]>
Commit | Line | Data |
---|---|---|
2abae5f1 | 1 | /* Conversion of SESE regions to Polyhedra. |
23a5b65a | 2 | Copyright (C) 2009-2014 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" | |
4d648807 | 36 | #include "tree.h" |
2fb9a547 AM |
37 | #include "basic-block.h" |
38 | #include "tree-ssa-alias.h" | |
39 | #include "internal-fn.h" | |
40 | #include "gimple-expr.h" | |
41 | #include "is-a.h" | |
18f429e2 | 42 | #include "gimple.h" |
5be5c238 | 43 | #include "gimple-iterator.h" |
18f429e2 AM |
44 | #include "gimplify.h" |
45 | #include "gimplify-me.h" | |
442b4905 AM |
46 | #include "gimple-ssa.h" |
47 | #include "tree-cfg.h" | |
48 | #include "tree-phinodes.h" | |
49 | #include "ssa-iterators.h" | |
d8a2d370 | 50 | #include "stringpool.h" |
442b4905 | 51 | #include "tree-ssanames.h" |
e28030cf AM |
52 | #include "tree-ssa-loop-manip.h" |
53 | #include "tree-ssa-loop-niter.h" | |
442b4905 AM |
54 | #include "tree-ssa-loop.h" |
55 | #include "tree-into-ssa.h" | |
7a1c57d3 | 56 | #include "tree-pass.h" |
2abae5f1 SP |
57 | #include "cfgloop.h" |
58 | #include "tree-chrec.h" | |
59 | #include "tree-data-ref.h" | |
60 | #include "tree-scalar-evolution.h" | |
2abae5f1 | 61 | #include "domwalk.h" |
2abae5f1 | 62 | #include "sese.h" |
440917de | 63 | #include "tree-ssa-propagate.h" |
2abae5f1 SP |
64 | |
65 | #ifdef HAVE_cloog | |
e4effef7 | 66 | #include "expr.h" |
2abae5f1 | 67 | #include "graphite-poly.h" |
2abae5f1 SP |
68 | #include "graphite-sese-to-poly.h" |
69 | ||
33ad93b9 RG |
70 | |
71 | /* Assigns to RES the value of the INTEGER_CST T. */ | |
72 | ||
73 | static inline void | |
74 | tree_int_to_gmp (tree t, mpz_t res) | |
75 | { | |
76 | double_int di = tree_to_double_int (t); | |
77 | mpz_set_double_int (res, di, TYPE_UNSIGNED (TREE_TYPE (t))); | |
78 | } | |
79 | ||
159e4616 SP |
80 | /* Returns the index of the PHI argument defined in the outermost |
81 | loop. */ | |
2abae5f1 SP |
82 | |
83 | static size_t | |
159e4616 | 84 | phi_arg_in_outermost_loop (gimple phi) |
2abae5f1 SP |
85 | { |
86 | loop_p loop = gimple_bb (phi)->loop_father; | |
159e4616 | 87 | size_t i, res = 0; |
2abae5f1 SP |
88 | |
89 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
90 | if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src)) | |
159e4616 SP |
91 | { |
92 | loop = gimple_phi_arg_edge (phi, i)->src->loop_father; | |
93 | res = i; | |
94 | } | |
2abae5f1 | 95 | |
159e4616 | 96 | return res; |
2abae5f1 SP |
97 | } |
98 | ||
99 | /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position | |
100 | PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */ | |
101 | ||
102 | static void | |
103 | remove_simple_copy_phi (gimple_stmt_iterator *psi) | |
104 | { | |
105 | gimple phi = gsi_stmt (*psi); | |
106 | tree res = gimple_phi_result (phi); | |
159e4616 | 107 | size_t entry = phi_arg_in_outermost_loop (phi); |
2abae5f1 SP |
108 | tree init = gimple_phi_arg_def (phi, entry); |
109 | gimple stmt = gimple_build_assign (res, init); | |
110 | edge e = gimple_phi_arg_edge (phi, entry); | |
111 | ||
112 | remove_phi_node (psi, false); | |
113 | gsi_insert_on_edge_immediate (e, stmt); | |
2abae5f1 SP |
114 | } |
115 | ||
116 | /* Removes an invariant phi node at position PSI by inserting on the | |
117 | loop ENTRY edge the assignment RES = INIT. */ | |
118 | ||
119 | static void | |
120 | remove_invariant_phi (sese region, gimple_stmt_iterator *psi) | |
121 | { | |
122 | gimple phi = gsi_stmt (*psi); | |
123 | loop_p loop = loop_containing_stmt (phi); | |
124 | tree res = gimple_phi_result (phi); | |
125 | tree scev = scalar_evolution_in_region (region, loop, res); | |
159e4616 | 126 | size_t entry = phi_arg_in_outermost_loop (phi); |
2abae5f1 SP |
127 | edge e = gimple_phi_arg_edge (phi, entry); |
128 | tree var; | |
129 | gimple stmt; | |
355a7673 | 130 | gimple_seq stmts = NULL; |
2abae5f1 SP |
131 | |
132 | if (tree_contains_chrecs (scev, NULL)) | |
133 | scev = gimple_phi_arg_def (phi, entry); | |
134 | ||
135 | var = force_gimple_operand (scev, &stmts, true, NULL_TREE); | |
136 | stmt = gimple_build_assign (res, var); | |
137 | remove_phi_node (psi, false); | |
138 | ||
355a7673 | 139 | gimple_seq_add_stmt (&stmts, stmt); |
2abae5f1 SP |
140 | gsi_insert_seq_on_edge (e, stmts); |
141 | gsi_commit_edge_inserts (); | |
142 | SSA_NAME_DEF_STMT (res) = stmt; | |
143 | } | |
144 | ||
145 | /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */ | |
146 | ||
147 | static inline bool | |
148 | simple_copy_phi_p (gimple phi) | |
149 | { | |
150 | tree res; | |
151 | ||
152 | if (gimple_phi_num_args (phi) != 2) | |
153 | return false; | |
154 | ||
155 | res = gimple_phi_result (phi); | |
156 | return (res == gimple_phi_arg_def (phi, 0) | |
157 | || res == gimple_phi_arg_def (phi, 1)); | |
158 | } | |
159 | ||
160 | /* Returns true when the phi node at position PSI is a reduction phi | |
161 | node in REGION. Otherwise moves the pointer PSI to the next phi to | |
162 | be considered. */ | |
163 | ||
164 | static bool | |
165 | reduction_phi_p (sese region, gimple_stmt_iterator *psi) | |
166 | { | |
167 | loop_p loop; | |
2abae5f1 SP |
168 | gimple phi = gsi_stmt (*psi); |
169 | tree res = gimple_phi_result (phi); | |
170 | ||
2abae5f1 SP |
171 | loop = loop_containing_stmt (phi); |
172 | ||
173 | if (simple_copy_phi_p (phi)) | |
174 | { | |
a5a59b11 | 175 | /* PRE introduces phi nodes like these, for an example, |
2abae5f1 SP |
176 | see id-5.f in the fortran graphite testsuite: |
177 | ||
178 | # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)> | |
179 | */ | |
180 | remove_simple_copy_phi (psi); | |
181 | return false; | |
182 | } | |
183 | ||
87b28340 | 184 | if (scev_analyzable_p (res, region)) |
2abae5f1 | 185 | { |
87b28340 SP |
186 | tree scev = scalar_evolution_in_region (region, loop, res); |
187 | ||
188 | if (evolution_function_is_invariant_p (scev, loop->num)) | |
7cc4ff8d SP |
189 | remove_invariant_phi (region, psi); |
190 | else | |
191 | gsi_next (psi); | |
192 | ||
2abae5f1 SP |
193 | return false; |
194 | } | |
195 | ||
2abae5f1 SP |
196 | /* All the other cases are considered reductions. */ |
197 | return true; | |
198 | } | |
199 | ||
2abae5f1 SP |
200 | /* Store the GRAPHITE representation of BB. */ |
201 | ||
202 | static gimple_bb_p | |
9771b263 | 203 | new_gimple_bb (basic_block bb, vec<data_reference_p> drs) |
2abae5f1 SP |
204 | { |
205 | struct gimple_bb *gbb; | |
206 | ||
207 | gbb = XNEW (struct gimple_bb); | |
208 | bb->aux = gbb; | |
209 | GBB_BB (gbb) = bb; | |
210 | GBB_DATA_REFS (gbb) = drs; | |
9771b263 DN |
211 | GBB_CONDITIONS (gbb).create (0); |
212 | GBB_CONDITION_CASES (gbb).create (0); | |
2abae5f1 SP |
213 | |
214 | return gbb; | |
215 | } | |
216 | ||
1825f9a2 | 217 | static void |
9771b263 | 218 | free_data_refs_aux (vec<data_reference_p> datarefs) |
1825f9a2 LF |
219 | { |
220 | unsigned int i; | |
221 | struct data_reference *dr; | |
fb00d28e | 222 | |
9771b263 | 223 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
fb00d28e | 224 | if (dr->aux) |
1825f9a2 | 225 | { |
2b178a5f | 226 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); |
fb00d28e | 227 | |
04695783 | 228 | free (bap->alias_set); |
fb00d28e | 229 | |
2b178a5f | 230 | free (bap); |
1825f9a2 LF |
231 | dr->aux = NULL; |
232 | } | |
233 | } | |
2abae5f1 SP |
234 | /* Frees GBB. */ |
235 | ||
236 | static void | |
237 | free_gimple_bb (struct gimple_bb *gbb) | |
238 | { | |
1825f9a2 | 239 | free_data_refs_aux (GBB_DATA_REFS (gbb)); |
2abae5f1 SP |
240 | free_data_refs (GBB_DATA_REFS (gbb)); |
241 | ||
9771b263 DN |
242 | GBB_CONDITIONS (gbb).release (); |
243 | GBB_CONDITION_CASES (gbb).release (); | |
2abae5f1 SP |
244 | GBB_BB (gbb)->aux = 0; |
245 | XDELETE (gbb); | |
246 | } | |
247 | ||
248 | /* Deletes all gimple bbs in SCOP. */ | |
249 | ||
250 | static void | |
251 | remove_gbbs_in_scop (scop_p scop) | |
252 | { | |
253 | int i; | |
254 | poly_bb_p pbb; | |
255 | ||
9771b263 | 256 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
257 | free_gimple_bb (PBB_BLACK_BOX (pbb)); |
258 | } | |
259 | ||
260 | /* Deletes all scops in SCOPS. */ | |
261 | ||
262 | void | |
9771b263 | 263 | free_scops (vec<scop_p> scops) |
2abae5f1 SP |
264 | { |
265 | int i; | |
266 | scop_p scop; | |
267 | ||
9771b263 | 268 | FOR_EACH_VEC_ELT (scops, i, scop) |
2abae5f1 SP |
269 | { |
270 | remove_gbbs_in_scop (scop); | |
271 | free_sese (SCOP_REGION (scop)); | |
272 | free_scop (scop); | |
273 | } | |
274 | ||
9771b263 | 275 | scops.release (); |
2abae5f1 SP |
276 | } |
277 | ||
5c640e29 SP |
278 | /* Same as outermost_loop_in_sese, returns the outermost loop |
279 | containing BB in REGION, but makes sure that the returned loop | |
280 | belongs to the REGION, and so this returns the first loop in the | |
281 | REGION when the loop containing BB does not belong to REGION. */ | |
282 | ||
283 | static loop_p | |
284 | outermost_loop_in_sese_1 (sese region, basic_block bb) | |
285 | { | |
286 | loop_p nest = outermost_loop_in_sese (region, bb); | |
287 | ||
288 | if (loop_in_sese_p (nest, region)) | |
289 | return nest; | |
290 | ||
291 | /* When the basic block BB does not belong to a loop in the region, | |
292 | return the first loop in the region. */ | |
293 | nest = nest->inner; | |
294 | while (nest) | |
295 | if (loop_in_sese_p (nest, region)) | |
296 | break; | |
297 | else | |
298 | nest = nest->next; | |
299 | ||
300 | gcc_assert (nest); | |
301 | return nest; | |
302 | } | |
303 | ||
2abae5f1 SP |
304 | /* Generates a polyhedral black box only if the bb contains interesting |
305 | information. */ | |
306 | ||
efa21390 SP |
307 | static gimple_bb_p |
308 | try_generate_gimple_bb (scop_p scop, basic_block bb) | |
2abae5f1 | 309 | { |
9771b263 DN |
310 | vec<data_reference_p> drs; |
311 | drs.create (5); | |
5c640e29 SP |
312 | sese region = SCOP_REGION (scop); |
313 | loop_p nest = outermost_loop_in_sese_1 (region, bb); | |
2abae5f1 SP |
314 | gimple_stmt_iterator gsi; |
315 | ||
316 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
a3201927 AO |
317 | { |
318 | gimple stmt = gsi_stmt (gsi); | |
5c640e29 SP |
319 | loop_p loop; |
320 | ||
321 | if (is_gimple_debug (stmt)) | |
322 | continue; | |
323 | ||
324 | loop = loop_containing_stmt (stmt); | |
325 | if (!loop_in_sese_p (loop, region)) | |
326 | loop = nest; | |
327 | ||
328 | graphite_find_data_references_in_stmt (nest, loop, stmt, &drs); | |
a3201927 | 329 | } |
2abae5f1 | 330 | |
efa21390 | 331 | return new_gimple_bb (bb, drs); |
2abae5f1 SP |
332 | } |
333 | ||
334 | /* Returns true if all predecessors of BB, that are not dominated by BB, are | |
335 | marked in MAP. The predecessors dominated by BB are loop latches and will | |
336 | be handled after BB. */ | |
337 | ||
338 | static bool | |
339 | all_non_dominated_preds_marked_p (basic_block bb, sbitmap map) | |
340 | { | |
341 | edge e; | |
342 | edge_iterator ei; | |
343 | ||
344 | FOR_EACH_EDGE (e, ei, bb->preds) | |
d7c028c0 | 345 | if (!bitmap_bit_p (map, e->src->index) |
2abae5f1 SP |
346 | && !dominated_by_p (CDI_DOMINATORS, e->src, bb)) |
347 | return false; | |
348 | ||
349 | return true; | |
350 | } | |
351 | ||
352 | /* Compare the depth of two basic_block's P1 and P2. */ | |
353 | ||
354 | static int | |
355 | compare_bb_depths (const void *p1, const void *p2) | |
356 | { | |
357 | const_basic_block const bb1 = *(const_basic_block const*)p1; | |
358 | const_basic_block const bb2 = *(const_basic_block const*)p2; | |
359 | int d1 = loop_depth (bb1->loop_father); | |
360 | int d2 = loop_depth (bb2->loop_father); | |
361 | ||
362 | if (d1 < d2) | |
363 | return 1; | |
364 | ||
365 | if (d1 > d2) | |
366 | return -1; | |
367 | ||
368 | return 0; | |
369 | } | |
370 | ||
371 | /* Sort the basic blocks from DOM such that the first are the ones at | |
372 | a deepest loop level. */ | |
373 | ||
374 | static void | |
9771b263 | 375 | graphite_sort_dominated_info (vec<basic_block> dom) |
2abae5f1 | 376 | { |
9771b263 | 377 | dom.qsort (compare_bb_depths); |
2abae5f1 SP |
378 | } |
379 | ||
380 | /* Recursive helper function for build_scops_bbs. */ | |
381 | ||
382 | static void | |
efa21390 | 383 | build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb) |
2abae5f1 SP |
384 | { |
385 | sese region = SCOP_REGION (scop); | |
9771b263 | 386 | vec<basic_block> dom; |
efa21390 | 387 | poly_bb_p pbb; |
2abae5f1 | 388 | |
d7c028c0 | 389 | if (bitmap_bit_p (visited, bb->index) |
2abae5f1 SP |
390 | || !bb_in_sese_p (bb, region)) |
391 | return; | |
392 | ||
efa21390 | 393 | pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb)); |
9771b263 | 394 | SCOP_BBS (scop).safe_push (pbb); |
d7c028c0 | 395 | bitmap_set_bit (visited, bb->index); |
2abae5f1 SP |
396 | |
397 | dom = get_dominated_by (CDI_DOMINATORS, bb); | |
398 | ||
9771b263 | 399 | if (!dom.exists ()) |
2abae5f1 SP |
400 | return; |
401 | ||
402 | graphite_sort_dominated_info (dom); | |
403 | ||
9771b263 | 404 | while (!dom.is_empty ()) |
2abae5f1 SP |
405 | { |
406 | int i; | |
407 | basic_block dom_bb; | |
408 | ||
9771b263 | 409 | FOR_EACH_VEC_ELT (dom, i, dom_bb) |
2abae5f1 SP |
410 | if (all_non_dominated_preds_marked_p (dom_bb, visited)) |
411 | { | |
efa21390 | 412 | build_scop_bbs_1 (scop, visited, dom_bb); |
9771b263 | 413 | dom.unordered_remove (i); |
2abae5f1 SP |
414 | break; |
415 | } | |
416 | } | |
417 | ||
9771b263 | 418 | dom.release (); |
2abae5f1 SP |
419 | } |
420 | ||
421 | /* Gather the basic blocks belonging to the SCOP. */ | |
422 | ||
efa21390 SP |
423 | static void |
424 | build_scop_bbs (scop_p scop) | |
2abae5f1 | 425 | { |
8b1c6fd7 | 426 | sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
2abae5f1 SP |
427 | sese region = SCOP_REGION (scop); |
428 | ||
f61e445a | 429 | bitmap_clear (visited); |
efa21390 | 430 | build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region)); |
2abae5f1 SP |
431 | sbitmap_free (visited); |
432 | } | |
433 | ||
33ad93b9 RG |
434 | /* Return an ISL identifier for the polyhedral basic block PBB. */ |
435 | ||
436 | static isl_id * | |
437 | isl_id_for_pbb (scop_p s, poly_bb_p pbb) | |
438 | { | |
439 | char name[50]; | |
440 | snprintf (name, sizeof (name), "S_%d", pbb_index (pbb)); | |
441 | return isl_id_alloc (s->ctx, name, pbb); | |
442 | } | |
443 | ||
2abae5f1 SP |
444 | /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron. |
445 | We generate SCATTERING_DIMENSIONS scattering dimensions. | |
446 | ||
447 | CLooG 0.15.0 and previous versions require, that all | |
448 | scattering functions of one CloogProgram have the same number of | |
449 | scattering dimensions, therefore we allow to specify it. This | |
450 | should be removed in future versions of CLooG. | |
451 | ||
452 | The scattering polyhedron consists of these dimensions: scattering, | |
453 | loop_iterators, parameters. | |
454 | ||
455 | Example: | |
456 | ||
457 | | scattering_dimensions = 5 | |
458 | | used_scattering_dimensions = 3 | |
459 | | nb_iterators = 1 | |
460 | | scop_nb_params = 2 | |
461 | | | |
462 | | Schedule: | |
463 | | i | |
464 | | 4 5 | |
465 | | | |
466 | | Scattering polyhedron: | |
467 | | | |
468 | | scattering: {s1, s2, s3, s4, s5} | |
469 | | loop_iterators: {i} | |
470 | | parameters: {p1, p2} | |
471 | | | |
472 | | s1 s2 s3 s4 s5 i p1 p2 1 | |
473 | | 1 0 0 0 0 0 0 0 -4 = 0 | |
474 | | 0 1 0 0 0 -1 0 0 0 = 0 | |
475 | | 0 0 1 0 0 0 0 0 -5 = 0 */ | |
476 | ||
477 | static void | |
33ad93b9 | 478 | build_pbb_scattering_polyhedrons (isl_aff *static_sched, |
2abae5f1 SP |
479 | poly_bb_p pbb, int scattering_dimensions) |
480 | { | |
481 | int i; | |
2abae5f1 SP |
482 | int nb_iterators = pbb_dim_iter_domain (pbb); |
483 | int used_scattering_dimensions = nb_iterators * 2 + 1; | |
33ad93b9 RG |
484 | isl_int val; |
485 | isl_space *dc, *dm; | |
2abae5f1 SP |
486 | |
487 | gcc_assert (scattering_dimensions >= used_scattering_dimensions); | |
488 | ||
33ad93b9 | 489 | isl_int_init (val); |
2abae5f1 | 490 | |
33ad93b9 RG |
491 | dc = isl_set_get_space (pbb->domain); |
492 | dm = isl_space_add_dims (isl_space_from_domain (dc), | |
493 | isl_dim_out, scattering_dimensions); | |
494 | pbb->schedule = isl_map_universe (dm); | |
2abae5f1 SP |
495 | |
496 | for (i = 0; i < scattering_dimensions; i++) | |
497 | { | |
2abae5f1 SP |
498 | /* Textual order inside this loop. */ |
499 | if ((i % 2) == 0) | |
500 | { | |
33ad93b9 RG |
501 | isl_constraint *c = isl_equality_alloc |
502 | (isl_local_space_from_space (isl_map_get_space (pbb->schedule))); | |
503 | ||
504 | if (0 != isl_aff_get_coefficient (static_sched, isl_dim_in, | |
505 | i / 2, &val)) | |
506 | gcc_unreachable (); | |
507 | ||
508 | isl_int_neg (val, val); | |
509 | c = isl_constraint_set_constant (c, val); | |
510 | c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1); | |
511 | pbb->schedule = isl_map_add_constraint (pbb->schedule, c); | |
2abae5f1 SP |
512 | } |
513 | ||
514 | /* Iterations of this loop. */ | |
515 | else /* if ((i % 2) == 1) */ | |
516 | { | |
517 | int loop = (i - 1) / 2; | |
33ad93b9 RG |
518 | pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop, |
519 | isl_dim_out, i); | |
2abae5f1 | 520 | } |
2abae5f1 SP |
521 | } |
522 | ||
33ad93b9 | 523 | isl_int_clear (val); |
2abae5f1 | 524 | |
33ad93b9 | 525 | pbb->transformed = isl_map_copy (pbb->schedule); |
2abae5f1 SP |
526 | } |
527 | ||
528 | /* Build for BB the static schedule. | |
529 | ||
530 | The static schedule is a Dewey numbering of the abstract syntax | |
531 | tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification | |
532 | ||
533 | The following example informally defines the static schedule: | |
534 | ||
535 | A | |
536 | for (i: ...) | |
537 | { | |
538 | for (j: ...) | |
539 | { | |
540 | B | |
541 | C | |
542 | } | |
543 | ||
544 | for (k: ...) | |
545 | { | |
546 | D | |
547 | E | |
548 | } | |
549 | } | |
550 | F | |
551 | ||
552 | Static schedules for A to F: | |
553 | ||
554 | DEPTH | |
555 | 0 1 2 | |
556 | A 0 | |
557 | B 1 0 0 | |
558 | C 1 0 1 | |
559 | D 1 1 0 | |
560 | E 1 1 1 | |
561 | F 2 | |
562 | */ | |
563 | ||
564 | static void | |
565 | build_scop_scattering (scop_p scop) | |
566 | { | |
567 | int i; | |
568 | poly_bb_p pbb; | |
569 | gimple_bb_p previous_gbb = NULL; | |
33ad93b9 RG |
570 | isl_space *dc = isl_set_get_space (scop->context); |
571 | isl_aff *static_sched; | |
2abae5f1 | 572 | |
0fc822d0 | 573 | dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun)); |
33ad93b9 | 574 | static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc)); |
2abae5f1 SP |
575 | |
576 | /* We have to start schedules at 0 on the first component and | |
577 | because we cannot compare_prefix_loops against a previous loop, | |
578 | prefix will be equal to zero, and that index will be | |
579 | incremented before copying. */ | |
33ad93b9 | 580 | static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1); |
2abae5f1 | 581 | |
9771b263 | 582 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
583 | { |
584 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); | |
2abae5f1 SP |
585 | int prefix; |
586 | int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1; | |
587 | ||
588 | if (previous_gbb) | |
589 | prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb); | |
590 | else | |
591 | prefix = 0; | |
592 | ||
593 | previous_gbb = gbb; | |
2abae5f1 | 594 | |
33ad93b9 RG |
595 | static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, |
596 | prefix, 1); | |
597 | build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims); | |
598 | } | |
599 | ||
600 | isl_aff_free (static_sched); | |
601 | } | |
602 | ||
603 | static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space); | |
604 | ||
605 | /* Extract an affine expression from the chain of recurrence E. */ | |
2abae5f1 | 606 | |
33ad93b9 RG |
607 | static isl_pw_aff * |
608 | extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space) | |
609 | { | |
610 | isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space)); | |
611 | isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space)); | |
612 | isl_local_space *ls = isl_local_space_from_space (space); | |
0fc822d0 | 613 | unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1; |
33ad93b9 RG |
614 | isl_aff *loop = isl_aff_set_coefficient_si |
615 | (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1); | |
616 | isl_pw_aff *l = isl_pw_aff_from_aff (loop); | |
617 | ||
618 | /* Before multiplying, make sure that the result is affine. */ | |
619 | gcc_assert (isl_pw_aff_is_cst (rhs) | |
620 | || isl_pw_aff_is_cst (l)); | |
621 | ||
622 | return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l)); | |
623 | } | |
624 | ||
625 | /* Extract an affine expression from the mult_expr E. */ | |
626 | ||
627 | static isl_pw_aff * | |
628 | extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space) | |
629 | { | |
630 | isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0), | |
631 | isl_space_copy (space)); | |
632 | isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
2abae5f1 | 633 | |
33ad93b9 RG |
634 | if (!isl_pw_aff_is_cst (lhs) |
635 | && !isl_pw_aff_is_cst (rhs)) | |
636 | { | |
637 | isl_pw_aff_free (lhs); | |
638 | isl_pw_aff_free (rhs); | |
639 | return NULL; | |
2abae5f1 SP |
640 | } |
641 | ||
33ad93b9 | 642 | return isl_pw_aff_mul (lhs, rhs); |
2abae5f1 SP |
643 | } |
644 | ||
33ad93b9 | 645 | /* Return an ISL identifier from the name of the ssa_name E. */ |
2abae5f1 | 646 | |
33ad93b9 RG |
647 | static isl_id * |
648 | isl_id_for_ssa_name (scop_p s, tree e) | |
2abae5f1 | 649 | { |
33ad93b9 RG |
650 | const char *name = get_name (e); |
651 | isl_id *id; | |
652 | ||
653 | if (name) | |
654 | id = isl_id_alloc (s->ctx, name, e); | |
655 | else | |
656 | { | |
657 | char name1[50]; | |
658 | snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e)); | |
659 | id = isl_id_alloc (s->ctx, name1, e); | |
660 | } | |
2abae5f1 | 661 | |
33ad93b9 RG |
662 | return id; |
663 | } | |
2abae5f1 | 664 | |
33ad93b9 | 665 | /* Return an ISL identifier for the data reference DR. */ |
2abae5f1 | 666 | |
33ad93b9 RG |
667 | static isl_id * |
668 | isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED) | |
669 | { | |
670 | /* Data references all get the same isl_id. They need to be comparable | |
671 | and are distinguished through the first dimension, which contains the | |
672 | alias set number. */ | |
673 | return isl_id_alloc (s->ctx, "", 0); | |
2abae5f1 SP |
674 | } |
675 | ||
33ad93b9 | 676 | /* Extract an affine expression from the ssa_name E. */ |
2abae5f1 | 677 | |
33ad93b9 RG |
678 | static isl_pw_aff * |
679 | extract_affine_name (scop_p s, tree e, __isl_take isl_space *space) | |
2abae5f1 | 680 | { |
33ad93b9 RG |
681 | isl_aff *aff; |
682 | isl_set *dom; | |
683 | isl_id *id; | |
684 | int dimension; | |
685 | ||
686 | id = isl_id_for_ssa_name (s, e); | |
687 | dimension = isl_space_find_dim_by_id (space, isl_dim_param, id); | |
c3284718 | 688 | isl_id_free (id); |
33ad93b9 RG |
689 | dom = isl_set_universe (isl_space_copy (space)); |
690 | aff = isl_aff_zero_on_domain (isl_local_space_from_space (space)); | |
691 | aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1); | |
692 | return isl_pw_aff_alloc (dom, aff); | |
693 | } | |
2abae5f1 | 694 | |
33ad93b9 | 695 | /* Extract an affine expression from the gmp constant G. */ |
2abae5f1 | 696 | |
33ad93b9 RG |
697 | static isl_pw_aff * |
698 | extract_affine_gmp (mpz_t g, __isl_take isl_space *space) | |
699 | { | |
700 | isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space)); | |
701 | isl_aff *aff = isl_aff_zero_on_domain (ls); | |
702 | isl_set *dom = isl_set_universe (space); | |
703 | isl_int v; | |
2abae5f1 | 704 | |
33ad93b9 RG |
705 | isl_int_init (v); |
706 | isl_int_set_gmp (v, g); | |
707 | aff = isl_aff_add_constant (aff, v); | |
708 | isl_int_clear (v); | |
2abae5f1 | 709 | |
33ad93b9 | 710 | return isl_pw_aff_alloc (dom, aff); |
2abae5f1 SP |
711 | } |
712 | ||
33ad93b9 | 713 | /* Extract an affine expression from the integer_cst E. */ |
2abae5f1 | 714 | |
33ad93b9 RG |
715 | static isl_pw_aff * |
716 | extract_affine_int (tree e, __isl_take isl_space *space) | |
717 | { | |
718 | isl_pw_aff *res; | |
719 | mpz_t g; | |
720 | ||
721 | mpz_init (g); | |
722 | tree_int_to_gmp (e, g); | |
723 | res = extract_affine_gmp (g, space); | |
724 | mpz_clear (g); | |
725 | ||
726 | return res; | |
727 | } | |
728 | ||
729 | /* Compute pwaff mod 2^width. */ | |
730 | ||
731 | static isl_pw_aff * | |
732 | wrap (isl_pw_aff *pwaff, unsigned width) | |
2abae5f1 | 733 | { |
33ad93b9 | 734 | isl_int mod; |
2abae5f1 | 735 | |
33ad93b9 RG |
736 | isl_int_init (mod); |
737 | isl_int_set_si (mod, 1); | |
738 | isl_int_mul_2exp (mod, mod, width); | |
2abae5f1 | 739 | |
33ad93b9 | 740 | pwaff = isl_pw_aff_mod (pwaff, mod); |
2abae5f1 | 741 | |
33ad93b9 RG |
742 | isl_int_clear (mod); |
743 | ||
744 | return pwaff; | |
2abae5f1 SP |
745 | } |
746 | ||
2abae5f1 SP |
747 | /* When parameter NAME is in REGION, returns its index in SESE_PARAMS. |
748 | Otherwise returns -1. */ | |
749 | ||
750 | static inline int | |
751 | parameter_index_in_region_1 (tree name, sese region) | |
752 | { | |
753 | int i; | |
754 | tree p; | |
755 | ||
756 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
757 | ||
9771b263 | 758 | FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p) |
2abae5f1 SP |
759 | if (p == name) |
760 | return i; | |
761 | ||
762 | return -1; | |
763 | } | |
764 | ||
765 | /* When the parameter NAME is in REGION, returns its index in | |
766 | SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS | |
767 | and returns the index of NAME. */ | |
768 | ||
769 | static int | |
770 | parameter_index_in_region (tree name, sese region) | |
771 | { | |
772 | int i; | |
773 | ||
774 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
775 | ||
776 | i = parameter_index_in_region_1 (name, region); | |
777 | if (i != -1) | |
778 | return i; | |
779 | ||
780 | gcc_assert (SESE_ADD_PARAMS (region)); | |
781 | ||
9771b263 DN |
782 | i = SESE_PARAMS (region).length (); |
783 | SESE_PARAMS (region).safe_push (name); | |
2abae5f1 SP |
784 | return i; |
785 | } | |
786 | ||
33ad93b9 | 787 | /* Extract an affine expression from the tree E in the scop S. */ |
2abae5f1 | 788 | |
33ad93b9 RG |
789 | static isl_pw_aff * |
790 | extract_affine (scop_p s, tree e, __isl_take isl_space *space) | |
2abae5f1 | 791 | { |
33ad93b9 RG |
792 | isl_pw_aff *lhs, *rhs, *res; |
793 | tree type; | |
794 | ||
795 | if (e == chrec_dont_know) { | |
796 | isl_space_free (space); | |
797 | return NULL; | |
798 | } | |
2abae5f1 SP |
799 | |
800 | switch (TREE_CODE (e)) | |
801 | { | |
802 | case POLYNOMIAL_CHREC: | |
33ad93b9 | 803 | res = extract_affine_chrec (s, e, space); |
2abae5f1 SP |
804 | break; |
805 | ||
806 | case MULT_EXPR: | |
33ad93b9 | 807 | res = extract_affine_mul (s, e, space); |
2abae5f1 SP |
808 | break; |
809 | ||
810 | case PLUS_EXPR: | |
811 | case POINTER_PLUS_EXPR: | |
33ad93b9 RG |
812 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); |
813 | rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
814 | res = isl_pw_aff_add (lhs, rhs); | |
2abae5f1 SP |
815 | break; |
816 | ||
817 | case MINUS_EXPR: | |
33ad93b9 RG |
818 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); |
819 | rhs = extract_affine (s, TREE_OPERAND (e, 1), space); | |
820 | res = isl_pw_aff_sub (lhs, rhs); | |
821 | break; | |
2abae5f1 SP |
822 | |
823 | case NEGATE_EXPR: | |
33ad93b9 RG |
824 | case BIT_NOT_EXPR: |
825 | lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space)); | |
826 | rhs = extract_affine (s, integer_minus_one_node, space); | |
827 | res = isl_pw_aff_mul (lhs, rhs); | |
828 | break; | |
2abae5f1 | 829 | |
33ad93b9 RG |
830 | case SSA_NAME: |
831 | gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s))); | |
832 | res = extract_affine_name (s, e, space); | |
833 | break; | |
2abae5f1 | 834 | |
33ad93b9 RG |
835 | case INTEGER_CST: |
836 | res = extract_affine_int (e, space); | |
837 | /* No need to wrap a single integer. */ | |
838 | return res; | |
2abae5f1 | 839 | |
33ad93b9 RG |
840 | CASE_CONVERT: |
841 | case NON_LVALUE_EXPR: | |
842 | res = extract_affine (s, TREE_OPERAND (e, 0), space); | |
843 | break; | |
2abae5f1 | 844 | |
33ad93b9 RG |
845 | default: |
846 | gcc_unreachable (); | |
847 | break; | |
848 | } | |
2abae5f1 | 849 | |
33ad93b9 RG |
850 | type = TREE_TYPE (e); |
851 | if (TYPE_UNSIGNED (type)) | |
852 | res = wrap (res, TYPE_PRECISION (type)); | |
2abae5f1 | 853 | |
33ad93b9 RG |
854 | return res; |
855 | } | |
2abae5f1 | 856 | |
33ad93b9 RG |
857 | /* In the context of sese S, scan the expression E and translate it to |
858 | a linear expression C. When parsing a symbolic multiplication, K | |
859 | represents the constant multiplier of an expression containing | |
860 | parameters. */ | |
2abae5f1 | 861 | |
33ad93b9 RG |
862 | static void |
863 | scan_tree_for_params (sese s, tree e) | |
864 | { | |
865 | if (e == chrec_dont_know) | |
866 | return; | |
2abae5f1 | 867 | |
33ad93b9 RG |
868 | switch (TREE_CODE (e)) |
869 | { | |
870 | case POLYNOMIAL_CHREC: | |
871 | scan_tree_for_params (s, CHREC_LEFT (e)); | |
872 | break; | |
2abae5f1 | 873 | |
33ad93b9 RG |
874 | case MULT_EXPR: |
875 | if (chrec_contains_symbols (TREE_OPERAND (e, 0))) | |
876 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); | |
877 | else | |
878 | scan_tree_for_params (s, TREE_OPERAND (e, 1)); | |
879 | break; | |
2abae5f1 | 880 | |
33ad93b9 RG |
881 | case PLUS_EXPR: |
882 | case POINTER_PLUS_EXPR: | |
883 | case MINUS_EXPR: | |
884 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); | |
885 | scan_tree_for_params (s, TREE_OPERAND (e, 1)); | |
2abae5f1 SP |
886 | break; |
887 | ||
33ad93b9 RG |
888 | case NEGATE_EXPR: |
889 | case BIT_NOT_EXPR: | |
2abae5f1 SP |
890 | CASE_CONVERT: |
891 | case NON_LVALUE_EXPR: | |
33ad93b9 | 892 | scan_tree_for_params (s, TREE_OPERAND (e, 0)); |
2abae5f1 SP |
893 | break; |
894 | ||
33ad93b9 RG |
895 | case SSA_NAME: |
896 | parameter_index_in_region (e, s); | |
897 | break; | |
898 | ||
899 | case INTEGER_CST: | |
f4a2e571 SP |
900 | case ADDR_EXPR: |
901 | break; | |
902 | ||
2abae5f1 SP |
903 | default: |
904 | gcc_unreachable (); | |
905 | break; | |
906 | } | |
907 | } | |
908 | ||
2abae5f1 SP |
909 | /* Find parameters with respect to REGION in BB. We are looking in memory |
910 | access functions, conditions and loop bounds. */ | |
911 | ||
912 | static void | |
913 | find_params_in_bb (sese region, gimple_bb_p gbb) | |
914 | { | |
915 | int i; | |
54fc808a | 916 | unsigned j; |
2abae5f1 SP |
917 | data_reference_p dr; |
918 | gimple stmt; | |
919 | loop_p loop = GBB_BB (gbb)->loop_father; | |
2abae5f1 | 920 | |
54fc808a | 921 | /* Find parameters in the access functions of data references. */ |
9771b263 | 922 | FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr) |
54fc808a | 923 | for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++) |
33ad93b9 | 924 | scan_tree_for_params (region, DR_ACCESS_FN (dr, j)); |
2abae5f1 SP |
925 | |
926 | /* Find parameters in conditional statements. */ | |
9771b263 | 927 | FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt) |
2abae5f1 | 928 | { |
2abae5f1 SP |
929 | tree lhs = scalar_evolution_in_region (region, loop, |
930 | gimple_cond_lhs (stmt)); | |
931 | tree rhs = scalar_evolution_in_region (region, loop, | |
932 | gimple_cond_rhs (stmt)); | |
933 | ||
33ad93b9 RG |
934 | scan_tree_for_params (region, lhs); |
935 | scan_tree_for_params (region, rhs); | |
2abae5f1 SP |
936 | } |
937 | } | |
938 | ||
939 | /* Record the parameters used in the SCOP. A variable is a parameter | |
940 | in a scop if it does not vary during the execution of that scop. */ | |
941 | ||
942 | static void | |
943 | find_scop_parameters (scop_p scop) | |
944 | { | |
945 | poly_bb_p pbb; | |
946 | unsigned i; | |
947 | sese region = SCOP_REGION (scop); | |
948 | struct loop *loop; | |
33ad93b9 | 949 | int nbp; |
2abae5f1 SP |
950 | |
951 | /* Find the parameters used in the loop bounds. */ | |
9771b263 | 952 | FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop) |
2abae5f1 SP |
953 | { |
954 | tree nb_iters = number_of_latch_executions (loop); | |
955 | ||
956 | if (!chrec_contains_symbols (nb_iters)) | |
957 | continue; | |
958 | ||
959 | nb_iters = scalar_evolution_in_region (region, loop, nb_iters); | |
33ad93b9 | 960 | scan_tree_for_params (region, nb_iters); |
2abae5f1 SP |
961 | } |
962 | ||
2abae5f1 | 963 | /* Find the parameters used in data accesses. */ |
9771b263 | 964 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
965 | find_params_in_bb (region, PBB_BLACK_BOX (pbb)); |
966 | ||
33ad93b9 RG |
967 | nbp = sese_nb_params (region); |
968 | scop_set_nb_params (scop, nbp); | |
2abae5f1 | 969 | SESE_ADD_PARAMS (region) = false; |
62e475c5 | 970 | |
a5a59b11 | 971 | { |
33ad93b9 RG |
972 | tree e; |
973 | isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0); | |
a5a59b11 | 974 | |
9771b263 | 975 | FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e) |
33ad93b9 RG |
976 | space = isl_space_set_dim_id (space, isl_dim_param, i, |
977 | isl_id_for_ssa_name (scop, e)); | |
a5a59b11 | 978 | |
33ad93b9 | 979 | scop->context = isl_set_universe (space); |
a5a59b11 | 980 | } |
a5a59b11 SP |
981 | } |
982 | ||
2abae5f1 SP |
983 | /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives |
984 | the constraints for the surrounding loops. */ | |
985 | ||
986 | static void | |
987 | build_loop_iteration_domains (scop_p scop, struct loop *loop, | |
33ad93b9 RG |
988 | int nb, |
989 | isl_set *outer, isl_set **doms) | |
2abae5f1 | 990 | { |
2abae5f1 | 991 | tree nb_iters = number_of_latch_executions (loop); |
2abae5f1 SP |
992 | sese region = SCOP_REGION (scop); |
993 | ||
33ad93b9 RG |
994 | isl_set *inner = isl_set_copy (outer); |
995 | isl_space *space; | |
996 | isl_constraint *c; | |
997 | int pos = isl_set_dim (outer, isl_dim_set); | |
998 | isl_int v; | |
999 | mpz_t g; | |
1000 | ||
1001 | mpz_init (g); | |
1002 | isl_int_init (v); | |
1003 | ||
1004 | inner = isl_set_add_dims (inner, isl_dim_set, 1); | |
1005 | space = isl_set_get_space (inner); | |
2abae5f1 SP |
1006 | |
1007 | /* 0 <= loop_i */ | |
33ad93b9 RG |
1008 | c = isl_inequality_alloc |
1009 | (isl_local_space_from_space (isl_space_copy (space))); | |
1010 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1); | |
1011 | inner = isl_set_add_constraint (inner, c); | |
2abae5f1 | 1012 | |
33ad93b9 | 1013 | /* loop_i <= cst_nb_iters */ |
2abae5f1 SP |
1014 | if (TREE_CODE (nb_iters) == INTEGER_CST) |
1015 | { | |
33ad93b9 | 1016 | c = isl_inequality_alloc |
c3284718 | 1017 | (isl_local_space_from_space (isl_space_copy (space))); |
33ad93b9 RG |
1018 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1); |
1019 | tree_int_to_gmp (nb_iters, g); | |
1020 | isl_int_set_gmp (v, g); | |
1021 | c = isl_constraint_set_constant (c, v); | |
1022 | inner = isl_set_add_constraint (inner, c); | |
2abae5f1 | 1023 | } |
33ad93b9 RG |
1024 | |
1025 | /* loop_i <= expr_nb_iters */ | |
2abae5f1 SP |
1026 | else if (!chrec_contains_undetermined (nb_iters)) |
1027 | { | |
62e475c5 | 1028 | double_int nit; |
33ad93b9 RG |
1029 | isl_pw_aff *aff; |
1030 | isl_set *valid; | |
1031 | isl_local_space *ls; | |
1032 | isl_aff *al; | |
1033 | isl_set *le; | |
2abae5f1 | 1034 | |
2abae5f1 | 1035 | nb_iters = scalar_evolution_in_region (region, loop, nb_iters); |
33ad93b9 RG |
1036 | |
1037 | aff = extract_affine (scop, nb_iters, isl_set_get_space (inner)); | |
1038 | valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff)); | |
1039 | valid = isl_set_project_out (valid, isl_dim_set, 0, | |
1040 | isl_set_dim (valid, isl_dim_set)); | |
1041 | scop->context = isl_set_intersect (scop->context, valid); | |
1042 | ||
1043 | ls = isl_local_space_from_space (isl_space_copy (space)); | |
1044 | al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls), | |
1045 | isl_dim_in, pos, 1); | |
1046 | le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al), | |
1047 | isl_pw_aff_copy (aff)); | |
1048 | inner = isl_set_intersect (inner, le); | |
2abae5f1 | 1049 | |
652c4c71 | 1050 | if (max_stmt_executions (loop, &nit)) |
33ad93b9 RG |
1051 | { |
1052 | /* Insert in the context the constraints from the | |
1053 | estimation of the number of iterations NIT and the | |
1054 | symbolic number of iterations (involving parameter | |
1055 | names) NB_ITERS. First, build the affine expression | |
1056 | "NIT - NB_ITERS" and then say that it is positive, | |
1057 | i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */ | |
1058 | isl_pw_aff *approx; | |
1059 | mpz_t g; | |
1060 | isl_set *x; | |
1061 | isl_constraint *c; | |
1062 | ||
1063 | mpz_init (g); | |
1064 | mpz_set_double_int (g, nit, false); | |
1065 | mpz_sub_ui (g, g, 1); | |
1066 | approx = extract_affine_gmp (g, isl_set_get_space (inner)); | |
1067 | x = isl_pw_aff_ge_set (approx, aff); | |
1068 | x = isl_set_project_out (x, isl_dim_set, 0, | |
1069 | isl_set_dim (x, isl_dim_set)); | |
1070 | scop->context = isl_set_intersect (scop->context, x); | |
1071 | ||
1072 | c = isl_inequality_alloc | |
1073 | (isl_local_space_from_space (isl_space_copy (space))); | |
1074 | c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1); | |
1075 | isl_int_set_gmp (v, g); | |
1076 | mpz_clear (g); | |
1077 | c = isl_constraint_set_constant (c, v); | |
1078 | inner = isl_set_add_constraint (inner, c); | |
1079 | } | |
2b91f098 RB |
1080 | else |
1081 | isl_pw_aff_free (aff); | |
2abae5f1 SP |
1082 | } |
1083 | else | |
1084 | gcc_unreachable (); | |
1085 | ||
1086 | if (loop->inner && loop_in_sese_p (loop->inner, region)) | |
33ad93b9 RG |
1087 | build_loop_iteration_domains (scop, loop->inner, nb + 1, |
1088 | isl_set_copy (inner), doms); | |
2abae5f1 SP |
1089 | |
1090 | if (nb != 0 | |
1091 | && loop->next | |
1092 | && loop_in_sese_p (loop->next, region)) | |
33ad93b9 RG |
1093 | build_loop_iteration_domains (scop, loop->next, nb, |
1094 | isl_set_copy (outer), doms); | |
2abae5f1 | 1095 | |
33ad93b9 | 1096 | doms[loop->num] = inner; |
2abae5f1 | 1097 | |
33ad93b9 RG |
1098 | isl_set_free (outer); |
1099 | isl_space_free (space); | |
1100 | isl_int_clear (v); | |
1101 | mpz_clear (g); | |
2abae5f1 SP |
1102 | } |
1103 | ||
1104 | /* Returns a linear expression for tree T evaluated in PBB. */ | |
1105 | ||
33ad93b9 RG |
1106 | static isl_pw_aff * |
1107 | create_pw_aff_from_tree (poly_bb_p pbb, tree t) | |
2abae5f1 | 1108 | { |
33ad93b9 | 1109 | scop_p scop = PBB_SCOP (pbb); |
2abae5f1 | 1110 | |
33ad93b9 | 1111 | t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t); |
2abae5f1 SP |
1112 | gcc_assert (!automatically_generated_chrec_p (t)); |
1113 | ||
33ad93b9 | 1114 | return extract_affine (scop, t, isl_set_get_space (pbb->domain)); |
2abae5f1 SP |
1115 | } |
1116 | ||
33ad93b9 RG |
1117 | /* Add conditional statement STMT to pbb. CODE is used as the comparison |
1118 | operator. This allows us to invert the condition or to handle | |
1119 | inequalities. */ | |
2abae5f1 SP |
1120 | |
1121 | static void | |
33ad93b9 | 1122 | add_condition_to_pbb (poly_bb_p pbb, gimple stmt, enum tree_code code) |
2abae5f1 | 1123 | { |
33ad93b9 RG |
1124 | isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt)); |
1125 | isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt)); | |
1126 | isl_set *cond; | |
2abae5f1 | 1127 | |
33ad93b9 | 1128 | switch (code) |
2abae5f1 | 1129 | { |
33ad93b9 RG |
1130 | case LT_EXPR: |
1131 | cond = isl_pw_aff_lt_set (lhs, rhs); | |
1132 | break; | |
2abae5f1 | 1133 | |
33ad93b9 RG |
1134 | case GT_EXPR: |
1135 | cond = isl_pw_aff_gt_set (lhs, rhs); | |
1136 | break; | |
2abae5f1 | 1137 | |
33ad93b9 RG |
1138 | case LE_EXPR: |
1139 | cond = isl_pw_aff_le_set (lhs, rhs); | |
1140 | break; | |
2abae5f1 | 1141 | |
33ad93b9 RG |
1142 | case GE_EXPR: |
1143 | cond = isl_pw_aff_ge_set (lhs, rhs); | |
1144 | break; | |
2abae5f1 | 1145 | |
33ad93b9 RG |
1146 | case EQ_EXPR: |
1147 | cond = isl_pw_aff_eq_set (lhs, rhs); | |
1148 | break; | |
2abae5f1 | 1149 | |
33ad93b9 RG |
1150 | case NE_EXPR: |
1151 | cond = isl_pw_aff_ne_set (lhs, rhs); | |
1152 | break; | |
2abae5f1 | 1153 | |
33ad93b9 | 1154 | default: |
c3284718 RS |
1155 | isl_pw_aff_free (lhs); |
1156 | isl_pw_aff_free (rhs); | |
33ad93b9 | 1157 | return; |
2abae5f1 | 1158 | } |
33ad93b9 RG |
1159 | |
1160 | cond = isl_set_coalesce (cond); | |
1161 | cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain)); | |
1162 | pbb->domain = isl_set_intersect (pbb->domain, cond); | |
2abae5f1 SP |
1163 | } |
1164 | ||
1165 | /* Add conditions to the domain of PBB. */ | |
1166 | ||
1167 | static void | |
1168 | add_conditions_to_domain (poly_bb_p pbb) | |
1169 | { | |
1170 | unsigned int i; | |
1171 | gimple stmt; | |
1172 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); | |
2abae5f1 | 1173 | |
9771b263 | 1174 | if (GBB_CONDITIONS (gbb).is_empty ()) |
2abae5f1 SP |
1175 | return; |
1176 | ||
9771b263 | 1177 | FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt) |
2abae5f1 SP |
1178 | switch (gimple_code (stmt)) |
1179 | { | |
1180 | case GIMPLE_COND: | |
1181 | { | |
1182 | enum tree_code code = gimple_cond_code (stmt); | |
1183 | ||
1184 | /* The conditions for ELSE-branches are inverted. */ | |
9771b263 | 1185 | if (!GBB_CONDITION_CASES (gbb)[i]) |
2abae5f1 SP |
1186 | code = invert_tree_comparison (code, false); |
1187 | ||
1188 | add_condition_to_pbb (pbb, stmt, code); | |
1189 | break; | |
1190 | } | |
1191 | ||
1192 | case GIMPLE_SWITCH: | |
33ad93b9 | 1193 | /* Switch statements are not supported right now - fall through. */ |
2abae5f1 SP |
1194 | |
1195 | default: | |
1196 | gcc_unreachable (); | |
1197 | break; | |
1198 | } | |
1199 | } | |
1200 | ||
efa21390 SP |
1201 | /* Traverses all the GBBs of the SCOP and add their constraints to the |
1202 | iteration domains. */ | |
1203 | ||
1204 | static void | |
1205 | add_conditions_to_constraints (scop_p scop) | |
1206 | { | |
1207 | int i; | |
1208 | poly_bb_p pbb; | |
1209 | ||
9771b263 | 1210 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
efa21390 SP |
1211 | add_conditions_to_domain (pbb); |
1212 | } | |
1213 | ||
c12e2a5f SP |
1214 | /* Returns a COND_EXPR statement when BB has a single predecessor, the |
1215 | edge between BB and its predecessor is not a loop exit edge, and | |
1216 | the last statement of the single predecessor is a COND_EXPR. */ | |
2abae5f1 SP |
1217 | |
1218 | static gimple | |
c12e2a5f | 1219 | single_pred_cond_non_loop_exit (basic_block bb) |
2abae5f1 SP |
1220 | { |
1221 | if (single_pred_p (bb)) | |
1222 | { | |
1223 | edge e = single_pred_edge (bb); | |
1224 | basic_block pred = e->src; | |
c12e2a5f SP |
1225 | gimple stmt; |
1226 | ||
1227 | if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father)) | |
1228 | return NULL; | |
1229 | ||
1230 | stmt = last_stmt (pred); | |
2abae5f1 SP |
1231 | |
1232 | if (stmt && gimple_code (stmt) == GIMPLE_COND) | |
1233 | return stmt; | |
1234 | } | |
c12e2a5f | 1235 | |
2abae5f1 SP |
1236 | return NULL; |
1237 | } | |
1238 | ||
4d9192b5 TS |
1239 | class sese_dom_walker : public dom_walker |
1240 | { | |
1241 | public: | |
1242 | sese_dom_walker (cdi_direction, sese); | |
4d9192b5 TS |
1243 | |
1244 | virtual void before_dom_children (basic_block); | |
1245 | virtual void after_dom_children (basic_block); | |
1246 | ||
1247 | private: | |
00f96dc9 | 1248 | auto_vec<gimple, 3> m_conditions, m_cases; |
65d3284b | 1249 | sese m_region; |
4d9192b5 TS |
1250 | }; |
1251 | ||
1252 | sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region) | |
65d3284b | 1253 | : dom_walker (direction), m_region (region) |
4d9192b5 | 1254 | { |
4d9192b5 TS |
1255 | } |
1256 | ||
2abae5f1 SP |
1257 | /* Call-back for dom_walk executed before visiting the dominated |
1258 | blocks. */ | |
1259 | ||
4d9192b5 TS |
1260 | void |
1261 | sese_dom_walker::before_dom_children (basic_block bb) | |
2abae5f1 | 1262 | { |
072edf07 SP |
1263 | gimple_bb_p gbb; |
1264 | gimple stmt; | |
2abae5f1 | 1265 | |
65d3284b | 1266 | if (!bb_in_sese_p (bb, m_region)) |
2abae5f1 SP |
1267 | return; |
1268 | ||
c12e2a5f | 1269 | stmt = single_pred_cond_non_loop_exit (bb); |
072edf07 | 1270 | |
2abae5f1 SP |
1271 | if (stmt) |
1272 | { | |
1273 | edge e = single_pred_edge (bb); | |
1274 | ||
65d3284b | 1275 | m_conditions.safe_push (stmt); |
2abae5f1 SP |
1276 | |
1277 | if (e->flags & EDGE_TRUE_VALUE) | |
65d3284b | 1278 | m_cases.safe_push (stmt); |
2abae5f1 | 1279 | else |
65d3284b | 1280 | m_cases.safe_push (NULL); |
2abae5f1 SP |
1281 | } |
1282 | ||
072edf07 SP |
1283 | gbb = gbb_from_bb (bb); |
1284 | ||
2abae5f1 SP |
1285 | if (gbb) |
1286 | { | |
65d3284b RS |
1287 | GBB_CONDITIONS (gbb) = m_conditions.copy (); |
1288 | GBB_CONDITION_CASES (gbb) = m_cases.copy (); | |
2abae5f1 SP |
1289 | } |
1290 | } | |
1291 | ||
1292 | /* Call-back for dom_walk executed after visiting the dominated | |
1293 | blocks. */ | |
1294 | ||
4d9192b5 TS |
1295 | void |
1296 | sese_dom_walker::after_dom_children (basic_block bb) | |
2abae5f1 | 1297 | { |
65d3284b | 1298 | if (!bb_in_sese_p (bb, m_region)) |
2abae5f1 SP |
1299 | return; |
1300 | ||
c12e2a5f | 1301 | if (single_pred_cond_non_loop_exit (bb)) |
2abae5f1 | 1302 | { |
65d3284b RS |
1303 | m_conditions.pop (); |
1304 | m_cases.pop (); | |
2abae5f1 SP |
1305 | } |
1306 | } | |
1307 | ||
2abae5f1 SP |
1308 | /* Add constraints on the possible values of parameter P from the type |
1309 | of P. */ | |
1310 | ||
1311 | static void | |
33ad93b9 | 1312 | add_param_constraints (scop_p scop, graphite_dim_t p) |
2abae5f1 | 1313 | { |
9771b263 | 1314 | tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p]; |
2abae5f1 | 1315 | tree type = TREE_TYPE (parameter); |
3640d64c SP |
1316 | tree lb = NULL_TREE; |
1317 | tree ub = NULL_TREE; | |
2abae5f1 | 1318 | |
697f511d SP |
1319 | if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type)) |
1320 | lb = lower_bound_in_type (type, type); | |
1321 | else | |
1322 | lb = TYPE_MIN_VALUE (type); | |
1323 | ||
1324 | if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type)) | |
1325 | ub = upper_bound_in_type (type, type); | |
1326 | else | |
1327 | ub = TYPE_MAX_VALUE (type); | |
2abae5f1 SP |
1328 | |
1329 | if (lb) | |
1330 | { | |
33ad93b9 RG |
1331 | isl_space *space = isl_set_get_space (scop->context); |
1332 | isl_constraint *c; | |
1333 | mpz_t g; | |
1334 | isl_int v; | |
1335 | ||
1336 | c = isl_inequality_alloc (isl_local_space_from_space (space)); | |
1337 | mpz_init (g); | |
1338 | isl_int_init (v); | |
1339 | tree_int_to_gmp (lb, g); | |
1340 | isl_int_set_gmp (v, g); | |
1341 | isl_int_neg (v, v); | |
1342 | mpz_clear (g); | |
1343 | c = isl_constraint_set_constant (c, v); | |
1344 | isl_int_clear (v); | |
1345 | c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1); | |
1346 | ||
1347 | scop->context = isl_set_add_constraint (scop->context, c); | |
2abae5f1 SP |
1348 | } |
1349 | ||
1350 | if (ub) | |
1351 | { | |
33ad93b9 RG |
1352 | isl_space *space = isl_set_get_space (scop->context); |
1353 | isl_constraint *c; | |
1354 | mpz_t g; | |
1355 | isl_int v; | |
1356 | ||
1357 | c = isl_inequality_alloc (isl_local_space_from_space (space)); | |
1358 | ||
1359 | mpz_init (g); | |
1360 | isl_int_init (v); | |
1361 | tree_int_to_gmp (ub, g); | |
1362 | isl_int_set_gmp (v, g); | |
1363 | mpz_clear (g); | |
1364 | c = isl_constraint_set_constant (c, v); | |
1365 | isl_int_clear (v); | |
1366 | c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1); | |
1367 | ||
1368 | scop->context = isl_set_add_constraint (scop->context, c); | |
2abae5f1 SP |
1369 | } |
1370 | } | |
1371 | ||
1372 | /* Build the context of the SCOP. The context usually contains extra | |
1373 | constraints that are added to the iteration domains that constrain | |
1374 | some parameters. */ | |
1375 | ||
1376 | static void | |
1377 | build_scop_context (scop_p scop) | |
1378 | { | |
2abae5f1 SP |
1379 | graphite_dim_t p, n = scop_nb_params (scop); |
1380 | ||
2abae5f1 | 1381 | for (p = 0; p < n; p++) |
33ad93b9 | 1382 | add_param_constraints (scop, p); |
2abae5f1 SP |
1383 | } |
1384 | ||
1385 | /* Build the iteration domains: the loops belonging to the current | |
1386 | SCOP, and that vary for the execution of the current basic block. | |
1387 | Returns false if there is no loop in SCOP. */ | |
1388 | ||
1389 | static void | |
1390 | build_scop_iteration_domain (scop_p scop) | |
1391 | { | |
1392 | struct loop *loop; | |
1393 | sese region = SCOP_REGION (scop); | |
1394 | int i; | |
2abae5f1 | 1395 | poly_bb_p pbb; |
0fc822d0 | 1396 | int nb_loops = number_of_loops (cfun); |
33ad93b9 | 1397 | isl_set **doms = XCNEWVEC (isl_set *, nb_loops); |
2abae5f1 | 1398 | |
9771b263 | 1399 | FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop) |
2abae5f1 | 1400 | if (!loop_in_sese_p (loop_outer (loop), region)) |
33ad93b9 RG |
1401 | build_loop_iteration_domains (scop, loop, 0, |
1402 | isl_set_copy (scop->context), doms); | |
2abae5f1 | 1403 | |
9771b263 | 1404 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
33ad93b9 RG |
1405 | { |
1406 | loop = pbb_loop (pbb); | |
1407 | ||
1408 | if (doms[loop->num]) | |
1409 | pbb->domain = isl_set_copy (doms[loop->num]); | |
1410 | else | |
1411 | pbb->domain = isl_set_copy (scop->context); | |
1412 | ||
1413 | pbb->domain = isl_set_set_tuple_id (pbb->domain, | |
1414 | isl_id_for_pbb (scop, pbb)); | |
1415 | } | |
2abae5f1 | 1416 | |
6c6f84d7 | 1417 | for (i = 0; i < nb_loops; i++) |
33ad93b9 RG |
1418 | if (doms[i]) |
1419 | isl_set_free (doms[i]); | |
2abae5f1 | 1420 | |
33ad93b9 | 1421 | free (doms); |
2abae5f1 SP |
1422 | } |
1423 | ||
1424 | /* Add a constrain to the ACCESSES polyhedron for the alias set of | |
1425 | data reference DR. ACCESSP_NB_DIMS is the dimension of the | |
1426 | ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration | |
1427 | domain. */ | |
1428 | ||
33ad93b9 RG |
1429 | static isl_map * |
1430 | pdr_add_alias_set (isl_map *acc, data_reference_p dr) | |
2abae5f1 | 1431 | { |
33ad93b9 | 1432 | isl_constraint *c; |
2abae5f1 | 1433 | int alias_set_num = 0; |
2b178a5f | 1434 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); |
2abae5f1 | 1435 | |
fb00d28e | 1436 | if (bap && bap->alias_set) |
2b178a5f | 1437 | alias_set_num = *(bap->alias_set); |
2abae5f1 | 1438 | |
33ad93b9 RG |
1439 | c = isl_equality_alloc |
1440 | (isl_local_space_from_space (isl_map_get_space (acc))); | |
1441 | c = isl_constraint_set_constant_si (c, -alias_set_num); | |
1442 | c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1); | |
1443 | ||
1444 | return isl_map_add_constraint (acc, c); | |
1445 | } | |
1446 | ||
1447 | /* Assign the affine expression INDEX to the output dimension POS of | |
1448 | MAP and return the result. */ | |
1449 | ||
1450 | static isl_map * | |
1451 | set_index (isl_map *map, int pos, isl_pw_aff *index) | |
1452 | { | |
1453 | isl_map *index_map; | |
1454 | int len = isl_map_dim (map, isl_dim_out); | |
1455 | isl_id *id; | |
1456 | ||
1457 | index_map = isl_map_from_pw_aff (index); | |
1458 | index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos); | |
1459 | index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1); | |
2abae5f1 | 1460 | |
33ad93b9 RG |
1461 | id = isl_map_get_tuple_id (map, isl_dim_out); |
1462 | index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id); | |
1463 | id = isl_map_get_tuple_id (map, isl_dim_in); | |
1464 | index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id); | |
2abae5f1 | 1465 | |
33ad93b9 | 1466 | return isl_map_intersect (map, index_map); |
2abae5f1 SP |
1467 | } |
1468 | ||
1469 | /* Add to ACCESSES polyhedron equalities defining the access functions | |
1470 | to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES | |
1471 | polyhedron, DOM_NB_DIMS is the dimension of the iteration domain. | |
1472 | PBB is the poly_bb_p that contains the data reference DR. */ | |
1473 | ||
33ad93b9 RG |
1474 | static isl_map * |
1475 | pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb) | |
2abae5f1 SP |
1476 | { |
1477 | int i, nb_subscripts = DR_NUM_DIMENSIONS (dr); | |
2abae5f1 | 1478 | scop_p scop = PBB_SCOP (pbb); |
2abae5f1 SP |
1479 | |
1480 | for (i = 0; i < nb_subscripts; i++) | |
1481 | { | |
33ad93b9 | 1482 | isl_pw_aff *aff; |
2abae5f1 SP |
1483 | tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i); |
1484 | ||
33ad93b9 RG |
1485 | aff = extract_affine (scop, afn, |
1486 | isl_space_domain (isl_map_get_space (acc))); | |
1487 | acc = set_index (acc, i + 1, aff); | |
2abae5f1 SP |
1488 | } |
1489 | ||
33ad93b9 | 1490 | return acc; |
2abae5f1 SP |
1491 | } |
1492 | ||
1493 | /* Add constrains representing the size of the accessed data to the | |
66096911 SP |
1494 | ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the |
1495 | ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration | |
2abae5f1 SP |
1496 | domain. */ |
1497 | ||
33ad93b9 RG |
1498 | static isl_set * |
1499 | pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr) | |
2abae5f1 SP |
1500 | { |
1501 | tree ref = DR_REF (dr); | |
1502 | int i, nb_subscripts = DR_NUM_DIMENSIONS (dr); | |
2abae5f1 | 1503 | |
98f3eb1f | 1504 | for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0)) |
2abae5f1 | 1505 | { |
98f3eb1f | 1506 | tree low, high; |
2abae5f1 | 1507 | |
98f3eb1f | 1508 | if (TREE_CODE (ref) != ARRAY_REF) |
2abae5f1 SP |
1509 | break; |
1510 | ||
98f3eb1f | 1511 | low = array_ref_low_bound (ref); |
98f3eb1f AM |
1512 | high = array_ref_up_bound (ref); |
1513 | ||
33ad93b9 RG |
1514 | /* XXX The PPL code dealt separately with |
1515 | subscript - low >= 0 and high - subscript >= 0 in case one of | |
1516 | the two bounds isn't known. Do the same here? */ | |
1517 | ||
9541ffee | 1518 | if (tree_fits_shwi_p (low) |
33ad93b9 | 1519 | && high |
9541ffee | 1520 | && tree_fits_shwi_p (high) |
3899a0b2 SP |
1521 | /* 1-element arrays at end of structures may extend over |
1522 | their declared size. */ | |
1523 | && !(array_at_struct_end_p (ref) | |
1524 | && operand_equal_p (low, high, 0))) | |
98f3eb1f | 1525 | { |
33ad93b9 RG |
1526 | isl_id *id; |
1527 | isl_aff *aff; | |
1528 | isl_set *univ, *lbs, *ubs; | |
1529 | isl_pw_aff *index; | |
1530 | isl_space *space; | |
1531 | isl_set *valid; | |
1532 | isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent)); | |
1533 | isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent)); | |
1534 | ||
1535 | /* high >= 0 */ | |
1536 | valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub)); | |
1537 | valid = isl_set_project_out (valid, isl_dim_set, 0, | |
1538 | isl_set_dim (valid, isl_dim_set)); | |
1539 | scop->context = isl_set_intersect (scop->context, valid); | |
1540 | ||
1541 | space = isl_set_get_space (extent); | |
1542 | aff = isl_aff_zero_on_domain (isl_local_space_from_space (space)); | |
1543 | aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1); | |
1544 | univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff))); | |
1545 | index = isl_pw_aff_alloc (univ, aff); | |
1546 | ||
1547 | id = isl_set_get_tuple_id (extent); | |
1548 | lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id)); | |
1549 | ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id); | |
1550 | ||
1551 | /* low <= sub_i <= high */ | |
1552 | lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb); | |
1553 | ubs = isl_pw_aff_le_set (index, ub); | |
1554 | extent = isl_set_intersect (extent, lbs); | |
1555 | extent = isl_set_intersect (extent, ubs); | |
98f3eb1f | 1556 | } |
2abae5f1 | 1557 | } |
33ad93b9 RG |
1558 | |
1559 | return extent; | |
2abae5f1 SP |
1560 | } |
1561 | ||
1562 | /* Build data accesses for DR in PBB. */ | |
1563 | ||
1564 | static void | |
1565 | build_poly_dr (data_reference_p dr, poly_bb_p pbb) | |
1566 | { | |
1825f9a2 | 1567 | int dr_base_object_set; |
33ad93b9 RG |
1568 | isl_map *acc; |
1569 | isl_set *extent; | |
1570 | scop_p scop = PBB_SCOP (pbb); | |
2abae5f1 | 1571 | |
33ad93b9 RG |
1572 | { |
1573 | isl_space *dc = isl_set_get_space (pbb->domain); | |
1574 | int nb_out = 1 + DR_NUM_DIMENSIONS (dr); | |
1575 | isl_space *space = isl_space_add_dims (isl_space_from_domain (dc), | |
1576 | isl_dim_out, nb_out); | |
2abae5f1 | 1577 | |
33ad93b9 RG |
1578 | acc = isl_map_universe (space); |
1579 | acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr)); | |
1580 | } | |
2abae5f1 | 1581 | |
33ad93b9 RG |
1582 | acc = pdr_add_alias_set (acc, dr); |
1583 | acc = pdr_add_memory_accesses (acc, dr, pbb); | |
2abae5f1 | 1584 | |
33ad93b9 RG |
1585 | { |
1586 | isl_id *id = isl_id_for_dr (scop, dr); | |
1587 | int nb = 1 + DR_NUM_DIMENSIONS (dr); | |
1588 | isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb); | |
1589 | int alias_set_num = 0; | |
1590 | base_alias_pair *bap = (base_alias_pair *)(dr->aux); | |
1591 | ||
1592 | if (bap && bap->alias_set) | |
1593 | alias_set_num = *(bap->alias_set); | |
1594 | ||
1595 | space = isl_space_set_tuple_id (space, isl_dim_set, id); | |
1596 | extent = isl_set_nat_universe (space); | |
1597 | extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num); | |
1598 | extent = pdr_add_data_dimensions (extent, scop, dr); | |
1599 | } | |
2abae5f1 | 1600 | |
6e44d26e SP |
1601 | gcc_assert (dr->aux); |
1602 | dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set; | |
1825f9a2 | 1603 | |
33ad93b9 | 1604 | new_poly_dr (pbb, dr_base_object_set, |
6e44d26e | 1605 | DR_IS_READ (dr) ? PDR_READ : PDR_WRITE, |
33ad93b9 | 1606 | dr, DR_NUM_DIMENSIONS (dr), acc, extent); |
1825f9a2 | 1607 | } |
2abae5f1 | 1608 | |
2e5a7cbf | 1609 | /* Write to FILE the alias graph of data references in DIMACS format. */ |
cd43e5d7 LF |
1610 | |
1611 | static inline bool | |
1612 | write_alias_graph_to_ascii_dimacs (FILE *file, char *comment, | |
9771b263 | 1613 | vec<data_reference_p> drs) |
cd43e5d7 | 1614 | { |
9771b263 | 1615 | int num_vertex = drs.length (); |
cd43e5d7 LF |
1616 | int edge_num = 0; |
1617 | data_reference_p dr1, dr2; | |
1618 | int i, j; | |
1619 | ||
1620 | if (num_vertex == 0) | |
1621 | return true; | |
1622 | ||
9771b263 DN |
1623 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1624 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1625 | if (dr_may_alias_p (dr1, dr2, true)) |
cd43e5d7 LF |
1626 | edge_num++; |
1627 | ||
1628 | fprintf (file, "$\n"); | |
1629 | ||
1630 | if (comment) | |
1631 | fprintf (file, "c %s\n", comment); | |
1632 | ||
1633 | fprintf (file, "p edge %d %d\n", num_vertex, edge_num); | |
1634 | ||
9771b263 DN |
1635 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1636 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1637 | if (dr_may_alias_p (dr1, dr2, true)) |
cd43e5d7 LF |
1638 | fprintf (file, "e %d %d\n", i + 1, j + 1); |
1639 | ||
1640 | return true; | |
1641 | } | |
1642 | ||
2e5a7cbf RU |
1643 | /* Write to FILE the alias graph of data references in DOT format. */ |
1644 | ||
1645 | static inline bool | |
1646 | write_alias_graph_to_ascii_dot (FILE *file, char *comment, | |
9771b263 | 1647 | vec<data_reference_p> drs) |
2e5a7cbf | 1648 | { |
9771b263 | 1649 | int num_vertex = drs.length (); |
2e5a7cbf RU |
1650 | data_reference_p dr1, dr2; |
1651 | int i, j; | |
1652 | ||
1653 | if (num_vertex == 0) | |
1654 | return true; | |
1655 | ||
1656 | fprintf (file, "$\n"); | |
1657 | ||
1658 | if (comment) | |
1659 | fprintf (file, "c %s\n", comment); | |
1660 | ||
1661 | /* First print all the vertices. */ | |
9771b263 | 1662 | FOR_EACH_VEC_ELT (drs, i, dr1) |
2e5a7cbf RU |
1663 | fprintf (file, "n%d;\n", i); |
1664 | ||
9771b263 DN |
1665 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1666 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1667 | if (dr_may_alias_p (dr1, dr2, true)) |
2e5a7cbf RU |
1668 | fprintf (file, "n%d n%d\n", i, j); |
1669 | ||
1670 | return true; | |
1671 | } | |
1672 | ||
1673 | /* Write to FILE the alias graph of data references in ECC format. */ | |
1674 | ||
1675 | static inline bool | |
1676 | write_alias_graph_to_ascii_ecc (FILE *file, char *comment, | |
9771b263 | 1677 | vec<data_reference_p> drs) |
2e5a7cbf | 1678 | { |
9771b263 | 1679 | int num_vertex = drs.length (); |
2e5a7cbf RU |
1680 | data_reference_p dr1, dr2; |
1681 | int i, j; | |
1682 | ||
1683 | if (num_vertex == 0) | |
1684 | return true; | |
1685 | ||
1686 | fprintf (file, "$\n"); | |
1687 | ||
1688 | if (comment) | |
1689 | fprintf (file, "c %s\n", comment); | |
1690 | ||
9771b263 DN |
1691 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1692 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1693 | if (dr_may_alias_p (dr1, dr2, true)) |
2e5a7cbf RU |
1694 | fprintf (file, "%d %d\n", i, j); |
1695 | ||
1696 | return true; | |
1697 | } | |
1698 | ||
2b178a5f LF |
1699 | /* Check if DR1 and DR2 are in the same object set. */ |
1700 | ||
1701 | static bool | |
1702 | dr_same_base_object_p (const struct data_reference *dr1, | |
1703 | const struct data_reference *dr2) | |
1704 | { | |
1705 | return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0); | |
1706 | } | |
2e5a7cbf RU |
1707 | |
1708 | /* Uses DFS component number as representative of alias-sets. Also tests for | |
1709 | optimality by verifying if every connected component is a clique. Returns | |
1710 | true (1) if the above test is true, and false (0) otherwise. */ | |
1711 | ||
1712 | static int | |
9771b263 | 1713 | build_alias_set_optimal_p (vec<data_reference_p> drs) |
2abae5f1 | 1714 | { |
9771b263 | 1715 | int num_vertices = drs.length (); |
2e5a7cbf | 1716 | struct graph *g = new_graph (num_vertices); |
2abae5f1 SP |
1717 | data_reference_p dr1, dr2; |
1718 | int i, j; | |
2e5a7cbf RU |
1719 | int num_connected_components; |
1720 | int v_indx1, v_indx2, num_vertices_in_component; | |
1721 | int *all_vertices; | |
1722 | int *vertices; | |
1723 | struct graph_edge *e; | |
917f481a SP |
1724 | int this_component_is_clique; |
1725 | int all_components_are_cliques = 1; | |
2abae5f1 | 1726 | |
9771b263 DN |
1727 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1728 | for (j = i+1; drs.iterate (j, &dr2); j++) | |
02f5d6c5 | 1729 | if (dr_may_alias_p (dr1, dr2, true)) |
2abae5f1 SP |
1730 | { |
1731 | add_edge (g, i, j); | |
1732 | add_edge (g, j, i); | |
1733 | } | |
1734 | ||
2e5a7cbf RU |
1735 | all_vertices = XNEWVEC (int, num_vertices); |
1736 | vertices = XNEWVEC (int, num_vertices); | |
1737 | for (i = 0; i < num_vertices; i++) | |
1738 | all_vertices[i] = i; | |
1739 | ||
2b178a5f LF |
1740 | num_connected_components = graphds_dfs (g, all_vertices, num_vertices, |
1741 | NULL, true, NULL); | |
1742 | for (i = 0; i < g->n_vertices; i++) | |
1743 | { | |
9771b263 | 1744 | data_reference_p dr = drs[i]; |
2b178a5f | 1745 | base_alias_pair *bap; |
fb00d28e | 1746 | |
6e44d26e SP |
1747 | gcc_assert (dr->aux); |
1748 | bap = (base_alias_pair *)(dr->aux); | |
fb00d28e | 1749 | |
2b178a5f LF |
1750 | bap->alias_set = XNEW (int); |
1751 | *(bap->alias_set) = g->vertices[i].component + 1; | |
1752 | } | |
1753 | ||
2e5a7cbf RU |
1754 | /* Verify if the DFS numbering results in optimal solution. */ |
1755 | for (i = 0; i < num_connected_components; i++) | |
1756 | { | |
1757 | num_vertices_in_component = 0; | |
1758 | /* Get all vertices whose DFS component number is the same as i. */ | |
1759 | for (j = 0; j < num_vertices; j++) | |
1760 | if (g->vertices[j].component == i) | |
1761 | vertices[num_vertices_in_component++] = j; | |
1762 | ||
1763 | /* Now test if the vertices in 'vertices' form a clique, by testing | |
1764 | for edges among each pair. */ | |
1765 | this_component_is_clique = 1; | |
1766 | for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++) | |
1767 | { | |
1768 | for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++) | |
1769 | { | |
1770 | /* Check if the two vertices are connected by iterating | |
1771 | through all the edges which have one of these are source. */ | |
1772 | e = g->vertices[vertices[v_indx2]].pred; | |
1773 | while (e) | |
1774 | { | |
1775 | if (e->src == vertices[v_indx1]) | |
1776 | break; | |
1777 | e = e->pred_next; | |
1778 | } | |
1779 | if (!e) | |
1780 | { | |
1781 | this_component_is_clique = 0; | |
1782 | break; | |
1783 | } | |
1784 | } | |
1785 | if (!this_component_is_clique) | |
1786 | all_components_are_cliques = 0; | |
1787 | } | |
1788 | } | |
2abae5f1 | 1789 | |
2e5a7cbf RU |
1790 | free (all_vertices); |
1791 | free (vertices); | |
2abae5f1 | 1792 | free_graph (g); |
2e5a7cbf | 1793 | return all_components_are_cliques; |
2abae5f1 SP |
1794 | } |
1795 | ||
efa21390 | 1796 | /* Group each data reference in DRS with its base object set num. */ |
1825f9a2 LF |
1797 | |
1798 | static void | |
9771b263 | 1799 | build_base_obj_set_for_drs (vec<data_reference_p> drs) |
1825f9a2 | 1800 | { |
9771b263 | 1801 | int num_vertex = drs.length (); |
2b178a5f LF |
1802 | struct graph *g = new_graph (num_vertex); |
1803 | data_reference_p dr1, dr2; | |
1804 | int i, j; | |
2b178a5f LF |
1805 | int *queue; |
1806 | ||
9771b263 DN |
1807 | FOR_EACH_VEC_ELT (drs, i, dr1) |
1808 | for (j = i + 1; drs.iterate (j, &dr2); j++) | |
2b178a5f LF |
1809 | if (dr_same_base_object_p (dr1, dr2)) |
1810 | { | |
1811 | add_edge (g, i, j); | |
1812 | add_edge (g, j, i); | |
1813 | } | |
1814 | ||
1815 | queue = XNEWVEC (int, num_vertex); | |
1816 | for (i = 0; i < num_vertex; i++) | |
1817 | queue[i] = i; | |
1818 | ||
fb00d28e | 1819 | graphds_dfs (g, queue, num_vertex, NULL, true, NULL); |
2b178a5f LF |
1820 | |
1821 | for (i = 0; i < g->n_vertices; i++) | |
1822 | { | |
9771b263 | 1823 | data_reference_p dr = drs[i]; |
2b178a5f | 1824 | base_alias_pair *bap; |
fb00d28e | 1825 | |
6e44d26e SP |
1826 | gcc_assert (dr->aux); |
1827 | bap = (base_alias_pair *)(dr->aux); | |
fb00d28e | 1828 | |
2b178a5f LF |
1829 | bap->base_obj_set = g->vertices[i].component + 1; |
1830 | } | |
1831 | ||
1832 | free (queue); | |
1833 | free_graph (g); | |
1825f9a2 LF |
1834 | } |
1835 | ||
2abae5f1 SP |
1836 | /* Build the data references for PBB. */ |
1837 | ||
1838 | static void | |
1839 | build_pbb_drs (poly_bb_p pbb) | |
1840 | { | |
1841 | int j; | |
1842 | data_reference_p dr; | |
9771b263 | 1843 | vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb)); |
2abae5f1 | 1844 | |
9771b263 | 1845 | FOR_EACH_VEC_ELT (gbb_drs, j, dr) |
2abae5f1 SP |
1846 | build_poly_dr (dr, pbb); |
1847 | } | |
1848 | ||
0d5ef2a9 SP |
1849 | /* Dump to file the alias graphs for the data references in DRS. */ |
1850 | ||
1851 | static void | |
9771b263 | 1852 | dump_alias_graphs (vec<data_reference_p> drs) |
0d5ef2a9 SP |
1853 | { |
1854 | char comment[100]; | |
1855 | FILE *file_dimacs, *file_ecc, *file_dot; | |
1856 | ||
1857 | file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab"); | |
1858 | if (file_dimacs) | |
1859 | { | |
1860 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1861 | current_function_name ()); | |
1862 | write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs); | |
1863 | fclose (file_dimacs); | |
1864 | } | |
1865 | ||
1866 | file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab"); | |
1867 | if (file_ecc) | |
1868 | { | |
1869 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1870 | current_function_name ()); | |
1871 | write_alias_graph_to_ascii_ecc (file_ecc, comment, drs); | |
1872 | fclose (file_ecc); | |
1873 | } | |
1874 | ||
1875 | file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab"); | |
1876 | if (file_dot) | |
1877 | { | |
1878 | snprintf (comment, sizeof (comment), "%s %s", main_input_filename, | |
1879 | current_function_name ()); | |
1880 | write_alias_graph_to_ascii_dot (file_dot, comment, drs); | |
1881 | fclose (file_dot); | |
1882 | } | |
1883 | } | |
1884 | ||
2abae5f1 SP |
1885 | /* Build data references in SCOP. */ |
1886 | ||
1887 | static void | |
1888 | build_scop_drs (scop_p scop) | |
1889 | { | |
64393e40 | 1890 | int i, j; |
2abae5f1 | 1891 | poly_bb_p pbb; |
64393e40 | 1892 | data_reference_p dr; |
00f96dc9 | 1893 | auto_vec<data_reference_p, 3> drs; |
64393e40 | 1894 | |
efa21390 SP |
1895 | /* Remove all the PBBs that do not have data references: these basic |
1896 | blocks are not handled in the polyhedral representation. */ | |
9771b263 DN |
1897 | for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++) |
1898 | if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ()) | |
278b1a1d | 1899 | { |
7470b8fc | 1900 | free_gimple_bb (PBB_BLACK_BOX (pbb)); |
62e0a1ed | 1901 | free_poly_bb (pbb); |
9771b263 | 1902 | SCOP_BBS (scop).ordered_remove (i); |
278b1a1d SP |
1903 | i--; |
1904 | } | |
efa21390 | 1905 | |
9771b263 DN |
1906 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
1907 | for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++) | |
1908 | drs.safe_push (dr); | |
64393e40 | 1909 | |
9771b263 | 1910 | FOR_EACH_VEC_ELT (drs, i, dr) |
2b178a5f LF |
1911 | dr->aux = XNEW (base_alias_pair); |
1912 | ||
1913 | if (!build_alias_set_optimal_p (drs)) | |
1914 | { | |
1915 | /* TODO: Add support when building alias set is not optimal. */ | |
1916 | ; | |
1917 | } | |
1918 | ||
ee03cd20 | 1919 | build_base_obj_set_for_drs (drs); |
1825f9a2 | 1920 | |
cd43e5d7 LF |
1921 | /* When debugging, enable the following code. This cannot be used |
1922 | in production compilers. */ | |
0d5ef2a9 SP |
1923 | if (0) |
1924 | dump_alias_graphs (drs); | |
cd43e5d7 | 1925 | |
9771b263 | 1926 | drs.release (); |
2abae5f1 | 1927 | |
9771b263 | 1928 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
1929 | build_pbb_drs (pbb); |
1930 | } | |
1931 | ||
a0dd1440 SP |
1932 | /* Return a gsi at the position of the phi node STMT. */ |
1933 | ||
1934 | static gimple_stmt_iterator | |
1935 | gsi_for_phi_node (gimple stmt) | |
1936 | { | |
1937 | gimple_stmt_iterator psi; | |
1938 | basic_block bb = gimple_bb (stmt); | |
1939 | ||
1940 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1941 | if (stmt == gsi_stmt (psi)) | |
1942 | return psi; | |
1943 | ||
1944 | gcc_unreachable (); | |
1945 | return psi; | |
1946 | } | |
1947 | ||
278b1a1d SP |
1948 | /* Analyze all the data references of STMTS and add them to the |
1949 | GBB_DATA_REFS vector of BB. */ | |
1950 | ||
1951 | static void | |
9771b263 | 1952 | analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts) |
278b1a1d SP |
1953 | { |
1954 | loop_p nest; | |
278b1a1d SP |
1955 | gimple_bb_p gbb; |
1956 | gimple stmt; | |
1957 | int i; | |
5c640e29 | 1958 | sese region = SCOP_REGION (scop); |
278b1a1d | 1959 | |
5c640e29 | 1960 | if (!bb_in_sese_p (bb, region)) |
278b1a1d SP |
1961 | return; |
1962 | ||
5c640e29 | 1963 | nest = outermost_loop_in_sese_1 (region, bb); |
278b1a1d SP |
1964 | gbb = gbb_from_bb (bb); |
1965 | ||
9771b263 | 1966 | FOR_EACH_VEC_ELT (stmts, i, stmt) |
5c640e29 SP |
1967 | { |
1968 | loop_p loop; | |
1969 | ||
1970 | if (is_gimple_debug (stmt)) | |
1971 | continue; | |
1972 | ||
1973 | loop = loop_containing_stmt (stmt); | |
1974 | if (!loop_in_sese_p (loop, region)) | |
1975 | loop = nest; | |
1976 | ||
1977 | graphite_find_data_references_in_stmt (nest, loop, stmt, | |
278b1a1d | 1978 | &GBB_DATA_REFS (gbb)); |
5c640e29 | 1979 | } |
278b1a1d SP |
1980 | } |
1981 | ||
1982 | /* Insert STMT at the end of the STMTS sequence and then insert the | |
1983 | statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts | |
1984 | on STMTS. */ | |
1985 | ||
1986 | static void | |
1987 | insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts, | |
1988 | gimple_stmt_iterator insert_gsi) | |
1989 | { | |
1990 | gimple_stmt_iterator gsi; | |
00f96dc9 | 1991 | auto_vec<gimple, 3> x; |
278b1a1d | 1992 | |
355a7673 | 1993 | gimple_seq_add_stmt (&stmts, stmt); |
278b1a1d | 1994 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
9771b263 | 1995 | x.safe_push (gsi_stmt (gsi)); |
278b1a1d SP |
1996 | |
1997 | gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT); | |
1998 | analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x); | |
278b1a1d SP |
1999 | } |
2000 | ||
efa21390 | 2001 | /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */ |
2abae5f1 SP |
2002 | |
2003 | static void | |
278b1a1d | 2004 | insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt) |
2abae5f1 | 2005 | { |
2abae5f1 | 2006 | gimple_seq stmts; |
947121b8 | 2007 | gimple_stmt_iterator gsi; |
efa21390 | 2008 | tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE); |
2724573f | 2009 | gimple stmt = gimple_build_assign (unshare_expr (res), var); |
00f96dc9 | 2010 | auto_vec<gimple, 3> x; |
2abae5f1 | 2011 | |
355a7673 | 2012 | gimple_seq_add_stmt (&stmts, stmt); |
278b1a1d | 2013 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
9771b263 | 2014 | x.safe_push (gsi_stmt (gsi)); |
947121b8 | 2015 | |
5fed5769 | 2016 | if (gimple_code (after_stmt) == GIMPLE_PHI) |
947121b8 | 2017 | { |
5fed5769 | 2018 | gsi = gsi_after_labels (gimple_bb (after_stmt)); |
947121b8 SP |
2019 | gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT); |
2020 | } | |
2021 | else | |
2022 | { | |
5fed5769 | 2023 | gsi = gsi_for_stmt (after_stmt); |
947121b8 SP |
2024 | gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); |
2025 | } | |
278b1a1d SP |
2026 | |
2027 | analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x); | |
2abae5f1 SP |
2028 | } |
2029 | ||
efa21390 SP |
2030 | /* Creates a poly_bb_p for basic_block BB from the existing PBB. */ |
2031 | ||
2032 | static void | |
2033 | new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb) | |
2034 | { | |
9771b263 DN |
2035 | vec<data_reference_p> drs; |
2036 | drs.create (3); | |
efa21390 SP |
2037 | gimple_bb_p gbb = PBB_BLACK_BOX (pbb); |
2038 | gimple_bb_p gbb1 = new_gimple_bb (bb, drs); | |
2039 | poly_bb_p pbb1 = new_poly_bb (scop, gbb1); | |
9771b263 | 2040 | int index, n = SCOP_BBS (scop).length (); |
efa21390 SP |
2041 | |
2042 | /* The INDEX of PBB in SCOP_BBS. */ | |
2043 | for (index = 0; index < n; index++) | |
9771b263 | 2044 | if (SCOP_BBS (scop)[index] == pbb) |
efa21390 SP |
2045 | break; |
2046 | ||
33ad93b9 | 2047 | pbb1->domain = isl_set_copy (pbb->domain); |
38013f25 | 2048 | |
efa21390 | 2049 | GBB_PBB (gbb1) = pbb1; |
9771b263 DN |
2050 | GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy (); |
2051 | GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy (); | |
2052 | SCOP_BBS (scop).safe_insert (index + 1, pbb1); | |
efa21390 SP |
2053 | } |
2054 | ||
2abae5f1 SP |
2055 | /* Insert on edge E the assignment "RES := EXPR". */ |
2056 | ||
2057 | static void | |
efa21390 | 2058 | insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr) |
2abae5f1 SP |
2059 | { |
2060 | gimple_stmt_iterator gsi; | |
355a7673 | 2061 | gimple_seq stmts = NULL; |
2abae5f1 | 2062 | tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE); |
2724573f | 2063 | gimple stmt = gimple_build_assign (unshare_expr (res), var); |
efa21390 | 2064 | basic_block bb; |
00f96dc9 | 2065 | auto_vec<gimple, 3> x; |
2abae5f1 | 2066 | |
355a7673 | 2067 | gimple_seq_add_stmt (&stmts, stmt); |
278b1a1d | 2068 | for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi)) |
9771b263 | 2069 | x.safe_push (gsi_stmt (gsi)); |
278b1a1d | 2070 | |
2abae5f1 SP |
2071 | gsi_insert_seq_on_edge (e, stmts); |
2072 | gsi_commit_edge_inserts (); | |
efa21390 SP |
2073 | bb = gimple_bb (stmt); |
2074 | ||
2075 | if (!bb_in_sese_p (bb, SCOP_REGION (scop))) | |
2076 | return; | |
2077 | ||
2078 | if (!gbb_from_bb (bb)) | |
2079 | new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb); | |
278b1a1d SP |
2080 | |
2081 | analyze_drs_in_stmts (scop, bb, x); | |
2abae5f1 SP |
2082 | } |
2083 | ||
2084 | /* Creates a zero dimension array of the same type as VAR. */ | |
2085 | ||
2086 | static tree | |
63858ac6 | 2087 | create_zero_dim_array (tree var, const char *base_name) |
2abae5f1 SP |
2088 | { |
2089 | tree index_type = build_index_type (integer_zero_node); | |
2090 | tree elt_type = TREE_TYPE (var); | |
2091 | tree array_type = build_array_type (elt_type, index_type); | |
63858ac6 | 2092 | tree base = create_tmp_var (array_type, base_name); |
2abae5f1 | 2093 | |
2abae5f1 SP |
2094 | return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE, |
2095 | NULL_TREE); | |
2096 | } | |
2097 | ||
2098 | /* Returns true when PHI is a loop close phi node. */ | |
2099 | ||
2100 | static bool | |
2101 | scalar_close_phi_node_p (gimple phi) | |
2102 | { | |
a0dd1440 | 2103 | if (gimple_code (phi) != GIMPLE_PHI |
ea057359 | 2104 | || virtual_operand_p (gimple_phi_result (phi))) |
2abae5f1 SP |
2105 | return false; |
2106 | ||
79d03cf8 SP |
2107 | /* Note that loop close phi nodes should have a single argument |
2108 | because we translated the representation into a canonical form | |
2109 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
2abae5f1 SP |
2110 | return (gimple_phi_num_args (phi) == 1); |
2111 | } | |
2112 | ||
1c2a7491 SP |
2113 | /* For a definition DEF in REGION, propagates the expression EXPR in |
2114 | all the uses of DEF outside REGION. */ | |
2115 | ||
2116 | static void | |
2117 | propagate_expr_outside_region (tree def, tree expr, sese region) | |
2118 | { | |
2119 | imm_use_iterator imm_iter; | |
2120 | gimple use_stmt; | |
2121 | gimple_seq stmts; | |
2122 | bool replaced_once = false; | |
2123 | ||
ab756588 | 2124 | gcc_assert (TREE_CODE (def) == SSA_NAME); |
1c2a7491 SP |
2125 | |
2126 | expr = force_gimple_operand (unshare_expr (expr), &stmts, true, | |
2127 | NULL_TREE); | |
2128 | ||
2129 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2130 | if (!is_gimple_debug (use_stmt) | |
2131 | && !bb_in_sese_p (gimple_bb (use_stmt), region)) | |
2132 | { | |
2133 | ssa_op_iter iter; | |
2134 | use_operand_p use_p; | |
2135 | ||
2136 | FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES) | |
2137 | if (operand_equal_p (def, USE_FROM_PTR (use_p), 0) | |
2138 | && (replaced_once = true)) | |
2139 | replace_exp (use_p, expr); | |
2140 | ||
2141 | update_stmt (use_stmt); | |
2142 | } | |
2143 | ||
2144 | if (replaced_once) | |
2145 | { | |
2146 | gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts); | |
2147 | gsi_commit_edge_inserts (); | |
2148 | } | |
2149 | } | |
2150 | ||
2abae5f1 SP |
2151 | /* Rewrite out of SSA the reduction phi node at PSI by creating a zero |
2152 | dimension array for it. */ | |
2153 | ||
2154 | static void | |
efa21390 | 2155 | rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi) |
2abae5f1 | 2156 | { |
efa21390 | 2157 | sese region = SCOP_REGION (scop); |
2abae5f1 SP |
2158 | gimple phi = gsi_stmt (*psi); |
2159 | tree res = gimple_phi_result (phi); | |
8af6d9cd SP |
2160 | basic_block bb = gimple_bb (phi); |
2161 | gimple_stmt_iterator gsi = gsi_after_labels (bb); | |
2abae5f1 | 2162 | tree arg = gimple_phi_arg_def (phi, 0); |
8af6d9cd | 2163 | gimple stmt; |
2abae5f1 | 2164 | |
79d03cf8 SP |
2165 | /* Note that loop close phi nodes should have a single argument |
2166 | because we translated the representation into a canonical form | |
2167 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
2168 | gcc_assert (gimple_phi_num_args (phi) == 1); | |
2169 | ||
8af6d9cd | 2170 | /* The phi node can be a non close phi node, when its argument is |
974335d6 | 2171 | invariant, or a default definition. */ |
8af6d9cd | 2172 | if (is_gimple_min_invariant (arg) |
974335d6 | 2173 | || SSA_NAME_IS_DEFAULT_DEF (arg)) |
ab756588 SP |
2174 | { |
2175 | propagate_expr_outside_region (res, arg, region); | |
2176 | gsi_next (psi); | |
2177 | return; | |
2178 | } | |
1c2a7491 | 2179 | |
9707eeb0 SP |
2180 | else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father) |
2181 | { | |
2182 | propagate_expr_outside_region (res, arg, region); | |
2183 | stmt = gimple_build_assign (res, arg); | |
2184 | remove_phi_node (psi, false); | |
2185 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
9707eeb0 SP |
2186 | return; |
2187 | } | |
2188 | ||
1c2a7491 SP |
2189 | /* If res is scev analyzable and is not a scalar value, it is safe |
2190 | to ignore the close phi node: it will be code generated in the | |
2191 | out of Graphite pass. */ | |
2192 | else if (scev_analyzable_p (res, region)) | |
2193 | { | |
2194 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res)); | |
2195 | tree scev; | |
2196 | ||
2197 | if (!loop_in_sese_p (loop, region)) | |
2198 | { | |
2199 | loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg)); | |
2200 | scev = scalar_evolution_in_region (region, loop, arg); | |
2201 | scev = compute_overall_effect_of_inner_loop (loop, scev); | |
2202 | } | |
2203 | else | |
ab756588 | 2204 | scev = scalar_evolution_in_region (region, loop, res); |
1c2a7491 SP |
2205 | |
2206 | if (tree_does_not_contain_chrecs (scev)) | |
2207 | propagate_expr_outside_region (res, scev, region); | |
2208 | ||
2209 | gsi_next (psi); | |
2210 | return; | |
2211 | } | |
c880097d | 2212 | else |
8af6d9cd | 2213 | { |
070ecdfd | 2214 | tree zero_dim_array = create_zero_dim_array (res, "Close_Phi"); |
8af6d9cd | 2215 | |
2724573f | 2216 | stmt = gimple_build_assign (res, unshare_expr (zero_dim_array)); |
8af6d9cd | 2217 | |
3dd2dd57 | 2218 | if (TREE_CODE (arg) == SSA_NAME) |
278b1a1d | 2219 | insert_out_of_ssa_copy (scop, zero_dim_array, arg, |
efa21390 | 2220 | SSA_NAME_DEF_STMT (arg)); |
8af6d9cd | 2221 | else |
efa21390 | 2222 | insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb), |
8af6d9cd SP |
2223 | zero_dim_array, arg); |
2224 | } | |
2abae5f1 SP |
2225 | |
2226 | remove_phi_node (psi, false); | |
2abae5f1 | 2227 | SSA_NAME_DEF_STMT (res) = stmt; |
278b1a1d SP |
2228 | |
2229 | insert_stmts (scop, stmt, NULL, gsi_after_labels (bb)); | |
2abae5f1 SP |
2230 | } |
2231 | ||
2232 | /* Rewrite out of SSA the reduction phi node at PSI by creating a zero | |
2233 | dimension array for it. */ | |
2234 | ||
2235 | static void | |
efa21390 | 2236 | rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi) |
2abae5f1 SP |
2237 | { |
2238 | size_t i; | |
2239 | gimple phi = gsi_stmt (*psi); | |
2240 | basic_block bb = gimple_bb (phi); | |
2241 | tree res = gimple_phi_result (phi); | |
070ecdfd | 2242 | tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa"); |
2abae5f1 | 2243 | gimple stmt; |
2abae5f1 SP |
2244 | |
2245 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2246 | { | |
2247 | tree arg = gimple_phi_arg_def (phi, i); | |
4aa9a167 | 2248 | edge e = gimple_phi_arg_edge (phi, i); |
2abae5f1 | 2249 | |
4aa9a167 SP |
2250 | /* Avoid the insertion of code in the loop latch to please the |
2251 | pattern matching of the vectorizer. */ | |
320532a8 SP |
2252 | if (TREE_CODE (arg) == SSA_NAME |
2253 | && e->src == bb->loop_father->latch) | |
278b1a1d | 2254 | insert_out_of_ssa_copy (scop, zero_dim_array, arg, |
efa21390 | 2255 | SSA_NAME_DEF_STMT (arg)); |
2abae5f1 | 2256 | else |
efa21390 | 2257 | insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg); |
2abae5f1 SP |
2258 | } |
2259 | ||
2724573f | 2260 | stmt = gimple_build_assign (res, unshare_expr (zero_dim_array)); |
2abae5f1 | 2261 | remove_phi_node (psi, false); |
2724573f | 2262 | insert_stmts (scop, stmt, NULL, gsi_after_labels (bb)); |
2abae5f1 SP |
2263 | } |
2264 | ||
d3e7b889 SP |
2265 | /* Rewrite the degenerate phi node at position PSI from the degenerate |
2266 | form "x = phi (y, y, ..., y)" to "x = y". */ | |
2267 | ||
2268 | static void | |
2269 | rewrite_degenerate_phi (gimple_stmt_iterator *psi) | |
2270 | { | |
2271 | tree rhs; | |
2272 | gimple stmt; | |
2273 | gimple_stmt_iterator gsi; | |
2274 | gimple phi = gsi_stmt (*psi); | |
2275 | tree res = gimple_phi_result (phi); | |
2276 | basic_block bb; | |
2277 | ||
d3e7b889 SP |
2278 | bb = gimple_bb (phi); |
2279 | rhs = degenerate_phi_result (phi); | |
2280 | gcc_assert (rhs); | |
2281 | ||
2282 | stmt = gimple_build_assign (res, rhs); | |
2283 | remove_phi_node (psi, false); | |
d3e7b889 SP |
2284 | |
2285 | gsi = gsi_after_labels (bb); | |
2286 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
2287 | } | |
2288 | ||
9773d730 SP |
2289 | /* Rewrite out of SSA all the reduction phi nodes of SCOP. */ |
2290 | ||
efa21390 | 2291 | static void |
9773d730 SP |
2292 | rewrite_reductions_out_of_ssa (scop_p scop) |
2293 | { | |
2294 | basic_block bb; | |
2295 | gimple_stmt_iterator psi; | |
2296 | sese region = SCOP_REGION (scop); | |
2297 | ||
11cd3bed | 2298 | FOR_EACH_BB_FN (bb, cfun) |
9773d730 SP |
2299 | if (bb_in_sese_p (bb, region)) |
2300 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi);) | |
2301 | { | |
d3e7b889 SP |
2302 | gimple phi = gsi_stmt (psi); |
2303 | ||
ea057359 | 2304 | if (virtual_operand_p (gimple_phi_result (phi))) |
c2bc669e SP |
2305 | { |
2306 | gsi_next (&psi); | |
2307 | continue; | |
2308 | } | |
2309 | ||
d3e7b889 SP |
2310 | if (gimple_phi_num_args (phi) > 1 |
2311 | && degenerate_phi_result (phi)) | |
2312 | rewrite_degenerate_phi (&psi); | |
2313 | ||
2314 | else if (scalar_close_phi_node_p (phi)) | |
efa21390 | 2315 | rewrite_close_phi_out_of_ssa (scop, &psi); |
d3e7b889 | 2316 | |
9773d730 | 2317 | else if (reduction_phi_p (region, &psi)) |
efa21390 | 2318 | rewrite_phi_out_of_ssa (scop, &psi); |
9773d730 SP |
2319 | } |
2320 | ||
2321 | update_ssa (TODO_update_ssa); | |
2322 | #ifdef ENABLE_CHECKING | |
2323 | verify_loop_closed_ssa (true); | |
2324 | #endif | |
2325 | } | |
2326 | ||
5dcc64d9 SP |
2327 | /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory |
2328 | read from ZERO_DIM_ARRAY. */ | |
2329 | ||
2330 | static void | |
278b1a1d | 2331 | rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array, |
efa21390 | 2332 | tree def, gimple use_stmt) |
5dcc64d9 | 2333 | { |
070ecdfd RG |
2334 | gimple name_stmt; |
2335 | tree name; | |
5dcc64d9 SP |
2336 | ssa_op_iter iter; |
2337 | use_operand_p use_p; | |
5dcc64d9 | 2338 | |
c7dc2fab | 2339 | gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI); |
5dcc64d9 | 2340 | |
070ecdfd RG |
2341 | name = copy_ssa_name (def, NULL); |
2342 | name_stmt = gimple_build_assign (name, zero_dim_array); | |
2343 | ||
c7dc2fab | 2344 | gimple_assign_set_lhs (name_stmt, name); |
278b1a1d | 2345 | insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt)); |
5dcc64d9 | 2346 | |
c7dc2fab SP |
2347 | FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES) |
2348 | if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)) | |
2349 | replace_exp (use_p, name); | |
5dcc64d9 SP |
2350 | |
2351 | update_stmt (use_stmt); | |
2352 | } | |
2353 | ||
70a2ae0f SP |
2354 | /* For every definition DEF in the SCOP that is used outside the scop, |
2355 | insert a closing-scop definition in the basic block just after this | |
2356 | SCOP. */ | |
2357 | ||
2358 | static void | |
2359 | handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt) | |
2360 | { | |
2361 | tree var = create_tmp_reg (TREE_TYPE (def), NULL); | |
2362 | tree new_name = make_ssa_name (var, stmt); | |
2363 | bool needs_copy = false; | |
2364 | use_operand_p use_p; | |
2365 | imm_use_iterator imm_iter; | |
2366 | gimple use_stmt; | |
2367 | sese region = SCOP_REGION (scop); | |
2368 | ||
2369 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2370 | { | |
2371 | if (!bb_in_sese_p (gimple_bb (use_stmt), region)) | |
2372 | { | |
2373 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
2374 | { | |
2375 | SET_USE (use_p, new_name); | |
2376 | } | |
2377 | update_stmt (use_stmt); | |
2378 | needs_copy = true; | |
2379 | } | |
2380 | } | |
2381 | ||
2382 | /* Insert in the empty BB just after the scop a use of DEF such | |
2383 | that the rewrite of cross_bb_scalar_dependences won't insert | |
2384 | arrays everywhere else. */ | |
2385 | if (needs_copy) | |
2386 | { | |
2387 | gimple assign = gimple_build_assign (new_name, def); | |
2388 | gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest); | |
2389 | ||
70a2ae0f SP |
2390 | update_stmt (assign); |
2391 | gsi_insert_before (&psi, assign, GSI_SAME_STMT); | |
2392 | } | |
2393 | } | |
2394 | ||
9773d730 | 2395 | /* Rewrite the scalar dependences crossing the boundary of the BB |
5d737345 SP |
2396 | containing STMT with an array. Return true when something has been |
2397 | changed. */ | |
9773d730 | 2398 | |
5d737345 | 2399 | static bool |
70a2ae0f | 2400 | rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi) |
9773d730 | 2401 | { |
70a2ae0f | 2402 | sese region = SCOP_REGION (scop); |
9773d730 SP |
2403 | gimple stmt = gsi_stmt (*gsi); |
2404 | imm_use_iterator imm_iter; | |
2405 | tree def; | |
2406 | basic_block def_bb; | |
2407 | tree zero_dim_array = NULL_TREE; | |
2408 | gimple use_stmt; | |
5d737345 | 2409 | bool res = false; |
9773d730 | 2410 | |
dba9acfa SP |
2411 | switch (gimple_code (stmt)) |
2412 | { | |
2413 | case GIMPLE_ASSIGN: | |
2414 | def = gimple_assign_lhs (stmt); | |
2415 | break; | |
2416 | ||
2417 | case GIMPLE_CALL: | |
2418 | def = gimple_call_lhs (stmt); | |
2419 | break; | |
2420 | ||
2421 | default: | |
5d737345 | 2422 | return false; |
dba9acfa | 2423 | } |
9773d730 | 2424 | |
b4c8119f SP |
2425 | if (!def |
2426 | || !is_gimple_reg (def)) | |
5d737345 | 2427 | return false; |
9773d730 | 2428 | |
1c2a7491 SP |
2429 | if (scev_analyzable_p (def, region)) |
2430 | { | |
2431 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def)); | |
2432 | tree scev = scalar_evolution_in_region (region, loop, def); | |
2433 | ||
5d737345 SP |
2434 | if (tree_contains_chrecs (scev, NULL)) |
2435 | return false; | |
1c2a7491 | 2436 | |
5d737345 SP |
2437 | propagate_expr_outside_region (def, scev, region); |
2438 | return true; | |
1c2a7491 SP |
2439 | } |
2440 | ||
9773d730 SP |
2441 | def_bb = gimple_bb (stmt); |
2442 | ||
70a2ae0f SP |
2443 | handle_scalar_deps_crossing_scop_limits (scop, def, stmt); |
2444 | ||
9773d730 | 2445 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) |
5d737345 SP |
2446 | if (gimple_code (use_stmt) == GIMPLE_PHI |
2447 | && (res = true)) | |
5dcc64d9 | 2448 | { |
ab756588 | 2449 | gimple_stmt_iterator psi = gsi_for_stmt (use_stmt); |
9773d730 | 2450 | |
ab756588 | 2451 | if (scalar_close_phi_node_p (gsi_stmt (psi))) |
efa21390 | 2452 | rewrite_close_phi_out_of_ssa (scop, &psi); |
ab756588 | 2453 | else |
efa21390 | 2454 | rewrite_phi_out_of_ssa (scop, &psi); |
ab756588 SP |
2455 | } |
2456 | ||
2457 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
2458 | if (gimple_code (use_stmt) != GIMPLE_PHI | |
2459 | && def_bb != gimple_bb (use_stmt) | |
5d737345 SP |
2460 | && !is_gimple_debug (use_stmt) |
2461 | && (res = true)) | |
ab756588 | 2462 | { |
5dcc64d9 SP |
2463 | if (!zero_dim_array) |
2464 | { | |
63858ac6 | 2465 | zero_dim_array = create_zero_dim_array |
070ecdfd | 2466 | (def, "Cross_BB_scalar_dependence"); |
278b1a1d | 2467 | insert_out_of_ssa_copy (scop, zero_dim_array, def, |
5fed5769 | 2468 | SSA_NAME_DEF_STMT (def)); |
5dcc64d9 SP |
2469 | gsi_next (gsi); |
2470 | } | |
2471 | ||
278b1a1d | 2472 | rewrite_cross_bb_scalar_dependence (scop, zero_dim_array, |
efa21390 | 2473 | def, use_stmt); |
5dcc64d9 | 2474 | } |
5d737345 SP |
2475 | |
2476 | return res; | |
5dcc64d9 SP |
2477 | } |
2478 | ||
ee646fc6 SP |
2479 | /* Rewrite out of SSA all the reduction phi nodes of SCOP. */ |
2480 | ||
efa21390 | 2481 | static void |
ee646fc6 SP |
2482 | rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop) |
2483 | { | |
2484 | basic_block bb; | |
2485 | gimple_stmt_iterator psi; | |
2486 | sese region = SCOP_REGION (scop); | |
5d737345 | 2487 | bool changed = false; |
5dcc64d9 | 2488 | |
70a2ae0f | 2489 | /* Create an extra empty BB after the scop. */ |
844e904d | 2490 | split_edge (SESE_EXIT (region)); |
70a2ae0f | 2491 | |
11cd3bed | 2492 | FOR_EACH_BB_FN (bb, cfun) |
5dcc64d9 SP |
2493 | if (bb_in_sese_p (bb, region)) |
2494 | for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
70a2ae0f | 2495 | changed |= rewrite_cross_bb_scalar_deps (scop, &psi); |
5dcc64d9 | 2496 | |
5d737345 SP |
2497 | if (changed) |
2498 | { | |
2499 | scev_reset_htab (); | |
2500 | update_ssa (TODO_update_ssa); | |
5dcc64d9 | 2501 | #ifdef ENABLE_CHECKING |
5d737345 | 2502 | verify_loop_closed_ssa (true); |
5dcc64d9 | 2503 | #endif |
5d737345 | 2504 | } |
2abae5f1 SP |
2505 | } |
2506 | ||
2507 | /* Returns the number of pbbs that are in loops contained in SCOP. */ | |
2508 | ||
2509 | static int | |
2510 | nb_pbbs_in_loops (scop_p scop) | |
2511 | { | |
2512 | int i; | |
2513 | poly_bb_p pbb; | |
2514 | int res = 0; | |
2515 | ||
9771b263 | 2516 | FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb) |
2abae5f1 SP |
2517 | if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop))) |
2518 | res++; | |
2519 | ||
2520 | return res; | |
2521 | } | |
2522 | ||
60d2a8c3 SP |
2523 | /* Return the number of data references in BB that write in |
2524 | memory. */ | |
2525 | ||
2526 | static int | |
2527 | nb_data_writes_in_bb (basic_block bb) | |
2528 | { | |
2529 | int res = 0; | |
2530 | gimple_stmt_iterator gsi; | |
2531 | ||
2532 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2533 | if (gimple_vdef (gsi_stmt (gsi))) | |
2534 | res++; | |
2535 | ||
2536 | return res; | |
2537 | } | |
2538 | ||
efa21390 SP |
2539 | /* Splits at STMT the basic block BB represented as PBB in the |
2540 | polyhedral form. */ | |
2541 | ||
2542 | static edge | |
2543 | split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt) | |
2544 | { | |
2545 | edge e1 = split_block (bb, stmt); | |
2546 | new_pbb_from_pbb (scop, pbb, e1->dest); | |
2547 | return e1; | |
2548 | } | |
2549 | ||
2550 | /* Splits STMT out of its current BB. This is done for reduction | |
2551 | statements for which we want to ignore data dependences. */ | |
a0dd1440 SP |
2552 | |
2553 | static basic_block | |
efa21390 | 2554 | split_reduction_stmt (scop_p scop, gimple stmt) |
a0dd1440 | 2555 | { |
a0dd1440 | 2556 | basic_block bb = gimple_bb (stmt); |
efa21390 | 2557 | poly_bb_p pbb = pbb_from_bb (bb); |
278b1a1d | 2558 | gimple_bb_p gbb = gbb_from_bb (bb); |
efa21390 | 2559 | edge e1; |
278b1a1d SP |
2560 | int i; |
2561 | data_reference_p dr; | |
a0dd1440 | 2562 | |
60d2a8c3 SP |
2563 | /* Do not split basic blocks with no writes to memory: the reduction |
2564 | will be the only write to memory. */ | |
c513da01 SP |
2565 | if (nb_data_writes_in_bb (bb) == 0 |
2566 | /* Or if we have already marked BB as a reduction. */ | |
2567 | || PBB_IS_REDUCTION (pbb_from_bb (bb))) | |
60d2a8c3 SP |
2568 | return bb; |
2569 | ||
efa21390 | 2570 | e1 = split_pbb (scop, pbb, bb, stmt); |
a0dd1440 | 2571 | |
efa21390 SP |
2572 | /* Split once more only when the reduction stmt is not the only one |
2573 | left in the original BB. */ | |
2574 | if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb))) | |
2575 | { | |
2576 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
2577 | gsi_prev (&gsi); | |
2578 | e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi)); | |
2579 | } | |
a0dd1440 | 2580 | |
278b1a1d SP |
2581 | /* A part of the data references will end in a different basic block |
2582 | after the split: move the DRs from the original GBB to the newly | |
2583 | created GBB1. */ | |
9771b263 | 2584 | FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr) |
278b1a1d SP |
2585 | { |
2586 | basic_block bb1 = gimple_bb (DR_STMT (dr)); | |
2587 | ||
2588 | if (bb1 != bb) | |
2589 | { | |
2590 | gimple_bb_p gbb1 = gbb_from_bb (bb1); | |
9771b263 DN |
2591 | GBB_DATA_REFS (gbb1).safe_push (dr); |
2592 | GBB_DATA_REFS (gbb).ordered_remove (i); | |
278b1a1d SP |
2593 | i--; |
2594 | } | |
2595 | } | |
2596 | ||
efa21390 | 2597 | return e1->dest; |
a0dd1440 SP |
2598 | } |
2599 | ||
2600 | /* Return true when stmt is a reduction operation. */ | |
2601 | ||
2602 | static inline bool | |
2603 | is_reduction_operation_p (gimple stmt) | |
2604 | { | |
0596e97f AH |
2605 | enum tree_code code; |
2606 | ||
2607 | gcc_assert (is_gimple_assign (stmt)); | |
2608 | code = gimple_assign_rhs_code (stmt); | |
2609 | ||
a0dd1440 | 2610 | return flag_associative_math |
0596e97f AH |
2611 | && commutative_tree_code (code) |
2612 | && associative_tree_code (code); | |
a0dd1440 SP |
2613 | } |
2614 | ||
2615 | /* Returns true when PHI contains an argument ARG. */ | |
2616 | ||
2617 | static bool | |
2618 | phi_contains_arg (gimple phi, tree arg) | |
2619 | { | |
2620 | size_t i; | |
2621 | ||
2622 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2623 | if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0)) | |
2624 | return true; | |
2625 | ||
2626 | return false; | |
2627 | } | |
2628 | ||
2629 | /* Return a loop phi node that corresponds to a reduction containing LHS. */ | |
2630 | ||
2631 | static gimple | |
2632 | follow_ssa_with_commutative_ops (tree arg, tree lhs) | |
2633 | { | |
2634 | gimple stmt; | |
2635 | ||
2636 | if (TREE_CODE (arg) != SSA_NAME) | |
2637 | return NULL; | |
2638 | ||
2639 | stmt = SSA_NAME_DEF_STMT (arg); | |
2640 | ||
a84a556d SP |
2641 | if (gimple_code (stmt) == GIMPLE_NOP |
2642 | || gimple_code (stmt) == GIMPLE_CALL) | |
403ebc7e SP |
2643 | return NULL; |
2644 | ||
a0dd1440 SP |
2645 | if (gimple_code (stmt) == GIMPLE_PHI) |
2646 | { | |
2647 | if (phi_contains_arg (stmt, lhs)) | |
2648 | return stmt; | |
2649 | return NULL; | |
2650 | } | |
2651 | ||
0596e97f AH |
2652 | if (!is_gimple_assign (stmt)) |
2653 | return NULL; | |
2654 | ||
a0dd1440 SP |
2655 | if (gimple_num_ops (stmt) == 2) |
2656 | return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs); | |
2657 | ||
2658 | if (is_reduction_operation_p (stmt)) | |
2659 | { | |
2660 | gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs); | |
2661 | ||
2662 | return res ? res : | |
2663 | follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs); | |
2664 | } | |
2665 | ||
2666 | return NULL; | |
2667 | } | |
2668 | ||
2669 | /* Detect commutative and associative scalar reductions starting at | |
c880097d | 2670 | the STMT. Return the phi node of the reduction cycle, or NULL. */ |
a0dd1440 SP |
2671 | |
2672 | static gimple | |
2673 | detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg, | |
9771b263 DN |
2674 | vec<gimple> *in, |
2675 | vec<gimple> *out) | |
a0dd1440 SP |
2676 | { |
2677 | gimple phi = follow_ssa_with_commutative_ops (arg, lhs); | |
2678 | ||
c880097d SP |
2679 | if (!phi) |
2680 | return NULL; | |
a0dd1440 | 2681 | |
9771b263 DN |
2682 | in->safe_push (stmt); |
2683 | out->safe_push (stmt); | |
c880097d | 2684 | return phi; |
a0dd1440 SP |
2685 | } |
2686 | ||
2687 | /* Detect commutative and associative scalar reductions starting at | |
3a7086cc | 2688 | STMT. Return the phi node of the reduction cycle, or NULL. */ |
a0dd1440 SP |
2689 | |
2690 | static gimple | |
9771b263 DN |
2691 | detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in, |
2692 | vec<gimple> *out) | |
a0dd1440 SP |
2693 | { |
2694 | tree lhs = gimple_assign_lhs (stmt); | |
2695 | ||
2696 | if (gimple_num_ops (stmt) == 2) | |
2697 | return detect_commutative_reduction_arg (lhs, stmt, | |
2698 | gimple_assign_rhs1 (stmt), | |
2699 | in, out); | |
2700 | ||
2701 | if (is_reduction_operation_p (stmt)) | |
2702 | { | |
2703 | gimple res = detect_commutative_reduction_arg (lhs, stmt, | |
2704 | gimple_assign_rhs1 (stmt), | |
2705 | in, out); | |
2706 | return res ? res | |
2707 | : detect_commutative_reduction_arg (lhs, stmt, | |
2708 | gimple_assign_rhs2 (stmt), | |
2709 | in, out); | |
2710 | } | |
2711 | ||
2712 | return NULL; | |
2713 | } | |
2714 | ||
2715 | /* Return a loop phi node that corresponds to a reduction containing LHS. */ | |
2716 | ||
2717 | static gimple | |
2718 | follow_inital_value_to_phi (tree arg, tree lhs) | |
2719 | { | |
2720 | gimple stmt; | |
2721 | ||
2722 | if (!arg || TREE_CODE (arg) != SSA_NAME) | |
2723 | return NULL; | |
2724 | ||
2725 | stmt = SSA_NAME_DEF_STMT (arg); | |
2726 | ||
2727 | if (gimple_code (stmt) == GIMPLE_PHI | |
2728 | && phi_contains_arg (stmt, lhs)) | |
2729 | return stmt; | |
2730 | ||
2731 | return NULL; | |
2732 | } | |
2733 | ||
2734 | ||
073a8998 | 2735 | /* Return the argument of the loop PHI that is the initial value coming |
a0dd1440 SP |
2736 | from outside the loop. */ |
2737 | ||
2738 | static edge | |
2739 | edge_initial_value_for_loop_phi (gimple phi) | |
2740 | { | |
2741 | size_t i; | |
2742 | ||
2743 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2744 | { | |
2745 | edge e = gimple_phi_arg_edge (phi, i); | |
2746 | ||
2747 | if (loop_depth (e->src->loop_father) | |
2748 | < loop_depth (e->dest->loop_father)) | |
2749 | return e; | |
2750 | } | |
2751 | ||
2752 | return NULL; | |
2753 | } | |
2754 | ||
073a8998 | 2755 | /* Return the argument of the loop PHI that is the initial value coming |
a0dd1440 SP |
2756 | from outside the loop. */ |
2757 | ||
2758 | static tree | |
2759 | initial_value_for_loop_phi (gimple phi) | |
2760 | { | |
2761 | size_t i; | |
2762 | ||
2763 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2764 | { | |
2765 | edge e = gimple_phi_arg_edge (phi, i); | |
2766 | ||
2767 | if (loop_depth (e->src->loop_father) | |
2768 | < loop_depth (e->dest->loop_father)) | |
2769 | return gimple_phi_arg_def (phi, i); | |
2770 | } | |
2771 | ||
2772 | return NULL_TREE; | |
2773 | } | |
2774 | ||
479c1fb3 SP |
2775 | /* Returns true when DEF is used outside the reduction cycle of |
2776 | LOOP_PHI. */ | |
2777 | ||
2778 | static bool | |
2779 | used_outside_reduction (tree def, gimple loop_phi) | |
2780 | { | |
2781 | use_operand_p use_p; | |
2782 | imm_use_iterator imm_iter; | |
2783 | loop_p loop = loop_containing_stmt (loop_phi); | |
2784 | ||
2785 | /* In LOOP, DEF should be used only in LOOP_PHI. */ | |
2786 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2787 | { | |
2788 | gimple stmt = USE_STMT (use_p); | |
2789 | ||
2790 | if (stmt != loop_phi | |
2791 | && !is_gimple_debug (stmt) | |
2792 | && flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
2793 | return true; | |
2794 | } | |
2795 | ||
2796 | return false; | |
2797 | } | |
2798 | ||
a30e5345 SP |
2799 | /* Detect commutative and associative scalar reductions belonging to |
2800 | the SCOP starting at the loop closed phi node STMT. Return the phi | |
2801 | node of the reduction cycle, or NULL. */ | |
a0dd1440 SP |
2802 | |
2803 | static gimple | |
9771b263 DN |
2804 | detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in, |
2805 | vec<gimple> *out) | |
a0dd1440 SP |
2806 | { |
2807 | if (scalar_close_phi_node_p (stmt)) | |
2808 | { | |
479c1fb3 SP |
2809 | gimple def, loop_phi, phi, close_phi = stmt; |
2810 | tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0); | |
c880097d SP |
2811 | |
2812 | if (TREE_CODE (arg) != SSA_NAME) | |
2813 | return NULL; | |
2814 | ||
79d03cf8 SP |
2815 | /* Note that loop close phi nodes should have a single argument |
2816 | because we translated the representation into a canonical form | |
2817 | before Graphite: see canonicalize_loop_closed_ssa_form. */ | |
479c1fb3 | 2818 | gcc_assert (gimple_phi_num_args (close_phi) == 1); |
79d03cf8 | 2819 | |
c880097d | 2820 | def = SSA_NAME_DEF_STMT (arg); |
479c1fb3 SP |
2821 | if (!stmt_in_sese_p (def, SCOP_REGION (scop)) |
2822 | || !(loop_phi = detect_commutative_reduction (scop, def, in, out))) | |
a30e5345 SP |
2823 | return NULL; |
2824 | ||
479c1fb3 SP |
2825 | lhs = gimple_phi_result (close_phi); |
2826 | init = initial_value_for_loop_phi (loop_phi); | |
2827 | phi = follow_inital_value_to_phi (init, lhs); | |
a0dd1440 | 2828 | |
479c1fb3 SP |
2829 | if (phi && (used_outside_reduction (lhs, phi) |
2830 | || !has_single_use (gimple_phi_result (phi)))) | |
a0dd1440 | 2831 | return NULL; |
479c1fb3 | 2832 | |
9771b263 DN |
2833 | in->safe_push (loop_phi); |
2834 | out->safe_push (close_phi); | |
479c1fb3 | 2835 | return phi; |
a0dd1440 SP |
2836 | } |
2837 | ||
2838 | if (gimple_code (stmt) == GIMPLE_ASSIGN) | |
2839 | return detect_commutative_reduction_assign (stmt, in, out); | |
2840 | ||
2841 | return NULL; | |
2842 | } | |
2843 | ||
2844 | /* Translate the scalar reduction statement STMT to an array RED | |
2845 | knowing that its recursive phi node is LOOP_PHI. */ | |
2846 | ||
2847 | static void | |
278b1a1d SP |
2848 | translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red, |
2849 | gimple stmt, gimple loop_phi) | |
a0dd1440 | 2850 | { |
a0dd1440 | 2851 | tree res = gimple_phi_result (loop_phi); |
50034a36 | 2852 | gimple assign = gimple_build_assign (res, unshare_expr (red)); |
278b1a1d | 2853 | gimple_stmt_iterator gsi; |
a0dd1440 | 2854 | |
278b1a1d | 2855 | insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi))); |
a0dd1440 | 2856 | |
50034a36 | 2857 | assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt)); |
278b1a1d SP |
2858 | gsi = gsi_for_stmt (stmt); |
2859 | gsi_next (&gsi); | |
2860 | insert_stmts (scop, assign, NULL, gsi); | |
a0dd1440 SP |
2861 | } |
2862 | ||
a4681954 SP |
2863 | /* Removes the PHI node and resets all the debug stmts that are using |
2864 | the PHI_RESULT. */ | |
2865 | ||
2866 | static void | |
2867 | remove_phi (gimple phi) | |
2868 | { | |
2869 | imm_use_iterator imm_iter; | |
2870 | tree def; | |
2871 | use_operand_p use_p; | |
2872 | gimple_stmt_iterator gsi; | |
00f96dc9 | 2873 | auto_vec<gimple, 3> update; |
a4681954 SP |
2874 | unsigned int i; |
2875 | gimple stmt; | |
2876 | ||
2877 | def = PHI_RESULT (phi); | |
2878 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2879 | { | |
2880 | stmt = USE_STMT (use_p); | |
2881 | ||
2882 | if (is_gimple_debug (stmt)) | |
2883 | { | |
2884 | gimple_debug_bind_reset_value (stmt); | |
9771b263 | 2885 | update.safe_push (stmt); |
a4681954 SP |
2886 | } |
2887 | } | |
2888 | ||
9771b263 | 2889 | FOR_EACH_VEC_ELT (update, i, stmt) |
a4681954 SP |
2890 | update_stmt (stmt); |
2891 | ||
a4681954 SP |
2892 | gsi = gsi_for_phi_node (phi); |
2893 | remove_phi_node (&gsi, false); | |
2894 | } | |
2895 | ||
7c48ea69 SP |
2896 | /* Helper function for for_each_index. For each INDEX of the data |
2897 | reference REF, returns true when its indices are valid in the loop | |
2898 | nest LOOP passed in as DATA. */ | |
2899 | ||
2900 | static bool | |
2901 | dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data) | |
2902 | { | |
2903 | loop_p loop; | |
2904 | basic_block header, def_bb; | |
2905 | gimple stmt; | |
2906 | ||
2907 | if (TREE_CODE (*index) != SSA_NAME) | |
2908 | return true; | |
2909 | ||
2910 | loop = *((loop_p *) data); | |
2911 | header = loop->header; | |
2912 | stmt = SSA_NAME_DEF_STMT (*index); | |
2913 | ||
2914 | if (!stmt) | |
2915 | return true; | |
2916 | ||
2917 | def_bb = gimple_bb (stmt); | |
2918 | ||
2919 | if (!def_bb) | |
2920 | return true; | |
2921 | ||
2922 | return dominated_by_p (CDI_DOMINATORS, header, def_bb); | |
2923 | } | |
2924 | ||
50034a36 SP |
2925 | /* When the result of a CLOSE_PHI is written to a memory location, |
2926 | return a pointer to that memory reference, otherwise return | |
2927 | NULL_TREE. */ | |
2928 | ||
2929 | static tree | |
2930 | close_phi_written_to_memory (gimple close_phi) | |
2931 | { | |
2932 | imm_use_iterator imm_iter; | |
50034a36 SP |
2933 | use_operand_p use_p; |
2934 | gimple stmt; | |
7c48ea69 | 2935 | tree res, def = gimple_phi_result (close_phi); |
50034a36 SP |
2936 | |
2937 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
2938 | if ((stmt = USE_STMT (use_p)) | |
2939 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
7c48ea69 SP |
2940 | && (res = gimple_assign_lhs (stmt))) |
2941 | { | |
2942 | switch (TREE_CODE (res)) | |
2943 | { | |
2944 | case VAR_DECL: | |
2945 | case PARM_DECL: | |
2946 | case RESULT_DECL: | |
2947 | return res; | |
2948 | ||
2949 | case ARRAY_REF: | |
2950 | case MEM_REF: | |
2951 | { | |
2952 | tree arg = gimple_phi_arg_def (close_phi, 0); | |
2953 | loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg)); | |
2954 | ||
2955 | /* FIXME: this restriction is for id-{24,25}.f and | |
2956 | could be handled by duplicating the computation of | |
2957 | array indices before the loop of the close_phi. */ | |
2958 | if (for_each_index (&res, dr_indices_valid_in_loop, &nest)) | |
2959 | return res; | |
2960 | } | |
2961 | /* Fallthru. */ | |
50034a36 | 2962 | |
7c48ea69 SP |
2963 | default: |
2964 | continue; | |
2965 | } | |
2966 | } | |
50034a36 SP |
2967 | return NULL_TREE; |
2968 | } | |
2969 | ||
a0dd1440 SP |
2970 | /* Rewrite out of SSA the reduction described by the loop phi nodes |
2971 | IN, and the close phi nodes OUT. IN and OUT are structured by loop | |
2972 | levels like this: | |
2973 | ||
2974 | IN: stmt, loop_n, ..., loop_0 | |
2975 | OUT: stmt, close_n, ..., close_0 | |
2976 | ||
2977 | the first element is the reduction statement, and the next elements | |
2978 | are the loop and close phi nodes of each of the outer loops. */ | |
2979 | ||
2980 | static void | |
efa21390 | 2981 | translate_scalar_reduction_to_array (scop_p scop, |
9771b263 DN |
2982 | vec<gimple> in, |
2983 | vec<gimple> out) | |
a0dd1440 | 2984 | { |
a0dd1440 | 2985 | gimple loop_phi; |
9771b263 DN |
2986 | unsigned int i = out.length () - 1; |
2987 | tree red = close_phi_written_to_memory (out[i]); | |
a0dd1440 | 2988 | |
9771b263 | 2989 | FOR_EACH_VEC_ELT (in, i, loop_phi) |
a0dd1440 | 2990 | { |
9771b263 | 2991 | gimple close_phi = out[i]; |
a0dd1440 SP |
2992 | |
2993 | if (i == 0) | |
2994 | { | |
2995 | gimple stmt = loop_phi; | |
efa21390 SP |
2996 | basic_block bb = split_reduction_stmt (scop, stmt); |
2997 | poly_bb_p pbb = pbb_from_bb (bb); | |
2998 | PBB_IS_REDUCTION (pbb) = true; | |
a0dd1440 SP |
2999 | gcc_assert (close_phi == loop_phi); |
3000 | ||
50034a36 SP |
3001 | if (!red) |
3002 | red = create_zero_dim_array | |
3003 | (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction"); | |
3004 | ||
9771b263 | 3005 | translate_scalar_reduction_to_array_for_stmt (scop, red, stmt, in[1]); |
a0dd1440 SP |
3006 | continue; |
3007 | } | |
3008 | ||
9771b263 | 3009 | if (i == in.length () - 1) |
a0dd1440 | 3010 | { |
50034a36 SP |
3011 | insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi), |
3012 | unshare_expr (red), close_phi); | |
5fed5769 | 3013 | insert_out_of_ssa_copy_on_edge |
efa21390 | 3014 | (scop, edge_initial_value_for_loop_phi (loop_phi), |
50034a36 | 3015 | unshare_expr (red), initial_value_for_loop_phi (loop_phi)); |
a0dd1440 SP |
3016 | } |
3017 | ||
a4681954 SP |
3018 | remove_phi (loop_phi); |
3019 | remove_phi (close_phi); | |
a0dd1440 SP |
3020 | } |
3021 | } | |
3022 | ||
5d737345 SP |
3023 | /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns |
3024 | true when something has been changed. */ | |
a0dd1440 | 3025 | |
5d737345 | 3026 | static bool |
efa21390 SP |
3027 | rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop, |
3028 | gimple close_phi) | |
a0dd1440 | 3029 | { |
5d737345 | 3030 | bool res; |
00f96dc9 TS |
3031 | auto_vec<gimple, 10> in; |
3032 | auto_vec<gimple, 10> out; | |
a0dd1440 | 3033 | |
a30e5345 | 3034 | detect_commutative_reduction (scop, close_phi, &in, &out); |
9771b263 | 3035 | res = in.length () > 1; |
5d737345 | 3036 | if (res) |
efa21390 | 3037 | translate_scalar_reduction_to_array (scop, in, out); |
a0dd1440 | 3038 | |
5d737345 | 3039 | return res; |
a0dd1440 SP |
3040 | } |
3041 | ||
5d737345 SP |
3042 | /* Rewrites all the commutative reductions from LOOP out of SSA. |
3043 | Returns true when something has been changed. */ | |
a0dd1440 | 3044 | |
5d737345 | 3045 | static bool |
efa21390 SP |
3046 | rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop, |
3047 | loop_p loop) | |
a0dd1440 SP |
3048 | { |
3049 | gimple_stmt_iterator gsi; | |
3050 | edge exit = single_exit (loop); | |
4ee23fa8 | 3051 | tree res; |
5d737345 | 3052 | bool changed = false; |
a0dd1440 SP |
3053 | |
3054 | if (!exit) | |
5d737345 | 3055 | return false; |
a0dd1440 SP |
3056 | |
3057 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4ee23fa8 | 3058 | if ((res = gimple_phi_result (gsi_stmt (gsi))) |
ea057359 | 3059 | && !virtual_operand_p (res) |
efa21390 | 3060 | && !scev_analyzable_p (res, SCOP_REGION (scop))) |
5d737345 | 3061 | changed |= rewrite_commutative_reductions_out_of_ssa_close_phi |
efa21390 | 3062 | (scop, gsi_stmt (gsi)); |
5d737345 SP |
3063 | |
3064 | return changed; | |
a0dd1440 SP |
3065 | } |
3066 | ||
3067 | /* Rewrites all the commutative reductions from SCOP out of SSA. */ | |
3068 | ||
efa21390 SP |
3069 | static void |
3070 | rewrite_commutative_reductions_out_of_ssa (scop_p scop) | |
a0dd1440 | 3071 | { |
a0dd1440 | 3072 | loop_p loop; |
5d737345 | 3073 | bool changed = false; |
efa21390 | 3074 | sese region = SCOP_REGION (scop); |
cc588970 | 3075 | |
f0bd40b1 | 3076 | FOR_EACH_LOOP (loop, 0) |
a0dd1440 | 3077 | if (loop_in_sese_p (loop, region)) |
efa21390 | 3078 | changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop); |
6c4499b6 | 3079 | |
5d737345 SP |
3080 | if (changed) |
3081 | { | |
3082 | scev_reset_htab (); | |
3083 | gsi_commit_edge_inserts (); | |
3084 | update_ssa (TODO_update_ssa); | |
6c4499b6 | 3085 | #ifdef ENABLE_CHECKING |
5d737345 | 3086 | verify_loop_closed_ssa (true); |
6c4499b6 | 3087 | #endif |
5d737345 | 3088 | } |
a0dd1440 SP |
3089 | } |
3090 | ||
68d3ff90 TG |
3091 | /* Can all ivs be represented by a signed integer? |
3092 | As CLooG might generate negative values in its expressions, signed loop ivs | |
3093 | are required in the backend. */ | |
072edf07 | 3094 | |
68d3ff90 TG |
3095 | static bool |
3096 | scop_ivs_can_be_represented (scop_p scop) | |
3097 | { | |
68d3ff90 | 3098 | loop_p loop; |
a0d1afb3 | 3099 | gimple_stmt_iterator psi; |
f5843d08 | 3100 | bool result = true; |
68d3ff90 | 3101 | |
f0bd40b1 | 3102 | FOR_EACH_LOOP (loop, 0) |
68d3ff90 | 3103 | { |
68d3ff90 TG |
3104 | if (!loop_in_sese_p (loop, SCOP_REGION (scop))) |
3105 | continue; | |
3106 | ||
a0d1afb3 SP |
3107 | for (psi = gsi_start_phis (loop->header); |
3108 | !gsi_end_p (psi); gsi_next (&psi)) | |
3109 | { | |
3110 | gimple phi = gsi_stmt (psi); | |
3111 | tree res = PHI_RESULT (phi); | |
3112 | tree type = TREE_TYPE (res); | |
68d3ff90 | 3113 | |
a0d1afb3 | 3114 | if (TYPE_UNSIGNED (type) |
0d82a1c8 | 3115 | && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node)) |
f5843d08 RG |
3116 | { |
3117 | result = false; | |
3118 | break; | |
3119 | } | |
a0d1afb3 | 3120 | } |
f5843d08 | 3121 | if (!result) |
f0bd40b1 | 3122 | break; |
68d3ff90 TG |
3123 | } |
3124 | ||
f5843d08 | 3125 | return result; |
68d3ff90 TG |
3126 | } |
3127 | ||
2abae5f1 SP |
3128 | /* Builds the polyhedral representation for a SESE region. */ |
3129 | ||
e84aaa33 | 3130 | void |
2abae5f1 SP |
3131 | build_poly_scop (scop_p scop) |
3132 | { | |
3133 | sese region = SCOP_REGION (scop); | |
4e7dd376 | 3134 | graphite_dim_t max_dim; |
a0dd1440 | 3135 | |
efa21390 | 3136 | build_scop_bbs (scop); |
2abae5f1 SP |
3137 | |
3138 | /* FIXME: This restriction is needed to avoid a problem in CLooG. | |
3139 | Once CLooG is fixed, remove this guard. Anyways, it makes no | |
3140 | sense to optimize a scop containing only PBBs that do not belong | |
3141 | to any loops. */ | |
3142 | if (nb_pbbs_in_loops (scop) == 0) | |
e84aaa33 | 3143 | return; |
2abae5f1 | 3144 | |
68d3ff90 | 3145 | if (!scop_ivs_can_be_represented (scop)) |
e84aaa33 | 3146 | return; |
68d3ff90 | 3147 | |
ac53c069 SP |
3148 | if (flag_associative_math) |
3149 | rewrite_commutative_reductions_out_of_ssa (scop); | |
3150 | ||
2abae5f1 | 3151 | build_sese_loop_nests (region); |
4d9192b5 TS |
3152 | /* Record all conditions in REGION. */ |
3153 | sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr); | |
2abae5f1 SP |
3154 | find_scop_parameters (scop); |
3155 | ||
4e7dd376 SP |
3156 | max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS); |
3157 | if (scop_nb_params (scop) > max_dim) | |
e84aaa33 | 3158 | return; |
4e7dd376 | 3159 | |
2abae5f1 SP |
3160 | build_scop_iteration_domain (scop); |
3161 | build_scop_context (scop); | |
2abae5f1 | 3162 | add_conditions_to_constraints (scop); |
efa21390 SP |
3163 | |
3164 | /* Rewrite out of SSA only after having translated the | |
3165 | representation to the polyhedral representation to avoid scev | |
3166 | analysis failures. That means that these functions will insert | |
3167 | new data references that they create in the right place. */ | |
efa21390 SP |
3168 | rewrite_reductions_out_of_ssa (scop); |
3169 | rewrite_cross_bb_scalar_deps_out_of_ssa (scop); | |
3170 | ||
3171 | build_scop_drs (scop); | |
a36d12e2 | 3172 | scop_to_lst (scop); |
2abae5f1 | 3173 | build_scop_scattering (scop); |
2abae5f1 | 3174 | |
e84aaa33 SP |
3175 | /* This SCoP has been translated to the polyhedral |
3176 | representation. */ | |
3177 | POLY_SCOP_P (scop) = true; | |
2abae5f1 | 3178 | } |
2abae5f1 | 3179 | #endif |