]>
Commit | Line | Data |
---|---|---|
2abae5f1 | 1 | /* Conversion of SESE regions to Polyhedra. |
d1e082c2 | 2 | Copyright (C) 2009-2013 Free Software Foundation, Inc. |
2abae5f1 SP |
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" | |
33ad93b9 RG |
22 | |
23 | #ifdef HAVE_cloog | |
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> | |
29 | #include <cloog/cloog.h> | |
30 | #include <cloog/cloog.h> | |
31 | #include <cloog/isl/domain.h> | |
32 | #endif | |
33 | ||
2abae5f1 SP |
34 | #include "system.h" |
35 | #include "coretypes.h" | |
2abae5f1 | 36 | #include "tree-flow.h" |
7a1c57d3 | 37 | #include "tree-pass.h" |
2abae5f1 SP |
38 | #include "cfgloop.h" |
39 | #include "tree-chrec.h" | |
40 | #include "tree-data-ref.h" | |
41 | #include "tree-scalar-evolution.h" | |
2abae5f1 | 42 | #include "domwalk.h" |
2abae5f1 SP |
43 | #include "sese.h" |
44 | ||
45 | #ifdef HAVE_cloog | |
2abae5f1 | 46 | #include "graphite-poly.h" |
2abae5f1 SP |
47 | #include "graphite-sese-to-poly.h" |
48 | ||
33ad93b9 RG |
49 | |
50 | /* Assigns to RES the value of the INTEGER_CST T. */ | |
51 | ||
52 | static inline void | |
53 | tree_int_to_gmp (tree t, mpz_t res) | |
54 | { | |
55 | double_int di = tree_to_double_int (t); | |
56 | mpz_set_double_int (res, di, TYPE_UNSIGNED (TREE_TYPE (t))); | |
57 | } | |
58 | ||
159e4616 SP |
59 | /* Returns the index of the PHI argument defined in the outermost |
60 | loop. */ | |
2abae5f1 SP |
61 | |
62 | static size_t | |
159e4616 | 63 | phi_arg_in_outermost_loop (gimple phi) |
2abae5f1 SP |
64 | { |
65 | loop_p loop = gimple_bb (phi)->loop_father; | |
159e4616 | 66 | size_t i, res = 0; |
2abae5f1 SP |
67 | |
68 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
69 | if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src)) | |
159e4616 SP |
70 | { |
71 | loop = gimple_phi_arg_edge (phi, i)->src->loop_father; | |
72 | res = i; | |
73 | } | |
2abae5f1 | 74 | |
159e4616 | 75 | return res; |
2abae5f1 SP |
76 | } |
77 | ||
78 | /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position | |
79 | PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */ | |
80 | ||
81 | static void | |
82 | remove_simple_copy_phi (gimple_stmt_iterator *psi) | |
83 | { | |
84 | gimple phi = gsi_stmt (*psi); | |
85 | tree res = gimple_phi_result (phi); | |
159e4616 | 86 | size_t entry = phi_arg_in_outermost_loop (phi); |
2abae5f1 SP |
87 | tree init = gimple_phi_arg_def (phi, entry); |
88 | gimple stmt = gimple_build_assign (res, init); | |
89 | edge e = gimple_phi_arg_edge (phi, entry); | |
90 | ||
91 | remove_phi_node (psi, false); | |
92 | gsi_insert_on_edge_immediate (e, stmt); | |
93 | SSA_NAME_DEF_STMT (res) = stmt; | |
94 | } | |
95 | ||
96 | /* Removes an invariant phi node at position PSI by inserting on the | |
97 | loop ENTRY edge the assignment RES = INIT. */ | |
98 | ||
99 | static void | |
100 | remove_invariant_phi (sese region, gimple_stmt_iterator *psi) | |
101 | { | |
102 | gimple phi = gsi_stmt (*psi); | |
103 | loop_p loop = loop_containing_stmt (phi); | |
104 | tree res = gimple_phi_result (phi); | |
105 | tree scev = scalar_evolution_in_region (region, loop, res); | |
159e4616 | 106 | size_t entry = phi_arg_in_outermost_loop (phi); |
2abae5f1 SP |
107 | edge e = gimple_phi_arg_edge (phi, entry); |
108 | tree var; | |
109 | gimple stmt; | |
355a7673 | 110 | gimple_seq stmts = NULL; |
2abae5f1 SP |
111 | |
112 | if (tree_contains_chrecs (scev, NULL)) | |
113 | scev = gimple_phi_arg_def (phi, entry); | |
114 | ||
115 | var = force_gimple_operand (scev, &stmts, true, NULL_TREE); | |
116 | stmt = gimple_build_assign (res, var); | |
117 | remove_phi_node (psi, false); | |
118 | ||
355a7673 | 119 | gimple_seq_add_stmt (&stmts, stmt); |
2abae5f1 SP |
120 | gsi_insert_seq_on_edge (e, stmts); |
121 | gsi_commit_edge_inserts (); | |
122 | SSA_NAME_DEF_STMT (res) = stmt; | |
123 | } | |
124 | ||
125 | /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */ | |
126 | ||
127 | static inline bool | |
128 | simple_copy_phi_p (gimple phi) | |
129 | { | |
130 | tree res; | |
131 | ||
132 | if (gimple_phi_num_args (phi) != 2) | |
133 | return false; | |
134 | ||
135 | res = gimple_phi_result (phi); | |
136 | return (res == gimple_phi_arg_def (phi, 0) | |
137 | || res == gimple_phi_arg_def (phi, 1)); | |
138 | } | |
139 | ||
140 | /* Returns true when the phi node at position PSI is a reduction phi | |
141 | node in REGION. Otherwise moves the pointer PSI to the next phi to | |
142 | be considered. */ | |
143 | ||
144 | static bool | |
145 | reduction_phi_p (sese region, gimple_stmt_iterator *psi) | |
146 | { | |
147 | loop_p loop; | |
2abae5f1 SP |
148 | gimple phi = gsi_stmt (*psi); |
149 | tree res = gimple_phi_result (phi); | |
150 | ||
2abae5f1 SP |
151 | loop = loop_containing_stmt (phi); |
152 | ||
153 | if (simple_copy_phi_p (phi)) | |
154 | { | |
a5a59b11 | 155 | /* PRE introduces phi nodes like these, for an example, |
2abae5f1 SP |
156 | see id-5.f in the fortran graphite testsuite: |
157 | ||
158 | # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)> | |
159 | */ | |
160 | remove_simple_copy_phi (psi); | |
161 | return false; | |
162 | } | |
163 | ||
87b28340 | 164 | if (scev_analyzable_p (res, region)) |
2abae5f1 | 165 | { |
87b28340 SP |
166 | tree scev = scalar_evolution_in_region (region, loop, res); |
167 | ||
168 | if (evolution_function_is_invariant_p (scev, loop->num)) | |
7cc4ff8d SP |
169 | remove_invariant_phi (region, psi); |
170 | else | |
171 | gsi_next (psi); | |
172 | ||
2abae5f1 SP |
173 | return false; |
174 | } | |
175 | ||
2abae5f1 SP |
176 | /* All the other cases are considered reductions. */ |
177 | return true; | |
178 | } | |
179 | ||
2abae5f1 SP |
180 | /* Store the GRAPHITE representation of BB. */ |
181 | ||
182 | static gimple_bb_p | |
9771b263 | 183 | new_gimple_bb (basic_block bb, vec<data_reference_p> drs) |
2abae5f1 SP |
184 | { |
185 | struct gimple_bb *gbb; | |
186 | ||
187 | gbb = XNEW (struct gimple_bb); | |
188 | bb->aux = gbb; | |
189 | GBB_BB (gbb) = bb; | |
190 | GBB_DATA_REFS (gbb) = drs; | |
9771b263 DN |
191 | GBB_CONDITIONS (gbb).create (0); |
192 | GBB_CONDITION_CASES (gbb).create (0); | |
2abae5f1 SP |
193 | |
194 | return gbb; | |
195 | } | |
196 | ||
1825f9a2 | 197 | static void |
9771b263 | 198 | free_data_refs_aux (vec<data_reference_p> datarefs) |
1825f9a2 LF |
199 | { |
200 | unsigned int i; | |
201 | struct data_reference *dr; | |
fb00d28e | 202 | |
9771b263 | 203 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
fb00d28e | 204 | if (dr->aux) |
1825f9a2 | 205 | { |
2b178a5f | 206 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); |
fb00d28e | 207 | |
04695783 | 208 | free (bap->alias_set); |
fb00d28e | 209 | |
2b178a5f | 210 | free (bap); |
1825f9a2 LF |
211 | dr->aux = NULL; |
212 | } | |
213 | } | |
2abae5f1 SP |
214 | /* Frees GBB. */ |
215 | ||
216 | static void | |
217 | free_gimple_bb (struct gimple_bb *gbb) | |
218 | { | |
1825f9a2 | 219 | free_data_refs_aux (GBB_DATA_REFS (gbb)); |
2abae5f1 SP |
220 | free_data_refs (GBB_DATA_REFS (gbb)); |
221 | ||
9771b263 DN |
222 | GBB_CONDITIONS (gbb).release (); |
223 | GBB_CONDITION_CASES (gbb).release (); | |
2abae5f1 SP |
224 | GBB_BB (gbb)->aux = 0; |
225 | XDELETE (gbb); | |
226 | } | |
227 | ||
228 | /* Deletes all gimple bbs in SCOP. */ | |
229 | ||
230 | static void | |
231 | remove_gbbs_in_scop (scop_p scop) | |
232 | { | |
233 | int i; | |
234 | poly_bb_p pbb; | |
235 | ||
9771b263 | 236 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
237 | free_gimple_bb (PBB_BLACK_BOX (pbb)); |
238 | } | |
239 | ||
240 | /* Deletes all scops in SCOPS. */ | |
241 | ||
242 | void | |
9771b263 | 243 | free_scops (vec<scop_p> scops) |
2abae5f1 SP |
244 | { |
245 | int i; | |
246 | scop_p scop; | |
247 | ||
9771b263 | 248 | FOR_EACH_VEC_ELT (scops, i, scop) |
2abae5f1 SP |
249 | { |
250 | remove_gbbs_in_scop (scop); | |
251 | free_sese (SCOP_REGION (scop)); | |
252 | free_scop (scop); | |
253 | } | |
254 | ||
9771b263 | 255 | scops.release (); |
2abae5f1 SP |
256 | } |
257 | ||
5c640e29 SP |
258 | /* Same as outermost_loop_in_sese, returns the outermost loop |
259 | containing BB in REGION, but makes sure that the returned loop | |
260 | belongs to the REGION, and so this returns the first loop in the | |
261 | REGION when the loop containing BB does not belong to REGION. */ | |
262 | ||
263 | static loop_p | |
264 | outermost_loop_in_sese_1 (sese region, basic_block bb) | |
265 | { | |
266 | loop_p nest = outermost_loop_in_sese (region, bb); | |
267 | ||
268 | if (loop_in_sese_p (nest, region)) | |
269 | return nest; | |
270 | ||
271 | /* When the basic block BB does not belong to a loop in the region, | |
272 | return the first loop in the region. */ | |
273 | nest = nest->inner; | |
274 | while (nest) | |
275 | if (loop_in_sese_p (nest, region)) | |
276 | break; | |
277 | else | |
278 | nest = nest->next; | |
279 | ||
280 | gcc_assert (nest); | |
281 | return nest; | |
282 | } | |
283 | ||
2abae5f1 SP |
284 | /* Generates a polyhedral black box only if the bb contains interesting |
285 | information. */ | |
286 | ||
efa21390 SP |
287 | static gimple_bb_p |
288 | try_generate_gimple_bb (scop_p scop, basic_block bb) | |
2abae5f1 | 289 | { |
9771b263 DN |
290 | vec<data_reference_p> drs; |
291 | drs.create (5); | |
5c640e29 SP |
292 | sese region = SCOP_REGION (scop); |
293 | loop_p nest = outermost_loop_in_sese_1 (region, bb); | |
2abae5f1 SP |
294 | gimple_stmt_iterator gsi; |
295 | ||
296 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
a3201927 AO |
297 | { |
298 | gimple stmt = gsi_stmt (gsi); | |
5c640e29 SP |
299 | loop_p loop; |
300 | ||
301 | if (is_gimple_debug (stmt)) | |
302 | continue; | |
303 | ||
304 | loop = loop_containing_stmt (stmt); | |
305 | if (!loop_in_sese_p (loop, region)) | |
306 | loop = nest; | |
307 | ||
308 | graphite_find_data_references_in_stmt (nest, loop, stmt, &drs); | |
a3201927 | 309 | } |
2abae5f1 | 310 | |
efa21390 | 311 | return new_gimple_bb (bb, drs); |
2abae5f1 SP |
312 | } |
313 | ||
314 | /* Returns true if all predecessors of BB, that are not dominated by BB, are | |
315 | marked in MAP. The predecessors dominated by BB are loop latches and will | |
316 | be handled after BB. */ | |
317 | ||
318 | static bool | |
319 | all_non_dominated_preds_marked_p (basic_block bb, sbitmap map) | |
320 | { | |
321 | edge e; | |
322 | edge_iterator ei; | |
323 | ||
324 | FOR_EACH_EDGE (e, ei, bb->preds) | |
d7c028c0 | 325 | if (!bitmap_bit_p (map, e->src->index) |
2abae5f1 SP |
326 | && !dominated_by_p (CDI_DOMINATORS, e->src, bb)) |
327 | return false; | |
328 | ||
329 | return true; | |
330 | } | |
331 | ||
332 | /* Compare the depth of two basic_block's P1 and P2. */ | |
333 | ||
334 | static int | |
335 | compare_bb_depths (const void *p1, const void *p2) | |
336 | { | |
337 | const_basic_block const bb1 = *(const_basic_block const*)p1; | |
338 | const_basic_block const bb2 = *(const_basic_block const*)p2; | |
339 | int d1 = loop_depth (bb1->loop_father); | |
340 | int d2 = loop_depth (bb2->loop_father); | |
341 | ||
342 | if (d1 < d2) | |
343 | return 1; | |
344 | ||
345 | if (d1 > d2) | |
346 | return -1; | |
347 | ||
348 | return 0; | |
349 | } | |
350 | ||
351 | /* Sort the basic blocks from DOM such that the first are the ones at | |
352 | a deepest loop level. */ | |
353 | ||
354 | static void | |
9771b263 | 355 | graphite_sort_dominated_info (vec<basic_block> dom) |
2abae5f1 | 356 | { |
9771b263 | 357 | dom.qsort (compare_bb_depths); |
2abae5f1 SP |
358 | } |
359 | ||
360 | /* Recursive helper function for build_scops_bbs. */ | |
361 | ||
362 | static void | |
efa21390 | 363 | build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb) |
2abae5f1 SP |
364 | { |
365 | sese region = SCOP_REGION (scop); | |
9771b263 | 366 | vec<basic_block> dom; |
efa21390 | 367 | poly_bb_p pbb; |
2abae5f1 | 368 | |
d7c028c0 | 369 | if (bitmap_bit_p (visited, bb->index) |
2abae5f1 SP |
370 | || !bb_in_sese_p (bb, region)) |
371 | return; | |
372 | ||
efa21390 | 373 | pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb)); |
9771b263 | 374 | SCOP_BBS (scop).safe_push (pbb); |
d7c028c0 | 375 | bitmap_set_bit (visited, bb->index); |
2abae5f1 SP |
376 | |
377 | dom = get_dominated_by (CDI_DOMINATORS, bb); | |
378 | ||
9771b263 | 379 | if (!dom.exists ()) |
2abae5f1 SP |
380 | return; |
381 | ||
382 | graphite_sort_dominated_info (dom); | |
383 | ||
9771b263 | 384 | while (!dom.is_empty ()) |
2abae5f1 SP |
385 | { |
386 | int i; | |
387 | basic_block dom_bb; | |
388 | ||
9771b263 | 389 | FOR_EACH_VEC_ELT (dom, i, dom_bb) |
2abae5f1 SP |
390 | if (all_non_dominated_preds_marked_p (dom_bb, visited)) |
391 | { | |
efa21390 | 392 | build_scop_bbs_1 (scop, visited, dom_bb); |
9771b263 | 393 | dom.unordered_remove (i); |
2abae5f1 SP |
394 | break; |
395 | } | |
396 | } | |
397 | ||
9771b263 | 398 | dom.release (); |
2abae5f1 SP |
399 | } |
400 | ||
401 | /* Gather the basic blocks belonging to the SCOP. */ | |
402 | ||
efa21390 SP |
403 | static void |
404 | build_scop_bbs (scop_p scop) | |
2abae5f1 SP |
405 | { |
406 | sbitmap visited = sbitmap_alloc (last_basic_block); | |
407 | sese region = SCOP_REGION (scop); | |
408 | ||
f61e445a | 409 | bitmap_clear (visited); |
efa21390 | 410 | build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region)); |
2abae5f1 SP |
411 | sbitmap_free (visited); |
412 | } | |
413 | ||
33ad93b9 RG |
414 | /* Return an ISL identifier for the polyhedral basic block PBB. */ |
415 | ||
416 | static isl_id * | |
417 | isl_id_for_pbb (scop_p s, poly_bb_p pbb) | |
418 | { | |
419 | char name[50]; | |
420 | snprintf (name, sizeof (name), "S_%d", pbb_index (pbb)); | |
421 | return isl_id_alloc (s->ctx, name, pbb); | |
422 | } | |
423 | ||
2abae5f1 SP |
424 | /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron. |
425 | We generate SCATTERING_DIMENSIONS scattering dimensions. | |
426 | ||
427 | CLooG 0.15.0 and previous versions require, that all | |
428 | scattering functions of one CloogProgram have the same number of | |
429 | scattering dimensions, therefore we allow to specify it. This | |
430 | should be removed in future versions of CLooG. | |
431 | ||
432 | The scattering polyhedron consists of these dimensions: scattering, | |
433 | loop_iterators, parameters. | |
434 | ||
435 | Example: | |
436 | ||
437 | | scattering_dimensions = 5 | |
438 | | used_scattering_dimensions = 3 | |
439 | | nb_iterators = 1 | |
440 | | scop_nb_params = 2 | |
441 | | | |
442 | | Schedule: | |
443 | | i | |
444 | | 4 5 | |
445 | | | |
446 | | Scattering polyhedron: | |
447 | | | |
448 | | scattering: {s1, s2, s3, s4, s5} | |
449 | | loop_iterators: {i} | |
450 | | parameters: {p1, p2} | |
451 | | | |
452 | | s1 s2 s3 s4 s5 i p1 p2 1 | |
453 | | 1 0 0 0 0 0 0 0 -4 = 0 | |
454 | | 0 1 0 0 0 -1 0 0 0 = 0 | |
455 | | 0 0 1 0 0 0 0 0 -5 = 0 */ | |
456 | ||
457 | static void | |
33ad93b9 | 458 | build_pbb_scattering_polyhedrons (isl_aff *static_sched, |
2abae5f1 SP |
459 | poly_bb_p pbb, int scattering_dimensions) |
460 | { | |
461 | int i; | |
2abae5f1 SP |
462 | int nb_iterators = pbb_dim_iter_domain (pbb); |
463 | int used_scattering_dimensions = nb_iterators * 2 + 1; | |
33ad93b9 RG |
464 | isl_int val; |
465 | isl_space *dc, *dm; | |
2abae5f1 SP |
466 | |
467 | gcc_assert (scattering_dimensions >= used_scattering_dimensions); | |
468 | ||
33ad93b9 | 469 | isl_int_init (val); |
2abae5f1 | 470 | |
33ad93b9 RG |
471 | dc = isl_set_get_space (pbb->domain); |
472 | dm = isl_space_add_dims (isl_space_from_domain (dc), | |
473 | isl_dim_out, scattering_dimensions); | |
474 | pbb->schedule = isl_map_universe (dm); | |
2abae5f1 SP |
475 | |
476 | for (i = 0; i < scattering_dimensions; i++) | |
477 | { | |
2abae5f1 SP |
478 | /* Textual order inside this loop. */ |
479 | if ((i % 2) == 0) | |
480 | { | |
33ad93b9 RG |
481 | isl_constraint *c = isl_equality_alloc |
482 | (isl_local_space_from_space (isl_map_get_space (pbb->schedule))); | |
483 | ||
484 | if (0 != isl_aff_get_coefficient (static_sched, isl_dim_in, | |
485 | i / 2, &val)) | |
486 | gcc_unreachable (); | |
487 | ||
488 | isl_int_neg (val, val); | |
489 | c = isl_constraint_set_constant (c, val); | |
490 | c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1); | |
491 | pbb->schedule = isl_map_add_constraint (pbb->schedule, c); | |
2abae5f1 SP |
492 | } |
493 | ||
494 | /* Iterations of this loop. */ | |
495 | else /* if ((i % 2) == 1) */ | |
496 | { | |
497 | int loop = (i - 1) / 2; | |
33ad93b9 RG |
498 | pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop, |
499 | isl_dim_out, i); | |
2abae5f1 | 500 | } |
2abae5f1 SP |
501 | } |
502 | ||
33ad93b9 | 503 | isl_int_clear (val); |
2abae5f1 | 504 | |
33ad93b9 | 505 | pbb->transformed = isl_map_copy (pbb->schedule); |
2abae5f1 SP |
506 | } |
507 | ||
508 | /* Build for BB the static schedule. | |
509 | ||
510 | The static schedule is a Dewey numbering of the abstract syntax | |
511 | tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification | |
512 | ||
513 | The following example informally defines the static schedule: | |
514 | ||
515 | A | |
516 | for (i: ...) | |
517 | { | |
518 | for (j: ...) | |
519 | { | |
520 | B | |
521 | C | |
522 | } | |
523 | ||
524 | for (k: ...) | |
525 | { | |
526 | D | |
527 | E | |
528 | } | |
529 | } | |
530 | F | |
531 | ||
532 | Static schedules for A to F: | |
533 | ||
534 | DEPTH | |
535 | 0 1 2 | |
536 | A 0 | |
537 | B 1 0 0 | |
538 | C 1 0 1 | |
539 | D 1 1 0 | |
540 | E 1 1 1 | |
541 | F 2 | |
542 | */ | |
543 | ||
544 | static void | |
545 | build_scop_scattering (scop_p scop) | |
546 | { | |
547 | int i; | |
548 | poly_bb_p pbb; | |
549 | gimple_bb_p previous_gbb = NULL; | |
33ad93b9 RG |
550 | isl_space *dc = isl_set_get_space (scop->context); |
551 | isl_aff *static_sched; | |
2abae5f1 | 552 | |
0fc822d0 | 553 | dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun)); |
33ad93b9 | 554 | static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc)); |
2abae5f1 SP |
555 | |
556 | /* We have to start schedules at 0 on the first component and | |
557 | because we cannot compare_prefix_loops against a previous loop, | |
558 | prefix will be equal to zero, and that index will be | |
559 | incremented before copying. */ | |
33ad93b9 | 560 | static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1); |
2abae5f1 | 561 | |
9771b263 | 562 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
563 | { |
564 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); | |
2abae5f1 SP |
565 | int prefix; |
566 | int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1; | |
567 | ||
568 | if (previous_gbb) | |
569 | prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb); | |
570 | else | |
571 | prefix = 0; | |
572 | ||
573 | previous_gbb = gbb; | |
2abae5f1 | 574 | |
33ad93b9 RG |
575 | static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, |
576 | prefix, 1); | |
577 | build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims); | |
578 | } | |
579 | ||
580 | isl_aff_free (static_sched); | |
581 | } | |
582 | ||
583 | static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space); | |
584 | ||
585 | /* Extract an affine expression from the chain of recurrence E. */ | |
2abae5f1 | 586 | |
33ad93b9 RG |
587 | static isl_pw_aff * |
588 | extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space) | |
589 | { | |
590 | isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space)); | |
591 | isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space)); | |
592 | isl_local_space *ls = isl_local_space_from_space (space); | |
0fc822d0 | 593 | unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1; |
33ad93b9 RG |
594 | isl_aff *loop = isl_aff_set_coefficient_si |
595 | (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1); | |
596 | isl_pw_aff *l = isl_pw_aff_from_aff (loop); | |
597 | ||
598 | /* Before multiplying, make sure that the result is affine. */ | |
599 | gcc_assert (isl_pw_aff_is_cst (rhs) | |
600 | || isl_pw_aff_is_cst (l)); | |
601 | ||
602 | return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l)); | |
603 | } | |
604 | ||
605 | /* Extract an affine expression from the mult_expr E. */ | |
606 | ||
607 | static isl_pw_aff * | |
608 | extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space) | |
609 | { | |
610 | isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0), | |
611 | isl_space_copy (space)); | |
612 | isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
2abae5f1 | 613 | |
33ad93b9 RG |
614 | if (!isl_pw_aff_is_cst (lhs) |
615 | && !isl_pw_aff_is_cst (rhs)) | |
616 | { | |
617 | isl_pw_aff_free (lhs); | |
618 | isl_pw_aff_free (rhs); | |
619 | return NULL; | |
2abae5f1 SP |
620 | } |
621 | ||
33ad93b9 | 622 | return isl_pw_aff_mul (lhs, rhs); |
2abae5f1 SP |
623 | } |
624 | ||
33ad93b9 | 625 | /* Return an ISL identifier from the name of the ssa_name E. */ |
2abae5f1 | 626 | |
33ad93b9 RG |
627 | static isl_id * |
628 | isl_id_for_ssa_name (scop_p s, tree e) | |
2abae5f1 | 629 | { |
33ad93b9 RG |
630 | const char *name = get_name (e); |
631 | isl_id *id; | |
632 | ||
633 | if (name) | |
634 | id = isl_id_alloc (s->ctx, name, e); | |
635 | else | |
636 | { | |
637 | char name1[50]; | |
638 | snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e)); | |
639 | id = isl_id_alloc (s->ctx, name1, e); | |
640 | } | |
2abae5f1 | 641 | |
33ad93b9 RG |
642 | return id; |
643 | } | |
2abae5f1 | 644 | |
33ad93b9 | 645 | /* Return an ISL identifier for the data reference DR. */ |
2abae5f1 | 646 | |
33ad93b9 RG |
647 | static isl_id * |
648 | isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED) | |
649 | { | |
650 | /* Data references all get the same isl_id. They need to be comparable | |
651 | and are distinguished through the first dimension, which contains the | |
652 | alias set number. */ | |
653 | return isl_id_alloc (s->ctx, "", 0); | |
2abae5f1 SP |
654 | } |
655 | ||
33ad93b9 | 656 | /* Extract an affine expression from the ssa_name E. */ |
2abae5f1 | 657 | |
33ad93b9 RG |
658 | static isl_pw_aff * |
659 | extract_affine_name (scop_p s, tree e, __isl_take isl_space *space) | |
2abae5f1 | 660 | { |
33ad93b9 RG |
661 | isl_aff *aff; |
662 | isl_set *dom; | |
663 | isl_id *id; | |
664 | int dimension; | |
665 | ||
666 | id = isl_id_for_ssa_name (s, e); | |
667 | dimension = isl_space_find_dim_by_id (space, isl_dim_param, id); | |
668 | isl_id_free(id); | |
669 | dom = isl_set_universe (isl_space_copy (space)); | |
670 | aff = isl_aff_zero_on_domain (isl_local_space_from_space (space)); | |
671 | aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1); | |
672 | return isl_pw_aff_alloc (dom, aff); | |
673 | } | |
2abae5f1 | 674 | |
33ad93b9 | 675 | /* Extract an affine expression from the gmp constant G. */ |
2abae5f1 | 676 | |
33ad93b9 RG |
677 | static isl_pw_aff * |
678 | extract_affine_gmp (mpz_t g, __isl_take isl_space *space) | |
679 | { | |
680 | isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space)); | |
681 | isl_aff *aff = isl_aff_zero_on_domain (ls); | |
682 | isl_set *dom = isl_set_universe (space); | |
683 | isl_int v; | |
2abae5f1 | 684 | |
33ad93b9 RG |
685 | isl_int_init (v); |
686 | isl_int_set_gmp (v, g); | |
687 | aff = isl_aff_add_constant (aff, v); | |
688 | isl_int_clear (v); | |
2abae5f1 | 689 | |
33ad93b9 | 690 | return isl_pw_aff_alloc (dom, aff); |
2abae5f1 SP |
691 | } |
692 | ||
33ad93b9 | 693 | /* Extract an affine expression from the integer_cst E. */ |
2abae5f1 | 694 | |
33ad93b9 RG |
695 | static isl_pw_aff * |
696 | extract_affine_int (tree e, __isl_take isl_space *space) | |
697 | { | |
698 | isl_pw_aff *res; | |
699 | mpz_t g; | |
700 | ||
701 | mpz_init (g); | |
702 | tree_int_to_gmp (e, g); | |
703 | res = extract_affine_gmp (g, space); | |
704 | mpz_clear (g); | |
705 | ||
706 | return res; | |
707 | } | |
708 | ||
709 | /* Compute pwaff mod 2^width. */ | |
710 | ||
711 | static isl_pw_aff * | |
712 | wrap (isl_pw_aff *pwaff, unsigned width) | |
2abae5f1 | 713 | { |
33ad93b9 | 714 | isl_int mod; |
2abae5f1 | 715 | |
33ad93b9 RG |
716 | isl_int_init (mod); |
717 | isl_int_set_si (mod, 1); | |
718 | isl_int_mul_2exp (mod, mod, width); | |
2abae5f1 | 719 | |
33ad93b9 | 720 | pwaff = isl_pw_aff_mod (pwaff, mod); |
2abae5f1 | 721 | |
33ad93b9 RG |
722 | isl_int_clear (mod); |
723 | ||
724 | return pwaff; | |
2abae5f1 SP |
725 | } |
726 | ||
2abae5f1 SP |
727 | /* When parameter NAME is in REGION, returns its index in SESE_PARAMS. |
728 | Otherwise returns -1. */ | |
729 | ||
730 | static inline int | |
731 | parameter_index_in_region_1 (tree name, sese region) | |
732 | { | |
733 | int i; | |
734 | tree p; | |
735 | ||
736 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
737 | ||
9771b263 | 738 | FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p) |
2abae5f1 SP |
739 | if (p == name) |
740 | return i; | |
741 | ||
742 | return -1; | |
743 | } | |
744 | ||
745 | /* When the parameter NAME is in REGION, returns its index in | |
746 | SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS | |
747 | and returns the index of NAME. */ | |
748 | ||
749 | static int | |
750 | parameter_index_in_region (tree name, sese region) | |
751 | { | |
752 | int i; | |
753 | ||
754 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
755 | ||
756 | i = parameter_index_in_region_1 (name, region); | |
757 | if (i != -1) | |
758 | return i; | |
759 | ||
760 | gcc_assert (SESE_ADD_PARAMS (region)); | |
761 | ||
9771b263 DN |
762 | i = SESE_PARAMS (region).length (); |
763 | SESE_PARAMS (region).safe_push (name); | |
2abae5f1 SP |
764 | return i; |
765 | } | |
766 | ||
33ad93b9 | 767 | /* Extract an affine expression from the tree E in the scop S. */ |
2abae5f1 | 768 | |
33ad93b9 RG |
769 | static isl_pw_aff * |
770 | extract_affine (scop_p s, tree e, __isl_take isl_space *space) | |
2abae5f1 | 771 | { |
33ad93b9 RG |
772 | isl_pw_aff *lhs, *rhs, *res; |
773 | tree type; | |
774 | ||
775 | if (e == chrec_dont_know) { | |
776 | isl_space_free (space); | |
777 | return NULL; | |
778 | } | |
2abae5f1 SP |
779 | |
780 | switch (TREE_CODE (e)) | |
781 | { | |
782 | case POLYNOMIAL_CHREC: | |
33ad93b9 | 783 | res = extract_affine_chrec (s, e, space); |
2abae5f1 SP |
784 | break; |
785 | ||
786 | case MULT_EXPR: | |
33ad93b9 | 787 | res = extract_affine_mul (s, e, space); |
2abae5f1 SP |
788 | break; |
789 | ||
790 | case PLUS_EXPR: | |
791 | case POINTER_PLUS_EXPR: | |
33ad93b9 RG |
792 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); |
793 | rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
794 | res = isl_pw_aff_add (lhs, rhs); | |
2abae5f1 SP |
795 | break; |
796 | ||
797 | case MINUS_EXPR: | |
33ad93b9 RG |
798 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); |
799 | rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
800 | res = isl_pw_aff_sub (lhs, rhs); | |
801 | break; | |
2abae5f1 SP |
802 | |
803 | case NEGATE_EXPR: | |
33ad93b9 RG |
804 | case BIT_NOT_EXPR: |
805 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); | |
806 | rhs = extract_affine (s, integer_minus_one_node, space); | |
807 | res = isl_pw_aff_mul (lhs, rhs); | |
808 | break; | |
2abae5f1 | 809 | |
33ad93b9 RG |
810 | case SSA_NAME: |
811 | gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s))); | |
812 | res = extract_affine_name (s, e, space); | |
813 | break; | |
2abae5f1 | 814 | |
33ad93b9 RG |
815 | case INTEGER_CST: |
816 | res = extract_affine_int (e, space); | |
817 | /* No need to wrap a single integer. */ | |
818 | return res; | |
2abae5f1 | 819 | |
33ad93b9 RG |
820 | CASE_CONVERT: |
821 | case NON_LVALUE_EXPR: | |
822 | res = extract_affine (s, TREE_OPERAND (e, 0), space); | |
823 | break; | |
2abae5f1 | 824 | |
33ad93b9 RG |
825 | default: |
826 | gcc_unreachable (); | |
827 | break; | |
828 | } | |
2abae5f1 | 829 | |
33ad93b9 RG |
830 | type = TREE_TYPE (e); |
831 | if (TYPE_UNSIGNED (type)) | |
832 | res = wrap (res, TYPE_PRECISION (type)); | |
2abae5f1 | 833 | |
33ad93b9 RG |
834 | return res; |
835 | } | |
2abae5f1 | 836 | |
33ad93b9 RG |
837 | /* In the context of sese S, scan the expression E and translate it to |
838 | a linear expression C. When parsing a symbolic multiplication, K | |
839 | represents the constant multiplier of an expression containing | |
840 | parameters. */ | |
2abae5f1 | 841 | |
33ad93b9 RG |
842 | static void |
843 | scan_tree_for_params (sese s, tree e) | |
844 | { | |
845 | if (e == chrec_dont_know) | |
846 | return; | |
2abae5f1 | 847 | |
33ad93b9 RG |
848 | switch (TREE_CODE (e)) |
849 | { | |
850 | case POLYNOMIAL_CHREC: | |
851 | scan_tree_for_params (s, CHREC_LEFT (e)); | |
852 | break; | |
2abae5f1 | 853 | |
33ad93b9 RG |
854 | case MULT_EXPR: |
855 | if (chrec_contains_symbols (TREE_OPERAND (e, 0))) | |
856 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); | |
857 | else | |
858 | scan_tree_for_params (s, TREE_OPERAND (e, 1)); | |
859 | break; | |
2abae5f1 | 860 | |
33ad93b9 RG |
861 | case PLUS_EXPR: |
862 | case POINTER_PLUS_EXPR: | |
863 | case MINUS_EXPR: | |
864 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); | |
865 | scan_tree_for_params (s, TREE_OPERAND (e, 1)); | |
2abae5f1 SP |
866 | break; |
867 | ||
33ad93b9 RG |
868 | case NEGATE_EXPR: |
869 | case BIT_NOT_EXPR: | |
2abae5f1 SP |
870 | CASE_CONVERT: |
871 | case NON_LVALUE_EXPR: | |
33ad93b9 | 872 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); |
2abae5f1 SP |
873 | break; |
874 | ||
33ad93b9 RG |
875 | case SSA_NAME: |
876 | parameter_index_in_region (e, s); | |
877 | break; | |
878 | ||
879 | case INTEGER_CST: | |
f4a2e571 SP |
880 | case ADDR_EXPR: |
881 | break; | |
882 | ||
2abae5f1 SP |
883 | default: |
884 | gcc_unreachable (); | |
885 | break; | |
886 | } | |
887 | } | |
888 | ||
2abae5f1 SP |
889 | /* Find parameters with respect to REGION in BB. We are looking in memory |
890 | access functions, conditions and loop bounds. */ | |
891 | ||
892 | static void | |
893 | find_params_in_bb (sese region, gimple_bb_p gbb) | |
894 | { | |
895 | int i; | |
54fc808a | 896 | unsigned j; |
2abae5f1 SP |
897 | data_reference_p dr; |
898 | gimple stmt; | |
899 | loop_p loop = GBB_BB (gbb)->loop_father; | |
2abae5f1 | 900 | |
54fc808a | 901 | /* Find parameters in the access functions of data references. */ |
9771b263 | 902 | FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr) |
54fc808a | 903 | for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++) |
33ad93b9 | 904 | scan_tree_for_params (region, DR_ACCESS_FN (dr, j)); |
2abae5f1 SP |
905 | |
906 | /* Find parameters in conditional statements. */ | |
9771b263 | 907 | FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt) |
2abae5f1 | 908 | { |
2abae5f1 SP |
909 | tree lhs = scalar_evolution_in_region (region, loop, |
910 | gimple_cond_lhs (stmt)); | |
911 | tree rhs = scalar_evolution_in_region (region, loop, | |
912 | gimple_cond_rhs (stmt)); | |
913 | ||
33ad93b9 RG |
914 | scan_tree_for_params (region, lhs); |
915 | scan_tree_for_params (region, rhs); | |
2abae5f1 SP |
916 | } |
917 | } | |
918 | ||
919 | /* Record the parameters used in the SCOP. A variable is a parameter | |
920 | in a scop if it does not vary during the execution of that scop. */ | |
921 | ||
922 | static void | |
923 | find_scop_parameters (scop_p scop) | |
924 | { | |
925 | poly_bb_p pbb; | |
926 | unsigned i; | |
927 | sese region = SCOP_REGION (scop); | |
928 | struct loop *loop; | |
33ad93b9 | 929 | int nbp; |
2abae5f1 SP |
930 | |
931 | /* Find the parameters used in the loop bounds. */ | |
9771b263 | 932 | FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop) |
2abae5f1 SP |
933 | { |
934 | tree nb_iters = number_of_latch_executions (loop); | |
935 | ||
936 | if (!chrec_contains_symbols (nb_iters)) | |
937 | continue; | |
938 | ||
939 | nb_iters = scalar_evolution_in_region (region, loop, nb_iters); | |
33ad93b9 | 940 | scan_tree_for_params (region, nb_iters); |
2abae5f1 SP |
941 | } |
942 | ||
2abae5f1 | 943 | /* Find the parameters used in data accesses. */ |
9771b263 | 944 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
945 | find_params_in_bb (region, PBB_BLACK_BOX (pbb)); |
946 | ||
33ad93b9 RG |
947 | nbp = sese_nb_params (region); |
948 | scop_set_nb_params (scop, nbp); | |
2abae5f1 | 949 | SESE_ADD_PARAMS (region) = false; |
62e475c5 | 950 | |
a5a59b11 | 951 | { |
33ad93b9 RG |
952 | tree e; |
953 | isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0); | |
a5a59b11 | 954 | |
9771b263 | 955 | FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e) |
33ad93b9 RG |
956 | space = isl_space_set_dim_id (space, isl_dim_param, i, |
957 | isl_id_for_ssa_name (scop, e)); | |
a5a59b11 | 958 | |
33ad93b9 | 959 | scop->context = isl_set_universe (space); |
a5a59b11 | 960 | } |
a5a59b11 SP |
961 | } |
962 | ||
2abae5f1 SP |
963 | /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives |
964 | the constraints for the surrounding loops. */ | |
965 | ||
966 | static void | |
967 | build_loop_iteration_domains (scop_p scop, struct loop *loop, | |
33ad93b9 RG |
968 | int nb, |
969 | isl_set *outer, isl_set **doms) | |
2abae5f1 | 970 | { |
2abae5f1 | 971 | tree nb_iters = number_of_latch_executions (loop); |
2abae5f1 SP |
972 | sese region = SCOP_REGION (scop); |
973 | ||
33ad93b9 RG |
974 | isl_set *inner = isl_set_copy (outer); |
975 | isl_space *space; | |
976 | isl_constraint *c; | |
977 | int pos = isl_set_dim (outer, isl_dim_set); | |
978 | isl_int v; | |
979 | mpz_t g; | |
980 | ||
981 | mpz_init (g); | |
982 | isl_int_init (v); | |
983 | ||
984 | inner = isl_set_add_dims (inner, isl_dim_set, 1); | |
985 | space = isl_set_get_space (inner); | |
2abae5f1 SP |
986 | |
987 | /* 0 <= loop_i */ | |
33ad93b9 RG |
988 | c = isl_inequality_alloc |
989 | (isl_local_space_from_space (isl_space_copy (space))); | |
990 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1); | |
991 | inner = isl_set_add_constraint (inner, c); | |
2abae5f1 | 992 | |
33ad93b9 | 993 | /* loop_i <= cst_nb_iters */ |
2abae5f1 SP |
994 | if (TREE_CODE (nb_iters) == INTEGER_CST) |
995 | { | |
33ad93b9 RG |
996 | c = isl_inequality_alloc |
997 | (isl_local_space_from_space(isl_space_copy (space))); | |
998 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1); | |
999 | tree_int_to_gmp (nb_iters, g); | |
1000 | isl_int_set_gmp (v, g); | |
1001 | c = isl_constraint_set_constant (c, v); | |
1002 | inner = isl_set_add_constraint (inner, c); | |
2abae5f1 | 1003 | } |
33ad93b9 RG |
1004 | |
1005 | /* loop_i <= expr_nb_iters */ | |
2abae5f1 SP |
1006 | else if (!chrec_contains_undetermined (nb_iters)) |
1007 | { | |
62e475c5 | 1008 | double_int nit; |
33ad93b9 RG |
1009 | isl_pw_aff *aff; |
1010 | isl_set *valid; | |
1011 | isl_local_space *ls; | |
1012 | isl_aff *al; | |
1013 | isl_set *le; | |
2abae5f1 | 1014 | |
2abae5f1 | 1015 | nb_iters = scalar_evolution_in_region (region, loop, nb_iters); |
33ad93b9 RG |
1016 | |
1017 | aff = extract_affine (scop, nb_iters, isl_set_get_space (inner)); | |
1018 | valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff)); | |
1019 | valid = isl_set_project_out (valid, isl_dim_set, 0, | |
1020 | isl_set_dim (valid, isl_dim_set)); | |
1021 | scop->context = isl_set_intersect (scop->context, valid); | |
1022 | ||
1023 | ls = isl_local_space_from_space (isl_space_copy (space)); | |
1024 | al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls), | |
1025 | isl_dim_in, pos, 1); | |
1026 | le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al), | |
1027 | isl_pw_aff_copy (aff)); | |
1028 | inner = isl_set_intersect (inner, le); | |
2abae5f1 | 1029 | |
652c4c71 | 1030 | if (max_stmt_executions (loop, &nit)) |
33ad93b9 RG |
1031 | { |
1032 | /* Insert in the context the constraints from the | |
1033 | estimation of the number of iterations NIT and the | |
1034 | symbolic number of iterations (involving parameter | |
1035 | names) NB_ITERS. First, build the affine expression | |
1036 | "NIT - NB_ITERS" and then say that it is positive, | |
1037 | i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */ | |
1038 | isl_pw_aff *approx; | |
1039 | mpz_t g; | |
1040 | isl_set *x; | |
1041 | isl_constraint *c; | |
1042 | ||
1043 | mpz_init (g); | |
1044 | mpz_set_double_int (g, nit, false); | |
1045 | mpz_sub_ui (g, g, 1); | |
1046 | approx = extract_affine_gmp (g, isl_set_get_space (inner)); | |
1047 | x = isl_pw_aff_ge_set (approx, aff); | |
1048 | x = isl_set_project_out (x, isl_dim_set, 0, | |
1049 | isl_set_dim (x, isl_dim_set)); | |
1050 | scop->context = isl_set_intersect (scop->context, x); | |
1051 | ||
1052 | c = isl_inequality_alloc | |
1053 | (isl_local_space_from_space (isl_space_copy (space))); | |
1054 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1); | |
1055 | isl_int_set_gmp (v, g); | |
1056 | mpz_clear (g); | |
1057 | c = isl_constraint_set_constant (c, v); | |
1058 | inner = isl_set_add_constraint (inner, c); | |
1059 | } | |
2b91f098 RB |
1060 | else |
1061 | isl_pw_aff_free (aff); | |
2abae5f1 SP |
1062 | } |
1063 | else | |
1064 | gcc_unreachable (); | |
1065 | ||
1066 | if (loop->inner && loop_in_sese_p (loop->inner, region)) | |
33ad93b9 RG |
1067 | build_loop_iteration_domains (scop, loop->inner, nb + 1, |
1068 | isl_set_copy (inner), doms); | |
2abae5f1 SP |
1069 | |
1070 | if (nb != 0 | |
1071 | && loop->next | |
1072 | && loop_in_sese_p (loop->next, region)) | |
33ad93b9 RG |
1073 | build_loop_iteration_domains (scop, loop->next, nb, |
1074 | isl_set_copy (outer), doms); | |
2abae5f1 | 1075 | |
33ad93b9 | 1076 | doms[loop->num] = inner; |
2abae5f1 | 1077 | |
33ad93b9 RG |
1078 | isl_set_free (outer); |
1079 | isl_space_free (space); | |
1080 | isl_int_clear (v); | |
1081 | mpz_clear (g); | |
2abae5f1 SP |
1082 | } |
1083 | ||
1084 | /* Returns a linear expression for tree T evaluated in PBB. */ | |
1085 | ||
33ad93b9 RG |
1086 | static isl_pw_aff * |
1087 | create_pw_aff_from_tree (poly_bb_p pbb, tree t) | |
2abae5f1 | 1088 | { |
33ad93b9 | 1089 | scop_p scop = PBB_SCOP (pbb); |
2abae5f1 | 1090 | |
33ad93b9 | 1091 | t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t); |
2abae5f1 SP |
1092 | gcc_assert (!automatically_generated_chrec_p (t)); |
1093 | ||
33ad93b9 | 1094 | return extract_affine (scop, t, isl_set_get_space (pbb->domain)); |
2abae5f1 SP |
1095 | } |
1096 | ||
33ad93b9 RG |
1097 | /* Add conditional statement STMT to pbb. CODE is used as the comparison |
1098 | operator. This allows us to invert the condition or to handle | |
1099 | inequalities. */ | |
2abae5f1 SP |
1100 | |
1101 | static void | |
33ad93b9 | 1102 | add_condition_to_pbb (poly_bb_p pbb, gimple stmt, enum tree_code code) |
2abae5f1 | 1103 | { |
33ad93b9 RG |
1104 | isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt)); |
1105 | isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt)); | |
1106 | isl_set *cond; | |
2abae5f1 | 1107 | |
33ad93b9 | 1108 | switch (code) |
2abae5f1 | 1109 | { |
33ad93b9 RG |
1110 | case LT_EXPR: |
1111 | cond = isl_pw_aff_lt_set (lhs, rhs); | |
1112 | break; | |
2abae5f1 | 1113 | |
33ad93b9 RG |
1114 | case GT_EXPR: |
1115 | cond = isl_pw_aff_gt_set (lhs, rhs); | |
1116 | break; | |
2abae5f1 | 1117 | |
33ad93b9 RG |
1118 | case LE_EXPR: |
1119 | cond = isl_pw_aff_le_set (lhs, rhs); | |
1120 | break; | |
2abae5f1 | 1121 | |
33ad93b9 RG |
1122 | case GE_EXPR: |
1123 | cond = isl_pw_aff_ge_set (lhs, rhs); | |
1124 | break; | |
2abae5f1 | 1125 | |
33ad93b9 RG |
1126 | case EQ_EXPR: |
1127 | cond = isl_pw_aff_eq_set (lhs, rhs); | |
1128 | break; | |
2abae5f1 | 1129 | |
33ad93b9 RG |
1130 | case NE_EXPR: |
1131 | cond = isl_pw_aff_ne_set (lhs, rhs); | |
1132 | break; | |
2abae5f1 | 1133 | |
33ad93b9 RG |
1134 | default: |
1135 | isl_pw_aff_free(lhs); | |
1136 | isl_pw_aff_free(rhs); | |
1137 | return; | |
2abae5f1 | 1138 | } |
33ad93b9 RG |
1139 | |
1140 | cond = isl_set_coalesce (cond); | |
1141 | cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain)); | |
1142 | pbb->domain = isl_set_intersect (pbb->domain, cond); | |
2abae5f1 SP |
1143 | } |
1144 | ||
1145 | /* Add conditions to the domain of PBB. */ | |
1146 | ||
1147 | static void | |
1148 | add_conditions_to_domain (poly_bb_p pbb) | |
1149 | { | |
1150 | unsigned int i; | |
1151 | gimple stmt; | |
1152 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); | |
2abae5f1 | 1153 | |
9771b263 | 1154 | if (GBB_CONDITIONS (gbb).is_empty ()) |
2abae5f1 SP |
1155 | return; |
1156 | ||
9771b263 | 1157 | FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt) |
2abae5f1 SP |
1158 | switch (gimple_code (stmt)) |
1159 | { | |
1160 | case GIMPLE_COND: | |
1161 | { | |
1162 | enum tree_code code = gimple_cond_code (stmt); | |
1163 | ||
1164 | /* The conditions for ELSE-branches are inverted. */ | |
9771b263 | 1165 | if (!GBB_CONDITION_CASES (gbb)[i]) |
2abae5f1 SP |
1166 | code = invert_tree_comparison (code, false); |
1167 | ||
1168 | add_condition_to_pbb (pbb, stmt, code); | |
1169 | break; | |
1170 | } | |
1171 | ||
1172 | case GIMPLE_SWITCH: | |
33ad93b9 | 1173 | /* Switch statements are not supported right now - fall through. */ |
2abae5f1 SP |
1174 | |
1175 | default: | |
1176 | gcc_unreachable (); | |
1177 | break; | |
1178 | } | |
1179 | } | |
1180 | ||
efa21390 SP |
1181 | /* Traverses all the GBBs of the SCOP and add their constraints to the |
1182 | iteration domains. */ | |
1183 | ||
1184 | static void | |
1185 | add_conditions_to_constraints (scop_p scop) | |
1186 | { | |
1187 | int i; | |
1188 | poly_bb_p pbb; | |
1189 | ||
9771b263 | 1190 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
efa21390 SP |
1191 | add_conditions_to_domain (pbb); |
1192 | } | |
1193 | ||
2abae5f1 SP |
1194 | /* Structure used to pass data to dom_walk. */ |
1195 | ||
1196 | struct bsc | |
1197 | { | |
9771b263 | 1198 | vec<gimple> *conditions, *cases; |
2abae5f1 SP |
1199 | sese region; |
1200 | }; | |
1201 | ||
c12e2a5f SP |
1202 | /* Returns a COND_EXPR statement when BB has a single predecessor, the |
1203 | edge between BB and its predecessor is not a loop exit edge, and | |
1204 | the last statement of the single predecessor is a COND_EXPR. */ | |
2abae5f1 SP |
1205 | |
1206 | static gimple | |
c12e2a5f | 1207 | single_pred_cond_non_loop_exit (basic_block bb) |
2abae5f1 SP |
1208 | { |
1209 | if (single_pred_p (bb)) | |
1210 | { | |
1211 | edge e = single_pred_edge (bb); | |
1212 | basic_block pred = e->src; | |
c12e2a5f SP |
1213 | gimple stmt; |
1214 | ||
1215 | if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father)) | |
1216 | return NULL; | |
1217 | ||
1218 | stmt = last_stmt (pred); | |
2abae5f1 SP |
1219 | |
1220 | if (stmt && gimple_code (stmt) == GIMPLE_COND) | |
1221 | return stmt; | |
1222 | } | |
c12e2a5f | 1223 | |
2abae5f1 SP |
1224 | return NULL; |
1225 | } | |
1226 | ||
1227 | /* Call-back for dom_walk executed before visiting the dominated | |
1228 | blocks. */ | |
1229 | ||
1230 | static void | |
1231 | build_sese_conditions_before (struct dom_walk_data *dw_data, | |
1232 | basic_block bb) | |
1233 | { | |
1234 | struct bsc *data = (struct bsc *) dw_data->global_data; | |
9771b263 DN |
1235 | vec<gimple> *conditions = data->conditions; |
1236 | vec<gimple> *cases = data->cases; | |
072edf07 SP |
1237 | gimple_bb_p gbb; |
1238 | gimple stmt; | |
2abae5f1 SP |
1239 | |
1240 | if (!bb_in_sese_p (bb, data->region)) | |
1241 | return; | |
1242 | ||
c12e2a5f | 1243 | stmt = single_pred_cond_non_loop_exit (bb); |
072edf07 | 1244 | |
2abae5f1 SP |
1245 | if (stmt) |
1246 | { | |
1247 | edge e = single_pred_edge (bb); | |
1248 | ||
9771b263 | 1249 | conditions->safe_push (stmt); |
2abae5f1 SP |
1250 | |
1251 | if (e->flags & EDGE_TRUE_VALUE) | |
9771b263 | 1252 | cases->safe_push (stmt); |
2abae5f1 | 1253 | else |
9771b263 | 1254 | cases->safe_push (NULL); |
2abae5f1 SP |
1255 | } |
1256 | ||
072edf07 SP |
1257 | gbb = gbb_from_bb (bb); |
1258 | ||
2abae5f1 SP |
1259 | if (gbb) |
1260 | { | |
9771b263 DN |
1261 | GBB_CONDITIONS (gbb) = conditions->copy (); |
1262 | GBB_CONDITION_CASES (gbb) = cases->copy (); | |
2abae5f1 SP |
1263 | } |
1264 | } | |
1265 | ||
1266 | /* Call-back for dom_walk executed after visiting the dominated | |
1267 | blocks. */ | |
1268 | ||
1269 | static void | |
1270 | build_sese_conditions_after (struct dom_walk_data *dw_data, | |
1271 | basic_block bb) | |
1272 | { | |
1273 | struct bsc *data = (struct bsc *) dw_data->global_data; | |
9771b263 DN |
1274 | vec<gimple> *conditions = data->conditions; |
1275 | vec<gimple> *cases = data->cases; | |
2abae5f1 SP |
1276 | |
1277 | if (!bb_in_sese_p (bb, data->region)) | |
1278 | return; | |
1279 | ||
c12e2a5f | 1280 | if (single_pred_cond_non_loop_exit (bb)) |
2abae5f1 | 1281 | { |
9771b263 DN |
1282 | conditions->pop (); |
1283 | cases->pop (); | |
2abae5f1 SP |
1284 | } |
1285 | } | |
1286 | ||
1287 | /* Record all conditions in REGION. */ | |
1288 | ||
1289 | static void | |
1290 | build_sese_conditions (sese region) | |
1291 | { | |
1292 | struct dom_walk_data walk_data; | |
9771b263 DN |
1293 | vec<gimple> conditions; |
1294 | conditions.create (3); | |
1295 | vec<gimple> cases; | |
1296 | cases.create (3); | |
2abae5f1 SP |
1297 | struct bsc data; |
1298 | ||
1299 | data.conditions = &conditions; | |
1300 | data.cases = &cases; | |
1301 | data.region = region; | |
1302 | ||
1303 | walk_data.dom_direction = CDI_DOMINATORS; | |
1304 | walk_data.initialize_block_local_data = NULL; | |
1305 | walk_data.before_dom_children = build_sese_conditions_before; | |
1306 | walk_data.after_dom_children = build_sese_conditions_after; | |
1307 | walk_data.global_data = &data; | |
1308 | walk_data.block_local_data_size = 0; | |
1309 | ||
1310 | init_walk_dominator_tree (&walk_data); | |
1311 | walk_dominator_tree (&walk_data, SESE_ENTRY_BB (region)); | |
1312 | fini_walk_dominator_tree (&walk_data); | |
1313 | ||
9771b263 DN |
1314 | conditions.release (); |
1315 | cases.release (); | |
2abae5f1 SP |
1316 | } |
1317 | ||
2abae5f1 SP |
1318 | /* Add constraints on the possible values of parameter P from the type |
1319 | of P. */ | |
1320 | ||
1321 | static void | |
33ad93b9 | 1322 | add_param_constraints (scop_p scop, graphite_dim_t p) |
2abae5f1 | 1323 | { |
9771b263 | 1324 | tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p]; |
2abae5f1 | 1325 | tree type = TREE_TYPE (parameter); |
3640d64c SP |
1326 | tree lb = NULL_TREE; |
1327 | tree ub = NULL_TREE; | |
2abae5f1 | 1328 | |
697f511d SP |
1329 | if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type)) |
1330 | lb = lower_bound_in_type (type, type); | |
1331 | else | |
1332 | lb = TYPE_MIN_VALUE (type); | |
1333 | ||
1334 | if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type)) | |
1335 | ub = upper_bound_in_type (type, type); | |
1336 | else | |
1337 | ub = TYPE_MAX_VALUE (type); | |
2abae5f1 SP |
1338 | |
1339 | if (lb) | |
1340 | { | |
33ad93b9 RG |
1341 | isl_space *space = isl_set_get_space (scop->context); |
1342 | isl_constraint *c; | |
1343 | mpz_t g; | |
1344 | isl_int v; | |
1345 | ||
1346 | c = isl_inequality_alloc (isl_local_space_from_space (space)); | |
1347 | mpz_init (g); | |
1348 | isl_int_init (v); | |
1349 | tree_int_to_gmp (lb, g); | |
1350 | isl_int_set_gmp (v, g); | |
1351 | isl_int_neg (v, v); | |
1352 | mpz_clear (g); | |
1353 | c = isl_constraint_set_constant (c, v); | |
1354 | isl_int_clear (v); | |
1355 | c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1); | |
1356 | ||
1357 | scop->context = isl_set_add_constraint (scop->context, c); | |
2abae5f1 SP |
1358 | } |
1359 | ||
1360 | if (ub) | |
1361 | { | |
33ad93b9 RG |
1362 | isl_space *space = isl_set_get_space (scop->context); |
1363 | isl_constraint *c; | |
1364 | mpz_t g; | |
1365 | isl_int v; | |
1366 | ||
1367 | c = isl_inequality_alloc (isl_local_space_from_space (space)); | |
1368 | ||
1369 | mpz_init (g); | |
1370 | isl_int_init (v); | |
1371 | tree_int_to_gmp (ub, g); | |
1372 | isl_int_set_gmp (v, g); | |
1373 | mpz_clear (g); | |
1374 | c = isl_constraint_set_constant (c, v); | |
1375 | isl_int_clear (v); | |
1376 | c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1); | |
1377 | ||
1378 | scop->context = isl_set_add_constraint (scop->context, c); | |
2abae5f1 SP |
1379 | } |
1380 | } | |
1381 | ||
1382 | /* Build the context of the SCOP. The context usually contains extra | |
1383 | constraints that are added to the iteration domains that constrain | |
1384 | some parameters. */ | |
1385 | ||
1386 | static void | |
1387 | build_scop_context (scop_p scop) | |
1388 | { | |
2abae5f1 SP |
1389 | graphite_dim_t p, n = scop_nb_params (scop); |
1390 | ||
2abae5f1 | 1391 | for (p = 0; p < n; p++) |
33ad93b9 | 1392 | add_param_constraints (scop, p); |
2abae5f1 SP |
1393 | } |
1394 | ||
1395 | /* Build the iteration domains: the loops belonging to the current | |
1396 | SCOP, and that vary for the execution of the current basic block. | |
1397 | Returns false if there is no loop in SCOP. */ | |
1398 | ||
1399 | static void | |
1400 | build_scop_iteration_domain (scop_p scop) | |
1401 | { | |
1402 | struct loop *loop; | |
1403 | sese region = SCOP_REGION (scop); | |
1404 | int i; | |
2abae5f1 | 1405 | poly_bb_p pbb; |
0fc822d0 | 1406 | int nb_loops = number_of_loops (cfun); |
33ad93b9 | 1407 | isl_set **doms = XCNEWVEC (isl_set *, nb_loops); |
2abae5f1 | 1408 | |
9771b263 | 1409 | FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop) |
2abae5f1 | 1410 | if (!loop_in_sese_p (loop_outer (loop), region)) |
33ad93b9 RG |
1411 | build_loop_iteration_domains (scop, loop, 0, |
1412 | isl_set_copy (scop->context), doms); | |
2abae5f1 | 1413 | |
9771b263 | 1414 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
33ad93b9 RG |
1415 | { |
1416 | loop = pbb_loop (pbb); | |
1417 | ||
1418 | if (doms[loop->num]) | |
1419 | pbb->domain = isl_set_copy (doms[loop->num]); | |
1420 | else | |
1421 | pbb->domain = isl_set_copy (scop->context); | |
1422 | ||
1423 | pbb->domain = isl_set_set_tuple_id (pbb->domain, | |
1424 | isl_id_for_pbb (scop, pbb)); | |
1425 | } | |
2abae5f1 | 1426 | |
6c6f84d7 | 1427 | for (i = 0; i < nb_loops; i++) |
33ad93b9 RG |
1428 | if (doms[i]) |
1429 | isl_set_free (doms[i]); | |
2abae5f1 | 1430 | |
33ad93b9 | 1431 | free (doms); |
2abae5f1 SP |
1432 | } |
1433 | ||
1434 | /* Add a constrain to the ACCESSES polyhedron for the alias set of | |
1435 | data reference DR. ACCESSP_NB_DIMS is the dimension of the | |
1436 | ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration | |
1437 | domain. */ | |
1438 | ||
33ad93b9 RG |
1439 | static isl_map * |
1440 | pdr_add_alias_set (isl_map *acc, data_reference_p dr) | |
2abae5f1 | 1441 | { |
33ad93b9 | 1442 | isl_constraint *c; |
2abae5f1 | 1443 | int alias_set_num = 0; |
2b178a5f | 1444 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); |
2abae5f1 | 1445 | |
fb00d28e | 1446 | if (bap && bap->alias_set) |
2b178a5f | 1447 | alias_set_num = *(bap->alias_set); |
2abae5f1 | 1448 | |
33ad93b9 RG |
1449 | c = isl_equality_alloc |
1450 | (isl_local_space_from_space (isl_map_get_space (acc))); | |
1451 | c = isl_constraint_set_constant_si (c, -alias_set_num); | |
1452 | c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1); | |
1453 | ||
1454 | return isl_map_add_constraint (acc, c); | |
1455 | } | |
1456 | ||
1457 | /* Assign the affine expression INDEX to the output dimension POS of | |
1458 | MAP and return the result. */ | |
1459 | ||
1460 | static isl_map * | |
1461 | set_index (isl_map *map, int pos, isl_pw_aff *index) | |
1462 | { | |
1463 | isl_map *index_map; | |
1464 | int len = isl_map_dim (map, isl_dim_out); | |
1465 | isl_id *id; | |
1466 | ||
1467 | index_map = isl_map_from_pw_aff (index); | |
1468 | index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos); | |
1469 | index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1); | |
2abae5f1 | 1470 | |
33ad93b9 RG |
1471 | id = isl_map_get_tuple_id (map, isl_dim_out); |
1472 | index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id); | |
1473 | id = isl_map_get_tuple_id (map, isl_dim_in); | |
1474 | index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id); | |
2abae5f1 | 1475 | |
33ad93b9 | 1476 | return isl_map_intersect (map, index_map); |
2abae5f1 SP |
1477 | } |
1478 | ||
1479 | /* Add to ACCESSES polyhedron equalities defining the access functions | |
1480 | to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES | |
1481 | polyhedron, DOM_NB_DIMS is the dimension of the iteration domain. | |
1482 | PBB is the poly_bb_p that contains the data reference DR. */ | |
1483 | ||
33ad93b9 RG |
1484 | static isl_map * |
1485 | pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb) | |
2abae5f1 SP |
1486 | { |
1487 | int i, nb_subscripts = DR_NUM_DIMENSIONS (dr); | |
2abae5f1 | 1488 | scop_p scop = PBB_SCOP (pbb); |
2abae5f1 SP |
1489 | |
1490 | for (i = 0; i < nb_subscripts; i++) | |
1491 | { | |
33ad93b9 | 1492 | isl_pw_aff *aff; |
2abae5f1 SP |
1493 | tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i); |
1494 | ||
33ad93b9 RG |
1495 | aff = extract_affine (scop, afn, |
1496 | isl_space_domain (isl_map_get_space (acc))); | |
1497 | acc = set_index (acc, i + 1, aff); | |
2abae5f1 SP |
1498 | } |
1499 | ||
33ad93b9 | 1500 | return acc; |
2abae5f1 SP |
1501 | } |
1502 | ||
1503 | /* Add constrains representing the size of the accessed data to the | |
66096911 SP |
1504 | ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the |
1505 | ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration | |
2abae5f1 SP |
1506 | domain. */ |
1507 | ||
33ad93b9 RG |
1508 | static isl_set * |
1509 | pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr) | |
2abae5f1 SP |
1510 | { |
1511 | tree ref = DR_REF (dr); | |
1512 | int i, nb_subscripts = DR_NUM_DIMENSIONS (dr); | |
2abae5f1 | 1513 | |
98f3eb1f | 1514 | for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0)) |
2abae5f1 | 1515 | { |
98f3eb1f | 1516 | tree low, high; |
2abae5f1 | 1517 | |
98f3eb1f | 1518 | if (TREE_CODE (ref) != ARRAY_REF) |
2abae5f1 SP |
1519 | break; |
1520 | ||
98f3eb1f | 1521 | low = array_ref_low_bound (ref); |
98f3eb1f AM |
1522 | high = array_ref_up_bound (ref); |
1523 | ||
33ad93b9 RG |
1524 | /* XXX The PPL code dealt separately with |
1525 | subscript - low >= 0 and high - subscript >= 0 in case one of | |
1526 | the two bounds isn't known. Do the same here? */ | |
1527 | ||
1528 | if (host_integerp (low, 0) | |
1529 | && high | |
1530 | && host_integerp (high, 0) | |
3899a0b2 SP |
1531 | /* 1-element arrays at end of structures may extend over |
1532 | their declared size. */ | |
1533 | && !(array_at_struct_end_p (ref) | |
1534 | && operand_equal_p (low, high, 0))) | |
98f3eb1f | 1535 | { |
33ad93b9 RG |
1536 | isl_id *id; |
1537 | isl_aff *aff; | |
1538 | isl_set *univ, *lbs, *ubs; | |
1539 | isl_pw_aff *index; | |
1540 | isl_space *space; | |
1541 | isl_set *valid; | |
1542 | isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent)); | |
1543 | isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent)); | |
1544 | ||
1545 | /* high >= 0 */ | |
1546 | valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub)); | |
1547 | valid = isl_set_project_out (valid, isl_dim_set, 0, | |
1548 | isl_set_dim (valid, isl_dim_set)); | |
1549 | scop->context = isl_set_intersect (scop->context, valid); | |
1550 | ||
1551 | space = isl_set_get_space (extent); | |
1552 | aff = isl_aff_zero_on_domain (isl_local_space_from_space (space)); | |
1553 | aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1); | |
1554 | univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff))); | |
1555 | index = isl_pw_aff_alloc (univ, aff); | |
1556 | ||
1557 | id = isl_set_get_tuple_id (extent); | |
1558 | lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id)); | |
1559 | ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id); | |
1560 | ||
1561 | /* low <= sub_i <= high */ | |
1562 | lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb); | |
1563 | ubs = isl_pw_aff_le_set (index, ub); | |
1564 | extent = isl_set_intersect (extent, lbs); | |
1565 | extent = isl_set_intersect (extent, ubs); | |
98f3eb1f | 1566 | } |
2abae5f1 | 1567 | } |
33ad93b9 RG |
1568 | |
1569 | return extent; | |
2abae5f1 SP |
1570 | } |
1571 | ||
1572 | /* Build data accesses for DR in PBB. */ | |
1573 | ||
1574 | static void | |
1575 | build_poly_dr (data_reference_p dr, poly_bb_p pbb) | |
1576 | { | |
1825f9a2 | 1577 | int dr_base_object_set; |
33ad93b9 RG |
1578 | isl_map *acc; |
1579 | isl_set *extent; | |
1580 | scop_p scop = PBB_SCOP (pbb); | |
2abae5f1 | 1581 | |
33ad93b9 RG |
1582 | { |
1583 | isl_space *dc = isl_set_get_space (pbb->domain); | |
1584 | int nb_out = 1 + DR_NUM_DIMENSIONS (dr); | |
1585 | isl_space *space = isl_space_add_dims (isl_space_from_domain (dc), | |
1586 | isl_dim_out, nb_out); | |
2abae5f1 | 1587 | |
33ad93b9 RG |
1588 | acc = isl_map_universe (space); |
1589 | acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr)); | |
1590 | } | |
2abae5f1 | 1591 | |
33ad93b9 RG |
1592 | acc = pdr_add_alias_set (acc, dr); |
1593 | acc = pdr_add_memory_accesses (acc, dr, pbb); | |
2abae5f1 | 1594 | |
33ad93b9 RG |
1595 | { |
1596 | isl_id *id = isl_id_for_dr (scop, dr); | |
1597 | int nb = 1 + DR_NUM_DIMENSIONS (dr); | |
1598 | isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb); | |
1599 | int alias_set_num = 0; | |
1600 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); | |
1601 | ||
1602 | if (bap && bap->alias_set) | |
1603 | alias_set_num = *(bap->alias_set); | |
1604 | ||
1605 | space = isl_space_set_tuple_id (space, isl_dim_set, id); | |
1606 | extent = isl_set_nat_universe (space); | |
1607 | extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num); | |
1608 | extent = pdr_add_data_dimensions (extent, scop, dr); | |
1609 | } | |
2abae5f1 | 1610 | |
6e44d26e SP |
1611 | gcc_assert (dr->aux); |
1612 | dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set; | |
1825f9a2 | 1613 | |
33ad93b9 | 1614 | new_poly_dr (pbb, dr_base_object_set, |
6e44d26e | 1615 | DR_IS_READ (dr) ? PDR_READ : PDR_WRITE, |
33ad93b9 | 1616 | dr, DR_NUM_DIMENSIONS (dr), acc, extent); |
1825f9a2 | 1617 | } |
2abae5f1 | 1618 | |
2e5a7cbf | 1619 | /* Write to FILE the alias graph of data references in DIMACS format. */ |
cd43e5d7 LF |
1620 | |
1621 | static inline bool | |
1622 | write_alias_graph_to_ascii_dimacs (FILE *file, char *comment, | |
9771b263 | 1623 | vec<data_reference_p> drs) |
cd43e5d7 | 1624 | { |
9771b263 | 1625 | int num_vertex = drs.length (); |
cd43e5d7 LF |
1626 | int edge_num = 0; |
1627 | data_reference_p dr1, dr2; | |
1628 | int i, j; | |
1629 | ||
1630 | if (num_vertex == 0) | |
1631 | return true; | |
1632 | ||
9771b263 DN |
1633 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1634 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1635 | if (dr_may_alias_p (dr1, dr2, true)) |
cd43e5d7 LF |
1636 | edge_num++; |
1637 | ||
1638 | fprintf (file, "$\n"); | |
1639 | ||
1640 | if (comment) | |
1641 | fprintf (file, "c %s\n", comment); | |
1642 | ||
1643 | fprintf (file, "p edge %d %d\n", num_vertex, edge_num); | |
1644 | ||
9771b263 DN |
1645 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1646 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1647 | if (dr_may_alias_p (dr1, dr2, true)) |
cd43e5d7 LF |
1648 | fprintf (file, "e %d %d\n", i + 1, j + 1); |
1649 | ||
1650 | return true; | |
1651 | } | |
1652 | ||
2e5a7cbf RU |
1653 | /* Write to FILE the alias graph of data references in DOT format. */ |
1654 | ||
1655 | static inline bool | |
1656 | write_alias_graph_to_ascii_dot (FILE *file, char *comment, | |
9771b263 | 1657 | vec<data_reference_p> drs) |
2e5a7cbf | 1658 | { |
9771b263 | 1659 | int num_vertex = drs.length (); |
2e5a7cbf RU |
1660 | data_reference_p dr1, dr2; |
1661 | int i, j; | |
1662 | ||
1663 | if (num_vertex == 0) | |
1664 | return true; | |
1665 | ||
1666 | fprintf (file, "$\n"); | |
1667 | ||
1668 | if (comment) | |
1669 | fprintf (file, "c %s\n", comment); | |
1670 | ||
1671 | /* First print all the vertices. */ | |
9771b263 | 1672 | FOR_EACH_VEC_ELT (drs, i, dr1) |
2e5a7cbf RU |
1673 | fprintf (file, "n%d;\n", i); |
1674 | ||
9771b263 DN |
1675 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1676 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1677 | if (dr_may_alias_p (dr1, dr2, true)) |
2e5a7cbf RU |
1678 | fprintf (file, "n%d n%d\n", i, j); |
1679 | ||
1680 | return true; | |
1681 | } | |
1682 | ||
1683 | /* Write to FILE the alias graph of data references in ECC format. */ | |
1684 | ||
1685 | static inline bool | |
1686 | write_alias_graph_to_ascii_ecc (FILE *file, char *comment, | |
9771b263 | 1687 | vec<data_reference_p> drs) |
2e5a7cbf | 1688 | { |
9771b263 | 1689 | int num_vertex = drs.length (); |
2e5a7cbf RU |
1690 | data_reference_p dr1, dr2; |
1691 | int i, j; | |
1692 | ||
1693 | if (num_vertex == 0) | |
1694 | return true; | |
1695 | ||
1696 | fprintf (file, "$\n"); | |
1697 | ||
1698 | if (comment) | |
1699 | fprintf (file, "c %s\n", comment); | |
1700 | ||
9771b263 DN |
1701 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1702 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1703 | if (dr_may_alias_p (dr1, dr2, true)) |
2e5a7cbf RU |
1704 | fprintf (file, "%d %d\n", i, j); |
1705 | ||
1706 | return true; | |
1707 | } | |
1708 | ||
2b178a5f LF |
1709 | /* Check if DR1 and DR2 are in the same object set. */ |
1710 | ||
1711 | static bool | |
1712 | dr_same_base_object_p (const struct data_reference *dr1, | |
1713 | const struct data_reference *dr2) | |
1714 | { | |
1715 | return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0); | |
1716 | } | |
2e5a7cbf RU |
1717 | |
1718 | /* Uses DFS component number as representative of alias-sets. Also tests for | |
1719 | optimality by verifying if every connected component is a clique. Returns | |
1720 | true (1) if the above test is true, and false (0) otherwise. */ | |
1721 | ||
1722 | static int | |
9771b263 | 1723 | build_alias_set_optimal_p (vec<data_reference_p> drs) |
2abae5f1 | 1724 | { |
9771b263 | 1725 | int num_vertices = drs.length (); |
2e5a7cbf | 1726 | struct graph *g = new_graph (num_vertices); |
2abae5f1 SP |
1727 | data_reference_p dr1, dr2; |
1728 | int i, j; | |
2e5a7cbf RU |
1729 | int num_connected_components; |
1730 | int v_indx1, v_indx2, num_vertices_in_component; | |
1731 | int *all_vertices; | |
1732 | int *vertices; | |
1733 | struct graph_edge *e; | |
917f481a SP |
1734 | int this_component_is_clique; |
1735 | int all_components_are_cliques = 1; | |
2abae5f1 | 1736 | |
9771b263 DN |
1737 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1738 | for (j = i+1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1739 | if (dr_may_alias_p (dr1, dr2, true)) |
2abae5f1 SP |
1740 | { |
1741 | add_edge (g, i, j); | |
1742 | add_edge (g, j, i); | |
1743 | } | |
1744 | ||
2e5a7cbf RU |
1745 | all_vertices = XNEWVEC (int, num_vertices); |
1746 | vertices = XNEWVEC (int, num_vertices); | |
1747 | for (i = 0; i < num_vertices; i++) | |
1748 | all_vertices[i] = i; | |
1749 | ||
2b178a5f LF |
1750 | num_connected_components = graphds_dfs (g, all_vertices, num_vertices, |
1751 | NULL, true, NULL); | |
1752 | for (i = 0; i < g->n_vertices; i++) | |
1753 | { | |
9771b263 | 1754 | data_reference_p dr = drs[i]; |
2b178a5f | 1755 | base_alias_pair *bap; |
fb00d28e | 1756 | |
6e44d26e SP |
1757 | gcc_assert (dr->aux); |
1758 | bap = (base_alias_pair *)(dr->aux); | |
fb00d28e | 1759 | |
2b178a5f LF |
1760 | bap->alias_set = XNEW (int); |
1761 | *(bap->alias_set) = g->vertices[i].component + 1; | |
1762 | } | |
1763 | ||
2e5a7cbf RU |
1764 | /* Verify if the DFS numbering results in optimal solution. */ |
1765 | for (i = 0; i < num_connected_components; i++) | |
1766 | { | |
1767 | num_vertices_in_component = 0; | |
1768 | /* Get all vertices whose DFS component number is the same as i. */ | |
1769 | for (j = 0; j < num_vertices; j++) | |
1770 | if (g->vertices[j].component == i) | |
1771 | vertices[num_vertices_in_component++] = j; | |
1772 | ||
1773 | /* Now test if the vertices in 'vertices' form a clique, by testing | |
1774 | for edges among each pair. */ | |
1775 | this_component_is_clique = 1; | |
1776 | for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++) | |
1777 | { | |
1778 | for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++) | |
1779 | { | |
1780 | /* Check if the two vertices are connected by iterating | |
1781 | through all the edges which have one of these are source. */ | |
1782 | e = g->vertices[vertices[v_indx2]].pred; | |
1783 | while (e) | |
1784 | { | |
1785 | if (e->src == vertices[v_indx1]) | |
1786 | break; | |
1787 | e = e->pred_next; | |
1788 | } | |
1789 | if (!e) | |
1790 | { | |
1791 | this_component_is_clique = 0; | |
1792 | break; | |
1793 | } | |
1794 | } | |
1795 | if (!this_component_is_clique) | |
1796 | all_components_are_cliques = 0; | |
1797 | } | |
1798 | } | |
2abae5f1 | 1799 | |
2e5a7cbf RU |
1800 | free (all_vertices); |
1801 | free (vertices); | |
2abae5f1 | 1802 | free_graph (g); |
2e5a7cbf | 1803 | return all_components_are_cliques; |
2abae5f1 SP |
1804 | } |
1805 | ||
efa21390 | 1806 | /* Group each data reference in DRS with its base object set num. */ |
1825f9a2 LF |
1807 | |
1808 | static void | |
9771b263 | 1809 | build_base_obj_set_for_drs (vec<data_reference_p> drs) |
1825f9a2 | 1810 | { |
9771b263 | 1811 | int num_vertex = drs.length (); |
2b178a5f LF |
1812 | struct graph *g = new_graph (num_vertex); |
1813 | data_reference_p dr1, dr2; | |
1814 | int i, j; | |
2b178a5f LF |
1815 | int *queue; |
1816 | ||
9771b263 DN |
1817 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1818 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
2b178a5f LF |
1819 | if (dr_same_base_object_p (dr1, dr2)) |
1820 | { | |
1821 | add_edge (g, i, j); | |
1822 | add_edge (g, j, i); | |
1823 | } | |
1824 | ||
1825 | queue = XNEWVEC (int, num_vertex); | |
1826 | for (i = 0; i < num_vertex; i++) | |
1827 | queue[i] = i; | |
1828 | ||
fb00d28e | 1829 | graphds_dfs (g, queue, num_vertex, NULL, true, NULL); |
2b178a5f LF |
1830 | |
1831 | for (i = 0; i < g->n_vertices; i++) | |
1832 | { | |
9771b263 | 1833 | data_reference_p dr = drs[i]; |
2b178a5f | 1834 | base_alias_pair *bap; |
fb00d28e | 1835 | |
6e44d26e SP |
1836 | gcc_assert (dr->aux); |
1837 | bap = (base_alias_pair *)(dr->aux); | |
fb00d28e | 1838 | |
2b178a5f LF |
1839 | bap->base_obj_set = g->vertices[i].component + 1; |
1840 | } | |
1841 | ||
1842 | free (queue); | |
1843 | free_graph (g); | |
1825f9a2 LF |
1844 | } |
1845 | ||
2abae5f1 SP |
1846 | /* Build the data references for PBB. */ |
1847 | ||
1848 | static void | |
1849 | build_pbb_drs (poly_bb_p pbb) | |
1850 | { | |
1851 | int j; | |
1852 | data_reference_p dr; | |
9771b263 | 1853 | vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb)); |
2abae5f1 | 1854 | |
9771b263 | 1855 | FOR_EACH_VEC_ELT (gbb_drs, j, dr) |
2abae5f1 SP |
1856 | build_poly_dr (dr, pbb); |
1857 | } | |
1858 | ||
0d5ef2a9 SP |
1859 | /* Dump to file the alias graphs for the data references in DRS. */ |
1860 | ||
1861 | static void | |
9771b263 | 1862 | dump_alias_graphs (vec<data_reference_p> drs) |
0d5ef2a9 SP |
1863 | { |
1864 | char comment[100]; | |
1865 | FILE *file_dimacs, *file_ecc, *file_dot; | |
1866 | ||
1867 | file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab"); | |
1868 | if (file_dimacs) | |
1869 | { | |
1870 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1871 | current_function_name ()); | |
1872 | write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs); | |
1873 | fclose (file_dimacs); | |
1874 | } | |
1875 | ||
1876 | file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab"); | |
1877 | if (file_ecc) | |
1878 | { | |
1879 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1880 | current_function_name ()); | |
1881 | write_alias_graph_to_ascii_ecc (file_ecc, comment, drs); | |
1882 | fclose (file_ecc); | |
1883 | } | |
1884 | ||
1885 | file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab"); | |
1886 | if (file_dot) | |
1887 | { | |
1888 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1889 | current_function_name ()); | |
1890 | write_alias_graph_to_ascii_dot (file_dot, comment, drs); | |
1891 | fclose (file_dot); | |
1892 | } | |
1893 | } | |
1894 | ||
2abae5f1 SP |
1895 | /* Build data references in SCOP. */ |
1896 | ||
1897 | static void | |
1898 | build_scop_drs (scop_p scop) | |
1899 | { | |
64393e40 | 1900 | int i, j; |
2abae5f1 | 1901 | poly_bb_p pbb; |
64393e40 | 1902 | data_reference_p dr; |
9771b263 DN |
1903 | vec<data_reference_p> drs; |
1904 | drs.create (3); | |
64393e40 | 1905 | |
efa21390 SP |
1906 | /* Remove all the PBBs that do not have data references: these basic |
1907 | blocks are not handled in the polyhedral representation. */ | |
9771b263 DN |
1908 | for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++) |
1909 | if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ()) | |
278b1a1d | 1910 | { |
7470b8fc | 1911 | free_gimple_bb (PBB_BLACK_BOX (pbb)); |
62e0a1ed | 1912 | free_poly_bb (pbb); |
9771b263 | 1913 | SCOP_BBS (scop).ordered_remove (i); |
278b1a1d SP |
1914 | i--; |
1915 | } | |
efa21390 | 1916 | |
9771b263 DN |
1917 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
1918 | for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++) | |
1919 | drs.safe_push (dr); | |
64393e40 | 1920 | |
9771b263 | 1921 | FOR_EACH_VEC_ELT (drs, i, dr) |
2b178a5f LF |
1922 | dr->aux = XNEW (base_alias_pair); |
1923 | ||
1924 | if (!build_alias_set_optimal_p (drs)) | |
1925 | { | |
1926 | /* TODO: Add support when building alias set is not optimal. */ | |
1927 | ; | |
1928 | } | |
1929 | ||
ee03cd20 | 1930 | build_base_obj_set_for_drs (drs); |
1825f9a2 | 1931 | |
cd43e5d7 LF |
1932 | /* When debugging, enable the following code. This cannot be used |
1933 | in production compilers. */ | |
0d5ef2a9 SP |
1934 | if (0) |
1935 | dump_alias_graphs (drs); | |
cd43e5d7 | 1936 | |
9771b263 | 1937 | drs.release (); |
2abae5f1 | 1938 | |
9771b263 | 1939 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
1940 | build_pbb_drs (pbb); |
1941 | } | |
1942 | ||
a0dd1440 SP |
1943 | /* Return a gsi at the position of the phi node STMT. */ |
1944 | ||
1945 | static gimple_stmt_iterator | |
1946 | gsi_for_phi_node (gimple stmt) | |
1947 | { | |
1948 | gimple_stmt_iterator psi; | |
1949 | basic_block bb = gimple_bb (stmt); | |
1950 | ||
1951 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1952 | if (stmt == gsi_stmt (psi)) | |
1953 | return psi; | |
1954 | ||
1955 | gcc_unreachable (); | |
1956 | return psi; | |
1957 | } | |
1958 | ||
278b1a1d SP |
1959 | /* Analyze all the data references of STMTS and add them to the |
1960 | GBB_DATA_REFS vector of BB. */ | |
1961 | ||
1962 | static void | |
9771b263 | 1963 | analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts) |
278b1a1d SP |
1964 | { |
1965 | loop_p nest; | |
278b1a1d SP |
1966 | gimple_bb_p gbb; |
1967 | gimple stmt; | |
1968 | int i; | |
5c640e29 | 1969 | sese region = SCOP_REGION (scop); |
278b1a1d | 1970 | |
5c640e29 | 1971 | if (!bb_in_sese_p (bb, region)) |
278b1a1d SP |
1972 | return; |
1973 | ||
5c640e29 | 1974 | nest = outermost_loop_in_sese_1 (region, bb); |
278b1a1d SP |
1975 | gbb = gbb_from_bb (bb); |
1976 | ||
9771b263 | 1977 | FOR_EACH_VEC_ELT (stmts, i, stmt) |
5c640e29 SP |
1978 | { |
1979 | loop_p loop; | |
1980 | ||
1981 | if (is_gimple_debug (stmt)) | |
1982 | continue; | |
1983 | ||
1984 | loop = loop_containing_stmt (stmt); | |
1985 | if (!loop_in_sese_p (loop, region)) | |
1986 | loop = nest; | |
1987 | ||
1988 | graphite_find_data_references_in_stmt (nest, loop, stmt, | |
278b1a1d | 1989 | &GBB_DATA_REFS (gbb)); |
5c640e29 | 1990 | } |
278b1a1d SP |
1991 | } |
1992 | ||
1993 | /* Insert STMT at the end of the STMTS sequence and then insert the | |
1994 | statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts | |
1995 | on STMTS. */ | |
1996 | ||
1997 | static void | |
1998 | insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts, | |
1999 | gimple_stmt_iterator insert_gsi) | |
2000 | { | |
2001 | gimple_stmt_iterator gsi; | |
9771b263 DN |
2002 | vec<gimple> x; |
2003 | x.create (3); | |
278b1a1d | 2004 | |
355a7673 | 2005 | gimple_seq_add_stmt (&stmts, stmt); |
278b1a1d | 2006 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
9771b263 | 2007 | x.safe_push (gsi_stmt (gsi)); |
278b1a1d SP |
2008 | |
2009 | gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT); | |
2010 | analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x); | |
9771b263 | 2011 | x.release (); |
278b1a1d SP |
2012 | } |
2013 | ||
efa21390 | 2014 | /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */ |
2abae5f1 SP |
2015 | |
2016 | static void | |
278b1a1d | 2017 | insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt) |
2abae5f1 | 2018 | { |
2abae5f1 | 2019 | gimple_seq stmts; |
947121b8 | 2020 | gimple_stmt_iterator gsi; |
efa21390 | 2021 | tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE); |
2724573f | 2022 | gimple stmt = gimple_build_assign (unshare_expr (res), var); |
9771b263 DN |
2023 | vec<gimple> x; |
2024 | x.create (3); | |
2abae5f1 | 2025 | |
355a7673 | 2026 | gimple_seq_add_stmt (&stmts, stmt); |
278b1a1d | 2027 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
9771b263 | 2028 | x.safe_push (gsi_stmt (gsi)); |
947121b8 | 2029 | |
5fed5769 | 2030 | if (gimple_code (after_stmt) == GIMPLE_PHI) |
947121b8 | 2031 | { |
5fed5769 | 2032 | gsi = gsi_after_labels (gimple_bb (after_stmt)); |
947121b8 SP |
2033 | gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT); |
2034 | } | |
2035 | else | |
2036 | { | |
5fed5769 | 2037 | gsi = gsi_for_stmt (after_stmt); |
947121b8 SP |
2038 | gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); |
2039 | } | |
278b1a1d SP |
2040 | |
2041 | analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x); | |
9771b263 | 2042 | x.release (); |
2abae5f1 SP |
2043 | } |
2044 | ||
efa21390 SP |
2045 | /* Creates a poly_bb_p for basic_block BB from the existing PBB. */ |
2046 | ||
2047 | static void | |
2048 | new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb) | |
2049 | { | |
9771b263 DN |
2050 | vec<data_reference_p> drs; |
2051 | drs.create (3); | |
efa21390 SP |
2052 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); |
2053 | gimple_bb_p gbb1 = new_gimple_bb (bb, drs); | |
2054 | poly_bb_p pbb1 = new_poly_bb (scop, gbb1); | |
9771b263 | 2055 | int index, n = SCOP_BBS (scop).length (); |
efa21390 SP |
2056 | |
2057 | /* The INDEX of PBB in SCOP_BBS. */ | |
2058 | for (index = 0; index < n; index++) | |
9771b263 | 2059 | if (SCOP_BBS (scop)[index] == pbb) |
efa21390 SP |
2060 | break; |
2061 | ||
33ad93b9 | 2062 | pbb1->domain = isl_set_copy (pbb->domain); |
38013f25 | 2063 | |
efa21390 | 2064 | GBB_PBB (gbb1) = pbb1; |
9771b263 DN |
2065 | GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy (); |
2066 | GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy (); | |
2067 | SCOP_BBS (scop).safe_insert (index + 1, pbb1); | |
efa21390 SP |
2068 | } |
2069 | ||
2abae5f1 SP |
2070 | /* Insert on edge E the assignment "RES := EXPR". */ |
2071 | ||
2072 | static void | |
efa21390 | 2073 | insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr) |
2abae5f1 SP |
2074 | { |
2075 | gimple_stmt_iterator gsi; | |
355a7673 | 2076 | gimple_seq stmts = NULL; |
2abae5f1 | 2077 | tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE); |
2724573f | 2078 | gimple stmt = gimple_build_assign (unshare_expr (res), var); |
efa21390 | 2079 | basic_block bb; |
9771b263 DN |
2080 | vec<gimple> x; |
2081 | x.create (3); | |
2abae5f1 | 2082 | |
355a7673 | 2083 | gimple_seq_add_stmt (&stmts, stmt); |
278b1a1d | 2084 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
9771b263 | 2085 | x.safe_push (gsi_stmt (gsi)); |
278b1a1d | 2086 | |
2abae5f1 SP |
2087 | gsi_insert_seq_on_edge (e, stmts); |
2088 | gsi_commit_edge_inserts (); | |
efa21390 SP |
2089 | bb = gimple_bb (stmt); |
2090 | ||
2091 | if (!bb_in_sese_p (bb, SCOP_REGION (scop))) | |
2092 | return; | |
2093 | ||
2094 | if (!gbb_from_bb (bb)) | |
2095 | new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb); | |
278b1a1d SP |
2096 | |
2097 | analyze_drs_in_stmts (scop, bb, x); | |
9771b263 | 2098 | x.release (); |
2abae5f1 SP |
2099 | } |
2100 | ||
2101 | /* Creates a zero dimension array of the same type as VAR. */ | |
2102 | ||
2103 | static tree | |
63858ac6 | 2104 | create_zero_dim_array (tree var, const char *base_name) |
2abae5f1 SP |
2105 | { |
2106 | tree index_type = build_index_type (integer_zero_node); | |
2107 | tree elt_type = TREE_TYPE (var); | |
2108 | tree array_type = build_array_type (elt_type, index_type); | |
63858ac6 | 2109 | tree base = create_tmp_var (array_type, base_name); |
2abae5f1 | 2110 | |
2abae5f1 SP |
2111 | return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE, |
2112 | NULL_TREE); | |
2113 | } | |
2114 | ||
2115 | /* Returns true when PHI is a loop close phi node. */ | |
2116 | ||
2117 | static bool | |
2118 | scalar_close_phi_node_p (gimple phi) | |
2119 | { | |
a0dd1440 | 2120 | if (gimple_code (phi) != GIMPLE_PHI |
ea057359 | 2121 | || virtual_operand_p (gimple_phi_result (phi))) |
2abae5f1 SP |
2122 | return false; |
2123 | ||
79d03cf8 SP |
2124 | /* Note that loop close phi nodes should have a single argument |
2125 | because we translated the representation into a canonical form | |
2126 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
2abae5f1 SP |
2127 | return (gimple_phi_num_args (phi) == 1); |
2128 | } | |
2129 | ||
1c2a7491 SP |
2130 | /* For a definition DEF in REGION, propagates the expression EXPR in |
2131 | all the uses of DEF outside REGION. */ | |
2132 | ||
2133 | static void | |
2134 | propagate_expr_outside_region (tree def, tree expr, sese region) | |
2135 | { | |
2136 | imm_use_iterator imm_iter; | |
2137 | gimple use_stmt; | |
2138 | gimple_seq stmts; | |
2139 | bool replaced_once = false; | |
2140 | ||
ab756588 | 2141 | gcc_assert (TREE_CODE (def) == SSA_NAME); |
1c2a7491 SP |
2142 | |
2143 | expr = force_gimple_operand (unshare_expr (expr), &stmts, true, | |
2144 | NULL_TREE); | |
2145 | ||
2146 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2147 | if (!is_gimple_debug (use_stmt) | |
2148 | && !bb_in_sese_p (gimple_bb (use_stmt), region)) | |
2149 | { | |
2150 | ssa_op_iter iter; | |
2151 | use_operand_p use_p; | |
2152 | ||
2153 | FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES) | |
2154 | if (operand_equal_p (def, USE_FROM_PTR (use_p), 0) | |
2155 | && (replaced_once = true)) | |
2156 | replace_exp (use_p, expr); | |
2157 | ||
2158 | update_stmt (use_stmt); | |
2159 | } | |
2160 | ||
2161 | if (replaced_once) | |
2162 | { | |
2163 | gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts); | |
2164 | gsi_commit_edge_inserts (); | |
2165 | } | |
2166 | } | |
2167 | ||
2abae5f1 SP |
2168 | /* Rewrite out of SSA the reduction phi node at PSI by creating a zero |
2169 | dimension array for it. */ | |
2170 | ||
2171 | static void | |
efa21390 | 2172 | rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi) |
2abae5f1 | 2173 | { |
efa21390 | 2174 | sese region = SCOP_REGION (scop); |
2abae5f1 SP |
2175 | gimple phi = gsi_stmt (*psi); |
2176 | tree res = gimple_phi_result (phi); | |
8af6d9cd SP |
2177 | basic_block bb = gimple_bb (phi); |
2178 | gimple_stmt_iterator gsi = gsi_after_labels (bb); | |
2abae5f1 | 2179 | tree arg = gimple_phi_arg_def (phi, 0); |
8af6d9cd | 2180 | gimple stmt; |
2abae5f1 | 2181 | |
79d03cf8 SP |
2182 | /* Note that loop close phi nodes should have a single argument |
2183 | because we translated the representation into a canonical form | |
2184 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
2185 | gcc_assert (gimple_phi_num_args (phi) == 1); | |
2186 | ||
8af6d9cd | 2187 | /* The phi node can be a non close phi node, when its argument is |
974335d6 | 2188 | invariant, or a default definition. */ |
8af6d9cd | 2189 | if (is_gimple_min_invariant (arg) |
974335d6 | 2190 | || SSA_NAME_IS_DEFAULT_DEF (arg)) |
ab756588 SP |
2191 | { |
2192 | propagate_expr_outside_region (res, arg, region); | |
2193 | gsi_next (psi); | |
2194 | return; | |
2195 | } | |
1c2a7491 | 2196 | |
9707eeb0 SP |
2197 | else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father) |
2198 | { | |
2199 | propagate_expr_outside_region (res, arg, region); | |
2200 | stmt = gimple_build_assign (res, arg); | |
2201 | remove_phi_node (psi, false); | |
2202 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
2203 | SSA_NAME_DEF_STMT (res) = stmt; | |
2204 | return; | |
2205 | } | |
2206 | ||
1c2a7491 SP |
2207 | /* If res is scev analyzable and is not a scalar value, it is safe |
2208 | to ignore the close phi node: it will be code generated in the | |
2209 | out of Graphite pass. */ | |
2210 | else if (scev_analyzable_p (res, region)) | |
2211 | { | |
2212 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res)); | |
2213 | tree scev; | |
2214 | ||
2215 | if (!loop_in_sese_p (loop, region)) | |
2216 | { | |
2217 | loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg)); | |
2218 | scev = scalar_evolution_in_region (region, loop, arg); | |
2219 | scev = compute_overall_effect_of_inner_loop (loop, scev); | |
2220 | } | |
2221 | else | |
ab756588 | 2222 | scev = scalar_evolution_in_region (region, loop, res); |
1c2a7491 SP |
2223 | |
2224 | if (tree_does_not_contain_chrecs (scev)) | |
2225 | propagate_expr_outside_region (res, scev, region); | |
2226 | ||
2227 | gsi_next (psi); | |
2228 | return; | |
2229 | } | |
c880097d | 2230 | else |
8af6d9cd | 2231 | { |
070ecdfd | 2232 | tree zero_dim_array = create_zero_dim_array (res, "Close_Phi"); |
8af6d9cd | 2233 | |
2724573f | 2234 | stmt = gimple_build_assign (res, unshare_expr (zero_dim_array)); |
8af6d9cd | 2235 | |
3dd2dd57 | 2236 | if (TREE_CODE (arg) == SSA_NAME) |
278b1a1d | 2237 | insert_out_of_ssa_copy (scop, zero_dim_array, arg, |
efa21390 | 2238 | SSA_NAME_DEF_STMT (arg)); |
8af6d9cd | 2239 | else |
efa21390 | 2240 | insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb), |
8af6d9cd SP |
2241 | zero_dim_array, arg); |
2242 | } | |
2abae5f1 SP |
2243 | |
2244 | remove_phi_node (psi, false); | |
2abae5f1 | 2245 | SSA_NAME_DEF_STMT (res) = stmt; |
278b1a1d SP |
2246 | |
2247 | insert_stmts (scop, stmt, NULL, gsi_after_labels (bb)); | |
2abae5f1 SP |
2248 | } |
2249 | ||
2250 | /* Rewrite out of SSA the reduction phi node at PSI by creating a zero | |
2251 | dimension array for it. */ | |
2252 | ||
2253 | static void | |
efa21390 | 2254 | rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi) |
2abae5f1 SP |
2255 | { |
2256 | size_t i; | |
2257 | gimple phi = gsi_stmt (*psi); | |
2258 | basic_block bb = gimple_bb (phi); | |
2259 | tree res = gimple_phi_result (phi); | |
070ecdfd | 2260 | tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa"); |
2abae5f1 | 2261 | gimple stmt; |
2abae5f1 SP |
2262 | |
2263 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2264 | { | |
2265 | tree arg = gimple_phi_arg_def (phi, i); | |
4aa9a167 | 2266 | edge e = gimple_phi_arg_edge (phi, i); |
2abae5f1 | 2267 | |
4aa9a167 SP |
2268 | /* Avoid the insertion of code in the loop latch to please the |
2269 | pattern matching of the vectorizer. */ | |
320532a8 SP |
2270 | if (TREE_CODE (arg) == SSA_NAME |
2271 | && e->src == bb->loop_father->latch) | |
278b1a1d | 2272 | insert_out_of_ssa_copy (scop, zero_dim_array, arg, |
efa21390 | 2273 | SSA_NAME_DEF_STMT (arg)); |
2abae5f1 | 2274 | else |
efa21390 | 2275 | insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg); |
2abae5f1 SP |
2276 | } |
2277 | ||
2724573f | 2278 | stmt = gimple_build_assign (res, unshare_expr (zero_dim_array)); |
2abae5f1 SP |
2279 | remove_phi_node (psi, false); |
2280 | SSA_NAME_DEF_STMT (res) = stmt; | |
2724573f | 2281 | insert_stmts (scop, stmt, NULL, gsi_after_labels (bb)); |
2abae5f1 SP |
2282 | } |
2283 | ||
d3e7b889 SP |
2284 | /* Rewrite the degenerate phi node at position PSI from the degenerate |
2285 | form "x = phi (y, y, ..., y)" to "x = y". */ | |
2286 | ||
2287 | static void | |
2288 | rewrite_degenerate_phi (gimple_stmt_iterator *psi) | |
2289 | { | |
2290 | tree rhs; | |
2291 | gimple stmt; | |
2292 | gimple_stmt_iterator gsi; | |
2293 | gimple phi = gsi_stmt (*psi); | |
2294 | tree res = gimple_phi_result (phi); | |
2295 | basic_block bb; | |
2296 | ||
d3e7b889 SP |
2297 | bb = gimple_bb (phi); |
2298 | rhs = degenerate_phi_result (phi); | |
2299 | gcc_assert (rhs); | |
2300 | ||
2301 | stmt = gimple_build_assign (res, rhs); | |
2302 | remove_phi_node (psi, false); | |
2303 | SSA_NAME_DEF_STMT (res) = stmt; | |
2304 | ||
2305 | gsi = gsi_after_labels (bb); | |
2306 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
2307 | } | |
2308 | ||
9773d730 SP |
2309 | /* Rewrite out of SSA all the reduction phi nodes of SCOP. */ |
2310 | ||
efa21390 | 2311 | static void |
9773d730 SP |
2312 | rewrite_reductions_out_of_ssa (scop_p scop) |
2313 | { | |
2314 | basic_block bb; | |
2315 | gimple_stmt_iterator psi; | |
2316 | sese region = SCOP_REGION (scop); | |
2317 | ||
2318 | FOR_EACH_BB (bb) | |
2319 | if (bb_in_sese_p (bb, region)) | |
2320 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi);) | |
2321 | { | |
d3e7b889 SP |
2322 | gimple phi = gsi_stmt (psi); |
2323 | ||
ea057359 | 2324 | if (virtual_operand_p (gimple_phi_result (phi))) |
c2bc669e SP |
2325 | { |
2326 | gsi_next (&psi); | |
2327 | continue; | |
2328 | } | |
2329 | ||
d3e7b889 SP |
2330 | if (gimple_phi_num_args (phi) > 1 |
2331 | && degenerate_phi_result (phi)) | |
2332 | rewrite_degenerate_phi (&psi); | |
2333 | ||
2334 | else if (scalar_close_phi_node_p (phi)) | |
efa21390 | 2335 | rewrite_close_phi_out_of_ssa (scop, &psi); |
d3e7b889 | 2336 | |
9773d730 | 2337 | else if (reduction_phi_p (region, &psi)) |
efa21390 | 2338 | rewrite_phi_out_of_ssa (scop, &psi); |
9773d730 SP |
2339 | } |
2340 | ||
2341 | update_ssa (TODO_update_ssa); | |
2342 | #ifdef ENABLE_CHECKING | |
2343 | verify_loop_closed_ssa (true); | |
2344 | #endif | |
2345 | } | |
2346 | ||
5dcc64d9 SP |
2347 | /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory |
2348 | read from ZERO_DIM_ARRAY. */ | |
2349 | ||
2350 | static void | |
278b1a1d | 2351 | rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array, |
efa21390 | 2352 | tree def, gimple use_stmt) |
5dcc64d9 | 2353 | { |
070ecdfd RG |
2354 | gimple name_stmt; |
2355 | tree name; | |
5dcc64d9 SP |
2356 | ssa_op_iter iter; |
2357 | use_operand_p use_p; | |
5dcc64d9 | 2358 | |
c7dc2fab | 2359 | gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI); |
5dcc64d9 | 2360 | |
070ecdfd RG |
2361 | name = copy_ssa_name (def, NULL); |
2362 | name_stmt = gimple_build_assign (name, zero_dim_array); | |
2363 | ||
c7dc2fab | 2364 | gimple_assign_set_lhs (name_stmt, name); |
278b1a1d | 2365 | insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt)); |
5dcc64d9 | 2366 | |
c7dc2fab SP |
2367 | FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES) |
2368 | if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)) | |
2369 | replace_exp (use_p, name); | |
5dcc64d9 SP |
2370 | |
2371 | update_stmt (use_stmt); | |
2372 | } | |
2373 | ||
70a2ae0f SP |
2374 | /* For every definition DEF in the SCOP that is used outside the scop, |
2375 | insert a closing-scop definition in the basic block just after this | |
2376 | SCOP. */ | |
2377 | ||
2378 | static void | |
2379 | handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt) | |
2380 | { | |
2381 | tree var = create_tmp_reg (TREE_TYPE (def), NULL); | |
2382 | tree new_name = make_ssa_name (var, stmt); | |
2383 | bool needs_copy = false; | |
2384 | use_operand_p use_p; | |
2385 | imm_use_iterator imm_iter; | |
2386 | gimple use_stmt; | |
2387 | sese region = SCOP_REGION (scop); | |
2388 | ||
2389 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2390 | { | |
2391 | if (!bb_in_sese_p (gimple_bb (use_stmt), region)) | |
2392 | { | |
2393 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
2394 | { | |
2395 | SET_USE (use_p, new_name); | |
2396 | } | |
2397 | update_stmt (use_stmt); | |
2398 | needs_copy = true; | |
2399 | } | |
2400 | } | |
2401 | ||
2402 | /* Insert in the empty BB just after the scop a use of DEF such | |
2403 | that the rewrite of cross_bb_scalar_dependences won't insert | |
2404 | arrays everywhere else. */ | |
2405 | if (needs_copy) | |
2406 | { | |
2407 | gimple assign = gimple_build_assign (new_name, def); | |
2408 | gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest); | |
2409 | ||
70a2ae0f SP |
2410 | SSA_NAME_DEF_STMT (new_name) = assign; |
2411 | update_stmt (assign); | |
2412 | gsi_insert_before (&psi, assign, GSI_SAME_STMT); | |
2413 | } | |
2414 | } | |
2415 | ||
9773d730 | 2416 | /* Rewrite the scalar dependences crossing the boundary of the BB |
5d737345 SP |
2417 | containing STMT with an array. Return true when something has been |
2418 | changed. */ | |
9773d730 | 2419 | |
5d737345 | 2420 | static bool |
70a2ae0f | 2421 | rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi) |
9773d730 | 2422 | { |
70a2ae0f | 2423 | sese region = SCOP_REGION (scop); |
9773d730 SP |
2424 | gimple stmt = gsi_stmt (*gsi); |
2425 | imm_use_iterator imm_iter; | |
2426 | tree def; | |
2427 | basic_block def_bb; | |
2428 | tree zero_dim_array = NULL_TREE; | |
2429 | gimple use_stmt; | |
5d737345 | 2430 | bool res = false; |
9773d730 | 2431 | |
dba9acfa SP |
2432 | switch (gimple_code (stmt)) |
2433 | { | |
2434 | case GIMPLE_ASSIGN: | |
2435 | def = gimple_assign_lhs (stmt); | |
2436 | break; | |
2437 | ||
2438 | case GIMPLE_CALL: | |
2439 | def = gimple_call_lhs (stmt); | |
2440 | break; | |
2441 | ||
2442 | default: | |
5d737345 | 2443 | return false; |
dba9acfa | 2444 | } |
9773d730 | 2445 | |
b4c8119f SP |
2446 | if (!def |
2447 | || !is_gimple_reg (def)) | |
5d737345 | 2448 | return false; |
9773d730 | 2449 | |
1c2a7491 SP |
2450 | if (scev_analyzable_p (def, region)) |
2451 | { | |
2452 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def)); | |
2453 | tree scev = scalar_evolution_in_region (region, loop, def); | |
2454 | ||
5d737345 SP |
2455 | if (tree_contains_chrecs (scev, NULL)) |
2456 | return false; | |
1c2a7491 | 2457 | |
5d737345 SP |
2458 | propagate_expr_outside_region (def, scev, region); |
2459 | return true; | |
1c2a7491 SP |
2460 | } |
2461 | ||
9773d730 SP |
2462 | def_bb = gimple_bb (stmt); |
2463 | ||
70a2ae0f SP |
2464 | handle_scalar_deps_crossing_scop_limits (scop, def, stmt); |
2465 | ||
9773d730 | 2466 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) |
5d737345 SP |
2467 | if (gimple_code (use_stmt) == GIMPLE_PHI |
2468 | && (res = true)) | |
5dcc64d9 | 2469 | { |
ab756588 | 2470 | gimple_stmt_iterator psi = gsi_for_stmt (use_stmt); |
9773d730 | 2471 | |
ab756588 | 2472 | if (scalar_close_phi_node_p (gsi_stmt (psi))) |
efa21390 | 2473 | rewrite_close_phi_out_of_ssa (scop, &psi); |
ab756588 | 2474 | else |
efa21390 | 2475 | rewrite_phi_out_of_ssa (scop, &psi); |
ab756588 SP |
2476 | } |
2477 | ||
2478 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2479 | if (gimple_code (use_stmt) != GIMPLE_PHI | |
2480 | && def_bb != gimple_bb (use_stmt) | |
5d737345 SP |
2481 | && !is_gimple_debug (use_stmt) |
2482 | && (res = true)) | |
ab756588 | 2483 | { |
5dcc64d9 SP |
2484 | if (!zero_dim_array) |
2485 | { | |
63858ac6 | 2486 | zero_dim_array = create_zero_dim_array |
070ecdfd | 2487 | (def, "Cross_BB_scalar_dependence"); |
278b1a1d | 2488 | insert_out_of_ssa_copy (scop, zero_dim_array, def, |
5fed5769 | 2489 | SSA_NAME_DEF_STMT (def)); |
5dcc64d9 SP |
2490 | gsi_next (gsi); |
2491 | } | |
2492 | ||
278b1a1d | 2493 | rewrite_cross_bb_scalar_dependence (scop, zero_dim_array, |
efa21390 | 2494 | def, use_stmt); |
5dcc64d9 | 2495 | } |
5d737345 SP |
2496 | |
2497 | return res; | |
5dcc64d9 SP |
2498 | } |
2499 | ||
ee646fc6 SP |
2500 | /* Rewrite out of SSA all the reduction phi nodes of SCOP. */ |
2501 | ||
efa21390 | 2502 | static void |
ee646fc6 SP |
2503 | rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop) |
2504 | { | |
2505 | basic_block bb; | |
2506 | gimple_stmt_iterator psi; | |
2507 | sese region = SCOP_REGION (scop); | |
5d737345 | 2508 | bool changed = false; |
5dcc64d9 | 2509 | |
70a2ae0f | 2510 | /* Create an extra empty BB after the scop. */ |
844e904d | 2511 | split_edge (SESE_EXIT (region)); |
70a2ae0f | 2512 | |
5dcc64d9 SP |
2513 | FOR_EACH_BB (bb) |
2514 | if (bb_in_sese_p (bb, region)) | |
2515 | for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
70a2ae0f | 2516 | changed |= rewrite_cross_bb_scalar_deps (scop, &psi); |
5dcc64d9 | 2517 | |
5d737345 SP |
2518 | if (changed) |
2519 | { | |
2520 | scev_reset_htab (); | |
2521 | update_ssa (TODO_update_ssa); | |
5dcc64d9 | 2522 | #ifdef ENABLE_CHECKING |
5d737345 | 2523 | verify_loop_closed_ssa (true); |
5dcc64d9 | 2524 | #endif |
5d737345 | 2525 | } |
2abae5f1 SP |
2526 | } |
2527 | ||
2528 | /* Returns the number of pbbs that are in loops contained in SCOP. */ | |
2529 | ||
2530 | static int | |
2531 | nb_pbbs_in_loops (scop_p scop) | |
2532 | { | |
2533 | int i; | |
2534 | poly_bb_p pbb; | |
2535 | int res = 0; | |
2536 | ||
9771b263 | 2537 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
2538 | if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop))) |
2539 | res++; | |
2540 | ||
2541 | return res; | |
2542 | } | |
2543 | ||
60d2a8c3 SP |
2544 | /* Return the number of data references in BB that write in |
2545 | memory. */ | |
2546 | ||
2547 | static int | |
2548 | nb_data_writes_in_bb (basic_block bb) | |
2549 | { | |
2550 | int res = 0; | |
2551 | gimple_stmt_iterator gsi; | |
2552 | ||
2553 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2554 | if (gimple_vdef (gsi_stmt (gsi))) | |
2555 | res++; | |
2556 | ||
2557 | return res; | |
2558 | } | |
2559 | ||
efa21390 SP |
2560 | /* Splits at STMT the basic block BB represented as PBB in the |
2561 | polyhedral form. */ | |
2562 | ||
2563 | static edge | |
2564 | split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt) | |
2565 | { | |
2566 | edge e1 = split_block (bb, stmt); | |
2567 | new_pbb_from_pbb (scop, pbb, e1->dest); | |
2568 | return e1; | |
2569 | } | |
2570 | ||
2571 | /* Splits STMT out of its current BB. This is done for reduction | |
2572 | statements for which we want to ignore data dependences. */ | |
a0dd1440 SP |
2573 | |
2574 | static basic_block | |
efa21390 | 2575 | split_reduction_stmt (scop_p scop, gimple stmt) |
a0dd1440 | 2576 | { |
a0dd1440 | 2577 | basic_block bb = gimple_bb (stmt); |
efa21390 | 2578 | poly_bb_p pbb = pbb_from_bb (bb); |
278b1a1d | 2579 | gimple_bb_p gbb = gbb_from_bb (bb); |
efa21390 | 2580 | edge e1; |
278b1a1d SP |
2581 | int i; |
2582 | data_reference_p dr; | |
a0dd1440 | 2583 | |
60d2a8c3 SP |
2584 | /* Do not split basic blocks with no writes to memory: the reduction |
2585 | will be the only write to memory. */ | |
c513da01 SP |
2586 | if (nb_data_writes_in_bb (bb) == 0 |
2587 | /* Or if we have already marked BB as a reduction. */ | |
2588 | || PBB_IS_REDUCTION (pbb_from_bb (bb))) | |
60d2a8c3 SP |
2589 | return bb; |
2590 | ||
efa21390 | 2591 | e1 = split_pbb (scop, pbb, bb, stmt); |
a0dd1440 | 2592 | |
efa21390 SP |
2593 | /* Split once more only when the reduction stmt is not the only one |
2594 | left in the original BB. */ | |
2595 | if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb))) | |
2596 | { | |
2597 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
2598 | gsi_prev (&gsi); | |
2599 | e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi)); | |
2600 | } | |
a0dd1440 | 2601 | |
278b1a1d SP |
2602 | /* A part of the data references will end in a different basic block |
2603 | after the split: move the DRs from the original GBB to the newly | |
2604 | created GBB1. */ | |
9771b263 | 2605 | FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr) |
278b1a1d SP |
2606 | { |
2607 | basic_block bb1 = gimple_bb (DR_STMT (dr)); | |
2608 | ||
2609 | if (bb1 != bb) | |
2610 | { | |
2611 | gimple_bb_p gbb1 = gbb_from_bb (bb1); | |
9771b263 DN |
2612 | GBB_DATA_REFS (gbb1).safe_push (dr); |
2613 | GBB_DATA_REFS (gbb).ordered_remove (i); | |
278b1a1d SP |
2614 | i--; |
2615 | } | |
2616 | } | |
2617 | ||
efa21390 | 2618 | return e1->dest; |
a0dd1440 SP |
2619 | } |
2620 | ||
2621 | /* Return true when stmt is a reduction operation. */ | |
2622 | ||
2623 | static inline bool | |
2624 | is_reduction_operation_p (gimple stmt) | |
2625 | { | |
0596e97f AH |
2626 | enum tree_code code; |
2627 | ||
2628 | gcc_assert (is_gimple_assign (stmt)); | |
2629 | code = gimple_assign_rhs_code (stmt); | |
2630 | ||
a0dd1440 | 2631 | return flag_associative_math |
0596e97f AH |
2632 | && commutative_tree_code (code) |
2633 | && associative_tree_code (code); | |
a0dd1440 SP |
2634 | } |
2635 | ||
2636 | /* Returns true when PHI contains an argument ARG. */ | |
2637 | ||
2638 | static bool | |
2639 | phi_contains_arg (gimple phi, tree arg) | |
2640 | { | |
2641 | size_t i; | |
2642 | ||
2643 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2644 | if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0)) | |
2645 | return true; | |
2646 | ||
2647 | return false; | |
2648 | } | |
2649 | ||
2650 | /* Return a loop phi node that corresponds to a reduction containing LHS. */ | |
2651 | ||
2652 | static gimple | |
2653 | follow_ssa_with_commutative_ops (tree arg, tree lhs) | |
2654 | { | |
2655 | gimple stmt; | |
2656 | ||
2657 | if (TREE_CODE (arg) != SSA_NAME) | |
2658 | return NULL; | |
2659 | ||
2660 | stmt = SSA_NAME_DEF_STMT (arg); | |
2661 | ||
a84a556d SP |
2662 | if (gimple_code (stmt) == GIMPLE_NOP |
2663 | || gimple_code (stmt) == GIMPLE_CALL) | |
403ebc7e SP |
2664 | return NULL; |
2665 | ||
a0dd1440 SP |
2666 | if (gimple_code (stmt) == GIMPLE_PHI) |
2667 | { | |
2668 | if (phi_contains_arg (stmt, lhs)) | |
2669 | return stmt; | |
2670 | return NULL; | |
2671 | } | |
2672 | ||
0596e97f AH |
2673 | if (!is_gimple_assign (stmt)) |
2674 | return NULL; | |
2675 | ||
a0dd1440 SP |
2676 | if (gimple_num_ops (stmt) == 2) |
2677 | return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs); | |
2678 | ||
2679 | if (is_reduction_operation_p (stmt)) | |
2680 | { | |
2681 | gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs); | |
2682 | ||
2683 | return res ? res : | |
2684 | follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs); | |
2685 | } | |
2686 | ||
2687 | return NULL; | |
2688 | } | |
2689 | ||
2690 | /* Detect commutative and associative scalar reductions starting at | |
c880097d | 2691 | the STMT. Return the phi node of the reduction cycle, or NULL. */ |
a0dd1440 SP |
2692 | |
2693 | static gimple | |
2694 | detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg, | |
9771b263 DN |
2695 | vec<gimple> *in, |
2696 | vec<gimple> *out) | |
a0dd1440 SP |
2697 | { |
2698 | gimple phi = follow_ssa_with_commutative_ops (arg, lhs); | |
2699 | ||
c880097d SP |
2700 | if (!phi) |
2701 | return NULL; | |
a0dd1440 | 2702 | |
9771b263 DN |
2703 | in->safe_push (stmt); |
2704 | out->safe_push (stmt); | |
c880097d | 2705 | return phi; |
a0dd1440 SP |
2706 | } |
2707 | ||
2708 | /* Detect commutative and associative scalar reductions starting at | |
3a7086cc | 2709 | STMT. Return the phi node of the reduction cycle, or NULL. */ |
a0dd1440 SP |
2710 | |
2711 | static gimple | |
9771b263 DN |
2712 | detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in, |
2713 | vec<gimple> *out) | |
a0dd1440 SP |
2714 | { |
2715 | tree lhs = gimple_assign_lhs (stmt); | |
2716 | ||
2717 | if (gimple_num_ops (stmt) == 2) | |
2718 | return detect_commutative_reduction_arg (lhs, stmt, | |
2719 | gimple_assign_rhs1 (stmt), | |
2720 | in, out); | |
2721 | ||
2722 | if (is_reduction_operation_p (stmt)) | |
2723 | { | |
2724 | gimple res = detect_commutative_reduction_arg (lhs, stmt, | |
2725 | gimple_assign_rhs1 (stmt), | |
2726 | in, out); | |
2727 | return res ? res | |
2728 | : detect_commutative_reduction_arg (lhs, stmt, | |
2729 | gimple_assign_rhs2 (stmt), | |
2730 | in, out); | |
2731 | } | |
2732 | ||
2733 | return NULL; | |
2734 | } | |
2735 | ||
2736 | /* Return a loop phi node that corresponds to a reduction containing LHS. */ | |
2737 | ||
2738 | static gimple | |
2739 | follow_inital_value_to_phi (tree arg, tree lhs) | |
2740 | { | |
2741 | gimple stmt; | |
2742 | ||
2743 | if (!arg || TREE_CODE (arg) != SSA_NAME) | |
2744 | return NULL; | |
2745 | ||
2746 | stmt = SSA_NAME_DEF_STMT (arg); | |
2747 | ||
2748 | if (gimple_code (stmt) == GIMPLE_PHI | |
2749 | && phi_contains_arg (stmt, lhs)) | |
2750 | return stmt; | |
2751 | ||
2752 | return NULL; | |
2753 | } | |
2754 | ||
2755 | ||
073a8998 | 2756 | /* Return the argument of the loop PHI that is the initial value coming |
a0dd1440 SP |
2757 | from outside the loop. */ |
2758 | ||
2759 | static edge | |
2760 | edge_initial_value_for_loop_phi (gimple phi) | |
2761 | { | |
2762 | size_t i; | |
2763 | ||
2764 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2765 | { | |
2766 | edge e = gimple_phi_arg_edge (phi, i); | |
2767 | ||
2768 | if (loop_depth (e->src->loop_father) | |
2769 | < loop_depth (e->dest->loop_father)) | |
2770 | return e; | |
2771 | } | |
2772 | ||
2773 | return NULL; | |
2774 | } | |
2775 | ||
073a8998 | 2776 | /* Return the argument of the loop PHI that is the initial value coming |
a0dd1440 SP |
2777 | from outside the loop. */ |
2778 | ||
2779 | static tree | |
2780 | initial_value_for_loop_phi (gimple phi) | |
2781 | { | |
2782 | size_t i; | |
2783 | ||
2784 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2785 | { | |
2786 | edge e = gimple_phi_arg_edge (phi, i); | |
2787 | ||
2788 | if (loop_depth (e->src->loop_father) | |
2789 | < loop_depth (e->dest->loop_father)) | |
2790 | return gimple_phi_arg_def (phi, i); | |
2791 | } | |
2792 | ||
2793 | return NULL_TREE; | |
2794 | } | |
2795 | ||
479c1fb3 SP |
2796 | /* Returns true when DEF is used outside the reduction cycle of |
2797 | LOOP_PHI. */ | |
2798 | ||
2799 | static bool | |
2800 | used_outside_reduction (tree def, gimple loop_phi) | |
2801 | { | |
2802 | use_operand_p use_p; | |
2803 | imm_use_iterator imm_iter; | |
2804 | loop_p loop = loop_containing_stmt (loop_phi); | |
2805 | ||
2806 | /* In LOOP, DEF should be used only in LOOP_PHI. */ | |
2807 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2808 | { | |
2809 | gimple stmt = USE_STMT (use_p); | |
2810 | ||
2811 | if (stmt != loop_phi | |
2812 | && !is_gimple_debug (stmt) | |
2813 | && flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
2814 | return true; | |
2815 | } | |
2816 | ||
2817 | return false; | |
2818 | } | |
2819 | ||
a30e5345 SP |
2820 | /* Detect commutative and associative scalar reductions belonging to |
2821 | the SCOP starting at the loop closed phi node STMT. Return the phi | |
2822 | node of the reduction cycle, or NULL. */ | |
a0dd1440 SP |
2823 | |
2824 | static gimple | |
9771b263 DN |
2825 | detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in, |
2826 | vec<gimple> *out) | |
a0dd1440 SP |
2827 | { |
2828 | if (scalar_close_phi_node_p (stmt)) | |
2829 | { | |
479c1fb3 SP |
2830 | gimple def, loop_phi, phi, close_phi = stmt; |
2831 | tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0); | |
c880097d SP |
2832 | |
2833 | if (TREE_CODE (arg) != SSA_NAME) | |
2834 | return NULL; | |
2835 | ||
79d03cf8 SP |
2836 | /* Note that loop close phi nodes should have a single argument |
2837 | because we translated the representation into a canonical form | |
2838 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
479c1fb3 | 2839 | gcc_assert (gimple_phi_num_args (close_phi) == 1); |
79d03cf8 | 2840 | |
c880097d | 2841 | def = SSA_NAME_DEF_STMT (arg); |
479c1fb3 SP |
2842 | if (!stmt_in_sese_p (def, SCOP_REGION (scop)) |
2843 | || !(loop_phi = detect_commutative_reduction (scop, def, in, out))) | |
a30e5345 SP |
2844 | return NULL; |
2845 | ||
479c1fb3 SP |
2846 | lhs = gimple_phi_result (close_phi); |
2847 | init = initial_value_for_loop_phi (loop_phi); | |
2848 | phi = follow_inital_value_to_phi (init, lhs); | |
a0dd1440 | 2849 | |
479c1fb3 SP |
2850 | if (phi && (used_outside_reduction (lhs, phi) |
2851 | || !has_single_use (gimple_phi_result (phi)))) | |
a0dd1440 | 2852 | return NULL; |
479c1fb3 | 2853 | |
9771b263 DN |
2854 | in->safe_push (loop_phi); |
2855 | out->safe_push (close_phi); | |
479c1fb3 | 2856 | return phi; |
a0dd1440 SP |
2857 | } |
2858 | ||
2859 | if (gimple_code (stmt) == GIMPLE_ASSIGN) | |
2860 | return detect_commutative_reduction_assign (stmt, in, out); | |
2861 | ||
2862 | return NULL; | |
2863 | } | |
2864 | ||
2865 | /* Translate the scalar reduction statement STMT to an array RED | |
2866 | knowing that its recursive phi node is LOOP_PHI. */ | |
2867 | ||
2868 | static void | |
278b1a1d SP |
2869 | translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red, |
2870 | gimple stmt, gimple loop_phi) | |
a0dd1440 | 2871 | { |
a0dd1440 | 2872 | tree res = gimple_phi_result (loop_phi); |
50034a36 | 2873 | gimple assign = gimple_build_assign (res, unshare_expr (red)); |
278b1a1d | 2874 | gimple_stmt_iterator gsi; |
a0dd1440 | 2875 | |
278b1a1d | 2876 | insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi))); |
a0dd1440 | 2877 | |
50034a36 | 2878 | assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt)); |
278b1a1d SP |
2879 | gsi = gsi_for_stmt (stmt); |
2880 | gsi_next (&gsi); | |
2881 | insert_stmts (scop, assign, NULL, gsi); | |
a0dd1440 SP |
2882 | } |
2883 | ||
a4681954 SP |
2884 | /* Removes the PHI node and resets all the debug stmts that are using |
2885 | the PHI_RESULT. */ | |
2886 | ||
2887 | static void | |
2888 | remove_phi (gimple phi) | |
2889 | { | |
2890 | imm_use_iterator imm_iter; | |
2891 | tree def; | |
2892 | use_operand_p use_p; | |
2893 | gimple_stmt_iterator gsi; | |
9771b263 DN |
2894 | vec<gimple> update; |
2895 | update.create (3); | |
a4681954 SP |
2896 | unsigned int i; |
2897 | gimple stmt; | |
2898 | ||
2899 | def = PHI_RESULT (phi); | |
2900 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2901 | { | |
2902 | stmt = USE_STMT (use_p); | |
2903 | ||
2904 | if (is_gimple_debug (stmt)) | |
2905 | { | |
2906 | gimple_debug_bind_reset_value (stmt); | |
9771b263 | 2907 | update.safe_push (stmt); |
a4681954 SP |
2908 | } |
2909 | } | |
2910 | ||
9771b263 | 2911 | FOR_EACH_VEC_ELT (update, i, stmt) |
a4681954 SP |
2912 | update_stmt (stmt); |
2913 | ||
9771b263 | 2914 | update.release (); |
a4681954 SP |
2915 | |
2916 | gsi = gsi_for_phi_node (phi); | |
2917 | remove_phi_node (&gsi, false); | |
2918 | } | |
2919 | ||
7c48ea69 SP |
2920 | /* Helper function for for_each_index. For each INDEX of the data |
2921 | reference REF, returns true when its indices are valid in the loop | |
2922 | nest LOOP passed in as DATA. */ | |
2923 | ||
2924 | static bool | |
2925 | dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data) | |
2926 | { | |
2927 | loop_p loop; | |
2928 | basic_block header, def_bb; | |
2929 | gimple stmt; | |
2930 | ||
2931 | if (TREE_CODE (*index) != SSA_NAME) | |
2932 | return true; | |
2933 | ||
2934 | loop = *((loop_p *) data); | |
2935 | header = loop->header; | |
2936 | stmt = SSA_NAME_DEF_STMT (*index); | |
2937 | ||
2938 | if (!stmt) | |
2939 | return true; | |
2940 | ||
2941 | def_bb = gimple_bb (stmt); | |
2942 | ||
2943 | if (!def_bb) | |
2944 | return true; | |
2945 | ||
2946 | return dominated_by_p (CDI_DOMINATORS, header, def_bb); | |
2947 | } | |
2948 | ||
50034a36 SP |
2949 | /* When the result of a CLOSE_PHI is written to a memory location, |
2950 | return a pointer to that memory reference, otherwise return | |
2951 | NULL_TREE. */ | |
2952 | ||
2953 | static tree | |
2954 | close_phi_written_to_memory (gimple close_phi) | |
2955 | { | |
2956 | imm_use_iterator imm_iter; | |
50034a36 SP |
2957 | use_operand_p use_p; |
2958 | gimple stmt; | |
7c48ea69 | 2959 | tree res, def = gimple_phi_result (close_phi); |
50034a36 SP |
2960 | |
2961 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2962 | if ((stmt = USE_STMT (use_p)) | |
2963 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
7c48ea69 SP |
2964 | && (res = gimple_assign_lhs (stmt))) |
2965 | { | |
2966 | switch (TREE_CODE (res)) | |
2967 | { | |
2968 | case VAR_DECL: | |
2969 | case PARM_DECL: | |
2970 | case RESULT_DECL: | |
2971 | return res; | |
2972 | ||
2973 | case ARRAY_REF: | |
2974 | case MEM_REF: | |
2975 | { | |
2976 | tree arg = gimple_phi_arg_def (close_phi, 0); | |
2977 | loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg)); | |
2978 | ||
2979 | /* FIXME: this restriction is for id-{24,25}.f and | |
2980 | could be handled by duplicating the computation of | |
2981 | array indices before the loop of the close_phi. */ | |
2982 | if (for_each_index (&res, dr_indices_valid_in_loop, &nest)) | |
2983 | return res; | |
2984 | } | |
2985 | /* Fallthru. */ | |
50034a36 | 2986 | |
7c48ea69 SP |
2987 | default: |
2988 | continue; | |
2989 | } | |
2990 | } | |
50034a36 SP |
2991 | return NULL_TREE; |
2992 | } | |
2993 | ||
a0dd1440 SP |
2994 | /* Rewrite out of SSA the reduction described by the loop phi nodes |
2995 | IN, and the close phi nodes OUT. IN and OUT are structured by loop | |
2996 | levels like this: | |
2997 | ||
2998 | IN: stmt, loop_n, ..., loop_0 | |
2999 | OUT: stmt, close_n, ..., close_0 | |
3000 | ||
3001 | the first element is the reduction statement, and the next elements | |
3002 | are the loop and close phi nodes of each of the outer loops. */ | |
3003 | ||
3004 | static void | |
efa21390 | 3005 | translate_scalar_reduction_to_array (scop_p scop, |
9771b263 DN |
3006 | vec<gimple> in, |
3007 | vec<gimple> out) | |
a0dd1440 | 3008 | { |
a0dd1440 | 3009 | gimple loop_phi; |
9771b263 DN |
3010 | unsigned int i = out.length () - 1; |
3011 | tree red = close_phi_written_to_memory (out[i]); | |
a0dd1440 | 3012 | |
9771b263 | 3013 | FOR_EACH_VEC_ELT (in, i, loop_phi) |
a0dd1440 | 3014 | { |
9771b263 | 3015 | gimple close_phi = out[i]; |
a0dd1440 SP |
3016 | |
3017 | if (i == 0) | |
3018 | { | |
3019 | gimple stmt = loop_phi; | |
efa21390 SP |
3020 | basic_block bb = split_reduction_stmt (scop, stmt); |
3021 | poly_bb_p pbb = pbb_from_bb (bb); | |
3022 | PBB_IS_REDUCTION (pbb) = true; | |
a0dd1440 SP |
3023 | gcc_assert (close_phi == loop_phi); |
3024 | ||
50034a36 SP |
3025 | if (!red) |
3026 | red = create_zero_dim_array | |
3027 | (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction"); | |
3028 | ||
9771b263 | 3029 | translate_scalar_reduction_to_array_for_stmt (scop, red, stmt, in[1]); |
a0dd1440 SP |
3030 | continue; |
3031 | } | |
3032 | ||
9771b263 | 3033 | if (i == in.length () - 1) |
a0dd1440 | 3034 | { |
50034a36 SP |
3035 | insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi), |
3036 | unshare_expr (red), close_phi); | |
5fed5769 | 3037 | insert_out_of_ssa_copy_on_edge |
efa21390 | 3038 | (scop, edge_initial_value_for_loop_phi (loop_phi), |
50034a36 | 3039 | unshare_expr (red), initial_value_for_loop_phi (loop_phi)); |
a0dd1440 SP |
3040 | } |
3041 | ||
a4681954 SP |
3042 | remove_phi (loop_phi); |
3043 | remove_phi (close_phi); | |
a0dd1440 SP |
3044 | } |
3045 | } | |
3046 | ||
5d737345 SP |
3047 | /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns |
3048 | true when something has been changed. */ | |
a0dd1440 | 3049 | |
5d737345 | 3050 | static bool |
efa21390 SP |
3051 | rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop, |
3052 | gimple close_phi) | |
a0dd1440 | 3053 | { |
5d737345 | 3054 | bool res; |
9771b263 DN |
3055 | vec<gimple> in; |
3056 | in.create (10); | |
3057 | vec<gimple> out; | |
3058 | out.create (10); | |
a0dd1440 | 3059 | |
a30e5345 | 3060 | detect_commutative_reduction (scop, close_phi, &in, &out); |
9771b263 | 3061 | res = in.length () > 1; |
5d737345 | 3062 | if (res) |
efa21390 | 3063 | translate_scalar_reduction_to_array (scop, in, out); |
a0dd1440 | 3064 | |
9771b263 DN |
3065 | in.release (); |
3066 | out.release (); | |
5d737345 | 3067 | return res; |
a0dd1440 SP |
3068 | } |
3069 | ||
5d737345 SP |
3070 | /* Rewrites all the commutative reductions from LOOP out of SSA. |
3071 | Returns true when something has been changed. */ | |
a0dd1440 | 3072 | |
5d737345 | 3073 | static bool |
efa21390 SP |
3074 | rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop, |
3075 | loop_p loop) | |
a0dd1440 SP |
3076 | { |
3077 | gimple_stmt_iterator gsi; | |
3078 | edge exit = single_exit (loop); | |
4ee23fa8 | 3079 | tree res; |
5d737345 | 3080 | bool changed = false; |
a0dd1440 SP |
3081 | |
3082 | if (!exit) | |
5d737345 | 3083 | return false; |
a0dd1440 SP |
3084 | |
3085 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4ee23fa8 | 3086 | if ((res = gimple_phi_result (gsi_stmt (gsi))) |
ea057359 | 3087 | && !virtual_operand_p (res) |
efa21390 | 3088 | && !scev_analyzable_p (res, SCOP_REGION (scop))) |
5d737345 | 3089 | changed |= rewrite_commutative_reductions_out_of_ssa_close_phi |
efa21390 | 3090 | (scop, gsi_stmt (gsi)); |
5d737345 SP |
3091 | |
3092 | return changed; | |
a0dd1440 SP |
3093 | } |
3094 | ||
3095 | /* Rewrites all the commutative reductions from SCOP out of SSA. */ | |
3096 | ||
efa21390 SP |
3097 | static void |
3098 | rewrite_commutative_reductions_out_of_ssa (scop_p scop) | |
a0dd1440 SP |
3099 | { |
3100 | loop_iterator li; | |
3101 | loop_p loop; | |
5d737345 | 3102 | bool changed = false; |
efa21390 | 3103 | sese region = SCOP_REGION (scop); |
cc588970 | 3104 | |
a0dd1440 SP |
3105 | FOR_EACH_LOOP (li, loop, 0) |
3106 | if (loop_in_sese_p (loop, region)) | |
efa21390 | 3107 | changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop); |
6c4499b6 | 3108 | |
5d737345 SP |
3109 | if (changed) |
3110 | { | |
3111 | scev_reset_htab (); | |
3112 | gsi_commit_edge_inserts (); | |
3113 | update_ssa (TODO_update_ssa); | |
6c4499b6 | 3114 | #ifdef ENABLE_CHECKING |
5d737345 | 3115 | verify_loop_closed_ssa (true); |
6c4499b6 | 3116 | #endif |
5d737345 | 3117 | } |
a0dd1440 SP |
3118 | } |
3119 | ||
68d3ff90 TG |
3120 | /* Can all ivs be represented by a signed integer? |
3121 | As CLooG might generate negative values in its expressions, signed loop ivs | |
3122 | are required in the backend. */ | |
072edf07 | 3123 | |
68d3ff90 TG |
3124 | static bool |
3125 | scop_ivs_can_be_represented (scop_p scop) | |
3126 | { | |
3127 | loop_iterator li; | |
3128 | loop_p loop; | |
a0d1afb3 | 3129 | gimple_stmt_iterator psi; |
f5843d08 | 3130 | bool result = true; |
68d3ff90 TG |
3131 | |
3132 | FOR_EACH_LOOP (li, loop, 0) | |
3133 | { | |
68d3ff90 TG |
3134 | if (!loop_in_sese_p (loop, SCOP_REGION (scop))) |
3135 | continue; | |
3136 | ||
a0d1afb3 SP |
3137 | for (psi = gsi_start_phis (loop->header); |
3138 | !gsi_end_p (psi); gsi_next (&psi)) | |
3139 | { | |
3140 | gimple phi = gsi_stmt (psi); | |
3141 | tree res = PHI_RESULT (phi); | |
3142 | tree type = TREE_TYPE (res); | |
68d3ff90 | 3143 | |
a0d1afb3 | 3144 | if (TYPE_UNSIGNED (type) |
0d82a1c8 | 3145 | && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node)) |
f5843d08 RG |
3146 | { |
3147 | result = false; | |
3148 | break; | |
3149 | } | |
a0d1afb3 | 3150 | } |
f5843d08 RG |
3151 | if (!result) |
3152 | FOR_EACH_LOOP_BREAK (li); | |
68d3ff90 TG |
3153 | } |
3154 | ||
f5843d08 | 3155 | return result; |
68d3ff90 TG |
3156 | } |
3157 | ||
2abae5f1 SP |
3158 | /* Builds the polyhedral representation for a SESE region. */ |
3159 | ||
e84aaa33 | 3160 | void |
2abae5f1 SP |
3161 | build_poly_scop (scop_p scop) |
3162 | { | |
3163 | sese region = SCOP_REGION (scop); | |
4e7dd376 | 3164 | graphite_dim_t max_dim; |
a0dd1440 | 3165 | |
efa21390 | 3166 | build_scop_bbs (scop); |
2abae5f1 SP |
3167 | |
3168 | /* FIXME: This restriction is needed to avoid a problem in CLooG. | |
3169 | Once CLooG is fixed, remove this guard. Anyways, it makes no | |
3170 | sense to optimize a scop containing only PBBs that do not belong | |
3171 | to any loops. */ | |
3172 | if (nb_pbbs_in_loops (scop) == 0) | |
e84aaa33 | 3173 | return; |
2abae5f1 | 3174 | |
68d3ff90 | 3175 | if (!scop_ivs_can_be_represented (scop)) |
e84aaa33 | 3176 | return; |
68d3ff90 | 3177 | |
ac53c069 SP |
3178 | if (flag_associative_math) |
3179 | rewrite_commutative_reductions_out_of_ssa (scop); | |
3180 | ||
2abae5f1 SP |
3181 | build_sese_loop_nests (region); |
3182 | build_sese_conditions (region); | |
3183 | find_scop_parameters (scop); | |
3184 | ||
4e7dd376 SP |
3185 | max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS); |
3186 | if (scop_nb_params (scop) > max_dim) | |
e84aaa33 | 3187 | return; |
4e7dd376 | 3188 | |
2abae5f1 SP |
3189 | build_scop_iteration_domain (scop); |
3190 | build_scop_context (scop); | |
2abae5f1 | 3191 | add_conditions_to_constraints (scop); |
efa21390 SP |
3192 | |
3193 | /* Rewrite out of SSA only after having translated the | |
3194 | representation to the polyhedral representation to avoid scev | |
3195 | analysis failures. That means that these functions will insert | |
3196 | new data references that they create in the right place. */ | |
efa21390 SP |
3197 | rewrite_reductions_out_of_ssa (scop); |
3198 | rewrite_cross_bb_scalar_deps_out_of_ssa (scop); | |
3199 | ||
3200 | build_scop_drs (scop); | |
a36d12e2 | 3201 | scop_to_lst (scop); |
2abae5f1 | 3202 | build_scop_scattering (scop); |
2abae5f1 | 3203 | |
e84aaa33 SP |
3204 | /* This SCoP has been translated to the polyhedral |
3205 | representation. */ | |
3206 | POLY_SCOP_P (scop) = true; | |
2abae5f1 | 3207 | } |
2abae5f1 | 3208 | #endif |