]>
Commit | Line | Data |
---|---|---|
629b3d75 MJ |
1 | /* General types and functions that are uselful for processing of OpenMP, |
2 | OpenACC and similar directivers at various stages of compilation. | |
3 | ||
99dee823 | 4 | Copyright (C) 2005-2021 Free Software Foundation, Inc. |
629b3d75 MJ |
5 | |
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | /* Find an OMP clause of type KIND within CLAUSES. */ | |
23 | ||
24 | #include "config.h" | |
25 | #include "system.h" | |
26 | #include "coretypes.h" | |
27 | #include "backend.h" | |
28 | #include "target.h" | |
29 | #include "tree.h" | |
30 | #include "gimple.h" | |
31 | #include "ssa.h" | |
32 | #include "diagnostic-core.h" | |
33 | #include "fold-const.h" | |
34 | #include "langhooks.h" | |
35 | #include "omp-general.h" | |
314e6352 ML |
36 | #include "stringpool.h" |
37 | #include "attribs.h" | |
135df52c JJ |
38 | #include "gimplify.h" |
39 | #include "cgraph.h" | |
a895e6d7 | 40 | #include "alloc-pool.h" |
135df52c | 41 | #include "symbol-summary.h" |
135df52c | 42 | #include "tree-pass.h" |
9ba66bf5 | 43 | #include "omp-device-properties.h" |
f1f862ae | 44 | #include "tree-iterator.h" |
f165ef89 JJ |
45 | #include "data-streamer.h" |
46 | #include "streamer-hooks.h" | |
629b3d75 | 47 | |
28567c40 JJ |
48 | enum omp_requires omp_requires_mask; |
49 | ||
629b3d75 MJ |
50 | tree |
51 | omp_find_clause (tree clauses, enum omp_clause_code kind) | |
52 | { | |
53 | for (; clauses ; clauses = OMP_CLAUSE_CHAIN (clauses)) | |
54 | if (OMP_CLAUSE_CODE (clauses) == kind) | |
55 | return clauses; | |
56 | ||
57 | return NULL_TREE; | |
58 | } | |
59 | ||
08c14aaa TB |
60 | /* True if OpenMP should regard this DECL as being a scalar which has Fortran's |
61 | allocatable or pointer attribute. */ | |
62 | bool | |
63 | omp_is_allocatable_or_ptr (tree decl) | |
64 | { | |
65 | return lang_hooks.decls.omp_is_allocatable_or_ptr (decl); | |
66 | } | |
67 | ||
a2c26c50 TB |
68 | /* Check whether this DECL belongs to a Fortran optional argument. |
69 | With 'for_present_check' set to false, decls which are optional parameters | |
70 | themselve are returned as tree - or a NULL_TREE otherwise. Those decls are | |
71 | always pointers. With 'for_present_check' set to true, the decl for checking | |
72 | whether an argument is present is returned; for arguments with value | |
73 | attribute this is the hidden argument and of BOOLEAN_TYPE. If the decl is | |
74 | unrelated to optional arguments, NULL_TREE is returned. */ | |
73a28634 | 75 | |
a2c26c50 TB |
76 | tree |
77 | omp_check_optional_argument (tree decl, bool for_present_check) | |
73a28634 | 78 | { |
a2c26c50 | 79 | return lang_hooks.decls.omp_check_optional_argument (decl, for_present_check); |
73a28634 KCY |
80 | } |
81 | ||
22e6b327 TS |
82 | /* True if OpenMP should privatize what this DECL points to rather |
83 | than the DECL itself. */ | |
629b3d75 MJ |
84 | |
85 | bool | |
22e6b327 | 86 | omp_privatize_by_reference (tree decl) |
629b3d75 MJ |
87 | { |
88 | return lang_hooks.decls.omp_privatize_by_reference (decl); | |
89 | } | |
90 | ||
031c5c8b MJ |
91 | /* Adjust *COND_CODE and *N2 so that the former is either LT_EXPR or GT_EXPR, |
92 | given that V is the loop index variable and STEP is loop step. */ | |
629b3d75 MJ |
93 | |
94 | void | |
031c5c8b MJ |
95 | omp_adjust_for_condition (location_t loc, enum tree_code *cond_code, tree *n2, |
96 | tree v, tree step) | |
629b3d75 MJ |
97 | { |
98 | switch (*cond_code) | |
99 | { | |
100 | case LT_EXPR: | |
101 | case GT_EXPR: | |
031c5c8b MJ |
102 | break; |
103 | ||
629b3d75 | 104 | case NE_EXPR: |
031c5c8b MJ |
105 | gcc_assert (TREE_CODE (step) == INTEGER_CST); |
106 | if (TREE_CODE (TREE_TYPE (v)) == INTEGER_TYPE) | |
107 | { | |
108 | if (integer_onep (step)) | |
109 | *cond_code = LT_EXPR; | |
110 | else | |
111 | { | |
112 | gcc_assert (integer_minus_onep (step)); | |
113 | *cond_code = GT_EXPR; | |
114 | } | |
115 | } | |
116 | else | |
117 | { | |
118 | tree unit = TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (v))); | |
119 | gcc_assert (TREE_CODE (unit) == INTEGER_CST); | |
120 | if (tree_int_cst_equal (unit, step)) | |
121 | *cond_code = LT_EXPR; | |
122 | else | |
123 | { | |
124 | gcc_assert (wi::neg (wi::to_widest (unit)) | |
125 | == wi::to_widest (step)); | |
126 | *cond_code = GT_EXPR; | |
127 | } | |
128 | } | |
129 | ||
629b3d75 | 130 | break; |
031c5c8b | 131 | |
629b3d75 MJ |
132 | case LE_EXPR: |
133 | if (POINTER_TYPE_P (TREE_TYPE (*n2))) | |
134 | *n2 = fold_build_pointer_plus_hwi_loc (loc, *n2, 1); | |
135 | else | |
136 | *n2 = fold_build2_loc (loc, PLUS_EXPR, TREE_TYPE (*n2), *n2, | |
137 | build_int_cst (TREE_TYPE (*n2), 1)); | |
138 | *cond_code = LT_EXPR; | |
139 | break; | |
140 | case GE_EXPR: | |
141 | if (POINTER_TYPE_P (TREE_TYPE (*n2))) | |
142 | *n2 = fold_build_pointer_plus_hwi_loc (loc, *n2, -1); | |
143 | else | |
144 | *n2 = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (*n2), *n2, | |
145 | build_int_cst (TREE_TYPE (*n2), 1)); | |
146 | *cond_code = GT_EXPR; | |
147 | break; | |
148 | default: | |
149 | gcc_unreachable (); | |
150 | } | |
151 | } | |
152 | ||
153 | /* Return the looping step from INCR, extracted from the step of a gimple omp | |
154 | for statement. */ | |
155 | ||
156 | tree | |
157 | omp_get_for_step_from_incr (location_t loc, tree incr) | |
158 | { | |
159 | tree step; | |
160 | switch (TREE_CODE (incr)) | |
161 | { | |
162 | case PLUS_EXPR: | |
163 | step = TREE_OPERAND (incr, 1); | |
164 | break; | |
165 | case POINTER_PLUS_EXPR: | |
166 | step = fold_convert (ssizetype, TREE_OPERAND (incr, 1)); | |
167 | break; | |
168 | case MINUS_EXPR: | |
169 | step = TREE_OPERAND (incr, 1); | |
170 | step = fold_build1_loc (loc, NEGATE_EXPR, TREE_TYPE (step), step); | |
171 | break; | |
172 | default: | |
173 | gcc_unreachable (); | |
174 | } | |
175 | return step; | |
176 | } | |
177 | ||
178 | /* Extract the header elements of parallel loop FOR_STMT and store | |
179 | them into *FD. */ | |
180 | ||
181 | void | |
182 | omp_extract_for_data (gomp_for *for_stmt, struct omp_for_data *fd, | |
183 | struct omp_for_data_loop *loops) | |
184 | { | |
185 | tree t, var, *collapse_iter, *collapse_count; | |
186 | tree count = NULL_TREE, iter_type = long_integer_type_node; | |
187 | struct omp_for_data_loop *loop; | |
188 | int i; | |
189 | struct omp_for_data_loop dummy_loop; | |
190 | location_t loc = gimple_location (for_stmt); | |
dfa6e5b4 | 191 | bool simd = gimple_omp_for_kind (for_stmt) == GF_OMP_FOR_KIND_SIMD; |
629b3d75 MJ |
192 | bool distribute = gimple_omp_for_kind (for_stmt) |
193 | == GF_OMP_FOR_KIND_DISTRIBUTE; | |
194 | bool taskloop = gimple_omp_for_kind (for_stmt) | |
195 | == GF_OMP_FOR_KIND_TASKLOOP; | |
196 | tree iterv, countv; | |
197 | ||
198 | fd->for_stmt = for_stmt; | |
199 | fd->pre = NULL; | |
629b3d75 MJ |
200 | fd->have_nowait = distribute || simd; |
201 | fd->have_ordered = false; | |
28567c40 | 202 | fd->have_reductemp = false; |
8221c30b | 203 | fd->have_pointer_condtemp = false; |
2f6bb511 JJ |
204 | fd->have_scantemp = false; |
205 | fd->have_nonctrl_scantemp = false; | |
1160ec9a | 206 | fd->non_rect = false; |
6c7ae8c5 | 207 | fd->lastprivate_conditional = 0; |
02889d23 | 208 | fd->tiling = NULL_TREE; |
629b3d75 MJ |
209 | fd->collapse = 1; |
210 | fd->ordered = 0; | |
aed3ab25 JJ |
211 | fd->first_nonrect = -1; |
212 | fd->last_nonrect = -1; | |
629b3d75 MJ |
213 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC; |
214 | fd->sched_modifiers = 0; | |
215 | fd->chunk_size = NULL_TREE; | |
216 | fd->simd_schedule = false; | |
79c12969 | 217 | fd->first_inner_iterations = NULL_TREE; |
5acef69f | 218 | fd->factor = NULL_TREE; |
f418bd4b | 219 | fd->adjn1 = NULL_TREE; |
629b3d75 MJ |
220 | collapse_iter = NULL; |
221 | collapse_count = NULL; | |
222 | ||
223 | for (t = gimple_omp_for_clauses (for_stmt); t ; t = OMP_CLAUSE_CHAIN (t)) | |
224 | switch (OMP_CLAUSE_CODE (t)) | |
225 | { | |
226 | case OMP_CLAUSE_NOWAIT: | |
227 | fd->have_nowait = true; | |
228 | break; | |
229 | case OMP_CLAUSE_ORDERED: | |
230 | fd->have_ordered = true; | |
231 | if (OMP_CLAUSE_ORDERED_EXPR (t)) | |
232 | fd->ordered = tree_to_shwi (OMP_CLAUSE_ORDERED_EXPR (t)); | |
233 | break; | |
234 | case OMP_CLAUSE_SCHEDULE: | |
235 | gcc_assert (!distribute && !taskloop); | |
236 | fd->sched_kind | |
237 | = (enum omp_clause_schedule_kind) | |
238 | (OMP_CLAUSE_SCHEDULE_KIND (t) & OMP_CLAUSE_SCHEDULE_MASK); | |
239 | fd->sched_modifiers = (OMP_CLAUSE_SCHEDULE_KIND (t) | |
240 | & ~OMP_CLAUSE_SCHEDULE_MASK); | |
241 | fd->chunk_size = OMP_CLAUSE_SCHEDULE_CHUNK_EXPR (t); | |
242 | fd->simd_schedule = OMP_CLAUSE_SCHEDULE_SIMD (t); | |
243 | break; | |
244 | case OMP_CLAUSE_DIST_SCHEDULE: | |
245 | gcc_assert (distribute); | |
246 | fd->chunk_size = OMP_CLAUSE_DIST_SCHEDULE_CHUNK_EXPR (t); | |
247 | break; | |
248 | case OMP_CLAUSE_COLLAPSE: | |
249 | fd->collapse = tree_to_shwi (OMP_CLAUSE_COLLAPSE_EXPR (t)); | |
250 | if (fd->collapse > 1) | |
251 | { | |
252 | collapse_iter = &OMP_CLAUSE_COLLAPSE_ITERVAR (t); | |
253 | collapse_count = &OMP_CLAUSE_COLLAPSE_COUNT (t); | |
254 | } | |
255 | break; | |
02889d23 CLT |
256 | case OMP_CLAUSE_TILE: |
257 | fd->tiling = OMP_CLAUSE_TILE_LIST (t); | |
258 | fd->collapse = list_length (fd->tiling); | |
259 | gcc_assert (fd->collapse); | |
260 | collapse_iter = &OMP_CLAUSE_TILE_ITERVAR (t); | |
261 | collapse_count = &OMP_CLAUSE_TILE_COUNT (t); | |
262 | break; | |
28567c40 JJ |
263 | case OMP_CLAUSE__REDUCTEMP_: |
264 | fd->have_reductemp = true; | |
6c7ae8c5 JJ |
265 | break; |
266 | case OMP_CLAUSE_LASTPRIVATE: | |
267 | if (OMP_CLAUSE_LASTPRIVATE_CONDITIONAL (t)) | |
268 | fd->lastprivate_conditional++; | |
269 | break; | |
8221c30b JJ |
270 | case OMP_CLAUSE__CONDTEMP_: |
271 | if (POINTER_TYPE_P (TREE_TYPE (OMP_CLAUSE_DECL (t)))) | |
272 | fd->have_pointer_condtemp = true; | |
273 | break; | |
2f6bb511 JJ |
274 | case OMP_CLAUSE__SCANTEMP_: |
275 | fd->have_scantemp = true; | |
276 | if (!OMP_CLAUSE__SCANTEMP__ALLOC (t) | |
277 | && !OMP_CLAUSE__SCANTEMP__CONTROL (t)) | |
278 | fd->have_nonctrl_scantemp = true; | |
279 | break; | |
629b3d75 MJ |
280 | default: |
281 | break; | |
282 | } | |
02889d23 CLT |
283 | |
284 | if (fd->collapse > 1 || fd->tiling) | |
285 | fd->loops = loops; | |
286 | else | |
287 | fd->loops = &fd->loop; | |
288 | ||
629b3d75 MJ |
289 | if (fd->ordered && fd->collapse == 1 && loops != NULL) |
290 | { | |
291 | fd->loops = loops; | |
292 | iterv = NULL_TREE; | |
293 | countv = NULL_TREE; | |
294 | collapse_iter = &iterv; | |
295 | collapse_count = &countv; | |
296 | } | |
297 | ||
298 | /* FIXME: for now map schedule(auto) to schedule(static). | |
299 | There should be analysis to determine whether all iterations | |
300 | are approximately the same amount of work (then schedule(static) | |
301 | is best) or if it varies (then schedule(dynamic,N) is better). */ | |
302 | if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_AUTO) | |
303 | { | |
304 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC; | |
305 | gcc_assert (fd->chunk_size == NULL); | |
306 | } | |
02889d23 | 307 | gcc_assert ((fd->collapse == 1 && !fd->tiling) || collapse_iter != NULL); |
629b3d75 MJ |
308 | if (taskloop) |
309 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_RUNTIME; | |
310 | if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_RUNTIME) | |
311 | gcc_assert (fd->chunk_size == NULL); | |
312 | else if (fd->chunk_size == NULL) | |
313 | { | |
314 | /* We only need to compute a default chunk size for ordered | |
315 | static loops and dynamic loops. */ | |
316 | if (fd->sched_kind != OMP_CLAUSE_SCHEDULE_STATIC | |
317 | || fd->have_ordered) | |
318 | fd->chunk_size = (fd->sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) | |
319 | ? integer_zero_node : integer_one_node; | |
320 | } | |
321 | ||
322 | int cnt = fd->ordered ? fd->ordered : fd->collapse; | |
c154b8bc JJ |
323 | int single_nonrect = -1; |
324 | tree single_nonrect_count = NULL_TREE; | |
325 | enum tree_code single_nonrect_cond_code = ERROR_MARK; | |
326 | for (i = 1; i < cnt; i++) | |
327 | { | |
328 | tree n1 = gimple_omp_for_initial (for_stmt, i); | |
329 | tree n2 = gimple_omp_for_final (for_stmt, i); | |
330 | if (TREE_CODE (n1) == TREE_VEC) | |
331 | { | |
332 | if (fd->non_rect) | |
333 | { | |
334 | single_nonrect = -1; | |
335 | break; | |
336 | } | |
337 | for (int j = i - 1; j >= 0; j--) | |
338 | if (TREE_VEC_ELT (n1, 0) == gimple_omp_for_index (for_stmt, j)) | |
339 | { | |
340 | single_nonrect = j; | |
341 | break; | |
342 | } | |
343 | fd->non_rect = true; | |
344 | } | |
345 | else if (TREE_CODE (n2) == TREE_VEC) | |
346 | { | |
347 | if (fd->non_rect) | |
348 | { | |
349 | single_nonrect = -1; | |
350 | break; | |
351 | } | |
352 | for (int j = i - 1; j >= 0; j--) | |
353 | if (TREE_VEC_ELT (n2, 0) == gimple_omp_for_index (for_stmt, j)) | |
354 | { | |
355 | single_nonrect = j; | |
356 | break; | |
357 | } | |
358 | fd->non_rect = true; | |
359 | } | |
360 | } | |
629b3d75 MJ |
361 | for (i = 0; i < cnt; i++) |
362 | { | |
02889d23 CLT |
363 | if (i == 0 |
364 | && fd->collapse == 1 | |
365 | && !fd->tiling | |
366 | && (fd->ordered == 0 || loops == NULL)) | |
629b3d75 MJ |
367 | loop = &fd->loop; |
368 | else if (loops != NULL) | |
369 | loop = loops + i; | |
370 | else | |
371 | loop = &dummy_loop; | |
372 | ||
373 | loop->v = gimple_omp_for_index (for_stmt, i); | |
374 | gcc_assert (SSA_VAR_P (loop->v)); | |
375 | gcc_assert (TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE | |
376 | || TREE_CODE (TREE_TYPE (loop->v)) == POINTER_TYPE); | |
377 | var = TREE_CODE (loop->v) == SSA_NAME ? SSA_NAME_VAR (loop->v) : loop->v; | |
378 | loop->n1 = gimple_omp_for_initial (for_stmt, i); | |
1160ec9a JJ |
379 | loop->m1 = NULL_TREE; |
380 | loop->m2 = NULL_TREE; | |
381 | loop->outer = 0; | |
aed3ab25 | 382 | loop->non_rect_referenced = false; |
1160ec9a JJ |
383 | if (TREE_CODE (loop->n1) == TREE_VEC) |
384 | { | |
385 | for (int j = i - 1; j >= 0; j--) | |
386 | if (TREE_VEC_ELT (loop->n1, 0) == gimple_omp_for_index (for_stmt, j)) | |
387 | { | |
388 | loop->outer = i - j; | |
aed3ab25 JJ |
389 | if (loops != NULL) |
390 | loops[j].non_rect_referenced = true; | |
391 | if (fd->first_nonrect == -1 || fd->first_nonrect > j) | |
392 | fd->first_nonrect = j; | |
1160ec9a JJ |
393 | break; |
394 | } | |
395 | gcc_assert (loop->outer); | |
396 | loop->m1 = TREE_VEC_ELT (loop->n1, 1); | |
397 | loop->n1 = TREE_VEC_ELT (loop->n1, 2); | |
398 | fd->non_rect = true; | |
aed3ab25 | 399 | fd->last_nonrect = i; |
1160ec9a | 400 | } |
629b3d75 MJ |
401 | |
402 | loop->cond_code = gimple_omp_for_cond (for_stmt, i); | |
403 | loop->n2 = gimple_omp_for_final (for_stmt, i); | |
28567c40 JJ |
404 | gcc_assert (loop->cond_code != NE_EXPR |
405 | || (gimple_omp_for_kind (for_stmt) | |
406 | != GF_OMP_FOR_KIND_OACC_LOOP)); | |
1160ec9a JJ |
407 | if (TREE_CODE (loop->n2) == TREE_VEC) |
408 | { | |
409 | if (loop->outer) | |
410 | gcc_assert (TREE_VEC_ELT (loop->n2, 0) | |
411 | == gimple_omp_for_index (for_stmt, i - loop->outer)); | |
412 | else | |
413 | for (int j = i - 1; j >= 0; j--) | |
414 | if (TREE_VEC_ELT (loop->n2, 0) == gimple_omp_for_index (for_stmt, j)) | |
415 | { | |
416 | loop->outer = i - j; | |
aed3ab25 JJ |
417 | if (loops != NULL) |
418 | loops[j].non_rect_referenced = true; | |
419 | if (fd->first_nonrect == -1 || fd->first_nonrect > j) | |
420 | fd->first_nonrect = j; | |
1160ec9a JJ |
421 | break; |
422 | } | |
423 | gcc_assert (loop->outer); | |
424 | loop->m2 = TREE_VEC_ELT (loop->n2, 1); | |
425 | loop->n2 = TREE_VEC_ELT (loop->n2, 2); | |
426 | fd->non_rect = true; | |
aed3ab25 | 427 | fd->last_nonrect = i; |
1160ec9a | 428 | } |
629b3d75 MJ |
429 | |
430 | t = gimple_omp_for_incr (for_stmt, i); | |
431 | gcc_assert (TREE_OPERAND (t, 0) == var); | |
432 | loop->step = omp_get_for_step_from_incr (loc, t); | |
433 | ||
031c5c8b MJ |
434 | omp_adjust_for_condition (loc, &loop->cond_code, &loop->n2, loop->v, |
435 | loop->step); | |
28567c40 | 436 | |
629b3d75 MJ |
437 | if (simd |
438 | || (fd->sched_kind == OMP_CLAUSE_SCHEDULE_STATIC | |
439 | && !fd->have_ordered)) | |
440 | { | |
02889d23 | 441 | if (fd->collapse == 1 && !fd->tiling) |
629b3d75 MJ |
442 | iter_type = TREE_TYPE (loop->v); |
443 | else if (i == 0 | |
444 | || TYPE_PRECISION (iter_type) | |
445 | < TYPE_PRECISION (TREE_TYPE (loop->v))) | |
446 | iter_type | |
447 | = build_nonstandard_integer_type | |
448 | (TYPE_PRECISION (TREE_TYPE (loop->v)), 1); | |
449 | } | |
450 | else if (iter_type != long_long_unsigned_type_node) | |
451 | { | |
452 | if (POINTER_TYPE_P (TREE_TYPE (loop->v))) | |
453 | iter_type = long_long_unsigned_type_node; | |
454 | else if (TYPE_UNSIGNED (TREE_TYPE (loop->v)) | |
455 | && TYPE_PRECISION (TREE_TYPE (loop->v)) | |
456 | >= TYPE_PRECISION (iter_type)) | |
457 | { | |
458 | tree n; | |
459 | ||
460 | if (loop->cond_code == LT_EXPR) | |
28567c40 JJ |
461 | n = fold_build2_loc (loc, PLUS_EXPR, TREE_TYPE (loop->v), |
462 | loop->n2, loop->step); | |
629b3d75 MJ |
463 | else |
464 | n = loop->n1; | |
2e47c8c6 JJ |
465 | if (loop->m1 |
466 | || loop->m2 | |
467 | || TREE_CODE (n) != INTEGER_CST | |
629b3d75 MJ |
468 | || tree_int_cst_lt (TYPE_MAX_VALUE (iter_type), n)) |
469 | iter_type = long_long_unsigned_type_node; | |
470 | } | |
471 | else if (TYPE_PRECISION (TREE_TYPE (loop->v)) | |
472 | > TYPE_PRECISION (iter_type)) | |
473 | { | |
474 | tree n1, n2; | |
475 | ||
476 | if (loop->cond_code == LT_EXPR) | |
477 | { | |
478 | n1 = loop->n1; | |
28567c40 JJ |
479 | n2 = fold_build2_loc (loc, PLUS_EXPR, TREE_TYPE (loop->v), |
480 | loop->n2, loop->step); | |
629b3d75 MJ |
481 | } |
482 | else | |
483 | { | |
28567c40 JJ |
484 | n1 = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (loop->v), |
485 | loop->n2, loop->step); | |
629b3d75 MJ |
486 | n2 = loop->n1; |
487 | } | |
2e47c8c6 JJ |
488 | if (loop->m1 |
489 | || loop->m2 | |
490 | || TREE_CODE (n1) != INTEGER_CST | |
629b3d75 MJ |
491 | || TREE_CODE (n2) != INTEGER_CST |
492 | || !tree_int_cst_lt (TYPE_MIN_VALUE (iter_type), n1) | |
493 | || !tree_int_cst_lt (n2, TYPE_MAX_VALUE (iter_type))) | |
494 | iter_type = long_long_unsigned_type_node; | |
495 | } | |
496 | } | |
497 | ||
498 | if (i >= fd->collapse) | |
499 | continue; | |
500 | ||
501 | if (collapse_count && *collapse_count == NULL) | |
502 | { | |
c154b8bc JJ |
503 | if (count && integer_zerop (count)) |
504 | continue; | |
505 | tree n1first = NULL_TREE, n2first = NULL_TREE; | |
506 | tree n1last = NULL_TREE, n2last = NULL_TREE; | |
507 | tree ostep = NULL_TREE; | |
1160ec9a | 508 | if (loop->m1 || loop->m2) |
c154b8bc JJ |
509 | { |
510 | if (count == NULL_TREE) | |
511 | continue; | |
512 | if (single_nonrect == -1 | |
513 | || (loop->m1 && TREE_CODE (loop->m1) != INTEGER_CST) | |
f418bd4b JJ |
514 | || (loop->m2 && TREE_CODE (loop->m2) != INTEGER_CST) |
515 | || TREE_CODE (loop->n1) != INTEGER_CST | |
516 | || TREE_CODE (loop->n2) != INTEGER_CST | |
517 | || TREE_CODE (loop->step) != INTEGER_CST) | |
c154b8bc JJ |
518 | { |
519 | count = NULL_TREE; | |
520 | continue; | |
521 | } | |
522 | tree var = gimple_omp_for_initial (for_stmt, single_nonrect); | |
523 | tree itype = TREE_TYPE (var); | |
524 | tree first = gimple_omp_for_initial (for_stmt, single_nonrect); | |
525 | t = gimple_omp_for_incr (for_stmt, single_nonrect); | |
526 | ostep = omp_get_for_step_from_incr (loc, t); | |
527 | t = fold_binary (MINUS_EXPR, long_long_unsigned_type_node, | |
528 | single_nonrect_count, | |
529 | build_one_cst (long_long_unsigned_type_node)); | |
530 | t = fold_convert (itype, t); | |
531 | first = fold_convert (itype, first); | |
532 | ostep = fold_convert (itype, ostep); | |
533 | tree last = fold_binary (PLUS_EXPR, itype, first, | |
534 | fold_binary (MULT_EXPR, itype, t, | |
535 | ostep)); | |
536 | if (TREE_CODE (first) != INTEGER_CST | |
537 | || TREE_CODE (last) != INTEGER_CST) | |
538 | { | |
539 | count = NULL_TREE; | |
540 | continue; | |
541 | } | |
542 | if (loop->m1) | |
543 | { | |
544 | tree m1 = fold_convert (itype, loop->m1); | |
545 | tree n1 = fold_convert (itype, loop->n1); | |
546 | n1first = fold_binary (PLUS_EXPR, itype, | |
547 | fold_binary (MULT_EXPR, itype, | |
548 | first, m1), n1); | |
549 | n1last = fold_binary (PLUS_EXPR, itype, | |
550 | fold_binary (MULT_EXPR, itype, | |
551 | last, m1), n1); | |
552 | } | |
553 | else | |
554 | n1first = n1last = loop->n1; | |
555 | if (loop->m2) | |
556 | { | |
557 | tree n2 = fold_convert (itype, loop->n2); | |
558 | tree m2 = fold_convert (itype, loop->m2); | |
559 | n2first = fold_binary (PLUS_EXPR, itype, | |
560 | fold_binary (MULT_EXPR, itype, | |
561 | first, m2), n2); | |
562 | n2last = fold_binary (PLUS_EXPR, itype, | |
563 | fold_binary (MULT_EXPR, itype, | |
564 | last, m2), n2); | |
565 | } | |
566 | else | |
567 | n2first = n2last = loop->n2; | |
568 | n1first = fold_convert (TREE_TYPE (loop->v), n1first); | |
569 | n2first = fold_convert (TREE_TYPE (loop->v), n2first); | |
570 | n1last = fold_convert (TREE_TYPE (loop->v), n1last); | |
571 | n2last = fold_convert (TREE_TYPE (loop->v), n2last); | |
572 | t = fold_binary (loop->cond_code, boolean_type_node, | |
573 | n1first, n2first); | |
574 | tree t2 = fold_binary (loop->cond_code, boolean_type_node, | |
575 | n1last, n2last); | |
576 | if (t && t2 && integer_nonzerop (t) && integer_nonzerop (t2)) | |
577 | /* All outer loop iterators have at least one inner loop | |
578 | iteration. Try to compute the count at compile time. */ | |
579 | t = NULL_TREE; | |
580 | else if (t && t2 && integer_zerop (t) && integer_zerop (t2)) | |
581 | /* No iterations of the inner loop. count will be set to | |
582 | zero cst below. */; | |
f418bd4b JJ |
583 | else if (TYPE_UNSIGNED (itype) |
584 | || t == NULL_TREE | |
585 | || t2 == NULL_TREE | |
586 | || TREE_CODE (t) != INTEGER_CST | |
587 | || TREE_CODE (t2) != INTEGER_CST) | |
c154b8bc JJ |
588 | { |
589 | /* Punt (for now). */ | |
590 | count = NULL_TREE; | |
591 | continue; | |
592 | } | |
f418bd4b JJ |
593 | else |
594 | { | |
595 | /* Some iterations of the outer loop have zero iterations | |
596 | of the inner loop, while others have at least one. | |
597 | In this case, we need to adjust one of those outer | |
598 | loop bounds. If ADJ_FIRST, we need to adjust outer n1 | |
599 | (first), otherwise outer n2 (last). */ | |
600 | bool adj_first = integer_zerop (t); | |
601 | tree n1 = fold_convert (itype, loop->n1); | |
602 | tree n2 = fold_convert (itype, loop->n2); | |
603 | tree m1 = loop->m1 ? fold_convert (itype, loop->m1) | |
604 | : build_zero_cst (itype); | |
605 | tree m2 = loop->m2 ? fold_convert (itype, loop->m2) | |
606 | : build_zero_cst (itype); | |
607 | t = fold_binary (MINUS_EXPR, itype, n1, n2); | |
608 | t2 = fold_binary (MINUS_EXPR, itype, m2, m1); | |
609 | t = fold_binary (TRUNC_DIV_EXPR, itype, t, t2); | |
610 | t2 = fold_binary (MINUS_EXPR, itype, t, first); | |
611 | t2 = fold_binary (TRUNC_MOD_EXPR, itype, t2, ostep); | |
612 | t = fold_binary (MINUS_EXPR, itype, t, t2); | |
613 | tree n1cur | |
614 | = fold_binary (PLUS_EXPR, itype, n1, | |
615 | fold_binary (MULT_EXPR, itype, m1, t)); | |
616 | tree n2cur | |
617 | = fold_binary (PLUS_EXPR, itype, n2, | |
618 | fold_binary (MULT_EXPR, itype, m2, t)); | |
619 | t2 = fold_binary (loop->cond_code, boolean_type_node, | |
620 | n1cur, n2cur); | |
621 | tree t3 = fold_binary (MULT_EXPR, itype, m1, ostep); | |
622 | tree t4 = fold_binary (MULT_EXPR, itype, m2, ostep); | |
623 | tree diff; | |
624 | if (adj_first) | |
625 | { | |
626 | tree new_first; | |
627 | if (integer_nonzerop (t2)) | |
628 | { | |
629 | new_first = t; | |
630 | n1first = n1cur; | |
631 | n2first = n2cur; | |
632 | if (flag_checking) | |
633 | { | |
634 | t3 = fold_binary (MINUS_EXPR, itype, n1cur, t3); | |
635 | t4 = fold_binary (MINUS_EXPR, itype, n2cur, t4); | |
636 | t3 = fold_binary (loop->cond_code, | |
637 | boolean_type_node, t3, t4); | |
638 | gcc_assert (integer_zerop (t3)); | |
639 | } | |
640 | } | |
641 | else | |
642 | { | |
643 | t3 = fold_binary (PLUS_EXPR, itype, n1cur, t3); | |
644 | t4 = fold_binary (PLUS_EXPR, itype, n2cur, t4); | |
645 | new_first = fold_binary (PLUS_EXPR, itype, t, ostep); | |
646 | n1first = t3; | |
647 | n2first = t4; | |
648 | if (flag_checking) | |
649 | { | |
650 | t3 = fold_binary (loop->cond_code, | |
651 | boolean_type_node, t3, t4); | |
652 | gcc_assert (integer_nonzerop (t3)); | |
653 | } | |
654 | } | |
655 | diff = fold_binary (MINUS_EXPR, itype, new_first, first); | |
656 | first = new_first; | |
657 | fd->adjn1 = first; | |
658 | } | |
659 | else | |
660 | { | |
661 | tree new_last; | |
662 | if (integer_zerop (t2)) | |
663 | { | |
664 | t3 = fold_binary (MINUS_EXPR, itype, n1cur, t3); | |
665 | t4 = fold_binary (MINUS_EXPR, itype, n2cur, t4); | |
666 | new_last = fold_binary (MINUS_EXPR, itype, t, ostep); | |
667 | n1last = t3; | |
668 | n2last = t4; | |
669 | if (flag_checking) | |
670 | { | |
671 | t3 = fold_binary (loop->cond_code, | |
672 | boolean_type_node, t3, t4); | |
673 | gcc_assert (integer_nonzerop (t3)); | |
674 | } | |
675 | } | |
676 | else | |
677 | { | |
678 | new_last = t; | |
679 | n1last = n1cur; | |
680 | n2last = n2cur; | |
681 | if (flag_checking) | |
682 | { | |
683 | t3 = fold_binary (PLUS_EXPR, itype, n1cur, t3); | |
684 | t4 = fold_binary (PLUS_EXPR, itype, n2cur, t4); | |
685 | t3 = fold_binary (loop->cond_code, | |
686 | boolean_type_node, t3, t4); | |
687 | gcc_assert (integer_zerop (t3)); | |
688 | } | |
689 | } | |
690 | diff = fold_binary (MINUS_EXPR, itype, last, new_last); | |
691 | } | |
692 | if (TYPE_UNSIGNED (itype) | |
693 | && single_nonrect_cond_code == GT_EXPR) | |
694 | diff = fold_binary (TRUNC_DIV_EXPR, itype, | |
695 | fold_unary (NEGATE_EXPR, itype, diff), | |
696 | fold_unary (NEGATE_EXPR, itype, | |
697 | ostep)); | |
698 | else | |
699 | diff = fold_binary (TRUNC_DIV_EXPR, itype, diff, ostep); | |
700 | diff = fold_convert (long_long_unsigned_type_node, diff); | |
701 | single_nonrect_count | |
702 | = fold_binary (MINUS_EXPR, long_long_unsigned_type_node, | |
703 | single_nonrect_count, diff); | |
704 | t = NULL_TREE; | |
705 | } | |
c154b8bc | 706 | } |
1160ec9a JJ |
707 | else |
708 | t = fold_binary (loop->cond_code, boolean_type_node, | |
709 | fold_convert (TREE_TYPE (loop->v), loop->n1), | |
710 | fold_convert (TREE_TYPE (loop->v), loop->n2)); | |
629b3d75 MJ |
711 | if (t && integer_zerop (t)) |
712 | count = build_zero_cst (long_long_unsigned_type_node); | |
713 | else if ((i == 0 || count != NULL_TREE) | |
714 | && TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE | |
715 | && TREE_CONSTANT (loop->n1) | |
716 | && TREE_CONSTANT (loop->n2) | |
717 | && TREE_CODE (loop->step) == INTEGER_CST) | |
718 | { | |
719 | tree itype = TREE_TYPE (loop->v); | |
720 | ||
721 | if (POINTER_TYPE_P (itype)) | |
722 | itype = signed_type_for (itype); | |
723 | t = build_int_cst (itype, (loop->cond_code == LT_EXPR ? -1 : 1)); | |
c154b8bc JJ |
724 | t = fold_build2 (PLUS_EXPR, itype, |
725 | fold_convert (itype, loop->step), t); | |
726 | tree n1 = loop->n1; | |
727 | tree n2 = loop->n2; | |
728 | if (loop->m1 || loop->m2) | |
28567c40 | 729 | { |
c154b8bc | 730 | gcc_assert (single_nonrect != -1); |
79c12969 JJ |
731 | n1 = n1first; |
732 | n2 = n2first; | |
28567c40 | 733 | } |
c154b8bc JJ |
734 | t = fold_build2 (PLUS_EXPR, itype, t, fold_convert (itype, n2)); |
735 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
736 | tree step = fold_convert_loc (loc, itype, loop->step); | |
737 | if (TYPE_UNSIGNED (itype) && loop->cond_code == GT_EXPR) | |
738 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
739 | fold_build1 (NEGATE_EXPR, itype, t), | |
740 | fold_build1 (NEGATE_EXPR, itype, step)); | |
629b3d75 | 741 | else |
c154b8bc JJ |
742 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
743 | tree llutype = long_long_unsigned_type_node; | |
744 | t = fold_convert (llutype, t); | |
745 | if (loop->m1 || loop->m2) | |
746 | { | |
747 | /* t is number of iterations of inner loop at either first | |
748 | or last value of the outer iterator (the one with fewer | |
749 | iterations). | |
750 | Compute t2 = ((m2 - m1) * ostep) / step | |
c154b8bc JJ |
751 | and niters = outer_count * t |
752 | + t2 * ((outer_count - 1) * outer_count / 2) | |
753 | */ | |
754 | tree m1 = loop->m1 ? loop->m1 : integer_zero_node; | |
755 | tree m2 = loop->m2 ? loop->m2 : integer_zero_node; | |
756 | m1 = fold_convert (itype, m1); | |
757 | m2 = fold_convert (itype, m2); | |
79c12969 | 758 | tree t2 = fold_build2 (MINUS_EXPR, itype, m2, m1); |
c154b8bc JJ |
759 | t2 = fold_build2 (MULT_EXPR, itype, t2, ostep); |
760 | if (TYPE_UNSIGNED (itype) && loop->cond_code == GT_EXPR) | |
761 | t2 = fold_build2 (TRUNC_DIV_EXPR, itype, | |
762 | fold_build1 (NEGATE_EXPR, itype, t2), | |
763 | fold_build1 (NEGATE_EXPR, itype, step)); | |
764 | else | |
765 | t2 = fold_build2 (TRUNC_DIV_EXPR, itype, t2, step); | |
766 | t2 = fold_convert (llutype, t2); | |
79c12969 | 767 | fd->first_inner_iterations = t; |
5acef69f | 768 | fd->factor = t2; |
c154b8bc JJ |
769 | t = fold_build2 (MULT_EXPR, llutype, t, |
770 | single_nonrect_count); | |
771 | tree t3 = fold_build2 (MINUS_EXPR, llutype, | |
772 | single_nonrect_count, | |
773 | build_one_cst (llutype)); | |
774 | t3 = fold_build2 (MULT_EXPR, llutype, t3, | |
775 | single_nonrect_count); | |
776 | t3 = fold_build2 (TRUNC_DIV_EXPR, llutype, t3, | |
777 | build_int_cst (llutype, 2)); | |
778 | t2 = fold_build2 (MULT_EXPR, llutype, t2, t3); | |
779 | t = fold_build2 (PLUS_EXPR, llutype, t, t2); | |
780 | } | |
781 | if (i == single_nonrect) | |
782 | { | |
783 | if (integer_zerop (t) || TREE_CODE (t) != INTEGER_CST) | |
784 | count = t; | |
785 | else | |
786 | { | |
787 | single_nonrect_count = t; | |
788 | single_nonrect_cond_code = loop->cond_code; | |
789 | if (count == NULL_TREE) | |
790 | count = build_one_cst (llutype); | |
791 | } | |
792 | } | |
793 | else if (count != NULL_TREE) | |
794 | count = fold_build2 (MULT_EXPR, llutype, count, t); | |
629b3d75 MJ |
795 | else |
796 | count = t; | |
797 | if (TREE_CODE (count) != INTEGER_CST) | |
798 | count = NULL_TREE; | |
799 | } | |
800 | else if (count && !integer_zerop (count)) | |
801 | count = NULL_TREE; | |
802 | } | |
803 | } | |
804 | ||
805 | if (count | |
806 | && !simd | |
807 | && (fd->sched_kind != OMP_CLAUSE_SCHEDULE_STATIC | |
808 | || fd->have_ordered)) | |
809 | { | |
810 | if (!tree_int_cst_lt (count, TYPE_MAX_VALUE (long_integer_type_node))) | |
811 | iter_type = long_long_unsigned_type_node; | |
812 | else | |
813 | iter_type = long_integer_type_node; | |
814 | } | |
815 | else if (collapse_iter && *collapse_iter != NULL) | |
816 | iter_type = TREE_TYPE (*collapse_iter); | |
817 | fd->iter_type = iter_type; | |
818 | if (collapse_iter && *collapse_iter == NULL) | |
819 | *collapse_iter = create_tmp_var (iter_type, ".iter"); | |
820 | if (collapse_count && *collapse_count == NULL) | |
821 | { | |
822 | if (count) | |
5acef69f JJ |
823 | { |
824 | *collapse_count = fold_convert_loc (loc, iter_type, count); | |
79c12969 | 825 | if (fd->first_inner_iterations && fd->factor) |
5acef69f | 826 | { |
f418bd4b | 827 | t = make_tree_vec (4); |
5acef69f | 828 | TREE_VEC_ELT (t, 0) = *collapse_count; |
79c12969 | 829 | TREE_VEC_ELT (t, 1) = fd->first_inner_iterations; |
5acef69f | 830 | TREE_VEC_ELT (t, 2) = fd->factor; |
f418bd4b | 831 | TREE_VEC_ELT (t, 3) = fd->adjn1; |
5acef69f JJ |
832 | *collapse_count = t; |
833 | } | |
834 | } | |
629b3d75 MJ |
835 | else |
836 | *collapse_count = create_tmp_var (iter_type, ".count"); | |
837 | } | |
838 | ||
02889d23 | 839 | if (fd->collapse > 1 || fd->tiling || (fd->ordered && loops)) |
629b3d75 MJ |
840 | { |
841 | fd->loop.v = *collapse_iter; | |
842 | fd->loop.n1 = build_int_cst (TREE_TYPE (fd->loop.v), 0); | |
843 | fd->loop.n2 = *collapse_count; | |
5acef69f JJ |
844 | if (TREE_CODE (fd->loop.n2) == TREE_VEC) |
845 | { | |
846 | gcc_assert (fd->non_rect); | |
79c12969 | 847 | fd->first_inner_iterations = TREE_VEC_ELT (fd->loop.n2, 1); |
5acef69f | 848 | fd->factor = TREE_VEC_ELT (fd->loop.n2, 2); |
f418bd4b | 849 | fd->adjn1 = TREE_VEC_ELT (fd->loop.n2, 3); |
5acef69f JJ |
850 | fd->loop.n2 = TREE_VEC_ELT (fd->loop.n2, 0); |
851 | } | |
629b3d75 | 852 | fd->loop.step = build_int_cst (TREE_TYPE (fd->loop.v), 1); |
1160ec9a JJ |
853 | fd->loop.m1 = NULL_TREE; |
854 | fd->loop.m2 = NULL_TREE; | |
855 | fd->loop.outer = 0; | |
629b3d75 MJ |
856 | fd->loop.cond_code = LT_EXPR; |
857 | } | |
858 | else if (loops) | |
859 | loops[0] = fd->loop; | |
860 | } | |
861 | ||
862 | /* Build a call to GOMP_barrier. */ | |
863 | ||
864 | gimple * | |
865 | omp_build_barrier (tree lhs) | |
866 | { | |
867 | tree fndecl = builtin_decl_explicit (lhs ? BUILT_IN_GOMP_BARRIER_CANCEL | |
868 | : BUILT_IN_GOMP_BARRIER); | |
869 | gcall *g = gimple_build_call (fndecl, 0); | |
870 | if (lhs) | |
871 | gimple_call_set_lhs (g, lhs); | |
872 | return g; | |
873 | } | |
874 | ||
f1f862ae JJ |
875 | /* Find OMP_FOR resp. OMP_SIMD with non-NULL OMP_FOR_INIT. Also, fill in pdata |
876 | array, pdata[0] non-NULL if there is anything non-trivial in between, | |
877 | pdata[1] is address of OMP_PARALLEL in between if any, pdata[2] is address | |
878 | of OMP_FOR in between if any and pdata[3] is address of the inner | |
879 | OMP_FOR/OMP_SIMD. */ | |
880 | ||
881 | tree | |
882 | find_combined_omp_for (tree *tp, int *walk_subtrees, void *data) | |
883 | { | |
884 | tree **pdata = (tree **) data; | |
885 | *walk_subtrees = 0; | |
886 | switch (TREE_CODE (*tp)) | |
887 | { | |
888 | case OMP_FOR: | |
889 | if (OMP_FOR_INIT (*tp) != NULL_TREE) | |
890 | { | |
891 | pdata[3] = tp; | |
892 | return *tp; | |
893 | } | |
894 | pdata[2] = tp; | |
895 | *walk_subtrees = 1; | |
896 | break; | |
897 | case OMP_SIMD: | |
898 | if (OMP_FOR_INIT (*tp) != NULL_TREE) | |
899 | { | |
900 | pdata[3] = tp; | |
901 | return *tp; | |
902 | } | |
903 | break; | |
904 | case BIND_EXPR: | |
905 | if (BIND_EXPR_VARS (*tp) | |
906 | || (BIND_EXPR_BLOCK (*tp) | |
907 | && BLOCK_VARS (BIND_EXPR_BLOCK (*tp)))) | |
908 | pdata[0] = tp; | |
909 | *walk_subtrees = 1; | |
910 | break; | |
911 | case STATEMENT_LIST: | |
912 | if (!tsi_one_before_end_p (tsi_start (*tp))) | |
913 | pdata[0] = tp; | |
914 | *walk_subtrees = 1; | |
915 | break; | |
916 | case TRY_FINALLY_EXPR: | |
917 | pdata[0] = tp; | |
918 | *walk_subtrees = 1; | |
919 | break; | |
920 | case OMP_PARALLEL: | |
921 | pdata[1] = tp; | |
922 | *walk_subtrees = 1; | |
923 | break; | |
924 | default: | |
925 | break; | |
926 | } | |
927 | return NULL_TREE; | |
928 | } | |
929 | ||
629b3d75 MJ |
930 | /* Return maximum possible vectorization factor for the target. */ |
931 | ||
9d2f08ab | 932 | poly_uint64 |
629b3d75 MJ |
933 | omp_max_vf (void) |
934 | { | |
935 | if (!optimize | |
936 | || optimize_debug | |
937 | || !flag_tree_loop_optimize | |
938 | || (!flag_tree_loop_vectorize | |
26d476cd | 939 | && global_options_set.x_flag_tree_loop_vectorize)) |
629b3d75 MJ |
940 | return 1; |
941 | ||
e021fb86 RS |
942 | auto_vector_modes modes; |
943 | targetm.vectorize.autovectorize_vector_modes (&modes, true); | |
944 | if (!modes.is_empty ()) | |
629b3d75 | 945 | { |
86e36728 | 946 | poly_uint64 vf = 0; |
e021fb86 RS |
947 | for (unsigned int i = 0; i < modes.length (); ++i) |
948 | /* The returned modes use the smallest element size (and thus | |
949 | the largest nunits) for the vectorization approach that they | |
950 | represent. */ | |
951 | vf = ordered_max (vf, GET_MODE_NUNITS (modes[i])); | |
86e36728 | 952 | return vf; |
629b3d75 | 953 | } |
86e36728 RS |
954 | |
955 | machine_mode vqimode = targetm.vectorize.preferred_simd_mode (QImode); | |
956 | if (GET_MODE_CLASS (vqimode) == MODE_VECTOR_INT) | |
957 | return GET_MODE_NUNITS (vqimode); | |
958 | ||
959 | return 1; | |
629b3d75 MJ |
960 | } |
961 | ||
962 | /* Return maximum SIMT width if offloading may target SIMT hardware. */ | |
963 | ||
964 | int | |
965 | omp_max_simt_vf (void) | |
966 | { | |
967 | if (!optimize) | |
968 | return 0; | |
969 | if (ENABLE_OFFLOADING) | |
01914336 | 970 | for (const char *c = getenv ("OFFLOAD_TARGET_NAMES"); c;) |
629b3d75 | 971 | { |
6ba3079d | 972 | if (startswith (c, "nvptx")) |
629b3d75 | 973 | return 32; |
9ba66bf5 | 974 | else if ((c = strchr (c, ':'))) |
629b3d75 MJ |
975 | c++; |
976 | } | |
977 | return 0; | |
978 | } | |
979 | ||
135df52c JJ |
980 | /* Store the construct selectors as tree codes from last to first, |
981 | return their number. */ | |
982 | ||
983 | int | |
984 | omp_constructor_traits_to_codes (tree ctx, enum tree_code *constructs) | |
985 | { | |
986 | int nconstructs = list_length (ctx); | |
987 | int i = nconstructs - 1; | |
988 | for (tree t2 = ctx; t2; t2 = TREE_CHAIN (t2), i--) | |
989 | { | |
990 | const char *sel = IDENTIFIER_POINTER (TREE_PURPOSE (t2)); | |
991 | if (!strcmp (sel, "target")) | |
992 | constructs[i] = OMP_TARGET; | |
993 | else if (!strcmp (sel, "teams")) | |
994 | constructs[i] = OMP_TEAMS; | |
995 | else if (!strcmp (sel, "parallel")) | |
996 | constructs[i] = OMP_PARALLEL; | |
997 | else if (!strcmp (sel, "for") || !strcmp (sel, "do")) | |
998 | constructs[i] = OMP_FOR; | |
999 | else if (!strcmp (sel, "simd")) | |
1000 | constructs[i] = OMP_SIMD; | |
1001 | else | |
1002 | gcc_unreachable (); | |
1003 | } | |
1004 | gcc_assert (i == -1); | |
1005 | return nconstructs; | |
1006 | } | |
1007 | ||
9ba66bf5 JJ |
1008 | /* Return true if PROP is possibly present in one of the offloading target's |
1009 | OpenMP contexts. The format of PROPS string is always offloading target's | |
1010 | name terminated by '\0', followed by properties for that offloading | |
1011 | target separated by '\0' and terminated by another '\0'. The strings | |
1012 | are created from omp-device-properties installed files of all configured | |
1013 | offloading targets. */ | |
1014 | ||
1015 | static bool | |
1016 | omp_offload_device_kind_arch_isa (const char *props, const char *prop) | |
1017 | { | |
1018 | const char *names = getenv ("OFFLOAD_TARGET_NAMES"); | |
1019 | if (names == NULL || *names == '\0') | |
1020 | return false; | |
1021 | while (*props != '\0') | |
1022 | { | |
1023 | size_t name_len = strlen (props); | |
1024 | bool matches = false; | |
1025 | for (const char *c = names; c; ) | |
1026 | { | |
1027 | if (strncmp (props, c, name_len) == 0 | |
1028 | && (c[name_len] == '\0' | |
1029 | || c[name_len] == ':' | |
1030 | || c[name_len] == '=')) | |
1031 | { | |
1032 | matches = true; | |
1033 | break; | |
1034 | } | |
1035 | else if ((c = strchr (c, ':'))) | |
1036 | c++; | |
1037 | } | |
1038 | props = props + name_len + 1; | |
1039 | while (*props != '\0') | |
1040 | { | |
1041 | if (matches && strcmp (props, prop) == 0) | |
1042 | return true; | |
1043 | props = strchr (props, '\0') + 1; | |
1044 | } | |
1045 | props++; | |
1046 | } | |
1047 | return false; | |
1048 | } | |
1049 | ||
1050 | /* Return true if the current code location is or might be offloaded. | |
1051 | Return true in declare target functions, or when nested in a target | |
1052 | region or when unsure, return false otherwise. */ | |
1053 | ||
1054 | static bool | |
1055 | omp_maybe_offloaded (void) | |
1056 | { | |
c56684fd MJ |
1057 | if (!ENABLE_OFFLOADING) |
1058 | return false; | |
1059 | const char *names = getenv ("OFFLOAD_TARGET_NAMES"); | |
1060 | if (names == NULL || *names == '\0') | |
1061 | return false; | |
1062 | ||
9ba66bf5 JJ |
1063 | if (symtab->state == PARSING) |
1064 | /* Maybe. */ | |
1065 | return true; | |
7a50e708 JJ |
1066 | if (cfun && cfun->after_inlining) |
1067 | return false; | |
9ba66bf5 JJ |
1068 | if (current_function_decl |
1069 | && lookup_attribute ("omp declare target", | |
1070 | DECL_ATTRIBUTES (current_function_decl))) | |
1071 | return true; | |
1072 | if (cfun && (cfun->curr_properties & PROP_gimple_any) == 0) | |
1073 | { | |
1074 | enum tree_code construct = OMP_TARGET; | |
d0c464d2 | 1075 | if (omp_construct_selector_matches (&construct, 1, NULL)) |
9ba66bf5 JJ |
1076 | return true; |
1077 | } | |
1078 | return false; | |
1079 | } | |
1080 | ||
b2417b59 JJ |
1081 | /* Return a name from PROP, a property in selectors accepting |
1082 | name lists. */ | |
1083 | ||
1084 | static const char * | |
1085 | omp_context_name_list_prop (tree prop) | |
1086 | { | |
1087 | if (TREE_PURPOSE (prop)) | |
1088 | return IDENTIFIER_POINTER (TREE_PURPOSE (prop)); | |
1089 | else | |
1090 | { | |
1091 | const char *ret = TREE_STRING_POINTER (TREE_VALUE (prop)); | |
1092 | if ((size_t) TREE_STRING_LENGTH (TREE_VALUE (prop)) == strlen (ret) + 1) | |
1093 | return ret; | |
1094 | return NULL; | |
1095 | } | |
1096 | } | |
1097 | ||
135df52c JJ |
1098 | /* Return 1 if context selector matches the current OpenMP context, 0 |
1099 | if it does not and -1 if it is unknown and need to be determined later. | |
1100 | Some properties can be checked right away during parsing (this routine), | |
1101 | others need to wait until the whole TU is parsed, others need to wait until | |
1102 | IPA, others until vectorization. */ | |
1103 | ||
1104 | int | |
1105 | omp_context_selector_matches (tree ctx) | |
1106 | { | |
1107 | int ret = 1; | |
1108 | for (tree t1 = ctx; t1; t1 = TREE_CHAIN (t1)) | |
1109 | { | |
1110 | char set = IDENTIFIER_POINTER (TREE_PURPOSE (t1))[0]; | |
1111 | if (set == 'c') | |
1112 | { | |
1113 | /* For now, ignore the construct set. While something can be | |
1114 | determined already during parsing, we don't know until end of TU | |
1115 | whether additional constructs aren't added through declare variant | |
1116 | unless "omp declare variant variant" attribute exists already | |
1117 | (so in most of the cases), and we'd need to maintain set of | |
1118 | surrounding OpenMP constructs, which is better handled during | |
1119 | gimplification. */ | |
7a50e708 | 1120 | if (symtab->state == PARSING) |
135df52c JJ |
1121 | { |
1122 | ret = -1; | |
1123 | continue; | |
1124 | } | |
1125 | ||
1126 | enum tree_code constructs[5]; | |
1127 | int nconstructs | |
1128 | = omp_constructor_traits_to_codes (TREE_VALUE (t1), constructs); | |
7a50e708 JJ |
1129 | |
1130 | if (cfun && (cfun->curr_properties & PROP_gimple_any) != 0) | |
1131 | { | |
1132 | if (!cfun->after_inlining) | |
1133 | { | |
1134 | ret = -1; | |
1135 | continue; | |
1136 | } | |
1137 | int i; | |
1138 | for (i = 0; i < nconstructs; ++i) | |
1139 | if (constructs[i] == OMP_SIMD) | |
1140 | break; | |
1141 | if (i < nconstructs) | |
1142 | { | |
1143 | ret = -1; | |
1144 | continue; | |
1145 | } | |
1146 | /* If there is no simd, assume it is ok after IPA, | |
1147 | constructs should have been checked before. */ | |
1148 | continue; | |
1149 | } | |
1150 | ||
d0c464d2 JJ |
1151 | int r = omp_construct_selector_matches (constructs, nconstructs, |
1152 | NULL); | |
135df52c JJ |
1153 | if (r == 0) |
1154 | return 0; | |
1155 | if (r == -1) | |
1156 | ret = -1; | |
1157 | continue; | |
1158 | } | |
1159 | for (tree t2 = TREE_VALUE (t1); t2; t2 = TREE_CHAIN (t2)) | |
1160 | { | |
1161 | const char *sel = IDENTIFIER_POINTER (TREE_PURPOSE (t2)); | |
1162 | switch (*sel) | |
1163 | { | |
1164 | case 'v': | |
1165 | if (set == 'i' && !strcmp (sel, "vendor")) | |
1166 | for (tree t3 = TREE_VALUE (t2); t3; t3 = TREE_CHAIN (t3)) | |
1167 | { | |
b2417b59 JJ |
1168 | const char *prop = omp_context_name_list_prop (t3); |
1169 | if (prop == NULL) | |
1170 | return 0; | |
1171 | if ((!strcmp (prop, " score") && TREE_PURPOSE (t3)) | |
1172 | || !strcmp (prop, "gnu")) | |
135df52c JJ |
1173 | continue; |
1174 | return 0; | |
1175 | } | |
1176 | break; | |
1177 | case 'e': | |
1178 | if (set == 'i' && !strcmp (sel, "extension")) | |
1179 | /* We don't support any extensions right now. */ | |
1180 | return 0; | |
1181 | break; | |
1182 | case 'a': | |
1183 | if (set == 'i' && !strcmp (sel, "atomic_default_mem_order")) | |
1184 | { | |
7a50e708 JJ |
1185 | if (cfun && (cfun->curr_properties & PROP_gimple_any) != 0) |
1186 | break; | |
1187 | ||
135df52c JJ |
1188 | enum omp_memory_order omo |
1189 | = ((enum omp_memory_order) | |
1190 | (omp_requires_mask | |
1191 | & OMP_REQUIRES_ATOMIC_DEFAULT_MEM_ORDER)); | |
1192 | if (omo == OMP_MEMORY_ORDER_UNSPECIFIED) | |
1193 | { | |
1194 | /* We don't know yet, until end of TU. */ | |
1195 | if (symtab->state == PARSING) | |
1196 | { | |
1197 | ret = -1; | |
1198 | break; | |
1199 | } | |
1200 | else | |
1201 | omo = OMP_MEMORY_ORDER_RELAXED; | |
1202 | } | |
1203 | tree t3 = TREE_VALUE (t2); | |
1204 | const char *prop = IDENTIFIER_POINTER (TREE_PURPOSE (t3)); | |
1205 | if (!strcmp (prop, " score")) | |
1206 | { | |
1207 | t3 = TREE_CHAIN (t3); | |
1208 | prop = IDENTIFIER_POINTER (TREE_PURPOSE (t3)); | |
1209 | } | |
1210 | if (!strcmp (prop, "relaxed") | |
1211 | && omo != OMP_MEMORY_ORDER_RELAXED) | |
1212 | return 0; | |
1213 | else if (!strcmp (prop, "seq_cst") | |
1214 | && omo != OMP_MEMORY_ORDER_SEQ_CST) | |
1215 | return 0; | |
1216 | else if (!strcmp (prop, "acq_rel") | |
1217 | && omo != OMP_MEMORY_ORDER_ACQ_REL) | |
1218 | return 0; | |
1219 | } | |
1220 | if (set == 'd' && !strcmp (sel, "arch")) | |
9ba66bf5 JJ |
1221 | for (tree t3 = TREE_VALUE (t2); t3; t3 = TREE_CHAIN (t3)) |
1222 | { | |
b2417b59 JJ |
1223 | const char *arch = omp_context_name_list_prop (t3); |
1224 | if (arch == NULL) | |
1225 | return 0; | |
9ba66bf5 JJ |
1226 | int r = 0; |
1227 | if (targetm.omp.device_kind_arch_isa != NULL) | |
1228 | r = targetm.omp.device_kind_arch_isa (omp_device_arch, | |
1229 | arch); | |
1230 | if (r == 0 || (r == -1 && symtab->state != PARSING)) | |
1231 | { | |
1232 | /* If we are or might be in a target region or | |
1233 | declare target function, need to take into account | |
1234 | also offloading values. */ | |
1235 | if (!omp_maybe_offloaded ()) | |
1236 | return 0; | |
9ba66bf5 JJ |
1237 | if (ENABLE_OFFLOADING) |
1238 | { | |
1239 | const char *arches = omp_offload_device_arch; | |
1240 | if (omp_offload_device_kind_arch_isa (arches, | |
1241 | arch)) | |
1242 | { | |
1243 | ret = -1; | |
1244 | continue; | |
1245 | } | |
1246 | } | |
1247 | return 0; | |
1248 | } | |
1249 | else if (r == -1) | |
1250 | ret = -1; | |
1251 | /* If arch matches on the host, it still might not match | |
1252 | in the offloading region. */ | |
1253 | else if (omp_maybe_offloaded ()) | |
1254 | ret = -1; | |
1255 | } | |
135df52c JJ |
1256 | break; |
1257 | case 'u': | |
1258 | if (set == 'i' && !strcmp (sel, "unified_address")) | |
1259 | { | |
7a50e708 JJ |
1260 | if (cfun && (cfun->curr_properties & PROP_gimple_any) != 0) |
1261 | break; | |
1262 | ||
135df52c JJ |
1263 | if ((omp_requires_mask & OMP_REQUIRES_UNIFIED_ADDRESS) == 0) |
1264 | { | |
1265 | if (symtab->state == PARSING) | |
1266 | ret = -1; | |
1267 | else | |
1268 | return 0; | |
1269 | } | |
1270 | break; | |
1271 | } | |
1272 | if (set == 'i' && !strcmp (sel, "unified_shared_memory")) | |
1273 | { | |
7a50e708 JJ |
1274 | if (cfun && (cfun->curr_properties & PROP_gimple_any) != 0) |
1275 | break; | |
1276 | ||
135df52c JJ |
1277 | if ((omp_requires_mask |
1278 | & OMP_REQUIRES_UNIFIED_SHARED_MEMORY) == 0) | |
1279 | { | |
1280 | if (symtab->state == PARSING) | |
1281 | ret = -1; | |
1282 | else | |
1283 | return 0; | |
1284 | } | |
1285 | break; | |
1286 | } | |
1287 | break; | |
1288 | case 'd': | |
1289 | if (set == 'i' && !strcmp (sel, "dynamic_allocators")) | |
1290 | { | |
7a50e708 JJ |
1291 | if (cfun && (cfun->curr_properties & PROP_gimple_any) != 0) |
1292 | break; | |
1293 | ||
135df52c JJ |
1294 | if ((omp_requires_mask |
1295 | & OMP_REQUIRES_DYNAMIC_ALLOCATORS) == 0) | |
1296 | { | |
1297 | if (symtab->state == PARSING) | |
1298 | ret = -1; | |
1299 | else | |
1300 | return 0; | |
1301 | } | |
1302 | break; | |
1303 | } | |
1304 | break; | |
1305 | case 'r': | |
1306 | if (set == 'i' && !strcmp (sel, "reverse_offload")) | |
1307 | { | |
7a50e708 JJ |
1308 | if (cfun && (cfun->curr_properties & PROP_gimple_any) != 0) |
1309 | break; | |
1310 | ||
135df52c JJ |
1311 | if ((omp_requires_mask & OMP_REQUIRES_REVERSE_OFFLOAD) == 0) |
1312 | { | |
1313 | if (symtab->state == PARSING) | |
1314 | ret = -1; | |
1315 | else | |
1316 | return 0; | |
1317 | } | |
1318 | break; | |
1319 | } | |
1320 | break; | |
1321 | case 'k': | |
1322 | if (set == 'd' && !strcmp (sel, "kind")) | |
1323 | for (tree t3 = TREE_VALUE (t2); t3; t3 = TREE_CHAIN (t3)) | |
1324 | { | |
b2417b59 JJ |
1325 | const char *prop = omp_context_name_list_prop (t3); |
1326 | if (prop == NULL) | |
1327 | return 0; | |
135df52c JJ |
1328 | if (!strcmp (prop, "any")) |
1329 | continue; | |
135df52c JJ |
1330 | if (!strcmp (prop, "host")) |
1331 | { | |
9ba66bf5 | 1332 | if (omp_maybe_offloaded ()) |
135df52c JJ |
1333 | ret = -1; |
1334 | continue; | |
1335 | } | |
1336 | if (!strcmp (prop, "nohost")) | |
1337 | { | |
9ba66bf5 | 1338 | if (omp_maybe_offloaded ()) |
135df52c JJ |
1339 | ret = -1; |
1340 | else | |
1341 | return 0; | |
1342 | continue; | |
1343 | } | |
9ba66bf5 JJ |
1344 | int r = 0; |
1345 | if (targetm.omp.device_kind_arch_isa != NULL) | |
1346 | r = targetm.omp.device_kind_arch_isa (omp_device_kind, | |
1347 | prop); | |
1348 | else | |
1349 | r = strcmp (prop, "cpu") == 0; | |
1350 | if (r == 0 || (r == -1 && symtab->state != PARSING)) | |
135df52c | 1351 | { |
9ba66bf5 JJ |
1352 | /* If we are or might be in a target region or |
1353 | declare target function, need to take into account | |
1354 | also offloading values. */ | |
1355 | if (!omp_maybe_offloaded ()) | |
1356 | return 0; | |
9ba66bf5 JJ |
1357 | if (ENABLE_OFFLOADING) |
1358 | { | |
1359 | const char *kinds = omp_offload_device_kind; | |
1360 | if (omp_offload_device_kind_arch_isa (kinds, prop)) | |
1361 | { | |
1362 | ret = -1; | |
1363 | continue; | |
1364 | } | |
1365 | } | |
1366 | return 0; | |
135df52c | 1367 | } |
9ba66bf5 JJ |
1368 | else if (r == -1) |
1369 | ret = -1; | |
1370 | /* If kind matches on the host, it still might not match | |
1371 | in the offloading region. */ | |
1372 | else if (omp_maybe_offloaded ()) | |
1373 | ret = -1; | |
135df52c JJ |
1374 | } |
1375 | break; | |
1376 | case 'i': | |
1377 | if (set == 'd' && !strcmp (sel, "isa")) | |
9ba66bf5 JJ |
1378 | for (tree t3 = TREE_VALUE (t2); t3; t3 = TREE_CHAIN (t3)) |
1379 | { | |
b2417b59 JJ |
1380 | const char *isa = omp_context_name_list_prop (t3); |
1381 | if (isa == NULL) | |
1382 | return 0; | |
9ba66bf5 JJ |
1383 | int r = 0; |
1384 | if (targetm.omp.device_kind_arch_isa != NULL) | |
1385 | r = targetm.omp.device_kind_arch_isa (omp_device_isa, | |
1386 | isa); | |
1387 | if (r == 0 || (r == -1 && symtab->state != PARSING)) | |
1388 | { | |
0227ffa9 JJ |
1389 | /* If isa is valid on the target, but not in the |
1390 | current function and current function has | |
1391 | #pragma omp declare simd on it, some simd clones | |
1392 | might have the isa added later on. */ | |
1393 | if (r == -1 | |
7a50e708 JJ |
1394 | && targetm.simd_clone.compute_vecsize_and_simdlen |
1395 | && (cfun == NULL || !cfun->after_inlining)) | |
0227ffa9 JJ |
1396 | { |
1397 | tree attrs | |
1398 | = DECL_ATTRIBUTES (current_function_decl); | |
1399 | if (lookup_attribute ("omp declare simd", attrs)) | |
1400 | { | |
1401 | ret = -1; | |
1402 | continue; | |
1403 | } | |
1404 | } | |
9ba66bf5 JJ |
1405 | /* If we are or might be in a target region or |
1406 | declare target function, need to take into account | |
1407 | also offloading values. */ | |
1408 | if (!omp_maybe_offloaded ()) | |
1409 | return 0; | |
1410 | if (ENABLE_OFFLOADING) | |
1411 | { | |
1412 | const char *isas = omp_offload_device_isa; | |
1413 | if (omp_offload_device_kind_arch_isa (isas, isa)) | |
1414 | { | |
1415 | ret = -1; | |
1416 | continue; | |
1417 | } | |
1418 | } | |
1419 | return 0; | |
1420 | } | |
1421 | else if (r == -1) | |
1422 | ret = -1; | |
1423 | /* If isa matches on the host, it still might not match | |
1424 | in the offloading region. */ | |
1425 | else if (omp_maybe_offloaded ()) | |
1426 | ret = -1; | |
1427 | } | |
135df52c JJ |
1428 | break; |
1429 | case 'c': | |
1430 | if (set == 'u' && !strcmp (sel, "condition")) | |
1431 | for (tree t3 = TREE_VALUE (t2); t3; t3 = TREE_CHAIN (t3)) | |
1432 | if (TREE_PURPOSE (t3) == NULL_TREE) | |
1433 | { | |
1434 | if (integer_zerop (TREE_VALUE (t3))) | |
1435 | return 0; | |
1436 | if (integer_nonzerop (TREE_VALUE (t3))) | |
1437 | break; | |
1438 | ret = -1; | |
1439 | } | |
1440 | break; | |
1441 | default: | |
1442 | break; | |
1443 | } | |
1444 | } | |
1445 | } | |
1446 | return ret; | |
1447 | } | |
1448 | ||
917dd789 JJ |
1449 | /* Compare construct={simd} CLAUSES1 with CLAUSES2, return 0/-1/1/2 as |
1450 | in omp_context_selector_set_compare. */ | |
1451 | ||
1452 | static int | |
1453 | omp_construct_simd_compare (tree clauses1, tree clauses2) | |
1454 | { | |
1455 | if (clauses1 == NULL_TREE) | |
1456 | return clauses2 == NULL_TREE ? 0 : -1; | |
1457 | if (clauses2 == NULL_TREE) | |
1458 | return 1; | |
1459 | ||
1460 | int r = 0; | |
1461 | struct declare_variant_simd_data { | |
1462 | bool inbranch, notinbranch; | |
1463 | tree simdlen; | |
1464 | auto_vec<tree,16> data_sharing; | |
1465 | auto_vec<tree,16> aligned; | |
1466 | declare_variant_simd_data () | |
1467 | : inbranch(false), notinbranch(false), simdlen(NULL_TREE) {} | |
1468 | } data[2]; | |
1469 | unsigned int i; | |
1470 | for (i = 0; i < 2; i++) | |
1471 | for (tree c = i ? clauses2 : clauses1; c; c = OMP_CLAUSE_CHAIN (c)) | |
1472 | { | |
1473 | vec<tree> *v; | |
1474 | switch (OMP_CLAUSE_CODE (c)) | |
1475 | { | |
1476 | case OMP_CLAUSE_INBRANCH: | |
1477 | data[i].inbranch = true; | |
1478 | continue; | |
1479 | case OMP_CLAUSE_NOTINBRANCH: | |
1480 | data[i].notinbranch = true; | |
1481 | continue; | |
1482 | case OMP_CLAUSE_SIMDLEN: | |
1483 | data[i].simdlen = OMP_CLAUSE_SIMDLEN_EXPR (c); | |
1484 | continue; | |
1485 | case OMP_CLAUSE_UNIFORM: | |
1486 | case OMP_CLAUSE_LINEAR: | |
1487 | v = &data[i].data_sharing; | |
1488 | break; | |
1489 | case OMP_CLAUSE_ALIGNED: | |
1490 | v = &data[i].aligned; | |
1491 | break; | |
1492 | default: | |
1493 | gcc_unreachable (); | |
1494 | } | |
1495 | unsigned HOST_WIDE_INT argno = tree_to_uhwi (OMP_CLAUSE_DECL (c)); | |
1496 | if (argno >= v->length ()) | |
cb3874dc | 1497 | v->safe_grow_cleared (argno + 1, true); |
917dd789 JJ |
1498 | (*v)[argno] = c; |
1499 | } | |
1500 | /* Here, r is used as a bitmask, 2 is set if CLAUSES1 has something | |
1501 | CLAUSES2 doesn't, 1 is set if CLAUSES2 has something CLAUSES1 | |
1502 | doesn't. Thus, r == 3 implies return value 2, r == 1 implies | |
1503 | -1, r == 2 implies 1 and r == 0 implies 0. */ | |
1504 | if (data[0].inbranch != data[1].inbranch) | |
1505 | r |= data[0].inbranch ? 2 : 1; | |
1506 | if (data[0].notinbranch != data[1].notinbranch) | |
1507 | r |= data[0].notinbranch ? 2 : 1; | |
1508 | if (!simple_cst_equal (data[0].simdlen, data[1].simdlen)) | |
1509 | { | |
1510 | if (data[0].simdlen && data[1].simdlen) | |
1511 | return 2; | |
1512 | r |= data[0].simdlen ? 2 : 1; | |
1513 | } | |
1514 | if (data[0].data_sharing.length () < data[1].data_sharing.length () | |
1515 | || data[0].aligned.length () < data[1].aligned.length ()) | |
1516 | r |= 1; | |
1517 | tree c1, c2; | |
1518 | FOR_EACH_VEC_ELT (data[0].data_sharing, i, c1) | |
1519 | { | |
1520 | c2 = (i < data[1].data_sharing.length () | |
1521 | ? data[1].data_sharing[i] : NULL_TREE); | |
1522 | if ((c1 == NULL_TREE) != (c2 == NULL_TREE)) | |
1523 | { | |
1524 | r |= c1 != NULL_TREE ? 2 : 1; | |
1525 | continue; | |
1526 | } | |
1527 | if (c1 == NULL_TREE) | |
1528 | continue; | |
1529 | if (OMP_CLAUSE_CODE (c1) != OMP_CLAUSE_CODE (c2)) | |
1530 | return 2; | |
1531 | if (OMP_CLAUSE_CODE (c1) != OMP_CLAUSE_LINEAR) | |
1532 | continue; | |
1533 | if (OMP_CLAUSE_LINEAR_VARIABLE_STRIDE (c1) | |
1534 | != OMP_CLAUSE_LINEAR_VARIABLE_STRIDE (c2)) | |
1535 | return 2; | |
1536 | if (OMP_CLAUSE_LINEAR_KIND (c1) != OMP_CLAUSE_LINEAR_KIND (c2)) | |
1537 | return 2; | |
1538 | if (!simple_cst_equal (OMP_CLAUSE_LINEAR_STEP (c1), | |
1539 | OMP_CLAUSE_LINEAR_STEP (c2))) | |
1540 | return 2; | |
1541 | } | |
1542 | FOR_EACH_VEC_ELT (data[0].aligned, i, c1) | |
1543 | { | |
1544 | c2 = i < data[1].aligned.length () ? data[1].aligned[i] : NULL_TREE; | |
1545 | if ((c1 == NULL_TREE) != (c2 == NULL_TREE)) | |
1546 | { | |
1547 | r |= c1 != NULL_TREE ? 2 : 1; | |
1548 | continue; | |
1549 | } | |
1550 | if (c1 == NULL_TREE) | |
1551 | continue; | |
1552 | if (!simple_cst_equal (OMP_CLAUSE_ALIGNED_ALIGNMENT (c1), | |
1553 | OMP_CLAUSE_ALIGNED_ALIGNMENT (c2))) | |
1554 | return 2; | |
1555 | } | |
1556 | switch (r) | |
1557 | { | |
1558 | case 0: return 0; | |
1559 | case 1: return -1; | |
1560 | case 2: return 1; | |
1561 | case 3: return 2; | |
1562 | default: gcc_unreachable (); | |
1563 | } | |
1564 | } | |
1565 | ||
1566 | /* Compare properties of selectors SEL from SET other than construct. | |
1567 | Return 0/-1/1/2 as in omp_context_selector_set_compare. | |
1568 | Unlike set names or selector names, properties can have duplicates. */ | |
1569 | ||
1570 | static int | |
1571 | omp_context_selector_props_compare (const char *set, const char *sel, | |
1572 | tree ctx1, tree ctx2) | |
1573 | { | |
1574 | int ret = 0; | |
1575 | for (int pass = 0; pass < 2; pass++) | |
1576 | for (tree t1 = pass ? ctx2 : ctx1; t1; t1 = TREE_CHAIN (t1)) | |
1577 | { | |
1578 | tree t2; | |
1579 | for (t2 = pass ? ctx1 : ctx2; t2; t2 = TREE_CHAIN (t2)) | |
1580 | if (TREE_PURPOSE (t1) == TREE_PURPOSE (t2)) | |
1581 | { | |
1582 | if (TREE_PURPOSE (t1) == NULL_TREE) | |
1583 | { | |
1584 | if (set[0] == 'u' && strcmp (sel, "condition") == 0) | |
1585 | { | |
1586 | if (integer_zerop (TREE_VALUE (t1)) | |
1587 | != integer_zerop (TREE_VALUE (t2))) | |
1588 | return 2; | |
1589 | break; | |
1590 | } | |
1591 | if (simple_cst_equal (TREE_VALUE (t1), TREE_VALUE (t2))) | |
1592 | break; | |
1593 | } | |
1594 | else if (strcmp (IDENTIFIER_POINTER (TREE_PURPOSE (t1)), | |
1595 | " score") == 0) | |
1596 | { | |
1597 | if (!simple_cst_equal (TREE_VALUE (t1), TREE_VALUE (t2))) | |
1598 | return 2; | |
1599 | break; | |
1600 | } | |
1601 | else | |
1602 | break; | |
1603 | } | |
b2417b59 JJ |
1604 | else if (TREE_PURPOSE (t1) |
1605 | && TREE_PURPOSE (t2) == NULL_TREE | |
1606 | && TREE_CODE (TREE_VALUE (t2)) == STRING_CST) | |
1607 | { | |
1608 | const char *p1 = omp_context_name_list_prop (t1); | |
1609 | const char *p2 = omp_context_name_list_prop (t2); | |
1610 | if (p2 | |
1611 | && strcmp (p1, p2) == 0 | |
1612 | && strcmp (p1, " score")) | |
1613 | break; | |
1614 | } | |
1615 | else if (TREE_PURPOSE (t1) == NULL_TREE | |
1616 | && TREE_PURPOSE (t2) | |
1617 | && TREE_CODE (TREE_VALUE (t1)) == STRING_CST) | |
1618 | { | |
1619 | const char *p1 = omp_context_name_list_prop (t1); | |
1620 | const char *p2 = omp_context_name_list_prop (t2); | |
1621 | if (p1 | |
1622 | && strcmp (p1, p2) == 0 | |
1623 | && strcmp (p1, " score")) | |
1624 | break; | |
1625 | } | |
917dd789 JJ |
1626 | if (t2 == NULL_TREE) |
1627 | { | |
1628 | int r = pass ? -1 : 1; | |
1629 | if (ret && ret != r) | |
1630 | return 2; | |
1631 | else if (pass) | |
1632 | return r; | |
1633 | else | |
1634 | { | |
1635 | ret = r; | |
1636 | break; | |
1637 | } | |
1638 | } | |
1639 | } | |
1640 | return ret; | |
1641 | } | |
1642 | ||
1643 | /* Compare single context selector sets CTX1 and CTX2 with SET name. | |
1644 | Return 0 if CTX1 is equal to CTX2, | |
1645 | -1 if CTX1 is a strict subset of CTX2, | |
1646 | 1 if CTX2 is a strict subset of CTX1, or | |
1647 | 2 if neither context is a subset of another one. */ | |
1648 | ||
1649 | int | |
1650 | omp_context_selector_set_compare (const char *set, tree ctx1, tree ctx2) | |
1651 | { | |
1652 | bool swapped = false; | |
1653 | int ret = 0; | |
1654 | int len1 = list_length (ctx1); | |
1655 | int len2 = list_length (ctx2); | |
1656 | int cnt = 0; | |
1657 | if (len1 < len2) | |
1658 | { | |
1659 | swapped = true; | |
1660 | std::swap (ctx1, ctx2); | |
1661 | std::swap (len1, len2); | |
1662 | } | |
1663 | if (set[0] == 'c') | |
1664 | { | |
1665 | tree t1; | |
1666 | tree t2 = ctx2; | |
1667 | tree simd = get_identifier ("simd"); | |
1668 | /* Handle construct set specially. In this case the order | |
1669 | of the selector matters too. */ | |
1670 | for (t1 = ctx1; t1; t1 = TREE_CHAIN (t1)) | |
1671 | if (TREE_PURPOSE (t1) == TREE_PURPOSE (t2)) | |
1672 | { | |
1673 | int r = 0; | |
1674 | if (TREE_PURPOSE (t1) == simd) | |
1675 | r = omp_construct_simd_compare (TREE_VALUE (t1), | |
1676 | TREE_VALUE (t2)); | |
1677 | if (r == 2 || (ret && r && (ret < 0) != (r < 0))) | |
1678 | return 2; | |
1679 | if (ret == 0) | |
1680 | ret = r; | |
1681 | t2 = TREE_CHAIN (t2); | |
1682 | if (t2 == NULL_TREE) | |
1683 | { | |
1684 | t1 = TREE_CHAIN (t1); | |
1685 | break; | |
1686 | } | |
1687 | } | |
1688 | else if (ret < 0) | |
1689 | return 2; | |
1690 | else | |
1691 | ret = 1; | |
1692 | if (t2 != NULL_TREE) | |
1693 | return 2; | |
1694 | if (t1 != NULL_TREE) | |
1695 | { | |
1696 | if (ret < 0) | |
1697 | return 2; | |
1698 | ret = 1; | |
1699 | } | |
1700 | if (ret == 0) | |
1701 | return 0; | |
1702 | return swapped ? -ret : ret; | |
1703 | } | |
1704 | for (tree t1 = ctx1; t1; t1 = TREE_CHAIN (t1)) | |
1705 | { | |
1706 | tree t2; | |
1707 | for (t2 = ctx2; t2; t2 = TREE_CHAIN (t2)) | |
1708 | if (TREE_PURPOSE (t1) == TREE_PURPOSE (t2)) | |
1709 | { | |
1710 | const char *sel = IDENTIFIER_POINTER (TREE_PURPOSE (t1)); | |
1711 | int r = omp_context_selector_props_compare (set, sel, | |
1712 | TREE_VALUE (t1), | |
1713 | TREE_VALUE (t2)); | |
1714 | if (r == 2 || (ret && r && (ret < 0) != (r < 0))) | |
1715 | return 2; | |
1716 | if (ret == 0) | |
1717 | ret = r; | |
1718 | cnt++; | |
1719 | break; | |
1720 | } | |
1721 | if (t2 == NULL_TREE) | |
1722 | { | |
1723 | if (ret == -1) | |
1724 | return 2; | |
1725 | ret = 1; | |
1726 | } | |
1727 | } | |
1728 | if (cnt < len2) | |
1729 | return 2; | |
1730 | if (ret == 0) | |
1731 | return 0; | |
1732 | return swapped ? -ret : ret; | |
1733 | } | |
1734 | ||
1735 | /* Compare whole context selector specification CTX1 and CTX2. | |
1736 | Return 0 if CTX1 is equal to CTX2, | |
1737 | -1 if CTX1 is a strict subset of CTX2, | |
1738 | 1 if CTX2 is a strict subset of CTX1, or | |
1739 | 2 if neither context is a subset of another one. */ | |
1740 | ||
1741 | static int | |
1742 | omp_context_selector_compare (tree ctx1, tree ctx2) | |
1743 | { | |
1744 | bool swapped = false; | |
1745 | int ret = 0; | |
1746 | int len1 = list_length (ctx1); | |
1747 | int len2 = list_length (ctx2); | |
1748 | int cnt = 0; | |
1749 | if (len1 < len2) | |
1750 | { | |
1751 | swapped = true; | |
1752 | std::swap (ctx1, ctx2); | |
1753 | std::swap (len1, len2); | |
1754 | } | |
1755 | for (tree t1 = ctx1; t1; t1 = TREE_CHAIN (t1)) | |
1756 | { | |
1757 | tree t2; | |
1758 | for (t2 = ctx2; t2; t2 = TREE_CHAIN (t2)) | |
1759 | if (TREE_PURPOSE (t1) == TREE_PURPOSE (t2)) | |
1760 | { | |
1761 | const char *set = IDENTIFIER_POINTER (TREE_PURPOSE (t1)); | |
1762 | int r = omp_context_selector_set_compare (set, TREE_VALUE (t1), | |
1763 | TREE_VALUE (t2)); | |
1764 | if (r == 2 || (ret && r && (ret < 0) != (r < 0))) | |
1765 | return 2; | |
1766 | if (ret == 0) | |
1767 | ret = r; | |
1768 | cnt++; | |
1769 | break; | |
1770 | } | |
1771 | if (t2 == NULL_TREE) | |
1772 | { | |
1773 | if (ret == -1) | |
1774 | return 2; | |
1775 | ret = 1; | |
1776 | } | |
1777 | } | |
1778 | if (cnt < len2) | |
1779 | return 2; | |
1780 | if (ret == 0) | |
1781 | return 0; | |
1782 | return swapped ? -ret : ret; | |
1783 | } | |
1784 | ||
d0c464d2 JJ |
1785 | /* From context selector CTX, return trait-selector with name SEL in |
1786 | trait-selector-set with name SET if any, or NULL_TREE if not found. | |
1787 | If SEL is NULL, return the list of trait-selectors in SET. */ | |
1788 | ||
1789 | tree | |
1790 | omp_get_context_selector (tree ctx, const char *set, const char *sel) | |
1791 | { | |
1792 | tree setid = get_identifier (set); | |
1793 | tree selid = sel ? get_identifier (sel) : NULL_TREE; | |
1794 | for (tree t1 = ctx; t1; t1 = TREE_CHAIN (t1)) | |
1795 | if (TREE_PURPOSE (t1) == setid) | |
1796 | { | |
1797 | if (sel == NULL) | |
1798 | return TREE_VALUE (t1); | |
1799 | for (tree t2 = TREE_VALUE (t1); t2; t2 = TREE_CHAIN (t2)) | |
1800 | if (TREE_PURPOSE (t2) == selid) | |
1801 | return t2; | |
1802 | } | |
1803 | return NULL_TREE; | |
1804 | } | |
1805 | ||
1806 | /* Compute *SCORE for context selector CTX. Return true if the score | |
1807 | would be different depending on whether it is a declare simd clone or | |
1808 | not. DECLARE_SIMD should be true for the case when it would be | |
1809 | a declare simd clone. */ | |
1810 | ||
1811 | static bool | |
1812 | omp_context_compute_score (tree ctx, widest_int *score, bool declare_simd) | |
1813 | { | |
1814 | tree construct = omp_get_context_selector (ctx, "construct", NULL); | |
1815 | bool has_kind = omp_get_context_selector (ctx, "device", "kind"); | |
1816 | bool has_arch = omp_get_context_selector (ctx, "device", "arch"); | |
1817 | bool has_isa = omp_get_context_selector (ctx, "device", "isa"); | |
1818 | bool ret = false; | |
1819 | *score = 1; | |
1820 | for (tree t1 = ctx; t1; t1 = TREE_CHAIN (t1)) | |
0227ffa9 JJ |
1821 | if (TREE_VALUE (t1) != construct) |
1822 | for (tree t2 = TREE_VALUE (t1); t2; t2 = TREE_CHAIN (t2)) | |
1823 | if (tree t3 = TREE_VALUE (t2)) | |
1824 | if (TREE_PURPOSE (t3) | |
1825 | && strcmp (IDENTIFIER_POINTER (TREE_PURPOSE (t3)), " score") == 0 | |
1826 | && TREE_CODE (TREE_VALUE (t3)) == INTEGER_CST) | |
1827 | *score += wi::to_widest (TREE_VALUE (t3)); | |
d0c464d2 JJ |
1828 | if (construct || has_kind || has_arch || has_isa) |
1829 | { | |
1830 | int scores[12]; | |
1831 | enum tree_code constructs[5]; | |
1832 | int nconstructs = 0; | |
1833 | if (construct) | |
1834 | nconstructs = omp_constructor_traits_to_codes (construct, constructs); | |
1835 | if (omp_construct_selector_matches (constructs, nconstructs, scores) | |
1836 | == 2) | |
1837 | ret = true; | |
1838 | int b = declare_simd ? nconstructs + 1 : 0; | |
1839 | if (scores[b + nconstructs] + 4U < score->get_precision ()) | |
1840 | { | |
1841 | for (int n = 0; n < nconstructs; ++n) | |
1842 | { | |
1843 | if (scores[b + n] < 0) | |
1844 | { | |
0227ffa9 | 1845 | *score = -1; |
d0c464d2 JJ |
1846 | return ret; |
1847 | } | |
1848 | *score += wi::shifted_mask <widest_int> (scores[b + n], 1, false); | |
1849 | } | |
1850 | if (has_kind) | |
1851 | *score += wi::shifted_mask <widest_int> (scores[b + nconstructs], | |
1852 | 1, false); | |
1853 | if (has_arch) | |
1854 | *score += wi::shifted_mask <widest_int> (scores[b + nconstructs] + 1, | |
1855 | 1, false); | |
1856 | if (has_isa) | |
1857 | *score += wi::shifted_mask <widest_int> (scores[b + nconstructs] + 2, | |
1858 | 1, false); | |
1859 | } | |
1860 | else /* FIXME: Implement this. */ | |
1861 | gcc_unreachable (); | |
1862 | } | |
1863 | return ret; | |
1864 | } | |
1865 | ||
7a50e708 JJ |
1866 | /* Class describing a single variant. */ |
1867 | struct GTY(()) omp_declare_variant_entry { | |
1868 | /* NODE of the variant. */ | |
1869 | cgraph_node *variant; | |
1870 | /* Score if not in declare simd clone. */ | |
1871 | widest_int score; | |
1872 | /* Score if in declare simd clone. */ | |
1873 | widest_int score_in_declare_simd_clone; | |
1874 | /* Context selector for the variant. */ | |
1875 | tree ctx; | |
1876 | /* True if the context selector is known to match already. */ | |
1877 | bool matches; | |
1878 | }; | |
1879 | ||
1880 | /* Class describing a function with variants. */ | |
1881 | struct GTY((for_user)) omp_declare_variant_base_entry { | |
1882 | /* NODE of the base function. */ | |
1883 | cgraph_node *base; | |
1884 | /* NODE of the artificial function created for the deferred variant | |
1885 | resolution. */ | |
1886 | cgraph_node *node; | |
1887 | /* Vector of the variants. */ | |
1888 | vec<omp_declare_variant_entry, va_gc> *variants; | |
1889 | }; | |
1890 | ||
1891 | struct omp_declare_variant_hasher | |
1892 | : ggc_ptr_hash<omp_declare_variant_base_entry> { | |
1893 | static hashval_t hash (omp_declare_variant_base_entry *); | |
1894 | static bool equal (omp_declare_variant_base_entry *, | |
1895 | omp_declare_variant_base_entry *); | |
1896 | }; | |
1897 | ||
1898 | hashval_t | |
1899 | omp_declare_variant_hasher::hash (omp_declare_variant_base_entry *x) | |
1900 | { | |
1901 | inchash::hash hstate; | |
1902 | hstate.add_int (DECL_UID (x->base->decl)); | |
1903 | hstate.add_int (x->variants->length ()); | |
1904 | omp_declare_variant_entry *variant; | |
1905 | unsigned int i; | |
1906 | FOR_EACH_VEC_SAFE_ELT (x->variants, i, variant) | |
1907 | { | |
1908 | hstate.add_int (DECL_UID (variant->variant->decl)); | |
1909 | hstate.add_wide_int (variant->score); | |
1910 | hstate.add_wide_int (variant->score_in_declare_simd_clone); | |
1911 | hstate.add_ptr (variant->ctx); | |
1912 | hstate.add_int (variant->matches); | |
1913 | } | |
1914 | return hstate.end (); | |
1915 | } | |
1916 | ||
1917 | bool | |
1918 | omp_declare_variant_hasher::equal (omp_declare_variant_base_entry *x, | |
1919 | omp_declare_variant_base_entry *y) | |
1920 | { | |
1921 | if (x->base != y->base | |
1922 | || x->variants->length () != y->variants->length ()) | |
1923 | return false; | |
1924 | omp_declare_variant_entry *variant; | |
1925 | unsigned int i; | |
1926 | FOR_EACH_VEC_SAFE_ELT (x->variants, i, variant) | |
1927 | if (variant->variant != (*y->variants)[i].variant | |
1928 | || variant->score != (*y->variants)[i].score | |
1929 | || (variant->score_in_declare_simd_clone | |
1930 | != (*y->variants)[i].score_in_declare_simd_clone) | |
1931 | || variant->ctx != (*y->variants)[i].ctx | |
1932 | || variant->matches != (*y->variants)[i].matches) | |
1933 | return false; | |
1934 | return true; | |
1935 | } | |
1936 | ||
1937 | static GTY(()) hash_table<omp_declare_variant_hasher> *omp_declare_variants; | |
1938 | ||
1939 | struct omp_declare_variant_alt_hasher | |
1940 | : ggc_ptr_hash<omp_declare_variant_base_entry> { | |
1941 | static hashval_t hash (omp_declare_variant_base_entry *); | |
1942 | static bool equal (omp_declare_variant_base_entry *, | |
1943 | omp_declare_variant_base_entry *); | |
1944 | }; | |
1945 | ||
1946 | hashval_t | |
1947 | omp_declare_variant_alt_hasher::hash (omp_declare_variant_base_entry *x) | |
1948 | { | |
1949 | return DECL_UID (x->node->decl); | |
1950 | } | |
1951 | ||
1952 | bool | |
1953 | omp_declare_variant_alt_hasher::equal (omp_declare_variant_base_entry *x, | |
1954 | omp_declare_variant_base_entry *y) | |
1955 | { | |
1956 | return x->node == y->node; | |
1957 | } | |
1958 | ||
1959 | static GTY(()) hash_table<omp_declare_variant_alt_hasher> | |
1960 | *omp_declare_variant_alt; | |
1961 | ||
1962 | /* Try to resolve declare variant after gimplification. */ | |
1963 | ||
1964 | static tree | |
1965 | omp_resolve_late_declare_variant (tree alt) | |
1966 | { | |
1967 | cgraph_node *node = cgraph_node::get (alt); | |
1968 | cgraph_node *cur_node = cgraph_node::get (cfun->decl); | |
1969 | if (node == NULL | |
1970 | || !node->declare_variant_alt | |
1971 | || !cfun->after_inlining) | |
1972 | return alt; | |
1973 | ||
1974 | omp_declare_variant_base_entry entry; | |
1975 | entry.base = NULL; | |
1976 | entry.node = node; | |
1977 | entry.variants = NULL; | |
1978 | omp_declare_variant_base_entry *entryp | |
1979 | = omp_declare_variant_alt->find_with_hash (&entry, DECL_UID (alt)); | |
1980 | ||
1981 | unsigned int i, j; | |
1982 | omp_declare_variant_entry *varentry1, *varentry2; | |
1983 | auto_vec <bool, 16> matches; | |
1984 | unsigned int nmatches = 0; | |
1985 | FOR_EACH_VEC_SAFE_ELT (entryp->variants, i, varentry1) | |
1986 | { | |
1987 | if (varentry1->matches) | |
1988 | { | |
1989 | /* This has been checked to be ok already. */ | |
1990 | matches.safe_push (true); | |
1991 | nmatches++; | |
1992 | continue; | |
1993 | } | |
1994 | switch (omp_context_selector_matches (varentry1->ctx)) | |
1995 | { | |
1996 | case 0: | |
1997 | matches.safe_push (false); | |
1998 | break; | |
1999 | case -1: | |
2000 | return alt; | |
2001 | default: | |
2002 | matches.safe_push (true); | |
2003 | nmatches++; | |
2004 | break; | |
2005 | } | |
2006 | } | |
2007 | ||
2008 | if (nmatches == 0) | |
2009 | return entryp->base->decl; | |
2010 | ||
2011 | /* A context selector that is a strict subset of another context selector | |
2012 | has a score of zero. */ | |
2013 | FOR_EACH_VEC_SAFE_ELT (entryp->variants, i, varentry1) | |
2014 | if (matches[i]) | |
2015 | { | |
2016 | for (j = i + 1; | |
2017 | vec_safe_iterate (entryp->variants, j, &varentry2); ++j) | |
2018 | if (matches[j]) | |
2019 | { | |
2020 | int r = omp_context_selector_compare (varentry1->ctx, | |
2021 | varentry2->ctx); | |
2022 | if (r == -1) | |
2023 | { | |
2024 | /* ctx1 is a strict subset of ctx2, ignore ctx1. */ | |
2025 | matches[i] = false; | |
2026 | break; | |
2027 | } | |
2028 | else if (r == 1) | |
2029 | /* ctx2 is a strict subset of ctx1, remove ctx2. */ | |
2030 | matches[j] = false; | |
2031 | } | |
2032 | } | |
2033 | ||
2034 | widest_int max_score = -1; | |
2035 | varentry2 = NULL; | |
2036 | FOR_EACH_VEC_SAFE_ELT (entryp->variants, i, varentry1) | |
2037 | if (matches[i]) | |
2038 | { | |
2039 | widest_int score | |
2040 | = (cur_node->simdclone ? varentry1->score_in_declare_simd_clone | |
2041 | : varentry1->score); | |
2042 | if (score > max_score) | |
2043 | { | |
2044 | max_score = score; | |
2045 | varentry2 = varentry1; | |
2046 | } | |
2047 | } | |
2048 | return varentry2->variant->decl; | |
2049 | } | |
2050 | ||
baff22c4 JJ |
2051 | /* Hook to adjust hash tables on cgraph_node removal. */ |
2052 | ||
2053 | static void | |
2054 | omp_declare_variant_remove_hook (struct cgraph_node *node, void *) | |
2055 | { | |
2056 | if (!node->declare_variant_alt) | |
2057 | return; | |
2058 | ||
2059 | /* Drop this hash table completely. */ | |
2060 | omp_declare_variants = NULL; | |
2061 | /* And remove node from the other hash table. */ | |
2062 | if (omp_declare_variant_alt) | |
2063 | { | |
2064 | omp_declare_variant_base_entry entry; | |
2065 | entry.base = NULL; | |
2066 | entry.node = node; | |
2067 | entry.variants = NULL; | |
2068 | omp_declare_variant_alt->remove_elt_with_hash (&entry, | |
2069 | DECL_UID (node->decl)); | |
2070 | } | |
2071 | } | |
2072 | ||
135df52c JJ |
2073 | /* Try to resolve declare variant, return the variant decl if it should |
2074 | be used instead of base, or base otherwise. */ | |
2075 | ||
2076 | tree | |
2077 | omp_resolve_declare_variant (tree base) | |
2078 | { | |
d0c464d2 | 2079 | tree variant1 = NULL_TREE, variant2 = NULL_TREE; |
7a50e708 JJ |
2080 | if (cfun && (cfun->curr_properties & PROP_gimple_any) != 0) |
2081 | return omp_resolve_late_declare_variant (base); | |
2082 | ||
917dd789 | 2083 | auto_vec <tree, 16> variants; |
0227ffa9 JJ |
2084 | auto_vec <bool, 16> defer; |
2085 | bool any_deferred = false; | |
135df52c JJ |
2086 | for (tree attr = DECL_ATTRIBUTES (base); attr; attr = TREE_CHAIN (attr)) |
2087 | { | |
2088 | attr = lookup_attribute ("omp declare variant base", attr); | |
2089 | if (attr == NULL_TREE) | |
2090 | break; | |
917dd789 JJ |
2091 | if (TREE_CODE (TREE_PURPOSE (TREE_VALUE (attr))) != FUNCTION_DECL) |
2092 | continue; | |
baff22c4 JJ |
2093 | cgraph_node *node = cgraph_node::get (base); |
2094 | /* If this is already a magic decl created by this function, | |
2095 | don't process it again. */ | |
2096 | if (node && node->declare_variant_alt) | |
2097 | return base; | |
135df52c JJ |
2098 | switch (omp_context_selector_matches (TREE_VALUE (TREE_VALUE (attr)))) |
2099 | { | |
2100 | case 0: | |
2101 | /* No match, ignore. */ | |
2102 | break; | |
2103 | case -1: | |
2104 | /* Needs to be deferred. */ | |
0227ffa9 JJ |
2105 | any_deferred = true; |
2106 | variants.safe_push (attr); | |
2107 | defer.safe_push (true); | |
2108 | break; | |
135df52c | 2109 | default: |
917dd789 | 2110 | variants.safe_push (attr); |
0227ffa9 JJ |
2111 | defer.safe_push (false); |
2112 | break; | |
135df52c JJ |
2113 | } |
2114 | } | |
917dd789 JJ |
2115 | if (variants.length () == 0) |
2116 | return base; | |
0227ffa9 JJ |
2117 | |
2118 | if (any_deferred) | |
2119 | { | |
2120 | widest_int max_score1 = 0; | |
2121 | widest_int max_score2 = 0; | |
2122 | bool first = true; | |
2123 | unsigned int i; | |
2124 | tree attr1, attr2; | |
7a50e708 JJ |
2125 | omp_declare_variant_base_entry entry; |
2126 | entry.base = cgraph_node::get_create (base); | |
2127 | entry.node = NULL; | |
2128 | vec_alloc (entry.variants, variants.length ()); | |
0227ffa9 JJ |
2129 | FOR_EACH_VEC_ELT (variants, i, attr1) |
2130 | { | |
2131 | widest_int score1; | |
2132 | widest_int score2; | |
2133 | bool need_two; | |
2134 | tree ctx = TREE_VALUE (TREE_VALUE (attr1)); | |
2135 | need_two = omp_context_compute_score (ctx, &score1, false); | |
2136 | if (need_two) | |
2137 | omp_context_compute_score (ctx, &score2, true); | |
2138 | else | |
2139 | score2 = score1; | |
2140 | if (first) | |
2141 | { | |
2142 | first = false; | |
2143 | max_score1 = score1; | |
2144 | max_score2 = score2; | |
2145 | if (!defer[i]) | |
2146 | { | |
2147 | variant1 = attr1; | |
2148 | variant2 = attr1; | |
2149 | } | |
2150 | } | |
2151 | else | |
2152 | { | |
2153 | if (max_score1 == score1) | |
2154 | variant1 = NULL_TREE; | |
2155 | else if (score1 > max_score1) | |
2156 | { | |
2157 | max_score1 = score1; | |
2158 | variant1 = defer[i] ? NULL_TREE : attr1; | |
2159 | } | |
2160 | if (max_score2 == score2) | |
2161 | variant2 = NULL_TREE; | |
2162 | else if (score2 > max_score2) | |
2163 | { | |
2164 | max_score2 = score2; | |
2165 | variant2 = defer[i] ? NULL_TREE : attr1; | |
2166 | } | |
2167 | } | |
7a50e708 JJ |
2168 | omp_declare_variant_entry varentry; |
2169 | varentry.variant | |
2170 | = cgraph_node::get_create (TREE_PURPOSE (TREE_VALUE (attr1))); | |
2171 | varentry.score = score1; | |
2172 | varentry.score_in_declare_simd_clone = score2; | |
2173 | varentry.ctx = ctx; | |
2174 | varentry.matches = !defer[i]; | |
2175 | entry.variants->quick_push (varentry); | |
0227ffa9 JJ |
2176 | } |
2177 | ||
2178 | /* If there is a clear winner variant with the score which is not | |
2179 | deferred, verify it is not a strict subset of any other context | |
2180 | selector and if it is not, it is the best alternative no matter | |
2181 | whether the others do or don't match. */ | |
2182 | if (variant1 && variant1 == variant2) | |
2183 | { | |
2184 | tree ctx1 = TREE_VALUE (TREE_VALUE (variant1)); | |
2185 | FOR_EACH_VEC_ELT (variants, i, attr2) | |
2186 | { | |
2187 | if (attr2 == variant1) | |
2188 | continue; | |
2189 | tree ctx2 = TREE_VALUE (TREE_VALUE (attr2)); | |
2190 | int r = omp_context_selector_compare (ctx1, ctx2); | |
2191 | if (r == -1) | |
2192 | { | |
2193 | /* The winner is a strict subset of ctx2, can't | |
2194 | decide now. */ | |
2195 | variant1 = NULL_TREE; | |
2196 | break; | |
2197 | } | |
2198 | } | |
2199 | if (variant1) | |
7a50e708 JJ |
2200 | { |
2201 | vec_free (entry.variants); | |
2202 | return TREE_PURPOSE (TREE_VALUE (variant1)); | |
2203 | } | |
2204 | } | |
2205 | ||
baff22c4 | 2206 | static struct cgraph_node_hook_list *node_removal_hook_holder; |
3d0675f3 | 2207 | if (!node_removal_hook_holder) |
baff22c4 JJ |
2208 | node_removal_hook_holder |
2209 | = symtab->add_cgraph_removal_hook (omp_declare_variant_remove_hook, | |
2210 | NULL); | |
2211 | ||
7a50e708 JJ |
2212 | if (omp_declare_variants == NULL) |
2213 | omp_declare_variants | |
2214 | = hash_table<omp_declare_variant_hasher>::create_ggc (64); | |
2215 | omp_declare_variant_base_entry **slot | |
2216 | = omp_declare_variants->find_slot (&entry, INSERT); | |
2217 | if (*slot != NULL) | |
2218 | { | |
2219 | vec_free (entry.variants); | |
2220 | return (*slot)->node->decl; | |
0227ffa9 JJ |
2221 | } |
2222 | ||
7a50e708 JJ |
2223 | *slot = ggc_cleared_alloc<omp_declare_variant_base_entry> (); |
2224 | (*slot)->base = entry.base; | |
2225 | (*slot)->node = entry.base; | |
2226 | (*slot)->variants = entry.variants; | |
2227 | tree alt = build_decl (DECL_SOURCE_LOCATION (base), FUNCTION_DECL, | |
2228 | DECL_NAME (base), TREE_TYPE (base)); | |
2229 | DECL_ARTIFICIAL (alt) = 1; | |
2230 | DECL_IGNORED_P (alt) = 1; | |
2231 | TREE_STATIC (alt) = 1; | |
2232 | tree attributes = DECL_ATTRIBUTES (base); | |
2233 | if (lookup_attribute ("noipa", attributes) == NULL) | |
2234 | { | |
2235 | attributes = tree_cons (get_identifier ("noipa"), NULL, attributes); | |
2236 | if (lookup_attribute ("noinline", attributes) == NULL) | |
2237 | attributes = tree_cons (get_identifier ("noinline"), NULL, | |
2238 | attributes); | |
2239 | if (lookup_attribute ("noclone", attributes) == NULL) | |
2240 | attributes = tree_cons (get_identifier ("noclone"), NULL, | |
2241 | attributes); | |
2242 | if (lookup_attribute ("no_icf", attributes) == NULL) | |
2243 | attributes = tree_cons (get_identifier ("no_icf"), NULL, | |
2244 | attributes); | |
2245 | } | |
2246 | DECL_ATTRIBUTES (alt) = attributes; | |
2247 | DECL_INITIAL (alt) = error_mark_node; | |
2248 | (*slot)->node = cgraph_node::create (alt); | |
2249 | (*slot)->node->declare_variant_alt = 1; | |
2250 | (*slot)->node->create_reference (entry.base, IPA_REF_ADDR); | |
2251 | omp_declare_variant_entry *varentry; | |
2252 | FOR_EACH_VEC_SAFE_ELT (entry.variants, i, varentry) | |
2253 | (*slot)->node->create_reference (varentry->variant, IPA_REF_ADDR); | |
2254 | if (omp_declare_variant_alt == NULL) | |
2255 | omp_declare_variant_alt | |
2256 | = hash_table<omp_declare_variant_alt_hasher>::create_ggc (64); | |
2257 | *omp_declare_variant_alt->find_slot_with_hash (*slot, DECL_UID (alt), | |
2258 | INSERT) = *slot; | |
2259 | return alt; | |
0227ffa9 JJ |
2260 | } |
2261 | ||
917dd789 JJ |
2262 | if (variants.length () == 1) |
2263 | return TREE_PURPOSE (TREE_VALUE (variants[0])); | |
2264 | ||
7a50e708 JJ |
2265 | /* A context selector that is a strict subset of another context selector |
2266 | has a score of zero. */ | |
917dd789 JJ |
2267 | tree attr1, attr2; |
2268 | unsigned int i, j; | |
2269 | FOR_EACH_VEC_ELT (variants, i, attr1) | |
2270 | if (attr1) | |
2271 | { | |
2272 | tree ctx1 = TREE_VALUE (TREE_VALUE (attr1)); | |
2273 | FOR_EACH_VEC_ELT_FROM (variants, j, attr2, i + 1) | |
2274 | if (attr2) | |
2275 | { | |
2276 | tree ctx2 = TREE_VALUE (TREE_VALUE (attr2)); | |
2277 | int r = omp_context_selector_compare (ctx1, ctx2); | |
2278 | if (r == -1) | |
2279 | { | |
2280 | /* ctx1 is a strict subset of ctx2, remove | |
2281 | attr1 from the vector. */ | |
2282 | variants[i] = NULL_TREE; | |
2283 | break; | |
2284 | } | |
2285 | else if (r == 1) | |
2286 | /* ctx2 is a strict subset of ctx1, remove attr2 | |
2287 | from the vector. */ | |
2288 | variants[j] = NULL_TREE; | |
2289 | } | |
2290 | } | |
d0c464d2 JJ |
2291 | widest_int max_score1 = 0; |
2292 | widest_int max_score2 = 0; | |
2293 | bool first = true; | |
917dd789 JJ |
2294 | FOR_EACH_VEC_ELT (variants, i, attr1) |
2295 | if (attr1) | |
2296 | { | |
d0c464d2 JJ |
2297 | if (variant1) |
2298 | { | |
2299 | widest_int score1; | |
2300 | widest_int score2; | |
2301 | bool need_two; | |
2302 | tree ctx; | |
2303 | if (first) | |
2304 | { | |
2305 | first = false; | |
2306 | ctx = TREE_VALUE (TREE_VALUE (variant1)); | |
2307 | need_two = omp_context_compute_score (ctx, &max_score1, false); | |
2308 | if (need_two) | |
2309 | omp_context_compute_score (ctx, &max_score2, true); | |
2310 | else | |
2311 | max_score2 = max_score1; | |
2312 | } | |
2313 | ctx = TREE_VALUE (TREE_VALUE (attr1)); | |
2314 | need_two = omp_context_compute_score (ctx, &score1, false); | |
2315 | if (need_two) | |
2316 | omp_context_compute_score (ctx, &score2, true); | |
2317 | else | |
2318 | score2 = score1; | |
2319 | if (score1 > max_score1) | |
2320 | { | |
2321 | max_score1 = score1; | |
2322 | variant1 = attr1; | |
2323 | } | |
2324 | if (score2 > max_score2) | |
2325 | { | |
2326 | max_score2 = score2; | |
2327 | variant2 = attr1; | |
2328 | } | |
2329 | } | |
2330 | else | |
2331 | { | |
2332 | variant1 = attr1; | |
2333 | variant2 = attr1; | |
2334 | } | |
917dd789 | 2335 | } |
d0c464d2 JJ |
2336 | /* If there is a disagreement on which variant has the highest score |
2337 | depending on whether it will be in a declare simd clone or not, | |
2338 | punt for now and defer until after IPA where we will know that. */ | |
2339 | return ((variant1 && variant1 == variant2) | |
2340 | ? TREE_PURPOSE (TREE_VALUE (variant1)) : base); | |
135df52c JJ |
2341 | } |
2342 | ||
f165ef89 JJ |
2343 | void |
2344 | omp_lto_output_declare_variant_alt (lto_simple_output_block *ob, | |
2345 | cgraph_node *node, | |
2346 | lto_symtab_encoder_t encoder) | |
2347 | { | |
2348 | gcc_assert (node->declare_variant_alt); | |
2349 | ||
2350 | omp_declare_variant_base_entry entry; | |
2351 | entry.base = NULL; | |
2352 | entry.node = node; | |
2353 | entry.variants = NULL; | |
2354 | omp_declare_variant_base_entry *entryp | |
2355 | = omp_declare_variant_alt->find_with_hash (&entry, DECL_UID (node->decl)); | |
2356 | gcc_assert (entryp); | |
2357 | ||
2358 | int nbase = lto_symtab_encoder_lookup (encoder, entryp->base); | |
2359 | gcc_assert (nbase != LCC_NOT_FOUND); | |
2360 | streamer_write_hwi_stream (ob->main_stream, nbase); | |
2361 | ||
2362 | streamer_write_hwi_stream (ob->main_stream, entryp->variants->length ()); | |
2363 | ||
2364 | unsigned int i; | |
2365 | omp_declare_variant_entry *varentry; | |
2366 | FOR_EACH_VEC_SAFE_ELT (entryp->variants, i, varentry) | |
2367 | { | |
2368 | int nvar = lto_symtab_encoder_lookup (encoder, varentry->variant); | |
2369 | gcc_assert (nvar != LCC_NOT_FOUND); | |
2370 | streamer_write_hwi_stream (ob->main_stream, nvar); | |
2371 | ||
2372 | for (widest_int *w = &varentry->score; ; | |
2373 | w = &varentry->score_in_declare_simd_clone) | |
2374 | { | |
2375 | unsigned len = w->get_len (); | |
2376 | streamer_write_hwi_stream (ob->main_stream, len); | |
2377 | const HOST_WIDE_INT *val = w->get_val (); | |
2378 | for (unsigned j = 0; j < len; j++) | |
2379 | streamer_write_hwi_stream (ob->main_stream, val[j]); | |
2380 | if (w == &varentry->score_in_declare_simd_clone) | |
2381 | break; | |
2382 | } | |
2383 | ||
2384 | HOST_WIDE_INT cnt = -1; | |
2385 | HOST_WIDE_INT i = varentry->matches ? 1 : 0; | |
2386 | for (tree attr = DECL_ATTRIBUTES (entryp->base->decl); | |
2387 | attr; attr = TREE_CHAIN (attr), i += 2) | |
2388 | { | |
2389 | attr = lookup_attribute ("omp declare variant base", attr); | |
2390 | if (attr == NULL_TREE) | |
2391 | break; | |
2392 | ||
2393 | if (varentry->ctx == TREE_VALUE (TREE_VALUE (attr))) | |
2394 | { | |
2395 | cnt = i; | |
2396 | break; | |
2397 | } | |
2398 | } | |
2399 | ||
2400 | gcc_assert (cnt != -1); | |
2401 | streamer_write_hwi_stream (ob->main_stream, cnt); | |
2402 | } | |
2403 | } | |
2404 | ||
2405 | void | |
2406 | omp_lto_input_declare_variant_alt (lto_input_block *ib, cgraph_node *node, | |
2407 | vec<symtab_node *> nodes) | |
2408 | { | |
2409 | gcc_assert (node->declare_variant_alt); | |
2410 | omp_declare_variant_base_entry *entryp | |
2411 | = ggc_cleared_alloc<omp_declare_variant_base_entry> (); | |
2412 | entryp->base = dyn_cast<cgraph_node *> (nodes[streamer_read_hwi (ib)]); | |
2413 | entryp->node = node; | |
2414 | unsigned int len = streamer_read_hwi (ib); | |
2415 | vec_alloc (entryp->variants, len); | |
2416 | ||
2417 | for (unsigned int i = 0; i < len; i++) | |
2418 | { | |
2419 | omp_declare_variant_entry varentry; | |
2420 | varentry.variant | |
2421 | = dyn_cast<cgraph_node *> (nodes[streamer_read_hwi (ib)]); | |
2422 | for (widest_int *w = &varentry.score; ; | |
2423 | w = &varentry.score_in_declare_simd_clone) | |
2424 | { | |
2425 | unsigned len2 = streamer_read_hwi (ib); | |
2426 | HOST_WIDE_INT arr[WIDE_INT_MAX_ELTS]; | |
2427 | gcc_assert (len2 <= WIDE_INT_MAX_ELTS); | |
2428 | for (unsigned int j = 0; j < len2; j++) | |
2429 | arr[j] = streamer_read_hwi (ib); | |
2430 | *w = widest_int::from_array (arr, len2, true); | |
2431 | if (w == &varentry.score_in_declare_simd_clone) | |
2432 | break; | |
2433 | } | |
2434 | ||
2435 | HOST_WIDE_INT cnt = streamer_read_hwi (ib); | |
2436 | HOST_WIDE_INT j = 0; | |
2437 | varentry.ctx = NULL_TREE; | |
2438 | varentry.matches = (cnt & 1) ? true : false; | |
2439 | cnt &= ~HOST_WIDE_INT_1; | |
2440 | for (tree attr = DECL_ATTRIBUTES (entryp->base->decl); | |
2441 | attr; attr = TREE_CHAIN (attr), j += 2) | |
2442 | { | |
2443 | attr = lookup_attribute ("omp declare variant base", attr); | |
2444 | if (attr == NULL_TREE) | |
2445 | break; | |
2446 | ||
2447 | if (cnt == j) | |
2448 | { | |
2449 | varentry.ctx = TREE_VALUE (TREE_VALUE (attr)); | |
2450 | break; | |
2451 | } | |
2452 | } | |
2453 | gcc_assert (varentry.ctx != NULL_TREE); | |
2454 | entryp->variants->quick_push (varentry); | |
2455 | } | |
2456 | if (omp_declare_variant_alt == NULL) | |
2457 | omp_declare_variant_alt | |
2458 | = hash_table<omp_declare_variant_alt_hasher>::create_ggc (64); | |
2459 | *omp_declare_variant_alt->find_slot_with_hash (entryp, DECL_UID (node->decl), | |
2460 | INSERT) = entryp; | |
2461 | } | |
135df52c | 2462 | |
629b3d75 MJ |
2463 | /* Encode an oacc launch argument. This matches the GOMP_LAUNCH_PACK |
2464 | macro on gomp-constants.h. We do not check for overflow. */ | |
2465 | ||
2466 | tree | |
2467 | oacc_launch_pack (unsigned code, tree device, unsigned op) | |
2468 | { | |
2469 | tree res; | |
2470 | ||
2471 | res = build_int_cst (unsigned_type_node, GOMP_LAUNCH_PACK (code, 0, op)); | |
2472 | if (device) | |
2473 | { | |
2474 | device = fold_build2 (LSHIFT_EXPR, unsigned_type_node, | |
2475 | device, build_int_cst (unsigned_type_node, | |
2476 | GOMP_LAUNCH_DEVICE_SHIFT)); | |
2477 | res = fold_build2 (BIT_IOR_EXPR, unsigned_type_node, res, device); | |
2478 | } | |
2479 | return res; | |
2480 | } | |
2481 | ||
2482 | /* FIXME: What is the following comment for? */ | |
2483 | /* Look for compute grid dimension clauses and convert to an attribute | |
2484 | attached to FN. This permits the target-side code to (a) massage | |
2485 | the dimensions, (b) emit that data and (c) optimize. Non-constant | |
2486 | dimensions are pushed onto ARGS. | |
2487 | ||
2488 | The attribute value is a TREE_LIST. A set of dimensions is | |
2489 | represented as a list of INTEGER_CST. Those that are runtime | |
2490 | exprs are represented as an INTEGER_CST of zero. | |
2491 | ||
01914336 | 2492 | TODO: Normally the attribute will just contain a single such list. If |
629b3d75 MJ |
2493 | however it contains a list of lists, this will represent the use of |
2494 | device_type. Each member of the outer list is an assoc list of | |
2495 | dimensions, keyed by the device type. The first entry will be the | |
2496 | default. Well, that's the plan. */ | |
2497 | ||
2498 | /* Replace any existing oacc fn attribute with updated dimensions. */ | |
2499 | ||
68034b1b TS |
2500 | /* Variant working on a list of attributes. */ |
2501 | ||
2502 | tree | |
2503 | oacc_replace_fn_attrib_attr (tree attribs, tree dims) | |
629b3d75 MJ |
2504 | { |
2505 | tree ident = get_identifier (OACC_FN_ATTRIB); | |
629b3d75 MJ |
2506 | |
2507 | /* If we happen to be present as the first attrib, drop it. */ | |
2508 | if (attribs && TREE_PURPOSE (attribs) == ident) | |
2509 | attribs = TREE_CHAIN (attribs); | |
68034b1b TS |
2510 | return tree_cons (ident, dims, attribs); |
2511 | } | |
2512 | ||
2513 | /* Variant working on a function decl. */ | |
2514 | ||
2515 | void | |
2516 | oacc_replace_fn_attrib (tree fn, tree dims) | |
2517 | { | |
2518 | DECL_ATTRIBUTES (fn) | |
2519 | = oacc_replace_fn_attrib_attr (DECL_ATTRIBUTES (fn), dims); | |
629b3d75 MJ |
2520 | } |
2521 | ||
2522 | /* Scan CLAUSES for launch dimensions and attach them to the oacc | |
2523 | function attribute. Push any that are non-constant onto the ARGS | |
25651634 | 2524 | list, along with an appropriate GOMP_LAUNCH_DIM tag. */ |
629b3d75 MJ |
2525 | |
2526 | void | |
25651634 | 2527 | oacc_set_fn_attrib (tree fn, tree clauses, vec<tree> *args) |
629b3d75 MJ |
2528 | { |
2529 | /* Must match GOMP_DIM ordering. */ | |
2530 | static const omp_clause_code ids[] | |
2531 | = { OMP_CLAUSE_NUM_GANGS, OMP_CLAUSE_NUM_WORKERS, | |
2532 | OMP_CLAUSE_VECTOR_LENGTH }; | |
2533 | unsigned ix; | |
2534 | tree dims[GOMP_DIM_MAX]; | |
2535 | ||
2536 | tree attr = NULL_TREE; | |
2537 | unsigned non_const = 0; | |
2538 | ||
2539 | for (ix = GOMP_DIM_MAX; ix--;) | |
2540 | { | |
2541 | tree clause = omp_find_clause (clauses, ids[ix]); | |
2542 | tree dim = NULL_TREE; | |
2543 | ||
2544 | if (clause) | |
2545 | dim = OMP_CLAUSE_EXPR (clause, ids[ix]); | |
2546 | dims[ix] = dim; | |
2547 | if (dim && TREE_CODE (dim) != INTEGER_CST) | |
2548 | { | |
2549 | dim = integer_zero_node; | |
2550 | non_const |= GOMP_DIM_MASK (ix); | |
2551 | } | |
2552 | attr = tree_cons (NULL_TREE, dim, attr); | |
629b3d75 MJ |
2553 | } |
2554 | ||
2555 | oacc_replace_fn_attrib (fn, attr); | |
2556 | ||
2557 | if (non_const) | |
2558 | { | |
2559 | /* Push a dynamic argument set. */ | |
2560 | args->safe_push (oacc_launch_pack (GOMP_LAUNCH_DIM, | |
2561 | NULL_TREE, non_const)); | |
2562 | for (unsigned ix = 0; ix != GOMP_DIM_MAX; ix++) | |
2563 | if (non_const & GOMP_DIM_MASK (ix)) | |
2564 | args->safe_push (dims[ix]); | |
2565 | } | |
2566 | } | |
2567 | ||
5bf04509 TS |
2568 | /* Verify OpenACC routine clauses. |
2569 | ||
b48f44bf TS |
2570 | Returns 0 if FNDECL should be marked with an OpenACC 'routine' directive, 1 |
2571 | if it has already been marked in compatible way, and -1 if incompatible. | |
5bf04509 TS |
2572 | Upon returning, the chain of clauses will contain exactly one clause |
2573 | specifying the level of parallelism. */ | |
2574 | ||
b48f44bf TS |
2575 | int |
2576 | oacc_verify_routine_clauses (tree fndecl, tree *clauses, location_t loc, | |
2577 | const char *routine_str) | |
5bf04509 TS |
2578 | { |
2579 | tree c_level = NULL_TREE; | |
a61f6afb | 2580 | tree c_nohost = NULL_TREE; |
5bf04509 TS |
2581 | tree c_p = NULL_TREE; |
2582 | for (tree c = *clauses; c; c_p = c, c = OMP_CLAUSE_CHAIN (c)) | |
2583 | switch (OMP_CLAUSE_CODE (c)) | |
2584 | { | |
2585 | case OMP_CLAUSE_GANG: | |
2586 | case OMP_CLAUSE_WORKER: | |
2587 | case OMP_CLAUSE_VECTOR: | |
2588 | case OMP_CLAUSE_SEQ: | |
2589 | if (c_level == NULL_TREE) | |
2590 | c_level = c; | |
2591 | else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_CODE (c_level)) | |
2592 | { | |
2593 | /* This has already been diagnosed in the front ends. */ | |
2594 | /* Drop the duplicate clause. */ | |
2595 | gcc_checking_assert (c_p != NULL_TREE); | |
2596 | OMP_CLAUSE_CHAIN (c_p) = OMP_CLAUSE_CHAIN (c); | |
2597 | c = c_p; | |
2598 | } | |
2599 | else | |
2600 | { | |
2601 | error_at (OMP_CLAUSE_LOCATION (c), | |
2602 | "%qs specifies a conflicting level of parallelism", | |
2603 | omp_clause_code_name[OMP_CLAUSE_CODE (c)]); | |
2604 | inform (OMP_CLAUSE_LOCATION (c_level), | |
2605 | "... to the previous %qs clause here", | |
2606 | omp_clause_code_name[OMP_CLAUSE_CODE (c_level)]); | |
2607 | /* Drop the conflicting clause. */ | |
2608 | gcc_checking_assert (c_p != NULL_TREE); | |
2609 | OMP_CLAUSE_CHAIN (c_p) = OMP_CLAUSE_CHAIN (c); | |
2610 | c = c_p; | |
2611 | } | |
2612 | break; | |
a61f6afb TS |
2613 | case OMP_CLAUSE_NOHOST: |
2614 | /* Don't worry about duplicate clauses here. */ | |
2615 | c_nohost = c; | |
2616 | break; | |
5bf04509 TS |
2617 | default: |
2618 | gcc_unreachable (); | |
2619 | } | |
2620 | if (c_level == NULL_TREE) | |
2621 | { | |
2622 | /* Default to an implicit 'seq' clause. */ | |
2623 | c_level = build_omp_clause (loc, OMP_CLAUSE_SEQ); | |
2624 | OMP_CLAUSE_CHAIN (c_level) = *clauses; | |
2625 | *clauses = c_level; | |
2626 | } | |
b48f44bf TS |
2627 | /* In *clauses, we now have exactly one clause specifying the level of |
2628 | parallelism. */ | |
2629 | ||
2630 | tree attr | |
2631 | = lookup_attribute ("omp declare target", DECL_ATTRIBUTES (fndecl)); | |
2632 | if (attr != NULL_TREE) | |
2633 | { | |
ff3f862b TS |
2634 | /* Diagnose if "#pragma omp declare target" has also been applied. */ |
2635 | if (TREE_VALUE (attr) == NULL_TREE) | |
2636 | { | |
2637 | /* See <https://gcc.gnu.org/PR93465>; the semantics of combining | |
2638 | OpenACC and OpenMP 'target' are not clear. */ | |
2639 | error_at (loc, | |
2640 | "cannot apply %<%s%> to %qD, which has also been" | |
2641 | " marked with an OpenMP 'declare target' directive", | |
2642 | routine_str, fndecl); | |
2643 | /* Incompatible. */ | |
2644 | return -1; | |
2645 | } | |
2646 | ||
b48f44bf TS |
2647 | /* If a "#pragma acc routine" has already been applied, just verify |
2648 | this one for compatibility. */ | |
2649 | /* Collect previous directive's clauses. */ | |
2650 | tree c_level_p = NULL_TREE; | |
a61f6afb | 2651 | tree c_nohost_p = NULL_TREE; |
b48f44bf TS |
2652 | for (tree c = TREE_VALUE (attr); c; c = OMP_CLAUSE_CHAIN (c)) |
2653 | switch (OMP_CLAUSE_CODE (c)) | |
2654 | { | |
2655 | case OMP_CLAUSE_GANG: | |
2656 | case OMP_CLAUSE_WORKER: | |
2657 | case OMP_CLAUSE_VECTOR: | |
2658 | case OMP_CLAUSE_SEQ: | |
2659 | gcc_checking_assert (c_level_p == NULL_TREE); | |
2660 | c_level_p = c; | |
2661 | break; | |
a61f6afb TS |
2662 | case OMP_CLAUSE_NOHOST: |
2663 | gcc_checking_assert (c_nohost_p == NULL_TREE); | |
2664 | c_nohost_p = c; | |
2665 | break; | |
b48f44bf TS |
2666 | default: |
2667 | gcc_unreachable (); | |
2668 | } | |
2669 | gcc_checking_assert (c_level_p != NULL_TREE); | |
2670 | /* ..., and compare to current directive's, which we've already collected | |
2671 | above. */ | |
2672 | tree c_diag; | |
2673 | tree c_diag_p; | |
2674 | /* Matching level of parallelism? */ | |
2675 | if (OMP_CLAUSE_CODE (c_level) != OMP_CLAUSE_CODE (c_level_p)) | |
2676 | { | |
2677 | c_diag = c_level; | |
2678 | c_diag_p = c_level_p; | |
2679 | goto incompatible; | |
2680 | } | |
a61f6afb TS |
2681 | /* Matching 'nohost' clauses? */ |
2682 | if ((c_nohost == NULL_TREE) != (c_nohost_p == NULL_TREE)) | |
2683 | { | |
2684 | c_diag = c_nohost; | |
2685 | c_diag_p = c_nohost_p; | |
2686 | goto incompatible; | |
2687 | } | |
b48f44bf TS |
2688 | /* Compatible. */ |
2689 | return 1; | |
2690 | ||
2691 | incompatible: | |
2692 | if (c_diag != NULL_TREE) | |
2693 | error_at (OMP_CLAUSE_LOCATION (c_diag), | |
2694 | "incompatible %qs clause when applying" | |
2695 | " %<%s%> to %qD, which has already been" | |
2696 | " marked with an OpenACC 'routine' directive", | |
2697 | omp_clause_code_name[OMP_CLAUSE_CODE (c_diag)], | |
2698 | routine_str, fndecl); | |
2699 | else if (c_diag_p != NULL_TREE) | |
2700 | error_at (loc, | |
2701 | "missing %qs clause when applying" | |
2702 | " %<%s%> to %qD, which has already been" | |
2703 | " marked with an OpenACC 'routine' directive", | |
2704 | omp_clause_code_name[OMP_CLAUSE_CODE (c_diag_p)], | |
2705 | routine_str, fndecl); | |
2706 | else | |
2707 | gcc_unreachable (); | |
2708 | if (c_diag_p != NULL_TREE) | |
2709 | inform (OMP_CLAUSE_LOCATION (c_diag_p), | |
2710 | "... with %qs clause here", | |
2711 | omp_clause_code_name[OMP_CLAUSE_CODE (c_diag_p)]); | |
2712 | else | |
2713 | { | |
2714 | /* In the front ends, we don't preserve location information for the | |
2715 | OpenACC routine directive itself. However, that of c_level_p | |
2716 | should be close. */ | |
2717 | location_t loc_routine = OMP_CLAUSE_LOCATION (c_level_p); | |
2718 | inform (loc_routine, "... without %qs clause near to here", | |
2719 | omp_clause_code_name[OMP_CLAUSE_CODE (c_diag)]); | |
2720 | } | |
2721 | /* Incompatible. */ | |
2722 | return -1; | |
2723 | } | |
2724 | ||
2725 | return 0; | |
5bf04509 TS |
2726 | } |
2727 | ||
2728 | /* Process the OpenACC 'routine' directive clauses to generate an attribute | |
2729 | for the level of parallelism. All dimensions have a size of zero | |
629b3d75 MJ |
2730 | (dynamic). TREE_PURPOSE is set to indicate whether that dimension |
2731 | can have a loop partitioned on it. non-zero indicates | |
2732 | yes, zero indicates no. By construction once a non-zero has been | |
2733 | reached, further inner dimensions must also be non-zero. We set | |
2734 | TREE_VALUE to zero for the dimensions that may be partitioned and | |
2735 | 1 for the other ones -- if a loop is (erroneously) spawned at | |
2736 | an outer level, we don't want to try and partition it. */ | |
2737 | ||
2738 | tree | |
2739 | oacc_build_routine_dims (tree clauses) | |
2740 | { | |
2741 | /* Must match GOMP_DIM ordering. */ | |
01914336 MJ |
2742 | static const omp_clause_code ids[] |
2743 | = {OMP_CLAUSE_GANG, OMP_CLAUSE_WORKER, OMP_CLAUSE_VECTOR, OMP_CLAUSE_SEQ}; | |
629b3d75 MJ |
2744 | int ix; |
2745 | int level = -1; | |
2746 | ||
2747 | for (; clauses; clauses = OMP_CLAUSE_CHAIN (clauses)) | |
2748 | for (ix = GOMP_DIM_MAX + 1; ix--;) | |
2749 | if (OMP_CLAUSE_CODE (clauses) == ids[ix]) | |
2750 | { | |
629b3d75 MJ |
2751 | level = ix; |
2752 | break; | |
2753 | } | |
5bf04509 | 2754 | gcc_checking_assert (level >= 0); |
629b3d75 MJ |
2755 | |
2756 | tree dims = NULL_TREE; | |
2757 | ||
2758 | for (ix = GOMP_DIM_MAX; ix--;) | |
2759 | dims = tree_cons (build_int_cst (boolean_type_node, ix >= level), | |
2760 | build_int_cst (integer_type_node, ix < level), dims); | |
2761 | ||
2762 | return dims; | |
2763 | } | |
2764 | ||
2765 | /* Retrieve the oacc function attrib and return it. Non-oacc | |
2766 | functions will return NULL. */ | |
2767 | ||
2768 | tree | |
2769 | oacc_get_fn_attrib (tree fn) | |
2770 | { | |
2771 | return lookup_attribute (OACC_FN_ATTRIB, DECL_ATTRIBUTES (fn)); | |
2772 | } | |
2773 | ||
46dbeb40 TV |
2774 | /* Return true if FN is an OpenMP or OpenACC offloading function. */ |
2775 | ||
2776 | bool | |
2777 | offloading_function_p (tree fn) | |
2778 | { | |
2779 | tree attrs = DECL_ATTRIBUTES (fn); | |
2780 | return (lookup_attribute ("omp declare target", attrs) | |
2781 | || lookup_attribute ("omp target entrypoint", attrs)); | |
2782 | } | |
2783 | ||
629b3d75 MJ |
2784 | /* Extract an oacc execution dimension from FN. FN must be an |
2785 | offloaded function or routine that has already had its execution | |
2786 | dimensions lowered to the target-specific values. */ | |
2787 | ||
2788 | int | |
2789 | oacc_get_fn_dim_size (tree fn, int axis) | |
2790 | { | |
2791 | tree attrs = oacc_get_fn_attrib (fn); | |
2792 | ||
2793 | gcc_assert (axis < GOMP_DIM_MAX); | |
2794 | ||
2795 | tree dims = TREE_VALUE (attrs); | |
2796 | while (axis--) | |
2797 | dims = TREE_CHAIN (dims); | |
2798 | ||
2799 | int size = TREE_INT_CST_LOW (TREE_VALUE (dims)); | |
2800 | ||
2801 | return size; | |
2802 | } | |
2803 | ||
2804 | /* Extract the dimension axis from an IFN_GOACC_DIM_POS or | |
2805 | IFN_GOACC_DIM_SIZE call. */ | |
2806 | ||
2807 | int | |
2808 | oacc_get_ifn_dim_arg (const gimple *stmt) | |
2809 | { | |
2810 | gcc_checking_assert (gimple_call_internal_fn (stmt) == IFN_GOACC_DIM_SIZE | |
2811 | || gimple_call_internal_fn (stmt) == IFN_GOACC_DIM_POS); | |
2812 | tree arg = gimple_call_arg (stmt, 0); | |
2813 | HOST_WIDE_INT axis = TREE_INT_CST_LOW (arg); | |
2814 | ||
2815 | gcc_checking_assert (axis >= 0 && axis < GOMP_DIM_MAX); | |
2816 | return (int) axis; | |
2817 | } | |
7a50e708 JJ |
2818 | |
2819 | #include "gt-omp-general.h" |