loop and close phi nodes of each of the outer loops. */ | |
2981 | ||
2982 | static void | |
8c6b3774 | 2983 | translate_scalar_reduction_to_array (scop_p scop, |
f1f41a6c | 2984 | vec<gimple> in, |
2985 | vec<gimple> out) | |
30f4f4a6 | 2986 | { |
30f4f4a6 | 2987 | gimple loop_phi; |
f1f41a6c | 2988 | unsigned int i = out.length () - 1; |
2989 | tree red = close_phi_written_to_memory (out[i]); | |
30f4f4a6 | 2990 | |
f1f41a6c | 2991 | FOR_EACH_VEC_ELT (in, i, loop_phi) |
30f4f4a6 | 2992 | { |
f1f41a6c | 2993 | gimple close_phi = out[i]; |
30f4f4a6 | 2994 | |
2995 | if (i == 0) | |
2996 | { | |
2997 | gimple stmt = loop_phi; | |
8c6b3774 | 2998 | basic_block bb = split_reduction_stmt (scop, stmt); |
2999 | poly_bb_p pbb = pbb_from_bb (bb); | |
3000 | PBB_IS_REDUCTION (pbb) = true; | |
30f4f4a6 | 3001 | gcc_assert (close_phi == loop_phi); |
3002 | ||
53b5bc41 | 3003 | if (!red) |
3004 | red = create_zero_dim_array | |
3005 | (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction"); | |
3006 | ||
f1f41a6c | 3007 | translate_scalar_reduction_to_array_for_stmt (scop, red, stmt, in[1]); |
30f4f4a6 | 3008 | continue; |
3009 | } | |
3010 | ||
f1f41a6c | 3011 | if (i == in.length () - 1) |
30f4f4a6 | 3012 | { |
53b5bc41 | 3013 | insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi), |
3014 | unshare_expr (red), close_phi); | |
39a34dd8 | 3015 | insert_out_of_ssa_copy_on_edge |
8c6b3774 | 3016 | (scop, edge_initial_value_for_loop_phi (loop_phi), |
53b5bc41 | 3017 | unshare_expr (red), initial_value_for_loop_phi (loop_phi)); |
30f4f4a6 | 3018 | } |
3019 | ||
eae8f2a1 | 3020 | remove_phi (loop_phi); |
3021 | remove_phi (close_phi); | |
30f4f4a6 | 3022 | } |
3023 | } | |
3024 | ||
c9a67530 | 3025 | /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns |
3026 | true when something has been changed. */ | |
30f4f4a6 | 3027 | |
c9a67530 | 3028 | static bool |
8c6b3774 | 3029 | rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop, |
3030 | gimple close_phi) | |
30f4f4a6 | 3031 | { |
c9a67530 | 3032 | bool res; |
4997014d | 3033 | auto_vec<gimple, 10> in; |
3034 | auto_vec<gimple, 10> out; | |
30f4f4a6 | 3035 | |
93514494 | 3036 | detect_commutative_reduction (scop, close_phi, &in, &out); |
f1f41a6c | 3037 | res = in.length () > 1; |
c9a67530 | 3038 | if (res) |
8c6b3774 | 3039 | translate_scalar_reduction_to_array (scop, in, out); |
30f4f4a6 | 3040 | |
c9a67530 | 3041 | return res; |
30f4f4a6 | 3042 | } |
3043 | ||
c9a67530 | 3044 | /* Rewrites all the commutative reductions from LOOP out of SSA. |
3045 | Returns true when something has been changed. */ | |
30f4f4a6 | 3046 | |
c9a67530 | 3047 | static bool |
8c6b3774 | 3048 | rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop, |
3049 | loop_p loop) | |
30f4f4a6 | 3050 | { |
3051 | gimple_stmt_iterator gsi; | |
3052 | edge exit = single_exit (loop); | |
72085fb7 | 3053 | tree res; |
c9a67530 | 3054 | bool changed = false; |
30f4f4a6 | 3055 | |
3056 | if (!exit) | |
c9a67530 | 3057 | return false; |
30f4f4a6 | 3058 | |
3059 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
72085fb7 | 3060 | if ((res = gimple_phi_result (gsi_stmt (gsi))) |
7c782c9b | 3061 | && !virtual_operand_p (res) |
8c6b3774 | 3062 | && !scev_analyzable_p (res, SCOP_REGION (scop))) |
c9a67530 | 3063 | changed |= rewrite_commutative_reductions_out_of_ssa_close_phi |
8c6b3774 | 3064 | (scop, gsi_stmt (gsi)); |
c9a67530 | 3065 | |
3066 | return changed; | |
30f4f4a6 | 3067 | } |
3068 | ||
3069 | /* Rewrites all the commutative reductions from SCOP out of SSA. */ | |
3070 | ||
8c6b3774 | 3071 | static void |
3072 | rewrite_commutative_reductions_out_of_ssa (scop_p scop) | |
30f4f4a6 | 3073 | { |
30f4f4a6 | 3074 | loop_p loop; |
c9a67530 | 3075 | bool changed = false; |
8c6b3774 | 3076 | sese region = SCOP_REGION (scop); |
8643dd0a | 3077 | |
f21d4d00 | 3078 | FOR_EACH_LOOP (loop, 0) |
30f4f4a6 | 3079 | if (loop_in_sese_p (loop, region)) |
8c6b3774 | 3080 | changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop); |
ff010926 | 3081 | |
c9a67530 | 3082 | if (changed) |
3083 | { | |
3084 | scev_reset_htab (); | |
3085 | gsi_commit_edge_inserts (); | |
3086 | update_ssa (TODO_update_ssa); | |
ff010926 | 3087 | #ifdef ENABLE_CHECKING |
c9a67530 | 3088 | verify_loop_closed_ssa (true); |
ff010926 | 3089 | #endif |
c9a67530 | 3090 | } |
30f4f4a6 | 3091 | } |
3092 | ||
b43389bd | 3093 | /* Can all ivs be represented by a signed integer? |
3094 | As CLooG might generate negative values in its expressions, signed loop ivs | |
3095 | are required in the backend. */ | |
d3746d81 | 3096 | |
b43389bd | 3097 | static bool |
3098 | scop_ivs_can_be_represented (scop_p scop) | |
3099 | { | |
b43389bd | 3100 | loop_p loop; |
18046220 | 3101 | gimple_stmt_iterator psi; |
83b709f2 | 3102 | bool result = true; |
b43389bd | 3103 | |
f21d4d00 | 3104 | FOR_EACH_LOOP (loop, 0) |
b43389bd | 3105 | { |
b43389bd | 3106 | if (!loop_in_sese_p (loop, SCOP_REGION (scop))) |
3107 | continue; | |
3108 | ||
18046220 | 3109 | for (psi = gsi_start_phis (loop->header); |
3110 | !gsi_end_p (psi); gsi_next (&psi)) | |
3111 | { | |
3112 | gimple phi = gsi_stmt (psi); | |
3113 | tree res = PHI_RESULT (phi); | |
3114 | tree type = TREE_TYPE (res); | |
b43389bd | 3115 | |
18046220 | 3116 | if (TYPE_UNSIGNED (type) |
a0553bff | 3117 | && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node)) |
83b709f2 | 3118 | { |
3119 | result = false; | |
3120 | break; | |
3121 | } | |
18046220 | 3122 | } |
83b709f2 | 3123 | if (!result) |
f21d4d00 | 3124 | break; |
b43389bd | 3125 | } |
3126 | ||
83b709f2 | 3127 | return result; |
b43389bd | 3128 | } |
3129 | ||
c6bb733d | 3130 | /* Builds the polyhedral representation for a SESE region. */ |
3131 | ||
f49215ce | 3132 | void |
c6bb733d | 3133 | build_poly_scop (scop_p scop) |
3134 | { | |
3135 | sese region = SCOP_REGION (scop); | |
c2e502a5 | 3136 | graphite_dim_t max_dim; |
30f4f4a6 | 3137 | |
8c6b3774 | 3138 | build_scop_bbs (scop); |
c6bb733d | 3139 | |
3140 | /* FIXME: This restriction is needed to avoid a problem in CLooG. | |
3141 | Once CLooG is fixed, remove this guard. Anyways, it makes no | |
3142 | sense to optimize a scop containing only PBBs that do not belong | |
3143 | to any loops. */ | |
3144 | if (nb_pbbs_in_loops (scop) == 0) | |
f49215ce | 3145 | return; |
c6bb733d | 3146 | |
b43389bd | 3147 | if (!scop_ivs_can_be_represented (scop)) |
f49215ce | 3148 | return; |
b43389bd | 3149 | |
c5409e1f | 3150 | if (flag_associative_math) |
3151 | rewrite_commutative_reductions_out_of_ssa (scop); | |
3152 | ||
c6bb733d | 3153 | build_sese_loop_nests (region); |
54c91640 | 3154 | /* Record all conditions in REGION. */ |
3155 | sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr); | |
c6bb733d | 3156 | find_scop_parameters (scop); |
3157 | ||
c2e502a5 | 3158 | max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS); |
3159 | if (scop_nb_params (scop) > max_dim) | |
f49215ce | 3160 | return; |
c2e502a5 | 3161 | |
c6bb733d | 3162 | build_scop_iteration_domain (scop); |
3163 | build_scop_context (scop); | |
c6bb733d | 3164 | add_conditions_to_constraints (scop); |
8c6b3774 | 3165 | |
3166 | /* Rewrite out of SSA only after having translated the | |
3167 | representation to the polyhedral representation to avoid scev | |
3168 | analysis failures. That means that these functions will insert | |
3169 | new data references that they create in the right place. */ | |
8c6b3774 | 3170 | rewrite_reductions_out_of_ssa (scop); |
3171 | rewrite_cross_bb_scalar_deps_out_of_ssa (scop); | |
3172 | ||
3173 | build_scop_drs (scop); | |
f77385d3 | 3174 | scop_to_lst (scop); |
c6bb733d | 3175 | build_scop_scattering (scop); |
c6bb733d | 3176 | |
f49215ce | 3177 | /* This SCoP has been translated to the polyhedral |
3178 | representation. */ | |
3179 | POLY_SCOP_P (scop) = true; | |
c6bb733d | 3180 | } |
c6bb733d | 3181 | #endif |