]>
Commit | Line | Data |
---|---|---|
1e8e9920 | 1 | /* Lowering pass for OpenMP directives. Converts OpenMP directives |
2 | into explicit calls to the runtime library (libgomp) and data | |
3 | marshalling to implement data sharing and copying clauses. | |
4 | Contributed by Diego Novillo <dnovillo@redhat.com> | |
5 | ||
711789cc | 6 | Copyright (C) 2005-2013 Free Software Foundation, Inc. |
1e8e9920 | 7 | |
8 | This file is part of GCC. | |
9 | ||
10 | GCC is free software; you can redistribute it and/or modify it under | |
11 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 12 | Software Foundation; either version 3, or (at your option) any later |
1e8e9920 | 13 | version. |
14 | ||
15 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
16 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
17 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
18 | for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 21 | along with GCC; see the file COPYING3. If not see |
22 | <http://www.gnu.org/licenses/>. */ | |
1e8e9920 | 23 | |
24 | #include "config.h" | |
25 | #include "system.h" | |
26 | #include "coretypes.h" | |
27 | #include "tm.h" | |
28 | #include "tree.h" | |
9ed99284 | 29 | #include "stringpool.h" |
30 | #include "stor-layout.h" | |
1e8e9920 | 31 | #include "rtl.h" |
bc61cadb | 32 | #include "pointer-set.h" |
33 | #include "basic-block.h" | |
34 | #include "tree-ssa-alias.h" | |
35 | #include "internal-fn.h" | |
36 | #include "gimple-fold.h" | |
37 | #include "gimple-expr.h" | |
38 | #include "is-a.h" | |
e795d6e1 | 39 | #include "gimple.h" |
a8783bee | 40 | #include "gimplify.h" |
dcf1a1ec | 41 | #include "gimple-iterator.h" |
e795d6e1 | 42 | #include "gimplify-me.h" |
dcf1a1ec | 43 | #include "gimple-walk.h" |
75a70cf9 | 44 | #include "tree-iterator.h" |
1e8e9920 | 45 | #include "tree-inline.h" |
46 | #include "langhooks.h" | |
852f689e | 47 | #include "diagnostic-core.h" |
073c1fd5 | 48 | #include "gimple-ssa.h" |
49 | #include "cgraph.h" | |
50 | #include "tree-cfg.h" | |
51 | #include "tree-phinodes.h" | |
52 | #include "ssa-iterators.h" | |
53 | #include "tree-ssanames.h" | |
54 | #include "tree-into-ssa.h" | |
9ed99284 | 55 | #include "expr.h" |
073c1fd5 | 56 | #include "tree-dfa.h" |
69ee5dbb | 57 | #include "tree-ssa.h" |
1e8e9920 | 58 | #include "flags.h" |
59 | #include "function.h" | |
60 | #include "expr.h" | |
1e8e9920 | 61 | #include "tree-pass.h" |
1e8e9920 | 62 | #include "except.h" |
e3022db7 | 63 | #include "splay-tree.h" |
cb7f680b | 64 | #include "optabs.h" |
65 | #include "cfgloop.h" | |
3d483a94 | 66 | #include "target.h" |
7740abd8 | 67 | #include "omp-low.h" |
424a4a92 | 68 | #include "gimple-low.h" |
69 | #include "tree-cfgcleanup.h" | |
d09768a4 | 70 | #include "pretty-print.h" |
71 | #include "ipa-prop.h" | |
e797f49f | 72 | #include "tree-nested.h" |
d09768a4 | 73 | #include "tree-eh.h" |
1e8e9920 | 74 | |
75a70cf9 | 75 | |
48e1416a | 76 | /* Lowering of OpenMP parallel and workshare constructs proceeds in two |
1e8e9920 | 77 | phases. The first phase scans the function looking for OMP statements |
78 | and then for variables that must be replaced to satisfy data sharing | |
79 | clauses. The second phase expands code for the constructs, as well as | |
334ec2d8 | 80 | re-gimplifying things when variables have been replaced with complex |
1e8e9920 | 81 | expressions. |
82 | ||
d134bccc | 83 | Final code generation is done by pass_expand_omp. The flowgraph is |
84 | scanned for parallel regions which are then moved to a new | |
85 | function, to be invoked by the thread library. */ | |
1e8e9920 | 86 | |
7740abd8 | 87 | /* Parallel region information. Every parallel and workshare |
88 | directive is enclosed between two markers, the OMP_* directive | |
89 | and a corresponding OMP_RETURN statement. */ | |
90 | ||
91 | struct omp_region | |
92 | { | |
93 | /* The enclosing region. */ | |
94 | struct omp_region *outer; | |
95 | ||
96 | /* First child region. */ | |
97 | struct omp_region *inner; | |
98 | ||
99 | /* Next peer region. */ | |
100 | struct omp_region *next; | |
101 | ||
102 | /* Block containing the omp directive as its last stmt. */ | |
103 | basic_block entry; | |
104 | ||
105 | /* Block containing the OMP_RETURN as its last stmt. */ | |
106 | basic_block exit; | |
107 | ||
108 | /* Block containing the OMP_CONTINUE as its last stmt. */ | |
109 | basic_block cont; | |
110 | ||
111 | /* If this is a combined parallel+workshare region, this is a list | |
112 | of additional arguments needed by the combined parallel+workshare | |
113 | library call. */ | |
114 | vec<tree, va_gc> *ws_args; | |
115 | ||
116 | /* The code for the omp directive of this region. */ | |
117 | enum gimple_code type; | |
118 | ||
119 | /* Schedule kind, only used for OMP_FOR type regions. */ | |
120 | enum omp_clause_schedule_kind sched_kind; | |
121 | ||
122 | /* True if this is a combined parallel+workshare region. */ | |
123 | bool is_combined_parallel; | |
124 | }; | |
125 | ||
1e8e9920 | 126 | /* Context structure. Used to store information about each parallel |
127 | directive in the code. */ | |
128 | ||
129 | typedef struct omp_context | |
130 | { | |
131 | /* This field must be at the beginning, as we do "inheritance": Some | |
132 | callback functions for tree-inline.c (e.g., omp_copy_decl) | |
133 | receive a copy_body_data pointer that is up-casted to an | |
134 | omp_context pointer. */ | |
135 | copy_body_data cb; | |
136 | ||
137 | /* The tree of contexts corresponding to the encountered constructs. */ | |
138 | struct omp_context *outer; | |
75a70cf9 | 139 | gimple stmt; |
1e8e9920 | 140 | |
48e1416a | 141 | /* Map variables to fields in a structure that allows communication |
1e8e9920 | 142 | between sending and receiving threads. */ |
143 | splay_tree field_map; | |
144 | tree record_type; | |
145 | tree sender_decl; | |
146 | tree receiver_decl; | |
147 | ||
fd6481cf | 148 | /* These are used just by task contexts, if task firstprivate fn is |
149 | needed. srecord_type is used to communicate from the thread | |
150 | that encountered the task construct to task firstprivate fn, | |
151 | record_type is allocated by GOMP_task, initialized by task firstprivate | |
152 | fn and passed to the task body fn. */ | |
153 | splay_tree sfield_map; | |
154 | tree srecord_type; | |
155 | ||
1e8e9920 | 156 | /* A chain of variables to add to the top-level block surrounding the |
157 | construct. In the case of a parallel, this is in the child function. */ | |
158 | tree block_vars; | |
159 | ||
bc7bff74 | 160 | /* Label to which GOMP_cancel{,llation_point} and explicit and implicit |
161 | barriers should jump to during omplower pass. */ | |
162 | tree cancel_label; | |
163 | ||
1e8e9920 | 164 | /* What to do with variables with implicitly determined sharing |
165 | attributes. */ | |
166 | enum omp_clause_default_kind default_kind; | |
167 | ||
168 | /* Nesting depth of this context. Used to beautify error messages re | |
169 | invalid gotos. The outermost ctx is depth 1, with depth 0 being | |
170 | reserved for the main body of the function. */ | |
171 | int depth; | |
172 | ||
1e8e9920 | 173 | /* True if this parallel directive is nested within another. */ |
174 | bool is_nested; | |
bc7bff74 | 175 | |
176 | /* True if this construct can be cancelled. */ | |
177 | bool cancellable; | |
1e8e9920 | 178 | } omp_context; |
179 | ||
180 | ||
fd6481cf | 181 | struct omp_for_data_loop |
182 | { | |
183 | tree v, n1, n2, step; | |
184 | enum tree_code cond_code; | |
185 | }; | |
186 | ||
773c5ba7 | 187 | /* A structure describing the main elements of a parallel loop. */ |
1e8e9920 | 188 | |
773c5ba7 | 189 | struct omp_for_data |
1e8e9920 | 190 | { |
fd6481cf | 191 | struct omp_for_data_loop loop; |
75a70cf9 | 192 | tree chunk_size; |
193 | gimple for_stmt; | |
fd6481cf | 194 | tree pre, iter_type; |
195 | int collapse; | |
1e8e9920 | 196 | bool have_nowait, have_ordered; |
197 | enum omp_clause_schedule_kind sched_kind; | |
fd6481cf | 198 | struct omp_for_data_loop *loops; |
1e8e9920 | 199 | }; |
200 | ||
773c5ba7 | 201 | |
1e8e9920 | 202 | static splay_tree all_contexts; |
fd6481cf | 203 | static int taskreg_nesting_level; |
bc7bff74 | 204 | static int target_nesting_level; |
7740abd8 | 205 | static struct omp_region *root_omp_region; |
fd6481cf | 206 | static bitmap task_shared_vars; |
1e8e9920 | 207 | |
ab129075 | 208 | static void scan_omp (gimple_seq *, omp_context *); |
75a70cf9 | 209 | static tree scan_omp_1_op (tree *, int *, void *); |
210 | ||
211 | #define WALK_SUBSTMTS \ | |
212 | case GIMPLE_BIND: \ | |
213 | case GIMPLE_TRY: \ | |
214 | case GIMPLE_CATCH: \ | |
215 | case GIMPLE_EH_FILTER: \ | |
4c0315d0 | 216 | case GIMPLE_TRANSACTION: \ |
75a70cf9 | 217 | /* The sub-statements for these should be walked. */ \ |
218 | *handled_ops_p = false; \ | |
219 | break; | |
220 | ||
221 | /* Convenience function for calling scan_omp_1_op on tree operands. */ | |
222 | ||
223 | static inline tree | |
224 | scan_omp_op (tree *tp, omp_context *ctx) | |
225 | { | |
226 | struct walk_stmt_info wi; | |
227 | ||
228 | memset (&wi, 0, sizeof (wi)); | |
229 | wi.info = ctx; | |
230 | wi.want_locations = true; | |
231 | ||
232 | return walk_tree (tp, scan_omp_1_op, &wi, NULL); | |
233 | } | |
234 | ||
e3a19533 | 235 | static void lower_omp (gimple_seq *, omp_context *); |
f49d7bb5 | 236 | static tree lookup_decl_in_outer_ctx (tree, omp_context *); |
237 | static tree maybe_lookup_decl_in_outer_ctx (tree, omp_context *); | |
1e8e9920 | 238 | |
239 | /* Find an OpenMP clause of type KIND within CLAUSES. */ | |
240 | ||
79acaae1 | 241 | tree |
590c3166 | 242 | find_omp_clause (tree clauses, enum omp_clause_code kind) |
1e8e9920 | 243 | { |
244 | for (; clauses ; clauses = OMP_CLAUSE_CHAIN (clauses)) | |
55d6e7cd | 245 | if (OMP_CLAUSE_CODE (clauses) == kind) |
1e8e9920 | 246 | return clauses; |
247 | ||
248 | return NULL_TREE; | |
249 | } | |
250 | ||
251 | /* Return true if CTX is for an omp parallel. */ | |
252 | ||
253 | static inline bool | |
254 | is_parallel_ctx (omp_context *ctx) | |
255 | { | |
75a70cf9 | 256 | return gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL; |
1e8e9920 | 257 | } |
258 | ||
773c5ba7 | 259 | |
fd6481cf | 260 | /* Return true if CTX is for an omp task. */ |
261 | ||
262 | static inline bool | |
263 | is_task_ctx (omp_context *ctx) | |
264 | { | |
75a70cf9 | 265 | return gimple_code (ctx->stmt) == GIMPLE_OMP_TASK; |
fd6481cf | 266 | } |
267 | ||
268 | ||
269 | /* Return true if CTX is for an omp parallel or omp task. */ | |
270 | ||
271 | static inline bool | |
272 | is_taskreg_ctx (omp_context *ctx) | |
273 | { | |
75a70cf9 | 274 | return gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL |
275 | || gimple_code (ctx->stmt) == GIMPLE_OMP_TASK; | |
fd6481cf | 276 | } |
277 | ||
278 | ||
773c5ba7 | 279 | /* Return true if REGION is a combined parallel+workshare region. */ |
1e8e9920 | 280 | |
281 | static inline bool | |
773c5ba7 | 282 | is_combined_parallel (struct omp_region *region) |
283 | { | |
284 | return region->is_combined_parallel; | |
285 | } | |
286 | ||
287 | ||
288 | /* Extract the header elements of parallel loop FOR_STMT and store | |
289 | them into *FD. */ | |
290 | ||
291 | static void | |
75a70cf9 | 292 | extract_omp_for_data (gimple for_stmt, struct omp_for_data *fd, |
fd6481cf | 293 | struct omp_for_data_loop *loops) |
773c5ba7 | 294 | { |
fd6481cf | 295 | tree t, var, *collapse_iter, *collapse_count; |
296 | tree count = NULL_TREE, iter_type = long_integer_type_node; | |
297 | struct omp_for_data_loop *loop; | |
298 | int i; | |
299 | struct omp_for_data_loop dummy_loop; | |
389dd41b | 300 | location_t loc = gimple_location (for_stmt); |
f2697631 | 301 | bool simd = gimple_omp_for_kind (for_stmt) & GF_OMP_FOR_KIND_SIMD; |
bc7bff74 | 302 | bool distribute = gimple_omp_for_kind (for_stmt) |
303 | == GF_OMP_FOR_KIND_DISTRIBUTE; | |
773c5ba7 | 304 | |
305 | fd->for_stmt = for_stmt; | |
306 | fd->pre = NULL; | |
75a70cf9 | 307 | fd->collapse = gimple_omp_for_collapse (for_stmt); |
fd6481cf | 308 | if (fd->collapse > 1) |
309 | fd->loops = loops; | |
310 | else | |
311 | fd->loops = &fd->loop; | |
773c5ba7 | 312 | |
bc7bff74 | 313 | fd->have_nowait = distribute || simd; |
314 | fd->have_ordered = false; | |
773c5ba7 | 315 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC; |
316 | fd->chunk_size = NULL_TREE; | |
fd6481cf | 317 | collapse_iter = NULL; |
318 | collapse_count = NULL; | |
773c5ba7 | 319 | |
75a70cf9 | 320 | for (t = gimple_omp_for_clauses (for_stmt); t ; t = OMP_CLAUSE_CHAIN (t)) |
55d6e7cd | 321 | switch (OMP_CLAUSE_CODE (t)) |
773c5ba7 | 322 | { |
323 | case OMP_CLAUSE_NOWAIT: | |
324 | fd->have_nowait = true; | |
325 | break; | |
326 | case OMP_CLAUSE_ORDERED: | |
327 | fd->have_ordered = true; | |
328 | break; | |
329 | case OMP_CLAUSE_SCHEDULE: | |
bc7bff74 | 330 | gcc_assert (!distribute); |
773c5ba7 | 331 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_KIND (t); |
332 | fd->chunk_size = OMP_CLAUSE_SCHEDULE_CHUNK_EXPR (t); | |
333 | break; | |
bc7bff74 | 334 | case OMP_CLAUSE_DIST_SCHEDULE: |
335 | gcc_assert (distribute); | |
336 | fd->chunk_size = OMP_CLAUSE_DIST_SCHEDULE_CHUNK_EXPR (t); | |
337 | break; | |
fd6481cf | 338 | case OMP_CLAUSE_COLLAPSE: |
339 | if (fd->collapse > 1) | |
340 | { | |
341 | collapse_iter = &OMP_CLAUSE_COLLAPSE_ITERVAR (t); | |
342 | collapse_count = &OMP_CLAUSE_COLLAPSE_COUNT (t); | |
343 | } | |
773c5ba7 | 344 | default: |
345 | break; | |
346 | } | |
347 | ||
fd6481cf | 348 | /* FIXME: for now map schedule(auto) to schedule(static). |
349 | There should be analysis to determine whether all iterations | |
350 | are approximately the same amount of work (then schedule(static) | |
bde357c8 | 351 | is best) or if it varies (then schedule(dynamic,N) is better). */ |
fd6481cf | 352 | if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_AUTO) |
353 | { | |
354 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC; | |
355 | gcc_assert (fd->chunk_size == NULL); | |
356 | } | |
357 | gcc_assert (fd->collapse == 1 || collapse_iter != NULL); | |
773c5ba7 | 358 | if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_RUNTIME) |
359 | gcc_assert (fd->chunk_size == NULL); | |
360 | else if (fd->chunk_size == NULL) | |
361 | { | |
362 | /* We only need to compute a default chunk size for ordered | |
363 | static loops and dynamic loops. */ | |
fd6481cf | 364 | if (fd->sched_kind != OMP_CLAUSE_SCHEDULE_STATIC |
bc7bff74 | 365 | || fd->have_ordered) |
773c5ba7 | 366 | fd->chunk_size = (fd->sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) |
367 | ? integer_zero_node : integer_one_node; | |
368 | } | |
fd6481cf | 369 | |
370 | for (i = 0; i < fd->collapse; i++) | |
371 | { | |
372 | if (fd->collapse == 1) | |
373 | loop = &fd->loop; | |
374 | else if (loops != NULL) | |
375 | loop = loops + i; | |
376 | else | |
377 | loop = &dummy_loop; | |
378 | ||
75a70cf9 | 379 | loop->v = gimple_omp_for_index (for_stmt, i); |
fd6481cf | 380 | gcc_assert (SSA_VAR_P (loop->v)); |
381 | gcc_assert (TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE | |
382 | || TREE_CODE (TREE_TYPE (loop->v)) == POINTER_TYPE); | |
383 | var = TREE_CODE (loop->v) == SSA_NAME ? SSA_NAME_VAR (loop->v) : loop->v; | |
75a70cf9 | 384 | loop->n1 = gimple_omp_for_initial (for_stmt, i); |
fd6481cf | 385 | |
75a70cf9 | 386 | loop->cond_code = gimple_omp_for_cond (for_stmt, i); |
387 | loop->n2 = gimple_omp_for_final (for_stmt, i); | |
fd6481cf | 388 | switch (loop->cond_code) |
389 | { | |
390 | case LT_EXPR: | |
391 | case GT_EXPR: | |
392 | break; | |
f2697631 | 393 | case NE_EXPR: |
394 | gcc_assert (gimple_omp_for_kind (for_stmt) | |
395 | == GF_OMP_FOR_KIND_CILKSIMD); | |
396 | break; | |
fd6481cf | 397 | case LE_EXPR: |
398 | if (POINTER_TYPE_P (TREE_TYPE (loop->n2))) | |
2cc66f2a | 399 | loop->n2 = fold_build_pointer_plus_hwi_loc (loc, loop->n2, 1); |
fd6481cf | 400 | else |
389dd41b | 401 | loop->n2 = fold_build2_loc (loc, |
402 | PLUS_EXPR, TREE_TYPE (loop->n2), loop->n2, | |
fd6481cf | 403 | build_int_cst (TREE_TYPE (loop->n2), 1)); |
404 | loop->cond_code = LT_EXPR; | |
405 | break; | |
406 | case GE_EXPR: | |
407 | if (POINTER_TYPE_P (TREE_TYPE (loop->n2))) | |
2cc66f2a | 408 | loop->n2 = fold_build_pointer_plus_hwi_loc (loc, loop->n2, -1); |
fd6481cf | 409 | else |
389dd41b | 410 | loop->n2 = fold_build2_loc (loc, |
411 | MINUS_EXPR, TREE_TYPE (loop->n2), loop->n2, | |
fd6481cf | 412 | build_int_cst (TREE_TYPE (loop->n2), 1)); |
413 | loop->cond_code = GT_EXPR; | |
414 | break; | |
415 | default: | |
416 | gcc_unreachable (); | |
417 | } | |
418 | ||
75a70cf9 | 419 | t = gimple_omp_for_incr (for_stmt, i); |
fd6481cf | 420 | gcc_assert (TREE_OPERAND (t, 0) == var); |
421 | switch (TREE_CODE (t)) | |
422 | { | |
423 | case PLUS_EXPR: | |
fd6481cf | 424 | loop->step = TREE_OPERAND (t, 1); |
425 | break; | |
85d86b55 | 426 | case POINTER_PLUS_EXPR: |
427 | loop->step = fold_convert (ssizetype, TREE_OPERAND (t, 1)); | |
428 | break; | |
fd6481cf | 429 | case MINUS_EXPR: |
430 | loop->step = TREE_OPERAND (t, 1); | |
389dd41b | 431 | loop->step = fold_build1_loc (loc, |
432 | NEGATE_EXPR, TREE_TYPE (loop->step), | |
fd6481cf | 433 | loop->step); |
434 | break; | |
435 | default: | |
436 | gcc_unreachable (); | |
437 | } | |
438 | ||
bc7bff74 | 439 | if (simd |
440 | || (fd->sched_kind == OMP_CLAUSE_SCHEDULE_STATIC | |
441 | && !fd->have_ordered)) | |
3d483a94 | 442 | { |
443 | if (fd->collapse == 1) | |
444 | iter_type = TREE_TYPE (loop->v); | |
445 | else if (i == 0 | |
446 | || TYPE_PRECISION (iter_type) | |
447 | < TYPE_PRECISION (TREE_TYPE (loop->v))) | |
448 | iter_type | |
449 | = build_nonstandard_integer_type | |
bc7bff74 | 450 | (TYPE_PRECISION (TREE_TYPE (loop->v)), 1); |
3d483a94 | 451 | } |
452 | else if (iter_type != long_long_unsigned_type_node) | |
fd6481cf | 453 | { |
454 | if (POINTER_TYPE_P (TREE_TYPE (loop->v))) | |
455 | iter_type = long_long_unsigned_type_node; | |
456 | else if (TYPE_UNSIGNED (TREE_TYPE (loop->v)) | |
457 | && TYPE_PRECISION (TREE_TYPE (loop->v)) | |
458 | >= TYPE_PRECISION (iter_type)) | |
459 | { | |
460 | tree n; | |
461 | ||
462 | if (loop->cond_code == LT_EXPR) | |
389dd41b | 463 | n = fold_build2_loc (loc, |
464 | PLUS_EXPR, TREE_TYPE (loop->v), | |
fd6481cf | 465 | loop->n2, loop->step); |
466 | else | |
467 | n = loop->n1; | |
468 | if (TREE_CODE (n) != INTEGER_CST | |
469 | || tree_int_cst_lt (TYPE_MAX_VALUE (iter_type), n)) | |
470 | iter_type = long_long_unsigned_type_node; | |
471 | } | |
472 | else if (TYPE_PRECISION (TREE_TYPE (loop->v)) | |
473 | > TYPE_PRECISION (iter_type)) | |
474 | { | |
475 | tree n1, n2; | |
476 | ||
477 | if (loop->cond_code == LT_EXPR) | |
478 | { | |
479 | n1 = loop->n1; | |
389dd41b | 480 | n2 = fold_build2_loc (loc, |
481 | PLUS_EXPR, TREE_TYPE (loop->v), | |
fd6481cf | 482 | loop->n2, loop->step); |
483 | } | |
484 | else | |
485 | { | |
389dd41b | 486 | n1 = fold_build2_loc (loc, |
487 | MINUS_EXPR, TREE_TYPE (loop->v), | |
fd6481cf | 488 | loop->n2, loop->step); |
489 | n2 = loop->n1; | |
490 | } | |
491 | if (TREE_CODE (n1) != INTEGER_CST | |
492 | || TREE_CODE (n2) != INTEGER_CST | |
493 | || !tree_int_cst_lt (TYPE_MIN_VALUE (iter_type), n1) | |
494 | || !tree_int_cst_lt (n2, TYPE_MAX_VALUE (iter_type))) | |
495 | iter_type = long_long_unsigned_type_node; | |
496 | } | |
497 | } | |
498 | ||
499 | if (collapse_count && *collapse_count == NULL) | |
500 | { | |
8e6b4515 | 501 | t = fold_binary (loop->cond_code, boolean_type_node, |
502 | fold_convert (TREE_TYPE (loop->v), loop->n1), | |
503 | fold_convert (TREE_TYPE (loop->v), loop->n2)); | |
504 | if (t && integer_zerop (t)) | |
505 | count = build_zero_cst (long_long_unsigned_type_node); | |
506 | else if ((i == 0 || count != NULL_TREE) | |
507 | && TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE | |
508 | && TREE_CONSTANT (loop->n1) | |
509 | && TREE_CONSTANT (loop->n2) | |
510 | && TREE_CODE (loop->step) == INTEGER_CST) | |
fd6481cf | 511 | { |
512 | tree itype = TREE_TYPE (loop->v); | |
513 | ||
514 | if (POINTER_TYPE_P (itype)) | |
3cea8318 | 515 | itype = signed_type_for (itype); |
fd6481cf | 516 | t = build_int_cst (itype, (loop->cond_code == LT_EXPR ? -1 : 1)); |
389dd41b | 517 | t = fold_build2_loc (loc, |
518 | PLUS_EXPR, itype, | |
519 | fold_convert_loc (loc, itype, loop->step), t); | |
520 | t = fold_build2_loc (loc, PLUS_EXPR, itype, t, | |
521 | fold_convert_loc (loc, itype, loop->n2)); | |
522 | t = fold_build2_loc (loc, MINUS_EXPR, itype, t, | |
523 | fold_convert_loc (loc, itype, loop->n1)); | |
fd6481cf | 524 | if (TYPE_UNSIGNED (itype) && loop->cond_code == GT_EXPR) |
389dd41b | 525 | t = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, |
526 | fold_build1_loc (loc, NEGATE_EXPR, itype, t), | |
527 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
528 | fold_convert_loc (loc, itype, | |
529 | loop->step))); | |
fd6481cf | 530 | else |
389dd41b | 531 | t = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, t, |
532 | fold_convert_loc (loc, itype, loop->step)); | |
533 | t = fold_convert_loc (loc, long_long_unsigned_type_node, t); | |
fd6481cf | 534 | if (count != NULL_TREE) |
389dd41b | 535 | count = fold_build2_loc (loc, |
536 | MULT_EXPR, long_long_unsigned_type_node, | |
fd6481cf | 537 | count, t); |
538 | else | |
539 | count = t; | |
540 | if (TREE_CODE (count) != INTEGER_CST) | |
541 | count = NULL_TREE; | |
542 | } | |
8e6b4515 | 543 | else if (count && !integer_zerop (count)) |
fd6481cf | 544 | count = NULL_TREE; |
545 | } | |
546 | } | |
547 | ||
3d483a94 | 548 | if (count |
bc7bff74 | 549 | && !simd |
550 | && (fd->sched_kind != OMP_CLAUSE_SCHEDULE_STATIC | |
551 | || fd->have_ordered)) | |
fd6481cf | 552 | { |
553 | if (!tree_int_cst_lt (count, TYPE_MAX_VALUE (long_integer_type_node))) | |
554 | iter_type = long_long_unsigned_type_node; | |
555 | else | |
556 | iter_type = long_integer_type_node; | |
557 | } | |
558 | else if (collapse_iter && *collapse_iter != NULL) | |
559 | iter_type = TREE_TYPE (*collapse_iter); | |
560 | fd->iter_type = iter_type; | |
561 | if (collapse_iter && *collapse_iter == NULL) | |
562 | *collapse_iter = create_tmp_var (iter_type, ".iter"); | |
563 | if (collapse_count && *collapse_count == NULL) | |
564 | { | |
565 | if (count) | |
389dd41b | 566 | *collapse_count = fold_convert_loc (loc, iter_type, count); |
fd6481cf | 567 | else |
568 | *collapse_count = create_tmp_var (iter_type, ".count"); | |
569 | } | |
570 | ||
571 | if (fd->collapse > 1) | |
572 | { | |
573 | fd->loop.v = *collapse_iter; | |
574 | fd->loop.n1 = build_int_cst (TREE_TYPE (fd->loop.v), 0); | |
575 | fd->loop.n2 = *collapse_count; | |
576 | fd->loop.step = build_int_cst (TREE_TYPE (fd->loop.v), 1); | |
577 | fd->loop.cond_code = LT_EXPR; | |
578 | } | |
773c5ba7 | 579 | } |
580 | ||
581 | ||
582 | /* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB | |
583 | is the immediate dominator of PAR_ENTRY_BB, return true if there | |
584 | are no data dependencies that would prevent expanding the parallel | |
585 | directive at PAR_ENTRY_BB as a combined parallel+workshare region. | |
586 | ||
587 | When expanding a combined parallel+workshare region, the call to | |
588 | the child function may need additional arguments in the case of | |
75a70cf9 | 589 | GIMPLE_OMP_FOR regions. In some cases, these arguments are |
590 | computed out of variables passed in from the parent to the child | |
591 | via 'struct .omp_data_s'. For instance: | |
773c5ba7 | 592 | |
593 | #pragma omp parallel for schedule (guided, i * 4) | |
594 | for (j ...) | |
595 | ||
596 | Is lowered into: | |
597 | ||
598 | # BLOCK 2 (PAR_ENTRY_BB) | |
599 | .omp_data_o.i = i; | |
600 | #pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598) | |
48e1416a | 601 | |
773c5ba7 | 602 | # BLOCK 3 (WS_ENTRY_BB) |
603 | .omp_data_i = &.omp_data_o; | |
604 | D.1667 = .omp_data_i->i; | |
605 | D.1598 = D.1667 * 4; | |
606 | #pragma omp for schedule (guided, D.1598) | |
607 | ||
608 | When we outline the parallel region, the call to the child function | |
609 | 'bar.omp_fn.0' will need the value D.1598 in its argument list, but | |
610 | that value is computed *after* the call site. So, in principle we | |
611 | cannot do the transformation. | |
612 | ||
613 | To see whether the code in WS_ENTRY_BB blocks the combined | |
614 | parallel+workshare call, we collect all the variables used in the | |
75a70cf9 | 615 | GIMPLE_OMP_FOR header check whether they appear on the LHS of any |
773c5ba7 | 616 | statement in WS_ENTRY_BB. If so, then we cannot emit the combined |
617 | call. | |
618 | ||
619 | FIXME. If we had the SSA form built at this point, we could merely | |
620 | hoist the code in block 3 into block 2 and be done with it. But at | |
621 | this point we don't have dataflow information and though we could | |
622 | hack something up here, it is really not worth the aggravation. */ | |
623 | ||
624 | static bool | |
f018d957 | 625 | workshare_safe_to_combine_p (basic_block ws_entry_bb) |
773c5ba7 | 626 | { |
627 | struct omp_for_data fd; | |
f018d957 | 628 | gimple ws_stmt = last_stmt (ws_entry_bb); |
773c5ba7 | 629 | |
75a70cf9 | 630 | if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) |
773c5ba7 | 631 | return true; |
632 | ||
75a70cf9 | 633 | gcc_assert (gimple_code (ws_stmt) == GIMPLE_OMP_FOR); |
773c5ba7 | 634 | |
fd6481cf | 635 | extract_omp_for_data (ws_stmt, &fd, NULL); |
636 | ||
637 | if (fd.collapse > 1 && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
638 | return false; | |
639 | if (fd.iter_type != long_integer_type_node) | |
640 | return false; | |
773c5ba7 | 641 | |
642 | /* FIXME. We give up too easily here. If any of these arguments | |
643 | are not constants, they will likely involve variables that have | |
644 | been mapped into fields of .omp_data_s for sharing with the child | |
645 | function. With appropriate data flow, it would be possible to | |
646 | see through this. */ | |
fd6481cf | 647 | if (!is_gimple_min_invariant (fd.loop.n1) |
648 | || !is_gimple_min_invariant (fd.loop.n2) | |
649 | || !is_gimple_min_invariant (fd.loop.step) | |
773c5ba7 | 650 | || (fd.chunk_size && !is_gimple_min_invariant (fd.chunk_size))) |
651 | return false; | |
652 | ||
653 | return true; | |
654 | } | |
655 | ||
656 | ||
657 | /* Collect additional arguments needed to emit a combined | |
658 | parallel+workshare call. WS_STMT is the workshare directive being | |
659 | expanded. */ | |
660 | ||
f1f41a6c | 661 | static vec<tree, va_gc> * |
bc7bff74 | 662 | get_ws_args_for (gimple par_stmt, gimple ws_stmt) |
773c5ba7 | 663 | { |
664 | tree t; | |
389dd41b | 665 | location_t loc = gimple_location (ws_stmt); |
f1f41a6c | 666 | vec<tree, va_gc> *ws_args; |
773c5ba7 | 667 | |
75a70cf9 | 668 | if (gimple_code (ws_stmt) == GIMPLE_OMP_FOR) |
773c5ba7 | 669 | { |
670 | struct omp_for_data fd; | |
bc7bff74 | 671 | tree n1, n2; |
773c5ba7 | 672 | |
fd6481cf | 673 | extract_omp_for_data (ws_stmt, &fd, NULL); |
bc7bff74 | 674 | n1 = fd.loop.n1; |
675 | n2 = fd.loop.n2; | |
676 | ||
677 | if (gimple_omp_for_combined_into_p (ws_stmt)) | |
678 | { | |
679 | tree innerc | |
680 | = find_omp_clause (gimple_omp_parallel_clauses (par_stmt), | |
681 | OMP_CLAUSE__LOOPTEMP_); | |
682 | gcc_assert (innerc); | |
683 | n1 = OMP_CLAUSE_DECL (innerc); | |
684 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
685 | OMP_CLAUSE__LOOPTEMP_); | |
686 | gcc_assert (innerc); | |
687 | n2 = OMP_CLAUSE_DECL (innerc); | |
688 | } | |
773c5ba7 | 689 | |
f1f41a6c | 690 | vec_alloc (ws_args, 3 + (fd.chunk_size != 0)); |
773c5ba7 | 691 | |
bc7bff74 | 692 | t = fold_convert_loc (loc, long_integer_type_node, n1); |
f1f41a6c | 693 | ws_args->quick_push (t); |
773c5ba7 | 694 | |
bc7bff74 | 695 | t = fold_convert_loc (loc, long_integer_type_node, n2); |
f1f41a6c | 696 | ws_args->quick_push (t); |
773c5ba7 | 697 | |
414c3a2c | 698 | t = fold_convert_loc (loc, long_integer_type_node, fd.loop.step); |
f1f41a6c | 699 | ws_args->quick_push (t); |
414c3a2c | 700 | |
701 | if (fd.chunk_size) | |
702 | { | |
703 | t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size); | |
f1f41a6c | 704 | ws_args->quick_push (t); |
414c3a2c | 705 | } |
773c5ba7 | 706 | |
707 | return ws_args; | |
708 | } | |
75a70cf9 | 709 | else if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) |
773c5ba7 | 710 | { |
ac6e3339 | 711 | /* Number of sections is equal to the number of edges from the |
75a70cf9 | 712 | GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to |
713 | the exit of the sections region. */ | |
714 | basic_block bb = single_succ (gimple_bb (ws_stmt)); | |
ac6e3339 | 715 | t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1); |
f1f41a6c | 716 | vec_alloc (ws_args, 1); |
717 | ws_args->quick_push (t); | |
414c3a2c | 718 | return ws_args; |
773c5ba7 | 719 | } |
720 | ||
721 | gcc_unreachable (); | |
722 | } | |
723 | ||
724 | ||
725 | /* Discover whether REGION is a combined parallel+workshare region. */ | |
726 | ||
727 | static void | |
728 | determine_parallel_type (struct omp_region *region) | |
1e8e9920 | 729 | { |
773c5ba7 | 730 | basic_block par_entry_bb, par_exit_bb; |
731 | basic_block ws_entry_bb, ws_exit_bb; | |
732 | ||
03ed154b | 733 | if (region == NULL || region->inner == NULL |
ac6e3339 | 734 | || region->exit == NULL || region->inner->exit == NULL |
735 | || region->inner->cont == NULL) | |
773c5ba7 | 736 | return; |
737 | ||
738 | /* We only support parallel+for and parallel+sections. */ | |
75a70cf9 | 739 | if (region->type != GIMPLE_OMP_PARALLEL |
740 | || (region->inner->type != GIMPLE_OMP_FOR | |
741 | && region->inner->type != GIMPLE_OMP_SECTIONS)) | |
773c5ba7 | 742 | return; |
743 | ||
744 | /* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and | |
745 | WS_EXIT_BB -> PAR_EXIT_BB. */ | |
61e47ac8 | 746 | par_entry_bb = region->entry; |
747 | par_exit_bb = region->exit; | |
748 | ws_entry_bb = region->inner->entry; | |
749 | ws_exit_bb = region->inner->exit; | |
773c5ba7 | 750 | |
751 | if (single_succ (par_entry_bb) == ws_entry_bb | |
752 | && single_succ (ws_exit_bb) == par_exit_bb | |
f018d957 | 753 | && workshare_safe_to_combine_p (ws_entry_bb) |
75a70cf9 | 754 | && (gimple_omp_parallel_combined_p (last_stmt (par_entry_bb)) |
de7ef844 | 755 | || (last_and_only_stmt (ws_entry_bb) |
756 | && last_and_only_stmt (par_exit_bb)))) | |
773c5ba7 | 757 | { |
bc7bff74 | 758 | gimple par_stmt = last_stmt (par_entry_bb); |
75a70cf9 | 759 | gimple ws_stmt = last_stmt (ws_entry_bb); |
61e47ac8 | 760 | |
75a70cf9 | 761 | if (region->inner->type == GIMPLE_OMP_FOR) |
773c5ba7 | 762 | { |
763 | /* If this is a combined parallel loop, we need to determine | |
764 | whether or not to use the combined library calls. There | |
765 | are two cases where we do not apply the transformation: | |
766 | static loops and any kind of ordered loop. In the first | |
767 | case, we already open code the loop so there is no need | |
768 | to do anything else. In the latter case, the combined | |
769 | parallel loop call would still need extra synchronization | |
770 | to implement ordered semantics, so there would not be any | |
771 | gain in using the combined call. */ | |
75a70cf9 | 772 | tree clauses = gimple_omp_for_clauses (ws_stmt); |
773c5ba7 | 773 | tree c = find_omp_clause (clauses, OMP_CLAUSE_SCHEDULE); |
774 | if (c == NULL | |
775 | || OMP_CLAUSE_SCHEDULE_KIND (c) == OMP_CLAUSE_SCHEDULE_STATIC | |
776 | || find_omp_clause (clauses, OMP_CLAUSE_ORDERED)) | |
777 | { | |
778 | region->is_combined_parallel = false; | |
779 | region->inner->is_combined_parallel = false; | |
780 | return; | |
781 | } | |
782 | } | |
783 | ||
784 | region->is_combined_parallel = true; | |
785 | region->inner->is_combined_parallel = true; | |
bc7bff74 | 786 | region->ws_args = get_ws_args_for (par_stmt, ws_stmt); |
773c5ba7 | 787 | } |
1e8e9920 | 788 | } |
789 | ||
773c5ba7 | 790 | |
1e8e9920 | 791 | /* Return true if EXPR is variable sized. */ |
792 | ||
793 | static inline bool | |
1f1872fd | 794 | is_variable_sized (const_tree expr) |
1e8e9920 | 795 | { |
796 | return !TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (expr))); | |
797 | } | |
798 | ||
799 | /* Return true if DECL is a reference type. */ | |
800 | ||
801 | static inline bool | |
802 | is_reference (tree decl) | |
803 | { | |
804 | return lang_hooks.decls.omp_privatize_by_reference (decl); | |
805 | } | |
806 | ||
807 | /* Lookup variables in the decl or field splay trees. The "maybe" form | |
808 | allows for the variable form to not have been entered, otherwise we | |
809 | assert that the variable must have been entered. */ | |
810 | ||
811 | static inline tree | |
812 | lookup_decl (tree var, omp_context *ctx) | |
813 | { | |
e3022db7 | 814 | tree *n; |
815 | n = (tree *) pointer_map_contains (ctx->cb.decl_map, var); | |
816 | return *n; | |
1e8e9920 | 817 | } |
818 | ||
819 | static inline tree | |
e8a588af | 820 | maybe_lookup_decl (const_tree var, omp_context *ctx) |
1e8e9920 | 821 | { |
e3022db7 | 822 | tree *n; |
823 | n = (tree *) pointer_map_contains (ctx->cb.decl_map, var); | |
824 | return n ? *n : NULL_TREE; | |
1e8e9920 | 825 | } |
826 | ||
827 | static inline tree | |
828 | lookup_field (tree var, omp_context *ctx) | |
829 | { | |
830 | splay_tree_node n; | |
831 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) var); | |
832 | return (tree) n->value; | |
833 | } | |
834 | ||
fd6481cf | 835 | static inline tree |
836 | lookup_sfield (tree var, omp_context *ctx) | |
837 | { | |
838 | splay_tree_node n; | |
839 | n = splay_tree_lookup (ctx->sfield_map | |
840 | ? ctx->sfield_map : ctx->field_map, | |
841 | (splay_tree_key) var); | |
842 | return (tree) n->value; | |
843 | } | |
844 | ||
1e8e9920 | 845 | static inline tree |
846 | maybe_lookup_field (tree var, omp_context *ctx) | |
847 | { | |
848 | splay_tree_node n; | |
849 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) var); | |
850 | return n ? (tree) n->value : NULL_TREE; | |
851 | } | |
852 | ||
e8a588af | 853 | /* Return true if DECL should be copied by pointer. SHARED_CTX is |
854 | the parallel context if DECL is to be shared. */ | |
1e8e9920 | 855 | |
856 | static bool | |
fd6481cf | 857 | use_pointer_for_field (tree decl, omp_context *shared_ctx) |
1e8e9920 | 858 | { |
859 | if (AGGREGATE_TYPE_P (TREE_TYPE (decl))) | |
860 | return true; | |
861 | ||
554f2707 | 862 | /* We can only use copy-in/copy-out semantics for shared variables |
1e8e9920 | 863 | when we know the value is not accessible from an outer scope. */ |
e8a588af | 864 | if (shared_ctx) |
1e8e9920 | 865 | { |
866 | /* ??? Trivially accessible from anywhere. But why would we even | |
867 | be passing an address in this case? Should we simply assert | |
868 | this to be false, or should we have a cleanup pass that removes | |
869 | these from the list of mappings? */ | |
870 | if (TREE_STATIC (decl) || DECL_EXTERNAL (decl)) | |
871 | return true; | |
872 | ||
873 | /* For variables with DECL_HAS_VALUE_EXPR_P set, we cannot tell | |
874 | without analyzing the expression whether or not its location | |
875 | is accessible to anyone else. In the case of nested parallel | |
876 | regions it certainly may be. */ | |
df2c34fc | 877 | if (TREE_CODE (decl) != RESULT_DECL && DECL_HAS_VALUE_EXPR_P (decl)) |
1e8e9920 | 878 | return true; |
879 | ||
880 | /* Do not use copy-in/copy-out for variables that have their | |
881 | address taken. */ | |
882 | if (TREE_ADDRESSABLE (decl)) | |
883 | return true; | |
e8a588af | 884 | |
b8214689 | 885 | /* lower_send_shared_vars only uses copy-in, but not copy-out |
886 | for these. */ | |
887 | if (TREE_READONLY (decl) | |
888 | || ((TREE_CODE (decl) == RESULT_DECL | |
889 | || TREE_CODE (decl) == PARM_DECL) | |
890 | && DECL_BY_REFERENCE (decl))) | |
891 | return false; | |
892 | ||
e8a588af | 893 | /* Disallow copy-in/out in nested parallel if |
894 | decl is shared in outer parallel, otherwise | |
895 | each thread could store the shared variable | |
896 | in its own copy-in location, making the | |
897 | variable no longer really shared. */ | |
b8214689 | 898 | if (shared_ctx->is_nested) |
e8a588af | 899 | { |
900 | omp_context *up; | |
901 | ||
902 | for (up = shared_ctx->outer; up; up = up->outer) | |
0cb159ec | 903 | if (is_taskreg_ctx (up) && maybe_lookup_decl (decl, up)) |
e8a588af | 904 | break; |
905 | ||
0cb159ec | 906 | if (up) |
e8a588af | 907 | { |
908 | tree c; | |
909 | ||
75a70cf9 | 910 | for (c = gimple_omp_taskreg_clauses (up->stmt); |
e8a588af | 911 | c; c = OMP_CLAUSE_CHAIN (c)) |
912 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_SHARED | |
913 | && OMP_CLAUSE_DECL (c) == decl) | |
914 | break; | |
915 | ||
916 | if (c) | |
784ad964 | 917 | goto maybe_mark_addressable_and_ret; |
e8a588af | 918 | } |
919 | } | |
fd6481cf | 920 | |
b8214689 | 921 | /* For tasks avoid using copy-in/out. As tasks can be |
fd6481cf | 922 | deferred or executed in different thread, when GOMP_task |
923 | returns, the task hasn't necessarily terminated. */ | |
b8214689 | 924 | if (is_task_ctx (shared_ctx)) |
fd6481cf | 925 | { |
784ad964 | 926 | tree outer; |
927 | maybe_mark_addressable_and_ret: | |
928 | outer = maybe_lookup_decl_in_outer_ctx (decl, shared_ctx); | |
fd6481cf | 929 | if (is_gimple_reg (outer)) |
930 | { | |
931 | /* Taking address of OUTER in lower_send_shared_vars | |
932 | might need regimplification of everything that uses the | |
933 | variable. */ | |
934 | if (!task_shared_vars) | |
935 | task_shared_vars = BITMAP_ALLOC (NULL); | |
936 | bitmap_set_bit (task_shared_vars, DECL_UID (outer)); | |
937 | TREE_ADDRESSABLE (outer) = 1; | |
938 | } | |
939 | return true; | |
940 | } | |
1e8e9920 | 941 | } |
942 | ||
943 | return false; | |
944 | } | |
945 | ||
79acaae1 | 946 | /* Construct a new automatic decl similar to VAR. */ |
947 | ||
948 | static tree | |
949 | omp_copy_decl_2 (tree var, tree name, tree type, omp_context *ctx) | |
950 | { | |
951 | tree copy = copy_var_decl (var, name, type); | |
952 | ||
953 | DECL_CONTEXT (copy) = current_function_decl; | |
1767a056 | 954 | DECL_CHAIN (copy) = ctx->block_vars; |
1e8e9920 | 955 | ctx->block_vars = copy; |
956 | ||
957 | return copy; | |
958 | } | |
959 | ||
960 | static tree | |
961 | omp_copy_decl_1 (tree var, omp_context *ctx) | |
962 | { | |
963 | return omp_copy_decl_2 (var, DECL_NAME (var), TREE_TYPE (var), ctx); | |
964 | } | |
965 | ||
445d06b6 | 966 | /* Build COMPONENT_REF and set TREE_THIS_VOLATILE and TREE_READONLY on it |
967 | as appropriate. */ | |
968 | static tree | |
969 | omp_build_component_ref (tree obj, tree field) | |
970 | { | |
971 | tree ret = build3 (COMPONENT_REF, TREE_TYPE (field), obj, field, NULL); | |
972 | if (TREE_THIS_VOLATILE (field)) | |
973 | TREE_THIS_VOLATILE (ret) |= 1; | |
974 | if (TREE_READONLY (field)) | |
975 | TREE_READONLY (ret) |= 1; | |
976 | return ret; | |
977 | } | |
978 | ||
1e8e9920 | 979 | /* Build tree nodes to access the field for VAR on the receiver side. */ |
980 | ||
981 | static tree | |
982 | build_receiver_ref (tree var, bool by_ref, omp_context *ctx) | |
983 | { | |
984 | tree x, field = lookup_field (var, ctx); | |
985 | ||
986 | /* If the receiver record type was remapped in the child function, | |
987 | remap the field into the new record type. */ | |
988 | x = maybe_lookup_field (field, ctx); | |
989 | if (x != NULL) | |
990 | field = x; | |
991 | ||
182cf5a9 | 992 | x = build_simple_mem_ref (ctx->receiver_decl); |
445d06b6 | 993 | x = omp_build_component_ref (x, field); |
1e8e9920 | 994 | if (by_ref) |
182cf5a9 | 995 | x = build_simple_mem_ref (x); |
1e8e9920 | 996 | |
997 | return x; | |
998 | } | |
999 | ||
1000 | /* Build tree nodes to access VAR in the scope outer to CTX. In the case | |
1001 | of a parallel, this is a component reference; for workshare constructs | |
1002 | this is some variable. */ | |
1003 | ||
1004 | static tree | |
1005 | build_outer_var_ref (tree var, omp_context *ctx) | |
1006 | { | |
1007 | tree x; | |
1008 | ||
f49d7bb5 | 1009 | if (is_global_var (maybe_lookup_decl_in_outer_ctx (var, ctx))) |
1e8e9920 | 1010 | x = var; |
1011 | else if (is_variable_sized (var)) | |
1012 | { | |
1013 | x = TREE_OPERAND (DECL_VALUE_EXPR (var), 0); | |
1014 | x = build_outer_var_ref (x, ctx); | |
182cf5a9 | 1015 | x = build_simple_mem_ref (x); |
1e8e9920 | 1016 | } |
fd6481cf | 1017 | else if (is_taskreg_ctx (ctx)) |
1e8e9920 | 1018 | { |
e8a588af | 1019 | bool by_ref = use_pointer_for_field (var, NULL); |
1e8e9920 | 1020 | x = build_receiver_ref (var, by_ref, ctx); |
1021 | } | |
3d483a94 | 1022 | else if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR |
f2697631 | 1023 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_KIND_SIMD) |
3d483a94 | 1024 | { |
1025 | /* #pragma omp simd isn't a worksharing construct, and can reference even | |
1026 | private vars in its linear etc. clauses. */ | |
1027 | x = NULL_TREE; | |
1028 | if (ctx->outer && is_taskreg_ctx (ctx)) | |
1029 | x = lookup_decl (var, ctx->outer); | |
1030 | else if (ctx->outer) | |
84cb1020 | 1031 | x = maybe_lookup_decl_in_outer_ctx (var, ctx); |
3d483a94 | 1032 | if (x == NULL_TREE) |
1033 | x = var; | |
1034 | } | |
1e8e9920 | 1035 | else if (ctx->outer) |
1036 | x = lookup_decl (var, ctx->outer); | |
9438af57 | 1037 | else if (is_reference (var)) |
1038 | /* This can happen with orphaned constructs. If var is reference, it is | |
1039 | possible it is shared and as such valid. */ | |
1040 | x = var; | |
1e8e9920 | 1041 | else |
1042 | gcc_unreachable (); | |
1043 | ||
1044 | if (is_reference (var)) | |
182cf5a9 | 1045 | x = build_simple_mem_ref (x); |
1e8e9920 | 1046 | |
1047 | return x; | |
1048 | } | |
1049 | ||
1050 | /* Build tree nodes to access the field for VAR on the sender side. */ | |
1051 | ||
1052 | static tree | |
1053 | build_sender_ref (tree var, omp_context *ctx) | |
1054 | { | |
fd6481cf | 1055 | tree field = lookup_sfield (var, ctx); |
445d06b6 | 1056 | return omp_build_component_ref (ctx->sender_decl, field); |
1e8e9920 | 1057 | } |
1058 | ||
1059 | /* Add a new field for VAR inside the structure CTX->SENDER_DECL. */ | |
1060 | ||
1061 | static void | |
fd6481cf | 1062 | install_var_field (tree var, bool by_ref, int mask, omp_context *ctx) |
1e8e9920 | 1063 | { |
fd6481cf | 1064 | tree field, type, sfield = NULL_TREE; |
1e8e9920 | 1065 | |
fd6481cf | 1066 | gcc_assert ((mask & 1) == 0 |
1067 | || !splay_tree_lookup (ctx->field_map, (splay_tree_key) var)); | |
1068 | gcc_assert ((mask & 2) == 0 || !ctx->sfield_map | |
1069 | || !splay_tree_lookup (ctx->sfield_map, (splay_tree_key) var)); | |
1e8e9920 | 1070 | |
1071 | type = TREE_TYPE (var); | |
bc7bff74 | 1072 | if (mask & 4) |
1073 | { | |
1074 | gcc_assert (TREE_CODE (type) == ARRAY_TYPE); | |
1075 | type = build_pointer_type (build_pointer_type (type)); | |
1076 | } | |
1077 | else if (by_ref) | |
1e8e9920 | 1078 | type = build_pointer_type (type); |
fd6481cf | 1079 | else if ((mask & 3) == 1 && is_reference (var)) |
1080 | type = TREE_TYPE (type); | |
1e8e9920 | 1081 | |
e60a6f7b | 1082 | field = build_decl (DECL_SOURCE_LOCATION (var), |
1083 | FIELD_DECL, DECL_NAME (var), type); | |
1e8e9920 | 1084 | |
1085 | /* Remember what variable this field was created for. This does have a | |
1086 | side effect of making dwarf2out ignore this member, so for helpful | |
1087 | debugging we clear it later in delete_omp_context. */ | |
1088 | DECL_ABSTRACT_ORIGIN (field) = var; | |
fd6481cf | 1089 | if (type == TREE_TYPE (var)) |
1090 | { | |
1091 | DECL_ALIGN (field) = DECL_ALIGN (var); | |
1092 | DECL_USER_ALIGN (field) = DECL_USER_ALIGN (var); | |
1093 | TREE_THIS_VOLATILE (field) = TREE_THIS_VOLATILE (var); | |
1094 | } | |
1095 | else | |
1096 | DECL_ALIGN (field) = TYPE_ALIGN (type); | |
1e8e9920 | 1097 | |
fd6481cf | 1098 | if ((mask & 3) == 3) |
1099 | { | |
1100 | insert_field_into_struct (ctx->record_type, field); | |
1101 | if (ctx->srecord_type) | |
1102 | { | |
e60a6f7b | 1103 | sfield = build_decl (DECL_SOURCE_LOCATION (var), |
1104 | FIELD_DECL, DECL_NAME (var), type); | |
fd6481cf | 1105 | DECL_ABSTRACT_ORIGIN (sfield) = var; |
1106 | DECL_ALIGN (sfield) = DECL_ALIGN (field); | |
1107 | DECL_USER_ALIGN (sfield) = DECL_USER_ALIGN (field); | |
1108 | TREE_THIS_VOLATILE (sfield) = TREE_THIS_VOLATILE (field); | |
1109 | insert_field_into_struct (ctx->srecord_type, sfield); | |
1110 | } | |
1111 | } | |
1112 | else | |
1113 | { | |
1114 | if (ctx->srecord_type == NULL_TREE) | |
1115 | { | |
1116 | tree t; | |
1117 | ||
1118 | ctx->srecord_type = lang_hooks.types.make_type (RECORD_TYPE); | |
1119 | ctx->sfield_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
1120 | for (t = TYPE_FIELDS (ctx->record_type); t ; t = TREE_CHAIN (t)) | |
1121 | { | |
e60a6f7b | 1122 | sfield = build_decl (DECL_SOURCE_LOCATION (var), |
1123 | FIELD_DECL, DECL_NAME (t), TREE_TYPE (t)); | |
fd6481cf | 1124 | DECL_ABSTRACT_ORIGIN (sfield) = DECL_ABSTRACT_ORIGIN (t); |
1125 | insert_field_into_struct (ctx->srecord_type, sfield); | |
1126 | splay_tree_insert (ctx->sfield_map, | |
1127 | (splay_tree_key) DECL_ABSTRACT_ORIGIN (t), | |
1128 | (splay_tree_value) sfield); | |
1129 | } | |
1130 | } | |
1131 | sfield = field; | |
1132 | insert_field_into_struct ((mask & 1) ? ctx->record_type | |
1133 | : ctx->srecord_type, field); | |
1134 | } | |
1e8e9920 | 1135 | |
fd6481cf | 1136 | if (mask & 1) |
1137 | splay_tree_insert (ctx->field_map, (splay_tree_key) var, | |
1138 | (splay_tree_value) field); | |
1139 | if ((mask & 2) && ctx->sfield_map) | |
1140 | splay_tree_insert (ctx->sfield_map, (splay_tree_key) var, | |
1141 | (splay_tree_value) sfield); | |
1e8e9920 | 1142 | } |
1143 | ||
1144 | static tree | |
1145 | install_var_local (tree var, omp_context *ctx) | |
1146 | { | |
1147 | tree new_var = omp_copy_decl_1 (var, ctx); | |
1148 | insert_decl_map (&ctx->cb, var, new_var); | |
1149 | return new_var; | |
1150 | } | |
1151 | ||
1152 | /* Adjust the replacement for DECL in CTX for the new context. This means | |
1153 | copying the DECL_VALUE_EXPR, and fixing up the type. */ | |
1154 | ||
1155 | static void | |
1156 | fixup_remapped_decl (tree decl, omp_context *ctx, bool private_debug) | |
1157 | { | |
1158 | tree new_decl, size; | |
1159 | ||
1160 | new_decl = lookup_decl (decl, ctx); | |
1161 | ||
1162 | TREE_TYPE (new_decl) = remap_type (TREE_TYPE (decl), &ctx->cb); | |
1163 | ||
1164 | if ((!TREE_CONSTANT (DECL_SIZE (new_decl)) || private_debug) | |
1165 | && DECL_HAS_VALUE_EXPR_P (decl)) | |
1166 | { | |
1167 | tree ve = DECL_VALUE_EXPR (decl); | |
75a70cf9 | 1168 | walk_tree (&ve, copy_tree_body_r, &ctx->cb, NULL); |
1e8e9920 | 1169 | SET_DECL_VALUE_EXPR (new_decl, ve); |
1170 | DECL_HAS_VALUE_EXPR_P (new_decl) = 1; | |
1171 | } | |
1172 | ||
1173 | if (!TREE_CONSTANT (DECL_SIZE (new_decl))) | |
1174 | { | |
1175 | size = remap_decl (DECL_SIZE (decl), &ctx->cb); | |
1176 | if (size == error_mark_node) | |
1177 | size = TYPE_SIZE (TREE_TYPE (new_decl)); | |
1178 | DECL_SIZE (new_decl) = size; | |
1179 | ||
1180 | size = remap_decl (DECL_SIZE_UNIT (decl), &ctx->cb); | |
1181 | if (size == error_mark_node) | |
1182 | size = TYPE_SIZE_UNIT (TREE_TYPE (new_decl)); | |
1183 | DECL_SIZE_UNIT (new_decl) = size; | |
1184 | } | |
1185 | } | |
1186 | ||
1187 | /* The callback for remap_decl. Search all containing contexts for a | |
1188 | mapping of the variable; this avoids having to duplicate the splay | |
1189 | tree ahead of time. We know a mapping doesn't already exist in the | |
1190 | given context. Create new mappings to implement default semantics. */ | |
1191 | ||
1192 | static tree | |
1193 | omp_copy_decl (tree var, copy_body_data *cb) | |
1194 | { | |
1195 | omp_context *ctx = (omp_context *) cb; | |
1196 | tree new_var; | |
1197 | ||
1e8e9920 | 1198 | if (TREE_CODE (var) == LABEL_DECL) |
1199 | { | |
e60a6f7b | 1200 | new_var = create_artificial_label (DECL_SOURCE_LOCATION (var)); |
773c5ba7 | 1201 | DECL_CONTEXT (new_var) = current_function_decl; |
1e8e9920 | 1202 | insert_decl_map (&ctx->cb, var, new_var); |
1203 | return new_var; | |
1204 | } | |
1205 | ||
fd6481cf | 1206 | while (!is_taskreg_ctx (ctx)) |
1e8e9920 | 1207 | { |
1208 | ctx = ctx->outer; | |
1209 | if (ctx == NULL) | |
1210 | return var; | |
1211 | new_var = maybe_lookup_decl (var, ctx); | |
1212 | if (new_var) | |
1213 | return new_var; | |
1214 | } | |
1215 | ||
f49d7bb5 | 1216 | if (is_global_var (var) || decl_function_context (var) != ctx->cb.src_fn) |
1217 | return var; | |
1218 | ||
1e8e9920 | 1219 | return error_mark_node; |
1220 | } | |
1221 | ||
773c5ba7 | 1222 | |
1223 | /* Return the parallel region associated with STMT. */ | |
1224 | ||
773c5ba7 | 1225 | /* Debugging dumps for parallel regions. */ |
1226 | void dump_omp_region (FILE *, struct omp_region *, int); | |
1227 | void debug_omp_region (struct omp_region *); | |
1228 | void debug_all_omp_regions (void); | |
1229 | ||
1230 | /* Dump the parallel region tree rooted at REGION. */ | |
1231 | ||
1232 | void | |
1233 | dump_omp_region (FILE *file, struct omp_region *region, int indent) | |
1234 | { | |
61e47ac8 | 1235 | fprintf (file, "%*sbb %d: %s\n", indent, "", region->entry->index, |
75a70cf9 | 1236 | gimple_code_name[region->type]); |
773c5ba7 | 1237 | |
1238 | if (region->inner) | |
1239 | dump_omp_region (file, region->inner, indent + 4); | |
1240 | ||
61e47ac8 | 1241 | if (region->cont) |
1242 | { | |
75a70cf9 | 1243 | fprintf (file, "%*sbb %d: GIMPLE_OMP_CONTINUE\n", indent, "", |
61e47ac8 | 1244 | region->cont->index); |
1245 | } | |
48e1416a | 1246 | |
773c5ba7 | 1247 | if (region->exit) |
75a70cf9 | 1248 | fprintf (file, "%*sbb %d: GIMPLE_OMP_RETURN\n", indent, "", |
61e47ac8 | 1249 | region->exit->index); |
773c5ba7 | 1250 | else |
61e47ac8 | 1251 | fprintf (file, "%*s[no exit marker]\n", indent, ""); |
773c5ba7 | 1252 | |
1253 | if (region->next) | |
61e47ac8 | 1254 | dump_omp_region (file, region->next, indent); |
773c5ba7 | 1255 | } |
1256 | ||
4b987fac | 1257 | DEBUG_FUNCTION void |
773c5ba7 | 1258 | debug_omp_region (struct omp_region *region) |
1259 | { | |
1260 | dump_omp_region (stderr, region, 0); | |
1261 | } | |
1262 | ||
4b987fac | 1263 | DEBUG_FUNCTION void |
773c5ba7 | 1264 | debug_all_omp_regions (void) |
1265 | { | |
1266 | dump_omp_region (stderr, root_omp_region, 0); | |
1267 | } | |
1268 | ||
1269 | ||
1270 | /* Create a new parallel region starting at STMT inside region PARENT. */ | |
1271 | ||
7740abd8 | 1272 | static struct omp_region * |
75a70cf9 | 1273 | new_omp_region (basic_block bb, enum gimple_code type, |
1274 | struct omp_region *parent) | |
773c5ba7 | 1275 | { |
4077bf7a | 1276 | struct omp_region *region = XCNEW (struct omp_region); |
773c5ba7 | 1277 | |
1278 | region->outer = parent; | |
61e47ac8 | 1279 | region->entry = bb; |
1280 | region->type = type; | |
773c5ba7 | 1281 | |
1282 | if (parent) | |
1283 | { | |
1284 | /* This is a nested region. Add it to the list of inner | |
1285 | regions in PARENT. */ | |
1286 | region->next = parent->inner; | |
1287 | parent->inner = region; | |
1288 | } | |
61e47ac8 | 1289 | else |
773c5ba7 | 1290 | { |
1291 | /* This is a toplevel region. Add it to the list of toplevel | |
1292 | regions in ROOT_OMP_REGION. */ | |
1293 | region->next = root_omp_region; | |
1294 | root_omp_region = region; | |
1295 | } | |
61e47ac8 | 1296 | |
1297 | return region; | |
1298 | } | |
1299 | ||
1300 | /* Release the memory associated with the region tree rooted at REGION. */ | |
1301 | ||
1302 | static void | |
1303 | free_omp_region_1 (struct omp_region *region) | |
1304 | { | |
1305 | struct omp_region *i, *n; | |
1306 | ||
1307 | for (i = region->inner; i ; i = n) | |
773c5ba7 | 1308 | { |
61e47ac8 | 1309 | n = i->next; |
1310 | free_omp_region_1 (i); | |
773c5ba7 | 1311 | } |
1312 | ||
61e47ac8 | 1313 | free (region); |
1314 | } | |
773c5ba7 | 1315 | |
61e47ac8 | 1316 | /* Release the memory for the entire omp region tree. */ |
1317 | ||
1318 | void | |
1319 | free_omp_regions (void) | |
1320 | { | |
1321 | struct omp_region *r, *n; | |
1322 | for (r = root_omp_region; r ; r = n) | |
1323 | { | |
1324 | n = r->next; | |
1325 | free_omp_region_1 (r); | |
1326 | } | |
1327 | root_omp_region = NULL; | |
773c5ba7 | 1328 | } |
1329 | ||
1330 | ||
1e8e9920 | 1331 | /* Create a new context, with OUTER_CTX being the surrounding context. */ |
1332 | ||
1333 | static omp_context * | |
75a70cf9 | 1334 | new_omp_context (gimple stmt, omp_context *outer_ctx) |
1e8e9920 | 1335 | { |
1336 | omp_context *ctx = XCNEW (omp_context); | |
1337 | ||
1338 | splay_tree_insert (all_contexts, (splay_tree_key) stmt, | |
1339 | (splay_tree_value) ctx); | |
1340 | ctx->stmt = stmt; | |
1341 | ||
1342 | if (outer_ctx) | |
1343 | { | |
1344 | ctx->outer = outer_ctx; | |
1345 | ctx->cb = outer_ctx->cb; | |
1346 | ctx->cb.block = NULL; | |
1347 | ctx->depth = outer_ctx->depth + 1; | |
1348 | } | |
1349 | else | |
1350 | { | |
1351 | ctx->cb.src_fn = current_function_decl; | |
1352 | ctx->cb.dst_fn = current_function_decl; | |
53f79206 | 1353 | ctx->cb.src_node = cgraph_get_node (current_function_decl); |
1354 | gcc_checking_assert (ctx->cb.src_node); | |
1e8e9920 | 1355 | ctx->cb.dst_node = ctx->cb.src_node; |
1356 | ctx->cb.src_cfun = cfun; | |
1357 | ctx->cb.copy_decl = omp_copy_decl; | |
e38def9c | 1358 | ctx->cb.eh_lp_nr = 0; |
1e8e9920 | 1359 | ctx->cb.transform_call_graph_edges = CB_CGE_MOVE; |
1360 | ctx->depth = 1; | |
1361 | } | |
1362 | ||
e3022db7 | 1363 | ctx->cb.decl_map = pointer_map_create (); |
1e8e9920 | 1364 | |
1365 | return ctx; | |
1366 | } | |
1367 | ||
75a70cf9 | 1368 | static gimple_seq maybe_catch_exception (gimple_seq); |
f6430caa | 1369 | |
1370 | /* Finalize task copyfn. */ | |
1371 | ||
1372 | static void | |
75a70cf9 | 1373 | finalize_task_copyfn (gimple task_stmt) |
f6430caa | 1374 | { |
1375 | struct function *child_cfun; | |
9078126c | 1376 | tree child_fn; |
e3a19533 | 1377 | gimple_seq seq = NULL, new_seq; |
75a70cf9 | 1378 | gimple bind; |
f6430caa | 1379 | |
75a70cf9 | 1380 | child_fn = gimple_omp_task_copy_fn (task_stmt); |
f6430caa | 1381 | if (child_fn == NULL_TREE) |
1382 | return; | |
1383 | ||
1384 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
82b40354 | 1385 | DECL_STRUCT_FUNCTION (child_fn)->curr_properties = cfun->curr_properties; |
f6430caa | 1386 | |
f6430caa | 1387 | push_cfun (child_cfun); |
7e3aae05 | 1388 | bind = gimplify_body (child_fn, false); |
75a70cf9 | 1389 | gimple_seq_add_stmt (&seq, bind); |
1390 | new_seq = maybe_catch_exception (seq); | |
1391 | if (new_seq != seq) | |
1392 | { | |
1393 | bind = gimple_build_bind (NULL, new_seq, NULL); | |
e3a19533 | 1394 | seq = NULL; |
75a70cf9 | 1395 | gimple_seq_add_stmt (&seq, bind); |
1396 | } | |
1397 | gimple_set_body (child_fn, seq); | |
f6430caa | 1398 | pop_cfun (); |
f6430caa | 1399 | |
82b40354 | 1400 | /* Inform the callgraph about the new function. */ |
f6430caa | 1401 | cgraph_add_new_function (child_fn, false); |
1402 | } | |
1403 | ||
1e8e9920 | 1404 | /* Destroy a omp_context data structures. Called through the splay tree |
1405 | value delete callback. */ | |
1406 | ||
1407 | static void | |
1408 | delete_omp_context (splay_tree_value value) | |
1409 | { | |
1410 | omp_context *ctx = (omp_context *) value; | |
1411 | ||
e3022db7 | 1412 | pointer_map_destroy (ctx->cb.decl_map); |
1e8e9920 | 1413 | |
1414 | if (ctx->field_map) | |
1415 | splay_tree_delete (ctx->field_map); | |
fd6481cf | 1416 | if (ctx->sfield_map) |
1417 | splay_tree_delete (ctx->sfield_map); | |
1e8e9920 | 1418 | |
1419 | /* We hijacked DECL_ABSTRACT_ORIGIN earlier. We need to clear it before | |
1420 | it produces corrupt debug information. */ | |
1421 | if (ctx->record_type) | |
1422 | { | |
1423 | tree t; | |
1767a056 | 1424 | for (t = TYPE_FIELDS (ctx->record_type); t ; t = DECL_CHAIN (t)) |
1e8e9920 | 1425 | DECL_ABSTRACT_ORIGIN (t) = NULL; |
1426 | } | |
fd6481cf | 1427 | if (ctx->srecord_type) |
1428 | { | |
1429 | tree t; | |
1767a056 | 1430 | for (t = TYPE_FIELDS (ctx->srecord_type); t ; t = DECL_CHAIN (t)) |
fd6481cf | 1431 | DECL_ABSTRACT_ORIGIN (t) = NULL; |
1432 | } | |
1e8e9920 | 1433 | |
f6430caa | 1434 | if (is_task_ctx (ctx)) |
1435 | finalize_task_copyfn (ctx->stmt); | |
1436 | ||
1e8e9920 | 1437 | XDELETE (ctx); |
1438 | } | |
1439 | ||
1440 | /* Fix up RECEIVER_DECL with a type that has been remapped to the child | |
1441 | context. */ | |
1442 | ||
1443 | static void | |
1444 | fixup_child_record_type (omp_context *ctx) | |
1445 | { | |
1446 | tree f, type = ctx->record_type; | |
1447 | ||
1448 | /* ??? It isn't sufficient to just call remap_type here, because | |
1449 | variably_modified_type_p doesn't work the way we expect for | |
1450 | record types. Testing each field for whether it needs remapping | |
1451 | and creating a new record by hand works, however. */ | |
1767a056 | 1452 | for (f = TYPE_FIELDS (type); f ; f = DECL_CHAIN (f)) |
1e8e9920 | 1453 | if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn)) |
1454 | break; | |
1455 | if (f) | |
1456 | { | |
1457 | tree name, new_fields = NULL; | |
1458 | ||
1459 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
1460 | name = DECL_NAME (TYPE_NAME (ctx->record_type)); | |
e60a6f7b | 1461 | name = build_decl (DECL_SOURCE_LOCATION (ctx->receiver_decl), |
1462 | TYPE_DECL, name, type); | |
1e8e9920 | 1463 | TYPE_NAME (type) = name; |
1464 | ||
1767a056 | 1465 | for (f = TYPE_FIELDS (ctx->record_type); f ; f = DECL_CHAIN (f)) |
1e8e9920 | 1466 | { |
1467 | tree new_f = copy_node (f); | |
1468 | DECL_CONTEXT (new_f) = type; | |
1469 | TREE_TYPE (new_f) = remap_type (TREE_TYPE (f), &ctx->cb); | |
1767a056 | 1470 | DECL_CHAIN (new_f) = new_fields; |
75a70cf9 | 1471 | walk_tree (&DECL_SIZE (new_f), copy_tree_body_r, &ctx->cb, NULL); |
1472 | walk_tree (&DECL_SIZE_UNIT (new_f), copy_tree_body_r, | |
1473 | &ctx->cb, NULL); | |
1474 | walk_tree (&DECL_FIELD_OFFSET (new_f), copy_tree_body_r, | |
1475 | &ctx->cb, NULL); | |
1e8e9920 | 1476 | new_fields = new_f; |
1477 | ||
1478 | /* Arrange to be able to look up the receiver field | |
1479 | given the sender field. */ | |
1480 | splay_tree_insert (ctx->field_map, (splay_tree_key) f, | |
1481 | (splay_tree_value) new_f); | |
1482 | } | |
1483 | TYPE_FIELDS (type) = nreverse (new_fields); | |
1484 | layout_type (type); | |
1485 | } | |
1486 | ||
1487 | TREE_TYPE (ctx->receiver_decl) = build_pointer_type (type); | |
1488 | } | |
1489 | ||
1490 | /* Instantiate decls as necessary in CTX to satisfy the data sharing | |
1491 | specified by CLAUSES. */ | |
1492 | ||
1493 | static void | |
1494 | scan_sharing_clauses (tree clauses, omp_context *ctx) | |
1495 | { | |
1496 | tree c, decl; | |
1497 | bool scan_array_reductions = false; | |
1498 | ||
1499 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
1500 | { | |
1501 | bool by_ref; | |
1502 | ||
55d6e7cd | 1503 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 1504 | { |
1505 | case OMP_CLAUSE_PRIVATE: | |
1506 | decl = OMP_CLAUSE_DECL (c); | |
fd6481cf | 1507 | if (OMP_CLAUSE_PRIVATE_OUTER_REF (c)) |
1508 | goto do_private; | |
1509 | else if (!is_variable_sized (decl)) | |
1e8e9920 | 1510 | install_var_local (decl, ctx); |
1511 | break; | |
1512 | ||
1513 | case OMP_CLAUSE_SHARED: | |
bc7bff74 | 1514 | /* Ignore shared directives in teams construct. */ |
1515 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
1516 | break; | |
fd6481cf | 1517 | gcc_assert (is_taskreg_ctx (ctx)); |
1e8e9920 | 1518 | decl = OMP_CLAUSE_DECL (c); |
e7327393 | 1519 | gcc_assert (!COMPLETE_TYPE_P (TREE_TYPE (decl)) |
1520 | || !is_variable_sized (decl)); | |
f49d7bb5 | 1521 | /* Global variables don't need to be copied, |
1522 | the receiver side will use them directly. */ | |
1523 | if (is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx))) | |
1524 | break; | |
fd6481cf | 1525 | by_ref = use_pointer_for_field (decl, ctx); |
1e8e9920 | 1526 | if (! TREE_READONLY (decl) |
1527 | || TREE_ADDRESSABLE (decl) | |
1528 | || by_ref | |
1529 | || is_reference (decl)) | |
1530 | { | |
fd6481cf | 1531 | install_var_field (decl, by_ref, 3, ctx); |
1e8e9920 | 1532 | install_var_local (decl, ctx); |
1533 | break; | |
1534 | } | |
1535 | /* We don't need to copy const scalar vars back. */ | |
55d6e7cd | 1536 | OMP_CLAUSE_SET_CODE (c, OMP_CLAUSE_FIRSTPRIVATE); |
1e8e9920 | 1537 | goto do_private; |
1538 | ||
1539 | case OMP_CLAUSE_LASTPRIVATE: | |
1540 | /* Let the corresponding firstprivate clause create | |
1541 | the variable. */ | |
1542 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) | |
1543 | break; | |
1544 | /* FALLTHRU */ | |
1545 | ||
1546 | case OMP_CLAUSE_FIRSTPRIVATE: | |
1547 | case OMP_CLAUSE_REDUCTION: | |
3d483a94 | 1548 | case OMP_CLAUSE_LINEAR: |
1e8e9920 | 1549 | decl = OMP_CLAUSE_DECL (c); |
1550 | do_private: | |
1551 | if (is_variable_sized (decl)) | |
1e8e9920 | 1552 | { |
fd6481cf | 1553 | if (is_task_ctx (ctx)) |
1554 | install_var_field (decl, false, 1, ctx); | |
1555 | break; | |
1556 | } | |
1557 | else if (is_taskreg_ctx (ctx)) | |
1558 | { | |
1559 | bool global | |
1560 | = is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)); | |
e8a588af | 1561 | by_ref = use_pointer_for_field (decl, NULL); |
fd6481cf | 1562 | |
1563 | if (is_task_ctx (ctx) | |
1564 | && (global || by_ref || is_reference (decl))) | |
1565 | { | |
1566 | install_var_field (decl, false, 1, ctx); | |
1567 | if (!global) | |
1568 | install_var_field (decl, by_ref, 2, ctx); | |
1569 | } | |
1570 | else if (!global) | |
1571 | install_var_field (decl, by_ref, 3, ctx); | |
1e8e9920 | 1572 | } |
1573 | install_var_local (decl, ctx); | |
1574 | break; | |
1575 | ||
bc7bff74 | 1576 | case OMP_CLAUSE__LOOPTEMP_: |
1577 | gcc_assert (is_parallel_ctx (ctx)); | |
1578 | decl = OMP_CLAUSE_DECL (c); | |
1579 | install_var_field (decl, false, 3, ctx); | |
1580 | install_var_local (decl, ctx); | |
1581 | break; | |
1582 | ||
1e8e9920 | 1583 | case OMP_CLAUSE_COPYPRIVATE: |
1e8e9920 | 1584 | case OMP_CLAUSE_COPYIN: |
1585 | decl = OMP_CLAUSE_DECL (c); | |
e8a588af | 1586 | by_ref = use_pointer_for_field (decl, NULL); |
fd6481cf | 1587 | install_var_field (decl, by_ref, 3, ctx); |
1e8e9920 | 1588 | break; |
1589 | ||
1590 | case OMP_CLAUSE_DEFAULT: | |
1591 | ctx->default_kind = OMP_CLAUSE_DEFAULT_KIND (c); | |
1592 | break; | |
1593 | ||
2169f33b | 1594 | case OMP_CLAUSE_FINAL: |
1e8e9920 | 1595 | case OMP_CLAUSE_IF: |
1596 | case OMP_CLAUSE_NUM_THREADS: | |
bc7bff74 | 1597 | case OMP_CLAUSE_NUM_TEAMS: |
1598 | case OMP_CLAUSE_THREAD_LIMIT: | |
1599 | case OMP_CLAUSE_DEVICE: | |
1e8e9920 | 1600 | case OMP_CLAUSE_SCHEDULE: |
bc7bff74 | 1601 | case OMP_CLAUSE_DIST_SCHEDULE: |
1602 | case OMP_CLAUSE_DEPEND: | |
1e8e9920 | 1603 | if (ctx->outer) |
75a70cf9 | 1604 | scan_omp_op (&OMP_CLAUSE_OPERAND (c, 0), ctx->outer); |
1e8e9920 | 1605 | break; |
1606 | ||
bc7bff74 | 1607 | case OMP_CLAUSE_TO: |
1608 | case OMP_CLAUSE_FROM: | |
1609 | case OMP_CLAUSE_MAP: | |
1610 | if (ctx->outer) | |
1611 | scan_omp_op (&OMP_CLAUSE_SIZE (c), ctx->outer); | |
1612 | decl = OMP_CLAUSE_DECL (c); | |
1613 | /* Global variables with "omp declare target" attribute | |
1614 | don't need to be copied, the receiver side will use them | |
1615 | directly. */ | |
1616 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
1617 | && DECL_P (decl) | |
1618 | && is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)) | |
1619 | && lookup_attribute ("omp declare target", | |
1620 | DECL_ATTRIBUTES (decl))) | |
1621 | break; | |
1622 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
1623 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER) | |
1624 | { | |
1625 | /* Ignore OMP_CLAUSE_MAP_POINTER kind for arrays in | |
1626 | #pragma omp target data, there is nothing to map for | |
1627 | those. */ | |
1628 | if (gimple_omp_target_kind (ctx->stmt) == GF_OMP_TARGET_KIND_DATA | |
1629 | && !POINTER_TYPE_P (TREE_TYPE (decl))) | |
1630 | break; | |
1631 | } | |
1632 | if (DECL_P (decl)) | |
1633 | { | |
1634 | if (DECL_SIZE (decl) | |
1635 | && TREE_CODE (DECL_SIZE (decl)) != INTEGER_CST) | |
1636 | { | |
1637 | tree decl2 = DECL_VALUE_EXPR (decl); | |
1638 | gcc_assert (TREE_CODE (decl2) == INDIRECT_REF); | |
1639 | decl2 = TREE_OPERAND (decl2, 0); | |
1640 | gcc_assert (DECL_P (decl2)); | |
1641 | install_var_field (decl2, true, 3, ctx); | |
1642 | install_var_local (decl2, ctx); | |
1643 | install_var_local (decl, ctx); | |
1644 | } | |
1645 | else | |
1646 | { | |
1647 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
1648 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
1649 | && !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) | |
1650 | && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE) | |
1651 | install_var_field (decl, true, 7, ctx); | |
1652 | else | |
1653 | install_var_field (decl, true, 3, ctx); | |
1654 | if (gimple_omp_target_kind (ctx->stmt) | |
1655 | == GF_OMP_TARGET_KIND_REGION) | |
1656 | install_var_local (decl, ctx); | |
1657 | } | |
1658 | } | |
1659 | else | |
1660 | { | |
1661 | tree base = get_base_address (decl); | |
1662 | tree nc = OMP_CLAUSE_CHAIN (c); | |
1663 | if (DECL_P (base) | |
1664 | && nc != NULL_TREE | |
1665 | && OMP_CLAUSE_CODE (nc) == OMP_CLAUSE_MAP | |
1666 | && OMP_CLAUSE_DECL (nc) == base | |
1667 | && OMP_CLAUSE_MAP_KIND (nc) == OMP_CLAUSE_MAP_POINTER | |
1668 | && integer_zerop (OMP_CLAUSE_SIZE (nc))) | |
1669 | { | |
1670 | OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) = 1; | |
1671 | OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (nc) = 1; | |
1672 | } | |
1673 | else | |
1674 | { | |
1675 | gcc_assert (!splay_tree_lookup (ctx->field_map, | |
1676 | (splay_tree_key) decl)); | |
1677 | tree field | |
1678 | = build_decl (OMP_CLAUSE_LOCATION (c), | |
1679 | FIELD_DECL, NULL_TREE, ptr_type_node); | |
1680 | DECL_ALIGN (field) = TYPE_ALIGN (ptr_type_node); | |
1681 | insert_field_into_struct (ctx->record_type, field); | |
1682 | splay_tree_insert (ctx->field_map, (splay_tree_key) decl, | |
1683 | (splay_tree_value) field); | |
1684 | } | |
1685 | } | |
1686 | break; | |
1687 | ||
1e8e9920 | 1688 | case OMP_CLAUSE_NOWAIT: |
1689 | case OMP_CLAUSE_ORDERED: | |
fd6481cf | 1690 | case OMP_CLAUSE_COLLAPSE: |
1691 | case OMP_CLAUSE_UNTIED: | |
2169f33b | 1692 | case OMP_CLAUSE_MERGEABLE: |
bc7bff74 | 1693 | case OMP_CLAUSE_PROC_BIND: |
3d483a94 | 1694 | case OMP_CLAUSE_SAFELEN: |
1e8e9920 | 1695 | break; |
1696 | ||
bc7bff74 | 1697 | case OMP_CLAUSE_ALIGNED: |
1698 | decl = OMP_CLAUSE_DECL (c); | |
1699 | if (is_global_var (decl) | |
1700 | && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE) | |
1701 | install_var_local (decl, ctx); | |
1702 | break; | |
1703 | ||
1e8e9920 | 1704 | default: |
1705 | gcc_unreachable (); | |
1706 | } | |
1707 | } | |
1708 | ||
1709 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
1710 | { | |
55d6e7cd | 1711 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 1712 | { |
1713 | case OMP_CLAUSE_LASTPRIVATE: | |
1714 | /* Let the corresponding firstprivate clause create | |
1715 | the variable. */ | |
75a70cf9 | 1716 | if (OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)) |
fd6481cf | 1717 | scan_array_reductions = true; |
1e8e9920 | 1718 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) |
1719 | break; | |
1720 | /* FALLTHRU */ | |
1721 | ||
1722 | case OMP_CLAUSE_PRIVATE: | |
1723 | case OMP_CLAUSE_FIRSTPRIVATE: | |
1724 | case OMP_CLAUSE_REDUCTION: | |
3d483a94 | 1725 | case OMP_CLAUSE_LINEAR: |
1e8e9920 | 1726 | decl = OMP_CLAUSE_DECL (c); |
1727 | if (is_variable_sized (decl)) | |
1728 | install_var_local (decl, ctx); | |
1729 | fixup_remapped_decl (decl, ctx, | |
55d6e7cd | 1730 | OMP_CLAUSE_CODE (c) == OMP_CLAUSE_PRIVATE |
1e8e9920 | 1731 | && OMP_CLAUSE_PRIVATE_DEBUG (c)); |
55d6e7cd | 1732 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION |
1e8e9920 | 1733 | && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) |
1734 | scan_array_reductions = true; | |
1735 | break; | |
1736 | ||
1737 | case OMP_CLAUSE_SHARED: | |
bc7bff74 | 1738 | /* Ignore shared directives in teams construct. */ |
1739 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
1740 | break; | |
1e8e9920 | 1741 | decl = OMP_CLAUSE_DECL (c); |
f49d7bb5 | 1742 | if (! is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx))) |
1743 | fixup_remapped_decl (decl, ctx, false); | |
1e8e9920 | 1744 | break; |
1745 | ||
bc7bff74 | 1746 | case OMP_CLAUSE_MAP: |
1747 | if (gimple_omp_target_kind (ctx->stmt) == GF_OMP_TARGET_KIND_DATA) | |
1748 | break; | |
1749 | decl = OMP_CLAUSE_DECL (c); | |
1750 | if (DECL_P (decl) | |
1751 | && is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)) | |
1752 | && lookup_attribute ("omp declare target", | |
1753 | DECL_ATTRIBUTES (decl))) | |
1754 | break; | |
1755 | if (DECL_P (decl)) | |
1756 | { | |
1757 | if (OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
1758 | && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE | |
1759 | && !COMPLETE_TYPE_P (TREE_TYPE (decl))) | |
1760 | { | |
1761 | tree new_decl = lookup_decl (decl, ctx); | |
1762 | TREE_TYPE (new_decl) | |
1763 | = remap_type (TREE_TYPE (decl), &ctx->cb); | |
1764 | } | |
1765 | else if (DECL_SIZE (decl) | |
1766 | && TREE_CODE (DECL_SIZE (decl)) != INTEGER_CST) | |
1767 | { | |
1768 | tree decl2 = DECL_VALUE_EXPR (decl); | |
1769 | gcc_assert (TREE_CODE (decl2) == INDIRECT_REF); | |
1770 | decl2 = TREE_OPERAND (decl2, 0); | |
1771 | gcc_assert (DECL_P (decl2)); | |
1772 | fixup_remapped_decl (decl2, ctx, false); | |
1773 | fixup_remapped_decl (decl, ctx, true); | |
1774 | } | |
1775 | else | |
1776 | fixup_remapped_decl (decl, ctx, false); | |
1777 | } | |
1778 | break; | |
1779 | ||
1e8e9920 | 1780 | case OMP_CLAUSE_COPYPRIVATE: |
1781 | case OMP_CLAUSE_COPYIN: | |
1782 | case OMP_CLAUSE_DEFAULT: | |
1783 | case OMP_CLAUSE_IF: | |
1784 | case OMP_CLAUSE_NUM_THREADS: | |
bc7bff74 | 1785 | case OMP_CLAUSE_NUM_TEAMS: |
1786 | case OMP_CLAUSE_THREAD_LIMIT: | |
1787 | case OMP_CLAUSE_DEVICE: | |
1e8e9920 | 1788 | case OMP_CLAUSE_SCHEDULE: |
bc7bff74 | 1789 | case OMP_CLAUSE_DIST_SCHEDULE: |
1e8e9920 | 1790 | case OMP_CLAUSE_NOWAIT: |
1791 | case OMP_CLAUSE_ORDERED: | |
fd6481cf | 1792 | case OMP_CLAUSE_COLLAPSE: |
1793 | case OMP_CLAUSE_UNTIED: | |
2169f33b | 1794 | case OMP_CLAUSE_FINAL: |
1795 | case OMP_CLAUSE_MERGEABLE: | |
bc7bff74 | 1796 | case OMP_CLAUSE_PROC_BIND: |
3d483a94 | 1797 | case OMP_CLAUSE_SAFELEN: |
bc7bff74 | 1798 | case OMP_CLAUSE_ALIGNED: |
1799 | case OMP_CLAUSE_DEPEND: | |
1800 | case OMP_CLAUSE__LOOPTEMP_: | |
1801 | case OMP_CLAUSE_TO: | |
1802 | case OMP_CLAUSE_FROM: | |
1e8e9920 | 1803 | break; |
1804 | ||
1805 | default: | |
1806 | gcc_unreachable (); | |
1807 | } | |
1808 | } | |
1809 | ||
1810 | if (scan_array_reductions) | |
1811 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
55d6e7cd | 1812 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION |
1e8e9920 | 1813 | && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) |
1814 | { | |
ab129075 | 1815 | scan_omp (&OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c), ctx); |
1816 | scan_omp (&OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c), ctx); | |
1e8e9920 | 1817 | } |
fd6481cf | 1818 | else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE |
75a70cf9 | 1819 | && OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)) |
ab129075 | 1820 | scan_omp (&OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c), ctx); |
1e8e9920 | 1821 | } |
1822 | ||
1823 | /* Create a new name for omp child function. Returns an identifier. */ | |
1824 | ||
1825 | static GTY(()) unsigned int tmp_ompfn_id_num; | |
1826 | ||
1827 | static tree | |
fd6481cf | 1828 | create_omp_child_function_name (bool task_copy) |
1e8e9920 | 1829 | { |
a70a5e2c | 1830 | return (clone_function_name (current_function_decl, |
1831 | task_copy ? "_omp_cpyfn" : "_omp_fn")); | |
1e8e9920 | 1832 | } |
1833 | ||
1834 | /* Build a decl for the omp child function. It'll not contain a body | |
1835 | yet, just the bare decl. */ | |
1836 | ||
1837 | static void | |
fd6481cf | 1838 | create_omp_child_function (omp_context *ctx, bool task_copy) |
1e8e9920 | 1839 | { |
1840 | tree decl, type, name, t; | |
1841 | ||
fd6481cf | 1842 | name = create_omp_child_function_name (task_copy); |
1843 | if (task_copy) | |
1844 | type = build_function_type_list (void_type_node, ptr_type_node, | |
1845 | ptr_type_node, NULL_TREE); | |
1846 | else | |
1847 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
1e8e9920 | 1848 | |
e60a6f7b | 1849 | decl = build_decl (gimple_location (ctx->stmt), |
1850 | FUNCTION_DECL, name, type); | |
1e8e9920 | 1851 | |
fd6481cf | 1852 | if (!task_copy) |
1853 | ctx->cb.dst_fn = decl; | |
1854 | else | |
75a70cf9 | 1855 | gimple_omp_task_set_copy_fn (ctx->stmt, decl); |
1e8e9920 | 1856 | |
1857 | TREE_STATIC (decl) = 1; | |
1858 | TREE_USED (decl) = 1; | |
1859 | DECL_ARTIFICIAL (decl) = 1; | |
84bfaaeb | 1860 | DECL_NAMELESS (decl) = 1; |
1e8e9920 | 1861 | DECL_IGNORED_P (decl) = 0; |
1862 | TREE_PUBLIC (decl) = 0; | |
1863 | DECL_UNINLINABLE (decl) = 1; | |
1864 | DECL_EXTERNAL (decl) = 0; | |
1865 | DECL_CONTEXT (decl) = NULL_TREE; | |
773c5ba7 | 1866 | DECL_INITIAL (decl) = make_node (BLOCK); |
bc7bff74 | 1867 | bool target_p = false; |
1868 | if (lookup_attribute ("omp declare target", | |
1869 | DECL_ATTRIBUTES (current_function_decl))) | |
1870 | target_p = true; | |
1871 | else | |
1872 | { | |
1873 | omp_context *octx; | |
1874 | for (octx = ctx; octx; octx = octx->outer) | |
1875 | if (gimple_code (octx->stmt) == GIMPLE_OMP_TARGET | |
1876 | && gimple_omp_target_kind (octx->stmt) | |
1877 | == GF_OMP_TARGET_KIND_REGION) | |
1878 | { | |
1879 | target_p = true; | |
1880 | break; | |
1881 | } | |
1882 | } | |
1883 | if (target_p) | |
1884 | DECL_ATTRIBUTES (decl) | |
1885 | = tree_cons (get_identifier ("omp declare target"), | |
1886 | NULL_TREE, DECL_ATTRIBUTES (decl)); | |
1e8e9920 | 1887 | |
e60a6f7b | 1888 | t = build_decl (DECL_SOURCE_LOCATION (decl), |
1889 | RESULT_DECL, NULL_TREE, void_type_node); | |
1e8e9920 | 1890 | DECL_ARTIFICIAL (t) = 1; |
1891 | DECL_IGNORED_P (t) = 1; | |
8e5b4ed6 | 1892 | DECL_CONTEXT (t) = decl; |
1e8e9920 | 1893 | DECL_RESULT (decl) = t; |
1894 | ||
e60a6f7b | 1895 | t = build_decl (DECL_SOURCE_LOCATION (decl), |
1896 | PARM_DECL, get_identifier (".omp_data_i"), ptr_type_node); | |
1e8e9920 | 1897 | DECL_ARTIFICIAL (t) = 1; |
84bfaaeb | 1898 | DECL_NAMELESS (t) = 1; |
1e8e9920 | 1899 | DECL_ARG_TYPE (t) = ptr_type_node; |
773c5ba7 | 1900 | DECL_CONTEXT (t) = current_function_decl; |
1e8e9920 | 1901 | TREE_USED (t) = 1; |
1902 | DECL_ARGUMENTS (decl) = t; | |
fd6481cf | 1903 | if (!task_copy) |
1904 | ctx->receiver_decl = t; | |
1905 | else | |
1906 | { | |
e60a6f7b | 1907 | t = build_decl (DECL_SOURCE_LOCATION (decl), |
1908 | PARM_DECL, get_identifier (".omp_data_o"), | |
fd6481cf | 1909 | ptr_type_node); |
1910 | DECL_ARTIFICIAL (t) = 1; | |
84bfaaeb | 1911 | DECL_NAMELESS (t) = 1; |
fd6481cf | 1912 | DECL_ARG_TYPE (t) = ptr_type_node; |
1913 | DECL_CONTEXT (t) = current_function_decl; | |
1914 | TREE_USED (t) = 1; | |
86f2ad37 | 1915 | TREE_ADDRESSABLE (t) = 1; |
1767a056 | 1916 | DECL_CHAIN (t) = DECL_ARGUMENTS (decl); |
fd6481cf | 1917 | DECL_ARGUMENTS (decl) = t; |
1918 | } | |
1e8e9920 | 1919 | |
48e1416a | 1920 | /* Allocate memory for the function structure. The call to |
773c5ba7 | 1921 | allocate_struct_function clobbers CFUN, so we need to restore |
1e8e9920 | 1922 | it afterward. */ |
87d4aa85 | 1923 | push_struct_function (decl); |
75a70cf9 | 1924 | cfun->function_end_locus = gimple_location (ctx->stmt); |
87d4aa85 | 1925 | pop_cfun (); |
1e8e9920 | 1926 | } |
1927 | ||
bc7bff74 | 1928 | /* Callback for walk_gimple_seq. Check if combined parallel |
1929 | contains gimple_omp_for_combined_into_p OMP_FOR. */ | |
1930 | ||
1931 | static tree | |
1932 | find_combined_for (gimple_stmt_iterator *gsi_p, | |
1933 | bool *handled_ops_p, | |
1934 | struct walk_stmt_info *wi) | |
1935 | { | |
1936 | gimple stmt = gsi_stmt (*gsi_p); | |
1937 | ||
1938 | *handled_ops_p = true; | |
1939 | switch (gimple_code (stmt)) | |
1940 | { | |
1941 | WALK_SUBSTMTS; | |
1942 | ||
1943 | case GIMPLE_OMP_FOR: | |
1944 | if (gimple_omp_for_combined_into_p (stmt) | |
1945 | && gimple_omp_for_kind (stmt) == GF_OMP_FOR_KIND_FOR) | |
1946 | { | |
1947 | wi->info = stmt; | |
1948 | return integer_zero_node; | |
1949 | } | |
1950 | break; | |
1951 | default: | |
1952 | break; | |
1953 | } | |
1954 | return NULL; | |
1955 | } | |
1956 | ||
1e8e9920 | 1957 | /* Scan an OpenMP parallel directive. */ |
1958 | ||
1959 | static void | |
75a70cf9 | 1960 | scan_omp_parallel (gimple_stmt_iterator *gsi, omp_context *outer_ctx) |
1e8e9920 | 1961 | { |
1962 | omp_context *ctx; | |
1963 | tree name; | |
75a70cf9 | 1964 | gimple stmt = gsi_stmt (*gsi); |
1e8e9920 | 1965 | |
1966 | /* Ignore parallel directives with empty bodies, unless there | |
1967 | are copyin clauses. */ | |
1968 | if (optimize > 0 | |
75a70cf9 | 1969 | && empty_body_p (gimple_omp_body (stmt)) |
1970 | && find_omp_clause (gimple_omp_parallel_clauses (stmt), | |
1971 | OMP_CLAUSE_COPYIN) == NULL) | |
1e8e9920 | 1972 | { |
75a70cf9 | 1973 | gsi_replace (gsi, gimple_build_nop (), false); |
1e8e9920 | 1974 | return; |
1975 | } | |
1976 | ||
bc7bff74 | 1977 | if (gimple_omp_parallel_combined_p (stmt)) |
1978 | { | |
1979 | gimple for_stmt; | |
1980 | struct walk_stmt_info wi; | |
1981 | ||
1982 | memset (&wi, 0, sizeof (wi)); | |
1983 | wi.val_only = true; | |
1984 | walk_gimple_seq (gimple_omp_body (stmt), | |
1985 | find_combined_for, NULL, &wi); | |
1986 | for_stmt = (gimple) wi.info; | |
1987 | if (for_stmt) | |
1988 | { | |
1989 | struct omp_for_data fd; | |
1990 | extract_omp_for_data (for_stmt, &fd, NULL); | |
1991 | /* We need two temporaries with fd.loop.v type (istart/iend) | |
1992 | and then (fd.collapse - 1) temporaries with the same | |
1993 | type for count2 ... countN-1 vars if not constant. */ | |
1994 | size_t count = 2, i; | |
1995 | tree type = fd.iter_type; | |
1996 | if (fd.collapse > 1 | |
1997 | && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
1998 | count += fd.collapse - 1; | |
1999 | for (i = 0; i < count; i++) | |
2000 | { | |
2001 | tree temp = create_tmp_var (type, NULL); | |
2002 | tree c = build_omp_clause (UNKNOWN_LOCATION, | |
2003 | OMP_CLAUSE__LOOPTEMP_); | |
2004 | OMP_CLAUSE_DECL (c) = temp; | |
2005 | OMP_CLAUSE_CHAIN (c) = gimple_omp_parallel_clauses (stmt); | |
2006 | gimple_omp_parallel_set_clauses (stmt, c); | |
2007 | } | |
2008 | } | |
2009 | } | |
2010 | ||
75a70cf9 | 2011 | ctx = new_omp_context (stmt, outer_ctx); |
fd6481cf | 2012 | if (taskreg_nesting_level > 1) |
773c5ba7 | 2013 | ctx->is_nested = true; |
1e8e9920 | 2014 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); |
1e8e9920 | 2015 | ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED; |
2016 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
1e8e9920 | 2017 | name = create_tmp_var_name (".omp_data_s"); |
e60a6f7b | 2018 | name = build_decl (gimple_location (stmt), |
2019 | TYPE_DECL, name, ctx->record_type); | |
84bfaaeb | 2020 | DECL_ARTIFICIAL (name) = 1; |
2021 | DECL_NAMELESS (name) = 1; | |
1e8e9920 | 2022 | TYPE_NAME (ctx->record_type) = name; |
fd6481cf | 2023 | create_omp_child_function (ctx, false); |
75a70cf9 | 2024 | gimple_omp_parallel_set_child_fn (stmt, ctx->cb.dst_fn); |
1e8e9920 | 2025 | |
75a70cf9 | 2026 | scan_sharing_clauses (gimple_omp_parallel_clauses (stmt), ctx); |
ab129075 | 2027 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2028 | |
2029 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2030 | ctx->record_type = ctx->receiver_decl = NULL; | |
2031 | else | |
2032 | { | |
2033 | layout_type (ctx->record_type); | |
2034 | fixup_child_record_type (ctx); | |
2035 | } | |
2036 | } | |
2037 | ||
fd6481cf | 2038 | /* Scan an OpenMP task directive. */ |
2039 | ||
2040 | static void | |
75a70cf9 | 2041 | scan_omp_task (gimple_stmt_iterator *gsi, omp_context *outer_ctx) |
fd6481cf | 2042 | { |
2043 | omp_context *ctx; | |
75a70cf9 | 2044 | tree name, t; |
2045 | gimple stmt = gsi_stmt (*gsi); | |
389dd41b | 2046 | location_t loc = gimple_location (stmt); |
fd6481cf | 2047 | |
2048 | /* Ignore task directives with empty bodies. */ | |
2049 | if (optimize > 0 | |
75a70cf9 | 2050 | && empty_body_p (gimple_omp_body (stmt))) |
fd6481cf | 2051 | { |
75a70cf9 | 2052 | gsi_replace (gsi, gimple_build_nop (), false); |
fd6481cf | 2053 | return; |
2054 | } | |
2055 | ||
75a70cf9 | 2056 | ctx = new_omp_context (stmt, outer_ctx); |
fd6481cf | 2057 | if (taskreg_nesting_level > 1) |
2058 | ctx->is_nested = true; | |
2059 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
2060 | ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED; | |
2061 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
2062 | name = create_tmp_var_name (".omp_data_s"); | |
e60a6f7b | 2063 | name = build_decl (gimple_location (stmt), |
2064 | TYPE_DECL, name, ctx->record_type); | |
84bfaaeb | 2065 | DECL_ARTIFICIAL (name) = 1; |
2066 | DECL_NAMELESS (name) = 1; | |
fd6481cf | 2067 | TYPE_NAME (ctx->record_type) = name; |
2068 | create_omp_child_function (ctx, false); | |
75a70cf9 | 2069 | gimple_omp_task_set_child_fn (stmt, ctx->cb.dst_fn); |
fd6481cf | 2070 | |
75a70cf9 | 2071 | scan_sharing_clauses (gimple_omp_task_clauses (stmt), ctx); |
fd6481cf | 2072 | |
2073 | if (ctx->srecord_type) | |
2074 | { | |
2075 | name = create_tmp_var_name (".omp_data_a"); | |
e60a6f7b | 2076 | name = build_decl (gimple_location (stmt), |
2077 | TYPE_DECL, name, ctx->srecord_type); | |
84bfaaeb | 2078 | DECL_ARTIFICIAL (name) = 1; |
2079 | DECL_NAMELESS (name) = 1; | |
fd6481cf | 2080 | TYPE_NAME (ctx->srecord_type) = name; |
2081 | create_omp_child_function (ctx, true); | |
2082 | } | |
2083 | ||
ab129075 | 2084 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
fd6481cf | 2085 | |
2086 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2087 | { | |
2088 | ctx->record_type = ctx->receiver_decl = NULL; | |
75a70cf9 | 2089 | t = build_int_cst (long_integer_type_node, 0); |
2090 | gimple_omp_task_set_arg_size (stmt, t); | |
2091 | t = build_int_cst (long_integer_type_node, 1); | |
2092 | gimple_omp_task_set_arg_align (stmt, t); | |
fd6481cf | 2093 | } |
2094 | else | |
2095 | { | |
2096 | tree *p, vla_fields = NULL_TREE, *q = &vla_fields; | |
2097 | /* Move VLA fields to the end. */ | |
2098 | p = &TYPE_FIELDS (ctx->record_type); | |
2099 | while (*p) | |
2100 | if (!TYPE_SIZE_UNIT (TREE_TYPE (*p)) | |
2101 | || ! TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (*p)))) | |
2102 | { | |
2103 | *q = *p; | |
2104 | *p = TREE_CHAIN (*p); | |
2105 | TREE_CHAIN (*q) = NULL_TREE; | |
2106 | q = &TREE_CHAIN (*q); | |
2107 | } | |
2108 | else | |
1767a056 | 2109 | p = &DECL_CHAIN (*p); |
fd6481cf | 2110 | *p = vla_fields; |
2111 | layout_type (ctx->record_type); | |
2112 | fixup_child_record_type (ctx); | |
2113 | if (ctx->srecord_type) | |
2114 | layout_type (ctx->srecord_type); | |
389dd41b | 2115 | t = fold_convert_loc (loc, long_integer_type_node, |
fd6481cf | 2116 | TYPE_SIZE_UNIT (ctx->record_type)); |
75a70cf9 | 2117 | gimple_omp_task_set_arg_size (stmt, t); |
2118 | t = build_int_cst (long_integer_type_node, | |
fd6481cf | 2119 | TYPE_ALIGN_UNIT (ctx->record_type)); |
75a70cf9 | 2120 | gimple_omp_task_set_arg_align (stmt, t); |
fd6481cf | 2121 | } |
2122 | } | |
2123 | ||
1e8e9920 | 2124 | |
773c5ba7 | 2125 | /* Scan an OpenMP loop directive. */ |
1e8e9920 | 2126 | |
2127 | static void | |
75a70cf9 | 2128 | scan_omp_for (gimple stmt, omp_context *outer_ctx) |
1e8e9920 | 2129 | { |
773c5ba7 | 2130 | omp_context *ctx; |
75a70cf9 | 2131 | size_t i; |
1e8e9920 | 2132 | |
773c5ba7 | 2133 | ctx = new_omp_context (stmt, outer_ctx); |
1e8e9920 | 2134 | |
75a70cf9 | 2135 | scan_sharing_clauses (gimple_omp_for_clauses (stmt), ctx); |
1e8e9920 | 2136 | |
ab129075 | 2137 | scan_omp (gimple_omp_for_pre_body_ptr (stmt), ctx); |
75a70cf9 | 2138 | for (i = 0; i < gimple_omp_for_collapse (stmt); i++) |
fd6481cf | 2139 | { |
75a70cf9 | 2140 | scan_omp_op (gimple_omp_for_index_ptr (stmt, i), ctx); |
2141 | scan_omp_op (gimple_omp_for_initial_ptr (stmt, i), ctx); | |
2142 | scan_omp_op (gimple_omp_for_final_ptr (stmt, i), ctx); | |
2143 | scan_omp_op (gimple_omp_for_incr_ptr (stmt, i), ctx); | |
fd6481cf | 2144 | } |
ab129075 | 2145 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2146 | } |
2147 | ||
2148 | /* Scan an OpenMP sections directive. */ | |
2149 | ||
2150 | static void | |
75a70cf9 | 2151 | scan_omp_sections (gimple stmt, omp_context *outer_ctx) |
1e8e9920 | 2152 | { |
1e8e9920 | 2153 | omp_context *ctx; |
2154 | ||
2155 | ctx = new_omp_context (stmt, outer_ctx); | |
75a70cf9 | 2156 | scan_sharing_clauses (gimple_omp_sections_clauses (stmt), ctx); |
ab129075 | 2157 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2158 | } |
2159 | ||
2160 | /* Scan an OpenMP single directive. */ | |
2161 | ||
2162 | static void | |
75a70cf9 | 2163 | scan_omp_single (gimple stmt, omp_context *outer_ctx) |
1e8e9920 | 2164 | { |
1e8e9920 | 2165 | omp_context *ctx; |
2166 | tree name; | |
2167 | ||
2168 | ctx = new_omp_context (stmt, outer_ctx); | |
2169 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
2170 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
2171 | name = create_tmp_var_name (".omp_copy_s"); | |
e60a6f7b | 2172 | name = build_decl (gimple_location (stmt), |
2173 | TYPE_DECL, name, ctx->record_type); | |
1e8e9920 | 2174 | TYPE_NAME (ctx->record_type) = name; |
2175 | ||
75a70cf9 | 2176 | scan_sharing_clauses (gimple_omp_single_clauses (stmt), ctx); |
ab129075 | 2177 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2178 | |
2179 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2180 | ctx->record_type = NULL; | |
2181 | else | |
2182 | layout_type (ctx->record_type); | |
2183 | } | |
2184 | ||
bc7bff74 | 2185 | /* Scan an OpenMP target{, data, update} directive. */ |
2186 | ||
2187 | static void | |
2188 | scan_omp_target (gimple stmt, omp_context *outer_ctx) | |
2189 | { | |
2190 | omp_context *ctx; | |
2191 | tree name; | |
2192 | int kind = gimple_omp_target_kind (stmt); | |
2193 | ||
2194 | ctx = new_omp_context (stmt, outer_ctx); | |
2195 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
2196 | ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED; | |
2197 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
2198 | name = create_tmp_var_name (".omp_data_t"); | |
2199 | name = build_decl (gimple_location (stmt), | |
2200 | TYPE_DECL, name, ctx->record_type); | |
2201 | DECL_ARTIFICIAL (name) = 1; | |
2202 | DECL_NAMELESS (name) = 1; | |
2203 | TYPE_NAME (ctx->record_type) = name; | |
2204 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
2205 | { | |
2206 | create_omp_child_function (ctx, false); | |
2207 | gimple_omp_target_set_child_fn (stmt, ctx->cb.dst_fn); | |
2208 | } | |
2209 | ||
2210 | scan_sharing_clauses (gimple_omp_target_clauses (stmt), ctx); | |
2211 | scan_omp (gimple_omp_body_ptr (stmt), ctx); | |
2212 | ||
2213 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2214 | ctx->record_type = ctx->receiver_decl = NULL; | |
2215 | else | |
2216 | { | |
2217 | TYPE_FIELDS (ctx->record_type) | |
2218 | = nreverse (TYPE_FIELDS (ctx->record_type)); | |
2219 | #ifdef ENABLE_CHECKING | |
2220 | tree field; | |
2221 | unsigned int align = DECL_ALIGN (TYPE_FIELDS (ctx->record_type)); | |
2222 | for (field = TYPE_FIELDS (ctx->record_type); | |
2223 | field; | |
2224 | field = DECL_CHAIN (field)) | |
2225 | gcc_assert (DECL_ALIGN (field) == align); | |
2226 | #endif | |
2227 | layout_type (ctx->record_type); | |
2228 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
2229 | fixup_child_record_type (ctx); | |
2230 | } | |
2231 | } | |
2232 | ||
2233 | /* Scan an OpenMP teams directive. */ | |
2234 | ||
2235 | static void | |
2236 | scan_omp_teams (gimple stmt, omp_context *outer_ctx) | |
2237 | { | |
2238 | omp_context *ctx = new_omp_context (stmt, outer_ctx); | |
2239 | scan_sharing_clauses (gimple_omp_teams_clauses (stmt), ctx); | |
2240 | scan_omp (gimple_omp_body_ptr (stmt), ctx); | |
2241 | } | |
1e8e9920 | 2242 | |
c1d127dd | 2243 | /* Check OpenMP nesting restrictions. */ |
ab129075 | 2244 | static bool |
2245 | check_omp_nesting_restrictions (gimple stmt, omp_context *ctx) | |
c1d127dd | 2246 | { |
3d483a94 | 2247 | if (ctx != NULL) |
2248 | { | |
2249 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR | |
f2697631 | 2250 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_KIND_SIMD) |
3d483a94 | 2251 | { |
2252 | error_at (gimple_location (stmt), | |
2253 | "OpenMP constructs may not be nested inside simd region"); | |
2254 | return false; | |
2255 | } | |
bc7bff74 | 2256 | else if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) |
2257 | { | |
2258 | if ((gimple_code (stmt) != GIMPLE_OMP_FOR | |
2259 | || (gimple_omp_for_kind (stmt) | |
2260 | != GF_OMP_FOR_KIND_DISTRIBUTE)) | |
2261 | && gimple_code (stmt) != GIMPLE_OMP_PARALLEL) | |
2262 | { | |
2263 | error_at (gimple_location (stmt), | |
2264 | "only distribute or parallel constructs are allowed to " | |
2265 | "be closely nested inside teams construct"); | |
2266 | return false; | |
2267 | } | |
2268 | } | |
3d483a94 | 2269 | } |
75a70cf9 | 2270 | switch (gimple_code (stmt)) |
c1d127dd | 2271 | { |
75a70cf9 | 2272 | case GIMPLE_OMP_FOR: |
f2697631 | 2273 | if (gimple_omp_for_kind (stmt) & GF_OMP_FOR_KIND_SIMD) |
3d483a94 | 2274 | return true; |
bc7bff74 | 2275 | if (gimple_omp_for_kind (stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
2276 | { | |
2277 | if (ctx != NULL && gimple_code (ctx->stmt) != GIMPLE_OMP_TEAMS) | |
2278 | { | |
2279 | error_at (gimple_location (stmt), | |
2280 | "distribute construct must be closely nested inside " | |
2281 | "teams construct"); | |
2282 | return false; | |
2283 | } | |
2284 | return true; | |
2285 | } | |
2286 | /* FALLTHRU */ | |
2287 | case GIMPLE_CALL: | |
2288 | if (is_gimple_call (stmt) | |
2289 | && (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2290 | == BUILT_IN_GOMP_CANCEL | |
2291 | || DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2292 | == BUILT_IN_GOMP_CANCELLATION_POINT)) | |
2293 | { | |
2294 | const char *bad = NULL; | |
2295 | const char *kind = NULL; | |
2296 | if (ctx == NULL) | |
2297 | { | |
2298 | error_at (gimple_location (stmt), "orphaned %qs construct", | |
2299 | DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2300 | == BUILT_IN_GOMP_CANCEL | |
2301 | ? "#pragma omp cancel" | |
2302 | : "#pragma omp cancellation point"); | |
2303 | return false; | |
2304 | } | |
35ec552a | 2305 | switch (tree_fits_shwi_p (gimple_call_arg (stmt, 0)) |
fcb97e84 | 2306 | ? tree_to_shwi (gimple_call_arg (stmt, 0)) |
bc7bff74 | 2307 | : 0) |
2308 | { | |
2309 | case 1: | |
2310 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_PARALLEL) | |
2311 | bad = "#pragma omp parallel"; | |
2312 | else if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2313 | == BUILT_IN_GOMP_CANCEL | |
2314 | && !integer_zerop (gimple_call_arg (stmt, 1))) | |
2315 | ctx->cancellable = true; | |
2316 | kind = "parallel"; | |
2317 | break; | |
2318 | case 2: | |
2319 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_FOR | |
2320 | || gimple_omp_for_kind (ctx->stmt) != GF_OMP_FOR_KIND_FOR) | |
2321 | bad = "#pragma omp for"; | |
2322 | else if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2323 | == BUILT_IN_GOMP_CANCEL | |
2324 | && !integer_zerop (gimple_call_arg (stmt, 1))) | |
2325 | { | |
2326 | ctx->cancellable = true; | |
2327 | if (find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
2328 | OMP_CLAUSE_NOWAIT)) | |
2329 | warning_at (gimple_location (stmt), 0, | |
2330 | "%<#pragma omp cancel for%> inside " | |
2331 | "%<nowait%> for construct"); | |
2332 | if (find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
2333 | OMP_CLAUSE_ORDERED)) | |
2334 | warning_at (gimple_location (stmt), 0, | |
2335 | "%<#pragma omp cancel for%> inside " | |
2336 | "%<ordered%> for construct"); | |
2337 | } | |
2338 | kind = "for"; | |
2339 | break; | |
2340 | case 4: | |
2341 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_SECTIONS | |
2342 | && gimple_code (ctx->stmt) != GIMPLE_OMP_SECTION) | |
2343 | bad = "#pragma omp sections"; | |
2344 | else if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2345 | == BUILT_IN_GOMP_CANCEL | |
2346 | && !integer_zerop (gimple_call_arg (stmt, 1))) | |
2347 | { | |
2348 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_SECTIONS) | |
2349 | { | |
2350 | ctx->cancellable = true; | |
2351 | if (find_omp_clause (gimple_omp_sections_clauses | |
2352 | (ctx->stmt), | |
2353 | OMP_CLAUSE_NOWAIT)) | |
2354 | warning_at (gimple_location (stmt), 0, | |
2355 | "%<#pragma omp cancel sections%> inside " | |
2356 | "%<nowait%> sections construct"); | |
2357 | } | |
2358 | else | |
2359 | { | |
2360 | gcc_assert (ctx->outer | |
2361 | && gimple_code (ctx->outer->stmt) | |
2362 | == GIMPLE_OMP_SECTIONS); | |
2363 | ctx->outer->cancellable = true; | |
2364 | if (find_omp_clause (gimple_omp_sections_clauses | |
2365 | (ctx->outer->stmt), | |
2366 | OMP_CLAUSE_NOWAIT)) | |
2367 | warning_at (gimple_location (stmt), 0, | |
2368 | "%<#pragma omp cancel sections%> inside " | |
2369 | "%<nowait%> sections construct"); | |
2370 | } | |
2371 | } | |
2372 | kind = "sections"; | |
2373 | break; | |
2374 | case 8: | |
2375 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_TASK) | |
2376 | bad = "#pragma omp task"; | |
2377 | else | |
2378 | ctx->cancellable = true; | |
2379 | kind = "taskgroup"; | |
2380 | break; | |
2381 | default: | |
2382 | error_at (gimple_location (stmt), "invalid arguments"); | |
2383 | return false; | |
2384 | } | |
2385 | if (bad) | |
2386 | { | |
2387 | error_at (gimple_location (stmt), | |
2388 | "%<%s %s%> construct not closely nested inside of %qs", | |
2389 | DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2390 | == BUILT_IN_GOMP_CANCEL | |
2391 | ? "#pragma omp cancel" | |
2392 | : "#pragma omp cancellation point", kind, bad); | |
2393 | return false; | |
2394 | } | |
2395 | } | |
3d483a94 | 2396 | /* FALLTHRU */ |
75a70cf9 | 2397 | case GIMPLE_OMP_SECTIONS: |
2398 | case GIMPLE_OMP_SINGLE: | |
c1d127dd | 2399 | for (; ctx != NULL; ctx = ctx->outer) |
75a70cf9 | 2400 | switch (gimple_code (ctx->stmt)) |
c1d127dd | 2401 | { |
75a70cf9 | 2402 | case GIMPLE_OMP_FOR: |
2403 | case GIMPLE_OMP_SECTIONS: | |
2404 | case GIMPLE_OMP_SINGLE: | |
2405 | case GIMPLE_OMP_ORDERED: | |
2406 | case GIMPLE_OMP_MASTER: | |
2407 | case GIMPLE_OMP_TASK: | |
bc7bff74 | 2408 | case GIMPLE_OMP_CRITICAL: |
75a70cf9 | 2409 | if (is_gimple_call (stmt)) |
fd6481cf | 2410 | { |
bc7bff74 | 2411 | if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) |
2412 | != BUILT_IN_GOMP_BARRIER) | |
2413 | return true; | |
ab129075 | 2414 | error_at (gimple_location (stmt), |
2415 | "barrier region may not be closely nested inside " | |
2416 | "of work-sharing, critical, ordered, master or " | |
2417 | "explicit task region"); | |
2418 | return false; | |
fd6481cf | 2419 | } |
ab129075 | 2420 | error_at (gimple_location (stmt), |
2421 | "work-sharing region may not be closely nested inside " | |
2422 | "of work-sharing, critical, ordered, master or explicit " | |
2423 | "task region"); | |
2424 | return false; | |
75a70cf9 | 2425 | case GIMPLE_OMP_PARALLEL: |
ab129075 | 2426 | return true; |
c1d127dd | 2427 | default: |
2428 | break; | |
2429 | } | |
2430 | break; | |
75a70cf9 | 2431 | case GIMPLE_OMP_MASTER: |
c1d127dd | 2432 | for (; ctx != NULL; ctx = ctx->outer) |
75a70cf9 | 2433 | switch (gimple_code (ctx->stmt)) |
c1d127dd | 2434 | { |
75a70cf9 | 2435 | case GIMPLE_OMP_FOR: |
2436 | case GIMPLE_OMP_SECTIONS: | |
2437 | case GIMPLE_OMP_SINGLE: | |
2438 | case GIMPLE_OMP_TASK: | |
ab129075 | 2439 | error_at (gimple_location (stmt), |
2440 | "master region may not be closely nested inside " | |
2441 | "of work-sharing or explicit task region"); | |
2442 | return false; | |
75a70cf9 | 2443 | case GIMPLE_OMP_PARALLEL: |
ab129075 | 2444 | return true; |
c1d127dd | 2445 | default: |
2446 | break; | |
2447 | } | |
2448 | break; | |
75a70cf9 | 2449 | case GIMPLE_OMP_ORDERED: |
c1d127dd | 2450 | for (; ctx != NULL; ctx = ctx->outer) |
75a70cf9 | 2451 | switch (gimple_code (ctx->stmt)) |
c1d127dd | 2452 | { |
75a70cf9 | 2453 | case GIMPLE_OMP_CRITICAL: |
2454 | case GIMPLE_OMP_TASK: | |
ab129075 | 2455 | error_at (gimple_location (stmt), |
2456 | "ordered region may not be closely nested inside " | |
2457 | "of critical or explicit task region"); | |
2458 | return false; | |
75a70cf9 | 2459 | case GIMPLE_OMP_FOR: |
2460 | if (find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
c1d127dd | 2461 | OMP_CLAUSE_ORDERED) == NULL) |
ab129075 | 2462 | { |
2463 | error_at (gimple_location (stmt), | |
2464 | "ordered region must be closely nested inside " | |
c1d127dd | 2465 | "a loop region with an ordered clause"); |
ab129075 | 2466 | return false; |
2467 | } | |
2468 | return true; | |
75a70cf9 | 2469 | case GIMPLE_OMP_PARALLEL: |
bc7bff74 | 2470 | error_at (gimple_location (stmt), |
2471 | "ordered region must be closely nested inside " | |
2472 | "a loop region with an ordered clause"); | |
2473 | return false; | |
c1d127dd | 2474 | default: |
2475 | break; | |
2476 | } | |
2477 | break; | |
75a70cf9 | 2478 | case GIMPLE_OMP_CRITICAL: |
c1d127dd | 2479 | for (; ctx != NULL; ctx = ctx->outer) |
75a70cf9 | 2480 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_CRITICAL |
2481 | && (gimple_omp_critical_name (stmt) | |
2482 | == gimple_omp_critical_name (ctx->stmt))) | |
c1d127dd | 2483 | { |
ab129075 | 2484 | error_at (gimple_location (stmt), |
2485 | "critical region may not be nested inside a critical " | |
2486 | "region with the same name"); | |
2487 | return false; | |
c1d127dd | 2488 | } |
2489 | break; | |
bc7bff74 | 2490 | case GIMPLE_OMP_TEAMS: |
2491 | if (ctx == NULL | |
2492 | || gimple_code (ctx->stmt) != GIMPLE_OMP_TARGET | |
2493 | || gimple_omp_target_kind (ctx->stmt) != GF_OMP_TARGET_KIND_REGION) | |
2494 | { | |
2495 | error_at (gimple_location (stmt), | |
2496 | "teams construct not closely nested inside of target " | |
2497 | "region"); | |
2498 | return false; | |
2499 | } | |
2500 | break; | |
c1d127dd | 2501 | default: |
2502 | break; | |
2503 | } | |
ab129075 | 2504 | return true; |
c1d127dd | 2505 | } |
2506 | ||
2507 | ||
75a70cf9 | 2508 | /* Helper function scan_omp. |
2509 | ||
2510 | Callback for walk_tree or operators in walk_gimple_stmt used to | |
2511 | scan for OpenMP directives in TP. */ | |
1e8e9920 | 2512 | |
2513 | static tree | |
75a70cf9 | 2514 | scan_omp_1_op (tree *tp, int *walk_subtrees, void *data) |
1e8e9920 | 2515 | { |
4077bf7a | 2516 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; |
2517 | omp_context *ctx = (omp_context *) wi->info; | |
1e8e9920 | 2518 | tree t = *tp; |
2519 | ||
75a70cf9 | 2520 | switch (TREE_CODE (t)) |
2521 | { | |
2522 | case VAR_DECL: | |
2523 | case PARM_DECL: | |
2524 | case LABEL_DECL: | |
2525 | case RESULT_DECL: | |
2526 | if (ctx) | |
2527 | *tp = remap_decl (t, &ctx->cb); | |
2528 | break; | |
2529 | ||
2530 | default: | |
2531 | if (ctx && TYPE_P (t)) | |
2532 | *tp = remap_type (t, &ctx->cb); | |
2533 | else if (!DECL_P (t)) | |
7cf869dd | 2534 | { |
2535 | *walk_subtrees = 1; | |
2536 | if (ctx) | |
182cf5a9 | 2537 | { |
2538 | tree tem = remap_type (TREE_TYPE (t), &ctx->cb); | |
2539 | if (tem != TREE_TYPE (t)) | |
2540 | { | |
2541 | if (TREE_CODE (t) == INTEGER_CST) | |
2542 | *tp = build_int_cst_wide (tem, | |
2543 | TREE_INT_CST_LOW (t), | |
2544 | TREE_INT_CST_HIGH (t)); | |
2545 | else | |
2546 | TREE_TYPE (t) = tem; | |
2547 | } | |
2548 | } | |
7cf869dd | 2549 | } |
75a70cf9 | 2550 | break; |
2551 | } | |
2552 | ||
2553 | return NULL_TREE; | |
2554 | } | |
2555 | ||
f2697631 | 2556 | /* Return true if FNDECL is a setjmp or a longjmp. */ |
2557 | ||
2558 | static bool | |
2559 | setjmp_or_longjmp_p (const_tree fndecl) | |
2560 | { | |
2561 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL | |
2562 | && (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_SETJMP | |
2563 | || DECL_FUNCTION_CODE (fndecl) == BUILT_IN_LONGJMP)) | |
2564 | return true; | |
2565 | ||
2566 | tree declname = DECL_NAME (fndecl); | |
2567 | if (!declname) | |
2568 | return false; | |
2569 | const char *name = IDENTIFIER_POINTER (declname); | |
2570 | return !strcmp (name, "setjmp") || !strcmp (name, "longjmp"); | |
2571 | } | |
2572 | ||
75a70cf9 | 2573 | |
2574 | /* Helper function for scan_omp. | |
2575 | ||
2576 | Callback for walk_gimple_stmt used to scan for OpenMP directives in | |
2577 | the current statement in GSI. */ | |
2578 | ||
2579 | static tree | |
2580 | scan_omp_1_stmt (gimple_stmt_iterator *gsi, bool *handled_ops_p, | |
2581 | struct walk_stmt_info *wi) | |
2582 | { | |
2583 | gimple stmt = gsi_stmt (*gsi); | |
2584 | omp_context *ctx = (omp_context *) wi->info; | |
2585 | ||
2586 | if (gimple_has_location (stmt)) | |
2587 | input_location = gimple_location (stmt); | |
1e8e9920 | 2588 | |
c1d127dd | 2589 | /* Check the OpenMP nesting restrictions. */ |
bc7bff74 | 2590 | bool remove = false; |
2591 | if (is_gimple_omp (stmt)) | |
2592 | remove = !check_omp_nesting_restrictions (stmt, ctx); | |
2593 | else if (is_gimple_call (stmt)) | |
2594 | { | |
2595 | tree fndecl = gimple_call_fndecl (stmt); | |
f2697631 | 2596 | if (fndecl) |
2597 | { | |
2598 | if (setjmp_or_longjmp_p (fndecl) | |
2599 | && ctx | |
2600 | && gimple_code (ctx->stmt) == GIMPLE_OMP_FOR | |
2601 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_KIND_SIMD) | |
2602 | { | |
2603 | remove = true; | |
2604 | error_at (gimple_location (stmt), | |
2605 | "setjmp/longjmp inside simd construct"); | |
2606 | } | |
2607 | else if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
2608 | switch (DECL_FUNCTION_CODE (fndecl)) | |
2609 | { | |
2610 | case BUILT_IN_GOMP_BARRIER: | |
2611 | case BUILT_IN_GOMP_CANCEL: | |
2612 | case BUILT_IN_GOMP_CANCELLATION_POINT: | |
2613 | case BUILT_IN_GOMP_TASKYIELD: | |
2614 | case BUILT_IN_GOMP_TASKWAIT: | |
2615 | case BUILT_IN_GOMP_TASKGROUP_START: | |
2616 | case BUILT_IN_GOMP_TASKGROUP_END: | |
2617 | remove = !check_omp_nesting_restrictions (stmt, ctx); | |
2618 | break; | |
2619 | default: | |
2620 | break; | |
2621 | } | |
2622 | } | |
bc7bff74 | 2623 | } |
2624 | if (remove) | |
2625 | { | |
2626 | stmt = gimple_build_nop (); | |
2627 | gsi_replace (gsi, stmt, false); | |
fd6481cf | 2628 | } |
c1d127dd | 2629 | |
75a70cf9 | 2630 | *handled_ops_p = true; |
2631 | ||
2632 | switch (gimple_code (stmt)) | |
1e8e9920 | 2633 | { |
75a70cf9 | 2634 | case GIMPLE_OMP_PARALLEL: |
fd6481cf | 2635 | taskreg_nesting_level++; |
75a70cf9 | 2636 | scan_omp_parallel (gsi, ctx); |
fd6481cf | 2637 | taskreg_nesting_level--; |
2638 | break; | |
2639 | ||
75a70cf9 | 2640 | case GIMPLE_OMP_TASK: |
fd6481cf | 2641 | taskreg_nesting_level++; |
75a70cf9 | 2642 | scan_omp_task (gsi, ctx); |
fd6481cf | 2643 | taskreg_nesting_level--; |
1e8e9920 | 2644 | break; |
2645 | ||
75a70cf9 | 2646 | case GIMPLE_OMP_FOR: |
2647 | scan_omp_for (stmt, ctx); | |
1e8e9920 | 2648 | break; |
2649 | ||
75a70cf9 | 2650 | case GIMPLE_OMP_SECTIONS: |
2651 | scan_omp_sections (stmt, ctx); | |
1e8e9920 | 2652 | break; |
2653 | ||
75a70cf9 | 2654 | case GIMPLE_OMP_SINGLE: |
2655 | scan_omp_single (stmt, ctx); | |
1e8e9920 | 2656 | break; |
2657 | ||
75a70cf9 | 2658 | case GIMPLE_OMP_SECTION: |
2659 | case GIMPLE_OMP_MASTER: | |
bc7bff74 | 2660 | case GIMPLE_OMP_TASKGROUP: |
75a70cf9 | 2661 | case GIMPLE_OMP_ORDERED: |
2662 | case GIMPLE_OMP_CRITICAL: | |
2663 | ctx = new_omp_context (stmt, ctx); | |
ab129075 | 2664 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2665 | break; |
2666 | ||
bc7bff74 | 2667 | case GIMPLE_OMP_TARGET: |
2668 | scan_omp_target (stmt, ctx); | |
2669 | break; | |
2670 | ||
2671 | case GIMPLE_OMP_TEAMS: | |
2672 | scan_omp_teams (stmt, ctx); | |
2673 | break; | |
2674 | ||
75a70cf9 | 2675 | case GIMPLE_BIND: |
1e8e9920 | 2676 | { |
2677 | tree var; | |
1e8e9920 | 2678 | |
75a70cf9 | 2679 | *handled_ops_p = false; |
2680 | if (ctx) | |
1767a056 | 2681 | for (var = gimple_bind_vars (stmt); var ; var = DECL_CHAIN (var)) |
75a70cf9 | 2682 | insert_decl_map (&ctx->cb, var, var); |
1e8e9920 | 2683 | } |
2684 | break; | |
1e8e9920 | 2685 | default: |
75a70cf9 | 2686 | *handled_ops_p = false; |
1e8e9920 | 2687 | break; |
2688 | } | |
2689 | ||
2690 | return NULL_TREE; | |
2691 | } | |
2692 | ||
2693 | ||
75a70cf9 | 2694 | /* Scan all the statements starting at the current statement. CTX |
2695 | contains context information about the OpenMP directives and | |
2696 | clauses found during the scan. */ | |
1e8e9920 | 2697 | |
2698 | static void | |
ab129075 | 2699 | scan_omp (gimple_seq *body_p, omp_context *ctx) |
1e8e9920 | 2700 | { |
2701 | location_t saved_location; | |
2702 | struct walk_stmt_info wi; | |
2703 | ||
2704 | memset (&wi, 0, sizeof (wi)); | |
1e8e9920 | 2705 | wi.info = ctx; |
1e8e9920 | 2706 | wi.want_locations = true; |
2707 | ||
2708 | saved_location = input_location; | |
ab129075 | 2709 | walk_gimple_seq_mod (body_p, scan_omp_1_stmt, scan_omp_1_op, &wi); |
1e8e9920 | 2710 | input_location = saved_location; |
2711 | } | |
2712 | \f | |
2713 | /* Re-gimplification and code generation routines. */ | |
2714 | ||
2715 | /* Build a call to GOMP_barrier. */ | |
2716 | ||
bc7bff74 | 2717 | static gimple |
2718 | build_omp_barrier (tree lhs) | |
2719 | { | |
2720 | tree fndecl = builtin_decl_explicit (lhs ? BUILT_IN_GOMP_BARRIER_CANCEL | |
2721 | : BUILT_IN_GOMP_BARRIER); | |
2722 | gimple g = gimple_build_call (fndecl, 0); | |
2723 | if (lhs) | |
2724 | gimple_call_set_lhs (g, lhs); | |
2725 | return g; | |
1e8e9920 | 2726 | } |
2727 | ||
2728 | /* If a context was created for STMT when it was scanned, return it. */ | |
2729 | ||
2730 | static omp_context * | |
75a70cf9 | 2731 | maybe_lookup_ctx (gimple stmt) |
1e8e9920 | 2732 | { |
2733 | splay_tree_node n; | |
2734 | n = splay_tree_lookup (all_contexts, (splay_tree_key) stmt); | |
2735 | return n ? (omp_context *) n->value : NULL; | |
2736 | } | |
2737 | ||
773c5ba7 | 2738 | |
2739 | /* Find the mapping for DECL in CTX or the immediately enclosing | |
2740 | context that has a mapping for DECL. | |
2741 | ||
2742 | If CTX is a nested parallel directive, we may have to use the decl | |
2743 | mappings created in CTX's parent context. Suppose that we have the | |
2744 | following parallel nesting (variable UIDs showed for clarity): | |
2745 | ||
2746 | iD.1562 = 0; | |
2747 | #omp parallel shared(iD.1562) -> outer parallel | |
2748 | iD.1562 = iD.1562 + 1; | |
2749 | ||
2750 | #omp parallel shared (iD.1562) -> inner parallel | |
2751 | iD.1562 = iD.1562 - 1; | |
2752 | ||
2753 | Each parallel structure will create a distinct .omp_data_s structure | |
2754 | for copying iD.1562 in/out of the directive: | |
2755 | ||
2756 | outer parallel .omp_data_s.1.i -> iD.1562 | |
2757 | inner parallel .omp_data_s.2.i -> iD.1562 | |
2758 | ||
2759 | A shared variable mapping will produce a copy-out operation before | |
2760 | the parallel directive and a copy-in operation after it. So, in | |
2761 | this case we would have: | |
2762 | ||
2763 | iD.1562 = 0; | |
2764 | .omp_data_o.1.i = iD.1562; | |
2765 | #omp parallel shared(iD.1562) -> outer parallel | |
2766 | .omp_data_i.1 = &.omp_data_o.1 | |
2767 | .omp_data_i.1->i = .omp_data_i.1->i + 1; | |
2768 | ||
2769 | .omp_data_o.2.i = iD.1562; -> ** | |
2770 | #omp parallel shared(iD.1562) -> inner parallel | |
2771 | .omp_data_i.2 = &.omp_data_o.2 | |
2772 | .omp_data_i.2->i = .omp_data_i.2->i - 1; | |
2773 | ||
2774 | ||
2775 | ** This is a problem. The symbol iD.1562 cannot be referenced | |
2776 | inside the body of the outer parallel region. But since we are | |
2777 | emitting this copy operation while expanding the inner parallel | |
2778 | directive, we need to access the CTX structure of the outer | |
2779 | parallel directive to get the correct mapping: | |
2780 | ||
2781 | .omp_data_o.2.i = .omp_data_i.1->i | |
2782 | ||
2783 | Since there may be other workshare or parallel directives enclosing | |
2784 | the parallel directive, it may be necessary to walk up the context | |
2785 | parent chain. This is not a problem in general because nested | |
2786 | parallelism happens only rarely. */ | |
2787 | ||
2788 | static tree | |
2789 | lookup_decl_in_outer_ctx (tree decl, omp_context *ctx) | |
2790 | { | |
2791 | tree t; | |
2792 | omp_context *up; | |
2793 | ||
773c5ba7 | 2794 | for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer) |
2795 | t = maybe_lookup_decl (decl, up); | |
2796 | ||
87b31375 | 2797 | gcc_assert (!ctx->is_nested || t || is_global_var (decl)); |
773c5ba7 | 2798 | |
c37594c7 | 2799 | return t ? t : decl; |
773c5ba7 | 2800 | } |
2801 | ||
2802 | ||
f49d7bb5 | 2803 | /* Similar to lookup_decl_in_outer_ctx, but return DECL if not found |
2804 | in outer contexts. */ | |
2805 | ||
2806 | static tree | |
2807 | maybe_lookup_decl_in_outer_ctx (tree decl, omp_context *ctx) | |
2808 | { | |
2809 | tree t = NULL; | |
2810 | omp_context *up; | |
2811 | ||
87b31375 | 2812 | for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer) |
2813 | t = maybe_lookup_decl (decl, up); | |
f49d7bb5 | 2814 | |
2815 | return t ? t : decl; | |
2816 | } | |
2817 | ||
2818 | ||
1e8e9920 | 2819 | /* Construct the initialization value for reduction CLAUSE. */ |
2820 | ||
2821 | tree | |
2822 | omp_reduction_init (tree clause, tree type) | |
2823 | { | |
389dd41b | 2824 | location_t loc = OMP_CLAUSE_LOCATION (clause); |
1e8e9920 | 2825 | switch (OMP_CLAUSE_REDUCTION_CODE (clause)) |
2826 | { | |
2827 | case PLUS_EXPR: | |
2828 | case MINUS_EXPR: | |
2829 | case BIT_IOR_EXPR: | |
2830 | case BIT_XOR_EXPR: | |
2831 | case TRUTH_OR_EXPR: | |
2832 | case TRUTH_ORIF_EXPR: | |
2833 | case TRUTH_XOR_EXPR: | |
2834 | case NE_EXPR: | |
385f3f36 | 2835 | return build_zero_cst (type); |
1e8e9920 | 2836 | |
2837 | case MULT_EXPR: | |
2838 | case TRUTH_AND_EXPR: | |
2839 | case TRUTH_ANDIF_EXPR: | |
2840 | case EQ_EXPR: | |
389dd41b | 2841 | return fold_convert_loc (loc, type, integer_one_node); |
1e8e9920 | 2842 | |
2843 | case BIT_AND_EXPR: | |
389dd41b | 2844 | return fold_convert_loc (loc, type, integer_minus_one_node); |
1e8e9920 | 2845 | |
2846 | case MAX_EXPR: | |
2847 | if (SCALAR_FLOAT_TYPE_P (type)) | |
2848 | { | |
2849 | REAL_VALUE_TYPE max, min; | |
2850 | if (HONOR_INFINITIES (TYPE_MODE (type))) | |
2851 | { | |
2852 | real_inf (&max); | |
2853 | real_arithmetic (&min, NEGATE_EXPR, &max, NULL); | |
2854 | } | |
2855 | else | |
2856 | real_maxval (&min, 1, TYPE_MODE (type)); | |
2857 | return build_real (type, min); | |
2858 | } | |
2859 | else | |
2860 | { | |
2861 | gcc_assert (INTEGRAL_TYPE_P (type)); | |
2862 | return TYPE_MIN_VALUE (type); | |
2863 | } | |
2864 | ||
2865 | case MIN_EXPR: | |
2866 | if (SCALAR_FLOAT_TYPE_P (type)) | |
2867 | { | |
2868 | REAL_VALUE_TYPE max; | |
2869 | if (HONOR_INFINITIES (TYPE_MODE (type))) | |
2870 | real_inf (&max); | |
2871 | else | |
2872 | real_maxval (&max, 0, TYPE_MODE (type)); | |
2873 | return build_real (type, max); | |
2874 | } | |
2875 | else | |
2876 | { | |
2877 | gcc_assert (INTEGRAL_TYPE_P (type)); | |
2878 | return TYPE_MAX_VALUE (type); | |
2879 | } | |
2880 | ||
2881 | default: | |
2882 | gcc_unreachable (); | |
2883 | } | |
2884 | } | |
2885 | ||
bc7bff74 | 2886 | /* Return alignment to be assumed for var in CLAUSE, which should be |
2887 | OMP_CLAUSE_ALIGNED. */ | |
2888 | ||
2889 | static tree | |
2890 | omp_clause_aligned_alignment (tree clause) | |
2891 | { | |
2892 | if (OMP_CLAUSE_ALIGNED_ALIGNMENT (clause)) | |
2893 | return OMP_CLAUSE_ALIGNED_ALIGNMENT (clause); | |
2894 | ||
2895 | /* Otherwise return implementation defined alignment. */ | |
2896 | unsigned int al = 1; | |
2897 | enum machine_mode mode, vmode; | |
2898 | int vs = targetm.vectorize.autovectorize_vector_sizes (); | |
2899 | if (vs) | |
2900 | vs = 1 << floor_log2 (vs); | |
2901 | static enum mode_class classes[] | |
2902 | = { MODE_INT, MODE_VECTOR_INT, MODE_FLOAT, MODE_VECTOR_FLOAT }; | |
2903 | for (int i = 0; i < 4; i += 2) | |
2904 | for (mode = GET_CLASS_NARROWEST_MODE (classes[i]); | |
2905 | mode != VOIDmode; | |
2906 | mode = GET_MODE_WIDER_MODE (mode)) | |
2907 | { | |
2908 | vmode = targetm.vectorize.preferred_simd_mode (mode); | |
2909 | if (GET_MODE_CLASS (vmode) != classes[i + 1]) | |
2910 | continue; | |
2911 | while (vs | |
2912 | && GET_MODE_SIZE (vmode) < vs | |
2913 | && GET_MODE_2XWIDER_MODE (vmode) != VOIDmode) | |
2914 | vmode = GET_MODE_2XWIDER_MODE (vmode); | |
2915 | ||
2916 | tree type = lang_hooks.types.type_for_mode (mode, 1); | |
2917 | if (type == NULL_TREE || TYPE_MODE (type) != mode) | |
2918 | continue; | |
2919 | type = build_vector_type (type, GET_MODE_SIZE (vmode) | |
2920 | / GET_MODE_SIZE (mode)); | |
2921 | if (TYPE_MODE (type) != vmode) | |
2922 | continue; | |
2923 | if (TYPE_ALIGN_UNIT (type) > al) | |
2924 | al = TYPE_ALIGN_UNIT (type); | |
2925 | } | |
2926 | return build_int_cst (integer_type_node, al); | |
2927 | } | |
2928 | ||
3d483a94 | 2929 | /* Return maximum possible vectorization factor for the target. */ |
2930 | ||
2931 | static int | |
2932 | omp_max_vf (void) | |
2933 | { | |
2934 | if (!optimize | |
2935 | || optimize_debug | |
043115ec | 2936 | || (!flag_tree_loop_vectorize |
2937 | && (global_options_set.x_flag_tree_loop_vectorize | |
2938 | || global_options_set.x_flag_tree_vectorize))) | |
3d483a94 | 2939 | return 1; |
2940 | ||
2941 | int vs = targetm.vectorize.autovectorize_vector_sizes (); | |
2942 | if (vs) | |
2943 | { | |
2944 | vs = 1 << floor_log2 (vs); | |
2945 | return vs; | |
2946 | } | |
2947 | enum machine_mode vqimode = targetm.vectorize.preferred_simd_mode (QImode); | |
2948 | if (GET_MODE_CLASS (vqimode) == MODE_VECTOR_INT) | |
2949 | return GET_MODE_NUNITS (vqimode); | |
2950 | return 1; | |
2951 | } | |
2952 | ||
2953 | /* Helper function of lower_rec_input_clauses, used for #pragma omp simd | |
2954 | privatization. */ | |
2955 | ||
2956 | static bool | |
2957 | lower_rec_simd_input_clauses (tree new_var, omp_context *ctx, int &max_vf, | |
2958 | tree &idx, tree &lane, tree &ivar, tree &lvar) | |
2959 | { | |
2960 | if (max_vf == 0) | |
2961 | { | |
2962 | max_vf = omp_max_vf (); | |
2963 | if (max_vf > 1) | |
2964 | { | |
2965 | tree c = find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
2966 | OMP_CLAUSE_SAFELEN); | |
2967 | if (c | |
2968 | && compare_tree_int (OMP_CLAUSE_SAFELEN_EXPR (c), max_vf) == -1) | |
fcb97e84 | 2969 | max_vf = tree_to_shwi (OMP_CLAUSE_SAFELEN_EXPR (c)); |
3d483a94 | 2970 | } |
2971 | if (max_vf > 1) | |
2972 | { | |
2973 | idx = create_tmp_var (unsigned_type_node, NULL); | |
2974 | lane = create_tmp_var (unsigned_type_node, NULL); | |
2975 | } | |
2976 | } | |
2977 | if (max_vf == 1) | |
2978 | return false; | |
2979 | ||
2980 | tree atype = build_array_type_nelts (TREE_TYPE (new_var), max_vf); | |
2981 | tree avar = create_tmp_var_raw (atype, NULL); | |
2982 | if (TREE_ADDRESSABLE (new_var)) | |
2983 | TREE_ADDRESSABLE (avar) = 1; | |
2984 | DECL_ATTRIBUTES (avar) | |
2985 | = tree_cons (get_identifier ("omp simd array"), NULL, | |
2986 | DECL_ATTRIBUTES (avar)); | |
2987 | gimple_add_tmp_var (avar); | |
2988 | ivar = build4 (ARRAY_REF, TREE_TYPE (new_var), avar, idx, | |
2989 | NULL_TREE, NULL_TREE); | |
2990 | lvar = build4 (ARRAY_REF, TREE_TYPE (new_var), avar, lane, | |
2991 | NULL_TREE, NULL_TREE); | |
bc7bff74 | 2992 | if (DECL_P (new_var)) |
2993 | { | |
2994 | SET_DECL_VALUE_EXPR (new_var, lvar); | |
2995 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
2996 | } | |
3d483a94 | 2997 | return true; |
2998 | } | |
2999 | ||
1e8e9920 | 3000 | /* Generate code to implement the input clauses, FIRSTPRIVATE and COPYIN, |
3001 | from the receiver (aka child) side and initializers for REFERENCE_TYPE | |
3002 | private variables. Initialization statements go in ILIST, while calls | |
3003 | to destructors go in DLIST. */ | |
3004 | ||
3005 | static void | |
75a70cf9 | 3006 | lower_rec_input_clauses (tree clauses, gimple_seq *ilist, gimple_seq *dlist, |
bc7bff74 | 3007 | omp_context *ctx, struct omp_for_data *fd) |
1e8e9920 | 3008 | { |
c2f47e15 | 3009 | tree c, dtor, copyin_seq, x, ptr; |
1e8e9920 | 3010 | bool copyin_by_ref = false; |
f49d7bb5 | 3011 | bool lastprivate_firstprivate = false; |
bc7bff74 | 3012 | bool reduction_omp_orig_ref = false; |
1e8e9920 | 3013 | int pass; |
3d483a94 | 3014 | bool is_simd = (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR |
f2697631 | 3015 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_KIND_SIMD); |
3d483a94 | 3016 | int max_vf = 0; |
3017 | tree lane = NULL_TREE, idx = NULL_TREE; | |
3018 | tree ivar = NULL_TREE, lvar = NULL_TREE; | |
3019 | gimple_seq llist[2] = { NULL, NULL }; | |
1e8e9920 | 3020 | |
1e8e9920 | 3021 | copyin_seq = NULL; |
3022 | ||
3d483a94 | 3023 | /* Set max_vf=1 (which will later enforce safelen=1) in simd loops |
3024 | with data sharing clauses referencing variable sized vars. That | |
3025 | is unnecessarily hard to support and very unlikely to result in | |
3026 | vectorized code anyway. */ | |
3027 | if (is_simd) | |
3028 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
3029 | switch (OMP_CLAUSE_CODE (c)) | |
3030 | { | |
3031 | case OMP_CLAUSE_REDUCTION: | |
3d483a94 | 3032 | case OMP_CLAUSE_PRIVATE: |
3033 | case OMP_CLAUSE_FIRSTPRIVATE: | |
3034 | case OMP_CLAUSE_LASTPRIVATE: | |
3035 | case OMP_CLAUSE_LINEAR: | |
3036 | if (is_variable_sized (OMP_CLAUSE_DECL (c))) | |
3037 | max_vf = 1; | |
3038 | break; | |
3039 | default: | |
3040 | continue; | |
3041 | } | |
3042 | ||
1e8e9920 | 3043 | /* Do all the fixed sized types in the first pass, and the variable sized |
3044 | types in the second pass. This makes sure that the scalar arguments to | |
48e1416a | 3045 | the variable sized types are processed before we use them in the |
1e8e9920 | 3046 | variable sized operations. */ |
3047 | for (pass = 0; pass < 2; ++pass) | |
3048 | { | |
3049 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
3050 | { | |
55d6e7cd | 3051 | enum omp_clause_code c_kind = OMP_CLAUSE_CODE (c); |
1e8e9920 | 3052 | tree var, new_var; |
3053 | bool by_ref; | |
389dd41b | 3054 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 3055 | |
3056 | switch (c_kind) | |
3057 | { | |
3058 | case OMP_CLAUSE_PRIVATE: | |
3059 | if (OMP_CLAUSE_PRIVATE_DEBUG (c)) | |
3060 | continue; | |
3061 | break; | |
3062 | case OMP_CLAUSE_SHARED: | |
bc7bff74 | 3063 | /* Ignore shared directives in teams construct. */ |
3064 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
3065 | continue; | |
f49d7bb5 | 3066 | if (maybe_lookup_decl (OMP_CLAUSE_DECL (c), ctx) == NULL) |
3067 | { | |
3068 | gcc_assert (is_global_var (OMP_CLAUSE_DECL (c))); | |
3069 | continue; | |
3070 | } | |
1e8e9920 | 3071 | case OMP_CLAUSE_FIRSTPRIVATE: |
1e8e9920 | 3072 | case OMP_CLAUSE_COPYIN: |
bc7bff74 | 3073 | case OMP_CLAUSE_LINEAR: |
3074 | break; | |
1e8e9920 | 3075 | case OMP_CLAUSE_REDUCTION: |
bc7bff74 | 3076 | if (OMP_CLAUSE_REDUCTION_OMP_ORIG_REF (c)) |
3077 | reduction_omp_orig_ref = true; | |
1e8e9920 | 3078 | break; |
bc7bff74 | 3079 | case OMP_CLAUSE__LOOPTEMP_: |
3080 | /* Handle _looptemp_ clauses only on parallel. */ | |
3081 | if (fd) | |
3082 | continue; | |
3d483a94 | 3083 | break; |
df2c34fc | 3084 | case OMP_CLAUSE_LASTPRIVATE: |
f49d7bb5 | 3085 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) |
3086 | { | |
3087 | lastprivate_firstprivate = true; | |
3088 | if (pass != 0) | |
3089 | continue; | |
3090 | } | |
df2c34fc | 3091 | break; |
bc7bff74 | 3092 | case OMP_CLAUSE_ALIGNED: |
3093 | if (pass == 0) | |
3094 | continue; | |
3095 | var = OMP_CLAUSE_DECL (c); | |
3096 | if (TREE_CODE (TREE_TYPE (var)) == POINTER_TYPE | |
3097 | && !is_global_var (var)) | |
3098 | { | |
3099 | new_var = maybe_lookup_decl (var, ctx); | |
3100 | if (new_var == NULL_TREE) | |
3101 | new_var = maybe_lookup_decl_in_outer_ctx (var, ctx); | |
3102 | x = builtin_decl_explicit (BUILT_IN_ASSUME_ALIGNED); | |
3103 | x = build_call_expr_loc (clause_loc, x, 2, new_var, | |
3104 | omp_clause_aligned_alignment (c)); | |
3105 | x = fold_convert_loc (clause_loc, TREE_TYPE (new_var), x); | |
3106 | x = build2 (MODIFY_EXPR, TREE_TYPE (new_var), new_var, x); | |
3107 | gimplify_and_add (x, ilist); | |
3108 | } | |
3109 | else if (TREE_CODE (TREE_TYPE (var)) == ARRAY_TYPE | |
3110 | && is_global_var (var)) | |
3111 | { | |
3112 | tree ptype = build_pointer_type (TREE_TYPE (var)), t, t2; | |
3113 | new_var = lookup_decl (var, ctx); | |
3114 | t = maybe_lookup_decl_in_outer_ctx (var, ctx); | |
3115 | t = build_fold_addr_expr_loc (clause_loc, t); | |
3116 | t2 = builtin_decl_explicit (BUILT_IN_ASSUME_ALIGNED); | |
3117 | t = build_call_expr_loc (clause_loc, t2, 2, t, | |
3118 | omp_clause_aligned_alignment (c)); | |
3119 | t = fold_convert_loc (clause_loc, ptype, t); | |
3120 | x = create_tmp_var (ptype, NULL); | |
3121 | t = build2 (MODIFY_EXPR, ptype, x, t); | |
3122 | gimplify_and_add (t, ilist); | |
3123 | t = build_simple_mem_ref_loc (clause_loc, x); | |
3124 | SET_DECL_VALUE_EXPR (new_var, t); | |
3125 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
3126 | } | |
3127 | continue; | |
1e8e9920 | 3128 | default: |
3129 | continue; | |
3130 | } | |
3131 | ||
3132 | new_var = var = OMP_CLAUSE_DECL (c); | |
3133 | if (c_kind != OMP_CLAUSE_COPYIN) | |
3134 | new_var = lookup_decl (var, ctx); | |
3135 | ||
3136 | if (c_kind == OMP_CLAUSE_SHARED || c_kind == OMP_CLAUSE_COPYIN) | |
3137 | { | |
3138 | if (pass != 0) | |
3139 | continue; | |
3140 | } | |
1e8e9920 | 3141 | else if (is_variable_sized (var)) |
3142 | { | |
773c5ba7 | 3143 | /* For variable sized types, we need to allocate the |
3144 | actual storage here. Call alloca and store the | |
3145 | result in the pointer decl that we created elsewhere. */ | |
1e8e9920 | 3146 | if (pass == 0) |
3147 | continue; | |
3148 | ||
fd6481cf | 3149 | if (c_kind != OMP_CLAUSE_FIRSTPRIVATE || !is_task_ctx (ctx)) |
3150 | { | |
75a70cf9 | 3151 | gimple stmt; |
b9a16870 | 3152 | tree tmp, atmp; |
75a70cf9 | 3153 | |
fd6481cf | 3154 | ptr = DECL_VALUE_EXPR (new_var); |
3155 | gcc_assert (TREE_CODE (ptr) == INDIRECT_REF); | |
3156 | ptr = TREE_OPERAND (ptr, 0); | |
3157 | gcc_assert (DECL_P (ptr)); | |
3158 | x = TYPE_SIZE_UNIT (TREE_TYPE (new_var)); | |
75a70cf9 | 3159 | |
3160 | /* void *tmp = __builtin_alloca */ | |
b9a16870 | 3161 | atmp = builtin_decl_explicit (BUILT_IN_ALLOCA); |
3162 | stmt = gimple_build_call (atmp, 1, x); | |
75a70cf9 | 3163 | tmp = create_tmp_var_raw (ptr_type_node, NULL); |
3164 | gimple_add_tmp_var (tmp); | |
3165 | gimple_call_set_lhs (stmt, tmp); | |
3166 | ||
3167 | gimple_seq_add_stmt (ilist, stmt); | |
3168 | ||
389dd41b | 3169 | x = fold_convert_loc (clause_loc, TREE_TYPE (ptr), tmp); |
75a70cf9 | 3170 | gimplify_assign (ptr, x, ilist); |
fd6481cf | 3171 | } |
1e8e9920 | 3172 | } |
1e8e9920 | 3173 | else if (is_reference (var)) |
3174 | { | |
773c5ba7 | 3175 | /* For references that are being privatized for Fortran, |
3176 | allocate new backing storage for the new pointer | |
3177 | variable. This allows us to avoid changing all the | |
3178 | code that expects a pointer to something that expects | |
bc7bff74 | 3179 | a direct variable. */ |
1e8e9920 | 3180 | if (pass == 0) |
3181 | continue; | |
3182 | ||
3183 | x = TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (new_var))); | |
fd6481cf | 3184 | if (c_kind == OMP_CLAUSE_FIRSTPRIVATE && is_task_ctx (ctx)) |
3185 | { | |
3186 | x = build_receiver_ref (var, false, ctx); | |
389dd41b | 3187 | x = build_fold_addr_expr_loc (clause_loc, x); |
fd6481cf | 3188 | } |
3189 | else if (TREE_CONSTANT (x)) | |
1e8e9920 | 3190 | { |
09d1c205 | 3191 | /* For reduction with placeholder in SIMD loop, |
3192 | defer adding the initialization of the reference, | |
3193 | because if we decide to use SIMD array for it, | |
3194 | the initilization could cause expansion ICE. */ | |
3195 | if (c_kind == OMP_CLAUSE_REDUCTION | |
3196 | && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c) | |
3197 | && is_simd) | |
3198 | x = NULL_TREE; | |
3199 | else | |
3200 | { | |
3201 | const char *name = NULL; | |
3202 | if (DECL_NAME (var)) | |
3203 | name = IDENTIFIER_POINTER (DECL_NAME (new_var)); | |
3204 | ||
3205 | x = create_tmp_var_raw (TREE_TYPE (TREE_TYPE (new_var)), | |
3206 | name); | |
3207 | gimple_add_tmp_var (x); | |
3208 | TREE_ADDRESSABLE (x) = 1; | |
3209 | x = build_fold_addr_expr_loc (clause_loc, x); | |
3210 | } | |
1e8e9920 | 3211 | } |
3212 | else | |
3213 | { | |
b9a16870 | 3214 | tree atmp = builtin_decl_explicit (BUILT_IN_ALLOCA); |
3215 | x = build_call_expr_loc (clause_loc, atmp, 1, x); | |
1e8e9920 | 3216 | } |
3217 | ||
09d1c205 | 3218 | if (x) |
3219 | { | |
3220 | x = fold_convert_loc (clause_loc, TREE_TYPE (new_var), x); | |
3221 | gimplify_assign (new_var, x, ilist); | |
3222 | } | |
1e8e9920 | 3223 | |
182cf5a9 | 3224 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); |
1e8e9920 | 3225 | } |
3226 | else if (c_kind == OMP_CLAUSE_REDUCTION | |
3227 | && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
3228 | { | |
3229 | if (pass == 0) | |
3230 | continue; | |
3231 | } | |
3232 | else if (pass != 0) | |
3233 | continue; | |
3234 | ||
55d6e7cd | 3235 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 3236 | { |
3237 | case OMP_CLAUSE_SHARED: | |
bc7bff74 | 3238 | /* Ignore shared directives in teams construct. */ |
3239 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
3240 | continue; | |
f49d7bb5 | 3241 | /* Shared global vars are just accessed directly. */ |
3242 | if (is_global_var (new_var)) | |
3243 | break; | |
1e8e9920 | 3244 | /* Set up the DECL_VALUE_EXPR for shared variables now. This |
3245 | needs to be delayed until after fixup_child_record_type so | |
3246 | that we get the correct type during the dereference. */ | |
e8a588af | 3247 | by_ref = use_pointer_for_field (var, ctx); |
1e8e9920 | 3248 | x = build_receiver_ref (var, by_ref, ctx); |
3249 | SET_DECL_VALUE_EXPR (new_var, x); | |
3250 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
3251 | ||
3252 | /* ??? If VAR is not passed by reference, and the variable | |
3253 | hasn't been initialized yet, then we'll get a warning for | |
3254 | the store into the omp_data_s structure. Ideally, we'd be | |
48e1416a | 3255 | able to notice this and not store anything at all, but |
1e8e9920 | 3256 | we're generating code too early. Suppress the warning. */ |
3257 | if (!by_ref) | |
3258 | TREE_NO_WARNING (var) = 1; | |
3259 | break; | |
3260 | ||
3261 | case OMP_CLAUSE_LASTPRIVATE: | |
3262 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) | |
3263 | break; | |
3264 | /* FALLTHRU */ | |
3265 | ||
3266 | case OMP_CLAUSE_PRIVATE: | |
fd6481cf | 3267 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_PRIVATE) |
3268 | x = build_outer_var_ref (var, ctx); | |
3269 | else if (OMP_CLAUSE_PRIVATE_OUTER_REF (c)) | |
3270 | { | |
3271 | if (is_task_ctx (ctx)) | |
3272 | x = build_receiver_ref (var, false, ctx); | |
3273 | else | |
3274 | x = build_outer_var_ref (var, ctx); | |
3275 | } | |
3276 | else | |
3277 | x = NULL; | |
3d483a94 | 3278 | do_private: |
bc7bff74 | 3279 | tree nx; |
3280 | nx = lang_hooks.decls.omp_clause_default_ctor (c, new_var, x); | |
3d483a94 | 3281 | if (is_simd) |
3282 | { | |
3283 | tree y = lang_hooks.decls.omp_clause_dtor (c, new_var); | |
bc7bff74 | 3284 | if ((TREE_ADDRESSABLE (new_var) || nx || y |
3d483a94 | 3285 | || OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE) |
3286 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3287 | idx, lane, ivar, lvar)) | |
3288 | { | |
bc7bff74 | 3289 | if (nx) |
3d483a94 | 3290 | x = lang_hooks.decls.omp_clause_default_ctor |
3291 | (c, unshare_expr (ivar), x); | |
bc7bff74 | 3292 | if (nx && x) |
3d483a94 | 3293 | gimplify_and_add (x, &llist[0]); |
3294 | if (y) | |
3295 | { | |
3296 | y = lang_hooks.decls.omp_clause_dtor (c, ivar); | |
3297 | if (y) | |
3298 | { | |
3299 | gimple_seq tseq = NULL; | |
3300 | ||
3301 | dtor = y; | |
3302 | gimplify_stmt (&dtor, &tseq); | |
3303 | gimple_seq_add_seq (&llist[1], tseq); | |
3304 | } | |
3305 | } | |
3306 | break; | |
3307 | } | |
3308 | } | |
bc7bff74 | 3309 | if (nx) |
3310 | gimplify_and_add (nx, ilist); | |
1e8e9920 | 3311 | /* FALLTHRU */ |
3312 | ||
3313 | do_dtor: | |
3314 | x = lang_hooks.decls.omp_clause_dtor (c, new_var); | |
3315 | if (x) | |
3316 | { | |
75a70cf9 | 3317 | gimple_seq tseq = NULL; |
3318 | ||
1e8e9920 | 3319 | dtor = x; |
75a70cf9 | 3320 | gimplify_stmt (&dtor, &tseq); |
e3a19533 | 3321 | gimple_seq_add_seq (dlist, tseq); |
1e8e9920 | 3322 | } |
3323 | break; | |
3324 | ||
3d483a94 | 3325 | case OMP_CLAUSE_LINEAR: |
3326 | if (!OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
3327 | goto do_firstprivate; | |
3328 | if (OMP_CLAUSE_LINEAR_NO_COPYOUT (c)) | |
3329 | x = NULL; | |
3330 | else | |
3331 | x = build_outer_var_ref (var, ctx); | |
3332 | goto do_private; | |
3333 | ||
1e8e9920 | 3334 | case OMP_CLAUSE_FIRSTPRIVATE: |
fd6481cf | 3335 | if (is_task_ctx (ctx)) |
3336 | { | |
3337 | if (is_reference (var) || is_variable_sized (var)) | |
3338 | goto do_dtor; | |
3339 | else if (is_global_var (maybe_lookup_decl_in_outer_ctx (var, | |
3340 | ctx)) | |
3341 | || use_pointer_for_field (var, NULL)) | |
3342 | { | |
3343 | x = build_receiver_ref (var, false, ctx); | |
3344 | SET_DECL_VALUE_EXPR (new_var, x); | |
3345 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
3346 | goto do_dtor; | |
3347 | } | |
3348 | } | |
3d483a94 | 3349 | do_firstprivate: |
1e8e9920 | 3350 | x = build_outer_var_ref (var, ctx); |
3d483a94 | 3351 | if (is_simd) |
3352 | { | |
bc7bff74 | 3353 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
3354 | && gimple_omp_for_combined_into_p (ctx->stmt)) | |
3355 | { | |
3356 | tree stept = POINTER_TYPE_P (TREE_TYPE (x)) | |
3357 | ? sizetype : TREE_TYPE (x); | |
3358 | tree t = fold_convert (stept, | |
3359 | OMP_CLAUSE_LINEAR_STEP (c)); | |
3360 | tree c = find_omp_clause (clauses, | |
3361 | OMP_CLAUSE__LOOPTEMP_); | |
3362 | gcc_assert (c); | |
3363 | tree l = OMP_CLAUSE_DECL (c); | |
3364 | if (fd->collapse == 1) | |
3365 | { | |
3366 | tree n1 = fd->loop.n1; | |
3367 | tree step = fd->loop.step; | |
3368 | tree itype = TREE_TYPE (l); | |
3369 | if (POINTER_TYPE_P (itype)) | |
3370 | itype = signed_type_for (itype); | |
3371 | l = fold_build2 (MINUS_EXPR, itype, l, n1); | |
3372 | if (TYPE_UNSIGNED (itype) | |
3373 | && fd->loop.cond_code == GT_EXPR) | |
3374 | l = fold_build2 (TRUNC_DIV_EXPR, itype, | |
3375 | fold_build1 (NEGATE_EXPR, | |
3376 | itype, l), | |
3377 | fold_build1 (NEGATE_EXPR, | |
3378 | itype, step)); | |
3379 | else | |
3380 | l = fold_build2 (TRUNC_DIV_EXPR, itype, l, step); | |
3381 | } | |
3382 | t = fold_build2 (MULT_EXPR, stept, | |
3383 | fold_convert (stept, l), t); | |
3384 | if (POINTER_TYPE_P (TREE_TYPE (x))) | |
3385 | x = fold_build2 (POINTER_PLUS_EXPR, | |
3386 | TREE_TYPE (x), x, t); | |
3387 | else | |
3388 | x = fold_build2 (PLUS_EXPR, TREE_TYPE (x), x, t); | |
3389 | } | |
3390 | ||
3d483a94 | 3391 | if ((OMP_CLAUSE_CODE (c) != OMP_CLAUSE_LINEAR |
3392 | || TREE_ADDRESSABLE (new_var)) | |
3393 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3394 | idx, lane, ivar, lvar)) | |
3395 | { | |
3396 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR) | |
3397 | { | |
3398 | tree iv = create_tmp_var (TREE_TYPE (new_var), NULL); | |
3399 | x = lang_hooks.decls.omp_clause_copy_ctor (c, iv, x); | |
3400 | gimplify_and_add (x, ilist); | |
3401 | gimple_stmt_iterator gsi | |
3402 | = gsi_start_1 (gimple_omp_body_ptr (ctx->stmt)); | |
3403 | gimple g | |
3404 | = gimple_build_assign (unshare_expr (lvar), iv); | |
3405 | gsi_insert_before_without_update (&gsi, g, | |
3406 | GSI_SAME_STMT); | |
3407 | tree stept = POINTER_TYPE_P (TREE_TYPE (x)) | |
3408 | ? sizetype : TREE_TYPE (x); | |
3409 | tree t = fold_convert (stept, | |
3410 | OMP_CLAUSE_LINEAR_STEP (c)); | |
3411 | enum tree_code code = PLUS_EXPR; | |
3412 | if (POINTER_TYPE_P (TREE_TYPE (new_var))) | |
3413 | code = POINTER_PLUS_EXPR; | |
3414 | g = gimple_build_assign_with_ops (code, iv, iv, t); | |
3415 | gsi_insert_before_without_update (&gsi, g, | |
3416 | GSI_SAME_STMT); | |
3417 | break; | |
3418 | } | |
3419 | x = lang_hooks.decls.omp_clause_copy_ctor | |
3420 | (c, unshare_expr (ivar), x); | |
3421 | gimplify_and_add (x, &llist[0]); | |
3422 | x = lang_hooks.decls.omp_clause_dtor (c, ivar); | |
3423 | if (x) | |
3424 | { | |
3425 | gimple_seq tseq = NULL; | |
3426 | ||
3427 | dtor = x; | |
3428 | gimplify_stmt (&dtor, &tseq); | |
3429 | gimple_seq_add_seq (&llist[1], tseq); | |
3430 | } | |
3431 | break; | |
3432 | } | |
3433 | } | |
1e8e9920 | 3434 | x = lang_hooks.decls.omp_clause_copy_ctor (c, new_var, x); |
3435 | gimplify_and_add (x, ilist); | |
3436 | goto do_dtor; | |
1e8e9920 | 3437 | |
bc7bff74 | 3438 | case OMP_CLAUSE__LOOPTEMP_: |
3439 | gcc_assert (is_parallel_ctx (ctx)); | |
3440 | x = build_outer_var_ref (var, ctx); | |
3441 | x = build2 (MODIFY_EXPR, TREE_TYPE (new_var), new_var, x); | |
3442 | gimplify_and_add (x, ilist); | |
3443 | break; | |
3444 | ||
1e8e9920 | 3445 | case OMP_CLAUSE_COPYIN: |
e8a588af | 3446 | by_ref = use_pointer_for_field (var, NULL); |
1e8e9920 | 3447 | x = build_receiver_ref (var, by_ref, ctx); |
3448 | x = lang_hooks.decls.omp_clause_assign_op (c, new_var, x); | |
3449 | append_to_statement_list (x, ©in_seq); | |
3450 | copyin_by_ref |= by_ref; | |
3451 | break; | |
3452 | ||
3453 | case OMP_CLAUSE_REDUCTION: | |
3454 | if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
3455 | { | |
fd6481cf | 3456 | tree placeholder = OMP_CLAUSE_REDUCTION_PLACEHOLDER (c); |
bc7bff74 | 3457 | gimple tseq; |
fd6481cf | 3458 | x = build_outer_var_ref (var, ctx); |
3459 | ||
bc7bff74 | 3460 | if (is_reference (var) |
3461 | && !useless_type_conversion_p (TREE_TYPE (placeholder), | |
3462 | TREE_TYPE (x))) | |
389dd41b | 3463 | x = build_fold_addr_expr_loc (clause_loc, x); |
fd6481cf | 3464 | SET_DECL_VALUE_EXPR (placeholder, x); |
3465 | DECL_HAS_VALUE_EXPR_P (placeholder) = 1; | |
bc7bff74 | 3466 | tree new_vard = new_var; |
3467 | if (is_reference (var)) | |
3468 | { | |
3469 | gcc_assert (TREE_CODE (new_var) == MEM_REF); | |
3470 | new_vard = TREE_OPERAND (new_var, 0); | |
3471 | gcc_assert (DECL_P (new_vard)); | |
3472 | } | |
3d483a94 | 3473 | if (is_simd |
3474 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3475 | idx, lane, ivar, lvar)) | |
3476 | { | |
bc7bff74 | 3477 | if (new_vard == new_var) |
3478 | { | |
3479 | gcc_assert (DECL_VALUE_EXPR (new_var) == lvar); | |
3480 | SET_DECL_VALUE_EXPR (new_var, ivar); | |
3481 | } | |
3482 | else | |
3483 | { | |
3484 | SET_DECL_VALUE_EXPR (new_vard, | |
3485 | build_fold_addr_expr (ivar)); | |
3486 | DECL_HAS_VALUE_EXPR_P (new_vard) = 1; | |
3487 | } | |
3488 | x = lang_hooks.decls.omp_clause_default_ctor | |
3489 | (c, unshare_expr (ivar), | |
3490 | build_outer_var_ref (var, ctx)); | |
3491 | if (x) | |
3492 | gimplify_and_add (x, &llist[0]); | |
3493 | if (OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c)) | |
3494 | { | |
3495 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c); | |
3496 | lower_omp (&tseq, ctx); | |
3497 | gimple_seq_add_seq (&llist[0], tseq); | |
3498 | } | |
3499 | OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c) = NULL; | |
3500 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c); | |
3501 | lower_omp (&tseq, ctx); | |
3502 | gimple_seq_add_seq (&llist[1], tseq); | |
3503 | OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c) = NULL; | |
3504 | DECL_HAS_VALUE_EXPR_P (placeholder) = 0; | |
3505 | if (new_vard == new_var) | |
3506 | SET_DECL_VALUE_EXPR (new_var, lvar); | |
3507 | else | |
3508 | SET_DECL_VALUE_EXPR (new_vard, | |
3509 | build_fold_addr_expr (lvar)); | |
3510 | x = lang_hooks.decls.omp_clause_dtor (c, ivar); | |
3511 | if (x) | |
3512 | { | |
3513 | tseq = NULL; | |
3514 | dtor = x; | |
3515 | gimplify_stmt (&dtor, &tseq); | |
3516 | gimple_seq_add_seq (&llist[1], tseq); | |
3517 | } | |
3518 | break; | |
3519 | } | |
09d1c205 | 3520 | /* If this is a reference to constant size reduction var |
3521 | with placeholder, we haven't emitted the initializer | |
3522 | for it because it is undesirable if SIMD arrays are used. | |
3523 | But if they aren't used, we need to emit the deferred | |
3524 | initialization now. */ | |
3525 | else if (is_reference (var) && is_simd) | |
3526 | { | |
3527 | tree z | |
3528 | = TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (new_vard))); | |
3529 | if (TREE_CONSTANT (z)) | |
3530 | { | |
3531 | const char *name = NULL; | |
3532 | if (DECL_NAME (var)) | |
3533 | name = IDENTIFIER_POINTER (DECL_NAME (new_vard)); | |
3534 | ||
3535 | z = create_tmp_var_raw | |
3536 | (TREE_TYPE (TREE_TYPE (new_vard)), name); | |
3537 | gimple_add_tmp_var (z); | |
3538 | TREE_ADDRESSABLE (z) = 1; | |
3539 | z = build_fold_addr_expr_loc (clause_loc, z); | |
3540 | gimplify_assign (new_vard, z, ilist); | |
3541 | } | |
3542 | } | |
bc7bff74 | 3543 | x = lang_hooks.decls.omp_clause_default_ctor |
3544 | (c, new_var, unshare_expr (x)); | |
3545 | if (x) | |
3546 | gimplify_and_add (x, ilist); | |
3547 | if (OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c)) | |
3548 | { | |
3549 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c); | |
3550 | lower_omp (&tseq, ctx); | |
3551 | gimple_seq_add_seq (ilist, tseq); | |
3552 | } | |
3553 | OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c) = NULL; | |
3554 | if (is_simd) | |
3555 | { | |
3556 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c); | |
3557 | lower_omp (&tseq, ctx); | |
3558 | gimple_seq_add_seq (dlist, tseq); | |
3559 | OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c) = NULL; | |
3560 | } | |
3561 | DECL_HAS_VALUE_EXPR_P (placeholder) = 0; | |
3562 | goto do_dtor; | |
3563 | } | |
3564 | else | |
3565 | { | |
3566 | x = omp_reduction_init (c, TREE_TYPE (new_var)); | |
3567 | gcc_assert (TREE_CODE (TREE_TYPE (new_var)) != ARRAY_TYPE); | |
3568 | if (is_simd | |
3569 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3570 | idx, lane, ivar, lvar)) | |
3571 | { | |
3572 | enum tree_code code = OMP_CLAUSE_REDUCTION_CODE (c); | |
3573 | tree ref = build_outer_var_ref (var, ctx); | |
3574 | ||
3575 | gimplify_assign (unshare_expr (ivar), x, &llist[0]); | |
3576 | ||
3577 | /* reduction(-:var) sums up the partial results, so it | |
3578 | acts identically to reduction(+:var). */ | |
3579 | if (code == MINUS_EXPR) | |
3580 | code = PLUS_EXPR; | |
3581 | ||
3582 | x = build2 (code, TREE_TYPE (ref), ref, ivar); | |
3d483a94 | 3583 | ref = build_outer_var_ref (var, ctx); |
3584 | gimplify_assign (ref, x, &llist[1]); | |
3585 | } | |
3586 | else | |
3587 | { | |
3588 | gimplify_assign (new_var, x, ilist); | |
3589 | if (is_simd) | |
3590 | gimplify_assign (build_outer_var_ref (var, ctx), | |
3591 | new_var, dlist); | |
3592 | } | |
1e8e9920 | 3593 | } |
3594 | break; | |
3595 | ||
3596 | default: | |
3597 | gcc_unreachable (); | |
3598 | } | |
3599 | } | |
3600 | } | |
3601 | ||
3d483a94 | 3602 | if (lane) |
3603 | { | |
3604 | tree uid = create_tmp_var (ptr_type_node, "simduid"); | |
8e1a382d | 3605 | /* Don't want uninit warnings on simduid, it is always uninitialized, |
3606 | but we use it not for the value, but for the DECL_UID only. */ | |
3607 | TREE_NO_WARNING (uid) = 1; | |
3d483a94 | 3608 | gimple g |
3609 | = gimple_build_call_internal (IFN_GOMP_SIMD_LANE, 1, uid); | |
3610 | gimple_call_set_lhs (g, lane); | |
3611 | gimple_stmt_iterator gsi = gsi_start_1 (gimple_omp_body_ptr (ctx->stmt)); | |
3612 | gsi_insert_before_without_update (&gsi, g, GSI_SAME_STMT); | |
3613 | c = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE__SIMDUID_); | |
3614 | OMP_CLAUSE__SIMDUID__DECL (c) = uid; | |
3615 | OMP_CLAUSE_CHAIN (c) = gimple_omp_for_clauses (ctx->stmt); | |
3616 | gimple_omp_for_set_clauses (ctx->stmt, c); | |
3617 | g = gimple_build_assign_with_ops (INTEGER_CST, lane, | |
3618 | build_int_cst (unsigned_type_node, 0), | |
3619 | NULL_TREE); | |
3620 | gimple_seq_add_stmt (ilist, g); | |
3621 | for (int i = 0; i < 2; i++) | |
3622 | if (llist[i]) | |
3623 | { | |
3624 | tree vf = create_tmp_var (unsigned_type_node, NULL); | |
3625 | g = gimple_build_call_internal (IFN_GOMP_SIMD_VF, 1, uid); | |
3626 | gimple_call_set_lhs (g, vf); | |
3627 | gimple_seq *seq = i == 0 ? ilist : dlist; | |
3628 | gimple_seq_add_stmt (seq, g); | |
3629 | tree t = build_int_cst (unsigned_type_node, 0); | |
3630 | g = gimple_build_assign_with_ops (INTEGER_CST, idx, t, NULL_TREE); | |
3631 | gimple_seq_add_stmt (seq, g); | |
3632 | tree body = create_artificial_label (UNKNOWN_LOCATION); | |
3633 | tree header = create_artificial_label (UNKNOWN_LOCATION); | |
3634 | tree end = create_artificial_label (UNKNOWN_LOCATION); | |
3635 | gimple_seq_add_stmt (seq, gimple_build_goto (header)); | |
3636 | gimple_seq_add_stmt (seq, gimple_build_label (body)); | |
3637 | gimple_seq_add_seq (seq, llist[i]); | |
3638 | t = build_int_cst (unsigned_type_node, 1); | |
3639 | g = gimple_build_assign_with_ops (PLUS_EXPR, idx, idx, t); | |
3640 | gimple_seq_add_stmt (seq, g); | |
3641 | gimple_seq_add_stmt (seq, gimple_build_label (header)); | |
3642 | g = gimple_build_cond (LT_EXPR, idx, vf, body, end); | |
3643 | gimple_seq_add_stmt (seq, g); | |
3644 | gimple_seq_add_stmt (seq, gimple_build_label (end)); | |
3645 | } | |
3646 | } | |
3647 | ||
1e8e9920 | 3648 | /* The copyin sequence is not to be executed by the main thread, since |
3649 | that would result in self-copies. Perhaps not visible to scalars, | |
3650 | but it certainly is to C++ operator=. */ | |
3651 | if (copyin_seq) | |
3652 | { | |
b9a16870 | 3653 | x = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM), |
3654 | 0); | |
1e8e9920 | 3655 | x = build2 (NE_EXPR, boolean_type_node, x, |
3656 | build_int_cst (TREE_TYPE (x), 0)); | |
3657 | x = build3 (COND_EXPR, void_type_node, x, copyin_seq, NULL); | |
3658 | gimplify_and_add (x, ilist); | |
3659 | } | |
3660 | ||
3661 | /* If any copyin variable is passed by reference, we must ensure the | |
3662 | master thread doesn't modify it before it is copied over in all | |
f49d7bb5 | 3663 | threads. Similarly for variables in both firstprivate and |
3664 | lastprivate clauses we need to ensure the lastprivate copying | |
bc7bff74 | 3665 | happens after firstprivate copying in all threads. And similarly |
3666 | for UDRs if initializer expression refers to omp_orig. */ | |
3667 | if (copyin_by_ref || lastprivate_firstprivate || reduction_omp_orig_ref) | |
3d483a94 | 3668 | { |
3669 | /* Don't add any barrier for #pragma omp simd or | |
3670 | #pragma omp distribute. */ | |
3671 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_FOR | |
f2697631 | 3672 | || gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_KIND_FOR) |
bc7bff74 | 3673 | gimple_seq_add_stmt (ilist, build_omp_barrier (NULL_TREE)); |
3d483a94 | 3674 | } |
3675 | ||
3676 | /* If max_vf is non-zero, then we can use only a vectorization factor | |
3677 | up to the max_vf we chose. So stick it into the safelen clause. */ | |
3678 | if (max_vf) | |
3679 | { | |
3680 | tree c = find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
3681 | OMP_CLAUSE_SAFELEN); | |
3682 | if (c == NULL_TREE | |
3683 | || compare_tree_int (OMP_CLAUSE_SAFELEN_EXPR (c), | |
3684 | max_vf) == 1) | |
3685 | { | |
3686 | c = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE_SAFELEN); | |
3687 | OMP_CLAUSE_SAFELEN_EXPR (c) = build_int_cst (integer_type_node, | |
3688 | max_vf); | |
3689 | OMP_CLAUSE_CHAIN (c) = gimple_omp_for_clauses (ctx->stmt); | |
3690 | gimple_omp_for_set_clauses (ctx->stmt, c); | |
3691 | } | |
3692 | } | |
1e8e9920 | 3693 | } |
3694 | ||
773c5ba7 | 3695 | |
1e8e9920 | 3696 | /* Generate code to implement the LASTPRIVATE clauses. This is used for |
3697 | both parallel and workshare constructs. PREDICATE may be NULL if it's | |
3698 | always true. */ | |
3699 | ||
3700 | static void | |
75a70cf9 | 3701 | lower_lastprivate_clauses (tree clauses, tree predicate, gimple_seq *stmt_list, |
bc7bff74 | 3702 | omp_context *ctx) |
1e8e9920 | 3703 | { |
3d483a94 | 3704 | tree x, c, label = NULL, orig_clauses = clauses; |
fd6481cf | 3705 | bool par_clauses = false; |
3d483a94 | 3706 | tree simduid = NULL, lastlane = NULL; |
1e8e9920 | 3707 | |
3d483a94 | 3708 | /* Early exit if there are no lastprivate or linear clauses. */ |
3709 | for (; clauses ; clauses = OMP_CLAUSE_CHAIN (clauses)) | |
3710 | if (OMP_CLAUSE_CODE (clauses) == OMP_CLAUSE_LASTPRIVATE | |
3711 | || (OMP_CLAUSE_CODE (clauses) == OMP_CLAUSE_LINEAR | |
3712 | && !OMP_CLAUSE_LINEAR_NO_COPYOUT (clauses))) | |
3713 | break; | |
1e8e9920 | 3714 | if (clauses == NULL) |
3715 | { | |
3716 | /* If this was a workshare clause, see if it had been combined | |
3717 | with its parallel. In that case, look for the clauses on the | |
3718 | parallel statement itself. */ | |
3719 | if (is_parallel_ctx (ctx)) | |
3720 | return; | |
3721 | ||
3722 | ctx = ctx->outer; | |
3723 | if (ctx == NULL || !is_parallel_ctx (ctx)) | |
3724 | return; | |
3725 | ||
75a70cf9 | 3726 | clauses = find_omp_clause (gimple_omp_parallel_clauses (ctx->stmt), |
1e8e9920 | 3727 | OMP_CLAUSE_LASTPRIVATE); |
3728 | if (clauses == NULL) | |
3729 | return; | |
fd6481cf | 3730 | par_clauses = true; |
1e8e9920 | 3731 | } |
3732 | ||
75a70cf9 | 3733 | if (predicate) |
3734 | { | |
3735 | gimple stmt; | |
3736 | tree label_true, arm1, arm2; | |
3737 | ||
e60a6f7b | 3738 | label = create_artificial_label (UNKNOWN_LOCATION); |
3739 | label_true = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 3740 | arm1 = TREE_OPERAND (predicate, 0); |
3741 | arm2 = TREE_OPERAND (predicate, 1); | |
3742 | gimplify_expr (&arm1, stmt_list, NULL, is_gimple_val, fb_rvalue); | |
3743 | gimplify_expr (&arm2, stmt_list, NULL, is_gimple_val, fb_rvalue); | |
3744 | stmt = gimple_build_cond (TREE_CODE (predicate), arm1, arm2, | |
3745 | label_true, label); | |
3746 | gimple_seq_add_stmt (stmt_list, stmt); | |
3747 | gimple_seq_add_stmt (stmt_list, gimple_build_label (label_true)); | |
3748 | } | |
1e8e9920 | 3749 | |
3d483a94 | 3750 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR |
f2697631 | 3751 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_KIND_SIMD) |
3d483a94 | 3752 | { |
3753 | simduid = find_omp_clause (orig_clauses, OMP_CLAUSE__SIMDUID_); | |
3754 | if (simduid) | |
3755 | simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); | |
3756 | } | |
3757 | ||
fd6481cf | 3758 | for (c = clauses; c ;) |
1e8e9920 | 3759 | { |
3760 | tree var, new_var; | |
389dd41b | 3761 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 3762 | |
3d483a94 | 3763 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE |
3764 | || (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
3765 | && !OMP_CLAUSE_LINEAR_NO_COPYOUT (c))) | |
fd6481cf | 3766 | { |
3767 | var = OMP_CLAUSE_DECL (c); | |
3768 | new_var = lookup_decl (var, ctx); | |
1e8e9920 | 3769 | |
3d483a94 | 3770 | if (simduid && DECL_HAS_VALUE_EXPR_P (new_var)) |
3771 | { | |
3772 | tree val = DECL_VALUE_EXPR (new_var); | |
3773 | if (TREE_CODE (val) == ARRAY_REF | |
3774 | && VAR_P (TREE_OPERAND (val, 0)) | |
3775 | && lookup_attribute ("omp simd array", | |
3776 | DECL_ATTRIBUTES (TREE_OPERAND (val, | |
3777 | 0)))) | |
3778 | { | |
3779 | if (lastlane == NULL) | |
3780 | { | |
3781 | lastlane = create_tmp_var (unsigned_type_node, NULL); | |
3782 | gimple g | |
3783 | = gimple_build_call_internal (IFN_GOMP_SIMD_LAST_LANE, | |
3784 | 2, simduid, | |
3785 | TREE_OPERAND (val, 1)); | |
3786 | gimple_call_set_lhs (g, lastlane); | |
3787 | gimple_seq_add_stmt (stmt_list, g); | |
3788 | } | |
3789 | new_var = build4 (ARRAY_REF, TREE_TYPE (val), | |
3790 | TREE_OPERAND (val, 0), lastlane, | |
3791 | NULL_TREE, NULL_TREE); | |
3792 | } | |
3793 | } | |
3794 | ||
3795 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE | |
3796 | && OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)) | |
75a70cf9 | 3797 | { |
e3a19533 | 3798 | lower_omp (&OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c), ctx); |
75a70cf9 | 3799 | gimple_seq_add_seq (stmt_list, |
3800 | OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)); | |
3d483a94 | 3801 | OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c) = NULL; |
75a70cf9 | 3802 | } |
1e8e9920 | 3803 | |
fd6481cf | 3804 | x = build_outer_var_ref (var, ctx); |
3805 | if (is_reference (var)) | |
182cf5a9 | 3806 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); |
fd6481cf | 3807 | x = lang_hooks.decls.omp_clause_assign_op (c, x, new_var); |
75a70cf9 | 3808 | gimplify_and_add (x, stmt_list); |
fd6481cf | 3809 | } |
3810 | c = OMP_CLAUSE_CHAIN (c); | |
3811 | if (c == NULL && !par_clauses) | |
3812 | { | |
3813 | /* If this was a workshare clause, see if it had been combined | |
3814 | with its parallel. In that case, continue looking for the | |
3815 | clauses also on the parallel statement itself. */ | |
3816 | if (is_parallel_ctx (ctx)) | |
3817 | break; | |
3818 | ||
3819 | ctx = ctx->outer; | |
3820 | if (ctx == NULL || !is_parallel_ctx (ctx)) | |
3821 | break; | |
3822 | ||
75a70cf9 | 3823 | c = find_omp_clause (gimple_omp_parallel_clauses (ctx->stmt), |
fd6481cf | 3824 | OMP_CLAUSE_LASTPRIVATE); |
3825 | par_clauses = true; | |
3826 | } | |
1e8e9920 | 3827 | } |
3828 | ||
75a70cf9 | 3829 | if (label) |
3830 | gimple_seq_add_stmt (stmt_list, gimple_build_label (label)); | |
1e8e9920 | 3831 | } |
3832 | ||
773c5ba7 | 3833 | |
1e8e9920 | 3834 | /* Generate code to implement the REDUCTION clauses. */ |
3835 | ||
3836 | static void | |
75a70cf9 | 3837 | lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx) |
1e8e9920 | 3838 | { |
75a70cf9 | 3839 | gimple_seq sub_seq = NULL; |
3840 | gimple stmt; | |
3841 | tree x, c; | |
1e8e9920 | 3842 | int count = 0; |
3843 | ||
3d483a94 | 3844 | /* SIMD reductions are handled in lower_rec_input_clauses. */ |
3845 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR | |
f2697631 | 3846 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_KIND_SIMD) |
3d483a94 | 3847 | return; |
3848 | ||
1e8e9920 | 3849 | /* First see if there is exactly one reduction clause. Use OMP_ATOMIC |
3850 | update in that case, otherwise use a lock. */ | |
3851 | for (c = clauses; c && count < 2; c = OMP_CLAUSE_CHAIN (c)) | |
55d6e7cd | 3852 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION) |
1e8e9920 | 3853 | { |
3854 | if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
3855 | { | |
bc7bff74 | 3856 | /* Never use OMP_ATOMIC for array reductions or UDRs. */ |
1e8e9920 | 3857 | count = -1; |
3858 | break; | |
3859 | } | |
3860 | count++; | |
3861 | } | |
3862 | ||
3863 | if (count == 0) | |
3864 | return; | |
3865 | ||
3866 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
3867 | { | |
3868 | tree var, ref, new_var; | |
3869 | enum tree_code code; | |
389dd41b | 3870 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 3871 | |
55d6e7cd | 3872 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION) |
1e8e9920 | 3873 | continue; |
3874 | ||
3875 | var = OMP_CLAUSE_DECL (c); | |
3876 | new_var = lookup_decl (var, ctx); | |
3877 | if (is_reference (var)) | |
182cf5a9 | 3878 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); |
1e8e9920 | 3879 | ref = build_outer_var_ref (var, ctx); |
3880 | code = OMP_CLAUSE_REDUCTION_CODE (c); | |
773c5ba7 | 3881 | |
3882 | /* reduction(-:var) sums up the partial results, so it acts | |
3883 | identically to reduction(+:var). */ | |
1e8e9920 | 3884 | if (code == MINUS_EXPR) |
3885 | code = PLUS_EXPR; | |
3886 | ||
3887 | if (count == 1) | |
3888 | { | |
389dd41b | 3889 | tree addr = build_fold_addr_expr_loc (clause_loc, ref); |
1e8e9920 | 3890 | |
3891 | addr = save_expr (addr); | |
3892 | ref = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (addr)), addr); | |
389dd41b | 3893 | x = fold_build2_loc (clause_loc, code, TREE_TYPE (ref), ref, new_var); |
1e8e9920 | 3894 | x = build2 (OMP_ATOMIC, void_type_node, addr, x); |
75a70cf9 | 3895 | gimplify_and_add (x, stmt_seqp); |
1e8e9920 | 3896 | return; |
3897 | } | |
3898 | ||
3899 | if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
3900 | { | |
3901 | tree placeholder = OMP_CLAUSE_REDUCTION_PLACEHOLDER (c); | |
3902 | ||
bc7bff74 | 3903 | if (is_reference (var) |
3904 | && !useless_type_conversion_p (TREE_TYPE (placeholder), | |
3905 | TREE_TYPE (ref))) | |
389dd41b | 3906 | ref = build_fold_addr_expr_loc (clause_loc, ref); |
1e8e9920 | 3907 | SET_DECL_VALUE_EXPR (placeholder, ref); |
3908 | DECL_HAS_VALUE_EXPR_P (placeholder) = 1; | |
e3a19533 | 3909 | lower_omp (&OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c), ctx); |
75a70cf9 | 3910 | gimple_seq_add_seq (&sub_seq, OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c)); |
3911 | OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c) = NULL; | |
1e8e9920 | 3912 | OMP_CLAUSE_REDUCTION_PLACEHOLDER (c) = NULL; |
3913 | } | |
3914 | else | |
3915 | { | |
3916 | x = build2 (code, TREE_TYPE (ref), ref, new_var); | |
3917 | ref = build_outer_var_ref (var, ctx); | |
75a70cf9 | 3918 | gimplify_assign (ref, x, &sub_seq); |
1e8e9920 | 3919 | } |
3920 | } | |
3921 | ||
b9a16870 | 3922 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START), |
3923 | 0); | |
75a70cf9 | 3924 | gimple_seq_add_stmt (stmt_seqp, stmt); |
1e8e9920 | 3925 | |
75a70cf9 | 3926 | gimple_seq_add_seq (stmt_seqp, sub_seq); |
1e8e9920 | 3927 | |
b9a16870 | 3928 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END), |
3929 | 0); | |
75a70cf9 | 3930 | gimple_seq_add_stmt (stmt_seqp, stmt); |
1e8e9920 | 3931 | } |
3932 | ||
773c5ba7 | 3933 | |
1e8e9920 | 3934 | /* Generate code to implement the COPYPRIVATE clauses. */ |
3935 | ||
3936 | static void | |
75a70cf9 | 3937 | lower_copyprivate_clauses (tree clauses, gimple_seq *slist, gimple_seq *rlist, |
1e8e9920 | 3938 | omp_context *ctx) |
3939 | { | |
3940 | tree c; | |
3941 | ||
3942 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
3943 | { | |
cb561506 | 3944 | tree var, new_var, ref, x; |
1e8e9920 | 3945 | bool by_ref; |
389dd41b | 3946 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 3947 | |
55d6e7cd | 3948 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYPRIVATE) |
1e8e9920 | 3949 | continue; |
3950 | ||
3951 | var = OMP_CLAUSE_DECL (c); | |
e8a588af | 3952 | by_ref = use_pointer_for_field (var, NULL); |
1e8e9920 | 3953 | |
3954 | ref = build_sender_ref (var, ctx); | |
cb561506 | 3955 | x = new_var = lookup_decl_in_outer_ctx (var, ctx); |
3956 | if (by_ref) | |
3957 | { | |
3958 | x = build_fold_addr_expr_loc (clause_loc, new_var); | |
3959 | x = fold_convert_loc (clause_loc, TREE_TYPE (ref), x); | |
3960 | } | |
75a70cf9 | 3961 | gimplify_assign (ref, x, slist); |
1e8e9920 | 3962 | |
cb561506 | 3963 | ref = build_receiver_ref (var, false, ctx); |
3964 | if (by_ref) | |
3965 | { | |
3966 | ref = fold_convert_loc (clause_loc, | |
3967 | build_pointer_type (TREE_TYPE (new_var)), | |
3968 | ref); | |
3969 | ref = build_fold_indirect_ref_loc (clause_loc, ref); | |
3970 | } | |
1e8e9920 | 3971 | if (is_reference (var)) |
3972 | { | |
cb561506 | 3973 | ref = fold_convert_loc (clause_loc, TREE_TYPE (new_var), ref); |
182cf5a9 | 3974 | ref = build_simple_mem_ref_loc (clause_loc, ref); |
3975 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); | |
1e8e9920 | 3976 | } |
cb561506 | 3977 | x = lang_hooks.decls.omp_clause_assign_op (c, new_var, ref); |
1e8e9920 | 3978 | gimplify_and_add (x, rlist); |
3979 | } | |
3980 | } | |
3981 | ||
773c5ba7 | 3982 | |
1e8e9920 | 3983 | /* Generate code to implement the clauses, FIRSTPRIVATE, COPYIN, LASTPRIVATE, |
3984 | and REDUCTION from the sender (aka parent) side. */ | |
3985 | ||
3986 | static void | |
75a70cf9 | 3987 | lower_send_clauses (tree clauses, gimple_seq *ilist, gimple_seq *olist, |
3988 | omp_context *ctx) | |
1e8e9920 | 3989 | { |
3990 | tree c; | |
3991 | ||
3992 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
3993 | { | |
773c5ba7 | 3994 | tree val, ref, x, var; |
1e8e9920 | 3995 | bool by_ref, do_in = false, do_out = false; |
389dd41b | 3996 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 3997 | |
55d6e7cd | 3998 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 3999 | { |
fd6481cf | 4000 | case OMP_CLAUSE_PRIVATE: |
4001 | if (OMP_CLAUSE_PRIVATE_OUTER_REF (c)) | |
4002 | break; | |
4003 | continue; | |
1e8e9920 | 4004 | case OMP_CLAUSE_FIRSTPRIVATE: |
4005 | case OMP_CLAUSE_COPYIN: | |
4006 | case OMP_CLAUSE_LASTPRIVATE: | |
4007 | case OMP_CLAUSE_REDUCTION: | |
bc7bff74 | 4008 | case OMP_CLAUSE__LOOPTEMP_: |
1e8e9920 | 4009 | break; |
4010 | default: | |
4011 | continue; | |
4012 | } | |
4013 | ||
87b31375 | 4014 | val = OMP_CLAUSE_DECL (c); |
4015 | var = lookup_decl_in_outer_ctx (val, ctx); | |
773c5ba7 | 4016 | |
f49d7bb5 | 4017 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYIN |
4018 | && is_global_var (var)) | |
4019 | continue; | |
1e8e9920 | 4020 | if (is_variable_sized (val)) |
4021 | continue; | |
e8a588af | 4022 | by_ref = use_pointer_for_field (val, NULL); |
1e8e9920 | 4023 | |
55d6e7cd | 4024 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 4025 | { |
fd6481cf | 4026 | case OMP_CLAUSE_PRIVATE: |
1e8e9920 | 4027 | case OMP_CLAUSE_FIRSTPRIVATE: |
4028 | case OMP_CLAUSE_COPYIN: | |
bc7bff74 | 4029 | case OMP_CLAUSE__LOOPTEMP_: |
1e8e9920 | 4030 | do_in = true; |
4031 | break; | |
4032 | ||
4033 | case OMP_CLAUSE_LASTPRIVATE: | |
4034 | if (by_ref || is_reference (val)) | |
4035 | { | |
4036 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) | |
4037 | continue; | |
4038 | do_in = true; | |
4039 | } | |
4040 | else | |
fd6481cf | 4041 | { |
4042 | do_out = true; | |
4043 | if (lang_hooks.decls.omp_private_outer_ref (val)) | |
4044 | do_in = true; | |
4045 | } | |
1e8e9920 | 4046 | break; |
4047 | ||
4048 | case OMP_CLAUSE_REDUCTION: | |
4049 | do_in = true; | |
4050 | do_out = !(by_ref || is_reference (val)); | |
4051 | break; | |
4052 | ||
4053 | default: | |
4054 | gcc_unreachable (); | |
4055 | } | |
4056 | ||
4057 | if (do_in) | |
4058 | { | |
4059 | ref = build_sender_ref (val, ctx); | |
389dd41b | 4060 | x = by_ref ? build_fold_addr_expr_loc (clause_loc, var) : var; |
75a70cf9 | 4061 | gimplify_assign (ref, x, ilist); |
fd6481cf | 4062 | if (is_task_ctx (ctx)) |
4063 | DECL_ABSTRACT_ORIGIN (TREE_OPERAND (ref, 1)) = NULL; | |
1e8e9920 | 4064 | } |
773c5ba7 | 4065 | |
1e8e9920 | 4066 | if (do_out) |
4067 | { | |
4068 | ref = build_sender_ref (val, ctx); | |
75a70cf9 | 4069 | gimplify_assign (var, ref, olist); |
1e8e9920 | 4070 | } |
4071 | } | |
4072 | } | |
4073 | ||
75a70cf9 | 4074 | /* Generate code to implement SHARED from the sender (aka parent) |
4075 | side. This is trickier, since GIMPLE_OMP_PARALLEL_CLAUSES doesn't | |
4076 | list things that got automatically shared. */ | |
1e8e9920 | 4077 | |
4078 | static void | |
75a70cf9 | 4079 | lower_send_shared_vars (gimple_seq *ilist, gimple_seq *olist, omp_context *ctx) |
1e8e9920 | 4080 | { |
fd6481cf | 4081 | tree var, ovar, nvar, f, x, record_type; |
1e8e9920 | 4082 | |
4083 | if (ctx->record_type == NULL) | |
4084 | return; | |
773c5ba7 | 4085 | |
fd6481cf | 4086 | record_type = ctx->srecord_type ? ctx->srecord_type : ctx->record_type; |
1767a056 | 4087 | for (f = TYPE_FIELDS (record_type); f ; f = DECL_CHAIN (f)) |
1e8e9920 | 4088 | { |
4089 | ovar = DECL_ABSTRACT_ORIGIN (f); | |
4090 | nvar = maybe_lookup_decl (ovar, ctx); | |
4091 | if (!nvar || !DECL_HAS_VALUE_EXPR_P (nvar)) | |
4092 | continue; | |
4093 | ||
773c5ba7 | 4094 | /* If CTX is a nested parallel directive. Find the immediately |
4095 | enclosing parallel or workshare construct that contains a | |
4096 | mapping for OVAR. */ | |
87b31375 | 4097 | var = lookup_decl_in_outer_ctx (ovar, ctx); |
773c5ba7 | 4098 | |
e8a588af | 4099 | if (use_pointer_for_field (ovar, ctx)) |
1e8e9920 | 4100 | { |
4101 | x = build_sender_ref (ovar, ctx); | |
773c5ba7 | 4102 | var = build_fold_addr_expr (var); |
75a70cf9 | 4103 | gimplify_assign (x, var, ilist); |
1e8e9920 | 4104 | } |
4105 | else | |
4106 | { | |
4107 | x = build_sender_ref (ovar, ctx); | |
75a70cf9 | 4108 | gimplify_assign (x, var, ilist); |
1e8e9920 | 4109 | |
d2263ebb | 4110 | if (!TREE_READONLY (var) |
4111 | /* We don't need to receive a new reference to a result | |
4112 | or parm decl. In fact we may not store to it as we will | |
4113 | invalidate any pending RSO and generate wrong gimple | |
4114 | during inlining. */ | |
4115 | && !((TREE_CODE (var) == RESULT_DECL | |
4116 | || TREE_CODE (var) == PARM_DECL) | |
4117 | && DECL_BY_REFERENCE (var))) | |
fd6481cf | 4118 | { |
4119 | x = build_sender_ref (ovar, ctx); | |
75a70cf9 | 4120 | gimplify_assign (var, x, olist); |
fd6481cf | 4121 | } |
1e8e9920 | 4122 | } |
4123 | } | |
4124 | } | |
4125 | ||
75a70cf9 | 4126 | |
4127 | /* A convenience function to build an empty GIMPLE_COND with just the | |
4128 | condition. */ | |
4129 | ||
4130 | static gimple | |
4131 | gimple_build_cond_empty (tree cond) | |
4132 | { | |
4133 | enum tree_code pred_code; | |
4134 | tree lhs, rhs; | |
4135 | ||
4136 | gimple_cond_get_ops_from_tree (cond, &pred_code, &lhs, &rhs); | |
4137 | return gimple_build_cond (pred_code, lhs, rhs, NULL_TREE, NULL_TREE); | |
4138 | } | |
4139 | ||
4140 | ||
48e1416a | 4141 | /* Build the function calls to GOMP_parallel_start etc to actually |
773c5ba7 | 4142 | generate the parallel operation. REGION is the parallel region |
4143 | being expanded. BB is the block where to insert the code. WS_ARGS | |
4144 | will be set if this is a call to a combined parallel+workshare | |
4145 | construct, it contains the list of additional arguments needed by | |
4146 | the workshare construct. */ | |
1e8e9920 | 4147 | |
4148 | static void | |
61e47ac8 | 4149 | expand_parallel_call (struct omp_region *region, basic_block bb, |
f1f41a6c | 4150 | gimple entry_stmt, vec<tree, va_gc> *ws_args) |
1e8e9920 | 4151 | { |
bc7bff74 | 4152 | tree t, t1, t2, val, cond, c, clauses, flags; |
75a70cf9 | 4153 | gimple_stmt_iterator gsi; |
4154 | gimple stmt; | |
b9a16870 | 4155 | enum built_in_function start_ix; |
4156 | int start_ix2; | |
389dd41b | 4157 | location_t clause_loc; |
f1f41a6c | 4158 | vec<tree, va_gc> *args; |
773c5ba7 | 4159 | |
75a70cf9 | 4160 | clauses = gimple_omp_parallel_clauses (entry_stmt); |
773c5ba7 | 4161 | |
bc7bff74 | 4162 | /* Determine what flavor of GOMP_parallel we will be |
773c5ba7 | 4163 | emitting. */ |
bc7bff74 | 4164 | start_ix = BUILT_IN_GOMP_PARALLEL; |
773c5ba7 | 4165 | if (is_combined_parallel (region)) |
4166 | { | |
61e47ac8 | 4167 | switch (region->inner->type) |
773c5ba7 | 4168 | { |
75a70cf9 | 4169 | case GIMPLE_OMP_FOR: |
fd6481cf | 4170 | gcc_assert (region->inner->sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); |
bc7bff74 | 4171 | start_ix2 = ((int)BUILT_IN_GOMP_PARALLEL_LOOP_STATIC |
b9a16870 | 4172 | + (region->inner->sched_kind |
4173 | == OMP_CLAUSE_SCHEDULE_RUNTIME | |
4174 | ? 3 : region->inner->sched_kind)); | |
4175 | start_ix = (enum built_in_function)start_ix2; | |
61e47ac8 | 4176 | break; |
75a70cf9 | 4177 | case GIMPLE_OMP_SECTIONS: |
bc7bff74 | 4178 | start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS; |
61e47ac8 | 4179 | break; |
4180 | default: | |
4181 | gcc_unreachable (); | |
773c5ba7 | 4182 | } |
773c5ba7 | 4183 | } |
1e8e9920 | 4184 | |
4185 | /* By default, the value of NUM_THREADS is zero (selected at run time) | |
4186 | and there is no conditional. */ | |
4187 | cond = NULL_TREE; | |
4188 | val = build_int_cst (unsigned_type_node, 0); | |
bc7bff74 | 4189 | flags = build_int_cst (unsigned_type_node, 0); |
1e8e9920 | 4190 | |
4191 | c = find_omp_clause (clauses, OMP_CLAUSE_IF); | |
4192 | if (c) | |
4193 | cond = OMP_CLAUSE_IF_EXPR (c); | |
4194 | ||
4195 | c = find_omp_clause (clauses, OMP_CLAUSE_NUM_THREADS); | |
4196 | if (c) | |
389dd41b | 4197 | { |
4198 | val = OMP_CLAUSE_NUM_THREADS_EXPR (c); | |
4199 | clause_loc = OMP_CLAUSE_LOCATION (c); | |
4200 | } | |
4201 | else | |
4202 | clause_loc = gimple_location (entry_stmt); | |
1e8e9920 | 4203 | |
bc7bff74 | 4204 | c = find_omp_clause (clauses, OMP_CLAUSE_PROC_BIND); |
4205 | if (c) | |
4206 | flags = build_int_cst (unsigned_type_node, OMP_CLAUSE_PROC_BIND_KIND (c)); | |
4207 | ||
1e8e9920 | 4208 | /* Ensure 'val' is of the correct type. */ |
389dd41b | 4209 | val = fold_convert_loc (clause_loc, unsigned_type_node, val); |
1e8e9920 | 4210 | |
4211 | /* If we found the clause 'if (cond)', build either | |
4212 | (cond != 0) or (cond ? val : 1u). */ | |
4213 | if (cond) | |
4214 | { | |
75a70cf9 | 4215 | gimple_stmt_iterator gsi; |
773c5ba7 | 4216 | |
4217 | cond = gimple_boolify (cond); | |
4218 | ||
1e8e9920 | 4219 | if (integer_zerop (val)) |
389dd41b | 4220 | val = fold_build2_loc (clause_loc, |
4221 | EQ_EXPR, unsigned_type_node, cond, | |
79acaae1 | 4222 | build_int_cst (TREE_TYPE (cond), 0)); |
1e8e9920 | 4223 | else |
773c5ba7 | 4224 | { |
4225 | basic_block cond_bb, then_bb, else_bb; | |
79acaae1 | 4226 | edge e, e_then, e_else; |
75a70cf9 | 4227 | tree tmp_then, tmp_else, tmp_join, tmp_var; |
79acaae1 | 4228 | |
4229 | tmp_var = create_tmp_var (TREE_TYPE (val), NULL); | |
4230 | if (gimple_in_ssa_p (cfun)) | |
4231 | { | |
75a70cf9 | 4232 | tmp_then = make_ssa_name (tmp_var, NULL); |
4233 | tmp_else = make_ssa_name (tmp_var, NULL); | |
4234 | tmp_join = make_ssa_name (tmp_var, NULL); | |
79acaae1 | 4235 | } |
4236 | else | |
4237 | { | |
4238 | tmp_then = tmp_var; | |
4239 | tmp_else = tmp_var; | |
4240 | tmp_join = tmp_var; | |
4241 | } | |
773c5ba7 | 4242 | |
773c5ba7 | 4243 | e = split_block (bb, NULL); |
4244 | cond_bb = e->src; | |
4245 | bb = e->dest; | |
4246 | remove_edge (e); | |
4247 | ||
4248 | then_bb = create_empty_bb (cond_bb); | |
4249 | else_bb = create_empty_bb (then_bb); | |
79acaae1 | 4250 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
4251 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
773c5ba7 | 4252 | |
75a70cf9 | 4253 | stmt = gimple_build_cond_empty (cond); |
4254 | gsi = gsi_start_bb (cond_bb); | |
4255 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
773c5ba7 | 4256 | |
75a70cf9 | 4257 | gsi = gsi_start_bb (then_bb); |
4258 | stmt = gimple_build_assign (tmp_then, val); | |
4259 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
773c5ba7 | 4260 | |
75a70cf9 | 4261 | gsi = gsi_start_bb (else_bb); |
4262 | stmt = gimple_build_assign | |
4263 | (tmp_else, build_int_cst (unsigned_type_node, 1)); | |
4264 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
773c5ba7 | 4265 | |
4266 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
4267 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
f6568ea4 | 4268 | if (current_loops) |
4269 | { | |
4270 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
4271 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
4272 | } | |
79acaae1 | 4273 | e_then = make_edge (then_bb, bb, EDGE_FALLTHRU); |
4274 | e_else = make_edge (else_bb, bb, EDGE_FALLTHRU); | |
773c5ba7 | 4275 | |
79acaae1 | 4276 | if (gimple_in_ssa_p (cfun)) |
4277 | { | |
75a70cf9 | 4278 | gimple phi = create_phi_node (tmp_join, bb); |
60d535d2 | 4279 | add_phi_arg (phi, tmp_then, e_then, UNKNOWN_LOCATION); |
4280 | add_phi_arg (phi, tmp_else, e_else, UNKNOWN_LOCATION); | |
79acaae1 | 4281 | } |
4282 | ||
4283 | val = tmp_join; | |
773c5ba7 | 4284 | } |
4285 | ||
75a70cf9 | 4286 | gsi = gsi_start_bb (bb); |
4287 | val = force_gimple_operand_gsi (&gsi, val, true, NULL_TREE, | |
4288 | false, GSI_CONTINUE_LINKING); | |
1e8e9920 | 4289 | } |
4290 | ||
75a70cf9 | 4291 | gsi = gsi_last_bb (bb); |
4292 | t = gimple_omp_parallel_data_arg (entry_stmt); | |
1e8e9920 | 4293 | if (t == NULL) |
c2f47e15 | 4294 | t1 = null_pointer_node; |
1e8e9920 | 4295 | else |
c2f47e15 | 4296 | t1 = build_fold_addr_expr (t); |
75a70cf9 | 4297 | t2 = build_fold_addr_expr (gimple_omp_parallel_child_fn (entry_stmt)); |
773c5ba7 | 4298 | |
bc7bff74 | 4299 | vec_alloc (args, 4 + vec_safe_length (ws_args)); |
f1f41a6c | 4300 | args->quick_push (t2); |
4301 | args->quick_push (t1); | |
4302 | args->quick_push (val); | |
4303 | if (ws_args) | |
4304 | args->splice (*ws_args); | |
bc7bff74 | 4305 | args->quick_push (flags); |
414c3a2c | 4306 | |
4307 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
b9a16870 | 4308 | builtin_decl_explicit (start_ix), args); |
773c5ba7 | 4309 | |
75a70cf9 | 4310 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4311 | false, GSI_CONTINUE_LINKING); | |
1e8e9920 | 4312 | } |
4313 | ||
773c5ba7 | 4314 | |
fd6481cf | 4315 | /* Build the function call to GOMP_task to actually |
4316 | generate the task operation. BB is the block where to insert the code. */ | |
4317 | ||
4318 | static void | |
75a70cf9 | 4319 | expand_task_call (basic_block bb, gimple entry_stmt) |
fd6481cf | 4320 | { |
bc7bff74 | 4321 | tree t, t1, t2, t3, flags, cond, c, c2, clauses, depend; |
75a70cf9 | 4322 | gimple_stmt_iterator gsi; |
389dd41b | 4323 | location_t loc = gimple_location (entry_stmt); |
fd6481cf | 4324 | |
75a70cf9 | 4325 | clauses = gimple_omp_task_clauses (entry_stmt); |
fd6481cf | 4326 | |
fd6481cf | 4327 | c = find_omp_clause (clauses, OMP_CLAUSE_IF); |
4328 | if (c) | |
4329 | cond = gimple_boolify (OMP_CLAUSE_IF_EXPR (c)); | |
4330 | else | |
4331 | cond = boolean_true_node; | |
4332 | ||
4333 | c = find_omp_clause (clauses, OMP_CLAUSE_UNTIED); | |
2169f33b | 4334 | c2 = find_omp_clause (clauses, OMP_CLAUSE_MERGEABLE); |
bc7bff74 | 4335 | depend = find_omp_clause (clauses, OMP_CLAUSE_DEPEND); |
2169f33b | 4336 | flags = build_int_cst (unsigned_type_node, |
bc7bff74 | 4337 | (c ? 1 : 0) + (c2 ? 4 : 0) + (depend ? 8 : 0)); |
2169f33b | 4338 | |
4339 | c = find_omp_clause (clauses, OMP_CLAUSE_FINAL); | |
4340 | if (c) | |
4341 | { | |
4342 | c = gimple_boolify (OMP_CLAUSE_FINAL_EXPR (c)); | |
4343 | c = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, c, | |
4344 | build_int_cst (unsigned_type_node, 2), | |
4345 | build_int_cst (unsigned_type_node, 0)); | |
4346 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, flags, c); | |
4347 | } | |
bc7bff74 | 4348 | if (depend) |
4349 | depend = OMP_CLAUSE_DECL (depend); | |
4350 | else | |
4351 | depend = build_int_cst (ptr_type_node, 0); | |
fd6481cf | 4352 | |
75a70cf9 | 4353 | gsi = gsi_last_bb (bb); |
4354 | t = gimple_omp_task_data_arg (entry_stmt); | |
fd6481cf | 4355 | if (t == NULL) |
4356 | t2 = null_pointer_node; | |
4357 | else | |
389dd41b | 4358 | t2 = build_fold_addr_expr_loc (loc, t); |
4359 | t1 = build_fold_addr_expr_loc (loc, gimple_omp_task_child_fn (entry_stmt)); | |
75a70cf9 | 4360 | t = gimple_omp_task_copy_fn (entry_stmt); |
fd6481cf | 4361 | if (t == NULL) |
4362 | t3 = null_pointer_node; | |
4363 | else | |
389dd41b | 4364 | t3 = build_fold_addr_expr_loc (loc, t); |
fd6481cf | 4365 | |
b9a16870 | 4366 | t = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASK), |
bc7bff74 | 4367 | 8, t1, t2, t3, |
75a70cf9 | 4368 | gimple_omp_task_arg_size (entry_stmt), |
bc7bff74 | 4369 | gimple_omp_task_arg_align (entry_stmt), cond, flags, |
4370 | depend); | |
fd6481cf | 4371 | |
75a70cf9 | 4372 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4373 | false, GSI_CONTINUE_LINKING); | |
fd6481cf | 4374 | } |
4375 | ||
4376 | ||
75a70cf9 | 4377 | /* If exceptions are enabled, wrap the statements in BODY in a MUST_NOT_THROW |
4378 | catch handler and return it. This prevents programs from violating the | |
4379 | structured block semantics with throws. */ | |
1e8e9920 | 4380 | |
75a70cf9 | 4381 | static gimple_seq |
4382 | maybe_catch_exception (gimple_seq body) | |
1e8e9920 | 4383 | { |
e38def9c | 4384 | gimple g; |
4385 | tree decl; | |
1e8e9920 | 4386 | |
4387 | if (!flag_exceptions) | |
75a70cf9 | 4388 | return body; |
1e8e9920 | 4389 | |
596981c8 | 4390 | if (lang_hooks.eh_protect_cleanup_actions != NULL) |
4391 | decl = lang_hooks.eh_protect_cleanup_actions (); | |
1e8e9920 | 4392 | else |
b9a16870 | 4393 | decl = builtin_decl_explicit (BUILT_IN_TRAP); |
75a70cf9 | 4394 | |
e38def9c | 4395 | g = gimple_build_eh_must_not_throw (decl); |
4396 | g = gimple_build_try (body, gimple_seq_alloc_with_stmt (g), | |
75a70cf9 | 4397 | GIMPLE_TRY_CATCH); |
1e8e9920 | 4398 | |
e38def9c | 4399 | return gimple_seq_alloc_with_stmt (g); |
1e8e9920 | 4400 | } |
4401 | ||
773c5ba7 | 4402 | /* Chain all the DECLs in LIST by their TREE_CHAIN fields. */ |
1e8e9920 | 4403 | |
773c5ba7 | 4404 | static tree |
f1f41a6c | 4405 | vec2chain (vec<tree, va_gc> *v) |
1e8e9920 | 4406 | { |
2ab2ce89 | 4407 | tree chain = NULL_TREE, t; |
4408 | unsigned ix; | |
1e8e9920 | 4409 | |
f1f41a6c | 4410 | FOR_EACH_VEC_SAFE_ELT_REVERSE (v, ix, t) |
773c5ba7 | 4411 | { |
1767a056 | 4412 | DECL_CHAIN (t) = chain; |
2ab2ce89 | 4413 | chain = t; |
773c5ba7 | 4414 | } |
1e8e9920 | 4415 | |
2ab2ce89 | 4416 | return chain; |
773c5ba7 | 4417 | } |
1e8e9920 | 4418 | |
1e8e9920 | 4419 | |
773c5ba7 | 4420 | /* Remove barriers in REGION->EXIT's block. Note that this is only |
75a70cf9 | 4421 | valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region |
4422 | is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that | |
4423 | left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be | |
773c5ba7 | 4424 | removed. */ |
1e8e9920 | 4425 | |
773c5ba7 | 4426 | static void |
4427 | remove_exit_barrier (struct omp_region *region) | |
4428 | { | |
75a70cf9 | 4429 | gimple_stmt_iterator gsi; |
773c5ba7 | 4430 | basic_block exit_bb; |
61e47ac8 | 4431 | edge_iterator ei; |
4432 | edge e; | |
75a70cf9 | 4433 | gimple stmt; |
4a04f4b4 | 4434 | int any_addressable_vars = -1; |
1e8e9920 | 4435 | |
61e47ac8 | 4436 | exit_bb = region->exit; |
1e8e9920 | 4437 | |
5056ba1a | 4438 | /* If the parallel region doesn't return, we don't have REGION->EXIT |
4439 | block at all. */ | |
4440 | if (! exit_bb) | |
4441 | return; | |
4442 | ||
75a70cf9 | 4443 | /* The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The |
4444 | workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of | |
61e47ac8 | 4445 | statements that can appear in between are extremely limited -- no |
4446 | memory operations at all. Here, we allow nothing at all, so the | |
75a70cf9 | 4447 | only thing we allow to precede this GIMPLE_OMP_RETURN is a label. */ |
4448 | gsi = gsi_last_bb (exit_bb); | |
4449 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
4450 | gsi_prev (&gsi); | |
4451 | if (!gsi_end_p (gsi) && gimple_code (gsi_stmt (gsi)) != GIMPLE_LABEL) | |
773c5ba7 | 4452 | return; |
1e8e9920 | 4453 | |
61e47ac8 | 4454 | FOR_EACH_EDGE (e, ei, exit_bb->preds) |
4455 | { | |
75a70cf9 | 4456 | gsi = gsi_last_bb (e->src); |
4457 | if (gsi_end_p (gsi)) | |
61e47ac8 | 4458 | continue; |
75a70cf9 | 4459 | stmt = gsi_stmt (gsi); |
4a04f4b4 | 4460 | if (gimple_code (stmt) == GIMPLE_OMP_RETURN |
4461 | && !gimple_omp_return_nowait_p (stmt)) | |
4462 | { | |
4463 | /* OpenMP 3.0 tasks unfortunately prevent this optimization | |
4464 | in many cases. If there could be tasks queued, the barrier | |
4465 | might be needed to let the tasks run before some local | |
4466 | variable of the parallel that the task uses as shared | |
4467 | runs out of scope. The task can be spawned either | |
4468 | from within current function (this would be easy to check) | |
4469 | or from some function it calls and gets passed an address | |
4470 | of such a variable. */ | |
4471 | if (any_addressable_vars < 0) | |
4472 | { | |
4473 | gimple parallel_stmt = last_stmt (region->entry); | |
4474 | tree child_fun = gimple_omp_parallel_child_fn (parallel_stmt); | |
2ab2ce89 | 4475 | tree local_decls, block, decl; |
4476 | unsigned ix; | |
4a04f4b4 | 4477 | |
4478 | any_addressable_vars = 0; | |
2ab2ce89 | 4479 | FOR_EACH_LOCAL_DECL (DECL_STRUCT_FUNCTION (child_fun), ix, decl) |
4480 | if (TREE_ADDRESSABLE (decl)) | |
4a04f4b4 | 4481 | { |
4482 | any_addressable_vars = 1; | |
4483 | break; | |
4484 | } | |
4485 | for (block = gimple_block (stmt); | |
4486 | !any_addressable_vars | |
4487 | && block | |
4488 | && TREE_CODE (block) == BLOCK; | |
4489 | block = BLOCK_SUPERCONTEXT (block)) | |
4490 | { | |
4491 | for (local_decls = BLOCK_VARS (block); | |
4492 | local_decls; | |
1767a056 | 4493 | local_decls = DECL_CHAIN (local_decls)) |
4a04f4b4 | 4494 | if (TREE_ADDRESSABLE (local_decls)) |
4495 | { | |
4496 | any_addressable_vars = 1; | |
4497 | break; | |
4498 | } | |
4499 | if (block == gimple_block (parallel_stmt)) | |
4500 | break; | |
4501 | } | |
4502 | } | |
4503 | if (!any_addressable_vars) | |
4504 | gimple_omp_return_set_nowait (stmt); | |
4505 | } | |
61e47ac8 | 4506 | } |
1e8e9920 | 4507 | } |
4508 | ||
61e47ac8 | 4509 | static void |
4510 | remove_exit_barriers (struct omp_region *region) | |
4511 | { | |
75a70cf9 | 4512 | if (region->type == GIMPLE_OMP_PARALLEL) |
61e47ac8 | 4513 | remove_exit_barrier (region); |
4514 | ||
4515 | if (region->inner) | |
4516 | { | |
4517 | region = region->inner; | |
4518 | remove_exit_barriers (region); | |
4519 | while (region->next) | |
4520 | { | |
4521 | region = region->next; | |
4522 | remove_exit_barriers (region); | |
4523 | } | |
4524 | } | |
4525 | } | |
773c5ba7 | 4526 | |
658b4427 | 4527 | /* Optimize omp_get_thread_num () and omp_get_num_threads () |
4528 | calls. These can't be declared as const functions, but | |
4529 | within one parallel body they are constant, so they can be | |
4530 | transformed there into __builtin_omp_get_{thread_num,num_threads} () | |
fd6481cf | 4531 | which are declared const. Similarly for task body, except |
4532 | that in untied task omp_get_thread_num () can change at any task | |
4533 | scheduling point. */ | |
658b4427 | 4534 | |
4535 | static void | |
75a70cf9 | 4536 | optimize_omp_library_calls (gimple entry_stmt) |
658b4427 | 4537 | { |
4538 | basic_block bb; | |
75a70cf9 | 4539 | gimple_stmt_iterator gsi; |
b9a16870 | 4540 | tree thr_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); |
4541 | tree thr_num_id = DECL_ASSEMBLER_NAME (thr_num_tree); | |
4542 | tree num_thr_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
4543 | tree num_thr_id = DECL_ASSEMBLER_NAME (num_thr_tree); | |
75a70cf9 | 4544 | bool untied_task = (gimple_code (entry_stmt) == GIMPLE_OMP_TASK |
4545 | && find_omp_clause (gimple_omp_task_clauses (entry_stmt), | |
fd6481cf | 4546 | OMP_CLAUSE_UNTIED) != NULL); |
658b4427 | 4547 | |
fc00614f | 4548 | FOR_EACH_BB_FN (bb, cfun) |
75a70cf9 | 4549 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
658b4427 | 4550 | { |
75a70cf9 | 4551 | gimple call = gsi_stmt (gsi); |
658b4427 | 4552 | tree decl; |
4553 | ||
75a70cf9 | 4554 | if (is_gimple_call (call) |
4555 | && (decl = gimple_call_fndecl (call)) | |
658b4427 | 4556 | && DECL_EXTERNAL (decl) |
4557 | && TREE_PUBLIC (decl) | |
4558 | && DECL_INITIAL (decl) == NULL) | |
4559 | { | |
4560 | tree built_in; | |
4561 | ||
4562 | if (DECL_NAME (decl) == thr_num_id) | |
fd6481cf | 4563 | { |
4564 | /* In #pragma omp task untied omp_get_thread_num () can change | |
4565 | during the execution of the task region. */ | |
4566 | if (untied_task) | |
4567 | continue; | |
b9a16870 | 4568 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); |
fd6481cf | 4569 | } |
658b4427 | 4570 | else if (DECL_NAME (decl) == num_thr_id) |
b9a16870 | 4571 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); |
658b4427 | 4572 | else |
4573 | continue; | |
4574 | ||
4575 | if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in) | |
75a70cf9 | 4576 | || gimple_call_num_args (call) != 0) |
658b4427 | 4577 | continue; |
4578 | ||
4579 | if (flag_exceptions && !TREE_NOTHROW (decl)) | |
4580 | continue; | |
4581 | ||
4582 | if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE | |
1ea6a73c | 4583 | || !types_compatible_p (TREE_TYPE (TREE_TYPE (decl)), |
4584 | TREE_TYPE (TREE_TYPE (built_in)))) | |
658b4427 | 4585 | continue; |
4586 | ||
0acacf9e | 4587 | gimple_call_set_fndecl (call, built_in); |
658b4427 | 4588 | } |
4589 | } | |
4590 | } | |
4591 | ||
8e6b4515 | 4592 | /* Callback for expand_omp_build_assign. Return non-NULL if *tp needs to be |
4593 | regimplified. */ | |
4594 | ||
4595 | static tree | |
4596 | expand_omp_regimplify_p (tree *tp, int *walk_subtrees, void *) | |
4597 | { | |
4598 | tree t = *tp; | |
4599 | ||
4600 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ | |
4601 | if (TREE_CODE (t) == VAR_DECL && DECL_HAS_VALUE_EXPR_P (t)) | |
4602 | return t; | |
4603 | ||
4604 | if (TREE_CODE (t) == ADDR_EXPR) | |
4605 | recompute_tree_invariant_for_addr_expr (t); | |
4606 | ||
4607 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); | |
4608 | return NULL_TREE; | |
4609 | } | |
4610 | ||
3d483a94 | 4611 | /* Prepend TO = FROM assignment before *GSI_P. */ |
4612 | ||
4613 | static void | |
4614 | expand_omp_build_assign (gimple_stmt_iterator *gsi_p, tree to, tree from) | |
4615 | { | |
4616 | bool simple_p = DECL_P (to) && TREE_ADDRESSABLE (to); | |
4617 | from = force_gimple_operand_gsi (gsi_p, from, simple_p, NULL_TREE, | |
4618 | true, GSI_SAME_STMT); | |
4619 | gimple stmt = gimple_build_assign (to, from); | |
4620 | gsi_insert_before (gsi_p, stmt, GSI_SAME_STMT); | |
4621 | if (walk_tree (&from, expand_omp_regimplify_p, NULL, NULL) | |
4622 | || walk_tree (&to, expand_omp_regimplify_p, NULL, NULL)) | |
4623 | { | |
4624 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
4625 | gimple_regimplify_operands (stmt, &gsi); | |
4626 | } | |
4627 | } | |
4628 | ||
fd6481cf | 4629 | /* Expand the OpenMP parallel or task directive starting at REGION. */ |
1e8e9920 | 4630 | |
4631 | static void | |
fd6481cf | 4632 | expand_omp_taskreg (struct omp_region *region) |
1e8e9920 | 4633 | { |
773c5ba7 | 4634 | basic_block entry_bb, exit_bb, new_bb; |
87d4aa85 | 4635 | struct function *child_cfun; |
414c3a2c | 4636 | tree child_fn, block, t; |
75a70cf9 | 4637 | gimple_stmt_iterator gsi; |
4638 | gimple entry_stmt, stmt; | |
773c5ba7 | 4639 | edge e; |
f1f41a6c | 4640 | vec<tree, va_gc> *ws_args; |
773c5ba7 | 4641 | |
61e47ac8 | 4642 | entry_stmt = last_stmt (region->entry); |
75a70cf9 | 4643 | child_fn = gimple_omp_taskreg_child_fn (entry_stmt); |
773c5ba7 | 4644 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); |
773c5ba7 | 4645 | |
61e47ac8 | 4646 | entry_bb = region->entry; |
4647 | exit_bb = region->exit; | |
773c5ba7 | 4648 | |
773c5ba7 | 4649 | if (is_combined_parallel (region)) |
61e47ac8 | 4650 | ws_args = region->ws_args; |
773c5ba7 | 4651 | else |
414c3a2c | 4652 | ws_args = NULL; |
1e8e9920 | 4653 | |
61e47ac8 | 4654 | if (child_cfun->cfg) |
1e8e9920 | 4655 | { |
773c5ba7 | 4656 | /* Due to inlining, it may happen that we have already outlined |
4657 | the region, in which case all we need to do is make the | |
4658 | sub-graph unreachable and emit the parallel call. */ | |
4659 | edge entry_succ_e, exit_succ_e; | |
75a70cf9 | 4660 | gimple_stmt_iterator gsi; |
773c5ba7 | 4661 | |
4662 | entry_succ_e = single_succ_edge (entry_bb); | |
773c5ba7 | 4663 | |
75a70cf9 | 4664 | gsi = gsi_last_bb (entry_bb); |
4665 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_PARALLEL | |
4666 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK); | |
4667 | gsi_remove (&gsi, true); | |
773c5ba7 | 4668 | |
4669 | new_bb = entry_bb; | |
03ed154b | 4670 | if (exit_bb) |
4671 | { | |
4672 | exit_succ_e = single_succ_edge (exit_bb); | |
4673 | make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU); | |
4674 | } | |
79acaae1 | 4675 | remove_edge_and_dominated_blocks (entry_succ_e); |
1e8e9920 | 4676 | } |
773c5ba7 | 4677 | else |
4678 | { | |
501bdd19 | 4679 | unsigned srcidx, dstidx, num; |
2ab2ce89 | 4680 | |
773c5ba7 | 4681 | /* If the parallel region needs data sent from the parent |
3480139d | 4682 | function, then the very first statement (except possible |
4683 | tree profile counter updates) of the parallel body | |
773c5ba7 | 4684 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since |
4685 | &.OMP_DATA_O is passed as an argument to the child function, | |
4686 | we need to replace it with the argument as seen by the child | |
4687 | function. | |
4688 | ||
4689 | In most cases, this will end up being the identity assignment | |
4690 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had | |
4691 | a function call that has been inlined, the original PARM_DECL | |
4692 | .OMP_DATA_I may have been converted into a different local | |
4693 | variable. In which case, we need to keep the assignment. */ | |
75a70cf9 | 4694 | if (gimple_omp_taskreg_data_arg (entry_stmt)) |
773c5ba7 | 4695 | { |
4696 | basic_block entry_succ_bb = single_succ (entry_bb); | |
75a70cf9 | 4697 | gimple_stmt_iterator gsi; |
4698 | tree arg, narg; | |
4699 | gimple parcopy_stmt = NULL; | |
1e8e9920 | 4700 | |
75a70cf9 | 4701 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) |
3480139d | 4702 | { |
75a70cf9 | 4703 | gimple stmt; |
3480139d | 4704 | |
75a70cf9 | 4705 | gcc_assert (!gsi_end_p (gsi)); |
4706 | stmt = gsi_stmt (gsi); | |
4707 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
cc6b725b | 4708 | continue; |
4709 | ||
75a70cf9 | 4710 | if (gimple_num_ops (stmt) == 2) |
3480139d | 4711 | { |
75a70cf9 | 4712 | tree arg = gimple_assign_rhs1 (stmt); |
4713 | ||
4714 | /* We're ignore the subcode because we're | |
4715 | effectively doing a STRIP_NOPS. */ | |
4716 | ||
4717 | if (TREE_CODE (arg) == ADDR_EXPR | |
4718 | && TREE_OPERAND (arg, 0) | |
4719 | == gimple_omp_taskreg_data_arg (entry_stmt)) | |
4720 | { | |
4721 | parcopy_stmt = stmt; | |
4722 | break; | |
4723 | } | |
3480139d | 4724 | } |
4725 | } | |
79acaae1 | 4726 | |
75a70cf9 | 4727 | gcc_assert (parcopy_stmt != NULL); |
79acaae1 | 4728 | arg = DECL_ARGUMENTS (child_fn); |
4729 | ||
4730 | if (!gimple_in_ssa_p (cfun)) | |
4731 | { | |
75a70cf9 | 4732 | if (gimple_assign_lhs (parcopy_stmt) == arg) |
4733 | gsi_remove (&gsi, true); | |
79acaae1 | 4734 | else |
75a70cf9 | 4735 | { |
4736 | /* ?? Is setting the subcode really necessary ?? */ | |
4737 | gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (arg)); | |
4738 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
4739 | } | |
79acaae1 | 4740 | } |
4741 | else | |
4742 | { | |
4743 | /* If we are in ssa form, we must load the value from the default | |
4744 | definition of the argument. That should not be defined now, | |
4745 | since the argument is not used uninitialized. */ | |
c6dfe037 | 4746 | gcc_assert (ssa_default_def (cfun, arg) == NULL); |
75a70cf9 | 4747 | narg = make_ssa_name (arg, gimple_build_nop ()); |
c6dfe037 | 4748 | set_ssa_default_def (cfun, arg, narg); |
75a70cf9 | 4749 | /* ?? Is setting the subcode really necessary ?? */ |
4750 | gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (narg)); | |
4751 | gimple_assign_set_rhs1 (parcopy_stmt, narg); | |
79acaae1 | 4752 | update_stmt (parcopy_stmt); |
4753 | } | |
773c5ba7 | 4754 | } |
4755 | ||
4756 | /* Declare local variables needed in CHILD_CFUN. */ | |
4757 | block = DECL_INITIAL (child_fn); | |
2ab2ce89 | 4758 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); |
e1a7ccb9 | 4759 | /* The gimplifier could record temporaries in parallel/task block |
4760 | rather than in containing function's local_decls chain, | |
4761 | which would mean cgraph missed finalizing them. Do it now. */ | |
1767a056 | 4762 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) |
e1a7ccb9 | 4763 | if (TREE_CODE (t) == VAR_DECL |
4764 | && TREE_STATIC (t) | |
4765 | && !DECL_EXTERNAL (t)) | |
4766 | varpool_finalize_decl (t); | |
75a70cf9 | 4767 | DECL_SAVED_TREE (child_fn) = NULL; |
e3a19533 | 4768 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ |
4769 | gimple_set_body (child_fn, NULL); | |
1d22f541 | 4770 | TREE_USED (block) = 1; |
773c5ba7 | 4771 | |
79acaae1 | 4772 | /* Reset DECL_CONTEXT on function arguments. */ |
1767a056 | 4773 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) |
773c5ba7 | 4774 | DECL_CONTEXT (t) = child_fn; |
4775 | ||
75a70cf9 | 4776 | /* Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK, |
4777 | so that it can be moved to the child function. */ | |
4778 | gsi = gsi_last_bb (entry_bb); | |
4779 | stmt = gsi_stmt (gsi); | |
4780 | gcc_assert (stmt && (gimple_code (stmt) == GIMPLE_OMP_PARALLEL | |
4781 | || gimple_code (stmt) == GIMPLE_OMP_TASK)); | |
4782 | gsi_remove (&gsi, true); | |
4783 | e = split_block (entry_bb, stmt); | |
773c5ba7 | 4784 | entry_bb = e->dest; |
4785 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
4786 | ||
75a70cf9 | 4787 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ |
5056ba1a | 4788 | if (exit_bb) |
4789 | { | |
75a70cf9 | 4790 | gsi = gsi_last_bb (exit_bb); |
4791 | gcc_assert (!gsi_end_p (gsi) | |
4792 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
4793 | stmt = gimple_build_return (NULL); | |
4794 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
4795 | gsi_remove (&gsi, true); | |
5056ba1a | 4796 | } |
79acaae1 | 4797 | |
4798 | /* Move the parallel region into CHILD_CFUN. */ | |
48e1416a | 4799 | |
79acaae1 | 4800 | if (gimple_in_ssa_p (cfun)) |
4801 | { | |
bcaa2770 | 4802 | init_tree_ssa (child_cfun); |
5084b2e4 | 4803 | init_ssa_operands (child_cfun); |
4804 | child_cfun->gimple_df->in_ssa_p = true; | |
1d22f541 | 4805 | block = NULL_TREE; |
79acaae1 | 4806 | } |
1d22f541 | 4807 | else |
75a70cf9 | 4808 | block = gimple_block (entry_stmt); |
1d22f541 | 4809 | |
4810 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
79acaae1 | 4811 | if (exit_bb) |
4812 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
04c2922b | 4813 | /* When the OMP expansion process cannot guarantee an up-to-date |
4814 | loop tree arrange for the child function to fixup loops. */ | |
4815 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
4816 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
79acaae1 | 4817 | |
1d22f541 | 4818 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ |
f1f41a6c | 4819 | num = vec_safe_length (child_cfun->local_decls); |
501bdd19 | 4820 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) |
4821 | { | |
f1f41a6c | 4822 | t = (*child_cfun->local_decls)[srcidx]; |
501bdd19 | 4823 | if (DECL_CONTEXT (t) == cfun->decl) |
4824 | continue; | |
4825 | if (srcidx != dstidx) | |
f1f41a6c | 4826 | (*child_cfun->local_decls)[dstidx] = t; |
501bdd19 | 4827 | dstidx++; |
4828 | } | |
4829 | if (dstidx != num) | |
f1f41a6c | 4830 | vec_safe_truncate (child_cfun->local_decls, dstidx); |
1d22f541 | 4831 | |
79acaae1 | 4832 | /* Inform the callgraph about the new function. */ |
82b40354 | 4833 | DECL_STRUCT_FUNCTION (child_fn)->curr_properties = cfun->curr_properties; |
79acaae1 | 4834 | cgraph_add_new_function (child_fn, true); |
4835 | ||
4836 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
4837 | fixed in a following pass. */ | |
4838 | push_cfun (child_cfun); | |
658b4427 | 4839 | if (optimize) |
fd6481cf | 4840 | optimize_omp_library_calls (entry_stmt); |
79acaae1 | 4841 | rebuild_cgraph_edges (); |
fbe86b1b | 4842 | |
4843 | /* Some EH regions might become dead, see PR34608. If | |
4844 | pass_cleanup_cfg isn't the first pass to happen with the | |
4845 | new child, these dead EH edges might cause problems. | |
4846 | Clean them up now. */ | |
4847 | if (flag_exceptions) | |
4848 | { | |
4849 | basic_block bb; | |
fbe86b1b | 4850 | bool changed = false; |
4851 | ||
fc00614f | 4852 | FOR_EACH_BB_FN (bb, cfun) |
75a70cf9 | 4853 | changed |= gimple_purge_dead_eh_edges (bb); |
fbe86b1b | 4854 | if (changed) |
4855 | cleanup_tree_cfg (); | |
fbe86b1b | 4856 | } |
dd277d48 | 4857 | if (gimple_in_ssa_p (cfun)) |
4858 | update_ssa (TODO_update_ssa); | |
79acaae1 | 4859 | pop_cfun (); |
773c5ba7 | 4860 | } |
48e1416a | 4861 | |
773c5ba7 | 4862 | /* Emit a library call to launch the children threads. */ |
75a70cf9 | 4863 | if (gimple_code (entry_stmt) == GIMPLE_OMP_PARALLEL) |
fd6481cf | 4864 | expand_parallel_call (region, new_bb, entry_stmt, ws_args); |
4865 | else | |
4866 | expand_task_call (new_bb, entry_stmt); | |
083152fb | 4867 | if (gimple_in_ssa_p (cfun)) |
4868 | update_ssa (TODO_update_ssa_only_virtuals); | |
1e8e9920 | 4869 | } |
4870 | ||
773c5ba7 | 4871 | |
3d483a94 | 4872 | /* Helper function for expand_omp_{for_*,simd}. If this is the outermost |
4873 | of the combined collapse > 1 loop constructs, generate code like: | |
4874 | if (__builtin_expect (N32 cond3 N31, 0)) goto ZERO_ITER_BB; | |
4875 | if (cond3 is <) | |
4876 | adj = STEP3 - 1; | |
4877 | else | |
4878 | adj = STEP3 + 1; | |
4879 | count3 = (adj + N32 - N31) / STEP3; | |
4880 | if (__builtin_expect (N22 cond2 N21, 0)) goto ZERO_ITER_BB; | |
4881 | if (cond2 is <) | |
4882 | adj = STEP2 - 1; | |
4883 | else | |
4884 | adj = STEP2 + 1; | |
4885 | count2 = (adj + N22 - N21) / STEP2; | |
4886 | if (__builtin_expect (N12 cond1 N11, 0)) goto ZERO_ITER_BB; | |
4887 | if (cond1 is <) | |
4888 | adj = STEP1 - 1; | |
4889 | else | |
4890 | adj = STEP1 + 1; | |
4891 | count1 = (adj + N12 - N11) / STEP1; | |
4892 | count = count1 * count2 * count3; | |
4893 | Furthermore, if ZERO_ITER_BB is NULL, create a BB which does: | |
4894 | count = 0; | |
bc7bff74 | 4895 | and set ZERO_ITER_BB to that bb. If this isn't the outermost |
4896 | of the combined loop constructs, just initialize COUNTS array | |
4897 | from the _looptemp_ clauses. */ | |
3d483a94 | 4898 | |
4899 | /* NOTE: It *could* be better to moosh all of the BBs together, | |
4900 | creating one larger BB with all the computation and the unexpected | |
4901 | jump at the end. I.e. | |
4902 | ||
4903 | bool zero3, zero2, zero1, zero; | |
4904 | ||
4905 | zero3 = N32 c3 N31; | |
4906 | count3 = (N32 - N31) /[cl] STEP3; | |
4907 | zero2 = N22 c2 N21; | |
4908 | count2 = (N22 - N21) /[cl] STEP2; | |
4909 | zero1 = N12 c1 N11; | |
4910 | count1 = (N12 - N11) /[cl] STEP1; | |
4911 | zero = zero3 || zero2 || zero1; | |
4912 | count = count1 * count2 * count3; | |
4913 | if (__builtin_expect(zero, false)) goto zero_iter_bb; | |
4914 | ||
4915 | After all, we expect the zero=false, and thus we expect to have to | |
4916 | evaluate all of the comparison expressions, so short-circuiting | |
4917 | oughtn't be a win. Since the condition isn't protecting a | |
4918 | denominator, we're not concerned about divide-by-zero, so we can | |
4919 | fully evaluate count even if a numerator turned out to be wrong. | |
4920 | ||
4921 | It seems like putting this all together would create much better | |
4922 | scheduling opportunities, and less pressure on the chip's branch | |
4923 | predictor. */ | |
4924 | ||
4925 | static void | |
4926 | expand_omp_for_init_counts (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
4927 | basic_block &entry_bb, tree *counts, | |
4928 | basic_block &zero_iter_bb, int &first_zero_iter, | |
4929 | basic_block &l2_dom_bb) | |
4930 | { | |
4931 | tree t, type = TREE_TYPE (fd->loop.v); | |
4932 | gimple stmt; | |
4933 | edge e, ne; | |
4934 | int i; | |
4935 | ||
4936 | /* Collapsed loops need work for expansion into SSA form. */ | |
4937 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
4938 | ||
bc7bff74 | 4939 | if (gimple_omp_for_combined_into_p (fd->for_stmt) |
4940 | && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
4941 | { | |
4942 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
4943 | isn't supposed to be handled, as the inner loop doesn't | |
4944 | use it. */ | |
4945 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
4946 | OMP_CLAUSE__LOOPTEMP_); | |
4947 | gcc_assert (innerc); | |
4948 | for (i = 0; i < fd->collapse; i++) | |
4949 | { | |
4950 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
4951 | OMP_CLAUSE__LOOPTEMP_); | |
4952 | gcc_assert (innerc); | |
4953 | if (i) | |
4954 | counts[i] = OMP_CLAUSE_DECL (innerc); | |
4955 | else | |
4956 | counts[0] = NULL_TREE; | |
4957 | } | |
4958 | return; | |
4959 | } | |
4960 | ||
3d483a94 | 4961 | for (i = 0; i < fd->collapse; i++) |
4962 | { | |
4963 | tree itype = TREE_TYPE (fd->loops[i].v); | |
4964 | ||
4965 | if (SSA_VAR_P (fd->loop.n2) | |
4966 | && ((t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
4967 | fold_convert (itype, fd->loops[i].n1), | |
4968 | fold_convert (itype, fd->loops[i].n2))) | |
4969 | == NULL_TREE || !integer_onep (t))) | |
4970 | { | |
4971 | tree n1, n2; | |
4972 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
4973 | n1 = force_gimple_operand_gsi (gsi, n1, true, NULL_TREE, | |
4974 | true, GSI_SAME_STMT); | |
4975 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
4976 | n2 = force_gimple_operand_gsi (gsi, n2, true, NULL_TREE, | |
4977 | true, GSI_SAME_STMT); | |
4978 | stmt = gimple_build_cond (fd->loops[i].cond_code, n1, n2, | |
4979 | NULL_TREE, NULL_TREE); | |
4980 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
4981 | if (walk_tree (gimple_cond_lhs_ptr (stmt), | |
4982 | expand_omp_regimplify_p, NULL, NULL) | |
4983 | || walk_tree (gimple_cond_rhs_ptr (stmt), | |
4984 | expand_omp_regimplify_p, NULL, NULL)) | |
4985 | { | |
4986 | *gsi = gsi_for_stmt (stmt); | |
4987 | gimple_regimplify_operands (stmt, gsi); | |
4988 | } | |
4989 | e = split_block (entry_bb, stmt); | |
4990 | if (zero_iter_bb == NULL) | |
4991 | { | |
4992 | first_zero_iter = i; | |
4993 | zero_iter_bb = create_empty_bb (entry_bb); | |
4994 | if (current_loops) | |
4995 | add_bb_to_loop (zero_iter_bb, entry_bb->loop_father); | |
4996 | *gsi = gsi_after_labels (zero_iter_bb); | |
4997 | stmt = gimple_build_assign (fd->loop.n2, | |
4998 | build_zero_cst (type)); | |
4999 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
5000 | set_immediate_dominator (CDI_DOMINATORS, zero_iter_bb, | |
5001 | entry_bb); | |
5002 | } | |
5003 | ne = make_edge (entry_bb, zero_iter_bb, EDGE_FALSE_VALUE); | |
5004 | ne->probability = REG_BR_PROB_BASE / 2000 - 1; | |
5005 | e->flags = EDGE_TRUE_VALUE; | |
5006 | e->probability = REG_BR_PROB_BASE - ne->probability; | |
5007 | if (l2_dom_bb == NULL) | |
5008 | l2_dom_bb = entry_bb; | |
5009 | entry_bb = e->dest; | |
5010 | *gsi = gsi_last_bb (entry_bb); | |
5011 | } | |
5012 | ||
5013 | if (POINTER_TYPE_P (itype)) | |
5014 | itype = signed_type_for (itype); | |
5015 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
5016 | ? -1 : 1)); | |
5017 | t = fold_build2 (PLUS_EXPR, itype, | |
5018 | fold_convert (itype, fd->loops[i].step), t); | |
5019 | t = fold_build2 (PLUS_EXPR, itype, t, | |
5020 | fold_convert (itype, fd->loops[i].n2)); | |
5021 | t = fold_build2 (MINUS_EXPR, itype, t, | |
5022 | fold_convert (itype, fd->loops[i].n1)); | |
5023 | /* ?? We could probably use CEIL_DIV_EXPR instead of | |
5024 | TRUNC_DIV_EXPR and adjusting by hand. Unless we can't | |
5025 | generate the same code in the end because generically we | |
5026 | don't know that the values involved must be negative for | |
5027 | GT?? */ | |
5028 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
5029 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
5030 | fold_build1 (NEGATE_EXPR, itype, t), | |
5031 | fold_build1 (NEGATE_EXPR, itype, | |
5032 | fold_convert (itype, | |
5033 | fd->loops[i].step))); | |
5034 | else | |
5035 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
5036 | fold_convert (itype, fd->loops[i].step)); | |
5037 | t = fold_convert (type, t); | |
5038 | if (TREE_CODE (t) == INTEGER_CST) | |
5039 | counts[i] = t; | |
5040 | else | |
5041 | { | |
5042 | counts[i] = create_tmp_reg (type, ".count"); | |
5043 | expand_omp_build_assign (gsi, counts[i], t); | |
5044 | } | |
5045 | if (SSA_VAR_P (fd->loop.n2)) | |
5046 | { | |
5047 | if (i == 0) | |
5048 | t = counts[0]; | |
5049 | else | |
5050 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, counts[i]); | |
5051 | expand_omp_build_assign (gsi, fd->loop.n2, t); | |
5052 | } | |
5053 | } | |
5054 | } | |
5055 | ||
5056 | ||
5057 | /* Helper function for expand_omp_{for_*,simd}. Generate code like: | |
5058 | T = V; | |
5059 | V3 = N31 + (T % count3) * STEP3; | |
5060 | T = T / count3; | |
5061 | V2 = N21 + (T % count2) * STEP2; | |
5062 | T = T / count2; | |
5063 | V1 = N11 + T * STEP1; | |
bc7bff74 | 5064 | if this loop doesn't have an inner loop construct combined with it. |
5065 | If it does have an inner loop construct combined with it and the | |
5066 | iteration count isn't known constant, store values from counts array | |
5067 | into its _looptemp_ temporaries instead. */ | |
3d483a94 | 5068 | |
5069 | static void | |
5070 | expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
bc7bff74 | 5071 | tree *counts, gimple inner_stmt, tree startvar) |
3d483a94 | 5072 | { |
5073 | int i; | |
bc7bff74 | 5074 | if (gimple_omp_for_combined_p (fd->for_stmt)) |
5075 | { | |
5076 | /* If fd->loop.n2 is constant, then no propagation of the counts | |
5077 | is needed, they are constant. */ | |
5078 | if (TREE_CODE (fd->loop.n2) == INTEGER_CST) | |
5079 | return; | |
5080 | ||
5081 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
5082 | ? gimple_omp_parallel_clauses (inner_stmt) | |
5083 | : gimple_omp_for_clauses (inner_stmt); | |
5084 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
5085 | isn't supposed to be handled, as the inner loop doesn't | |
5086 | use it. */ | |
5087 | tree innerc = find_omp_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
5088 | gcc_assert (innerc); | |
5089 | for (i = 0; i < fd->collapse; i++) | |
5090 | { | |
5091 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5092 | OMP_CLAUSE__LOOPTEMP_); | |
5093 | gcc_assert (innerc); | |
5094 | if (i) | |
5095 | { | |
5096 | tree tem = OMP_CLAUSE_DECL (innerc); | |
5097 | tree t = fold_convert (TREE_TYPE (tem), counts[i]); | |
5098 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
5099 | false, GSI_CONTINUE_LINKING); | |
5100 | gimple stmt = gimple_build_assign (tem, t); | |
5101 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
5102 | } | |
5103 | } | |
5104 | return; | |
5105 | } | |
5106 | ||
3d483a94 | 5107 | tree type = TREE_TYPE (fd->loop.v); |
5108 | tree tem = create_tmp_reg (type, ".tem"); | |
5109 | gimple stmt = gimple_build_assign (tem, startvar); | |
5110 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
5111 | ||
5112 | for (i = fd->collapse - 1; i >= 0; i--) | |
5113 | { | |
5114 | tree vtype = TREE_TYPE (fd->loops[i].v), itype, t; | |
5115 | itype = vtype; | |
5116 | if (POINTER_TYPE_P (vtype)) | |
5117 | itype = signed_type_for (vtype); | |
5118 | if (i != 0) | |
5119 | t = fold_build2 (TRUNC_MOD_EXPR, type, tem, counts[i]); | |
5120 | else | |
5121 | t = tem; | |
5122 | t = fold_convert (itype, t); | |
5123 | t = fold_build2 (MULT_EXPR, itype, t, | |
5124 | fold_convert (itype, fd->loops[i].step)); | |
5125 | if (POINTER_TYPE_P (vtype)) | |
5126 | t = fold_build_pointer_plus (fd->loops[i].n1, t); | |
5127 | else | |
5128 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); | |
5129 | t = force_gimple_operand_gsi (gsi, t, | |
5130 | DECL_P (fd->loops[i].v) | |
5131 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
5132 | NULL_TREE, false, | |
5133 | GSI_CONTINUE_LINKING); | |
5134 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
5135 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
5136 | if (i != 0) | |
5137 | { | |
5138 | t = fold_build2 (TRUNC_DIV_EXPR, type, tem, counts[i]); | |
5139 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
5140 | false, GSI_CONTINUE_LINKING); | |
5141 | stmt = gimple_build_assign (tem, t); | |
5142 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
5143 | } | |
5144 | } | |
5145 | } | |
5146 | ||
5147 | ||
5148 | /* Helper function for expand_omp_for_*. Generate code like: | |
5149 | L10: | |
5150 | V3 += STEP3; | |
5151 | if (V3 cond3 N32) goto BODY_BB; else goto L11; | |
5152 | L11: | |
5153 | V3 = N31; | |
5154 | V2 += STEP2; | |
5155 | if (V2 cond2 N22) goto BODY_BB; else goto L12; | |
5156 | L12: | |
5157 | V2 = N21; | |
5158 | V1 += STEP1; | |
5159 | goto BODY_BB; */ | |
5160 | ||
5161 | static basic_block | |
5162 | extract_omp_for_update_vars (struct omp_for_data *fd, basic_block cont_bb, | |
5163 | basic_block body_bb) | |
5164 | { | |
5165 | basic_block last_bb, bb, collapse_bb = NULL; | |
5166 | int i; | |
5167 | gimple_stmt_iterator gsi; | |
5168 | edge e; | |
5169 | tree t; | |
5170 | gimple stmt; | |
5171 | ||
5172 | last_bb = cont_bb; | |
5173 | for (i = fd->collapse - 1; i >= 0; i--) | |
5174 | { | |
5175 | tree vtype = TREE_TYPE (fd->loops[i].v); | |
5176 | ||
5177 | bb = create_empty_bb (last_bb); | |
5178 | if (current_loops) | |
5179 | add_bb_to_loop (bb, last_bb->loop_father); | |
5180 | gsi = gsi_start_bb (bb); | |
5181 | ||
5182 | if (i < fd->collapse - 1) | |
5183 | { | |
5184 | e = make_edge (last_bb, bb, EDGE_FALSE_VALUE); | |
5185 | e->probability = REG_BR_PROB_BASE / 8; | |
5186 | ||
5187 | t = fd->loops[i + 1].n1; | |
5188 | t = force_gimple_operand_gsi (&gsi, t, | |
5189 | DECL_P (fd->loops[i + 1].v) | |
5190 | && TREE_ADDRESSABLE (fd->loops[i | |
5191 | + 1].v), | |
5192 | NULL_TREE, false, | |
5193 | GSI_CONTINUE_LINKING); | |
5194 | stmt = gimple_build_assign (fd->loops[i + 1].v, t); | |
5195 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
5196 | } | |
5197 | else | |
5198 | collapse_bb = bb; | |
5199 | ||
5200 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); | |
5201 | ||
5202 | if (POINTER_TYPE_P (vtype)) | |
5203 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); | |
5204 | else | |
5205 | t = fold_build2 (PLUS_EXPR, vtype, fd->loops[i].v, fd->loops[i].step); | |
5206 | t = force_gimple_operand_gsi (&gsi, t, | |
5207 | DECL_P (fd->loops[i].v) | |
5208 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
5209 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5210 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
5211 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
5212 | ||
5213 | if (i > 0) | |
5214 | { | |
5215 | t = fd->loops[i].n2; | |
5216 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5217 | false, GSI_CONTINUE_LINKING); | |
5218 | tree v = fd->loops[i].v; | |
5219 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
5220 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
5221 | false, GSI_CONTINUE_LINKING); | |
5222 | t = fold_build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
5223 | stmt = gimple_build_cond_empty (t); | |
5224 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
5225 | e = make_edge (bb, body_bb, EDGE_TRUE_VALUE); | |
5226 | e->probability = REG_BR_PROB_BASE * 7 / 8; | |
5227 | } | |
5228 | else | |
5229 | make_edge (bb, body_bb, EDGE_FALLTHRU); | |
5230 | last_bb = bb; | |
5231 | } | |
5232 | ||
5233 | return collapse_bb; | |
5234 | } | |
5235 | ||
5236 | ||
773c5ba7 | 5237 | /* A subroutine of expand_omp_for. Generate code for a parallel |
1e8e9920 | 5238 | loop with any schedule. Given parameters: |
5239 | ||
5240 | for (V = N1; V cond N2; V += STEP) BODY; | |
5241 | ||
5242 | where COND is "<" or ">", we generate pseudocode | |
5243 | ||
5244 | more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0); | |
773c5ba7 | 5245 | if (more) goto L0; else goto L3; |
1e8e9920 | 5246 | L0: |
5247 | V = istart0; | |
5248 | iend = iend0; | |
5249 | L1: | |
5250 | BODY; | |
5251 | V += STEP; | |
773c5ba7 | 5252 | if (V cond iend) goto L1; else goto L2; |
1e8e9920 | 5253 | L2: |
773c5ba7 | 5254 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; |
5255 | L3: | |
1e8e9920 | 5256 | |
773c5ba7 | 5257 | If this is a combined omp parallel loop, instead of the call to |
fd6481cf | 5258 | GOMP_loop_foo_start, we call GOMP_loop_foo_next. |
bc7bff74 | 5259 | If this is gimple_omp_for_combined_p loop, then instead of assigning |
5260 | V and iend in L0 we assign the first two _looptemp_ clause decls of the | |
5261 | inner GIMPLE_OMP_FOR and V += STEP; and | |
5262 | if (V cond iend) goto L1; else goto L2; are removed. | |
fd6481cf | 5263 | |
5264 | For collapsed loops, given parameters: | |
5265 | collapse(3) | |
5266 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
5267 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
5268 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
5269 | BODY; | |
5270 | ||
5271 | we generate pseudocode | |
5272 | ||
8e6b4515 | 5273 | if (__builtin_expect (N32 cond3 N31, 0)) goto Z0; |
fd6481cf | 5274 | if (cond3 is <) |
5275 | adj = STEP3 - 1; | |
5276 | else | |
5277 | adj = STEP3 + 1; | |
5278 | count3 = (adj + N32 - N31) / STEP3; | |
8e6b4515 | 5279 | if (__builtin_expect (N22 cond2 N21, 0)) goto Z0; |
fd6481cf | 5280 | if (cond2 is <) |
5281 | adj = STEP2 - 1; | |
5282 | else | |
5283 | adj = STEP2 + 1; | |
5284 | count2 = (adj + N22 - N21) / STEP2; | |
8e6b4515 | 5285 | if (__builtin_expect (N12 cond1 N11, 0)) goto Z0; |
fd6481cf | 5286 | if (cond1 is <) |
5287 | adj = STEP1 - 1; | |
5288 | else | |
5289 | adj = STEP1 + 1; | |
5290 | count1 = (adj + N12 - N11) / STEP1; | |
5291 | count = count1 * count2 * count3; | |
8e6b4515 | 5292 | goto Z1; |
5293 | Z0: | |
5294 | count = 0; | |
5295 | Z1: | |
fd6481cf | 5296 | more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0); |
5297 | if (more) goto L0; else goto L3; | |
5298 | L0: | |
5299 | V = istart0; | |
5300 | T = V; | |
5301 | V3 = N31 + (T % count3) * STEP3; | |
5302 | T = T / count3; | |
5303 | V2 = N21 + (T % count2) * STEP2; | |
5304 | T = T / count2; | |
5305 | V1 = N11 + T * STEP1; | |
5306 | iend = iend0; | |
5307 | L1: | |
5308 | BODY; | |
5309 | V += 1; | |
5310 | if (V < iend) goto L10; else goto L2; | |
5311 | L10: | |
5312 | V3 += STEP3; | |
5313 | if (V3 cond3 N32) goto L1; else goto L11; | |
5314 | L11: | |
5315 | V3 = N31; | |
5316 | V2 += STEP2; | |
5317 | if (V2 cond2 N22) goto L1; else goto L12; | |
5318 | L12: | |
5319 | V2 = N21; | |
5320 | V1 += STEP1; | |
5321 | goto L1; | |
5322 | L2: | |
5323 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
5324 | L3: | |
5325 | ||
5326 | */ | |
1e8e9920 | 5327 | |
61e47ac8 | 5328 | static void |
773c5ba7 | 5329 | expand_omp_for_generic (struct omp_region *region, |
5330 | struct omp_for_data *fd, | |
1e8e9920 | 5331 | enum built_in_function start_fn, |
bc7bff74 | 5332 | enum built_in_function next_fn, |
5333 | gimple inner_stmt) | |
1e8e9920 | 5334 | { |
75a70cf9 | 5335 | tree type, istart0, iend0, iend; |
fd6481cf | 5336 | tree t, vmain, vback, bias = NULL_TREE; |
5337 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, collapse_bb; | |
03ed154b | 5338 | basic_block l2_bb = NULL, l3_bb = NULL; |
75a70cf9 | 5339 | gimple_stmt_iterator gsi; |
5340 | gimple stmt; | |
773c5ba7 | 5341 | bool in_combined_parallel = is_combined_parallel (region); |
ac6e3339 | 5342 | bool broken_loop = region->cont == NULL; |
79acaae1 | 5343 | edge e, ne; |
fd6481cf | 5344 | tree *counts = NULL; |
5345 | int i; | |
ac6e3339 | 5346 | |
5347 | gcc_assert (!broken_loop || !in_combined_parallel); | |
fd6481cf | 5348 | gcc_assert (fd->iter_type == long_integer_type_node |
5349 | || !in_combined_parallel); | |
1e8e9920 | 5350 | |
fd6481cf | 5351 | type = TREE_TYPE (fd->loop.v); |
5352 | istart0 = create_tmp_var (fd->iter_type, ".istart0"); | |
5353 | iend0 = create_tmp_var (fd->iter_type, ".iend0"); | |
6d63fc03 | 5354 | TREE_ADDRESSABLE (istart0) = 1; |
5355 | TREE_ADDRESSABLE (iend0) = 1; | |
1e8e9920 | 5356 | |
fd6481cf | 5357 | /* See if we need to bias by LLONG_MIN. */ |
5358 | if (fd->iter_type == long_long_unsigned_type_node | |
5359 | && TREE_CODE (type) == INTEGER_TYPE | |
5360 | && !TYPE_UNSIGNED (type)) | |
5361 | { | |
5362 | tree n1, n2; | |
5363 | ||
5364 | if (fd->loop.cond_code == LT_EXPR) | |
5365 | { | |
5366 | n1 = fd->loop.n1; | |
5367 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5368 | } | |
5369 | else | |
5370 | { | |
5371 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5372 | n2 = fd->loop.n1; | |
5373 | } | |
5374 | if (TREE_CODE (n1) != INTEGER_CST | |
5375 | || TREE_CODE (n2) != INTEGER_CST | |
5376 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
5377 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
5378 | } | |
5379 | ||
61e47ac8 | 5380 | entry_bb = region->entry; |
03ed154b | 5381 | cont_bb = region->cont; |
fd6481cf | 5382 | collapse_bb = NULL; |
ac6e3339 | 5383 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
5384 | gcc_assert (broken_loop | |
5385 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
5386 | l0_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
5387 | l1_bb = single_succ (l0_bb); | |
5388 | if (!broken_loop) | |
03ed154b | 5389 | { |
5390 | l2_bb = create_empty_bb (cont_bb); | |
ac6e3339 | 5391 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb); |
5392 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
03ed154b | 5393 | } |
ac6e3339 | 5394 | else |
5395 | l2_bb = NULL; | |
5396 | l3_bb = BRANCH_EDGE (entry_bb)->dest; | |
5397 | exit_bb = region->exit; | |
773c5ba7 | 5398 | |
75a70cf9 | 5399 | gsi = gsi_last_bb (entry_bb); |
fd6481cf | 5400 | |
75a70cf9 | 5401 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
fd6481cf | 5402 | if (fd->collapse > 1) |
5403 | { | |
8e6b4515 | 5404 | int first_zero_iter = -1; |
3d483a94 | 5405 | basic_block zero_iter_bb = NULL, l2_dom_bb = NULL; |
8e6b4515 | 5406 | |
3d483a94 | 5407 | counts = XALLOCAVEC (tree, fd->collapse); |
5408 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5409 | zero_iter_bb, first_zero_iter, | |
5410 | l2_dom_bb); | |
fd6481cf | 5411 | |
8e6b4515 | 5412 | if (zero_iter_bb) |
5413 | { | |
5414 | /* Some counts[i] vars might be uninitialized if | |
5415 | some loop has zero iterations. But the body shouldn't | |
5416 | be executed in that case, so just avoid uninit warnings. */ | |
5417 | for (i = first_zero_iter; i < fd->collapse; i++) | |
5418 | if (SSA_VAR_P (counts[i])) | |
5419 | TREE_NO_WARNING (counts[i]) = 1; | |
5420 | gsi_prev (&gsi); | |
5421 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
5422 | entry_bb = e->dest; | |
5423 | make_edge (zero_iter_bb, entry_bb, EDGE_FALLTHRU); | |
5424 | gsi = gsi_last_bb (entry_bb); | |
5425 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, | |
5426 | get_immediate_dominator (CDI_DOMINATORS, | |
5427 | zero_iter_bb)); | |
5428 | } | |
fd6481cf | 5429 | } |
79acaae1 | 5430 | if (in_combined_parallel) |
5431 | { | |
5432 | /* In a combined parallel loop, emit a call to | |
5433 | GOMP_loop_foo_next. */ | |
b9a16870 | 5434 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, |
79acaae1 | 5435 | build_fold_addr_expr (istart0), |
5436 | build_fold_addr_expr (iend0)); | |
5437 | } | |
5438 | else | |
1e8e9920 | 5439 | { |
c2f47e15 | 5440 | tree t0, t1, t2, t3, t4; |
773c5ba7 | 5441 | /* If this is not a combined parallel loop, emit a call to |
5442 | GOMP_loop_foo_start in ENTRY_BB. */ | |
c2f47e15 | 5443 | t4 = build_fold_addr_expr (iend0); |
5444 | t3 = build_fold_addr_expr (istart0); | |
fd6481cf | 5445 | t2 = fold_convert (fd->iter_type, fd->loop.step); |
3d483a94 | 5446 | t1 = fd->loop.n2; |
5447 | t0 = fd->loop.n1; | |
bc7bff74 | 5448 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) |
5449 | { | |
5450 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5451 | OMP_CLAUSE__LOOPTEMP_); | |
5452 | gcc_assert (innerc); | |
5453 | t0 = OMP_CLAUSE_DECL (innerc); | |
5454 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5455 | OMP_CLAUSE__LOOPTEMP_); | |
5456 | gcc_assert (innerc); | |
5457 | t1 = OMP_CLAUSE_DECL (innerc); | |
5458 | } | |
3d483a94 | 5459 | if (POINTER_TYPE_P (TREE_TYPE (t0)) |
5460 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
5461 | != TYPE_PRECISION (fd->iter_type)) | |
c799f233 | 5462 | { |
5463 | /* Avoid casting pointers to integer of a different size. */ | |
3cea8318 | 5464 | tree itype = signed_type_for (type); |
3d483a94 | 5465 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); |
5466 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
c799f233 | 5467 | } |
5468 | else | |
5469 | { | |
3d483a94 | 5470 | t1 = fold_convert (fd->iter_type, t1); |
5471 | t0 = fold_convert (fd->iter_type, t0); | |
c799f233 | 5472 | } |
fd6481cf | 5473 | if (bias) |
1e8e9920 | 5474 | { |
fd6481cf | 5475 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); |
5476 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
5477 | } | |
5478 | if (fd->iter_type == long_integer_type_node) | |
5479 | { | |
5480 | if (fd->chunk_size) | |
5481 | { | |
5482 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
b9a16870 | 5483 | t = build_call_expr (builtin_decl_explicit (start_fn), |
5484 | 6, t0, t1, t2, t, t3, t4); | |
fd6481cf | 5485 | } |
5486 | else | |
b9a16870 | 5487 | t = build_call_expr (builtin_decl_explicit (start_fn), |
5488 | 5, t0, t1, t2, t3, t4); | |
1e8e9920 | 5489 | } |
c2f47e15 | 5490 | else |
fd6481cf | 5491 | { |
5492 | tree t5; | |
5493 | tree c_bool_type; | |
b9a16870 | 5494 | tree bfn_decl; |
fd6481cf | 5495 | |
5496 | /* The GOMP_loop_ull_*start functions have additional boolean | |
5497 | argument, true for < loops and false for > loops. | |
5498 | In Fortran, the C bool type can be different from | |
5499 | boolean_type_node. */ | |
b9a16870 | 5500 | bfn_decl = builtin_decl_explicit (start_fn); |
5501 | c_bool_type = TREE_TYPE (TREE_TYPE (bfn_decl)); | |
fd6481cf | 5502 | t5 = build_int_cst (c_bool_type, |
5503 | fd->loop.cond_code == LT_EXPR ? 1 : 0); | |
5504 | if (fd->chunk_size) | |
5505 | { | |
b9a16870 | 5506 | tree bfn_decl = builtin_decl_explicit (start_fn); |
fd6481cf | 5507 | t = fold_convert (fd->iter_type, fd->chunk_size); |
b9a16870 | 5508 | t = build_call_expr (bfn_decl, 7, t5, t0, t1, t2, t, t3, t4); |
fd6481cf | 5509 | } |
5510 | else | |
b9a16870 | 5511 | t = build_call_expr (builtin_decl_explicit (start_fn), |
5512 | 6, t5, t0, t1, t2, t3, t4); | |
fd6481cf | 5513 | } |
1e8e9920 | 5514 | } |
fd6481cf | 5515 | if (TREE_TYPE (t) != boolean_type_node) |
5516 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
5517 | t, build_int_cst (TREE_TYPE (t), 0)); | |
75a70cf9 | 5518 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5519 | true, GSI_SAME_STMT); | |
5520 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
79acaae1 | 5521 | |
75a70cf9 | 5522 | /* Remove the GIMPLE_OMP_FOR statement. */ |
5523 | gsi_remove (&gsi, true); | |
1e8e9920 | 5524 | |
773c5ba7 | 5525 | /* Iteration setup for sequential loop goes in L0_BB. */ |
3d483a94 | 5526 | tree startvar = fd->loop.v; |
5527 | tree endvar = NULL_TREE; | |
5528 | ||
bc7bff74 | 5529 | if (gimple_omp_for_combined_p (fd->for_stmt)) |
5530 | { | |
5531 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_FOR | |
5532 | && gimple_omp_for_kind (inner_stmt) | |
5533 | == GF_OMP_FOR_KIND_SIMD); | |
5534 | tree innerc = find_omp_clause (gimple_omp_for_clauses (inner_stmt), | |
5535 | OMP_CLAUSE__LOOPTEMP_); | |
5536 | gcc_assert (innerc); | |
5537 | startvar = OMP_CLAUSE_DECL (innerc); | |
5538 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5539 | OMP_CLAUSE__LOOPTEMP_); | |
5540 | gcc_assert (innerc); | |
5541 | endvar = OMP_CLAUSE_DECL (innerc); | |
5542 | } | |
5543 | ||
75a70cf9 | 5544 | gsi = gsi_start_bb (l0_bb); |
1efcacec | 5545 | t = istart0; |
fd6481cf | 5546 | if (bias) |
1efcacec | 5547 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); |
3d483a94 | 5548 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
5549 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
5550 | t = fold_convert (TREE_TYPE (startvar), t); | |
4abecb72 | 5551 | t = force_gimple_operand_gsi (&gsi, t, |
3d483a94 | 5552 | DECL_P (startvar) |
5553 | && TREE_ADDRESSABLE (startvar), | |
4abecb72 | 5554 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
3d483a94 | 5555 | stmt = gimple_build_assign (startvar, t); |
75a70cf9 | 5556 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
1e8e9920 | 5557 | |
1efcacec | 5558 | t = iend0; |
fd6481cf | 5559 | if (bias) |
1efcacec | 5560 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); |
3d483a94 | 5561 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
5562 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
5563 | t = fold_convert (TREE_TYPE (startvar), t); | |
75a70cf9 | 5564 | iend = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5565 | false, GSI_CONTINUE_LINKING); | |
3d483a94 | 5566 | if (endvar) |
fd6481cf | 5567 | { |
3d483a94 | 5568 | stmt = gimple_build_assign (endvar, iend); |
75a70cf9 | 5569 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
fd6481cf | 5570 | } |
3d483a94 | 5571 | if (fd->collapse > 1) |
bc7bff74 | 5572 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); |
773c5ba7 | 5573 | |
ac6e3339 | 5574 | if (!broken_loop) |
03ed154b | 5575 | { |
ac6e3339 | 5576 | /* Code to control the increment and predicate for the sequential |
5577 | loop goes in the CONT_BB. */ | |
75a70cf9 | 5578 | gsi = gsi_last_bb (cont_bb); |
5579 | stmt = gsi_stmt (gsi); | |
5580 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
5581 | vmain = gimple_omp_continue_control_use (stmt); | |
5582 | vback = gimple_omp_continue_control_def (stmt); | |
79acaae1 | 5583 | |
bc7bff74 | 5584 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
3d483a94 | 5585 | { |
5586 | if (POINTER_TYPE_P (type)) | |
5587 | t = fold_build_pointer_plus (vmain, fd->loop.step); | |
5588 | else | |
5589 | t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step); | |
5590 | t = force_gimple_operand_gsi (&gsi, t, | |
5591 | DECL_P (vback) | |
5592 | && TREE_ADDRESSABLE (vback), | |
5593 | NULL_TREE, true, GSI_SAME_STMT); | |
5594 | stmt = gimple_build_assign (vback, t); | |
5595 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
5596 | ||
5597 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
5598 | DECL_P (vback) && TREE_ADDRESSABLE (vback) ? t : vback, | |
5599 | iend); | |
5600 | stmt = gimple_build_cond_empty (t); | |
5601 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
5602 | } | |
773c5ba7 | 5603 | |
75a70cf9 | 5604 | /* Remove GIMPLE_OMP_CONTINUE. */ |
5605 | gsi_remove (&gsi, true); | |
773c5ba7 | 5606 | |
bc7bff74 | 5607 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) |
3d483a94 | 5608 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, l1_bb); |
fd6481cf | 5609 | |
ac6e3339 | 5610 | /* Emit code to get the next parallel iteration in L2_BB. */ |
75a70cf9 | 5611 | gsi = gsi_start_bb (l2_bb); |
773c5ba7 | 5612 | |
b9a16870 | 5613 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, |
ac6e3339 | 5614 | build_fold_addr_expr (istart0), |
5615 | build_fold_addr_expr (iend0)); | |
75a70cf9 | 5616 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5617 | false, GSI_CONTINUE_LINKING); | |
fd6481cf | 5618 | if (TREE_TYPE (t) != boolean_type_node) |
5619 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
5620 | t, build_int_cst (TREE_TYPE (t), 0)); | |
75a70cf9 | 5621 | stmt = gimple_build_cond_empty (t); |
5622 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
ac6e3339 | 5623 | } |
1e8e9920 | 5624 | |
61e47ac8 | 5625 | /* Add the loop cleanup function. */ |
75a70cf9 | 5626 | gsi = gsi_last_bb (exit_bb); |
5627 | if (gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
b9a16870 | 5628 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); |
bc7bff74 | 5629 | else if (gimple_omp_return_lhs (gsi_stmt (gsi))) |
5630 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
61e47ac8 | 5631 | else |
b9a16870 | 5632 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); |
75a70cf9 | 5633 | stmt = gimple_build_call (t, 0); |
bc7bff74 | 5634 | if (gimple_omp_return_lhs (gsi_stmt (gsi))) |
5635 | gimple_call_set_lhs (stmt, gimple_omp_return_lhs (gsi_stmt (gsi))); | |
75a70cf9 | 5636 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); |
5637 | gsi_remove (&gsi, true); | |
773c5ba7 | 5638 | |
5639 | /* Connect the new blocks. */ | |
79acaae1 | 5640 | find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE; |
5641 | find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE; | |
1e8e9920 | 5642 | |
ac6e3339 | 5643 | if (!broken_loop) |
5644 | { | |
75a70cf9 | 5645 | gimple_seq phis; |
5646 | ||
79acaae1 | 5647 | e = find_edge (cont_bb, l3_bb); |
5648 | ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE); | |
5649 | ||
75a70cf9 | 5650 | phis = phi_nodes (l3_bb); |
5651 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) | |
5652 | { | |
5653 | gimple phi = gsi_stmt (gsi); | |
5654 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne), | |
5655 | PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
5656 | } | |
79acaae1 | 5657 | remove_edge (e); |
5658 | ||
ac6e3339 | 5659 | make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE); |
f6568ea4 | 5660 | if (current_loops) |
5661 | add_bb_to_loop (l2_bb, cont_bb->loop_father); | |
3d483a94 | 5662 | e = find_edge (cont_bb, l1_bb); |
bc7bff74 | 5663 | if (gimple_omp_for_combined_p (fd->for_stmt)) |
5664 | { | |
5665 | remove_edge (e); | |
5666 | e = NULL; | |
5667 | } | |
3d483a94 | 5668 | else if (fd->collapse > 1) |
fd6481cf | 5669 | { |
fd6481cf | 5670 | remove_edge (e); |
5671 | e = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
5672 | } | |
5673 | else | |
3d483a94 | 5674 | e->flags = EDGE_TRUE_VALUE; |
5675 | if (e) | |
fd6481cf | 5676 | { |
3d483a94 | 5677 | e->probability = REG_BR_PROB_BASE * 7 / 8; |
5678 | find_edge (cont_bb, l2_bb)->probability = REG_BR_PROB_BASE / 8; | |
5679 | } | |
5680 | else | |
5681 | { | |
5682 | e = find_edge (cont_bb, l2_bb); | |
5683 | e->flags = EDGE_FALLTHRU; | |
fd6481cf | 5684 | } |
ac6e3339 | 5685 | make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE); |
79acaae1 | 5686 | |
5687 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, | |
5688 | recompute_dominator (CDI_DOMINATORS, l2_bb)); | |
5689 | set_immediate_dominator (CDI_DOMINATORS, l3_bb, | |
5690 | recompute_dominator (CDI_DOMINATORS, l3_bb)); | |
5691 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, | |
5692 | recompute_dominator (CDI_DOMINATORS, l0_bb)); | |
5693 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, | |
5694 | recompute_dominator (CDI_DOMINATORS, l1_bb)); | |
04c2922b | 5695 | |
5696 | struct loop *outer_loop = alloc_loop (); | |
5697 | outer_loop->header = l0_bb; | |
5698 | outer_loop->latch = l2_bb; | |
5699 | add_loop (outer_loop, l0_bb->loop_father); | |
5700 | ||
bc7bff74 | 5701 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
3d483a94 | 5702 | { |
5703 | struct loop *loop = alloc_loop (); | |
5704 | loop->header = l1_bb; | |
5705 | /* The loop may have multiple latches. */ | |
5706 | add_loop (loop, outer_loop); | |
5707 | } | |
ac6e3339 | 5708 | } |
1e8e9920 | 5709 | } |
5710 | ||
5711 | ||
773c5ba7 | 5712 | /* A subroutine of expand_omp_for. Generate code for a parallel |
5713 | loop with static schedule and no specified chunk size. Given | |
5714 | parameters: | |
1e8e9920 | 5715 | |
5716 | for (V = N1; V cond N2; V += STEP) BODY; | |
5717 | ||
5718 | where COND is "<" or ">", we generate pseudocode | |
5719 | ||
8e6b4515 | 5720 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; |
1e8e9920 | 5721 | if (cond is <) |
5722 | adj = STEP - 1; | |
5723 | else | |
5724 | adj = STEP + 1; | |
fd6481cf | 5725 | if ((__typeof (V)) -1 > 0 && cond is >) |
5726 | n = -(adj + N2 - N1) / -STEP; | |
5727 | else | |
5728 | n = (adj + N2 - N1) / STEP; | |
1e8e9920 | 5729 | q = n / nthreads; |
31712e83 | 5730 | tt = n % nthreads; |
5731 | if (threadid < tt) goto L3; else goto L4; | |
5732 | L3: | |
5733 | tt = 0; | |
5734 | q = q + 1; | |
5735 | L4: | |
5736 | s0 = q * threadid + tt; | |
5737 | e0 = s0 + q; | |
79acaae1 | 5738 | V = s0 * STEP + N1; |
1e8e9920 | 5739 | if (s0 >= e0) goto L2; else goto L0; |
5740 | L0: | |
1e8e9920 | 5741 | e = e0 * STEP + N1; |
5742 | L1: | |
5743 | BODY; | |
5744 | V += STEP; | |
5745 | if (V cond e) goto L1; | |
1e8e9920 | 5746 | L2: |
5747 | */ | |
5748 | ||
61e47ac8 | 5749 | static void |
773c5ba7 | 5750 | expand_omp_for_static_nochunk (struct omp_region *region, |
bc7bff74 | 5751 | struct omp_for_data *fd, |
5752 | gimple inner_stmt) | |
1e8e9920 | 5753 | { |
31712e83 | 5754 | tree n, q, s0, e0, e, t, tt, nthreads, threadid; |
fd6481cf | 5755 | tree type, itype, vmain, vback; |
31712e83 | 5756 | basic_block entry_bb, second_bb, third_bb, exit_bb, seq_start_bb; |
bc7bff74 | 5757 | basic_block body_bb, cont_bb, collapse_bb = NULL; |
61e47ac8 | 5758 | basic_block fin_bb; |
75a70cf9 | 5759 | gimple_stmt_iterator gsi; |
5760 | gimple stmt; | |
31712e83 | 5761 | edge ep; |
bc7bff74 | 5762 | enum built_in_function get_num_threads = BUILT_IN_OMP_GET_NUM_THREADS; |
5763 | enum built_in_function get_thread_num = BUILT_IN_OMP_GET_THREAD_NUM; | |
5764 | bool broken_loop = region->cont == NULL; | |
5765 | tree *counts = NULL; | |
5766 | tree n1, n2, step; | |
1e8e9920 | 5767 | |
fd6481cf | 5768 | itype = type = TREE_TYPE (fd->loop.v); |
5769 | if (POINTER_TYPE_P (type)) | |
3cea8318 | 5770 | itype = signed_type_for (type); |
1e8e9920 | 5771 | |
61e47ac8 | 5772 | entry_bb = region->entry; |
61e47ac8 | 5773 | cont_bb = region->cont; |
ac6e3339 | 5774 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
bc7bff74 | 5775 | fin_bb = BRANCH_EDGE (entry_bb)->dest; |
5776 | gcc_assert (broken_loop | |
5777 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
ac6e3339 | 5778 | seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb)); |
5779 | body_bb = single_succ (seq_start_bb); | |
bc7bff74 | 5780 | if (!broken_loop) |
5781 | { | |
5782 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); | |
5783 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
5784 | } | |
61e47ac8 | 5785 | exit_bb = region->exit; |
5786 | ||
773c5ba7 | 5787 | /* Iteration space partitioning goes in ENTRY_BB. */ |
75a70cf9 | 5788 | gsi = gsi_last_bb (entry_bb); |
5789 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
61e47ac8 | 5790 | |
bc7bff74 | 5791 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
5792 | { | |
5793 | get_num_threads = BUILT_IN_OMP_GET_NUM_TEAMS; | |
5794 | get_thread_num = BUILT_IN_OMP_GET_TEAM_NUM; | |
5795 | } | |
5796 | ||
5797 | if (fd->collapse > 1) | |
5798 | { | |
5799 | int first_zero_iter = -1; | |
5800 | basic_block l2_dom_bb = NULL; | |
5801 | ||
5802 | counts = XALLOCAVEC (tree, fd->collapse); | |
5803 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5804 | fin_bb, first_zero_iter, | |
5805 | l2_dom_bb); | |
5806 | t = NULL_TREE; | |
5807 | } | |
5808 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5809 | t = integer_one_node; | |
5810 | else | |
5811 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
5812 | fold_convert (type, fd->loop.n1), | |
5813 | fold_convert (type, fd->loop.n2)); | |
5814 | if (fd->collapse == 1 | |
5815 | && TYPE_UNSIGNED (type) | |
8e6b4515 | 5816 | && (t == NULL_TREE || !integer_onep (t))) |
5817 | { | |
8e6b4515 | 5818 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); |
5819 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
5820 | true, GSI_SAME_STMT); | |
5821 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
5822 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
5823 | true, GSI_SAME_STMT); | |
5824 | stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, | |
5825 | NULL_TREE, NULL_TREE); | |
5826 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
5827 | if (walk_tree (gimple_cond_lhs_ptr (stmt), | |
5828 | expand_omp_regimplify_p, NULL, NULL) | |
5829 | || walk_tree (gimple_cond_rhs_ptr (stmt), | |
5830 | expand_omp_regimplify_p, NULL, NULL)) | |
5831 | { | |
5832 | gsi = gsi_for_stmt (stmt); | |
5833 | gimple_regimplify_operands (stmt, &gsi); | |
5834 | } | |
5835 | ep = split_block (entry_bb, stmt); | |
5836 | ep->flags = EDGE_TRUE_VALUE; | |
5837 | entry_bb = ep->dest; | |
5838 | ep->probability = REG_BR_PROB_BASE - (REG_BR_PROB_BASE / 2000 - 1); | |
5839 | ep = make_edge (ep->src, fin_bb, EDGE_FALSE_VALUE); | |
5840 | ep->probability = REG_BR_PROB_BASE / 2000 - 1; | |
5841 | if (gimple_in_ssa_p (cfun)) | |
5842 | { | |
5843 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; | |
5844 | for (gsi = gsi_start_phis (fin_bb); | |
5845 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
5846 | { | |
5847 | gimple phi = gsi_stmt (gsi); | |
5848 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
5849 | ep, UNKNOWN_LOCATION); | |
5850 | } | |
5851 | } | |
5852 | gsi = gsi_last_bb (entry_bb); | |
5853 | } | |
5854 | ||
bc7bff74 | 5855 | t = build_call_expr (builtin_decl_explicit (get_num_threads), 0); |
fd6481cf | 5856 | t = fold_convert (itype, t); |
75a70cf9 | 5857 | nthreads = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5858 | true, GSI_SAME_STMT); | |
48e1416a | 5859 | |
bc7bff74 | 5860 | t = build_call_expr (builtin_decl_explicit (get_thread_num), 0); |
fd6481cf | 5861 | t = fold_convert (itype, t); |
75a70cf9 | 5862 | threadid = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5863 | true, GSI_SAME_STMT); | |
1e8e9920 | 5864 | |
bc7bff74 | 5865 | n1 = fd->loop.n1; |
5866 | n2 = fd->loop.n2; | |
5867 | step = fd->loop.step; | |
5868 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5869 | { | |
5870 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5871 | OMP_CLAUSE__LOOPTEMP_); | |
5872 | gcc_assert (innerc); | |
5873 | n1 = OMP_CLAUSE_DECL (innerc); | |
5874 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5875 | OMP_CLAUSE__LOOPTEMP_); | |
5876 | gcc_assert (innerc); | |
5877 | n2 = OMP_CLAUSE_DECL (innerc); | |
5878 | } | |
5879 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
5880 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5881 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
5882 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5883 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
5884 | true, NULL_TREE, true, GSI_SAME_STMT); | |
fd6481cf | 5885 | |
5886 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
bc7bff74 | 5887 | t = fold_build2 (PLUS_EXPR, itype, step, t); |
5888 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
5889 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
fd6481cf | 5890 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) |
5891 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
5892 | fold_build1 (NEGATE_EXPR, itype, t), | |
bc7bff74 | 5893 | fold_build1 (NEGATE_EXPR, itype, step)); |
fd6481cf | 5894 | else |
bc7bff74 | 5895 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
fd6481cf | 5896 | t = fold_convert (itype, t); |
75a70cf9 | 5897 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
1e8e9920 | 5898 | |
072f7ab1 | 5899 | q = create_tmp_reg (itype, "q"); |
fd6481cf | 5900 | t = fold_build2 (TRUNC_DIV_EXPR, itype, n, nthreads); |
31712e83 | 5901 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); |
5902 | gsi_insert_before (&gsi, gimple_build_assign (q, t), GSI_SAME_STMT); | |
5903 | ||
072f7ab1 | 5904 | tt = create_tmp_reg (itype, "tt"); |
31712e83 | 5905 | t = fold_build2 (TRUNC_MOD_EXPR, itype, n, nthreads); |
5906 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
5907 | gsi_insert_before (&gsi, gimple_build_assign (tt, t), GSI_SAME_STMT); | |
1e8e9920 | 5908 | |
31712e83 | 5909 | t = build2 (LT_EXPR, boolean_type_node, threadid, tt); |
5910 | stmt = gimple_build_cond_empty (t); | |
5911 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
5912 | ||
5913 | second_bb = split_block (entry_bb, stmt)->dest; | |
5914 | gsi = gsi_last_bb (second_bb); | |
5915 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
5916 | ||
5917 | gsi_insert_before (&gsi, gimple_build_assign (tt, build_int_cst (itype, 0)), | |
5918 | GSI_SAME_STMT); | |
5919 | stmt = gimple_build_assign_with_ops (PLUS_EXPR, q, q, | |
5920 | build_int_cst (itype, 1)); | |
5921 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
5922 | ||
5923 | third_bb = split_block (second_bb, stmt)->dest; | |
5924 | gsi = gsi_last_bb (third_bb); | |
5925 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
1e8e9920 | 5926 | |
fd6481cf | 5927 | t = build2 (MULT_EXPR, itype, q, threadid); |
31712e83 | 5928 | t = build2 (PLUS_EXPR, itype, t, tt); |
75a70cf9 | 5929 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
1e8e9920 | 5930 | |
fd6481cf | 5931 | t = fold_build2 (PLUS_EXPR, itype, s0, q); |
75a70cf9 | 5932 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
79acaae1 | 5933 | |
1e8e9920 | 5934 | t = build2 (GE_EXPR, boolean_type_node, s0, e0); |
75a70cf9 | 5935 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); |
773c5ba7 | 5936 | |
75a70cf9 | 5937 | /* Remove the GIMPLE_OMP_FOR statement. */ |
5938 | gsi_remove (&gsi, true); | |
773c5ba7 | 5939 | |
5940 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
75a70cf9 | 5941 | gsi = gsi_start_bb (seq_start_bb); |
1e8e9920 | 5942 | |
bc7bff74 | 5943 | tree startvar = fd->loop.v; |
5944 | tree endvar = NULL_TREE; | |
5945 | ||
5946 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5947 | { | |
5948 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
5949 | ? gimple_omp_parallel_clauses (inner_stmt) | |
5950 | : gimple_omp_for_clauses (inner_stmt); | |
5951 | tree innerc = find_omp_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
5952 | gcc_assert (innerc); | |
5953 | startvar = OMP_CLAUSE_DECL (innerc); | |
5954 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5955 | OMP_CLAUSE__LOOPTEMP_); | |
5956 | gcc_assert (innerc); | |
5957 | endvar = OMP_CLAUSE_DECL (innerc); | |
5958 | } | |
fd6481cf | 5959 | t = fold_convert (itype, s0); |
bc7bff74 | 5960 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 5961 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 5962 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 5963 | else |
bc7bff74 | 5964 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
5965 | t = fold_convert (TREE_TYPE (startvar), t); | |
4abecb72 | 5966 | t = force_gimple_operand_gsi (&gsi, t, |
bc7bff74 | 5967 | DECL_P (startvar) |
5968 | && TREE_ADDRESSABLE (startvar), | |
4abecb72 | 5969 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
bc7bff74 | 5970 | stmt = gimple_build_assign (startvar, t); |
75a70cf9 | 5971 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
48e1416a | 5972 | |
fd6481cf | 5973 | t = fold_convert (itype, e0); |
bc7bff74 | 5974 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 5975 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 5976 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 5977 | else |
bc7bff74 | 5978 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
5979 | t = fold_convert (TREE_TYPE (startvar), t); | |
75a70cf9 | 5980 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5981 | false, GSI_CONTINUE_LINKING); | |
bc7bff74 | 5982 | if (endvar) |
5983 | { | |
5984 | stmt = gimple_build_assign (endvar, e); | |
5985 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
5986 | } | |
5987 | if (fd->collapse > 1) | |
5988 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); | |
1e8e9920 | 5989 | |
bc7bff74 | 5990 | if (!broken_loop) |
5991 | { | |
5992 | /* The code controlling the sequential loop replaces the | |
5993 | GIMPLE_OMP_CONTINUE. */ | |
5994 | gsi = gsi_last_bb (cont_bb); | |
5995 | stmt = gsi_stmt (gsi); | |
5996 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
5997 | vmain = gimple_omp_continue_control_use (stmt); | |
5998 | vback = gimple_omp_continue_control_def (stmt); | |
79acaae1 | 5999 | |
bc7bff74 | 6000 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
6001 | { | |
6002 | if (POINTER_TYPE_P (type)) | |
6003 | t = fold_build_pointer_plus (vmain, step); | |
6004 | else | |
6005 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
6006 | t = force_gimple_operand_gsi (&gsi, t, | |
6007 | DECL_P (vback) | |
6008 | && TREE_ADDRESSABLE (vback), | |
6009 | NULL_TREE, true, GSI_SAME_STMT); | |
6010 | stmt = gimple_build_assign (vback, t); | |
6011 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
79acaae1 | 6012 | |
bc7bff74 | 6013 | t = build2 (fd->loop.cond_code, boolean_type_node, |
6014 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
6015 | ? t : vback, e); | |
6016 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
6017 | } | |
1e8e9920 | 6018 | |
bc7bff74 | 6019 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ |
6020 | gsi_remove (&gsi, true); | |
6021 | ||
6022 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
6023 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, body_bb); | |
6024 | } | |
773c5ba7 | 6025 | |
75a70cf9 | 6026 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ |
6027 | gsi = gsi_last_bb (exit_bb); | |
6028 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
bc7bff74 | 6029 | { |
6030 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
6031 | gsi_insert_after (&gsi, build_omp_barrier (t), GSI_SAME_STMT); | |
6032 | } | |
75a70cf9 | 6033 | gsi_remove (&gsi, true); |
773c5ba7 | 6034 | |
6035 | /* Connect all the blocks. */ | |
31712e83 | 6036 | ep = make_edge (entry_bb, third_bb, EDGE_FALSE_VALUE); |
6037 | ep->probability = REG_BR_PROB_BASE / 4 * 3; | |
6038 | ep = find_edge (entry_bb, second_bb); | |
6039 | ep->flags = EDGE_TRUE_VALUE; | |
6040 | ep->probability = REG_BR_PROB_BASE / 4; | |
6041 | find_edge (third_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE; | |
6042 | find_edge (third_bb, fin_bb)->flags = EDGE_TRUE_VALUE; | |
79acaae1 | 6043 | |
bc7bff74 | 6044 | if (!broken_loop) |
6045 | { | |
6046 | ep = find_edge (cont_bb, body_bb); | |
6047 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6048 | { | |
6049 | remove_edge (ep); | |
6050 | ep = NULL; | |
6051 | } | |
6052 | else if (fd->collapse > 1) | |
6053 | { | |
6054 | remove_edge (ep); | |
6055 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
6056 | } | |
6057 | else | |
6058 | ep->flags = EDGE_TRUE_VALUE; | |
6059 | find_edge (cont_bb, fin_bb)->flags | |
6060 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
6061 | } | |
48e1416a | 6062 | |
31712e83 | 6063 | set_immediate_dominator (CDI_DOMINATORS, second_bb, entry_bb); |
6064 | set_immediate_dominator (CDI_DOMINATORS, third_bb, entry_bb); | |
6065 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, third_bb); | |
bc7bff74 | 6066 | |
79acaae1 | 6067 | set_immediate_dominator (CDI_DOMINATORS, body_bb, |
6068 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
6069 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
6070 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
04c2922b | 6071 | |
bc7bff74 | 6072 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) |
6073 | { | |
6074 | struct loop *loop = alloc_loop (); | |
6075 | loop->header = body_bb; | |
6076 | if (collapse_bb == NULL) | |
6077 | loop->latch = cont_bb; | |
6078 | add_loop (loop, body_bb->loop_father); | |
6079 | } | |
1e8e9920 | 6080 | } |
6081 | ||
773c5ba7 | 6082 | |
6083 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
6084 | loop with static schedule and a specified chunk size. Given | |
6085 | parameters: | |
1e8e9920 | 6086 | |
6087 | for (V = N1; V cond N2; V += STEP) BODY; | |
6088 | ||
6089 | where COND is "<" or ">", we generate pseudocode | |
6090 | ||
8e6b4515 | 6091 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; |
1e8e9920 | 6092 | if (cond is <) |
6093 | adj = STEP - 1; | |
6094 | else | |
6095 | adj = STEP + 1; | |
fd6481cf | 6096 | if ((__typeof (V)) -1 > 0 && cond is >) |
6097 | n = -(adj + N2 - N1) / -STEP; | |
6098 | else | |
6099 | n = (adj + N2 - N1) / STEP; | |
1e8e9920 | 6100 | trip = 0; |
79acaae1 | 6101 | V = threadid * CHUNK * STEP + N1; -- this extra definition of V is |
6102 | here so that V is defined | |
6103 | if the loop is not entered | |
1e8e9920 | 6104 | L0: |
6105 | s0 = (trip * nthreads + threadid) * CHUNK; | |
6106 | e0 = min(s0 + CHUNK, n); | |
6107 | if (s0 < n) goto L1; else goto L4; | |
6108 | L1: | |
6109 | V = s0 * STEP + N1; | |
6110 | e = e0 * STEP + N1; | |
6111 | L2: | |
6112 | BODY; | |
6113 | V += STEP; | |
6114 | if (V cond e) goto L2; else goto L3; | |
6115 | L3: | |
6116 | trip += 1; | |
6117 | goto L0; | |
6118 | L4: | |
1e8e9920 | 6119 | */ |
6120 | ||
61e47ac8 | 6121 | static void |
bc7bff74 | 6122 | expand_omp_for_static_chunk (struct omp_region *region, |
6123 | struct omp_for_data *fd, gimple inner_stmt) | |
1e8e9920 | 6124 | { |
75a70cf9 | 6125 | tree n, s0, e0, e, t; |
79acaae1 | 6126 | tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid; |
75a70cf9 | 6127 | tree type, itype, v_main, v_back, v_extra; |
773c5ba7 | 6128 | basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb; |
bc7bff74 | 6129 | basic_block trip_update_bb = NULL, cont_bb, collapse_bb = NULL, fin_bb; |
75a70cf9 | 6130 | gimple_stmt_iterator si; |
6131 | gimple stmt; | |
6132 | edge se; | |
bc7bff74 | 6133 | enum built_in_function get_num_threads = BUILT_IN_OMP_GET_NUM_THREADS; |
6134 | enum built_in_function get_thread_num = BUILT_IN_OMP_GET_THREAD_NUM; | |
6135 | bool broken_loop = region->cont == NULL; | |
6136 | tree *counts = NULL; | |
6137 | tree n1, n2, step; | |
1e8e9920 | 6138 | |
fd6481cf | 6139 | itype = type = TREE_TYPE (fd->loop.v); |
6140 | if (POINTER_TYPE_P (type)) | |
3cea8318 | 6141 | itype = signed_type_for (type); |
1e8e9920 | 6142 | |
61e47ac8 | 6143 | entry_bb = region->entry; |
ac6e3339 | 6144 | se = split_block (entry_bb, last_stmt (entry_bb)); |
6145 | entry_bb = se->src; | |
6146 | iter_part_bb = se->dest; | |
61e47ac8 | 6147 | cont_bb = region->cont; |
ac6e3339 | 6148 | gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2); |
bc7bff74 | 6149 | fin_bb = BRANCH_EDGE (iter_part_bb)->dest; |
6150 | gcc_assert (broken_loop | |
6151 | || fin_bb == FALLTHRU_EDGE (cont_bb)->dest); | |
ac6e3339 | 6152 | seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb)); |
6153 | body_bb = single_succ (seq_start_bb); | |
bc7bff74 | 6154 | if (!broken_loop) |
6155 | { | |
6156 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); | |
6157 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
6158 | trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb)); | |
6159 | } | |
61e47ac8 | 6160 | exit_bb = region->exit; |
773c5ba7 | 6161 | |
773c5ba7 | 6162 | /* Trip and adjustment setup goes in ENTRY_BB. */ |
75a70cf9 | 6163 | si = gsi_last_bb (entry_bb); |
6164 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_FOR); | |
773c5ba7 | 6165 | |
bc7bff74 | 6166 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
6167 | { | |
6168 | get_num_threads = BUILT_IN_OMP_GET_NUM_TEAMS; | |
6169 | get_thread_num = BUILT_IN_OMP_GET_TEAM_NUM; | |
6170 | } | |
6171 | ||
6172 | if (fd->collapse > 1) | |
6173 | { | |
6174 | int first_zero_iter = -1; | |
6175 | basic_block l2_dom_bb = NULL; | |
6176 | ||
6177 | counts = XALLOCAVEC (tree, fd->collapse); | |
6178 | expand_omp_for_init_counts (fd, &si, entry_bb, counts, | |
6179 | fin_bb, first_zero_iter, | |
6180 | l2_dom_bb); | |
6181 | t = NULL_TREE; | |
6182 | } | |
6183 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6184 | t = integer_one_node; | |
6185 | else | |
6186 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
6187 | fold_convert (type, fd->loop.n1), | |
6188 | fold_convert (type, fd->loop.n2)); | |
6189 | if (fd->collapse == 1 | |
6190 | && TYPE_UNSIGNED (type) | |
8e6b4515 | 6191 | && (t == NULL_TREE || !integer_onep (t))) |
6192 | { | |
8e6b4515 | 6193 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); |
6194 | n1 = force_gimple_operand_gsi (&si, n1, true, NULL_TREE, | |
6195 | true, GSI_SAME_STMT); | |
6196 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
6197 | n2 = force_gimple_operand_gsi (&si, n2, true, NULL_TREE, | |
6198 | true, GSI_SAME_STMT); | |
6199 | stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, | |
6200 | NULL_TREE, NULL_TREE); | |
6201 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
6202 | if (walk_tree (gimple_cond_lhs_ptr (stmt), | |
6203 | expand_omp_regimplify_p, NULL, NULL) | |
6204 | || walk_tree (gimple_cond_rhs_ptr (stmt), | |
6205 | expand_omp_regimplify_p, NULL, NULL)) | |
6206 | { | |
6207 | si = gsi_for_stmt (stmt); | |
6208 | gimple_regimplify_operands (stmt, &si); | |
6209 | } | |
6210 | se = split_block (entry_bb, stmt); | |
6211 | se->flags = EDGE_TRUE_VALUE; | |
6212 | entry_bb = se->dest; | |
6213 | se->probability = REG_BR_PROB_BASE - (REG_BR_PROB_BASE / 2000 - 1); | |
6214 | se = make_edge (se->src, fin_bb, EDGE_FALSE_VALUE); | |
6215 | se->probability = REG_BR_PROB_BASE / 2000 - 1; | |
6216 | if (gimple_in_ssa_p (cfun)) | |
6217 | { | |
6218 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; | |
6219 | for (si = gsi_start_phis (fin_bb); | |
6220 | !gsi_end_p (si); gsi_next (&si)) | |
6221 | { | |
6222 | gimple phi = gsi_stmt (si); | |
6223 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
6224 | se, UNKNOWN_LOCATION); | |
6225 | } | |
6226 | } | |
6227 | si = gsi_last_bb (entry_bb); | |
6228 | } | |
6229 | ||
bc7bff74 | 6230 | t = build_call_expr (builtin_decl_explicit (get_num_threads), 0); |
fd6481cf | 6231 | t = fold_convert (itype, t); |
75a70cf9 | 6232 | nthreads = force_gimple_operand_gsi (&si, t, true, NULL_TREE, |
6233 | true, GSI_SAME_STMT); | |
48e1416a | 6234 | |
bc7bff74 | 6235 | t = build_call_expr (builtin_decl_explicit (get_thread_num), 0); |
fd6481cf | 6236 | t = fold_convert (itype, t); |
75a70cf9 | 6237 | threadid = force_gimple_operand_gsi (&si, t, true, NULL_TREE, |
6238 | true, GSI_SAME_STMT); | |
79acaae1 | 6239 | |
bc7bff74 | 6240 | n1 = fd->loop.n1; |
6241 | n2 = fd->loop.n2; | |
6242 | step = fd->loop.step; | |
6243 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6244 | { | |
6245 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6246 | OMP_CLAUSE__LOOPTEMP_); | |
6247 | gcc_assert (innerc); | |
6248 | n1 = OMP_CLAUSE_DECL (innerc); | |
6249 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
6250 | OMP_CLAUSE__LOOPTEMP_); | |
6251 | gcc_assert (innerc); | |
6252 | n2 = OMP_CLAUSE_DECL (innerc); | |
6253 | } | |
6254 | n1 = force_gimple_operand_gsi (&si, fold_convert (type, n1), | |
6255 | true, NULL_TREE, true, GSI_SAME_STMT); | |
6256 | n2 = force_gimple_operand_gsi (&si, fold_convert (itype, n2), | |
6257 | true, NULL_TREE, true, GSI_SAME_STMT); | |
6258 | step = force_gimple_operand_gsi (&si, fold_convert (itype, step), | |
6259 | true, NULL_TREE, true, GSI_SAME_STMT); | |
79acaae1 | 6260 | fd->chunk_size |
75a70cf9 | 6261 | = force_gimple_operand_gsi (&si, fold_convert (itype, fd->chunk_size), |
6262 | true, NULL_TREE, true, GSI_SAME_STMT); | |
fd6481cf | 6263 | |
6264 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
bc7bff74 | 6265 | t = fold_build2 (PLUS_EXPR, itype, step, t); |
6266 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
6267 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
fd6481cf | 6268 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) |
6269 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6270 | fold_build1 (NEGATE_EXPR, itype, t), | |
bc7bff74 | 6271 | fold_build1 (NEGATE_EXPR, itype, step)); |
fd6481cf | 6272 | else |
bc7bff74 | 6273 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
fd6481cf | 6274 | t = fold_convert (itype, t); |
75a70cf9 | 6275 | n = force_gimple_operand_gsi (&si, t, true, NULL_TREE, |
6276 | true, GSI_SAME_STMT); | |
79acaae1 | 6277 | |
083152fb | 6278 | trip_var = create_tmp_reg (itype, ".trip"); |
79acaae1 | 6279 | if (gimple_in_ssa_p (cfun)) |
6280 | { | |
75a70cf9 | 6281 | trip_init = make_ssa_name (trip_var, NULL); |
6282 | trip_main = make_ssa_name (trip_var, NULL); | |
6283 | trip_back = make_ssa_name (trip_var, NULL); | |
79acaae1 | 6284 | } |
1e8e9920 | 6285 | else |
79acaae1 | 6286 | { |
6287 | trip_init = trip_var; | |
6288 | trip_main = trip_var; | |
6289 | trip_back = trip_var; | |
6290 | } | |
1e8e9920 | 6291 | |
75a70cf9 | 6292 | stmt = gimple_build_assign (trip_init, build_int_cst (itype, 0)); |
6293 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
773c5ba7 | 6294 | |
fd6481cf | 6295 | t = fold_build2 (MULT_EXPR, itype, threadid, fd->chunk_size); |
bc7bff74 | 6296 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6297 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6298 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6299 | else |
bc7bff74 | 6300 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
75a70cf9 | 6301 | v_extra = force_gimple_operand_gsi (&si, t, true, NULL_TREE, |
6302 | true, GSI_SAME_STMT); | |
79acaae1 | 6303 | |
75a70cf9 | 6304 | /* Remove the GIMPLE_OMP_FOR. */ |
6305 | gsi_remove (&si, true); | |
773c5ba7 | 6306 | |
6307 | /* Iteration space partitioning goes in ITER_PART_BB. */ | |
75a70cf9 | 6308 | si = gsi_last_bb (iter_part_bb); |
1e8e9920 | 6309 | |
fd6481cf | 6310 | t = fold_build2 (MULT_EXPR, itype, trip_main, nthreads); |
6311 | t = fold_build2 (PLUS_EXPR, itype, t, threadid); | |
6312 | t = fold_build2 (MULT_EXPR, itype, t, fd->chunk_size); | |
75a70cf9 | 6313 | s0 = force_gimple_operand_gsi (&si, t, true, NULL_TREE, |
6314 | false, GSI_CONTINUE_LINKING); | |
1e8e9920 | 6315 | |
fd6481cf | 6316 | t = fold_build2 (PLUS_EXPR, itype, s0, fd->chunk_size); |
6317 | t = fold_build2 (MIN_EXPR, itype, t, n); | |
75a70cf9 | 6318 | e0 = force_gimple_operand_gsi (&si, t, true, NULL_TREE, |
6319 | false, GSI_CONTINUE_LINKING); | |
1e8e9920 | 6320 | |
6321 | t = build2 (LT_EXPR, boolean_type_node, s0, n); | |
75a70cf9 | 6322 | gsi_insert_after (&si, gimple_build_cond_empty (t), GSI_CONTINUE_LINKING); |
773c5ba7 | 6323 | |
6324 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
75a70cf9 | 6325 | si = gsi_start_bb (seq_start_bb); |
1e8e9920 | 6326 | |
bc7bff74 | 6327 | tree startvar = fd->loop.v; |
6328 | tree endvar = NULL_TREE; | |
6329 | ||
6330 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6331 | { | |
6332 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
6333 | ? gimple_omp_parallel_clauses (inner_stmt) | |
6334 | : gimple_omp_for_clauses (inner_stmt); | |
6335 | tree innerc = find_omp_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
6336 | gcc_assert (innerc); | |
6337 | startvar = OMP_CLAUSE_DECL (innerc); | |
6338 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
6339 | OMP_CLAUSE__LOOPTEMP_); | |
6340 | gcc_assert (innerc); | |
6341 | endvar = OMP_CLAUSE_DECL (innerc); | |
6342 | } | |
6343 | ||
fd6481cf | 6344 | t = fold_convert (itype, s0); |
bc7bff74 | 6345 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6346 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6347 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6348 | else |
bc7bff74 | 6349 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6350 | t = fold_convert (TREE_TYPE (startvar), t); | |
4abecb72 | 6351 | t = force_gimple_operand_gsi (&si, t, |
bc7bff74 | 6352 | DECL_P (startvar) |
6353 | && TREE_ADDRESSABLE (startvar), | |
4abecb72 | 6354 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
bc7bff74 | 6355 | stmt = gimple_build_assign (startvar, t); |
75a70cf9 | 6356 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); |
1e8e9920 | 6357 | |
fd6481cf | 6358 | t = fold_convert (itype, e0); |
bc7bff74 | 6359 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6360 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6361 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6362 | else |
bc7bff74 | 6363 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6364 | t = fold_convert (TREE_TYPE (startvar), t); | |
75a70cf9 | 6365 | e = force_gimple_operand_gsi (&si, t, true, NULL_TREE, |
6366 | false, GSI_CONTINUE_LINKING); | |
bc7bff74 | 6367 | if (endvar) |
6368 | { | |
6369 | stmt = gimple_build_assign (endvar, e); | |
6370 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); | |
6371 | } | |
6372 | if (fd->collapse > 1) | |
6373 | expand_omp_for_init_vars (fd, &si, counts, inner_stmt, startvar); | |
6374 | ||
6375 | if (!broken_loop) | |
6376 | { | |
6377 | /* The code controlling the sequential loop goes in CONT_BB, | |
6378 | replacing the GIMPLE_OMP_CONTINUE. */ | |
6379 | si = gsi_last_bb (cont_bb); | |
6380 | stmt = gsi_stmt (si); | |
6381 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
6382 | v_main = gimple_omp_continue_control_use (stmt); | |
6383 | v_back = gimple_omp_continue_control_def (stmt); | |
1e8e9920 | 6384 | |
bc7bff74 | 6385 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
6386 | { | |
6387 | if (POINTER_TYPE_P (type)) | |
6388 | t = fold_build_pointer_plus (v_main, step); | |
6389 | else | |
6390 | t = fold_build2 (PLUS_EXPR, type, v_main, step); | |
6391 | if (DECL_P (v_back) && TREE_ADDRESSABLE (v_back)) | |
6392 | t = force_gimple_operand_gsi (&si, t, true, NULL_TREE, | |
6393 | true, GSI_SAME_STMT); | |
6394 | stmt = gimple_build_assign (v_back, t); | |
6395 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
79acaae1 | 6396 | |
bc7bff74 | 6397 | t = build2 (fd->loop.cond_code, boolean_type_node, |
6398 | DECL_P (v_back) && TREE_ADDRESSABLE (v_back) | |
6399 | ? t : v_back, e); | |
6400 | gsi_insert_before (&si, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
6401 | } | |
79acaae1 | 6402 | |
bc7bff74 | 6403 | /* Remove GIMPLE_OMP_CONTINUE. */ |
6404 | gsi_remove (&si, true); | |
48e1416a | 6405 | |
bc7bff74 | 6406 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) |
6407 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, body_bb); | |
773c5ba7 | 6408 | |
bc7bff74 | 6409 | /* Trip update code goes into TRIP_UPDATE_BB. */ |
6410 | si = gsi_start_bb (trip_update_bb); | |
1e8e9920 | 6411 | |
bc7bff74 | 6412 | t = build_int_cst (itype, 1); |
6413 | t = build2 (PLUS_EXPR, itype, trip_main, t); | |
6414 | stmt = gimple_build_assign (trip_back, t); | |
6415 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); | |
6416 | } | |
1e8e9920 | 6417 | |
75a70cf9 | 6418 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ |
6419 | si = gsi_last_bb (exit_bb); | |
6420 | if (!gimple_omp_return_nowait_p (gsi_stmt (si))) | |
bc7bff74 | 6421 | { |
6422 | t = gimple_omp_return_lhs (gsi_stmt (si)); | |
6423 | gsi_insert_after (&si, build_omp_barrier (t), GSI_SAME_STMT); | |
6424 | } | |
75a70cf9 | 6425 | gsi_remove (&si, true); |
1e8e9920 | 6426 | |
773c5ba7 | 6427 | /* Connect the new blocks. */ |
ac6e3339 | 6428 | find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE; |
6429 | find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE; | |
79acaae1 | 6430 | |
bc7bff74 | 6431 | if (!broken_loop) |
6432 | { | |
6433 | se = find_edge (cont_bb, body_bb); | |
6434 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6435 | { | |
6436 | remove_edge (se); | |
6437 | se = NULL; | |
6438 | } | |
6439 | else if (fd->collapse > 1) | |
6440 | { | |
6441 | remove_edge (se); | |
6442 | se = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
6443 | } | |
6444 | else | |
6445 | se->flags = EDGE_TRUE_VALUE; | |
6446 | find_edge (cont_bb, trip_update_bb)->flags | |
6447 | = se ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
79acaae1 | 6448 | |
bc7bff74 | 6449 | redirect_edge_and_branch (single_succ_edge (trip_update_bb), iter_part_bb); |
6450 | } | |
79acaae1 | 6451 | |
6452 | if (gimple_in_ssa_p (cfun)) | |
6453 | { | |
75a70cf9 | 6454 | gimple_stmt_iterator psi; |
6455 | gimple phi; | |
6456 | edge re, ene; | |
f1f41a6c | 6457 | edge_var_map_vector *head; |
75a70cf9 | 6458 | edge_var_map *vm; |
6459 | size_t i; | |
6460 | ||
bc7bff74 | 6461 | gcc_assert (fd->collapse == 1 && !broken_loop); |
6462 | ||
79acaae1 | 6463 | /* When we redirect the edge from trip_update_bb to iter_part_bb, we |
6464 | remove arguments of the phi nodes in fin_bb. We need to create | |
6465 | appropriate phi nodes in iter_part_bb instead. */ | |
6466 | se = single_pred_edge (fin_bb); | |
6467 | re = single_succ_edge (trip_update_bb); | |
75a70cf9 | 6468 | head = redirect_edge_var_map_vector (re); |
79acaae1 | 6469 | ene = single_succ_edge (entry_bb); |
6470 | ||
75a70cf9 | 6471 | psi = gsi_start_phis (fin_bb); |
f1f41a6c | 6472 | for (i = 0; !gsi_end_p (psi) && head->iterate (i, &vm); |
75a70cf9 | 6473 | gsi_next (&psi), ++i) |
79acaae1 | 6474 | { |
75a70cf9 | 6475 | gimple nphi; |
efbcb6de | 6476 | source_location locus; |
75a70cf9 | 6477 | |
6478 | phi = gsi_stmt (psi); | |
6479 | t = gimple_phi_result (phi); | |
6480 | gcc_assert (t == redirect_edge_var_map_result (vm)); | |
79acaae1 | 6481 | nphi = create_phi_node (t, iter_part_bb); |
79acaae1 | 6482 | |
6483 | t = PHI_ARG_DEF_FROM_EDGE (phi, se); | |
efbcb6de | 6484 | locus = gimple_phi_arg_location_from_edge (phi, se); |
6485 | ||
fd6481cf | 6486 | /* A special case -- fd->loop.v is not yet computed in |
6487 | iter_part_bb, we need to use v_extra instead. */ | |
6488 | if (t == fd->loop.v) | |
79acaae1 | 6489 | t = v_extra; |
60d535d2 | 6490 | add_phi_arg (nphi, t, ene, locus); |
efbcb6de | 6491 | locus = redirect_edge_var_map_location (vm); |
60d535d2 | 6492 | add_phi_arg (nphi, redirect_edge_var_map_def (vm), re, locus); |
75a70cf9 | 6493 | } |
f1f41a6c | 6494 | gcc_assert (!gsi_end_p (psi) && i == head->length ()); |
75a70cf9 | 6495 | redirect_edge_var_map_clear (re); |
6496 | while (1) | |
6497 | { | |
6498 | psi = gsi_start_phis (fin_bb); | |
6499 | if (gsi_end_p (psi)) | |
6500 | break; | |
6501 | remove_phi_node (&psi, false); | |
79acaae1 | 6502 | } |
79acaae1 | 6503 | |
6504 | /* Make phi node for trip. */ | |
6505 | phi = create_phi_node (trip_main, iter_part_bb); | |
efbcb6de | 6506 | add_phi_arg (phi, trip_back, single_succ_edge (trip_update_bb), |
60d535d2 | 6507 | UNKNOWN_LOCATION); |
efbcb6de | 6508 | add_phi_arg (phi, trip_init, single_succ_edge (entry_bb), |
60d535d2 | 6509 | UNKNOWN_LOCATION); |
79acaae1 | 6510 | } |
6511 | ||
bc7bff74 | 6512 | if (!broken_loop) |
6513 | set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb); | |
79acaae1 | 6514 | set_immediate_dominator (CDI_DOMINATORS, iter_part_bb, |
6515 | recompute_dominator (CDI_DOMINATORS, iter_part_bb)); | |
6516 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
6517 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
6518 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, | |
6519 | recompute_dominator (CDI_DOMINATORS, seq_start_bb)); | |
6520 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
6521 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
04c2922b | 6522 | |
bc7bff74 | 6523 | if (!broken_loop) |
6524 | { | |
6525 | struct loop *trip_loop = alloc_loop (); | |
6526 | trip_loop->header = iter_part_bb; | |
6527 | trip_loop->latch = trip_update_bb; | |
6528 | add_loop (trip_loop, iter_part_bb->loop_father); | |
04c2922b | 6529 | |
bc7bff74 | 6530 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
6531 | { | |
6532 | struct loop *loop = alloc_loop (); | |
6533 | loop->header = body_bb; | |
33ee4d72 | 6534 | if (collapse_bb == NULL) |
6535 | loop->latch = cont_bb; | |
bc7bff74 | 6536 | add_loop (loop, trip_loop); |
6537 | } | |
6538 | } | |
1e8e9920 | 6539 | } |
6540 | ||
bc7bff74 | 6541 | |
3d483a94 | 6542 | /* A subroutine of expand_omp_for. Generate code for a simd non-worksharing |
6543 | loop. Given parameters: | |
6544 | ||
6545 | for (V = N1; V cond N2; V += STEP) BODY; | |
6546 | ||
6547 | where COND is "<" or ">", we generate pseudocode | |
6548 | ||
6549 | V = N1; | |
6550 | goto L1; | |
6551 | L0: | |
6552 | BODY; | |
6553 | V += STEP; | |
6554 | L1: | |
6555 | if (V cond N2) goto L0; else goto L2; | |
6556 | L2: | |
6557 | ||
6558 | For collapsed loops, given parameters: | |
6559 | collapse(3) | |
6560 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
6561 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
6562 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
6563 | BODY; | |
6564 | ||
6565 | we generate pseudocode | |
6566 | ||
6567 | if (cond3 is <) | |
6568 | adj = STEP3 - 1; | |
6569 | else | |
6570 | adj = STEP3 + 1; | |
6571 | count3 = (adj + N32 - N31) / STEP3; | |
6572 | if (cond2 is <) | |
6573 | adj = STEP2 - 1; | |
6574 | else | |
6575 | adj = STEP2 + 1; | |
6576 | count2 = (adj + N22 - N21) / STEP2; | |
6577 | if (cond1 is <) | |
6578 | adj = STEP1 - 1; | |
6579 | else | |
6580 | adj = STEP1 + 1; | |
6581 | count1 = (adj + N12 - N11) / STEP1; | |
6582 | count = count1 * count2 * count3; | |
6583 | V = 0; | |
6584 | V1 = N11; | |
6585 | V2 = N21; | |
6586 | V3 = N31; | |
6587 | goto L1; | |
6588 | L0: | |
6589 | BODY; | |
6590 | V += 1; | |
6591 | V3 += STEP3; | |
6592 | V2 += (V3 cond3 N32) ? 0 : STEP2; | |
6593 | V3 = (V3 cond3 N32) ? V3 : N31; | |
6594 | V1 += (V2 cond2 N22) ? 0 : STEP1; | |
6595 | V2 = (V2 cond2 N22) ? V2 : N21; | |
6596 | L1: | |
6597 | if (V < count) goto L0; else goto L2; | |
6598 | L2: | |
6599 | ||
6600 | */ | |
6601 | ||
6602 | static void | |
6603 | expand_omp_simd (struct omp_region *region, struct omp_for_data *fd) | |
6604 | { | |
6605 | tree type, t; | |
6606 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, l2_bb, l2_dom_bb; | |
6607 | gimple_stmt_iterator gsi; | |
6608 | gimple stmt; | |
6609 | bool broken_loop = region->cont == NULL; | |
6610 | edge e, ne; | |
6611 | tree *counts = NULL; | |
6612 | int i; | |
6613 | tree safelen = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6614 | OMP_CLAUSE_SAFELEN); | |
6615 | tree simduid = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6616 | OMP_CLAUSE__SIMDUID_); | |
bc7bff74 | 6617 | tree n1, n2; |
3d483a94 | 6618 | |
6619 | type = TREE_TYPE (fd->loop.v); | |
6620 | entry_bb = region->entry; | |
6621 | cont_bb = region->cont; | |
6622 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
6623 | gcc_assert (broken_loop | |
6624 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
6625 | l0_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
6626 | if (!broken_loop) | |
6627 | { | |
6628 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l0_bb); | |
6629 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
6630 | l1_bb = split_block (cont_bb, last_stmt (cont_bb))->dest; | |
6631 | l2_bb = BRANCH_EDGE (entry_bb)->dest; | |
6632 | } | |
6633 | else | |
6634 | { | |
6635 | BRANCH_EDGE (entry_bb)->flags &= ~EDGE_ABNORMAL; | |
6636 | l1_bb = split_edge (BRANCH_EDGE (entry_bb)); | |
6637 | l2_bb = single_succ (l1_bb); | |
6638 | } | |
6639 | exit_bb = region->exit; | |
6640 | l2_dom_bb = NULL; | |
6641 | ||
6642 | gsi = gsi_last_bb (entry_bb); | |
6643 | ||
6644 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
6645 | /* Not needed in SSA form right now. */ | |
6646 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
6647 | if (fd->collapse > 1) | |
6648 | { | |
6649 | int first_zero_iter = -1; | |
6650 | basic_block zero_iter_bb = l2_bb; | |
6651 | ||
6652 | counts = XALLOCAVEC (tree, fd->collapse); | |
6653 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
6654 | zero_iter_bb, first_zero_iter, | |
6655 | l2_dom_bb); | |
6656 | } | |
6657 | if (l2_dom_bb == NULL) | |
6658 | l2_dom_bb = l1_bb; | |
6659 | ||
bc7bff74 | 6660 | n1 = fd->loop.n1; |
3d483a94 | 6661 | n2 = fd->loop.n2; |
bc7bff74 | 6662 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) |
6663 | { | |
6664 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6665 | OMP_CLAUSE__LOOPTEMP_); | |
6666 | gcc_assert (innerc); | |
6667 | n1 = OMP_CLAUSE_DECL (innerc); | |
6668 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
6669 | OMP_CLAUSE__LOOPTEMP_); | |
6670 | gcc_assert (innerc); | |
6671 | n2 = OMP_CLAUSE_DECL (innerc); | |
6672 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6673 | fold_convert (type, n1)); | |
6674 | if (fd->collapse > 1) | |
6675 | { | |
6676 | gsi_prev (&gsi); | |
6677 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, n1); | |
6678 | gsi_next (&gsi); | |
6679 | } | |
6680 | } | |
3d483a94 | 6681 | else |
6682 | { | |
6683 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6684 | fold_convert (type, fd->loop.n1)); | |
6685 | if (fd->collapse > 1) | |
6686 | for (i = 0; i < fd->collapse; i++) | |
6687 | { | |
6688 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6689 | if (POINTER_TYPE_P (itype)) | |
6690 | itype = signed_type_for (itype); | |
6691 | t = fold_convert (TREE_TYPE (fd->loops[i].v), fd->loops[i].n1); | |
6692 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6693 | } | |
6694 | } | |
6695 | ||
6696 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
6697 | gsi_remove (&gsi, true); | |
6698 | ||
6699 | if (!broken_loop) | |
6700 | { | |
6701 | /* Code to control the increment goes in the CONT_BB. */ | |
6702 | gsi = gsi_last_bb (cont_bb); | |
6703 | stmt = gsi_stmt (gsi); | |
6704 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
6705 | ||
6706 | if (POINTER_TYPE_P (type)) | |
6707 | t = fold_build_pointer_plus (fd->loop.v, fd->loop.step); | |
6708 | else | |
6709 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, fd->loop.step); | |
6710 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
6711 | ||
6712 | if (fd->collapse > 1) | |
6713 | { | |
6714 | i = fd->collapse - 1; | |
6715 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
6716 | { | |
6717 | t = fold_convert (sizetype, fd->loops[i].step); | |
6718 | t = fold_build_pointer_plus (fd->loops[i].v, t); | |
6719 | } | |
6720 | else | |
6721 | { | |
6722 | t = fold_convert (TREE_TYPE (fd->loops[i].v), | |
6723 | fd->loops[i].step); | |
6724 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
6725 | fd->loops[i].v, t); | |
6726 | } | |
6727 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6728 | ||
6729 | for (i = fd->collapse - 1; i > 0; i--) | |
6730 | { | |
6731 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6732 | tree itype2 = TREE_TYPE (fd->loops[i - 1].v); | |
6733 | if (POINTER_TYPE_P (itype2)) | |
6734 | itype2 = signed_type_for (itype2); | |
6735 | t = build3 (COND_EXPR, itype2, | |
6736 | build2 (fd->loops[i].cond_code, boolean_type_node, | |
6737 | fd->loops[i].v, | |
6738 | fold_convert (itype, fd->loops[i].n2)), | |
6739 | build_int_cst (itype2, 0), | |
6740 | fold_convert (itype2, fd->loops[i - 1].step)); | |
6741 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i - 1].v))) | |
6742 | t = fold_build_pointer_plus (fd->loops[i - 1].v, t); | |
6743 | else | |
6744 | t = fold_build2 (PLUS_EXPR, itype2, fd->loops[i - 1].v, t); | |
6745 | expand_omp_build_assign (&gsi, fd->loops[i - 1].v, t); | |
6746 | ||
6747 | t = build3 (COND_EXPR, itype, | |
6748 | build2 (fd->loops[i].cond_code, boolean_type_node, | |
6749 | fd->loops[i].v, | |
6750 | fold_convert (itype, fd->loops[i].n2)), | |
6751 | fd->loops[i].v, | |
6752 | fold_convert (itype, fd->loops[i].n1)); | |
6753 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6754 | } | |
6755 | } | |
6756 | ||
6757 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
6758 | gsi_remove (&gsi, true); | |
6759 | } | |
6760 | ||
6761 | /* Emit the condition in L1_BB. */ | |
6762 | gsi = gsi_start_bb (l1_bb); | |
6763 | ||
6764 | t = fold_convert (type, n2); | |
6765 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6766 | false, GSI_CONTINUE_LINKING); | |
6767 | t = build2 (fd->loop.cond_code, boolean_type_node, fd->loop.v, t); | |
6768 | stmt = gimple_build_cond_empty (t); | |
6769 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
6770 | if (walk_tree (gimple_cond_lhs_ptr (stmt), expand_omp_regimplify_p, | |
6771 | NULL, NULL) | |
6772 | || walk_tree (gimple_cond_rhs_ptr (stmt), expand_omp_regimplify_p, | |
6773 | NULL, NULL)) | |
6774 | { | |
6775 | gsi = gsi_for_stmt (stmt); | |
6776 | gimple_regimplify_operands (stmt, &gsi); | |
6777 | } | |
6778 | ||
6779 | /* Remove GIMPLE_OMP_RETURN. */ | |
6780 | gsi = gsi_last_bb (exit_bb); | |
6781 | gsi_remove (&gsi, true); | |
6782 | ||
6783 | /* Connect the new blocks. */ | |
6784 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
6785 | ||
6786 | if (!broken_loop) | |
6787 | { | |
6788 | remove_edge (BRANCH_EDGE (entry_bb)); | |
6789 | make_edge (entry_bb, l1_bb, EDGE_FALLTHRU); | |
6790 | ||
6791 | e = BRANCH_EDGE (l1_bb); | |
6792 | ne = FALLTHRU_EDGE (l1_bb); | |
6793 | e->flags = EDGE_TRUE_VALUE; | |
6794 | } | |
6795 | else | |
6796 | { | |
6797 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
6798 | ||
6799 | ne = single_succ_edge (l1_bb); | |
6800 | e = make_edge (l1_bb, l0_bb, EDGE_TRUE_VALUE); | |
6801 | ||
6802 | } | |
6803 | ne->flags = EDGE_FALSE_VALUE; | |
6804 | e->probability = REG_BR_PROB_BASE * 7 / 8; | |
6805 | ne->probability = REG_BR_PROB_BASE / 8; | |
6806 | ||
6807 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, entry_bb); | |
6808 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, l2_dom_bb); | |
6809 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, l1_bb); | |
6810 | ||
6811 | if (!broken_loop) | |
6812 | { | |
6813 | struct loop *loop = alloc_loop (); | |
6814 | loop->header = l1_bb; | |
33ee4d72 | 6815 | loop->latch = cont_bb; |
3d483a94 | 6816 | add_loop (loop, l1_bb->loop_father); |
6817 | if (safelen == NULL_TREE) | |
6818 | loop->safelen = INT_MAX; | |
6819 | else | |
6820 | { | |
6821 | safelen = OMP_CLAUSE_SAFELEN_EXPR (safelen); | |
cd4547bf | 6822 | if (!tree_fits_uhwi_p (safelen) |
aa59f000 | 6823 | || tree_to_uhwi (safelen) > INT_MAX) |
3d483a94 | 6824 | loop->safelen = INT_MAX; |
6825 | else | |
6a0712d4 | 6826 | loop->safelen = tree_to_uhwi (safelen); |
3d483a94 | 6827 | if (loop->safelen == 1) |
6828 | loop->safelen = 0; | |
6829 | } | |
6830 | if (simduid) | |
6831 | { | |
6832 | loop->simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); | |
6833 | cfun->has_simduid_loops = true; | |
6834 | } | |
043115ec | 6835 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize |
3d483a94 | 6836 | the loop. */ |
043115ec | 6837 | if ((flag_tree_loop_vectorize |
6838 | || (!global_options_set.x_flag_tree_loop_vectorize | |
6839 | && !global_options_set.x_flag_tree_vectorize)) | |
3d483a94 | 6840 | && loop->safelen > 1) |
6841 | { | |
6842 | loop->force_vect = true; | |
6843 | cfun->has_force_vect_loops = true; | |
6844 | } | |
6845 | } | |
6846 | } | |
6847 | ||
1e8e9920 | 6848 | |
773c5ba7 | 6849 | /* Expand the OpenMP loop defined by REGION. */ |
1e8e9920 | 6850 | |
773c5ba7 | 6851 | static void |
bc7bff74 | 6852 | expand_omp_for (struct omp_region *region, gimple inner_stmt) |
773c5ba7 | 6853 | { |
6854 | struct omp_for_data fd; | |
fd6481cf | 6855 | struct omp_for_data_loop *loops; |
1e8e9920 | 6856 | |
fd6481cf | 6857 | loops |
6858 | = (struct omp_for_data_loop *) | |
75a70cf9 | 6859 | alloca (gimple_omp_for_collapse (last_stmt (region->entry)) |
fd6481cf | 6860 | * sizeof (struct omp_for_data_loop)); |
fd6481cf | 6861 | extract_omp_for_data (last_stmt (region->entry), &fd, loops); |
f77459c5 | 6862 | region->sched_kind = fd.sched_kind; |
1e8e9920 | 6863 | |
b3a3ddec | 6864 | gcc_assert (EDGE_COUNT (region->entry->succs) == 2); |
6865 | BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
6866 | FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
6867 | if (region->cont) | |
6868 | { | |
6869 | gcc_assert (EDGE_COUNT (region->cont->succs) == 2); | |
6870 | BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
6871 | FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
6872 | } | |
04c2922b | 6873 | else |
75de4aa2 | 6874 | /* If there isn't a continue then this is a degerate case where |
04c2922b | 6875 | the introduction of abnormal edges during lowering will prevent |
6876 | original loops from being detected. Fix that up. */ | |
6877 | loops_state_set (LOOPS_NEED_FIXUP); | |
b3a3ddec | 6878 | |
f2697631 | 6879 | if (gimple_omp_for_kind (fd.for_stmt) & GF_OMP_FOR_KIND_SIMD) |
3d483a94 | 6880 | expand_omp_simd (region, &fd); |
6881 | else if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC | |
bc7bff74 | 6882 | && !fd.have_ordered) |
1e8e9920 | 6883 | { |
6884 | if (fd.chunk_size == NULL) | |
bc7bff74 | 6885 | expand_omp_for_static_nochunk (region, &fd, inner_stmt); |
1e8e9920 | 6886 | else |
bc7bff74 | 6887 | expand_omp_for_static_chunk (region, &fd, inner_stmt); |
1e8e9920 | 6888 | } |
6889 | else | |
6890 | { | |
fd6481cf | 6891 | int fn_index, start_ix, next_ix; |
6892 | ||
3d483a94 | 6893 | gcc_assert (gimple_omp_for_kind (fd.for_stmt) |
6894 | == GF_OMP_FOR_KIND_FOR); | |
0416ca72 | 6895 | if (fd.chunk_size == NULL |
6896 | && fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) | |
6897 | fd.chunk_size = integer_zero_node; | |
fd6481cf | 6898 | gcc_assert (fd.sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); |
6899 | fn_index = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_RUNTIME) | |
75a70cf9 | 6900 | ? 3 : fd.sched_kind; |
fd6481cf | 6901 | fn_index += fd.have_ordered * 4; |
b9a16870 | 6902 | start_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_START) + fn_index; |
6903 | next_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_NEXT) + fn_index; | |
fd6481cf | 6904 | if (fd.iter_type == long_long_unsigned_type_node) |
6905 | { | |
b9a16870 | 6906 | start_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_START |
6907 | - (int)BUILT_IN_GOMP_LOOP_STATIC_START); | |
6908 | next_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_NEXT | |
6909 | - (int)BUILT_IN_GOMP_LOOP_STATIC_NEXT); | |
fd6481cf | 6910 | } |
b9c74b4d | 6911 | expand_omp_for_generic (region, &fd, (enum built_in_function) start_ix, |
bc7bff74 | 6912 | (enum built_in_function) next_ix, inner_stmt); |
1e8e9920 | 6913 | } |
28c92cbb | 6914 | |
083152fb | 6915 | if (gimple_in_ssa_p (cfun)) |
6916 | update_ssa (TODO_update_ssa_only_virtuals); | |
1e8e9920 | 6917 | } |
6918 | ||
1e8e9920 | 6919 | |
6920 | /* Expand code for an OpenMP sections directive. In pseudo code, we generate | |
6921 | ||
1e8e9920 | 6922 | v = GOMP_sections_start (n); |
6923 | L0: | |
6924 | switch (v) | |
6925 | { | |
6926 | case 0: | |
6927 | goto L2; | |
6928 | case 1: | |
6929 | section 1; | |
6930 | goto L1; | |
6931 | case 2: | |
6932 | ... | |
6933 | case n: | |
6934 | ... | |
1e8e9920 | 6935 | default: |
6936 | abort (); | |
6937 | } | |
6938 | L1: | |
6939 | v = GOMP_sections_next (); | |
6940 | goto L0; | |
6941 | L2: | |
6942 | reduction; | |
6943 | ||
773c5ba7 | 6944 | If this is a combined parallel sections, replace the call to |
79acaae1 | 6945 | GOMP_sections_start with call to GOMP_sections_next. */ |
1e8e9920 | 6946 | |
6947 | static void | |
773c5ba7 | 6948 | expand_omp_sections (struct omp_region *region) |
1e8e9920 | 6949 | { |
f018d957 | 6950 | tree t, u, vin = NULL, vmain, vnext, l2; |
75a70cf9 | 6951 | unsigned len; |
ac6e3339 | 6952 | basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb; |
75a70cf9 | 6953 | gimple_stmt_iterator si, switch_si; |
6954 | gimple sections_stmt, stmt, cont; | |
9884aaf8 | 6955 | edge_iterator ei; |
6956 | edge e; | |
61e47ac8 | 6957 | struct omp_region *inner; |
75a70cf9 | 6958 | unsigned i, casei; |
ac6e3339 | 6959 | bool exit_reachable = region->cont != NULL; |
1e8e9920 | 6960 | |
d244d9de | 6961 | gcc_assert (region->exit != NULL); |
61e47ac8 | 6962 | entry_bb = region->entry; |
ac6e3339 | 6963 | l0_bb = single_succ (entry_bb); |
61e47ac8 | 6964 | l1_bb = region->cont; |
ac6e3339 | 6965 | l2_bb = region->exit; |
d244d9de | 6966 | if (single_pred_p (l2_bb) && single_pred (l2_bb) == l0_bb) |
6967 | l2 = gimple_block_label (l2_bb); | |
6968 | else | |
03ed154b | 6969 | { |
d244d9de | 6970 | /* This can happen if there are reductions. */ |
6971 | len = EDGE_COUNT (l0_bb->succs); | |
6972 | gcc_assert (len > 0); | |
6973 | e = EDGE_SUCC (l0_bb, len - 1); | |
6974 | si = gsi_last_bb (e->dest); | |
6975 | l2 = NULL_TREE; | |
6976 | if (gsi_end_p (si) | |
6977 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
6978 | l2 = gimple_block_label (e->dest); | |
9884aaf8 | 6979 | else |
d244d9de | 6980 | FOR_EACH_EDGE (e, ei, l0_bb->succs) |
6981 | { | |
6982 | si = gsi_last_bb (e->dest); | |
6983 | if (gsi_end_p (si) | |
6984 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
9884aaf8 | 6985 | { |
d244d9de | 6986 | l2 = gimple_block_label (e->dest); |
6987 | break; | |
9884aaf8 | 6988 | } |
d244d9de | 6989 | } |
03ed154b | 6990 | } |
d244d9de | 6991 | if (exit_reachable) |
6992 | default_bb = create_empty_bb (l1_bb->prev_bb); | |
03ed154b | 6993 | else |
d244d9de | 6994 | default_bb = create_empty_bb (l0_bb); |
773c5ba7 | 6995 | |
6996 | /* We will build a switch() with enough cases for all the | |
75a70cf9 | 6997 | GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work |
773c5ba7 | 6998 | and a default case to abort if something goes wrong. */ |
ac6e3339 | 6999 | len = EDGE_COUNT (l0_bb->succs); |
75a70cf9 | 7000 | |
f1f41a6c | 7001 | /* Use vec::quick_push on label_vec throughout, since we know the size |
75a70cf9 | 7002 | in advance. */ |
c2078b80 | 7003 | auto_vec<tree> label_vec (len); |
1e8e9920 | 7004 | |
61e47ac8 | 7005 | /* The call to GOMP_sections_start goes in ENTRY_BB, replacing the |
75a70cf9 | 7006 | GIMPLE_OMP_SECTIONS statement. */ |
7007 | si = gsi_last_bb (entry_bb); | |
7008 | sections_stmt = gsi_stmt (si); | |
7009 | gcc_assert (gimple_code (sections_stmt) == GIMPLE_OMP_SECTIONS); | |
7010 | vin = gimple_omp_sections_control (sections_stmt); | |
773c5ba7 | 7011 | if (!is_combined_parallel (region)) |
1e8e9920 | 7012 | { |
773c5ba7 | 7013 | /* If we are not inside a combined parallel+sections region, |
7014 | call GOMP_sections_start. */ | |
39cb6d68 | 7015 | t = build_int_cst (unsigned_type_node, len - 1); |
b9a16870 | 7016 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_START); |
75a70cf9 | 7017 | stmt = gimple_build_call (u, 1, t); |
1e8e9920 | 7018 | } |
79acaae1 | 7019 | else |
7020 | { | |
7021 | /* Otherwise, call GOMP_sections_next. */ | |
b9a16870 | 7022 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); |
75a70cf9 | 7023 | stmt = gimple_build_call (u, 0); |
79acaae1 | 7024 | } |
75a70cf9 | 7025 | gimple_call_set_lhs (stmt, vin); |
7026 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
7027 | gsi_remove (&si, true); | |
7028 | ||
7029 | /* The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in | |
7030 | L0_BB. */ | |
7031 | switch_si = gsi_last_bb (l0_bb); | |
7032 | gcc_assert (gimple_code (gsi_stmt (switch_si)) == GIMPLE_OMP_SECTIONS_SWITCH); | |
79acaae1 | 7033 | if (exit_reachable) |
7034 | { | |
7035 | cont = last_stmt (l1_bb); | |
75a70cf9 | 7036 | gcc_assert (gimple_code (cont) == GIMPLE_OMP_CONTINUE); |
7037 | vmain = gimple_omp_continue_control_use (cont); | |
7038 | vnext = gimple_omp_continue_control_def (cont); | |
79acaae1 | 7039 | } |
7040 | else | |
7041 | { | |
7042 | vmain = vin; | |
7043 | vnext = NULL_TREE; | |
7044 | } | |
1e8e9920 | 7045 | |
d244d9de | 7046 | t = build_case_label (build_int_cst (unsigned_type_node, 0), NULL, l2); |
f1f41a6c | 7047 | label_vec.quick_push (t); |
d244d9de | 7048 | i = 1; |
03ed154b | 7049 | |
75a70cf9 | 7050 | /* Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR. */ |
ac6e3339 | 7051 | for (inner = region->inner, casei = 1; |
7052 | inner; | |
7053 | inner = inner->next, i++, casei++) | |
1e8e9920 | 7054 | { |
773c5ba7 | 7055 | basic_block s_entry_bb, s_exit_bb; |
7056 | ||
9884aaf8 | 7057 | /* Skip optional reduction region. */ |
75a70cf9 | 7058 | if (inner->type == GIMPLE_OMP_ATOMIC_LOAD) |
9884aaf8 | 7059 | { |
7060 | --i; | |
7061 | --casei; | |
7062 | continue; | |
7063 | } | |
7064 | ||
61e47ac8 | 7065 | s_entry_bb = inner->entry; |
7066 | s_exit_bb = inner->exit; | |
1e8e9920 | 7067 | |
75a70cf9 | 7068 | t = gimple_block_label (s_entry_bb); |
ac6e3339 | 7069 | u = build_int_cst (unsigned_type_node, casei); |
b6e3dd65 | 7070 | u = build_case_label (u, NULL, t); |
f1f41a6c | 7071 | label_vec.quick_push (u); |
61e47ac8 | 7072 | |
75a70cf9 | 7073 | si = gsi_last_bb (s_entry_bb); |
7074 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SECTION); | |
7075 | gcc_assert (i < len || gimple_omp_section_last_p (gsi_stmt (si))); | |
7076 | gsi_remove (&si, true); | |
61e47ac8 | 7077 | single_succ_edge (s_entry_bb)->flags = EDGE_FALLTHRU; |
03ed154b | 7078 | |
7079 | if (s_exit_bb == NULL) | |
7080 | continue; | |
7081 | ||
75a70cf9 | 7082 | si = gsi_last_bb (s_exit_bb); |
7083 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); | |
7084 | gsi_remove (&si, true); | |
03ed154b | 7085 | |
773c5ba7 | 7086 | single_succ_edge (s_exit_bb)->flags = EDGE_FALLTHRU; |
1e8e9920 | 7087 | } |
7088 | ||
773c5ba7 | 7089 | /* Error handling code goes in DEFAULT_BB. */ |
75a70cf9 | 7090 | t = gimple_block_label (default_bb); |
b6e3dd65 | 7091 | u = build_case_label (NULL, NULL, t); |
61e47ac8 | 7092 | make_edge (l0_bb, default_bb, 0); |
f6568ea4 | 7093 | if (current_loops) |
04c2922b | 7094 | add_bb_to_loop (default_bb, current_loops->tree_root); |
1e8e9920 | 7095 | |
49a70175 | 7096 | stmt = gimple_build_switch (vmain, u, label_vec); |
75a70cf9 | 7097 | gsi_insert_after (&switch_si, stmt, GSI_SAME_STMT); |
7098 | gsi_remove (&switch_si, true); | |
75a70cf9 | 7099 | |
7100 | si = gsi_start_bb (default_bb); | |
b9a16870 | 7101 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_TRAP), 0); |
75a70cf9 | 7102 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); |
773c5ba7 | 7103 | |
ac6e3339 | 7104 | if (exit_reachable) |
03ed154b | 7105 | { |
b9a16870 | 7106 | tree bfn_decl; |
7107 | ||
ac6e3339 | 7108 | /* Code to get the next section goes in L1_BB. */ |
75a70cf9 | 7109 | si = gsi_last_bb (l1_bb); |
7110 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CONTINUE); | |
1e8e9920 | 7111 | |
b9a16870 | 7112 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); |
7113 | stmt = gimple_build_call (bfn_decl, 0); | |
75a70cf9 | 7114 | gimple_call_set_lhs (stmt, vnext); |
7115 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
7116 | gsi_remove (&si, true); | |
773c5ba7 | 7117 | |
ac6e3339 | 7118 | single_succ_edge (l1_bb)->flags = EDGE_FALLTHRU; |
03ed154b | 7119 | } |
773c5ba7 | 7120 | |
d244d9de | 7121 | /* Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB. */ |
7122 | si = gsi_last_bb (l2_bb); | |
7123 | if (gimple_omp_return_nowait_p (gsi_stmt (si))) | |
7124 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_NOWAIT); | |
bc7bff74 | 7125 | else if (gimple_omp_return_lhs (gsi_stmt (si))) |
7126 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_CANCEL); | |
d244d9de | 7127 | else |
7128 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END); | |
7129 | stmt = gimple_build_call (t, 0); | |
bc7bff74 | 7130 | if (gimple_omp_return_lhs (gsi_stmt (si))) |
7131 | gimple_call_set_lhs (stmt, gimple_omp_return_lhs (gsi_stmt (si))); | |
d244d9de | 7132 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); |
7133 | gsi_remove (&si, true); | |
7134 | ||
79acaae1 | 7135 | set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb); |
773c5ba7 | 7136 | } |
1e8e9920 | 7137 | |
1e8e9920 | 7138 | |
61e47ac8 | 7139 | /* Expand code for an OpenMP single directive. We've already expanded |
7140 | much of the code, here we simply place the GOMP_barrier call. */ | |
7141 | ||
7142 | static void | |
7143 | expand_omp_single (struct omp_region *region) | |
7144 | { | |
7145 | basic_block entry_bb, exit_bb; | |
75a70cf9 | 7146 | gimple_stmt_iterator si; |
61e47ac8 | 7147 | |
7148 | entry_bb = region->entry; | |
7149 | exit_bb = region->exit; | |
7150 | ||
75a70cf9 | 7151 | si = gsi_last_bb (entry_bb); |
75a70cf9 | 7152 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE); |
7153 | gsi_remove (&si, true); | |
61e47ac8 | 7154 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; |
7155 | ||
75a70cf9 | 7156 | si = gsi_last_bb (exit_bb); |
bc7bff74 | 7157 | if (!gimple_omp_return_nowait_p (gsi_stmt (si))) |
7158 | { | |
7159 | tree t = gimple_omp_return_lhs (gsi_stmt (si)); | |
7160 | gsi_insert_after (&si, build_omp_barrier (t), GSI_SAME_STMT); | |
7161 | } | |
75a70cf9 | 7162 | gsi_remove (&si, true); |
61e47ac8 | 7163 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; |
7164 | } | |
7165 | ||
7166 | ||
7167 | /* Generic expansion for OpenMP synchronization directives: master, | |
7168 | ordered and critical. All we need to do here is remove the entry | |
7169 | and exit markers for REGION. */ | |
773c5ba7 | 7170 | |
7171 | static void | |
7172 | expand_omp_synch (struct omp_region *region) | |
7173 | { | |
7174 | basic_block entry_bb, exit_bb; | |
75a70cf9 | 7175 | gimple_stmt_iterator si; |
773c5ba7 | 7176 | |
61e47ac8 | 7177 | entry_bb = region->entry; |
7178 | exit_bb = region->exit; | |
773c5ba7 | 7179 | |
75a70cf9 | 7180 | si = gsi_last_bb (entry_bb); |
7181 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE | |
7182 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASTER | |
bc7bff74 | 7183 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TASKGROUP |
75a70cf9 | 7184 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ORDERED |
bc7bff74 | 7185 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CRITICAL |
7186 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS); | |
75a70cf9 | 7187 | gsi_remove (&si, true); |
773c5ba7 | 7188 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; |
7189 | ||
03ed154b | 7190 | if (exit_bb) |
7191 | { | |
75a70cf9 | 7192 | si = gsi_last_bb (exit_bb); |
7193 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); | |
7194 | gsi_remove (&si, true); | |
03ed154b | 7195 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; |
7196 | } | |
773c5ba7 | 7197 | } |
1e8e9920 | 7198 | |
2169f33b | 7199 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
7200 | operation as a normal volatile load. */ | |
7201 | ||
7202 | static bool | |
3ec11c49 | 7203 | expand_omp_atomic_load (basic_block load_bb, tree addr, |
7204 | tree loaded_val, int index) | |
2169f33b | 7205 | { |
3ec11c49 | 7206 | enum built_in_function tmpbase; |
7207 | gimple_stmt_iterator gsi; | |
7208 | basic_block store_bb; | |
7209 | location_t loc; | |
7210 | gimple stmt; | |
7211 | tree decl, call, type, itype; | |
7212 | ||
7213 | gsi = gsi_last_bb (load_bb); | |
7214 | stmt = gsi_stmt (gsi); | |
7215 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
7216 | loc = gimple_location (stmt); | |
7217 | ||
7218 | /* ??? If the target does not implement atomic_load_optab[mode], and mode | |
7219 | is smaller than word size, then expand_atomic_load assumes that the load | |
7220 | is atomic. We could avoid the builtin entirely in this case. */ | |
7221 | ||
7222 | tmpbase = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
7223 | decl = builtin_decl_explicit (tmpbase); | |
7224 | if (decl == NULL_TREE) | |
7225 | return false; | |
7226 | ||
7227 | type = TREE_TYPE (loaded_val); | |
7228 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
7229 | ||
7230 | call = build_call_expr_loc (loc, decl, 2, addr, | |
bc7bff74 | 7231 | build_int_cst (NULL, |
7232 | gimple_omp_atomic_seq_cst_p (stmt) | |
7233 | ? MEMMODEL_SEQ_CST | |
7234 | : MEMMODEL_RELAXED)); | |
3ec11c49 | 7235 | if (!useless_type_conversion_p (type, itype)) |
7236 | call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
7237 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
7238 | ||
7239 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
7240 | gsi_remove (&gsi, true); | |
7241 | ||
7242 | store_bb = single_succ (load_bb); | |
7243 | gsi = gsi_last_bb (store_bb); | |
7244 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); | |
7245 | gsi_remove (&gsi, true); | |
7246 | ||
7247 | if (gimple_in_ssa_p (cfun)) | |
7248 | update_ssa (TODO_update_ssa_no_phi); | |
7249 | ||
7250 | return true; | |
2169f33b | 7251 | } |
7252 | ||
7253 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
7254 | operation as a normal volatile store. */ | |
7255 | ||
7256 | static bool | |
3ec11c49 | 7257 | expand_omp_atomic_store (basic_block load_bb, tree addr, |
7258 | tree loaded_val, tree stored_val, int index) | |
2169f33b | 7259 | { |
3ec11c49 | 7260 | enum built_in_function tmpbase; |
7261 | gimple_stmt_iterator gsi; | |
7262 | basic_block store_bb = single_succ (load_bb); | |
7263 | location_t loc; | |
7264 | gimple stmt; | |
7265 | tree decl, call, type, itype; | |
7266 | enum machine_mode imode; | |
7267 | bool exchange; | |
7268 | ||
7269 | gsi = gsi_last_bb (load_bb); | |
7270 | stmt = gsi_stmt (gsi); | |
7271 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
7272 | ||
7273 | /* If the load value is needed, then this isn't a store but an exchange. */ | |
7274 | exchange = gimple_omp_atomic_need_value_p (stmt); | |
7275 | ||
7276 | gsi = gsi_last_bb (store_bb); | |
7277 | stmt = gsi_stmt (gsi); | |
7278 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_STORE); | |
7279 | loc = gimple_location (stmt); | |
7280 | ||
7281 | /* ??? If the target does not implement atomic_store_optab[mode], and mode | |
7282 | is smaller than word size, then expand_atomic_store assumes that the store | |
7283 | is atomic. We could avoid the builtin entirely in this case. */ | |
7284 | ||
7285 | tmpbase = (exchange ? BUILT_IN_ATOMIC_EXCHANGE_N : BUILT_IN_ATOMIC_STORE_N); | |
7286 | tmpbase = (enum built_in_function) ((int) tmpbase + index + 1); | |
7287 | decl = builtin_decl_explicit (tmpbase); | |
7288 | if (decl == NULL_TREE) | |
7289 | return false; | |
7290 | ||
7291 | type = TREE_TYPE (stored_val); | |
7292 | ||
7293 | /* Dig out the type of the function's second argument. */ | |
7294 | itype = TREE_TYPE (decl); | |
7295 | itype = TYPE_ARG_TYPES (itype); | |
7296 | itype = TREE_CHAIN (itype); | |
7297 | itype = TREE_VALUE (itype); | |
7298 | imode = TYPE_MODE (itype); | |
7299 | ||
7300 | if (exchange && !can_atomic_exchange_p (imode, true)) | |
7301 | return false; | |
7302 | ||
7303 | if (!useless_type_conversion_p (itype, type)) | |
7304 | stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val); | |
7305 | call = build_call_expr_loc (loc, decl, 3, addr, stored_val, | |
bc7bff74 | 7306 | build_int_cst (NULL, |
7307 | gimple_omp_atomic_seq_cst_p (stmt) | |
7308 | ? MEMMODEL_SEQ_CST | |
7309 | : MEMMODEL_RELAXED)); | |
3ec11c49 | 7310 | if (exchange) |
7311 | { | |
7312 | if (!useless_type_conversion_p (type, itype)) | |
7313 | call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
7314 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
7315 | } | |
7316 | ||
7317 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
7318 | gsi_remove (&gsi, true); | |
7319 | ||
7320 | /* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */ | |
7321 | gsi = gsi_last_bb (load_bb); | |
7322 | gsi_remove (&gsi, true); | |
7323 | ||
7324 | if (gimple_in_ssa_p (cfun)) | |
7325 | update_ssa (TODO_update_ssa_no_phi); | |
7326 | ||
7327 | return true; | |
2169f33b | 7328 | } |
7329 | ||
cb7f680b | 7330 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
1cd6e20d | 7331 | operation as a __atomic_fetch_op builtin. INDEX is log2 of the |
cb7f680b | 7332 | size of the data type, and thus usable to find the index of the builtin |
7333 | decl. Returns false if the expression is not of the proper form. */ | |
7334 | ||
7335 | static bool | |
7336 | expand_omp_atomic_fetch_op (basic_block load_bb, | |
7337 | tree addr, tree loaded_val, | |
7338 | tree stored_val, int index) | |
7339 | { | |
b9a16870 | 7340 | enum built_in_function oldbase, newbase, tmpbase; |
cb7f680b | 7341 | tree decl, itype, call; |
2169f33b | 7342 | tree lhs, rhs; |
cb7f680b | 7343 | basic_block store_bb = single_succ (load_bb); |
75a70cf9 | 7344 | gimple_stmt_iterator gsi; |
7345 | gimple stmt; | |
389dd41b | 7346 | location_t loc; |
1cd6e20d | 7347 | enum tree_code code; |
2169f33b | 7348 | bool need_old, need_new; |
1cd6e20d | 7349 | enum machine_mode imode; |
bc7bff74 | 7350 | bool seq_cst; |
cb7f680b | 7351 | |
7352 | /* We expect to find the following sequences: | |
48e1416a | 7353 | |
cb7f680b | 7354 | load_bb: |
75a70cf9 | 7355 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) |
cb7f680b | 7356 | |
7357 | store_bb: | |
7358 | val = tmp OP something; (or: something OP tmp) | |
48e1416a | 7359 | GIMPLE_OMP_STORE (val) |
cb7f680b | 7360 | |
48e1416a | 7361 | ???FIXME: Allow a more flexible sequence. |
cb7f680b | 7362 | Perhaps use data flow to pick the statements. |
48e1416a | 7363 | |
cb7f680b | 7364 | */ |
7365 | ||
75a70cf9 | 7366 | gsi = gsi_after_labels (store_bb); |
7367 | stmt = gsi_stmt (gsi); | |
389dd41b | 7368 | loc = gimple_location (stmt); |
75a70cf9 | 7369 | if (!is_gimple_assign (stmt)) |
cb7f680b | 7370 | return false; |
75a70cf9 | 7371 | gsi_next (&gsi); |
7372 | if (gimple_code (gsi_stmt (gsi)) != GIMPLE_OMP_ATOMIC_STORE) | |
cb7f680b | 7373 | return false; |
2169f33b | 7374 | need_new = gimple_omp_atomic_need_value_p (gsi_stmt (gsi)); |
7375 | need_old = gimple_omp_atomic_need_value_p (last_stmt (load_bb)); | |
bc7bff74 | 7376 | seq_cst = gimple_omp_atomic_seq_cst_p (last_stmt (load_bb)); |
2169f33b | 7377 | gcc_checking_assert (!need_old || !need_new); |
cb7f680b | 7378 | |
75a70cf9 | 7379 | if (!operand_equal_p (gimple_assign_lhs (stmt), stored_val, 0)) |
cb7f680b | 7380 | return false; |
7381 | ||
cb7f680b | 7382 | /* Check for one of the supported fetch-op operations. */ |
1cd6e20d | 7383 | code = gimple_assign_rhs_code (stmt); |
7384 | switch (code) | |
cb7f680b | 7385 | { |
7386 | case PLUS_EXPR: | |
7387 | case POINTER_PLUS_EXPR: | |
1cd6e20d | 7388 | oldbase = BUILT_IN_ATOMIC_FETCH_ADD_N; |
7389 | newbase = BUILT_IN_ATOMIC_ADD_FETCH_N; | |
cb7f680b | 7390 | break; |
7391 | case MINUS_EXPR: | |
1cd6e20d | 7392 | oldbase = BUILT_IN_ATOMIC_FETCH_SUB_N; |
7393 | newbase = BUILT_IN_ATOMIC_SUB_FETCH_N; | |
cb7f680b | 7394 | break; |
7395 | case BIT_AND_EXPR: | |
1cd6e20d | 7396 | oldbase = BUILT_IN_ATOMIC_FETCH_AND_N; |
7397 | newbase = BUILT_IN_ATOMIC_AND_FETCH_N; | |
cb7f680b | 7398 | break; |
7399 | case BIT_IOR_EXPR: | |
1cd6e20d | 7400 | oldbase = BUILT_IN_ATOMIC_FETCH_OR_N; |
7401 | newbase = BUILT_IN_ATOMIC_OR_FETCH_N; | |
cb7f680b | 7402 | break; |
7403 | case BIT_XOR_EXPR: | |
1cd6e20d | 7404 | oldbase = BUILT_IN_ATOMIC_FETCH_XOR_N; |
7405 | newbase = BUILT_IN_ATOMIC_XOR_FETCH_N; | |
cb7f680b | 7406 | break; |
7407 | default: | |
7408 | return false; | |
7409 | } | |
1cd6e20d | 7410 | |
cb7f680b | 7411 | /* Make sure the expression is of the proper form. */ |
75a70cf9 | 7412 | if (operand_equal_p (gimple_assign_rhs1 (stmt), loaded_val, 0)) |
7413 | rhs = gimple_assign_rhs2 (stmt); | |
7414 | else if (commutative_tree_code (gimple_assign_rhs_code (stmt)) | |
7415 | && operand_equal_p (gimple_assign_rhs2 (stmt), loaded_val, 0)) | |
7416 | rhs = gimple_assign_rhs1 (stmt); | |
cb7f680b | 7417 | else |
7418 | return false; | |
7419 | ||
b9a16870 | 7420 | tmpbase = ((enum built_in_function) |
7421 | ((need_new ? newbase : oldbase) + index + 1)); | |
7422 | decl = builtin_decl_explicit (tmpbase); | |
0f94f46b | 7423 | if (decl == NULL_TREE) |
7424 | return false; | |
cb7f680b | 7425 | itype = TREE_TYPE (TREE_TYPE (decl)); |
1cd6e20d | 7426 | imode = TYPE_MODE (itype); |
cb7f680b | 7427 | |
1cd6e20d | 7428 | /* We could test all of the various optabs involved, but the fact of the |
7429 | matter is that (with the exception of i486 vs i586 and xadd) all targets | |
7430 | that support any atomic operaton optab also implements compare-and-swap. | |
7431 | Let optabs.c take care of expanding any compare-and-swap loop. */ | |
29139cdc | 7432 | if (!can_compare_and_swap_p (imode, true)) |
cb7f680b | 7433 | return false; |
7434 | ||
75a70cf9 | 7435 | gsi = gsi_last_bb (load_bb); |
7436 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_LOAD); | |
1cd6e20d | 7437 | |
7438 | /* OpenMP does not imply any barrier-like semantics on its atomic ops. | |
7439 | It only requires that the operation happen atomically. Thus we can | |
7440 | use the RELAXED memory model. */ | |
7441 | call = build_call_expr_loc (loc, decl, 3, addr, | |
7442 | fold_convert_loc (loc, itype, rhs), | |
bc7bff74 | 7443 | build_int_cst (NULL, |
7444 | seq_cst ? MEMMODEL_SEQ_CST | |
7445 | : MEMMODEL_RELAXED)); | |
1cd6e20d | 7446 | |
2169f33b | 7447 | if (need_old || need_new) |
7448 | { | |
7449 | lhs = need_old ? loaded_val : stored_val; | |
7450 | call = fold_convert_loc (loc, TREE_TYPE (lhs), call); | |
7451 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, lhs, call); | |
7452 | } | |
7453 | else | |
7454 | call = fold_convert_loc (loc, void_type_node, call); | |
75a70cf9 | 7455 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); |
7456 | gsi_remove (&gsi, true); | |
cb7f680b | 7457 | |
75a70cf9 | 7458 | gsi = gsi_last_bb (store_bb); |
7459 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); | |
7460 | gsi_remove (&gsi, true); | |
7461 | gsi = gsi_last_bb (store_bb); | |
7462 | gsi_remove (&gsi, true); | |
cb7f680b | 7463 | |
7464 | if (gimple_in_ssa_p (cfun)) | |
7465 | update_ssa (TODO_update_ssa_no_phi); | |
7466 | ||
7467 | return true; | |
7468 | } | |
7469 | ||
7470 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
7471 | ||
7472 | oldval = *addr; | |
7473 | repeat: | |
7474 | newval = rhs; // with oldval replacing *addr in rhs | |
7475 | oldval = __sync_val_compare_and_swap (addr, oldval, newval); | |
7476 | if (oldval != newval) | |
7477 | goto repeat; | |
7478 | ||
7479 | INDEX is log2 of the size of the data type, and thus usable to find the | |
7480 | index of the builtin decl. */ | |
7481 | ||
7482 | static bool | |
7483 | expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb, | |
7484 | tree addr, tree loaded_val, tree stored_val, | |
7485 | int index) | |
7486 | { | |
790368c5 | 7487 | tree loadedi, storedi, initial, new_storedi, old_vali; |
cb7f680b | 7488 | tree type, itype, cmpxchg, iaddr; |
75a70cf9 | 7489 | gimple_stmt_iterator si; |
cb7f680b | 7490 | basic_block loop_header = single_succ (load_bb); |
75a70cf9 | 7491 | gimple phi, stmt; |
cb7f680b | 7492 | edge e; |
b9a16870 | 7493 | enum built_in_function fncode; |
cb7f680b | 7494 | |
1cd6e20d | 7495 | /* ??? We need a non-pointer interface to __atomic_compare_exchange in |
7496 | order to use the RELAXED memory model effectively. */ | |
b9a16870 | 7497 | fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N |
7498 | + index + 1); | |
7499 | cmpxchg = builtin_decl_explicit (fncode); | |
0f94f46b | 7500 | if (cmpxchg == NULL_TREE) |
7501 | return false; | |
cb7f680b | 7502 | type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr))); |
7503 | itype = TREE_TYPE (TREE_TYPE (cmpxchg)); | |
7504 | ||
29139cdc | 7505 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true)) |
cb7f680b | 7506 | return false; |
7507 | ||
75a70cf9 | 7508 | /* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */ |
7509 | si = gsi_last_bb (load_bb); | |
7510 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); | |
7511 | ||
790368c5 | 7512 | /* For floating-point values, we'll need to view-convert them to integers |
7513 | so that we can perform the atomic compare and swap. Simplify the | |
7514 | following code by always setting up the "i"ntegral variables. */ | |
7515 | if (!INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type)) | |
7516 | { | |
75a70cf9 | 7517 | tree iaddr_val; |
7518 | ||
072f7ab1 | 7519 | iaddr = create_tmp_reg (build_pointer_type_for_mode (itype, ptr_mode, |
7520 | true), NULL); | |
75a70cf9 | 7521 | iaddr_val |
7522 | = force_gimple_operand_gsi (&si, | |
7523 | fold_convert (TREE_TYPE (iaddr), addr), | |
7524 | false, NULL_TREE, true, GSI_SAME_STMT); | |
7525 | stmt = gimple_build_assign (iaddr, iaddr_val); | |
7526 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
790368c5 | 7527 | loadedi = create_tmp_var (itype, NULL); |
7528 | if (gimple_in_ssa_p (cfun)) | |
b03e5397 | 7529 | loadedi = make_ssa_name (loadedi, NULL); |
790368c5 | 7530 | } |
7531 | else | |
7532 | { | |
7533 | iaddr = addr; | |
7534 | loadedi = loaded_val; | |
7535 | } | |
75a70cf9 | 7536 | |
182cf5a9 | 7537 | initial |
7538 | = force_gimple_operand_gsi (&si, | |
7539 | build2 (MEM_REF, TREE_TYPE (TREE_TYPE (iaddr)), | |
7540 | iaddr, | |
7541 | build_int_cst (TREE_TYPE (iaddr), 0)), | |
7542 | true, NULL_TREE, true, GSI_SAME_STMT); | |
790368c5 | 7543 | |
7544 | /* Move the value to the LOADEDI temporary. */ | |
cb7f680b | 7545 | if (gimple_in_ssa_p (cfun)) |
7546 | { | |
75a70cf9 | 7547 | gcc_assert (gimple_seq_empty_p (phi_nodes (loop_header))); |
790368c5 | 7548 | phi = create_phi_node (loadedi, loop_header); |
cb7f680b | 7549 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)), |
7550 | initial); | |
7551 | } | |
7552 | else | |
75a70cf9 | 7553 | gsi_insert_before (&si, |
7554 | gimple_build_assign (loadedi, initial), | |
7555 | GSI_SAME_STMT); | |
790368c5 | 7556 | if (loadedi != loaded_val) |
7557 | { | |
75a70cf9 | 7558 | gimple_stmt_iterator gsi2; |
7559 | tree x; | |
790368c5 | 7560 | |
7561 | x = build1 (VIEW_CONVERT_EXPR, type, loadedi); | |
75a70cf9 | 7562 | gsi2 = gsi_start_bb (loop_header); |
790368c5 | 7563 | if (gimple_in_ssa_p (cfun)) |
7564 | { | |
75a70cf9 | 7565 | gimple stmt; |
7566 | x = force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
7567 | true, GSI_SAME_STMT); | |
7568 | stmt = gimple_build_assign (loaded_val, x); | |
7569 | gsi_insert_before (&gsi2, stmt, GSI_SAME_STMT); | |
790368c5 | 7570 | } |
7571 | else | |
7572 | { | |
75a70cf9 | 7573 | x = build2 (MODIFY_EXPR, TREE_TYPE (loaded_val), loaded_val, x); |
7574 | force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
7575 | true, GSI_SAME_STMT); | |
790368c5 | 7576 | } |
7577 | } | |
75a70cf9 | 7578 | gsi_remove (&si, true); |
cb7f680b | 7579 | |
75a70cf9 | 7580 | si = gsi_last_bb (store_bb); |
7581 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); | |
cb7f680b | 7582 | |
790368c5 | 7583 | if (iaddr == addr) |
7584 | storedi = stored_val; | |
cb7f680b | 7585 | else |
790368c5 | 7586 | storedi = |
75a70cf9 | 7587 | force_gimple_operand_gsi (&si, |
790368c5 | 7588 | build1 (VIEW_CONVERT_EXPR, itype, |
7589 | stored_val), true, NULL_TREE, true, | |
75a70cf9 | 7590 | GSI_SAME_STMT); |
cb7f680b | 7591 | |
7592 | /* Build the compare&swap statement. */ | |
7593 | new_storedi = build_call_expr (cmpxchg, 3, iaddr, loadedi, storedi); | |
75a70cf9 | 7594 | new_storedi = force_gimple_operand_gsi (&si, |
87f9ffa4 | 7595 | fold_convert (TREE_TYPE (loadedi), |
7596 | new_storedi), | |
cb7f680b | 7597 | true, NULL_TREE, |
75a70cf9 | 7598 | true, GSI_SAME_STMT); |
cb7f680b | 7599 | |
7600 | if (gimple_in_ssa_p (cfun)) | |
7601 | old_vali = loadedi; | |
7602 | else | |
7603 | { | |
87f9ffa4 | 7604 | old_vali = create_tmp_var (TREE_TYPE (loadedi), NULL); |
75a70cf9 | 7605 | stmt = gimple_build_assign (old_vali, loadedi); |
7606 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
cb7f680b | 7607 | |
75a70cf9 | 7608 | stmt = gimple_build_assign (loadedi, new_storedi); |
7609 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
cb7f680b | 7610 | } |
7611 | ||
7612 | /* Note that we always perform the comparison as an integer, even for | |
48e1416a | 7613 | floating point. This allows the atomic operation to properly |
cb7f680b | 7614 | succeed even with NaNs and -0.0. */ |
75a70cf9 | 7615 | stmt = gimple_build_cond_empty |
7616 | (build2 (NE_EXPR, boolean_type_node, | |
7617 | new_storedi, old_vali)); | |
7618 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
cb7f680b | 7619 | |
7620 | /* Update cfg. */ | |
7621 | e = single_succ_edge (store_bb); | |
7622 | e->flags &= ~EDGE_FALLTHRU; | |
7623 | e->flags |= EDGE_FALSE_VALUE; | |
7624 | ||
7625 | e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE); | |
7626 | ||
790368c5 | 7627 | /* Copy the new value to loadedi (we already did that before the condition |
cb7f680b | 7628 | if we are not in SSA). */ |
7629 | if (gimple_in_ssa_p (cfun)) | |
7630 | { | |
75a70cf9 | 7631 | phi = gimple_seq_first_stmt (phi_nodes (loop_header)); |
790368c5 | 7632 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_storedi); |
cb7f680b | 7633 | } |
7634 | ||
75a70cf9 | 7635 | /* Remove GIMPLE_OMP_ATOMIC_STORE. */ |
7636 | gsi_remove (&si, true); | |
cb7f680b | 7637 | |
04c2922b | 7638 | struct loop *loop = alloc_loop (); |
7639 | loop->header = loop_header; | |
5f037457 | 7640 | loop->latch = store_bb; |
04c2922b | 7641 | add_loop (loop, loop_header->loop_father); |
7642 | ||
cb7f680b | 7643 | if (gimple_in_ssa_p (cfun)) |
7644 | update_ssa (TODO_update_ssa_no_phi); | |
7645 | ||
7646 | return true; | |
7647 | } | |
7648 | ||
7649 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
7650 | ||
7651 | GOMP_atomic_start (); | |
7652 | *addr = rhs; | |
7653 | GOMP_atomic_end (); | |
7654 | ||
7655 | The result is not globally atomic, but works so long as all parallel | |
7656 | references are within #pragma omp atomic directives. According to | |
7657 | responses received from omp@openmp.org, appears to be within spec. | |
7658 | Which makes sense, since that's how several other compilers handle | |
48e1416a | 7659 | this situation as well. |
75a70cf9 | 7660 | LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're |
7661 | expanding. STORED_VAL is the operand of the matching | |
7662 | GIMPLE_OMP_ATOMIC_STORE. | |
cb7f680b | 7663 | |
48e1416a | 7664 | We replace |
7665 | GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with | |
cb7f680b | 7666 | loaded_val = *addr; |
7667 | ||
7668 | and replace | |
3ec11c49 | 7669 | GIMPLE_OMP_ATOMIC_STORE (stored_val) with |
48e1416a | 7670 | *addr = stored_val; |
cb7f680b | 7671 | */ |
7672 | ||
7673 | static bool | |
7674 | expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb, | |
7675 | tree addr, tree loaded_val, tree stored_val) | |
7676 | { | |
75a70cf9 | 7677 | gimple_stmt_iterator si; |
7678 | gimple stmt; | |
cb7f680b | 7679 | tree t; |
7680 | ||
75a70cf9 | 7681 | si = gsi_last_bb (load_bb); |
7682 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); | |
cb7f680b | 7683 | |
b9a16870 | 7684 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START); |
414c3a2c | 7685 | t = build_call_expr (t, 0); |
75a70cf9 | 7686 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); |
cb7f680b | 7687 | |
182cf5a9 | 7688 | stmt = gimple_build_assign (loaded_val, build_simple_mem_ref (addr)); |
75a70cf9 | 7689 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
7690 | gsi_remove (&si, true); | |
cb7f680b | 7691 | |
75a70cf9 | 7692 | si = gsi_last_bb (store_bb); |
7693 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); | |
cb7f680b | 7694 | |
182cf5a9 | 7695 | stmt = gimple_build_assign (build_simple_mem_ref (unshare_expr (addr)), |
7696 | stored_val); | |
75a70cf9 | 7697 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
cb7f680b | 7698 | |
b9a16870 | 7699 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END); |
414c3a2c | 7700 | t = build_call_expr (t, 0); |
75a70cf9 | 7701 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); |
7702 | gsi_remove (&si, true); | |
cb7f680b | 7703 | |
7704 | if (gimple_in_ssa_p (cfun)) | |
7705 | update_ssa (TODO_update_ssa_no_phi); | |
7706 | return true; | |
7707 | } | |
7708 | ||
48e1416a | 7709 | /* Expand an GIMPLE_OMP_ATOMIC statement. We try to expand |
7710 | using expand_omp_atomic_fetch_op. If it failed, we try to | |
cb7f680b | 7711 | call expand_omp_atomic_pipeline, and if it fails too, the |
7712 | ultimate fallback is wrapping the operation in a mutex | |
48e1416a | 7713 | (expand_omp_atomic_mutex). REGION is the atomic region built |
7714 | by build_omp_regions_1(). */ | |
cb7f680b | 7715 | |
7716 | static void | |
7717 | expand_omp_atomic (struct omp_region *region) | |
7718 | { | |
7719 | basic_block load_bb = region->entry, store_bb = region->exit; | |
75a70cf9 | 7720 | gimple load = last_stmt (load_bb), store = last_stmt (store_bb); |
7721 | tree loaded_val = gimple_omp_atomic_load_lhs (load); | |
7722 | tree addr = gimple_omp_atomic_load_rhs (load); | |
7723 | tree stored_val = gimple_omp_atomic_store_val (store); | |
cb7f680b | 7724 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr))); |
7725 | HOST_WIDE_INT index; | |
7726 | ||
7727 | /* Make sure the type is one of the supported sizes. */ | |
6a0712d4 | 7728 | index = tree_to_uhwi (TYPE_SIZE_UNIT (type)); |
cb7f680b | 7729 | index = exact_log2 (index); |
7730 | if (index >= 0 && index <= 4) | |
7731 | { | |
7732 | unsigned int align = TYPE_ALIGN_UNIT (type); | |
7733 | ||
7734 | /* __sync builtins require strict data alignment. */ | |
dcf7024c | 7735 | if (exact_log2 (align) >= index) |
cb7f680b | 7736 | { |
3ec11c49 | 7737 | /* Atomic load. */ |
2169f33b | 7738 | if (loaded_val == stored_val |
7739 | && (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_INT | |
7740 | || GET_MODE_CLASS (TYPE_MODE (type)) == MODE_FLOAT) | |
7741 | && GET_MODE_BITSIZE (TYPE_MODE (type)) <= BITS_PER_WORD | |
3ec11c49 | 7742 | && expand_omp_atomic_load (load_bb, addr, loaded_val, index)) |
2169f33b | 7743 | return; |
7744 | ||
3ec11c49 | 7745 | /* Atomic store. */ |
2169f33b | 7746 | if ((GET_MODE_CLASS (TYPE_MODE (type)) == MODE_INT |
7747 | || GET_MODE_CLASS (TYPE_MODE (type)) == MODE_FLOAT) | |
7748 | && GET_MODE_BITSIZE (TYPE_MODE (type)) <= BITS_PER_WORD | |
7749 | && store_bb == single_succ (load_bb) | |
7750 | && first_stmt (store_bb) == store | |
3ec11c49 | 7751 | && expand_omp_atomic_store (load_bb, addr, loaded_val, |
7752 | stored_val, index)) | |
2169f33b | 7753 | return; |
7754 | ||
cb7f680b | 7755 | /* When possible, use specialized atomic update functions. */ |
7756 | if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type)) | |
3ec11c49 | 7757 | && store_bb == single_succ (load_bb) |
7758 | && expand_omp_atomic_fetch_op (load_bb, addr, | |
7759 | loaded_val, stored_val, index)) | |
7760 | return; | |
cb7f680b | 7761 | |
7762 | /* If we don't have specialized __sync builtins, try and implement | |
7763 | as a compare and swap loop. */ | |
7764 | if (expand_omp_atomic_pipeline (load_bb, store_bb, addr, | |
7765 | loaded_val, stored_val, index)) | |
7766 | return; | |
7767 | } | |
7768 | } | |
7769 | ||
7770 | /* The ultimate fallback is wrapping the operation in a mutex. */ | |
7771 | expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val); | |
7772 | } | |
7773 | ||
1e8e9920 | 7774 | |
bc7bff74 | 7775 | /* Expand the OpenMP target{, data, update} directive starting at REGION. */ |
7776 | ||
7777 | static void | |
7778 | expand_omp_target (struct omp_region *region) | |
7779 | { | |
7780 | basic_block entry_bb, exit_bb, new_bb; | |
7781 | struct function *child_cfun = NULL; | |
7782 | tree child_fn = NULL_TREE, block, t; | |
7783 | gimple_stmt_iterator gsi; | |
7784 | gimple entry_stmt, stmt; | |
7785 | edge e; | |
7786 | ||
7787 | entry_stmt = last_stmt (region->entry); | |
7788 | new_bb = region->entry; | |
7789 | int kind = gimple_omp_target_kind (entry_stmt); | |
7790 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
7791 | { | |
7792 | child_fn = gimple_omp_target_child_fn (entry_stmt); | |
7793 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
7794 | } | |
7795 | ||
7796 | entry_bb = region->entry; | |
7797 | exit_bb = region->exit; | |
7798 | ||
7799 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
7800 | { | |
7801 | unsigned srcidx, dstidx, num; | |
7802 | ||
7803 | /* If the target region needs data sent from the parent | |
7804 | function, then the very first statement (except possible | |
7805 | tree profile counter updates) of the parallel body | |
7806 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
7807 | &.OMP_DATA_O is passed as an argument to the child function, | |
7808 | we need to replace it with the argument as seen by the child | |
7809 | function. | |
7810 | ||
7811 | In most cases, this will end up being the identity assignment | |
7812 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had | |
7813 | a function call that has been inlined, the original PARM_DECL | |
7814 | .OMP_DATA_I may have been converted into a different local | |
7815 | variable. In which case, we need to keep the assignment. */ | |
7816 | if (gimple_omp_target_data_arg (entry_stmt)) | |
7817 | { | |
7818 | basic_block entry_succ_bb = single_succ (entry_bb); | |
7819 | gimple_stmt_iterator gsi; | |
7820 | tree arg; | |
7821 | gimple tgtcopy_stmt = NULL; | |
7822 | tree sender | |
7823 | = TREE_VEC_ELT (gimple_omp_target_data_arg (entry_stmt), 0); | |
7824 | ||
7825 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
7826 | { | |
7827 | gcc_assert (!gsi_end_p (gsi)); | |
7828 | stmt = gsi_stmt (gsi); | |
7829 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
7830 | continue; | |
7831 | ||
7832 | if (gimple_num_ops (stmt) == 2) | |
7833 | { | |
7834 | tree arg = gimple_assign_rhs1 (stmt); | |
7835 | ||
7836 | /* We're ignoring the subcode because we're | |
7837 | effectively doing a STRIP_NOPS. */ | |
7838 | ||
7839 | if (TREE_CODE (arg) == ADDR_EXPR | |
7840 | && TREE_OPERAND (arg, 0) == sender) | |
7841 | { | |
7842 | tgtcopy_stmt = stmt; | |
7843 | break; | |
7844 | } | |
7845 | } | |
7846 | } | |
7847 | ||
7848 | gcc_assert (tgtcopy_stmt != NULL); | |
7849 | arg = DECL_ARGUMENTS (child_fn); | |
7850 | ||
7851 | gcc_assert (gimple_assign_lhs (tgtcopy_stmt) == arg); | |
7852 | gsi_remove (&gsi, true); | |
7853 | } | |
7854 | ||
7855 | /* Declare local variables needed in CHILD_CFUN. */ | |
7856 | block = DECL_INITIAL (child_fn); | |
7857 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
7858 | /* The gimplifier could record temporaries in target block | |
7859 | rather than in containing function's local_decls chain, | |
7860 | which would mean cgraph missed finalizing them. Do it now. */ | |
7861 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
7862 | if (TREE_CODE (t) == VAR_DECL | |
7863 | && TREE_STATIC (t) | |
7864 | && !DECL_EXTERNAL (t)) | |
7865 | varpool_finalize_decl (t); | |
7866 | DECL_SAVED_TREE (child_fn) = NULL; | |
7867 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
7868 | gimple_set_body (child_fn, NULL); | |
7869 | TREE_USED (block) = 1; | |
7870 | ||
7871 | /* Reset DECL_CONTEXT on function arguments. */ | |
7872 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
7873 | DECL_CONTEXT (t) = child_fn; | |
7874 | ||
7875 | /* Split ENTRY_BB at GIMPLE_OMP_TARGET, | |
7876 | so that it can be moved to the child function. */ | |
7877 | gsi = gsi_last_bb (entry_bb); | |
7878 | stmt = gsi_stmt (gsi); | |
7879 | gcc_assert (stmt && gimple_code (stmt) == GIMPLE_OMP_TARGET | |
7880 | && gimple_omp_target_kind (stmt) | |
7881 | == GF_OMP_TARGET_KIND_REGION); | |
7882 | gsi_remove (&gsi, true); | |
7883 | e = split_block (entry_bb, stmt); | |
7884 | entry_bb = e->dest; | |
7885 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
7886 | ||
7887 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ | |
7888 | if (exit_bb) | |
7889 | { | |
7890 | gsi = gsi_last_bb (exit_bb); | |
7891 | gcc_assert (!gsi_end_p (gsi) | |
7892 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
7893 | stmt = gimple_build_return (NULL); | |
7894 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
7895 | gsi_remove (&gsi, true); | |
7896 | } | |
7897 | ||
7898 | /* Move the target region into CHILD_CFUN. */ | |
7899 | ||
7900 | block = gimple_block (entry_stmt); | |
7901 | ||
7902 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
7903 | if (exit_bb) | |
7904 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
7905 | /* When the OMP expansion process cannot guarantee an up-to-date | |
7906 | loop tree arrange for the child function to fixup loops. */ | |
7907 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
7908 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
7909 | ||
7910 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
7911 | num = vec_safe_length (child_cfun->local_decls); | |
7912 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
7913 | { | |
7914 | t = (*child_cfun->local_decls)[srcidx]; | |
7915 | if (DECL_CONTEXT (t) == cfun->decl) | |
7916 | continue; | |
7917 | if (srcidx != dstidx) | |
7918 | (*child_cfun->local_decls)[dstidx] = t; | |
7919 | dstidx++; | |
7920 | } | |
7921 | if (dstidx != num) | |
7922 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
7923 | ||
7924 | /* Inform the callgraph about the new function. */ | |
7925 | DECL_STRUCT_FUNCTION (child_fn)->curr_properties = cfun->curr_properties; | |
7926 | cgraph_add_new_function (child_fn, true); | |
7927 | ||
7928 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
7929 | fixed in a following pass. */ | |
7930 | push_cfun (child_cfun); | |
7931 | rebuild_cgraph_edges (); | |
7932 | ||
7933 | /* Some EH regions might become dead, see PR34608. If | |
7934 | pass_cleanup_cfg isn't the first pass to happen with the | |
7935 | new child, these dead EH edges might cause problems. | |
7936 | Clean them up now. */ | |
7937 | if (flag_exceptions) | |
7938 | { | |
7939 | basic_block bb; | |
7940 | bool changed = false; | |
7941 | ||
fc00614f | 7942 | FOR_EACH_BB_FN (bb, cfun) |
bc7bff74 | 7943 | changed |= gimple_purge_dead_eh_edges (bb); |
7944 | if (changed) | |
7945 | cleanup_tree_cfg (); | |
7946 | } | |
7947 | pop_cfun (); | |
7948 | } | |
7949 | ||
7950 | /* Emit a library call to launch the target region, or do data | |
7951 | transfers. */ | |
7952 | tree t1, t2, t3, t4, device, cond, c, clauses; | |
7953 | enum built_in_function start_ix; | |
7954 | location_t clause_loc; | |
7955 | ||
7956 | clauses = gimple_omp_target_clauses (entry_stmt); | |
7957 | ||
7958 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
7959 | start_ix = BUILT_IN_GOMP_TARGET; | |
7960 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
7961 | start_ix = BUILT_IN_GOMP_TARGET_DATA; | |
7962 | else | |
7963 | start_ix = BUILT_IN_GOMP_TARGET_UPDATE; | |
7964 | ||
7965 | /* By default, the value of DEVICE is -1 (let runtime library choose) | |
7966 | and there is no conditional. */ | |
7967 | cond = NULL_TREE; | |
7968 | device = build_int_cst (integer_type_node, -1); | |
7969 | ||
7970 | c = find_omp_clause (clauses, OMP_CLAUSE_IF); | |
7971 | if (c) | |
7972 | cond = OMP_CLAUSE_IF_EXPR (c); | |
7973 | ||
7974 | c = find_omp_clause (clauses, OMP_CLAUSE_DEVICE); | |
7975 | if (c) | |
7976 | { | |
7977 | device = OMP_CLAUSE_DEVICE_ID (c); | |
7978 | clause_loc = OMP_CLAUSE_LOCATION (c); | |
7979 | } | |
7980 | else | |
7981 | clause_loc = gimple_location (entry_stmt); | |
7982 | ||
7983 | /* Ensure 'device' is of the correct type. */ | |
7984 | device = fold_convert_loc (clause_loc, integer_type_node, device); | |
7985 | ||
7986 | /* If we found the clause 'if (cond)', build | |
7987 | (cond ? device : -2). */ | |
7988 | if (cond) | |
7989 | { | |
7990 | cond = gimple_boolify (cond); | |
7991 | ||
7992 | basic_block cond_bb, then_bb, else_bb; | |
7993 | edge e; | |
7994 | tree tmp_var; | |
7995 | ||
7996 | tmp_var = create_tmp_var (TREE_TYPE (device), NULL); | |
7997 | if (kind != GF_OMP_TARGET_KIND_REGION) | |
7998 | { | |
7999 | gsi = gsi_last_bb (new_bb); | |
8000 | gsi_prev (&gsi); | |
8001 | e = split_block (new_bb, gsi_stmt (gsi)); | |
8002 | } | |
8003 | else | |
8004 | e = split_block (new_bb, NULL); | |
8005 | cond_bb = e->src; | |
8006 | new_bb = e->dest; | |
8007 | remove_edge (e); | |
8008 | ||
8009 | then_bb = create_empty_bb (cond_bb); | |
8010 | else_bb = create_empty_bb (then_bb); | |
8011 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
8012 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
8013 | ||
8014 | stmt = gimple_build_cond_empty (cond); | |
8015 | gsi = gsi_last_bb (cond_bb); | |
8016 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
8017 | ||
8018 | gsi = gsi_start_bb (then_bb); | |
8019 | stmt = gimple_build_assign (tmp_var, device); | |
8020 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
8021 | ||
8022 | gsi = gsi_start_bb (else_bb); | |
8023 | stmt = gimple_build_assign (tmp_var, | |
8024 | build_int_cst (integer_type_node, -2)); | |
8025 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
8026 | ||
8027 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
8028 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
8029 | if (current_loops) | |
8030 | { | |
8031 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
8032 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
8033 | } | |
8034 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); | |
8035 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); | |
8036 | ||
8037 | device = tmp_var; | |
8038 | } | |
8039 | ||
8040 | gsi = gsi_last_bb (new_bb); | |
8041 | t = gimple_omp_target_data_arg (entry_stmt); | |
8042 | if (t == NULL) | |
8043 | { | |
8044 | t1 = size_zero_node; | |
8045 | t2 = build_zero_cst (ptr_type_node); | |
8046 | t3 = t2; | |
8047 | t4 = t2; | |
8048 | } | |
8049 | else | |
8050 | { | |
8051 | t1 = TYPE_MAX_VALUE (TYPE_DOMAIN (TREE_TYPE (TREE_VEC_ELT (t, 1)))); | |
8052 | t1 = size_binop (PLUS_EXPR, t1, size_int (1)); | |
8053 | t2 = build_fold_addr_expr (TREE_VEC_ELT (t, 0)); | |
8054 | t3 = build_fold_addr_expr (TREE_VEC_ELT (t, 1)); | |
8055 | t4 = build_fold_addr_expr (TREE_VEC_ELT (t, 2)); | |
8056 | } | |
8057 | ||
8058 | gimple g; | |
8059 | /* FIXME: This will be address of | |
8060 | extern char __OPENMP_TARGET__[] __attribute__((visibility ("hidden"))) | |
8061 | symbol, as soon as the linker plugin is able to create it for us. */ | |
8062 | tree openmp_target = build_zero_cst (ptr_type_node); | |
8063 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
8064 | { | |
8065 | tree fnaddr = build_fold_addr_expr (child_fn); | |
8066 | g = gimple_build_call (builtin_decl_explicit (start_ix), 7, | |
8067 | device, fnaddr, openmp_target, t1, t2, t3, t4); | |
8068 | } | |
8069 | else | |
8070 | g = gimple_build_call (builtin_decl_explicit (start_ix), 6, | |
8071 | device, openmp_target, t1, t2, t3, t4); | |
8072 | gimple_set_location (g, gimple_location (entry_stmt)); | |
8073 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
8074 | if (kind != GF_OMP_TARGET_KIND_REGION) | |
8075 | { | |
8076 | g = gsi_stmt (gsi); | |
8077 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_TARGET); | |
8078 | gsi_remove (&gsi, true); | |
8079 | } | |
8080 | if (kind == GF_OMP_TARGET_KIND_DATA && region->exit) | |
8081 | { | |
8082 | gsi = gsi_last_bb (region->exit); | |
8083 | g = gsi_stmt (gsi); | |
8084 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_RETURN); | |
8085 | gsi_remove (&gsi, true); | |
8086 | } | |
8087 | } | |
8088 | ||
8089 | ||
8090 | /* Expand the parallel region tree rooted at REGION. Expansion | |
8091 | proceeds in depth-first order. Innermost regions are expanded | |
8092 | first. This way, parallel regions that require a new function to | |
75a70cf9 | 8093 | be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any |
773c5ba7 | 8094 | internal dependencies in their body. */ |
8095 | ||
8096 | static void | |
8097 | expand_omp (struct omp_region *region) | |
8098 | { | |
8099 | while (region) | |
8100 | { | |
1d22f541 | 8101 | location_t saved_location; |
bc7bff74 | 8102 | gimple inner_stmt = NULL; |
1d22f541 | 8103 | |
d1d5b012 | 8104 | /* First, determine whether this is a combined parallel+workshare |
8105 | region. */ | |
75a70cf9 | 8106 | if (region->type == GIMPLE_OMP_PARALLEL) |
d1d5b012 | 8107 | determine_parallel_type (region); |
8108 | ||
bc7bff74 | 8109 | if (region->type == GIMPLE_OMP_FOR |
8110 | && gimple_omp_for_combined_p (last_stmt (region->entry))) | |
8111 | inner_stmt = last_stmt (region->inner->entry); | |
8112 | ||
773c5ba7 | 8113 | if (region->inner) |
8114 | expand_omp (region->inner); | |
8115 | ||
1d22f541 | 8116 | saved_location = input_location; |
75a70cf9 | 8117 | if (gimple_has_location (last_stmt (region->entry))) |
8118 | input_location = gimple_location (last_stmt (region->entry)); | |
1d22f541 | 8119 | |
61e47ac8 | 8120 | switch (region->type) |
773c5ba7 | 8121 | { |
75a70cf9 | 8122 | case GIMPLE_OMP_PARALLEL: |
8123 | case GIMPLE_OMP_TASK: | |
fd6481cf | 8124 | expand_omp_taskreg (region); |
8125 | break; | |
8126 | ||
75a70cf9 | 8127 | case GIMPLE_OMP_FOR: |
bc7bff74 | 8128 | expand_omp_for (region, inner_stmt); |
61e47ac8 | 8129 | break; |
773c5ba7 | 8130 | |
75a70cf9 | 8131 | case GIMPLE_OMP_SECTIONS: |
61e47ac8 | 8132 | expand_omp_sections (region); |
8133 | break; | |
773c5ba7 | 8134 | |
75a70cf9 | 8135 | case GIMPLE_OMP_SECTION: |
61e47ac8 | 8136 | /* Individual omp sections are handled together with their |
75a70cf9 | 8137 | parent GIMPLE_OMP_SECTIONS region. */ |
61e47ac8 | 8138 | break; |
773c5ba7 | 8139 | |
75a70cf9 | 8140 | case GIMPLE_OMP_SINGLE: |
61e47ac8 | 8141 | expand_omp_single (region); |
8142 | break; | |
773c5ba7 | 8143 | |
75a70cf9 | 8144 | case GIMPLE_OMP_MASTER: |
bc7bff74 | 8145 | case GIMPLE_OMP_TASKGROUP: |
75a70cf9 | 8146 | case GIMPLE_OMP_ORDERED: |
8147 | case GIMPLE_OMP_CRITICAL: | |
bc7bff74 | 8148 | case GIMPLE_OMP_TEAMS: |
61e47ac8 | 8149 | expand_omp_synch (region); |
8150 | break; | |
773c5ba7 | 8151 | |
75a70cf9 | 8152 | case GIMPLE_OMP_ATOMIC_LOAD: |
cb7f680b | 8153 | expand_omp_atomic (region); |
8154 | break; | |
8155 | ||
bc7bff74 | 8156 | case GIMPLE_OMP_TARGET: |
8157 | expand_omp_target (region); | |
8158 | break; | |
8159 | ||
61e47ac8 | 8160 | default: |
8161 | gcc_unreachable (); | |
8162 | } | |
cc5982dc | 8163 | |
1d22f541 | 8164 | input_location = saved_location; |
773c5ba7 | 8165 | region = region->next; |
8166 | } | |
8167 | } | |
8168 | ||
8169 | ||
8170 | /* Helper for build_omp_regions. Scan the dominator tree starting at | |
28c92cbb | 8171 | block BB. PARENT is the region that contains BB. If SINGLE_TREE is |
8172 | true, the function ends once a single tree is built (otherwise, whole | |
8173 | forest of OMP constructs may be built). */ | |
773c5ba7 | 8174 | |
8175 | static void | |
28c92cbb | 8176 | build_omp_regions_1 (basic_block bb, struct omp_region *parent, |
8177 | bool single_tree) | |
773c5ba7 | 8178 | { |
75a70cf9 | 8179 | gimple_stmt_iterator gsi; |
8180 | gimple stmt; | |
773c5ba7 | 8181 | basic_block son; |
8182 | ||
75a70cf9 | 8183 | gsi = gsi_last_bb (bb); |
8184 | if (!gsi_end_p (gsi) && is_gimple_omp (gsi_stmt (gsi))) | |
773c5ba7 | 8185 | { |
8186 | struct omp_region *region; | |
75a70cf9 | 8187 | enum gimple_code code; |
773c5ba7 | 8188 | |
75a70cf9 | 8189 | stmt = gsi_stmt (gsi); |
8190 | code = gimple_code (stmt); | |
8191 | if (code == GIMPLE_OMP_RETURN) | |
773c5ba7 | 8192 | { |
8193 | /* STMT is the return point out of region PARENT. Mark it | |
8194 | as the exit point and make PARENT the immediately | |
8195 | enclosing region. */ | |
8196 | gcc_assert (parent); | |
8197 | region = parent; | |
61e47ac8 | 8198 | region->exit = bb; |
773c5ba7 | 8199 | parent = parent->outer; |
773c5ba7 | 8200 | } |
75a70cf9 | 8201 | else if (code == GIMPLE_OMP_ATOMIC_STORE) |
cb7f680b | 8202 | { |
75a70cf9 | 8203 | /* GIMPLE_OMP_ATOMIC_STORE is analoguous to |
8204 | GIMPLE_OMP_RETURN, but matches with | |
8205 | GIMPLE_OMP_ATOMIC_LOAD. */ | |
cb7f680b | 8206 | gcc_assert (parent); |
75a70cf9 | 8207 | gcc_assert (parent->type == GIMPLE_OMP_ATOMIC_LOAD); |
cb7f680b | 8208 | region = parent; |
8209 | region->exit = bb; | |
8210 | parent = parent->outer; | |
8211 | } | |
8212 | ||
75a70cf9 | 8213 | else if (code == GIMPLE_OMP_CONTINUE) |
61e47ac8 | 8214 | { |
8215 | gcc_assert (parent); | |
8216 | parent->cont = bb; | |
8217 | } | |
75a70cf9 | 8218 | else if (code == GIMPLE_OMP_SECTIONS_SWITCH) |
ac6e3339 | 8219 | { |
75a70cf9 | 8220 | /* GIMPLE_OMP_SECTIONS_SWITCH is part of |
8221 | GIMPLE_OMP_SECTIONS, and we do nothing for it. */ | |
8222 | ; | |
ac6e3339 | 8223 | } |
bc7bff74 | 8224 | else if (code == GIMPLE_OMP_TARGET |
8225 | && gimple_omp_target_kind (stmt) == GF_OMP_TARGET_KIND_UPDATE) | |
8226 | new_omp_region (bb, code, parent); | |
773c5ba7 | 8227 | else |
8228 | { | |
8229 | /* Otherwise, this directive becomes the parent for a new | |
8230 | region. */ | |
61e47ac8 | 8231 | region = new_omp_region (bb, code, parent); |
773c5ba7 | 8232 | parent = region; |
8233 | } | |
773c5ba7 | 8234 | } |
8235 | ||
28c92cbb | 8236 | if (single_tree && !parent) |
8237 | return; | |
8238 | ||
773c5ba7 | 8239 | for (son = first_dom_son (CDI_DOMINATORS, bb); |
8240 | son; | |
8241 | son = next_dom_son (CDI_DOMINATORS, son)) | |
28c92cbb | 8242 | build_omp_regions_1 (son, parent, single_tree); |
8243 | } | |
8244 | ||
8245 | /* Builds the tree of OMP regions rooted at ROOT, storing it to | |
8246 | root_omp_region. */ | |
8247 | ||
8248 | static void | |
8249 | build_omp_regions_root (basic_block root) | |
8250 | { | |
8251 | gcc_assert (root_omp_region == NULL); | |
8252 | build_omp_regions_1 (root, NULL, true); | |
8253 | gcc_assert (root_omp_region != NULL); | |
773c5ba7 | 8254 | } |
8255 | ||
28c92cbb | 8256 | /* Expands omp construct (and its subconstructs) starting in HEAD. */ |
8257 | ||
8258 | void | |
8259 | omp_expand_local (basic_block head) | |
8260 | { | |
8261 | build_omp_regions_root (head); | |
8262 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
8263 | { | |
8264 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
8265 | dump_omp_region (dump_file, root_omp_region, 0); | |
8266 | fprintf (dump_file, "\n"); | |
8267 | } | |
8268 | ||
8269 | remove_exit_barriers (root_omp_region); | |
8270 | expand_omp (root_omp_region); | |
8271 | ||
8272 | free_omp_regions (); | |
8273 | } | |
773c5ba7 | 8274 | |
8275 | /* Scan the CFG and build a tree of OMP regions. Return the root of | |
8276 | the OMP region tree. */ | |
8277 | ||
8278 | static void | |
8279 | build_omp_regions (void) | |
8280 | { | |
61e47ac8 | 8281 | gcc_assert (root_omp_region == NULL); |
773c5ba7 | 8282 | calculate_dominance_info (CDI_DOMINATORS); |
34154e27 | 8283 | build_omp_regions_1 (ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, false); |
773c5ba7 | 8284 | } |
8285 | ||
773c5ba7 | 8286 | /* Main entry point for expanding OMP-GIMPLE into runtime calls. */ |
8287 | ||
2a1990e9 | 8288 | static unsigned int |
773c5ba7 | 8289 | execute_expand_omp (void) |
8290 | { | |
8291 | build_omp_regions (); | |
8292 | ||
61e47ac8 | 8293 | if (!root_omp_region) |
8294 | return 0; | |
773c5ba7 | 8295 | |
61e47ac8 | 8296 | if (dump_file) |
8297 | { | |
8298 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
8299 | dump_omp_region (dump_file, root_omp_region, 0); | |
8300 | fprintf (dump_file, "\n"); | |
773c5ba7 | 8301 | } |
61e47ac8 | 8302 | |
8303 | remove_exit_barriers (root_omp_region); | |
8304 | ||
8305 | expand_omp (root_omp_region); | |
8306 | ||
61e47ac8 | 8307 | cleanup_tree_cfg (); |
8308 | ||
8309 | free_omp_regions (); | |
8310 | ||
2a1990e9 | 8311 | return 0; |
773c5ba7 | 8312 | } |
8313 | ||
79acaae1 | 8314 | /* OMP expansion -- the default pass, run before creation of SSA form. */ |
8315 | ||
773c5ba7 | 8316 | static bool |
8317 | gate_expand_omp (void) | |
8318 | { | |
f2697631 | 8319 | return ((flag_openmp != 0 || flag_openmp_simd != 0 |
8320 | || flag_enable_cilkplus != 0) && !seen_error ()); | |
773c5ba7 | 8321 | } |
8322 | ||
cbe8bda8 | 8323 | namespace { |
8324 | ||
8325 | const pass_data pass_data_expand_omp = | |
8326 | { | |
8327 | GIMPLE_PASS, /* type */ | |
8328 | "ompexp", /* name */ | |
8329 | OPTGROUP_NONE, /* optinfo_flags */ | |
8330 | true, /* has_gate */ | |
8331 | true, /* has_execute */ | |
8332 | TV_NONE, /* tv_id */ | |
8333 | PROP_gimple_any, /* properties_required */ | |
8334 | 0, /* properties_provided */ | |
8335 | 0, /* properties_destroyed */ | |
8336 | 0, /* todo_flags_start */ | |
8337 | 0, /* todo_flags_finish */ | |
773c5ba7 | 8338 | }; |
cbe8bda8 | 8339 | |
8340 | class pass_expand_omp : public gimple_opt_pass | |
8341 | { | |
8342 | public: | |
9af5ce0c | 8343 | pass_expand_omp (gcc::context *ctxt) |
8344 | : gimple_opt_pass (pass_data_expand_omp, ctxt) | |
cbe8bda8 | 8345 | {} |
8346 | ||
8347 | /* opt_pass methods: */ | |
8348 | bool gate () { return gate_expand_omp (); } | |
8349 | unsigned int execute () { return execute_expand_omp (); } | |
8350 | ||
8351 | }; // class pass_expand_omp | |
8352 | ||
8353 | } // anon namespace | |
8354 | ||
8355 | gimple_opt_pass * | |
8356 | make_pass_expand_omp (gcc::context *ctxt) | |
8357 | { | |
8358 | return new pass_expand_omp (ctxt); | |
8359 | } | |
773c5ba7 | 8360 | \f |
8361 | /* Routines to lower OpenMP directives into OMP-GIMPLE. */ | |
8362 | ||
bc7bff74 | 8363 | /* If ctx is a worksharing context inside of a cancellable parallel |
8364 | region and it isn't nowait, add lhs to its GIMPLE_OMP_RETURN | |
8365 | and conditional branch to parallel's cancel_label to handle | |
8366 | cancellation in the implicit barrier. */ | |
8367 | ||
8368 | static void | |
8369 | maybe_add_implicit_barrier_cancel (omp_context *ctx, gimple_seq *body) | |
8370 | { | |
8371 | gimple omp_return = gimple_seq_last_stmt (*body); | |
8372 | gcc_assert (gimple_code (omp_return) == GIMPLE_OMP_RETURN); | |
8373 | if (gimple_omp_return_nowait_p (omp_return)) | |
8374 | return; | |
8375 | if (ctx->outer | |
8376 | && gimple_code (ctx->outer->stmt) == GIMPLE_OMP_PARALLEL | |
8377 | && ctx->outer->cancellable) | |
8378 | { | |
8379 | tree lhs = create_tmp_var (boolean_type_node, NULL); | |
8380 | gimple_omp_return_set_lhs (omp_return, lhs); | |
8381 | tree fallthru_label = create_artificial_label (UNKNOWN_LOCATION); | |
8382 | gimple g = gimple_build_cond (NE_EXPR, lhs, boolean_false_node, | |
8383 | ctx->outer->cancel_label, fallthru_label); | |
8384 | gimple_seq_add_stmt (body, g); | |
8385 | gimple_seq_add_stmt (body, gimple_build_label (fallthru_label)); | |
8386 | } | |
8387 | } | |
8388 | ||
75a70cf9 | 8389 | /* Lower the OpenMP sections directive in the current statement in GSI_P. |
8390 | CTX is the enclosing OMP context for the current statement. */ | |
773c5ba7 | 8391 | |
8392 | static void | |
75a70cf9 | 8393 | lower_omp_sections (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
773c5ba7 | 8394 | { |
75a70cf9 | 8395 | tree block, control; |
8396 | gimple_stmt_iterator tgsi; | |
75a70cf9 | 8397 | gimple stmt, new_stmt, bind, t; |
e3a19533 | 8398 | gimple_seq ilist, dlist, olist, new_body; |
773c5ba7 | 8399 | |
75a70cf9 | 8400 | stmt = gsi_stmt (*gsi_p); |
773c5ba7 | 8401 | |
8a4a28a8 | 8402 | push_gimplify_context (); |
773c5ba7 | 8403 | |
8404 | dlist = NULL; | |
8405 | ilist = NULL; | |
75a70cf9 | 8406 | lower_rec_input_clauses (gimple_omp_sections_clauses (stmt), |
bc7bff74 | 8407 | &ilist, &dlist, ctx, NULL); |
773c5ba7 | 8408 | |
e3a19533 | 8409 | new_body = gimple_omp_body (stmt); |
8410 | gimple_omp_set_body (stmt, NULL); | |
8411 | tgsi = gsi_start (new_body); | |
8412 | for (; !gsi_end_p (tgsi); gsi_next (&tgsi)) | |
773c5ba7 | 8413 | { |
8414 | omp_context *sctx; | |
75a70cf9 | 8415 | gimple sec_start; |
773c5ba7 | 8416 | |
75a70cf9 | 8417 | sec_start = gsi_stmt (tgsi); |
773c5ba7 | 8418 | sctx = maybe_lookup_ctx (sec_start); |
8419 | gcc_assert (sctx); | |
8420 | ||
e3a19533 | 8421 | lower_omp (gimple_omp_body_ptr (sec_start), sctx); |
8422 | gsi_insert_seq_after (&tgsi, gimple_omp_body (sec_start), | |
8423 | GSI_CONTINUE_LINKING); | |
75a70cf9 | 8424 | gimple_omp_set_body (sec_start, NULL); |
773c5ba7 | 8425 | |
e3a19533 | 8426 | if (gsi_one_before_end_p (tgsi)) |
773c5ba7 | 8427 | { |
75a70cf9 | 8428 | gimple_seq l = NULL; |
8429 | lower_lastprivate_clauses (gimple_omp_sections_clauses (stmt), NULL, | |
773c5ba7 | 8430 | &l, ctx); |
e3a19533 | 8431 | gsi_insert_seq_after (&tgsi, l, GSI_CONTINUE_LINKING); |
75a70cf9 | 8432 | gimple_omp_section_set_last (sec_start); |
773c5ba7 | 8433 | } |
48e1416a | 8434 | |
e3a19533 | 8435 | gsi_insert_after (&tgsi, gimple_build_omp_return (false), |
8436 | GSI_CONTINUE_LINKING); | |
773c5ba7 | 8437 | } |
1e8e9920 | 8438 | |
8439 | block = make_node (BLOCK); | |
e3a19533 | 8440 | bind = gimple_build_bind (NULL, new_body, block); |
1e8e9920 | 8441 | |
75a70cf9 | 8442 | olist = NULL; |
8443 | lower_reduction_clauses (gimple_omp_sections_clauses (stmt), &olist, ctx); | |
773c5ba7 | 8444 | |
1d22f541 | 8445 | block = make_node (BLOCK); |
75a70cf9 | 8446 | new_stmt = gimple_build_bind (NULL, NULL, block); |
e3a19533 | 8447 | gsi_replace (gsi_p, new_stmt, true); |
773c5ba7 | 8448 | |
1d22f541 | 8449 | pop_gimplify_context (new_stmt); |
75a70cf9 | 8450 | gimple_bind_append_vars (new_stmt, ctx->block_vars); |
8451 | BLOCK_VARS (block) = gimple_bind_vars (bind); | |
1d22f541 | 8452 | if (BLOCK_VARS (block)) |
8453 | TREE_USED (block) = 1; | |
8454 | ||
75a70cf9 | 8455 | new_body = NULL; |
8456 | gimple_seq_add_seq (&new_body, ilist); | |
8457 | gimple_seq_add_stmt (&new_body, stmt); | |
8458 | gimple_seq_add_stmt (&new_body, gimple_build_omp_sections_switch ()); | |
8459 | gimple_seq_add_stmt (&new_body, bind); | |
61e47ac8 | 8460 | |
ac6e3339 | 8461 | control = create_tmp_var (unsigned_type_node, ".section"); |
75a70cf9 | 8462 | t = gimple_build_omp_continue (control, control); |
8463 | gimple_omp_sections_set_control (stmt, control); | |
8464 | gimple_seq_add_stmt (&new_body, t); | |
61e47ac8 | 8465 | |
75a70cf9 | 8466 | gimple_seq_add_seq (&new_body, olist); |
bc7bff74 | 8467 | if (ctx->cancellable) |
8468 | gimple_seq_add_stmt (&new_body, gimple_build_label (ctx->cancel_label)); | |
75a70cf9 | 8469 | gimple_seq_add_seq (&new_body, dlist); |
773c5ba7 | 8470 | |
75a70cf9 | 8471 | new_body = maybe_catch_exception (new_body); |
aade31a0 | 8472 | |
75a70cf9 | 8473 | t = gimple_build_omp_return |
8474 | (!!find_omp_clause (gimple_omp_sections_clauses (stmt), | |
8475 | OMP_CLAUSE_NOWAIT)); | |
8476 | gimple_seq_add_stmt (&new_body, t); | |
bc7bff74 | 8477 | maybe_add_implicit_barrier_cancel (ctx, &new_body); |
61e47ac8 | 8478 | |
75a70cf9 | 8479 | gimple_bind_set_body (new_stmt, new_body); |
1e8e9920 | 8480 | } |
8481 | ||
8482 | ||
773c5ba7 | 8483 | /* A subroutine of lower_omp_single. Expand the simple form of |
75a70cf9 | 8484 | a GIMPLE_OMP_SINGLE, without a copyprivate clause: |
1e8e9920 | 8485 | |
8486 | if (GOMP_single_start ()) | |
8487 | BODY; | |
8488 | [ GOMP_barrier (); ] -> unless 'nowait' is present. | |
773c5ba7 | 8489 | |
8490 | FIXME. It may be better to delay expanding the logic of this until | |
8491 | pass_expand_omp. The expanded logic may make the job more difficult | |
8492 | to a synchronization analysis pass. */ | |
1e8e9920 | 8493 | |
8494 | static void | |
75a70cf9 | 8495 | lower_omp_single_simple (gimple single_stmt, gimple_seq *pre_p) |
1e8e9920 | 8496 | { |
e60a6f7b | 8497 | location_t loc = gimple_location (single_stmt); |
8498 | tree tlabel = create_artificial_label (loc); | |
8499 | tree flabel = create_artificial_label (loc); | |
75a70cf9 | 8500 | gimple call, cond; |
8501 | tree lhs, decl; | |
8502 | ||
b9a16870 | 8503 | decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_START); |
75a70cf9 | 8504 | lhs = create_tmp_var (TREE_TYPE (TREE_TYPE (decl)), NULL); |
8505 | call = gimple_build_call (decl, 0); | |
8506 | gimple_call_set_lhs (call, lhs); | |
8507 | gimple_seq_add_stmt (pre_p, call); | |
8508 | ||
8509 | cond = gimple_build_cond (EQ_EXPR, lhs, | |
389dd41b | 8510 | fold_convert_loc (loc, TREE_TYPE (lhs), |
8511 | boolean_true_node), | |
75a70cf9 | 8512 | tlabel, flabel); |
8513 | gimple_seq_add_stmt (pre_p, cond); | |
8514 | gimple_seq_add_stmt (pre_p, gimple_build_label (tlabel)); | |
8515 | gimple_seq_add_seq (pre_p, gimple_omp_body (single_stmt)); | |
8516 | gimple_seq_add_stmt (pre_p, gimple_build_label (flabel)); | |
1e8e9920 | 8517 | } |
8518 | ||
773c5ba7 | 8519 | |
8520 | /* A subroutine of lower_omp_single. Expand the simple form of | |
75a70cf9 | 8521 | a GIMPLE_OMP_SINGLE, with a copyprivate clause: |
1e8e9920 | 8522 | |
8523 | #pragma omp single copyprivate (a, b, c) | |
8524 | ||
8525 | Create a new structure to hold copies of 'a', 'b' and 'c' and emit: | |
8526 | ||
8527 | { | |
8528 | if ((copyout_p = GOMP_single_copy_start ()) == NULL) | |
8529 | { | |
8530 | BODY; | |
8531 | copyout.a = a; | |
8532 | copyout.b = b; | |
8533 | copyout.c = c; | |
8534 | GOMP_single_copy_end (©out); | |
8535 | } | |
8536 | else | |
8537 | { | |
8538 | a = copyout_p->a; | |
8539 | b = copyout_p->b; | |
8540 | c = copyout_p->c; | |
8541 | } | |
8542 | GOMP_barrier (); | |
8543 | } | |
773c5ba7 | 8544 | |
8545 | FIXME. It may be better to delay expanding the logic of this until | |
8546 | pass_expand_omp. The expanded logic may make the job more difficult | |
8547 | to a synchronization analysis pass. */ | |
1e8e9920 | 8548 | |
8549 | static void | |
75a70cf9 | 8550 | lower_omp_single_copy (gimple single_stmt, gimple_seq *pre_p, omp_context *ctx) |
1e8e9920 | 8551 | { |
b9a16870 | 8552 | tree ptr_type, t, l0, l1, l2, bfn_decl; |
75a70cf9 | 8553 | gimple_seq copyin_seq; |
e60a6f7b | 8554 | location_t loc = gimple_location (single_stmt); |
1e8e9920 | 8555 | |
8556 | ctx->sender_decl = create_tmp_var (ctx->record_type, ".omp_copy_o"); | |
8557 | ||
8558 | ptr_type = build_pointer_type (ctx->record_type); | |
8559 | ctx->receiver_decl = create_tmp_var (ptr_type, ".omp_copy_i"); | |
8560 | ||
e60a6f7b | 8561 | l0 = create_artificial_label (loc); |
8562 | l1 = create_artificial_label (loc); | |
8563 | l2 = create_artificial_label (loc); | |
1e8e9920 | 8564 | |
b9a16870 | 8565 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_COPY_START); |
8566 | t = build_call_expr_loc (loc, bfn_decl, 0); | |
389dd41b | 8567 | t = fold_convert_loc (loc, ptr_type, t); |
75a70cf9 | 8568 | gimplify_assign (ctx->receiver_decl, t, pre_p); |
1e8e9920 | 8569 | |
8570 | t = build2 (EQ_EXPR, boolean_type_node, ctx->receiver_decl, | |
8571 | build_int_cst (ptr_type, 0)); | |
8572 | t = build3 (COND_EXPR, void_type_node, t, | |
8573 | build_and_jump (&l0), build_and_jump (&l1)); | |
8574 | gimplify_and_add (t, pre_p); | |
8575 | ||
75a70cf9 | 8576 | gimple_seq_add_stmt (pre_p, gimple_build_label (l0)); |
1e8e9920 | 8577 | |
75a70cf9 | 8578 | gimple_seq_add_seq (pre_p, gimple_omp_body (single_stmt)); |
1e8e9920 | 8579 | |
8580 | copyin_seq = NULL; | |
75a70cf9 | 8581 | lower_copyprivate_clauses (gimple_omp_single_clauses (single_stmt), pre_p, |
1e8e9920 | 8582 | ©in_seq, ctx); |
8583 | ||
389dd41b | 8584 | t = build_fold_addr_expr_loc (loc, ctx->sender_decl); |
b9a16870 | 8585 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_COPY_END); |
8586 | t = build_call_expr_loc (loc, bfn_decl, 1, t); | |
1e8e9920 | 8587 | gimplify_and_add (t, pre_p); |
8588 | ||
8589 | t = build_and_jump (&l2); | |
8590 | gimplify_and_add (t, pre_p); | |
8591 | ||
75a70cf9 | 8592 | gimple_seq_add_stmt (pre_p, gimple_build_label (l1)); |
1e8e9920 | 8593 | |
75a70cf9 | 8594 | gimple_seq_add_seq (pre_p, copyin_seq); |
1e8e9920 | 8595 | |
75a70cf9 | 8596 | gimple_seq_add_stmt (pre_p, gimple_build_label (l2)); |
1e8e9920 | 8597 | } |
8598 | ||
773c5ba7 | 8599 | |
1e8e9920 | 8600 | /* Expand code for an OpenMP single directive. */ |
8601 | ||
8602 | static void | |
75a70cf9 | 8603 | lower_omp_single (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 8604 | { |
75a70cf9 | 8605 | tree block; |
8606 | gimple t, bind, single_stmt = gsi_stmt (*gsi_p); | |
bc7bff74 | 8607 | gimple_seq bind_body, bind_body_tail = NULL, dlist; |
1e8e9920 | 8608 | |
8a4a28a8 | 8609 | push_gimplify_context (); |
1e8e9920 | 8610 | |
e3a19533 | 8611 | block = make_node (BLOCK); |
8612 | bind = gimple_build_bind (NULL, NULL, block); | |
8613 | gsi_replace (gsi_p, bind, true); | |
75a70cf9 | 8614 | bind_body = NULL; |
e3a19533 | 8615 | dlist = NULL; |
75a70cf9 | 8616 | lower_rec_input_clauses (gimple_omp_single_clauses (single_stmt), |
bc7bff74 | 8617 | &bind_body, &dlist, ctx, NULL); |
e3a19533 | 8618 | lower_omp (gimple_omp_body_ptr (single_stmt), ctx); |
1e8e9920 | 8619 | |
75a70cf9 | 8620 | gimple_seq_add_stmt (&bind_body, single_stmt); |
1e8e9920 | 8621 | |
8622 | if (ctx->record_type) | |
75a70cf9 | 8623 | lower_omp_single_copy (single_stmt, &bind_body, ctx); |
1e8e9920 | 8624 | else |
75a70cf9 | 8625 | lower_omp_single_simple (single_stmt, &bind_body); |
8626 | ||
8627 | gimple_omp_set_body (single_stmt, NULL); | |
1e8e9920 | 8628 | |
75a70cf9 | 8629 | gimple_seq_add_seq (&bind_body, dlist); |
61e47ac8 | 8630 | |
75a70cf9 | 8631 | bind_body = maybe_catch_exception (bind_body); |
61e47ac8 | 8632 | |
48e1416a | 8633 | t = gimple_build_omp_return |
75a70cf9 | 8634 | (!!find_omp_clause (gimple_omp_single_clauses (single_stmt), |
8635 | OMP_CLAUSE_NOWAIT)); | |
bc7bff74 | 8636 | gimple_seq_add_stmt (&bind_body_tail, t); |
8637 | maybe_add_implicit_barrier_cancel (ctx, &bind_body_tail); | |
8638 | if (ctx->record_type) | |
8639 | { | |
8640 | gimple_stmt_iterator gsi = gsi_start (bind_body_tail); | |
8641 | tree clobber = build_constructor (ctx->record_type, NULL); | |
8642 | TREE_THIS_VOLATILE (clobber) = 1; | |
8643 | gsi_insert_after (&gsi, gimple_build_assign (ctx->sender_decl, | |
8644 | clobber), GSI_SAME_STMT); | |
8645 | } | |
8646 | gimple_seq_add_seq (&bind_body, bind_body_tail); | |
e3a19533 | 8647 | gimple_bind_set_body (bind, bind_body); |
61e47ac8 | 8648 | |
1e8e9920 | 8649 | pop_gimplify_context (bind); |
773c5ba7 | 8650 | |
75a70cf9 | 8651 | gimple_bind_append_vars (bind, ctx->block_vars); |
8652 | BLOCK_VARS (block) = ctx->block_vars; | |
1d22f541 | 8653 | if (BLOCK_VARS (block)) |
8654 | TREE_USED (block) = 1; | |
1e8e9920 | 8655 | } |
8656 | ||
773c5ba7 | 8657 | |
1e8e9920 | 8658 | /* Expand code for an OpenMP master directive. */ |
8659 | ||
8660 | static void | |
75a70cf9 | 8661 | lower_omp_master (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 8662 | { |
b9a16870 | 8663 | tree block, lab = NULL, x, bfn_decl; |
75a70cf9 | 8664 | gimple stmt = gsi_stmt (*gsi_p), bind; |
389dd41b | 8665 | location_t loc = gimple_location (stmt); |
75a70cf9 | 8666 | gimple_seq tseq; |
1e8e9920 | 8667 | |
8a4a28a8 | 8668 | push_gimplify_context (); |
1e8e9920 | 8669 | |
8670 | block = make_node (BLOCK); | |
e3a19533 | 8671 | bind = gimple_build_bind (NULL, NULL, block); |
8672 | gsi_replace (gsi_p, bind, true); | |
8673 | gimple_bind_add_stmt (bind, stmt); | |
61e47ac8 | 8674 | |
b9a16870 | 8675 | bfn_decl = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); |
8676 | x = build_call_expr_loc (loc, bfn_decl, 0); | |
1e8e9920 | 8677 | x = build2 (EQ_EXPR, boolean_type_node, x, integer_zero_node); |
8678 | x = build3 (COND_EXPR, void_type_node, x, NULL, build_and_jump (&lab)); | |
75a70cf9 | 8679 | tseq = NULL; |
8680 | gimplify_and_add (x, &tseq); | |
8681 | gimple_bind_add_seq (bind, tseq); | |
1e8e9920 | 8682 | |
e3a19533 | 8683 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
75a70cf9 | 8684 | gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt))); |
8685 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
8686 | gimple_omp_set_body (stmt, NULL); | |
1e8e9920 | 8687 | |
75a70cf9 | 8688 | gimple_bind_add_stmt (bind, gimple_build_label (lab)); |
61e47ac8 | 8689 | |
75a70cf9 | 8690 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); |
61e47ac8 | 8691 | |
1e8e9920 | 8692 | pop_gimplify_context (bind); |
773c5ba7 | 8693 | |
75a70cf9 | 8694 | gimple_bind_append_vars (bind, ctx->block_vars); |
8695 | BLOCK_VARS (block) = ctx->block_vars; | |
1e8e9920 | 8696 | } |
8697 | ||
773c5ba7 | 8698 | |
bc7bff74 | 8699 | /* Expand code for an OpenMP taskgroup directive. */ |
8700 | ||
8701 | static void | |
8702 | lower_omp_taskgroup (gimple_stmt_iterator *gsi_p, omp_context *ctx) | |
8703 | { | |
8704 | gimple stmt = gsi_stmt (*gsi_p), bind, x; | |
8705 | tree block = make_node (BLOCK); | |
8706 | ||
8707 | bind = gimple_build_bind (NULL, NULL, block); | |
8708 | gsi_replace (gsi_p, bind, true); | |
8709 | gimple_bind_add_stmt (bind, stmt); | |
8710 | ||
8711 | x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_TASKGROUP_START), | |
8712 | 0); | |
8713 | gimple_bind_add_stmt (bind, x); | |
8714 | ||
8715 | lower_omp (gimple_omp_body_ptr (stmt), ctx); | |
8716 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
8717 | gimple_omp_set_body (stmt, NULL); | |
8718 | ||
8719 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); | |
8720 | ||
8721 | gimple_bind_append_vars (bind, ctx->block_vars); | |
8722 | BLOCK_VARS (block) = ctx->block_vars; | |
8723 | } | |
8724 | ||
8725 | ||
1e8e9920 | 8726 | /* Expand code for an OpenMP ordered directive. */ |
8727 | ||
8728 | static void | |
75a70cf9 | 8729 | lower_omp_ordered (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 8730 | { |
75a70cf9 | 8731 | tree block; |
8732 | gimple stmt = gsi_stmt (*gsi_p), bind, x; | |
1e8e9920 | 8733 | |
8a4a28a8 | 8734 | push_gimplify_context (); |
1e8e9920 | 8735 | |
8736 | block = make_node (BLOCK); | |
e3a19533 | 8737 | bind = gimple_build_bind (NULL, NULL, block); |
8738 | gsi_replace (gsi_p, bind, true); | |
8739 | gimple_bind_add_stmt (bind, stmt); | |
61e47ac8 | 8740 | |
b9a16870 | 8741 | x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ORDERED_START), |
8742 | 0); | |
75a70cf9 | 8743 | gimple_bind_add_stmt (bind, x); |
1e8e9920 | 8744 | |
e3a19533 | 8745 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
75a70cf9 | 8746 | gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt))); |
8747 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
8748 | gimple_omp_set_body (stmt, NULL); | |
1e8e9920 | 8749 | |
b9a16870 | 8750 | x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ORDERED_END), 0); |
75a70cf9 | 8751 | gimple_bind_add_stmt (bind, x); |
61e47ac8 | 8752 | |
75a70cf9 | 8753 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); |
61e47ac8 | 8754 | |
1e8e9920 | 8755 | pop_gimplify_context (bind); |
773c5ba7 | 8756 | |
75a70cf9 | 8757 | gimple_bind_append_vars (bind, ctx->block_vars); |
8758 | BLOCK_VARS (block) = gimple_bind_vars (bind); | |
1e8e9920 | 8759 | } |
8760 | ||
1e8e9920 | 8761 | |
75a70cf9 | 8762 | /* Gimplify a GIMPLE_OMP_CRITICAL statement. This is a relatively simple |
1e8e9920 | 8763 | substitution of a couple of function calls. But in the NAMED case, |
8764 | requires that languages coordinate a symbol name. It is therefore | |
8765 | best put here in common code. */ | |
8766 | ||
8767 | static GTY((param1_is (tree), param2_is (tree))) | |
8768 | splay_tree critical_name_mutexes; | |
8769 | ||
8770 | static void | |
75a70cf9 | 8771 | lower_omp_critical (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 8772 | { |
75a70cf9 | 8773 | tree block; |
8774 | tree name, lock, unlock; | |
8775 | gimple stmt = gsi_stmt (*gsi_p), bind; | |
389dd41b | 8776 | location_t loc = gimple_location (stmt); |
75a70cf9 | 8777 | gimple_seq tbody; |
1e8e9920 | 8778 | |
75a70cf9 | 8779 | name = gimple_omp_critical_name (stmt); |
1e8e9920 | 8780 | if (name) |
8781 | { | |
c2f47e15 | 8782 | tree decl; |
1e8e9920 | 8783 | splay_tree_node n; |
8784 | ||
8785 | if (!critical_name_mutexes) | |
8786 | critical_name_mutexes | |
ba72912a | 8787 | = splay_tree_new_ggc (splay_tree_compare_pointers, |
8788 | ggc_alloc_splay_tree_tree_node_tree_node_splay_tree_s, | |
8789 | ggc_alloc_splay_tree_tree_node_tree_node_splay_tree_node_s); | |
1e8e9920 | 8790 | |
8791 | n = splay_tree_lookup (critical_name_mutexes, (splay_tree_key) name); | |
8792 | if (n == NULL) | |
8793 | { | |
8794 | char *new_str; | |
8795 | ||
8796 | decl = create_tmp_var_raw (ptr_type_node, NULL); | |
8797 | ||
8798 | new_str = ACONCAT ((".gomp_critical_user_", | |
8799 | IDENTIFIER_POINTER (name), NULL)); | |
8800 | DECL_NAME (decl) = get_identifier (new_str); | |
8801 | TREE_PUBLIC (decl) = 1; | |
8802 | TREE_STATIC (decl) = 1; | |
8803 | DECL_COMMON (decl) = 1; | |
8804 | DECL_ARTIFICIAL (decl) = 1; | |
8805 | DECL_IGNORED_P (decl) = 1; | |
1d416bd7 | 8806 | varpool_finalize_decl (decl); |
1e8e9920 | 8807 | |
8808 | splay_tree_insert (critical_name_mutexes, (splay_tree_key) name, | |
8809 | (splay_tree_value) decl); | |
8810 | } | |
8811 | else | |
8812 | decl = (tree) n->value; | |
8813 | ||
b9a16870 | 8814 | lock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_NAME_START); |
389dd41b | 8815 | lock = build_call_expr_loc (loc, lock, 1, build_fold_addr_expr_loc (loc, decl)); |
1e8e9920 | 8816 | |
b9a16870 | 8817 | unlock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_NAME_END); |
389dd41b | 8818 | unlock = build_call_expr_loc (loc, unlock, 1, |
8819 | build_fold_addr_expr_loc (loc, decl)); | |
1e8e9920 | 8820 | } |
8821 | else | |
8822 | { | |
b9a16870 | 8823 | lock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_START); |
389dd41b | 8824 | lock = build_call_expr_loc (loc, lock, 0); |
1e8e9920 | 8825 | |
b9a16870 | 8826 | unlock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_END); |
389dd41b | 8827 | unlock = build_call_expr_loc (loc, unlock, 0); |
1e8e9920 | 8828 | } |
8829 | ||
8a4a28a8 | 8830 | push_gimplify_context (); |
1e8e9920 | 8831 | |
8832 | block = make_node (BLOCK); | |
e3a19533 | 8833 | bind = gimple_build_bind (NULL, NULL, block); |
8834 | gsi_replace (gsi_p, bind, true); | |
8835 | gimple_bind_add_stmt (bind, stmt); | |
61e47ac8 | 8836 | |
75a70cf9 | 8837 | tbody = gimple_bind_body (bind); |
8838 | gimplify_and_add (lock, &tbody); | |
8839 | gimple_bind_set_body (bind, tbody); | |
1e8e9920 | 8840 | |
e3a19533 | 8841 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
75a70cf9 | 8842 | gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt))); |
8843 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
8844 | gimple_omp_set_body (stmt, NULL); | |
1e8e9920 | 8845 | |
75a70cf9 | 8846 | tbody = gimple_bind_body (bind); |
8847 | gimplify_and_add (unlock, &tbody); | |
8848 | gimple_bind_set_body (bind, tbody); | |
61e47ac8 | 8849 | |
75a70cf9 | 8850 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); |
1e8e9920 | 8851 | |
8852 | pop_gimplify_context (bind); | |
75a70cf9 | 8853 | gimple_bind_append_vars (bind, ctx->block_vars); |
8854 | BLOCK_VARS (block) = gimple_bind_vars (bind); | |
773c5ba7 | 8855 | } |
8856 | ||
8857 | ||
8858 | /* A subroutine of lower_omp_for. Generate code to emit the predicate | |
8859 | for a lastprivate clause. Given a loop control predicate of (V | |
8860 | cond N2), we gate the clause on (!(V cond N2)). The lowered form | |
1e4afe3c | 8861 | is appended to *DLIST, iterator initialization is appended to |
8862 | *BODY_P. */ | |
773c5ba7 | 8863 | |
8864 | static void | |
75a70cf9 | 8865 | lower_omp_for_lastprivate (struct omp_for_data *fd, gimple_seq *body_p, |
8866 | gimple_seq *dlist, struct omp_context *ctx) | |
773c5ba7 | 8867 | { |
75a70cf9 | 8868 | tree clauses, cond, vinit; |
773c5ba7 | 8869 | enum tree_code cond_code; |
75a70cf9 | 8870 | gimple_seq stmts; |
48e1416a | 8871 | |
fd6481cf | 8872 | cond_code = fd->loop.cond_code; |
773c5ba7 | 8873 | cond_code = cond_code == LT_EXPR ? GE_EXPR : LE_EXPR; |
8874 | ||
8875 | /* When possible, use a strict equality expression. This can let VRP | |
8876 | type optimizations deduce the value and remove a copy. */ | |
35ec552a | 8877 | if (tree_fits_shwi_p (fd->loop.step)) |
773c5ba7 | 8878 | { |
8c53c46c | 8879 | HOST_WIDE_INT step = tree_to_shwi (fd->loop.step); |
773c5ba7 | 8880 | if (step == 1 || step == -1) |
8881 | cond_code = EQ_EXPR; | |
8882 | } | |
8883 | ||
fd6481cf | 8884 | cond = build2 (cond_code, boolean_type_node, fd->loop.v, fd->loop.n2); |
773c5ba7 | 8885 | |
75a70cf9 | 8886 | clauses = gimple_omp_for_clauses (fd->for_stmt); |
1e4afe3c | 8887 | stmts = NULL; |
8888 | lower_lastprivate_clauses (clauses, cond, &stmts, ctx); | |
75a70cf9 | 8889 | if (!gimple_seq_empty_p (stmts)) |
1e4afe3c | 8890 | { |
75a70cf9 | 8891 | gimple_seq_add_seq (&stmts, *dlist); |
fd6481cf | 8892 | *dlist = stmts; |
1e4afe3c | 8893 | |
8894 | /* Optimize: v = 0; is usually cheaper than v = some_other_constant. */ | |
fd6481cf | 8895 | vinit = fd->loop.n1; |
1e4afe3c | 8896 | if (cond_code == EQ_EXPR |
35ec552a | 8897 | && tree_fits_shwi_p (fd->loop.n2) |
fd6481cf | 8898 | && ! integer_zerop (fd->loop.n2)) |
8899 | vinit = build_int_cst (TREE_TYPE (fd->loop.v), 0); | |
3d483a94 | 8900 | else |
8901 | vinit = unshare_expr (vinit); | |
1e4afe3c | 8902 | |
8903 | /* Initialize the iterator variable, so that threads that don't execute | |
8904 | any iterations don't execute the lastprivate clauses by accident. */ | |
75a70cf9 | 8905 | gimplify_assign (fd->loop.v, vinit, body_p); |
1e4afe3c | 8906 | } |
773c5ba7 | 8907 | } |
8908 | ||
8909 | ||
8910 | /* Lower code for an OpenMP loop directive. */ | |
8911 | ||
8912 | static void | |
75a70cf9 | 8913 | lower_omp_for (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
773c5ba7 | 8914 | { |
75a70cf9 | 8915 | tree *rhs_p, block; |
bc7bff74 | 8916 | struct omp_for_data fd, *fdp = NULL; |
75a70cf9 | 8917 | gimple stmt = gsi_stmt (*gsi_p), new_stmt; |
f018d957 | 8918 | gimple_seq omp_for_body, body, dlist; |
75a70cf9 | 8919 | size_t i; |
773c5ba7 | 8920 | |
8a4a28a8 | 8921 | push_gimplify_context (); |
773c5ba7 | 8922 | |
e3a19533 | 8923 | lower_omp (gimple_omp_for_pre_body_ptr (stmt), ctx); |
773c5ba7 | 8924 | |
1d22f541 | 8925 | block = make_node (BLOCK); |
75a70cf9 | 8926 | new_stmt = gimple_build_bind (NULL, NULL, block); |
e3a19533 | 8927 | /* Replace at gsi right away, so that 'stmt' is no member |
8928 | of a sequence anymore as we're going to add to to a different | |
8929 | one below. */ | |
8930 | gsi_replace (gsi_p, new_stmt, true); | |
1d22f541 | 8931 | |
773c5ba7 | 8932 | /* Move declaration of temporaries in the loop body before we make |
8933 | it go away. */ | |
75a70cf9 | 8934 | omp_for_body = gimple_omp_body (stmt); |
8935 | if (!gimple_seq_empty_p (omp_for_body) | |
8936 | && gimple_code (gimple_seq_first_stmt (omp_for_body)) == GIMPLE_BIND) | |
8937 | { | |
8938 | tree vars = gimple_bind_vars (gimple_seq_first_stmt (omp_for_body)); | |
8939 | gimple_bind_append_vars (new_stmt, vars); | |
8940 | } | |
773c5ba7 | 8941 | |
bc7bff74 | 8942 | if (gimple_omp_for_combined_into_p (stmt)) |
8943 | { | |
8944 | extract_omp_for_data (stmt, &fd, NULL); | |
8945 | fdp = &fd; | |
8946 | ||
8947 | /* We need two temporaries with fd.loop.v type (istart/iend) | |
8948 | and then (fd.collapse - 1) temporaries with the same | |
8949 | type for count2 ... countN-1 vars if not constant. */ | |
8950 | size_t count = 2; | |
8951 | tree type = fd.iter_type; | |
8952 | if (fd.collapse > 1 | |
8953 | && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
8954 | count += fd.collapse - 1; | |
8955 | bool parallel_for = gimple_omp_for_kind (stmt) == GF_OMP_FOR_KIND_FOR; | |
8956 | tree outerc = NULL, *pc = gimple_omp_for_clauses_ptr (stmt); | |
8957 | tree clauses = *pc; | |
8958 | if (parallel_for) | |
8959 | outerc | |
8960 | = find_omp_clause (gimple_omp_parallel_clauses (ctx->outer->stmt), | |
8961 | OMP_CLAUSE__LOOPTEMP_); | |
8962 | for (i = 0; i < count; i++) | |
8963 | { | |
8964 | tree temp; | |
8965 | if (parallel_for) | |
8966 | { | |
8967 | gcc_assert (outerc); | |
8968 | temp = lookup_decl (OMP_CLAUSE_DECL (outerc), ctx->outer); | |
8969 | outerc = find_omp_clause (OMP_CLAUSE_CHAIN (outerc), | |
8970 | OMP_CLAUSE__LOOPTEMP_); | |
8971 | } | |
8972 | else | |
8973 | temp = create_tmp_var (type, NULL); | |
8974 | *pc = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE__LOOPTEMP_); | |
8975 | OMP_CLAUSE_DECL (*pc) = temp; | |
8976 | pc = &OMP_CLAUSE_CHAIN (*pc); | |
8977 | } | |
8978 | *pc = clauses; | |
8979 | } | |
8980 | ||
75a70cf9 | 8981 | /* The pre-body and input clauses go before the lowered GIMPLE_OMP_FOR. */ |
773c5ba7 | 8982 | dlist = NULL; |
75a70cf9 | 8983 | body = NULL; |
bc7bff74 | 8984 | lower_rec_input_clauses (gimple_omp_for_clauses (stmt), &body, &dlist, ctx, |
8985 | fdp); | |
75a70cf9 | 8986 | gimple_seq_add_seq (&body, gimple_omp_for_pre_body (stmt)); |
773c5ba7 | 8987 | |
3d483a94 | 8988 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
8989 | ||
773c5ba7 | 8990 | /* Lower the header expressions. At this point, we can assume that |
8991 | the header is of the form: | |
8992 | ||
8993 | #pragma omp for (V = VAL1; V {<|>|<=|>=} VAL2; V = V [+-] VAL3) | |
8994 | ||
8995 | We just need to make sure that VAL1, VAL2 and VAL3 are lowered | |
8996 | using the .omp_data_s mapping, if needed. */ | |
75a70cf9 | 8997 | for (i = 0; i < gimple_omp_for_collapse (stmt); i++) |
fd6481cf | 8998 | { |
75a70cf9 | 8999 | rhs_p = gimple_omp_for_initial_ptr (stmt, i); |
fd6481cf | 9000 | if (!is_gimple_min_invariant (*rhs_p)) |
75a70cf9 | 9001 | *rhs_p = get_formal_tmp_var (*rhs_p, &body); |
fd6481cf | 9002 | |
75a70cf9 | 9003 | rhs_p = gimple_omp_for_final_ptr (stmt, i); |
fd6481cf | 9004 | if (!is_gimple_min_invariant (*rhs_p)) |
75a70cf9 | 9005 | *rhs_p = get_formal_tmp_var (*rhs_p, &body); |
fd6481cf | 9006 | |
75a70cf9 | 9007 | rhs_p = &TREE_OPERAND (gimple_omp_for_incr (stmt, i), 1); |
fd6481cf | 9008 | if (!is_gimple_min_invariant (*rhs_p)) |
75a70cf9 | 9009 | *rhs_p = get_formal_tmp_var (*rhs_p, &body); |
fd6481cf | 9010 | } |
773c5ba7 | 9011 | |
9012 | /* Once lowered, extract the bounds and clauses. */ | |
fd6481cf | 9013 | extract_omp_for_data (stmt, &fd, NULL); |
773c5ba7 | 9014 | |
75a70cf9 | 9015 | lower_omp_for_lastprivate (&fd, &body, &dlist, ctx); |
773c5ba7 | 9016 | |
75a70cf9 | 9017 | gimple_seq_add_stmt (&body, stmt); |
9018 | gimple_seq_add_seq (&body, gimple_omp_body (stmt)); | |
61e47ac8 | 9019 | |
75a70cf9 | 9020 | gimple_seq_add_stmt (&body, gimple_build_omp_continue (fd.loop.v, |
9021 | fd.loop.v)); | |
61e47ac8 | 9022 | |
773c5ba7 | 9023 | /* After the loop, add exit clauses. */ |
75a70cf9 | 9024 | lower_reduction_clauses (gimple_omp_for_clauses (stmt), &body, ctx); |
bc7bff74 | 9025 | |
9026 | if (ctx->cancellable) | |
9027 | gimple_seq_add_stmt (&body, gimple_build_label (ctx->cancel_label)); | |
9028 | ||
75a70cf9 | 9029 | gimple_seq_add_seq (&body, dlist); |
773c5ba7 | 9030 | |
75a70cf9 | 9031 | body = maybe_catch_exception (body); |
aade31a0 | 9032 | |
61e47ac8 | 9033 | /* Region exit marker goes at the end of the loop body. */ |
75a70cf9 | 9034 | gimple_seq_add_stmt (&body, gimple_build_omp_return (fd.have_nowait)); |
bc7bff74 | 9035 | maybe_add_implicit_barrier_cancel (ctx, &body); |
1d22f541 | 9036 | pop_gimplify_context (new_stmt); |
75a70cf9 | 9037 | |
9038 | gimple_bind_append_vars (new_stmt, ctx->block_vars); | |
9039 | BLOCK_VARS (block) = gimple_bind_vars (new_stmt); | |
1d22f541 | 9040 | if (BLOCK_VARS (block)) |
9041 | TREE_USED (block) = 1; | |
773c5ba7 | 9042 | |
75a70cf9 | 9043 | gimple_bind_set_body (new_stmt, body); |
9044 | gimple_omp_set_body (stmt, NULL); | |
9045 | gimple_omp_for_set_pre_body (stmt, NULL); | |
1e8e9920 | 9046 | } |
9047 | ||
48e1416a | 9048 | /* Callback for walk_stmts. Check if the current statement only contains |
75a70cf9 | 9049 | GIMPLE_OMP_FOR or GIMPLE_OMP_PARALLEL. */ |
de7ef844 | 9050 | |
9051 | static tree | |
75a70cf9 | 9052 | check_combined_parallel (gimple_stmt_iterator *gsi_p, |
9053 | bool *handled_ops_p, | |
9054 | struct walk_stmt_info *wi) | |
de7ef844 | 9055 | { |
4077bf7a | 9056 | int *info = (int *) wi->info; |
75a70cf9 | 9057 | gimple stmt = gsi_stmt (*gsi_p); |
de7ef844 | 9058 | |
75a70cf9 | 9059 | *handled_ops_p = true; |
9060 | switch (gimple_code (stmt)) | |
de7ef844 | 9061 | { |
75a70cf9 | 9062 | WALK_SUBSTMTS; |
9063 | ||
9064 | case GIMPLE_OMP_FOR: | |
9065 | case GIMPLE_OMP_SECTIONS: | |
de7ef844 | 9066 | *info = *info == 0 ? 1 : -1; |
9067 | break; | |
9068 | default: | |
9069 | *info = -1; | |
9070 | break; | |
9071 | } | |
9072 | return NULL; | |
9073 | } | |
773c5ba7 | 9074 | |
fd6481cf | 9075 | struct omp_taskcopy_context |
9076 | { | |
9077 | /* This field must be at the beginning, as we do "inheritance": Some | |
9078 | callback functions for tree-inline.c (e.g., omp_copy_decl) | |
9079 | receive a copy_body_data pointer that is up-casted to an | |
9080 | omp_context pointer. */ | |
9081 | copy_body_data cb; | |
9082 | omp_context *ctx; | |
9083 | }; | |
9084 | ||
9085 | static tree | |
9086 | task_copyfn_copy_decl (tree var, copy_body_data *cb) | |
9087 | { | |
9088 | struct omp_taskcopy_context *tcctx = (struct omp_taskcopy_context *) cb; | |
9089 | ||
9090 | if (splay_tree_lookup (tcctx->ctx->sfield_map, (splay_tree_key) var)) | |
9091 | return create_tmp_var (TREE_TYPE (var), NULL); | |
9092 | ||
9093 | return var; | |
9094 | } | |
9095 | ||
9096 | static tree | |
9097 | task_copyfn_remap_type (struct omp_taskcopy_context *tcctx, tree orig_type) | |
9098 | { | |
9099 | tree name, new_fields = NULL, type, f; | |
9100 | ||
9101 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
9102 | name = DECL_NAME (TYPE_NAME (orig_type)); | |
e60a6f7b | 9103 | name = build_decl (gimple_location (tcctx->ctx->stmt), |
9104 | TYPE_DECL, name, type); | |
fd6481cf | 9105 | TYPE_NAME (type) = name; |
9106 | ||
9107 | for (f = TYPE_FIELDS (orig_type); f ; f = TREE_CHAIN (f)) | |
9108 | { | |
9109 | tree new_f = copy_node (f); | |
9110 | DECL_CONTEXT (new_f) = type; | |
9111 | TREE_TYPE (new_f) = remap_type (TREE_TYPE (f), &tcctx->cb); | |
9112 | TREE_CHAIN (new_f) = new_fields; | |
75a70cf9 | 9113 | walk_tree (&DECL_SIZE (new_f), copy_tree_body_r, &tcctx->cb, NULL); |
9114 | walk_tree (&DECL_SIZE_UNIT (new_f), copy_tree_body_r, &tcctx->cb, NULL); | |
9115 | walk_tree (&DECL_FIELD_OFFSET (new_f), copy_tree_body_r, | |
9116 | &tcctx->cb, NULL); | |
fd6481cf | 9117 | new_fields = new_f; |
9118 | *pointer_map_insert (tcctx->cb.decl_map, f) = new_f; | |
9119 | } | |
9120 | TYPE_FIELDS (type) = nreverse (new_fields); | |
9121 | layout_type (type); | |
9122 | return type; | |
9123 | } | |
9124 | ||
9125 | /* Create task copyfn. */ | |
9126 | ||
9127 | static void | |
75a70cf9 | 9128 | create_task_copyfn (gimple task_stmt, omp_context *ctx) |
fd6481cf | 9129 | { |
9130 | struct function *child_cfun; | |
9131 | tree child_fn, t, c, src, dst, f, sf, arg, sarg, decl; | |
9132 | tree record_type, srecord_type, bind, list; | |
9133 | bool record_needs_remap = false, srecord_needs_remap = false; | |
9134 | splay_tree_node n; | |
9135 | struct omp_taskcopy_context tcctx; | |
389dd41b | 9136 | location_t loc = gimple_location (task_stmt); |
fd6481cf | 9137 | |
75a70cf9 | 9138 | child_fn = gimple_omp_task_copy_fn (task_stmt); |
fd6481cf | 9139 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); |
9140 | gcc_assert (child_cfun->cfg == NULL); | |
fd6481cf | 9141 | DECL_SAVED_TREE (child_fn) = alloc_stmt_list (); |
9142 | ||
9143 | /* Reset DECL_CONTEXT on function arguments. */ | |
1767a056 | 9144 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) |
fd6481cf | 9145 | DECL_CONTEXT (t) = child_fn; |
9146 | ||
9147 | /* Populate the function. */ | |
8a4a28a8 | 9148 | push_gimplify_context (); |
9078126c | 9149 | push_cfun (child_cfun); |
fd6481cf | 9150 | |
9151 | bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, NULL); | |
9152 | TREE_SIDE_EFFECTS (bind) = 1; | |
9153 | list = NULL; | |
9154 | DECL_SAVED_TREE (child_fn) = bind; | |
75a70cf9 | 9155 | DECL_SOURCE_LOCATION (child_fn) = gimple_location (task_stmt); |
fd6481cf | 9156 | |
9157 | /* Remap src and dst argument types if needed. */ | |
9158 | record_type = ctx->record_type; | |
9159 | srecord_type = ctx->srecord_type; | |
1767a056 | 9160 | for (f = TYPE_FIELDS (record_type); f ; f = DECL_CHAIN (f)) |
fd6481cf | 9161 | if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn)) |
9162 | { | |
9163 | record_needs_remap = true; | |
9164 | break; | |
9165 | } | |
1767a056 | 9166 | for (f = TYPE_FIELDS (srecord_type); f ; f = DECL_CHAIN (f)) |
fd6481cf | 9167 | if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn)) |
9168 | { | |
9169 | srecord_needs_remap = true; | |
9170 | break; | |
9171 | } | |
9172 | ||
9173 | if (record_needs_remap || srecord_needs_remap) | |
9174 | { | |
9175 | memset (&tcctx, '\0', sizeof (tcctx)); | |
9176 | tcctx.cb.src_fn = ctx->cb.src_fn; | |
9177 | tcctx.cb.dst_fn = child_fn; | |
53f79206 | 9178 | tcctx.cb.src_node = cgraph_get_node (tcctx.cb.src_fn); |
9179 | gcc_checking_assert (tcctx.cb.src_node); | |
fd6481cf | 9180 | tcctx.cb.dst_node = tcctx.cb.src_node; |
9181 | tcctx.cb.src_cfun = ctx->cb.src_cfun; | |
9182 | tcctx.cb.copy_decl = task_copyfn_copy_decl; | |
e38def9c | 9183 | tcctx.cb.eh_lp_nr = 0; |
fd6481cf | 9184 | tcctx.cb.transform_call_graph_edges = CB_CGE_MOVE; |
9185 | tcctx.cb.decl_map = pointer_map_create (); | |
9186 | tcctx.ctx = ctx; | |
9187 | ||
9188 | if (record_needs_remap) | |
9189 | record_type = task_copyfn_remap_type (&tcctx, record_type); | |
9190 | if (srecord_needs_remap) | |
9191 | srecord_type = task_copyfn_remap_type (&tcctx, srecord_type); | |
9192 | } | |
9193 | else | |
9194 | tcctx.cb.decl_map = NULL; | |
9195 | ||
fd6481cf | 9196 | arg = DECL_ARGUMENTS (child_fn); |
9197 | TREE_TYPE (arg) = build_pointer_type (record_type); | |
1767a056 | 9198 | sarg = DECL_CHAIN (arg); |
fd6481cf | 9199 | TREE_TYPE (sarg) = build_pointer_type (srecord_type); |
9200 | ||
9201 | /* First pass: initialize temporaries used in record_type and srecord_type | |
9202 | sizes and field offsets. */ | |
9203 | if (tcctx.cb.decl_map) | |
75a70cf9 | 9204 | for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c)) |
fd6481cf | 9205 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_FIRSTPRIVATE) |
9206 | { | |
9207 | tree *p; | |
9208 | ||
9209 | decl = OMP_CLAUSE_DECL (c); | |
9210 | p = (tree *) pointer_map_contains (tcctx.cb.decl_map, decl); | |
9211 | if (p == NULL) | |
9212 | continue; | |
9213 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); | |
9214 | sf = (tree) n->value; | |
9215 | sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf); | |
182cf5a9 | 9216 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9217 | src = omp_build_component_ref (src, sf); |
75a70cf9 | 9218 | t = build2 (MODIFY_EXPR, TREE_TYPE (*p), *p, src); |
fd6481cf | 9219 | append_to_statement_list (t, &list); |
9220 | } | |
9221 | ||
9222 | /* Second pass: copy shared var pointers and copy construct non-VLA | |
9223 | firstprivate vars. */ | |
75a70cf9 | 9224 | for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c)) |
fd6481cf | 9225 | switch (OMP_CLAUSE_CODE (c)) |
9226 | { | |
9227 | case OMP_CLAUSE_SHARED: | |
9228 | decl = OMP_CLAUSE_DECL (c); | |
9229 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9230 | if (n == NULL) | |
9231 | break; | |
9232 | f = (tree) n->value; | |
9233 | if (tcctx.cb.decl_map) | |
9234 | f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f); | |
9235 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); | |
9236 | sf = (tree) n->value; | |
9237 | if (tcctx.cb.decl_map) | |
9238 | sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf); | |
182cf5a9 | 9239 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9240 | src = omp_build_component_ref (src, sf); |
182cf5a9 | 9241 | dst = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9242 | dst = omp_build_component_ref (dst, f); |
75a70cf9 | 9243 | t = build2 (MODIFY_EXPR, TREE_TYPE (dst), dst, src); |
fd6481cf | 9244 | append_to_statement_list (t, &list); |
9245 | break; | |
9246 | case OMP_CLAUSE_FIRSTPRIVATE: | |
9247 | decl = OMP_CLAUSE_DECL (c); | |
9248 | if (is_variable_sized (decl)) | |
9249 | break; | |
9250 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9251 | if (n == NULL) | |
9252 | break; | |
9253 | f = (tree) n->value; | |
9254 | if (tcctx.cb.decl_map) | |
9255 | f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f); | |
9256 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); | |
9257 | if (n != NULL) | |
9258 | { | |
9259 | sf = (tree) n->value; | |
9260 | if (tcctx.cb.decl_map) | |
9261 | sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf); | |
182cf5a9 | 9262 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9263 | src = omp_build_component_ref (src, sf); |
fd6481cf | 9264 | if (use_pointer_for_field (decl, NULL) || is_reference (decl)) |
182cf5a9 | 9265 | src = build_simple_mem_ref_loc (loc, src); |
fd6481cf | 9266 | } |
9267 | else | |
9268 | src = decl; | |
182cf5a9 | 9269 | dst = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9270 | dst = omp_build_component_ref (dst, f); |
fd6481cf | 9271 | t = lang_hooks.decls.omp_clause_copy_ctor (c, dst, src); |
9272 | append_to_statement_list (t, &list); | |
9273 | break; | |
9274 | case OMP_CLAUSE_PRIVATE: | |
9275 | if (! OMP_CLAUSE_PRIVATE_OUTER_REF (c)) | |
9276 | break; | |
9277 | decl = OMP_CLAUSE_DECL (c); | |
9278 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9279 | f = (tree) n->value; | |
9280 | if (tcctx.cb.decl_map) | |
9281 | f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f); | |
9282 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); | |
9283 | if (n != NULL) | |
9284 | { | |
9285 | sf = (tree) n->value; | |
9286 | if (tcctx.cb.decl_map) | |
9287 | sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf); | |
182cf5a9 | 9288 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9289 | src = omp_build_component_ref (src, sf); |
fd6481cf | 9290 | if (use_pointer_for_field (decl, NULL)) |
182cf5a9 | 9291 | src = build_simple_mem_ref_loc (loc, src); |
fd6481cf | 9292 | } |
9293 | else | |
9294 | src = decl; | |
182cf5a9 | 9295 | dst = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9296 | dst = omp_build_component_ref (dst, f); |
75a70cf9 | 9297 | t = build2 (MODIFY_EXPR, TREE_TYPE (dst), dst, src); |
fd6481cf | 9298 | append_to_statement_list (t, &list); |
9299 | break; | |
9300 | default: | |
9301 | break; | |
9302 | } | |
9303 | ||
9304 | /* Last pass: handle VLA firstprivates. */ | |
9305 | if (tcctx.cb.decl_map) | |
75a70cf9 | 9306 | for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c)) |
fd6481cf | 9307 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_FIRSTPRIVATE) |
9308 | { | |
9309 | tree ind, ptr, df; | |
9310 | ||
9311 | decl = OMP_CLAUSE_DECL (c); | |
9312 | if (!is_variable_sized (decl)) | |
9313 | continue; | |
9314 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9315 | if (n == NULL) | |
9316 | continue; | |
9317 | f = (tree) n->value; | |
9318 | f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f); | |
9319 | gcc_assert (DECL_HAS_VALUE_EXPR_P (decl)); | |
9320 | ind = DECL_VALUE_EXPR (decl); | |
9321 | gcc_assert (TREE_CODE (ind) == INDIRECT_REF); | |
9322 | gcc_assert (DECL_P (TREE_OPERAND (ind, 0))); | |
9323 | n = splay_tree_lookup (ctx->sfield_map, | |
9324 | (splay_tree_key) TREE_OPERAND (ind, 0)); | |
9325 | sf = (tree) n->value; | |
9326 | sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf); | |
182cf5a9 | 9327 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9328 | src = omp_build_component_ref (src, sf); |
182cf5a9 | 9329 | src = build_simple_mem_ref_loc (loc, src); |
9330 | dst = build_simple_mem_ref_loc (loc, arg); | |
445d06b6 | 9331 | dst = omp_build_component_ref (dst, f); |
fd6481cf | 9332 | t = lang_hooks.decls.omp_clause_copy_ctor (c, dst, src); |
9333 | append_to_statement_list (t, &list); | |
9334 | n = splay_tree_lookup (ctx->field_map, | |
9335 | (splay_tree_key) TREE_OPERAND (ind, 0)); | |
9336 | df = (tree) n->value; | |
9337 | df = *(tree *) pointer_map_contains (tcctx.cb.decl_map, df); | |
182cf5a9 | 9338 | ptr = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9339 | ptr = omp_build_component_ref (ptr, df); |
75a70cf9 | 9340 | t = build2 (MODIFY_EXPR, TREE_TYPE (ptr), ptr, |
389dd41b | 9341 | build_fold_addr_expr_loc (loc, dst)); |
fd6481cf | 9342 | append_to_statement_list (t, &list); |
9343 | } | |
9344 | ||
9345 | t = build1 (RETURN_EXPR, void_type_node, NULL); | |
9346 | append_to_statement_list (t, &list); | |
9347 | ||
9348 | if (tcctx.cb.decl_map) | |
9349 | pointer_map_destroy (tcctx.cb.decl_map); | |
9350 | pop_gimplify_context (NULL); | |
9351 | BIND_EXPR_BODY (bind) = list; | |
9352 | pop_cfun (); | |
fd6481cf | 9353 | } |
9354 | ||
bc7bff74 | 9355 | static void |
9356 | lower_depend_clauses (gimple stmt, gimple_seq *iseq, gimple_seq *oseq) | |
9357 | { | |
9358 | tree c, clauses; | |
9359 | gimple g; | |
9360 | size_t n_in = 0, n_out = 0, idx = 2, i; | |
9361 | ||
9362 | clauses = find_omp_clause (gimple_omp_task_clauses (stmt), | |
9363 | OMP_CLAUSE_DEPEND); | |
9364 | gcc_assert (clauses); | |
9365 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
9366 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_DEPEND) | |
9367 | switch (OMP_CLAUSE_DEPEND_KIND (c)) | |
9368 | { | |
9369 | case OMP_CLAUSE_DEPEND_IN: | |
9370 | n_in++; | |
9371 | break; | |
9372 | case OMP_CLAUSE_DEPEND_OUT: | |
9373 | case OMP_CLAUSE_DEPEND_INOUT: | |
9374 | n_out++; | |
9375 | break; | |
9376 | default: | |
9377 | gcc_unreachable (); | |
9378 | } | |
9379 | tree type = build_array_type_nelts (ptr_type_node, n_in + n_out + 2); | |
9380 | tree array = create_tmp_var (type, NULL); | |
9381 | tree r = build4 (ARRAY_REF, ptr_type_node, array, size_int (0), NULL_TREE, | |
9382 | NULL_TREE); | |
9383 | g = gimple_build_assign (r, build_int_cst (ptr_type_node, n_in + n_out)); | |
9384 | gimple_seq_add_stmt (iseq, g); | |
9385 | r = build4 (ARRAY_REF, ptr_type_node, array, size_int (1), NULL_TREE, | |
9386 | NULL_TREE); | |
9387 | g = gimple_build_assign (r, build_int_cst (ptr_type_node, n_out)); | |
9388 | gimple_seq_add_stmt (iseq, g); | |
9389 | for (i = 0; i < 2; i++) | |
9390 | { | |
9391 | if ((i ? n_in : n_out) == 0) | |
9392 | continue; | |
9393 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
9394 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_DEPEND | |
9395 | && ((OMP_CLAUSE_DEPEND_KIND (c) != OMP_CLAUSE_DEPEND_IN) ^ i)) | |
9396 | { | |
9397 | tree t = OMP_CLAUSE_DECL (c); | |
9398 | t = fold_convert (ptr_type_node, t); | |
9399 | gimplify_expr (&t, iseq, NULL, is_gimple_val, fb_rvalue); | |
9400 | r = build4 (ARRAY_REF, ptr_type_node, array, size_int (idx++), | |
9401 | NULL_TREE, NULL_TREE); | |
9402 | g = gimple_build_assign (r, t); | |
9403 | gimple_seq_add_stmt (iseq, g); | |
9404 | } | |
9405 | } | |
9406 | tree *p = gimple_omp_task_clauses_ptr (stmt); | |
9407 | c = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE_DEPEND); | |
9408 | OMP_CLAUSE_DECL (c) = build_fold_addr_expr (array); | |
9409 | OMP_CLAUSE_CHAIN (c) = *p; | |
9410 | *p = c; | |
9411 | tree clobber = build_constructor (type, NULL); | |
9412 | TREE_THIS_VOLATILE (clobber) = 1; | |
9413 | g = gimple_build_assign (array, clobber); | |
9414 | gimple_seq_add_stmt (oseq, g); | |
9415 | } | |
9416 | ||
75a70cf9 | 9417 | /* Lower the OpenMP parallel or task directive in the current statement |
9418 | in GSI_P. CTX holds context information for the directive. */ | |
773c5ba7 | 9419 | |
9420 | static void | |
75a70cf9 | 9421 | lower_omp_taskreg (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
773c5ba7 | 9422 | { |
75a70cf9 | 9423 | tree clauses; |
9424 | tree child_fn, t; | |
9425 | gimple stmt = gsi_stmt (*gsi_p); | |
bc7bff74 | 9426 | gimple par_bind, bind, dep_bind = NULL; |
9427 | gimple_seq par_body, olist, ilist, par_olist, par_rlist, par_ilist, new_body; | |
389dd41b | 9428 | location_t loc = gimple_location (stmt); |
773c5ba7 | 9429 | |
75a70cf9 | 9430 | clauses = gimple_omp_taskreg_clauses (stmt); |
9431 | par_bind = gimple_seq_first_stmt (gimple_omp_body (stmt)); | |
9432 | par_body = gimple_bind_body (par_bind); | |
773c5ba7 | 9433 | child_fn = ctx->cb.dst_fn; |
75a70cf9 | 9434 | if (gimple_code (stmt) == GIMPLE_OMP_PARALLEL |
9435 | && !gimple_omp_parallel_combined_p (stmt)) | |
de7ef844 | 9436 | { |
9437 | struct walk_stmt_info wi; | |
9438 | int ws_num = 0; | |
9439 | ||
9440 | memset (&wi, 0, sizeof (wi)); | |
de7ef844 | 9441 | wi.info = &ws_num; |
9442 | wi.val_only = true; | |
75a70cf9 | 9443 | walk_gimple_seq (par_body, check_combined_parallel, NULL, &wi); |
de7ef844 | 9444 | if (ws_num == 1) |
75a70cf9 | 9445 | gimple_omp_parallel_set_combined_p (stmt, true); |
de7ef844 | 9446 | } |
bc7bff74 | 9447 | gimple_seq dep_ilist = NULL; |
9448 | gimple_seq dep_olist = NULL; | |
9449 | if (gimple_code (stmt) == GIMPLE_OMP_TASK | |
9450 | && find_omp_clause (clauses, OMP_CLAUSE_DEPEND)) | |
9451 | { | |
8a4a28a8 | 9452 | push_gimplify_context (); |
bc7bff74 | 9453 | dep_bind = gimple_build_bind (NULL, NULL, make_node (BLOCK)); |
9454 | lower_depend_clauses (stmt, &dep_ilist, &dep_olist); | |
9455 | } | |
9456 | ||
fd6481cf | 9457 | if (ctx->srecord_type) |
9458 | create_task_copyfn (stmt, ctx); | |
773c5ba7 | 9459 | |
8a4a28a8 | 9460 | push_gimplify_context (); |
773c5ba7 | 9461 | |
75a70cf9 | 9462 | par_olist = NULL; |
9463 | par_ilist = NULL; | |
bc7bff74 | 9464 | par_rlist = NULL; |
9465 | lower_rec_input_clauses (clauses, &par_ilist, &par_olist, ctx, NULL); | |
e3a19533 | 9466 | lower_omp (&par_body, ctx); |
75a70cf9 | 9467 | if (gimple_code (stmt) == GIMPLE_OMP_PARALLEL) |
bc7bff74 | 9468 | lower_reduction_clauses (clauses, &par_rlist, ctx); |
773c5ba7 | 9469 | |
9470 | /* Declare all the variables created by mapping and the variables | |
9471 | declared in the scope of the parallel body. */ | |
9472 | record_vars_into (ctx->block_vars, child_fn); | |
75a70cf9 | 9473 | record_vars_into (gimple_bind_vars (par_bind), child_fn); |
773c5ba7 | 9474 | |
9475 | if (ctx->record_type) | |
9476 | { | |
fd6481cf | 9477 | ctx->sender_decl |
9478 | = create_tmp_var (ctx->srecord_type ? ctx->srecord_type | |
9479 | : ctx->record_type, ".omp_data_o"); | |
84bfaaeb | 9480 | DECL_NAMELESS (ctx->sender_decl) = 1; |
86f2ad37 | 9481 | TREE_ADDRESSABLE (ctx->sender_decl) = 1; |
75a70cf9 | 9482 | gimple_omp_taskreg_set_data_arg (stmt, ctx->sender_decl); |
773c5ba7 | 9483 | } |
9484 | ||
75a70cf9 | 9485 | olist = NULL; |
9486 | ilist = NULL; | |
773c5ba7 | 9487 | lower_send_clauses (clauses, &ilist, &olist, ctx); |
9488 | lower_send_shared_vars (&ilist, &olist, ctx); | |
9489 | ||
bc7bff74 | 9490 | if (ctx->record_type) |
9491 | { | |
9492 | tree clobber = build_constructor (TREE_TYPE (ctx->sender_decl), NULL); | |
9493 | TREE_THIS_VOLATILE (clobber) = 1; | |
9494 | gimple_seq_add_stmt (&olist, gimple_build_assign (ctx->sender_decl, | |
9495 | clobber)); | |
9496 | } | |
9497 | ||
773c5ba7 | 9498 | /* Once all the expansions are done, sequence all the different |
75a70cf9 | 9499 | fragments inside gimple_omp_body. */ |
773c5ba7 | 9500 | |
75a70cf9 | 9501 | new_body = NULL; |
773c5ba7 | 9502 | |
9503 | if (ctx->record_type) | |
9504 | { | |
389dd41b | 9505 | t = build_fold_addr_expr_loc (loc, ctx->sender_decl); |
cc6b725b | 9506 | /* fixup_child_record_type might have changed receiver_decl's type. */ |
389dd41b | 9507 | t = fold_convert_loc (loc, TREE_TYPE (ctx->receiver_decl), t); |
75a70cf9 | 9508 | gimple_seq_add_stmt (&new_body, |
9509 | gimple_build_assign (ctx->receiver_decl, t)); | |
773c5ba7 | 9510 | } |
9511 | ||
75a70cf9 | 9512 | gimple_seq_add_seq (&new_body, par_ilist); |
9513 | gimple_seq_add_seq (&new_body, par_body); | |
bc7bff74 | 9514 | gimple_seq_add_seq (&new_body, par_rlist); |
9515 | if (ctx->cancellable) | |
9516 | gimple_seq_add_stmt (&new_body, gimple_build_label (ctx->cancel_label)); | |
75a70cf9 | 9517 | gimple_seq_add_seq (&new_body, par_olist); |
9518 | new_body = maybe_catch_exception (new_body); | |
9519 | gimple_seq_add_stmt (&new_body, gimple_build_omp_return (false)); | |
9520 | gimple_omp_set_body (stmt, new_body); | |
773c5ba7 | 9521 | |
75a70cf9 | 9522 | bind = gimple_build_bind (NULL, NULL, gimple_bind_block (par_bind)); |
bc7bff74 | 9523 | gsi_replace (gsi_p, dep_bind ? dep_bind : bind, true); |
9524 | gimple_bind_add_seq (bind, ilist); | |
9525 | gimple_bind_add_stmt (bind, stmt); | |
9526 | gimple_bind_add_seq (bind, olist); | |
9527 | ||
9528 | pop_gimplify_context (NULL); | |
9529 | ||
9530 | if (dep_bind) | |
9531 | { | |
9532 | gimple_bind_add_seq (dep_bind, dep_ilist); | |
9533 | gimple_bind_add_stmt (dep_bind, bind); | |
9534 | gimple_bind_add_seq (dep_bind, dep_olist); | |
9535 | pop_gimplify_context (dep_bind); | |
9536 | } | |
9537 | } | |
9538 | ||
9539 | /* Lower the OpenMP target directive in the current statement | |
9540 | in GSI_P. CTX holds context information for the directive. */ | |
9541 | ||
9542 | static void | |
9543 | lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx) | |
9544 | { | |
9545 | tree clauses; | |
9546 | tree child_fn, t, c; | |
9547 | gimple stmt = gsi_stmt (*gsi_p); | |
9548 | gimple tgt_bind = NULL, bind; | |
9549 | gimple_seq tgt_body = NULL, olist, ilist, new_body; | |
bc7bff74 | 9550 | location_t loc = gimple_location (stmt); |
9551 | int kind = gimple_omp_target_kind (stmt); | |
9552 | unsigned int map_cnt = 0; | |
9553 | ||
9554 | clauses = gimple_omp_target_clauses (stmt); | |
9555 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
9556 | { | |
9557 | tgt_bind = gimple_seq_first_stmt (gimple_omp_body (stmt)); | |
9558 | tgt_body = gimple_bind_body (tgt_bind); | |
9559 | } | |
9560 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
9561 | tgt_body = gimple_omp_body (stmt); | |
9562 | child_fn = ctx->cb.dst_fn; | |
9563 | ||
8a4a28a8 | 9564 | push_gimplify_context (); |
bc7bff74 | 9565 | |
9566 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
9567 | switch (OMP_CLAUSE_CODE (c)) | |
9568 | { | |
9569 | tree var, x; | |
9570 | ||
9571 | default: | |
9572 | break; | |
9573 | case OMP_CLAUSE_MAP: | |
9574 | case OMP_CLAUSE_TO: | |
9575 | case OMP_CLAUSE_FROM: | |
9576 | var = OMP_CLAUSE_DECL (c); | |
9577 | if (!DECL_P (var)) | |
9578 | { | |
9579 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_MAP | |
9580 | || !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c)) | |
9581 | map_cnt++; | |
9582 | continue; | |
9583 | } | |
9584 | ||
9585 | if (DECL_SIZE (var) | |
9586 | && TREE_CODE (DECL_SIZE (var)) != INTEGER_CST) | |
9587 | { | |
9588 | tree var2 = DECL_VALUE_EXPR (var); | |
9589 | gcc_assert (TREE_CODE (var2) == INDIRECT_REF); | |
9590 | var2 = TREE_OPERAND (var2, 0); | |
9591 | gcc_assert (DECL_P (var2)); | |
9592 | var = var2; | |
9593 | } | |
9594 | ||
9595 | if (!maybe_lookup_field (var, ctx)) | |
9596 | continue; | |
9597 | ||
9598 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
9599 | { | |
9600 | x = build_receiver_ref (var, true, ctx); | |
9601 | tree new_var = lookup_decl (var, ctx); | |
9602 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
9603 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
9604 | && !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) | |
9605 | && TREE_CODE (TREE_TYPE (var)) == ARRAY_TYPE) | |
9606 | x = build_simple_mem_ref (x); | |
9607 | SET_DECL_VALUE_EXPR (new_var, x); | |
9608 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
9609 | } | |
9610 | map_cnt++; | |
9611 | } | |
9612 | ||
9613 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
9614 | { | |
9615 | target_nesting_level++; | |
9616 | lower_omp (&tgt_body, ctx); | |
9617 | target_nesting_level--; | |
9618 | } | |
9619 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
9620 | lower_omp (&tgt_body, ctx); | |
9621 | ||
9622 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
9623 | { | |
9624 | /* Declare all the variables created by mapping and the variables | |
9625 | declared in the scope of the target body. */ | |
9626 | record_vars_into (ctx->block_vars, child_fn); | |
9627 | record_vars_into (gimple_bind_vars (tgt_bind), child_fn); | |
9628 | } | |
9629 | ||
9630 | olist = NULL; | |
9631 | ilist = NULL; | |
9632 | if (ctx->record_type) | |
9633 | { | |
9634 | ctx->sender_decl | |
9635 | = create_tmp_var (ctx->record_type, ".omp_data_arr"); | |
9636 | DECL_NAMELESS (ctx->sender_decl) = 1; | |
9637 | TREE_ADDRESSABLE (ctx->sender_decl) = 1; | |
9638 | t = make_tree_vec (3); | |
9639 | TREE_VEC_ELT (t, 0) = ctx->sender_decl; | |
9640 | TREE_VEC_ELT (t, 1) | |
9641 | = create_tmp_var (build_array_type_nelts (size_type_node, map_cnt), | |
9642 | ".omp_data_sizes"); | |
9643 | DECL_NAMELESS (TREE_VEC_ELT (t, 1)) = 1; | |
9644 | TREE_ADDRESSABLE (TREE_VEC_ELT (t, 1)) = 1; | |
9645 | TREE_STATIC (TREE_VEC_ELT (t, 1)) = 1; | |
9646 | TREE_VEC_ELT (t, 2) | |
9647 | = create_tmp_var (build_array_type_nelts (unsigned_char_type_node, | |
9648 | map_cnt), | |
9649 | ".omp_data_kinds"); | |
9650 | DECL_NAMELESS (TREE_VEC_ELT (t, 2)) = 1; | |
9651 | TREE_ADDRESSABLE (TREE_VEC_ELT (t, 2)) = 1; | |
9652 | TREE_STATIC (TREE_VEC_ELT (t, 2)) = 1; | |
9653 | gimple_omp_target_set_data_arg (stmt, t); | |
9654 | ||
9655 | vec<constructor_elt, va_gc> *vsize; | |
9656 | vec<constructor_elt, va_gc> *vkind; | |
9657 | vec_alloc (vsize, map_cnt); | |
9658 | vec_alloc (vkind, map_cnt); | |
9659 | unsigned int map_idx = 0; | |
9660 | ||
9661 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
9662 | switch (OMP_CLAUSE_CODE (c)) | |
9663 | { | |
9664 | tree ovar, nc; | |
9665 | ||
9666 | default: | |
9667 | break; | |
9668 | case OMP_CLAUSE_MAP: | |
9669 | case OMP_CLAUSE_TO: | |
9670 | case OMP_CLAUSE_FROM: | |
9671 | nc = c; | |
9672 | ovar = OMP_CLAUSE_DECL (c); | |
9673 | if (!DECL_P (ovar)) | |
9674 | { | |
9675 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
9676 | && OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c)) | |
9677 | { | |
9678 | gcc_checking_assert (OMP_CLAUSE_DECL (OMP_CLAUSE_CHAIN (c)) | |
9679 | == get_base_address (ovar)); | |
9680 | nc = OMP_CLAUSE_CHAIN (c); | |
9681 | ovar = OMP_CLAUSE_DECL (nc); | |
9682 | } | |
9683 | else | |
9684 | { | |
9685 | tree x = build_sender_ref (ovar, ctx); | |
9686 | tree v | |
9687 | = build_fold_addr_expr_with_type (ovar, ptr_type_node); | |
9688 | gimplify_assign (x, v, &ilist); | |
9689 | nc = NULL_TREE; | |
9690 | } | |
9691 | } | |
9692 | else | |
9693 | { | |
9694 | if (DECL_SIZE (ovar) | |
9695 | && TREE_CODE (DECL_SIZE (ovar)) != INTEGER_CST) | |
9696 | { | |
9697 | tree ovar2 = DECL_VALUE_EXPR (ovar); | |
9698 | gcc_assert (TREE_CODE (ovar2) == INDIRECT_REF); | |
9699 | ovar2 = TREE_OPERAND (ovar2, 0); | |
9700 | gcc_assert (DECL_P (ovar2)); | |
9701 | ovar = ovar2; | |
9702 | } | |
9703 | if (!maybe_lookup_field (ovar, ctx)) | |
9704 | continue; | |
9705 | } | |
9706 | ||
9707 | if (nc) | |
9708 | { | |
9709 | tree var = lookup_decl_in_outer_ctx (ovar, ctx); | |
9710 | tree x = build_sender_ref (ovar, ctx); | |
9711 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
9712 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
9713 | && !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) | |
9714 | && TREE_CODE (TREE_TYPE (ovar)) == ARRAY_TYPE) | |
9715 | { | |
9716 | gcc_assert (kind == GF_OMP_TARGET_KIND_REGION); | |
9717 | tree avar | |
9718 | = create_tmp_var (TREE_TYPE (TREE_TYPE (x)), NULL); | |
9719 | mark_addressable (avar); | |
9720 | gimplify_assign (avar, build_fold_addr_expr (var), &ilist); | |
9721 | avar = build_fold_addr_expr (avar); | |
9722 | gimplify_assign (x, avar, &ilist); | |
9723 | } | |
9724 | else if (is_gimple_reg (var)) | |
9725 | { | |
9726 | gcc_assert (kind == GF_OMP_TARGET_KIND_REGION); | |
9727 | tree avar = create_tmp_var (TREE_TYPE (var), NULL); | |
9728 | mark_addressable (avar); | |
9729 | if (OMP_CLAUSE_MAP_KIND (c) != OMP_CLAUSE_MAP_ALLOC | |
9730 | && OMP_CLAUSE_MAP_KIND (c) != OMP_CLAUSE_MAP_FROM) | |
9731 | gimplify_assign (avar, var, &ilist); | |
9732 | avar = build_fold_addr_expr (avar); | |
9733 | gimplify_assign (x, avar, &ilist); | |
9734 | if ((OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_FROM | |
9735 | || OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_TOFROM) | |
9736 | && !TYPE_READONLY (TREE_TYPE (var))) | |
9737 | { | |
9738 | x = build_sender_ref (ovar, ctx); | |
9739 | x = build_simple_mem_ref (x); | |
9740 | gimplify_assign (var, x, &olist); | |
9741 | } | |
9742 | } | |
9743 | else | |
9744 | { | |
9745 | var = build_fold_addr_expr (var); | |
9746 | gimplify_assign (x, var, &ilist); | |
9747 | } | |
9748 | } | |
9749 | tree s = OMP_CLAUSE_SIZE (c); | |
9750 | if (s == NULL_TREE) | |
9751 | s = TYPE_SIZE_UNIT (TREE_TYPE (ovar)); | |
9752 | s = fold_convert (size_type_node, s); | |
9753 | tree purpose = size_int (map_idx++); | |
9754 | CONSTRUCTOR_APPEND_ELT (vsize, purpose, s); | |
9755 | if (TREE_CODE (s) != INTEGER_CST) | |
9756 | TREE_STATIC (TREE_VEC_ELT (t, 1)) = 0; | |
9757 | ||
9758 | unsigned char tkind = 0; | |
9759 | switch (OMP_CLAUSE_CODE (c)) | |
9760 | { | |
9761 | case OMP_CLAUSE_MAP: | |
9762 | tkind = OMP_CLAUSE_MAP_KIND (c); | |
9763 | break; | |
9764 | case OMP_CLAUSE_TO: | |
9765 | tkind = OMP_CLAUSE_MAP_TO; | |
9766 | break; | |
9767 | case OMP_CLAUSE_FROM: | |
9768 | tkind = OMP_CLAUSE_MAP_FROM; | |
9769 | break; | |
9770 | default: | |
9771 | gcc_unreachable (); | |
9772 | } | |
9773 | unsigned int talign = TYPE_ALIGN_UNIT (TREE_TYPE (ovar)); | |
9774 | if (DECL_P (ovar) && DECL_ALIGN_UNIT (ovar) > talign) | |
9775 | talign = DECL_ALIGN_UNIT (ovar); | |
9776 | talign = ceil_log2 (talign); | |
9777 | tkind |= talign << 3; | |
9778 | CONSTRUCTOR_APPEND_ELT (vkind, purpose, | |
9779 | build_int_cst (unsigned_char_type_node, | |
9780 | tkind)); | |
9781 | if (nc && nc != c) | |
9782 | c = nc; | |
9783 | } | |
9784 | ||
9785 | gcc_assert (map_idx == map_cnt); | |
9786 | ||
9787 | DECL_INITIAL (TREE_VEC_ELT (t, 1)) | |
9788 | = build_constructor (TREE_TYPE (TREE_VEC_ELT (t, 1)), vsize); | |
9789 | DECL_INITIAL (TREE_VEC_ELT (t, 2)) | |
9790 | = build_constructor (TREE_TYPE (TREE_VEC_ELT (t, 2)), vkind); | |
9791 | if (!TREE_STATIC (TREE_VEC_ELT (t, 1))) | |
9792 | { | |
9793 | gimple_seq initlist = NULL; | |
9794 | force_gimple_operand (build1 (DECL_EXPR, void_type_node, | |
9795 | TREE_VEC_ELT (t, 1)), | |
9796 | &initlist, true, NULL_TREE); | |
9797 | gimple_seq_add_seq (&ilist, initlist); | |
9798 | } | |
9799 | ||
9800 | tree clobber = build_constructor (ctx->record_type, NULL); | |
9801 | TREE_THIS_VOLATILE (clobber) = 1; | |
9802 | gimple_seq_add_stmt (&olist, gimple_build_assign (ctx->sender_decl, | |
9803 | clobber)); | |
9804 | } | |
9805 | ||
9806 | /* Once all the expansions are done, sequence all the different | |
9807 | fragments inside gimple_omp_body. */ | |
9808 | ||
9809 | new_body = NULL; | |
9810 | ||
9811 | if (ctx->record_type && kind == GF_OMP_TARGET_KIND_REGION) | |
9812 | { | |
9813 | t = build_fold_addr_expr_loc (loc, ctx->sender_decl); | |
9814 | /* fixup_child_record_type might have changed receiver_decl's type. */ | |
9815 | t = fold_convert_loc (loc, TREE_TYPE (ctx->receiver_decl), t); | |
9816 | gimple_seq_add_stmt (&new_body, | |
9817 | gimple_build_assign (ctx->receiver_decl, t)); | |
9818 | } | |
9819 | ||
9820 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
9821 | { | |
9822 | gimple_seq_add_seq (&new_body, tgt_body); | |
9823 | new_body = maybe_catch_exception (new_body); | |
9824 | } | |
9825 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
9826 | new_body = tgt_body; | |
9827 | if (kind != GF_OMP_TARGET_KIND_UPDATE) | |
9828 | { | |
9829 | gimple_seq_add_stmt (&new_body, gimple_build_omp_return (false)); | |
9830 | gimple_omp_set_body (stmt, new_body); | |
9831 | } | |
9832 | ||
9833 | bind = gimple_build_bind (NULL, NULL, | |
9834 | tgt_bind ? gimple_bind_block (tgt_bind) | |
9835 | : NULL_TREE); | |
75a70cf9 | 9836 | gsi_replace (gsi_p, bind, true); |
e3a19533 | 9837 | gimple_bind_add_seq (bind, ilist); |
9838 | gimple_bind_add_stmt (bind, stmt); | |
9839 | gimple_bind_add_seq (bind, olist); | |
773c5ba7 | 9840 | |
75a70cf9 | 9841 | pop_gimplify_context (NULL); |
773c5ba7 | 9842 | } |
9843 | ||
bc7bff74 | 9844 | /* Expand code for an OpenMP teams directive. */ |
9845 | ||
9846 | static void | |
9847 | lower_omp_teams (gimple_stmt_iterator *gsi_p, omp_context *ctx) | |
9848 | { | |
9849 | gimple teams_stmt = gsi_stmt (*gsi_p); | |
8a4a28a8 | 9850 | push_gimplify_context (); |
bc7bff74 | 9851 | |
9852 | tree block = make_node (BLOCK); | |
9853 | gimple bind = gimple_build_bind (NULL, NULL, block); | |
9854 | gsi_replace (gsi_p, bind, true); | |
9855 | gimple_seq bind_body = NULL; | |
9856 | gimple_seq dlist = NULL; | |
9857 | gimple_seq olist = NULL; | |
9858 | ||
9859 | tree num_teams = find_omp_clause (gimple_omp_teams_clauses (teams_stmt), | |
9860 | OMP_CLAUSE_NUM_TEAMS); | |
9861 | if (num_teams == NULL_TREE) | |
9862 | num_teams = build_int_cst (unsigned_type_node, 0); | |
9863 | else | |
9864 | { | |
9865 | num_teams = OMP_CLAUSE_NUM_TEAMS_EXPR (num_teams); | |
9866 | num_teams = fold_convert (unsigned_type_node, num_teams); | |
9867 | gimplify_expr (&num_teams, &bind_body, NULL, is_gimple_val, fb_rvalue); | |
9868 | } | |
9869 | tree thread_limit = find_omp_clause (gimple_omp_teams_clauses (teams_stmt), | |
9870 | OMP_CLAUSE_THREAD_LIMIT); | |
9871 | if (thread_limit == NULL_TREE) | |
9872 | thread_limit = build_int_cst (unsigned_type_node, 0); | |
9873 | else | |
9874 | { | |
9875 | thread_limit = OMP_CLAUSE_THREAD_LIMIT_EXPR (thread_limit); | |
9876 | thread_limit = fold_convert (unsigned_type_node, thread_limit); | |
9877 | gimplify_expr (&thread_limit, &bind_body, NULL, is_gimple_val, | |
9878 | fb_rvalue); | |
9879 | } | |
9880 | ||
9881 | lower_rec_input_clauses (gimple_omp_teams_clauses (teams_stmt), | |
9882 | &bind_body, &dlist, ctx, NULL); | |
9883 | lower_omp (gimple_omp_body_ptr (teams_stmt), ctx); | |
9884 | lower_reduction_clauses (gimple_omp_teams_clauses (teams_stmt), &olist, ctx); | |
9885 | gimple_seq_add_stmt (&bind_body, teams_stmt); | |
9886 | ||
9887 | location_t loc = gimple_location (teams_stmt); | |
9888 | tree decl = builtin_decl_explicit (BUILT_IN_GOMP_TEAMS); | |
9889 | gimple call = gimple_build_call (decl, 2, num_teams, thread_limit); | |
9890 | gimple_set_location (call, loc); | |
9891 | gimple_seq_add_stmt (&bind_body, call); | |
9892 | ||
9893 | gimple_seq_add_seq (&bind_body, gimple_omp_body (teams_stmt)); | |
9894 | gimple_omp_set_body (teams_stmt, NULL); | |
9895 | gimple_seq_add_seq (&bind_body, olist); | |
9896 | gimple_seq_add_seq (&bind_body, dlist); | |
9897 | gimple_seq_add_stmt (&bind_body, gimple_build_omp_return (true)); | |
9898 | gimple_bind_set_body (bind, bind_body); | |
9899 | ||
9900 | pop_gimplify_context (bind); | |
9901 | ||
9902 | gimple_bind_append_vars (bind, ctx->block_vars); | |
9903 | BLOCK_VARS (block) = ctx->block_vars; | |
9904 | if (BLOCK_VARS (block)) | |
9905 | TREE_USED (block) = 1; | |
9906 | } | |
9907 | ||
9908 | ||
a4890dc9 | 9909 | /* Callback for lower_omp_1. Return non-NULL if *tp needs to be |
75a70cf9 | 9910 | regimplified. If DATA is non-NULL, lower_omp_1 is outside |
9911 | of OpenMP context, but with task_shared_vars set. */ | |
46515c92 | 9912 | |
9913 | static tree | |
75a70cf9 | 9914 | lower_omp_regimplify_p (tree *tp, int *walk_subtrees, |
9915 | void *data) | |
46515c92 | 9916 | { |
a4890dc9 | 9917 | tree t = *tp; |
46515c92 | 9918 | |
a4890dc9 | 9919 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ |
75a70cf9 | 9920 | if (TREE_CODE (t) == VAR_DECL && data == NULL && DECL_HAS_VALUE_EXPR_P (t)) |
9f49e155 | 9921 | return t; |
9922 | ||
9923 | if (task_shared_vars | |
9924 | && DECL_P (t) | |
9925 | && bitmap_bit_p (task_shared_vars, DECL_UID (t))) | |
a4890dc9 | 9926 | return t; |
46515c92 | 9927 | |
a4890dc9 | 9928 | /* If a global variable has been privatized, TREE_CONSTANT on |
9929 | ADDR_EXPR might be wrong. */ | |
75a70cf9 | 9930 | if (data == NULL && TREE_CODE (t) == ADDR_EXPR) |
a4890dc9 | 9931 | recompute_tree_invariant_for_addr_expr (t); |
46515c92 | 9932 | |
a4890dc9 | 9933 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); |
9934 | return NULL_TREE; | |
46515c92 | 9935 | } |
773c5ba7 | 9936 | |
a4890dc9 | 9937 | static void |
75a70cf9 | 9938 | lower_omp_1 (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 9939 | { |
75a70cf9 | 9940 | gimple stmt = gsi_stmt (*gsi_p); |
9941 | struct walk_stmt_info wi; | |
1e8e9920 | 9942 | |
75a70cf9 | 9943 | if (gimple_has_location (stmt)) |
9944 | input_location = gimple_location (stmt); | |
a4890dc9 | 9945 | |
75a70cf9 | 9946 | if (task_shared_vars) |
9947 | memset (&wi, '\0', sizeof (wi)); | |
a4890dc9 | 9948 | |
773c5ba7 | 9949 | /* If we have issued syntax errors, avoid doing any heavy lifting. |
9950 | Just replace the OpenMP directives with a NOP to avoid | |
9951 | confusing RTL expansion. */ | |
852f689e | 9952 | if (seen_error () && is_gimple_omp (stmt)) |
773c5ba7 | 9953 | { |
75a70cf9 | 9954 | gsi_replace (gsi_p, gimple_build_nop (), true); |
a4890dc9 | 9955 | return; |
773c5ba7 | 9956 | } |
9957 | ||
75a70cf9 | 9958 | switch (gimple_code (stmt)) |
1e8e9920 | 9959 | { |
75a70cf9 | 9960 | case GIMPLE_COND: |
fd6481cf | 9961 | if ((ctx || task_shared_vars) |
75a70cf9 | 9962 | && (walk_tree (gimple_cond_lhs_ptr (stmt), lower_omp_regimplify_p, |
9963 | ctx ? NULL : &wi, NULL) | |
9964 | || walk_tree (gimple_cond_rhs_ptr (stmt), lower_omp_regimplify_p, | |
9965 | ctx ? NULL : &wi, NULL))) | |
9966 | gimple_regimplify_operands (stmt, gsi_p); | |
a4890dc9 | 9967 | break; |
75a70cf9 | 9968 | case GIMPLE_CATCH: |
e3a19533 | 9969 | lower_omp (gimple_catch_handler_ptr (stmt), ctx); |
a4890dc9 | 9970 | break; |
75a70cf9 | 9971 | case GIMPLE_EH_FILTER: |
e3a19533 | 9972 | lower_omp (gimple_eh_filter_failure_ptr (stmt), ctx); |
a4890dc9 | 9973 | break; |
75a70cf9 | 9974 | case GIMPLE_TRY: |
e3a19533 | 9975 | lower_omp (gimple_try_eval_ptr (stmt), ctx); |
9976 | lower_omp (gimple_try_cleanup_ptr (stmt), ctx); | |
a4890dc9 | 9977 | break; |
35215227 | 9978 | case GIMPLE_TRANSACTION: |
9979 | lower_omp (gimple_transaction_body_ptr (stmt), ctx); | |
9980 | break; | |
75a70cf9 | 9981 | case GIMPLE_BIND: |
e3a19533 | 9982 | lower_omp (gimple_bind_body_ptr (stmt), ctx); |
a4890dc9 | 9983 | break; |
75a70cf9 | 9984 | case GIMPLE_OMP_PARALLEL: |
9985 | case GIMPLE_OMP_TASK: | |
9986 | ctx = maybe_lookup_ctx (stmt); | |
bc7bff74 | 9987 | gcc_assert (ctx); |
9988 | if (ctx->cancellable) | |
9989 | ctx->cancel_label = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 9990 | lower_omp_taskreg (gsi_p, ctx); |
a4890dc9 | 9991 | break; |
75a70cf9 | 9992 | case GIMPLE_OMP_FOR: |
9993 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 9994 | gcc_assert (ctx); |
bc7bff74 | 9995 | if (ctx->cancellable) |
9996 | ctx->cancel_label = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 9997 | lower_omp_for (gsi_p, ctx); |
1e8e9920 | 9998 | break; |
75a70cf9 | 9999 | case GIMPLE_OMP_SECTIONS: |
10000 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10001 | gcc_assert (ctx); |
bc7bff74 | 10002 | if (ctx->cancellable) |
10003 | ctx->cancel_label = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 10004 | lower_omp_sections (gsi_p, ctx); |
1e8e9920 | 10005 | break; |
75a70cf9 | 10006 | case GIMPLE_OMP_SINGLE: |
10007 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10008 | gcc_assert (ctx); |
75a70cf9 | 10009 | lower_omp_single (gsi_p, ctx); |
1e8e9920 | 10010 | break; |
75a70cf9 | 10011 | case GIMPLE_OMP_MASTER: |
10012 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10013 | gcc_assert (ctx); |
75a70cf9 | 10014 | lower_omp_master (gsi_p, ctx); |
1e8e9920 | 10015 | break; |
bc7bff74 | 10016 | case GIMPLE_OMP_TASKGROUP: |
10017 | ctx = maybe_lookup_ctx (stmt); | |
10018 | gcc_assert (ctx); | |
10019 | lower_omp_taskgroup (gsi_p, ctx); | |
10020 | break; | |
75a70cf9 | 10021 | case GIMPLE_OMP_ORDERED: |
10022 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10023 | gcc_assert (ctx); |
75a70cf9 | 10024 | lower_omp_ordered (gsi_p, ctx); |
1e8e9920 | 10025 | break; |
75a70cf9 | 10026 | case GIMPLE_OMP_CRITICAL: |
10027 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10028 | gcc_assert (ctx); |
75a70cf9 | 10029 | lower_omp_critical (gsi_p, ctx); |
10030 | break; | |
10031 | case GIMPLE_OMP_ATOMIC_LOAD: | |
10032 | if ((ctx || task_shared_vars) | |
10033 | && walk_tree (gimple_omp_atomic_load_rhs_ptr (stmt), | |
10034 | lower_omp_regimplify_p, ctx ? NULL : &wi, NULL)) | |
10035 | gimple_regimplify_operands (stmt, gsi_p); | |
1e8e9920 | 10036 | break; |
bc7bff74 | 10037 | case GIMPLE_OMP_TARGET: |
10038 | ctx = maybe_lookup_ctx (stmt); | |
10039 | gcc_assert (ctx); | |
10040 | lower_omp_target (gsi_p, ctx); | |
10041 | break; | |
10042 | case GIMPLE_OMP_TEAMS: | |
10043 | ctx = maybe_lookup_ctx (stmt); | |
10044 | gcc_assert (ctx); | |
10045 | lower_omp_teams (gsi_p, ctx); | |
10046 | break; | |
10047 | case GIMPLE_CALL: | |
10048 | tree fndecl; | |
10049 | fndecl = gimple_call_fndecl (stmt); | |
10050 | if (fndecl | |
10051 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
10052 | switch (DECL_FUNCTION_CODE (fndecl)) | |
10053 | { | |
10054 | case BUILT_IN_GOMP_BARRIER: | |
10055 | if (ctx == NULL) | |
10056 | break; | |
10057 | /* FALLTHRU */ | |
10058 | case BUILT_IN_GOMP_CANCEL: | |
10059 | case BUILT_IN_GOMP_CANCELLATION_POINT: | |
10060 | omp_context *cctx; | |
10061 | cctx = ctx; | |
10062 | if (gimple_code (cctx->stmt) == GIMPLE_OMP_SECTION) | |
10063 | cctx = cctx->outer; | |
10064 | gcc_assert (gimple_call_lhs (stmt) == NULL_TREE); | |
10065 | if (!cctx->cancellable) | |
10066 | { | |
10067 | if (DECL_FUNCTION_CODE (fndecl) | |
10068 | == BUILT_IN_GOMP_CANCELLATION_POINT) | |
10069 | { | |
10070 | stmt = gimple_build_nop (); | |
10071 | gsi_replace (gsi_p, stmt, false); | |
10072 | } | |
10073 | break; | |
10074 | } | |
10075 | tree lhs; | |
10076 | lhs = create_tmp_var (boolean_type_node, NULL); | |
10077 | if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_GOMP_BARRIER) | |
10078 | { | |
10079 | fndecl = builtin_decl_explicit (BUILT_IN_GOMP_BARRIER_CANCEL); | |
10080 | gimple_call_set_fndecl (stmt, fndecl); | |
10081 | gimple_call_set_fntype (stmt, TREE_TYPE (fndecl)); | |
10082 | } | |
10083 | gimple_call_set_lhs (stmt, lhs); | |
10084 | tree fallthru_label; | |
10085 | fallthru_label = create_artificial_label (UNKNOWN_LOCATION); | |
10086 | gimple g; | |
10087 | g = gimple_build_label (fallthru_label); | |
10088 | gsi_insert_after (gsi_p, g, GSI_SAME_STMT); | |
10089 | g = gimple_build_cond (NE_EXPR, lhs, boolean_false_node, | |
10090 | cctx->cancel_label, fallthru_label); | |
10091 | gsi_insert_after (gsi_p, g, GSI_SAME_STMT); | |
10092 | break; | |
10093 | default: | |
10094 | break; | |
10095 | } | |
10096 | /* FALLTHRU */ | |
a4890dc9 | 10097 | default: |
fd6481cf | 10098 | if ((ctx || task_shared_vars) |
75a70cf9 | 10099 | && walk_gimple_op (stmt, lower_omp_regimplify_p, |
10100 | ctx ? NULL : &wi)) | |
10101 | gimple_regimplify_operands (stmt, gsi_p); | |
1e8e9920 | 10102 | break; |
1e8e9920 | 10103 | } |
1e8e9920 | 10104 | } |
10105 | ||
10106 | static void | |
e3a19533 | 10107 | lower_omp (gimple_seq *body, omp_context *ctx) |
1e8e9920 | 10108 | { |
1d22f541 | 10109 | location_t saved_location = input_location; |
e3a19533 | 10110 | gimple_stmt_iterator gsi; |
10111 | for (gsi = gsi_start (*body); !gsi_end_p (gsi); gsi_next (&gsi)) | |
75a70cf9 | 10112 | lower_omp_1 (&gsi, ctx); |
bc7bff74 | 10113 | /* Inside target region we haven't called fold_stmt during gimplification, |
10114 | because it can break code by adding decl references that weren't in the | |
10115 | source. Call fold_stmt now. */ | |
10116 | if (target_nesting_level) | |
10117 | for (gsi = gsi_start (*body); !gsi_end_p (gsi); gsi_next (&gsi)) | |
10118 | fold_stmt (&gsi); | |
1d22f541 | 10119 | input_location = saved_location; |
1e8e9920 | 10120 | } |
10121 | \f | |
10122 | /* Main entry point. */ | |
10123 | ||
2a1990e9 | 10124 | static unsigned int |
1e8e9920 | 10125 | execute_lower_omp (void) |
10126 | { | |
75a70cf9 | 10127 | gimple_seq body; |
10128 | ||
41709826 | 10129 | /* This pass always runs, to provide PROP_gimple_lomp. |
10130 | But there is nothing to do unless -fopenmp is given. */ | |
f2697631 | 10131 | if (flag_openmp == 0 && flag_openmp_simd == 0 && flag_enable_cilkplus == 0) |
41709826 | 10132 | return 0; |
10133 | ||
1e8e9920 | 10134 | all_contexts = splay_tree_new (splay_tree_compare_pointers, 0, |
10135 | delete_omp_context); | |
10136 | ||
75a70cf9 | 10137 | body = gimple_body (current_function_decl); |
ab129075 | 10138 | scan_omp (&body, NULL); |
fd6481cf | 10139 | gcc_assert (taskreg_nesting_level == 0); |
1e8e9920 | 10140 | |
10141 | if (all_contexts->root) | |
fd6481cf | 10142 | { |
10143 | if (task_shared_vars) | |
8a4a28a8 | 10144 | push_gimplify_context (); |
e3a19533 | 10145 | lower_omp (&body, NULL); |
fd6481cf | 10146 | if (task_shared_vars) |
10147 | pop_gimplify_context (NULL); | |
10148 | } | |
1e8e9920 | 10149 | |
773c5ba7 | 10150 | if (all_contexts) |
10151 | { | |
10152 | splay_tree_delete (all_contexts); | |
10153 | all_contexts = NULL; | |
10154 | } | |
fd6481cf | 10155 | BITMAP_FREE (task_shared_vars); |
2a1990e9 | 10156 | return 0; |
1e8e9920 | 10157 | } |
10158 | ||
cbe8bda8 | 10159 | namespace { |
10160 | ||
10161 | const pass_data pass_data_lower_omp = | |
10162 | { | |
10163 | GIMPLE_PASS, /* type */ | |
10164 | "omplower", /* name */ | |
10165 | OPTGROUP_NONE, /* optinfo_flags */ | |
10166 | false, /* has_gate */ | |
10167 | true, /* has_execute */ | |
10168 | TV_NONE, /* tv_id */ | |
10169 | PROP_gimple_any, /* properties_required */ | |
10170 | PROP_gimple_lomp, /* properties_provided */ | |
10171 | 0, /* properties_destroyed */ | |
10172 | 0, /* todo_flags_start */ | |
10173 | 0, /* todo_flags_finish */ | |
1e8e9920 | 10174 | }; |
cbe8bda8 | 10175 | |
10176 | class pass_lower_omp : public gimple_opt_pass | |
10177 | { | |
10178 | public: | |
9af5ce0c | 10179 | pass_lower_omp (gcc::context *ctxt) |
10180 | : gimple_opt_pass (pass_data_lower_omp, ctxt) | |
cbe8bda8 | 10181 | {} |
10182 | ||
10183 | /* opt_pass methods: */ | |
10184 | unsigned int execute () { return execute_lower_omp (); } | |
10185 | ||
10186 | }; // class pass_lower_omp | |
10187 | ||
10188 | } // anon namespace | |
10189 | ||
10190 | gimple_opt_pass * | |
10191 | make_pass_lower_omp (gcc::context *ctxt) | |
10192 | { | |
10193 | return new pass_lower_omp (ctxt); | |
10194 | } | |
1e8e9920 | 10195 | \f |
10196 | /* The following is a utility to diagnose OpenMP structured block violations. | |
61e47ac8 | 10197 | It is not part of the "omplower" pass, as that's invoked too late. It |
10198 | should be invoked by the respective front ends after gimplification. */ | |
1e8e9920 | 10199 | |
10200 | static splay_tree all_labels; | |
10201 | ||
10202 | /* Check for mismatched contexts and generate an error if needed. Return | |
10203 | true if an error is detected. */ | |
10204 | ||
10205 | static bool | |
75a70cf9 | 10206 | diagnose_sb_0 (gimple_stmt_iterator *gsi_p, |
10207 | gimple branch_ctx, gimple label_ctx) | |
1e8e9920 | 10208 | { |
75a70cf9 | 10209 | if (label_ctx == branch_ctx) |
1e8e9920 | 10210 | return false; |
10211 | ||
48e1416a | 10212 | |
75a70cf9 | 10213 | /* |
10214 | Previously we kept track of the label's entire context in diagnose_sb_[12] | |
10215 | so we could traverse it and issue a correct "exit" or "enter" error | |
10216 | message upon a structured block violation. | |
10217 | ||
10218 | We built the context by building a list with tree_cons'ing, but there is | |
10219 | no easy counterpart in gimple tuples. It seems like far too much work | |
10220 | for issuing exit/enter error messages. If someone really misses the | |
10221 | distinct error message... patches welcome. | |
10222 | */ | |
48e1416a | 10223 | |
75a70cf9 | 10224 | #if 0 |
1e8e9920 | 10225 | /* Try to avoid confusing the user by producing and error message |
f0b5f617 | 10226 | with correct "exit" or "enter" verbiage. We prefer "exit" |
1e8e9920 | 10227 | unless we can show that LABEL_CTX is nested within BRANCH_CTX. */ |
10228 | if (branch_ctx == NULL) | |
10229 | exit_p = false; | |
10230 | else | |
10231 | { | |
10232 | while (label_ctx) | |
10233 | { | |
10234 | if (TREE_VALUE (label_ctx) == branch_ctx) | |
10235 | { | |
10236 | exit_p = false; | |
10237 | break; | |
10238 | } | |
10239 | label_ctx = TREE_CHAIN (label_ctx); | |
10240 | } | |
10241 | } | |
10242 | ||
10243 | if (exit_p) | |
10244 | error ("invalid exit from OpenMP structured block"); | |
10245 | else | |
10246 | error ("invalid entry to OpenMP structured block"); | |
75a70cf9 | 10247 | #endif |
1e8e9920 | 10248 | |
f2697631 | 10249 | bool cilkplus_block = false; |
10250 | if (flag_enable_cilkplus) | |
10251 | { | |
10252 | if ((branch_ctx | |
10253 | && gimple_code (branch_ctx) == GIMPLE_OMP_FOR | |
10254 | && gimple_omp_for_kind (branch_ctx) == GF_OMP_FOR_KIND_CILKSIMD) | |
10255 | || (gimple_code (label_ctx) == GIMPLE_OMP_FOR | |
10256 | && gimple_omp_for_kind (label_ctx) == GF_OMP_FOR_KIND_CILKSIMD)) | |
10257 | cilkplus_block = true; | |
10258 | } | |
10259 | ||
75a70cf9 | 10260 | /* If it's obvious we have an invalid entry, be specific about the error. */ |
10261 | if (branch_ctx == NULL) | |
f2697631 | 10262 | { |
10263 | if (cilkplus_block) | |
10264 | error ("invalid entry to Cilk Plus structured block"); | |
10265 | else | |
10266 | error ("invalid entry to OpenMP structured block"); | |
10267 | } | |
75a70cf9 | 10268 | else |
f2697631 | 10269 | { |
10270 | /* Otherwise, be vague and lazy, but efficient. */ | |
10271 | if (cilkplus_block) | |
10272 | error ("invalid branch to/from a Cilk Plus structured block"); | |
10273 | else | |
10274 | error ("invalid branch to/from an OpenMP structured block"); | |
10275 | } | |
75a70cf9 | 10276 | |
10277 | gsi_replace (gsi_p, gimple_build_nop (), false); | |
1e8e9920 | 10278 | return true; |
10279 | } | |
10280 | ||
10281 | /* Pass 1: Create a minimal tree of OpenMP structured blocks, and record | |
75a70cf9 | 10282 | where each label is found. */ |
1e8e9920 | 10283 | |
10284 | static tree | |
75a70cf9 | 10285 | diagnose_sb_1 (gimple_stmt_iterator *gsi_p, bool *handled_ops_p, |
10286 | struct walk_stmt_info *wi) | |
1e8e9920 | 10287 | { |
75a70cf9 | 10288 | gimple context = (gimple) wi->info; |
10289 | gimple inner_context; | |
10290 | gimple stmt = gsi_stmt (*gsi_p); | |
1e8e9920 | 10291 | |
75a70cf9 | 10292 | *handled_ops_p = true; |
10293 | ||
10294 | switch (gimple_code (stmt)) | |
1e8e9920 | 10295 | { |
75a70cf9 | 10296 | WALK_SUBSTMTS; |
48e1416a | 10297 | |
75a70cf9 | 10298 | case GIMPLE_OMP_PARALLEL: |
10299 | case GIMPLE_OMP_TASK: | |
10300 | case GIMPLE_OMP_SECTIONS: | |
10301 | case GIMPLE_OMP_SINGLE: | |
10302 | case GIMPLE_OMP_SECTION: | |
10303 | case GIMPLE_OMP_MASTER: | |
10304 | case GIMPLE_OMP_ORDERED: | |
10305 | case GIMPLE_OMP_CRITICAL: | |
bc7bff74 | 10306 | case GIMPLE_OMP_TARGET: |
10307 | case GIMPLE_OMP_TEAMS: | |
10308 | case GIMPLE_OMP_TASKGROUP: | |
75a70cf9 | 10309 | /* The minimal context here is just the current OMP construct. */ |
10310 | inner_context = stmt; | |
1e8e9920 | 10311 | wi->info = inner_context; |
75a70cf9 | 10312 | walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_1, NULL, wi); |
1e8e9920 | 10313 | wi->info = context; |
10314 | break; | |
10315 | ||
75a70cf9 | 10316 | case GIMPLE_OMP_FOR: |
10317 | inner_context = stmt; | |
1e8e9920 | 10318 | wi->info = inner_context; |
75a70cf9 | 10319 | /* gimple_omp_for_{index,initial,final} are all DECLs; no need to |
10320 | walk them. */ | |
10321 | walk_gimple_seq (gimple_omp_for_pre_body (stmt), | |
10322 | diagnose_sb_1, NULL, wi); | |
10323 | walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_1, NULL, wi); | |
1e8e9920 | 10324 | wi->info = context; |
10325 | break; | |
10326 | ||
75a70cf9 | 10327 | case GIMPLE_LABEL: |
10328 | splay_tree_insert (all_labels, (splay_tree_key) gimple_label_label (stmt), | |
1e8e9920 | 10329 | (splay_tree_value) context); |
10330 | break; | |
10331 | ||
10332 | default: | |
10333 | break; | |
10334 | } | |
10335 | ||
10336 | return NULL_TREE; | |
10337 | } | |
10338 | ||
10339 | /* Pass 2: Check each branch and see if its context differs from that of | |
10340 | the destination label's context. */ | |
10341 | ||
10342 | static tree | |
75a70cf9 | 10343 | diagnose_sb_2 (gimple_stmt_iterator *gsi_p, bool *handled_ops_p, |
10344 | struct walk_stmt_info *wi) | |
1e8e9920 | 10345 | { |
75a70cf9 | 10346 | gimple context = (gimple) wi->info; |
1e8e9920 | 10347 | splay_tree_node n; |
75a70cf9 | 10348 | gimple stmt = gsi_stmt (*gsi_p); |
1e8e9920 | 10349 | |
75a70cf9 | 10350 | *handled_ops_p = true; |
10351 | ||
10352 | switch (gimple_code (stmt)) | |
1e8e9920 | 10353 | { |
75a70cf9 | 10354 | WALK_SUBSTMTS; |
10355 | ||
10356 | case GIMPLE_OMP_PARALLEL: | |
10357 | case GIMPLE_OMP_TASK: | |
10358 | case GIMPLE_OMP_SECTIONS: | |
10359 | case GIMPLE_OMP_SINGLE: | |
10360 | case GIMPLE_OMP_SECTION: | |
10361 | case GIMPLE_OMP_MASTER: | |
10362 | case GIMPLE_OMP_ORDERED: | |
10363 | case GIMPLE_OMP_CRITICAL: | |
bc7bff74 | 10364 | case GIMPLE_OMP_TARGET: |
10365 | case GIMPLE_OMP_TEAMS: | |
10366 | case GIMPLE_OMP_TASKGROUP: | |
75a70cf9 | 10367 | wi->info = stmt; |
e3a19533 | 10368 | walk_gimple_seq_mod (gimple_omp_body_ptr (stmt), diagnose_sb_2, NULL, wi); |
1e8e9920 | 10369 | wi->info = context; |
10370 | break; | |
10371 | ||
75a70cf9 | 10372 | case GIMPLE_OMP_FOR: |
10373 | wi->info = stmt; | |
10374 | /* gimple_omp_for_{index,initial,final} are all DECLs; no need to | |
10375 | walk them. */ | |
e3a19533 | 10376 | walk_gimple_seq_mod (gimple_omp_for_pre_body_ptr (stmt), |
10377 | diagnose_sb_2, NULL, wi); | |
10378 | walk_gimple_seq_mod (gimple_omp_body_ptr (stmt), diagnose_sb_2, NULL, wi); | |
1e8e9920 | 10379 | wi->info = context; |
10380 | break; | |
10381 | ||
0e1818e7 | 10382 | case GIMPLE_COND: |
10383 | { | |
10384 | tree lab = gimple_cond_true_label (stmt); | |
10385 | if (lab) | |
10386 | { | |
10387 | n = splay_tree_lookup (all_labels, | |
10388 | (splay_tree_key) lab); | |
10389 | diagnose_sb_0 (gsi_p, context, | |
10390 | n ? (gimple) n->value : NULL); | |
10391 | } | |
10392 | lab = gimple_cond_false_label (stmt); | |
10393 | if (lab) | |
10394 | { | |
10395 | n = splay_tree_lookup (all_labels, | |
10396 | (splay_tree_key) lab); | |
10397 | diagnose_sb_0 (gsi_p, context, | |
10398 | n ? (gimple) n->value : NULL); | |
10399 | } | |
10400 | } | |
10401 | break; | |
10402 | ||
75a70cf9 | 10403 | case GIMPLE_GOTO: |
1e8e9920 | 10404 | { |
75a70cf9 | 10405 | tree lab = gimple_goto_dest (stmt); |
1e8e9920 | 10406 | if (TREE_CODE (lab) != LABEL_DECL) |
10407 | break; | |
10408 | ||
10409 | n = splay_tree_lookup (all_labels, (splay_tree_key) lab); | |
75a70cf9 | 10410 | diagnose_sb_0 (gsi_p, context, n ? (gimple) n->value : NULL); |
1e8e9920 | 10411 | } |
10412 | break; | |
10413 | ||
75a70cf9 | 10414 | case GIMPLE_SWITCH: |
1e8e9920 | 10415 | { |
75a70cf9 | 10416 | unsigned int i; |
10417 | for (i = 0; i < gimple_switch_num_labels (stmt); ++i) | |
1e8e9920 | 10418 | { |
75a70cf9 | 10419 | tree lab = CASE_LABEL (gimple_switch_label (stmt, i)); |
1e8e9920 | 10420 | n = splay_tree_lookup (all_labels, (splay_tree_key) lab); |
75a70cf9 | 10421 | if (n && diagnose_sb_0 (gsi_p, context, (gimple) n->value)) |
1e8e9920 | 10422 | break; |
10423 | } | |
10424 | } | |
10425 | break; | |
10426 | ||
75a70cf9 | 10427 | case GIMPLE_RETURN: |
10428 | diagnose_sb_0 (gsi_p, context, NULL); | |
1e8e9920 | 10429 | break; |
10430 | ||
10431 | default: | |
10432 | break; | |
10433 | } | |
10434 | ||
10435 | return NULL_TREE; | |
10436 | } | |
10437 | ||
7740abd8 | 10438 | /* Called from tree-cfg.c::make_edges to create cfg edges for all GIMPLE_OMP |
10439 | codes. */ | |
10440 | bool | |
10441 | make_gimple_omp_edges (basic_block bb, struct omp_region **region) | |
10442 | { | |
10443 | gimple last = last_stmt (bb); | |
10444 | enum gimple_code code = gimple_code (last); | |
10445 | struct omp_region *cur_region = *region; | |
10446 | bool fallthru = false; | |
10447 | ||
10448 | switch (code) | |
10449 | { | |
10450 | case GIMPLE_OMP_PARALLEL: | |
10451 | case GIMPLE_OMP_TASK: | |
10452 | case GIMPLE_OMP_FOR: | |
10453 | case GIMPLE_OMP_SINGLE: | |
10454 | case GIMPLE_OMP_TEAMS: | |
10455 | case GIMPLE_OMP_MASTER: | |
10456 | case GIMPLE_OMP_TASKGROUP: | |
10457 | case GIMPLE_OMP_ORDERED: | |
10458 | case GIMPLE_OMP_CRITICAL: | |
10459 | case GIMPLE_OMP_SECTION: | |
10460 | cur_region = new_omp_region (bb, code, cur_region); | |
10461 | fallthru = true; | |
10462 | break; | |
10463 | ||
10464 | case GIMPLE_OMP_TARGET: | |
10465 | cur_region = new_omp_region (bb, code, cur_region); | |
10466 | fallthru = true; | |
10467 | if (gimple_omp_target_kind (last) == GF_OMP_TARGET_KIND_UPDATE) | |
10468 | cur_region = cur_region->outer; | |
10469 | break; | |
10470 | ||
10471 | case GIMPLE_OMP_SECTIONS: | |
10472 | cur_region = new_omp_region (bb, code, cur_region); | |
10473 | fallthru = true; | |
10474 | break; | |
10475 | ||
10476 | case GIMPLE_OMP_SECTIONS_SWITCH: | |
10477 | fallthru = false; | |
10478 | break; | |
10479 | ||
10480 | case GIMPLE_OMP_ATOMIC_LOAD: | |
10481 | case GIMPLE_OMP_ATOMIC_STORE: | |
10482 | fallthru = true; | |
10483 | break; | |
10484 | ||
10485 | case GIMPLE_OMP_RETURN: | |
10486 | /* In the case of a GIMPLE_OMP_SECTION, the edge will go | |
10487 | somewhere other than the next block. This will be | |
10488 | created later. */ | |
10489 | cur_region->exit = bb; | |
10490 | fallthru = cur_region->type != GIMPLE_OMP_SECTION; | |
10491 | cur_region = cur_region->outer; | |
10492 | break; | |
10493 | ||
10494 | case GIMPLE_OMP_CONTINUE: | |
10495 | cur_region->cont = bb; | |
10496 | switch (cur_region->type) | |
10497 | { | |
10498 | case GIMPLE_OMP_FOR: | |
10499 | /* Mark all GIMPLE_OMP_FOR and GIMPLE_OMP_CONTINUE | |
10500 | succs edges as abnormal to prevent splitting | |
10501 | them. */ | |
10502 | single_succ_edge (cur_region->entry)->flags |= EDGE_ABNORMAL; | |
10503 | /* Make the loopback edge. */ | |
10504 | make_edge (bb, single_succ (cur_region->entry), | |
10505 | EDGE_ABNORMAL); | |
10506 | ||
10507 | /* Create an edge from GIMPLE_OMP_FOR to exit, which | |
10508 | corresponds to the case that the body of the loop | |
10509 | is not executed at all. */ | |
10510 | make_edge (cur_region->entry, bb->next_bb, EDGE_ABNORMAL); | |
10511 | make_edge (bb, bb->next_bb, EDGE_FALLTHRU | EDGE_ABNORMAL); | |
10512 | fallthru = false; | |
10513 | break; | |
10514 | ||
10515 | case GIMPLE_OMP_SECTIONS: | |
10516 | /* Wire up the edges into and out of the nested sections. */ | |
10517 | { | |
10518 | basic_block switch_bb = single_succ (cur_region->entry); | |
10519 | ||
10520 | struct omp_region *i; | |
10521 | for (i = cur_region->inner; i ; i = i->next) | |
10522 | { | |
10523 | gcc_assert (i->type == GIMPLE_OMP_SECTION); | |
10524 | make_edge (switch_bb, i->entry, 0); | |
10525 | make_edge (i->exit, bb, EDGE_FALLTHRU); | |
10526 | } | |
10527 | ||
10528 | /* Make the loopback edge to the block with | |
10529 | GIMPLE_OMP_SECTIONS_SWITCH. */ | |
10530 | make_edge (bb, switch_bb, 0); | |
10531 | ||
10532 | /* Make the edge from the switch to exit. */ | |
10533 | make_edge (switch_bb, bb->next_bb, 0); | |
10534 | fallthru = false; | |
10535 | } | |
10536 | break; | |
10537 | ||
10538 | default: | |
10539 | gcc_unreachable (); | |
10540 | } | |
10541 | break; | |
10542 | ||
10543 | default: | |
10544 | gcc_unreachable (); | |
10545 | } | |
10546 | ||
10547 | if (*region != cur_region) | |
10548 | *region = cur_region; | |
10549 | ||
10550 | return fallthru; | |
10551 | } | |
10552 | ||
bfec3452 | 10553 | static unsigned int |
10554 | diagnose_omp_structured_block_errors (void) | |
1e8e9920 | 10555 | { |
1e8e9920 | 10556 | struct walk_stmt_info wi; |
bfec3452 | 10557 | gimple_seq body = gimple_body (current_function_decl); |
1e8e9920 | 10558 | |
10559 | all_labels = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
10560 | ||
10561 | memset (&wi, 0, sizeof (wi)); | |
75a70cf9 | 10562 | walk_gimple_seq (body, diagnose_sb_1, NULL, &wi); |
1e8e9920 | 10563 | |
10564 | memset (&wi, 0, sizeof (wi)); | |
1e8e9920 | 10565 | wi.want_locations = true; |
e3a19533 | 10566 | walk_gimple_seq_mod (&body, diagnose_sb_2, NULL, &wi); |
10567 | ||
10568 | gimple_set_body (current_function_decl, body); | |
1e8e9920 | 10569 | |
10570 | splay_tree_delete (all_labels); | |
10571 | all_labels = NULL; | |
10572 | ||
bfec3452 | 10573 | return 0; |
1e8e9920 | 10574 | } |
10575 | ||
bfec3452 | 10576 | static bool |
10577 | gate_diagnose_omp_blocks (void) | |
10578 | { | |
f2697631 | 10579 | return flag_openmp || flag_enable_cilkplus; |
bfec3452 | 10580 | } |
10581 | ||
cbe8bda8 | 10582 | namespace { |
10583 | ||
10584 | const pass_data pass_data_diagnose_omp_blocks = | |
10585 | { | |
10586 | GIMPLE_PASS, /* type */ | |
10587 | "*diagnose_omp_blocks", /* name */ | |
10588 | OPTGROUP_NONE, /* optinfo_flags */ | |
10589 | true, /* has_gate */ | |
10590 | true, /* has_execute */ | |
10591 | TV_NONE, /* tv_id */ | |
10592 | PROP_gimple_any, /* properties_required */ | |
10593 | 0, /* properties_provided */ | |
10594 | 0, /* properties_destroyed */ | |
10595 | 0, /* todo_flags_start */ | |
10596 | 0, /* todo_flags_finish */ | |
bfec3452 | 10597 | }; |
10598 | ||
cbe8bda8 | 10599 | class pass_diagnose_omp_blocks : public gimple_opt_pass |
10600 | { | |
10601 | public: | |
9af5ce0c | 10602 | pass_diagnose_omp_blocks (gcc::context *ctxt) |
10603 | : gimple_opt_pass (pass_data_diagnose_omp_blocks, ctxt) | |
cbe8bda8 | 10604 | {} |
10605 | ||
10606 | /* opt_pass methods: */ | |
10607 | bool gate () { return gate_diagnose_omp_blocks (); } | |
10608 | unsigned int execute () { | |
10609 | return diagnose_omp_structured_block_errors (); | |
10610 | } | |
10611 | ||
10612 | }; // class pass_diagnose_omp_blocks | |
10613 | ||
10614 | } // anon namespace | |
10615 | ||
10616 | gimple_opt_pass * | |
10617 | make_pass_diagnose_omp_blocks (gcc::context *ctxt) | |
10618 | { | |
10619 | return new pass_diagnose_omp_blocks (ctxt); | |
10620 | } | |
d09768a4 | 10621 | \f |
10622 | /* SIMD clone supporting code. */ | |
10623 | ||
10624 | /* Allocate a fresh `simd_clone' and return it. NARGS is the number | |
10625 | of arguments to reserve space for. */ | |
10626 | ||
10627 | static struct cgraph_simd_clone * | |
10628 | simd_clone_struct_alloc (int nargs) | |
10629 | { | |
10630 | struct cgraph_simd_clone *clone_info; | |
10631 | size_t len = (sizeof (struct cgraph_simd_clone) | |
10632 | + nargs * sizeof (struct cgraph_simd_clone_arg)); | |
10633 | clone_info = (struct cgraph_simd_clone *) | |
10634 | ggc_internal_cleared_alloc_stat (len PASS_MEM_STAT); | |
10635 | return clone_info; | |
10636 | } | |
10637 | ||
10638 | /* Make a copy of the `struct cgraph_simd_clone' in FROM to TO. */ | |
10639 | ||
10640 | static inline void | |
10641 | simd_clone_struct_copy (struct cgraph_simd_clone *to, | |
10642 | struct cgraph_simd_clone *from) | |
10643 | { | |
10644 | memcpy (to, from, (sizeof (struct cgraph_simd_clone) | |
10645 | + from->nargs * sizeof (struct cgraph_simd_clone_arg))); | |
10646 | } | |
10647 | ||
10648 | /* Return vector of parameter types of function FNDECL. This uses | |
10649 | TYPE_ARG_TYPES if available, otherwise falls back to types of | |
10650 | DECL_ARGUMENTS types. */ | |
10651 | ||
10652 | vec<tree> | |
10653 | simd_clone_vector_of_formal_parm_types (tree fndecl) | |
10654 | { | |
10655 | if (TYPE_ARG_TYPES (TREE_TYPE (fndecl))) | |
10656 | return ipa_get_vector_of_formal_parm_types (TREE_TYPE (fndecl)); | |
10657 | vec<tree> args = ipa_get_vector_of_formal_parms (fndecl); | |
10658 | unsigned int i; | |
10659 | tree arg; | |
10660 | FOR_EACH_VEC_ELT (args, i, arg) | |
10661 | args[i] = TREE_TYPE (args[i]); | |
10662 | return args; | |
10663 | } | |
10664 | ||
10665 | /* Given a simd function in NODE, extract the simd specific | |
10666 | information from the OMP clauses passed in CLAUSES, and return | |
10667 | the struct cgraph_simd_clone * if it should be cloned. *INBRANCH_SPECIFIED | |
10668 | is set to TRUE if the `inbranch' or `notinbranch' clause specified, | |
10669 | otherwise set to FALSE. */ | |
10670 | ||
10671 | static struct cgraph_simd_clone * | |
10672 | simd_clone_clauses_extract (struct cgraph_node *node, tree clauses, | |
10673 | bool *inbranch_specified) | |
10674 | { | |
10675 | vec<tree> args = simd_clone_vector_of_formal_parm_types (node->decl); | |
10676 | tree t; | |
10677 | int n; | |
10678 | *inbranch_specified = false; | |
10679 | ||
10680 | n = args.length (); | |
10681 | if (n > 0 && args.last () == void_type_node) | |
10682 | n--; | |
10683 | ||
10684 | /* To distinguish from an OpenMP simd clone, Cilk Plus functions to | |
10685 | be cloned have a distinctive artificial label in addition to "omp | |
10686 | declare simd". */ | |
10687 | bool cilk_clone | |
10688 | = (flag_enable_cilkplus | |
10689 | && lookup_attribute ("cilk plus elemental", | |
10690 | DECL_ATTRIBUTES (node->decl))); | |
10691 | ||
10692 | /* Allocate one more than needed just in case this is an in-branch | |
10693 | clone which will require a mask argument. */ | |
10694 | struct cgraph_simd_clone *clone_info = simd_clone_struct_alloc (n + 1); | |
10695 | clone_info->nargs = n; | |
10696 | clone_info->cilk_elemental = cilk_clone; | |
10697 | ||
10698 | if (!clauses) | |
10699 | { | |
10700 | args.release (); | |
10701 | return clone_info; | |
10702 | } | |
10703 | clauses = TREE_VALUE (clauses); | |
10704 | if (!clauses || TREE_CODE (clauses) != OMP_CLAUSE) | |
10705 | return clone_info; | |
10706 | ||
10707 | for (t = clauses; t; t = OMP_CLAUSE_CHAIN (t)) | |
10708 | { | |
10709 | switch (OMP_CLAUSE_CODE (t)) | |
10710 | { | |
10711 | case OMP_CLAUSE_INBRANCH: | |
10712 | clone_info->inbranch = 1; | |
10713 | *inbranch_specified = true; | |
10714 | break; | |
10715 | case OMP_CLAUSE_NOTINBRANCH: | |
10716 | clone_info->inbranch = 0; | |
10717 | *inbranch_specified = true; | |
10718 | break; | |
10719 | case OMP_CLAUSE_SIMDLEN: | |
10720 | clone_info->simdlen | |
10721 | = TREE_INT_CST_LOW (OMP_CLAUSE_SIMDLEN_EXPR (t)); | |
10722 | break; | |
10723 | case OMP_CLAUSE_LINEAR: | |
10724 | { | |
10725 | tree decl = OMP_CLAUSE_DECL (t); | |
10726 | tree step = OMP_CLAUSE_LINEAR_STEP (t); | |
10727 | int argno = TREE_INT_CST_LOW (decl); | |
10728 | if (OMP_CLAUSE_LINEAR_VARIABLE_STRIDE (t)) | |
10729 | { | |
10730 | clone_info->args[argno].arg_type | |
10731 | = SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP; | |
10732 | clone_info->args[argno].linear_step = tree_to_shwi (step); | |
10733 | gcc_assert (clone_info->args[argno].linear_step >= 0 | |
10734 | && clone_info->args[argno].linear_step < n); | |
10735 | } | |
10736 | else | |
10737 | { | |
10738 | if (POINTER_TYPE_P (args[argno])) | |
10739 | step = fold_convert (ssizetype, step); | |
10740 | if (!tree_fits_shwi_p (step)) | |
10741 | { | |
10742 | warning_at (OMP_CLAUSE_LOCATION (t), 0, | |
10743 | "ignoring large linear step"); | |
10744 | args.release (); | |
10745 | return NULL; | |
10746 | } | |
10747 | else if (integer_zerop (step)) | |
10748 | { | |
10749 | warning_at (OMP_CLAUSE_LOCATION (t), 0, | |
10750 | "ignoring zero linear step"); | |
10751 | args.release (); | |
10752 | return NULL; | |
10753 | } | |
10754 | else | |
10755 | { | |
10756 | clone_info->args[argno].arg_type | |
10757 | = SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP; | |
10758 | clone_info->args[argno].linear_step = tree_to_shwi (step); | |
10759 | } | |
10760 | } | |
10761 | break; | |
10762 | } | |
10763 | case OMP_CLAUSE_UNIFORM: | |
10764 | { | |
10765 | tree decl = OMP_CLAUSE_DECL (t); | |
10766 | int argno = tree_to_uhwi (decl); | |
10767 | clone_info->args[argno].arg_type | |
10768 | = SIMD_CLONE_ARG_TYPE_UNIFORM; | |
10769 | break; | |
10770 | } | |
10771 | case OMP_CLAUSE_ALIGNED: | |
10772 | { | |
10773 | tree decl = OMP_CLAUSE_DECL (t); | |
10774 | int argno = tree_to_uhwi (decl); | |
10775 | clone_info->args[argno].alignment | |
10776 | = TREE_INT_CST_LOW (OMP_CLAUSE_ALIGNED_ALIGNMENT (t)); | |
10777 | break; | |
10778 | } | |
10779 | default: | |
10780 | break; | |
10781 | } | |
10782 | } | |
10783 | args.release (); | |
10784 | return clone_info; | |
10785 | } | |
10786 | ||
10787 | /* Given a SIMD clone in NODE, calculate the characteristic data | |
10788 | type and return the coresponding type. The characteristic data | |
10789 | type is computed as described in the Intel Vector ABI. */ | |
10790 | ||
10791 | static tree | |
10792 | simd_clone_compute_base_data_type (struct cgraph_node *node, | |
10793 | struct cgraph_simd_clone *clone_info) | |
10794 | { | |
10795 | tree type = integer_type_node; | |
10796 | tree fndecl = node->decl; | |
10797 | ||
10798 | /* a) For non-void function, the characteristic data type is the | |
10799 | return type. */ | |
10800 | if (TREE_CODE (TREE_TYPE (TREE_TYPE (fndecl))) != VOID_TYPE) | |
10801 | type = TREE_TYPE (TREE_TYPE (fndecl)); | |
10802 | ||
10803 | /* b) If the function has any non-uniform, non-linear parameters, | |
10804 | then the characteristic data type is the type of the first | |
10805 | such parameter. */ | |
10806 | else | |
10807 | { | |
10808 | vec<tree> map = simd_clone_vector_of_formal_parm_types (fndecl); | |
10809 | for (unsigned int i = 0; i < clone_info->nargs; ++i) | |
10810 | if (clone_info->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
10811 | { | |
10812 | type = map[i]; | |
10813 | break; | |
10814 | } | |
10815 | map.release (); | |
10816 | } | |
10817 | ||
10818 | /* c) If the characteristic data type determined by a) or b) above | |
10819 | is struct, union, or class type which is pass-by-value (except | |
10820 | for the type that maps to the built-in complex data type), the | |
10821 | characteristic data type is int. */ | |
10822 | if (RECORD_OR_UNION_TYPE_P (type) | |
10823 | && !aggregate_value_p (type, NULL) | |
10824 | && TREE_CODE (type) != COMPLEX_TYPE) | |
10825 | return integer_type_node; | |
10826 | ||
10827 | /* d) If none of the above three classes is applicable, the | |
10828 | characteristic data type is int. */ | |
10829 | ||
10830 | return type; | |
10831 | ||
10832 | /* e) For Intel Xeon Phi native and offload compilation, if the | |
10833 | resulting characteristic data type is 8-bit or 16-bit integer | |
10834 | data type, the characteristic data type is int. */ | |
10835 | /* Well, we don't handle Xeon Phi yet. */ | |
10836 | } | |
10837 | ||
10838 | static tree | |
10839 | simd_clone_mangle (struct cgraph_node *node, | |
10840 | struct cgraph_simd_clone *clone_info) | |
10841 | { | |
10842 | char vecsize_mangle = clone_info->vecsize_mangle; | |
10843 | char mask = clone_info->inbranch ? 'M' : 'N'; | |
10844 | unsigned int simdlen = clone_info->simdlen; | |
10845 | unsigned int n; | |
10846 | pretty_printer pp; | |
10847 | ||
10848 | gcc_assert (vecsize_mangle && simdlen); | |
10849 | ||
10850 | pp_string (&pp, "_ZGV"); | |
10851 | pp_character (&pp, vecsize_mangle); | |
10852 | pp_character (&pp, mask); | |
10853 | pp_decimal_int (&pp, simdlen); | |
10854 | ||
10855 | for (n = 0; n < clone_info->nargs; ++n) | |
10856 | { | |
10857 | struct cgraph_simd_clone_arg arg = clone_info->args[n]; | |
10858 | ||
10859 | if (arg.arg_type == SIMD_CLONE_ARG_TYPE_UNIFORM) | |
10860 | pp_character (&pp, 'u'); | |
10861 | else if (arg.arg_type == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
10862 | { | |
10863 | gcc_assert (arg.linear_step != 0); | |
10864 | pp_character (&pp, 'l'); | |
10865 | if (arg.linear_step > 1) | |
10866 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
10867 | else if (arg.linear_step < 0) | |
10868 | { | |
10869 | pp_character (&pp, 'n'); | |
10870 | pp_unsigned_wide_integer (&pp, (-(unsigned HOST_WIDE_INT) | |
10871 | arg.linear_step)); | |
10872 | } | |
10873 | } | |
10874 | else if (arg.arg_type == SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP) | |
10875 | { | |
10876 | pp_character (&pp, 's'); | |
10877 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
10878 | } | |
10879 | else | |
10880 | pp_character (&pp, 'v'); | |
10881 | if (arg.alignment) | |
10882 | { | |
10883 | pp_character (&pp, 'a'); | |
10884 | pp_decimal_int (&pp, arg.alignment); | |
10885 | } | |
10886 | } | |
10887 | ||
10888 | pp_underscore (&pp); | |
10889 | pp_string (&pp, | |
10890 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (node->decl))); | |
10891 | const char *str = pp_formatted_text (&pp); | |
10892 | ||
10893 | /* If there already is a SIMD clone with the same mangled name, don't | |
10894 | add another one. This can happen e.g. for | |
10895 | #pragma omp declare simd | |
10896 | #pragma omp declare simd simdlen(8) | |
10897 | int foo (int, int); | |
10898 | if the simdlen is assumed to be 8 for the first one, etc. */ | |
10899 | for (struct cgraph_node *clone = node->simd_clones; clone; | |
10900 | clone = clone->simdclone->next_clone) | |
10901 | if (strcmp (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (clone->decl)), | |
10902 | str) == 0) | |
10903 | return NULL_TREE; | |
10904 | ||
10905 | return get_identifier (str); | |
10906 | } | |
10907 | ||
10908 | /* Create a simd clone of OLD_NODE and return it. */ | |
10909 | ||
10910 | static struct cgraph_node * | |
10911 | simd_clone_create (struct cgraph_node *old_node) | |
10912 | { | |
10913 | struct cgraph_node *new_node; | |
10914 | if (old_node->definition) | |
ea0695f9 | 10915 | { |
10916 | if (!cgraph_function_with_gimple_body_p (old_node)) | |
10917 | return NULL; | |
10918 | cgraph_get_body (old_node); | |
10919 | new_node = cgraph_function_versioning (old_node, vNULL, NULL, NULL, | |
10920 | false, NULL, NULL, "simdclone"); | |
10921 | } | |
d09768a4 | 10922 | else |
10923 | { | |
10924 | tree old_decl = old_node->decl; | |
10925 | tree new_decl = copy_node (old_node->decl); | |
10926 | DECL_NAME (new_decl) = clone_function_name (old_decl, "simdclone"); | |
10927 | SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl)); | |
10928 | SET_DECL_RTL (new_decl, NULL); | |
10929 | DECL_STATIC_CONSTRUCTOR (new_decl) = 0; | |
10930 | DECL_STATIC_DESTRUCTOR (new_decl) = 0; | |
10931 | new_node | |
10932 | = cgraph_copy_node_for_versioning (old_node, new_decl, vNULL, NULL); | |
10933 | cgraph_call_function_insertion_hooks (new_node); | |
10934 | } | |
10935 | if (new_node == NULL) | |
10936 | return new_node; | |
10937 | ||
10938 | TREE_PUBLIC (new_node->decl) = TREE_PUBLIC (old_node->decl); | |
10939 | ||
10940 | /* The function cgraph_function_versioning () will force the new | |
10941 | symbol local. Undo this, and inherit external visability from | |
10942 | the old node. */ | |
10943 | new_node->local.local = old_node->local.local; | |
10944 | new_node->externally_visible = old_node->externally_visible; | |
10945 | ||
10946 | return new_node; | |
10947 | } | |
10948 | ||
10949 | /* Adjust the return type of the given function to its appropriate | |
10950 | vector counterpart. Returns a simd array to be used throughout the | |
10951 | function as a return value. */ | |
10952 | ||
10953 | static tree | |
10954 | simd_clone_adjust_return_type (struct cgraph_node *node) | |
10955 | { | |
10956 | tree fndecl = node->decl; | |
10957 | tree orig_rettype = TREE_TYPE (TREE_TYPE (fndecl)); | |
10958 | unsigned int veclen; | |
10959 | tree t; | |
10960 | ||
10961 | /* Adjust the function return type. */ | |
10962 | if (orig_rettype == void_type_node) | |
10963 | return NULL_TREE; | |
10964 | TREE_TYPE (fndecl) = build_distinct_type_copy (TREE_TYPE (fndecl)); | |
10965 | if (INTEGRAL_TYPE_P (TREE_TYPE (TREE_TYPE (fndecl))) | |
10966 | || POINTER_TYPE_P (TREE_TYPE (TREE_TYPE (fndecl)))) | |
10967 | veclen = node->simdclone->vecsize_int; | |
10968 | else | |
10969 | veclen = node->simdclone->vecsize_float; | |
10970 | veclen /= GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (TREE_TYPE (fndecl)))); | |
10971 | if (veclen > node->simdclone->simdlen) | |
10972 | veclen = node->simdclone->simdlen; | |
10973 | if (veclen == node->simdclone->simdlen) | |
10974 | TREE_TYPE (TREE_TYPE (fndecl)) | |
10975 | = build_vector_type (TREE_TYPE (TREE_TYPE (fndecl)), | |
10976 | node->simdclone->simdlen); | |
10977 | else | |
10978 | { | |
10979 | t = build_vector_type (TREE_TYPE (TREE_TYPE (fndecl)), veclen); | |
10980 | t = build_array_type_nelts (t, node->simdclone->simdlen / veclen); | |
10981 | TREE_TYPE (TREE_TYPE (fndecl)) = t; | |
10982 | } | |
10983 | if (!node->definition) | |
10984 | return NULL_TREE; | |
10985 | ||
10986 | t = DECL_RESULT (fndecl); | |
10987 | /* Adjust the DECL_RESULT. */ | |
10988 | gcc_assert (TREE_TYPE (t) != void_type_node); | |
10989 | TREE_TYPE (t) = TREE_TYPE (TREE_TYPE (fndecl)); | |
10990 | relayout_decl (t); | |
10991 | ||
10992 | tree atype = build_array_type_nelts (orig_rettype, | |
10993 | node->simdclone->simdlen); | |
10994 | if (veclen != node->simdclone->simdlen) | |
10995 | return build1 (VIEW_CONVERT_EXPR, atype, t); | |
10996 | ||
10997 | /* Set up a SIMD array to use as the return value. */ | |
10998 | tree retval = create_tmp_var_raw (atype, "retval"); | |
10999 | gimple_add_tmp_var (retval); | |
11000 | return retval; | |
11001 | } | |
11002 | ||
11003 | /* Each vector argument has a corresponding array to be used locally | |
11004 | as part of the eventual loop. Create such temporary array and | |
11005 | return it. | |
11006 | ||
11007 | PREFIX is the prefix to be used for the temporary. | |
11008 | ||
11009 | TYPE is the inner element type. | |
11010 | ||
11011 | SIMDLEN is the number of elements. */ | |
11012 | ||
11013 | static tree | |
11014 | create_tmp_simd_array (const char *prefix, tree type, int simdlen) | |
11015 | { | |
11016 | tree atype = build_array_type_nelts (type, simdlen); | |
11017 | tree avar = create_tmp_var_raw (atype, prefix); | |
11018 | gimple_add_tmp_var (avar); | |
11019 | return avar; | |
11020 | } | |
11021 | ||
11022 | /* Modify the function argument types to their corresponding vector | |
11023 | counterparts if appropriate. Also, create one array for each simd | |
11024 | argument to be used locally when using the function arguments as | |
11025 | part of the loop. | |
11026 | ||
11027 | NODE is the function whose arguments are to be adjusted. | |
11028 | ||
11029 | Returns an adjustment vector that will be filled describing how the | |
11030 | argument types will be adjusted. */ | |
11031 | ||
11032 | static ipa_parm_adjustment_vec | |
11033 | simd_clone_adjust_argument_types (struct cgraph_node *node) | |
11034 | { | |
11035 | vec<tree> args; | |
11036 | ipa_parm_adjustment_vec adjustments; | |
11037 | ||
11038 | if (node->definition) | |
11039 | args = ipa_get_vector_of_formal_parms (node->decl); | |
11040 | else | |
11041 | args = simd_clone_vector_of_formal_parm_types (node->decl); | |
11042 | adjustments.create (args.length ()); | |
11043 | unsigned i, j, veclen; | |
11044 | struct ipa_parm_adjustment adj; | |
11045 | for (i = 0; i < node->simdclone->nargs; ++i) | |
11046 | { | |
11047 | memset (&adj, 0, sizeof (adj)); | |
11048 | tree parm = args[i]; | |
11049 | tree parm_type = node->definition ? TREE_TYPE (parm) : parm; | |
11050 | adj.base_index = i; | |
11051 | adj.base = parm; | |
11052 | ||
11053 | node->simdclone->args[i].orig_arg = node->definition ? parm : NULL_TREE; | |
11054 | node->simdclone->args[i].orig_type = parm_type; | |
11055 | ||
11056 | if (node->simdclone->args[i].arg_type != SIMD_CLONE_ARG_TYPE_VECTOR) | |
11057 | { | |
11058 | /* No adjustment necessary for scalar arguments. */ | |
11059 | adj.op = IPA_PARM_OP_COPY; | |
11060 | } | |
11061 | else | |
11062 | { | |
11063 | if (INTEGRAL_TYPE_P (parm_type) || POINTER_TYPE_P (parm_type)) | |
11064 | veclen = node->simdclone->vecsize_int; | |
11065 | else | |
11066 | veclen = node->simdclone->vecsize_float; | |
11067 | veclen /= GET_MODE_BITSIZE (TYPE_MODE (parm_type)); | |
11068 | if (veclen > node->simdclone->simdlen) | |
11069 | veclen = node->simdclone->simdlen; | |
11070 | adj.arg_prefix = "simd"; | |
11071 | adj.type = build_vector_type (parm_type, veclen); | |
11072 | node->simdclone->args[i].vector_type = adj.type; | |
11073 | for (j = veclen; j < node->simdclone->simdlen; j += veclen) | |
11074 | { | |
11075 | adjustments.safe_push (adj); | |
11076 | if (j == veclen) | |
11077 | { | |
11078 | memset (&adj, 0, sizeof (adj)); | |
11079 | adj.op = IPA_PARM_OP_NEW; | |
11080 | adj.arg_prefix = "simd"; | |
11081 | adj.base_index = i; | |
11082 | adj.type = node->simdclone->args[i].vector_type; | |
11083 | } | |
11084 | } | |
11085 | ||
11086 | if (node->definition) | |
11087 | node->simdclone->args[i].simd_array | |
11088 | = create_tmp_simd_array (IDENTIFIER_POINTER (DECL_NAME (parm)), | |
11089 | parm_type, node->simdclone->simdlen); | |
11090 | } | |
11091 | adjustments.safe_push (adj); | |
11092 | } | |
11093 | ||
11094 | if (node->simdclone->inbranch) | |
11095 | { | |
11096 | tree base_type | |
11097 | = simd_clone_compute_base_data_type (node->simdclone->origin, | |
11098 | node->simdclone); | |
11099 | ||
11100 | memset (&adj, 0, sizeof (adj)); | |
11101 | adj.op = IPA_PARM_OP_NEW; | |
11102 | adj.arg_prefix = "mask"; | |
11103 | ||
11104 | adj.base_index = i; | |
11105 | if (INTEGRAL_TYPE_P (base_type) || POINTER_TYPE_P (base_type)) | |
11106 | veclen = node->simdclone->vecsize_int; | |
11107 | else | |
11108 | veclen = node->simdclone->vecsize_float; | |
11109 | veclen /= GET_MODE_BITSIZE (TYPE_MODE (base_type)); | |
11110 | if (veclen > node->simdclone->simdlen) | |
11111 | veclen = node->simdclone->simdlen; | |
11112 | adj.type = build_vector_type (base_type, veclen); | |
11113 | adjustments.safe_push (adj); | |
11114 | ||
11115 | for (j = veclen; j < node->simdclone->simdlen; j += veclen) | |
11116 | adjustments.safe_push (adj); | |
11117 | ||
11118 | /* We have previously allocated one extra entry for the mask. Use | |
11119 | it and fill it. */ | |
11120 | struct cgraph_simd_clone *sc = node->simdclone; | |
11121 | sc->nargs++; | |
11122 | if (node->definition) | |
11123 | { | |
11124 | sc->args[i].orig_arg | |
11125 | = build_decl (UNKNOWN_LOCATION, PARM_DECL, NULL, base_type); | |
11126 | sc->args[i].simd_array | |
11127 | = create_tmp_simd_array ("mask", base_type, sc->simdlen); | |
11128 | } | |
11129 | sc->args[i].orig_type = base_type; | |
11130 | sc->args[i].arg_type = SIMD_CLONE_ARG_TYPE_MASK; | |
11131 | } | |
11132 | ||
11133 | if (node->definition) | |
11134 | ipa_modify_formal_parameters (node->decl, adjustments); | |
11135 | else | |
11136 | { | |
11137 | tree new_arg_types = NULL_TREE, new_reversed; | |
11138 | bool last_parm_void = false; | |
11139 | if (args.length () > 0 && args.last () == void_type_node) | |
11140 | last_parm_void = true; | |
11141 | ||
11142 | gcc_assert (TYPE_ARG_TYPES (TREE_TYPE (node->decl))); | |
11143 | j = adjustments.length (); | |
11144 | for (i = 0; i < j; i++) | |
11145 | { | |
11146 | struct ipa_parm_adjustment *adj = &adjustments[i]; | |
11147 | tree ptype; | |
11148 | if (adj->op == IPA_PARM_OP_COPY) | |
11149 | ptype = args[adj->base_index]; | |
11150 | else | |
11151 | ptype = adj->type; | |
11152 | new_arg_types = tree_cons (NULL_TREE, ptype, new_arg_types); | |
11153 | } | |
11154 | new_reversed = nreverse (new_arg_types); | |
11155 | if (last_parm_void) | |
11156 | { | |
11157 | if (new_reversed) | |
11158 | TREE_CHAIN (new_arg_types) = void_list_node; | |
11159 | else | |
11160 | new_reversed = void_list_node; | |
11161 | } | |
11162 | ||
11163 | tree new_type = build_distinct_type_copy (TREE_TYPE (node->decl)); | |
11164 | TYPE_ARG_TYPES (new_type) = new_reversed; | |
11165 | TREE_TYPE (node->decl) = new_type; | |
11166 | ||
11167 | adjustments.release (); | |
11168 | } | |
11169 | args.release (); | |
11170 | return adjustments; | |
11171 | } | |
11172 | ||
11173 | /* Initialize and copy the function arguments in NODE to their | |
11174 | corresponding local simd arrays. Returns a fresh gimple_seq with | |
11175 | the instruction sequence generated. */ | |
11176 | ||
11177 | static gimple_seq | |
11178 | simd_clone_init_simd_arrays (struct cgraph_node *node, | |
11179 | ipa_parm_adjustment_vec adjustments) | |
11180 | { | |
11181 | gimple_seq seq = NULL; | |
11182 | unsigned i = 0, j = 0, k; | |
11183 | ||
11184 | for (tree arg = DECL_ARGUMENTS (node->decl); | |
11185 | arg; | |
11186 | arg = DECL_CHAIN (arg), i++, j++) | |
11187 | { | |
11188 | if (adjustments[j].op == IPA_PARM_OP_COPY) | |
11189 | continue; | |
11190 | ||
11191 | node->simdclone->args[i].vector_arg = arg; | |
11192 | ||
11193 | tree array = node->simdclone->args[i].simd_array; | |
11194 | if (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg)) == node->simdclone->simdlen) | |
11195 | { | |
11196 | tree ptype = build_pointer_type (TREE_TYPE (TREE_TYPE (array))); | |
11197 | tree ptr = build_fold_addr_expr (array); | |
11198 | tree t = build2 (MEM_REF, TREE_TYPE (arg), ptr, | |
11199 | build_int_cst (ptype, 0)); | |
11200 | t = build2 (MODIFY_EXPR, TREE_TYPE (t), t, arg); | |
11201 | gimplify_and_add (t, &seq); | |
11202 | } | |
11203 | else | |
11204 | { | |
11205 | unsigned int simdlen = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg)); | |
11206 | tree ptype = build_pointer_type (TREE_TYPE (TREE_TYPE (array))); | |
11207 | for (k = 0; k < node->simdclone->simdlen; k += simdlen) | |
11208 | { | |
11209 | tree ptr = build_fold_addr_expr (array); | |
11210 | int elemsize; | |
11211 | if (k) | |
11212 | { | |
11213 | arg = DECL_CHAIN (arg); | |
11214 | j++; | |
11215 | } | |
11216 | elemsize | |
11217 | = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (TREE_TYPE (arg)))); | |
11218 | tree t = build2 (MEM_REF, TREE_TYPE (arg), ptr, | |
11219 | build_int_cst (ptype, k * elemsize)); | |
11220 | t = build2 (MODIFY_EXPR, TREE_TYPE (t), t, arg); | |
11221 | gimplify_and_add (t, &seq); | |
11222 | } | |
11223 | } | |
11224 | } | |
11225 | return seq; | |
11226 | } | |
11227 | ||
11228 | /* Callback info for ipa_simd_modify_stmt_ops below. */ | |
11229 | ||
11230 | struct modify_stmt_info { | |
11231 | ipa_parm_adjustment_vec adjustments; | |
11232 | gimple stmt; | |
11233 | /* True if the parent statement was modified by | |
11234 | ipa_simd_modify_stmt_ops. */ | |
11235 | bool modified; | |
11236 | }; | |
11237 | ||
11238 | /* Callback for walk_gimple_op. | |
11239 | ||
11240 | Adjust operands from a given statement as specified in the | |
11241 | adjustments vector in the callback data. */ | |
11242 | ||
11243 | static tree | |
11244 | ipa_simd_modify_stmt_ops (tree *tp, int *walk_subtrees, void *data) | |
11245 | { | |
11246 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; | |
11247 | if (!SSA_VAR_P (*tp)) | |
11248 | { | |
11249 | /* Make sure we treat subtrees as a RHS. This makes sure that | |
11250 | when examining the `*foo' in *foo=x, the `foo' get treated as | |
11251 | a use properly. */ | |
11252 | wi->is_lhs = false; | |
11253 | wi->val_only = true; | |
11254 | if (TYPE_P (*tp)) | |
11255 | *walk_subtrees = 0; | |
11256 | return NULL_TREE; | |
11257 | } | |
11258 | struct modify_stmt_info *info = (struct modify_stmt_info *) wi->info; | |
11259 | struct ipa_parm_adjustment *cand | |
11260 | = ipa_get_adjustment_candidate (&tp, NULL, info->adjustments, true); | |
11261 | if (!cand) | |
11262 | return NULL_TREE; | |
11263 | ||
11264 | tree t = *tp; | |
11265 | tree repl = make_ssa_name (TREE_TYPE (t), NULL); | |
11266 | ||
11267 | gimple stmt; | |
11268 | gimple_stmt_iterator gsi = gsi_for_stmt (info->stmt); | |
11269 | if (wi->is_lhs) | |
11270 | { | |
11271 | stmt = gimple_build_assign (unshare_expr (cand->new_decl), repl); | |
11272 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
11273 | SSA_NAME_DEF_STMT (repl) = info->stmt; | |
11274 | } | |
11275 | else | |
11276 | { | |
11277 | /* You'd think we could skip the extra SSA variable when | |
11278 | wi->val_only=true, but we may have `*var' which will get | |
11279 | replaced into `*var_array[iter]' and will likely be something | |
11280 | not gimple. */ | |
11281 | stmt = gimple_build_assign (repl, unshare_expr (cand->new_decl)); | |
11282 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
11283 | } | |
11284 | ||
11285 | if (!useless_type_conversion_p (TREE_TYPE (*tp), TREE_TYPE (repl))) | |
11286 | { | |
11287 | tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*tp), repl); | |
11288 | *tp = vce; | |
11289 | } | |
11290 | else | |
11291 | *tp = repl; | |
11292 | ||
11293 | info->modified = true; | |
11294 | wi->is_lhs = false; | |
11295 | wi->val_only = true; | |
11296 | return NULL_TREE; | |
11297 | } | |
11298 | ||
11299 | /* Traverse the function body and perform all modifications as | |
11300 | described in ADJUSTMENTS. At function return, ADJUSTMENTS will be | |
11301 | modified such that the replacement/reduction value will now be an | |
11302 | offset into the corresponding simd_array. | |
11303 | ||
11304 | This function will replace all function argument uses with their | |
11305 | corresponding simd array elements, and ajust the return values | |
11306 | accordingly. */ | |
11307 | ||
11308 | static void | |
11309 | ipa_simd_modify_function_body (struct cgraph_node *node, | |
11310 | ipa_parm_adjustment_vec adjustments, | |
11311 | tree retval_array, tree iter) | |
11312 | { | |
11313 | basic_block bb; | |
11314 | unsigned int i, j; | |
11315 | ||
11316 | /* Re-use the adjustments array, but this time use it to replace | |
11317 | every function argument use to an offset into the corresponding | |
11318 | simd_array. */ | |
11319 | for (i = 0, j = 0; i < node->simdclone->nargs; ++i, ++j) | |
11320 | { | |
11321 | if (!node->simdclone->args[i].vector_arg) | |
11322 | continue; | |
11323 | ||
11324 | tree basetype = TREE_TYPE (node->simdclone->args[i].orig_arg); | |
11325 | tree vectype = TREE_TYPE (node->simdclone->args[i].vector_arg); | |
11326 | adjustments[j].new_decl | |
11327 | = build4 (ARRAY_REF, | |
11328 | basetype, | |
11329 | node->simdclone->args[i].simd_array, | |
11330 | iter, | |
11331 | NULL_TREE, NULL_TREE); | |
11332 | if (adjustments[j].op == IPA_PARM_OP_NONE | |
11333 | && TYPE_VECTOR_SUBPARTS (vectype) < node->simdclone->simdlen) | |
11334 | j += node->simdclone->simdlen / TYPE_VECTOR_SUBPARTS (vectype) - 1; | |
11335 | } | |
11336 | ||
11337 | struct modify_stmt_info info; | |
11338 | info.adjustments = adjustments; | |
11339 | ||
11340 | FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl)) | |
11341 | { | |
11342 | gimple_stmt_iterator gsi; | |
11343 | ||
11344 | gsi = gsi_start_bb (bb); | |
11345 | while (!gsi_end_p (gsi)) | |
11346 | { | |
11347 | gimple stmt = gsi_stmt (gsi); | |
11348 | info.stmt = stmt; | |
11349 | struct walk_stmt_info wi; | |
11350 | ||
11351 | memset (&wi, 0, sizeof (wi)); | |
11352 | info.modified = false; | |
11353 | wi.info = &info; | |
11354 | walk_gimple_op (stmt, ipa_simd_modify_stmt_ops, &wi); | |
11355 | ||
11356 | if (gimple_code (stmt) == GIMPLE_RETURN) | |
11357 | { | |
11358 | tree retval = gimple_return_retval (stmt); | |
11359 | if (!retval) | |
11360 | { | |
11361 | gsi_remove (&gsi, true); | |
11362 | continue; | |
11363 | } | |
11364 | ||
11365 | /* Replace `return foo' with `retval_array[iter] = foo'. */ | |
11366 | tree ref = build4 (ARRAY_REF, TREE_TYPE (retval), | |
11367 | retval_array, iter, NULL, NULL); | |
11368 | stmt = gimple_build_assign (ref, retval); | |
11369 | gsi_replace (&gsi, stmt, true); | |
11370 | info.modified = true; | |
11371 | } | |
11372 | ||
11373 | if (info.modified) | |
11374 | { | |
11375 | update_stmt (stmt); | |
11376 | if (maybe_clean_eh_stmt (stmt)) | |
11377 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
11378 | } | |
11379 | gsi_next (&gsi); | |
11380 | } | |
11381 | } | |
11382 | } | |
11383 | ||
11384 | /* Adjust the argument types in NODE to their appropriate vector | |
11385 | counterparts. */ | |
11386 | ||
11387 | static void | |
11388 | simd_clone_adjust (struct cgraph_node *node) | |
11389 | { | |
11390 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); | |
11391 | ||
11392 | targetm.simd_clone.adjust (node); | |
11393 | ||
11394 | tree retval = simd_clone_adjust_return_type (node); | |
11395 | ipa_parm_adjustment_vec adjustments | |
11396 | = simd_clone_adjust_argument_types (node); | |
11397 | ||
11398 | push_gimplify_context (); | |
11399 | ||
11400 | gimple_seq seq = simd_clone_init_simd_arrays (node, adjustments); | |
11401 | ||
11402 | /* Adjust all uses of vector arguments accordingly. Adjust all | |
11403 | return values accordingly. */ | |
11404 | tree iter = create_tmp_var (unsigned_type_node, "iter"); | |
11405 | tree iter1 = make_ssa_name (iter, NULL); | |
11406 | tree iter2 = make_ssa_name (iter, NULL); | |
11407 | ipa_simd_modify_function_body (node, adjustments, retval, iter1); | |
11408 | ||
11409 | /* Initialize the iteration variable. */ | |
11410 | basic_block entry_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
11411 | basic_block body_bb = split_block_after_labels (entry_bb)->dest; | |
11412 | gimple_stmt_iterator gsi = gsi_after_labels (entry_bb); | |
11413 | /* Insert the SIMD array and iv initialization at function | |
11414 | entry. */ | |
11415 | gsi_insert_seq_before (&gsi, seq, GSI_NEW_STMT); | |
11416 | ||
11417 | pop_gimplify_context (NULL); | |
11418 | ||
11419 | /* Create a new BB right before the original exit BB, to hold the | |
11420 | iteration increment and the condition/branch. */ | |
11421 | basic_block orig_exit = EDGE_PRED (EXIT_BLOCK_PTR_FOR_FN (cfun), 0)->src; | |
11422 | basic_block incr_bb = create_empty_bb (orig_exit); | |
11423 | /* The succ of orig_exit was EXIT_BLOCK_PTR_FOR_FN (cfun), with an empty | |
11424 | flag. Set it now to be a FALLTHRU_EDGE. */ | |
11425 | gcc_assert (EDGE_COUNT (orig_exit->succs) == 1); | |
11426 | EDGE_SUCC (orig_exit, 0)->flags |= EDGE_FALLTHRU; | |
11427 | for (unsigned i = 0; | |
11428 | i < EDGE_COUNT (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds); ++i) | |
11429 | { | |
11430 | edge e = EDGE_PRED (EXIT_BLOCK_PTR_FOR_FN (cfun), i); | |
11431 | redirect_edge_succ (e, incr_bb); | |
11432 | } | |
11433 | edge e = make_edge (incr_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0); | |
11434 | e->probability = REG_BR_PROB_BASE; | |
11435 | gsi = gsi_last_bb (incr_bb); | |
11436 | gimple g = gimple_build_assign_with_ops (PLUS_EXPR, iter2, iter1, | |
11437 | build_int_cst (unsigned_type_node, | |
11438 | 1)); | |
11439 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
11440 | ||
11441 | /* Mostly annotate the loop for the vectorizer (the rest is done below). */ | |
11442 | struct loop *loop = alloc_loop (); | |
11443 | cfun->has_force_vect_loops = true; | |
11444 | loop->safelen = node->simdclone->simdlen; | |
11445 | loop->force_vect = true; | |
11446 | loop->header = body_bb; | |
11447 | add_bb_to_loop (incr_bb, loop); | |
11448 | ||
11449 | /* Branch around the body if the mask applies. */ | |
11450 | if (node->simdclone->inbranch) | |
11451 | { | |
11452 | gimple_stmt_iterator gsi = gsi_last_bb (loop->header); | |
11453 | tree mask_array | |
11454 | = node->simdclone->args[node->simdclone->nargs - 1].simd_array; | |
11455 | tree mask = make_ssa_name (TREE_TYPE (TREE_TYPE (mask_array)), NULL); | |
11456 | tree aref = build4 (ARRAY_REF, | |
11457 | TREE_TYPE (TREE_TYPE (mask_array)), | |
11458 | mask_array, iter1, | |
11459 | NULL, NULL); | |
11460 | g = gimple_build_assign (mask, aref); | |
11461 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
11462 | int bitsize = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (aref))); | |
11463 | if (!INTEGRAL_TYPE_P (TREE_TYPE (aref))) | |
11464 | { | |
11465 | aref = build1 (VIEW_CONVERT_EXPR, | |
11466 | build_nonstandard_integer_type (bitsize, 0), mask); | |
11467 | mask = make_ssa_name (TREE_TYPE (aref), NULL); | |
11468 | g = gimple_build_assign (mask, aref); | |
11469 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
11470 | } | |
11471 | ||
11472 | g = gimple_build_cond (EQ_EXPR, mask, build_zero_cst (TREE_TYPE (mask)), | |
11473 | NULL, NULL); | |
11474 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
11475 | make_edge (loop->header, incr_bb, EDGE_TRUE_VALUE); | |
11476 | FALLTHRU_EDGE (loop->header)->flags = EDGE_FALSE_VALUE; | |
11477 | } | |
11478 | ||
11479 | /* Generate the condition. */ | |
11480 | g = gimple_build_cond (LT_EXPR, | |
11481 | iter2, | |
11482 | build_int_cst (unsigned_type_node, | |
11483 | node->simdclone->simdlen), | |
11484 | NULL, NULL); | |
11485 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
11486 | e = split_block (incr_bb, gsi_stmt (gsi)); | |
11487 | basic_block latch_bb = e->dest; | |
11488 | basic_block new_exit_bb = e->dest; | |
11489 | new_exit_bb = split_block (latch_bb, NULL)->dest; | |
11490 | loop->latch = latch_bb; | |
11491 | ||
11492 | redirect_edge_succ (FALLTHRU_EDGE (latch_bb), body_bb); | |
11493 | ||
11494 | make_edge (incr_bb, new_exit_bb, EDGE_FALSE_VALUE); | |
11495 | /* The successor of incr_bb is already pointing to latch_bb; just | |
11496 | change the flags. | |
11497 | make_edge (incr_bb, latch_bb, EDGE_TRUE_VALUE); */ | |
11498 | FALLTHRU_EDGE (incr_bb)->flags = EDGE_TRUE_VALUE; | |
11499 | ||
11500 | gimple phi = create_phi_node (iter1, body_bb); | |
11501 | edge preheader_edge = find_edge (entry_bb, body_bb); | |
11502 | edge latch_edge = single_succ_edge (latch_bb); | |
11503 | add_phi_arg (phi, build_zero_cst (unsigned_type_node), preheader_edge, | |
11504 | UNKNOWN_LOCATION); | |
11505 | add_phi_arg (phi, iter2, latch_edge, UNKNOWN_LOCATION); | |
11506 | ||
11507 | /* Generate the new return. */ | |
11508 | gsi = gsi_last_bb (new_exit_bb); | |
11509 | if (retval | |
11510 | && TREE_CODE (retval) == VIEW_CONVERT_EXPR | |
11511 | && TREE_CODE (TREE_OPERAND (retval, 0)) == RESULT_DECL) | |
11512 | retval = TREE_OPERAND (retval, 0); | |
11513 | else if (retval) | |
11514 | { | |
11515 | retval = build1 (VIEW_CONVERT_EXPR, | |
11516 | TREE_TYPE (TREE_TYPE (node->decl)), | |
11517 | retval); | |
11518 | retval = force_gimple_operand_gsi (&gsi, retval, true, NULL, | |
11519 | false, GSI_CONTINUE_LINKING); | |
11520 | } | |
11521 | g = gimple_build_return (retval); | |
11522 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
11523 | ||
11524 | /* Handle aligned clauses by replacing default defs of the aligned | |
11525 | uniform args with __builtin_assume_aligned (arg_N(D), alignment) | |
11526 | lhs. Handle linear by adding PHIs. */ | |
11527 | for (unsigned i = 0; i < node->simdclone->nargs; i++) | |
11528 | if (node->simdclone->args[i].alignment | |
11529 | && node->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_UNIFORM | |
11530 | && (node->simdclone->args[i].alignment | |
11531 | & (node->simdclone->args[i].alignment - 1)) == 0 | |
11532 | && TREE_CODE (TREE_TYPE (node->simdclone->args[i].orig_arg)) | |
11533 | == POINTER_TYPE) | |
11534 | { | |
11535 | unsigned int alignment = node->simdclone->args[i].alignment; | |
11536 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
11537 | tree def = ssa_default_def (cfun, orig_arg); | |
11538 | if (!has_zero_uses (def)) | |
11539 | { | |
11540 | tree fn = builtin_decl_explicit (BUILT_IN_ASSUME_ALIGNED); | |
11541 | gimple_seq seq = NULL; | |
11542 | bool need_cvt = false; | |
11543 | gimple call | |
11544 | = gimple_build_call (fn, 2, def, size_int (alignment)); | |
11545 | g = call; | |
11546 | if (!useless_type_conversion_p (TREE_TYPE (orig_arg), | |
11547 | ptr_type_node)) | |
11548 | need_cvt = true; | |
11549 | tree t = make_ssa_name (need_cvt ? ptr_type_node : orig_arg, NULL); | |
11550 | gimple_call_set_lhs (g, t); | |
11551 | gimple_seq_add_stmt_without_update (&seq, g); | |
11552 | if (need_cvt) | |
11553 | { | |
11554 | t = make_ssa_name (orig_arg, NULL); | |
11555 | g = gimple_build_assign_with_ops (NOP_EXPR, t, | |
11556 | gimple_call_lhs (g), | |
11557 | NULL_TREE); | |
11558 | gimple_seq_add_stmt_without_update (&seq, g); | |
11559 | } | |
11560 | gsi_insert_seq_on_edge_immediate | |
11561 | (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)), seq); | |
11562 | ||
11563 | entry_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
11564 | int freq = compute_call_stmt_bb_frequency (current_function_decl, | |
11565 | entry_bb); | |
11566 | cgraph_create_edge (node, cgraph_get_create_node (fn), | |
11567 | call, entry_bb->count, freq); | |
11568 | ||
11569 | imm_use_iterator iter; | |
11570 | use_operand_p use_p; | |
11571 | gimple use_stmt; | |
11572 | tree repl = gimple_get_lhs (g); | |
11573 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, def) | |
11574 | if (is_gimple_debug (use_stmt) || use_stmt == call) | |
11575 | continue; | |
11576 | else | |
11577 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
11578 | SET_USE (use_p, repl); | |
11579 | } | |
11580 | } | |
11581 | else if (node->simdclone->args[i].arg_type | |
11582 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
11583 | { | |
11584 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
11585 | tree def = ssa_default_def (cfun, orig_arg); | |
11586 | gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
11587 | || POINTER_TYPE_P (TREE_TYPE (orig_arg))); | |
11588 | if (!has_zero_uses (def)) | |
11589 | { | |
11590 | iter1 = make_ssa_name (orig_arg, NULL); | |
11591 | iter2 = make_ssa_name (orig_arg, NULL); | |
11592 | phi = create_phi_node (iter1, body_bb); | |
11593 | add_phi_arg (phi, def, preheader_edge, UNKNOWN_LOCATION); | |
11594 | add_phi_arg (phi, iter2, latch_edge, UNKNOWN_LOCATION); | |
11595 | enum tree_code code = INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
11596 | ? PLUS_EXPR : POINTER_PLUS_EXPR; | |
11597 | tree addtype = INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
11598 | ? TREE_TYPE (orig_arg) : sizetype; | |
11599 | tree addcst | |
11600 | = build_int_cst (addtype, node->simdclone->args[i].linear_step); | |
11601 | g = gimple_build_assign_with_ops (code, iter2, iter1, addcst); | |
11602 | gsi = gsi_last_bb (incr_bb); | |
11603 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
11604 | ||
11605 | imm_use_iterator iter; | |
11606 | use_operand_p use_p; | |
11607 | gimple use_stmt; | |
11608 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, def) | |
11609 | if (use_stmt == phi) | |
11610 | continue; | |
11611 | else | |
11612 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
11613 | SET_USE (use_p, iter1); | |
11614 | } | |
11615 | } | |
11616 | ||
11617 | calculate_dominance_info (CDI_DOMINATORS); | |
11618 | add_loop (loop, loop->header->loop_father); | |
11619 | update_ssa (TODO_update_ssa); | |
11620 | ||
11621 | pop_cfun (); | |
11622 | } | |
11623 | ||
11624 | /* If the function in NODE is tagged as an elemental SIMD function, | |
11625 | create the appropriate SIMD clones. */ | |
11626 | ||
11627 | static void | |
11628 | expand_simd_clones (struct cgraph_node *node) | |
11629 | { | |
d09768a4 | 11630 | tree attr = lookup_attribute ("omp declare simd", |
11631 | DECL_ATTRIBUTES (node->decl)); | |
ea0695f9 | 11632 | if (attr == NULL_TREE |
11633 | || node->global.inlined_to | |
11634 | || lookup_attribute ("noclone", DECL_ATTRIBUTES (node->decl))) | |
d09768a4 | 11635 | return; |
ea0695f9 | 11636 | |
d09768a4 | 11637 | /* Ignore |
11638 | #pragma omp declare simd | |
11639 | extern int foo (); | |
11640 | in C, there we don't know the argument types at all. */ | |
11641 | if (!node->definition | |
11642 | && TYPE_ARG_TYPES (TREE_TYPE (node->decl)) == NULL_TREE) | |
11643 | return; | |
11644 | ||
11645 | do | |
11646 | { | |
11647 | /* Start with parsing the "omp declare simd" attribute(s). */ | |
11648 | bool inbranch_clause_specified; | |
11649 | struct cgraph_simd_clone *clone_info | |
11650 | = simd_clone_clauses_extract (node, TREE_VALUE (attr), | |
11651 | &inbranch_clause_specified); | |
11652 | if (clone_info == NULL) | |
11653 | continue; | |
11654 | ||
11655 | int orig_simdlen = clone_info->simdlen; | |
11656 | tree base_type = simd_clone_compute_base_data_type (node, clone_info); | |
11657 | /* The target can return 0 (no simd clones should be created), | |
11658 | 1 (just one ISA of simd clones should be created) or higher | |
11659 | count of ISA variants. In that case, clone_info is initialized | |
11660 | for the first ISA variant. */ | |
11661 | int count | |
11662 | = targetm.simd_clone.compute_vecsize_and_simdlen (node, clone_info, | |
11663 | base_type, 0); | |
11664 | if (count == 0) | |
11665 | continue; | |
11666 | ||
11667 | /* Loop over all COUNT ISA variants, and if !INBRANCH_CLAUSE_SPECIFIED, | |
11668 | also create one inbranch and one !inbranch clone of it. */ | |
11669 | for (int i = 0; i < count * 2; i++) | |
11670 | { | |
11671 | struct cgraph_simd_clone *clone = clone_info; | |
11672 | if (inbranch_clause_specified && (i & 1) != 0) | |
11673 | continue; | |
11674 | ||
11675 | if (i != 0) | |
11676 | { | |
11677 | clone = simd_clone_struct_alloc (clone_info->nargs | |
11678 | - clone_info->inbranch | |
11679 | + ((i & 1) != 0)); | |
11680 | simd_clone_struct_copy (clone, clone_info); | |
11681 | /* Undo changes targetm.simd_clone.compute_vecsize_and_simdlen | |
11682 | and simd_clone_adjust_argument_types did to the first | |
11683 | clone's info. */ | |
11684 | clone->nargs -= clone_info->inbranch; | |
11685 | clone->simdlen = orig_simdlen; | |
11686 | /* And call the target hook again to get the right ISA. */ | |
11687 | targetm.simd_clone.compute_vecsize_and_simdlen (node, clone, | |
11688 | base_type, | |
11689 | i / 2); | |
11690 | if ((i & 1) != 0) | |
11691 | clone->inbranch = 1; | |
11692 | } | |
11693 | ||
11694 | /* simd_clone_mangle might fail if such a clone has been created | |
11695 | already. */ | |
11696 | tree id = simd_clone_mangle (node, clone); | |
11697 | if (id == NULL_TREE) | |
11698 | continue; | |
11699 | ||
11700 | /* Only when we are sure we want to create the clone actually | |
11701 | clone the function (or definitions) or create another | |
11702 | extern FUNCTION_DECL (for prototypes without definitions). */ | |
11703 | struct cgraph_node *n = simd_clone_create (node); | |
11704 | if (n == NULL) | |
11705 | continue; | |
11706 | ||
11707 | n->simdclone = clone; | |
11708 | clone->origin = node; | |
11709 | clone->next_clone = NULL; | |
11710 | if (node->simd_clones == NULL) | |
11711 | { | |
11712 | clone->prev_clone = n; | |
11713 | node->simd_clones = n; | |
11714 | } | |
11715 | else | |
11716 | { | |
11717 | clone->prev_clone = node->simd_clones->simdclone->prev_clone; | |
11718 | clone->prev_clone->simdclone->next_clone = n; | |
11719 | node->simd_clones->simdclone->prev_clone = n; | |
11720 | } | |
11721 | change_decl_assembler_name (n->decl, id); | |
11722 | /* And finally adjust the return type, parameters and for | |
11723 | definitions also function body. */ | |
11724 | if (node->definition) | |
11725 | simd_clone_adjust (n); | |
11726 | else | |
11727 | { | |
11728 | simd_clone_adjust_return_type (n); | |
11729 | simd_clone_adjust_argument_types (n); | |
11730 | } | |
11731 | } | |
11732 | } | |
11733 | while ((attr = lookup_attribute ("omp declare simd", TREE_CHAIN (attr)))); | |
11734 | } | |
11735 | ||
11736 | /* Entry point for IPA simd clone creation pass. */ | |
11737 | ||
11738 | static unsigned int | |
11739 | ipa_omp_simd_clone (void) | |
11740 | { | |
11741 | struct cgraph_node *node; | |
11742 | FOR_EACH_FUNCTION (node) | |
11743 | expand_simd_clones (node); | |
11744 | return 0; | |
11745 | } | |
11746 | ||
11747 | namespace { | |
11748 | ||
11749 | const pass_data pass_data_omp_simd_clone = | |
11750 | { | |
11751 | SIMPLE_IPA_PASS, /* type */ | |
11752 | "simdclone", /* name */ | |
11753 | OPTGROUP_NONE, /* optinfo_flags */ | |
11754 | true, /* has_gate */ | |
11755 | true, /* has_execute */ | |
11756 | TV_NONE, /* tv_id */ | |
11757 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
11758 | 0, /* properties_provided */ | |
11759 | 0, /* properties_destroyed */ | |
11760 | 0, /* todo_flags_start */ | |
11761 | 0, /* todo_flags_finish */ | |
11762 | }; | |
11763 | ||
11764 | class pass_omp_simd_clone : public simple_ipa_opt_pass | |
11765 | { | |
11766 | public: | |
11767 | pass_omp_simd_clone(gcc::context *ctxt) | |
11768 | : simple_ipa_opt_pass(pass_data_omp_simd_clone, ctxt) | |
11769 | {} | |
11770 | ||
11771 | /* opt_pass methods: */ | |
ea0695f9 | 11772 | bool gate () { return ((flag_openmp || flag_openmp_simd |
11773 | || flag_enable_cilkplus || (in_lto_p && !flag_wpa)) | |
11774 | && (targetm.simd_clone.compute_vecsize_and_simdlen | |
11775 | != NULL)); } | |
d09768a4 | 11776 | unsigned int execute () { return ipa_omp_simd_clone (); } |
11777 | }; | |
11778 | ||
11779 | } // anon namespace | |
11780 | ||
11781 | simple_ipa_opt_pass * | |
11782 | make_pass_omp_simd_clone (gcc::context *ctxt) | |
11783 | { | |
11784 | return new pass_omp_simd_clone (ctxt); | |
11785 | } | |
cbe8bda8 | 11786 | |
1e8e9920 | 11787 | #include "gt-omp-low.h" |