]>
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 | ||
3aea1f79 | 6 | Copyright (C) 2005-2014 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" |
94ea8568 | 32 | #include "predict.h" |
33 | #include "vec.h" | |
34 | #include "hashtab.h" | |
35 | #include "hash-set.h" | |
36 | #include "machmode.h" | |
37 | #include "hard-reg-set.h" | |
38 | #include "input.h" | |
39 | #include "function.h" | |
40 | #include "dominance.h" | |
41 | #include "cfg.h" | |
42 | #include "cfganal.h" | |
bc61cadb | 43 | #include "basic-block.h" |
44 | #include "tree-ssa-alias.h" | |
45 | #include "internal-fn.h" | |
46 | #include "gimple-fold.h" | |
47 | #include "gimple-expr.h" | |
48 | #include "is-a.h" | |
75a70cf9 | 49 | #include "gimple.h" |
a8783bee | 50 | #include "gimplify.h" |
dcf1a1ec | 51 | #include "gimple-iterator.h" |
e795d6e1 | 52 | #include "gimplify-me.h" |
dcf1a1ec | 53 | #include "gimple-walk.h" |
75a70cf9 | 54 | #include "tree-iterator.h" |
1e8e9920 | 55 | #include "tree-inline.h" |
56 | #include "langhooks.h" | |
852f689e | 57 | #include "diagnostic-core.h" |
073c1fd5 | 58 | #include "gimple-ssa.h" |
1140c305 | 59 | #include "hash-map.h" |
60 | #include "plugin-api.h" | |
61 | #include "ipa-ref.h" | |
073c1fd5 | 62 | #include "cgraph.h" |
63 | #include "tree-cfg.h" | |
64 | #include "tree-phinodes.h" | |
65 | #include "ssa-iterators.h" | |
66 | #include "tree-ssanames.h" | |
67 | #include "tree-into-ssa.h" | |
9ed99284 | 68 | #include "expr.h" |
073c1fd5 | 69 | #include "tree-dfa.h" |
69ee5dbb | 70 | #include "tree-ssa.h" |
1e8e9920 | 71 | #include "flags.h" |
1e8e9920 | 72 | #include "expr.h" |
1e8e9920 | 73 | #include "tree-pass.h" |
1e8e9920 | 74 | #include "except.h" |
e3022db7 | 75 | #include "splay-tree.h" |
34517c64 | 76 | #include "insn-codes.h" |
cb7f680b | 77 | #include "optabs.h" |
78 | #include "cfgloop.h" | |
3d483a94 | 79 | #include "target.h" |
dccabdd1 | 80 | #include "common/common-target.h" |
7740abd8 | 81 | #include "omp-low.h" |
424a4a92 | 82 | #include "gimple-low.h" |
83 | #include "tree-cfgcleanup.h" | |
d09768a4 | 84 | #include "pretty-print.h" |
1140c305 | 85 | #include "alloc-pool.h" |
d09768a4 | 86 | #include "ipa-prop.h" |
e797f49f | 87 | #include "tree-nested.h" |
d09768a4 | 88 | #include "tree-eh.h" |
40750995 | 89 | #include "cilk.h" |
b0c5e347 | 90 | #include "context.h" |
dccabdd1 | 91 | #include "lto-section-names.h" |
1e8e9920 | 92 | |
75a70cf9 | 93 | |
48e1416a | 94 | /* Lowering of OpenMP parallel and workshare constructs proceeds in two |
1e8e9920 | 95 | phases. The first phase scans the function looking for OMP statements |
96 | and then for variables that must be replaced to satisfy data sharing | |
97 | clauses. The second phase expands code for the constructs, as well as | |
334ec2d8 | 98 | re-gimplifying things when variables have been replaced with complex |
1e8e9920 | 99 | expressions. |
100 | ||
d134bccc | 101 | Final code generation is done by pass_expand_omp. The flowgraph is |
102 | scanned for parallel regions which are then moved to a new | |
103 | function, to be invoked by the thread library. */ | |
1e8e9920 | 104 | |
7740abd8 | 105 | /* Parallel region information. Every parallel and workshare |
106 | directive is enclosed between two markers, the OMP_* directive | |
107 | and a corresponding OMP_RETURN statement. */ | |
108 | ||
109 | struct omp_region | |
110 | { | |
111 | /* The enclosing region. */ | |
112 | struct omp_region *outer; | |
113 | ||
114 | /* First child region. */ | |
115 | struct omp_region *inner; | |
116 | ||
117 | /* Next peer region. */ | |
118 | struct omp_region *next; | |
119 | ||
120 | /* Block containing the omp directive as its last stmt. */ | |
121 | basic_block entry; | |
122 | ||
123 | /* Block containing the OMP_RETURN as its last stmt. */ | |
124 | basic_block exit; | |
125 | ||
126 | /* Block containing the OMP_CONTINUE as its last stmt. */ | |
127 | basic_block cont; | |
128 | ||
129 | /* If this is a combined parallel+workshare region, this is a list | |
130 | of additional arguments needed by the combined parallel+workshare | |
131 | library call. */ | |
132 | vec<tree, va_gc> *ws_args; | |
133 | ||
134 | /* The code for the omp directive of this region. */ | |
135 | enum gimple_code type; | |
136 | ||
137 | /* Schedule kind, only used for OMP_FOR type regions. */ | |
138 | enum omp_clause_schedule_kind sched_kind; | |
139 | ||
140 | /* True if this is a combined parallel+workshare region. */ | |
141 | bool is_combined_parallel; | |
142 | }; | |
143 | ||
1e8e9920 | 144 | /* Context structure. Used to store information about each parallel |
145 | directive in the code. */ | |
146 | ||
147 | typedef struct omp_context | |
148 | { | |
149 | /* This field must be at the beginning, as we do "inheritance": Some | |
150 | callback functions for tree-inline.c (e.g., omp_copy_decl) | |
151 | receive a copy_body_data pointer that is up-casted to an | |
152 | omp_context pointer. */ | |
153 | copy_body_data cb; | |
154 | ||
155 | /* The tree of contexts corresponding to the encountered constructs. */ | |
156 | struct omp_context *outer; | |
75a70cf9 | 157 | gimple stmt; |
1e8e9920 | 158 | |
48e1416a | 159 | /* Map variables to fields in a structure that allows communication |
1e8e9920 | 160 | between sending and receiving threads. */ |
161 | splay_tree field_map; | |
162 | tree record_type; | |
163 | tree sender_decl; | |
164 | tree receiver_decl; | |
165 | ||
fd6481cf | 166 | /* These are used just by task contexts, if task firstprivate fn is |
167 | needed. srecord_type is used to communicate from the thread | |
168 | that encountered the task construct to task firstprivate fn, | |
169 | record_type is allocated by GOMP_task, initialized by task firstprivate | |
170 | fn and passed to the task body fn. */ | |
171 | splay_tree sfield_map; | |
172 | tree srecord_type; | |
173 | ||
1e8e9920 | 174 | /* A chain of variables to add to the top-level block surrounding the |
175 | construct. In the case of a parallel, this is in the child function. */ | |
176 | tree block_vars; | |
177 | ||
bc7bff74 | 178 | /* Label to which GOMP_cancel{,llation_point} and explicit and implicit |
179 | barriers should jump to during omplower pass. */ | |
180 | tree cancel_label; | |
181 | ||
1e8e9920 | 182 | /* What to do with variables with implicitly determined sharing |
183 | attributes. */ | |
184 | enum omp_clause_default_kind default_kind; | |
185 | ||
186 | /* Nesting depth of this context. Used to beautify error messages re | |
187 | invalid gotos. The outermost ctx is depth 1, with depth 0 being | |
188 | reserved for the main body of the function. */ | |
189 | int depth; | |
190 | ||
1e8e9920 | 191 | /* True if this parallel directive is nested within another. */ |
192 | bool is_nested; | |
bc7bff74 | 193 | |
194 | /* True if this construct can be cancelled. */ | |
195 | bool cancellable; | |
1e8e9920 | 196 | } omp_context; |
197 | ||
198 | ||
fd6481cf | 199 | struct omp_for_data_loop |
200 | { | |
201 | tree v, n1, n2, step; | |
202 | enum tree_code cond_code; | |
203 | }; | |
204 | ||
773c5ba7 | 205 | /* A structure describing the main elements of a parallel loop. */ |
1e8e9920 | 206 | |
773c5ba7 | 207 | struct omp_for_data |
1e8e9920 | 208 | { |
fd6481cf | 209 | struct omp_for_data_loop loop; |
75a70cf9 | 210 | tree chunk_size; |
1a91d914 | 211 | gomp_for *for_stmt; |
fd6481cf | 212 | tree pre, iter_type; |
213 | int collapse; | |
1e8e9920 | 214 | bool have_nowait, have_ordered; |
215 | enum omp_clause_schedule_kind sched_kind; | |
fd6481cf | 216 | struct omp_for_data_loop *loops; |
1e8e9920 | 217 | }; |
218 | ||
773c5ba7 | 219 | |
1e8e9920 | 220 | static splay_tree all_contexts; |
fd6481cf | 221 | static int taskreg_nesting_level; |
bc7bff74 | 222 | static int target_nesting_level; |
7740abd8 | 223 | static struct omp_region *root_omp_region; |
fd6481cf | 224 | static bitmap task_shared_vars; |
37eaded9 | 225 | static vec<omp_context *> taskreg_contexts; |
1e8e9920 | 226 | |
ab129075 | 227 | static void scan_omp (gimple_seq *, omp_context *); |
75a70cf9 | 228 | static tree scan_omp_1_op (tree *, int *, void *); |
229 | ||
230 | #define WALK_SUBSTMTS \ | |
231 | case GIMPLE_BIND: \ | |
232 | case GIMPLE_TRY: \ | |
233 | case GIMPLE_CATCH: \ | |
234 | case GIMPLE_EH_FILTER: \ | |
4c0315d0 | 235 | case GIMPLE_TRANSACTION: \ |
75a70cf9 | 236 | /* The sub-statements for these should be walked. */ \ |
237 | *handled_ops_p = false; \ | |
238 | break; | |
239 | ||
dccabdd1 | 240 | /* Holds offload tables with decls. */ |
241 | vec<tree, va_gc> *offload_funcs, *offload_vars; | |
242 | ||
75a70cf9 | 243 | /* Convenience function for calling scan_omp_1_op on tree operands. */ |
244 | ||
245 | static inline tree | |
246 | scan_omp_op (tree *tp, omp_context *ctx) | |
247 | { | |
248 | struct walk_stmt_info wi; | |
249 | ||
250 | memset (&wi, 0, sizeof (wi)); | |
251 | wi.info = ctx; | |
252 | wi.want_locations = true; | |
253 | ||
254 | return walk_tree (tp, scan_omp_1_op, &wi, NULL); | |
255 | } | |
256 | ||
e3a19533 | 257 | static void lower_omp (gimple_seq *, omp_context *); |
f49d7bb5 | 258 | static tree lookup_decl_in_outer_ctx (tree, omp_context *); |
259 | static tree maybe_lookup_decl_in_outer_ctx (tree, omp_context *); | |
1e8e9920 | 260 | |
261 | /* Find an OpenMP clause of type KIND within CLAUSES. */ | |
262 | ||
79acaae1 | 263 | tree |
590c3166 | 264 | find_omp_clause (tree clauses, enum omp_clause_code kind) |
1e8e9920 | 265 | { |
266 | for (; clauses ; clauses = OMP_CLAUSE_CHAIN (clauses)) | |
55d6e7cd | 267 | if (OMP_CLAUSE_CODE (clauses) == kind) |
1e8e9920 | 268 | return clauses; |
269 | ||
270 | return NULL_TREE; | |
271 | } | |
272 | ||
273 | /* Return true if CTX is for an omp parallel. */ | |
274 | ||
275 | static inline bool | |
276 | is_parallel_ctx (omp_context *ctx) | |
277 | { | |
75a70cf9 | 278 | return gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL; |
1e8e9920 | 279 | } |
280 | ||
773c5ba7 | 281 | |
b0c5e347 | 282 | /* Return true if CTX is for an omp target region. */ |
283 | ||
284 | static inline bool | |
285 | is_targetreg_ctx (omp_context *ctx) | |
286 | { | |
287 | return gimple_code (ctx->stmt) == GIMPLE_OMP_TARGET | |
288 | && gimple_omp_target_kind (ctx->stmt) == GF_OMP_TARGET_KIND_REGION; | |
289 | } | |
290 | ||
291 | ||
fd6481cf | 292 | /* Return true if CTX is for an omp task. */ |
293 | ||
294 | static inline bool | |
295 | is_task_ctx (omp_context *ctx) | |
296 | { | |
75a70cf9 | 297 | return gimple_code (ctx->stmt) == GIMPLE_OMP_TASK; |
fd6481cf | 298 | } |
299 | ||
300 | ||
301 | /* Return true if CTX is for an omp parallel or omp task. */ | |
302 | ||
303 | static inline bool | |
304 | is_taskreg_ctx (omp_context *ctx) | |
305 | { | |
75a70cf9 | 306 | return gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL |
307 | || gimple_code (ctx->stmt) == GIMPLE_OMP_TASK; | |
fd6481cf | 308 | } |
309 | ||
310 | ||
773c5ba7 | 311 | /* Return true if REGION is a combined parallel+workshare region. */ |
1e8e9920 | 312 | |
313 | static inline bool | |
773c5ba7 | 314 | is_combined_parallel (struct omp_region *region) |
315 | { | |
316 | return region->is_combined_parallel; | |
317 | } | |
318 | ||
319 | ||
320 | /* Extract the header elements of parallel loop FOR_STMT and store | |
321 | them into *FD. */ | |
322 | ||
323 | static void | |
1a91d914 | 324 | extract_omp_for_data (gomp_for *for_stmt, struct omp_for_data *fd, |
fd6481cf | 325 | struct omp_for_data_loop *loops) |
773c5ba7 | 326 | { |
fd6481cf | 327 | tree t, var, *collapse_iter, *collapse_count; |
328 | tree count = NULL_TREE, iter_type = long_integer_type_node; | |
329 | struct omp_for_data_loop *loop; | |
330 | int i; | |
331 | struct omp_for_data_loop dummy_loop; | |
389dd41b | 332 | location_t loc = gimple_location (for_stmt); |
10c55644 | 333 | bool simd = gimple_omp_for_kind (for_stmt) & GF_OMP_FOR_SIMD; |
bc7bff74 | 334 | bool distribute = gimple_omp_for_kind (for_stmt) |
335 | == GF_OMP_FOR_KIND_DISTRIBUTE; | |
773c5ba7 | 336 | |
337 | fd->for_stmt = for_stmt; | |
338 | fd->pre = NULL; | |
75a70cf9 | 339 | fd->collapse = gimple_omp_for_collapse (for_stmt); |
fd6481cf | 340 | if (fd->collapse > 1) |
341 | fd->loops = loops; | |
342 | else | |
343 | fd->loops = &fd->loop; | |
773c5ba7 | 344 | |
bc7bff74 | 345 | fd->have_nowait = distribute || simd; |
346 | fd->have_ordered = false; | |
773c5ba7 | 347 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC; |
348 | fd->chunk_size = NULL_TREE; | |
40750995 | 349 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_CILKFOR) |
350 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_CILKFOR; | |
fd6481cf | 351 | collapse_iter = NULL; |
352 | collapse_count = NULL; | |
773c5ba7 | 353 | |
75a70cf9 | 354 | for (t = gimple_omp_for_clauses (for_stmt); t ; t = OMP_CLAUSE_CHAIN (t)) |
55d6e7cd | 355 | switch (OMP_CLAUSE_CODE (t)) |
773c5ba7 | 356 | { |
357 | case OMP_CLAUSE_NOWAIT: | |
358 | fd->have_nowait = true; | |
359 | break; | |
360 | case OMP_CLAUSE_ORDERED: | |
361 | fd->have_ordered = true; | |
362 | break; | |
363 | case OMP_CLAUSE_SCHEDULE: | |
bc7bff74 | 364 | gcc_assert (!distribute); |
773c5ba7 | 365 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_KIND (t); |
366 | fd->chunk_size = OMP_CLAUSE_SCHEDULE_CHUNK_EXPR (t); | |
367 | break; | |
bc7bff74 | 368 | case OMP_CLAUSE_DIST_SCHEDULE: |
369 | gcc_assert (distribute); | |
370 | fd->chunk_size = OMP_CLAUSE_DIST_SCHEDULE_CHUNK_EXPR (t); | |
371 | break; | |
fd6481cf | 372 | case OMP_CLAUSE_COLLAPSE: |
373 | if (fd->collapse > 1) | |
374 | { | |
375 | collapse_iter = &OMP_CLAUSE_COLLAPSE_ITERVAR (t); | |
376 | collapse_count = &OMP_CLAUSE_COLLAPSE_COUNT (t); | |
377 | } | |
757abb46 | 378 | break; |
773c5ba7 | 379 | default: |
380 | break; | |
381 | } | |
382 | ||
fd6481cf | 383 | /* FIXME: for now map schedule(auto) to schedule(static). |
384 | There should be analysis to determine whether all iterations | |
385 | are approximately the same amount of work (then schedule(static) | |
bde357c8 | 386 | is best) or if it varies (then schedule(dynamic,N) is better). */ |
fd6481cf | 387 | if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_AUTO) |
388 | { | |
389 | fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC; | |
390 | gcc_assert (fd->chunk_size == NULL); | |
391 | } | |
392 | gcc_assert (fd->collapse == 1 || collapse_iter != NULL); | |
773c5ba7 | 393 | if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_RUNTIME) |
394 | gcc_assert (fd->chunk_size == NULL); | |
395 | else if (fd->chunk_size == NULL) | |
396 | { | |
397 | /* We only need to compute a default chunk size for ordered | |
398 | static loops and dynamic loops. */ | |
fd6481cf | 399 | if (fd->sched_kind != OMP_CLAUSE_SCHEDULE_STATIC |
bc7bff74 | 400 | || fd->have_ordered) |
773c5ba7 | 401 | fd->chunk_size = (fd->sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) |
402 | ? integer_zero_node : integer_one_node; | |
403 | } | |
fd6481cf | 404 | |
405 | for (i = 0; i < fd->collapse; i++) | |
406 | { | |
407 | if (fd->collapse == 1) | |
408 | loop = &fd->loop; | |
409 | else if (loops != NULL) | |
410 | loop = loops + i; | |
411 | else | |
412 | loop = &dummy_loop; | |
413 | ||
75a70cf9 | 414 | loop->v = gimple_omp_for_index (for_stmt, i); |
fd6481cf | 415 | gcc_assert (SSA_VAR_P (loop->v)); |
416 | gcc_assert (TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE | |
417 | || TREE_CODE (TREE_TYPE (loop->v)) == POINTER_TYPE); | |
418 | var = TREE_CODE (loop->v) == SSA_NAME ? SSA_NAME_VAR (loop->v) : loop->v; | |
75a70cf9 | 419 | loop->n1 = gimple_omp_for_initial (for_stmt, i); |
fd6481cf | 420 | |
75a70cf9 | 421 | loop->cond_code = gimple_omp_for_cond (for_stmt, i); |
422 | loop->n2 = gimple_omp_for_final (for_stmt, i); | |
fd6481cf | 423 | switch (loop->cond_code) |
424 | { | |
425 | case LT_EXPR: | |
426 | case GT_EXPR: | |
427 | break; | |
f2697631 | 428 | case NE_EXPR: |
429 | gcc_assert (gimple_omp_for_kind (for_stmt) | |
40750995 | 430 | == GF_OMP_FOR_KIND_CILKSIMD |
431 | || (gimple_omp_for_kind (for_stmt) | |
432 | == GF_OMP_FOR_KIND_CILKFOR)); | |
f2697631 | 433 | break; |
fd6481cf | 434 | case LE_EXPR: |
435 | if (POINTER_TYPE_P (TREE_TYPE (loop->n2))) | |
2cc66f2a | 436 | loop->n2 = fold_build_pointer_plus_hwi_loc (loc, loop->n2, 1); |
fd6481cf | 437 | else |
389dd41b | 438 | loop->n2 = fold_build2_loc (loc, |
439 | PLUS_EXPR, TREE_TYPE (loop->n2), loop->n2, | |
fd6481cf | 440 | build_int_cst (TREE_TYPE (loop->n2), 1)); |
441 | loop->cond_code = LT_EXPR; | |
442 | break; | |
443 | case GE_EXPR: | |
444 | if (POINTER_TYPE_P (TREE_TYPE (loop->n2))) | |
2cc66f2a | 445 | loop->n2 = fold_build_pointer_plus_hwi_loc (loc, loop->n2, -1); |
fd6481cf | 446 | else |
389dd41b | 447 | loop->n2 = fold_build2_loc (loc, |
448 | MINUS_EXPR, TREE_TYPE (loop->n2), loop->n2, | |
fd6481cf | 449 | build_int_cst (TREE_TYPE (loop->n2), 1)); |
450 | loop->cond_code = GT_EXPR; | |
451 | break; | |
452 | default: | |
453 | gcc_unreachable (); | |
454 | } | |
455 | ||
75a70cf9 | 456 | t = gimple_omp_for_incr (for_stmt, i); |
fd6481cf | 457 | gcc_assert (TREE_OPERAND (t, 0) == var); |
458 | switch (TREE_CODE (t)) | |
459 | { | |
460 | case PLUS_EXPR: | |
fd6481cf | 461 | loop->step = TREE_OPERAND (t, 1); |
462 | break; | |
85d86b55 | 463 | case POINTER_PLUS_EXPR: |
464 | loop->step = fold_convert (ssizetype, TREE_OPERAND (t, 1)); | |
465 | break; | |
fd6481cf | 466 | case MINUS_EXPR: |
467 | loop->step = TREE_OPERAND (t, 1); | |
389dd41b | 468 | loop->step = fold_build1_loc (loc, |
469 | NEGATE_EXPR, TREE_TYPE (loop->step), | |
fd6481cf | 470 | loop->step); |
471 | break; | |
472 | default: | |
473 | gcc_unreachable (); | |
474 | } | |
475 | ||
bc7bff74 | 476 | if (simd |
477 | || (fd->sched_kind == OMP_CLAUSE_SCHEDULE_STATIC | |
478 | && !fd->have_ordered)) | |
3d483a94 | 479 | { |
480 | if (fd->collapse == 1) | |
481 | iter_type = TREE_TYPE (loop->v); | |
482 | else if (i == 0 | |
483 | || TYPE_PRECISION (iter_type) | |
484 | < TYPE_PRECISION (TREE_TYPE (loop->v))) | |
485 | iter_type | |
486 | = build_nonstandard_integer_type | |
bc7bff74 | 487 | (TYPE_PRECISION (TREE_TYPE (loop->v)), 1); |
3d483a94 | 488 | } |
489 | else if (iter_type != long_long_unsigned_type_node) | |
fd6481cf | 490 | { |
491 | if (POINTER_TYPE_P (TREE_TYPE (loop->v))) | |
492 | iter_type = long_long_unsigned_type_node; | |
493 | else if (TYPE_UNSIGNED (TREE_TYPE (loop->v)) | |
494 | && TYPE_PRECISION (TREE_TYPE (loop->v)) | |
495 | >= TYPE_PRECISION (iter_type)) | |
496 | { | |
497 | tree n; | |
498 | ||
499 | if (loop->cond_code == LT_EXPR) | |
389dd41b | 500 | n = fold_build2_loc (loc, |
501 | PLUS_EXPR, TREE_TYPE (loop->v), | |
fd6481cf | 502 | loop->n2, loop->step); |
503 | else | |
504 | n = loop->n1; | |
505 | if (TREE_CODE (n) != INTEGER_CST | |
506 | || tree_int_cst_lt (TYPE_MAX_VALUE (iter_type), n)) | |
507 | iter_type = long_long_unsigned_type_node; | |
508 | } | |
509 | else if (TYPE_PRECISION (TREE_TYPE (loop->v)) | |
510 | > TYPE_PRECISION (iter_type)) | |
511 | { | |
512 | tree n1, n2; | |
513 | ||
514 | if (loop->cond_code == LT_EXPR) | |
515 | { | |
516 | n1 = loop->n1; | |
389dd41b | 517 | n2 = fold_build2_loc (loc, |
518 | PLUS_EXPR, TREE_TYPE (loop->v), | |
fd6481cf | 519 | loop->n2, loop->step); |
520 | } | |
521 | else | |
522 | { | |
389dd41b | 523 | n1 = fold_build2_loc (loc, |
524 | MINUS_EXPR, TREE_TYPE (loop->v), | |
fd6481cf | 525 | loop->n2, loop->step); |
526 | n2 = loop->n1; | |
527 | } | |
528 | if (TREE_CODE (n1) != INTEGER_CST | |
529 | || TREE_CODE (n2) != INTEGER_CST | |
530 | || !tree_int_cst_lt (TYPE_MIN_VALUE (iter_type), n1) | |
531 | || !tree_int_cst_lt (n2, TYPE_MAX_VALUE (iter_type))) | |
532 | iter_type = long_long_unsigned_type_node; | |
533 | } | |
534 | } | |
535 | ||
536 | if (collapse_count && *collapse_count == NULL) | |
537 | { | |
8e6b4515 | 538 | t = fold_binary (loop->cond_code, boolean_type_node, |
539 | fold_convert (TREE_TYPE (loop->v), loop->n1), | |
540 | fold_convert (TREE_TYPE (loop->v), loop->n2)); | |
541 | if (t && integer_zerop (t)) | |
542 | count = build_zero_cst (long_long_unsigned_type_node); | |
543 | else if ((i == 0 || count != NULL_TREE) | |
544 | && TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE | |
545 | && TREE_CONSTANT (loop->n1) | |
546 | && TREE_CONSTANT (loop->n2) | |
547 | && TREE_CODE (loop->step) == INTEGER_CST) | |
fd6481cf | 548 | { |
549 | tree itype = TREE_TYPE (loop->v); | |
550 | ||
551 | if (POINTER_TYPE_P (itype)) | |
3cea8318 | 552 | itype = signed_type_for (itype); |
fd6481cf | 553 | t = build_int_cst (itype, (loop->cond_code == LT_EXPR ? -1 : 1)); |
389dd41b | 554 | t = fold_build2_loc (loc, |
555 | PLUS_EXPR, itype, | |
556 | fold_convert_loc (loc, itype, loop->step), t); | |
557 | t = fold_build2_loc (loc, PLUS_EXPR, itype, t, | |
558 | fold_convert_loc (loc, itype, loop->n2)); | |
559 | t = fold_build2_loc (loc, MINUS_EXPR, itype, t, | |
560 | fold_convert_loc (loc, itype, loop->n1)); | |
fd6481cf | 561 | if (TYPE_UNSIGNED (itype) && loop->cond_code == GT_EXPR) |
389dd41b | 562 | t = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, |
563 | fold_build1_loc (loc, NEGATE_EXPR, itype, t), | |
564 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
565 | fold_convert_loc (loc, itype, | |
566 | loop->step))); | |
fd6481cf | 567 | else |
389dd41b | 568 | t = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, t, |
569 | fold_convert_loc (loc, itype, loop->step)); | |
570 | t = fold_convert_loc (loc, long_long_unsigned_type_node, t); | |
fd6481cf | 571 | if (count != NULL_TREE) |
389dd41b | 572 | count = fold_build2_loc (loc, |
573 | MULT_EXPR, long_long_unsigned_type_node, | |
fd6481cf | 574 | count, t); |
575 | else | |
576 | count = t; | |
577 | if (TREE_CODE (count) != INTEGER_CST) | |
578 | count = NULL_TREE; | |
579 | } | |
8e6b4515 | 580 | else if (count && !integer_zerop (count)) |
fd6481cf | 581 | count = NULL_TREE; |
582 | } | |
583 | } | |
584 | ||
3d483a94 | 585 | if (count |
bc7bff74 | 586 | && !simd |
587 | && (fd->sched_kind != OMP_CLAUSE_SCHEDULE_STATIC | |
588 | || fd->have_ordered)) | |
fd6481cf | 589 | { |
590 | if (!tree_int_cst_lt (count, TYPE_MAX_VALUE (long_integer_type_node))) | |
591 | iter_type = long_long_unsigned_type_node; | |
592 | else | |
593 | iter_type = long_integer_type_node; | |
594 | } | |
595 | else if (collapse_iter && *collapse_iter != NULL) | |
596 | iter_type = TREE_TYPE (*collapse_iter); | |
597 | fd->iter_type = iter_type; | |
598 | if (collapse_iter && *collapse_iter == NULL) | |
599 | *collapse_iter = create_tmp_var (iter_type, ".iter"); | |
600 | if (collapse_count && *collapse_count == NULL) | |
601 | { | |
602 | if (count) | |
389dd41b | 603 | *collapse_count = fold_convert_loc (loc, iter_type, count); |
fd6481cf | 604 | else |
605 | *collapse_count = create_tmp_var (iter_type, ".count"); | |
606 | } | |
607 | ||
608 | if (fd->collapse > 1) | |
609 | { | |
610 | fd->loop.v = *collapse_iter; | |
611 | fd->loop.n1 = build_int_cst (TREE_TYPE (fd->loop.v), 0); | |
612 | fd->loop.n2 = *collapse_count; | |
613 | fd->loop.step = build_int_cst (TREE_TYPE (fd->loop.v), 1); | |
614 | fd->loop.cond_code = LT_EXPR; | |
615 | } | |
773c5ba7 | 616 | } |
617 | ||
618 | ||
619 | /* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB | |
620 | is the immediate dominator of PAR_ENTRY_BB, return true if there | |
621 | are no data dependencies that would prevent expanding the parallel | |
622 | directive at PAR_ENTRY_BB as a combined parallel+workshare region. | |
623 | ||
624 | When expanding a combined parallel+workshare region, the call to | |
625 | the child function may need additional arguments in the case of | |
75a70cf9 | 626 | GIMPLE_OMP_FOR regions. In some cases, these arguments are |
627 | computed out of variables passed in from the parent to the child | |
628 | via 'struct .omp_data_s'. For instance: | |
773c5ba7 | 629 | |
630 | #pragma omp parallel for schedule (guided, i * 4) | |
631 | for (j ...) | |
632 | ||
633 | Is lowered into: | |
634 | ||
635 | # BLOCK 2 (PAR_ENTRY_BB) | |
636 | .omp_data_o.i = i; | |
637 | #pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598) | |
48e1416a | 638 | |
773c5ba7 | 639 | # BLOCK 3 (WS_ENTRY_BB) |
640 | .omp_data_i = &.omp_data_o; | |
641 | D.1667 = .omp_data_i->i; | |
642 | D.1598 = D.1667 * 4; | |
643 | #pragma omp for schedule (guided, D.1598) | |
644 | ||
645 | When we outline the parallel region, the call to the child function | |
646 | 'bar.omp_fn.0' will need the value D.1598 in its argument list, but | |
647 | that value is computed *after* the call site. So, in principle we | |
648 | cannot do the transformation. | |
649 | ||
650 | To see whether the code in WS_ENTRY_BB blocks the combined | |
651 | parallel+workshare call, we collect all the variables used in the | |
75a70cf9 | 652 | GIMPLE_OMP_FOR header check whether they appear on the LHS of any |
773c5ba7 | 653 | statement in WS_ENTRY_BB. If so, then we cannot emit the combined |
654 | call. | |
655 | ||
656 | FIXME. If we had the SSA form built at this point, we could merely | |
657 | hoist the code in block 3 into block 2 and be done with it. But at | |
658 | this point we don't have dataflow information and though we could | |
659 | hack something up here, it is really not worth the aggravation. */ | |
660 | ||
661 | static bool | |
f018d957 | 662 | workshare_safe_to_combine_p (basic_block ws_entry_bb) |
773c5ba7 | 663 | { |
664 | struct omp_for_data fd; | |
f018d957 | 665 | gimple ws_stmt = last_stmt (ws_entry_bb); |
773c5ba7 | 666 | |
75a70cf9 | 667 | if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) |
773c5ba7 | 668 | return true; |
669 | ||
75a70cf9 | 670 | gcc_assert (gimple_code (ws_stmt) == GIMPLE_OMP_FOR); |
773c5ba7 | 671 | |
1a91d914 | 672 | extract_omp_for_data (as_a <gomp_for *> (ws_stmt), &fd, NULL); |
fd6481cf | 673 | |
674 | if (fd.collapse > 1 && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
675 | return false; | |
676 | if (fd.iter_type != long_integer_type_node) | |
677 | return false; | |
773c5ba7 | 678 | |
679 | /* FIXME. We give up too easily here. If any of these arguments | |
680 | are not constants, they will likely involve variables that have | |
681 | been mapped into fields of .omp_data_s for sharing with the child | |
682 | function. With appropriate data flow, it would be possible to | |
683 | see through this. */ | |
fd6481cf | 684 | if (!is_gimple_min_invariant (fd.loop.n1) |
685 | || !is_gimple_min_invariant (fd.loop.n2) | |
686 | || !is_gimple_min_invariant (fd.loop.step) | |
773c5ba7 | 687 | || (fd.chunk_size && !is_gimple_min_invariant (fd.chunk_size))) |
688 | return false; | |
689 | ||
690 | return true; | |
691 | } | |
692 | ||
693 | ||
694 | /* Collect additional arguments needed to emit a combined | |
695 | parallel+workshare call. WS_STMT is the workshare directive being | |
696 | expanded. */ | |
697 | ||
f1f41a6c | 698 | static vec<tree, va_gc> * |
bc7bff74 | 699 | get_ws_args_for (gimple par_stmt, gimple ws_stmt) |
773c5ba7 | 700 | { |
701 | tree t; | |
389dd41b | 702 | location_t loc = gimple_location (ws_stmt); |
f1f41a6c | 703 | vec<tree, va_gc> *ws_args; |
773c5ba7 | 704 | |
1a91d914 | 705 | if (gomp_for *for_stmt = dyn_cast <gomp_for *> (ws_stmt)) |
773c5ba7 | 706 | { |
707 | struct omp_for_data fd; | |
bc7bff74 | 708 | tree n1, n2; |
773c5ba7 | 709 | |
1a91d914 | 710 | extract_omp_for_data (for_stmt, &fd, NULL); |
bc7bff74 | 711 | n1 = fd.loop.n1; |
712 | n2 = fd.loop.n2; | |
713 | ||
1a91d914 | 714 | if (gimple_omp_for_combined_into_p (for_stmt)) |
bc7bff74 | 715 | { |
716 | tree innerc | |
717 | = find_omp_clause (gimple_omp_parallel_clauses (par_stmt), | |
718 | OMP_CLAUSE__LOOPTEMP_); | |
719 | gcc_assert (innerc); | |
720 | n1 = OMP_CLAUSE_DECL (innerc); | |
721 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
722 | OMP_CLAUSE__LOOPTEMP_); | |
723 | gcc_assert (innerc); | |
724 | n2 = OMP_CLAUSE_DECL (innerc); | |
725 | } | |
773c5ba7 | 726 | |
f1f41a6c | 727 | vec_alloc (ws_args, 3 + (fd.chunk_size != 0)); |
773c5ba7 | 728 | |
bc7bff74 | 729 | t = fold_convert_loc (loc, long_integer_type_node, n1); |
f1f41a6c | 730 | ws_args->quick_push (t); |
773c5ba7 | 731 | |
bc7bff74 | 732 | t = fold_convert_loc (loc, long_integer_type_node, n2); |
f1f41a6c | 733 | ws_args->quick_push (t); |
773c5ba7 | 734 | |
414c3a2c | 735 | t = fold_convert_loc (loc, long_integer_type_node, fd.loop.step); |
f1f41a6c | 736 | ws_args->quick_push (t); |
414c3a2c | 737 | |
738 | if (fd.chunk_size) | |
739 | { | |
740 | t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size); | |
f1f41a6c | 741 | ws_args->quick_push (t); |
414c3a2c | 742 | } |
773c5ba7 | 743 | |
744 | return ws_args; | |
745 | } | |
75a70cf9 | 746 | else if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) |
773c5ba7 | 747 | { |
ac6e3339 | 748 | /* Number of sections is equal to the number of edges from the |
75a70cf9 | 749 | GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to |
750 | the exit of the sections region. */ | |
751 | basic_block bb = single_succ (gimple_bb (ws_stmt)); | |
ac6e3339 | 752 | t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1); |
f1f41a6c | 753 | vec_alloc (ws_args, 1); |
754 | ws_args->quick_push (t); | |
414c3a2c | 755 | return ws_args; |
773c5ba7 | 756 | } |
757 | ||
758 | gcc_unreachable (); | |
759 | } | |
760 | ||
761 | ||
762 | /* Discover whether REGION is a combined parallel+workshare region. */ | |
763 | ||
764 | static void | |
765 | determine_parallel_type (struct omp_region *region) | |
1e8e9920 | 766 | { |
773c5ba7 | 767 | basic_block par_entry_bb, par_exit_bb; |
768 | basic_block ws_entry_bb, ws_exit_bb; | |
769 | ||
03ed154b | 770 | if (region == NULL || region->inner == NULL |
ac6e3339 | 771 | || region->exit == NULL || region->inner->exit == NULL |
772 | || region->inner->cont == NULL) | |
773c5ba7 | 773 | return; |
774 | ||
775 | /* We only support parallel+for and parallel+sections. */ | |
75a70cf9 | 776 | if (region->type != GIMPLE_OMP_PARALLEL |
777 | || (region->inner->type != GIMPLE_OMP_FOR | |
778 | && region->inner->type != GIMPLE_OMP_SECTIONS)) | |
773c5ba7 | 779 | return; |
780 | ||
781 | /* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and | |
782 | WS_EXIT_BB -> PAR_EXIT_BB. */ | |
61e47ac8 | 783 | par_entry_bb = region->entry; |
784 | par_exit_bb = region->exit; | |
785 | ws_entry_bb = region->inner->entry; | |
786 | ws_exit_bb = region->inner->exit; | |
773c5ba7 | 787 | |
788 | if (single_succ (par_entry_bb) == ws_entry_bb | |
789 | && single_succ (ws_exit_bb) == par_exit_bb | |
f018d957 | 790 | && workshare_safe_to_combine_p (ws_entry_bb) |
75a70cf9 | 791 | && (gimple_omp_parallel_combined_p (last_stmt (par_entry_bb)) |
de7ef844 | 792 | || (last_and_only_stmt (ws_entry_bb) |
793 | && last_and_only_stmt (par_exit_bb)))) | |
773c5ba7 | 794 | { |
bc7bff74 | 795 | gimple par_stmt = last_stmt (par_entry_bb); |
75a70cf9 | 796 | gimple ws_stmt = last_stmt (ws_entry_bb); |
61e47ac8 | 797 | |
75a70cf9 | 798 | if (region->inner->type == GIMPLE_OMP_FOR) |
773c5ba7 | 799 | { |
800 | /* If this is a combined parallel loop, we need to determine | |
801 | whether or not to use the combined library calls. There | |
802 | are two cases where we do not apply the transformation: | |
803 | static loops and any kind of ordered loop. In the first | |
804 | case, we already open code the loop so there is no need | |
805 | to do anything else. In the latter case, the combined | |
806 | parallel loop call would still need extra synchronization | |
807 | to implement ordered semantics, so there would not be any | |
808 | gain in using the combined call. */ | |
75a70cf9 | 809 | tree clauses = gimple_omp_for_clauses (ws_stmt); |
773c5ba7 | 810 | tree c = find_omp_clause (clauses, OMP_CLAUSE_SCHEDULE); |
811 | if (c == NULL | |
812 | || OMP_CLAUSE_SCHEDULE_KIND (c) == OMP_CLAUSE_SCHEDULE_STATIC | |
813 | || find_omp_clause (clauses, OMP_CLAUSE_ORDERED)) | |
814 | { | |
815 | region->is_combined_parallel = false; | |
816 | region->inner->is_combined_parallel = false; | |
817 | return; | |
818 | } | |
819 | } | |
820 | ||
821 | region->is_combined_parallel = true; | |
822 | region->inner->is_combined_parallel = true; | |
bc7bff74 | 823 | region->ws_args = get_ws_args_for (par_stmt, ws_stmt); |
773c5ba7 | 824 | } |
1e8e9920 | 825 | } |
826 | ||
773c5ba7 | 827 | |
1e8e9920 | 828 | /* Return true if EXPR is variable sized. */ |
829 | ||
830 | static inline bool | |
1f1872fd | 831 | is_variable_sized (const_tree expr) |
1e8e9920 | 832 | { |
833 | return !TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (expr))); | |
834 | } | |
835 | ||
836 | /* Return true if DECL is a reference type. */ | |
837 | ||
838 | static inline bool | |
839 | is_reference (tree decl) | |
840 | { | |
841 | return lang_hooks.decls.omp_privatize_by_reference (decl); | |
842 | } | |
843 | ||
844 | /* Lookup variables in the decl or field splay trees. The "maybe" form | |
845 | allows for the variable form to not have been entered, otherwise we | |
846 | assert that the variable must have been entered. */ | |
847 | ||
848 | static inline tree | |
849 | lookup_decl (tree var, omp_context *ctx) | |
850 | { | |
06ecf488 | 851 | tree *n = ctx->cb.decl_map->get (var); |
e3022db7 | 852 | return *n; |
1e8e9920 | 853 | } |
854 | ||
855 | static inline tree | |
e8a588af | 856 | maybe_lookup_decl (const_tree var, omp_context *ctx) |
1e8e9920 | 857 | { |
06ecf488 | 858 | tree *n = ctx->cb.decl_map->get (const_cast<tree> (var)); |
e3022db7 | 859 | return n ? *n : NULL_TREE; |
1e8e9920 | 860 | } |
861 | ||
862 | static inline tree | |
863 | lookup_field (tree var, omp_context *ctx) | |
864 | { | |
865 | splay_tree_node n; | |
866 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) var); | |
867 | return (tree) n->value; | |
868 | } | |
869 | ||
fd6481cf | 870 | static inline tree |
871 | lookup_sfield (tree var, omp_context *ctx) | |
872 | { | |
873 | splay_tree_node n; | |
874 | n = splay_tree_lookup (ctx->sfield_map | |
875 | ? ctx->sfield_map : ctx->field_map, | |
876 | (splay_tree_key) var); | |
877 | return (tree) n->value; | |
878 | } | |
879 | ||
1e8e9920 | 880 | static inline tree |
881 | maybe_lookup_field (tree var, omp_context *ctx) | |
882 | { | |
883 | splay_tree_node n; | |
884 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) var); | |
885 | return n ? (tree) n->value : NULL_TREE; | |
886 | } | |
887 | ||
e8a588af | 888 | /* Return true if DECL should be copied by pointer. SHARED_CTX is |
889 | the parallel context if DECL is to be shared. */ | |
1e8e9920 | 890 | |
891 | static bool | |
fd6481cf | 892 | use_pointer_for_field (tree decl, omp_context *shared_ctx) |
1e8e9920 | 893 | { |
894 | if (AGGREGATE_TYPE_P (TREE_TYPE (decl))) | |
895 | return true; | |
896 | ||
554f2707 | 897 | /* We can only use copy-in/copy-out semantics for shared variables |
1e8e9920 | 898 | when we know the value is not accessible from an outer scope. */ |
e8a588af | 899 | if (shared_ctx) |
1e8e9920 | 900 | { |
901 | /* ??? Trivially accessible from anywhere. But why would we even | |
902 | be passing an address in this case? Should we simply assert | |
903 | this to be false, or should we have a cleanup pass that removes | |
904 | these from the list of mappings? */ | |
905 | if (TREE_STATIC (decl) || DECL_EXTERNAL (decl)) | |
906 | return true; | |
907 | ||
908 | /* For variables with DECL_HAS_VALUE_EXPR_P set, we cannot tell | |
909 | without analyzing the expression whether or not its location | |
910 | is accessible to anyone else. In the case of nested parallel | |
911 | regions it certainly may be. */ | |
df2c34fc | 912 | if (TREE_CODE (decl) != RESULT_DECL && DECL_HAS_VALUE_EXPR_P (decl)) |
1e8e9920 | 913 | return true; |
914 | ||
915 | /* Do not use copy-in/copy-out for variables that have their | |
916 | address taken. */ | |
917 | if (TREE_ADDRESSABLE (decl)) | |
918 | return true; | |
e8a588af | 919 | |
b8214689 | 920 | /* lower_send_shared_vars only uses copy-in, but not copy-out |
921 | for these. */ | |
922 | if (TREE_READONLY (decl) | |
923 | || ((TREE_CODE (decl) == RESULT_DECL | |
924 | || TREE_CODE (decl) == PARM_DECL) | |
925 | && DECL_BY_REFERENCE (decl))) | |
926 | return false; | |
927 | ||
e8a588af | 928 | /* Disallow copy-in/out in nested parallel if |
929 | decl is shared in outer parallel, otherwise | |
930 | each thread could store the shared variable | |
931 | in its own copy-in location, making the | |
932 | variable no longer really shared. */ | |
b8214689 | 933 | if (shared_ctx->is_nested) |
e8a588af | 934 | { |
935 | omp_context *up; | |
936 | ||
937 | for (up = shared_ctx->outer; up; up = up->outer) | |
0cb159ec | 938 | if (is_taskreg_ctx (up) && maybe_lookup_decl (decl, up)) |
e8a588af | 939 | break; |
940 | ||
0cb159ec | 941 | if (up) |
e8a588af | 942 | { |
943 | tree c; | |
944 | ||
75a70cf9 | 945 | for (c = gimple_omp_taskreg_clauses (up->stmt); |
e8a588af | 946 | c; c = OMP_CLAUSE_CHAIN (c)) |
947 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_SHARED | |
948 | && OMP_CLAUSE_DECL (c) == decl) | |
949 | break; | |
950 | ||
951 | if (c) | |
784ad964 | 952 | goto maybe_mark_addressable_and_ret; |
e8a588af | 953 | } |
954 | } | |
fd6481cf | 955 | |
b8214689 | 956 | /* For tasks avoid using copy-in/out. As tasks can be |
fd6481cf | 957 | deferred or executed in different thread, when GOMP_task |
958 | returns, the task hasn't necessarily terminated. */ | |
b8214689 | 959 | if (is_task_ctx (shared_ctx)) |
fd6481cf | 960 | { |
784ad964 | 961 | tree outer; |
962 | maybe_mark_addressable_and_ret: | |
963 | outer = maybe_lookup_decl_in_outer_ctx (decl, shared_ctx); | |
fd6481cf | 964 | if (is_gimple_reg (outer)) |
965 | { | |
966 | /* Taking address of OUTER in lower_send_shared_vars | |
967 | might need regimplification of everything that uses the | |
968 | variable. */ | |
969 | if (!task_shared_vars) | |
970 | task_shared_vars = BITMAP_ALLOC (NULL); | |
971 | bitmap_set_bit (task_shared_vars, DECL_UID (outer)); | |
972 | TREE_ADDRESSABLE (outer) = 1; | |
973 | } | |
974 | return true; | |
975 | } | |
1e8e9920 | 976 | } |
977 | ||
978 | return false; | |
979 | } | |
980 | ||
79acaae1 | 981 | /* Construct a new automatic decl similar to VAR. */ |
982 | ||
983 | static tree | |
984 | omp_copy_decl_2 (tree var, tree name, tree type, omp_context *ctx) | |
985 | { | |
986 | tree copy = copy_var_decl (var, name, type); | |
987 | ||
988 | DECL_CONTEXT (copy) = current_function_decl; | |
1767a056 | 989 | DECL_CHAIN (copy) = ctx->block_vars; |
1e8e9920 | 990 | ctx->block_vars = copy; |
991 | ||
992 | return copy; | |
993 | } | |
994 | ||
995 | static tree | |
996 | omp_copy_decl_1 (tree var, omp_context *ctx) | |
997 | { | |
998 | return omp_copy_decl_2 (var, DECL_NAME (var), TREE_TYPE (var), ctx); | |
999 | } | |
1000 | ||
445d06b6 | 1001 | /* Build COMPONENT_REF and set TREE_THIS_VOLATILE and TREE_READONLY on it |
1002 | as appropriate. */ | |
1003 | static tree | |
1004 | omp_build_component_ref (tree obj, tree field) | |
1005 | { | |
1006 | tree ret = build3 (COMPONENT_REF, TREE_TYPE (field), obj, field, NULL); | |
1007 | if (TREE_THIS_VOLATILE (field)) | |
1008 | TREE_THIS_VOLATILE (ret) |= 1; | |
1009 | if (TREE_READONLY (field)) | |
1010 | TREE_READONLY (ret) |= 1; | |
1011 | return ret; | |
1012 | } | |
1013 | ||
1e8e9920 | 1014 | /* Build tree nodes to access the field for VAR on the receiver side. */ |
1015 | ||
1016 | static tree | |
1017 | build_receiver_ref (tree var, bool by_ref, omp_context *ctx) | |
1018 | { | |
1019 | tree x, field = lookup_field (var, ctx); | |
1020 | ||
1021 | /* If the receiver record type was remapped in the child function, | |
1022 | remap the field into the new record type. */ | |
1023 | x = maybe_lookup_field (field, ctx); | |
1024 | if (x != NULL) | |
1025 | field = x; | |
1026 | ||
182cf5a9 | 1027 | x = build_simple_mem_ref (ctx->receiver_decl); |
445d06b6 | 1028 | x = omp_build_component_ref (x, field); |
1e8e9920 | 1029 | if (by_ref) |
182cf5a9 | 1030 | x = build_simple_mem_ref (x); |
1e8e9920 | 1031 | |
1032 | return x; | |
1033 | } | |
1034 | ||
1035 | /* Build tree nodes to access VAR in the scope outer to CTX. In the case | |
1036 | of a parallel, this is a component reference; for workshare constructs | |
1037 | this is some variable. */ | |
1038 | ||
1039 | static tree | |
1040 | build_outer_var_ref (tree var, omp_context *ctx) | |
1041 | { | |
1042 | tree x; | |
1043 | ||
f49d7bb5 | 1044 | if (is_global_var (maybe_lookup_decl_in_outer_ctx (var, ctx))) |
1e8e9920 | 1045 | x = var; |
1046 | else if (is_variable_sized (var)) | |
1047 | { | |
1048 | x = TREE_OPERAND (DECL_VALUE_EXPR (var), 0); | |
1049 | x = build_outer_var_ref (x, ctx); | |
182cf5a9 | 1050 | x = build_simple_mem_ref (x); |
1e8e9920 | 1051 | } |
fd6481cf | 1052 | else if (is_taskreg_ctx (ctx)) |
1e8e9920 | 1053 | { |
e8a588af | 1054 | bool by_ref = use_pointer_for_field (var, NULL); |
1e8e9920 | 1055 | x = build_receiver_ref (var, by_ref, ctx); |
1056 | } | |
3d483a94 | 1057 | else if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR |
10c55644 | 1058 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_SIMD) |
3d483a94 | 1059 | { |
1060 | /* #pragma omp simd isn't a worksharing construct, and can reference even | |
1061 | private vars in its linear etc. clauses. */ | |
1062 | x = NULL_TREE; | |
1063 | if (ctx->outer && is_taskreg_ctx (ctx)) | |
1064 | x = lookup_decl (var, ctx->outer); | |
1065 | else if (ctx->outer) | |
84cb1020 | 1066 | x = maybe_lookup_decl_in_outer_ctx (var, ctx); |
3d483a94 | 1067 | if (x == NULL_TREE) |
1068 | x = var; | |
1069 | } | |
1e8e9920 | 1070 | else if (ctx->outer) |
1071 | x = lookup_decl (var, ctx->outer); | |
9438af57 | 1072 | else if (is_reference (var)) |
1073 | /* This can happen with orphaned constructs. If var is reference, it is | |
1074 | possible it is shared and as such valid. */ | |
1075 | x = var; | |
1e8e9920 | 1076 | else |
1077 | gcc_unreachable (); | |
1078 | ||
1079 | if (is_reference (var)) | |
182cf5a9 | 1080 | x = build_simple_mem_ref (x); |
1e8e9920 | 1081 | |
1082 | return x; | |
1083 | } | |
1084 | ||
1085 | /* Build tree nodes to access the field for VAR on the sender side. */ | |
1086 | ||
1087 | static tree | |
1088 | build_sender_ref (tree var, omp_context *ctx) | |
1089 | { | |
fd6481cf | 1090 | tree field = lookup_sfield (var, ctx); |
445d06b6 | 1091 | return omp_build_component_ref (ctx->sender_decl, field); |
1e8e9920 | 1092 | } |
1093 | ||
1094 | /* Add a new field for VAR inside the structure CTX->SENDER_DECL. */ | |
1095 | ||
1096 | static void | |
fd6481cf | 1097 | install_var_field (tree var, bool by_ref, int mask, omp_context *ctx) |
1e8e9920 | 1098 | { |
fd6481cf | 1099 | tree field, type, sfield = NULL_TREE; |
1e8e9920 | 1100 | |
fd6481cf | 1101 | gcc_assert ((mask & 1) == 0 |
1102 | || !splay_tree_lookup (ctx->field_map, (splay_tree_key) var)); | |
1103 | gcc_assert ((mask & 2) == 0 || !ctx->sfield_map | |
1104 | || !splay_tree_lookup (ctx->sfield_map, (splay_tree_key) var)); | |
1e8e9920 | 1105 | |
1106 | type = TREE_TYPE (var); | |
bc7bff74 | 1107 | if (mask & 4) |
1108 | { | |
1109 | gcc_assert (TREE_CODE (type) == ARRAY_TYPE); | |
1110 | type = build_pointer_type (build_pointer_type (type)); | |
1111 | } | |
1112 | else if (by_ref) | |
1e8e9920 | 1113 | type = build_pointer_type (type); |
fd6481cf | 1114 | else if ((mask & 3) == 1 && is_reference (var)) |
1115 | type = TREE_TYPE (type); | |
1e8e9920 | 1116 | |
e60a6f7b | 1117 | field = build_decl (DECL_SOURCE_LOCATION (var), |
1118 | FIELD_DECL, DECL_NAME (var), type); | |
1e8e9920 | 1119 | |
1120 | /* Remember what variable this field was created for. This does have a | |
1121 | side effect of making dwarf2out ignore this member, so for helpful | |
1122 | debugging we clear it later in delete_omp_context. */ | |
1123 | DECL_ABSTRACT_ORIGIN (field) = var; | |
fd6481cf | 1124 | if (type == TREE_TYPE (var)) |
1125 | { | |
1126 | DECL_ALIGN (field) = DECL_ALIGN (var); | |
1127 | DECL_USER_ALIGN (field) = DECL_USER_ALIGN (var); | |
1128 | TREE_THIS_VOLATILE (field) = TREE_THIS_VOLATILE (var); | |
1129 | } | |
1130 | else | |
1131 | DECL_ALIGN (field) = TYPE_ALIGN (type); | |
1e8e9920 | 1132 | |
fd6481cf | 1133 | if ((mask & 3) == 3) |
1134 | { | |
1135 | insert_field_into_struct (ctx->record_type, field); | |
1136 | if (ctx->srecord_type) | |
1137 | { | |
e60a6f7b | 1138 | sfield = build_decl (DECL_SOURCE_LOCATION (var), |
1139 | FIELD_DECL, DECL_NAME (var), type); | |
fd6481cf | 1140 | DECL_ABSTRACT_ORIGIN (sfield) = var; |
1141 | DECL_ALIGN (sfield) = DECL_ALIGN (field); | |
1142 | DECL_USER_ALIGN (sfield) = DECL_USER_ALIGN (field); | |
1143 | TREE_THIS_VOLATILE (sfield) = TREE_THIS_VOLATILE (field); | |
1144 | insert_field_into_struct (ctx->srecord_type, sfield); | |
1145 | } | |
1146 | } | |
1147 | else | |
1148 | { | |
1149 | if (ctx->srecord_type == NULL_TREE) | |
1150 | { | |
1151 | tree t; | |
1152 | ||
1153 | ctx->srecord_type = lang_hooks.types.make_type (RECORD_TYPE); | |
1154 | ctx->sfield_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
1155 | for (t = TYPE_FIELDS (ctx->record_type); t ; t = TREE_CHAIN (t)) | |
1156 | { | |
e60a6f7b | 1157 | sfield = build_decl (DECL_SOURCE_LOCATION (var), |
1158 | FIELD_DECL, DECL_NAME (t), TREE_TYPE (t)); | |
fd6481cf | 1159 | DECL_ABSTRACT_ORIGIN (sfield) = DECL_ABSTRACT_ORIGIN (t); |
1160 | insert_field_into_struct (ctx->srecord_type, sfield); | |
1161 | splay_tree_insert (ctx->sfield_map, | |
1162 | (splay_tree_key) DECL_ABSTRACT_ORIGIN (t), | |
1163 | (splay_tree_value) sfield); | |
1164 | } | |
1165 | } | |
1166 | sfield = field; | |
1167 | insert_field_into_struct ((mask & 1) ? ctx->record_type | |
1168 | : ctx->srecord_type, field); | |
1169 | } | |
1e8e9920 | 1170 | |
fd6481cf | 1171 | if (mask & 1) |
1172 | splay_tree_insert (ctx->field_map, (splay_tree_key) var, | |
1173 | (splay_tree_value) field); | |
1174 | if ((mask & 2) && ctx->sfield_map) | |
1175 | splay_tree_insert (ctx->sfield_map, (splay_tree_key) var, | |
1176 | (splay_tree_value) sfield); | |
1e8e9920 | 1177 | } |
1178 | ||
1179 | static tree | |
1180 | install_var_local (tree var, omp_context *ctx) | |
1181 | { | |
1182 | tree new_var = omp_copy_decl_1 (var, ctx); | |
1183 | insert_decl_map (&ctx->cb, var, new_var); | |
1184 | return new_var; | |
1185 | } | |
1186 | ||
1187 | /* Adjust the replacement for DECL in CTX for the new context. This means | |
1188 | copying the DECL_VALUE_EXPR, and fixing up the type. */ | |
1189 | ||
1190 | static void | |
1191 | fixup_remapped_decl (tree decl, omp_context *ctx, bool private_debug) | |
1192 | { | |
1193 | tree new_decl, size; | |
1194 | ||
1195 | new_decl = lookup_decl (decl, ctx); | |
1196 | ||
1197 | TREE_TYPE (new_decl) = remap_type (TREE_TYPE (decl), &ctx->cb); | |
1198 | ||
1199 | if ((!TREE_CONSTANT (DECL_SIZE (new_decl)) || private_debug) | |
1200 | && DECL_HAS_VALUE_EXPR_P (decl)) | |
1201 | { | |
1202 | tree ve = DECL_VALUE_EXPR (decl); | |
75a70cf9 | 1203 | walk_tree (&ve, copy_tree_body_r, &ctx->cb, NULL); |
1e8e9920 | 1204 | SET_DECL_VALUE_EXPR (new_decl, ve); |
1205 | DECL_HAS_VALUE_EXPR_P (new_decl) = 1; | |
1206 | } | |
1207 | ||
1208 | if (!TREE_CONSTANT (DECL_SIZE (new_decl))) | |
1209 | { | |
1210 | size = remap_decl (DECL_SIZE (decl), &ctx->cb); | |
1211 | if (size == error_mark_node) | |
1212 | size = TYPE_SIZE (TREE_TYPE (new_decl)); | |
1213 | DECL_SIZE (new_decl) = size; | |
1214 | ||
1215 | size = remap_decl (DECL_SIZE_UNIT (decl), &ctx->cb); | |
1216 | if (size == error_mark_node) | |
1217 | size = TYPE_SIZE_UNIT (TREE_TYPE (new_decl)); | |
1218 | DECL_SIZE_UNIT (new_decl) = size; | |
1219 | } | |
1220 | } | |
1221 | ||
1222 | /* The callback for remap_decl. Search all containing contexts for a | |
1223 | mapping of the variable; this avoids having to duplicate the splay | |
1224 | tree ahead of time. We know a mapping doesn't already exist in the | |
1225 | given context. Create new mappings to implement default semantics. */ | |
1226 | ||
1227 | static tree | |
1228 | omp_copy_decl (tree var, copy_body_data *cb) | |
1229 | { | |
1230 | omp_context *ctx = (omp_context *) cb; | |
1231 | tree new_var; | |
1232 | ||
1e8e9920 | 1233 | if (TREE_CODE (var) == LABEL_DECL) |
1234 | { | |
e60a6f7b | 1235 | new_var = create_artificial_label (DECL_SOURCE_LOCATION (var)); |
773c5ba7 | 1236 | DECL_CONTEXT (new_var) = current_function_decl; |
1e8e9920 | 1237 | insert_decl_map (&ctx->cb, var, new_var); |
1238 | return new_var; | |
1239 | } | |
1240 | ||
fd6481cf | 1241 | while (!is_taskreg_ctx (ctx)) |
1e8e9920 | 1242 | { |
1243 | ctx = ctx->outer; | |
1244 | if (ctx == NULL) | |
1245 | return var; | |
1246 | new_var = maybe_lookup_decl (var, ctx); | |
1247 | if (new_var) | |
1248 | return new_var; | |
1249 | } | |
1250 | ||
f49d7bb5 | 1251 | if (is_global_var (var) || decl_function_context (var) != ctx->cb.src_fn) |
1252 | return var; | |
1253 | ||
1e8e9920 | 1254 | return error_mark_node; |
1255 | } | |
1256 | ||
773c5ba7 | 1257 | |
773c5ba7 | 1258 | /* Debugging dumps for parallel regions. */ |
1259 | void dump_omp_region (FILE *, struct omp_region *, int); | |
1260 | void debug_omp_region (struct omp_region *); | |
1261 | void debug_all_omp_regions (void); | |
1262 | ||
1263 | /* Dump the parallel region tree rooted at REGION. */ | |
1264 | ||
1265 | void | |
1266 | dump_omp_region (FILE *file, struct omp_region *region, int indent) | |
1267 | { | |
61e47ac8 | 1268 | fprintf (file, "%*sbb %d: %s\n", indent, "", region->entry->index, |
75a70cf9 | 1269 | gimple_code_name[region->type]); |
773c5ba7 | 1270 | |
1271 | if (region->inner) | |
1272 | dump_omp_region (file, region->inner, indent + 4); | |
1273 | ||
61e47ac8 | 1274 | if (region->cont) |
1275 | { | |
75a70cf9 | 1276 | fprintf (file, "%*sbb %d: GIMPLE_OMP_CONTINUE\n", indent, "", |
61e47ac8 | 1277 | region->cont->index); |
1278 | } | |
48e1416a | 1279 | |
773c5ba7 | 1280 | if (region->exit) |
75a70cf9 | 1281 | fprintf (file, "%*sbb %d: GIMPLE_OMP_RETURN\n", indent, "", |
61e47ac8 | 1282 | region->exit->index); |
773c5ba7 | 1283 | else |
61e47ac8 | 1284 | fprintf (file, "%*s[no exit marker]\n", indent, ""); |
773c5ba7 | 1285 | |
1286 | if (region->next) | |
61e47ac8 | 1287 | dump_omp_region (file, region->next, indent); |
773c5ba7 | 1288 | } |
1289 | ||
4b987fac | 1290 | DEBUG_FUNCTION void |
773c5ba7 | 1291 | debug_omp_region (struct omp_region *region) |
1292 | { | |
1293 | dump_omp_region (stderr, region, 0); | |
1294 | } | |
1295 | ||
4b987fac | 1296 | DEBUG_FUNCTION void |
773c5ba7 | 1297 | debug_all_omp_regions (void) |
1298 | { | |
1299 | dump_omp_region (stderr, root_omp_region, 0); | |
1300 | } | |
1301 | ||
1302 | ||
1303 | /* Create a new parallel region starting at STMT inside region PARENT. */ | |
1304 | ||
7740abd8 | 1305 | static struct omp_region * |
75a70cf9 | 1306 | new_omp_region (basic_block bb, enum gimple_code type, |
1307 | struct omp_region *parent) | |
773c5ba7 | 1308 | { |
4077bf7a | 1309 | struct omp_region *region = XCNEW (struct omp_region); |
773c5ba7 | 1310 | |
1311 | region->outer = parent; | |
61e47ac8 | 1312 | region->entry = bb; |
1313 | region->type = type; | |
773c5ba7 | 1314 | |
1315 | if (parent) | |
1316 | { | |
1317 | /* This is a nested region. Add it to the list of inner | |
1318 | regions in PARENT. */ | |
1319 | region->next = parent->inner; | |
1320 | parent->inner = region; | |
1321 | } | |
61e47ac8 | 1322 | else |
773c5ba7 | 1323 | { |
1324 | /* This is a toplevel region. Add it to the list of toplevel | |
1325 | regions in ROOT_OMP_REGION. */ | |
1326 | region->next = root_omp_region; | |
1327 | root_omp_region = region; | |
1328 | } | |
61e47ac8 | 1329 | |
1330 | return region; | |
1331 | } | |
1332 | ||
1333 | /* Release the memory associated with the region tree rooted at REGION. */ | |
1334 | ||
1335 | static void | |
1336 | free_omp_region_1 (struct omp_region *region) | |
1337 | { | |
1338 | struct omp_region *i, *n; | |
1339 | ||
1340 | for (i = region->inner; i ; i = n) | |
773c5ba7 | 1341 | { |
61e47ac8 | 1342 | n = i->next; |
1343 | free_omp_region_1 (i); | |
773c5ba7 | 1344 | } |
1345 | ||
61e47ac8 | 1346 | free (region); |
1347 | } | |
773c5ba7 | 1348 | |
61e47ac8 | 1349 | /* Release the memory for the entire omp region tree. */ |
1350 | ||
1351 | void | |
1352 | free_omp_regions (void) | |
1353 | { | |
1354 | struct omp_region *r, *n; | |
1355 | for (r = root_omp_region; r ; r = n) | |
1356 | { | |
1357 | n = r->next; | |
1358 | free_omp_region_1 (r); | |
1359 | } | |
1360 | root_omp_region = NULL; | |
773c5ba7 | 1361 | } |
1362 | ||
1363 | ||
1e8e9920 | 1364 | /* Create a new context, with OUTER_CTX being the surrounding context. */ |
1365 | ||
1366 | static omp_context * | |
75a70cf9 | 1367 | new_omp_context (gimple stmt, omp_context *outer_ctx) |
1e8e9920 | 1368 | { |
1369 | omp_context *ctx = XCNEW (omp_context); | |
1370 | ||
1371 | splay_tree_insert (all_contexts, (splay_tree_key) stmt, | |
1372 | (splay_tree_value) ctx); | |
1373 | ctx->stmt = stmt; | |
1374 | ||
1375 | if (outer_ctx) | |
1376 | { | |
1377 | ctx->outer = outer_ctx; | |
1378 | ctx->cb = outer_ctx->cb; | |
1379 | ctx->cb.block = NULL; | |
1380 | ctx->depth = outer_ctx->depth + 1; | |
1381 | } | |
1382 | else | |
1383 | { | |
1384 | ctx->cb.src_fn = current_function_decl; | |
1385 | ctx->cb.dst_fn = current_function_decl; | |
415d1b9a | 1386 | ctx->cb.src_node = cgraph_node::get (current_function_decl); |
53f79206 | 1387 | gcc_checking_assert (ctx->cb.src_node); |
1e8e9920 | 1388 | ctx->cb.dst_node = ctx->cb.src_node; |
1389 | ctx->cb.src_cfun = cfun; | |
1390 | ctx->cb.copy_decl = omp_copy_decl; | |
e38def9c | 1391 | ctx->cb.eh_lp_nr = 0; |
1e8e9920 | 1392 | ctx->cb.transform_call_graph_edges = CB_CGE_MOVE; |
1393 | ctx->depth = 1; | |
1394 | } | |
1395 | ||
06ecf488 | 1396 | ctx->cb.decl_map = new hash_map<tree, tree>; |
1e8e9920 | 1397 | |
1398 | return ctx; | |
1399 | } | |
1400 | ||
75a70cf9 | 1401 | static gimple_seq maybe_catch_exception (gimple_seq); |
f6430caa | 1402 | |
1403 | /* Finalize task copyfn. */ | |
1404 | ||
1405 | static void | |
1a91d914 | 1406 | finalize_task_copyfn (gomp_task *task_stmt) |
f6430caa | 1407 | { |
1408 | struct function *child_cfun; | |
9078126c | 1409 | tree child_fn; |
e3a19533 | 1410 | gimple_seq seq = NULL, new_seq; |
1a91d914 | 1411 | gbind *bind; |
f6430caa | 1412 | |
75a70cf9 | 1413 | child_fn = gimple_omp_task_copy_fn (task_stmt); |
f6430caa | 1414 | if (child_fn == NULL_TREE) |
1415 | return; | |
1416 | ||
1417 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
82b40354 | 1418 | DECL_STRUCT_FUNCTION (child_fn)->curr_properties = cfun->curr_properties; |
f6430caa | 1419 | |
f6430caa | 1420 | push_cfun (child_cfun); |
7e3aae05 | 1421 | bind = gimplify_body (child_fn, false); |
75a70cf9 | 1422 | gimple_seq_add_stmt (&seq, bind); |
1423 | new_seq = maybe_catch_exception (seq); | |
1424 | if (new_seq != seq) | |
1425 | { | |
1426 | bind = gimple_build_bind (NULL, new_seq, NULL); | |
e3a19533 | 1427 | seq = NULL; |
75a70cf9 | 1428 | gimple_seq_add_stmt (&seq, bind); |
1429 | } | |
1430 | gimple_set_body (child_fn, seq); | |
f6430caa | 1431 | pop_cfun (); |
f6430caa | 1432 | |
82b40354 | 1433 | /* Inform the callgraph about the new function. */ |
415d1b9a | 1434 | cgraph_node::add_new_function (child_fn, false); |
f6430caa | 1435 | } |
1436 | ||
1e8e9920 | 1437 | /* Destroy a omp_context data structures. Called through the splay tree |
1438 | value delete callback. */ | |
1439 | ||
1440 | static void | |
1441 | delete_omp_context (splay_tree_value value) | |
1442 | { | |
1443 | omp_context *ctx = (omp_context *) value; | |
1444 | ||
06ecf488 | 1445 | delete ctx->cb.decl_map; |
1e8e9920 | 1446 | |
1447 | if (ctx->field_map) | |
1448 | splay_tree_delete (ctx->field_map); | |
fd6481cf | 1449 | if (ctx->sfield_map) |
1450 | splay_tree_delete (ctx->sfield_map); | |
1e8e9920 | 1451 | |
1452 | /* We hijacked DECL_ABSTRACT_ORIGIN earlier. We need to clear it before | |
1453 | it produces corrupt debug information. */ | |
1454 | if (ctx->record_type) | |
1455 | { | |
1456 | tree t; | |
1767a056 | 1457 | for (t = TYPE_FIELDS (ctx->record_type); t ; t = DECL_CHAIN (t)) |
1e8e9920 | 1458 | DECL_ABSTRACT_ORIGIN (t) = NULL; |
1459 | } | |
fd6481cf | 1460 | if (ctx->srecord_type) |
1461 | { | |
1462 | tree t; | |
1767a056 | 1463 | for (t = TYPE_FIELDS (ctx->srecord_type); t ; t = DECL_CHAIN (t)) |
fd6481cf | 1464 | DECL_ABSTRACT_ORIGIN (t) = NULL; |
1465 | } | |
1e8e9920 | 1466 | |
f6430caa | 1467 | if (is_task_ctx (ctx)) |
1a91d914 | 1468 | finalize_task_copyfn (as_a <gomp_task *> (ctx->stmt)); |
f6430caa | 1469 | |
1e8e9920 | 1470 | XDELETE (ctx); |
1471 | } | |
1472 | ||
1473 | /* Fix up RECEIVER_DECL with a type that has been remapped to the child | |
1474 | context. */ | |
1475 | ||
1476 | static void | |
1477 | fixup_child_record_type (omp_context *ctx) | |
1478 | { | |
1479 | tree f, type = ctx->record_type; | |
1480 | ||
1481 | /* ??? It isn't sufficient to just call remap_type here, because | |
1482 | variably_modified_type_p doesn't work the way we expect for | |
1483 | record types. Testing each field for whether it needs remapping | |
1484 | and creating a new record by hand works, however. */ | |
1767a056 | 1485 | for (f = TYPE_FIELDS (type); f ; f = DECL_CHAIN (f)) |
1e8e9920 | 1486 | if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn)) |
1487 | break; | |
1488 | if (f) | |
1489 | { | |
1490 | tree name, new_fields = NULL; | |
1491 | ||
1492 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
1493 | name = DECL_NAME (TYPE_NAME (ctx->record_type)); | |
e60a6f7b | 1494 | name = build_decl (DECL_SOURCE_LOCATION (ctx->receiver_decl), |
1495 | TYPE_DECL, name, type); | |
1e8e9920 | 1496 | TYPE_NAME (type) = name; |
1497 | ||
1767a056 | 1498 | for (f = TYPE_FIELDS (ctx->record_type); f ; f = DECL_CHAIN (f)) |
1e8e9920 | 1499 | { |
1500 | tree new_f = copy_node (f); | |
1501 | DECL_CONTEXT (new_f) = type; | |
1502 | TREE_TYPE (new_f) = remap_type (TREE_TYPE (f), &ctx->cb); | |
1767a056 | 1503 | DECL_CHAIN (new_f) = new_fields; |
75a70cf9 | 1504 | walk_tree (&DECL_SIZE (new_f), copy_tree_body_r, &ctx->cb, NULL); |
1505 | walk_tree (&DECL_SIZE_UNIT (new_f), copy_tree_body_r, | |
1506 | &ctx->cb, NULL); | |
1507 | walk_tree (&DECL_FIELD_OFFSET (new_f), copy_tree_body_r, | |
1508 | &ctx->cb, NULL); | |
1e8e9920 | 1509 | new_fields = new_f; |
1510 | ||
1511 | /* Arrange to be able to look up the receiver field | |
1512 | given the sender field. */ | |
1513 | splay_tree_insert (ctx->field_map, (splay_tree_key) f, | |
1514 | (splay_tree_value) new_f); | |
1515 | } | |
1516 | TYPE_FIELDS (type) = nreverse (new_fields); | |
1517 | layout_type (type); | |
1518 | } | |
1519 | ||
5455b100 | 1520 | TREE_TYPE (ctx->receiver_decl) |
1521 | = build_qualified_type (build_reference_type (type), TYPE_QUAL_RESTRICT); | |
1e8e9920 | 1522 | } |
1523 | ||
1524 | /* Instantiate decls as necessary in CTX to satisfy the data sharing | |
1525 | specified by CLAUSES. */ | |
1526 | ||
1527 | static void | |
1528 | scan_sharing_clauses (tree clauses, omp_context *ctx) | |
1529 | { | |
1530 | tree c, decl; | |
1531 | bool scan_array_reductions = false; | |
1532 | ||
1533 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
1534 | { | |
1535 | bool by_ref; | |
1536 | ||
55d6e7cd | 1537 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 1538 | { |
1539 | case OMP_CLAUSE_PRIVATE: | |
1540 | decl = OMP_CLAUSE_DECL (c); | |
fd6481cf | 1541 | if (OMP_CLAUSE_PRIVATE_OUTER_REF (c)) |
1542 | goto do_private; | |
1543 | else if (!is_variable_sized (decl)) | |
1e8e9920 | 1544 | install_var_local (decl, ctx); |
1545 | break; | |
1546 | ||
1547 | case OMP_CLAUSE_SHARED: | |
5fddcf34 | 1548 | decl = OMP_CLAUSE_DECL (c); |
bc7bff74 | 1549 | /* Ignore shared directives in teams construct. */ |
1550 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
5fddcf34 | 1551 | { |
1552 | /* Global variables don't need to be copied, | |
1553 | the receiver side will use them directly. */ | |
1554 | tree odecl = maybe_lookup_decl_in_outer_ctx (decl, ctx); | |
1555 | if (is_global_var (odecl)) | |
1556 | break; | |
1557 | insert_decl_map (&ctx->cb, decl, odecl); | |
1558 | break; | |
1559 | } | |
fd6481cf | 1560 | gcc_assert (is_taskreg_ctx (ctx)); |
e7327393 | 1561 | gcc_assert (!COMPLETE_TYPE_P (TREE_TYPE (decl)) |
1562 | || !is_variable_sized (decl)); | |
f49d7bb5 | 1563 | /* Global variables don't need to be copied, |
1564 | the receiver side will use them directly. */ | |
1565 | if (is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx))) | |
1566 | break; | |
fd6481cf | 1567 | by_ref = use_pointer_for_field (decl, ctx); |
1e8e9920 | 1568 | if (! TREE_READONLY (decl) |
1569 | || TREE_ADDRESSABLE (decl) | |
1570 | || by_ref | |
1571 | || is_reference (decl)) | |
1572 | { | |
fd6481cf | 1573 | install_var_field (decl, by_ref, 3, ctx); |
1e8e9920 | 1574 | install_var_local (decl, ctx); |
1575 | break; | |
1576 | } | |
1577 | /* We don't need to copy const scalar vars back. */ | |
55d6e7cd | 1578 | OMP_CLAUSE_SET_CODE (c, OMP_CLAUSE_FIRSTPRIVATE); |
1e8e9920 | 1579 | goto do_private; |
1580 | ||
1581 | case OMP_CLAUSE_LASTPRIVATE: | |
1582 | /* Let the corresponding firstprivate clause create | |
1583 | the variable. */ | |
1584 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) | |
1585 | break; | |
1586 | /* FALLTHRU */ | |
1587 | ||
1588 | case OMP_CLAUSE_FIRSTPRIVATE: | |
1589 | case OMP_CLAUSE_REDUCTION: | |
3d483a94 | 1590 | case OMP_CLAUSE_LINEAR: |
1e8e9920 | 1591 | decl = OMP_CLAUSE_DECL (c); |
1592 | do_private: | |
1593 | if (is_variable_sized (decl)) | |
1e8e9920 | 1594 | { |
fd6481cf | 1595 | if (is_task_ctx (ctx)) |
1596 | install_var_field (decl, false, 1, ctx); | |
1597 | break; | |
1598 | } | |
1599 | else if (is_taskreg_ctx (ctx)) | |
1600 | { | |
1601 | bool global | |
1602 | = is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)); | |
e8a588af | 1603 | by_ref = use_pointer_for_field (decl, NULL); |
fd6481cf | 1604 | |
1605 | if (is_task_ctx (ctx) | |
1606 | && (global || by_ref || is_reference (decl))) | |
1607 | { | |
1608 | install_var_field (decl, false, 1, ctx); | |
1609 | if (!global) | |
1610 | install_var_field (decl, by_ref, 2, ctx); | |
1611 | } | |
1612 | else if (!global) | |
1613 | install_var_field (decl, by_ref, 3, ctx); | |
1e8e9920 | 1614 | } |
1615 | install_var_local (decl, ctx); | |
1616 | break; | |
1617 | ||
bc7bff74 | 1618 | case OMP_CLAUSE__LOOPTEMP_: |
1619 | gcc_assert (is_parallel_ctx (ctx)); | |
1620 | decl = OMP_CLAUSE_DECL (c); | |
1621 | install_var_field (decl, false, 3, ctx); | |
1622 | install_var_local (decl, ctx); | |
1623 | break; | |
1624 | ||
1e8e9920 | 1625 | case OMP_CLAUSE_COPYPRIVATE: |
1e8e9920 | 1626 | case OMP_CLAUSE_COPYIN: |
1627 | decl = OMP_CLAUSE_DECL (c); | |
e8a588af | 1628 | by_ref = use_pointer_for_field (decl, NULL); |
fd6481cf | 1629 | install_var_field (decl, by_ref, 3, ctx); |
1e8e9920 | 1630 | break; |
1631 | ||
1632 | case OMP_CLAUSE_DEFAULT: | |
1633 | ctx->default_kind = OMP_CLAUSE_DEFAULT_KIND (c); | |
1634 | break; | |
1635 | ||
2169f33b | 1636 | case OMP_CLAUSE_FINAL: |
1e8e9920 | 1637 | case OMP_CLAUSE_IF: |
1638 | case OMP_CLAUSE_NUM_THREADS: | |
bc7bff74 | 1639 | case OMP_CLAUSE_NUM_TEAMS: |
1640 | case OMP_CLAUSE_THREAD_LIMIT: | |
1641 | case OMP_CLAUSE_DEVICE: | |
1e8e9920 | 1642 | case OMP_CLAUSE_SCHEDULE: |
bc7bff74 | 1643 | case OMP_CLAUSE_DIST_SCHEDULE: |
1644 | case OMP_CLAUSE_DEPEND: | |
40750995 | 1645 | case OMP_CLAUSE__CILK_FOR_COUNT_: |
1e8e9920 | 1646 | if (ctx->outer) |
75a70cf9 | 1647 | scan_omp_op (&OMP_CLAUSE_OPERAND (c, 0), ctx->outer); |
1e8e9920 | 1648 | break; |
1649 | ||
bc7bff74 | 1650 | case OMP_CLAUSE_TO: |
1651 | case OMP_CLAUSE_FROM: | |
1652 | case OMP_CLAUSE_MAP: | |
1653 | if (ctx->outer) | |
1654 | scan_omp_op (&OMP_CLAUSE_SIZE (c), ctx->outer); | |
1655 | decl = OMP_CLAUSE_DECL (c); | |
1656 | /* Global variables with "omp declare target" attribute | |
1657 | don't need to be copied, the receiver side will use them | |
1658 | directly. */ | |
1659 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
1660 | && DECL_P (decl) | |
1661 | && is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)) | |
b0c5e347 | 1662 | && varpool_node::get_create (decl)->offloadable) |
bc7bff74 | 1663 | break; |
1664 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
1665 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER) | |
1666 | { | |
1667 | /* Ignore OMP_CLAUSE_MAP_POINTER kind for arrays in | |
1668 | #pragma omp target data, there is nothing to map for | |
1669 | those. */ | |
1670 | if (gimple_omp_target_kind (ctx->stmt) == GF_OMP_TARGET_KIND_DATA | |
1671 | && !POINTER_TYPE_P (TREE_TYPE (decl))) | |
1672 | break; | |
1673 | } | |
1674 | if (DECL_P (decl)) | |
1675 | { | |
1676 | if (DECL_SIZE (decl) | |
1677 | && TREE_CODE (DECL_SIZE (decl)) != INTEGER_CST) | |
1678 | { | |
1679 | tree decl2 = DECL_VALUE_EXPR (decl); | |
1680 | gcc_assert (TREE_CODE (decl2) == INDIRECT_REF); | |
1681 | decl2 = TREE_OPERAND (decl2, 0); | |
1682 | gcc_assert (DECL_P (decl2)); | |
1683 | install_var_field (decl2, true, 3, ctx); | |
1684 | install_var_local (decl2, ctx); | |
1685 | install_var_local (decl, ctx); | |
1686 | } | |
1687 | else | |
1688 | { | |
1689 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
1690 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
1691 | && !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) | |
1692 | && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE) | |
1693 | install_var_field (decl, true, 7, ctx); | |
1694 | else | |
1695 | install_var_field (decl, true, 3, ctx); | |
1696 | if (gimple_omp_target_kind (ctx->stmt) | |
1697 | == GF_OMP_TARGET_KIND_REGION) | |
1698 | install_var_local (decl, ctx); | |
1699 | } | |
1700 | } | |
1701 | else | |
1702 | { | |
1703 | tree base = get_base_address (decl); | |
1704 | tree nc = OMP_CLAUSE_CHAIN (c); | |
1705 | if (DECL_P (base) | |
1706 | && nc != NULL_TREE | |
1707 | && OMP_CLAUSE_CODE (nc) == OMP_CLAUSE_MAP | |
1708 | && OMP_CLAUSE_DECL (nc) == base | |
1709 | && OMP_CLAUSE_MAP_KIND (nc) == OMP_CLAUSE_MAP_POINTER | |
1710 | && integer_zerop (OMP_CLAUSE_SIZE (nc))) | |
1711 | { | |
1712 | OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) = 1; | |
1713 | OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (nc) = 1; | |
1714 | } | |
1715 | else | |
1716 | { | |
691447ab | 1717 | if (ctx->outer) |
1718 | { | |
1719 | scan_omp_op (&OMP_CLAUSE_DECL (c), ctx->outer); | |
1720 | decl = OMP_CLAUSE_DECL (c); | |
1721 | } | |
bc7bff74 | 1722 | gcc_assert (!splay_tree_lookup (ctx->field_map, |
1723 | (splay_tree_key) decl)); | |
1724 | tree field | |
1725 | = build_decl (OMP_CLAUSE_LOCATION (c), | |
1726 | FIELD_DECL, NULL_TREE, ptr_type_node); | |
1727 | DECL_ALIGN (field) = TYPE_ALIGN (ptr_type_node); | |
1728 | insert_field_into_struct (ctx->record_type, field); | |
1729 | splay_tree_insert (ctx->field_map, (splay_tree_key) decl, | |
1730 | (splay_tree_value) field); | |
1731 | } | |
1732 | } | |
1733 | break; | |
1734 | ||
1e8e9920 | 1735 | case OMP_CLAUSE_NOWAIT: |
1736 | case OMP_CLAUSE_ORDERED: | |
fd6481cf | 1737 | case OMP_CLAUSE_COLLAPSE: |
1738 | case OMP_CLAUSE_UNTIED: | |
2169f33b | 1739 | case OMP_CLAUSE_MERGEABLE: |
bc7bff74 | 1740 | case OMP_CLAUSE_PROC_BIND: |
3d483a94 | 1741 | case OMP_CLAUSE_SAFELEN: |
1e8e9920 | 1742 | break; |
1743 | ||
bc7bff74 | 1744 | case OMP_CLAUSE_ALIGNED: |
1745 | decl = OMP_CLAUSE_DECL (c); | |
1746 | if (is_global_var (decl) | |
1747 | && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE) | |
1748 | install_var_local (decl, ctx); | |
1749 | break; | |
1750 | ||
1e8e9920 | 1751 | default: |
1752 | gcc_unreachable (); | |
1753 | } | |
1754 | } | |
1755 | ||
1756 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
1757 | { | |
55d6e7cd | 1758 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 1759 | { |
1760 | case OMP_CLAUSE_LASTPRIVATE: | |
1761 | /* Let the corresponding firstprivate clause create | |
1762 | the variable. */ | |
75a70cf9 | 1763 | if (OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)) |
fd6481cf | 1764 | scan_array_reductions = true; |
1e8e9920 | 1765 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) |
1766 | break; | |
1767 | /* FALLTHRU */ | |
1768 | ||
1769 | case OMP_CLAUSE_PRIVATE: | |
1770 | case OMP_CLAUSE_FIRSTPRIVATE: | |
1771 | case OMP_CLAUSE_REDUCTION: | |
3d483a94 | 1772 | case OMP_CLAUSE_LINEAR: |
1e8e9920 | 1773 | decl = OMP_CLAUSE_DECL (c); |
1774 | if (is_variable_sized (decl)) | |
1775 | install_var_local (decl, ctx); | |
1776 | fixup_remapped_decl (decl, ctx, | |
55d6e7cd | 1777 | OMP_CLAUSE_CODE (c) == OMP_CLAUSE_PRIVATE |
1e8e9920 | 1778 | && OMP_CLAUSE_PRIVATE_DEBUG (c)); |
55d6e7cd | 1779 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION |
1e8e9920 | 1780 | && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) |
1781 | scan_array_reductions = true; | |
2b536a17 | 1782 | else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
1783 | && OMP_CLAUSE_LINEAR_GIMPLE_SEQ (c)) | |
1784 | scan_array_reductions = true; | |
1e8e9920 | 1785 | break; |
1786 | ||
1787 | case OMP_CLAUSE_SHARED: | |
bc7bff74 | 1788 | /* Ignore shared directives in teams construct. */ |
1789 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
1790 | break; | |
1e8e9920 | 1791 | decl = OMP_CLAUSE_DECL (c); |
f49d7bb5 | 1792 | if (! is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx))) |
1793 | fixup_remapped_decl (decl, ctx, false); | |
1e8e9920 | 1794 | break; |
1795 | ||
bc7bff74 | 1796 | case OMP_CLAUSE_MAP: |
1797 | if (gimple_omp_target_kind (ctx->stmt) == GF_OMP_TARGET_KIND_DATA) | |
1798 | break; | |
1799 | decl = OMP_CLAUSE_DECL (c); | |
1800 | if (DECL_P (decl) | |
1801 | && is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)) | |
b0c5e347 | 1802 | && varpool_node::get_create (decl)->offloadable) |
bc7bff74 | 1803 | break; |
1804 | if (DECL_P (decl)) | |
1805 | { | |
1806 | if (OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
1807 | && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE | |
1808 | && !COMPLETE_TYPE_P (TREE_TYPE (decl))) | |
1809 | { | |
1810 | tree new_decl = lookup_decl (decl, ctx); | |
1811 | TREE_TYPE (new_decl) | |
1812 | = remap_type (TREE_TYPE (decl), &ctx->cb); | |
1813 | } | |
1814 | else if (DECL_SIZE (decl) | |
1815 | && TREE_CODE (DECL_SIZE (decl)) != INTEGER_CST) | |
1816 | { | |
1817 | tree decl2 = DECL_VALUE_EXPR (decl); | |
1818 | gcc_assert (TREE_CODE (decl2) == INDIRECT_REF); | |
1819 | decl2 = TREE_OPERAND (decl2, 0); | |
1820 | gcc_assert (DECL_P (decl2)); | |
1821 | fixup_remapped_decl (decl2, ctx, false); | |
1822 | fixup_remapped_decl (decl, ctx, true); | |
1823 | } | |
1824 | else | |
1825 | fixup_remapped_decl (decl, ctx, false); | |
1826 | } | |
1827 | break; | |
1828 | ||
1e8e9920 | 1829 | case OMP_CLAUSE_COPYPRIVATE: |
1830 | case OMP_CLAUSE_COPYIN: | |
1831 | case OMP_CLAUSE_DEFAULT: | |
1832 | case OMP_CLAUSE_IF: | |
1833 | case OMP_CLAUSE_NUM_THREADS: | |
bc7bff74 | 1834 | case OMP_CLAUSE_NUM_TEAMS: |
1835 | case OMP_CLAUSE_THREAD_LIMIT: | |
1836 | case OMP_CLAUSE_DEVICE: | |
1e8e9920 | 1837 | case OMP_CLAUSE_SCHEDULE: |
bc7bff74 | 1838 | case OMP_CLAUSE_DIST_SCHEDULE: |
1e8e9920 | 1839 | case OMP_CLAUSE_NOWAIT: |
1840 | case OMP_CLAUSE_ORDERED: | |
fd6481cf | 1841 | case OMP_CLAUSE_COLLAPSE: |
1842 | case OMP_CLAUSE_UNTIED: | |
2169f33b | 1843 | case OMP_CLAUSE_FINAL: |
1844 | case OMP_CLAUSE_MERGEABLE: | |
bc7bff74 | 1845 | case OMP_CLAUSE_PROC_BIND: |
3d483a94 | 1846 | case OMP_CLAUSE_SAFELEN: |
bc7bff74 | 1847 | case OMP_CLAUSE_ALIGNED: |
1848 | case OMP_CLAUSE_DEPEND: | |
1849 | case OMP_CLAUSE__LOOPTEMP_: | |
1850 | case OMP_CLAUSE_TO: | |
1851 | case OMP_CLAUSE_FROM: | |
40750995 | 1852 | case OMP_CLAUSE__CILK_FOR_COUNT_: |
1e8e9920 | 1853 | break; |
1854 | ||
1855 | default: | |
1856 | gcc_unreachable (); | |
1857 | } | |
1858 | } | |
1859 | ||
1860 | if (scan_array_reductions) | |
1861 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
55d6e7cd | 1862 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION |
1e8e9920 | 1863 | && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) |
1864 | { | |
ab129075 | 1865 | scan_omp (&OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c), ctx); |
1866 | scan_omp (&OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c), ctx); | |
1e8e9920 | 1867 | } |
fd6481cf | 1868 | else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE |
75a70cf9 | 1869 | && OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)) |
ab129075 | 1870 | scan_omp (&OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c), ctx); |
2b536a17 | 1871 | else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
1872 | && OMP_CLAUSE_LINEAR_GIMPLE_SEQ (c)) | |
1873 | scan_omp (&OMP_CLAUSE_LINEAR_GIMPLE_SEQ (c), ctx); | |
1e8e9920 | 1874 | } |
1875 | ||
40750995 | 1876 | /* Create a new name for omp child function. Returns an identifier. If |
1877 | IS_CILK_FOR is true then the suffix for the child function is | |
1878 | "_cilk_for_fn." */ | |
1e8e9920 | 1879 | |
1e8e9920 | 1880 | static tree |
40750995 | 1881 | create_omp_child_function_name (bool task_copy, bool is_cilk_for) |
1e8e9920 | 1882 | { |
40750995 | 1883 | if (is_cilk_for) |
1884 | return clone_function_name (current_function_decl, "_cilk_for_fn"); | |
1885 | return clone_function_name (current_function_decl, | |
1886 | task_copy ? "_omp_cpyfn" : "_omp_fn"); | |
1887 | } | |
1888 | ||
1889 | /* Returns the type of the induction variable for the child function for | |
1890 | _Cilk_for and the types for _high and _low variables based on TYPE. */ | |
1891 | ||
1892 | static tree | |
1893 | cilk_for_check_loop_diff_type (tree type) | |
1894 | { | |
1895 | if (TYPE_PRECISION (type) <= TYPE_PRECISION (uint32_type_node)) | |
1896 | { | |
1897 | if (TYPE_UNSIGNED (type)) | |
1898 | return uint32_type_node; | |
1899 | else | |
1900 | return integer_type_node; | |
1901 | } | |
1902 | else | |
1903 | { | |
1904 | if (TYPE_UNSIGNED (type)) | |
1905 | return uint64_type_node; | |
1906 | else | |
1907 | return long_long_integer_type_node; | |
1908 | } | |
1e8e9920 | 1909 | } |
1910 | ||
1911 | /* Build a decl for the omp child function. It'll not contain a body | |
1912 | yet, just the bare decl. */ | |
1913 | ||
1914 | static void | |
fd6481cf | 1915 | create_omp_child_function (omp_context *ctx, bool task_copy) |
1e8e9920 | 1916 | { |
1917 | tree decl, type, name, t; | |
1918 | ||
40750995 | 1919 | tree cilk_for_count |
1920 | = (flag_cilkplus && gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL) | |
1921 | ? find_omp_clause (gimple_omp_parallel_clauses (ctx->stmt), | |
1922 | OMP_CLAUSE__CILK_FOR_COUNT_) : NULL_TREE; | |
1923 | tree cilk_var_type = NULL_TREE; | |
1924 | ||
1925 | name = create_omp_child_function_name (task_copy, | |
1926 | cilk_for_count != NULL_TREE); | |
fd6481cf | 1927 | if (task_copy) |
1928 | type = build_function_type_list (void_type_node, ptr_type_node, | |
1929 | ptr_type_node, NULL_TREE); | |
40750995 | 1930 | else if (cilk_for_count) |
1931 | { | |
1932 | type = TREE_TYPE (OMP_CLAUSE_OPERAND (cilk_for_count, 0)); | |
1933 | cilk_var_type = cilk_for_check_loop_diff_type (type); | |
1934 | type = build_function_type_list (void_type_node, ptr_type_node, | |
1935 | cilk_var_type, cilk_var_type, NULL_TREE); | |
1936 | } | |
fd6481cf | 1937 | else |
1938 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
1e8e9920 | 1939 | |
40750995 | 1940 | decl = build_decl (gimple_location (ctx->stmt), FUNCTION_DECL, name, type); |
1e8e9920 | 1941 | |
fd6481cf | 1942 | if (!task_copy) |
1943 | ctx->cb.dst_fn = decl; | |
1944 | else | |
75a70cf9 | 1945 | gimple_omp_task_set_copy_fn (ctx->stmt, decl); |
1e8e9920 | 1946 | |
1947 | TREE_STATIC (decl) = 1; | |
1948 | TREE_USED (decl) = 1; | |
1949 | DECL_ARTIFICIAL (decl) = 1; | |
1950 | DECL_IGNORED_P (decl) = 0; | |
1951 | TREE_PUBLIC (decl) = 0; | |
1952 | DECL_UNINLINABLE (decl) = 1; | |
1953 | DECL_EXTERNAL (decl) = 0; | |
1954 | DECL_CONTEXT (decl) = NULL_TREE; | |
773c5ba7 | 1955 | DECL_INITIAL (decl) = make_node (BLOCK); |
b0c5e347 | 1956 | if (cgraph_node::get (current_function_decl)->offloadable) |
1957 | cgraph_node::get_create (decl)->offloadable = 1; | |
bc7bff74 | 1958 | else |
1959 | { | |
1960 | omp_context *octx; | |
1961 | for (octx = ctx; octx; octx = octx->outer) | |
b0c5e347 | 1962 | if (is_targetreg_ctx (octx)) |
bc7bff74 | 1963 | { |
b0c5e347 | 1964 | cgraph_node::get_create (decl)->offloadable = 1; |
7114ebdd | 1965 | #ifdef ENABLE_OFFLOADING |
b0c5e347 | 1966 | g->have_offload = true; |
7114ebdd | 1967 | #endif |
bc7bff74 | 1968 | break; |
1969 | } | |
1970 | } | |
1e8e9920 | 1971 | |
e60a6f7b | 1972 | t = build_decl (DECL_SOURCE_LOCATION (decl), |
1973 | RESULT_DECL, NULL_TREE, void_type_node); | |
1e8e9920 | 1974 | DECL_ARTIFICIAL (t) = 1; |
1975 | DECL_IGNORED_P (t) = 1; | |
8e5b4ed6 | 1976 | DECL_CONTEXT (t) = decl; |
1e8e9920 | 1977 | DECL_RESULT (decl) = t; |
1978 | ||
40750995 | 1979 | /* _Cilk_for's child function requires two extra parameters called |
1980 | __low and __high that are set the by Cilk runtime when it calls this | |
1981 | function. */ | |
1982 | if (cilk_for_count) | |
1983 | { | |
1984 | t = build_decl (DECL_SOURCE_LOCATION (decl), | |
1985 | PARM_DECL, get_identifier ("__high"), cilk_var_type); | |
1986 | DECL_ARTIFICIAL (t) = 1; | |
1987 | DECL_NAMELESS (t) = 1; | |
1988 | DECL_ARG_TYPE (t) = ptr_type_node; | |
1989 | DECL_CONTEXT (t) = current_function_decl; | |
1990 | TREE_USED (t) = 1; | |
1991 | DECL_CHAIN (t) = DECL_ARGUMENTS (decl); | |
1992 | DECL_ARGUMENTS (decl) = t; | |
1993 | ||
1994 | t = build_decl (DECL_SOURCE_LOCATION (decl), | |
1995 | PARM_DECL, get_identifier ("__low"), cilk_var_type); | |
1996 | DECL_ARTIFICIAL (t) = 1; | |
1997 | DECL_NAMELESS (t) = 1; | |
1998 | DECL_ARG_TYPE (t) = ptr_type_node; | |
1999 | DECL_CONTEXT (t) = current_function_decl; | |
2000 | TREE_USED (t) = 1; | |
2001 | DECL_CHAIN (t) = DECL_ARGUMENTS (decl); | |
2002 | DECL_ARGUMENTS (decl) = t; | |
2003 | } | |
2004 | ||
2005 | tree data_name = get_identifier (".omp_data_i"); | |
2006 | t = build_decl (DECL_SOURCE_LOCATION (decl), PARM_DECL, data_name, | |
2007 | ptr_type_node); | |
1e8e9920 | 2008 | DECL_ARTIFICIAL (t) = 1; |
84bfaaeb | 2009 | DECL_NAMELESS (t) = 1; |
1e8e9920 | 2010 | DECL_ARG_TYPE (t) = ptr_type_node; |
773c5ba7 | 2011 | DECL_CONTEXT (t) = current_function_decl; |
1e8e9920 | 2012 | TREE_USED (t) = 1; |
40750995 | 2013 | if (cilk_for_count) |
2014 | DECL_CHAIN (t) = DECL_ARGUMENTS (decl); | |
1e8e9920 | 2015 | DECL_ARGUMENTS (decl) = t; |
fd6481cf | 2016 | if (!task_copy) |
2017 | ctx->receiver_decl = t; | |
2018 | else | |
2019 | { | |
e60a6f7b | 2020 | t = build_decl (DECL_SOURCE_LOCATION (decl), |
2021 | PARM_DECL, get_identifier (".omp_data_o"), | |
fd6481cf | 2022 | ptr_type_node); |
2023 | DECL_ARTIFICIAL (t) = 1; | |
84bfaaeb | 2024 | DECL_NAMELESS (t) = 1; |
fd6481cf | 2025 | DECL_ARG_TYPE (t) = ptr_type_node; |
2026 | DECL_CONTEXT (t) = current_function_decl; | |
2027 | TREE_USED (t) = 1; | |
86f2ad37 | 2028 | TREE_ADDRESSABLE (t) = 1; |
1767a056 | 2029 | DECL_CHAIN (t) = DECL_ARGUMENTS (decl); |
fd6481cf | 2030 | DECL_ARGUMENTS (decl) = t; |
2031 | } | |
1e8e9920 | 2032 | |
48e1416a | 2033 | /* Allocate memory for the function structure. The call to |
773c5ba7 | 2034 | allocate_struct_function clobbers CFUN, so we need to restore |
1e8e9920 | 2035 | it afterward. */ |
87d4aa85 | 2036 | push_struct_function (decl); |
75a70cf9 | 2037 | cfun->function_end_locus = gimple_location (ctx->stmt); |
87d4aa85 | 2038 | pop_cfun (); |
1e8e9920 | 2039 | } |
2040 | ||
bc7bff74 | 2041 | /* Callback for walk_gimple_seq. Check if combined parallel |
2042 | contains gimple_omp_for_combined_into_p OMP_FOR. */ | |
2043 | ||
2044 | static tree | |
2045 | find_combined_for (gimple_stmt_iterator *gsi_p, | |
2046 | bool *handled_ops_p, | |
2047 | struct walk_stmt_info *wi) | |
2048 | { | |
2049 | gimple stmt = gsi_stmt (*gsi_p); | |
2050 | ||
2051 | *handled_ops_p = true; | |
2052 | switch (gimple_code (stmt)) | |
2053 | { | |
2054 | WALK_SUBSTMTS; | |
2055 | ||
2056 | case GIMPLE_OMP_FOR: | |
2057 | if (gimple_omp_for_combined_into_p (stmt) | |
2058 | && gimple_omp_for_kind (stmt) == GF_OMP_FOR_KIND_FOR) | |
2059 | { | |
2060 | wi->info = stmt; | |
2061 | return integer_zero_node; | |
2062 | } | |
2063 | break; | |
2064 | default: | |
2065 | break; | |
2066 | } | |
2067 | return NULL; | |
2068 | } | |
2069 | ||
1e8e9920 | 2070 | /* Scan an OpenMP parallel directive. */ |
2071 | ||
2072 | static void | |
75a70cf9 | 2073 | scan_omp_parallel (gimple_stmt_iterator *gsi, omp_context *outer_ctx) |
1e8e9920 | 2074 | { |
2075 | omp_context *ctx; | |
2076 | tree name; | |
1a91d914 | 2077 | gomp_parallel *stmt = as_a <gomp_parallel *> (gsi_stmt (*gsi)); |
1e8e9920 | 2078 | |
2079 | /* Ignore parallel directives with empty bodies, unless there | |
2080 | are copyin clauses. */ | |
2081 | if (optimize > 0 | |
75a70cf9 | 2082 | && empty_body_p (gimple_omp_body (stmt)) |
2083 | && find_omp_clause (gimple_omp_parallel_clauses (stmt), | |
2084 | OMP_CLAUSE_COPYIN) == NULL) | |
1e8e9920 | 2085 | { |
75a70cf9 | 2086 | gsi_replace (gsi, gimple_build_nop (), false); |
1e8e9920 | 2087 | return; |
2088 | } | |
2089 | ||
bc7bff74 | 2090 | if (gimple_omp_parallel_combined_p (stmt)) |
2091 | { | |
bc7bff74 | 2092 | struct walk_stmt_info wi; |
2093 | ||
2094 | memset (&wi, 0, sizeof (wi)); | |
2095 | wi.val_only = true; | |
2096 | walk_gimple_seq (gimple_omp_body (stmt), | |
2097 | find_combined_for, NULL, &wi); | |
1a91d914 | 2098 | if (wi.info) |
bc7bff74 | 2099 | { |
1a91d914 | 2100 | gomp_for *for_stmt = as_a <gomp_for *> ((gimple) wi.info); |
bc7bff74 | 2101 | struct omp_for_data fd; |
2102 | extract_omp_for_data (for_stmt, &fd, NULL); | |
2103 | /* We need two temporaries with fd.loop.v type (istart/iend) | |
2104 | and then (fd.collapse - 1) temporaries with the same | |
2105 | type for count2 ... countN-1 vars if not constant. */ | |
2106 | size_t count = 2, i; | |
2107 | tree type = fd.iter_type; | |
2108 | if (fd.collapse > 1 | |
2109 | && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
2110 | count += fd.collapse - 1; | |
2111 | for (i = 0; i < count; i++) | |
2112 | { | |
f9e245b2 | 2113 | tree temp = create_tmp_var (type); |
bc7bff74 | 2114 | tree c = build_omp_clause (UNKNOWN_LOCATION, |
2115 | OMP_CLAUSE__LOOPTEMP_); | |
691447ab | 2116 | insert_decl_map (&outer_ctx->cb, temp, temp); |
bc7bff74 | 2117 | OMP_CLAUSE_DECL (c) = temp; |
2118 | OMP_CLAUSE_CHAIN (c) = gimple_omp_parallel_clauses (stmt); | |
2119 | gimple_omp_parallel_set_clauses (stmt, c); | |
2120 | } | |
2121 | } | |
2122 | } | |
2123 | ||
75a70cf9 | 2124 | ctx = new_omp_context (stmt, outer_ctx); |
37eaded9 | 2125 | taskreg_contexts.safe_push (ctx); |
fd6481cf | 2126 | if (taskreg_nesting_level > 1) |
773c5ba7 | 2127 | ctx->is_nested = true; |
1e8e9920 | 2128 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); |
1e8e9920 | 2129 | ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED; |
2130 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
1e8e9920 | 2131 | name = create_tmp_var_name (".omp_data_s"); |
e60a6f7b | 2132 | name = build_decl (gimple_location (stmt), |
2133 | TYPE_DECL, name, ctx->record_type); | |
84bfaaeb | 2134 | DECL_ARTIFICIAL (name) = 1; |
2135 | DECL_NAMELESS (name) = 1; | |
1e8e9920 | 2136 | TYPE_NAME (ctx->record_type) = name; |
fd6481cf | 2137 | create_omp_child_function (ctx, false); |
75a70cf9 | 2138 | gimple_omp_parallel_set_child_fn (stmt, ctx->cb.dst_fn); |
1e8e9920 | 2139 | |
75a70cf9 | 2140 | scan_sharing_clauses (gimple_omp_parallel_clauses (stmt), ctx); |
ab129075 | 2141 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2142 | |
2143 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2144 | ctx->record_type = ctx->receiver_decl = NULL; | |
1e8e9920 | 2145 | } |
2146 | ||
fd6481cf | 2147 | /* Scan an OpenMP task directive. */ |
2148 | ||
2149 | static void | |
75a70cf9 | 2150 | scan_omp_task (gimple_stmt_iterator *gsi, omp_context *outer_ctx) |
fd6481cf | 2151 | { |
2152 | omp_context *ctx; | |
75a70cf9 | 2153 | tree name, t; |
1a91d914 | 2154 | gomp_task *stmt = as_a <gomp_task *> (gsi_stmt (*gsi)); |
fd6481cf | 2155 | |
2156 | /* Ignore task directives with empty bodies. */ | |
2157 | if (optimize > 0 | |
75a70cf9 | 2158 | && empty_body_p (gimple_omp_body (stmt))) |
fd6481cf | 2159 | { |
75a70cf9 | 2160 | gsi_replace (gsi, gimple_build_nop (), false); |
fd6481cf | 2161 | return; |
2162 | } | |
2163 | ||
75a70cf9 | 2164 | ctx = new_omp_context (stmt, outer_ctx); |
37eaded9 | 2165 | taskreg_contexts.safe_push (ctx); |
fd6481cf | 2166 | if (taskreg_nesting_level > 1) |
2167 | ctx->is_nested = true; | |
2168 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
2169 | ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED; | |
2170 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
2171 | name = create_tmp_var_name (".omp_data_s"); | |
e60a6f7b | 2172 | name = build_decl (gimple_location (stmt), |
2173 | TYPE_DECL, name, ctx->record_type); | |
84bfaaeb | 2174 | DECL_ARTIFICIAL (name) = 1; |
2175 | DECL_NAMELESS (name) = 1; | |
fd6481cf | 2176 | TYPE_NAME (ctx->record_type) = name; |
2177 | create_omp_child_function (ctx, false); | |
75a70cf9 | 2178 | gimple_omp_task_set_child_fn (stmt, ctx->cb.dst_fn); |
fd6481cf | 2179 | |
75a70cf9 | 2180 | scan_sharing_clauses (gimple_omp_task_clauses (stmt), ctx); |
fd6481cf | 2181 | |
2182 | if (ctx->srecord_type) | |
2183 | { | |
2184 | name = create_tmp_var_name (".omp_data_a"); | |
e60a6f7b | 2185 | name = build_decl (gimple_location (stmt), |
2186 | TYPE_DECL, name, ctx->srecord_type); | |
84bfaaeb | 2187 | DECL_ARTIFICIAL (name) = 1; |
2188 | DECL_NAMELESS (name) = 1; | |
fd6481cf | 2189 | TYPE_NAME (ctx->srecord_type) = name; |
2190 | create_omp_child_function (ctx, true); | |
2191 | } | |
2192 | ||
ab129075 | 2193 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
fd6481cf | 2194 | |
2195 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2196 | { | |
2197 | ctx->record_type = ctx->receiver_decl = NULL; | |
75a70cf9 | 2198 | t = build_int_cst (long_integer_type_node, 0); |
2199 | gimple_omp_task_set_arg_size (stmt, t); | |
2200 | t = build_int_cst (long_integer_type_node, 1); | |
2201 | gimple_omp_task_set_arg_align (stmt, t); | |
fd6481cf | 2202 | } |
37eaded9 | 2203 | } |
2204 | ||
2205 | ||
2206 | /* If any decls have been made addressable during scan_omp, | |
2207 | adjust their fields if needed, and layout record types | |
2208 | of parallel/task constructs. */ | |
2209 | ||
2210 | static void | |
2211 | finish_taskreg_scan (omp_context *ctx) | |
2212 | { | |
2213 | if (ctx->record_type == NULL_TREE) | |
2214 | return; | |
2215 | ||
2216 | /* If any task_shared_vars were needed, verify all | |
2217 | OMP_CLAUSE_SHARED clauses on GIMPLE_OMP_{PARALLEL,TASK} | |
2218 | statements if use_pointer_for_field hasn't changed | |
2219 | because of that. If it did, update field types now. */ | |
2220 | if (task_shared_vars) | |
2221 | { | |
2222 | tree c; | |
2223 | ||
2224 | for (c = gimple_omp_taskreg_clauses (ctx->stmt); | |
2225 | c; c = OMP_CLAUSE_CHAIN (c)) | |
2226 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_SHARED) | |
2227 | { | |
2228 | tree decl = OMP_CLAUSE_DECL (c); | |
2229 | ||
2230 | /* Global variables don't need to be copied, | |
2231 | the receiver side will use them directly. */ | |
2232 | if (is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx))) | |
2233 | continue; | |
2234 | if (!bitmap_bit_p (task_shared_vars, DECL_UID (decl)) | |
2235 | || !use_pointer_for_field (decl, ctx)) | |
2236 | continue; | |
2237 | tree field = lookup_field (decl, ctx); | |
2238 | if (TREE_CODE (TREE_TYPE (field)) == POINTER_TYPE | |
2239 | && TREE_TYPE (TREE_TYPE (field)) == TREE_TYPE (decl)) | |
2240 | continue; | |
2241 | TREE_TYPE (field) = build_pointer_type (TREE_TYPE (decl)); | |
2242 | TREE_THIS_VOLATILE (field) = 0; | |
2243 | DECL_USER_ALIGN (field) = 0; | |
2244 | DECL_ALIGN (field) = TYPE_ALIGN (TREE_TYPE (field)); | |
2245 | if (TYPE_ALIGN (ctx->record_type) < DECL_ALIGN (field)) | |
2246 | TYPE_ALIGN (ctx->record_type) = DECL_ALIGN (field); | |
2247 | if (ctx->srecord_type) | |
2248 | { | |
2249 | tree sfield = lookup_sfield (decl, ctx); | |
2250 | TREE_TYPE (sfield) = TREE_TYPE (field); | |
2251 | TREE_THIS_VOLATILE (sfield) = 0; | |
2252 | DECL_USER_ALIGN (sfield) = 0; | |
2253 | DECL_ALIGN (sfield) = DECL_ALIGN (field); | |
2254 | if (TYPE_ALIGN (ctx->srecord_type) < DECL_ALIGN (sfield)) | |
2255 | TYPE_ALIGN (ctx->srecord_type) = DECL_ALIGN (sfield); | |
2256 | } | |
2257 | } | |
2258 | } | |
2259 | ||
2260 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL) | |
2261 | { | |
2262 | layout_type (ctx->record_type); | |
2263 | fixup_child_record_type (ctx); | |
2264 | } | |
fd6481cf | 2265 | else |
2266 | { | |
37eaded9 | 2267 | location_t loc = gimple_location (ctx->stmt); |
fd6481cf | 2268 | tree *p, vla_fields = NULL_TREE, *q = &vla_fields; |
2269 | /* Move VLA fields to the end. */ | |
2270 | p = &TYPE_FIELDS (ctx->record_type); | |
2271 | while (*p) | |
2272 | if (!TYPE_SIZE_UNIT (TREE_TYPE (*p)) | |
2273 | || ! TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (*p)))) | |
2274 | { | |
2275 | *q = *p; | |
2276 | *p = TREE_CHAIN (*p); | |
2277 | TREE_CHAIN (*q) = NULL_TREE; | |
2278 | q = &TREE_CHAIN (*q); | |
2279 | } | |
2280 | else | |
1767a056 | 2281 | p = &DECL_CHAIN (*p); |
fd6481cf | 2282 | *p = vla_fields; |
2283 | layout_type (ctx->record_type); | |
2284 | fixup_child_record_type (ctx); | |
2285 | if (ctx->srecord_type) | |
2286 | layout_type (ctx->srecord_type); | |
37eaded9 | 2287 | tree t = fold_convert_loc (loc, long_integer_type_node, |
2288 | TYPE_SIZE_UNIT (ctx->record_type)); | |
2289 | gimple_omp_task_set_arg_size (ctx->stmt, t); | |
75a70cf9 | 2290 | t = build_int_cst (long_integer_type_node, |
fd6481cf | 2291 | TYPE_ALIGN_UNIT (ctx->record_type)); |
37eaded9 | 2292 | gimple_omp_task_set_arg_align (ctx->stmt, t); |
fd6481cf | 2293 | } |
2294 | } | |
2295 | ||
1e8e9920 | 2296 | |
773c5ba7 | 2297 | /* Scan an OpenMP loop directive. */ |
1e8e9920 | 2298 | |
2299 | static void | |
1a91d914 | 2300 | scan_omp_for (gomp_for *stmt, omp_context *outer_ctx) |
1e8e9920 | 2301 | { |
773c5ba7 | 2302 | omp_context *ctx; |
75a70cf9 | 2303 | size_t i; |
1e8e9920 | 2304 | |
773c5ba7 | 2305 | ctx = new_omp_context (stmt, outer_ctx); |
1e8e9920 | 2306 | |
75a70cf9 | 2307 | scan_sharing_clauses (gimple_omp_for_clauses (stmt), ctx); |
1e8e9920 | 2308 | |
ab129075 | 2309 | scan_omp (gimple_omp_for_pre_body_ptr (stmt), ctx); |
75a70cf9 | 2310 | for (i = 0; i < gimple_omp_for_collapse (stmt); i++) |
fd6481cf | 2311 | { |
75a70cf9 | 2312 | scan_omp_op (gimple_omp_for_index_ptr (stmt, i), ctx); |
2313 | scan_omp_op (gimple_omp_for_initial_ptr (stmt, i), ctx); | |
2314 | scan_omp_op (gimple_omp_for_final_ptr (stmt, i), ctx); | |
2315 | scan_omp_op (gimple_omp_for_incr_ptr (stmt, i), ctx); | |
fd6481cf | 2316 | } |
ab129075 | 2317 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2318 | } |
2319 | ||
2320 | /* Scan an OpenMP sections directive. */ | |
2321 | ||
2322 | static void | |
1a91d914 | 2323 | scan_omp_sections (gomp_sections *stmt, omp_context *outer_ctx) |
1e8e9920 | 2324 | { |
1e8e9920 | 2325 | omp_context *ctx; |
2326 | ||
2327 | ctx = new_omp_context (stmt, outer_ctx); | |
75a70cf9 | 2328 | scan_sharing_clauses (gimple_omp_sections_clauses (stmt), ctx); |
ab129075 | 2329 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2330 | } |
2331 | ||
2332 | /* Scan an OpenMP single directive. */ | |
2333 | ||
2334 | static void | |
1a91d914 | 2335 | scan_omp_single (gomp_single *stmt, omp_context *outer_ctx) |
1e8e9920 | 2336 | { |
1e8e9920 | 2337 | omp_context *ctx; |
2338 | tree name; | |
2339 | ||
2340 | ctx = new_omp_context (stmt, outer_ctx); | |
2341 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
2342 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
2343 | name = create_tmp_var_name (".omp_copy_s"); | |
e60a6f7b | 2344 | name = build_decl (gimple_location (stmt), |
2345 | TYPE_DECL, name, ctx->record_type); | |
1e8e9920 | 2346 | TYPE_NAME (ctx->record_type) = name; |
2347 | ||
75a70cf9 | 2348 | scan_sharing_clauses (gimple_omp_single_clauses (stmt), ctx); |
ab129075 | 2349 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2350 | |
2351 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2352 | ctx->record_type = NULL; | |
2353 | else | |
2354 | layout_type (ctx->record_type); | |
2355 | } | |
2356 | ||
bc7bff74 | 2357 | /* Scan an OpenMP target{, data, update} directive. */ |
2358 | ||
2359 | static void | |
1a91d914 | 2360 | scan_omp_target (gomp_target *stmt, omp_context *outer_ctx) |
bc7bff74 | 2361 | { |
2362 | omp_context *ctx; | |
2363 | tree name; | |
2364 | int kind = gimple_omp_target_kind (stmt); | |
2365 | ||
2366 | ctx = new_omp_context (stmt, outer_ctx); | |
2367 | ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
2368 | ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED; | |
2369 | ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE); | |
2370 | name = create_tmp_var_name (".omp_data_t"); | |
2371 | name = build_decl (gimple_location (stmt), | |
2372 | TYPE_DECL, name, ctx->record_type); | |
2373 | DECL_ARTIFICIAL (name) = 1; | |
2374 | DECL_NAMELESS (name) = 1; | |
2375 | TYPE_NAME (ctx->record_type) = name; | |
2376 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
2377 | { | |
2378 | create_omp_child_function (ctx, false); | |
2379 | gimple_omp_target_set_child_fn (stmt, ctx->cb.dst_fn); | |
2380 | } | |
2381 | ||
2382 | scan_sharing_clauses (gimple_omp_target_clauses (stmt), ctx); | |
2383 | scan_omp (gimple_omp_body_ptr (stmt), ctx); | |
2384 | ||
2385 | if (TYPE_FIELDS (ctx->record_type) == NULL) | |
2386 | ctx->record_type = ctx->receiver_decl = NULL; | |
2387 | else | |
2388 | { | |
2389 | TYPE_FIELDS (ctx->record_type) | |
2390 | = nreverse (TYPE_FIELDS (ctx->record_type)); | |
2391 | #ifdef ENABLE_CHECKING | |
2392 | tree field; | |
2393 | unsigned int align = DECL_ALIGN (TYPE_FIELDS (ctx->record_type)); | |
2394 | for (field = TYPE_FIELDS (ctx->record_type); | |
2395 | field; | |
2396 | field = DECL_CHAIN (field)) | |
2397 | gcc_assert (DECL_ALIGN (field) == align); | |
2398 | #endif | |
2399 | layout_type (ctx->record_type); | |
2400 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
2401 | fixup_child_record_type (ctx); | |
2402 | } | |
2403 | } | |
2404 | ||
2405 | /* Scan an OpenMP teams directive. */ | |
2406 | ||
2407 | static void | |
1a91d914 | 2408 | scan_omp_teams (gomp_teams *stmt, omp_context *outer_ctx) |
bc7bff74 | 2409 | { |
2410 | omp_context *ctx = new_omp_context (stmt, outer_ctx); | |
2411 | scan_sharing_clauses (gimple_omp_teams_clauses (stmt), ctx); | |
2412 | scan_omp (gimple_omp_body_ptr (stmt), ctx); | |
2413 | } | |
1e8e9920 | 2414 | |
c1d127dd | 2415 | /* Check OpenMP nesting restrictions. */ |
ab129075 | 2416 | static bool |
2417 | check_omp_nesting_restrictions (gimple stmt, omp_context *ctx) | |
c1d127dd | 2418 | { |
3d483a94 | 2419 | if (ctx != NULL) |
2420 | { | |
2421 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR | |
10c55644 | 2422 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_SIMD) |
3d483a94 | 2423 | { |
2424 | error_at (gimple_location (stmt), | |
2425 | "OpenMP constructs may not be nested inside simd region"); | |
2426 | return false; | |
2427 | } | |
bc7bff74 | 2428 | else if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) |
2429 | { | |
2430 | if ((gimple_code (stmt) != GIMPLE_OMP_FOR | |
2431 | || (gimple_omp_for_kind (stmt) | |
2432 | != GF_OMP_FOR_KIND_DISTRIBUTE)) | |
2433 | && gimple_code (stmt) != GIMPLE_OMP_PARALLEL) | |
2434 | { | |
2435 | error_at (gimple_location (stmt), | |
2436 | "only distribute or parallel constructs are allowed to " | |
2437 | "be closely nested inside teams construct"); | |
2438 | return false; | |
2439 | } | |
2440 | } | |
3d483a94 | 2441 | } |
75a70cf9 | 2442 | switch (gimple_code (stmt)) |
c1d127dd | 2443 | { |
75a70cf9 | 2444 | case GIMPLE_OMP_FOR: |
10c55644 | 2445 | if (gimple_omp_for_kind (stmt) & GF_OMP_FOR_SIMD) |
3d483a94 | 2446 | return true; |
bc7bff74 | 2447 | if (gimple_omp_for_kind (stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
2448 | { | |
2449 | if (ctx != NULL && gimple_code (ctx->stmt) != GIMPLE_OMP_TEAMS) | |
2450 | { | |
2451 | error_at (gimple_location (stmt), | |
2452 | "distribute construct must be closely nested inside " | |
2453 | "teams construct"); | |
2454 | return false; | |
2455 | } | |
2456 | return true; | |
2457 | } | |
2458 | /* FALLTHRU */ | |
2459 | case GIMPLE_CALL: | |
2460 | if (is_gimple_call (stmt) | |
2461 | && (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2462 | == BUILT_IN_GOMP_CANCEL | |
2463 | || DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2464 | == BUILT_IN_GOMP_CANCELLATION_POINT)) | |
2465 | { | |
2466 | const char *bad = NULL; | |
2467 | const char *kind = NULL; | |
2468 | if (ctx == NULL) | |
2469 | { | |
2470 | error_at (gimple_location (stmt), "orphaned %qs construct", | |
2471 | DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2472 | == BUILT_IN_GOMP_CANCEL | |
2473 | ? "#pragma omp cancel" | |
2474 | : "#pragma omp cancellation point"); | |
2475 | return false; | |
2476 | } | |
6b409616 | 2477 | switch (tree_fits_shwi_p (gimple_call_arg (stmt, 0)) |
2478 | ? tree_to_shwi (gimple_call_arg (stmt, 0)) | |
bc7bff74 | 2479 | : 0) |
2480 | { | |
2481 | case 1: | |
2482 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_PARALLEL) | |
2483 | bad = "#pragma omp parallel"; | |
2484 | else if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2485 | == BUILT_IN_GOMP_CANCEL | |
2486 | && !integer_zerop (gimple_call_arg (stmt, 1))) | |
2487 | ctx->cancellable = true; | |
2488 | kind = "parallel"; | |
2489 | break; | |
2490 | case 2: | |
2491 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_FOR | |
2492 | || gimple_omp_for_kind (ctx->stmt) != GF_OMP_FOR_KIND_FOR) | |
2493 | bad = "#pragma omp for"; | |
2494 | else if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2495 | == BUILT_IN_GOMP_CANCEL | |
2496 | && !integer_zerop (gimple_call_arg (stmt, 1))) | |
2497 | { | |
2498 | ctx->cancellable = true; | |
2499 | if (find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
2500 | OMP_CLAUSE_NOWAIT)) | |
2501 | warning_at (gimple_location (stmt), 0, | |
2502 | "%<#pragma omp cancel for%> inside " | |
2503 | "%<nowait%> for construct"); | |
2504 | if (find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
2505 | OMP_CLAUSE_ORDERED)) | |
2506 | warning_at (gimple_location (stmt), 0, | |
2507 | "%<#pragma omp cancel for%> inside " | |
2508 | "%<ordered%> for construct"); | |
2509 | } | |
2510 | kind = "for"; | |
2511 | break; | |
2512 | case 4: | |
2513 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_SECTIONS | |
2514 | && gimple_code (ctx->stmt) != GIMPLE_OMP_SECTION) | |
2515 | bad = "#pragma omp sections"; | |
2516 | else if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2517 | == BUILT_IN_GOMP_CANCEL | |
2518 | && !integer_zerop (gimple_call_arg (stmt, 1))) | |
2519 | { | |
2520 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_SECTIONS) | |
2521 | { | |
2522 | ctx->cancellable = true; | |
2523 | if (find_omp_clause (gimple_omp_sections_clauses | |
2524 | (ctx->stmt), | |
2525 | OMP_CLAUSE_NOWAIT)) | |
2526 | warning_at (gimple_location (stmt), 0, | |
2527 | "%<#pragma omp cancel sections%> inside " | |
2528 | "%<nowait%> sections construct"); | |
2529 | } | |
2530 | else | |
2531 | { | |
2532 | gcc_assert (ctx->outer | |
2533 | && gimple_code (ctx->outer->stmt) | |
2534 | == GIMPLE_OMP_SECTIONS); | |
2535 | ctx->outer->cancellable = true; | |
2536 | if (find_omp_clause (gimple_omp_sections_clauses | |
2537 | (ctx->outer->stmt), | |
2538 | OMP_CLAUSE_NOWAIT)) | |
2539 | warning_at (gimple_location (stmt), 0, | |
2540 | "%<#pragma omp cancel sections%> inside " | |
2541 | "%<nowait%> sections construct"); | |
2542 | } | |
2543 | } | |
2544 | kind = "sections"; | |
2545 | break; | |
2546 | case 8: | |
2547 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_TASK) | |
2548 | bad = "#pragma omp task"; | |
2549 | else | |
2550 | ctx->cancellable = true; | |
2551 | kind = "taskgroup"; | |
2552 | break; | |
2553 | default: | |
2554 | error_at (gimple_location (stmt), "invalid arguments"); | |
2555 | return false; | |
2556 | } | |
2557 | if (bad) | |
2558 | { | |
2559 | error_at (gimple_location (stmt), | |
2560 | "%<%s %s%> construct not closely nested inside of %qs", | |
2561 | DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) | |
2562 | == BUILT_IN_GOMP_CANCEL | |
2563 | ? "#pragma omp cancel" | |
2564 | : "#pragma omp cancellation point", kind, bad); | |
2565 | return false; | |
2566 | } | |
2567 | } | |
3d483a94 | 2568 | /* FALLTHRU */ |
75a70cf9 | 2569 | case GIMPLE_OMP_SECTIONS: |
2570 | case GIMPLE_OMP_SINGLE: | |
c1d127dd | 2571 | for (; ctx != NULL; ctx = ctx->outer) |
75a70cf9 | 2572 | switch (gimple_code (ctx->stmt)) |
c1d127dd | 2573 | { |
75a70cf9 | 2574 | case GIMPLE_OMP_FOR: |
2575 | case GIMPLE_OMP_SECTIONS: | |
2576 | case GIMPLE_OMP_SINGLE: | |
2577 | case GIMPLE_OMP_ORDERED: | |
2578 | case GIMPLE_OMP_MASTER: | |
2579 | case GIMPLE_OMP_TASK: | |
bc7bff74 | 2580 | case GIMPLE_OMP_CRITICAL: |
75a70cf9 | 2581 | if (is_gimple_call (stmt)) |
fd6481cf | 2582 | { |
bc7bff74 | 2583 | if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) |
2584 | != BUILT_IN_GOMP_BARRIER) | |
2585 | return true; | |
ab129075 | 2586 | error_at (gimple_location (stmt), |
2587 | "barrier region may not be closely nested inside " | |
2588 | "of work-sharing, critical, ordered, master or " | |
2589 | "explicit task region"); | |
2590 | return false; | |
fd6481cf | 2591 | } |
ab129075 | 2592 | error_at (gimple_location (stmt), |
2593 | "work-sharing region may not be closely nested inside " | |
2594 | "of work-sharing, critical, ordered, master or explicit " | |
2595 | "task region"); | |
2596 | return false; | |
75a70cf9 | 2597 | case GIMPLE_OMP_PARALLEL: |
ab129075 | 2598 | return true; |
c1d127dd | 2599 | default: |
2600 | break; | |
2601 | } | |
2602 | break; | |
75a70cf9 | 2603 | case GIMPLE_OMP_MASTER: |
c1d127dd | 2604 | for (; ctx != NULL; ctx = ctx->outer) |
75a70cf9 | 2605 | switch (gimple_code (ctx->stmt)) |
c1d127dd | 2606 | { |
75a70cf9 | 2607 | case GIMPLE_OMP_FOR: |
2608 | case GIMPLE_OMP_SECTIONS: | |
2609 | case GIMPLE_OMP_SINGLE: | |
2610 | case GIMPLE_OMP_TASK: | |
ab129075 | 2611 | error_at (gimple_location (stmt), |
2612 | "master region may not be closely nested inside " | |
2613 | "of work-sharing or explicit task region"); | |
2614 | return false; | |
75a70cf9 | 2615 | case GIMPLE_OMP_PARALLEL: |
ab129075 | 2616 | return true; |
c1d127dd | 2617 | default: |
2618 | break; | |
2619 | } | |
2620 | break; | |
75a70cf9 | 2621 | case GIMPLE_OMP_ORDERED: |
c1d127dd | 2622 | for (; ctx != NULL; ctx = ctx->outer) |
75a70cf9 | 2623 | switch (gimple_code (ctx->stmt)) |
c1d127dd | 2624 | { |
75a70cf9 | 2625 | case GIMPLE_OMP_CRITICAL: |
2626 | case GIMPLE_OMP_TASK: | |
ab129075 | 2627 | error_at (gimple_location (stmt), |
2628 | "ordered region may not be closely nested inside " | |
2629 | "of critical or explicit task region"); | |
2630 | return false; | |
75a70cf9 | 2631 | case GIMPLE_OMP_FOR: |
2632 | if (find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
c1d127dd | 2633 | OMP_CLAUSE_ORDERED) == NULL) |
ab129075 | 2634 | { |
2635 | error_at (gimple_location (stmt), | |
2636 | "ordered region must be closely nested inside " | |
c1d127dd | 2637 | "a loop region with an ordered clause"); |
ab129075 | 2638 | return false; |
2639 | } | |
2640 | return true; | |
75a70cf9 | 2641 | case GIMPLE_OMP_PARALLEL: |
bc7bff74 | 2642 | error_at (gimple_location (stmt), |
2643 | "ordered region must be closely nested inside " | |
2644 | "a loop region with an ordered clause"); | |
2645 | return false; | |
c1d127dd | 2646 | default: |
2647 | break; | |
2648 | } | |
2649 | break; | |
75a70cf9 | 2650 | case GIMPLE_OMP_CRITICAL: |
1a91d914 | 2651 | { |
2652 | tree this_stmt_name | |
2653 | = gimple_omp_critical_name (as_a <gomp_critical *> (stmt)); | |
2654 | for (; ctx != NULL; ctx = ctx->outer) | |
2655 | if (gomp_critical *other_crit | |
2656 | = dyn_cast <gomp_critical *> (ctx->stmt)) | |
2657 | if (this_stmt_name == gimple_omp_critical_name (other_crit)) | |
2658 | { | |
2659 | error_at (gimple_location (stmt), | |
2660 | "critical region may not be nested inside a critical " | |
2661 | "region with the same name"); | |
2662 | return false; | |
2663 | } | |
2664 | } | |
c1d127dd | 2665 | break; |
bc7bff74 | 2666 | case GIMPLE_OMP_TEAMS: |
2667 | if (ctx == NULL | |
2668 | || gimple_code (ctx->stmt) != GIMPLE_OMP_TARGET | |
2669 | || gimple_omp_target_kind (ctx->stmt) != GF_OMP_TARGET_KIND_REGION) | |
2670 | { | |
2671 | error_at (gimple_location (stmt), | |
2672 | "teams construct not closely nested inside of target " | |
2673 | "region"); | |
2674 | return false; | |
2675 | } | |
2676 | break; | |
691447ab | 2677 | case GIMPLE_OMP_TARGET: |
2678 | for (; ctx != NULL; ctx = ctx->outer) | |
b0c5e347 | 2679 | if (is_targetreg_ctx (ctx)) |
691447ab | 2680 | { |
2681 | const char *name; | |
2682 | switch (gimple_omp_target_kind (stmt)) | |
2683 | { | |
2684 | case GF_OMP_TARGET_KIND_REGION: name = "target"; break; | |
2685 | case GF_OMP_TARGET_KIND_DATA: name = "target data"; break; | |
2686 | case GF_OMP_TARGET_KIND_UPDATE: name = "target update"; break; | |
2687 | default: gcc_unreachable (); | |
2688 | } | |
2689 | warning_at (gimple_location (stmt), 0, | |
2690 | "%s construct inside of target region", name); | |
2691 | } | |
2692 | break; | |
c1d127dd | 2693 | default: |
2694 | break; | |
2695 | } | |
ab129075 | 2696 | return true; |
c1d127dd | 2697 | } |
2698 | ||
2699 | ||
75a70cf9 | 2700 | /* Helper function scan_omp. |
2701 | ||
2702 | Callback for walk_tree or operators in walk_gimple_stmt used to | |
2703 | scan for OpenMP directives in TP. */ | |
1e8e9920 | 2704 | |
2705 | static tree | |
75a70cf9 | 2706 | scan_omp_1_op (tree *tp, int *walk_subtrees, void *data) |
1e8e9920 | 2707 | { |
4077bf7a | 2708 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; |
2709 | omp_context *ctx = (omp_context *) wi->info; | |
1e8e9920 | 2710 | tree t = *tp; |
2711 | ||
75a70cf9 | 2712 | switch (TREE_CODE (t)) |
2713 | { | |
2714 | case VAR_DECL: | |
2715 | case PARM_DECL: | |
2716 | case LABEL_DECL: | |
2717 | case RESULT_DECL: | |
2718 | if (ctx) | |
2719 | *tp = remap_decl (t, &ctx->cb); | |
2720 | break; | |
2721 | ||
2722 | default: | |
2723 | if (ctx && TYPE_P (t)) | |
2724 | *tp = remap_type (t, &ctx->cb); | |
2725 | else if (!DECL_P (t)) | |
7cf869dd | 2726 | { |
2727 | *walk_subtrees = 1; | |
2728 | if (ctx) | |
182cf5a9 | 2729 | { |
2730 | tree tem = remap_type (TREE_TYPE (t), &ctx->cb); | |
2731 | if (tem != TREE_TYPE (t)) | |
2732 | { | |
2733 | if (TREE_CODE (t) == INTEGER_CST) | |
e913b5cd | 2734 | *tp = wide_int_to_tree (tem, t); |
182cf5a9 | 2735 | else |
2736 | TREE_TYPE (t) = tem; | |
2737 | } | |
2738 | } | |
7cf869dd | 2739 | } |
75a70cf9 | 2740 | break; |
2741 | } | |
2742 | ||
2743 | return NULL_TREE; | |
2744 | } | |
2745 | ||
f2697631 | 2746 | /* Return true if FNDECL is a setjmp or a longjmp. */ |
2747 | ||
2748 | static bool | |
2749 | setjmp_or_longjmp_p (const_tree fndecl) | |
2750 | { | |
2751 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL | |
2752 | && (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_SETJMP | |
2753 | || DECL_FUNCTION_CODE (fndecl) == BUILT_IN_LONGJMP)) | |
2754 | return true; | |
2755 | ||
2756 | tree declname = DECL_NAME (fndecl); | |
2757 | if (!declname) | |
2758 | return false; | |
2759 | const char *name = IDENTIFIER_POINTER (declname); | |
2760 | return !strcmp (name, "setjmp") || !strcmp (name, "longjmp"); | |
2761 | } | |
2762 | ||
75a70cf9 | 2763 | |
2764 | /* Helper function for scan_omp. | |
2765 | ||
2766 | Callback for walk_gimple_stmt used to scan for OpenMP directives in | |
2767 | the current statement in GSI. */ | |
2768 | ||
2769 | static tree | |
2770 | scan_omp_1_stmt (gimple_stmt_iterator *gsi, bool *handled_ops_p, | |
2771 | struct walk_stmt_info *wi) | |
2772 | { | |
2773 | gimple stmt = gsi_stmt (*gsi); | |
2774 | omp_context *ctx = (omp_context *) wi->info; | |
2775 | ||
2776 | if (gimple_has_location (stmt)) | |
2777 | input_location = gimple_location (stmt); | |
1e8e9920 | 2778 | |
c1d127dd | 2779 | /* Check the OpenMP nesting restrictions. */ |
bc7bff74 | 2780 | bool remove = false; |
2781 | if (is_gimple_omp (stmt)) | |
2782 | remove = !check_omp_nesting_restrictions (stmt, ctx); | |
2783 | else if (is_gimple_call (stmt)) | |
2784 | { | |
2785 | tree fndecl = gimple_call_fndecl (stmt); | |
f2697631 | 2786 | if (fndecl) |
2787 | { | |
2788 | if (setjmp_or_longjmp_p (fndecl) | |
2789 | && ctx | |
2790 | && gimple_code (ctx->stmt) == GIMPLE_OMP_FOR | |
10c55644 | 2791 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_SIMD) |
f2697631 | 2792 | { |
2793 | remove = true; | |
2794 | error_at (gimple_location (stmt), | |
2795 | "setjmp/longjmp inside simd construct"); | |
2796 | } | |
2797 | else if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
2798 | switch (DECL_FUNCTION_CODE (fndecl)) | |
2799 | { | |
2800 | case BUILT_IN_GOMP_BARRIER: | |
2801 | case BUILT_IN_GOMP_CANCEL: | |
2802 | case BUILT_IN_GOMP_CANCELLATION_POINT: | |
2803 | case BUILT_IN_GOMP_TASKYIELD: | |
2804 | case BUILT_IN_GOMP_TASKWAIT: | |
2805 | case BUILT_IN_GOMP_TASKGROUP_START: | |
2806 | case BUILT_IN_GOMP_TASKGROUP_END: | |
2807 | remove = !check_omp_nesting_restrictions (stmt, ctx); | |
2808 | break; | |
2809 | default: | |
2810 | break; | |
2811 | } | |
2812 | } | |
bc7bff74 | 2813 | } |
2814 | if (remove) | |
2815 | { | |
2816 | stmt = gimple_build_nop (); | |
2817 | gsi_replace (gsi, stmt, false); | |
fd6481cf | 2818 | } |
c1d127dd | 2819 | |
75a70cf9 | 2820 | *handled_ops_p = true; |
2821 | ||
2822 | switch (gimple_code (stmt)) | |
1e8e9920 | 2823 | { |
75a70cf9 | 2824 | case GIMPLE_OMP_PARALLEL: |
fd6481cf | 2825 | taskreg_nesting_level++; |
75a70cf9 | 2826 | scan_omp_parallel (gsi, ctx); |
fd6481cf | 2827 | taskreg_nesting_level--; |
2828 | break; | |
2829 | ||
75a70cf9 | 2830 | case GIMPLE_OMP_TASK: |
fd6481cf | 2831 | taskreg_nesting_level++; |
75a70cf9 | 2832 | scan_omp_task (gsi, ctx); |
fd6481cf | 2833 | taskreg_nesting_level--; |
1e8e9920 | 2834 | break; |
2835 | ||
75a70cf9 | 2836 | case GIMPLE_OMP_FOR: |
1a91d914 | 2837 | scan_omp_for (as_a <gomp_for *> (stmt), ctx); |
1e8e9920 | 2838 | break; |
2839 | ||
75a70cf9 | 2840 | case GIMPLE_OMP_SECTIONS: |
1a91d914 | 2841 | scan_omp_sections (as_a <gomp_sections *> (stmt), ctx); |
1e8e9920 | 2842 | break; |
2843 | ||
75a70cf9 | 2844 | case GIMPLE_OMP_SINGLE: |
1a91d914 | 2845 | scan_omp_single (as_a <gomp_single *> (stmt), ctx); |
1e8e9920 | 2846 | break; |
2847 | ||
75a70cf9 | 2848 | case GIMPLE_OMP_SECTION: |
2849 | case GIMPLE_OMP_MASTER: | |
bc7bff74 | 2850 | case GIMPLE_OMP_TASKGROUP: |
75a70cf9 | 2851 | case GIMPLE_OMP_ORDERED: |
2852 | case GIMPLE_OMP_CRITICAL: | |
2853 | ctx = new_omp_context (stmt, ctx); | |
ab129075 | 2854 | scan_omp (gimple_omp_body_ptr (stmt), ctx); |
1e8e9920 | 2855 | break; |
2856 | ||
bc7bff74 | 2857 | case GIMPLE_OMP_TARGET: |
1a91d914 | 2858 | scan_omp_target (as_a <gomp_target *> (stmt), ctx); |
bc7bff74 | 2859 | break; |
2860 | ||
2861 | case GIMPLE_OMP_TEAMS: | |
1a91d914 | 2862 | scan_omp_teams (as_a <gomp_teams *> (stmt), ctx); |
bc7bff74 | 2863 | break; |
2864 | ||
75a70cf9 | 2865 | case GIMPLE_BIND: |
1e8e9920 | 2866 | { |
2867 | tree var; | |
1e8e9920 | 2868 | |
75a70cf9 | 2869 | *handled_ops_p = false; |
2870 | if (ctx) | |
1a91d914 | 2871 | for (var = gimple_bind_vars (as_a <gbind *> (stmt)); |
2872 | var ; | |
2873 | var = DECL_CHAIN (var)) | |
75a70cf9 | 2874 | insert_decl_map (&ctx->cb, var, var); |
1e8e9920 | 2875 | } |
2876 | break; | |
1e8e9920 | 2877 | default: |
75a70cf9 | 2878 | *handled_ops_p = false; |
1e8e9920 | 2879 | break; |
2880 | } | |
2881 | ||
2882 | return NULL_TREE; | |
2883 | } | |
2884 | ||
2885 | ||
75a70cf9 | 2886 | /* Scan all the statements starting at the current statement. CTX |
2887 | contains context information about the OpenMP directives and | |
2888 | clauses found during the scan. */ | |
1e8e9920 | 2889 | |
2890 | static void | |
ab129075 | 2891 | scan_omp (gimple_seq *body_p, omp_context *ctx) |
1e8e9920 | 2892 | { |
2893 | location_t saved_location; | |
2894 | struct walk_stmt_info wi; | |
2895 | ||
2896 | memset (&wi, 0, sizeof (wi)); | |
1e8e9920 | 2897 | wi.info = ctx; |
1e8e9920 | 2898 | wi.want_locations = true; |
2899 | ||
2900 | saved_location = input_location; | |
ab129075 | 2901 | walk_gimple_seq_mod (body_p, scan_omp_1_stmt, scan_omp_1_op, &wi); |
1e8e9920 | 2902 | input_location = saved_location; |
2903 | } | |
2904 | \f | |
2905 | /* Re-gimplification and code generation routines. */ | |
2906 | ||
2907 | /* Build a call to GOMP_barrier. */ | |
2908 | ||
bc7bff74 | 2909 | static gimple |
2910 | build_omp_barrier (tree lhs) | |
2911 | { | |
2912 | tree fndecl = builtin_decl_explicit (lhs ? BUILT_IN_GOMP_BARRIER_CANCEL | |
2913 | : BUILT_IN_GOMP_BARRIER); | |
1a91d914 | 2914 | gcall *g = gimple_build_call (fndecl, 0); |
bc7bff74 | 2915 | if (lhs) |
2916 | gimple_call_set_lhs (g, lhs); | |
2917 | return g; | |
1e8e9920 | 2918 | } |
2919 | ||
2920 | /* If a context was created for STMT when it was scanned, return it. */ | |
2921 | ||
2922 | static omp_context * | |
75a70cf9 | 2923 | maybe_lookup_ctx (gimple stmt) |
1e8e9920 | 2924 | { |
2925 | splay_tree_node n; | |
2926 | n = splay_tree_lookup (all_contexts, (splay_tree_key) stmt); | |
2927 | return n ? (omp_context *) n->value : NULL; | |
2928 | } | |
2929 | ||
773c5ba7 | 2930 | |
2931 | /* Find the mapping for DECL in CTX or the immediately enclosing | |
2932 | context that has a mapping for DECL. | |
2933 | ||
2934 | If CTX is a nested parallel directive, we may have to use the decl | |
2935 | mappings created in CTX's parent context. Suppose that we have the | |
2936 | following parallel nesting (variable UIDs showed for clarity): | |
2937 | ||
2938 | iD.1562 = 0; | |
2939 | #omp parallel shared(iD.1562) -> outer parallel | |
2940 | iD.1562 = iD.1562 + 1; | |
2941 | ||
2942 | #omp parallel shared (iD.1562) -> inner parallel | |
2943 | iD.1562 = iD.1562 - 1; | |
2944 | ||
2945 | Each parallel structure will create a distinct .omp_data_s structure | |
2946 | for copying iD.1562 in/out of the directive: | |
2947 | ||
2948 | outer parallel .omp_data_s.1.i -> iD.1562 | |
2949 | inner parallel .omp_data_s.2.i -> iD.1562 | |
2950 | ||
2951 | A shared variable mapping will produce a copy-out operation before | |
2952 | the parallel directive and a copy-in operation after it. So, in | |
2953 | this case we would have: | |
2954 | ||
2955 | iD.1562 = 0; | |
2956 | .omp_data_o.1.i = iD.1562; | |
2957 | #omp parallel shared(iD.1562) -> outer parallel | |
2958 | .omp_data_i.1 = &.omp_data_o.1 | |
2959 | .omp_data_i.1->i = .omp_data_i.1->i + 1; | |
2960 | ||
2961 | .omp_data_o.2.i = iD.1562; -> ** | |
2962 | #omp parallel shared(iD.1562) -> inner parallel | |
2963 | .omp_data_i.2 = &.omp_data_o.2 | |
2964 | .omp_data_i.2->i = .omp_data_i.2->i - 1; | |
2965 | ||
2966 | ||
2967 | ** This is a problem. The symbol iD.1562 cannot be referenced | |
2968 | inside the body of the outer parallel region. But since we are | |
2969 | emitting this copy operation while expanding the inner parallel | |
2970 | directive, we need to access the CTX structure of the outer | |
2971 | parallel directive to get the correct mapping: | |
2972 | ||
2973 | .omp_data_o.2.i = .omp_data_i.1->i | |
2974 | ||
2975 | Since there may be other workshare or parallel directives enclosing | |
2976 | the parallel directive, it may be necessary to walk up the context | |
2977 | parent chain. This is not a problem in general because nested | |
2978 | parallelism happens only rarely. */ | |
2979 | ||
2980 | static tree | |
2981 | lookup_decl_in_outer_ctx (tree decl, omp_context *ctx) | |
2982 | { | |
2983 | tree t; | |
2984 | omp_context *up; | |
2985 | ||
773c5ba7 | 2986 | for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer) |
2987 | t = maybe_lookup_decl (decl, up); | |
2988 | ||
87b31375 | 2989 | gcc_assert (!ctx->is_nested || t || is_global_var (decl)); |
773c5ba7 | 2990 | |
c37594c7 | 2991 | return t ? t : decl; |
773c5ba7 | 2992 | } |
2993 | ||
2994 | ||
f49d7bb5 | 2995 | /* Similar to lookup_decl_in_outer_ctx, but return DECL if not found |
2996 | in outer contexts. */ | |
2997 | ||
2998 | static tree | |
2999 | maybe_lookup_decl_in_outer_ctx (tree decl, omp_context *ctx) | |
3000 | { | |
3001 | tree t = NULL; | |
3002 | omp_context *up; | |
3003 | ||
87b31375 | 3004 | for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer) |
3005 | t = maybe_lookup_decl (decl, up); | |
f49d7bb5 | 3006 | |
3007 | return t ? t : decl; | |
3008 | } | |
3009 | ||
3010 | ||
1e8e9920 | 3011 | /* Construct the initialization value for reduction CLAUSE. */ |
3012 | ||
3013 | tree | |
3014 | omp_reduction_init (tree clause, tree type) | |
3015 | { | |
389dd41b | 3016 | location_t loc = OMP_CLAUSE_LOCATION (clause); |
1e8e9920 | 3017 | switch (OMP_CLAUSE_REDUCTION_CODE (clause)) |
3018 | { | |
3019 | case PLUS_EXPR: | |
3020 | case MINUS_EXPR: | |
3021 | case BIT_IOR_EXPR: | |
3022 | case BIT_XOR_EXPR: | |
3023 | case TRUTH_OR_EXPR: | |
3024 | case TRUTH_ORIF_EXPR: | |
3025 | case TRUTH_XOR_EXPR: | |
3026 | case NE_EXPR: | |
385f3f36 | 3027 | return build_zero_cst (type); |
1e8e9920 | 3028 | |
3029 | case MULT_EXPR: | |
3030 | case TRUTH_AND_EXPR: | |
3031 | case TRUTH_ANDIF_EXPR: | |
3032 | case EQ_EXPR: | |
389dd41b | 3033 | return fold_convert_loc (loc, type, integer_one_node); |
1e8e9920 | 3034 | |
3035 | case BIT_AND_EXPR: | |
389dd41b | 3036 | return fold_convert_loc (loc, type, integer_minus_one_node); |
1e8e9920 | 3037 | |
3038 | case MAX_EXPR: | |
3039 | if (SCALAR_FLOAT_TYPE_P (type)) | |
3040 | { | |
3041 | REAL_VALUE_TYPE max, min; | |
3042 | if (HONOR_INFINITIES (TYPE_MODE (type))) | |
3043 | { | |
3044 | real_inf (&max); | |
3045 | real_arithmetic (&min, NEGATE_EXPR, &max, NULL); | |
3046 | } | |
3047 | else | |
3048 | real_maxval (&min, 1, TYPE_MODE (type)); | |
3049 | return build_real (type, min); | |
3050 | } | |
3051 | else | |
3052 | { | |
3053 | gcc_assert (INTEGRAL_TYPE_P (type)); | |
3054 | return TYPE_MIN_VALUE (type); | |
3055 | } | |
3056 | ||
3057 | case MIN_EXPR: | |
3058 | if (SCALAR_FLOAT_TYPE_P (type)) | |
3059 | { | |
3060 | REAL_VALUE_TYPE max; | |
3061 | if (HONOR_INFINITIES (TYPE_MODE (type))) | |
3062 | real_inf (&max); | |
3063 | else | |
3064 | real_maxval (&max, 0, TYPE_MODE (type)); | |
3065 | return build_real (type, max); | |
3066 | } | |
3067 | else | |
3068 | { | |
3069 | gcc_assert (INTEGRAL_TYPE_P (type)); | |
3070 | return TYPE_MAX_VALUE (type); | |
3071 | } | |
3072 | ||
3073 | default: | |
3074 | gcc_unreachable (); | |
3075 | } | |
3076 | } | |
3077 | ||
bc7bff74 | 3078 | /* Return alignment to be assumed for var in CLAUSE, which should be |
3079 | OMP_CLAUSE_ALIGNED. */ | |
3080 | ||
3081 | static tree | |
3082 | omp_clause_aligned_alignment (tree clause) | |
3083 | { | |
3084 | if (OMP_CLAUSE_ALIGNED_ALIGNMENT (clause)) | |
3085 | return OMP_CLAUSE_ALIGNED_ALIGNMENT (clause); | |
3086 | ||
3087 | /* Otherwise return implementation defined alignment. */ | |
3088 | unsigned int al = 1; | |
3754d046 | 3089 | machine_mode mode, vmode; |
bc7bff74 | 3090 | int vs = targetm.vectorize.autovectorize_vector_sizes (); |
3091 | if (vs) | |
3092 | vs = 1 << floor_log2 (vs); | |
3093 | static enum mode_class classes[] | |
3094 | = { MODE_INT, MODE_VECTOR_INT, MODE_FLOAT, MODE_VECTOR_FLOAT }; | |
3095 | for (int i = 0; i < 4; i += 2) | |
3096 | for (mode = GET_CLASS_NARROWEST_MODE (classes[i]); | |
3097 | mode != VOIDmode; | |
3098 | mode = GET_MODE_WIDER_MODE (mode)) | |
3099 | { | |
3100 | vmode = targetm.vectorize.preferred_simd_mode (mode); | |
3101 | if (GET_MODE_CLASS (vmode) != classes[i + 1]) | |
3102 | continue; | |
3103 | while (vs | |
3104 | && GET_MODE_SIZE (vmode) < vs | |
3105 | && GET_MODE_2XWIDER_MODE (vmode) != VOIDmode) | |
3106 | vmode = GET_MODE_2XWIDER_MODE (vmode); | |
3107 | ||
3108 | tree type = lang_hooks.types.type_for_mode (mode, 1); | |
3109 | if (type == NULL_TREE || TYPE_MODE (type) != mode) | |
3110 | continue; | |
3111 | type = build_vector_type (type, GET_MODE_SIZE (vmode) | |
3112 | / GET_MODE_SIZE (mode)); | |
3113 | if (TYPE_MODE (type) != vmode) | |
3114 | continue; | |
3115 | if (TYPE_ALIGN_UNIT (type) > al) | |
3116 | al = TYPE_ALIGN_UNIT (type); | |
3117 | } | |
3118 | return build_int_cst (integer_type_node, al); | |
3119 | } | |
3120 | ||
3d483a94 | 3121 | /* Return maximum possible vectorization factor for the target. */ |
3122 | ||
3123 | static int | |
3124 | omp_max_vf (void) | |
3125 | { | |
3126 | if (!optimize | |
3127 | || optimize_debug | |
ad45e43e | 3128 | || !flag_tree_loop_optimize |
043115ec | 3129 | || (!flag_tree_loop_vectorize |
3130 | && (global_options_set.x_flag_tree_loop_vectorize | |
3131 | || global_options_set.x_flag_tree_vectorize))) | |
3d483a94 | 3132 | return 1; |
3133 | ||
3134 | int vs = targetm.vectorize.autovectorize_vector_sizes (); | |
3135 | if (vs) | |
3136 | { | |
3137 | vs = 1 << floor_log2 (vs); | |
3138 | return vs; | |
3139 | } | |
3754d046 | 3140 | machine_mode vqimode = targetm.vectorize.preferred_simd_mode (QImode); |
3d483a94 | 3141 | if (GET_MODE_CLASS (vqimode) == MODE_VECTOR_INT) |
3142 | return GET_MODE_NUNITS (vqimode); | |
3143 | return 1; | |
3144 | } | |
3145 | ||
3146 | /* Helper function of lower_rec_input_clauses, used for #pragma omp simd | |
3147 | privatization. */ | |
3148 | ||
3149 | static bool | |
3150 | lower_rec_simd_input_clauses (tree new_var, omp_context *ctx, int &max_vf, | |
3151 | tree &idx, tree &lane, tree &ivar, tree &lvar) | |
3152 | { | |
3153 | if (max_vf == 0) | |
3154 | { | |
3155 | max_vf = omp_max_vf (); | |
3156 | if (max_vf > 1) | |
3157 | { | |
3158 | tree c = find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
3159 | OMP_CLAUSE_SAFELEN); | |
c3f3b68d | 3160 | if (c && TREE_CODE (OMP_CLAUSE_SAFELEN_EXPR (c)) != INTEGER_CST) |
3161 | max_vf = 1; | |
3162 | else if (c && compare_tree_int (OMP_CLAUSE_SAFELEN_EXPR (c), | |
3163 | max_vf) == -1) | |
d85a2013 | 3164 | max_vf = tree_to_shwi (OMP_CLAUSE_SAFELEN_EXPR (c)); |
3d483a94 | 3165 | } |
3166 | if (max_vf > 1) | |
3167 | { | |
f9e245b2 | 3168 | idx = create_tmp_var (unsigned_type_node); |
3169 | lane = create_tmp_var (unsigned_type_node); | |
3d483a94 | 3170 | } |
3171 | } | |
3172 | if (max_vf == 1) | |
3173 | return false; | |
3174 | ||
3175 | tree atype = build_array_type_nelts (TREE_TYPE (new_var), max_vf); | |
f9e245b2 | 3176 | tree avar = create_tmp_var_raw (atype); |
3d483a94 | 3177 | if (TREE_ADDRESSABLE (new_var)) |
3178 | TREE_ADDRESSABLE (avar) = 1; | |
3179 | DECL_ATTRIBUTES (avar) | |
3180 | = tree_cons (get_identifier ("omp simd array"), NULL, | |
3181 | DECL_ATTRIBUTES (avar)); | |
3182 | gimple_add_tmp_var (avar); | |
3183 | ivar = build4 (ARRAY_REF, TREE_TYPE (new_var), avar, idx, | |
3184 | NULL_TREE, NULL_TREE); | |
3185 | lvar = build4 (ARRAY_REF, TREE_TYPE (new_var), avar, lane, | |
3186 | NULL_TREE, NULL_TREE); | |
bc7bff74 | 3187 | if (DECL_P (new_var)) |
3188 | { | |
3189 | SET_DECL_VALUE_EXPR (new_var, lvar); | |
3190 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
3191 | } | |
3d483a94 | 3192 | return true; |
3193 | } | |
3194 | ||
2712b6de | 3195 | /* Helper function of lower_rec_input_clauses. For a reference |
3196 | in simd reduction, add an underlying variable it will reference. */ | |
3197 | ||
3198 | static void | |
3199 | handle_simd_reference (location_t loc, tree new_vard, gimple_seq *ilist) | |
3200 | { | |
3201 | tree z = TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (new_vard))); | |
3202 | if (TREE_CONSTANT (z)) | |
3203 | { | |
3204 | const char *name = NULL; | |
3205 | if (DECL_NAME (new_vard)) | |
3206 | name = IDENTIFIER_POINTER (DECL_NAME (new_vard)); | |
3207 | ||
3208 | z = create_tmp_var_raw (TREE_TYPE (TREE_TYPE (new_vard)), name); | |
3209 | gimple_add_tmp_var (z); | |
3210 | TREE_ADDRESSABLE (z) = 1; | |
3211 | z = build_fold_addr_expr_loc (loc, z); | |
3212 | gimplify_assign (new_vard, z, ilist); | |
3213 | } | |
3214 | } | |
3215 | ||
1e8e9920 | 3216 | /* Generate code to implement the input clauses, FIRSTPRIVATE and COPYIN, |
3217 | from the receiver (aka child) side and initializers for REFERENCE_TYPE | |
3218 | private variables. Initialization statements go in ILIST, while calls | |
3219 | to destructors go in DLIST. */ | |
3220 | ||
3221 | static void | |
75a70cf9 | 3222 | lower_rec_input_clauses (tree clauses, gimple_seq *ilist, gimple_seq *dlist, |
bc7bff74 | 3223 | omp_context *ctx, struct omp_for_data *fd) |
1e8e9920 | 3224 | { |
c2f47e15 | 3225 | tree c, dtor, copyin_seq, x, ptr; |
1e8e9920 | 3226 | bool copyin_by_ref = false; |
f49d7bb5 | 3227 | bool lastprivate_firstprivate = false; |
bc7bff74 | 3228 | bool reduction_omp_orig_ref = false; |
1e8e9920 | 3229 | int pass; |
3d483a94 | 3230 | bool is_simd = (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR |
10c55644 | 3231 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_SIMD); |
3d483a94 | 3232 | int max_vf = 0; |
3233 | tree lane = NULL_TREE, idx = NULL_TREE; | |
3234 | tree ivar = NULL_TREE, lvar = NULL_TREE; | |
3235 | gimple_seq llist[2] = { NULL, NULL }; | |
1e8e9920 | 3236 | |
1e8e9920 | 3237 | copyin_seq = NULL; |
3238 | ||
3d483a94 | 3239 | /* Set max_vf=1 (which will later enforce safelen=1) in simd loops |
3240 | with data sharing clauses referencing variable sized vars. That | |
3241 | is unnecessarily hard to support and very unlikely to result in | |
3242 | vectorized code anyway. */ | |
3243 | if (is_simd) | |
3244 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
3245 | switch (OMP_CLAUSE_CODE (c)) | |
3246 | { | |
9580cb79 | 3247 | case OMP_CLAUSE_LINEAR: |
3248 | if (OMP_CLAUSE_LINEAR_ARRAY (c)) | |
3249 | max_vf = 1; | |
3250 | /* FALLTHRU */ | |
3d483a94 | 3251 | case OMP_CLAUSE_REDUCTION: |
3d483a94 | 3252 | case OMP_CLAUSE_PRIVATE: |
3253 | case OMP_CLAUSE_FIRSTPRIVATE: | |
3254 | case OMP_CLAUSE_LASTPRIVATE: | |
3d483a94 | 3255 | if (is_variable_sized (OMP_CLAUSE_DECL (c))) |
3256 | max_vf = 1; | |
3257 | break; | |
3258 | default: | |
3259 | continue; | |
3260 | } | |
3261 | ||
1e8e9920 | 3262 | /* Do all the fixed sized types in the first pass, and the variable sized |
3263 | types in the second pass. This makes sure that the scalar arguments to | |
48e1416a | 3264 | the variable sized types are processed before we use them in the |
1e8e9920 | 3265 | variable sized operations. */ |
3266 | for (pass = 0; pass < 2; ++pass) | |
3267 | { | |
3268 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
3269 | { | |
55d6e7cd | 3270 | enum omp_clause_code c_kind = OMP_CLAUSE_CODE (c); |
1e8e9920 | 3271 | tree var, new_var; |
3272 | bool by_ref; | |
389dd41b | 3273 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 3274 | |
3275 | switch (c_kind) | |
3276 | { | |
3277 | case OMP_CLAUSE_PRIVATE: | |
3278 | if (OMP_CLAUSE_PRIVATE_DEBUG (c)) | |
3279 | continue; | |
3280 | break; | |
3281 | case OMP_CLAUSE_SHARED: | |
bc7bff74 | 3282 | /* Ignore shared directives in teams construct. */ |
3283 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
3284 | continue; | |
f49d7bb5 | 3285 | if (maybe_lookup_decl (OMP_CLAUSE_DECL (c), ctx) == NULL) |
3286 | { | |
3287 | gcc_assert (is_global_var (OMP_CLAUSE_DECL (c))); | |
3288 | continue; | |
3289 | } | |
1e8e9920 | 3290 | case OMP_CLAUSE_FIRSTPRIVATE: |
1e8e9920 | 3291 | case OMP_CLAUSE_COPYIN: |
bc7bff74 | 3292 | case OMP_CLAUSE_LINEAR: |
3293 | break; | |
1e8e9920 | 3294 | case OMP_CLAUSE_REDUCTION: |
bc7bff74 | 3295 | if (OMP_CLAUSE_REDUCTION_OMP_ORIG_REF (c)) |
3296 | reduction_omp_orig_ref = true; | |
1e8e9920 | 3297 | break; |
bc7bff74 | 3298 | case OMP_CLAUSE__LOOPTEMP_: |
3299 | /* Handle _looptemp_ clauses only on parallel. */ | |
3300 | if (fd) | |
3301 | continue; | |
3d483a94 | 3302 | break; |
df2c34fc | 3303 | case OMP_CLAUSE_LASTPRIVATE: |
f49d7bb5 | 3304 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) |
3305 | { | |
3306 | lastprivate_firstprivate = true; | |
3307 | if (pass != 0) | |
3308 | continue; | |
3309 | } | |
cf5f881f | 3310 | /* Even without corresponding firstprivate, if |
3311 | decl is Fortran allocatable, it needs outer var | |
3312 | reference. */ | |
3313 | else if (pass == 0 | |
3314 | && lang_hooks.decls.omp_private_outer_ref | |
3315 | (OMP_CLAUSE_DECL (c))) | |
3316 | lastprivate_firstprivate = true; | |
df2c34fc | 3317 | break; |
bc7bff74 | 3318 | case OMP_CLAUSE_ALIGNED: |
3319 | if (pass == 0) | |
3320 | continue; | |
3321 | var = OMP_CLAUSE_DECL (c); | |
3322 | if (TREE_CODE (TREE_TYPE (var)) == POINTER_TYPE | |
3323 | && !is_global_var (var)) | |
3324 | { | |
3325 | new_var = maybe_lookup_decl (var, ctx); | |
3326 | if (new_var == NULL_TREE) | |
3327 | new_var = maybe_lookup_decl_in_outer_ctx (var, ctx); | |
3328 | x = builtin_decl_explicit (BUILT_IN_ASSUME_ALIGNED); | |
3329 | x = build_call_expr_loc (clause_loc, x, 2, new_var, | |
3330 | omp_clause_aligned_alignment (c)); | |
3331 | x = fold_convert_loc (clause_loc, TREE_TYPE (new_var), x); | |
3332 | x = build2 (MODIFY_EXPR, TREE_TYPE (new_var), new_var, x); | |
3333 | gimplify_and_add (x, ilist); | |
3334 | } | |
3335 | else if (TREE_CODE (TREE_TYPE (var)) == ARRAY_TYPE | |
3336 | && is_global_var (var)) | |
3337 | { | |
3338 | tree ptype = build_pointer_type (TREE_TYPE (var)), t, t2; | |
3339 | new_var = lookup_decl (var, ctx); | |
3340 | t = maybe_lookup_decl_in_outer_ctx (var, ctx); | |
3341 | t = build_fold_addr_expr_loc (clause_loc, t); | |
3342 | t2 = builtin_decl_explicit (BUILT_IN_ASSUME_ALIGNED); | |
3343 | t = build_call_expr_loc (clause_loc, t2, 2, t, | |
3344 | omp_clause_aligned_alignment (c)); | |
3345 | t = fold_convert_loc (clause_loc, ptype, t); | |
f9e245b2 | 3346 | x = create_tmp_var (ptype); |
bc7bff74 | 3347 | t = build2 (MODIFY_EXPR, ptype, x, t); |
3348 | gimplify_and_add (t, ilist); | |
3349 | t = build_simple_mem_ref_loc (clause_loc, x); | |
3350 | SET_DECL_VALUE_EXPR (new_var, t); | |
3351 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
3352 | } | |
3353 | continue; | |
1e8e9920 | 3354 | default: |
3355 | continue; | |
3356 | } | |
3357 | ||
3358 | new_var = var = OMP_CLAUSE_DECL (c); | |
3359 | if (c_kind != OMP_CLAUSE_COPYIN) | |
3360 | new_var = lookup_decl (var, ctx); | |
3361 | ||
3362 | if (c_kind == OMP_CLAUSE_SHARED || c_kind == OMP_CLAUSE_COPYIN) | |
3363 | { | |
3364 | if (pass != 0) | |
3365 | continue; | |
3366 | } | |
1e8e9920 | 3367 | else if (is_variable_sized (var)) |
3368 | { | |
773c5ba7 | 3369 | /* For variable sized types, we need to allocate the |
3370 | actual storage here. Call alloca and store the | |
3371 | result in the pointer decl that we created elsewhere. */ | |
1e8e9920 | 3372 | if (pass == 0) |
3373 | continue; | |
3374 | ||
fd6481cf | 3375 | if (c_kind != OMP_CLAUSE_FIRSTPRIVATE || !is_task_ctx (ctx)) |
3376 | { | |
1a91d914 | 3377 | gcall *stmt; |
b9a16870 | 3378 | tree tmp, atmp; |
75a70cf9 | 3379 | |
fd6481cf | 3380 | ptr = DECL_VALUE_EXPR (new_var); |
3381 | gcc_assert (TREE_CODE (ptr) == INDIRECT_REF); | |
3382 | ptr = TREE_OPERAND (ptr, 0); | |
3383 | gcc_assert (DECL_P (ptr)); | |
3384 | x = TYPE_SIZE_UNIT (TREE_TYPE (new_var)); | |
75a70cf9 | 3385 | |
3386 | /* void *tmp = __builtin_alloca */ | |
b9a16870 | 3387 | atmp = builtin_decl_explicit (BUILT_IN_ALLOCA); |
3388 | stmt = gimple_build_call (atmp, 1, x); | |
f9e245b2 | 3389 | tmp = create_tmp_var_raw (ptr_type_node); |
75a70cf9 | 3390 | gimple_add_tmp_var (tmp); |
3391 | gimple_call_set_lhs (stmt, tmp); | |
3392 | ||
3393 | gimple_seq_add_stmt (ilist, stmt); | |
3394 | ||
389dd41b | 3395 | x = fold_convert_loc (clause_loc, TREE_TYPE (ptr), tmp); |
75a70cf9 | 3396 | gimplify_assign (ptr, x, ilist); |
fd6481cf | 3397 | } |
1e8e9920 | 3398 | } |
1e8e9920 | 3399 | else if (is_reference (var)) |
3400 | { | |
773c5ba7 | 3401 | /* For references that are being privatized for Fortran, |
3402 | allocate new backing storage for the new pointer | |
3403 | variable. This allows us to avoid changing all the | |
3404 | code that expects a pointer to something that expects | |
bc7bff74 | 3405 | a direct variable. */ |
1e8e9920 | 3406 | if (pass == 0) |
3407 | continue; | |
3408 | ||
3409 | x = TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (new_var))); | |
fd6481cf | 3410 | if (c_kind == OMP_CLAUSE_FIRSTPRIVATE && is_task_ctx (ctx)) |
3411 | { | |
3412 | x = build_receiver_ref (var, false, ctx); | |
389dd41b | 3413 | x = build_fold_addr_expr_loc (clause_loc, x); |
fd6481cf | 3414 | } |
3415 | else if (TREE_CONSTANT (x)) | |
1e8e9920 | 3416 | { |
2712b6de | 3417 | /* For reduction in SIMD loop, defer adding the |
3418 | initialization of the reference, because if we decide | |
3419 | to use SIMD array for it, the initilization could cause | |
3420 | expansion ICE. */ | |
3421 | if (c_kind == OMP_CLAUSE_REDUCTION && is_simd) | |
09d1c205 | 3422 | x = NULL_TREE; |
3423 | else | |
3424 | { | |
3425 | const char *name = NULL; | |
3426 | if (DECL_NAME (var)) | |
3427 | name = IDENTIFIER_POINTER (DECL_NAME (new_var)); | |
3428 | ||
3429 | x = create_tmp_var_raw (TREE_TYPE (TREE_TYPE (new_var)), | |
3430 | name); | |
3431 | gimple_add_tmp_var (x); | |
3432 | TREE_ADDRESSABLE (x) = 1; | |
3433 | x = build_fold_addr_expr_loc (clause_loc, x); | |
3434 | } | |
1e8e9920 | 3435 | } |
3436 | else | |
3437 | { | |
b9a16870 | 3438 | tree atmp = builtin_decl_explicit (BUILT_IN_ALLOCA); |
3439 | x = build_call_expr_loc (clause_loc, atmp, 1, x); | |
1e8e9920 | 3440 | } |
3441 | ||
09d1c205 | 3442 | if (x) |
3443 | { | |
3444 | x = fold_convert_loc (clause_loc, TREE_TYPE (new_var), x); | |
3445 | gimplify_assign (new_var, x, ilist); | |
3446 | } | |
1e8e9920 | 3447 | |
182cf5a9 | 3448 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); |
1e8e9920 | 3449 | } |
3450 | else if (c_kind == OMP_CLAUSE_REDUCTION | |
3451 | && OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
3452 | { | |
3453 | if (pass == 0) | |
3454 | continue; | |
3455 | } | |
3456 | else if (pass != 0) | |
3457 | continue; | |
3458 | ||
55d6e7cd | 3459 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 3460 | { |
3461 | case OMP_CLAUSE_SHARED: | |
bc7bff74 | 3462 | /* Ignore shared directives in teams construct. */ |
3463 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS) | |
3464 | continue; | |
f49d7bb5 | 3465 | /* Shared global vars are just accessed directly. */ |
3466 | if (is_global_var (new_var)) | |
3467 | break; | |
1e8e9920 | 3468 | /* Set up the DECL_VALUE_EXPR for shared variables now. This |
3469 | needs to be delayed until after fixup_child_record_type so | |
3470 | that we get the correct type during the dereference. */ | |
e8a588af | 3471 | by_ref = use_pointer_for_field (var, ctx); |
1e8e9920 | 3472 | x = build_receiver_ref (var, by_ref, ctx); |
3473 | SET_DECL_VALUE_EXPR (new_var, x); | |
3474 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
3475 | ||
3476 | /* ??? If VAR is not passed by reference, and the variable | |
3477 | hasn't been initialized yet, then we'll get a warning for | |
3478 | the store into the omp_data_s structure. Ideally, we'd be | |
48e1416a | 3479 | able to notice this and not store anything at all, but |
1e8e9920 | 3480 | we're generating code too early. Suppress the warning. */ |
3481 | if (!by_ref) | |
3482 | TREE_NO_WARNING (var) = 1; | |
3483 | break; | |
3484 | ||
3485 | case OMP_CLAUSE_LASTPRIVATE: | |
3486 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) | |
3487 | break; | |
3488 | /* FALLTHRU */ | |
3489 | ||
3490 | case OMP_CLAUSE_PRIVATE: | |
fd6481cf | 3491 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_PRIVATE) |
3492 | x = build_outer_var_ref (var, ctx); | |
3493 | else if (OMP_CLAUSE_PRIVATE_OUTER_REF (c)) | |
3494 | { | |
3495 | if (is_task_ctx (ctx)) | |
3496 | x = build_receiver_ref (var, false, ctx); | |
3497 | else | |
3498 | x = build_outer_var_ref (var, ctx); | |
3499 | } | |
3500 | else | |
3501 | x = NULL; | |
3d483a94 | 3502 | do_private: |
bc7bff74 | 3503 | tree nx; |
3504 | nx = lang_hooks.decls.omp_clause_default_ctor (c, new_var, x); | |
3d483a94 | 3505 | if (is_simd) |
3506 | { | |
3507 | tree y = lang_hooks.decls.omp_clause_dtor (c, new_var); | |
bc7bff74 | 3508 | if ((TREE_ADDRESSABLE (new_var) || nx || y |
3d483a94 | 3509 | || OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE) |
3510 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3511 | idx, lane, ivar, lvar)) | |
3512 | { | |
bc7bff74 | 3513 | if (nx) |
3d483a94 | 3514 | x = lang_hooks.decls.omp_clause_default_ctor |
3515 | (c, unshare_expr (ivar), x); | |
bc7bff74 | 3516 | if (nx && x) |
3d483a94 | 3517 | gimplify_and_add (x, &llist[0]); |
3518 | if (y) | |
3519 | { | |
3520 | y = lang_hooks.decls.omp_clause_dtor (c, ivar); | |
3521 | if (y) | |
3522 | { | |
3523 | gimple_seq tseq = NULL; | |
3524 | ||
3525 | dtor = y; | |
3526 | gimplify_stmt (&dtor, &tseq); | |
3527 | gimple_seq_add_seq (&llist[1], tseq); | |
3528 | } | |
3529 | } | |
3530 | break; | |
3531 | } | |
3532 | } | |
bc7bff74 | 3533 | if (nx) |
3534 | gimplify_and_add (nx, ilist); | |
1e8e9920 | 3535 | /* FALLTHRU */ |
3536 | ||
3537 | do_dtor: | |
3538 | x = lang_hooks.decls.omp_clause_dtor (c, new_var); | |
3539 | if (x) | |
3540 | { | |
75a70cf9 | 3541 | gimple_seq tseq = NULL; |
3542 | ||
1e8e9920 | 3543 | dtor = x; |
75a70cf9 | 3544 | gimplify_stmt (&dtor, &tseq); |
e3a19533 | 3545 | gimple_seq_add_seq (dlist, tseq); |
1e8e9920 | 3546 | } |
3547 | break; | |
3548 | ||
3d483a94 | 3549 | case OMP_CLAUSE_LINEAR: |
3550 | if (!OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
3551 | goto do_firstprivate; | |
3552 | if (OMP_CLAUSE_LINEAR_NO_COPYOUT (c)) | |
3553 | x = NULL; | |
3554 | else | |
3555 | x = build_outer_var_ref (var, ctx); | |
3556 | goto do_private; | |
3557 | ||
1e8e9920 | 3558 | case OMP_CLAUSE_FIRSTPRIVATE: |
fd6481cf | 3559 | if (is_task_ctx (ctx)) |
3560 | { | |
3561 | if (is_reference (var) || is_variable_sized (var)) | |
3562 | goto do_dtor; | |
3563 | else if (is_global_var (maybe_lookup_decl_in_outer_ctx (var, | |
3564 | ctx)) | |
3565 | || use_pointer_for_field (var, NULL)) | |
3566 | { | |
3567 | x = build_receiver_ref (var, false, ctx); | |
3568 | SET_DECL_VALUE_EXPR (new_var, x); | |
3569 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
3570 | goto do_dtor; | |
3571 | } | |
3572 | } | |
3d483a94 | 3573 | do_firstprivate: |
1e8e9920 | 3574 | x = build_outer_var_ref (var, ctx); |
3d483a94 | 3575 | if (is_simd) |
3576 | { | |
bc7bff74 | 3577 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
3578 | && gimple_omp_for_combined_into_p (ctx->stmt)) | |
3579 | { | |
9580cb79 | 3580 | tree t = OMP_CLAUSE_LINEAR_STEP (c); |
3581 | tree stept = TREE_TYPE (t); | |
3582 | tree ct = find_omp_clause (clauses, | |
3583 | OMP_CLAUSE__LOOPTEMP_); | |
3584 | gcc_assert (ct); | |
3585 | tree l = OMP_CLAUSE_DECL (ct); | |
e471cc6f | 3586 | tree n1 = fd->loop.n1; |
3587 | tree step = fd->loop.step; | |
3588 | tree itype = TREE_TYPE (l); | |
3589 | if (POINTER_TYPE_P (itype)) | |
3590 | itype = signed_type_for (itype); | |
3591 | l = fold_build2 (MINUS_EXPR, itype, l, n1); | |
3592 | if (TYPE_UNSIGNED (itype) | |
3593 | && fd->loop.cond_code == GT_EXPR) | |
3594 | l = fold_build2 (TRUNC_DIV_EXPR, itype, | |
3595 | fold_build1 (NEGATE_EXPR, itype, l), | |
3596 | fold_build1 (NEGATE_EXPR, | |
3597 | itype, step)); | |
3598 | else | |
3599 | l = fold_build2 (TRUNC_DIV_EXPR, itype, l, step); | |
bc7bff74 | 3600 | t = fold_build2 (MULT_EXPR, stept, |
3601 | fold_convert (stept, l), t); | |
9580cb79 | 3602 | |
3603 | if (OMP_CLAUSE_LINEAR_ARRAY (c)) | |
3604 | { | |
3605 | x = lang_hooks.decls.omp_clause_linear_ctor | |
3606 | (c, new_var, x, t); | |
3607 | gimplify_and_add (x, ilist); | |
3608 | goto do_dtor; | |
3609 | } | |
3610 | ||
bc7bff74 | 3611 | if (POINTER_TYPE_P (TREE_TYPE (x))) |
3612 | x = fold_build2 (POINTER_PLUS_EXPR, | |
3613 | TREE_TYPE (x), x, t); | |
3614 | else | |
3615 | x = fold_build2 (PLUS_EXPR, TREE_TYPE (x), x, t); | |
3616 | } | |
3617 | ||
3d483a94 | 3618 | if ((OMP_CLAUSE_CODE (c) != OMP_CLAUSE_LINEAR |
3619 | || TREE_ADDRESSABLE (new_var)) | |
3620 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3621 | idx, lane, ivar, lvar)) | |
3622 | { | |
3623 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR) | |
3624 | { | |
f9e245b2 | 3625 | tree iv = create_tmp_var (TREE_TYPE (new_var)); |
3d483a94 | 3626 | x = lang_hooks.decls.omp_clause_copy_ctor (c, iv, x); |
3627 | gimplify_and_add (x, ilist); | |
3628 | gimple_stmt_iterator gsi | |
3629 | = gsi_start_1 (gimple_omp_body_ptr (ctx->stmt)); | |
1a91d914 | 3630 | gassign *g |
3d483a94 | 3631 | = gimple_build_assign (unshare_expr (lvar), iv); |
3632 | gsi_insert_before_without_update (&gsi, g, | |
3633 | GSI_SAME_STMT); | |
9580cb79 | 3634 | tree t = OMP_CLAUSE_LINEAR_STEP (c); |
3d483a94 | 3635 | enum tree_code code = PLUS_EXPR; |
3636 | if (POINTER_TYPE_P (TREE_TYPE (new_var))) | |
3637 | code = POINTER_PLUS_EXPR; | |
3638 | g = gimple_build_assign_with_ops (code, iv, iv, t); | |
3639 | gsi_insert_before_without_update (&gsi, g, | |
3640 | GSI_SAME_STMT); | |
3641 | break; | |
3642 | } | |
3643 | x = lang_hooks.decls.omp_clause_copy_ctor | |
3644 | (c, unshare_expr (ivar), x); | |
3645 | gimplify_and_add (x, &llist[0]); | |
3646 | x = lang_hooks.decls.omp_clause_dtor (c, ivar); | |
3647 | if (x) | |
3648 | { | |
3649 | gimple_seq tseq = NULL; | |
3650 | ||
3651 | dtor = x; | |
3652 | gimplify_stmt (&dtor, &tseq); | |
3653 | gimple_seq_add_seq (&llist[1], tseq); | |
3654 | } | |
3655 | break; | |
3656 | } | |
3657 | } | |
1e8e9920 | 3658 | x = lang_hooks.decls.omp_clause_copy_ctor (c, new_var, x); |
3659 | gimplify_and_add (x, ilist); | |
3660 | goto do_dtor; | |
1e8e9920 | 3661 | |
bc7bff74 | 3662 | case OMP_CLAUSE__LOOPTEMP_: |
3663 | gcc_assert (is_parallel_ctx (ctx)); | |
3664 | x = build_outer_var_ref (var, ctx); | |
3665 | x = build2 (MODIFY_EXPR, TREE_TYPE (new_var), new_var, x); | |
3666 | gimplify_and_add (x, ilist); | |
3667 | break; | |
3668 | ||
1e8e9920 | 3669 | case OMP_CLAUSE_COPYIN: |
e8a588af | 3670 | by_ref = use_pointer_for_field (var, NULL); |
1e8e9920 | 3671 | x = build_receiver_ref (var, by_ref, ctx); |
3672 | x = lang_hooks.decls.omp_clause_assign_op (c, new_var, x); | |
3673 | append_to_statement_list (x, ©in_seq); | |
3674 | copyin_by_ref |= by_ref; | |
3675 | break; | |
3676 | ||
3677 | case OMP_CLAUSE_REDUCTION: | |
3678 | if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
3679 | { | |
fd6481cf | 3680 | tree placeholder = OMP_CLAUSE_REDUCTION_PLACEHOLDER (c); |
bc7bff74 | 3681 | gimple tseq; |
fd6481cf | 3682 | x = build_outer_var_ref (var, ctx); |
3683 | ||
bc7bff74 | 3684 | if (is_reference (var) |
3685 | && !useless_type_conversion_p (TREE_TYPE (placeholder), | |
3686 | TREE_TYPE (x))) | |
389dd41b | 3687 | x = build_fold_addr_expr_loc (clause_loc, x); |
fd6481cf | 3688 | SET_DECL_VALUE_EXPR (placeholder, x); |
3689 | DECL_HAS_VALUE_EXPR_P (placeholder) = 1; | |
bc7bff74 | 3690 | tree new_vard = new_var; |
3691 | if (is_reference (var)) | |
3692 | { | |
3693 | gcc_assert (TREE_CODE (new_var) == MEM_REF); | |
3694 | new_vard = TREE_OPERAND (new_var, 0); | |
3695 | gcc_assert (DECL_P (new_vard)); | |
3696 | } | |
3d483a94 | 3697 | if (is_simd |
3698 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3699 | idx, lane, ivar, lvar)) | |
3700 | { | |
bc7bff74 | 3701 | if (new_vard == new_var) |
3702 | { | |
3703 | gcc_assert (DECL_VALUE_EXPR (new_var) == lvar); | |
3704 | SET_DECL_VALUE_EXPR (new_var, ivar); | |
3705 | } | |
3706 | else | |
3707 | { | |
3708 | SET_DECL_VALUE_EXPR (new_vard, | |
3709 | build_fold_addr_expr (ivar)); | |
3710 | DECL_HAS_VALUE_EXPR_P (new_vard) = 1; | |
3711 | } | |
3712 | x = lang_hooks.decls.omp_clause_default_ctor | |
3713 | (c, unshare_expr (ivar), | |
3714 | build_outer_var_ref (var, ctx)); | |
3715 | if (x) | |
3716 | gimplify_and_add (x, &llist[0]); | |
3717 | if (OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c)) | |
3718 | { | |
3719 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c); | |
3720 | lower_omp (&tseq, ctx); | |
3721 | gimple_seq_add_seq (&llist[0], tseq); | |
3722 | } | |
3723 | OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c) = NULL; | |
3724 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c); | |
3725 | lower_omp (&tseq, ctx); | |
3726 | gimple_seq_add_seq (&llist[1], tseq); | |
3727 | OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c) = NULL; | |
3728 | DECL_HAS_VALUE_EXPR_P (placeholder) = 0; | |
3729 | if (new_vard == new_var) | |
3730 | SET_DECL_VALUE_EXPR (new_var, lvar); | |
3731 | else | |
3732 | SET_DECL_VALUE_EXPR (new_vard, | |
3733 | build_fold_addr_expr (lvar)); | |
3734 | x = lang_hooks.decls.omp_clause_dtor (c, ivar); | |
3735 | if (x) | |
3736 | { | |
3737 | tseq = NULL; | |
3738 | dtor = x; | |
3739 | gimplify_stmt (&dtor, &tseq); | |
3740 | gimple_seq_add_seq (&llist[1], tseq); | |
3741 | } | |
3742 | break; | |
3743 | } | |
09d1c205 | 3744 | /* If this is a reference to constant size reduction var |
3745 | with placeholder, we haven't emitted the initializer | |
3746 | for it because it is undesirable if SIMD arrays are used. | |
3747 | But if they aren't used, we need to emit the deferred | |
3748 | initialization now. */ | |
3749 | else if (is_reference (var) && is_simd) | |
2712b6de | 3750 | handle_simd_reference (clause_loc, new_vard, ilist); |
bc7bff74 | 3751 | x = lang_hooks.decls.omp_clause_default_ctor |
cf5f881f | 3752 | (c, unshare_expr (new_var), |
3753 | build_outer_var_ref (var, ctx)); | |
bc7bff74 | 3754 | if (x) |
3755 | gimplify_and_add (x, ilist); | |
3756 | if (OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c)) | |
3757 | { | |
3758 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c); | |
3759 | lower_omp (&tseq, ctx); | |
3760 | gimple_seq_add_seq (ilist, tseq); | |
3761 | } | |
75a70cf9 | 3762 | OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c) = NULL; |
bc7bff74 | 3763 | if (is_simd) |
3764 | { | |
3765 | tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c); | |
3766 | lower_omp (&tseq, ctx); | |
3767 | gimple_seq_add_seq (dlist, tseq); | |
3768 | OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c) = NULL; | |
3769 | } | |
fd6481cf | 3770 | DECL_HAS_VALUE_EXPR_P (placeholder) = 0; |
bc7bff74 | 3771 | goto do_dtor; |
1e8e9920 | 3772 | } |
3773 | else | |
3774 | { | |
3775 | x = omp_reduction_init (c, TREE_TYPE (new_var)); | |
3776 | gcc_assert (TREE_CODE (TREE_TYPE (new_var)) != ARRAY_TYPE); | |
c22ad515 | 3777 | enum tree_code code = OMP_CLAUSE_REDUCTION_CODE (c); |
3778 | ||
3779 | /* reduction(-:var) sums up the partial results, so it | |
3780 | acts identically to reduction(+:var). */ | |
3781 | if (code == MINUS_EXPR) | |
3782 | code = PLUS_EXPR; | |
3783 | ||
2712b6de | 3784 | tree new_vard = new_var; |
3785 | if (is_simd && is_reference (var)) | |
3786 | { | |
3787 | gcc_assert (TREE_CODE (new_var) == MEM_REF); | |
3788 | new_vard = TREE_OPERAND (new_var, 0); | |
3789 | gcc_assert (DECL_P (new_vard)); | |
3790 | } | |
3d483a94 | 3791 | if (is_simd |
3792 | && lower_rec_simd_input_clauses (new_var, ctx, max_vf, | |
3793 | idx, lane, ivar, lvar)) | |
3794 | { | |
3d483a94 | 3795 | tree ref = build_outer_var_ref (var, ctx); |
3796 | ||
3797 | gimplify_assign (unshare_expr (ivar), x, &llist[0]); | |
3798 | ||
3d483a94 | 3799 | x = build2 (code, TREE_TYPE (ref), ref, ivar); |
3800 | ref = build_outer_var_ref (var, ctx); | |
3801 | gimplify_assign (ref, x, &llist[1]); | |
2712b6de | 3802 | |
3803 | if (new_vard != new_var) | |
3804 | { | |
3805 | SET_DECL_VALUE_EXPR (new_vard, | |
3806 | build_fold_addr_expr (lvar)); | |
3807 | DECL_HAS_VALUE_EXPR_P (new_vard) = 1; | |
3808 | } | |
3d483a94 | 3809 | } |
3810 | else | |
3811 | { | |
2712b6de | 3812 | if (is_reference (var) && is_simd) |
3813 | handle_simd_reference (clause_loc, new_vard, ilist); | |
3d483a94 | 3814 | gimplify_assign (new_var, x, ilist); |
3815 | if (is_simd) | |
c22ad515 | 3816 | { |
3817 | tree ref = build_outer_var_ref (var, ctx); | |
3818 | ||
3819 | x = build2 (code, TREE_TYPE (ref), ref, new_var); | |
3820 | ref = build_outer_var_ref (var, ctx); | |
3821 | gimplify_assign (ref, x, dlist); | |
3822 | } | |
3d483a94 | 3823 | } |
1e8e9920 | 3824 | } |
3825 | break; | |
3826 | ||
3827 | default: | |
3828 | gcc_unreachable (); | |
3829 | } | |
3830 | } | |
3831 | } | |
3832 | ||
3d483a94 | 3833 | if (lane) |
3834 | { | |
3835 | tree uid = create_tmp_var (ptr_type_node, "simduid"); | |
8e1a382d | 3836 | /* Don't want uninit warnings on simduid, it is always uninitialized, |
3837 | but we use it not for the value, but for the DECL_UID only. */ | |
3838 | TREE_NO_WARNING (uid) = 1; | |
3d483a94 | 3839 | gimple g |
3840 | = gimple_build_call_internal (IFN_GOMP_SIMD_LANE, 1, uid); | |
3841 | gimple_call_set_lhs (g, lane); | |
3842 | gimple_stmt_iterator gsi = gsi_start_1 (gimple_omp_body_ptr (ctx->stmt)); | |
3843 | gsi_insert_before_without_update (&gsi, g, GSI_SAME_STMT); | |
3844 | c = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE__SIMDUID_); | |
3845 | OMP_CLAUSE__SIMDUID__DECL (c) = uid; | |
3846 | OMP_CLAUSE_CHAIN (c) = gimple_omp_for_clauses (ctx->stmt); | |
3847 | gimple_omp_for_set_clauses (ctx->stmt, c); | |
3848 | g = gimple_build_assign_with_ops (INTEGER_CST, lane, | |
806413d2 | 3849 | build_int_cst (unsigned_type_node, 0)); |
3d483a94 | 3850 | gimple_seq_add_stmt (ilist, g); |
3851 | for (int i = 0; i < 2; i++) | |
3852 | if (llist[i]) | |
3853 | { | |
f9e245b2 | 3854 | tree vf = create_tmp_var (unsigned_type_node); |
3d483a94 | 3855 | g = gimple_build_call_internal (IFN_GOMP_SIMD_VF, 1, uid); |
3856 | gimple_call_set_lhs (g, vf); | |
3857 | gimple_seq *seq = i == 0 ? ilist : dlist; | |
3858 | gimple_seq_add_stmt (seq, g); | |
3859 | tree t = build_int_cst (unsigned_type_node, 0); | |
806413d2 | 3860 | g = gimple_build_assign_with_ops (INTEGER_CST, idx, t); |
3d483a94 | 3861 | gimple_seq_add_stmt (seq, g); |
3862 | tree body = create_artificial_label (UNKNOWN_LOCATION); | |
3863 | tree header = create_artificial_label (UNKNOWN_LOCATION); | |
3864 | tree end = create_artificial_label (UNKNOWN_LOCATION); | |
3865 | gimple_seq_add_stmt (seq, gimple_build_goto (header)); | |
3866 | gimple_seq_add_stmt (seq, gimple_build_label (body)); | |
3867 | gimple_seq_add_seq (seq, llist[i]); | |
3868 | t = build_int_cst (unsigned_type_node, 1); | |
3869 | g = gimple_build_assign_with_ops (PLUS_EXPR, idx, idx, t); | |
3870 | gimple_seq_add_stmt (seq, g); | |
3871 | gimple_seq_add_stmt (seq, gimple_build_label (header)); | |
3872 | g = gimple_build_cond (LT_EXPR, idx, vf, body, end); | |
3873 | gimple_seq_add_stmt (seq, g); | |
3874 | gimple_seq_add_stmt (seq, gimple_build_label (end)); | |
3875 | } | |
3876 | } | |
3877 | ||
1e8e9920 | 3878 | /* The copyin sequence is not to be executed by the main thread, since |
3879 | that would result in self-copies. Perhaps not visible to scalars, | |
3880 | but it certainly is to C++ operator=. */ | |
3881 | if (copyin_seq) | |
3882 | { | |
b9a16870 | 3883 | x = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM), |
3884 | 0); | |
1e8e9920 | 3885 | x = build2 (NE_EXPR, boolean_type_node, x, |
3886 | build_int_cst (TREE_TYPE (x), 0)); | |
3887 | x = build3 (COND_EXPR, void_type_node, x, copyin_seq, NULL); | |
3888 | gimplify_and_add (x, ilist); | |
3889 | } | |
3890 | ||
3891 | /* If any copyin variable is passed by reference, we must ensure the | |
3892 | master thread doesn't modify it before it is copied over in all | |
f49d7bb5 | 3893 | threads. Similarly for variables in both firstprivate and |
3894 | lastprivate clauses we need to ensure the lastprivate copying | |
bc7bff74 | 3895 | happens after firstprivate copying in all threads. And similarly |
3896 | for UDRs if initializer expression refers to omp_orig. */ | |
3897 | if (copyin_by_ref || lastprivate_firstprivate || reduction_omp_orig_ref) | |
3d483a94 | 3898 | { |
3899 | /* Don't add any barrier for #pragma omp simd or | |
3900 | #pragma omp distribute. */ | |
3901 | if (gimple_code (ctx->stmt) != GIMPLE_OMP_FOR | |
7c01fa6c | 3902 | || gimple_omp_for_kind (ctx->stmt) == GF_OMP_FOR_KIND_FOR) |
bc7bff74 | 3903 | gimple_seq_add_stmt (ilist, build_omp_barrier (NULL_TREE)); |
3d483a94 | 3904 | } |
3905 | ||
3906 | /* If max_vf is non-zero, then we can use only a vectorization factor | |
3907 | up to the max_vf we chose. So stick it into the safelen clause. */ | |
3908 | if (max_vf) | |
3909 | { | |
3910 | tree c = find_omp_clause (gimple_omp_for_clauses (ctx->stmt), | |
3911 | OMP_CLAUSE_SAFELEN); | |
3912 | if (c == NULL_TREE | |
c3f3b68d | 3913 | || (TREE_CODE (OMP_CLAUSE_SAFELEN_EXPR (c)) == INTEGER_CST |
3914 | && compare_tree_int (OMP_CLAUSE_SAFELEN_EXPR (c), | |
3915 | max_vf) == 1)) | |
3d483a94 | 3916 | { |
3917 | c = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE_SAFELEN); | |
3918 | OMP_CLAUSE_SAFELEN_EXPR (c) = build_int_cst (integer_type_node, | |
3919 | max_vf); | |
3920 | OMP_CLAUSE_CHAIN (c) = gimple_omp_for_clauses (ctx->stmt); | |
3921 | gimple_omp_for_set_clauses (ctx->stmt, c); | |
3922 | } | |
3923 | } | |
1e8e9920 | 3924 | } |
3925 | ||
773c5ba7 | 3926 | |
1e8e9920 | 3927 | /* Generate code to implement the LASTPRIVATE clauses. This is used for |
3928 | both parallel and workshare constructs. PREDICATE may be NULL if it's | |
3929 | always true. */ | |
3930 | ||
3931 | static void | |
75a70cf9 | 3932 | lower_lastprivate_clauses (tree clauses, tree predicate, gimple_seq *stmt_list, |
bc7bff74 | 3933 | omp_context *ctx) |
1e8e9920 | 3934 | { |
3d483a94 | 3935 | tree x, c, label = NULL, orig_clauses = clauses; |
fd6481cf | 3936 | bool par_clauses = false; |
3d483a94 | 3937 | tree simduid = NULL, lastlane = NULL; |
1e8e9920 | 3938 | |
3d483a94 | 3939 | /* Early exit if there are no lastprivate or linear clauses. */ |
3940 | for (; clauses ; clauses = OMP_CLAUSE_CHAIN (clauses)) | |
3941 | if (OMP_CLAUSE_CODE (clauses) == OMP_CLAUSE_LASTPRIVATE | |
3942 | || (OMP_CLAUSE_CODE (clauses) == OMP_CLAUSE_LINEAR | |
3943 | && !OMP_CLAUSE_LINEAR_NO_COPYOUT (clauses))) | |
3944 | break; | |
1e8e9920 | 3945 | if (clauses == NULL) |
3946 | { | |
3947 | /* If this was a workshare clause, see if it had been combined | |
3948 | with its parallel. In that case, look for the clauses on the | |
3949 | parallel statement itself. */ | |
3950 | if (is_parallel_ctx (ctx)) | |
3951 | return; | |
3952 | ||
3953 | ctx = ctx->outer; | |
3954 | if (ctx == NULL || !is_parallel_ctx (ctx)) | |
3955 | return; | |
3956 | ||
75a70cf9 | 3957 | clauses = find_omp_clause (gimple_omp_parallel_clauses (ctx->stmt), |
1e8e9920 | 3958 | OMP_CLAUSE_LASTPRIVATE); |
3959 | if (clauses == NULL) | |
3960 | return; | |
fd6481cf | 3961 | par_clauses = true; |
1e8e9920 | 3962 | } |
3963 | ||
75a70cf9 | 3964 | if (predicate) |
3965 | { | |
1a91d914 | 3966 | gcond *stmt; |
75a70cf9 | 3967 | tree label_true, arm1, arm2; |
3968 | ||
e60a6f7b | 3969 | label = create_artificial_label (UNKNOWN_LOCATION); |
3970 | label_true = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 3971 | arm1 = TREE_OPERAND (predicate, 0); |
3972 | arm2 = TREE_OPERAND (predicate, 1); | |
3973 | gimplify_expr (&arm1, stmt_list, NULL, is_gimple_val, fb_rvalue); | |
3974 | gimplify_expr (&arm2, stmt_list, NULL, is_gimple_val, fb_rvalue); | |
3975 | stmt = gimple_build_cond (TREE_CODE (predicate), arm1, arm2, | |
3976 | label_true, label); | |
3977 | gimple_seq_add_stmt (stmt_list, stmt); | |
3978 | gimple_seq_add_stmt (stmt_list, gimple_build_label (label_true)); | |
3979 | } | |
1e8e9920 | 3980 | |
3d483a94 | 3981 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR |
10c55644 | 3982 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_SIMD) |
3d483a94 | 3983 | { |
3984 | simduid = find_omp_clause (orig_clauses, OMP_CLAUSE__SIMDUID_); | |
3985 | if (simduid) | |
3986 | simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); | |
3987 | } | |
3988 | ||
fd6481cf | 3989 | for (c = clauses; c ;) |
1e8e9920 | 3990 | { |
3991 | tree var, new_var; | |
389dd41b | 3992 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 3993 | |
3d483a94 | 3994 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE |
3995 | || (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
3996 | && !OMP_CLAUSE_LINEAR_NO_COPYOUT (c))) | |
fd6481cf | 3997 | { |
3998 | var = OMP_CLAUSE_DECL (c); | |
3999 | new_var = lookup_decl (var, ctx); | |
1e8e9920 | 4000 | |
3d483a94 | 4001 | if (simduid && DECL_HAS_VALUE_EXPR_P (new_var)) |
4002 | { | |
4003 | tree val = DECL_VALUE_EXPR (new_var); | |
4004 | if (TREE_CODE (val) == ARRAY_REF | |
4005 | && VAR_P (TREE_OPERAND (val, 0)) | |
4006 | && lookup_attribute ("omp simd array", | |
4007 | DECL_ATTRIBUTES (TREE_OPERAND (val, | |
4008 | 0)))) | |
4009 | { | |
4010 | if (lastlane == NULL) | |
4011 | { | |
f9e245b2 | 4012 | lastlane = create_tmp_var (unsigned_type_node); |
1a91d914 | 4013 | gcall *g |
3d483a94 | 4014 | = gimple_build_call_internal (IFN_GOMP_SIMD_LAST_LANE, |
4015 | 2, simduid, | |
4016 | TREE_OPERAND (val, 1)); | |
4017 | gimple_call_set_lhs (g, lastlane); | |
4018 | gimple_seq_add_stmt (stmt_list, g); | |
4019 | } | |
4020 | new_var = build4 (ARRAY_REF, TREE_TYPE (val), | |
4021 | TREE_OPERAND (val, 0), lastlane, | |
4022 | NULL_TREE, NULL_TREE); | |
4023 | } | |
4024 | } | |
4025 | ||
4026 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE | |
4027 | && OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)) | |
75a70cf9 | 4028 | { |
e3a19533 | 4029 | lower_omp (&OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c), ctx); |
75a70cf9 | 4030 | gimple_seq_add_seq (stmt_list, |
4031 | OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c)); | |
3d483a94 | 4032 | OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c) = NULL; |
75a70cf9 | 4033 | } |
2b536a17 | 4034 | else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
4035 | && OMP_CLAUSE_LINEAR_GIMPLE_SEQ (c)) | |
4036 | { | |
4037 | lower_omp (&OMP_CLAUSE_LINEAR_GIMPLE_SEQ (c), ctx); | |
4038 | gimple_seq_add_seq (stmt_list, | |
4039 | OMP_CLAUSE_LINEAR_GIMPLE_SEQ (c)); | |
4040 | OMP_CLAUSE_LINEAR_GIMPLE_SEQ (c) = NULL; | |
4041 | } | |
1e8e9920 | 4042 | |
fd6481cf | 4043 | x = build_outer_var_ref (var, ctx); |
4044 | if (is_reference (var)) | |
182cf5a9 | 4045 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); |
fd6481cf | 4046 | x = lang_hooks.decls.omp_clause_assign_op (c, x, new_var); |
75a70cf9 | 4047 | gimplify_and_add (x, stmt_list); |
fd6481cf | 4048 | } |
4049 | c = OMP_CLAUSE_CHAIN (c); | |
4050 | if (c == NULL && !par_clauses) | |
4051 | { | |
4052 | /* If this was a workshare clause, see if it had been combined | |
4053 | with its parallel. In that case, continue looking for the | |
4054 | clauses also on the parallel statement itself. */ | |
4055 | if (is_parallel_ctx (ctx)) | |
4056 | break; | |
4057 | ||
4058 | ctx = ctx->outer; | |
4059 | if (ctx == NULL || !is_parallel_ctx (ctx)) | |
4060 | break; | |
4061 | ||
75a70cf9 | 4062 | c = find_omp_clause (gimple_omp_parallel_clauses (ctx->stmt), |
fd6481cf | 4063 | OMP_CLAUSE_LASTPRIVATE); |
4064 | par_clauses = true; | |
4065 | } | |
1e8e9920 | 4066 | } |
4067 | ||
75a70cf9 | 4068 | if (label) |
4069 | gimple_seq_add_stmt (stmt_list, gimple_build_label (label)); | |
1e8e9920 | 4070 | } |
4071 | ||
773c5ba7 | 4072 | |
1e8e9920 | 4073 | /* Generate code to implement the REDUCTION clauses. */ |
4074 | ||
4075 | static void | |
75a70cf9 | 4076 | lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx) |
1e8e9920 | 4077 | { |
75a70cf9 | 4078 | gimple_seq sub_seq = NULL; |
4079 | gimple stmt; | |
4080 | tree x, c; | |
1e8e9920 | 4081 | int count = 0; |
4082 | ||
3d483a94 | 4083 | /* SIMD reductions are handled in lower_rec_input_clauses. */ |
4084 | if (gimple_code (ctx->stmt) == GIMPLE_OMP_FOR | |
10c55644 | 4085 | && gimple_omp_for_kind (ctx->stmt) & GF_OMP_FOR_SIMD) |
3d483a94 | 4086 | return; |
4087 | ||
1e8e9920 | 4088 | /* First see if there is exactly one reduction clause. Use OMP_ATOMIC |
4089 | update in that case, otherwise use a lock. */ | |
4090 | for (c = clauses; c && count < 2; c = OMP_CLAUSE_CHAIN (c)) | |
55d6e7cd | 4091 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION) |
1e8e9920 | 4092 | { |
4093 | if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
4094 | { | |
bc7bff74 | 4095 | /* Never use OMP_ATOMIC for array reductions or UDRs. */ |
1e8e9920 | 4096 | count = -1; |
4097 | break; | |
4098 | } | |
4099 | count++; | |
4100 | } | |
4101 | ||
4102 | if (count == 0) | |
4103 | return; | |
4104 | ||
4105 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
4106 | { | |
4107 | tree var, ref, new_var; | |
4108 | enum tree_code code; | |
389dd41b | 4109 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 4110 | |
55d6e7cd | 4111 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION) |
1e8e9920 | 4112 | continue; |
4113 | ||
4114 | var = OMP_CLAUSE_DECL (c); | |
4115 | new_var = lookup_decl (var, ctx); | |
4116 | if (is_reference (var)) | |
182cf5a9 | 4117 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); |
1e8e9920 | 4118 | ref = build_outer_var_ref (var, ctx); |
4119 | code = OMP_CLAUSE_REDUCTION_CODE (c); | |
773c5ba7 | 4120 | |
4121 | /* reduction(-:var) sums up the partial results, so it acts | |
4122 | identically to reduction(+:var). */ | |
1e8e9920 | 4123 | if (code == MINUS_EXPR) |
4124 | code = PLUS_EXPR; | |
4125 | ||
4126 | if (count == 1) | |
4127 | { | |
389dd41b | 4128 | tree addr = build_fold_addr_expr_loc (clause_loc, ref); |
1e8e9920 | 4129 | |
4130 | addr = save_expr (addr); | |
4131 | ref = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (addr)), addr); | |
389dd41b | 4132 | x = fold_build2_loc (clause_loc, code, TREE_TYPE (ref), ref, new_var); |
1e8e9920 | 4133 | x = build2 (OMP_ATOMIC, void_type_node, addr, x); |
75a70cf9 | 4134 | gimplify_and_add (x, stmt_seqp); |
1e8e9920 | 4135 | return; |
4136 | } | |
4137 | ||
4138 | if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c)) | |
4139 | { | |
4140 | tree placeholder = OMP_CLAUSE_REDUCTION_PLACEHOLDER (c); | |
4141 | ||
bc7bff74 | 4142 | if (is_reference (var) |
4143 | && !useless_type_conversion_p (TREE_TYPE (placeholder), | |
4144 | TREE_TYPE (ref))) | |
389dd41b | 4145 | ref = build_fold_addr_expr_loc (clause_loc, ref); |
1e8e9920 | 4146 | SET_DECL_VALUE_EXPR (placeholder, ref); |
4147 | DECL_HAS_VALUE_EXPR_P (placeholder) = 1; | |
e3a19533 | 4148 | lower_omp (&OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c), ctx); |
75a70cf9 | 4149 | gimple_seq_add_seq (&sub_seq, OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c)); |
4150 | OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c) = NULL; | |
1e8e9920 | 4151 | OMP_CLAUSE_REDUCTION_PLACEHOLDER (c) = NULL; |
4152 | } | |
4153 | else | |
4154 | { | |
4155 | x = build2 (code, TREE_TYPE (ref), ref, new_var); | |
4156 | ref = build_outer_var_ref (var, ctx); | |
75a70cf9 | 4157 | gimplify_assign (ref, x, &sub_seq); |
1e8e9920 | 4158 | } |
4159 | } | |
4160 | ||
b9a16870 | 4161 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START), |
4162 | 0); | |
75a70cf9 | 4163 | gimple_seq_add_stmt (stmt_seqp, stmt); |
1e8e9920 | 4164 | |
75a70cf9 | 4165 | gimple_seq_add_seq (stmt_seqp, sub_seq); |
1e8e9920 | 4166 | |
b9a16870 | 4167 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END), |
4168 | 0); | |
75a70cf9 | 4169 | gimple_seq_add_stmt (stmt_seqp, stmt); |
1e8e9920 | 4170 | } |
4171 | ||
773c5ba7 | 4172 | |
1e8e9920 | 4173 | /* Generate code to implement the COPYPRIVATE clauses. */ |
4174 | ||
4175 | static void | |
75a70cf9 | 4176 | lower_copyprivate_clauses (tree clauses, gimple_seq *slist, gimple_seq *rlist, |
1e8e9920 | 4177 | omp_context *ctx) |
4178 | { | |
4179 | tree c; | |
4180 | ||
4181 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
4182 | { | |
cb561506 | 4183 | tree var, new_var, ref, x; |
1e8e9920 | 4184 | bool by_ref; |
389dd41b | 4185 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 4186 | |
55d6e7cd | 4187 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYPRIVATE) |
1e8e9920 | 4188 | continue; |
4189 | ||
4190 | var = OMP_CLAUSE_DECL (c); | |
e8a588af | 4191 | by_ref = use_pointer_for_field (var, NULL); |
1e8e9920 | 4192 | |
4193 | ref = build_sender_ref (var, ctx); | |
cb561506 | 4194 | x = new_var = lookup_decl_in_outer_ctx (var, ctx); |
4195 | if (by_ref) | |
4196 | { | |
4197 | x = build_fold_addr_expr_loc (clause_loc, new_var); | |
4198 | x = fold_convert_loc (clause_loc, TREE_TYPE (ref), x); | |
4199 | } | |
75a70cf9 | 4200 | gimplify_assign (ref, x, slist); |
1e8e9920 | 4201 | |
cb561506 | 4202 | ref = build_receiver_ref (var, false, ctx); |
4203 | if (by_ref) | |
4204 | { | |
4205 | ref = fold_convert_loc (clause_loc, | |
4206 | build_pointer_type (TREE_TYPE (new_var)), | |
4207 | ref); | |
4208 | ref = build_fold_indirect_ref_loc (clause_loc, ref); | |
4209 | } | |
1e8e9920 | 4210 | if (is_reference (var)) |
4211 | { | |
cb561506 | 4212 | ref = fold_convert_loc (clause_loc, TREE_TYPE (new_var), ref); |
182cf5a9 | 4213 | ref = build_simple_mem_ref_loc (clause_loc, ref); |
4214 | new_var = build_simple_mem_ref_loc (clause_loc, new_var); | |
1e8e9920 | 4215 | } |
cb561506 | 4216 | x = lang_hooks.decls.omp_clause_assign_op (c, new_var, ref); |
1e8e9920 | 4217 | gimplify_and_add (x, rlist); |
4218 | } | |
4219 | } | |
4220 | ||
773c5ba7 | 4221 | |
1e8e9920 | 4222 | /* Generate code to implement the clauses, FIRSTPRIVATE, COPYIN, LASTPRIVATE, |
4223 | and REDUCTION from the sender (aka parent) side. */ | |
4224 | ||
4225 | static void | |
75a70cf9 | 4226 | lower_send_clauses (tree clauses, gimple_seq *ilist, gimple_seq *olist, |
4227 | omp_context *ctx) | |
1e8e9920 | 4228 | { |
4229 | tree c; | |
4230 | ||
4231 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
4232 | { | |
773c5ba7 | 4233 | tree val, ref, x, var; |
1e8e9920 | 4234 | bool by_ref, do_in = false, do_out = false; |
389dd41b | 4235 | location_t clause_loc = OMP_CLAUSE_LOCATION (c); |
1e8e9920 | 4236 | |
55d6e7cd | 4237 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 4238 | { |
fd6481cf | 4239 | case OMP_CLAUSE_PRIVATE: |
4240 | if (OMP_CLAUSE_PRIVATE_OUTER_REF (c)) | |
4241 | break; | |
4242 | continue; | |
1e8e9920 | 4243 | case OMP_CLAUSE_FIRSTPRIVATE: |
4244 | case OMP_CLAUSE_COPYIN: | |
4245 | case OMP_CLAUSE_LASTPRIVATE: | |
4246 | case OMP_CLAUSE_REDUCTION: | |
bc7bff74 | 4247 | case OMP_CLAUSE__LOOPTEMP_: |
1e8e9920 | 4248 | break; |
4249 | default: | |
4250 | continue; | |
4251 | } | |
4252 | ||
87b31375 | 4253 | val = OMP_CLAUSE_DECL (c); |
4254 | var = lookup_decl_in_outer_ctx (val, ctx); | |
773c5ba7 | 4255 | |
f49d7bb5 | 4256 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYIN |
4257 | && is_global_var (var)) | |
4258 | continue; | |
1e8e9920 | 4259 | if (is_variable_sized (val)) |
4260 | continue; | |
e8a588af | 4261 | by_ref = use_pointer_for_field (val, NULL); |
1e8e9920 | 4262 | |
55d6e7cd | 4263 | switch (OMP_CLAUSE_CODE (c)) |
1e8e9920 | 4264 | { |
fd6481cf | 4265 | case OMP_CLAUSE_PRIVATE: |
1e8e9920 | 4266 | case OMP_CLAUSE_FIRSTPRIVATE: |
4267 | case OMP_CLAUSE_COPYIN: | |
bc7bff74 | 4268 | case OMP_CLAUSE__LOOPTEMP_: |
1e8e9920 | 4269 | do_in = true; |
4270 | break; | |
4271 | ||
4272 | case OMP_CLAUSE_LASTPRIVATE: | |
4273 | if (by_ref || is_reference (val)) | |
4274 | { | |
4275 | if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c)) | |
4276 | continue; | |
4277 | do_in = true; | |
4278 | } | |
4279 | else | |
fd6481cf | 4280 | { |
4281 | do_out = true; | |
4282 | if (lang_hooks.decls.omp_private_outer_ref (val)) | |
4283 | do_in = true; | |
4284 | } | |
1e8e9920 | 4285 | break; |
4286 | ||
4287 | case OMP_CLAUSE_REDUCTION: | |
4288 | do_in = true; | |
4289 | do_out = !(by_ref || is_reference (val)); | |
4290 | break; | |
4291 | ||
4292 | default: | |
4293 | gcc_unreachable (); | |
4294 | } | |
4295 | ||
4296 | if (do_in) | |
4297 | { | |
4298 | ref = build_sender_ref (val, ctx); | |
389dd41b | 4299 | x = by_ref ? build_fold_addr_expr_loc (clause_loc, var) : var; |
75a70cf9 | 4300 | gimplify_assign (ref, x, ilist); |
fd6481cf | 4301 | if (is_task_ctx (ctx)) |
4302 | DECL_ABSTRACT_ORIGIN (TREE_OPERAND (ref, 1)) = NULL; | |
1e8e9920 | 4303 | } |
773c5ba7 | 4304 | |
1e8e9920 | 4305 | if (do_out) |
4306 | { | |
4307 | ref = build_sender_ref (val, ctx); | |
75a70cf9 | 4308 | gimplify_assign (var, ref, olist); |
1e8e9920 | 4309 | } |
4310 | } | |
4311 | } | |
4312 | ||
75a70cf9 | 4313 | /* Generate code to implement SHARED from the sender (aka parent) |
4314 | side. This is trickier, since GIMPLE_OMP_PARALLEL_CLAUSES doesn't | |
4315 | list things that got automatically shared. */ | |
1e8e9920 | 4316 | |
4317 | static void | |
75a70cf9 | 4318 | lower_send_shared_vars (gimple_seq *ilist, gimple_seq *olist, omp_context *ctx) |
1e8e9920 | 4319 | { |
fd6481cf | 4320 | tree var, ovar, nvar, f, x, record_type; |
1e8e9920 | 4321 | |
4322 | if (ctx->record_type == NULL) | |
4323 | return; | |
773c5ba7 | 4324 | |
fd6481cf | 4325 | record_type = ctx->srecord_type ? ctx->srecord_type : ctx->record_type; |
1767a056 | 4326 | for (f = TYPE_FIELDS (record_type); f ; f = DECL_CHAIN (f)) |
1e8e9920 | 4327 | { |
4328 | ovar = DECL_ABSTRACT_ORIGIN (f); | |
4329 | nvar = maybe_lookup_decl (ovar, ctx); | |
4330 | if (!nvar || !DECL_HAS_VALUE_EXPR_P (nvar)) | |
4331 | continue; | |
4332 | ||
773c5ba7 | 4333 | /* If CTX is a nested parallel directive. Find the immediately |
4334 | enclosing parallel or workshare construct that contains a | |
4335 | mapping for OVAR. */ | |
87b31375 | 4336 | var = lookup_decl_in_outer_ctx (ovar, ctx); |
773c5ba7 | 4337 | |
e8a588af | 4338 | if (use_pointer_for_field (ovar, ctx)) |
1e8e9920 | 4339 | { |
4340 | x = build_sender_ref (ovar, ctx); | |
773c5ba7 | 4341 | var = build_fold_addr_expr (var); |
75a70cf9 | 4342 | gimplify_assign (x, var, ilist); |
1e8e9920 | 4343 | } |
4344 | else | |
4345 | { | |
4346 | x = build_sender_ref (ovar, ctx); | |
75a70cf9 | 4347 | gimplify_assign (x, var, ilist); |
1e8e9920 | 4348 | |
d2263ebb | 4349 | if (!TREE_READONLY (var) |
4350 | /* We don't need to receive a new reference to a result | |
4351 | or parm decl. In fact we may not store to it as we will | |
4352 | invalidate any pending RSO and generate wrong gimple | |
4353 | during inlining. */ | |
4354 | && !((TREE_CODE (var) == RESULT_DECL | |
4355 | || TREE_CODE (var) == PARM_DECL) | |
4356 | && DECL_BY_REFERENCE (var))) | |
fd6481cf | 4357 | { |
4358 | x = build_sender_ref (ovar, ctx); | |
75a70cf9 | 4359 | gimplify_assign (var, x, olist); |
fd6481cf | 4360 | } |
1e8e9920 | 4361 | } |
4362 | } | |
4363 | } | |
4364 | ||
75a70cf9 | 4365 | |
4366 | /* A convenience function to build an empty GIMPLE_COND with just the | |
4367 | condition. */ | |
4368 | ||
1a91d914 | 4369 | static gcond * |
75a70cf9 | 4370 | gimple_build_cond_empty (tree cond) |
4371 | { | |
4372 | enum tree_code pred_code; | |
4373 | tree lhs, rhs; | |
4374 | ||
4375 | gimple_cond_get_ops_from_tree (cond, &pred_code, &lhs, &rhs); | |
4376 | return gimple_build_cond (pred_code, lhs, rhs, NULL_TREE, NULL_TREE); | |
4377 | } | |
4378 | ||
4379 | ||
48e1416a | 4380 | /* Build the function calls to GOMP_parallel_start etc to actually |
773c5ba7 | 4381 | generate the parallel operation. REGION is the parallel region |
4382 | being expanded. BB is the block where to insert the code. WS_ARGS | |
4383 | will be set if this is a call to a combined parallel+workshare | |
4384 | construct, it contains the list of additional arguments needed by | |
4385 | the workshare construct. */ | |
1e8e9920 | 4386 | |
4387 | static void | |
61e47ac8 | 4388 | expand_parallel_call (struct omp_region *region, basic_block bb, |
1a91d914 | 4389 | gomp_parallel *entry_stmt, |
4390 | vec<tree, va_gc> *ws_args) | |
1e8e9920 | 4391 | { |
bc7bff74 | 4392 | tree t, t1, t2, val, cond, c, clauses, flags; |
75a70cf9 | 4393 | gimple_stmt_iterator gsi; |
4394 | gimple stmt; | |
b9a16870 | 4395 | enum built_in_function start_ix; |
4396 | int start_ix2; | |
389dd41b | 4397 | location_t clause_loc; |
f1f41a6c | 4398 | vec<tree, va_gc> *args; |
773c5ba7 | 4399 | |
75a70cf9 | 4400 | clauses = gimple_omp_parallel_clauses (entry_stmt); |
773c5ba7 | 4401 | |
bc7bff74 | 4402 | /* Determine what flavor of GOMP_parallel we will be |
773c5ba7 | 4403 | emitting. */ |
bc7bff74 | 4404 | start_ix = BUILT_IN_GOMP_PARALLEL; |
773c5ba7 | 4405 | if (is_combined_parallel (region)) |
4406 | { | |
61e47ac8 | 4407 | switch (region->inner->type) |
773c5ba7 | 4408 | { |
75a70cf9 | 4409 | case GIMPLE_OMP_FOR: |
fd6481cf | 4410 | gcc_assert (region->inner->sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); |
bc7bff74 | 4411 | start_ix2 = ((int)BUILT_IN_GOMP_PARALLEL_LOOP_STATIC |
b9a16870 | 4412 | + (region->inner->sched_kind |
4413 | == OMP_CLAUSE_SCHEDULE_RUNTIME | |
4414 | ? 3 : region->inner->sched_kind)); | |
4415 | start_ix = (enum built_in_function)start_ix2; | |
61e47ac8 | 4416 | break; |
75a70cf9 | 4417 | case GIMPLE_OMP_SECTIONS: |
bc7bff74 | 4418 | start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS; |
61e47ac8 | 4419 | break; |
4420 | default: | |
4421 | gcc_unreachable (); | |
773c5ba7 | 4422 | } |
773c5ba7 | 4423 | } |
1e8e9920 | 4424 | |
4425 | /* By default, the value of NUM_THREADS is zero (selected at run time) | |
4426 | and there is no conditional. */ | |
4427 | cond = NULL_TREE; | |
4428 | val = build_int_cst (unsigned_type_node, 0); | |
bc7bff74 | 4429 | flags = build_int_cst (unsigned_type_node, 0); |
1e8e9920 | 4430 | |
4431 | c = find_omp_clause (clauses, OMP_CLAUSE_IF); | |
4432 | if (c) | |
4433 | cond = OMP_CLAUSE_IF_EXPR (c); | |
4434 | ||
4435 | c = find_omp_clause (clauses, OMP_CLAUSE_NUM_THREADS); | |
4436 | if (c) | |
389dd41b | 4437 | { |
4438 | val = OMP_CLAUSE_NUM_THREADS_EXPR (c); | |
4439 | clause_loc = OMP_CLAUSE_LOCATION (c); | |
4440 | } | |
4441 | else | |
4442 | clause_loc = gimple_location (entry_stmt); | |
1e8e9920 | 4443 | |
bc7bff74 | 4444 | c = find_omp_clause (clauses, OMP_CLAUSE_PROC_BIND); |
4445 | if (c) | |
4446 | flags = build_int_cst (unsigned_type_node, OMP_CLAUSE_PROC_BIND_KIND (c)); | |
4447 | ||
1e8e9920 | 4448 | /* Ensure 'val' is of the correct type. */ |
389dd41b | 4449 | val = fold_convert_loc (clause_loc, unsigned_type_node, val); |
1e8e9920 | 4450 | |
4451 | /* If we found the clause 'if (cond)', build either | |
4452 | (cond != 0) or (cond ? val : 1u). */ | |
4453 | if (cond) | |
4454 | { | |
773c5ba7 | 4455 | cond = gimple_boolify (cond); |
4456 | ||
1e8e9920 | 4457 | if (integer_zerop (val)) |
389dd41b | 4458 | val = fold_build2_loc (clause_loc, |
4459 | EQ_EXPR, unsigned_type_node, cond, | |
79acaae1 | 4460 | build_int_cst (TREE_TYPE (cond), 0)); |
1e8e9920 | 4461 | else |
773c5ba7 | 4462 | { |
4463 | basic_block cond_bb, then_bb, else_bb; | |
79acaae1 | 4464 | edge e, e_then, e_else; |
75a70cf9 | 4465 | tree tmp_then, tmp_else, tmp_join, tmp_var; |
79acaae1 | 4466 | |
f9e245b2 | 4467 | tmp_var = create_tmp_var (TREE_TYPE (val)); |
79acaae1 | 4468 | if (gimple_in_ssa_p (cfun)) |
4469 | { | |
f9e245b2 | 4470 | tmp_then = make_ssa_name (tmp_var); |
4471 | tmp_else = make_ssa_name (tmp_var); | |
4472 | tmp_join = make_ssa_name (tmp_var); | |
79acaae1 | 4473 | } |
4474 | else | |
4475 | { | |
4476 | tmp_then = tmp_var; | |
4477 | tmp_else = tmp_var; | |
4478 | tmp_join = tmp_var; | |
4479 | } | |
773c5ba7 | 4480 | |
773c5ba7 | 4481 | e = split_block (bb, NULL); |
4482 | cond_bb = e->src; | |
4483 | bb = e->dest; | |
4484 | remove_edge (e); | |
4485 | ||
4486 | then_bb = create_empty_bb (cond_bb); | |
4487 | else_bb = create_empty_bb (then_bb); | |
79acaae1 | 4488 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
4489 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
773c5ba7 | 4490 | |
75a70cf9 | 4491 | stmt = gimple_build_cond_empty (cond); |
4492 | gsi = gsi_start_bb (cond_bb); | |
4493 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
773c5ba7 | 4494 | |
75a70cf9 | 4495 | gsi = gsi_start_bb (then_bb); |
4496 | stmt = gimple_build_assign (tmp_then, val); | |
4497 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
773c5ba7 | 4498 | |
75a70cf9 | 4499 | gsi = gsi_start_bb (else_bb); |
4500 | stmt = gimple_build_assign | |
4501 | (tmp_else, build_int_cst (unsigned_type_node, 1)); | |
4502 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
773c5ba7 | 4503 | |
4504 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
4505 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
b3083327 | 4506 | add_bb_to_loop (then_bb, cond_bb->loop_father); |
4507 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
79acaae1 | 4508 | e_then = make_edge (then_bb, bb, EDGE_FALLTHRU); |
4509 | e_else = make_edge (else_bb, bb, EDGE_FALLTHRU); | |
773c5ba7 | 4510 | |
79acaae1 | 4511 | if (gimple_in_ssa_p (cfun)) |
4512 | { | |
1a91d914 | 4513 | gphi *phi = create_phi_node (tmp_join, bb); |
60d535d2 | 4514 | add_phi_arg (phi, tmp_then, e_then, UNKNOWN_LOCATION); |
4515 | add_phi_arg (phi, tmp_else, e_else, UNKNOWN_LOCATION); | |
79acaae1 | 4516 | } |
4517 | ||
4518 | val = tmp_join; | |
773c5ba7 | 4519 | } |
4520 | ||
75a70cf9 | 4521 | gsi = gsi_start_bb (bb); |
4522 | val = force_gimple_operand_gsi (&gsi, val, true, NULL_TREE, | |
4523 | false, GSI_CONTINUE_LINKING); | |
1e8e9920 | 4524 | } |
4525 | ||
75a70cf9 | 4526 | gsi = gsi_last_bb (bb); |
4527 | t = gimple_omp_parallel_data_arg (entry_stmt); | |
1e8e9920 | 4528 | if (t == NULL) |
c2f47e15 | 4529 | t1 = null_pointer_node; |
1e8e9920 | 4530 | else |
c2f47e15 | 4531 | t1 = build_fold_addr_expr (t); |
75a70cf9 | 4532 | t2 = build_fold_addr_expr (gimple_omp_parallel_child_fn (entry_stmt)); |
773c5ba7 | 4533 | |
bc7bff74 | 4534 | vec_alloc (args, 4 + vec_safe_length (ws_args)); |
f1f41a6c | 4535 | args->quick_push (t2); |
4536 | args->quick_push (t1); | |
4537 | args->quick_push (val); | |
4538 | if (ws_args) | |
4539 | args->splice (*ws_args); | |
bc7bff74 | 4540 | args->quick_push (flags); |
414c3a2c | 4541 | |
4542 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
b9a16870 | 4543 | builtin_decl_explicit (start_ix), args); |
773c5ba7 | 4544 | |
75a70cf9 | 4545 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4546 | false, GSI_CONTINUE_LINKING); | |
1e8e9920 | 4547 | } |
4548 | ||
40750995 | 4549 | /* Insert a function call whose name is FUNC_NAME with the information from |
4550 | ENTRY_STMT into the basic_block BB. */ | |
4551 | ||
4552 | static void | |
1a91d914 | 4553 | expand_cilk_for_call (basic_block bb, gomp_parallel *entry_stmt, |
40750995 | 4554 | vec <tree, va_gc> *ws_args) |
4555 | { | |
4556 | tree t, t1, t2; | |
4557 | gimple_stmt_iterator gsi; | |
4558 | vec <tree, va_gc> *args; | |
4559 | ||
4560 | gcc_assert (vec_safe_length (ws_args) == 2); | |
4561 | tree func_name = (*ws_args)[0]; | |
4562 | tree grain = (*ws_args)[1]; | |
4563 | ||
4564 | tree clauses = gimple_omp_parallel_clauses (entry_stmt); | |
4565 | tree count = find_omp_clause (clauses, OMP_CLAUSE__CILK_FOR_COUNT_); | |
4566 | gcc_assert (count != NULL_TREE); | |
4567 | count = OMP_CLAUSE_OPERAND (count, 0); | |
4568 | ||
4569 | gsi = gsi_last_bb (bb); | |
4570 | t = gimple_omp_parallel_data_arg (entry_stmt); | |
4571 | if (t == NULL) | |
4572 | t1 = null_pointer_node; | |
4573 | else | |
4574 | t1 = build_fold_addr_expr (t); | |
4575 | t2 = build_fold_addr_expr (gimple_omp_parallel_child_fn (entry_stmt)); | |
4576 | ||
4577 | vec_alloc (args, 4); | |
4578 | args->quick_push (t2); | |
4579 | args->quick_push (t1); | |
4580 | args->quick_push (count); | |
4581 | args->quick_push (grain); | |
4582 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, func_name, args); | |
4583 | ||
4584 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, false, | |
4585 | GSI_CONTINUE_LINKING); | |
4586 | } | |
773c5ba7 | 4587 | |
fd6481cf | 4588 | /* Build the function call to GOMP_task to actually |
4589 | generate the task operation. BB is the block where to insert the code. */ | |
4590 | ||
4591 | static void | |
1a91d914 | 4592 | expand_task_call (basic_block bb, gomp_task *entry_stmt) |
fd6481cf | 4593 | { |
bc7bff74 | 4594 | tree t, t1, t2, t3, flags, cond, c, c2, clauses, depend; |
75a70cf9 | 4595 | gimple_stmt_iterator gsi; |
389dd41b | 4596 | location_t loc = gimple_location (entry_stmt); |
fd6481cf | 4597 | |
75a70cf9 | 4598 | clauses = gimple_omp_task_clauses (entry_stmt); |
fd6481cf | 4599 | |
fd6481cf | 4600 | c = find_omp_clause (clauses, OMP_CLAUSE_IF); |
4601 | if (c) | |
4602 | cond = gimple_boolify (OMP_CLAUSE_IF_EXPR (c)); | |
4603 | else | |
4604 | cond = boolean_true_node; | |
4605 | ||
4606 | c = find_omp_clause (clauses, OMP_CLAUSE_UNTIED); | |
2169f33b | 4607 | c2 = find_omp_clause (clauses, OMP_CLAUSE_MERGEABLE); |
bc7bff74 | 4608 | depend = find_omp_clause (clauses, OMP_CLAUSE_DEPEND); |
2169f33b | 4609 | flags = build_int_cst (unsigned_type_node, |
bc7bff74 | 4610 | (c ? 1 : 0) + (c2 ? 4 : 0) + (depend ? 8 : 0)); |
2169f33b | 4611 | |
4612 | c = find_omp_clause (clauses, OMP_CLAUSE_FINAL); | |
4613 | if (c) | |
4614 | { | |
4615 | c = gimple_boolify (OMP_CLAUSE_FINAL_EXPR (c)); | |
4616 | c = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, c, | |
4617 | build_int_cst (unsigned_type_node, 2), | |
4618 | build_int_cst (unsigned_type_node, 0)); | |
4619 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, flags, c); | |
4620 | } | |
bc7bff74 | 4621 | if (depend) |
4622 | depend = OMP_CLAUSE_DECL (depend); | |
4623 | else | |
4624 | depend = build_int_cst (ptr_type_node, 0); | |
fd6481cf | 4625 | |
75a70cf9 | 4626 | gsi = gsi_last_bb (bb); |
4627 | t = gimple_omp_task_data_arg (entry_stmt); | |
fd6481cf | 4628 | if (t == NULL) |
4629 | t2 = null_pointer_node; | |
4630 | else | |
389dd41b | 4631 | t2 = build_fold_addr_expr_loc (loc, t); |
4632 | t1 = build_fold_addr_expr_loc (loc, gimple_omp_task_child_fn (entry_stmt)); | |
75a70cf9 | 4633 | t = gimple_omp_task_copy_fn (entry_stmt); |
fd6481cf | 4634 | if (t == NULL) |
4635 | t3 = null_pointer_node; | |
4636 | else | |
389dd41b | 4637 | t3 = build_fold_addr_expr_loc (loc, t); |
fd6481cf | 4638 | |
b9a16870 | 4639 | t = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASK), |
bc7bff74 | 4640 | 8, t1, t2, t3, |
75a70cf9 | 4641 | gimple_omp_task_arg_size (entry_stmt), |
bc7bff74 | 4642 | gimple_omp_task_arg_align (entry_stmt), cond, flags, |
4643 | depend); | |
fd6481cf | 4644 | |
75a70cf9 | 4645 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4646 | false, GSI_CONTINUE_LINKING); | |
fd6481cf | 4647 | } |
4648 | ||
4649 | ||
75a70cf9 | 4650 | /* If exceptions are enabled, wrap the statements in BODY in a MUST_NOT_THROW |
4651 | catch handler and return it. This prevents programs from violating the | |
4652 | structured block semantics with throws. */ | |
1e8e9920 | 4653 | |
75a70cf9 | 4654 | static gimple_seq |
4655 | maybe_catch_exception (gimple_seq body) | |
1e8e9920 | 4656 | { |
e38def9c | 4657 | gimple g; |
4658 | tree decl; | |
1e8e9920 | 4659 | |
4660 | if (!flag_exceptions) | |
75a70cf9 | 4661 | return body; |
1e8e9920 | 4662 | |
596981c8 | 4663 | if (lang_hooks.eh_protect_cleanup_actions != NULL) |
4664 | decl = lang_hooks.eh_protect_cleanup_actions (); | |
1e8e9920 | 4665 | else |
b9a16870 | 4666 | decl = builtin_decl_explicit (BUILT_IN_TRAP); |
75a70cf9 | 4667 | |
e38def9c | 4668 | g = gimple_build_eh_must_not_throw (decl); |
4669 | g = gimple_build_try (body, gimple_seq_alloc_with_stmt (g), | |
75a70cf9 | 4670 | GIMPLE_TRY_CATCH); |
1e8e9920 | 4671 | |
e38def9c | 4672 | return gimple_seq_alloc_with_stmt (g); |
1e8e9920 | 4673 | } |
4674 | ||
773c5ba7 | 4675 | /* Chain all the DECLs in LIST by their TREE_CHAIN fields. */ |
1e8e9920 | 4676 | |
773c5ba7 | 4677 | static tree |
f1f41a6c | 4678 | vec2chain (vec<tree, va_gc> *v) |
1e8e9920 | 4679 | { |
2ab2ce89 | 4680 | tree chain = NULL_TREE, t; |
4681 | unsigned ix; | |
1e8e9920 | 4682 | |
f1f41a6c | 4683 | FOR_EACH_VEC_SAFE_ELT_REVERSE (v, ix, t) |
773c5ba7 | 4684 | { |
1767a056 | 4685 | DECL_CHAIN (t) = chain; |
2ab2ce89 | 4686 | chain = t; |
773c5ba7 | 4687 | } |
1e8e9920 | 4688 | |
2ab2ce89 | 4689 | return chain; |
773c5ba7 | 4690 | } |
1e8e9920 | 4691 | |
1e8e9920 | 4692 | |
773c5ba7 | 4693 | /* Remove barriers in REGION->EXIT's block. Note that this is only |
75a70cf9 | 4694 | valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region |
4695 | is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that | |
4696 | left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be | |
773c5ba7 | 4697 | removed. */ |
1e8e9920 | 4698 | |
773c5ba7 | 4699 | static void |
4700 | remove_exit_barrier (struct omp_region *region) | |
4701 | { | |
75a70cf9 | 4702 | gimple_stmt_iterator gsi; |
773c5ba7 | 4703 | basic_block exit_bb; |
61e47ac8 | 4704 | edge_iterator ei; |
4705 | edge e; | |
75a70cf9 | 4706 | gimple stmt; |
4a04f4b4 | 4707 | int any_addressable_vars = -1; |
1e8e9920 | 4708 | |
61e47ac8 | 4709 | exit_bb = region->exit; |
1e8e9920 | 4710 | |
5056ba1a | 4711 | /* If the parallel region doesn't return, we don't have REGION->EXIT |
4712 | block at all. */ | |
4713 | if (! exit_bb) | |
4714 | return; | |
4715 | ||
75a70cf9 | 4716 | /* The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The |
4717 | workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of | |
61e47ac8 | 4718 | statements that can appear in between are extremely limited -- no |
4719 | memory operations at all. Here, we allow nothing at all, so the | |
75a70cf9 | 4720 | only thing we allow to precede this GIMPLE_OMP_RETURN is a label. */ |
4721 | gsi = gsi_last_bb (exit_bb); | |
4722 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
4723 | gsi_prev (&gsi); | |
4724 | if (!gsi_end_p (gsi) && gimple_code (gsi_stmt (gsi)) != GIMPLE_LABEL) | |
773c5ba7 | 4725 | return; |
1e8e9920 | 4726 | |
61e47ac8 | 4727 | FOR_EACH_EDGE (e, ei, exit_bb->preds) |
4728 | { | |
75a70cf9 | 4729 | gsi = gsi_last_bb (e->src); |
4730 | if (gsi_end_p (gsi)) | |
61e47ac8 | 4731 | continue; |
75a70cf9 | 4732 | stmt = gsi_stmt (gsi); |
4a04f4b4 | 4733 | if (gimple_code (stmt) == GIMPLE_OMP_RETURN |
4734 | && !gimple_omp_return_nowait_p (stmt)) | |
4735 | { | |
4736 | /* OpenMP 3.0 tasks unfortunately prevent this optimization | |
4737 | in many cases. If there could be tasks queued, the barrier | |
4738 | might be needed to let the tasks run before some local | |
4739 | variable of the parallel that the task uses as shared | |
4740 | runs out of scope. The task can be spawned either | |
4741 | from within current function (this would be easy to check) | |
4742 | or from some function it calls and gets passed an address | |
4743 | of such a variable. */ | |
4744 | if (any_addressable_vars < 0) | |
4745 | { | |
1a91d914 | 4746 | gomp_parallel *parallel_stmt |
4747 | = as_a <gomp_parallel *> (last_stmt (region->entry)); | |
4a04f4b4 | 4748 | tree child_fun = gimple_omp_parallel_child_fn (parallel_stmt); |
2ab2ce89 | 4749 | tree local_decls, block, decl; |
4750 | unsigned ix; | |
4a04f4b4 | 4751 | |
4752 | any_addressable_vars = 0; | |
2ab2ce89 | 4753 | FOR_EACH_LOCAL_DECL (DECL_STRUCT_FUNCTION (child_fun), ix, decl) |
4754 | if (TREE_ADDRESSABLE (decl)) | |
4a04f4b4 | 4755 | { |
4756 | any_addressable_vars = 1; | |
4757 | break; | |
4758 | } | |
4759 | for (block = gimple_block (stmt); | |
4760 | !any_addressable_vars | |
4761 | && block | |
4762 | && TREE_CODE (block) == BLOCK; | |
4763 | block = BLOCK_SUPERCONTEXT (block)) | |
4764 | { | |
4765 | for (local_decls = BLOCK_VARS (block); | |
4766 | local_decls; | |
1767a056 | 4767 | local_decls = DECL_CHAIN (local_decls)) |
4a04f4b4 | 4768 | if (TREE_ADDRESSABLE (local_decls)) |
4769 | { | |
4770 | any_addressable_vars = 1; | |
4771 | break; | |
4772 | } | |
4773 | if (block == gimple_block (parallel_stmt)) | |
4774 | break; | |
4775 | } | |
4776 | } | |
4777 | if (!any_addressable_vars) | |
4778 | gimple_omp_return_set_nowait (stmt); | |
4779 | } | |
61e47ac8 | 4780 | } |
1e8e9920 | 4781 | } |
4782 | ||
61e47ac8 | 4783 | static void |
4784 | remove_exit_barriers (struct omp_region *region) | |
4785 | { | |
75a70cf9 | 4786 | if (region->type == GIMPLE_OMP_PARALLEL) |
61e47ac8 | 4787 | remove_exit_barrier (region); |
4788 | ||
4789 | if (region->inner) | |
4790 | { | |
4791 | region = region->inner; | |
4792 | remove_exit_barriers (region); | |
4793 | while (region->next) | |
4794 | { | |
4795 | region = region->next; | |
4796 | remove_exit_barriers (region); | |
4797 | } | |
4798 | } | |
4799 | } | |
773c5ba7 | 4800 | |
658b4427 | 4801 | /* Optimize omp_get_thread_num () and omp_get_num_threads () |
4802 | calls. These can't be declared as const functions, but | |
4803 | within one parallel body they are constant, so they can be | |
4804 | transformed there into __builtin_omp_get_{thread_num,num_threads} () | |
fd6481cf | 4805 | which are declared const. Similarly for task body, except |
4806 | that in untied task omp_get_thread_num () can change at any task | |
4807 | scheduling point. */ | |
658b4427 | 4808 | |
4809 | static void | |
75a70cf9 | 4810 | optimize_omp_library_calls (gimple entry_stmt) |
658b4427 | 4811 | { |
4812 | basic_block bb; | |
75a70cf9 | 4813 | gimple_stmt_iterator gsi; |
b9a16870 | 4814 | tree thr_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); |
4815 | tree thr_num_id = DECL_ASSEMBLER_NAME (thr_num_tree); | |
4816 | tree num_thr_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
4817 | tree num_thr_id = DECL_ASSEMBLER_NAME (num_thr_tree); | |
75a70cf9 | 4818 | bool untied_task = (gimple_code (entry_stmt) == GIMPLE_OMP_TASK |
4819 | && find_omp_clause (gimple_omp_task_clauses (entry_stmt), | |
fd6481cf | 4820 | OMP_CLAUSE_UNTIED) != NULL); |
658b4427 | 4821 | |
fc00614f | 4822 | FOR_EACH_BB_FN (bb, cfun) |
75a70cf9 | 4823 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
658b4427 | 4824 | { |
75a70cf9 | 4825 | gimple call = gsi_stmt (gsi); |
658b4427 | 4826 | tree decl; |
4827 | ||
75a70cf9 | 4828 | if (is_gimple_call (call) |
4829 | && (decl = gimple_call_fndecl (call)) | |
658b4427 | 4830 | && DECL_EXTERNAL (decl) |
4831 | && TREE_PUBLIC (decl) | |
4832 | && DECL_INITIAL (decl) == NULL) | |
4833 | { | |
4834 | tree built_in; | |
4835 | ||
4836 | if (DECL_NAME (decl) == thr_num_id) | |
fd6481cf | 4837 | { |
4838 | /* In #pragma omp task untied omp_get_thread_num () can change | |
4839 | during the execution of the task region. */ | |
4840 | if (untied_task) | |
4841 | continue; | |
b9a16870 | 4842 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); |
fd6481cf | 4843 | } |
658b4427 | 4844 | else if (DECL_NAME (decl) == num_thr_id) |
b9a16870 | 4845 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); |
658b4427 | 4846 | else |
4847 | continue; | |
4848 | ||
4849 | if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in) | |
75a70cf9 | 4850 | || gimple_call_num_args (call) != 0) |
658b4427 | 4851 | continue; |
4852 | ||
4853 | if (flag_exceptions && !TREE_NOTHROW (decl)) | |
4854 | continue; | |
4855 | ||
4856 | if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE | |
1ea6a73c | 4857 | || !types_compatible_p (TREE_TYPE (TREE_TYPE (decl)), |
4858 | TREE_TYPE (TREE_TYPE (built_in)))) | |
658b4427 | 4859 | continue; |
4860 | ||
0acacf9e | 4861 | gimple_call_set_fndecl (call, built_in); |
658b4427 | 4862 | } |
4863 | } | |
4864 | } | |
4865 | ||
8e6b4515 | 4866 | /* Callback for expand_omp_build_assign. Return non-NULL if *tp needs to be |
4867 | regimplified. */ | |
4868 | ||
4869 | static tree | |
4870 | expand_omp_regimplify_p (tree *tp, int *walk_subtrees, void *) | |
4871 | { | |
4872 | tree t = *tp; | |
4873 | ||
4874 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ | |
4875 | if (TREE_CODE (t) == VAR_DECL && DECL_HAS_VALUE_EXPR_P (t)) | |
4876 | return t; | |
4877 | ||
4878 | if (TREE_CODE (t) == ADDR_EXPR) | |
4879 | recompute_tree_invariant_for_addr_expr (t); | |
4880 | ||
4881 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); | |
4882 | return NULL_TREE; | |
4883 | } | |
4884 | ||
3d483a94 | 4885 | /* Prepend TO = FROM assignment before *GSI_P. */ |
4886 | ||
4887 | static void | |
4888 | expand_omp_build_assign (gimple_stmt_iterator *gsi_p, tree to, tree from) | |
4889 | { | |
4890 | bool simple_p = DECL_P (to) && TREE_ADDRESSABLE (to); | |
4891 | from = force_gimple_operand_gsi (gsi_p, from, simple_p, NULL_TREE, | |
4892 | true, GSI_SAME_STMT); | |
4893 | gimple stmt = gimple_build_assign (to, from); | |
4894 | gsi_insert_before (gsi_p, stmt, GSI_SAME_STMT); | |
4895 | if (walk_tree (&from, expand_omp_regimplify_p, NULL, NULL) | |
4896 | || walk_tree (&to, expand_omp_regimplify_p, NULL, NULL)) | |
4897 | { | |
4898 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
4899 | gimple_regimplify_operands (stmt, &gsi); | |
4900 | } | |
4901 | } | |
4902 | ||
fd6481cf | 4903 | /* Expand the OpenMP parallel or task directive starting at REGION. */ |
1e8e9920 | 4904 | |
4905 | static void | |
fd6481cf | 4906 | expand_omp_taskreg (struct omp_region *region) |
1e8e9920 | 4907 | { |
773c5ba7 | 4908 | basic_block entry_bb, exit_bb, new_bb; |
87d4aa85 | 4909 | struct function *child_cfun; |
414c3a2c | 4910 | tree child_fn, block, t; |
75a70cf9 | 4911 | gimple_stmt_iterator gsi; |
4912 | gimple entry_stmt, stmt; | |
773c5ba7 | 4913 | edge e; |
f1f41a6c | 4914 | vec<tree, va_gc> *ws_args; |
773c5ba7 | 4915 | |
61e47ac8 | 4916 | entry_stmt = last_stmt (region->entry); |
75a70cf9 | 4917 | child_fn = gimple_omp_taskreg_child_fn (entry_stmt); |
773c5ba7 | 4918 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); |
773c5ba7 | 4919 | |
61e47ac8 | 4920 | entry_bb = region->entry; |
4921 | exit_bb = region->exit; | |
773c5ba7 | 4922 | |
40750995 | 4923 | bool is_cilk_for |
4924 | = (flag_cilkplus | |
4925 | && gimple_code (entry_stmt) == GIMPLE_OMP_PARALLEL | |
4926 | && find_omp_clause (gimple_omp_parallel_clauses (entry_stmt), | |
4927 | OMP_CLAUSE__CILK_FOR_COUNT_) != NULL_TREE); | |
4928 | ||
4929 | if (is_cilk_for) | |
4930 | /* If it is a _Cilk_for statement, it is modelled *like* a parallel for, | |
4931 | and the inner statement contains the name of the built-in function | |
4932 | and grain. */ | |
4933 | ws_args = region->inner->ws_args; | |
4934 | else if (is_combined_parallel (region)) | |
61e47ac8 | 4935 | ws_args = region->ws_args; |
773c5ba7 | 4936 | else |
414c3a2c | 4937 | ws_args = NULL; |
1e8e9920 | 4938 | |
61e47ac8 | 4939 | if (child_cfun->cfg) |
1e8e9920 | 4940 | { |
773c5ba7 | 4941 | /* Due to inlining, it may happen that we have already outlined |
4942 | the region, in which case all we need to do is make the | |
4943 | sub-graph unreachable and emit the parallel call. */ | |
4944 | edge entry_succ_e, exit_succ_e; | |
773c5ba7 | 4945 | |
4946 | entry_succ_e = single_succ_edge (entry_bb); | |
773c5ba7 | 4947 | |
75a70cf9 | 4948 | gsi = gsi_last_bb (entry_bb); |
4949 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_PARALLEL | |
4950 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK); | |
4951 | gsi_remove (&gsi, true); | |
773c5ba7 | 4952 | |
4953 | new_bb = entry_bb; | |
03ed154b | 4954 | if (exit_bb) |
4955 | { | |
4956 | exit_succ_e = single_succ_edge (exit_bb); | |
4957 | make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU); | |
4958 | } | |
79acaae1 | 4959 | remove_edge_and_dominated_blocks (entry_succ_e); |
1e8e9920 | 4960 | } |
773c5ba7 | 4961 | else |
4962 | { | |
501bdd19 | 4963 | unsigned srcidx, dstidx, num; |
2ab2ce89 | 4964 | |
773c5ba7 | 4965 | /* If the parallel region needs data sent from the parent |
3480139d | 4966 | function, then the very first statement (except possible |
4967 | tree profile counter updates) of the parallel body | |
773c5ba7 | 4968 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since |
4969 | &.OMP_DATA_O is passed as an argument to the child function, | |
4970 | we need to replace it with the argument as seen by the child | |
4971 | function. | |
4972 | ||
4973 | In most cases, this will end up being the identity assignment | |
4974 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had | |
4975 | a function call that has been inlined, the original PARM_DECL | |
4976 | .OMP_DATA_I may have been converted into a different local | |
4977 | variable. In which case, we need to keep the assignment. */ | |
75a70cf9 | 4978 | if (gimple_omp_taskreg_data_arg (entry_stmt)) |
773c5ba7 | 4979 | { |
4980 | basic_block entry_succ_bb = single_succ (entry_bb); | |
75a70cf9 | 4981 | tree arg, narg; |
4982 | gimple parcopy_stmt = NULL; | |
1e8e9920 | 4983 | |
75a70cf9 | 4984 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) |
3480139d | 4985 | { |
75a70cf9 | 4986 | gimple stmt; |
3480139d | 4987 | |
75a70cf9 | 4988 | gcc_assert (!gsi_end_p (gsi)); |
4989 | stmt = gsi_stmt (gsi); | |
4990 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
cc6b725b | 4991 | continue; |
4992 | ||
75a70cf9 | 4993 | if (gimple_num_ops (stmt) == 2) |
3480139d | 4994 | { |
75a70cf9 | 4995 | tree arg = gimple_assign_rhs1 (stmt); |
4996 | ||
4997 | /* We're ignore the subcode because we're | |
4998 | effectively doing a STRIP_NOPS. */ | |
4999 | ||
5000 | if (TREE_CODE (arg) == ADDR_EXPR | |
5001 | && TREE_OPERAND (arg, 0) | |
5002 | == gimple_omp_taskreg_data_arg (entry_stmt)) | |
5003 | { | |
5004 | parcopy_stmt = stmt; | |
5005 | break; | |
5006 | } | |
3480139d | 5007 | } |
5008 | } | |
79acaae1 | 5009 | |
75a70cf9 | 5010 | gcc_assert (parcopy_stmt != NULL); |
79acaae1 | 5011 | arg = DECL_ARGUMENTS (child_fn); |
5012 | ||
5013 | if (!gimple_in_ssa_p (cfun)) | |
5014 | { | |
75a70cf9 | 5015 | if (gimple_assign_lhs (parcopy_stmt) == arg) |
5016 | gsi_remove (&gsi, true); | |
79acaae1 | 5017 | else |
75a70cf9 | 5018 | { |
5019 | /* ?? Is setting the subcode really necessary ?? */ | |
5020 | gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (arg)); | |
5021 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
5022 | } | |
79acaae1 | 5023 | } |
5024 | else | |
5025 | { | |
5026 | /* If we are in ssa form, we must load the value from the default | |
5027 | definition of the argument. That should not be defined now, | |
5028 | since the argument is not used uninitialized. */ | |
c6dfe037 | 5029 | gcc_assert (ssa_default_def (cfun, arg) == NULL); |
75a70cf9 | 5030 | narg = make_ssa_name (arg, gimple_build_nop ()); |
c6dfe037 | 5031 | set_ssa_default_def (cfun, arg, narg); |
75a70cf9 | 5032 | /* ?? Is setting the subcode really necessary ?? */ |
5033 | gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (narg)); | |
5034 | gimple_assign_set_rhs1 (parcopy_stmt, narg); | |
79acaae1 | 5035 | update_stmt (parcopy_stmt); |
5036 | } | |
773c5ba7 | 5037 | } |
5038 | ||
5039 | /* Declare local variables needed in CHILD_CFUN. */ | |
5040 | block = DECL_INITIAL (child_fn); | |
2ab2ce89 | 5041 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); |
e1a7ccb9 | 5042 | /* The gimplifier could record temporaries in parallel/task block |
5043 | rather than in containing function's local_decls chain, | |
5044 | which would mean cgraph missed finalizing them. Do it now. */ | |
1767a056 | 5045 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) |
e1a7ccb9 | 5046 | if (TREE_CODE (t) == VAR_DECL |
5047 | && TREE_STATIC (t) | |
5048 | && !DECL_EXTERNAL (t)) | |
97221fd7 | 5049 | varpool_node::finalize_decl (t); |
75a70cf9 | 5050 | DECL_SAVED_TREE (child_fn) = NULL; |
e3a19533 | 5051 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ |
5052 | gimple_set_body (child_fn, NULL); | |
1d22f541 | 5053 | TREE_USED (block) = 1; |
773c5ba7 | 5054 | |
79acaae1 | 5055 | /* Reset DECL_CONTEXT on function arguments. */ |
1767a056 | 5056 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) |
773c5ba7 | 5057 | DECL_CONTEXT (t) = child_fn; |
5058 | ||
75a70cf9 | 5059 | /* Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK, |
5060 | so that it can be moved to the child function. */ | |
5061 | gsi = gsi_last_bb (entry_bb); | |
5062 | stmt = gsi_stmt (gsi); | |
5063 | gcc_assert (stmt && (gimple_code (stmt) == GIMPLE_OMP_PARALLEL | |
5064 | || gimple_code (stmt) == GIMPLE_OMP_TASK)); | |
5065 | gsi_remove (&gsi, true); | |
5066 | e = split_block (entry_bb, stmt); | |
773c5ba7 | 5067 | entry_bb = e->dest; |
5068 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
5069 | ||
75a70cf9 | 5070 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ |
5056ba1a | 5071 | if (exit_bb) |
5072 | { | |
75a70cf9 | 5073 | gsi = gsi_last_bb (exit_bb); |
5074 | gcc_assert (!gsi_end_p (gsi) | |
5075 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
5076 | stmt = gimple_build_return (NULL); | |
5077 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
5078 | gsi_remove (&gsi, true); | |
5056ba1a | 5079 | } |
79acaae1 | 5080 | |
5081 | /* Move the parallel region into CHILD_CFUN. */ | |
48e1416a | 5082 | |
79acaae1 | 5083 | if (gimple_in_ssa_p (cfun)) |
5084 | { | |
bcaa2770 | 5085 | init_tree_ssa (child_cfun); |
5084b2e4 | 5086 | init_ssa_operands (child_cfun); |
5087 | child_cfun->gimple_df->in_ssa_p = true; | |
1d22f541 | 5088 | block = NULL_TREE; |
79acaae1 | 5089 | } |
1d22f541 | 5090 | else |
75a70cf9 | 5091 | block = gimple_block (entry_stmt); |
1d22f541 | 5092 | |
5093 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
79acaae1 | 5094 | if (exit_bb) |
5095 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
04c2922b | 5096 | /* When the OMP expansion process cannot guarantee an up-to-date |
5097 | loop tree arrange for the child function to fixup loops. */ | |
5098 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
5099 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
79acaae1 | 5100 | |
1d22f541 | 5101 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ |
f1f41a6c | 5102 | num = vec_safe_length (child_cfun->local_decls); |
501bdd19 | 5103 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) |
5104 | { | |
f1f41a6c | 5105 | t = (*child_cfun->local_decls)[srcidx]; |
501bdd19 | 5106 | if (DECL_CONTEXT (t) == cfun->decl) |
5107 | continue; | |
5108 | if (srcidx != dstidx) | |
f1f41a6c | 5109 | (*child_cfun->local_decls)[dstidx] = t; |
501bdd19 | 5110 | dstidx++; |
5111 | } | |
5112 | if (dstidx != num) | |
f1f41a6c | 5113 | vec_safe_truncate (child_cfun->local_decls, dstidx); |
1d22f541 | 5114 | |
79acaae1 | 5115 | /* Inform the callgraph about the new function. */ |
82b40354 | 5116 | DECL_STRUCT_FUNCTION (child_fn)->curr_properties = cfun->curr_properties; |
415d1b9a | 5117 | cgraph_node::add_new_function (child_fn, true); |
79acaae1 | 5118 | |
5119 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
5120 | fixed in a following pass. */ | |
5121 | push_cfun (child_cfun); | |
658b4427 | 5122 | if (optimize) |
fd6481cf | 5123 | optimize_omp_library_calls (entry_stmt); |
35ee1c66 | 5124 | cgraph_edge::rebuild_edges (); |
fbe86b1b | 5125 | |
5126 | /* Some EH regions might become dead, see PR34608. If | |
5127 | pass_cleanup_cfg isn't the first pass to happen with the | |
5128 | new child, these dead EH edges might cause problems. | |
5129 | Clean them up now. */ | |
5130 | if (flag_exceptions) | |
5131 | { | |
5132 | basic_block bb; | |
fbe86b1b | 5133 | bool changed = false; |
5134 | ||
fc00614f | 5135 | FOR_EACH_BB_FN (bb, cfun) |
75a70cf9 | 5136 | changed |= gimple_purge_dead_eh_edges (bb); |
fbe86b1b | 5137 | if (changed) |
5138 | cleanup_tree_cfg (); | |
fbe86b1b | 5139 | } |
dd277d48 | 5140 | if (gimple_in_ssa_p (cfun)) |
5141 | update_ssa (TODO_update_ssa); | |
79acaae1 | 5142 | pop_cfun (); |
773c5ba7 | 5143 | } |
48e1416a | 5144 | |
773c5ba7 | 5145 | /* Emit a library call to launch the children threads. */ |
40750995 | 5146 | if (is_cilk_for) |
1a91d914 | 5147 | expand_cilk_for_call (new_bb, |
5148 | as_a <gomp_parallel *> (entry_stmt), ws_args); | |
40750995 | 5149 | else if (gimple_code (entry_stmt) == GIMPLE_OMP_PARALLEL) |
1a91d914 | 5150 | expand_parallel_call (region, new_bb, |
5151 | as_a <gomp_parallel *> (entry_stmt), ws_args); | |
fd6481cf | 5152 | else |
1a91d914 | 5153 | expand_task_call (new_bb, as_a <gomp_task *> (entry_stmt)); |
083152fb | 5154 | if (gimple_in_ssa_p (cfun)) |
5155 | update_ssa (TODO_update_ssa_only_virtuals); | |
1e8e9920 | 5156 | } |
5157 | ||
773c5ba7 | 5158 | |
3d483a94 | 5159 | /* Helper function for expand_omp_{for_*,simd}. If this is the outermost |
5160 | of the combined collapse > 1 loop constructs, generate code like: | |
5161 | if (__builtin_expect (N32 cond3 N31, 0)) goto ZERO_ITER_BB; | |
5162 | if (cond3 is <) | |
5163 | adj = STEP3 - 1; | |
5164 | else | |
5165 | adj = STEP3 + 1; | |
5166 | count3 = (adj + N32 - N31) / STEP3; | |
5167 | if (__builtin_expect (N22 cond2 N21, 0)) goto ZERO_ITER_BB; | |
5168 | if (cond2 is <) | |
5169 | adj = STEP2 - 1; | |
5170 | else | |
5171 | adj = STEP2 + 1; | |
5172 | count2 = (adj + N22 - N21) / STEP2; | |
5173 | if (__builtin_expect (N12 cond1 N11, 0)) goto ZERO_ITER_BB; | |
5174 | if (cond1 is <) | |
5175 | adj = STEP1 - 1; | |
5176 | else | |
5177 | adj = STEP1 + 1; | |
5178 | count1 = (adj + N12 - N11) / STEP1; | |
5179 | count = count1 * count2 * count3; | |
5180 | Furthermore, if ZERO_ITER_BB is NULL, create a BB which does: | |
5181 | count = 0; | |
bc7bff74 | 5182 | and set ZERO_ITER_BB to that bb. If this isn't the outermost |
5183 | of the combined loop constructs, just initialize COUNTS array | |
5184 | from the _looptemp_ clauses. */ | |
3d483a94 | 5185 | |
5186 | /* NOTE: It *could* be better to moosh all of the BBs together, | |
5187 | creating one larger BB with all the computation and the unexpected | |
5188 | jump at the end. I.e. | |
5189 | ||
5190 | bool zero3, zero2, zero1, zero; | |
5191 | ||
5192 | zero3 = N32 c3 N31; | |
5193 | count3 = (N32 - N31) /[cl] STEP3; | |
5194 | zero2 = N22 c2 N21; | |
5195 | count2 = (N22 - N21) /[cl] STEP2; | |
5196 | zero1 = N12 c1 N11; | |
5197 | count1 = (N12 - N11) /[cl] STEP1; | |
5198 | zero = zero3 || zero2 || zero1; | |
5199 | count = count1 * count2 * count3; | |
5200 | if (__builtin_expect(zero, false)) goto zero_iter_bb; | |
5201 | ||
5202 | After all, we expect the zero=false, and thus we expect to have to | |
5203 | evaluate all of the comparison expressions, so short-circuiting | |
5204 | oughtn't be a win. Since the condition isn't protecting a | |
5205 | denominator, we're not concerned about divide-by-zero, so we can | |
5206 | fully evaluate count even if a numerator turned out to be wrong. | |
5207 | ||
5208 | It seems like putting this all together would create much better | |
5209 | scheduling opportunities, and less pressure on the chip's branch | |
5210 | predictor. */ | |
5211 | ||
5212 | static void | |
5213 | expand_omp_for_init_counts (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
5214 | basic_block &entry_bb, tree *counts, | |
5215 | basic_block &zero_iter_bb, int &first_zero_iter, | |
5216 | basic_block &l2_dom_bb) | |
5217 | { | |
5218 | tree t, type = TREE_TYPE (fd->loop.v); | |
3d483a94 | 5219 | edge e, ne; |
5220 | int i; | |
5221 | ||
5222 | /* Collapsed loops need work for expansion into SSA form. */ | |
5223 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
5224 | ||
bc7bff74 | 5225 | if (gimple_omp_for_combined_into_p (fd->for_stmt) |
5226 | && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
5227 | { | |
5228 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
5229 | isn't supposed to be handled, as the inner loop doesn't | |
5230 | use it. */ | |
5231 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5232 | OMP_CLAUSE__LOOPTEMP_); | |
5233 | gcc_assert (innerc); | |
5234 | for (i = 0; i < fd->collapse; i++) | |
5235 | { | |
5236 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5237 | OMP_CLAUSE__LOOPTEMP_); | |
5238 | gcc_assert (innerc); | |
5239 | if (i) | |
5240 | counts[i] = OMP_CLAUSE_DECL (innerc); | |
5241 | else | |
5242 | counts[0] = NULL_TREE; | |
5243 | } | |
5244 | return; | |
5245 | } | |
5246 | ||
3d483a94 | 5247 | for (i = 0; i < fd->collapse; i++) |
5248 | { | |
5249 | tree itype = TREE_TYPE (fd->loops[i].v); | |
5250 | ||
5251 | if (SSA_VAR_P (fd->loop.n2) | |
5252 | && ((t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
5253 | fold_convert (itype, fd->loops[i].n1), | |
5254 | fold_convert (itype, fd->loops[i].n2))) | |
5255 | == NULL_TREE || !integer_onep (t))) | |
5256 | { | |
1a91d914 | 5257 | gcond *cond_stmt; |
3d483a94 | 5258 | tree n1, n2; |
5259 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
5260 | n1 = force_gimple_operand_gsi (gsi, n1, true, NULL_TREE, | |
5261 | true, GSI_SAME_STMT); | |
5262 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
5263 | n2 = force_gimple_operand_gsi (gsi, n2, true, NULL_TREE, | |
5264 | true, GSI_SAME_STMT); | |
1a91d914 | 5265 | cond_stmt = gimple_build_cond (fd->loops[i].cond_code, n1, n2, |
5266 | NULL_TREE, NULL_TREE); | |
5267 | gsi_insert_before (gsi, cond_stmt, GSI_SAME_STMT); | |
5268 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
3d483a94 | 5269 | expand_omp_regimplify_p, NULL, NULL) |
1a91d914 | 5270 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), |
3d483a94 | 5271 | expand_omp_regimplify_p, NULL, NULL)) |
5272 | { | |
1a91d914 | 5273 | *gsi = gsi_for_stmt (cond_stmt); |
5274 | gimple_regimplify_operands (cond_stmt, gsi); | |
3d483a94 | 5275 | } |
1a91d914 | 5276 | e = split_block (entry_bb, cond_stmt); |
3d483a94 | 5277 | if (zero_iter_bb == NULL) |
5278 | { | |
1a91d914 | 5279 | gassign *assign_stmt; |
3d483a94 | 5280 | first_zero_iter = i; |
5281 | zero_iter_bb = create_empty_bb (entry_bb); | |
b3083327 | 5282 | add_bb_to_loop (zero_iter_bb, entry_bb->loop_father); |
3d483a94 | 5283 | *gsi = gsi_after_labels (zero_iter_bb); |
1a91d914 | 5284 | assign_stmt = gimple_build_assign (fd->loop.n2, |
5285 | build_zero_cst (type)); | |
5286 | gsi_insert_before (gsi, assign_stmt, GSI_SAME_STMT); | |
3d483a94 | 5287 | set_immediate_dominator (CDI_DOMINATORS, zero_iter_bb, |
5288 | entry_bb); | |
5289 | } | |
5290 | ne = make_edge (entry_bb, zero_iter_bb, EDGE_FALSE_VALUE); | |
5291 | ne->probability = REG_BR_PROB_BASE / 2000 - 1; | |
5292 | e->flags = EDGE_TRUE_VALUE; | |
5293 | e->probability = REG_BR_PROB_BASE - ne->probability; | |
5294 | if (l2_dom_bb == NULL) | |
5295 | l2_dom_bb = entry_bb; | |
5296 | entry_bb = e->dest; | |
5297 | *gsi = gsi_last_bb (entry_bb); | |
5298 | } | |
5299 | ||
5300 | if (POINTER_TYPE_P (itype)) | |
5301 | itype = signed_type_for (itype); | |
5302 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
5303 | ? -1 : 1)); | |
5304 | t = fold_build2 (PLUS_EXPR, itype, | |
5305 | fold_convert (itype, fd->loops[i].step), t); | |
5306 | t = fold_build2 (PLUS_EXPR, itype, t, | |
5307 | fold_convert (itype, fd->loops[i].n2)); | |
5308 | t = fold_build2 (MINUS_EXPR, itype, t, | |
5309 | fold_convert (itype, fd->loops[i].n1)); | |
5310 | /* ?? We could probably use CEIL_DIV_EXPR instead of | |
5311 | TRUNC_DIV_EXPR and adjusting by hand. Unless we can't | |
5312 | generate the same code in the end because generically we | |
5313 | don't know that the values involved must be negative for | |
5314 | GT?? */ | |
5315 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
5316 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
5317 | fold_build1 (NEGATE_EXPR, itype, t), | |
5318 | fold_build1 (NEGATE_EXPR, itype, | |
5319 | fold_convert (itype, | |
5320 | fd->loops[i].step))); | |
5321 | else | |
5322 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
5323 | fold_convert (itype, fd->loops[i].step)); | |
5324 | t = fold_convert (type, t); | |
5325 | if (TREE_CODE (t) == INTEGER_CST) | |
5326 | counts[i] = t; | |
5327 | else | |
5328 | { | |
5329 | counts[i] = create_tmp_reg (type, ".count"); | |
5330 | expand_omp_build_assign (gsi, counts[i], t); | |
5331 | } | |
5332 | if (SSA_VAR_P (fd->loop.n2)) | |
5333 | { | |
5334 | if (i == 0) | |
5335 | t = counts[0]; | |
5336 | else | |
5337 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, counts[i]); | |
5338 | expand_omp_build_assign (gsi, fd->loop.n2, t); | |
5339 | } | |
5340 | } | |
5341 | } | |
5342 | ||
5343 | ||
5344 | /* Helper function for expand_omp_{for_*,simd}. Generate code like: | |
5345 | T = V; | |
5346 | V3 = N31 + (T % count3) * STEP3; | |
5347 | T = T / count3; | |
5348 | V2 = N21 + (T % count2) * STEP2; | |
5349 | T = T / count2; | |
5350 | V1 = N11 + T * STEP1; | |
bc7bff74 | 5351 | if this loop doesn't have an inner loop construct combined with it. |
5352 | If it does have an inner loop construct combined with it and the | |
5353 | iteration count isn't known constant, store values from counts array | |
5354 | into its _looptemp_ temporaries instead. */ | |
3d483a94 | 5355 | |
5356 | static void | |
5357 | expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
bc7bff74 | 5358 | tree *counts, gimple inner_stmt, tree startvar) |
3d483a94 | 5359 | { |
5360 | int i; | |
bc7bff74 | 5361 | if (gimple_omp_for_combined_p (fd->for_stmt)) |
5362 | { | |
5363 | /* If fd->loop.n2 is constant, then no propagation of the counts | |
5364 | is needed, they are constant. */ | |
5365 | if (TREE_CODE (fd->loop.n2) == INTEGER_CST) | |
5366 | return; | |
5367 | ||
5368 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
5369 | ? gimple_omp_parallel_clauses (inner_stmt) | |
5370 | : gimple_omp_for_clauses (inner_stmt); | |
5371 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
5372 | isn't supposed to be handled, as the inner loop doesn't | |
5373 | use it. */ | |
5374 | tree innerc = find_omp_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
5375 | gcc_assert (innerc); | |
5376 | for (i = 0; i < fd->collapse; i++) | |
5377 | { | |
5378 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5379 | OMP_CLAUSE__LOOPTEMP_); | |
5380 | gcc_assert (innerc); | |
5381 | if (i) | |
5382 | { | |
5383 | tree tem = OMP_CLAUSE_DECL (innerc); | |
5384 | tree t = fold_convert (TREE_TYPE (tem), counts[i]); | |
5385 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
5386 | false, GSI_CONTINUE_LINKING); | |
1a91d914 | 5387 | gassign *stmt = gimple_build_assign (tem, t); |
bc7bff74 | 5388 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); |
5389 | } | |
5390 | } | |
5391 | return; | |
5392 | } | |
5393 | ||
3d483a94 | 5394 | tree type = TREE_TYPE (fd->loop.v); |
5395 | tree tem = create_tmp_reg (type, ".tem"); | |
1a91d914 | 5396 | gassign *stmt = gimple_build_assign (tem, startvar); |
3d483a94 | 5397 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); |
5398 | ||
5399 | for (i = fd->collapse - 1; i >= 0; i--) | |
5400 | { | |
5401 | tree vtype = TREE_TYPE (fd->loops[i].v), itype, t; | |
5402 | itype = vtype; | |
5403 | if (POINTER_TYPE_P (vtype)) | |
5404 | itype = signed_type_for (vtype); | |
5405 | if (i != 0) | |
5406 | t = fold_build2 (TRUNC_MOD_EXPR, type, tem, counts[i]); | |
5407 | else | |
5408 | t = tem; | |
5409 | t = fold_convert (itype, t); | |
5410 | t = fold_build2 (MULT_EXPR, itype, t, | |
5411 | fold_convert (itype, fd->loops[i].step)); | |
5412 | if (POINTER_TYPE_P (vtype)) | |
5413 | t = fold_build_pointer_plus (fd->loops[i].n1, t); | |
5414 | else | |
5415 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); | |
5416 | t = force_gimple_operand_gsi (gsi, t, | |
5417 | DECL_P (fd->loops[i].v) | |
5418 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
5419 | NULL_TREE, false, | |
5420 | GSI_CONTINUE_LINKING); | |
5421 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
5422 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
5423 | if (i != 0) | |
5424 | { | |
5425 | t = fold_build2 (TRUNC_DIV_EXPR, type, tem, counts[i]); | |
5426 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
5427 | false, GSI_CONTINUE_LINKING); | |
5428 | stmt = gimple_build_assign (tem, t); | |
5429 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
5430 | } | |
5431 | } | |
5432 | } | |
5433 | ||
5434 | ||
5435 | /* Helper function for expand_omp_for_*. Generate code like: | |
5436 | L10: | |
5437 | V3 += STEP3; | |
5438 | if (V3 cond3 N32) goto BODY_BB; else goto L11; | |
5439 | L11: | |
5440 | V3 = N31; | |
5441 | V2 += STEP2; | |
5442 | if (V2 cond2 N22) goto BODY_BB; else goto L12; | |
5443 | L12: | |
5444 | V2 = N21; | |
5445 | V1 += STEP1; | |
5446 | goto BODY_BB; */ | |
5447 | ||
5448 | static basic_block | |
5449 | extract_omp_for_update_vars (struct omp_for_data *fd, basic_block cont_bb, | |
5450 | basic_block body_bb) | |
5451 | { | |
5452 | basic_block last_bb, bb, collapse_bb = NULL; | |
5453 | int i; | |
5454 | gimple_stmt_iterator gsi; | |
5455 | edge e; | |
5456 | tree t; | |
5457 | gimple stmt; | |
5458 | ||
5459 | last_bb = cont_bb; | |
5460 | for (i = fd->collapse - 1; i >= 0; i--) | |
5461 | { | |
5462 | tree vtype = TREE_TYPE (fd->loops[i].v); | |
5463 | ||
5464 | bb = create_empty_bb (last_bb); | |
b3083327 | 5465 | add_bb_to_loop (bb, last_bb->loop_father); |
3d483a94 | 5466 | gsi = gsi_start_bb (bb); |
5467 | ||
5468 | if (i < fd->collapse - 1) | |
5469 | { | |
5470 | e = make_edge (last_bb, bb, EDGE_FALSE_VALUE); | |
5471 | e->probability = REG_BR_PROB_BASE / 8; | |
5472 | ||
5473 | t = fd->loops[i + 1].n1; | |
5474 | t = force_gimple_operand_gsi (&gsi, t, | |
5475 | DECL_P (fd->loops[i + 1].v) | |
5476 | && TREE_ADDRESSABLE (fd->loops[i | |
5477 | + 1].v), | |
5478 | NULL_TREE, false, | |
5479 | GSI_CONTINUE_LINKING); | |
5480 | stmt = gimple_build_assign (fd->loops[i + 1].v, t); | |
5481 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
5482 | } | |
5483 | else | |
5484 | collapse_bb = bb; | |
5485 | ||
5486 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); | |
5487 | ||
5488 | if (POINTER_TYPE_P (vtype)) | |
5489 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); | |
5490 | else | |
5491 | t = fold_build2 (PLUS_EXPR, vtype, fd->loops[i].v, fd->loops[i].step); | |
5492 | t = force_gimple_operand_gsi (&gsi, t, | |
5493 | DECL_P (fd->loops[i].v) | |
5494 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
5495 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5496 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
5497 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
5498 | ||
5499 | if (i > 0) | |
5500 | { | |
5501 | t = fd->loops[i].n2; | |
5502 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5503 | false, GSI_CONTINUE_LINKING); | |
5504 | tree v = fd->loops[i].v; | |
5505 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
5506 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
5507 | false, GSI_CONTINUE_LINKING); | |
5508 | t = fold_build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
5509 | stmt = gimple_build_cond_empty (t); | |
5510 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
5511 | e = make_edge (bb, body_bb, EDGE_TRUE_VALUE); | |
5512 | e->probability = REG_BR_PROB_BASE * 7 / 8; | |
5513 | } | |
5514 | else | |
5515 | make_edge (bb, body_bb, EDGE_FALLTHRU); | |
5516 | last_bb = bb; | |
5517 | } | |
5518 | ||
5519 | return collapse_bb; | |
5520 | } | |
5521 | ||
5522 | ||
773c5ba7 | 5523 | /* A subroutine of expand_omp_for. Generate code for a parallel |
1e8e9920 | 5524 | loop with any schedule. Given parameters: |
5525 | ||
5526 | for (V = N1; V cond N2; V += STEP) BODY; | |
5527 | ||
5528 | where COND is "<" or ">", we generate pseudocode | |
5529 | ||
5530 | more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0); | |
773c5ba7 | 5531 | if (more) goto L0; else goto L3; |
1e8e9920 | 5532 | L0: |
5533 | V = istart0; | |
5534 | iend = iend0; | |
5535 | L1: | |
5536 | BODY; | |
5537 | V += STEP; | |
773c5ba7 | 5538 | if (V cond iend) goto L1; else goto L2; |
1e8e9920 | 5539 | L2: |
773c5ba7 | 5540 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; |
5541 | L3: | |
1e8e9920 | 5542 | |
773c5ba7 | 5543 | If this is a combined omp parallel loop, instead of the call to |
fd6481cf | 5544 | GOMP_loop_foo_start, we call GOMP_loop_foo_next. |
bc7bff74 | 5545 | If this is gimple_omp_for_combined_p loop, then instead of assigning |
5546 | V and iend in L0 we assign the first two _looptemp_ clause decls of the | |
5547 | inner GIMPLE_OMP_FOR and V += STEP; and | |
5548 | if (V cond iend) goto L1; else goto L2; are removed. | |
fd6481cf | 5549 | |
5550 | For collapsed loops, given parameters: | |
5551 | collapse(3) | |
5552 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
5553 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
5554 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
5555 | BODY; | |
5556 | ||
5557 | we generate pseudocode | |
5558 | ||
8e6b4515 | 5559 | if (__builtin_expect (N32 cond3 N31, 0)) goto Z0; |
fd6481cf | 5560 | if (cond3 is <) |
5561 | adj = STEP3 - 1; | |
5562 | else | |
5563 | adj = STEP3 + 1; | |
5564 | count3 = (adj + N32 - N31) / STEP3; | |
8e6b4515 | 5565 | if (__builtin_expect (N22 cond2 N21, 0)) goto Z0; |
fd6481cf | 5566 | if (cond2 is <) |
5567 | adj = STEP2 - 1; | |
5568 | else | |
5569 | adj = STEP2 + 1; | |
5570 | count2 = (adj + N22 - N21) / STEP2; | |
8e6b4515 | 5571 | if (__builtin_expect (N12 cond1 N11, 0)) goto Z0; |
fd6481cf | 5572 | if (cond1 is <) |
5573 | adj = STEP1 - 1; | |
5574 | else | |
5575 | adj = STEP1 + 1; | |
5576 | count1 = (adj + N12 - N11) / STEP1; | |
5577 | count = count1 * count2 * count3; | |
8e6b4515 | 5578 | goto Z1; |
5579 | Z0: | |
5580 | count = 0; | |
5581 | Z1: | |
fd6481cf | 5582 | more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0); |
5583 | if (more) goto L0; else goto L3; | |
5584 | L0: | |
5585 | V = istart0; | |
5586 | T = V; | |
5587 | V3 = N31 + (T % count3) * STEP3; | |
5588 | T = T / count3; | |
5589 | V2 = N21 + (T % count2) * STEP2; | |
5590 | T = T / count2; | |
5591 | V1 = N11 + T * STEP1; | |
5592 | iend = iend0; | |
5593 | L1: | |
5594 | BODY; | |
5595 | V += 1; | |
5596 | if (V < iend) goto L10; else goto L2; | |
5597 | L10: | |
5598 | V3 += STEP3; | |
5599 | if (V3 cond3 N32) goto L1; else goto L11; | |
5600 | L11: | |
5601 | V3 = N31; | |
5602 | V2 += STEP2; | |
5603 | if (V2 cond2 N22) goto L1; else goto L12; | |
5604 | L12: | |
5605 | V2 = N21; | |
5606 | V1 += STEP1; | |
5607 | goto L1; | |
5608 | L2: | |
5609 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
5610 | L3: | |
5611 | ||
5612 | */ | |
1e8e9920 | 5613 | |
61e47ac8 | 5614 | static void |
773c5ba7 | 5615 | expand_omp_for_generic (struct omp_region *region, |
5616 | struct omp_for_data *fd, | |
1e8e9920 | 5617 | enum built_in_function start_fn, |
bc7bff74 | 5618 | enum built_in_function next_fn, |
5619 | gimple inner_stmt) | |
1e8e9920 | 5620 | { |
75a70cf9 | 5621 | tree type, istart0, iend0, iend; |
fd6481cf | 5622 | tree t, vmain, vback, bias = NULL_TREE; |
5623 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, collapse_bb; | |
03ed154b | 5624 | basic_block l2_bb = NULL, l3_bb = NULL; |
75a70cf9 | 5625 | gimple_stmt_iterator gsi; |
1a91d914 | 5626 | gassign *assign_stmt; |
773c5ba7 | 5627 | bool in_combined_parallel = is_combined_parallel (region); |
ac6e3339 | 5628 | bool broken_loop = region->cont == NULL; |
79acaae1 | 5629 | edge e, ne; |
fd6481cf | 5630 | tree *counts = NULL; |
5631 | int i; | |
ac6e3339 | 5632 | |
5633 | gcc_assert (!broken_loop || !in_combined_parallel); | |
fd6481cf | 5634 | gcc_assert (fd->iter_type == long_integer_type_node |
5635 | || !in_combined_parallel); | |
1e8e9920 | 5636 | |
fd6481cf | 5637 | type = TREE_TYPE (fd->loop.v); |
5638 | istart0 = create_tmp_var (fd->iter_type, ".istart0"); | |
5639 | iend0 = create_tmp_var (fd->iter_type, ".iend0"); | |
6d63fc03 | 5640 | TREE_ADDRESSABLE (istart0) = 1; |
5641 | TREE_ADDRESSABLE (iend0) = 1; | |
1e8e9920 | 5642 | |
fd6481cf | 5643 | /* See if we need to bias by LLONG_MIN. */ |
5644 | if (fd->iter_type == long_long_unsigned_type_node | |
5645 | && TREE_CODE (type) == INTEGER_TYPE | |
5646 | && !TYPE_UNSIGNED (type)) | |
5647 | { | |
5648 | tree n1, n2; | |
5649 | ||
5650 | if (fd->loop.cond_code == LT_EXPR) | |
5651 | { | |
5652 | n1 = fd->loop.n1; | |
5653 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5654 | } | |
5655 | else | |
5656 | { | |
5657 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5658 | n2 = fd->loop.n1; | |
5659 | } | |
5660 | if (TREE_CODE (n1) != INTEGER_CST | |
5661 | || TREE_CODE (n2) != INTEGER_CST | |
5662 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
5663 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
5664 | } | |
5665 | ||
61e47ac8 | 5666 | entry_bb = region->entry; |
03ed154b | 5667 | cont_bb = region->cont; |
fd6481cf | 5668 | collapse_bb = NULL; |
ac6e3339 | 5669 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
5670 | gcc_assert (broken_loop | |
5671 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
5672 | l0_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
5673 | l1_bb = single_succ (l0_bb); | |
5674 | if (!broken_loop) | |
03ed154b | 5675 | { |
5676 | l2_bb = create_empty_bb (cont_bb); | |
ac6e3339 | 5677 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb); |
5678 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
03ed154b | 5679 | } |
ac6e3339 | 5680 | else |
5681 | l2_bb = NULL; | |
5682 | l3_bb = BRANCH_EDGE (entry_bb)->dest; | |
5683 | exit_bb = region->exit; | |
773c5ba7 | 5684 | |
75a70cf9 | 5685 | gsi = gsi_last_bb (entry_bb); |
fd6481cf | 5686 | |
75a70cf9 | 5687 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
fd6481cf | 5688 | if (fd->collapse > 1) |
5689 | { | |
8e6b4515 | 5690 | int first_zero_iter = -1; |
3d483a94 | 5691 | basic_block zero_iter_bb = NULL, l2_dom_bb = NULL; |
8e6b4515 | 5692 | |
3d483a94 | 5693 | counts = XALLOCAVEC (tree, fd->collapse); |
5694 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5695 | zero_iter_bb, first_zero_iter, | |
5696 | l2_dom_bb); | |
fd6481cf | 5697 | |
8e6b4515 | 5698 | if (zero_iter_bb) |
5699 | { | |
5700 | /* Some counts[i] vars might be uninitialized if | |
5701 | some loop has zero iterations. But the body shouldn't | |
5702 | be executed in that case, so just avoid uninit warnings. */ | |
5703 | for (i = first_zero_iter; i < fd->collapse; i++) | |
5704 | if (SSA_VAR_P (counts[i])) | |
5705 | TREE_NO_WARNING (counts[i]) = 1; | |
5706 | gsi_prev (&gsi); | |
5707 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
5708 | entry_bb = e->dest; | |
5709 | make_edge (zero_iter_bb, entry_bb, EDGE_FALLTHRU); | |
5710 | gsi = gsi_last_bb (entry_bb); | |
5711 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, | |
5712 | get_immediate_dominator (CDI_DOMINATORS, | |
5713 | zero_iter_bb)); | |
5714 | } | |
fd6481cf | 5715 | } |
79acaae1 | 5716 | if (in_combined_parallel) |
5717 | { | |
5718 | /* In a combined parallel loop, emit a call to | |
5719 | GOMP_loop_foo_next. */ | |
b9a16870 | 5720 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, |
79acaae1 | 5721 | build_fold_addr_expr (istart0), |
5722 | build_fold_addr_expr (iend0)); | |
5723 | } | |
5724 | else | |
1e8e9920 | 5725 | { |
c2f47e15 | 5726 | tree t0, t1, t2, t3, t4; |
773c5ba7 | 5727 | /* If this is not a combined parallel loop, emit a call to |
5728 | GOMP_loop_foo_start in ENTRY_BB. */ | |
c2f47e15 | 5729 | t4 = build_fold_addr_expr (iend0); |
5730 | t3 = build_fold_addr_expr (istart0); | |
fd6481cf | 5731 | t2 = fold_convert (fd->iter_type, fd->loop.step); |
3d483a94 | 5732 | t1 = fd->loop.n2; |
5733 | t0 = fd->loop.n1; | |
bc7bff74 | 5734 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) |
5735 | { | |
5736 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5737 | OMP_CLAUSE__LOOPTEMP_); | |
5738 | gcc_assert (innerc); | |
5739 | t0 = OMP_CLAUSE_DECL (innerc); | |
5740 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5741 | OMP_CLAUSE__LOOPTEMP_); | |
5742 | gcc_assert (innerc); | |
5743 | t1 = OMP_CLAUSE_DECL (innerc); | |
5744 | } | |
3d483a94 | 5745 | if (POINTER_TYPE_P (TREE_TYPE (t0)) |
5746 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
5747 | != TYPE_PRECISION (fd->iter_type)) | |
c799f233 | 5748 | { |
5749 | /* Avoid casting pointers to integer of a different size. */ | |
3cea8318 | 5750 | tree itype = signed_type_for (type); |
3d483a94 | 5751 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); |
5752 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
c799f233 | 5753 | } |
5754 | else | |
5755 | { | |
3d483a94 | 5756 | t1 = fold_convert (fd->iter_type, t1); |
5757 | t0 = fold_convert (fd->iter_type, t0); | |
c799f233 | 5758 | } |
fd6481cf | 5759 | if (bias) |
1e8e9920 | 5760 | { |
fd6481cf | 5761 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); |
5762 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
5763 | } | |
5764 | if (fd->iter_type == long_integer_type_node) | |
5765 | { | |
5766 | if (fd->chunk_size) | |
5767 | { | |
5768 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
b9a16870 | 5769 | t = build_call_expr (builtin_decl_explicit (start_fn), |
5770 | 6, t0, t1, t2, t, t3, t4); | |
fd6481cf | 5771 | } |
5772 | else | |
b9a16870 | 5773 | t = build_call_expr (builtin_decl_explicit (start_fn), |
5774 | 5, t0, t1, t2, t3, t4); | |
1e8e9920 | 5775 | } |
c2f47e15 | 5776 | else |
fd6481cf | 5777 | { |
5778 | tree t5; | |
5779 | tree c_bool_type; | |
b9a16870 | 5780 | tree bfn_decl; |
fd6481cf | 5781 | |
5782 | /* The GOMP_loop_ull_*start functions have additional boolean | |
5783 | argument, true for < loops and false for > loops. | |
5784 | In Fortran, the C bool type can be different from | |
5785 | boolean_type_node. */ | |
b9a16870 | 5786 | bfn_decl = builtin_decl_explicit (start_fn); |
5787 | c_bool_type = TREE_TYPE (TREE_TYPE (bfn_decl)); | |
fd6481cf | 5788 | t5 = build_int_cst (c_bool_type, |
5789 | fd->loop.cond_code == LT_EXPR ? 1 : 0); | |
5790 | if (fd->chunk_size) | |
5791 | { | |
b9a16870 | 5792 | tree bfn_decl = builtin_decl_explicit (start_fn); |
fd6481cf | 5793 | t = fold_convert (fd->iter_type, fd->chunk_size); |
b9a16870 | 5794 | t = build_call_expr (bfn_decl, 7, t5, t0, t1, t2, t, t3, t4); |
fd6481cf | 5795 | } |
5796 | else | |
b9a16870 | 5797 | t = build_call_expr (builtin_decl_explicit (start_fn), |
5798 | 6, t5, t0, t1, t2, t3, t4); | |
fd6481cf | 5799 | } |
1e8e9920 | 5800 | } |
fd6481cf | 5801 | if (TREE_TYPE (t) != boolean_type_node) |
5802 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
5803 | t, build_int_cst (TREE_TYPE (t), 0)); | |
75a70cf9 | 5804 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5805 | true, GSI_SAME_STMT); | |
5806 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
79acaae1 | 5807 | |
75a70cf9 | 5808 | /* Remove the GIMPLE_OMP_FOR statement. */ |
5809 | gsi_remove (&gsi, true); | |
1e8e9920 | 5810 | |
773c5ba7 | 5811 | /* Iteration setup for sequential loop goes in L0_BB. */ |
3d483a94 | 5812 | tree startvar = fd->loop.v; |
5813 | tree endvar = NULL_TREE; | |
5814 | ||
bc7bff74 | 5815 | if (gimple_omp_for_combined_p (fd->for_stmt)) |
5816 | { | |
5817 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_FOR | |
5818 | && gimple_omp_for_kind (inner_stmt) | |
5819 | == GF_OMP_FOR_KIND_SIMD); | |
5820 | tree innerc = find_omp_clause (gimple_omp_for_clauses (inner_stmt), | |
5821 | OMP_CLAUSE__LOOPTEMP_); | |
5822 | gcc_assert (innerc); | |
5823 | startvar = OMP_CLAUSE_DECL (innerc); | |
5824 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
5825 | OMP_CLAUSE__LOOPTEMP_); | |
5826 | gcc_assert (innerc); | |
5827 | endvar = OMP_CLAUSE_DECL (innerc); | |
5828 | } | |
5829 | ||
75a70cf9 | 5830 | gsi = gsi_start_bb (l0_bb); |
1efcacec | 5831 | t = istart0; |
fd6481cf | 5832 | if (bias) |
1efcacec | 5833 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); |
3d483a94 | 5834 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
5835 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
5836 | t = fold_convert (TREE_TYPE (startvar), t); | |
4abecb72 | 5837 | t = force_gimple_operand_gsi (&gsi, t, |
3d483a94 | 5838 | DECL_P (startvar) |
5839 | && TREE_ADDRESSABLE (startvar), | |
4abecb72 | 5840 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
1a91d914 | 5841 | assign_stmt = gimple_build_assign (startvar, t); |
5842 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
1e8e9920 | 5843 | |
1efcacec | 5844 | t = iend0; |
fd6481cf | 5845 | if (bias) |
1efcacec | 5846 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); |
3d483a94 | 5847 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
5848 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
5849 | t = fold_convert (TREE_TYPE (startvar), t); | |
75a70cf9 | 5850 | iend = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5851 | false, GSI_CONTINUE_LINKING); | |
3d483a94 | 5852 | if (endvar) |
fd6481cf | 5853 | { |
1a91d914 | 5854 | assign_stmt = gimple_build_assign (endvar, iend); |
5855 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
d44f2f7c | 5856 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (iend))) |
1a91d914 | 5857 | assign_stmt = gimple_build_assign (fd->loop.v, iend); |
d44f2f7c | 5858 | else |
1a91d914 | 5859 | assign_stmt = gimple_build_assign_with_ops (NOP_EXPR, fd->loop.v, iend); |
5860 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
fd6481cf | 5861 | } |
3d483a94 | 5862 | if (fd->collapse > 1) |
bc7bff74 | 5863 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); |
773c5ba7 | 5864 | |
ac6e3339 | 5865 | if (!broken_loop) |
03ed154b | 5866 | { |
ac6e3339 | 5867 | /* Code to control the increment and predicate for the sequential |
5868 | loop goes in the CONT_BB. */ | |
75a70cf9 | 5869 | gsi = gsi_last_bb (cont_bb); |
1a91d914 | 5870 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
5871 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
5872 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
5873 | vback = gimple_omp_continue_control_def (cont_stmt); | |
79acaae1 | 5874 | |
bc7bff74 | 5875 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
3d483a94 | 5876 | { |
5877 | if (POINTER_TYPE_P (type)) | |
5878 | t = fold_build_pointer_plus (vmain, fd->loop.step); | |
5879 | else | |
5880 | t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step); | |
5881 | t = force_gimple_operand_gsi (&gsi, t, | |
5882 | DECL_P (vback) | |
5883 | && TREE_ADDRESSABLE (vback), | |
5884 | NULL_TREE, true, GSI_SAME_STMT); | |
1a91d914 | 5885 | assign_stmt = gimple_build_assign (vback, t); |
5886 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
3d483a94 | 5887 | |
5888 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
5889 | DECL_P (vback) && TREE_ADDRESSABLE (vback) ? t : vback, | |
5890 | iend); | |
1a91d914 | 5891 | gcond *cond_stmt = gimple_build_cond_empty (t); |
5892 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
3d483a94 | 5893 | } |
773c5ba7 | 5894 | |
75a70cf9 | 5895 | /* Remove GIMPLE_OMP_CONTINUE. */ |
5896 | gsi_remove (&gsi, true); | |
773c5ba7 | 5897 | |
bc7bff74 | 5898 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) |
3d483a94 | 5899 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, l1_bb); |
fd6481cf | 5900 | |
ac6e3339 | 5901 | /* Emit code to get the next parallel iteration in L2_BB. */ |
75a70cf9 | 5902 | gsi = gsi_start_bb (l2_bb); |
773c5ba7 | 5903 | |
b9a16870 | 5904 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, |
ac6e3339 | 5905 | build_fold_addr_expr (istart0), |
5906 | build_fold_addr_expr (iend0)); | |
75a70cf9 | 5907 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5908 | false, GSI_CONTINUE_LINKING); | |
fd6481cf | 5909 | if (TREE_TYPE (t) != boolean_type_node) |
5910 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
5911 | t, build_int_cst (TREE_TYPE (t), 0)); | |
1a91d914 | 5912 | gcond *cond_stmt = gimple_build_cond_empty (t); |
5913 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
ac6e3339 | 5914 | } |
1e8e9920 | 5915 | |
61e47ac8 | 5916 | /* Add the loop cleanup function. */ |
75a70cf9 | 5917 | gsi = gsi_last_bb (exit_bb); |
5918 | if (gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
b9a16870 | 5919 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); |
bc7bff74 | 5920 | else if (gimple_omp_return_lhs (gsi_stmt (gsi))) |
5921 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
61e47ac8 | 5922 | else |
b9a16870 | 5923 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); |
1a91d914 | 5924 | gcall *call_stmt = gimple_build_call (t, 0); |
bc7bff74 | 5925 | if (gimple_omp_return_lhs (gsi_stmt (gsi))) |
1a91d914 | 5926 | gimple_call_set_lhs (call_stmt, gimple_omp_return_lhs (gsi_stmt (gsi))); |
5927 | gsi_insert_after (&gsi, call_stmt, GSI_SAME_STMT); | |
75a70cf9 | 5928 | gsi_remove (&gsi, true); |
773c5ba7 | 5929 | |
5930 | /* Connect the new blocks. */ | |
79acaae1 | 5931 | find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE; |
5932 | find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE; | |
1e8e9920 | 5933 | |
ac6e3339 | 5934 | if (!broken_loop) |
5935 | { | |
75a70cf9 | 5936 | gimple_seq phis; |
5937 | ||
79acaae1 | 5938 | e = find_edge (cont_bb, l3_bb); |
5939 | ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE); | |
5940 | ||
75a70cf9 | 5941 | phis = phi_nodes (l3_bb); |
5942 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) | |
5943 | { | |
5944 | gimple phi = gsi_stmt (gsi); | |
5945 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne), | |
5946 | PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
5947 | } | |
79acaae1 | 5948 | remove_edge (e); |
5949 | ||
ac6e3339 | 5950 | make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE); |
b3083327 | 5951 | add_bb_to_loop (l2_bb, cont_bb->loop_father); |
3d483a94 | 5952 | e = find_edge (cont_bb, l1_bb); |
bc7bff74 | 5953 | if (gimple_omp_for_combined_p (fd->for_stmt)) |
5954 | { | |
5955 | remove_edge (e); | |
5956 | e = NULL; | |
5957 | } | |
3d483a94 | 5958 | else if (fd->collapse > 1) |
fd6481cf | 5959 | { |
fd6481cf | 5960 | remove_edge (e); |
5961 | e = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
5962 | } | |
5963 | else | |
3d483a94 | 5964 | e->flags = EDGE_TRUE_VALUE; |
5965 | if (e) | |
fd6481cf | 5966 | { |
3d483a94 | 5967 | e->probability = REG_BR_PROB_BASE * 7 / 8; |
5968 | find_edge (cont_bb, l2_bb)->probability = REG_BR_PROB_BASE / 8; | |
5969 | } | |
5970 | else | |
5971 | { | |
5972 | e = find_edge (cont_bb, l2_bb); | |
5973 | e->flags = EDGE_FALLTHRU; | |
fd6481cf | 5974 | } |
ac6e3339 | 5975 | make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE); |
79acaae1 | 5976 | |
5977 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, | |
5978 | recompute_dominator (CDI_DOMINATORS, l2_bb)); | |
5979 | set_immediate_dominator (CDI_DOMINATORS, l3_bb, | |
5980 | recompute_dominator (CDI_DOMINATORS, l3_bb)); | |
5981 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, | |
5982 | recompute_dominator (CDI_DOMINATORS, l0_bb)); | |
5983 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, | |
5984 | recompute_dominator (CDI_DOMINATORS, l1_bb)); | |
04c2922b | 5985 | |
5986 | struct loop *outer_loop = alloc_loop (); | |
5987 | outer_loop->header = l0_bb; | |
5988 | outer_loop->latch = l2_bb; | |
5989 | add_loop (outer_loop, l0_bb->loop_father); | |
5990 | ||
bc7bff74 | 5991 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
3d483a94 | 5992 | { |
5993 | struct loop *loop = alloc_loop (); | |
5994 | loop->header = l1_bb; | |
5995 | /* The loop may have multiple latches. */ | |
5996 | add_loop (loop, outer_loop); | |
5997 | } | |
ac6e3339 | 5998 | } |
1e8e9920 | 5999 | } |
6000 | ||
6001 | ||
773c5ba7 | 6002 | /* A subroutine of expand_omp_for. Generate code for a parallel |
6003 | loop with static schedule and no specified chunk size. Given | |
6004 | parameters: | |
1e8e9920 | 6005 | |
6006 | for (V = N1; V cond N2; V += STEP) BODY; | |
6007 | ||
6008 | where COND is "<" or ">", we generate pseudocode | |
6009 | ||
8e6b4515 | 6010 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; |
1e8e9920 | 6011 | if (cond is <) |
6012 | adj = STEP - 1; | |
6013 | else | |
6014 | adj = STEP + 1; | |
fd6481cf | 6015 | if ((__typeof (V)) -1 > 0 && cond is >) |
6016 | n = -(adj + N2 - N1) / -STEP; | |
6017 | else | |
6018 | n = (adj + N2 - N1) / STEP; | |
1e8e9920 | 6019 | q = n / nthreads; |
31712e83 | 6020 | tt = n % nthreads; |
6021 | if (threadid < tt) goto L3; else goto L4; | |
6022 | L3: | |
6023 | tt = 0; | |
6024 | q = q + 1; | |
6025 | L4: | |
6026 | s0 = q * threadid + tt; | |
6027 | e0 = s0 + q; | |
79acaae1 | 6028 | V = s0 * STEP + N1; |
1e8e9920 | 6029 | if (s0 >= e0) goto L2; else goto L0; |
6030 | L0: | |
1e8e9920 | 6031 | e = e0 * STEP + N1; |
6032 | L1: | |
6033 | BODY; | |
6034 | V += STEP; | |
6035 | if (V cond e) goto L1; | |
1e8e9920 | 6036 | L2: |
6037 | */ | |
6038 | ||
61e47ac8 | 6039 | static void |
773c5ba7 | 6040 | expand_omp_for_static_nochunk (struct omp_region *region, |
bc7bff74 | 6041 | struct omp_for_data *fd, |
6042 | gimple inner_stmt) | |
1e8e9920 | 6043 | { |
31712e83 | 6044 | tree n, q, s0, e0, e, t, tt, nthreads, threadid; |
fd6481cf | 6045 | tree type, itype, vmain, vback; |
31712e83 | 6046 | basic_block entry_bb, second_bb, third_bb, exit_bb, seq_start_bb; |
bc7bff74 | 6047 | basic_block body_bb, cont_bb, collapse_bb = NULL; |
61e47ac8 | 6048 | basic_block fin_bb; |
75a70cf9 | 6049 | gimple_stmt_iterator gsi; |
31712e83 | 6050 | edge ep; |
bc7bff74 | 6051 | enum built_in_function get_num_threads = BUILT_IN_OMP_GET_NUM_THREADS; |
6052 | enum built_in_function get_thread_num = BUILT_IN_OMP_GET_THREAD_NUM; | |
6053 | bool broken_loop = region->cont == NULL; | |
6054 | tree *counts = NULL; | |
6055 | tree n1, n2, step; | |
1e8e9920 | 6056 | |
fd6481cf | 6057 | itype = type = TREE_TYPE (fd->loop.v); |
6058 | if (POINTER_TYPE_P (type)) | |
3cea8318 | 6059 | itype = signed_type_for (type); |
1e8e9920 | 6060 | |
61e47ac8 | 6061 | entry_bb = region->entry; |
61e47ac8 | 6062 | cont_bb = region->cont; |
ac6e3339 | 6063 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
bc7bff74 | 6064 | fin_bb = BRANCH_EDGE (entry_bb)->dest; |
6065 | gcc_assert (broken_loop | |
6066 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
ac6e3339 | 6067 | seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb)); |
6068 | body_bb = single_succ (seq_start_bb); | |
bc7bff74 | 6069 | if (!broken_loop) |
6070 | { | |
6071 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); | |
6072 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
6073 | } | |
61e47ac8 | 6074 | exit_bb = region->exit; |
6075 | ||
773c5ba7 | 6076 | /* Iteration space partitioning goes in ENTRY_BB. */ |
75a70cf9 | 6077 | gsi = gsi_last_bb (entry_bb); |
6078 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
61e47ac8 | 6079 | |
bc7bff74 | 6080 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
6081 | { | |
6082 | get_num_threads = BUILT_IN_OMP_GET_NUM_TEAMS; | |
6083 | get_thread_num = BUILT_IN_OMP_GET_TEAM_NUM; | |
6084 | } | |
6085 | ||
6086 | if (fd->collapse > 1) | |
6087 | { | |
6088 | int first_zero_iter = -1; | |
6089 | basic_block l2_dom_bb = NULL; | |
6090 | ||
6091 | counts = XALLOCAVEC (tree, fd->collapse); | |
6092 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
6093 | fin_bb, first_zero_iter, | |
6094 | l2_dom_bb); | |
6095 | t = NULL_TREE; | |
6096 | } | |
6097 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6098 | t = integer_one_node; | |
6099 | else | |
6100 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
6101 | fold_convert (type, fd->loop.n1), | |
6102 | fold_convert (type, fd->loop.n2)); | |
6103 | if (fd->collapse == 1 | |
6104 | && TYPE_UNSIGNED (type) | |
8e6b4515 | 6105 | && (t == NULL_TREE || !integer_onep (t))) |
6106 | { | |
8e6b4515 | 6107 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); |
6108 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
6109 | true, GSI_SAME_STMT); | |
6110 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
6111 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
6112 | true, GSI_SAME_STMT); | |
1a91d914 | 6113 | gcond *cond_stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, |
6114 | NULL_TREE, NULL_TREE); | |
6115 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
6116 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
8e6b4515 | 6117 | expand_omp_regimplify_p, NULL, NULL) |
1a91d914 | 6118 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), |
8e6b4515 | 6119 | expand_omp_regimplify_p, NULL, NULL)) |
6120 | { | |
1a91d914 | 6121 | gsi = gsi_for_stmt (cond_stmt); |
6122 | gimple_regimplify_operands (cond_stmt, &gsi); | |
8e6b4515 | 6123 | } |
1a91d914 | 6124 | ep = split_block (entry_bb, cond_stmt); |
8e6b4515 | 6125 | ep->flags = EDGE_TRUE_VALUE; |
6126 | entry_bb = ep->dest; | |
6127 | ep->probability = REG_BR_PROB_BASE - (REG_BR_PROB_BASE / 2000 - 1); | |
6128 | ep = make_edge (ep->src, fin_bb, EDGE_FALSE_VALUE); | |
6129 | ep->probability = REG_BR_PROB_BASE / 2000 - 1; | |
6130 | if (gimple_in_ssa_p (cfun)) | |
6131 | { | |
6132 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; | |
1a91d914 | 6133 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); |
6134 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
8e6b4515 | 6135 | { |
1a91d914 | 6136 | gphi *phi = gpi.phi (); |
8e6b4515 | 6137 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), |
6138 | ep, UNKNOWN_LOCATION); | |
6139 | } | |
6140 | } | |
6141 | gsi = gsi_last_bb (entry_bb); | |
6142 | } | |
6143 | ||
bc7bff74 | 6144 | t = build_call_expr (builtin_decl_explicit (get_num_threads), 0); |
fd6481cf | 6145 | t = fold_convert (itype, t); |
75a70cf9 | 6146 | nthreads = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6147 | true, GSI_SAME_STMT); | |
48e1416a | 6148 | |
bc7bff74 | 6149 | t = build_call_expr (builtin_decl_explicit (get_thread_num), 0); |
fd6481cf | 6150 | t = fold_convert (itype, t); |
75a70cf9 | 6151 | threadid = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6152 | true, GSI_SAME_STMT); | |
1e8e9920 | 6153 | |
bc7bff74 | 6154 | n1 = fd->loop.n1; |
6155 | n2 = fd->loop.n2; | |
6156 | step = fd->loop.step; | |
6157 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6158 | { | |
6159 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6160 | OMP_CLAUSE__LOOPTEMP_); | |
6161 | gcc_assert (innerc); | |
6162 | n1 = OMP_CLAUSE_DECL (innerc); | |
6163 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
6164 | OMP_CLAUSE__LOOPTEMP_); | |
6165 | gcc_assert (innerc); | |
6166 | n2 = OMP_CLAUSE_DECL (innerc); | |
6167 | } | |
6168 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
6169 | true, NULL_TREE, true, GSI_SAME_STMT); | |
6170 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
6171 | true, NULL_TREE, true, GSI_SAME_STMT); | |
6172 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
6173 | true, NULL_TREE, true, GSI_SAME_STMT); | |
fd6481cf | 6174 | |
6175 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
bc7bff74 | 6176 | t = fold_build2 (PLUS_EXPR, itype, step, t); |
6177 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
6178 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
fd6481cf | 6179 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) |
6180 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6181 | fold_build1 (NEGATE_EXPR, itype, t), | |
bc7bff74 | 6182 | fold_build1 (NEGATE_EXPR, itype, step)); |
fd6481cf | 6183 | else |
bc7bff74 | 6184 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
fd6481cf | 6185 | t = fold_convert (itype, t); |
75a70cf9 | 6186 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
1e8e9920 | 6187 | |
072f7ab1 | 6188 | q = create_tmp_reg (itype, "q"); |
fd6481cf | 6189 | t = fold_build2 (TRUNC_DIV_EXPR, itype, n, nthreads); |
31712e83 | 6190 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); |
6191 | gsi_insert_before (&gsi, gimple_build_assign (q, t), GSI_SAME_STMT); | |
6192 | ||
072f7ab1 | 6193 | tt = create_tmp_reg (itype, "tt"); |
31712e83 | 6194 | t = fold_build2 (TRUNC_MOD_EXPR, itype, n, nthreads); |
6195 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
6196 | gsi_insert_before (&gsi, gimple_build_assign (tt, t), GSI_SAME_STMT); | |
1e8e9920 | 6197 | |
31712e83 | 6198 | t = build2 (LT_EXPR, boolean_type_node, threadid, tt); |
1a91d914 | 6199 | gcond *cond_stmt = gimple_build_cond_empty (t); |
6200 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
31712e83 | 6201 | |
1a91d914 | 6202 | second_bb = split_block (entry_bb, cond_stmt)->dest; |
31712e83 | 6203 | gsi = gsi_last_bb (second_bb); |
6204 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
6205 | ||
6206 | gsi_insert_before (&gsi, gimple_build_assign (tt, build_int_cst (itype, 0)), | |
6207 | GSI_SAME_STMT); | |
1a91d914 | 6208 | gassign *assign_stmt |
6209 | = gimple_build_assign_with_ops (PLUS_EXPR, q, q, | |
6210 | build_int_cst (itype, 1)); | |
6211 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
31712e83 | 6212 | |
1a91d914 | 6213 | third_bb = split_block (second_bb, assign_stmt)->dest; |
31712e83 | 6214 | gsi = gsi_last_bb (third_bb); |
6215 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
1e8e9920 | 6216 | |
fd6481cf | 6217 | t = build2 (MULT_EXPR, itype, q, threadid); |
31712e83 | 6218 | t = build2 (PLUS_EXPR, itype, t, tt); |
75a70cf9 | 6219 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
1e8e9920 | 6220 | |
fd6481cf | 6221 | t = fold_build2 (PLUS_EXPR, itype, s0, q); |
75a70cf9 | 6222 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
79acaae1 | 6223 | |
1e8e9920 | 6224 | t = build2 (GE_EXPR, boolean_type_node, s0, e0); |
75a70cf9 | 6225 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); |
773c5ba7 | 6226 | |
75a70cf9 | 6227 | /* Remove the GIMPLE_OMP_FOR statement. */ |
6228 | gsi_remove (&gsi, true); | |
773c5ba7 | 6229 | |
6230 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
75a70cf9 | 6231 | gsi = gsi_start_bb (seq_start_bb); |
1e8e9920 | 6232 | |
bc7bff74 | 6233 | tree startvar = fd->loop.v; |
6234 | tree endvar = NULL_TREE; | |
6235 | ||
6236 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6237 | { | |
6238 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
6239 | ? gimple_omp_parallel_clauses (inner_stmt) | |
6240 | : gimple_omp_for_clauses (inner_stmt); | |
6241 | tree innerc = find_omp_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
6242 | gcc_assert (innerc); | |
6243 | startvar = OMP_CLAUSE_DECL (innerc); | |
6244 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
6245 | OMP_CLAUSE__LOOPTEMP_); | |
6246 | gcc_assert (innerc); | |
6247 | endvar = OMP_CLAUSE_DECL (innerc); | |
6248 | } | |
fd6481cf | 6249 | t = fold_convert (itype, s0); |
bc7bff74 | 6250 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6251 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6252 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6253 | else |
bc7bff74 | 6254 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6255 | t = fold_convert (TREE_TYPE (startvar), t); | |
4abecb72 | 6256 | t = force_gimple_operand_gsi (&gsi, t, |
bc7bff74 | 6257 | DECL_P (startvar) |
6258 | && TREE_ADDRESSABLE (startvar), | |
4abecb72 | 6259 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
1a91d914 | 6260 | assign_stmt = gimple_build_assign (startvar, t); |
6261 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
48e1416a | 6262 | |
fd6481cf | 6263 | t = fold_convert (itype, e0); |
bc7bff74 | 6264 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6265 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6266 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6267 | else |
bc7bff74 | 6268 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6269 | t = fold_convert (TREE_TYPE (startvar), t); | |
75a70cf9 | 6270 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6271 | false, GSI_CONTINUE_LINKING); | |
bc7bff74 | 6272 | if (endvar) |
6273 | { | |
1a91d914 | 6274 | assign_stmt = gimple_build_assign (endvar, e); |
6275 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
d44f2f7c | 6276 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) |
1a91d914 | 6277 | assign_stmt = gimple_build_assign (fd->loop.v, e); |
d44f2f7c | 6278 | else |
1a91d914 | 6279 | assign_stmt = gimple_build_assign_with_ops (NOP_EXPR, fd->loop.v, e); |
6280 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
bc7bff74 | 6281 | } |
6282 | if (fd->collapse > 1) | |
6283 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); | |
1e8e9920 | 6284 | |
bc7bff74 | 6285 | if (!broken_loop) |
6286 | { | |
6287 | /* The code controlling the sequential loop replaces the | |
6288 | GIMPLE_OMP_CONTINUE. */ | |
6289 | gsi = gsi_last_bb (cont_bb); | |
1a91d914 | 6290 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
6291 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
6292 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
6293 | vback = gimple_omp_continue_control_def (cont_stmt); | |
79acaae1 | 6294 | |
bc7bff74 | 6295 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
6296 | { | |
6297 | if (POINTER_TYPE_P (type)) | |
6298 | t = fold_build_pointer_plus (vmain, step); | |
6299 | else | |
6300 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
6301 | t = force_gimple_operand_gsi (&gsi, t, | |
6302 | DECL_P (vback) | |
6303 | && TREE_ADDRESSABLE (vback), | |
6304 | NULL_TREE, true, GSI_SAME_STMT); | |
1a91d914 | 6305 | assign_stmt = gimple_build_assign (vback, t); |
6306 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
79acaae1 | 6307 | |
bc7bff74 | 6308 | t = build2 (fd->loop.cond_code, boolean_type_node, |
6309 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
6310 | ? t : vback, e); | |
6311 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
6312 | } | |
1e8e9920 | 6313 | |
bc7bff74 | 6314 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ |
6315 | gsi_remove (&gsi, true); | |
6316 | ||
6317 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
6318 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, body_bb); | |
6319 | } | |
773c5ba7 | 6320 | |
75a70cf9 | 6321 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ |
6322 | gsi = gsi_last_bb (exit_bb); | |
6323 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
bc7bff74 | 6324 | { |
6325 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
6326 | gsi_insert_after (&gsi, build_omp_barrier (t), GSI_SAME_STMT); | |
6327 | } | |
75a70cf9 | 6328 | gsi_remove (&gsi, true); |
773c5ba7 | 6329 | |
6330 | /* Connect all the blocks. */ | |
31712e83 | 6331 | ep = make_edge (entry_bb, third_bb, EDGE_FALSE_VALUE); |
6332 | ep->probability = REG_BR_PROB_BASE / 4 * 3; | |
6333 | ep = find_edge (entry_bb, second_bb); | |
6334 | ep->flags = EDGE_TRUE_VALUE; | |
6335 | ep->probability = REG_BR_PROB_BASE / 4; | |
6336 | find_edge (third_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE; | |
6337 | find_edge (third_bb, fin_bb)->flags = EDGE_TRUE_VALUE; | |
79acaae1 | 6338 | |
bc7bff74 | 6339 | if (!broken_loop) |
6340 | { | |
6341 | ep = find_edge (cont_bb, body_bb); | |
6342 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6343 | { | |
6344 | remove_edge (ep); | |
6345 | ep = NULL; | |
6346 | } | |
6347 | else if (fd->collapse > 1) | |
6348 | { | |
6349 | remove_edge (ep); | |
6350 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
6351 | } | |
6352 | else | |
6353 | ep->flags = EDGE_TRUE_VALUE; | |
6354 | find_edge (cont_bb, fin_bb)->flags | |
6355 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
6356 | } | |
48e1416a | 6357 | |
31712e83 | 6358 | set_immediate_dominator (CDI_DOMINATORS, second_bb, entry_bb); |
6359 | set_immediate_dominator (CDI_DOMINATORS, third_bb, entry_bb); | |
6360 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, third_bb); | |
bc7bff74 | 6361 | |
79acaae1 | 6362 | set_immediate_dominator (CDI_DOMINATORS, body_bb, |
6363 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
6364 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
6365 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
04c2922b | 6366 | |
bc7bff74 | 6367 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) |
6368 | { | |
6369 | struct loop *loop = alloc_loop (); | |
6370 | loop->header = body_bb; | |
6371 | if (collapse_bb == NULL) | |
6372 | loop->latch = cont_bb; | |
6373 | add_loop (loop, body_bb->loop_father); | |
6374 | } | |
1e8e9920 | 6375 | } |
6376 | ||
773c5ba7 | 6377 | |
6378 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
6379 | loop with static schedule and a specified chunk size. Given | |
6380 | parameters: | |
1e8e9920 | 6381 | |
6382 | for (V = N1; V cond N2; V += STEP) BODY; | |
6383 | ||
6384 | where COND is "<" or ">", we generate pseudocode | |
6385 | ||
8e6b4515 | 6386 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; |
1e8e9920 | 6387 | if (cond is <) |
6388 | adj = STEP - 1; | |
6389 | else | |
6390 | adj = STEP + 1; | |
fd6481cf | 6391 | if ((__typeof (V)) -1 > 0 && cond is >) |
6392 | n = -(adj + N2 - N1) / -STEP; | |
6393 | else | |
6394 | n = (adj + N2 - N1) / STEP; | |
1e8e9920 | 6395 | trip = 0; |
79acaae1 | 6396 | V = threadid * CHUNK * STEP + N1; -- this extra definition of V is |
6397 | here so that V is defined | |
6398 | if the loop is not entered | |
1e8e9920 | 6399 | L0: |
6400 | s0 = (trip * nthreads + threadid) * CHUNK; | |
6401 | e0 = min(s0 + CHUNK, n); | |
6402 | if (s0 < n) goto L1; else goto L4; | |
6403 | L1: | |
6404 | V = s0 * STEP + N1; | |
6405 | e = e0 * STEP + N1; | |
6406 | L2: | |
6407 | BODY; | |
6408 | V += STEP; | |
6409 | if (V cond e) goto L2; else goto L3; | |
6410 | L3: | |
6411 | trip += 1; | |
6412 | goto L0; | |
6413 | L4: | |
1e8e9920 | 6414 | */ |
6415 | ||
61e47ac8 | 6416 | static void |
bc7bff74 | 6417 | expand_omp_for_static_chunk (struct omp_region *region, |
6418 | struct omp_for_data *fd, gimple inner_stmt) | |
1e8e9920 | 6419 | { |
75a70cf9 | 6420 | tree n, s0, e0, e, t; |
79acaae1 | 6421 | tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid; |
93481288 | 6422 | tree type, itype, vmain, vback, vextra; |
773c5ba7 | 6423 | basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb; |
bc7bff74 | 6424 | basic_block trip_update_bb = NULL, cont_bb, collapse_bb = NULL, fin_bb; |
93481288 | 6425 | gimple_stmt_iterator gsi; |
75a70cf9 | 6426 | edge se; |
bc7bff74 | 6427 | enum built_in_function get_num_threads = BUILT_IN_OMP_GET_NUM_THREADS; |
6428 | enum built_in_function get_thread_num = BUILT_IN_OMP_GET_THREAD_NUM; | |
6429 | bool broken_loop = region->cont == NULL; | |
6430 | tree *counts = NULL; | |
6431 | tree n1, n2, step; | |
1e8e9920 | 6432 | |
fd6481cf | 6433 | itype = type = TREE_TYPE (fd->loop.v); |
6434 | if (POINTER_TYPE_P (type)) | |
3cea8318 | 6435 | itype = signed_type_for (type); |
1e8e9920 | 6436 | |
61e47ac8 | 6437 | entry_bb = region->entry; |
ac6e3339 | 6438 | se = split_block (entry_bb, last_stmt (entry_bb)); |
6439 | entry_bb = se->src; | |
6440 | iter_part_bb = se->dest; | |
61e47ac8 | 6441 | cont_bb = region->cont; |
ac6e3339 | 6442 | gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2); |
bc7bff74 | 6443 | fin_bb = BRANCH_EDGE (iter_part_bb)->dest; |
6444 | gcc_assert (broken_loop | |
6445 | || fin_bb == FALLTHRU_EDGE (cont_bb)->dest); | |
ac6e3339 | 6446 | seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb)); |
6447 | body_bb = single_succ (seq_start_bb); | |
bc7bff74 | 6448 | if (!broken_loop) |
6449 | { | |
6450 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); | |
6451 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
6452 | trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb)); | |
6453 | } | |
61e47ac8 | 6454 | exit_bb = region->exit; |
773c5ba7 | 6455 | |
773c5ba7 | 6456 | /* Trip and adjustment setup goes in ENTRY_BB. */ |
93481288 | 6457 | gsi = gsi_last_bb (entry_bb); |
6458 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
773c5ba7 | 6459 | |
bc7bff74 | 6460 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
6461 | { | |
6462 | get_num_threads = BUILT_IN_OMP_GET_NUM_TEAMS; | |
6463 | get_thread_num = BUILT_IN_OMP_GET_TEAM_NUM; | |
6464 | } | |
6465 | ||
6466 | if (fd->collapse > 1) | |
6467 | { | |
6468 | int first_zero_iter = -1; | |
6469 | basic_block l2_dom_bb = NULL; | |
6470 | ||
6471 | counts = XALLOCAVEC (tree, fd->collapse); | |
93481288 | 6472 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, |
bc7bff74 | 6473 | fin_bb, first_zero_iter, |
6474 | l2_dom_bb); | |
6475 | t = NULL_TREE; | |
6476 | } | |
6477 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6478 | t = integer_one_node; | |
6479 | else | |
6480 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
6481 | fold_convert (type, fd->loop.n1), | |
6482 | fold_convert (type, fd->loop.n2)); | |
6483 | if (fd->collapse == 1 | |
6484 | && TYPE_UNSIGNED (type) | |
8e6b4515 | 6485 | && (t == NULL_TREE || !integer_onep (t))) |
6486 | { | |
8e6b4515 | 6487 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); |
93481288 | 6488 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, |
8e6b4515 | 6489 | true, GSI_SAME_STMT); |
6490 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
93481288 | 6491 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, |
8e6b4515 | 6492 | true, GSI_SAME_STMT); |
1a91d914 | 6493 | gcond *cond_stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, |
6494 | NULL_TREE, NULL_TREE); | |
6495 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
6496 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
8e6b4515 | 6497 | expand_omp_regimplify_p, NULL, NULL) |
1a91d914 | 6498 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), |
8e6b4515 | 6499 | expand_omp_regimplify_p, NULL, NULL)) |
6500 | { | |
1a91d914 | 6501 | gsi = gsi_for_stmt (cond_stmt); |
6502 | gimple_regimplify_operands (cond_stmt, &gsi); | |
8e6b4515 | 6503 | } |
1a91d914 | 6504 | se = split_block (entry_bb, cond_stmt); |
8e6b4515 | 6505 | se->flags = EDGE_TRUE_VALUE; |
6506 | entry_bb = se->dest; | |
6507 | se->probability = REG_BR_PROB_BASE - (REG_BR_PROB_BASE / 2000 - 1); | |
6508 | se = make_edge (se->src, fin_bb, EDGE_FALSE_VALUE); | |
6509 | se->probability = REG_BR_PROB_BASE / 2000 - 1; | |
6510 | if (gimple_in_ssa_p (cfun)) | |
6511 | { | |
6512 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; | |
1a91d914 | 6513 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); |
6514 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
8e6b4515 | 6515 | { |
1a91d914 | 6516 | gphi *phi = gpi.phi (); |
8e6b4515 | 6517 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), |
6518 | se, UNKNOWN_LOCATION); | |
6519 | } | |
6520 | } | |
93481288 | 6521 | gsi = gsi_last_bb (entry_bb); |
8e6b4515 | 6522 | } |
6523 | ||
bc7bff74 | 6524 | t = build_call_expr (builtin_decl_explicit (get_num_threads), 0); |
fd6481cf | 6525 | t = fold_convert (itype, t); |
93481288 | 6526 | nthreads = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
75a70cf9 | 6527 | true, GSI_SAME_STMT); |
48e1416a | 6528 | |
bc7bff74 | 6529 | t = build_call_expr (builtin_decl_explicit (get_thread_num), 0); |
fd6481cf | 6530 | t = fold_convert (itype, t); |
93481288 | 6531 | threadid = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
75a70cf9 | 6532 | true, GSI_SAME_STMT); |
79acaae1 | 6533 | |
bc7bff74 | 6534 | n1 = fd->loop.n1; |
6535 | n2 = fd->loop.n2; | |
6536 | step = fd->loop.step; | |
6537 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6538 | { | |
6539 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6540 | OMP_CLAUSE__LOOPTEMP_); | |
6541 | gcc_assert (innerc); | |
6542 | n1 = OMP_CLAUSE_DECL (innerc); | |
6543 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
6544 | OMP_CLAUSE__LOOPTEMP_); | |
6545 | gcc_assert (innerc); | |
6546 | n2 = OMP_CLAUSE_DECL (innerc); | |
6547 | } | |
93481288 | 6548 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), |
bc7bff74 | 6549 | true, NULL_TREE, true, GSI_SAME_STMT); |
93481288 | 6550 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), |
bc7bff74 | 6551 | true, NULL_TREE, true, GSI_SAME_STMT); |
93481288 | 6552 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), |
bc7bff74 | 6553 | true, NULL_TREE, true, GSI_SAME_STMT); |
79acaae1 | 6554 | fd->chunk_size |
93481288 | 6555 | = force_gimple_operand_gsi (&gsi, fold_convert (itype, fd->chunk_size), |
75a70cf9 | 6556 | true, NULL_TREE, true, GSI_SAME_STMT); |
fd6481cf | 6557 | |
6558 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
bc7bff74 | 6559 | t = fold_build2 (PLUS_EXPR, itype, step, t); |
6560 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
6561 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
fd6481cf | 6562 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) |
6563 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6564 | fold_build1 (NEGATE_EXPR, itype, t), | |
bc7bff74 | 6565 | fold_build1 (NEGATE_EXPR, itype, step)); |
fd6481cf | 6566 | else |
bc7bff74 | 6567 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
fd6481cf | 6568 | t = fold_convert (itype, t); |
93481288 | 6569 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
75a70cf9 | 6570 | true, GSI_SAME_STMT); |
79acaae1 | 6571 | |
083152fb | 6572 | trip_var = create_tmp_reg (itype, ".trip"); |
79acaae1 | 6573 | if (gimple_in_ssa_p (cfun)) |
6574 | { | |
f9e245b2 | 6575 | trip_init = make_ssa_name (trip_var); |
6576 | trip_main = make_ssa_name (trip_var); | |
6577 | trip_back = make_ssa_name (trip_var); | |
79acaae1 | 6578 | } |
1e8e9920 | 6579 | else |
79acaae1 | 6580 | { |
6581 | trip_init = trip_var; | |
6582 | trip_main = trip_var; | |
6583 | trip_back = trip_var; | |
6584 | } | |
1e8e9920 | 6585 | |
1a91d914 | 6586 | gassign *assign_stmt |
6587 | = gimple_build_assign (trip_init, build_int_cst (itype, 0)); | |
6588 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
773c5ba7 | 6589 | |
fd6481cf | 6590 | t = fold_build2 (MULT_EXPR, itype, threadid, fd->chunk_size); |
bc7bff74 | 6591 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6592 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6593 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6594 | else |
bc7bff74 | 6595 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
93481288 | 6596 | vextra = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6597 | true, GSI_SAME_STMT); | |
79acaae1 | 6598 | |
75a70cf9 | 6599 | /* Remove the GIMPLE_OMP_FOR. */ |
93481288 | 6600 | gsi_remove (&gsi, true); |
773c5ba7 | 6601 | |
6602 | /* Iteration space partitioning goes in ITER_PART_BB. */ | |
93481288 | 6603 | gsi = gsi_last_bb (iter_part_bb); |
1e8e9920 | 6604 | |
fd6481cf | 6605 | t = fold_build2 (MULT_EXPR, itype, trip_main, nthreads); |
6606 | t = fold_build2 (PLUS_EXPR, itype, t, threadid); | |
6607 | t = fold_build2 (MULT_EXPR, itype, t, fd->chunk_size); | |
93481288 | 6608 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
75a70cf9 | 6609 | false, GSI_CONTINUE_LINKING); |
1e8e9920 | 6610 | |
fd6481cf | 6611 | t = fold_build2 (PLUS_EXPR, itype, s0, fd->chunk_size); |
6612 | t = fold_build2 (MIN_EXPR, itype, t, n); | |
93481288 | 6613 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
75a70cf9 | 6614 | false, GSI_CONTINUE_LINKING); |
1e8e9920 | 6615 | |
6616 | t = build2 (LT_EXPR, boolean_type_node, s0, n); | |
93481288 | 6617 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_CONTINUE_LINKING); |
773c5ba7 | 6618 | |
6619 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
93481288 | 6620 | gsi = gsi_start_bb (seq_start_bb); |
1e8e9920 | 6621 | |
bc7bff74 | 6622 | tree startvar = fd->loop.v; |
6623 | tree endvar = NULL_TREE; | |
6624 | ||
6625 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6626 | { | |
6627 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
6628 | ? gimple_omp_parallel_clauses (inner_stmt) | |
6629 | : gimple_omp_for_clauses (inner_stmt); | |
6630 | tree innerc = find_omp_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
6631 | gcc_assert (innerc); | |
6632 | startvar = OMP_CLAUSE_DECL (innerc); | |
6633 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
6634 | OMP_CLAUSE__LOOPTEMP_); | |
6635 | gcc_assert (innerc); | |
6636 | endvar = OMP_CLAUSE_DECL (innerc); | |
6637 | } | |
6638 | ||
fd6481cf | 6639 | t = fold_convert (itype, s0); |
bc7bff74 | 6640 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6641 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6642 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6643 | else |
bc7bff74 | 6644 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6645 | t = fold_convert (TREE_TYPE (startvar), t); | |
93481288 | 6646 | t = force_gimple_operand_gsi (&gsi, t, |
bc7bff74 | 6647 | DECL_P (startvar) |
6648 | && TREE_ADDRESSABLE (startvar), | |
4abecb72 | 6649 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
1a91d914 | 6650 | assign_stmt = gimple_build_assign (startvar, t); |
6651 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
1e8e9920 | 6652 | |
fd6481cf | 6653 | t = fold_convert (itype, e0); |
bc7bff74 | 6654 | t = fold_build2 (MULT_EXPR, itype, t, step); |
fd6481cf | 6655 | if (POINTER_TYPE_P (type)) |
bc7bff74 | 6656 | t = fold_build_pointer_plus (n1, t); |
fd6481cf | 6657 | else |
bc7bff74 | 6658 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6659 | t = fold_convert (TREE_TYPE (startvar), t); | |
93481288 | 6660 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
75a70cf9 | 6661 | false, GSI_CONTINUE_LINKING); |
bc7bff74 | 6662 | if (endvar) |
6663 | { | |
1a91d914 | 6664 | assign_stmt = gimple_build_assign (endvar, e); |
6665 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
d44f2f7c | 6666 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) |
1a91d914 | 6667 | assign_stmt = gimple_build_assign (fd->loop.v, e); |
d44f2f7c | 6668 | else |
1a91d914 | 6669 | assign_stmt = gimple_build_assign_with_ops (NOP_EXPR, fd->loop.v, e); |
6670 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
bc7bff74 | 6671 | } |
6672 | if (fd->collapse > 1) | |
93481288 | 6673 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); |
bc7bff74 | 6674 | |
6675 | if (!broken_loop) | |
6676 | { | |
6677 | /* The code controlling the sequential loop goes in CONT_BB, | |
6678 | replacing the GIMPLE_OMP_CONTINUE. */ | |
93481288 | 6679 | gsi = gsi_last_bb (cont_bb); |
1a91d914 | 6680 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
6681 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
6682 | vback = gimple_omp_continue_control_def (cont_stmt); | |
1e8e9920 | 6683 | |
bc7bff74 | 6684 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
6685 | { | |
6686 | if (POINTER_TYPE_P (type)) | |
93481288 | 6687 | t = fold_build_pointer_plus (vmain, step); |
bc7bff74 | 6688 | else |
93481288 | 6689 | t = fold_build2 (PLUS_EXPR, type, vmain, step); |
6690 | if (DECL_P (vback) && TREE_ADDRESSABLE (vback)) | |
6691 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
bc7bff74 | 6692 | true, GSI_SAME_STMT); |
1a91d914 | 6693 | assign_stmt = gimple_build_assign (vback, t); |
6694 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
79acaae1 | 6695 | |
bc7bff74 | 6696 | t = build2 (fd->loop.cond_code, boolean_type_node, |
93481288 | 6697 | DECL_P (vback) && TREE_ADDRESSABLE (vback) |
6698 | ? t : vback, e); | |
6699 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
bc7bff74 | 6700 | } |
79acaae1 | 6701 | |
bc7bff74 | 6702 | /* Remove GIMPLE_OMP_CONTINUE. */ |
93481288 | 6703 | gsi_remove (&gsi, true); |
48e1416a | 6704 | |
bc7bff74 | 6705 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) |
6706 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, body_bb); | |
773c5ba7 | 6707 | |
bc7bff74 | 6708 | /* Trip update code goes into TRIP_UPDATE_BB. */ |
93481288 | 6709 | gsi = gsi_start_bb (trip_update_bb); |
1e8e9920 | 6710 | |
bc7bff74 | 6711 | t = build_int_cst (itype, 1); |
6712 | t = build2 (PLUS_EXPR, itype, trip_main, t); | |
1a91d914 | 6713 | assign_stmt = gimple_build_assign (trip_back, t); |
6714 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
bc7bff74 | 6715 | } |
1e8e9920 | 6716 | |
75a70cf9 | 6717 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ |
93481288 | 6718 | gsi = gsi_last_bb (exit_bb); |
6719 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
bc7bff74 | 6720 | { |
93481288 | 6721 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); |
6722 | gsi_insert_after (&gsi, build_omp_barrier (t), GSI_SAME_STMT); | |
bc7bff74 | 6723 | } |
93481288 | 6724 | gsi_remove (&gsi, true); |
1e8e9920 | 6725 | |
773c5ba7 | 6726 | /* Connect the new blocks. */ |
ac6e3339 | 6727 | find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE; |
6728 | find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE; | |
79acaae1 | 6729 | |
bc7bff74 | 6730 | if (!broken_loop) |
6731 | { | |
6732 | se = find_edge (cont_bb, body_bb); | |
6733 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6734 | { | |
6735 | remove_edge (se); | |
6736 | se = NULL; | |
6737 | } | |
6738 | else if (fd->collapse > 1) | |
6739 | { | |
6740 | remove_edge (se); | |
6741 | se = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
6742 | } | |
6743 | else | |
6744 | se->flags = EDGE_TRUE_VALUE; | |
6745 | find_edge (cont_bb, trip_update_bb)->flags | |
6746 | = se ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
79acaae1 | 6747 | |
bc7bff74 | 6748 | redirect_edge_and_branch (single_succ_edge (trip_update_bb), iter_part_bb); |
6749 | } | |
79acaae1 | 6750 | |
6751 | if (gimple_in_ssa_p (cfun)) | |
6752 | { | |
1a91d914 | 6753 | gphi_iterator psi; |
6754 | gphi *phi; | |
75a70cf9 | 6755 | edge re, ene; |
75a70cf9 | 6756 | edge_var_map *vm; |
6757 | size_t i; | |
6758 | ||
bc7bff74 | 6759 | gcc_assert (fd->collapse == 1 && !broken_loop); |
6760 | ||
79acaae1 | 6761 | /* When we redirect the edge from trip_update_bb to iter_part_bb, we |
6762 | remove arguments of the phi nodes in fin_bb. We need to create | |
6763 | appropriate phi nodes in iter_part_bb instead. */ | |
6764 | se = single_pred_edge (fin_bb); | |
6765 | re = single_succ_edge (trip_update_bb); | |
06ecf488 | 6766 | vec<edge_var_map> *head = redirect_edge_var_map_vector (re); |
79acaae1 | 6767 | ene = single_succ_edge (entry_bb); |
6768 | ||
75a70cf9 | 6769 | psi = gsi_start_phis (fin_bb); |
f1f41a6c | 6770 | for (i = 0; !gsi_end_p (psi) && head->iterate (i, &vm); |
75a70cf9 | 6771 | gsi_next (&psi), ++i) |
79acaae1 | 6772 | { |
1a91d914 | 6773 | gphi *nphi; |
efbcb6de | 6774 | source_location locus; |
75a70cf9 | 6775 | |
1a91d914 | 6776 | phi = psi.phi (); |
75a70cf9 | 6777 | t = gimple_phi_result (phi); |
6778 | gcc_assert (t == redirect_edge_var_map_result (vm)); | |
79acaae1 | 6779 | nphi = create_phi_node (t, iter_part_bb); |
79acaae1 | 6780 | |
6781 | t = PHI_ARG_DEF_FROM_EDGE (phi, se); | |
efbcb6de | 6782 | locus = gimple_phi_arg_location_from_edge (phi, se); |
6783 | ||
fd6481cf | 6784 | /* A special case -- fd->loop.v is not yet computed in |
93481288 | 6785 | iter_part_bb, we need to use vextra instead. */ |
fd6481cf | 6786 | if (t == fd->loop.v) |
93481288 | 6787 | t = vextra; |
60d535d2 | 6788 | add_phi_arg (nphi, t, ene, locus); |
efbcb6de | 6789 | locus = redirect_edge_var_map_location (vm); |
60d535d2 | 6790 | add_phi_arg (nphi, redirect_edge_var_map_def (vm), re, locus); |
75a70cf9 | 6791 | } |
f1f41a6c | 6792 | gcc_assert (!gsi_end_p (psi) && i == head->length ()); |
75a70cf9 | 6793 | redirect_edge_var_map_clear (re); |
6794 | while (1) | |
6795 | { | |
6796 | psi = gsi_start_phis (fin_bb); | |
6797 | if (gsi_end_p (psi)) | |
6798 | break; | |
6799 | remove_phi_node (&psi, false); | |
79acaae1 | 6800 | } |
79acaae1 | 6801 | |
6802 | /* Make phi node for trip. */ | |
6803 | phi = create_phi_node (trip_main, iter_part_bb); | |
efbcb6de | 6804 | add_phi_arg (phi, trip_back, single_succ_edge (trip_update_bb), |
60d535d2 | 6805 | UNKNOWN_LOCATION); |
efbcb6de | 6806 | add_phi_arg (phi, trip_init, single_succ_edge (entry_bb), |
60d535d2 | 6807 | UNKNOWN_LOCATION); |
79acaae1 | 6808 | } |
6809 | ||
bc7bff74 | 6810 | if (!broken_loop) |
6811 | set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb); | |
79acaae1 | 6812 | set_immediate_dominator (CDI_DOMINATORS, iter_part_bb, |
6813 | recompute_dominator (CDI_DOMINATORS, iter_part_bb)); | |
6814 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
6815 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
6816 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, | |
6817 | recompute_dominator (CDI_DOMINATORS, seq_start_bb)); | |
6818 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
6819 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
04c2922b | 6820 | |
bc7bff74 | 6821 | if (!broken_loop) |
6822 | { | |
6823 | struct loop *trip_loop = alloc_loop (); | |
6824 | trip_loop->header = iter_part_bb; | |
6825 | trip_loop->latch = trip_update_bb; | |
6826 | add_loop (trip_loop, iter_part_bb->loop_father); | |
04c2922b | 6827 | |
bc7bff74 | 6828 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
6829 | { | |
6830 | struct loop *loop = alloc_loop (); | |
6831 | loop->header = body_bb; | |
33ee4d72 | 6832 | if (collapse_bb == NULL) |
6833 | loop->latch = cont_bb; | |
bc7bff74 | 6834 | add_loop (loop, trip_loop); |
6835 | } | |
6836 | } | |
1e8e9920 | 6837 | } |
6838 | ||
40750995 | 6839 | /* A subroutine of expand_omp_for. Generate code for _Cilk_for loop. |
6840 | Given parameters: | |
6841 | for (V = N1; V cond N2; V += STEP) BODY; | |
6842 | ||
6843 | where COND is "<" or ">" or "!=", we generate pseudocode | |
6844 | ||
6845 | for (ind_var = low; ind_var < high; ind_var++) | |
6846 | { | |
6847 | V = n1 + (ind_var * STEP) | |
6848 | ||
6849 | <BODY> | |
6850 | } | |
6851 | ||
6852 | In the above pseudocode, low and high are function parameters of the | |
6853 | child function. In the function below, we are inserting a temp. | |
6854 | variable that will be making a call to two OMP functions that will not be | |
6855 | found in the body of _Cilk_for (since OMP_FOR cannot be mixed | |
6856 | with _Cilk_for). These functions are replaced with low and high | |
6857 | by the function that handles taskreg. */ | |
6858 | ||
6859 | ||
6860 | static void | |
6861 | expand_cilk_for (struct omp_region *region, struct omp_for_data *fd) | |
6862 | { | |
6863 | bool broken_loop = region->cont == NULL; | |
6864 | basic_block entry_bb = region->entry; | |
6865 | basic_block cont_bb = region->cont; | |
6866 | ||
6867 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
6868 | gcc_assert (broken_loop | |
6869 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
6870 | basic_block l0_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
6871 | basic_block l1_bb, l2_bb; | |
6872 | ||
6873 | if (!broken_loop) | |
6874 | { | |
6875 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l0_bb); | |
6876 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
6877 | l1_bb = split_block (cont_bb, last_stmt (cont_bb))->dest; | |
6878 | l2_bb = BRANCH_EDGE (entry_bb)->dest; | |
6879 | } | |
6880 | else | |
6881 | { | |
6882 | BRANCH_EDGE (entry_bb)->flags &= ~EDGE_ABNORMAL; | |
6883 | l1_bb = split_edge (BRANCH_EDGE (entry_bb)); | |
6884 | l2_bb = single_succ (l1_bb); | |
6885 | } | |
6886 | basic_block exit_bb = region->exit; | |
6887 | basic_block l2_dom_bb = NULL; | |
6888 | ||
6889 | gimple_stmt_iterator gsi = gsi_last_bb (entry_bb); | |
6890 | ||
6891 | /* Below statements until the "tree high_val = ..." are pseudo statements | |
6892 | used to pass information to be used by expand_omp_taskreg. | |
6893 | low_val and high_val will be replaced by the __low and __high | |
6894 | parameter from the child function. | |
6895 | ||
6896 | The call_exprs part is a place-holder, it is mainly used | |
6897 | to distinctly identify to the top-level part that this is | |
6898 | where we should put low and high (reasoning given in header | |
6899 | comment). */ | |
6900 | ||
6901 | tree child_fndecl | |
1a91d914 | 6902 | = gimple_omp_parallel_child_fn ( |
6903 | as_a <gomp_parallel *> (last_stmt (region->outer->entry))); | |
40750995 | 6904 | tree t, low_val = NULL_TREE, high_val = NULL_TREE; |
6905 | for (t = DECL_ARGUMENTS (child_fndecl); t; t = TREE_CHAIN (t)) | |
6906 | { | |
6907 | if (!strcmp (IDENTIFIER_POINTER (DECL_NAME (t)), "__high")) | |
6908 | high_val = t; | |
6909 | else if (!strcmp (IDENTIFIER_POINTER (DECL_NAME (t)), "__low")) | |
6910 | low_val = t; | |
6911 | } | |
6912 | gcc_assert (low_val && high_val); | |
6913 | ||
6914 | tree type = TREE_TYPE (low_val); | |
6915 | tree ind_var = create_tmp_reg (type, "__cilk_ind_var"); | |
6916 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
6917 | ||
6918 | /* Not needed in SSA form right now. */ | |
6919 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
6920 | if (l2_dom_bb == NULL) | |
6921 | l2_dom_bb = l1_bb; | |
6922 | ||
6923 | tree n1 = low_val; | |
6924 | tree n2 = high_val; | |
6925 | ||
6926 | gimple stmt = gimple_build_assign (ind_var, n1); | |
6927 | ||
6928 | /* Replace the GIMPLE_OMP_FOR statement. */ | |
6929 | gsi_replace (&gsi, stmt, true); | |
6930 | ||
6931 | if (!broken_loop) | |
6932 | { | |
6933 | /* Code to control the increment goes in the CONT_BB. */ | |
6934 | gsi = gsi_last_bb (cont_bb); | |
6935 | stmt = gsi_stmt (gsi); | |
6936 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
6937 | stmt = gimple_build_assign_with_ops (PLUS_EXPR, ind_var, ind_var, | |
6938 | build_one_cst (type)); | |
6939 | ||
6940 | /* Replace GIMPLE_OMP_CONTINUE. */ | |
6941 | gsi_replace (&gsi, stmt, true); | |
6942 | } | |
6943 | ||
6944 | /* Emit the condition in L1_BB. */ | |
6945 | gsi = gsi_after_labels (l1_bb); | |
6946 | t = fold_build2 (MULT_EXPR, TREE_TYPE (fd->loop.step), | |
6947 | fold_convert (TREE_TYPE (fd->loop.step), ind_var), | |
6948 | fd->loop.step); | |
6949 | if (POINTER_TYPE_P (TREE_TYPE (fd->loop.n1))) | |
6950 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (fd->loop.n1), | |
6951 | fd->loop.n1, fold_convert (sizetype, t)); | |
6952 | else | |
6953 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loop.n1), | |
6954 | fd->loop.n1, fold_convert (TREE_TYPE (fd->loop.n1), t)); | |
6955 | t = fold_convert (TREE_TYPE (fd->loop.v), t); | |
6956 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
6957 | ||
6958 | /* The condition is always '<' since the runtime will fill in the low | |
6959 | and high values. */ | |
6960 | stmt = gimple_build_cond (LT_EXPR, ind_var, n2, NULL_TREE, NULL_TREE); | |
6961 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
6962 | ||
6963 | /* Remove GIMPLE_OMP_RETURN. */ | |
6964 | gsi = gsi_last_bb (exit_bb); | |
6965 | gsi_remove (&gsi, true); | |
6966 | ||
6967 | /* Connect the new blocks. */ | |
6968 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
6969 | ||
6970 | edge e, ne; | |
6971 | if (!broken_loop) | |
6972 | { | |
6973 | remove_edge (BRANCH_EDGE (entry_bb)); | |
6974 | make_edge (entry_bb, l1_bb, EDGE_FALLTHRU); | |
6975 | ||
6976 | e = BRANCH_EDGE (l1_bb); | |
6977 | ne = FALLTHRU_EDGE (l1_bb); | |
6978 | e->flags = EDGE_TRUE_VALUE; | |
6979 | } | |
6980 | else | |
6981 | { | |
6982 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
6983 | ||
6984 | ne = single_succ_edge (l1_bb); | |
6985 | e = make_edge (l1_bb, l0_bb, EDGE_TRUE_VALUE); | |
6986 | ||
6987 | } | |
6988 | ne->flags = EDGE_FALSE_VALUE; | |
6989 | e->probability = REG_BR_PROB_BASE * 7 / 8; | |
6990 | ne->probability = REG_BR_PROB_BASE / 8; | |
6991 | ||
6992 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, entry_bb); | |
6993 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, l2_dom_bb); | |
6994 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, l1_bb); | |
6995 | ||
6996 | if (!broken_loop) | |
6997 | { | |
6998 | struct loop *loop = alloc_loop (); | |
6999 | loop->header = l1_bb; | |
7000 | loop->latch = cont_bb; | |
7001 | add_loop (loop, l1_bb->loop_father); | |
7002 | loop->safelen = INT_MAX; | |
7003 | } | |
7004 | ||
7005 | /* Pick the correct library function based on the precision of the | |
7006 | induction variable type. */ | |
7007 | tree lib_fun = NULL_TREE; | |
7008 | if (TYPE_PRECISION (type) == 32) | |
7009 | lib_fun = cilk_for_32_fndecl; | |
7010 | else if (TYPE_PRECISION (type) == 64) | |
7011 | lib_fun = cilk_for_64_fndecl; | |
7012 | else | |
7013 | gcc_unreachable (); | |
7014 | ||
7015 | gcc_assert (fd->sched_kind == OMP_CLAUSE_SCHEDULE_CILKFOR); | |
7016 | ||
7017 | /* WS_ARGS contains the library function flavor to call: | |
7018 | __libcilkrts_cilk_for_64 or __libcilkrts_cilk_for_32), and the | |
7019 | user-defined grain value. If the user does not define one, then zero | |
7020 | is passed in by the parser. */ | |
7021 | vec_alloc (region->ws_args, 2); | |
7022 | region->ws_args->quick_push (lib_fun); | |
7023 | region->ws_args->quick_push (fd->chunk_size); | |
7024 | } | |
bc7bff74 | 7025 | |
3d483a94 | 7026 | /* A subroutine of expand_omp_for. Generate code for a simd non-worksharing |
7027 | loop. Given parameters: | |
7028 | ||
7029 | for (V = N1; V cond N2; V += STEP) BODY; | |
7030 | ||
7031 | where COND is "<" or ">", we generate pseudocode | |
7032 | ||
7033 | V = N1; | |
7034 | goto L1; | |
7035 | L0: | |
7036 | BODY; | |
7037 | V += STEP; | |
7038 | L1: | |
7039 | if (V cond N2) goto L0; else goto L2; | |
7040 | L2: | |
7041 | ||
7042 | For collapsed loops, given parameters: | |
7043 | collapse(3) | |
7044 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
7045 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
7046 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
7047 | BODY; | |
7048 | ||
7049 | we generate pseudocode | |
7050 | ||
7051 | if (cond3 is <) | |
7052 | adj = STEP3 - 1; | |
7053 | else | |
7054 | adj = STEP3 + 1; | |
7055 | count3 = (adj + N32 - N31) / STEP3; | |
7056 | if (cond2 is <) | |
7057 | adj = STEP2 - 1; | |
7058 | else | |
7059 | adj = STEP2 + 1; | |
7060 | count2 = (adj + N22 - N21) / STEP2; | |
7061 | if (cond1 is <) | |
7062 | adj = STEP1 - 1; | |
7063 | else | |
7064 | adj = STEP1 + 1; | |
7065 | count1 = (adj + N12 - N11) / STEP1; | |
7066 | count = count1 * count2 * count3; | |
7067 | V = 0; | |
7068 | V1 = N11; | |
7069 | V2 = N21; | |
7070 | V3 = N31; | |
7071 | goto L1; | |
7072 | L0: | |
7073 | BODY; | |
7074 | V += 1; | |
7075 | V3 += STEP3; | |
7076 | V2 += (V3 cond3 N32) ? 0 : STEP2; | |
7077 | V3 = (V3 cond3 N32) ? V3 : N31; | |
7078 | V1 += (V2 cond2 N22) ? 0 : STEP1; | |
7079 | V2 = (V2 cond2 N22) ? V2 : N21; | |
7080 | L1: | |
7081 | if (V < count) goto L0; else goto L2; | |
7082 | L2: | |
7083 | ||
7084 | */ | |
7085 | ||
7086 | static void | |
7087 | expand_omp_simd (struct omp_region *region, struct omp_for_data *fd) | |
7088 | { | |
7089 | tree type, t; | |
7090 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, l2_bb, l2_dom_bb; | |
7091 | gimple_stmt_iterator gsi; | |
7092 | gimple stmt; | |
1a91d914 | 7093 | gcond *cond_stmt; |
3d483a94 | 7094 | bool broken_loop = region->cont == NULL; |
7095 | edge e, ne; | |
7096 | tree *counts = NULL; | |
7097 | int i; | |
7098 | tree safelen = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
7099 | OMP_CLAUSE_SAFELEN); | |
7100 | tree simduid = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
7101 | OMP_CLAUSE__SIMDUID_); | |
bc7bff74 | 7102 | tree n1, n2; |
3d483a94 | 7103 | |
7104 | type = TREE_TYPE (fd->loop.v); | |
7105 | entry_bb = region->entry; | |
7106 | cont_bb = region->cont; | |
7107 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
7108 | gcc_assert (broken_loop | |
7109 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
7110 | l0_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
7111 | if (!broken_loop) | |
7112 | { | |
7113 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l0_bb); | |
7114 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
7115 | l1_bb = split_block (cont_bb, last_stmt (cont_bb))->dest; | |
7116 | l2_bb = BRANCH_EDGE (entry_bb)->dest; | |
7117 | } | |
7118 | else | |
7119 | { | |
7120 | BRANCH_EDGE (entry_bb)->flags &= ~EDGE_ABNORMAL; | |
7121 | l1_bb = split_edge (BRANCH_EDGE (entry_bb)); | |
7122 | l2_bb = single_succ (l1_bb); | |
7123 | } | |
7124 | exit_bb = region->exit; | |
7125 | l2_dom_bb = NULL; | |
7126 | ||
7127 | gsi = gsi_last_bb (entry_bb); | |
7128 | ||
7129 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
7130 | /* Not needed in SSA form right now. */ | |
7131 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
7132 | if (fd->collapse > 1) | |
7133 | { | |
7134 | int first_zero_iter = -1; | |
7135 | basic_block zero_iter_bb = l2_bb; | |
7136 | ||
7137 | counts = XALLOCAVEC (tree, fd->collapse); | |
7138 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
7139 | zero_iter_bb, first_zero_iter, | |
7140 | l2_dom_bb); | |
7141 | } | |
7142 | if (l2_dom_bb == NULL) | |
7143 | l2_dom_bb = l1_bb; | |
7144 | ||
bc7bff74 | 7145 | n1 = fd->loop.n1; |
3d483a94 | 7146 | n2 = fd->loop.n2; |
bc7bff74 | 7147 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) |
7148 | { | |
7149 | tree innerc = find_omp_clause (gimple_omp_for_clauses (fd->for_stmt), | |
7150 | OMP_CLAUSE__LOOPTEMP_); | |
7151 | gcc_assert (innerc); | |
7152 | n1 = OMP_CLAUSE_DECL (innerc); | |
7153 | innerc = find_omp_clause (OMP_CLAUSE_CHAIN (innerc), | |
7154 | OMP_CLAUSE__LOOPTEMP_); | |
7155 | gcc_assert (innerc); | |
7156 | n2 = OMP_CLAUSE_DECL (innerc); | |
7157 | expand_omp_build_assign (&gsi, fd->loop.v, | |
7158 | fold_convert (type, n1)); | |
7159 | if (fd->collapse > 1) | |
7160 | { | |
7161 | gsi_prev (&gsi); | |
7162 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, n1); | |
7163 | gsi_next (&gsi); | |
7164 | } | |
7165 | } | |
3d483a94 | 7166 | else |
7167 | { | |
7168 | expand_omp_build_assign (&gsi, fd->loop.v, | |
7169 | fold_convert (type, fd->loop.n1)); | |
7170 | if (fd->collapse > 1) | |
7171 | for (i = 0; i < fd->collapse; i++) | |
7172 | { | |
7173 | tree itype = TREE_TYPE (fd->loops[i].v); | |
7174 | if (POINTER_TYPE_P (itype)) | |
7175 | itype = signed_type_for (itype); | |
7176 | t = fold_convert (TREE_TYPE (fd->loops[i].v), fd->loops[i].n1); | |
7177 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
7178 | } | |
7179 | } | |
7180 | ||
7181 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
7182 | gsi_remove (&gsi, true); | |
7183 | ||
7184 | if (!broken_loop) | |
7185 | { | |
7186 | /* Code to control the increment goes in the CONT_BB. */ | |
7187 | gsi = gsi_last_bb (cont_bb); | |
7188 | stmt = gsi_stmt (gsi); | |
7189 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
7190 | ||
7191 | if (POINTER_TYPE_P (type)) | |
7192 | t = fold_build_pointer_plus (fd->loop.v, fd->loop.step); | |
7193 | else | |
7194 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, fd->loop.step); | |
7195 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
7196 | ||
7197 | if (fd->collapse > 1) | |
7198 | { | |
7199 | i = fd->collapse - 1; | |
7200 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
7201 | { | |
7202 | t = fold_convert (sizetype, fd->loops[i].step); | |
7203 | t = fold_build_pointer_plus (fd->loops[i].v, t); | |
7204 | } | |
7205 | else | |
7206 | { | |
7207 | t = fold_convert (TREE_TYPE (fd->loops[i].v), | |
7208 | fd->loops[i].step); | |
7209 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
7210 | fd->loops[i].v, t); | |
7211 | } | |
7212 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
7213 | ||
7214 | for (i = fd->collapse - 1; i > 0; i--) | |
7215 | { | |
7216 | tree itype = TREE_TYPE (fd->loops[i].v); | |
7217 | tree itype2 = TREE_TYPE (fd->loops[i - 1].v); | |
7218 | if (POINTER_TYPE_P (itype2)) | |
7219 | itype2 = signed_type_for (itype2); | |
7220 | t = build3 (COND_EXPR, itype2, | |
7221 | build2 (fd->loops[i].cond_code, boolean_type_node, | |
7222 | fd->loops[i].v, | |
7223 | fold_convert (itype, fd->loops[i].n2)), | |
7224 | build_int_cst (itype2, 0), | |
7225 | fold_convert (itype2, fd->loops[i - 1].step)); | |
7226 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i - 1].v))) | |
7227 | t = fold_build_pointer_plus (fd->loops[i - 1].v, t); | |
7228 | else | |
7229 | t = fold_build2 (PLUS_EXPR, itype2, fd->loops[i - 1].v, t); | |
7230 | expand_omp_build_assign (&gsi, fd->loops[i - 1].v, t); | |
7231 | ||
7232 | t = build3 (COND_EXPR, itype, | |
7233 | build2 (fd->loops[i].cond_code, boolean_type_node, | |
7234 | fd->loops[i].v, | |
7235 | fold_convert (itype, fd->loops[i].n2)), | |
7236 | fd->loops[i].v, | |
7237 | fold_convert (itype, fd->loops[i].n1)); | |
7238 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
7239 | } | |
7240 | } | |
7241 | ||
7242 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
7243 | gsi_remove (&gsi, true); | |
7244 | } | |
7245 | ||
7246 | /* Emit the condition in L1_BB. */ | |
7247 | gsi = gsi_start_bb (l1_bb); | |
7248 | ||
7249 | t = fold_convert (type, n2); | |
7250 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
7251 | false, GSI_CONTINUE_LINKING); | |
7252 | t = build2 (fd->loop.cond_code, boolean_type_node, fd->loop.v, t); | |
1a91d914 | 7253 | cond_stmt = gimple_build_cond_empty (t); |
7254 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
7255 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
3d483a94 | 7256 | NULL, NULL) |
1a91d914 | 7257 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), expand_omp_regimplify_p, |
3d483a94 | 7258 | NULL, NULL)) |
7259 | { | |
1a91d914 | 7260 | gsi = gsi_for_stmt (cond_stmt); |
7261 | gimple_regimplify_operands (cond_stmt, &gsi); | |
3d483a94 | 7262 | } |
7263 | ||
7264 | /* Remove GIMPLE_OMP_RETURN. */ | |
7265 | gsi = gsi_last_bb (exit_bb); | |
7266 | gsi_remove (&gsi, true); | |
7267 | ||
7268 | /* Connect the new blocks. */ | |
7269 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
7270 | ||
7271 | if (!broken_loop) | |
7272 | { | |
7273 | remove_edge (BRANCH_EDGE (entry_bb)); | |
7274 | make_edge (entry_bb, l1_bb, EDGE_FALLTHRU); | |
7275 | ||
7276 | e = BRANCH_EDGE (l1_bb); | |
7277 | ne = FALLTHRU_EDGE (l1_bb); | |
7278 | e->flags = EDGE_TRUE_VALUE; | |
7279 | } | |
7280 | else | |
7281 | { | |
7282 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
7283 | ||
7284 | ne = single_succ_edge (l1_bb); | |
7285 | e = make_edge (l1_bb, l0_bb, EDGE_TRUE_VALUE); | |
7286 | ||
7287 | } | |
7288 | ne->flags = EDGE_FALSE_VALUE; | |
7289 | e->probability = REG_BR_PROB_BASE * 7 / 8; | |
7290 | ne->probability = REG_BR_PROB_BASE / 8; | |
7291 | ||
7292 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, entry_bb); | |
7293 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, l2_dom_bb); | |
7294 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, l1_bb); | |
7295 | ||
7296 | if (!broken_loop) | |
7297 | { | |
7298 | struct loop *loop = alloc_loop (); | |
7299 | loop->header = l1_bb; | |
33ee4d72 | 7300 | loop->latch = cont_bb; |
3d483a94 | 7301 | add_loop (loop, l1_bb->loop_father); |
7302 | if (safelen == NULL_TREE) | |
7303 | loop->safelen = INT_MAX; | |
7304 | else | |
7305 | { | |
7306 | safelen = OMP_CLAUSE_SAFELEN_EXPR (safelen); | |
c3f3b68d | 7307 | if (TREE_CODE (safelen) != INTEGER_CST) |
7308 | loop->safelen = 0; | |
7309 | else if (!tree_fits_uhwi_p (safelen) | |
7310 | || tree_to_uhwi (safelen) > INT_MAX) | |
3d483a94 | 7311 | loop->safelen = INT_MAX; |
7312 | else | |
d85a2013 | 7313 | loop->safelen = tree_to_uhwi (safelen); |
3d483a94 | 7314 | if (loop->safelen == 1) |
7315 | loop->safelen = 0; | |
7316 | } | |
7317 | if (simduid) | |
7318 | { | |
7319 | loop->simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); | |
7320 | cfun->has_simduid_loops = true; | |
7321 | } | |
043115ec | 7322 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize |
3d483a94 | 7323 | the loop. */ |
043115ec | 7324 | if ((flag_tree_loop_vectorize |
7325 | || (!global_options_set.x_flag_tree_loop_vectorize | |
7326 | && !global_options_set.x_flag_tree_vectorize)) | |
ad45e43e | 7327 | && flag_tree_loop_optimize |
3d483a94 | 7328 | && loop->safelen > 1) |
7329 | { | |
4c73695b | 7330 | loop->force_vectorize = true; |
7331 | cfun->has_force_vectorize_loops = true; | |
3d483a94 | 7332 | } |
7333 | } | |
7334 | } | |
7335 | ||
1e8e9920 | 7336 | |
773c5ba7 | 7337 | /* Expand the OpenMP loop defined by REGION. */ |
1e8e9920 | 7338 | |
773c5ba7 | 7339 | static void |
bc7bff74 | 7340 | expand_omp_for (struct omp_region *region, gimple inner_stmt) |
773c5ba7 | 7341 | { |
7342 | struct omp_for_data fd; | |
fd6481cf | 7343 | struct omp_for_data_loop *loops; |
1e8e9920 | 7344 | |
fd6481cf | 7345 | loops |
7346 | = (struct omp_for_data_loop *) | |
75a70cf9 | 7347 | alloca (gimple_omp_for_collapse (last_stmt (region->entry)) |
fd6481cf | 7348 | * sizeof (struct omp_for_data_loop)); |
1a91d914 | 7349 | extract_omp_for_data (as_a <gomp_for *> (last_stmt (region->entry)), |
7350 | &fd, loops); | |
f77459c5 | 7351 | region->sched_kind = fd.sched_kind; |
1e8e9920 | 7352 | |
b3a3ddec | 7353 | gcc_assert (EDGE_COUNT (region->entry->succs) == 2); |
7354 | BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
7355 | FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
7356 | if (region->cont) | |
7357 | { | |
7358 | gcc_assert (EDGE_COUNT (region->cont->succs) == 2); | |
7359 | BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
7360 | FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
7361 | } | |
04c2922b | 7362 | else |
75de4aa2 | 7363 | /* If there isn't a continue then this is a degerate case where |
04c2922b | 7364 | the introduction of abnormal edges during lowering will prevent |
7365 | original loops from being detected. Fix that up. */ | |
7366 | loops_state_set (LOOPS_NEED_FIXUP); | |
b3a3ddec | 7367 | |
10c55644 | 7368 | if (gimple_omp_for_kind (fd.for_stmt) & GF_OMP_FOR_SIMD) |
3d483a94 | 7369 | expand_omp_simd (region, &fd); |
40750995 | 7370 | else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_CILKFOR) |
7371 | expand_cilk_for (region, &fd); | |
3d483a94 | 7372 | else if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC |
bc7bff74 | 7373 | && !fd.have_ordered) |
1e8e9920 | 7374 | { |
7375 | if (fd.chunk_size == NULL) | |
bc7bff74 | 7376 | expand_omp_for_static_nochunk (region, &fd, inner_stmt); |
1e8e9920 | 7377 | else |
bc7bff74 | 7378 | expand_omp_for_static_chunk (region, &fd, inner_stmt); |
1e8e9920 | 7379 | } |
7380 | else | |
7381 | { | |
fd6481cf | 7382 | int fn_index, start_ix, next_ix; |
7383 | ||
3d483a94 | 7384 | gcc_assert (gimple_omp_for_kind (fd.for_stmt) |
7385 | == GF_OMP_FOR_KIND_FOR); | |
0416ca72 | 7386 | if (fd.chunk_size == NULL |
7387 | && fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) | |
7388 | fd.chunk_size = integer_zero_node; | |
fd6481cf | 7389 | gcc_assert (fd.sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); |
7390 | fn_index = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_RUNTIME) | |
75a70cf9 | 7391 | ? 3 : fd.sched_kind; |
fd6481cf | 7392 | fn_index += fd.have_ordered * 4; |
b9a16870 | 7393 | start_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_START) + fn_index; |
7394 | next_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_NEXT) + fn_index; | |
fd6481cf | 7395 | if (fd.iter_type == long_long_unsigned_type_node) |
7396 | { | |
b9a16870 | 7397 | start_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_START |
7398 | - (int)BUILT_IN_GOMP_LOOP_STATIC_START); | |
7399 | next_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_NEXT | |
7400 | - (int)BUILT_IN_GOMP_LOOP_STATIC_NEXT); | |
fd6481cf | 7401 | } |
b9c74b4d | 7402 | expand_omp_for_generic (region, &fd, (enum built_in_function) start_ix, |
bc7bff74 | 7403 | (enum built_in_function) next_ix, inner_stmt); |
1e8e9920 | 7404 | } |
28c92cbb | 7405 | |
083152fb | 7406 | if (gimple_in_ssa_p (cfun)) |
7407 | update_ssa (TODO_update_ssa_only_virtuals); | |
1e8e9920 | 7408 | } |
7409 | ||
1e8e9920 | 7410 | |
7411 | /* Expand code for an OpenMP sections directive. In pseudo code, we generate | |
7412 | ||
1e8e9920 | 7413 | v = GOMP_sections_start (n); |
7414 | L0: | |
7415 | switch (v) | |
7416 | { | |
7417 | case 0: | |
7418 | goto L2; | |
7419 | case 1: | |
7420 | section 1; | |
7421 | goto L1; | |
7422 | case 2: | |
7423 | ... | |
7424 | case n: | |
7425 | ... | |
1e8e9920 | 7426 | default: |
7427 | abort (); | |
7428 | } | |
7429 | L1: | |
7430 | v = GOMP_sections_next (); | |
7431 | goto L0; | |
7432 | L2: | |
7433 | reduction; | |
7434 | ||
773c5ba7 | 7435 | If this is a combined parallel sections, replace the call to |
79acaae1 | 7436 | GOMP_sections_start with call to GOMP_sections_next. */ |
1e8e9920 | 7437 | |
7438 | static void | |
773c5ba7 | 7439 | expand_omp_sections (struct omp_region *region) |
1e8e9920 | 7440 | { |
f018d957 | 7441 | tree t, u, vin = NULL, vmain, vnext, l2; |
75a70cf9 | 7442 | unsigned len; |
ac6e3339 | 7443 | basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb; |
75a70cf9 | 7444 | gimple_stmt_iterator si, switch_si; |
1a91d914 | 7445 | gomp_sections *sections_stmt; |
7446 | gimple stmt; | |
7447 | gomp_continue *cont; | |
9884aaf8 | 7448 | edge_iterator ei; |
7449 | edge e; | |
61e47ac8 | 7450 | struct omp_region *inner; |
75a70cf9 | 7451 | unsigned i, casei; |
ac6e3339 | 7452 | bool exit_reachable = region->cont != NULL; |
1e8e9920 | 7453 | |
d244d9de | 7454 | gcc_assert (region->exit != NULL); |
61e47ac8 | 7455 | entry_bb = region->entry; |
ac6e3339 | 7456 | l0_bb = single_succ (entry_bb); |
61e47ac8 | 7457 | l1_bb = region->cont; |
ac6e3339 | 7458 | l2_bb = region->exit; |
d244d9de | 7459 | if (single_pred_p (l2_bb) && single_pred (l2_bb) == l0_bb) |
7460 | l2 = gimple_block_label (l2_bb); | |
7461 | else | |
03ed154b | 7462 | { |
d244d9de | 7463 | /* This can happen if there are reductions. */ |
7464 | len = EDGE_COUNT (l0_bb->succs); | |
7465 | gcc_assert (len > 0); | |
7466 | e = EDGE_SUCC (l0_bb, len - 1); | |
7467 | si = gsi_last_bb (e->dest); | |
7468 | l2 = NULL_TREE; | |
7469 | if (gsi_end_p (si) | |
7470 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
7471 | l2 = gimple_block_label (e->dest); | |
9884aaf8 | 7472 | else |
d244d9de | 7473 | FOR_EACH_EDGE (e, ei, l0_bb->succs) |
7474 | { | |
7475 | si = gsi_last_bb (e->dest); | |
7476 | if (gsi_end_p (si) | |
7477 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
9884aaf8 | 7478 | { |
d244d9de | 7479 | l2 = gimple_block_label (e->dest); |
7480 | break; | |
9884aaf8 | 7481 | } |
d244d9de | 7482 | } |
03ed154b | 7483 | } |
d244d9de | 7484 | if (exit_reachable) |
7485 | default_bb = create_empty_bb (l1_bb->prev_bb); | |
03ed154b | 7486 | else |
d244d9de | 7487 | default_bb = create_empty_bb (l0_bb); |
773c5ba7 | 7488 | |
7489 | /* We will build a switch() with enough cases for all the | |
75a70cf9 | 7490 | GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work |
773c5ba7 | 7491 | and a default case to abort if something goes wrong. */ |
ac6e3339 | 7492 | len = EDGE_COUNT (l0_bb->succs); |
75a70cf9 | 7493 | |
f1f41a6c | 7494 | /* Use vec::quick_push on label_vec throughout, since we know the size |
75a70cf9 | 7495 | in advance. */ |
c2078b80 | 7496 | auto_vec<tree> label_vec (len); |
1e8e9920 | 7497 | |
61e47ac8 | 7498 | /* The call to GOMP_sections_start goes in ENTRY_BB, replacing the |
75a70cf9 | 7499 | GIMPLE_OMP_SECTIONS statement. */ |
7500 | si = gsi_last_bb (entry_bb); | |
1a91d914 | 7501 | sections_stmt = as_a <gomp_sections *> (gsi_stmt (si)); |
75a70cf9 | 7502 | gcc_assert (gimple_code (sections_stmt) == GIMPLE_OMP_SECTIONS); |
7503 | vin = gimple_omp_sections_control (sections_stmt); | |
773c5ba7 | 7504 | if (!is_combined_parallel (region)) |
1e8e9920 | 7505 | { |
773c5ba7 | 7506 | /* If we are not inside a combined parallel+sections region, |
7507 | call GOMP_sections_start. */ | |
39cb6d68 | 7508 | t = build_int_cst (unsigned_type_node, len - 1); |
b9a16870 | 7509 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_START); |
75a70cf9 | 7510 | stmt = gimple_build_call (u, 1, t); |
1e8e9920 | 7511 | } |
79acaae1 | 7512 | else |
7513 | { | |
7514 | /* Otherwise, call GOMP_sections_next. */ | |
b9a16870 | 7515 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); |
75a70cf9 | 7516 | stmt = gimple_build_call (u, 0); |
79acaae1 | 7517 | } |
75a70cf9 | 7518 | gimple_call_set_lhs (stmt, vin); |
7519 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
7520 | gsi_remove (&si, true); | |
7521 | ||
7522 | /* The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in | |
7523 | L0_BB. */ | |
7524 | switch_si = gsi_last_bb (l0_bb); | |
7525 | gcc_assert (gimple_code (gsi_stmt (switch_si)) == GIMPLE_OMP_SECTIONS_SWITCH); | |
79acaae1 | 7526 | if (exit_reachable) |
7527 | { | |
1a91d914 | 7528 | cont = as_a <gomp_continue *> (last_stmt (l1_bb)); |
75a70cf9 | 7529 | gcc_assert (gimple_code (cont) == GIMPLE_OMP_CONTINUE); |
7530 | vmain = gimple_omp_continue_control_use (cont); | |
7531 | vnext = gimple_omp_continue_control_def (cont); | |
79acaae1 | 7532 | } |
7533 | else | |
7534 | { | |
7535 | vmain = vin; | |
7536 | vnext = NULL_TREE; | |
7537 | } | |
1e8e9920 | 7538 | |
d244d9de | 7539 | t = build_case_label (build_int_cst (unsigned_type_node, 0), NULL, l2); |
f1f41a6c | 7540 | label_vec.quick_push (t); |
d244d9de | 7541 | i = 1; |
03ed154b | 7542 | |
75a70cf9 | 7543 | /* Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR. */ |
ac6e3339 | 7544 | for (inner = region->inner, casei = 1; |
7545 | inner; | |
7546 | inner = inner->next, i++, casei++) | |
1e8e9920 | 7547 | { |
773c5ba7 | 7548 | basic_block s_entry_bb, s_exit_bb; |
7549 | ||
9884aaf8 | 7550 | /* Skip optional reduction region. */ |
75a70cf9 | 7551 | if (inner->type == GIMPLE_OMP_ATOMIC_LOAD) |
9884aaf8 | 7552 | { |
7553 | --i; | |
7554 | --casei; | |
7555 | continue; | |
7556 | } | |
7557 | ||
61e47ac8 | 7558 | s_entry_bb = inner->entry; |
7559 | s_exit_bb = inner->exit; | |
1e8e9920 | 7560 | |
75a70cf9 | 7561 | t = gimple_block_label (s_entry_bb); |
ac6e3339 | 7562 | u = build_int_cst (unsigned_type_node, casei); |
b6e3dd65 | 7563 | u = build_case_label (u, NULL, t); |
f1f41a6c | 7564 | label_vec.quick_push (u); |
61e47ac8 | 7565 | |
75a70cf9 | 7566 | si = gsi_last_bb (s_entry_bb); |
7567 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SECTION); | |
7568 | gcc_assert (i < len || gimple_omp_section_last_p (gsi_stmt (si))); | |
7569 | gsi_remove (&si, true); | |
61e47ac8 | 7570 | single_succ_edge (s_entry_bb)->flags = EDGE_FALLTHRU; |
03ed154b | 7571 | |
7572 | if (s_exit_bb == NULL) | |
7573 | continue; | |
7574 | ||
75a70cf9 | 7575 | si = gsi_last_bb (s_exit_bb); |
7576 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); | |
7577 | gsi_remove (&si, true); | |
03ed154b | 7578 | |
773c5ba7 | 7579 | single_succ_edge (s_exit_bb)->flags = EDGE_FALLTHRU; |
1e8e9920 | 7580 | } |
7581 | ||
773c5ba7 | 7582 | /* Error handling code goes in DEFAULT_BB. */ |
75a70cf9 | 7583 | t = gimple_block_label (default_bb); |
b6e3dd65 | 7584 | u = build_case_label (NULL, NULL, t); |
61e47ac8 | 7585 | make_edge (l0_bb, default_bb, 0); |
b3083327 | 7586 | add_bb_to_loop (default_bb, current_loops->tree_root); |
1e8e9920 | 7587 | |
49a70175 | 7588 | stmt = gimple_build_switch (vmain, u, label_vec); |
75a70cf9 | 7589 | gsi_insert_after (&switch_si, stmt, GSI_SAME_STMT); |
7590 | gsi_remove (&switch_si, true); | |
75a70cf9 | 7591 | |
7592 | si = gsi_start_bb (default_bb); | |
b9a16870 | 7593 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_TRAP), 0); |
75a70cf9 | 7594 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); |
773c5ba7 | 7595 | |
ac6e3339 | 7596 | if (exit_reachable) |
03ed154b | 7597 | { |
b9a16870 | 7598 | tree bfn_decl; |
7599 | ||
ac6e3339 | 7600 | /* Code to get the next section goes in L1_BB. */ |
75a70cf9 | 7601 | si = gsi_last_bb (l1_bb); |
7602 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CONTINUE); | |
1e8e9920 | 7603 | |
b9a16870 | 7604 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); |
7605 | stmt = gimple_build_call (bfn_decl, 0); | |
75a70cf9 | 7606 | gimple_call_set_lhs (stmt, vnext); |
7607 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
7608 | gsi_remove (&si, true); | |
773c5ba7 | 7609 | |
ac6e3339 | 7610 | single_succ_edge (l1_bb)->flags = EDGE_FALLTHRU; |
03ed154b | 7611 | } |
773c5ba7 | 7612 | |
d244d9de | 7613 | /* Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB. */ |
7614 | si = gsi_last_bb (l2_bb); | |
7615 | if (gimple_omp_return_nowait_p (gsi_stmt (si))) | |
7616 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_NOWAIT); | |
bc7bff74 | 7617 | else if (gimple_omp_return_lhs (gsi_stmt (si))) |
7618 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_CANCEL); | |
d244d9de | 7619 | else |
7620 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END); | |
7621 | stmt = gimple_build_call (t, 0); | |
bc7bff74 | 7622 | if (gimple_omp_return_lhs (gsi_stmt (si))) |
7623 | gimple_call_set_lhs (stmt, gimple_omp_return_lhs (gsi_stmt (si))); | |
d244d9de | 7624 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); |
7625 | gsi_remove (&si, true); | |
7626 | ||
79acaae1 | 7627 | set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb); |
773c5ba7 | 7628 | } |
1e8e9920 | 7629 | |
1e8e9920 | 7630 | |
61e47ac8 | 7631 | /* Expand code for an OpenMP single directive. We've already expanded |
7632 | much of the code, here we simply place the GOMP_barrier call. */ | |
7633 | ||
7634 | static void | |
7635 | expand_omp_single (struct omp_region *region) | |
7636 | { | |
7637 | basic_block entry_bb, exit_bb; | |
75a70cf9 | 7638 | gimple_stmt_iterator si; |
61e47ac8 | 7639 | |
7640 | entry_bb = region->entry; | |
7641 | exit_bb = region->exit; | |
7642 | ||
75a70cf9 | 7643 | si = gsi_last_bb (entry_bb); |
75a70cf9 | 7644 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE); |
7645 | gsi_remove (&si, true); | |
61e47ac8 | 7646 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; |
7647 | ||
75a70cf9 | 7648 | si = gsi_last_bb (exit_bb); |
bc7bff74 | 7649 | if (!gimple_omp_return_nowait_p (gsi_stmt (si))) |
7650 | { | |
7651 | tree t = gimple_omp_return_lhs (gsi_stmt (si)); | |
7652 | gsi_insert_after (&si, build_omp_barrier (t), GSI_SAME_STMT); | |
7653 | } | |
75a70cf9 | 7654 | gsi_remove (&si, true); |
61e47ac8 | 7655 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; |
7656 | } | |
7657 | ||
7658 | ||
7659 | /* Generic expansion for OpenMP synchronization directives: master, | |
7660 | ordered and critical. All we need to do here is remove the entry | |
7661 | and exit markers for REGION. */ | |
773c5ba7 | 7662 | |
7663 | static void | |
7664 | expand_omp_synch (struct omp_region *region) | |
7665 | { | |
7666 | basic_block entry_bb, exit_bb; | |
75a70cf9 | 7667 | gimple_stmt_iterator si; |
773c5ba7 | 7668 | |
61e47ac8 | 7669 | entry_bb = region->entry; |
7670 | exit_bb = region->exit; | |
773c5ba7 | 7671 | |
75a70cf9 | 7672 | si = gsi_last_bb (entry_bb); |
7673 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE | |
7674 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASTER | |
bc7bff74 | 7675 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TASKGROUP |
75a70cf9 | 7676 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ORDERED |
bc7bff74 | 7677 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CRITICAL |
7678 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS); | |
75a70cf9 | 7679 | gsi_remove (&si, true); |
773c5ba7 | 7680 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; |
7681 | ||
03ed154b | 7682 | if (exit_bb) |
7683 | { | |
75a70cf9 | 7684 | si = gsi_last_bb (exit_bb); |
7685 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); | |
7686 | gsi_remove (&si, true); | |
03ed154b | 7687 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; |
7688 | } | |
773c5ba7 | 7689 | } |
1e8e9920 | 7690 | |
2169f33b | 7691 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
7692 | operation as a normal volatile load. */ | |
7693 | ||
7694 | static bool | |
3ec11c49 | 7695 | expand_omp_atomic_load (basic_block load_bb, tree addr, |
7696 | tree loaded_val, int index) | |
2169f33b | 7697 | { |
3ec11c49 | 7698 | enum built_in_function tmpbase; |
7699 | gimple_stmt_iterator gsi; | |
7700 | basic_block store_bb; | |
7701 | location_t loc; | |
7702 | gimple stmt; | |
7703 | tree decl, call, type, itype; | |
7704 | ||
7705 | gsi = gsi_last_bb (load_bb); | |
7706 | stmt = gsi_stmt (gsi); | |
7707 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
7708 | loc = gimple_location (stmt); | |
7709 | ||
7710 | /* ??? If the target does not implement atomic_load_optab[mode], and mode | |
7711 | is smaller than word size, then expand_atomic_load assumes that the load | |
7712 | is atomic. We could avoid the builtin entirely in this case. */ | |
7713 | ||
7714 | tmpbase = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
7715 | decl = builtin_decl_explicit (tmpbase); | |
7716 | if (decl == NULL_TREE) | |
7717 | return false; | |
7718 | ||
7719 | type = TREE_TYPE (loaded_val); | |
7720 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
7721 | ||
7722 | call = build_call_expr_loc (loc, decl, 2, addr, | |
bc7bff74 | 7723 | build_int_cst (NULL, |
7724 | gimple_omp_atomic_seq_cst_p (stmt) | |
7725 | ? MEMMODEL_SEQ_CST | |
7726 | : MEMMODEL_RELAXED)); | |
3ec11c49 | 7727 | if (!useless_type_conversion_p (type, itype)) |
7728 | call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
7729 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
7730 | ||
7731 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
7732 | gsi_remove (&gsi, true); | |
7733 | ||
7734 | store_bb = single_succ (load_bb); | |
7735 | gsi = gsi_last_bb (store_bb); | |
7736 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); | |
7737 | gsi_remove (&gsi, true); | |
7738 | ||
7739 | if (gimple_in_ssa_p (cfun)) | |
7740 | update_ssa (TODO_update_ssa_no_phi); | |
7741 | ||
7742 | return true; | |
2169f33b | 7743 | } |
7744 | ||
7745 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
7746 | operation as a normal volatile store. */ | |
7747 | ||
7748 | static bool | |
3ec11c49 | 7749 | expand_omp_atomic_store (basic_block load_bb, tree addr, |
7750 | tree loaded_val, tree stored_val, int index) | |
2169f33b | 7751 | { |
3ec11c49 | 7752 | enum built_in_function tmpbase; |
7753 | gimple_stmt_iterator gsi; | |
7754 | basic_block store_bb = single_succ (load_bb); | |
7755 | location_t loc; | |
7756 | gimple stmt; | |
7757 | tree decl, call, type, itype; | |
3754d046 | 7758 | machine_mode imode; |
3ec11c49 | 7759 | bool exchange; |
7760 | ||
7761 | gsi = gsi_last_bb (load_bb); | |
7762 | stmt = gsi_stmt (gsi); | |
7763 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
7764 | ||
7765 | /* If the load value is needed, then this isn't a store but an exchange. */ | |
7766 | exchange = gimple_omp_atomic_need_value_p (stmt); | |
7767 | ||
7768 | gsi = gsi_last_bb (store_bb); | |
7769 | stmt = gsi_stmt (gsi); | |
7770 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_STORE); | |
7771 | loc = gimple_location (stmt); | |
7772 | ||
7773 | /* ??? If the target does not implement atomic_store_optab[mode], and mode | |
7774 | is smaller than word size, then expand_atomic_store assumes that the store | |
7775 | is atomic. We could avoid the builtin entirely in this case. */ | |
7776 | ||
7777 | tmpbase = (exchange ? BUILT_IN_ATOMIC_EXCHANGE_N : BUILT_IN_ATOMIC_STORE_N); | |
7778 | tmpbase = (enum built_in_function) ((int) tmpbase + index + 1); | |
7779 | decl = builtin_decl_explicit (tmpbase); | |
7780 | if (decl == NULL_TREE) | |
7781 | return false; | |
7782 | ||
7783 | type = TREE_TYPE (stored_val); | |
7784 | ||
7785 | /* Dig out the type of the function's second argument. */ | |
7786 | itype = TREE_TYPE (decl); | |
7787 | itype = TYPE_ARG_TYPES (itype); | |
7788 | itype = TREE_CHAIN (itype); | |
7789 | itype = TREE_VALUE (itype); | |
7790 | imode = TYPE_MODE (itype); | |
7791 | ||
7792 | if (exchange && !can_atomic_exchange_p (imode, true)) | |
7793 | return false; | |
7794 | ||
7795 | if (!useless_type_conversion_p (itype, type)) | |
7796 | stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val); | |
7797 | call = build_call_expr_loc (loc, decl, 3, addr, stored_val, | |
bc7bff74 | 7798 | build_int_cst (NULL, |
7799 | gimple_omp_atomic_seq_cst_p (stmt) | |
7800 | ? MEMMODEL_SEQ_CST | |
7801 | : MEMMODEL_RELAXED)); | |
3ec11c49 | 7802 | if (exchange) |
7803 | { | |
7804 | if (!useless_type_conversion_p (type, itype)) | |
7805 | call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
7806 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
7807 | } | |
7808 | ||
7809 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
7810 | gsi_remove (&gsi, true); | |
7811 | ||
7812 | /* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */ | |
7813 | gsi = gsi_last_bb (load_bb); | |
7814 | gsi_remove (&gsi, true); | |
7815 | ||
7816 | if (gimple_in_ssa_p (cfun)) | |
7817 | update_ssa (TODO_update_ssa_no_phi); | |
7818 | ||
7819 | return true; | |
2169f33b | 7820 | } |
7821 | ||
cb7f680b | 7822 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
1cd6e20d | 7823 | operation as a __atomic_fetch_op builtin. INDEX is log2 of the |
cb7f680b | 7824 | size of the data type, and thus usable to find the index of the builtin |
7825 | decl. Returns false if the expression is not of the proper form. */ | |
7826 | ||
7827 | static bool | |
7828 | expand_omp_atomic_fetch_op (basic_block load_bb, | |
7829 | tree addr, tree loaded_val, | |
7830 | tree stored_val, int index) | |
7831 | { | |
b9a16870 | 7832 | enum built_in_function oldbase, newbase, tmpbase; |
cb7f680b | 7833 | tree decl, itype, call; |
2169f33b | 7834 | tree lhs, rhs; |
cb7f680b | 7835 | basic_block store_bb = single_succ (load_bb); |
75a70cf9 | 7836 | gimple_stmt_iterator gsi; |
7837 | gimple stmt; | |
389dd41b | 7838 | location_t loc; |
1cd6e20d | 7839 | enum tree_code code; |
2169f33b | 7840 | bool need_old, need_new; |
3754d046 | 7841 | machine_mode imode; |
bc7bff74 | 7842 | bool seq_cst; |
cb7f680b | 7843 | |
7844 | /* We expect to find the following sequences: | |
48e1416a | 7845 | |
cb7f680b | 7846 | load_bb: |
75a70cf9 | 7847 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) |
cb7f680b | 7848 | |
7849 | store_bb: | |
7850 | val = tmp OP something; (or: something OP tmp) | |
48e1416a | 7851 | GIMPLE_OMP_STORE (val) |
cb7f680b | 7852 | |
48e1416a | 7853 | ???FIXME: Allow a more flexible sequence. |
cb7f680b | 7854 | Perhaps use data flow to pick the statements. |
48e1416a | 7855 | |
cb7f680b | 7856 | */ |
7857 | ||
75a70cf9 | 7858 | gsi = gsi_after_labels (store_bb); |
7859 | stmt = gsi_stmt (gsi); | |
389dd41b | 7860 | loc = gimple_location (stmt); |
75a70cf9 | 7861 | if (!is_gimple_assign (stmt)) |
cb7f680b | 7862 | return false; |
75a70cf9 | 7863 | gsi_next (&gsi); |
7864 | if (gimple_code (gsi_stmt (gsi)) != GIMPLE_OMP_ATOMIC_STORE) | |
cb7f680b | 7865 | return false; |
2169f33b | 7866 | need_new = gimple_omp_atomic_need_value_p (gsi_stmt (gsi)); |
7867 | need_old = gimple_omp_atomic_need_value_p (last_stmt (load_bb)); | |
bc7bff74 | 7868 | seq_cst = gimple_omp_atomic_seq_cst_p (last_stmt (load_bb)); |
2169f33b | 7869 | gcc_checking_assert (!need_old || !need_new); |
cb7f680b | 7870 | |
75a70cf9 | 7871 | if (!operand_equal_p (gimple_assign_lhs (stmt), stored_val, 0)) |
cb7f680b | 7872 | return false; |
7873 | ||
cb7f680b | 7874 | /* Check for one of the supported fetch-op operations. */ |
1cd6e20d | 7875 | code = gimple_assign_rhs_code (stmt); |
7876 | switch (code) | |
cb7f680b | 7877 | { |
7878 | case PLUS_EXPR: | |
7879 | case POINTER_PLUS_EXPR: | |
1cd6e20d | 7880 | oldbase = BUILT_IN_ATOMIC_FETCH_ADD_N; |
7881 | newbase = BUILT_IN_ATOMIC_ADD_FETCH_N; | |
cb7f680b | 7882 | break; |
7883 | case MINUS_EXPR: | |
1cd6e20d | 7884 | oldbase = BUILT_IN_ATOMIC_FETCH_SUB_N; |
7885 | newbase = BUILT_IN_ATOMIC_SUB_FETCH_N; | |
cb7f680b | 7886 | break; |
7887 | case BIT_AND_EXPR: | |
1cd6e20d | 7888 | oldbase = BUILT_IN_ATOMIC_FETCH_AND_N; |
7889 | newbase = BUILT_IN_ATOMIC_AND_FETCH_N; | |
cb7f680b | 7890 | break; |
7891 | case BIT_IOR_EXPR: | |
1cd6e20d | 7892 | oldbase = BUILT_IN_ATOMIC_FETCH_OR_N; |
7893 | newbase = BUILT_IN_ATOMIC_OR_FETCH_N; | |
cb7f680b | 7894 | break; |
7895 | case BIT_XOR_EXPR: | |
1cd6e20d | 7896 | oldbase = BUILT_IN_ATOMIC_FETCH_XOR_N; |
7897 | newbase = BUILT_IN_ATOMIC_XOR_FETCH_N; | |
cb7f680b | 7898 | break; |
7899 | default: | |
7900 | return false; | |
7901 | } | |
1cd6e20d | 7902 | |
cb7f680b | 7903 | /* Make sure the expression is of the proper form. */ |
75a70cf9 | 7904 | if (operand_equal_p (gimple_assign_rhs1 (stmt), loaded_val, 0)) |
7905 | rhs = gimple_assign_rhs2 (stmt); | |
7906 | else if (commutative_tree_code (gimple_assign_rhs_code (stmt)) | |
7907 | && operand_equal_p (gimple_assign_rhs2 (stmt), loaded_val, 0)) | |
7908 | rhs = gimple_assign_rhs1 (stmt); | |
cb7f680b | 7909 | else |
7910 | return false; | |
7911 | ||
b9a16870 | 7912 | tmpbase = ((enum built_in_function) |
7913 | ((need_new ? newbase : oldbase) + index + 1)); | |
7914 | decl = builtin_decl_explicit (tmpbase); | |
0f94f46b | 7915 | if (decl == NULL_TREE) |
7916 | return false; | |
cb7f680b | 7917 | itype = TREE_TYPE (TREE_TYPE (decl)); |
1cd6e20d | 7918 | imode = TYPE_MODE (itype); |
cb7f680b | 7919 | |
1cd6e20d | 7920 | /* We could test all of the various optabs involved, but the fact of the |
7921 | matter is that (with the exception of i486 vs i586 and xadd) all targets | |
7922 | that support any atomic operaton optab also implements compare-and-swap. | |
7923 | Let optabs.c take care of expanding any compare-and-swap loop. */ | |
29139cdc | 7924 | if (!can_compare_and_swap_p (imode, true)) |
cb7f680b | 7925 | return false; |
7926 | ||
75a70cf9 | 7927 | gsi = gsi_last_bb (load_bb); |
7928 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_LOAD); | |
1cd6e20d | 7929 | |
7930 | /* OpenMP does not imply any barrier-like semantics on its atomic ops. | |
7931 | It only requires that the operation happen atomically. Thus we can | |
7932 | use the RELAXED memory model. */ | |
7933 | call = build_call_expr_loc (loc, decl, 3, addr, | |
7934 | fold_convert_loc (loc, itype, rhs), | |
bc7bff74 | 7935 | build_int_cst (NULL, |
7936 | seq_cst ? MEMMODEL_SEQ_CST | |
7937 | : MEMMODEL_RELAXED)); | |
1cd6e20d | 7938 | |
2169f33b | 7939 | if (need_old || need_new) |
7940 | { | |
7941 | lhs = need_old ? loaded_val : stored_val; | |
7942 | call = fold_convert_loc (loc, TREE_TYPE (lhs), call); | |
7943 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, lhs, call); | |
7944 | } | |
7945 | else | |
7946 | call = fold_convert_loc (loc, void_type_node, call); | |
75a70cf9 | 7947 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); |
7948 | gsi_remove (&gsi, true); | |
cb7f680b | 7949 | |
75a70cf9 | 7950 | gsi = gsi_last_bb (store_bb); |
7951 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); | |
7952 | gsi_remove (&gsi, true); | |
7953 | gsi = gsi_last_bb (store_bb); | |
7954 | gsi_remove (&gsi, true); | |
cb7f680b | 7955 | |
7956 | if (gimple_in_ssa_p (cfun)) | |
7957 | update_ssa (TODO_update_ssa_no_phi); | |
7958 | ||
7959 | return true; | |
7960 | } | |
7961 | ||
7962 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
7963 | ||
7964 | oldval = *addr; | |
7965 | repeat: | |
7966 | newval = rhs; // with oldval replacing *addr in rhs | |
7967 | oldval = __sync_val_compare_and_swap (addr, oldval, newval); | |
7968 | if (oldval != newval) | |
7969 | goto repeat; | |
7970 | ||
7971 | INDEX is log2 of the size of the data type, and thus usable to find the | |
7972 | index of the builtin decl. */ | |
7973 | ||
7974 | static bool | |
7975 | expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb, | |
7976 | tree addr, tree loaded_val, tree stored_val, | |
7977 | int index) | |
7978 | { | |
790368c5 | 7979 | tree loadedi, storedi, initial, new_storedi, old_vali; |
cb7f680b | 7980 | tree type, itype, cmpxchg, iaddr; |
75a70cf9 | 7981 | gimple_stmt_iterator si; |
cb7f680b | 7982 | basic_block loop_header = single_succ (load_bb); |
75a70cf9 | 7983 | gimple phi, stmt; |
cb7f680b | 7984 | edge e; |
b9a16870 | 7985 | enum built_in_function fncode; |
cb7f680b | 7986 | |
1cd6e20d | 7987 | /* ??? We need a non-pointer interface to __atomic_compare_exchange in |
7988 | order to use the RELAXED memory model effectively. */ | |
b9a16870 | 7989 | fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N |
7990 | + index + 1); | |
7991 | cmpxchg = builtin_decl_explicit (fncode); | |
0f94f46b | 7992 | if (cmpxchg == NULL_TREE) |
7993 | return false; | |
cb7f680b | 7994 | type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr))); |
7995 | itype = TREE_TYPE (TREE_TYPE (cmpxchg)); | |
7996 | ||
29139cdc | 7997 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true)) |
cb7f680b | 7998 | return false; |
7999 | ||
75a70cf9 | 8000 | /* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */ |
8001 | si = gsi_last_bb (load_bb); | |
8002 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); | |
8003 | ||
790368c5 | 8004 | /* For floating-point values, we'll need to view-convert them to integers |
8005 | so that we can perform the atomic compare and swap. Simplify the | |
8006 | following code by always setting up the "i"ntegral variables. */ | |
8007 | if (!INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type)) | |
8008 | { | |
75a70cf9 | 8009 | tree iaddr_val; |
8010 | ||
072f7ab1 | 8011 | iaddr = create_tmp_reg (build_pointer_type_for_mode (itype, ptr_mode, |
f9e245b2 | 8012 | true)); |
75a70cf9 | 8013 | iaddr_val |
8014 | = force_gimple_operand_gsi (&si, | |
8015 | fold_convert (TREE_TYPE (iaddr), addr), | |
8016 | false, NULL_TREE, true, GSI_SAME_STMT); | |
8017 | stmt = gimple_build_assign (iaddr, iaddr_val); | |
8018 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
f9e245b2 | 8019 | loadedi = create_tmp_var (itype); |
790368c5 | 8020 | if (gimple_in_ssa_p (cfun)) |
f9e245b2 | 8021 | loadedi = make_ssa_name (loadedi); |
790368c5 | 8022 | } |
8023 | else | |
8024 | { | |
8025 | iaddr = addr; | |
8026 | loadedi = loaded_val; | |
8027 | } | |
75a70cf9 | 8028 | |
2ed72821 | 8029 | fncode = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); |
8030 | tree loaddecl = builtin_decl_explicit (fncode); | |
8031 | if (loaddecl) | |
8032 | initial | |
8033 | = fold_convert (TREE_TYPE (TREE_TYPE (iaddr)), | |
8034 | build_call_expr (loaddecl, 2, iaddr, | |
8035 | build_int_cst (NULL_TREE, | |
8036 | MEMMODEL_RELAXED))); | |
8037 | else | |
8038 | initial = build2 (MEM_REF, TREE_TYPE (TREE_TYPE (iaddr)), iaddr, | |
8039 | build_int_cst (TREE_TYPE (iaddr), 0)); | |
8040 | ||
182cf5a9 | 8041 | initial |
2ed72821 | 8042 | = force_gimple_operand_gsi (&si, initial, true, NULL_TREE, true, |
8043 | GSI_SAME_STMT); | |
790368c5 | 8044 | |
8045 | /* Move the value to the LOADEDI temporary. */ | |
cb7f680b | 8046 | if (gimple_in_ssa_p (cfun)) |
8047 | { | |
75a70cf9 | 8048 | gcc_assert (gimple_seq_empty_p (phi_nodes (loop_header))); |
790368c5 | 8049 | phi = create_phi_node (loadedi, loop_header); |
cb7f680b | 8050 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)), |
8051 | initial); | |
8052 | } | |
8053 | else | |
75a70cf9 | 8054 | gsi_insert_before (&si, |
8055 | gimple_build_assign (loadedi, initial), | |
8056 | GSI_SAME_STMT); | |
790368c5 | 8057 | if (loadedi != loaded_val) |
8058 | { | |
75a70cf9 | 8059 | gimple_stmt_iterator gsi2; |
8060 | tree x; | |
790368c5 | 8061 | |
8062 | x = build1 (VIEW_CONVERT_EXPR, type, loadedi); | |
75a70cf9 | 8063 | gsi2 = gsi_start_bb (loop_header); |
790368c5 | 8064 | if (gimple_in_ssa_p (cfun)) |
8065 | { | |
1a91d914 | 8066 | gassign *stmt; |
75a70cf9 | 8067 | x = force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, |
8068 | true, GSI_SAME_STMT); | |
8069 | stmt = gimple_build_assign (loaded_val, x); | |
8070 | gsi_insert_before (&gsi2, stmt, GSI_SAME_STMT); | |
790368c5 | 8071 | } |
8072 | else | |
8073 | { | |
75a70cf9 | 8074 | x = build2 (MODIFY_EXPR, TREE_TYPE (loaded_val), loaded_val, x); |
8075 | force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
8076 | true, GSI_SAME_STMT); | |
790368c5 | 8077 | } |
8078 | } | |
75a70cf9 | 8079 | gsi_remove (&si, true); |
cb7f680b | 8080 | |
75a70cf9 | 8081 | si = gsi_last_bb (store_bb); |
8082 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); | |
cb7f680b | 8083 | |
790368c5 | 8084 | if (iaddr == addr) |
8085 | storedi = stored_val; | |
cb7f680b | 8086 | else |
790368c5 | 8087 | storedi = |
75a70cf9 | 8088 | force_gimple_operand_gsi (&si, |
790368c5 | 8089 | build1 (VIEW_CONVERT_EXPR, itype, |
8090 | stored_val), true, NULL_TREE, true, | |
75a70cf9 | 8091 | GSI_SAME_STMT); |
cb7f680b | 8092 | |
8093 | /* Build the compare&swap statement. */ | |
8094 | new_storedi = build_call_expr (cmpxchg, 3, iaddr, loadedi, storedi); | |
75a70cf9 | 8095 | new_storedi = force_gimple_operand_gsi (&si, |
87f9ffa4 | 8096 | fold_convert (TREE_TYPE (loadedi), |
8097 | new_storedi), | |
cb7f680b | 8098 | true, NULL_TREE, |
75a70cf9 | 8099 | true, GSI_SAME_STMT); |
cb7f680b | 8100 | |
8101 | if (gimple_in_ssa_p (cfun)) | |
8102 | old_vali = loadedi; | |
8103 | else | |
8104 | { | |
f9e245b2 | 8105 | old_vali = create_tmp_var (TREE_TYPE (loadedi)); |
75a70cf9 | 8106 | stmt = gimple_build_assign (old_vali, loadedi); |
8107 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
cb7f680b | 8108 | |
75a70cf9 | 8109 | stmt = gimple_build_assign (loadedi, new_storedi); |
8110 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
cb7f680b | 8111 | } |
8112 | ||
8113 | /* Note that we always perform the comparison as an integer, even for | |
48e1416a | 8114 | floating point. This allows the atomic operation to properly |
cb7f680b | 8115 | succeed even with NaNs and -0.0. */ |
75a70cf9 | 8116 | stmt = gimple_build_cond_empty |
8117 | (build2 (NE_EXPR, boolean_type_node, | |
8118 | new_storedi, old_vali)); | |
8119 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
cb7f680b | 8120 | |
8121 | /* Update cfg. */ | |
8122 | e = single_succ_edge (store_bb); | |
8123 | e->flags &= ~EDGE_FALLTHRU; | |
8124 | e->flags |= EDGE_FALSE_VALUE; | |
8125 | ||
8126 | e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE); | |
8127 | ||
790368c5 | 8128 | /* Copy the new value to loadedi (we already did that before the condition |
cb7f680b | 8129 | if we are not in SSA). */ |
8130 | if (gimple_in_ssa_p (cfun)) | |
8131 | { | |
75a70cf9 | 8132 | phi = gimple_seq_first_stmt (phi_nodes (loop_header)); |
790368c5 | 8133 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_storedi); |
cb7f680b | 8134 | } |
8135 | ||
75a70cf9 | 8136 | /* Remove GIMPLE_OMP_ATOMIC_STORE. */ |
8137 | gsi_remove (&si, true); | |
cb7f680b | 8138 | |
04c2922b | 8139 | struct loop *loop = alloc_loop (); |
8140 | loop->header = loop_header; | |
5f037457 | 8141 | loop->latch = store_bb; |
04c2922b | 8142 | add_loop (loop, loop_header->loop_father); |
8143 | ||
cb7f680b | 8144 | if (gimple_in_ssa_p (cfun)) |
8145 | update_ssa (TODO_update_ssa_no_phi); | |
8146 | ||
8147 | return true; | |
8148 | } | |
8149 | ||
8150 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
8151 | ||
8152 | GOMP_atomic_start (); | |
8153 | *addr = rhs; | |
8154 | GOMP_atomic_end (); | |
8155 | ||
8156 | The result is not globally atomic, but works so long as all parallel | |
8157 | references are within #pragma omp atomic directives. According to | |
8158 | responses received from omp@openmp.org, appears to be within spec. | |
8159 | Which makes sense, since that's how several other compilers handle | |
48e1416a | 8160 | this situation as well. |
75a70cf9 | 8161 | LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're |
8162 | expanding. STORED_VAL is the operand of the matching | |
8163 | GIMPLE_OMP_ATOMIC_STORE. | |
cb7f680b | 8164 | |
48e1416a | 8165 | We replace |
8166 | GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with | |
cb7f680b | 8167 | loaded_val = *addr; |
8168 | ||
8169 | and replace | |
3ec11c49 | 8170 | GIMPLE_OMP_ATOMIC_STORE (stored_val) with |
48e1416a | 8171 | *addr = stored_val; |
cb7f680b | 8172 | */ |
8173 | ||
8174 | static bool | |
8175 | expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb, | |
8176 | tree addr, tree loaded_val, tree stored_val) | |
8177 | { | |
75a70cf9 | 8178 | gimple_stmt_iterator si; |
1a91d914 | 8179 | gassign *stmt; |
cb7f680b | 8180 | tree t; |
8181 | ||
75a70cf9 | 8182 | si = gsi_last_bb (load_bb); |
8183 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); | |
cb7f680b | 8184 | |
b9a16870 | 8185 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START); |
414c3a2c | 8186 | t = build_call_expr (t, 0); |
75a70cf9 | 8187 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); |
cb7f680b | 8188 | |
182cf5a9 | 8189 | stmt = gimple_build_assign (loaded_val, build_simple_mem_ref (addr)); |
75a70cf9 | 8190 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
8191 | gsi_remove (&si, true); | |
cb7f680b | 8192 | |
75a70cf9 | 8193 | si = gsi_last_bb (store_bb); |
8194 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); | |
cb7f680b | 8195 | |
182cf5a9 | 8196 | stmt = gimple_build_assign (build_simple_mem_ref (unshare_expr (addr)), |
8197 | stored_val); | |
75a70cf9 | 8198 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
cb7f680b | 8199 | |
b9a16870 | 8200 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END); |
414c3a2c | 8201 | t = build_call_expr (t, 0); |
75a70cf9 | 8202 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); |
8203 | gsi_remove (&si, true); | |
cb7f680b | 8204 | |
8205 | if (gimple_in_ssa_p (cfun)) | |
8206 | update_ssa (TODO_update_ssa_no_phi); | |
8207 | return true; | |
8208 | } | |
8209 | ||
48e1416a | 8210 | /* Expand an GIMPLE_OMP_ATOMIC statement. We try to expand |
8211 | using expand_omp_atomic_fetch_op. If it failed, we try to | |
cb7f680b | 8212 | call expand_omp_atomic_pipeline, and if it fails too, the |
8213 | ultimate fallback is wrapping the operation in a mutex | |
48e1416a | 8214 | (expand_omp_atomic_mutex). REGION is the atomic region built |
8215 | by build_omp_regions_1(). */ | |
cb7f680b | 8216 | |
8217 | static void | |
8218 | expand_omp_atomic (struct omp_region *region) | |
8219 | { | |
8220 | basic_block load_bb = region->entry, store_bb = region->exit; | |
1a91d914 | 8221 | gomp_atomic_load *load = as_a <gomp_atomic_load *> (last_stmt (load_bb)); |
8222 | gomp_atomic_store *store = as_a <gomp_atomic_store *> (last_stmt (store_bb)); | |
75a70cf9 | 8223 | tree loaded_val = gimple_omp_atomic_load_lhs (load); |
8224 | tree addr = gimple_omp_atomic_load_rhs (load); | |
8225 | tree stored_val = gimple_omp_atomic_store_val (store); | |
cb7f680b | 8226 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr))); |
8227 | HOST_WIDE_INT index; | |
8228 | ||
8229 | /* Make sure the type is one of the supported sizes. */ | |
e913b5cd | 8230 | index = tree_to_uhwi (TYPE_SIZE_UNIT (type)); |
cb7f680b | 8231 | index = exact_log2 (index); |
8232 | if (index >= 0 && index <= 4) | |
8233 | { | |
8234 | unsigned int align = TYPE_ALIGN_UNIT (type); | |
8235 | ||
8236 | /* __sync builtins require strict data alignment. */ | |
dcf7024c | 8237 | if (exact_log2 (align) >= index) |
cb7f680b | 8238 | { |
3ec11c49 | 8239 | /* Atomic load. */ |
2169f33b | 8240 | if (loaded_val == stored_val |
8241 | && (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_INT | |
8242 | || GET_MODE_CLASS (TYPE_MODE (type)) == MODE_FLOAT) | |
8243 | && GET_MODE_BITSIZE (TYPE_MODE (type)) <= BITS_PER_WORD | |
3ec11c49 | 8244 | && expand_omp_atomic_load (load_bb, addr, loaded_val, index)) |
2169f33b | 8245 | return; |
8246 | ||
3ec11c49 | 8247 | /* Atomic store. */ |
2169f33b | 8248 | if ((GET_MODE_CLASS (TYPE_MODE (type)) == MODE_INT |
8249 | || GET_MODE_CLASS (TYPE_MODE (type)) == MODE_FLOAT) | |
8250 | && GET_MODE_BITSIZE (TYPE_MODE (type)) <= BITS_PER_WORD | |
8251 | && store_bb == single_succ (load_bb) | |
8252 | && first_stmt (store_bb) == store | |
3ec11c49 | 8253 | && expand_omp_atomic_store (load_bb, addr, loaded_val, |
8254 | stored_val, index)) | |
2169f33b | 8255 | return; |
8256 | ||
cb7f680b | 8257 | /* When possible, use specialized atomic update functions. */ |
8258 | if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type)) | |
3ec11c49 | 8259 | && store_bb == single_succ (load_bb) |
8260 | && expand_omp_atomic_fetch_op (load_bb, addr, | |
8261 | loaded_val, stored_val, index)) | |
8262 | return; | |
cb7f680b | 8263 | |
8264 | /* If we don't have specialized __sync builtins, try and implement | |
8265 | as a compare and swap loop. */ | |
8266 | if (expand_omp_atomic_pipeline (load_bb, store_bb, addr, | |
8267 | loaded_val, stored_val, index)) | |
8268 | return; | |
8269 | } | |
8270 | } | |
8271 | ||
8272 | /* The ultimate fallback is wrapping the operation in a mutex. */ | |
8273 | expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val); | |
8274 | } | |
8275 | ||
1e8e9920 | 8276 | |
bc7bff74 | 8277 | /* Expand the OpenMP target{, data, update} directive starting at REGION. */ |
773c5ba7 | 8278 | |
8279 | static void | |
bc7bff74 | 8280 | expand_omp_target (struct omp_region *region) |
773c5ba7 | 8281 | { |
bc7bff74 | 8282 | basic_block entry_bb, exit_bb, new_bb; |
8283 | struct function *child_cfun = NULL; | |
8284 | tree child_fn = NULL_TREE, block, t; | |
8285 | gimple_stmt_iterator gsi; | |
1a91d914 | 8286 | gomp_target *entry_stmt; |
8287 | gimple stmt; | |
bc7bff74 | 8288 | edge e; |
8289 | ||
1a91d914 | 8290 | entry_stmt = as_a <gomp_target *> (last_stmt (region->entry)); |
bc7bff74 | 8291 | new_bb = region->entry; |
8292 | int kind = gimple_omp_target_kind (entry_stmt); | |
8293 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
8294 | { | |
8295 | child_fn = gimple_omp_target_child_fn (entry_stmt); | |
8296 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
8297 | } | |
8298 | ||
8299 | entry_bb = region->entry; | |
8300 | exit_bb = region->exit; | |
8301 | ||
8302 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
8303 | { | |
8304 | unsigned srcidx, dstidx, num; | |
8305 | ||
8306 | /* If the target region needs data sent from the parent | |
8307 | function, then the very first statement (except possible | |
8308 | tree profile counter updates) of the parallel body | |
8309 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
8310 | &.OMP_DATA_O is passed as an argument to the child function, | |
8311 | we need to replace it with the argument as seen by the child | |
8312 | function. | |
8313 | ||
8314 | In most cases, this will end up being the identity assignment | |
8315 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had | |
8316 | a function call that has been inlined, the original PARM_DECL | |
8317 | .OMP_DATA_I may have been converted into a different local | |
8318 | variable. In which case, we need to keep the assignment. */ | |
8319 | if (gimple_omp_target_data_arg (entry_stmt)) | |
8320 | { | |
8321 | basic_block entry_succ_bb = single_succ (entry_bb); | |
8322 | gimple_stmt_iterator gsi; | |
8323 | tree arg; | |
8324 | gimple tgtcopy_stmt = NULL; | |
8325 | tree sender | |
8326 | = TREE_VEC_ELT (gimple_omp_target_data_arg (entry_stmt), 0); | |
8327 | ||
8328 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
8329 | { | |
8330 | gcc_assert (!gsi_end_p (gsi)); | |
8331 | stmt = gsi_stmt (gsi); | |
8332 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
8333 | continue; | |
8334 | ||
8335 | if (gimple_num_ops (stmt) == 2) | |
8336 | { | |
8337 | tree arg = gimple_assign_rhs1 (stmt); | |
8338 | ||
8339 | /* We're ignoring the subcode because we're | |
8340 | effectively doing a STRIP_NOPS. */ | |
8341 | ||
8342 | if (TREE_CODE (arg) == ADDR_EXPR | |
8343 | && TREE_OPERAND (arg, 0) == sender) | |
8344 | { | |
8345 | tgtcopy_stmt = stmt; | |
8346 | break; | |
8347 | } | |
8348 | } | |
8349 | } | |
8350 | ||
8351 | gcc_assert (tgtcopy_stmt != NULL); | |
8352 | arg = DECL_ARGUMENTS (child_fn); | |
8353 | ||
8354 | gcc_assert (gimple_assign_lhs (tgtcopy_stmt) == arg); | |
8355 | gsi_remove (&gsi, true); | |
8356 | } | |
8357 | ||
8358 | /* Declare local variables needed in CHILD_CFUN. */ | |
8359 | block = DECL_INITIAL (child_fn); | |
8360 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
8361 | /* The gimplifier could record temporaries in target block | |
8362 | rather than in containing function's local_decls chain, | |
8363 | which would mean cgraph missed finalizing them. Do it now. */ | |
8364 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
8365 | if (TREE_CODE (t) == VAR_DECL | |
8366 | && TREE_STATIC (t) | |
8367 | && !DECL_EXTERNAL (t)) | |
97221fd7 | 8368 | varpool_node::finalize_decl (t); |
bc7bff74 | 8369 | DECL_SAVED_TREE (child_fn) = NULL; |
8370 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
8371 | gimple_set_body (child_fn, NULL); | |
8372 | TREE_USED (block) = 1; | |
8373 | ||
8374 | /* Reset DECL_CONTEXT on function arguments. */ | |
8375 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
8376 | DECL_CONTEXT (t) = child_fn; | |
8377 | ||
8378 | /* Split ENTRY_BB at GIMPLE_OMP_TARGET, | |
8379 | so that it can be moved to the child function. */ | |
8380 | gsi = gsi_last_bb (entry_bb); | |
8381 | stmt = gsi_stmt (gsi); | |
8382 | gcc_assert (stmt && gimple_code (stmt) == GIMPLE_OMP_TARGET | |
8383 | && gimple_omp_target_kind (stmt) | |
8384 | == GF_OMP_TARGET_KIND_REGION); | |
8385 | gsi_remove (&gsi, true); | |
8386 | e = split_block (entry_bb, stmt); | |
8387 | entry_bb = e->dest; | |
8388 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
8389 | ||
8390 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ | |
8391 | if (exit_bb) | |
8392 | { | |
8393 | gsi = gsi_last_bb (exit_bb); | |
8394 | gcc_assert (!gsi_end_p (gsi) | |
8395 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
8396 | stmt = gimple_build_return (NULL); | |
8397 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
8398 | gsi_remove (&gsi, true); | |
8399 | } | |
8400 | ||
8401 | /* Move the target region into CHILD_CFUN. */ | |
8402 | ||
8403 | block = gimple_block (entry_stmt); | |
8404 | ||
8405 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
8406 | if (exit_bb) | |
8407 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
8408 | /* When the OMP expansion process cannot guarantee an up-to-date | |
8409 | loop tree arrange for the child function to fixup loops. */ | |
8410 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
8411 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
8412 | ||
8413 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
8414 | num = vec_safe_length (child_cfun->local_decls); | |
8415 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
8416 | { | |
8417 | t = (*child_cfun->local_decls)[srcidx]; | |
8418 | if (DECL_CONTEXT (t) == cfun->decl) | |
8419 | continue; | |
8420 | if (srcidx != dstidx) | |
8421 | (*child_cfun->local_decls)[dstidx] = t; | |
8422 | dstidx++; | |
8423 | } | |
8424 | if (dstidx != num) | |
8425 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
8426 | ||
8427 | /* Inform the callgraph about the new function. */ | |
8428 | DECL_STRUCT_FUNCTION (child_fn)->curr_properties = cfun->curr_properties; | |
415d1b9a | 8429 | cgraph_node::add_new_function (child_fn, true); |
bc7bff74 | 8430 | |
7114ebdd | 8431 | #ifdef ENABLE_OFFLOADING |
dccabdd1 | 8432 | /* Add the new function to the offload table. */ |
8433 | vec_safe_push (offload_funcs, child_fn); | |
7114ebdd | 8434 | #endif |
dccabdd1 | 8435 | |
bc7bff74 | 8436 | /* Fix the callgraph edges for child_cfun. Those for cfun will be |
8437 | fixed in a following pass. */ | |
8438 | push_cfun (child_cfun); | |
35ee1c66 | 8439 | cgraph_edge::rebuild_edges (); |
bc7bff74 | 8440 | |
7114ebdd | 8441 | #ifdef ENABLE_OFFLOADING |
b0c5e347 | 8442 | /* Prevent IPA from removing child_fn as unreachable, since there are no |
8443 | refs from the parent function to child_fn in offload LTO mode. */ | |
7114ebdd | 8444 | struct cgraph_node *node = cgraph_node::get (child_fn); |
b0c5e347 | 8445 | node->mark_force_output (); |
7114ebdd | 8446 | #endif |
b0c5e347 | 8447 | |
bc7bff74 | 8448 | /* Some EH regions might become dead, see PR34608. If |
8449 | pass_cleanup_cfg isn't the first pass to happen with the | |
8450 | new child, these dead EH edges might cause problems. | |
8451 | Clean them up now. */ | |
8452 | if (flag_exceptions) | |
8453 | { | |
8454 | basic_block bb; | |
8455 | bool changed = false; | |
8456 | ||
fc00614f | 8457 | FOR_EACH_BB_FN (bb, cfun) |
bc7bff74 | 8458 | changed |= gimple_purge_dead_eh_edges (bb); |
8459 | if (changed) | |
8460 | cleanup_tree_cfg (); | |
8461 | } | |
8462 | pop_cfun (); | |
8463 | } | |
8464 | ||
8465 | /* Emit a library call to launch the target region, or do data | |
8466 | transfers. */ | |
8467 | tree t1, t2, t3, t4, device, cond, c, clauses; | |
8468 | enum built_in_function start_ix; | |
8469 | location_t clause_loc; | |
8470 | ||
8471 | clauses = gimple_omp_target_clauses (entry_stmt); | |
8472 | ||
8473 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
8474 | start_ix = BUILT_IN_GOMP_TARGET; | |
8475 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
8476 | start_ix = BUILT_IN_GOMP_TARGET_DATA; | |
8477 | else | |
8478 | start_ix = BUILT_IN_GOMP_TARGET_UPDATE; | |
8479 | ||
8480 | /* By default, the value of DEVICE is -1 (let runtime library choose) | |
8481 | and there is no conditional. */ | |
8482 | cond = NULL_TREE; | |
8483 | device = build_int_cst (integer_type_node, -1); | |
8484 | ||
8485 | c = find_omp_clause (clauses, OMP_CLAUSE_IF); | |
8486 | if (c) | |
8487 | cond = OMP_CLAUSE_IF_EXPR (c); | |
8488 | ||
8489 | c = find_omp_clause (clauses, OMP_CLAUSE_DEVICE); | |
8490 | if (c) | |
8491 | { | |
8492 | device = OMP_CLAUSE_DEVICE_ID (c); | |
8493 | clause_loc = OMP_CLAUSE_LOCATION (c); | |
8494 | } | |
8495 | else | |
8496 | clause_loc = gimple_location (entry_stmt); | |
8497 | ||
8498 | /* Ensure 'device' is of the correct type. */ | |
8499 | device = fold_convert_loc (clause_loc, integer_type_node, device); | |
8500 | ||
8501 | /* If we found the clause 'if (cond)', build | |
8502 | (cond ? device : -2). */ | |
8503 | if (cond) | |
8504 | { | |
8505 | cond = gimple_boolify (cond); | |
8506 | ||
8507 | basic_block cond_bb, then_bb, else_bb; | |
8508 | edge e; | |
8509 | tree tmp_var; | |
8510 | ||
f9e245b2 | 8511 | tmp_var = create_tmp_var (TREE_TYPE (device)); |
bc7bff74 | 8512 | if (kind != GF_OMP_TARGET_KIND_REGION) |
8513 | { | |
8514 | gsi = gsi_last_bb (new_bb); | |
8515 | gsi_prev (&gsi); | |
8516 | e = split_block (new_bb, gsi_stmt (gsi)); | |
8517 | } | |
8518 | else | |
8519 | e = split_block (new_bb, NULL); | |
8520 | cond_bb = e->src; | |
8521 | new_bb = e->dest; | |
8522 | remove_edge (e); | |
8523 | ||
8524 | then_bb = create_empty_bb (cond_bb); | |
8525 | else_bb = create_empty_bb (then_bb); | |
8526 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
8527 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
8528 | ||
8529 | stmt = gimple_build_cond_empty (cond); | |
8530 | gsi = gsi_last_bb (cond_bb); | |
8531 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
8532 | ||
8533 | gsi = gsi_start_bb (then_bb); | |
8534 | stmt = gimple_build_assign (tmp_var, device); | |
8535 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
8536 | ||
8537 | gsi = gsi_start_bb (else_bb); | |
8538 | stmt = gimple_build_assign (tmp_var, | |
8539 | build_int_cst (integer_type_node, -2)); | |
8540 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
8541 | ||
8542 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
8543 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
b3083327 | 8544 | add_bb_to_loop (then_bb, cond_bb->loop_father); |
8545 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
bc7bff74 | 8546 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); |
8547 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); | |
8548 | ||
8549 | device = tmp_var; | |
8550 | } | |
8551 | ||
8552 | gsi = gsi_last_bb (new_bb); | |
8553 | t = gimple_omp_target_data_arg (entry_stmt); | |
8554 | if (t == NULL) | |
8555 | { | |
8556 | t1 = size_zero_node; | |
8557 | t2 = build_zero_cst (ptr_type_node); | |
8558 | t3 = t2; | |
8559 | t4 = t2; | |
8560 | } | |
8561 | else | |
8562 | { | |
8563 | t1 = TYPE_MAX_VALUE (TYPE_DOMAIN (TREE_TYPE (TREE_VEC_ELT (t, 1)))); | |
8564 | t1 = size_binop (PLUS_EXPR, t1, size_int (1)); | |
8565 | t2 = build_fold_addr_expr (TREE_VEC_ELT (t, 0)); | |
8566 | t3 = build_fold_addr_expr (TREE_VEC_ELT (t, 1)); | |
8567 | t4 = build_fold_addr_expr (TREE_VEC_ELT (t, 2)); | |
8568 | } | |
8569 | ||
8570 | gimple g; | |
8571 | /* FIXME: This will be address of | |
8572 | extern char __OPENMP_TARGET__[] __attribute__((visibility ("hidden"))) | |
8573 | symbol, as soon as the linker plugin is able to create it for us. */ | |
8574 | tree openmp_target = build_zero_cst (ptr_type_node); | |
8575 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
8576 | { | |
8577 | tree fnaddr = build_fold_addr_expr (child_fn); | |
8578 | g = gimple_build_call (builtin_decl_explicit (start_ix), 7, | |
8579 | device, fnaddr, openmp_target, t1, t2, t3, t4); | |
8580 | } | |
8581 | else | |
8582 | g = gimple_build_call (builtin_decl_explicit (start_ix), 6, | |
8583 | device, openmp_target, t1, t2, t3, t4); | |
8584 | gimple_set_location (g, gimple_location (entry_stmt)); | |
8585 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
8586 | if (kind != GF_OMP_TARGET_KIND_REGION) | |
8587 | { | |
8588 | g = gsi_stmt (gsi); | |
8589 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_TARGET); | |
8590 | gsi_remove (&gsi, true); | |
8591 | } | |
8592 | if (kind == GF_OMP_TARGET_KIND_DATA && region->exit) | |
8593 | { | |
8594 | gsi = gsi_last_bb (region->exit); | |
8595 | g = gsi_stmt (gsi); | |
8596 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_RETURN); | |
8597 | gsi_remove (&gsi, true); | |
8598 | } | |
8599 | } | |
8600 | ||
8601 | ||
8602 | /* Expand the parallel region tree rooted at REGION. Expansion | |
8603 | proceeds in depth-first order. Innermost regions are expanded | |
8604 | first. This way, parallel regions that require a new function to | |
75a70cf9 | 8605 | be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any |
773c5ba7 | 8606 | internal dependencies in their body. */ |
8607 | ||
8608 | static void | |
8609 | expand_omp (struct omp_region *region) | |
8610 | { | |
8611 | while (region) | |
8612 | { | |
1d22f541 | 8613 | location_t saved_location; |
bc7bff74 | 8614 | gimple inner_stmt = NULL; |
1d22f541 | 8615 | |
d1d5b012 | 8616 | /* First, determine whether this is a combined parallel+workshare |
8617 | region. */ | |
75a70cf9 | 8618 | if (region->type == GIMPLE_OMP_PARALLEL) |
d1d5b012 | 8619 | determine_parallel_type (region); |
8620 | ||
bc7bff74 | 8621 | if (region->type == GIMPLE_OMP_FOR |
8622 | && gimple_omp_for_combined_p (last_stmt (region->entry))) | |
8623 | inner_stmt = last_stmt (region->inner->entry); | |
8624 | ||
773c5ba7 | 8625 | if (region->inner) |
8626 | expand_omp (region->inner); | |
8627 | ||
1d22f541 | 8628 | saved_location = input_location; |
75a70cf9 | 8629 | if (gimple_has_location (last_stmt (region->entry))) |
8630 | input_location = gimple_location (last_stmt (region->entry)); | |
1d22f541 | 8631 | |
61e47ac8 | 8632 | switch (region->type) |
773c5ba7 | 8633 | { |
75a70cf9 | 8634 | case GIMPLE_OMP_PARALLEL: |
8635 | case GIMPLE_OMP_TASK: | |
fd6481cf | 8636 | expand_omp_taskreg (region); |
8637 | break; | |
8638 | ||
75a70cf9 | 8639 | case GIMPLE_OMP_FOR: |
bc7bff74 | 8640 | expand_omp_for (region, inner_stmt); |
61e47ac8 | 8641 | break; |
773c5ba7 | 8642 | |
75a70cf9 | 8643 | case GIMPLE_OMP_SECTIONS: |
61e47ac8 | 8644 | expand_omp_sections (region); |
8645 | break; | |
773c5ba7 | 8646 | |
75a70cf9 | 8647 | case GIMPLE_OMP_SECTION: |
61e47ac8 | 8648 | /* Individual omp sections are handled together with their |
75a70cf9 | 8649 | parent GIMPLE_OMP_SECTIONS region. */ |
61e47ac8 | 8650 | break; |
773c5ba7 | 8651 | |
75a70cf9 | 8652 | case GIMPLE_OMP_SINGLE: |
61e47ac8 | 8653 | expand_omp_single (region); |
8654 | break; | |
773c5ba7 | 8655 | |
75a70cf9 | 8656 | case GIMPLE_OMP_MASTER: |
bc7bff74 | 8657 | case GIMPLE_OMP_TASKGROUP: |
75a70cf9 | 8658 | case GIMPLE_OMP_ORDERED: |
8659 | case GIMPLE_OMP_CRITICAL: | |
bc7bff74 | 8660 | case GIMPLE_OMP_TEAMS: |
61e47ac8 | 8661 | expand_omp_synch (region); |
8662 | break; | |
773c5ba7 | 8663 | |
75a70cf9 | 8664 | case GIMPLE_OMP_ATOMIC_LOAD: |
cb7f680b | 8665 | expand_omp_atomic (region); |
8666 | break; | |
8667 | ||
bc7bff74 | 8668 | case GIMPLE_OMP_TARGET: |
8669 | expand_omp_target (region); | |
8670 | break; | |
8671 | ||
61e47ac8 | 8672 | default: |
8673 | gcc_unreachable (); | |
8674 | } | |
cc5982dc | 8675 | |
1d22f541 | 8676 | input_location = saved_location; |
773c5ba7 | 8677 | region = region->next; |
8678 | } | |
8679 | } | |
8680 | ||
8681 | ||
8682 | /* Helper for build_omp_regions. Scan the dominator tree starting at | |
28c92cbb | 8683 | block BB. PARENT is the region that contains BB. If SINGLE_TREE is |
8684 | true, the function ends once a single tree is built (otherwise, whole | |
8685 | forest of OMP constructs may be built). */ | |
773c5ba7 | 8686 | |
8687 | static void | |
28c92cbb | 8688 | build_omp_regions_1 (basic_block bb, struct omp_region *parent, |
8689 | bool single_tree) | |
773c5ba7 | 8690 | { |
75a70cf9 | 8691 | gimple_stmt_iterator gsi; |
8692 | gimple stmt; | |
773c5ba7 | 8693 | basic_block son; |
8694 | ||
75a70cf9 | 8695 | gsi = gsi_last_bb (bb); |
8696 | if (!gsi_end_p (gsi) && is_gimple_omp (gsi_stmt (gsi))) | |
773c5ba7 | 8697 | { |
8698 | struct omp_region *region; | |
75a70cf9 | 8699 | enum gimple_code code; |
773c5ba7 | 8700 | |
75a70cf9 | 8701 | stmt = gsi_stmt (gsi); |
8702 | code = gimple_code (stmt); | |
8703 | if (code == GIMPLE_OMP_RETURN) | |
773c5ba7 | 8704 | { |
8705 | /* STMT is the return point out of region PARENT. Mark it | |
8706 | as the exit point and make PARENT the immediately | |
8707 | enclosing region. */ | |
8708 | gcc_assert (parent); | |
8709 | region = parent; | |
61e47ac8 | 8710 | region->exit = bb; |
773c5ba7 | 8711 | parent = parent->outer; |
773c5ba7 | 8712 | } |
75a70cf9 | 8713 | else if (code == GIMPLE_OMP_ATOMIC_STORE) |
cb7f680b | 8714 | { |
75a70cf9 | 8715 | /* GIMPLE_OMP_ATOMIC_STORE is analoguous to |
8716 | GIMPLE_OMP_RETURN, but matches with | |
8717 | GIMPLE_OMP_ATOMIC_LOAD. */ | |
cb7f680b | 8718 | gcc_assert (parent); |
75a70cf9 | 8719 | gcc_assert (parent->type == GIMPLE_OMP_ATOMIC_LOAD); |
cb7f680b | 8720 | region = parent; |
8721 | region->exit = bb; | |
8722 | parent = parent->outer; | |
8723 | } | |
8724 | ||
75a70cf9 | 8725 | else if (code == GIMPLE_OMP_CONTINUE) |
61e47ac8 | 8726 | { |
8727 | gcc_assert (parent); | |
8728 | parent->cont = bb; | |
8729 | } | |
75a70cf9 | 8730 | else if (code == GIMPLE_OMP_SECTIONS_SWITCH) |
ac6e3339 | 8731 | { |
75a70cf9 | 8732 | /* GIMPLE_OMP_SECTIONS_SWITCH is part of |
8733 | GIMPLE_OMP_SECTIONS, and we do nothing for it. */ | |
8734 | ; | |
ac6e3339 | 8735 | } |
bc7bff74 | 8736 | else if (code == GIMPLE_OMP_TARGET |
8737 | && gimple_omp_target_kind (stmt) == GF_OMP_TARGET_KIND_UPDATE) | |
8738 | new_omp_region (bb, code, parent); | |
773c5ba7 | 8739 | else |
8740 | { | |
8741 | /* Otherwise, this directive becomes the parent for a new | |
8742 | region. */ | |
61e47ac8 | 8743 | region = new_omp_region (bb, code, parent); |
773c5ba7 | 8744 | parent = region; |
8745 | } | |
773c5ba7 | 8746 | } |
8747 | ||
28c92cbb | 8748 | if (single_tree && !parent) |
8749 | return; | |
8750 | ||
773c5ba7 | 8751 | for (son = first_dom_son (CDI_DOMINATORS, bb); |
8752 | son; | |
8753 | son = next_dom_son (CDI_DOMINATORS, son)) | |
28c92cbb | 8754 | build_omp_regions_1 (son, parent, single_tree); |
8755 | } | |
8756 | ||
8757 | /* Builds the tree of OMP regions rooted at ROOT, storing it to | |
8758 | root_omp_region. */ | |
8759 | ||
8760 | static void | |
8761 | build_omp_regions_root (basic_block root) | |
8762 | { | |
8763 | gcc_assert (root_omp_region == NULL); | |
8764 | build_omp_regions_1 (root, NULL, true); | |
8765 | gcc_assert (root_omp_region != NULL); | |
773c5ba7 | 8766 | } |
8767 | ||
28c92cbb | 8768 | /* Expands omp construct (and its subconstructs) starting in HEAD. */ |
8769 | ||
8770 | void | |
8771 | omp_expand_local (basic_block head) | |
8772 | { | |
8773 | build_omp_regions_root (head); | |
8774 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
8775 | { | |
8776 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
8777 | dump_omp_region (dump_file, root_omp_region, 0); | |
8778 | fprintf (dump_file, "\n"); | |
8779 | } | |
8780 | ||
8781 | remove_exit_barriers (root_omp_region); | |
8782 | expand_omp (root_omp_region); | |
8783 | ||
8784 | free_omp_regions (); | |
8785 | } | |
773c5ba7 | 8786 | |
8787 | /* Scan the CFG and build a tree of OMP regions. Return the root of | |
8788 | the OMP region tree. */ | |
8789 | ||
8790 | static void | |
8791 | build_omp_regions (void) | |
8792 | { | |
61e47ac8 | 8793 | gcc_assert (root_omp_region == NULL); |
773c5ba7 | 8794 | calculate_dominance_info (CDI_DOMINATORS); |
34154e27 | 8795 | build_omp_regions_1 (ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, false); |
773c5ba7 | 8796 | } |
8797 | ||
773c5ba7 | 8798 | /* Main entry point for expanding OMP-GIMPLE into runtime calls. */ |
8799 | ||
2a1990e9 | 8800 | static unsigned int |
773c5ba7 | 8801 | execute_expand_omp (void) |
8802 | { | |
8803 | build_omp_regions (); | |
8804 | ||
61e47ac8 | 8805 | if (!root_omp_region) |
8806 | return 0; | |
773c5ba7 | 8807 | |
61e47ac8 | 8808 | if (dump_file) |
8809 | { | |
8810 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
8811 | dump_omp_region (dump_file, root_omp_region, 0); | |
8812 | fprintf (dump_file, "\n"); | |
773c5ba7 | 8813 | } |
61e47ac8 | 8814 | |
8815 | remove_exit_barriers (root_omp_region); | |
8816 | ||
8817 | expand_omp (root_omp_region); | |
8818 | ||
61e47ac8 | 8819 | cleanup_tree_cfg (); |
8820 | ||
8821 | free_omp_regions (); | |
8822 | ||
2a1990e9 | 8823 | return 0; |
773c5ba7 | 8824 | } |
8825 | ||
79acaae1 | 8826 | /* OMP expansion -- the default pass, run before creation of SSA form. */ |
8827 | ||
cbe8bda8 | 8828 | namespace { |
8829 | ||
8830 | const pass_data pass_data_expand_omp = | |
8831 | { | |
8832 | GIMPLE_PASS, /* type */ | |
8833 | "ompexp", /* name */ | |
8834 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 8835 | TV_NONE, /* tv_id */ |
8836 | PROP_gimple_any, /* properties_required */ | |
8917c50b | 8837 | PROP_gimple_eomp, /* properties_provided */ |
cbe8bda8 | 8838 | 0, /* properties_destroyed */ |
8839 | 0, /* todo_flags_start */ | |
8840 | 0, /* todo_flags_finish */ | |
773c5ba7 | 8841 | }; |
cbe8bda8 | 8842 | |
8843 | class pass_expand_omp : public gimple_opt_pass | |
8844 | { | |
8845 | public: | |
9af5ce0c | 8846 | pass_expand_omp (gcc::context *ctxt) |
8847 | : gimple_opt_pass (pass_data_expand_omp, ctxt) | |
cbe8bda8 | 8848 | {} |
8849 | ||
8850 | /* opt_pass methods: */ | |
8917c50b | 8851 | virtual unsigned int execute (function *) |
31315c24 | 8852 | { |
8917c50b | 8853 | bool gate = ((flag_openmp != 0 || flag_openmp_simd != 0 |
8854 | || flag_cilkplus != 0) && !seen_error ()); | |
31315c24 | 8855 | |
8917c50b | 8856 | /* This pass always runs, to provide PROP_gimple_eomp. |
8857 | But there is nothing to do unless -fopenmp is given. */ | |
8858 | if (!gate) | |
8859 | return 0; | |
8860 | ||
8861 | return execute_expand_omp (); | |
8862 | } | |
cbe8bda8 | 8863 | |
8864 | }; // class pass_expand_omp | |
8865 | ||
8866 | } // anon namespace | |
8867 | ||
8868 | gimple_opt_pass * | |
8869 | make_pass_expand_omp (gcc::context *ctxt) | |
8870 | { | |
8871 | return new pass_expand_omp (ctxt); | |
8872 | } | |
8917c50b | 8873 | |
8874 | namespace { | |
8875 | ||
8876 | const pass_data pass_data_expand_omp_ssa = | |
8877 | { | |
8878 | GIMPLE_PASS, /* type */ | |
8879 | "ompexpssa", /* name */ | |
8880 | OPTGROUP_NONE, /* optinfo_flags */ | |
8881 | TV_NONE, /* tv_id */ | |
8882 | PROP_cfg | PROP_ssa, /* properties_required */ | |
8883 | PROP_gimple_eomp, /* properties_provided */ | |
8884 | 0, /* properties_destroyed */ | |
8885 | 0, /* todo_flags_start */ | |
8886 | TODO_cleanup_cfg | TODO_rebuild_alias, /* todo_flags_finish */ | |
8887 | }; | |
8888 | ||
8889 | class pass_expand_omp_ssa : public gimple_opt_pass | |
8890 | { | |
8891 | public: | |
8892 | pass_expand_omp_ssa (gcc::context *ctxt) | |
8893 | : gimple_opt_pass (pass_data_expand_omp_ssa, ctxt) | |
8894 | {} | |
8895 | ||
8896 | /* opt_pass methods: */ | |
8897 | virtual bool gate (function *fun) | |
8898 | { | |
8899 | return !(fun->curr_properties & PROP_gimple_eomp); | |
8900 | } | |
8901 | virtual unsigned int execute (function *) { return execute_expand_omp (); } | |
8902 | ||
8903 | }; // class pass_expand_omp_ssa | |
8904 | ||
8905 | } // anon namespace | |
8906 | ||
8907 | gimple_opt_pass * | |
8908 | make_pass_expand_omp_ssa (gcc::context *ctxt) | |
8909 | { | |
8910 | return new pass_expand_omp_ssa (ctxt); | |
8911 | } | |
773c5ba7 | 8912 | \f |
8913 | /* Routines to lower OpenMP directives into OMP-GIMPLE. */ | |
8914 | ||
bc7bff74 | 8915 | /* If ctx is a worksharing context inside of a cancellable parallel |
8916 | region and it isn't nowait, add lhs to its GIMPLE_OMP_RETURN | |
8917 | and conditional branch to parallel's cancel_label to handle | |
8918 | cancellation in the implicit barrier. */ | |
8919 | ||
8920 | static void | |
8921 | maybe_add_implicit_barrier_cancel (omp_context *ctx, gimple_seq *body) | |
8922 | { | |
8923 | gimple omp_return = gimple_seq_last_stmt (*body); | |
8924 | gcc_assert (gimple_code (omp_return) == GIMPLE_OMP_RETURN); | |
8925 | if (gimple_omp_return_nowait_p (omp_return)) | |
8926 | return; | |
8927 | if (ctx->outer | |
8928 | && gimple_code (ctx->outer->stmt) == GIMPLE_OMP_PARALLEL | |
8929 | && ctx->outer->cancellable) | |
8930 | { | |
15b28553 | 8931 | tree fndecl = builtin_decl_explicit (BUILT_IN_GOMP_CANCEL); |
8932 | tree c_bool_type = TREE_TYPE (TREE_TYPE (fndecl)); | |
f9e245b2 | 8933 | tree lhs = create_tmp_var (c_bool_type); |
bc7bff74 | 8934 | gimple_omp_return_set_lhs (omp_return, lhs); |
8935 | tree fallthru_label = create_artificial_label (UNKNOWN_LOCATION); | |
15b28553 | 8936 | gimple g = gimple_build_cond (NE_EXPR, lhs, |
8937 | fold_convert (c_bool_type, | |
8938 | boolean_false_node), | |
bc7bff74 | 8939 | ctx->outer->cancel_label, fallthru_label); |
8940 | gimple_seq_add_stmt (body, g); | |
8941 | gimple_seq_add_stmt (body, gimple_build_label (fallthru_label)); | |
8942 | } | |
8943 | } | |
8944 | ||
75a70cf9 | 8945 | /* Lower the OpenMP sections directive in the current statement in GSI_P. |
8946 | CTX is the enclosing OMP context for the current statement. */ | |
773c5ba7 | 8947 | |
8948 | static void | |
75a70cf9 | 8949 | lower_omp_sections (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
773c5ba7 | 8950 | { |
75a70cf9 | 8951 | tree block, control; |
8952 | gimple_stmt_iterator tgsi; | |
1a91d914 | 8953 | gomp_sections *stmt; |
8954 | gimple t; | |
8955 | gbind *new_stmt, *bind; | |
e3a19533 | 8956 | gimple_seq ilist, dlist, olist, new_body; |
773c5ba7 | 8957 | |
1a91d914 | 8958 | stmt = as_a <gomp_sections *> (gsi_stmt (*gsi_p)); |
773c5ba7 | 8959 | |
8a4a28a8 | 8960 | push_gimplify_context (); |
773c5ba7 | 8961 | |
8962 | dlist = NULL; | |
8963 | ilist = NULL; | |
75a70cf9 | 8964 | lower_rec_input_clauses (gimple_omp_sections_clauses (stmt), |
bc7bff74 | 8965 | &ilist, &dlist, ctx, NULL); |
773c5ba7 | 8966 | |
e3a19533 | 8967 | new_body = gimple_omp_body (stmt); |
8968 | gimple_omp_set_body (stmt, NULL); | |
8969 | tgsi = gsi_start (new_body); | |
8970 | for (; !gsi_end_p (tgsi); gsi_next (&tgsi)) | |
773c5ba7 | 8971 | { |
8972 | omp_context *sctx; | |
75a70cf9 | 8973 | gimple sec_start; |
773c5ba7 | 8974 | |
75a70cf9 | 8975 | sec_start = gsi_stmt (tgsi); |
773c5ba7 | 8976 | sctx = maybe_lookup_ctx (sec_start); |
8977 | gcc_assert (sctx); | |
8978 | ||
e3a19533 | 8979 | lower_omp (gimple_omp_body_ptr (sec_start), sctx); |
8980 | gsi_insert_seq_after (&tgsi, gimple_omp_body (sec_start), | |
8981 | GSI_CONTINUE_LINKING); | |
75a70cf9 | 8982 | gimple_omp_set_body (sec_start, NULL); |
773c5ba7 | 8983 | |
e3a19533 | 8984 | if (gsi_one_before_end_p (tgsi)) |
773c5ba7 | 8985 | { |
75a70cf9 | 8986 | gimple_seq l = NULL; |
8987 | lower_lastprivate_clauses (gimple_omp_sections_clauses (stmt), NULL, | |
773c5ba7 | 8988 | &l, ctx); |
e3a19533 | 8989 | gsi_insert_seq_after (&tgsi, l, GSI_CONTINUE_LINKING); |
75a70cf9 | 8990 | gimple_omp_section_set_last (sec_start); |
773c5ba7 | 8991 | } |
48e1416a | 8992 | |
e3a19533 | 8993 | gsi_insert_after (&tgsi, gimple_build_omp_return (false), |
8994 | GSI_CONTINUE_LINKING); | |
773c5ba7 | 8995 | } |
1e8e9920 | 8996 | |
8997 | block = make_node (BLOCK); | |
e3a19533 | 8998 | bind = gimple_build_bind (NULL, new_body, block); |
1e8e9920 | 8999 | |
75a70cf9 | 9000 | olist = NULL; |
9001 | lower_reduction_clauses (gimple_omp_sections_clauses (stmt), &olist, ctx); | |
773c5ba7 | 9002 | |
1d22f541 | 9003 | block = make_node (BLOCK); |
75a70cf9 | 9004 | new_stmt = gimple_build_bind (NULL, NULL, block); |
e3a19533 | 9005 | gsi_replace (gsi_p, new_stmt, true); |
773c5ba7 | 9006 | |
1d22f541 | 9007 | pop_gimplify_context (new_stmt); |
75a70cf9 | 9008 | gimple_bind_append_vars (new_stmt, ctx->block_vars); |
9009 | BLOCK_VARS (block) = gimple_bind_vars (bind); | |
1d22f541 | 9010 | if (BLOCK_VARS (block)) |
9011 | TREE_USED (block) = 1; | |
9012 | ||
75a70cf9 | 9013 | new_body = NULL; |
9014 | gimple_seq_add_seq (&new_body, ilist); | |
9015 | gimple_seq_add_stmt (&new_body, stmt); | |
9016 | gimple_seq_add_stmt (&new_body, gimple_build_omp_sections_switch ()); | |
9017 | gimple_seq_add_stmt (&new_body, bind); | |
61e47ac8 | 9018 | |
ac6e3339 | 9019 | control = create_tmp_var (unsigned_type_node, ".section"); |
75a70cf9 | 9020 | t = gimple_build_omp_continue (control, control); |
9021 | gimple_omp_sections_set_control (stmt, control); | |
9022 | gimple_seq_add_stmt (&new_body, t); | |
61e47ac8 | 9023 | |
75a70cf9 | 9024 | gimple_seq_add_seq (&new_body, olist); |
bc7bff74 | 9025 | if (ctx->cancellable) |
9026 | gimple_seq_add_stmt (&new_body, gimple_build_label (ctx->cancel_label)); | |
75a70cf9 | 9027 | gimple_seq_add_seq (&new_body, dlist); |
773c5ba7 | 9028 | |
75a70cf9 | 9029 | new_body = maybe_catch_exception (new_body); |
aade31a0 | 9030 | |
75a70cf9 | 9031 | t = gimple_build_omp_return |
9032 | (!!find_omp_clause (gimple_omp_sections_clauses (stmt), | |
9033 | OMP_CLAUSE_NOWAIT)); | |
9034 | gimple_seq_add_stmt (&new_body, t); | |
bc7bff74 | 9035 | maybe_add_implicit_barrier_cancel (ctx, &new_body); |
61e47ac8 | 9036 | |
75a70cf9 | 9037 | gimple_bind_set_body (new_stmt, new_body); |
1e8e9920 | 9038 | } |
9039 | ||
9040 | ||
773c5ba7 | 9041 | /* A subroutine of lower_omp_single. Expand the simple form of |
75a70cf9 | 9042 | a GIMPLE_OMP_SINGLE, without a copyprivate clause: |
1e8e9920 | 9043 | |
9044 | if (GOMP_single_start ()) | |
9045 | BODY; | |
9046 | [ GOMP_barrier (); ] -> unless 'nowait' is present. | |
773c5ba7 | 9047 | |
9048 | FIXME. It may be better to delay expanding the logic of this until | |
9049 | pass_expand_omp. The expanded logic may make the job more difficult | |
9050 | to a synchronization analysis pass. */ | |
1e8e9920 | 9051 | |
9052 | static void | |
1a91d914 | 9053 | lower_omp_single_simple (gomp_single *single_stmt, gimple_seq *pre_p) |
1e8e9920 | 9054 | { |
e60a6f7b | 9055 | location_t loc = gimple_location (single_stmt); |
9056 | tree tlabel = create_artificial_label (loc); | |
9057 | tree flabel = create_artificial_label (loc); | |
75a70cf9 | 9058 | gimple call, cond; |
9059 | tree lhs, decl; | |
9060 | ||
b9a16870 | 9061 | decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_START); |
f9e245b2 | 9062 | lhs = create_tmp_var (TREE_TYPE (TREE_TYPE (decl))); |
75a70cf9 | 9063 | call = gimple_build_call (decl, 0); |
9064 | gimple_call_set_lhs (call, lhs); | |
9065 | gimple_seq_add_stmt (pre_p, call); | |
9066 | ||
9067 | cond = gimple_build_cond (EQ_EXPR, lhs, | |
389dd41b | 9068 | fold_convert_loc (loc, TREE_TYPE (lhs), |
9069 | boolean_true_node), | |
75a70cf9 | 9070 | tlabel, flabel); |
9071 | gimple_seq_add_stmt (pre_p, cond); | |
9072 | gimple_seq_add_stmt (pre_p, gimple_build_label (tlabel)); | |
9073 | gimple_seq_add_seq (pre_p, gimple_omp_body (single_stmt)); | |
9074 | gimple_seq_add_stmt (pre_p, gimple_build_label (flabel)); | |
1e8e9920 | 9075 | } |
9076 | ||
773c5ba7 | 9077 | |
9078 | /* A subroutine of lower_omp_single. Expand the simple form of | |
75a70cf9 | 9079 | a GIMPLE_OMP_SINGLE, with a copyprivate clause: |
1e8e9920 | 9080 | |
9081 | #pragma omp single copyprivate (a, b, c) | |
9082 | ||
9083 | Create a new structure to hold copies of 'a', 'b' and 'c' and emit: | |
9084 | ||
9085 | { | |
9086 | if ((copyout_p = GOMP_single_copy_start ()) == NULL) | |
9087 | { | |
9088 | BODY; | |
9089 | copyout.a = a; | |
9090 | copyout.b = b; | |
9091 | copyout.c = c; | |
9092 | GOMP_single_copy_end (©out); | |
9093 | } | |
9094 | else | |
9095 | { | |
9096 | a = copyout_p->a; | |
9097 | b = copyout_p->b; | |
9098 | c = copyout_p->c; | |
9099 | } | |
9100 | GOMP_barrier (); | |
9101 | } | |
773c5ba7 | 9102 | |
9103 | FIXME. It may be better to delay expanding the logic of this until | |
9104 | pass_expand_omp. The expanded logic may make the job more difficult | |
9105 | to a synchronization analysis pass. */ | |
1e8e9920 | 9106 | |
9107 | static void | |
1a91d914 | 9108 | lower_omp_single_copy (gomp_single *single_stmt, gimple_seq *pre_p, |
9109 | omp_context *ctx) | |
1e8e9920 | 9110 | { |
b9a16870 | 9111 | tree ptr_type, t, l0, l1, l2, bfn_decl; |
75a70cf9 | 9112 | gimple_seq copyin_seq; |
e60a6f7b | 9113 | location_t loc = gimple_location (single_stmt); |
1e8e9920 | 9114 | |
9115 | ctx->sender_decl = create_tmp_var (ctx->record_type, ".omp_copy_o"); | |
9116 | ||
9117 | ptr_type = build_pointer_type (ctx->record_type); | |
9118 | ctx->receiver_decl = create_tmp_var (ptr_type, ".omp_copy_i"); | |
9119 | ||
e60a6f7b | 9120 | l0 = create_artificial_label (loc); |
9121 | l1 = create_artificial_label (loc); | |
9122 | l2 = create_artificial_label (loc); | |
1e8e9920 | 9123 | |
b9a16870 | 9124 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_COPY_START); |
9125 | t = build_call_expr_loc (loc, bfn_decl, 0); | |
389dd41b | 9126 | t = fold_convert_loc (loc, ptr_type, t); |
75a70cf9 | 9127 | gimplify_assign (ctx->receiver_decl, t, pre_p); |
1e8e9920 | 9128 | |
9129 | t = build2 (EQ_EXPR, boolean_type_node, ctx->receiver_decl, | |
9130 | build_int_cst (ptr_type, 0)); | |
9131 | t = build3 (COND_EXPR, void_type_node, t, | |
9132 | build_and_jump (&l0), build_and_jump (&l1)); | |
9133 | gimplify_and_add (t, pre_p); | |
9134 | ||
75a70cf9 | 9135 | gimple_seq_add_stmt (pre_p, gimple_build_label (l0)); |
1e8e9920 | 9136 | |
75a70cf9 | 9137 | gimple_seq_add_seq (pre_p, gimple_omp_body (single_stmt)); |
1e8e9920 | 9138 | |
9139 | copyin_seq = NULL; | |
75a70cf9 | 9140 | lower_copyprivate_clauses (gimple_omp_single_clauses (single_stmt), pre_p, |
1e8e9920 | 9141 | ©in_seq, ctx); |
9142 | ||
389dd41b | 9143 | t = build_fold_addr_expr_loc (loc, ctx->sender_decl); |
b9a16870 | 9144 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_COPY_END); |
9145 | t = build_call_expr_loc (loc, bfn_decl, 1, t); | |
1e8e9920 | 9146 | gimplify_and_add (t, pre_p); |
9147 | ||
9148 | t = build_and_jump (&l2); | |
9149 | gimplify_and_add (t, pre_p); | |
9150 | ||
75a70cf9 | 9151 | gimple_seq_add_stmt (pre_p, gimple_build_label (l1)); |
1e8e9920 | 9152 | |
75a70cf9 | 9153 | gimple_seq_add_seq (pre_p, copyin_seq); |
1e8e9920 | 9154 | |
75a70cf9 | 9155 | gimple_seq_add_stmt (pre_p, gimple_build_label (l2)); |
1e8e9920 | 9156 | } |
9157 | ||
773c5ba7 | 9158 | |
1e8e9920 | 9159 | /* Expand code for an OpenMP single directive. */ |
9160 | ||
9161 | static void | |
75a70cf9 | 9162 | lower_omp_single (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 9163 | { |
75a70cf9 | 9164 | tree block; |
1a91d914 | 9165 | gimple t; |
9166 | gomp_single *single_stmt = as_a <gomp_single *> (gsi_stmt (*gsi_p)); | |
9167 | gbind *bind; | |
bc7bff74 | 9168 | gimple_seq bind_body, bind_body_tail = NULL, dlist; |
1e8e9920 | 9169 | |
8a4a28a8 | 9170 | push_gimplify_context (); |
1e8e9920 | 9171 | |
e3a19533 | 9172 | block = make_node (BLOCK); |
9173 | bind = gimple_build_bind (NULL, NULL, block); | |
9174 | gsi_replace (gsi_p, bind, true); | |
75a70cf9 | 9175 | bind_body = NULL; |
e3a19533 | 9176 | dlist = NULL; |
75a70cf9 | 9177 | lower_rec_input_clauses (gimple_omp_single_clauses (single_stmt), |
bc7bff74 | 9178 | &bind_body, &dlist, ctx, NULL); |
e3a19533 | 9179 | lower_omp (gimple_omp_body_ptr (single_stmt), ctx); |
1e8e9920 | 9180 | |
75a70cf9 | 9181 | gimple_seq_add_stmt (&bind_body, single_stmt); |
1e8e9920 | 9182 | |
9183 | if (ctx->record_type) | |
75a70cf9 | 9184 | lower_omp_single_copy (single_stmt, &bind_body, ctx); |
1e8e9920 | 9185 | else |
75a70cf9 | 9186 | lower_omp_single_simple (single_stmt, &bind_body); |
9187 | ||
9188 | gimple_omp_set_body (single_stmt, NULL); | |
1e8e9920 | 9189 | |
75a70cf9 | 9190 | gimple_seq_add_seq (&bind_body, dlist); |
61e47ac8 | 9191 | |
75a70cf9 | 9192 | bind_body = maybe_catch_exception (bind_body); |
61e47ac8 | 9193 | |
48e1416a | 9194 | t = gimple_build_omp_return |
75a70cf9 | 9195 | (!!find_omp_clause (gimple_omp_single_clauses (single_stmt), |
9196 | OMP_CLAUSE_NOWAIT)); | |
bc7bff74 | 9197 | gimple_seq_add_stmt (&bind_body_tail, t); |
9198 | maybe_add_implicit_barrier_cancel (ctx, &bind_body_tail); | |
9199 | if (ctx->record_type) | |
9200 | { | |
9201 | gimple_stmt_iterator gsi = gsi_start (bind_body_tail); | |
9202 | tree clobber = build_constructor (ctx->record_type, NULL); | |
9203 | TREE_THIS_VOLATILE (clobber) = 1; | |
9204 | gsi_insert_after (&gsi, gimple_build_assign (ctx->sender_decl, | |
9205 | clobber), GSI_SAME_STMT); | |
9206 | } | |
9207 | gimple_seq_add_seq (&bind_body, bind_body_tail); | |
e3a19533 | 9208 | gimple_bind_set_body (bind, bind_body); |
61e47ac8 | 9209 | |
1e8e9920 | 9210 | pop_gimplify_context (bind); |
773c5ba7 | 9211 | |
75a70cf9 | 9212 | gimple_bind_append_vars (bind, ctx->block_vars); |
9213 | BLOCK_VARS (block) = ctx->block_vars; | |
1d22f541 | 9214 | if (BLOCK_VARS (block)) |
9215 | TREE_USED (block) = 1; | |
1e8e9920 | 9216 | } |
9217 | ||
773c5ba7 | 9218 | |
1e8e9920 | 9219 | /* Expand code for an OpenMP master directive. */ |
9220 | ||
9221 | static void | |
75a70cf9 | 9222 | lower_omp_master (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 9223 | { |
b9a16870 | 9224 | tree block, lab = NULL, x, bfn_decl; |
1a91d914 | 9225 | gimple stmt = gsi_stmt (*gsi_p); |
9226 | gbind *bind; | |
389dd41b | 9227 | location_t loc = gimple_location (stmt); |
75a70cf9 | 9228 | gimple_seq tseq; |
1e8e9920 | 9229 | |
8a4a28a8 | 9230 | push_gimplify_context (); |
1e8e9920 | 9231 | |
9232 | block = make_node (BLOCK); | |
e3a19533 | 9233 | bind = gimple_build_bind (NULL, NULL, block); |
9234 | gsi_replace (gsi_p, bind, true); | |
9235 | gimple_bind_add_stmt (bind, stmt); | |
61e47ac8 | 9236 | |
b9a16870 | 9237 | bfn_decl = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); |
9238 | x = build_call_expr_loc (loc, bfn_decl, 0); | |
1e8e9920 | 9239 | x = build2 (EQ_EXPR, boolean_type_node, x, integer_zero_node); |
9240 | x = build3 (COND_EXPR, void_type_node, x, NULL, build_and_jump (&lab)); | |
75a70cf9 | 9241 | tseq = NULL; |
9242 | gimplify_and_add (x, &tseq); | |
9243 | gimple_bind_add_seq (bind, tseq); | |
1e8e9920 | 9244 | |
e3a19533 | 9245 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
75a70cf9 | 9246 | gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt))); |
9247 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
9248 | gimple_omp_set_body (stmt, NULL); | |
1e8e9920 | 9249 | |
75a70cf9 | 9250 | gimple_bind_add_stmt (bind, gimple_build_label (lab)); |
61e47ac8 | 9251 | |
75a70cf9 | 9252 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); |
61e47ac8 | 9253 | |
1e8e9920 | 9254 | pop_gimplify_context (bind); |
773c5ba7 | 9255 | |
75a70cf9 | 9256 | gimple_bind_append_vars (bind, ctx->block_vars); |
9257 | BLOCK_VARS (block) = ctx->block_vars; | |
1e8e9920 | 9258 | } |
9259 | ||
773c5ba7 | 9260 | |
bc7bff74 | 9261 | /* Expand code for an OpenMP taskgroup directive. */ |
9262 | ||
9263 | static void | |
9264 | lower_omp_taskgroup (gimple_stmt_iterator *gsi_p, omp_context *ctx) | |
9265 | { | |
1a91d914 | 9266 | gimple stmt = gsi_stmt (*gsi_p); |
9267 | gcall *x; | |
9268 | gbind *bind; | |
bc7bff74 | 9269 | tree block = make_node (BLOCK); |
9270 | ||
9271 | bind = gimple_build_bind (NULL, NULL, block); | |
9272 | gsi_replace (gsi_p, bind, true); | |
9273 | gimple_bind_add_stmt (bind, stmt); | |
9274 | ||
9275 | x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_TASKGROUP_START), | |
9276 | 0); | |
9277 | gimple_bind_add_stmt (bind, x); | |
9278 | ||
9279 | lower_omp (gimple_omp_body_ptr (stmt), ctx); | |
9280 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
9281 | gimple_omp_set_body (stmt, NULL); | |
9282 | ||
9283 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); | |
9284 | ||
9285 | gimple_bind_append_vars (bind, ctx->block_vars); | |
9286 | BLOCK_VARS (block) = ctx->block_vars; | |
9287 | } | |
9288 | ||
9289 | ||
1e8e9920 | 9290 | /* Expand code for an OpenMP ordered directive. */ |
9291 | ||
9292 | static void | |
75a70cf9 | 9293 | lower_omp_ordered (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 9294 | { |
75a70cf9 | 9295 | tree block; |
1a91d914 | 9296 | gimple stmt = gsi_stmt (*gsi_p); |
9297 | gcall *x; | |
9298 | gbind *bind; | |
1e8e9920 | 9299 | |
8a4a28a8 | 9300 | push_gimplify_context (); |
1e8e9920 | 9301 | |
9302 | block = make_node (BLOCK); | |
e3a19533 | 9303 | bind = gimple_build_bind (NULL, NULL, block); |
9304 | gsi_replace (gsi_p, bind, true); | |
9305 | gimple_bind_add_stmt (bind, stmt); | |
61e47ac8 | 9306 | |
b9a16870 | 9307 | x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ORDERED_START), |
9308 | 0); | |
75a70cf9 | 9309 | gimple_bind_add_stmt (bind, x); |
1e8e9920 | 9310 | |
e3a19533 | 9311 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
75a70cf9 | 9312 | gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt))); |
9313 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
9314 | gimple_omp_set_body (stmt, NULL); | |
1e8e9920 | 9315 | |
b9a16870 | 9316 | x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ORDERED_END), 0); |
75a70cf9 | 9317 | gimple_bind_add_stmt (bind, x); |
61e47ac8 | 9318 | |
75a70cf9 | 9319 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); |
61e47ac8 | 9320 | |
1e8e9920 | 9321 | pop_gimplify_context (bind); |
773c5ba7 | 9322 | |
75a70cf9 | 9323 | gimple_bind_append_vars (bind, ctx->block_vars); |
9324 | BLOCK_VARS (block) = gimple_bind_vars (bind); | |
1e8e9920 | 9325 | } |
9326 | ||
1e8e9920 | 9327 | |
75a70cf9 | 9328 | /* Gimplify a GIMPLE_OMP_CRITICAL statement. This is a relatively simple |
1e8e9920 | 9329 | substitution of a couple of function calls. But in the NAMED case, |
9330 | requires that languages coordinate a symbol name. It is therefore | |
9331 | best put here in common code. */ | |
9332 | ||
d4786b13 | 9333 | static GTY(()) hash_map<tree, tree> *critical_name_mutexes; |
1e8e9920 | 9334 | |
9335 | static void | |
75a70cf9 | 9336 | lower_omp_critical (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 9337 | { |
75a70cf9 | 9338 | tree block; |
9339 | tree name, lock, unlock; | |
1a91d914 | 9340 | gomp_critical *stmt = as_a <gomp_critical *> (gsi_stmt (*gsi_p)); |
9341 | gbind *bind; | |
389dd41b | 9342 | location_t loc = gimple_location (stmt); |
75a70cf9 | 9343 | gimple_seq tbody; |
1e8e9920 | 9344 | |
75a70cf9 | 9345 | name = gimple_omp_critical_name (stmt); |
1e8e9920 | 9346 | if (name) |
9347 | { | |
c2f47e15 | 9348 | tree decl; |
1e8e9920 | 9349 | |
9350 | if (!critical_name_mutexes) | |
d4786b13 | 9351 | critical_name_mutexes = hash_map<tree, tree>::create_ggc (10); |
1e8e9920 | 9352 | |
d4786b13 | 9353 | tree *n = critical_name_mutexes->get (name); |
1e8e9920 | 9354 | if (n == NULL) |
9355 | { | |
9356 | char *new_str; | |
9357 | ||
f9e245b2 | 9358 | decl = create_tmp_var_raw (ptr_type_node); |
1e8e9920 | 9359 | |
9360 | new_str = ACONCAT ((".gomp_critical_user_", | |
9361 | IDENTIFIER_POINTER (name), NULL)); | |
9362 | DECL_NAME (decl) = get_identifier (new_str); | |
9363 | TREE_PUBLIC (decl) = 1; | |
9364 | TREE_STATIC (decl) = 1; | |
9365 | DECL_COMMON (decl) = 1; | |
9366 | DECL_ARTIFICIAL (decl) = 1; | |
9367 | DECL_IGNORED_P (decl) = 1; | |
b0c5e347 | 9368 | |
97221fd7 | 9369 | varpool_node::finalize_decl (decl); |
1e8e9920 | 9370 | |
d4786b13 | 9371 | critical_name_mutexes->put (name, decl); |
1e8e9920 | 9372 | } |
9373 | else | |
d4786b13 | 9374 | decl = *n; |
1e8e9920 | 9375 | |
351fb0ce | 9376 | /* If '#pragma omp critical' is inside target region or |
9377 | inside function marked as offloadable, the symbol must be | |
9378 | marked as offloadable too. */ | |
9379 | omp_context *octx; | |
9380 | if (cgraph_node::get (current_function_decl)->offloadable) | |
9381 | varpool_node::get_create (decl)->offloadable = 1; | |
9382 | else | |
9383 | for (octx = ctx->outer; octx; octx = octx->outer) | |
9384 | if (is_targetreg_ctx (octx)) | |
9385 | { | |
9386 | varpool_node::get_create (decl)->offloadable = 1; | |
9387 | break; | |
9388 | } | |
9389 | ||
b9a16870 | 9390 | lock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_NAME_START); |
389dd41b | 9391 | lock = build_call_expr_loc (loc, lock, 1, build_fold_addr_expr_loc (loc, decl)); |
1e8e9920 | 9392 | |
b9a16870 | 9393 | unlock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_NAME_END); |
389dd41b | 9394 | unlock = build_call_expr_loc (loc, unlock, 1, |
9395 | build_fold_addr_expr_loc (loc, decl)); | |
1e8e9920 | 9396 | } |
9397 | else | |
9398 | { | |
b9a16870 | 9399 | lock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_START); |
389dd41b | 9400 | lock = build_call_expr_loc (loc, lock, 0); |
1e8e9920 | 9401 | |
b9a16870 | 9402 | unlock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_END); |
389dd41b | 9403 | unlock = build_call_expr_loc (loc, unlock, 0); |
1e8e9920 | 9404 | } |
9405 | ||
8a4a28a8 | 9406 | push_gimplify_context (); |
1e8e9920 | 9407 | |
9408 | block = make_node (BLOCK); | |
e3a19533 | 9409 | bind = gimple_build_bind (NULL, NULL, block); |
9410 | gsi_replace (gsi_p, bind, true); | |
9411 | gimple_bind_add_stmt (bind, stmt); | |
61e47ac8 | 9412 | |
75a70cf9 | 9413 | tbody = gimple_bind_body (bind); |
9414 | gimplify_and_add (lock, &tbody); | |
9415 | gimple_bind_set_body (bind, tbody); | |
1e8e9920 | 9416 | |
e3a19533 | 9417 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
75a70cf9 | 9418 | gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt))); |
9419 | gimple_bind_add_seq (bind, gimple_omp_body (stmt)); | |
9420 | gimple_omp_set_body (stmt, NULL); | |
1e8e9920 | 9421 | |
75a70cf9 | 9422 | tbody = gimple_bind_body (bind); |
9423 | gimplify_and_add (unlock, &tbody); | |
9424 | gimple_bind_set_body (bind, tbody); | |
61e47ac8 | 9425 | |
75a70cf9 | 9426 | gimple_bind_add_stmt (bind, gimple_build_omp_return (true)); |
1e8e9920 | 9427 | |
9428 | pop_gimplify_context (bind); | |
75a70cf9 | 9429 | gimple_bind_append_vars (bind, ctx->block_vars); |
9430 | BLOCK_VARS (block) = gimple_bind_vars (bind); | |
773c5ba7 | 9431 | } |
9432 | ||
9433 | ||
9434 | /* A subroutine of lower_omp_for. Generate code to emit the predicate | |
9435 | for a lastprivate clause. Given a loop control predicate of (V | |
9436 | cond N2), we gate the clause on (!(V cond N2)). The lowered form | |
1e4afe3c | 9437 | is appended to *DLIST, iterator initialization is appended to |
9438 | *BODY_P. */ | |
773c5ba7 | 9439 | |
9440 | static void | |
75a70cf9 | 9441 | lower_omp_for_lastprivate (struct omp_for_data *fd, gimple_seq *body_p, |
9442 | gimple_seq *dlist, struct omp_context *ctx) | |
773c5ba7 | 9443 | { |
75a70cf9 | 9444 | tree clauses, cond, vinit; |
773c5ba7 | 9445 | enum tree_code cond_code; |
75a70cf9 | 9446 | gimple_seq stmts; |
48e1416a | 9447 | |
fd6481cf | 9448 | cond_code = fd->loop.cond_code; |
773c5ba7 | 9449 | cond_code = cond_code == LT_EXPR ? GE_EXPR : LE_EXPR; |
9450 | ||
9451 | /* When possible, use a strict equality expression. This can let VRP | |
9452 | type optimizations deduce the value and remove a copy. */ | |
e913b5cd | 9453 | if (tree_fits_shwi_p (fd->loop.step)) |
773c5ba7 | 9454 | { |
e913b5cd | 9455 | HOST_WIDE_INT step = tree_to_shwi (fd->loop.step); |
773c5ba7 | 9456 | if (step == 1 || step == -1) |
9457 | cond_code = EQ_EXPR; | |
9458 | } | |
9459 | ||
fd6481cf | 9460 | cond = build2 (cond_code, boolean_type_node, fd->loop.v, fd->loop.n2); |
773c5ba7 | 9461 | |
75a70cf9 | 9462 | clauses = gimple_omp_for_clauses (fd->for_stmt); |
1e4afe3c | 9463 | stmts = NULL; |
9464 | lower_lastprivate_clauses (clauses, cond, &stmts, ctx); | |
75a70cf9 | 9465 | if (!gimple_seq_empty_p (stmts)) |
1e4afe3c | 9466 | { |
75a70cf9 | 9467 | gimple_seq_add_seq (&stmts, *dlist); |
fd6481cf | 9468 | *dlist = stmts; |
1e4afe3c | 9469 | |
9470 | /* Optimize: v = 0; is usually cheaper than v = some_other_constant. */ | |
fd6481cf | 9471 | vinit = fd->loop.n1; |
1e4afe3c | 9472 | if (cond_code == EQ_EXPR |
e913b5cd | 9473 | && tree_fits_shwi_p (fd->loop.n2) |
fd6481cf | 9474 | && ! integer_zerop (fd->loop.n2)) |
9475 | vinit = build_int_cst (TREE_TYPE (fd->loop.v), 0); | |
3d483a94 | 9476 | else |
9477 | vinit = unshare_expr (vinit); | |
1e4afe3c | 9478 | |
9479 | /* Initialize the iterator variable, so that threads that don't execute | |
9480 | any iterations don't execute the lastprivate clauses by accident. */ | |
75a70cf9 | 9481 | gimplify_assign (fd->loop.v, vinit, body_p); |
1e4afe3c | 9482 | } |
773c5ba7 | 9483 | } |
9484 | ||
9485 | ||
9486 | /* Lower code for an OpenMP loop directive. */ | |
9487 | ||
9488 | static void | |
75a70cf9 | 9489 | lower_omp_for (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
773c5ba7 | 9490 | { |
75a70cf9 | 9491 | tree *rhs_p, block; |
bc7bff74 | 9492 | struct omp_for_data fd, *fdp = NULL; |
1a91d914 | 9493 | gomp_for *stmt = as_a <gomp_for *> (gsi_stmt (*gsi_p)); |
9494 | gbind *new_stmt; | |
f018d957 | 9495 | gimple_seq omp_for_body, body, dlist; |
75a70cf9 | 9496 | size_t i; |
773c5ba7 | 9497 | |
8a4a28a8 | 9498 | push_gimplify_context (); |
773c5ba7 | 9499 | |
e3a19533 | 9500 | lower_omp (gimple_omp_for_pre_body_ptr (stmt), ctx); |
773c5ba7 | 9501 | |
1d22f541 | 9502 | block = make_node (BLOCK); |
75a70cf9 | 9503 | new_stmt = gimple_build_bind (NULL, NULL, block); |
e3a19533 | 9504 | /* Replace at gsi right away, so that 'stmt' is no member |
9505 | of a sequence anymore as we're going to add to to a different | |
9506 | one below. */ | |
9507 | gsi_replace (gsi_p, new_stmt, true); | |
1d22f541 | 9508 | |
773c5ba7 | 9509 | /* Move declaration of temporaries in the loop body before we make |
9510 | it go away. */ | |
75a70cf9 | 9511 | omp_for_body = gimple_omp_body (stmt); |
9512 | if (!gimple_seq_empty_p (omp_for_body) | |
9513 | && gimple_code (gimple_seq_first_stmt (omp_for_body)) == GIMPLE_BIND) | |
9514 | { | |
1a91d914 | 9515 | gbind *inner_bind |
9516 | = as_a <gbind *> (gimple_seq_first_stmt (omp_for_body)); | |
6d5a8d89 | 9517 | tree vars = gimple_bind_vars (inner_bind); |
75a70cf9 | 9518 | gimple_bind_append_vars (new_stmt, vars); |
6d5a8d89 | 9519 | /* bind_vars/BLOCK_VARS are being moved to new_stmt/block, don't |
9520 | keep them on the inner_bind and it's block. */ | |
9521 | gimple_bind_set_vars (inner_bind, NULL_TREE); | |
9522 | if (gimple_bind_block (inner_bind)) | |
9523 | BLOCK_VARS (gimple_bind_block (inner_bind)) = NULL_TREE; | |
75a70cf9 | 9524 | } |
773c5ba7 | 9525 | |
bc7bff74 | 9526 | if (gimple_omp_for_combined_into_p (stmt)) |
9527 | { | |
9528 | extract_omp_for_data (stmt, &fd, NULL); | |
9529 | fdp = &fd; | |
9530 | ||
9531 | /* We need two temporaries with fd.loop.v type (istart/iend) | |
9532 | and then (fd.collapse - 1) temporaries with the same | |
9533 | type for count2 ... countN-1 vars if not constant. */ | |
9534 | size_t count = 2; | |
9535 | tree type = fd.iter_type; | |
9536 | if (fd.collapse > 1 | |
9537 | && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
9538 | count += fd.collapse - 1; | |
9539 | bool parallel_for = gimple_omp_for_kind (stmt) == GF_OMP_FOR_KIND_FOR; | |
9540 | tree outerc = NULL, *pc = gimple_omp_for_clauses_ptr (stmt); | |
9541 | tree clauses = *pc; | |
9542 | if (parallel_for) | |
9543 | outerc | |
9544 | = find_omp_clause (gimple_omp_parallel_clauses (ctx->outer->stmt), | |
9545 | OMP_CLAUSE__LOOPTEMP_); | |
9546 | for (i = 0; i < count; i++) | |
9547 | { | |
9548 | tree temp; | |
9549 | if (parallel_for) | |
9550 | { | |
9551 | gcc_assert (outerc); | |
9552 | temp = lookup_decl (OMP_CLAUSE_DECL (outerc), ctx->outer); | |
9553 | outerc = find_omp_clause (OMP_CLAUSE_CHAIN (outerc), | |
9554 | OMP_CLAUSE__LOOPTEMP_); | |
9555 | } | |
9556 | else | |
691447ab | 9557 | { |
f9e245b2 | 9558 | temp = create_tmp_var (type); |
691447ab | 9559 | insert_decl_map (&ctx->outer->cb, temp, temp); |
9560 | } | |
bc7bff74 | 9561 | *pc = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE__LOOPTEMP_); |
9562 | OMP_CLAUSE_DECL (*pc) = temp; | |
9563 | pc = &OMP_CLAUSE_CHAIN (*pc); | |
9564 | } | |
9565 | *pc = clauses; | |
9566 | } | |
9567 | ||
75a70cf9 | 9568 | /* The pre-body and input clauses go before the lowered GIMPLE_OMP_FOR. */ |
773c5ba7 | 9569 | dlist = NULL; |
75a70cf9 | 9570 | body = NULL; |
bc7bff74 | 9571 | lower_rec_input_clauses (gimple_omp_for_clauses (stmt), &body, &dlist, ctx, |
9572 | fdp); | |
75a70cf9 | 9573 | gimple_seq_add_seq (&body, gimple_omp_for_pre_body (stmt)); |
773c5ba7 | 9574 | |
3d483a94 | 9575 | lower_omp (gimple_omp_body_ptr (stmt), ctx); |
9576 | ||
773c5ba7 | 9577 | /* Lower the header expressions. At this point, we can assume that |
9578 | the header is of the form: | |
9579 | ||
9580 | #pragma omp for (V = VAL1; V {<|>|<=|>=} VAL2; V = V [+-] VAL3) | |
9581 | ||
9582 | We just need to make sure that VAL1, VAL2 and VAL3 are lowered | |
9583 | using the .omp_data_s mapping, if needed. */ | |
75a70cf9 | 9584 | for (i = 0; i < gimple_omp_for_collapse (stmt); i++) |
fd6481cf | 9585 | { |
75a70cf9 | 9586 | rhs_p = gimple_omp_for_initial_ptr (stmt, i); |
fd6481cf | 9587 | if (!is_gimple_min_invariant (*rhs_p)) |
75a70cf9 | 9588 | *rhs_p = get_formal_tmp_var (*rhs_p, &body); |
fd6481cf | 9589 | |
75a70cf9 | 9590 | rhs_p = gimple_omp_for_final_ptr (stmt, i); |
fd6481cf | 9591 | if (!is_gimple_min_invariant (*rhs_p)) |
75a70cf9 | 9592 | *rhs_p = get_formal_tmp_var (*rhs_p, &body); |
fd6481cf | 9593 | |
75a70cf9 | 9594 | rhs_p = &TREE_OPERAND (gimple_omp_for_incr (stmt, i), 1); |
fd6481cf | 9595 | if (!is_gimple_min_invariant (*rhs_p)) |
75a70cf9 | 9596 | *rhs_p = get_formal_tmp_var (*rhs_p, &body); |
fd6481cf | 9597 | } |
773c5ba7 | 9598 | |
9599 | /* Once lowered, extract the bounds and clauses. */ | |
fd6481cf | 9600 | extract_omp_for_data (stmt, &fd, NULL); |
773c5ba7 | 9601 | |
75a70cf9 | 9602 | lower_omp_for_lastprivate (&fd, &body, &dlist, ctx); |
773c5ba7 | 9603 | |
75a70cf9 | 9604 | gimple_seq_add_stmt (&body, stmt); |
9605 | gimple_seq_add_seq (&body, gimple_omp_body (stmt)); | |
61e47ac8 | 9606 | |
75a70cf9 | 9607 | gimple_seq_add_stmt (&body, gimple_build_omp_continue (fd.loop.v, |
9608 | fd.loop.v)); | |
61e47ac8 | 9609 | |
773c5ba7 | 9610 | /* After the loop, add exit clauses. */ |
75a70cf9 | 9611 | lower_reduction_clauses (gimple_omp_for_clauses (stmt), &body, ctx); |
bc7bff74 | 9612 | |
9613 | if (ctx->cancellable) | |
9614 | gimple_seq_add_stmt (&body, gimple_build_label (ctx->cancel_label)); | |
9615 | ||
75a70cf9 | 9616 | gimple_seq_add_seq (&body, dlist); |
773c5ba7 | 9617 | |
75a70cf9 | 9618 | body = maybe_catch_exception (body); |
aade31a0 | 9619 | |
61e47ac8 | 9620 | /* Region exit marker goes at the end of the loop body. */ |
75a70cf9 | 9621 | gimple_seq_add_stmt (&body, gimple_build_omp_return (fd.have_nowait)); |
bc7bff74 | 9622 | maybe_add_implicit_barrier_cancel (ctx, &body); |
1d22f541 | 9623 | pop_gimplify_context (new_stmt); |
75a70cf9 | 9624 | |
9625 | gimple_bind_append_vars (new_stmt, ctx->block_vars); | |
9626 | BLOCK_VARS (block) = gimple_bind_vars (new_stmt); | |
1d22f541 | 9627 | if (BLOCK_VARS (block)) |
9628 | TREE_USED (block) = 1; | |
773c5ba7 | 9629 | |
75a70cf9 | 9630 | gimple_bind_set_body (new_stmt, body); |
9631 | gimple_omp_set_body (stmt, NULL); | |
9632 | gimple_omp_for_set_pre_body (stmt, NULL); | |
1e8e9920 | 9633 | } |
9634 | ||
48e1416a | 9635 | /* Callback for walk_stmts. Check if the current statement only contains |
a7dfc4fe | 9636 | GIMPLE_OMP_FOR or GIMPLE_OMP_SECTIONS. */ |
de7ef844 | 9637 | |
9638 | static tree | |
75a70cf9 | 9639 | check_combined_parallel (gimple_stmt_iterator *gsi_p, |
9640 | bool *handled_ops_p, | |
9641 | struct walk_stmt_info *wi) | |
de7ef844 | 9642 | { |
4077bf7a | 9643 | int *info = (int *) wi->info; |
75a70cf9 | 9644 | gimple stmt = gsi_stmt (*gsi_p); |
de7ef844 | 9645 | |
75a70cf9 | 9646 | *handled_ops_p = true; |
9647 | switch (gimple_code (stmt)) | |
de7ef844 | 9648 | { |
75a70cf9 | 9649 | WALK_SUBSTMTS; |
9650 | ||
9651 | case GIMPLE_OMP_FOR: | |
9652 | case GIMPLE_OMP_SECTIONS: | |
de7ef844 | 9653 | *info = *info == 0 ? 1 : -1; |
9654 | break; | |
9655 | default: | |
9656 | *info = -1; | |
9657 | break; | |
9658 | } | |
9659 | return NULL; | |
9660 | } | |
773c5ba7 | 9661 | |
fd6481cf | 9662 | struct omp_taskcopy_context |
9663 | { | |
9664 | /* This field must be at the beginning, as we do "inheritance": Some | |
9665 | callback functions for tree-inline.c (e.g., omp_copy_decl) | |
9666 | receive a copy_body_data pointer that is up-casted to an | |
9667 | omp_context pointer. */ | |
9668 | copy_body_data cb; | |
9669 | omp_context *ctx; | |
9670 | }; | |
9671 | ||
9672 | static tree | |
9673 | task_copyfn_copy_decl (tree var, copy_body_data *cb) | |
9674 | { | |
9675 | struct omp_taskcopy_context *tcctx = (struct omp_taskcopy_context *) cb; | |
9676 | ||
9677 | if (splay_tree_lookup (tcctx->ctx->sfield_map, (splay_tree_key) var)) | |
f9e245b2 | 9678 | return create_tmp_var (TREE_TYPE (var)); |
fd6481cf | 9679 | |
9680 | return var; | |
9681 | } | |
9682 | ||
9683 | static tree | |
9684 | task_copyfn_remap_type (struct omp_taskcopy_context *tcctx, tree orig_type) | |
9685 | { | |
9686 | tree name, new_fields = NULL, type, f; | |
9687 | ||
9688 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
9689 | name = DECL_NAME (TYPE_NAME (orig_type)); | |
e60a6f7b | 9690 | name = build_decl (gimple_location (tcctx->ctx->stmt), |
9691 | TYPE_DECL, name, type); | |
fd6481cf | 9692 | TYPE_NAME (type) = name; |
9693 | ||
9694 | for (f = TYPE_FIELDS (orig_type); f ; f = TREE_CHAIN (f)) | |
9695 | { | |
9696 | tree new_f = copy_node (f); | |
9697 | DECL_CONTEXT (new_f) = type; | |
9698 | TREE_TYPE (new_f) = remap_type (TREE_TYPE (f), &tcctx->cb); | |
9699 | TREE_CHAIN (new_f) = new_fields; | |
75a70cf9 | 9700 | walk_tree (&DECL_SIZE (new_f), copy_tree_body_r, &tcctx->cb, NULL); |
9701 | walk_tree (&DECL_SIZE_UNIT (new_f), copy_tree_body_r, &tcctx->cb, NULL); | |
9702 | walk_tree (&DECL_FIELD_OFFSET (new_f), copy_tree_body_r, | |
9703 | &tcctx->cb, NULL); | |
fd6481cf | 9704 | new_fields = new_f; |
06ecf488 | 9705 | tcctx->cb.decl_map->put (f, new_f); |
fd6481cf | 9706 | } |
9707 | TYPE_FIELDS (type) = nreverse (new_fields); | |
9708 | layout_type (type); | |
9709 | return type; | |
9710 | } | |
9711 | ||
9712 | /* Create task copyfn. */ | |
9713 | ||
9714 | static void | |
1a91d914 | 9715 | create_task_copyfn (gomp_task *task_stmt, omp_context *ctx) |
fd6481cf | 9716 | { |
9717 | struct function *child_cfun; | |
9718 | tree child_fn, t, c, src, dst, f, sf, arg, sarg, decl; | |
9719 | tree record_type, srecord_type, bind, list; | |
9720 | bool record_needs_remap = false, srecord_needs_remap = false; | |
9721 | splay_tree_node n; | |
9722 | struct omp_taskcopy_context tcctx; | |
389dd41b | 9723 | location_t loc = gimple_location (task_stmt); |
fd6481cf | 9724 | |
75a70cf9 | 9725 | child_fn = gimple_omp_task_copy_fn (task_stmt); |
fd6481cf | 9726 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); |
9727 | gcc_assert (child_cfun->cfg == NULL); | |
fd6481cf | 9728 | DECL_SAVED_TREE (child_fn) = alloc_stmt_list (); |
9729 | ||
9730 | /* Reset DECL_CONTEXT on function arguments. */ | |
1767a056 | 9731 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) |
fd6481cf | 9732 | DECL_CONTEXT (t) = child_fn; |
9733 | ||
9734 | /* Populate the function. */ | |
8a4a28a8 | 9735 | push_gimplify_context (); |
9078126c | 9736 | push_cfun (child_cfun); |
fd6481cf | 9737 | |
9738 | bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, NULL); | |
9739 | TREE_SIDE_EFFECTS (bind) = 1; | |
9740 | list = NULL; | |
9741 | DECL_SAVED_TREE (child_fn) = bind; | |
75a70cf9 | 9742 | DECL_SOURCE_LOCATION (child_fn) = gimple_location (task_stmt); |
fd6481cf | 9743 | |
9744 | /* Remap src and dst argument types if needed. */ | |
9745 | record_type = ctx->record_type; | |
9746 | srecord_type = ctx->srecord_type; | |
1767a056 | 9747 | for (f = TYPE_FIELDS (record_type); f ; f = DECL_CHAIN (f)) |
fd6481cf | 9748 | if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn)) |
9749 | { | |
9750 | record_needs_remap = true; | |
9751 | break; | |
9752 | } | |
1767a056 | 9753 | for (f = TYPE_FIELDS (srecord_type); f ; f = DECL_CHAIN (f)) |
fd6481cf | 9754 | if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn)) |
9755 | { | |
9756 | srecord_needs_remap = true; | |
9757 | break; | |
9758 | } | |
9759 | ||
9760 | if (record_needs_remap || srecord_needs_remap) | |
9761 | { | |
9762 | memset (&tcctx, '\0', sizeof (tcctx)); | |
9763 | tcctx.cb.src_fn = ctx->cb.src_fn; | |
9764 | tcctx.cb.dst_fn = child_fn; | |
415d1b9a | 9765 | tcctx.cb.src_node = cgraph_node::get (tcctx.cb.src_fn); |
53f79206 | 9766 | gcc_checking_assert (tcctx.cb.src_node); |
fd6481cf | 9767 | tcctx.cb.dst_node = tcctx.cb.src_node; |
9768 | tcctx.cb.src_cfun = ctx->cb.src_cfun; | |
9769 | tcctx.cb.copy_decl = task_copyfn_copy_decl; | |
e38def9c | 9770 | tcctx.cb.eh_lp_nr = 0; |
fd6481cf | 9771 | tcctx.cb.transform_call_graph_edges = CB_CGE_MOVE; |
06ecf488 | 9772 | tcctx.cb.decl_map = new hash_map<tree, tree>; |
fd6481cf | 9773 | tcctx.ctx = ctx; |
9774 | ||
9775 | if (record_needs_remap) | |
9776 | record_type = task_copyfn_remap_type (&tcctx, record_type); | |
9777 | if (srecord_needs_remap) | |
9778 | srecord_type = task_copyfn_remap_type (&tcctx, srecord_type); | |
9779 | } | |
9780 | else | |
9781 | tcctx.cb.decl_map = NULL; | |
9782 | ||
fd6481cf | 9783 | arg = DECL_ARGUMENTS (child_fn); |
9784 | TREE_TYPE (arg) = build_pointer_type (record_type); | |
1767a056 | 9785 | sarg = DECL_CHAIN (arg); |
fd6481cf | 9786 | TREE_TYPE (sarg) = build_pointer_type (srecord_type); |
9787 | ||
9788 | /* First pass: initialize temporaries used in record_type and srecord_type | |
9789 | sizes and field offsets. */ | |
9790 | if (tcctx.cb.decl_map) | |
75a70cf9 | 9791 | for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c)) |
fd6481cf | 9792 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_FIRSTPRIVATE) |
9793 | { | |
9794 | tree *p; | |
9795 | ||
9796 | decl = OMP_CLAUSE_DECL (c); | |
06ecf488 | 9797 | p = tcctx.cb.decl_map->get (decl); |
fd6481cf | 9798 | if (p == NULL) |
9799 | continue; | |
9800 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); | |
9801 | sf = (tree) n->value; | |
06ecf488 | 9802 | sf = *tcctx.cb.decl_map->get (sf); |
182cf5a9 | 9803 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9804 | src = omp_build_component_ref (src, sf); |
75a70cf9 | 9805 | t = build2 (MODIFY_EXPR, TREE_TYPE (*p), *p, src); |
fd6481cf | 9806 | append_to_statement_list (t, &list); |
9807 | } | |
9808 | ||
9809 | /* Second pass: copy shared var pointers and copy construct non-VLA | |
9810 | firstprivate vars. */ | |
75a70cf9 | 9811 | for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c)) |
fd6481cf | 9812 | switch (OMP_CLAUSE_CODE (c)) |
9813 | { | |
9814 | case OMP_CLAUSE_SHARED: | |
9815 | decl = OMP_CLAUSE_DECL (c); | |
9816 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9817 | if (n == NULL) | |
9818 | break; | |
9819 | f = (tree) n->value; | |
9820 | if (tcctx.cb.decl_map) | |
06ecf488 | 9821 | f = *tcctx.cb.decl_map->get (f); |
fd6481cf | 9822 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); |
9823 | sf = (tree) n->value; | |
9824 | if (tcctx.cb.decl_map) | |
06ecf488 | 9825 | sf = *tcctx.cb.decl_map->get (sf); |
182cf5a9 | 9826 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9827 | src = omp_build_component_ref (src, sf); |
182cf5a9 | 9828 | dst = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9829 | dst = omp_build_component_ref (dst, f); |
75a70cf9 | 9830 | t = build2 (MODIFY_EXPR, TREE_TYPE (dst), dst, src); |
fd6481cf | 9831 | append_to_statement_list (t, &list); |
9832 | break; | |
9833 | case OMP_CLAUSE_FIRSTPRIVATE: | |
9834 | decl = OMP_CLAUSE_DECL (c); | |
9835 | if (is_variable_sized (decl)) | |
9836 | break; | |
9837 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9838 | if (n == NULL) | |
9839 | break; | |
9840 | f = (tree) n->value; | |
9841 | if (tcctx.cb.decl_map) | |
06ecf488 | 9842 | f = *tcctx.cb.decl_map->get (f); |
fd6481cf | 9843 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); |
9844 | if (n != NULL) | |
9845 | { | |
9846 | sf = (tree) n->value; | |
9847 | if (tcctx.cb.decl_map) | |
06ecf488 | 9848 | sf = *tcctx.cb.decl_map->get (sf); |
182cf5a9 | 9849 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9850 | src = omp_build_component_ref (src, sf); |
fd6481cf | 9851 | if (use_pointer_for_field (decl, NULL) || is_reference (decl)) |
182cf5a9 | 9852 | src = build_simple_mem_ref_loc (loc, src); |
fd6481cf | 9853 | } |
9854 | else | |
9855 | src = decl; | |
182cf5a9 | 9856 | dst = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9857 | dst = omp_build_component_ref (dst, f); |
fd6481cf | 9858 | t = lang_hooks.decls.omp_clause_copy_ctor (c, dst, src); |
9859 | append_to_statement_list (t, &list); | |
9860 | break; | |
9861 | case OMP_CLAUSE_PRIVATE: | |
9862 | if (! OMP_CLAUSE_PRIVATE_OUTER_REF (c)) | |
9863 | break; | |
9864 | decl = OMP_CLAUSE_DECL (c); | |
9865 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9866 | f = (tree) n->value; | |
9867 | if (tcctx.cb.decl_map) | |
06ecf488 | 9868 | f = *tcctx.cb.decl_map->get (f); |
fd6481cf | 9869 | n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl); |
9870 | if (n != NULL) | |
9871 | { | |
9872 | sf = (tree) n->value; | |
9873 | if (tcctx.cb.decl_map) | |
06ecf488 | 9874 | sf = *tcctx.cb.decl_map->get (sf); |
182cf5a9 | 9875 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9876 | src = omp_build_component_ref (src, sf); |
fd6481cf | 9877 | if (use_pointer_for_field (decl, NULL)) |
182cf5a9 | 9878 | src = build_simple_mem_ref_loc (loc, src); |
fd6481cf | 9879 | } |
9880 | else | |
9881 | src = decl; | |
182cf5a9 | 9882 | dst = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9883 | dst = omp_build_component_ref (dst, f); |
75a70cf9 | 9884 | t = build2 (MODIFY_EXPR, TREE_TYPE (dst), dst, src); |
fd6481cf | 9885 | append_to_statement_list (t, &list); |
9886 | break; | |
9887 | default: | |
9888 | break; | |
9889 | } | |
9890 | ||
9891 | /* Last pass: handle VLA firstprivates. */ | |
9892 | if (tcctx.cb.decl_map) | |
75a70cf9 | 9893 | for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c)) |
fd6481cf | 9894 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_FIRSTPRIVATE) |
9895 | { | |
9896 | tree ind, ptr, df; | |
9897 | ||
9898 | decl = OMP_CLAUSE_DECL (c); | |
9899 | if (!is_variable_sized (decl)) | |
9900 | continue; | |
9901 | n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl); | |
9902 | if (n == NULL) | |
9903 | continue; | |
9904 | f = (tree) n->value; | |
06ecf488 | 9905 | f = *tcctx.cb.decl_map->get (f); |
fd6481cf | 9906 | gcc_assert (DECL_HAS_VALUE_EXPR_P (decl)); |
9907 | ind = DECL_VALUE_EXPR (decl); | |
9908 | gcc_assert (TREE_CODE (ind) == INDIRECT_REF); | |
9909 | gcc_assert (DECL_P (TREE_OPERAND (ind, 0))); | |
9910 | n = splay_tree_lookup (ctx->sfield_map, | |
9911 | (splay_tree_key) TREE_OPERAND (ind, 0)); | |
9912 | sf = (tree) n->value; | |
06ecf488 | 9913 | sf = *tcctx.cb.decl_map->get (sf); |
182cf5a9 | 9914 | src = build_simple_mem_ref_loc (loc, sarg); |
445d06b6 | 9915 | src = omp_build_component_ref (src, sf); |
182cf5a9 | 9916 | src = build_simple_mem_ref_loc (loc, src); |
9917 | dst = build_simple_mem_ref_loc (loc, arg); | |
445d06b6 | 9918 | dst = omp_build_component_ref (dst, f); |
fd6481cf | 9919 | t = lang_hooks.decls.omp_clause_copy_ctor (c, dst, src); |
9920 | append_to_statement_list (t, &list); | |
9921 | n = splay_tree_lookup (ctx->field_map, | |
9922 | (splay_tree_key) TREE_OPERAND (ind, 0)); | |
9923 | df = (tree) n->value; | |
06ecf488 | 9924 | df = *tcctx.cb.decl_map->get (df); |
182cf5a9 | 9925 | ptr = build_simple_mem_ref_loc (loc, arg); |
445d06b6 | 9926 | ptr = omp_build_component_ref (ptr, df); |
75a70cf9 | 9927 | t = build2 (MODIFY_EXPR, TREE_TYPE (ptr), ptr, |
389dd41b | 9928 | build_fold_addr_expr_loc (loc, dst)); |
fd6481cf | 9929 | append_to_statement_list (t, &list); |
9930 | } | |
9931 | ||
9932 | t = build1 (RETURN_EXPR, void_type_node, NULL); | |
9933 | append_to_statement_list (t, &list); | |
9934 | ||
9935 | if (tcctx.cb.decl_map) | |
06ecf488 | 9936 | delete tcctx.cb.decl_map; |
fd6481cf | 9937 | pop_gimplify_context (NULL); |
9938 | BIND_EXPR_BODY (bind) = list; | |
9939 | pop_cfun (); | |
fd6481cf | 9940 | } |
9941 | ||
bc7bff74 | 9942 | static void |
9943 | lower_depend_clauses (gimple stmt, gimple_seq *iseq, gimple_seq *oseq) | |
9944 | { | |
9945 | tree c, clauses; | |
9946 | gimple g; | |
9947 | size_t n_in = 0, n_out = 0, idx = 2, i; | |
9948 | ||
9949 | clauses = find_omp_clause (gimple_omp_task_clauses (stmt), | |
9950 | OMP_CLAUSE_DEPEND); | |
9951 | gcc_assert (clauses); | |
9952 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
9953 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_DEPEND) | |
9954 | switch (OMP_CLAUSE_DEPEND_KIND (c)) | |
9955 | { | |
9956 | case OMP_CLAUSE_DEPEND_IN: | |
9957 | n_in++; | |
9958 | break; | |
9959 | case OMP_CLAUSE_DEPEND_OUT: | |
9960 | case OMP_CLAUSE_DEPEND_INOUT: | |
9961 | n_out++; | |
9962 | break; | |
9963 | default: | |
9964 | gcc_unreachable (); | |
9965 | } | |
9966 | tree type = build_array_type_nelts (ptr_type_node, n_in + n_out + 2); | |
f9e245b2 | 9967 | tree array = create_tmp_var (type); |
bc7bff74 | 9968 | tree r = build4 (ARRAY_REF, ptr_type_node, array, size_int (0), NULL_TREE, |
9969 | NULL_TREE); | |
9970 | g = gimple_build_assign (r, build_int_cst (ptr_type_node, n_in + n_out)); | |
9971 | gimple_seq_add_stmt (iseq, g); | |
9972 | r = build4 (ARRAY_REF, ptr_type_node, array, size_int (1), NULL_TREE, | |
9973 | NULL_TREE); | |
9974 | g = gimple_build_assign (r, build_int_cst (ptr_type_node, n_out)); | |
9975 | gimple_seq_add_stmt (iseq, g); | |
9976 | for (i = 0; i < 2; i++) | |
9977 | { | |
9978 | if ((i ? n_in : n_out) == 0) | |
9979 | continue; | |
9980 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
9981 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_DEPEND | |
9982 | && ((OMP_CLAUSE_DEPEND_KIND (c) != OMP_CLAUSE_DEPEND_IN) ^ i)) | |
9983 | { | |
9984 | tree t = OMP_CLAUSE_DECL (c); | |
9985 | t = fold_convert (ptr_type_node, t); | |
9986 | gimplify_expr (&t, iseq, NULL, is_gimple_val, fb_rvalue); | |
9987 | r = build4 (ARRAY_REF, ptr_type_node, array, size_int (idx++), | |
9988 | NULL_TREE, NULL_TREE); | |
9989 | g = gimple_build_assign (r, t); | |
9990 | gimple_seq_add_stmt (iseq, g); | |
9991 | } | |
9992 | } | |
9993 | tree *p = gimple_omp_task_clauses_ptr (stmt); | |
9994 | c = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE_DEPEND); | |
9995 | OMP_CLAUSE_DECL (c) = build_fold_addr_expr (array); | |
9996 | OMP_CLAUSE_CHAIN (c) = *p; | |
9997 | *p = c; | |
9998 | tree clobber = build_constructor (type, NULL); | |
9999 | TREE_THIS_VOLATILE (clobber) = 1; | |
10000 | g = gimple_build_assign (array, clobber); | |
10001 | gimple_seq_add_stmt (oseq, g); | |
10002 | } | |
10003 | ||
75a70cf9 | 10004 | /* Lower the OpenMP parallel or task directive in the current statement |
10005 | in GSI_P. CTX holds context information for the directive. */ | |
773c5ba7 | 10006 | |
10007 | static void | |
75a70cf9 | 10008 | lower_omp_taskreg (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
773c5ba7 | 10009 | { |
75a70cf9 | 10010 | tree clauses; |
10011 | tree child_fn, t; | |
10012 | gimple stmt = gsi_stmt (*gsi_p); | |
1a91d914 | 10013 | gbind *par_bind, *bind, *dep_bind = NULL; |
bc7bff74 | 10014 | gimple_seq par_body, olist, ilist, par_olist, par_rlist, par_ilist, new_body; |
389dd41b | 10015 | location_t loc = gimple_location (stmt); |
773c5ba7 | 10016 | |
75a70cf9 | 10017 | clauses = gimple_omp_taskreg_clauses (stmt); |
1a91d914 | 10018 | par_bind |
10019 | = as_a <gbind *> (gimple_seq_first_stmt (gimple_omp_body (stmt))); | |
75a70cf9 | 10020 | par_body = gimple_bind_body (par_bind); |
773c5ba7 | 10021 | child_fn = ctx->cb.dst_fn; |
75a70cf9 | 10022 | if (gimple_code (stmt) == GIMPLE_OMP_PARALLEL |
10023 | && !gimple_omp_parallel_combined_p (stmt)) | |
de7ef844 | 10024 | { |
10025 | struct walk_stmt_info wi; | |
10026 | int ws_num = 0; | |
10027 | ||
10028 | memset (&wi, 0, sizeof (wi)); | |
de7ef844 | 10029 | wi.info = &ws_num; |
10030 | wi.val_only = true; | |
75a70cf9 | 10031 | walk_gimple_seq (par_body, check_combined_parallel, NULL, &wi); |
de7ef844 | 10032 | if (ws_num == 1) |
75a70cf9 | 10033 | gimple_omp_parallel_set_combined_p (stmt, true); |
de7ef844 | 10034 | } |
bc7bff74 | 10035 | gimple_seq dep_ilist = NULL; |
10036 | gimple_seq dep_olist = NULL; | |
10037 | if (gimple_code (stmt) == GIMPLE_OMP_TASK | |
10038 | && find_omp_clause (clauses, OMP_CLAUSE_DEPEND)) | |
10039 | { | |
8a4a28a8 | 10040 | push_gimplify_context (); |
bc7bff74 | 10041 | dep_bind = gimple_build_bind (NULL, NULL, make_node (BLOCK)); |
10042 | lower_depend_clauses (stmt, &dep_ilist, &dep_olist); | |
10043 | } | |
10044 | ||
fd6481cf | 10045 | if (ctx->srecord_type) |
1a91d914 | 10046 | create_task_copyfn (as_a <gomp_task *> (stmt), ctx); |
773c5ba7 | 10047 | |
8a4a28a8 | 10048 | push_gimplify_context (); |
773c5ba7 | 10049 | |
75a70cf9 | 10050 | par_olist = NULL; |
10051 | par_ilist = NULL; | |
bc7bff74 | 10052 | par_rlist = NULL; |
10053 | lower_rec_input_clauses (clauses, &par_ilist, &par_olist, ctx, NULL); | |
e3a19533 | 10054 | lower_omp (&par_body, ctx); |
75a70cf9 | 10055 | if (gimple_code (stmt) == GIMPLE_OMP_PARALLEL) |
bc7bff74 | 10056 | lower_reduction_clauses (clauses, &par_rlist, ctx); |
773c5ba7 | 10057 | |
10058 | /* Declare all the variables created by mapping and the variables | |
10059 | declared in the scope of the parallel body. */ | |
10060 | record_vars_into (ctx->block_vars, child_fn); | |
75a70cf9 | 10061 | record_vars_into (gimple_bind_vars (par_bind), child_fn); |
773c5ba7 | 10062 | |
10063 | if (ctx->record_type) | |
10064 | { | |
fd6481cf | 10065 | ctx->sender_decl |
10066 | = create_tmp_var (ctx->srecord_type ? ctx->srecord_type | |
10067 | : ctx->record_type, ".omp_data_o"); | |
84bfaaeb | 10068 | DECL_NAMELESS (ctx->sender_decl) = 1; |
86f2ad37 | 10069 | TREE_ADDRESSABLE (ctx->sender_decl) = 1; |
75a70cf9 | 10070 | gimple_omp_taskreg_set_data_arg (stmt, ctx->sender_decl); |
773c5ba7 | 10071 | } |
10072 | ||
75a70cf9 | 10073 | olist = NULL; |
10074 | ilist = NULL; | |
773c5ba7 | 10075 | lower_send_clauses (clauses, &ilist, &olist, ctx); |
10076 | lower_send_shared_vars (&ilist, &olist, ctx); | |
10077 | ||
bc7bff74 | 10078 | if (ctx->record_type) |
10079 | { | |
10080 | tree clobber = build_constructor (TREE_TYPE (ctx->sender_decl), NULL); | |
10081 | TREE_THIS_VOLATILE (clobber) = 1; | |
10082 | gimple_seq_add_stmt (&olist, gimple_build_assign (ctx->sender_decl, | |
10083 | clobber)); | |
10084 | } | |
10085 | ||
773c5ba7 | 10086 | /* Once all the expansions are done, sequence all the different |
75a70cf9 | 10087 | fragments inside gimple_omp_body. */ |
773c5ba7 | 10088 | |
75a70cf9 | 10089 | new_body = NULL; |
773c5ba7 | 10090 | |
10091 | if (ctx->record_type) | |
10092 | { | |
389dd41b | 10093 | t = build_fold_addr_expr_loc (loc, ctx->sender_decl); |
cc6b725b | 10094 | /* fixup_child_record_type might have changed receiver_decl's type. */ |
389dd41b | 10095 | t = fold_convert_loc (loc, TREE_TYPE (ctx->receiver_decl), t); |
75a70cf9 | 10096 | gimple_seq_add_stmt (&new_body, |
10097 | gimple_build_assign (ctx->receiver_decl, t)); | |
773c5ba7 | 10098 | } |
10099 | ||
75a70cf9 | 10100 | gimple_seq_add_seq (&new_body, par_ilist); |
10101 | gimple_seq_add_seq (&new_body, par_body); | |
bc7bff74 | 10102 | gimple_seq_add_seq (&new_body, par_rlist); |
10103 | if (ctx->cancellable) | |
10104 | gimple_seq_add_stmt (&new_body, gimple_build_label (ctx->cancel_label)); | |
75a70cf9 | 10105 | gimple_seq_add_seq (&new_body, par_olist); |
10106 | new_body = maybe_catch_exception (new_body); | |
10107 | gimple_seq_add_stmt (&new_body, gimple_build_omp_return (false)); | |
10108 | gimple_omp_set_body (stmt, new_body); | |
773c5ba7 | 10109 | |
75a70cf9 | 10110 | bind = gimple_build_bind (NULL, NULL, gimple_bind_block (par_bind)); |
bc7bff74 | 10111 | gsi_replace (gsi_p, dep_bind ? dep_bind : bind, true); |
10112 | gimple_bind_add_seq (bind, ilist); | |
10113 | gimple_bind_add_stmt (bind, stmt); | |
10114 | gimple_bind_add_seq (bind, olist); | |
10115 | ||
10116 | pop_gimplify_context (NULL); | |
10117 | ||
10118 | if (dep_bind) | |
10119 | { | |
10120 | gimple_bind_add_seq (dep_bind, dep_ilist); | |
10121 | gimple_bind_add_stmt (dep_bind, bind); | |
10122 | gimple_bind_add_seq (dep_bind, dep_olist); | |
10123 | pop_gimplify_context (dep_bind); | |
10124 | } | |
10125 | } | |
10126 | ||
10127 | /* Lower the OpenMP target directive in the current statement | |
10128 | in GSI_P. CTX holds context information for the directive. */ | |
10129 | ||
10130 | static void | |
10131 | lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx) | |
10132 | { | |
10133 | tree clauses; | |
10134 | tree child_fn, t, c; | |
1a91d914 | 10135 | gomp_target *stmt = as_a <gomp_target *> (gsi_stmt (*gsi_p)); |
10136 | gbind *tgt_bind = NULL, *bind; | |
bc7bff74 | 10137 | gimple_seq tgt_body = NULL, olist, ilist, new_body; |
bc7bff74 | 10138 | location_t loc = gimple_location (stmt); |
10139 | int kind = gimple_omp_target_kind (stmt); | |
10140 | unsigned int map_cnt = 0; | |
10141 | ||
10142 | clauses = gimple_omp_target_clauses (stmt); | |
10143 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
10144 | { | |
1a91d914 | 10145 | tgt_bind = gimple_seq_first_stmt_as_a_bind (gimple_omp_body (stmt)); |
bc7bff74 | 10146 | tgt_body = gimple_bind_body (tgt_bind); |
10147 | } | |
10148 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
10149 | tgt_body = gimple_omp_body (stmt); | |
10150 | child_fn = ctx->cb.dst_fn; | |
10151 | ||
8a4a28a8 | 10152 | push_gimplify_context (); |
bc7bff74 | 10153 | |
10154 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
10155 | switch (OMP_CLAUSE_CODE (c)) | |
10156 | { | |
10157 | tree var, x; | |
10158 | ||
10159 | default: | |
10160 | break; | |
10161 | case OMP_CLAUSE_MAP: | |
10162 | case OMP_CLAUSE_TO: | |
10163 | case OMP_CLAUSE_FROM: | |
10164 | var = OMP_CLAUSE_DECL (c); | |
10165 | if (!DECL_P (var)) | |
10166 | { | |
10167 | if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_MAP | |
10168 | || !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c)) | |
10169 | map_cnt++; | |
10170 | continue; | |
10171 | } | |
10172 | ||
10173 | if (DECL_SIZE (var) | |
10174 | && TREE_CODE (DECL_SIZE (var)) != INTEGER_CST) | |
10175 | { | |
10176 | tree var2 = DECL_VALUE_EXPR (var); | |
10177 | gcc_assert (TREE_CODE (var2) == INDIRECT_REF); | |
10178 | var2 = TREE_OPERAND (var2, 0); | |
10179 | gcc_assert (DECL_P (var2)); | |
10180 | var = var2; | |
10181 | } | |
10182 | ||
10183 | if (!maybe_lookup_field (var, ctx)) | |
10184 | continue; | |
10185 | ||
10186 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
10187 | { | |
10188 | x = build_receiver_ref (var, true, ctx); | |
10189 | tree new_var = lookup_decl (var, ctx); | |
10190 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
10191 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
10192 | && !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) | |
10193 | && TREE_CODE (TREE_TYPE (var)) == ARRAY_TYPE) | |
10194 | x = build_simple_mem_ref (x); | |
10195 | SET_DECL_VALUE_EXPR (new_var, x); | |
10196 | DECL_HAS_VALUE_EXPR_P (new_var) = 1; | |
10197 | } | |
10198 | map_cnt++; | |
10199 | } | |
10200 | ||
10201 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
10202 | { | |
10203 | target_nesting_level++; | |
10204 | lower_omp (&tgt_body, ctx); | |
10205 | target_nesting_level--; | |
10206 | } | |
10207 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
10208 | lower_omp (&tgt_body, ctx); | |
10209 | ||
10210 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
10211 | { | |
10212 | /* Declare all the variables created by mapping and the variables | |
10213 | declared in the scope of the target body. */ | |
10214 | record_vars_into (ctx->block_vars, child_fn); | |
10215 | record_vars_into (gimple_bind_vars (tgt_bind), child_fn); | |
10216 | } | |
10217 | ||
10218 | olist = NULL; | |
10219 | ilist = NULL; | |
10220 | if (ctx->record_type) | |
10221 | { | |
10222 | ctx->sender_decl | |
10223 | = create_tmp_var (ctx->record_type, ".omp_data_arr"); | |
10224 | DECL_NAMELESS (ctx->sender_decl) = 1; | |
10225 | TREE_ADDRESSABLE (ctx->sender_decl) = 1; | |
10226 | t = make_tree_vec (3); | |
10227 | TREE_VEC_ELT (t, 0) = ctx->sender_decl; | |
10228 | TREE_VEC_ELT (t, 1) | |
10229 | = create_tmp_var (build_array_type_nelts (size_type_node, map_cnt), | |
10230 | ".omp_data_sizes"); | |
10231 | DECL_NAMELESS (TREE_VEC_ELT (t, 1)) = 1; | |
10232 | TREE_ADDRESSABLE (TREE_VEC_ELT (t, 1)) = 1; | |
10233 | TREE_STATIC (TREE_VEC_ELT (t, 1)) = 1; | |
10234 | TREE_VEC_ELT (t, 2) | |
10235 | = create_tmp_var (build_array_type_nelts (unsigned_char_type_node, | |
10236 | map_cnt), | |
10237 | ".omp_data_kinds"); | |
10238 | DECL_NAMELESS (TREE_VEC_ELT (t, 2)) = 1; | |
10239 | TREE_ADDRESSABLE (TREE_VEC_ELT (t, 2)) = 1; | |
10240 | TREE_STATIC (TREE_VEC_ELT (t, 2)) = 1; | |
10241 | gimple_omp_target_set_data_arg (stmt, t); | |
10242 | ||
10243 | vec<constructor_elt, va_gc> *vsize; | |
10244 | vec<constructor_elt, va_gc> *vkind; | |
10245 | vec_alloc (vsize, map_cnt); | |
10246 | vec_alloc (vkind, map_cnt); | |
10247 | unsigned int map_idx = 0; | |
10248 | ||
10249 | for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c)) | |
10250 | switch (OMP_CLAUSE_CODE (c)) | |
10251 | { | |
10252 | tree ovar, nc; | |
10253 | ||
10254 | default: | |
10255 | break; | |
10256 | case OMP_CLAUSE_MAP: | |
10257 | case OMP_CLAUSE_TO: | |
10258 | case OMP_CLAUSE_FROM: | |
10259 | nc = c; | |
10260 | ovar = OMP_CLAUSE_DECL (c); | |
10261 | if (!DECL_P (ovar)) | |
10262 | { | |
10263 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
10264 | && OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c)) | |
10265 | { | |
10266 | gcc_checking_assert (OMP_CLAUSE_DECL (OMP_CLAUSE_CHAIN (c)) | |
10267 | == get_base_address (ovar)); | |
10268 | nc = OMP_CLAUSE_CHAIN (c); | |
10269 | ovar = OMP_CLAUSE_DECL (nc); | |
10270 | } | |
10271 | else | |
10272 | { | |
10273 | tree x = build_sender_ref (ovar, ctx); | |
10274 | tree v | |
10275 | = build_fold_addr_expr_with_type (ovar, ptr_type_node); | |
10276 | gimplify_assign (x, v, &ilist); | |
10277 | nc = NULL_TREE; | |
10278 | } | |
10279 | } | |
10280 | else | |
10281 | { | |
10282 | if (DECL_SIZE (ovar) | |
10283 | && TREE_CODE (DECL_SIZE (ovar)) != INTEGER_CST) | |
10284 | { | |
10285 | tree ovar2 = DECL_VALUE_EXPR (ovar); | |
10286 | gcc_assert (TREE_CODE (ovar2) == INDIRECT_REF); | |
10287 | ovar2 = TREE_OPERAND (ovar2, 0); | |
10288 | gcc_assert (DECL_P (ovar2)); | |
10289 | ovar = ovar2; | |
10290 | } | |
10291 | if (!maybe_lookup_field (ovar, ctx)) | |
10292 | continue; | |
10293 | } | |
10294 | ||
4f3b8d78 | 10295 | unsigned int talign = TYPE_ALIGN_UNIT (TREE_TYPE (ovar)); |
10296 | if (DECL_P (ovar) && DECL_ALIGN_UNIT (ovar) > talign) | |
10297 | talign = DECL_ALIGN_UNIT (ovar); | |
bc7bff74 | 10298 | if (nc) |
10299 | { | |
10300 | tree var = lookup_decl_in_outer_ctx (ovar, ctx); | |
10301 | tree x = build_sender_ref (ovar, ctx); | |
10302 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_MAP | |
10303 | && OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_POINTER | |
10304 | && !OMP_CLAUSE_MAP_ZERO_BIAS_ARRAY_SECTION (c) | |
10305 | && TREE_CODE (TREE_TYPE (ovar)) == ARRAY_TYPE) | |
10306 | { | |
10307 | gcc_assert (kind == GF_OMP_TARGET_KIND_REGION); | |
10308 | tree avar | |
f9e245b2 | 10309 | = create_tmp_var (TREE_TYPE (TREE_TYPE (x))); |
bc7bff74 | 10310 | mark_addressable (avar); |
10311 | gimplify_assign (avar, build_fold_addr_expr (var), &ilist); | |
4f3b8d78 | 10312 | talign = DECL_ALIGN_UNIT (avar); |
bc7bff74 | 10313 | avar = build_fold_addr_expr (avar); |
10314 | gimplify_assign (x, avar, &ilist); | |
10315 | } | |
10316 | else if (is_gimple_reg (var)) | |
10317 | { | |
10318 | gcc_assert (kind == GF_OMP_TARGET_KIND_REGION); | |
f9e245b2 | 10319 | tree avar = create_tmp_var (TREE_TYPE (var)); |
bc7bff74 | 10320 | mark_addressable (avar); |
10321 | if (OMP_CLAUSE_MAP_KIND (c) != OMP_CLAUSE_MAP_ALLOC | |
10322 | && OMP_CLAUSE_MAP_KIND (c) != OMP_CLAUSE_MAP_FROM) | |
10323 | gimplify_assign (avar, var, &ilist); | |
10324 | avar = build_fold_addr_expr (avar); | |
10325 | gimplify_assign (x, avar, &ilist); | |
10326 | if ((OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_FROM | |
10327 | || OMP_CLAUSE_MAP_KIND (c) == OMP_CLAUSE_MAP_TOFROM) | |
10328 | && !TYPE_READONLY (TREE_TYPE (var))) | |
10329 | { | |
10330 | x = build_sender_ref (ovar, ctx); | |
10331 | x = build_simple_mem_ref (x); | |
10332 | gimplify_assign (var, x, &olist); | |
10333 | } | |
10334 | } | |
10335 | else | |
10336 | { | |
10337 | var = build_fold_addr_expr (var); | |
10338 | gimplify_assign (x, var, &ilist); | |
10339 | } | |
10340 | } | |
10341 | tree s = OMP_CLAUSE_SIZE (c); | |
10342 | if (s == NULL_TREE) | |
10343 | s = TYPE_SIZE_UNIT (TREE_TYPE (ovar)); | |
10344 | s = fold_convert (size_type_node, s); | |
10345 | tree purpose = size_int (map_idx++); | |
10346 | CONSTRUCTOR_APPEND_ELT (vsize, purpose, s); | |
10347 | if (TREE_CODE (s) != INTEGER_CST) | |
10348 | TREE_STATIC (TREE_VEC_ELT (t, 1)) = 0; | |
10349 | ||
10350 | unsigned char tkind = 0; | |
10351 | switch (OMP_CLAUSE_CODE (c)) | |
10352 | { | |
10353 | case OMP_CLAUSE_MAP: | |
10354 | tkind = OMP_CLAUSE_MAP_KIND (c); | |
10355 | break; | |
10356 | case OMP_CLAUSE_TO: | |
10357 | tkind = OMP_CLAUSE_MAP_TO; | |
10358 | break; | |
10359 | case OMP_CLAUSE_FROM: | |
10360 | tkind = OMP_CLAUSE_MAP_FROM; | |
10361 | break; | |
10362 | default: | |
10363 | gcc_unreachable (); | |
10364 | } | |
bc7bff74 | 10365 | talign = ceil_log2 (talign); |
10366 | tkind |= talign << 3; | |
10367 | CONSTRUCTOR_APPEND_ELT (vkind, purpose, | |
10368 | build_int_cst (unsigned_char_type_node, | |
10369 | tkind)); | |
10370 | if (nc && nc != c) | |
10371 | c = nc; | |
10372 | } | |
10373 | ||
10374 | gcc_assert (map_idx == map_cnt); | |
10375 | ||
10376 | DECL_INITIAL (TREE_VEC_ELT (t, 1)) | |
10377 | = build_constructor (TREE_TYPE (TREE_VEC_ELT (t, 1)), vsize); | |
10378 | DECL_INITIAL (TREE_VEC_ELT (t, 2)) | |
10379 | = build_constructor (TREE_TYPE (TREE_VEC_ELT (t, 2)), vkind); | |
10380 | if (!TREE_STATIC (TREE_VEC_ELT (t, 1))) | |
10381 | { | |
10382 | gimple_seq initlist = NULL; | |
10383 | force_gimple_operand (build1 (DECL_EXPR, void_type_node, | |
10384 | TREE_VEC_ELT (t, 1)), | |
10385 | &initlist, true, NULL_TREE); | |
10386 | gimple_seq_add_seq (&ilist, initlist); | |
d8e1e68d | 10387 | |
10388 | tree clobber = build_constructor (TREE_TYPE (TREE_VEC_ELT (t, 1)), | |
10389 | NULL); | |
10390 | TREE_THIS_VOLATILE (clobber) = 1; | |
10391 | gimple_seq_add_stmt (&olist, | |
10392 | gimple_build_assign (TREE_VEC_ELT (t, 1), | |
10393 | clobber)); | |
bc7bff74 | 10394 | } |
10395 | ||
10396 | tree clobber = build_constructor (ctx->record_type, NULL); | |
10397 | TREE_THIS_VOLATILE (clobber) = 1; | |
10398 | gimple_seq_add_stmt (&olist, gimple_build_assign (ctx->sender_decl, | |
10399 | clobber)); | |
10400 | } | |
10401 | ||
10402 | /* Once all the expansions are done, sequence all the different | |
10403 | fragments inside gimple_omp_body. */ | |
10404 | ||
10405 | new_body = NULL; | |
10406 | ||
10407 | if (ctx->record_type && kind == GF_OMP_TARGET_KIND_REGION) | |
10408 | { | |
10409 | t = build_fold_addr_expr_loc (loc, ctx->sender_decl); | |
10410 | /* fixup_child_record_type might have changed receiver_decl's type. */ | |
10411 | t = fold_convert_loc (loc, TREE_TYPE (ctx->receiver_decl), t); | |
10412 | gimple_seq_add_stmt (&new_body, | |
10413 | gimple_build_assign (ctx->receiver_decl, t)); | |
10414 | } | |
10415 | ||
10416 | if (kind == GF_OMP_TARGET_KIND_REGION) | |
10417 | { | |
10418 | gimple_seq_add_seq (&new_body, tgt_body); | |
10419 | new_body = maybe_catch_exception (new_body); | |
10420 | } | |
10421 | else if (kind == GF_OMP_TARGET_KIND_DATA) | |
10422 | new_body = tgt_body; | |
10423 | if (kind != GF_OMP_TARGET_KIND_UPDATE) | |
10424 | { | |
10425 | gimple_seq_add_stmt (&new_body, gimple_build_omp_return (false)); | |
10426 | gimple_omp_set_body (stmt, new_body); | |
10427 | } | |
10428 | ||
10429 | bind = gimple_build_bind (NULL, NULL, | |
10430 | tgt_bind ? gimple_bind_block (tgt_bind) | |
10431 | : NULL_TREE); | |
75a70cf9 | 10432 | gsi_replace (gsi_p, bind, true); |
e3a19533 | 10433 | gimple_bind_add_seq (bind, ilist); |
10434 | gimple_bind_add_stmt (bind, stmt); | |
10435 | gimple_bind_add_seq (bind, olist); | |
773c5ba7 | 10436 | |
75a70cf9 | 10437 | pop_gimplify_context (NULL); |
773c5ba7 | 10438 | } |
10439 | ||
bc7bff74 | 10440 | /* Expand code for an OpenMP teams directive. */ |
10441 | ||
10442 | static void | |
10443 | lower_omp_teams (gimple_stmt_iterator *gsi_p, omp_context *ctx) | |
10444 | { | |
1a91d914 | 10445 | gomp_teams *teams_stmt = as_a <gomp_teams *> (gsi_stmt (*gsi_p)); |
8a4a28a8 | 10446 | push_gimplify_context (); |
bc7bff74 | 10447 | |
10448 | tree block = make_node (BLOCK); | |
1a91d914 | 10449 | gbind *bind = gimple_build_bind (NULL, NULL, block); |
bc7bff74 | 10450 | gsi_replace (gsi_p, bind, true); |
10451 | gimple_seq bind_body = NULL; | |
10452 | gimple_seq dlist = NULL; | |
10453 | gimple_seq olist = NULL; | |
10454 | ||
10455 | tree num_teams = find_omp_clause (gimple_omp_teams_clauses (teams_stmt), | |
10456 | OMP_CLAUSE_NUM_TEAMS); | |
10457 | if (num_teams == NULL_TREE) | |
10458 | num_teams = build_int_cst (unsigned_type_node, 0); | |
10459 | else | |
10460 | { | |
10461 | num_teams = OMP_CLAUSE_NUM_TEAMS_EXPR (num_teams); | |
10462 | num_teams = fold_convert (unsigned_type_node, num_teams); | |
10463 | gimplify_expr (&num_teams, &bind_body, NULL, is_gimple_val, fb_rvalue); | |
10464 | } | |
10465 | tree thread_limit = find_omp_clause (gimple_omp_teams_clauses (teams_stmt), | |
10466 | OMP_CLAUSE_THREAD_LIMIT); | |
10467 | if (thread_limit == NULL_TREE) | |
10468 | thread_limit = build_int_cst (unsigned_type_node, 0); | |
10469 | else | |
10470 | { | |
10471 | thread_limit = OMP_CLAUSE_THREAD_LIMIT_EXPR (thread_limit); | |
10472 | thread_limit = fold_convert (unsigned_type_node, thread_limit); | |
10473 | gimplify_expr (&thread_limit, &bind_body, NULL, is_gimple_val, | |
10474 | fb_rvalue); | |
10475 | } | |
10476 | ||
10477 | lower_rec_input_clauses (gimple_omp_teams_clauses (teams_stmt), | |
10478 | &bind_body, &dlist, ctx, NULL); | |
10479 | lower_omp (gimple_omp_body_ptr (teams_stmt), ctx); | |
10480 | lower_reduction_clauses (gimple_omp_teams_clauses (teams_stmt), &olist, ctx); | |
10481 | gimple_seq_add_stmt (&bind_body, teams_stmt); | |
10482 | ||
10483 | location_t loc = gimple_location (teams_stmt); | |
10484 | tree decl = builtin_decl_explicit (BUILT_IN_GOMP_TEAMS); | |
10485 | gimple call = gimple_build_call (decl, 2, num_teams, thread_limit); | |
10486 | gimple_set_location (call, loc); | |
10487 | gimple_seq_add_stmt (&bind_body, call); | |
10488 | ||
10489 | gimple_seq_add_seq (&bind_body, gimple_omp_body (teams_stmt)); | |
10490 | gimple_omp_set_body (teams_stmt, NULL); | |
10491 | gimple_seq_add_seq (&bind_body, olist); | |
10492 | gimple_seq_add_seq (&bind_body, dlist); | |
10493 | gimple_seq_add_stmt (&bind_body, gimple_build_omp_return (true)); | |
10494 | gimple_bind_set_body (bind, bind_body); | |
10495 | ||
10496 | pop_gimplify_context (bind); | |
10497 | ||
10498 | gimple_bind_append_vars (bind, ctx->block_vars); | |
10499 | BLOCK_VARS (block) = ctx->block_vars; | |
10500 | if (BLOCK_VARS (block)) | |
10501 | TREE_USED (block) = 1; | |
10502 | } | |
10503 | ||
10504 | ||
a4890dc9 | 10505 | /* Callback for lower_omp_1. Return non-NULL if *tp needs to be |
75a70cf9 | 10506 | regimplified. If DATA is non-NULL, lower_omp_1 is outside |
10507 | of OpenMP context, but with task_shared_vars set. */ | |
46515c92 | 10508 | |
10509 | static tree | |
75a70cf9 | 10510 | lower_omp_regimplify_p (tree *tp, int *walk_subtrees, |
10511 | void *data) | |
46515c92 | 10512 | { |
a4890dc9 | 10513 | tree t = *tp; |
46515c92 | 10514 | |
a4890dc9 | 10515 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ |
75a70cf9 | 10516 | if (TREE_CODE (t) == VAR_DECL && data == NULL && DECL_HAS_VALUE_EXPR_P (t)) |
9f49e155 | 10517 | return t; |
10518 | ||
10519 | if (task_shared_vars | |
10520 | && DECL_P (t) | |
10521 | && bitmap_bit_p (task_shared_vars, DECL_UID (t))) | |
a4890dc9 | 10522 | return t; |
46515c92 | 10523 | |
a4890dc9 | 10524 | /* If a global variable has been privatized, TREE_CONSTANT on |
10525 | ADDR_EXPR might be wrong. */ | |
75a70cf9 | 10526 | if (data == NULL && TREE_CODE (t) == ADDR_EXPR) |
a4890dc9 | 10527 | recompute_tree_invariant_for_addr_expr (t); |
46515c92 | 10528 | |
a4890dc9 | 10529 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); |
10530 | return NULL_TREE; | |
46515c92 | 10531 | } |
773c5ba7 | 10532 | |
a4890dc9 | 10533 | static void |
75a70cf9 | 10534 | lower_omp_1 (gimple_stmt_iterator *gsi_p, omp_context *ctx) |
1e8e9920 | 10535 | { |
75a70cf9 | 10536 | gimple stmt = gsi_stmt (*gsi_p); |
10537 | struct walk_stmt_info wi; | |
1a91d914 | 10538 | gcall *call_stmt; |
1e8e9920 | 10539 | |
75a70cf9 | 10540 | if (gimple_has_location (stmt)) |
10541 | input_location = gimple_location (stmt); | |
a4890dc9 | 10542 | |
75a70cf9 | 10543 | if (task_shared_vars) |
10544 | memset (&wi, '\0', sizeof (wi)); | |
a4890dc9 | 10545 | |
773c5ba7 | 10546 | /* If we have issued syntax errors, avoid doing any heavy lifting. |
10547 | Just replace the OpenMP directives with a NOP to avoid | |
10548 | confusing RTL expansion. */ | |
852f689e | 10549 | if (seen_error () && is_gimple_omp (stmt)) |
773c5ba7 | 10550 | { |
75a70cf9 | 10551 | gsi_replace (gsi_p, gimple_build_nop (), true); |
a4890dc9 | 10552 | return; |
773c5ba7 | 10553 | } |
10554 | ||
75a70cf9 | 10555 | switch (gimple_code (stmt)) |
1e8e9920 | 10556 | { |
75a70cf9 | 10557 | case GIMPLE_COND: |
1a91d914 | 10558 | { |
10559 | gcond *cond_stmt = as_a <gcond *> (stmt); | |
10560 | if ((ctx || task_shared_vars) | |
10561 | && (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
10562 | lower_omp_regimplify_p, | |
10563 | ctx ? NULL : &wi, NULL) | |
10564 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
10565 | lower_omp_regimplify_p, | |
10566 | ctx ? NULL : &wi, NULL))) | |
10567 | gimple_regimplify_operands (cond_stmt, gsi_p); | |
10568 | } | |
a4890dc9 | 10569 | break; |
75a70cf9 | 10570 | case GIMPLE_CATCH: |
1a91d914 | 10571 | lower_omp (gimple_catch_handler_ptr (as_a <gcatch *> (stmt)), ctx); |
a4890dc9 | 10572 | break; |
75a70cf9 | 10573 | case GIMPLE_EH_FILTER: |
e3a19533 | 10574 | lower_omp (gimple_eh_filter_failure_ptr (stmt), ctx); |
a4890dc9 | 10575 | break; |
75a70cf9 | 10576 | case GIMPLE_TRY: |
e3a19533 | 10577 | lower_omp (gimple_try_eval_ptr (stmt), ctx); |
10578 | lower_omp (gimple_try_cleanup_ptr (stmt), ctx); | |
a4890dc9 | 10579 | break; |
35215227 | 10580 | case GIMPLE_TRANSACTION: |
1a91d914 | 10581 | lower_omp (gimple_transaction_body_ptr ( |
10582 | as_a <gtransaction *> (stmt)), | |
10583 | ctx); | |
35215227 | 10584 | break; |
75a70cf9 | 10585 | case GIMPLE_BIND: |
1a91d914 | 10586 | lower_omp (gimple_bind_body_ptr (as_a <gbind *> (stmt)), ctx); |
a4890dc9 | 10587 | break; |
75a70cf9 | 10588 | case GIMPLE_OMP_PARALLEL: |
10589 | case GIMPLE_OMP_TASK: | |
10590 | ctx = maybe_lookup_ctx (stmt); | |
bc7bff74 | 10591 | gcc_assert (ctx); |
10592 | if (ctx->cancellable) | |
10593 | ctx->cancel_label = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 10594 | lower_omp_taskreg (gsi_p, ctx); |
a4890dc9 | 10595 | break; |
75a70cf9 | 10596 | case GIMPLE_OMP_FOR: |
10597 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10598 | gcc_assert (ctx); |
bc7bff74 | 10599 | if (ctx->cancellable) |
10600 | ctx->cancel_label = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 10601 | lower_omp_for (gsi_p, ctx); |
1e8e9920 | 10602 | break; |
75a70cf9 | 10603 | case GIMPLE_OMP_SECTIONS: |
10604 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10605 | gcc_assert (ctx); |
bc7bff74 | 10606 | if (ctx->cancellable) |
10607 | ctx->cancel_label = create_artificial_label (UNKNOWN_LOCATION); | |
75a70cf9 | 10608 | lower_omp_sections (gsi_p, ctx); |
1e8e9920 | 10609 | break; |
75a70cf9 | 10610 | case GIMPLE_OMP_SINGLE: |
10611 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10612 | gcc_assert (ctx); |
75a70cf9 | 10613 | lower_omp_single (gsi_p, ctx); |
1e8e9920 | 10614 | break; |
75a70cf9 | 10615 | case GIMPLE_OMP_MASTER: |
10616 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10617 | gcc_assert (ctx); |
75a70cf9 | 10618 | lower_omp_master (gsi_p, ctx); |
1e8e9920 | 10619 | break; |
bc7bff74 | 10620 | case GIMPLE_OMP_TASKGROUP: |
10621 | ctx = maybe_lookup_ctx (stmt); | |
10622 | gcc_assert (ctx); | |
10623 | lower_omp_taskgroup (gsi_p, ctx); | |
10624 | break; | |
75a70cf9 | 10625 | case GIMPLE_OMP_ORDERED: |
10626 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10627 | gcc_assert (ctx); |
75a70cf9 | 10628 | lower_omp_ordered (gsi_p, ctx); |
1e8e9920 | 10629 | break; |
75a70cf9 | 10630 | case GIMPLE_OMP_CRITICAL: |
10631 | ctx = maybe_lookup_ctx (stmt); | |
1e8e9920 | 10632 | gcc_assert (ctx); |
75a70cf9 | 10633 | lower_omp_critical (gsi_p, ctx); |
10634 | break; | |
10635 | case GIMPLE_OMP_ATOMIC_LOAD: | |
10636 | if ((ctx || task_shared_vars) | |
1a91d914 | 10637 | && walk_tree (gimple_omp_atomic_load_rhs_ptr ( |
10638 | as_a <gomp_atomic_load *> (stmt)), | |
75a70cf9 | 10639 | lower_omp_regimplify_p, ctx ? NULL : &wi, NULL)) |
10640 | gimple_regimplify_operands (stmt, gsi_p); | |
1e8e9920 | 10641 | break; |
bc7bff74 | 10642 | case GIMPLE_OMP_TARGET: |
10643 | ctx = maybe_lookup_ctx (stmt); | |
10644 | gcc_assert (ctx); | |
10645 | lower_omp_target (gsi_p, ctx); | |
10646 | break; | |
10647 | case GIMPLE_OMP_TEAMS: | |
10648 | ctx = maybe_lookup_ctx (stmt); | |
10649 | gcc_assert (ctx); | |
10650 | lower_omp_teams (gsi_p, ctx); | |
10651 | break; | |
10652 | case GIMPLE_CALL: | |
10653 | tree fndecl; | |
1a91d914 | 10654 | call_stmt = as_a <gcall *> (stmt); |
10655 | fndecl = gimple_call_fndecl (call_stmt); | |
bc7bff74 | 10656 | if (fndecl |
10657 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
10658 | switch (DECL_FUNCTION_CODE (fndecl)) | |
10659 | { | |
10660 | case BUILT_IN_GOMP_BARRIER: | |
10661 | if (ctx == NULL) | |
10662 | break; | |
10663 | /* FALLTHRU */ | |
10664 | case BUILT_IN_GOMP_CANCEL: | |
10665 | case BUILT_IN_GOMP_CANCELLATION_POINT: | |
10666 | omp_context *cctx; | |
10667 | cctx = ctx; | |
10668 | if (gimple_code (cctx->stmt) == GIMPLE_OMP_SECTION) | |
10669 | cctx = cctx->outer; | |
1a91d914 | 10670 | gcc_assert (gimple_call_lhs (call_stmt) == NULL_TREE); |
bc7bff74 | 10671 | if (!cctx->cancellable) |
10672 | { | |
10673 | if (DECL_FUNCTION_CODE (fndecl) | |
10674 | == BUILT_IN_GOMP_CANCELLATION_POINT) | |
10675 | { | |
10676 | stmt = gimple_build_nop (); | |
10677 | gsi_replace (gsi_p, stmt, false); | |
10678 | } | |
10679 | break; | |
10680 | } | |
bc7bff74 | 10681 | if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_GOMP_BARRIER) |
10682 | { | |
10683 | fndecl = builtin_decl_explicit (BUILT_IN_GOMP_BARRIER_CANCEL); | |
1a91d914 | 10684 | gimple_call_set_fndecl (call_stmt, fndecl); |
10685 | gimple_call_set_fntype (call_stmt, TREE_TYPE (fndecl)); | |
bc7bff74 | 10686 | } |
15b28553 | 10687 | tree lhs; |
f9e245b2 | 10688 | lhs = create_tmp_var (TREE_TYPE (TREE_TYPE (fndecl))); |
1a91d914 | 10689 | gimple_call_set_lhs (call_stmt, lhs); |
bc7bff74 | 10690 | tree fallthru_label; |
10691 | fallthru_label = create_artificial_label (UNKNOWN_LOCATION); | |
10692 | gimple g; | |
10693 | g = gimple_build_label (fallthru_label); | |
10694 | gsi_insert_after (gsi_p, g, GSI_SAME_STMT); | |
15b28553 | 10695 | g = gimple_build_cond (NE_EXPR, lhs, |
10696 | fold_convert (TREE_TYPE (lhs), | |
10697 | boolean_false_node), | |
bc7bff74 | 10698 | cctx->cancel_label, fallthru_label); |
10699 | gsi_insert_after (gsi_p, g, GSI_SAME_STMT); | |
10700 | break; | |
10701 | default: | |
10702 | break; | |
10703 | } | |
10704 | /* FALLTHRU */ | |
a4890dc9 | 10705 | default: |
fd6481cf | 10706 | if ((ctx || task_shared_vars) |
75a70cf9 | 10707 | && walk_gimple_op (stmt, lower_omp_regimplify_p, |
10708 | ctx ? NULL : &wi)) | |
ef8cfd4d | 10709 | { |
10710 | /* Just remove clobbers, this should happen only if we have | |
10711 | "privatized" local addressable variables in SIMD regions, | |
10712 | the clobber isn't needed in that case and gimplifying address | |
10713 | of the ARRAY_REF into a pointer and creating MEM_REF based | |
10714 | clobber would create worse code than we get with the clobber | |
10715 | dropped. */ | |
10716 | if (gimple_clobber_p (stmt)) | |
10717 | { | |
10718 | gsi_replace (gsi_p, gimple_build_nop (), true); | |
10719 | break; | |
10720 | } | |
10721 | gimple_regimplify_operands (stmt, gsi_p); | |
10722 | } | |
1e8e9920 | 10723 | break; |
1e8e9920 | 10724 | } |
1e8e9920 | 10725 | } |
10726 | ||
10727 | static void | |
e3a19533 | 10728 | lower_omp (gimple_seq *body, omp_context *ctx) |
1e8e9920 | 10729 | { |
1d22f541 | 10730 | location_t saved_location = input_location; |
e3a19533 | 10731 | gimple_stmt_iterator gsi; |
10732 | for (gsi = gsi_start (*body); !gsi_end_p (gsi); gsi_next (&gsi)) | |
75a70cf9 | 10733 | lower_omp_1 (&gsi, ctx); |
c09f306d | 10734 | /* During gimplification, we have not always invoked fold_stmt |
10735 | (gimplify.c:maybe_fold_stmt); call it now. */ | |
bc7bff74 | 10736 | if (target_nesting_level) |
10737 | for (gsi = gsi_start (*body); !gsi_end_p (gsi); gsi_next (&gsi)) | |
10738 | fold_stmt (&gsi); | |
1d22f541 | 10739 | input_location = saved_location; |
1e8e9920 | 10740 | } |
10741 | \f | |
10742 | /* Main entry point. */ | |
10743 | ||
2a1990e9 | 10744 | static unsigned int |
1e8e9920 | 10745 | execute_lower_omp (void) |
10746 | { | |
75a70cf9 | 10747 | gimple_seq body; |
37eaded9 | 10748 | int i; |
10749 | omp_context *ctx; | |
75a70cf9 | 10750 | |
41709826 | 10751 | /* This pass always runs, to provide PROP_gimple_lomp. |
10752 | But there is nothing to do unless -fopenmp is given. */ | |
a89e6c15 | 10753 | if (flag_openmp == 0 && flag_openmp_simd == 0 && flag_cilkplus == 0) |
41709826 | 10754 | return 0; |
10755 | ||
1e8e9920 | 10756 | all_contexts = splay_tree_new (splay_tree_compare_pointers, 0, |
10757 | delete_omp_context); | |
10758 | ||
75a70cf9 | 10759 | body = gimple_body (current_function_decl); |
ab129075 | 10760 | scan_omp (&body, NULL); |
fd6481cf | 10761 | gcc_assert (taskreg_nesting_level == 0); |
37eaded9 | 10762 | FOR_EACH_VEC_ELT (taskreg_contexts, i, ctx) |
10763 | finish_taskreg_scan (ctx); | |
10764 | taskreg_contexts.release (); | |
1e8e9920 | 10765 | |
10766 | if (all_contexts->root) | |
fd6481cf | 10767 | { |
10768 | if (task_shared_vars) | |
8a4a28a8 | 10769 | push_gimplify_context (); |
e3a19533 | 10770 | lower_omp (&body, NULL); |
fd6481cf | 10771 | if (task_shared_vars) |
10772 | pop_gimplify_context (NULL); | |
10773 | } | |
1e8e9920 | 10774 | |
773c5ba7 | 10775 | if (all_contexts) |
10776 | { | |
10777 | splay_tree_delete (all_contexts); | |
10778 | all_contexts = NULL; | |
10779 | } | |
fd6481cf | 10780 | BITMAP_FREE (task_shared_vars); |
2a1990e9 | 10781 | return 0; |
1e8e9920 | 10782 | } |
10783 | ||
cbe8bda8 | 10784 | namespace { |
10785 | ||
10786 | const pass_data pass_data_lower_omp = | |
10787 | { | |
10788 | GIMPLE_PASS, /* type */ | |
10789 | "omplower", /* name */ | |
10790 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 10791 | TV_NONE, /* tv_id */ |
10792 | PROP_gimple_any, /* properties_required */ | |
10793 | PROP_gimple_lomp, /* properties_provided */ | |
10794 | 0, /* properties_destroyed */ | |
10795 | 0, /* todo_flags_start */ | |
10796 | 0, /* todo_flags_finish */ | |
1e8e9920 | 10797 | }; |
cbe8bda8 | 10798 | |
10799 | class pass_lower_omp : public gimple_opt_pass | |
10800 | { | |
10801 | public: | |
9af5ce0c | 10802 | pass_lower_omp (gcc::context *ctxt) |
10803 | : gimple_opt_pass (pass_data_lower_omp, ctxt) | |
cbe8bda8 | 10804 | {} |
10805 | ||
10806 | /* opt_pass methods: */ | |
65b0537f | 10807 | virtual unsigned int execute (function *) { return execute_lower_omp (); } |
cbe8bda8 | 10808 | |
10809 | }; // class pass_lower_omp | |
10810 | ||
10811 | } // anon namespace | |
10812 | ||
10813 | gimple_opt_pass * | |
10814 | make_pass_lower_omp (gcc::context *ctxt) | |
10815 | { | |
10816 | return new pass_lower_omp (ctxt); | |
10817 | } | |
1e8e9920 | 10818 | \f |
10819 | /* The following is a utility to diagnose OpenMP structured block violations. | |
61e47ac8 | 10820 | It is not part of the "omplower" pass, as that's invoked too late. It |
10821 | should be invoked by the respective front ends after gimplification. */ | |
1e8e9920 | 10822 | |
10823 | static splay_tree all_labels; | |
10824 | ||
10825 | /* Check for mismatched contexts and generate an error if needed. Return | |
10826 | true if an error is detected. */ | |
10827 | ||
10828 | static bool | |
75a70cf9 | 10829 | diagnose_sb_0 (gimple_stmt_iterator *gsi_p, |
10830 | gimple branch_ctx, gimple label_ctx) | |
1e8e9920 | 10831 | { |
75a70cf9 | 10832 | if (label_ctx == branch_ctx) |
1e8e9920 | 10833 | return false; |
10834 | ||
48e1416a | 10835 | |
75a70cf9 | 10836 | /* |
10837 | Previously we kept track of the label's entire context in diagnose_sb_[12] | |
10838 | so we could traverse it and issue a correct "exit" or "enter" error | |
10839 | message upon a structured block violation. | |
10840 | ||
10841 | We built the context by building a list with tree_cons'ing, but there is | |
10842 | no easy counterpart in gimple tuples. It seems like far too much work | |
10843 | for issuing exit/enter error messages. If someone really misses the | |
10844 | distinct error message... patches welcome. | |
10845 | */ | |
48e1416a | 10846 | |
75a70cf9 | 10847 | #if 0 |
1e8e9920 | 10848 | /* Try to avoid confusing the user by producing and error message |
f0b5f617 | 10849 | with correct "exit" or "enter" verbiage. We prefer "exit" |
1e8e9920 | 10850 | unless we can show that LABEL_CTX is nested within BRANCH_CTX. */ |
10851 | if (branch_ctx == NULL) | |
10852 | exit_p = false; | |
10853 | else | |
10854 | { | |
10855 | while (label_ctx) | |
10856 | { | |
10857 | if (TREE_VALUE (label_ctx) == branch_ctx) | |
10858 | { | |
10859 | exit_p = false; | |
10860 | break; | |
10861 | } | |
10862 | label_ctx = TREE_CHAIN (label_ctx); | |
10863 | } | |
10864 | } | |
10865 | ||
10866 | if (exit_p) | |
10867 | error ("invalid exit from OpenMP structured block"); | |
10868 | else | |
10869 | error ("invalid entry to OpenMP structured block"); | |
75a70cf9 | 10870 | #endif |
1e8e9920 | 10871 | |
f2697631 | 10872 | bool cilkplus_block = false; |
a89e6c15 | 10873 | if (flag_cilkplus) |
f2697631 | 10874 | { |
10875 | if ((branch_ctx | |
10876 | && gimple_code (branch_ctx) == GIMPLE_OMP_FOR | |
10877 | && gimple_omp_for_kind (branch_ctx) == GF_OMP_FOR_KIND_CILKSIMD) | |
a4eb3f73 | 10878 | || (label_ctx |
10879 | && gimple_code (label_ctx) == GIMPLE_OMP_FOR | |
f2697631 | 10880 | && gimple_omp_for_kind (label_ctx) == GF_OMP_FOR_KIND_CILKSIMD)) |
10881 | cilkplus_block = true; | |
10882 | } | |
10883 | ||
75a70cf9 | 10884 | /* If it's obvious we have an invalid entry, be specific about the error. */ |
10885 | if (branch_ctx == NULL) | |
f2697631 | 10886 | { |
10887 | if (cilkplus_block) | |
10888 | error ("invalid entry to Cilk Plus structured block"); | |
10889 | else | |
10890 | error ("invalid entry to OpenMP structured block"); | |
10891 | } | |
75a70cf9 | 10892 | else |
f2697631 | 10893 | { |
10894 | /* Otherwise, be vague and lazy, but efficient. */ | |
10895 | if (cilkplus_block) | |
10896 | error ("invalid branch to/from a Cilk Plus structured block"); | |
10897 | else | |
10898 | error ("invalid branch to/from an OpenMP structured block"); | |
10899 | } | |
75a70cf9 | 10900 | |
10901 | gsi_replace (gsi_p, gimple_build_nop (), false); | |
1e8e9920 | 10902 | return true; |
10903 | } | |
10904 | ||
10905 | /* Pass 1: Create a minimal tree of OpenMP structured blocks, and record | |
75a70cf9 | 10906 | where each label is found. */ |
1e8e9920 | 10907 | |
10908 | static tree | |
75a70cf9 | 10909 | diagnose_sb_1 (gimple_stmt_iterator *gsi_p, bool *handled_ops_p, |
10910 | struct walk_stmt_info *wi) | |
1e8e9920 | 10911 | { |
75a70cf9 | 10912 | gimple context = (gimple) wi->info; |
10913 | gimple inner_context; | |
10914 | gimple stmt = gsi_stmt (*gsi_p); | |
1e8e9920 | 10915 | |
75a70cf9 | 10916 | *handled_ops_p = true; |
10917 | ||
10918 | switch (gimple_code (stmt)) | |
1e8e9920 | 10919 | { |
75a70cf9 | 10920 | WALK_SUBSTMTS; |
48e1416a | 10921 | |
75a70cf9 | 10922 | case GIMPLE_OMP_PARALLEL: |
10923 | case GIMPLE_OMP_TASK: | |
10924 | case GIMPLE_OMP_SECTIONS: | |
10925 | case GIMPLE_OMP_SINGLE: | |
10926 | case GIMPLE_OMP_SECTION: | |
10927 | case GIMPLE_OMP_MASTER: | |
10928 | case GIMPLE_OMP_ORDERED: | |
10929 | case GIMPLE_OMP_CRITICAL: | |
bc7bff74 | 10930 | case GIMPLE_OMP_TARGET: |
10931 | case GIMPLE_OMP_TEAMS: | |
10932 | case GIMPLE_OMP_TASKGROUP: | |
75a70cf9 | 10933 | /* The minimal context here is just the current OMP construct. */ |
10934 | inner_context = stmt; | |
1e8e9920 | 10935 | wi->info = inner_context; |
75a70cf9 | 10936 | walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_1, NULL, wi); |
1e8e9920 | 10937 | wi->info = context; |
10938 | break; | |
10939 | ||
75a70cf9 | 10940 | case GIMPLE_OMP_FOR: |
10941 | inner_context = stmt; | |
1e8e9920 | 10942 | wi->info = inner_context; |
75a70cf9 | 10943 | /* gimple_omp_for_{index,initial,final} are all DECLs; no need to |
10944 | walk them. */ | |
10945 | walk_gimple_seq (gimple_omp_for_pre_body (stmt), | |
10946 | diagnose_sb_1, NULL, wi); | |
10947 | walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_1, NULL, wi); | |
1e8e9920 | 10948 | wi->info = context; |
10949 | break; | |
10950 | ||
75a70cf9 | 10951 | case GIMPLE_LABEL: |
1a91d914 | 10952 | splay_tree_insert (all_labels, |
10953 | (splay_tree_key) gimple_label_label ( | |
10954 | as_a <glabel *> (stmt)), | |
1e8e9920 | 10955 | (splay_tree_value) context); |
10956 | break; | |
10957 | ||
10958 | default: | |
10959 | break; | |
10960 | } | |
10961 | ||
10962 | return NULL_TREE; | |
10963 | } | |
10964 | ||
10965 | /* Pass 2: Check each branch and see if its context differs from that of | |
10966 | the destination label's context. */ | |
10967 | ||
10968 | static tree | |
75a70cf9 | 10969 | diagnose_sb_2 (gimple_stmt_iterator *gsi_p, bool *handled_ops_p, |
10970 | struct walk_stmt_info *wi) | |
1e8e9920 | 10971 | { |
75a70cf9 | 10972 | gimple context = (gimple) wi->info; |
1e8e9920 | 10973 | splay_tree_node n; |
75a70cf9 | 10974 | gimple stmt = gsi_stmt (*gsi_p); |
1e8e9920 | 10975 | |
75a70cf9 | 10976 | *handled_ops_p = true; |
10977 | ||
10978 | switch (gimple_code (stmt)) | |
1e8e9920 | 10979 | { |
75a70cf9 | 10980 | WALK_SUBSTMTS; |
10981 | ||
10982 | case GIMPLE_OMP_PARALLEL: | |
10983 | case GIMPLE_OMP_TASK: | |
10984 | case GIMPLE_OMP_SECTIONS: | |
10985 | case GIMPLE_OMP_SINGLE: | |
10986 | case GIMPLE_OMP_SECTION: | |
10987 | case GIMPLE_OMP_MASTER: | |
10988 | case GIMPLE_OMP_ORDERED: | |
10989 | case GIMPLE_OMP_CRITICAL: | |
bc7bff74 | 10990 | case GIMPLE_OMP_TARGET: |
10991 | case GIMPLE_OMP_TEAMS: | |
10992 | case GIMPLE_OMP_TASKGROUP: | |
75a70cf9 | 10993 | wi->info = stmt; |
e3a19533 | 10994 | walk_gimple_seq_mod (gimple_omp_body_ptr (stmt), diagnose_sb_2, NULL, wi); |
1e8e9920 | 10995 | wi->info = context; |
10996 | break; | |
10997 | ||
75a70cf9 | 10998 | case GIMPLE_OMP_FOR: |
10999 | wi->info = stmt; | |
11000 | /* gimple_omp_for_{index,initial,final} are all DECLs; no need to | |
11001 | walk them. */ | |
e3a19533 | 11002 | walk_gimple_seq_mod (gimple_omp_for_pre_body_ptr (stmt), |
11003 | diagnose_sb_2, NULL, wi); | |
11004 | walk_gimple_seq_mod (gimple_omp_body_ptr (stmt), diagnose_sb_2, NULL, wi); | |
1e8e9920 | 11005 | wi->info = context; |
11006 | break; | |
11007 | ||
0e1818e7 | 11008 | case GIMPLE_COND: |
11009 | { | |
1a91d914 | 11010 | gcond *cond_stmt = as_a <gcond *> (stmt); |
11011 | tree lab = gimple_cond_true_label (cond_stmt); | |
0e1818e7 | 11012 | if (lab) |
11013 | { | |
11014 | n = splay_tree_lookup (all_labels, | |
11015 | (splay_tree_key) lab); | |
11016 | diagnose_sb_0 (gsi_p, context, | |
11017 | n ? (gimple) n->value : NULL); | |
11018 | } | |
1a91d914 | 11019 | lab = gimple_cond_false_label (cond_stmt); |
0e1818e7 | 11020 | if (lab) |
11021 | { | |
11022 | n = splay_tree_lookup (all_labels, | |
11023 | (splay_tree_key) lab); | |
11024 | diagnose_sb_0 (gsi_p, context, | |
11025 | n ? (gimple) n->value : NULL); | |
11026 | } | |
11027 | } | |
11028 | break; | |
11029 | ||
75a70cf9 | 11030 | case GIMPLE_GOTO: |
1e8e9920 | 11031 | { |
75a70cf9 | 11032 | tree lab = gimple_goto_dest (stmt); |
1e8e9920 | 11033 | if (TREE_CODE (lab) != LABEL_DECL) |
11034 | break; | |
11035 | ||
11036 | n = splay_tree_lookup (all_labels, (splay_tree_key) lab); | |
75a70cf9 | 11037 | diagnose_sb_0 (gsi_p, context, n ? (gimple) n->value : NULL); |
1e8e9920 | 11038 | } |
11039 | break; | |
11040 | ||
75a70cf9 | 11041 | case GIMPLE_SWITCH: |
1e8e9920 | 11042 | { |
1a91d914 | 11043 | gswitch *switch_stmt = as_a <gswitch *> (stmt); |
75a70cf9 | 11044 | unsigned int i; |
1a91d914 | 11045 | for (i = 0; i < gimple_switch_num_labels (switch_stmt); ++i) |
1e8e9920 | 11046 | { |
1a91d914 | 11047 | tree lab = CASE_LABEL (gimple_switch_label (switch_stmt, i)); |
1e8e9920 | 11048 | n = splay_tree_lookup (all_labels, (splay_tree_key) lab); |
75a70cf9 | 11049 | if (n && diagnose_sb_0 (gsi_p, context, (gimple) n->value)) |
1e8e9920 | 11050 | break; |
11051 | } | |
11052 | } | |
11053 | break; | |
11054 | ||
75a70cf9 | 11055 | case GIMPLE_RETURN: |
11056 | diagnose_sb_0 (gsi_p, context, NULL); | |
1e8e9920 | 11057 | break; |
11058 | ||
11059 | default: | |
11060 | break; | |
11061 | } | |
11062 | ||
11063 | return NULL_TREE; | |
11064 | } | |
11065 | ||
7740abd8 | 11066 | /* Called from tree-cfg.c::make_edges to create cfg edges for all GIMPLE_OMP |
11067 | codes. */ | |
11068 | bool | |
b2c0e0b7 | 11069 | make_gimple_omp_edges (basic_block bb, struct omp_region **region, |
11070 | int *region_idx) | |
7740abd8 | 11071 | { |
11072 | gimple last = last_stmt (bb); | |
11073 | enum gimple_code code = gimple_code (last); | |
11074 | struct omp_region *cur_region = *region; | |
11075 | bool fallthru = false; | |
11076 | ||
11077 | switch (code) | |
11078 | { | |
11079 | case GIMPLE_OMP_PARALLEL: | |
11080 | case GIMPLE_OMP_TASK: | |
11081 | case GIMPLE_OMP_FOR: | |
11082 | case GIMPLE_OMP_SINGLE: | |
11083 | case GIMPLE_OMP_TEAMS: | |
11084 | case GIMPLE_OMP_MASTER: | |
11085 | case GIMPLE_OMP_TASKGROUP: | |
11086 | case GIMPLE_OMP_ORDERED: | |
11087 | case GIMPLE_OMP_CRITICAL: | |
11088 | case GIMPLE_OMP_SECTION: | |
11089 | cur_region = new_omp_region (bb, code, cur_region); | |
11090 | fallthru = true; | |
11091 | break; | |
11092 | ||
11093 | case GIMPLE_OMP_TARGET: | |
11094 | cur_region = new_omp_region (bb, code, cur_region); | |
11095 | fallthru = true; | |
11096 | if (gimple_omp_target_kind (last) == GF_OMP_TARGET_KIND_UPDATE) | |
11097 | cur_region = cur_region->outer; | |
11098 | break; | |
11099 | ||
11100 | case GIMPLE_OMP_SECTIONS: | |
11101 | cur_region = new_omp_region (bb, code, cur_region); | |
11102 | fallthru = true; | |
11103 | break; | |
11104 | ||
11105 | case GIMPLE_OMP_SECTIONS_SWITCH: | |
11106 | fallthru = false; | |
11107 | break; | |
11108 | ||
11109 | case GIMPLE_OMP_ATOMIC_LOAD: | |
11110 | case GIMPLE_OMP_ATOMIC_STORE: | |
11111 | fallthru = true; | |
11112 | break; | |
11113 | ||
11114 | case GIMPLE_OMP_RETURN: | |
11115 | /* In the case of a GIMPLE_OMP_SECTION, the edge will go | |
11116 | somewhere other than the next block. This will be | |
11117 | created later. */ | |
11118 | cur_region->exit = bb; | |
11119 | fallthru = cur_region->type != GIMPLE_OMP_SECTION; | |
11120 | cur_region = cur_region->outer; | |
11121 | break; | |
11122 | ||
11123 | case GIMPLE_OMP_CONTINUE: | |
11124 | cur_region->cont = bb; | |
11125 | switch (cur_region->type) | |
11126 | { | |
11127 | case GIMPLE_OMP_FOR: | |
11128 | /* Mark all GIMPLE_OMP_FOR and GIMPLE_OMP_CONTINUE | |
11129 | succs edges as abnormal to prevent splitting | |
11130 | them. */ | |
11131 | single_succ_edge (cur_region->entry)->flags |= EDGE_ABNORMAL; | |
11132 | /* Make the loopback edge. */ | |
11133 | make_edge (bb, single_succ (cur_region->entry), | |
11134 | EDGE_ABNORMAL); | |
11135 | ||
11136 | /* Create an edge from GIMPLE_OMP_FOR to exit, which | |
11137 | corresponds to the case that the body of the loop | |
11138 | is not executed at all. */ | |
11139 | make_edge (cur_region->entry, bb->next_bb, EDGE_ABNORMAL); | |
11140 | make_edge (bb, bb->next_bb, EDGE_FALLTHRU | EDGE_ABNORMAL); | |
11141 | fallthru = false; | |
11142 | break; | |
11143 | ||
11144 | case GIMPLE_OMP_SECTIONS: | |
11145 | /* Wire up the edges into and out of the nested sections. */ | |
11146 | { | |
11147 | basic_block switch_bb = single_succ (cur_region->entry); | |
11148 | ||
11149 | struct omp_region *i; | |
11150 | for (i = cur_region->inner; i ; i = i->next) | |
11151 | { | |
11152 | gcc_assert (i->type == GIMPLE_OMP_SECTION); | |
11153 | make_edge (switch_bb, i->entry, 0); | |
11154 | make_edge (i->exit, bb, EDGE_FALLTHRU); | |
11155 | } | |
11156 | ||
11157 | /* Make the loopback edge to the block with | |
11158 | GIMPLE_OMP_SECTIONS_SWITCH. */ | |
11159 | make_edge (bb, switch_bb, 0); | |
11160 | ||
11161 | /* Make the edge from the switch to exit. */ | |
11162 | make_edge (switch_bb, bb->next_bb, 0); | |
11163 | fallthru = false; | |
11164 | } | |
11165 | break; | |
11166 | ||
11167 | default: | |
11168 | gcc_unreachable (); | |
11169 | } | |
11170 | break; | |
11171 | ||
11172 | default: | |
11173 | gcc_unreachable (); | |
11174 | } | |
11175 | ||
11176 | if (*region != cur_region) | |
b2c0e0b7 | 11177 | { |
11178 | *region = cur_region; | |
11179 | if (cur_region) | |
11180 | *region_idx = cur_region->entry->index; | |
11181 | else | |
11182 | *region_idx = 0; | |
11183 | } | |
7740abd8 | 11184 | |
11185 | return fallthru; | |
11186 | } | |
11187 | ||
bfec3452 | 11188 | static unsigned int |
11189 | diagnose_omp_structured_block_errors (void) | |
1e8e9920 | 11190 | { |
1e8e9920 | 11191 | struct walk_stmt_info wi; |
bfec3452 | 11192 | gimple_seq body = gimple_body (current_function_decl); |
1e8e9920 | 11193 | |
11194 | all_labels = splay_tree_new (splay_tree_compare_pointers, 0, 0); | |
11195 | ||
11196 | memset (&wi, 0, sizeof (wi)); | |
75a70cf9 | 11197 | walk_gimple_seq (body, diagnose_sb_1, NULL, &wi); |
1e8e9920 | 11198 | |
11199 | memset (&wi, 0, sizeof (wi)); | |
1e8e9920 | 11200 | wi.want_locations = true; |
e3a19533 | 11201 | walk_gimple_seq_mod (&body, diagnose_sb_2, NULL, &wi); |
11202 | ||
11203 | gimple_set_body (current_function_decl, body); | |
1e8e9920 | 11204 | |
11205 | splay_tree_delete (all_labels); | |
11206 | all_labels = NULL; | |
11207 | ||
bfec3452 | 11208 | return 0; |
1e8e9920 | 11209 | } |
11210 | ||
cbe8bda8 | 11211 | namespace { |
11212 | ||
11213 | const pass_data pass_data_diagnose_omp_blocks = | |
11214 | { | |
11215 | GIMPLE_PASS, /* type */ | |
11216 | "*diagnose_omp_blocks", /* name */ | |
11217 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 11218 | TV_NONE, /* tv_id */ |
11219 | PROP_gimple_any, /* properties_required */ | |
11220 | 0, /* properties_provided */ | |
11221 | 0, /* properties_destroyed */ | |
11222 | 0, /* todo_flags_start */ | |
11223 | 0, /* todo_flags_finish */ | |
bfec3452 | 11224 | }; |
11225 | ||
cbe8bda8 | 11226 | class pass_diagnose_omp_blocks : public gimple_opt_pass |
11227 | { | |
11228 | public: | |
9af5ce0c | 11229 | pass_diagnose_omp_blocks (gcc::context *ctxt) |
11230 | : gimple_opt_pass (pass_data_diagnose_omp_blocks, ctxt) | |
cbe8bda8 | 11231 | {} |
11232 | ||
11233 | /* opt_pass methods: */ | |
31315c24 | 11234 | virtual bool gate (function *) { return flag_openmp || flag_cilkplus; } |
65b0537f | 11235 | virtual unsigned int execute (function *) |
11236 | { | |
11237 | return diagnose_omp_structured_block_errors (); | |
11238 | } | |
cbe8bda8 | 11239 | |
11240 | }; // class pass_diagnose_omp_blocks | |
11241 | ||
11242 | } // anon namespace | |
11243 | ||
11244 | gimple_opt_pass * | |
11245 | make_pass_diagnose_omp_blocks (gcc::context *ctxt) | |
11246 | { | |
11247 | return new pass_diagnose_omp_blocks (ctxt); | |
11248 | } | |
d09768a4 | 11249 | \f |
11250 | /* SIMD clone supporting code. */ | |
11251 | ||
11252 | /* Allocate a fresh `simd_clone' and return it. NARGS is the number | |
11253 | of arguments to reserve space for. */ | |
11254 | ||
11255 | static struct cgraph_simd_clone * | |
11256 | simd_clone_struct_alloc (int nargs) | |
11257 | { | |
11258 | struct cgraph_simd_clone *clone_info; | |
11259 | size_t len = (sizeof (struct cgraph_simd_clone) | |
11260 | + nargs * sizeof (struct cgraph_simd_clone_arg)); | |
11261 | clone_info = (struct cgraph_simd_clone *) | |
ea7d8c7a | 11262 | ggc_internal_cleared_alloc (len); |
d09768a4 | 11263 | return clone_info; |
11264 | } | |
11265 | ||
11266 | /* Make a copy of the `struct cgraph_simd_clone' in FROM to TO. */ | |
11267 | ||
11268 | static inline void | |
11269 | simd_clone_struct_copy (struct cgraph_simd_clone *to, | |
11270 | struct cgraph_simd_clone *from) | |
11271 | { | |
11272 | memcpy (to, from, (sizeof (struct cgraph_simd_clone) | |
46c0fd4b | 11273 | + ((from->nargs - from->inbranch) |
11274 | * sizeof (struct cgraph_simd_clone_arg)))); | |
d09768a4 | 11275 | } |
11276 | ||
11277 | /* Return vector of parameter types of function FNDECL. This uses | |
11278 | TYPE_ARG_TYPES if available, otherwise falls back to types of | |
11279 | DECL_ARGUMENTS types. */ | |
11280 | ||
11281 | vec<tree> | |
11282 | simd_clone_vector_of_formal_parm_types (tree fndecl) | |
11283 | { | |
11284 | if (TYPE_ARG_TYPES (TREE_TYPE (fndecl))) | |
11285 | return ipa_get_vector_of_formal_parm_types (TREE_TYPE (fndecl)); | |
11286 | vec<tree> args = ipa_get_vector_of_formal_parms (fndecl); | |
11287 | unsigned int i; | |
11288 | tree arg; | |
11289 | FOR_EACH_VEC_ELT (args, i, arg) | |
11290 | args[i] = TREE_TYPE (args[i]); | |
11291 | return args; | |
11292 | } | |
11293 | ||
11294 | /* Given a simd function in NODE, extract the simd specific | |
11295 | information from the OMP clauses passed in CLAUSES, and return | |
11296 | the struct cgraph_simd_clone * if it should be cloned. *INBRANCH_SPECIFIED | |
11297 | is set to TRUE if the `inbranch' or `notinbranch' clause specified, | |
11298 | otherwise set to FALSE. */ | |
11299 | ||
11300 | static struct cgraph_simd_clone * | |
11301 | simd_clone_clauses_extract (struct cgraph_node *node, tree clauses, | |
11302 | bool *inbranch_specified) | |
11303 | { | |
11304 | vec<tree> args = simd_clone_vector_of_formal_parm_types (node->decl); | |
11305 | tree t; | |
11306 | int n; | |
11307 | *inbranch_specified = false; | |
11308 | ||
11309 | n = args.length (); | |
11310 | if (n > 0 && args.last () == void_type_node) | |
11311 | n--; | |
11312 | ||
11313 | /* To distinguish from an OpenMP simd clone, Cilk Plus functions to | |
11314 | be cloned have a distinctive artificial label in addition to "omp | |
11315 | declare simd". */ | |
11316 | bool cilk_clone | |
a89e6c15 | 11317 | = (flag_cilkplus |
74acc703 | 11318 | && lookup_attribute ("cilk simd function", |
d09768a4 | 11319 | DECL_ATTRIBUTES (node->decl))); |
11320 | ||
11321 | /* Allocate one more than needed just in case this is an in-branch | |
11322 | clone which will require a mask argument. */ | |
11323 | struct cgraph_simd_clone *clone_info = simd_clone_struct_alloc (n + 1); | |
11324 | clone_info->nargs = n; | |
11325 | clone_info->cilk_elemental = cilk_clone; | |
11326 | ||
11327 | if (!clauses) | |
11328 | { | |
11329 | args.release (); | |
11330 | return clone_info; | |
11331 | } | |
11332 | clauses = TREE_VALUE (clauses); | |
11333 | if (!clauses || TREE_CODE (clauses) != OMP_CLAUSE) | |
11334 | return clone_info; | |
11335 | ||
11336 | for (t = clauses; t; t = OMP_CLAUSE_CHAIN (t)) | |
11337 | { | |
11338 | switch (OMP_CLAUSE_CODE (t)) | |
11339 | { | |
11340 | case OMP_CLAUSE_INBRANCH: | |
11341 | clone_info->inbranch = 1; | |
11342 | *inbranch_specified = true; | |
11343 | break; | |
11344 | case OMP_CLAUSE_NOTINBRANCH: | |
11345 | clone_info->inbranch = 0; | |
11346 | *inbranch_specified = true; | |
11347 | break; | |
11348 | case OMP_CLAUSE_SIMDLEN: | |
11349 | clone_info->simdlen | |
11350 | = TREE_INT_CST_LOW (OMP_CLAUSE_SIMDLEN_EXPR (t)); | |
11351 | break; | |
11352 | case OMP_CLAUSE_LINEAR: | |
11353 | { | |
11354 | tree decl = OMP_CLAUSE_DECL (t); | |
11355 | tree step = OMP_CLAUSE_LINEAR_STEP (t); | |
11356 | int argno = TREE_INT_CST_LOW (decl); | |
11357 | if (OMP_CLAUSE_LINEAR_VARIABLE_STRIDE (t)) | |
11358 | { | |
11359 | clone_info->args[argno].arg_type | |
11360 | = SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP; | |
11361 | clone_info->args[argno].linear_step = tree_to_shwi (step); | |
11362 | gcc_assert (clone_info->args[argno].linear_step >= 0 | |
11363 | && clone_info->args[argno].linear_step < n); | |
11364 | } | |
11365 | else | |
11366 | { | |
11367 | if (POINTER_TYPE_P (args[argno])) | |
11368 | step = fold_convert (ssizetype, step); | |
11369 | if (!tree_fits_shwi_p (step)) | |
11370 | { | |
11371 | warning_at (OMP_CLAUSE_LOCATION (t), 0, | |
11372 | "ignoring large linear step"); | |
11373 | args.release (); | |
11374 | return NULL; | |
11375 | } | |
11376 | else if (integer_zerop (step)) | |
11377 | { | |
11378 | warning_at (OMP_CLAUSE_LOCATION (t), 0, | |
11379 | "ignoring zero linear step"); | |
11380 | args.release (); | |
11381 | return NULL; | |
11382 | } | |
11383 | else | |
11384 | { | |
11385 | clone_info->args[argno].arg_type | |
11386 | = SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP; | |
11387 | clone_info->args[argno].linear_step = tree_to_shwi (step); | |
11388 | } | |
11389 | } | |
11390 | break; | |
11391 | } | |
11392 | case OMP_CLAUSE_UNIFORM: | |
11393 | { | |
11394 | tree decl = OMP_CLAUSE_DECL (t); | |
11395 | int argno = tree_to_uhwi (decl); | |
11396 | clone_info->args[argno].arg_type | |
11397 | = SIMD_CLONE_ARG_TYPE_UNIFORM; | |
11398 | break; | |
11399 | } | |
11400 | case OMP_CLAUSE_ALIGNED: | |
11401 | { | |
11402 | tree decl = OMP_CLAUSE_DECL (t); | |
11403 | int argno = tree_to_uhwi (decl); | |
11404 | clone_info->args[argno].alignment | |
11405 | = TREE_INT_CST_LOW (OMP_CLAUSE_ALIGNED_ALIGNMENT (t)); | |
11406 | break; | |
11407 | } | |
11408 | default: | |
11409 | break; | |
11410 | } | |
11411 | } | |
11412 | args.release (); | |
11413 | return clone_info; | |
11414 | } | |
11415 | ||
11416 | /* Given a SIMD clone in NODE, calculate the characteristic data | |
11417 | type and return the coresponding type. The characteristic data | |
11418 | type is computed as described in the Intel Vector ABI. */ | |
11419 | ||
11420 | static tree | |
11421 | simd_clone_compute_base_data_type (struct cgraph_node *node, | |
11422 | struct cgraph_simd_clone *clone_info) | |
11423 | { | |
11424 | tree type = integer_type_node; | |
11425 | tree fndecl = node->decl; | |
11426 | ||
11427 | /* a) For non-void function, the characteristic data type is the | |
11428 | return type. */ | |
11429 | if (TREE_CODE (TREE_TYPE (TREE_TYPE (fndecl))) != VOID_TYPE) | |
11430 | type = TREE_TYPE (TREE_TYPE (fndecl)); | |
11431 | ||
11432 | /* b) If the function has any non-uniform, non-linear parameters, | |
11433 | then the characteristic data type is the type of the first | |
11434 | such parameter. */ | |
11435 | else | |
11436 | { | |
11437 | vec<tree> map = simd_clone_vector_of_formal_parm_types (fndecl); | |
11438 | for (unsigned int i = 0; i < clone_info->nargs; ++i) | |
11439 | if (clone_info->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
11440 | { | |
11441 | type = map[i]; | |
11442 | break; | |
11443 | } | |
11444 | map.release (); | |
11445 | } | |
11446 | ||
11447 | /* c) If the characteristic data type determined by a) or b) above | |
11448 | is struct, union, or class type which is pass-by-value (except | |
11449 | for the type that maps to the built-in complex data type), the | |
11450 | characteristic data type is int. */ | |
11451 | if (RECORD_OR_UNION_TYPE_P (type) | |
11452 | && !aggregate_value_p (type, NULL) | |
11453 | && TREE_CODE (type) != COMPLEX_TYPE) | |
11454 | return integer_type_node; | |
11455 | ||
11456 | /* d) If none of the above three classes is applicable, the | |
11457 | characteristic data type is int. */ | |
11458 | ||
11459 | return type; | |
11460 | ||
11461 | /* e) For Intel Xeon Phi native and offload compilation, if the | |
11462 | resulting characteristic data type is 8-bit or 16-bit integer | |
11463 | data type, the characteristic data type is int. */ | |
11464 | /* Well, we don't handle Xeon Phi yet. */ | |
11465 | } | |
11466 | ||
11467 | static tree | |
11468 | simd_clone_mangle (struct cgraph_node *node, | |
11469 | struct cgraph_simd_clone *clone_info) | |
11470 | { | |
11471 | char vecsize_mangle = clone_info->vecsize_mangle; | |
11472 | char mask = clone_info->inbranch ? 'M' : 'N'; | |
11473 | unsigned int simdlen = clone_info->simdlen; | |
11474 | unsigned int n; | |
11475 | pretty_printer pp; | |
11476 | ||
11477 | gcc_assert (vecsize_mangle && simdlen); | |
11478 | ||
11479 | pp_string (&pp, "_ZGV"); | |
11480 | pp_character (&pp, vecsize_mangle); | |
11481 | pp_character (&pp, mask); | |
11482 | pp_decimal_int (&pp, simdlen); | |
11483 | ||
11484 | for (n = 0; n < clone_info->nargs; ++n) | |
11485 | { | |
11486 | struct cgraph_simd_clone_arg arg = clone_info->args[n]; | |
11487 | ||
11488 | if (arg.arg_type == SIMD_CLONE_ARG_TYPE_UNIFORM) | |
11489 | pp_character (&pp, 'u'); | |
11490 | else if (arg.arg_type == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
11491 | { | |
11492 | gcc_assert (arg.linear_step != 0); | |
11493 | pp_character (&pp, 'l'); | |
11494 | if (arg.linear_step > 1) | |
11495 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
11496 | else if (arg.linear_step < 0) | |
11497 | { | |
11498 | pp_character (&pp, 'n'); | |
11499 | pp_unsigned_wide_integer (&pp, (-(unsigned HOST_WIDE_INT) | |
11500 | arg.linear_step)); | |
11501 | } | |
11502 | } | |
11503 | else if (arg.arg_type == SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP) | |
11504 | { | |
11505 | pp_character (&pp, 's'); | |
11506 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
11507 | } | |
11508 | else | |
11509 | pp_character (&pp, 'v'); | |
11510 | if (arg.alignment) | |
11511 | { | |
11512 | pp_character (&pp, 'a'); | |
11513 | pp_decimal_int (&pp, arg.alignment); | |
11514 | } | |
11515 | } | |
11516 | ||
11517 | pp_underscore (&pp); | |
11518 | pp_string (&pp, | |
11519 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (node->decl))); | |
11520 | const char *str = pp_formatted_text (&pp); | |
11521 | ||
11522 | /* If there already is a SIMD clone with the same mangled name, don't | |
11523 | add another one. This can happen e.g. for | |
11524 | #pragma omp declare simd | |
11525 | #pragma omp declare simd simdlen(8) | |
11526 | int foo (int, int); | |
11527 | if the simdlen is assumed to be 8 for the first one, etc. */ | |
11528 | for (struct cgraph_node *clone = node->simd_clones; clone; | |
11529 | clone = clone->simdclone->next_clone) | |
11530 | if (strcmp (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (clone->decl)), | |
11531 | str) == 0) | |
11532 | return NULL_TREE; | |
11533 | ||
11534 | return get_identifier (str); | |
11535 | } | |
11536 | ||
11537 | /* Create a simd clone of OLD_NODE and return it. */ | |
11538 | ||
11539 | static struct cgraph_node * | |
11540 | simd_clone_create (struct cgraph_node *old_node) | |
11541 | { | |
11542 | struct cgraph_node *new_node; | |
11543 | if (old_node->definition) | |
ea0695f9 | 11544 | { |
415d1b9a | 11545 | if (!old_node->has_gimple_body_p ()) |
ea0695f9 | 11546 | return NULL; |
415d1b9a | 11547 | old_node->get_body (); |
11548 | new_node = old_node->create_version_clone_with_body (vNULL, NULL, NULL, | |
11549 | false, NULL, NULL, | |
11550 | "simdclone"); | |
ea0695f9 | 11551 | } |
d09768a4 | 11552 | else |
11553 | { | |
11554 | tree old_decl = old_node->decl; | |
11555 | tree new_decl = copy_node (old_node->decl); | |
11556 | DECL_NAME (new_decl) = clone_function_name (old_decl, "simdclone"); | |
11557 | SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl)); | |
11558 | SET_DECL_RTL (new_decl, NULL); | |
11559 | DECL_STATIC_CONSTRUCTOR (new_decl) = 0; | |
11560 | DECL_STATIC_DESTRUCTOR (new_decl) = 0; | |
415d1b9a | 11561 | new_node = old_node->create_version_clone (new_decl, vNULL, NULL); |
35ee1c66 | 11562 | symtab->call_cgraph_insertion_hooks (new_node); |
d09768a4 | 11563 | } |
11564 | if (new_node == NULL) | |
11565 | return new_node; | |
11566 | ||
11567 | TREE_PUBLIC (new_node->decl) = TREE_PUBLIC (old_node->decl); | |
11568 | ||
11569 | /* The function cgraph_function_versioning () will force the new | |
11570 | symbol local. Undo this, and inherit external visability from | |
11571 | the old node. */ | |
11572 | new_node->local.local = old_node->local.local; | |
11573 | new_node->externally_visible = old_node->externally_visible; | |
11574 | ||
11575 | return new_node; | |
11576 | } | |
11577 | ||
11578 | /* Adjust the return type of the given function to its appropriate | |
11579 | vector counterpart. Returns a simd array to be used throughout the | |
11580 | function as a return value. */ | |
11581 | ||
11582 | static tree | |
11583 | simd_clone_adjust_return_type (struct cgraph_node *node) | |
11584 | { | |
11585 | tree fndecl = node->decl; | |
11586 | tree orig_rettype = TREE_TYPE (TREE_TYPE (fndecl)); | |
11587 | unsigned int veclen; | |
11588 | tree t; | |
11589 | ||
11590 | /* Adjust the function return type. */ | |
11591 | if (orig_rettype == void_type_node) | |
11592 | return NULL_TREE; | |
11593 | TREE_TYPE (fndecl) = build_distinct_type_copy (TREE_TYPE (fndecl)); | |
11594 | if (INTEGRAL_TYPE_P (TREE_TYPE (TREE_TYPE (fndecl))) | |
11595 | || POINTER_TYPE_P (TREE_TYPE (TREE_TYPE (fndecl)))) | |
11596 | veclen = node->simdclone->vecsize_int; | |
11597 | else | |
11598 | veclen = node->simdclone->vecsize_float; | |
11599 | veclen /= GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (TREE_TYPE (fndecl)))); | |
11600 | if (veclen > node->simdclone->simdlen) | |
11601 | veclen = node->simdclone->simdlen; | |
11602 | if (veclen == node->simdclone->simdlen) | |
11603 | TREE_TYPE (TREE_TYPE (fndecl)) | |
11604 | = build_vector_type (TREE_TYPE (TREE_TYPE (fndecl)), | |
11605 | node->simdclone->simdlen); | |
11606 | else | |
11607 | { | |
11608 | t = build_vector_type (TREE_TYPE (TREE_TYPE (fndecl)), veclen); | |
11609 | t = build_array_type_nelts (t, node->simdclone->simdlen / veclen); | |
11610 | TREE_TYPE (TREE_TYPE (fndecl)) = t; | |
11611 | } | |
11612 | if (!node->definition) | |
11613 | return NULL_TREE; | |
11614 | ||
11615 | t = DECL_RESULT (fndecl); | |
11616 | /* Adjust the DECL_RESULT. */ | |
11617 | gcc_assert (TREE_TYPE (t) != void_type_node); | |
11618 | TREE_TYPE (t) = TREE_TYPE (TREE_TYPE (fndecl)); | |
11619 | relayout_decl (t); | |
11620 | ||
11621 | tree atype = build_array_type_nelts (orig_rettype, | |
11622 | node->simdclone->simdlen); | |
11623 | if (veclen != node->simdclone->simdlen) | |
11624 | return build1 (VIEW_CONVERT_EXPR, atype, t); | |
11625 | ||
11626 | /* Set up a SIMD array to use as the return value. */ | |
11627 | tree retval = create_tmp_var_raw (atype, "retval"); | |
11628 | gimple_add_tmp_var (retval); | |
11629 | return retval; | |
11630 | } | |
11631 | ||
11632 | /* Each vector argument has a corresponding array to be used locally | |
11633 | as part of the eventual loop. Create such temporary array and | |
11634 | return it. | |
11635 | ||
11636 | PREFIX is the prefix to be used for the temporary. | |
11637 | ||
11638 | TYPE is the inner element type. | |
11639 | ||
11640 | SIMDLEN is the number of elements. */ | |
11641 | ||
11642 | static tree | |
11643 | create_tmp_simd_array (const char *prefix, tree type, int simdlen) | |
11644 | { | |
11645 | tree atype = build_array_type_nelts (type, simdlen); | |
11646 | tree avar = create_tmp_var_raw (atype, prefix); | |
11647 | gimple_add_tmp_var (avar); | |
11648 | return avar; | |
11649 | } | |
11650 | ||
11651 | /* Modify the function argument types to their corresponding vector | |
11652 | counterparts if appropriate. Also, create one array for each simd | |
11653 | argument to be used locally when using the function arguments as | |
11654 | part of the loop. | |
11655 | ||
11656 | NODE is the function whose arguments are to be adjusted. | |
11657 | ||
11658 | Returns an adjustment vector that will be filled describing how the | |
11659 | argument types will be adjusted. */ | |
11660 | ||
11661 | static ipa_parm_adjustment_vec | |
11662 | simd_clone_adjust_argument_types (struct cgraph_node *node) | |
11663 | { | |
11664 | vec<tree> args; | |
11665 | ipa_parm_adjustment_vec adjustments; | |
11666 | ||
11667 | if (node->definition) | |
11668 | args = ipa_get_vector_of_formal_parms (node->decl); | |
11669 | else | |
11670 | args = simd_clone_vector_of_formal_parm_types (node->decl); | |
11671 | adjustments.create (args.length ()); | |
11672 | unsigned i, j, veclen; | |
11673 | struct ipa_parm_adjustment adj; | |
11674 | for (i = 0; i < node->simdclone->nargs; ++i) | |
11675 | { | |
11676 | memset (&adj, 0, sizeof (adj)); | |
11677 | tree parm = args[i]; | |
11678 | tree parm_type = node->definition ? TREE_TYPE (parm) : parm; | |
11679 | adj.base_index = i; | |
11680 | adj.base = parm; | |
11681 | ||
11682 | node->simdclone->args[i].orig_arg = node->definition ? parm : NULL_TREE; | |
11683 | node->simdclone->args[i].orig_type = parm_type; | |
11684 | ||
11685 | if (node->simdclone->args[i].arg_type != SIMD_CLONE_ARG_TYPE_VECTOR) | |
11686 | { | |
11687 | /* No adjustment necessary for scalar arguments. */ | |
11688 | adj.op = IPA_PARM_OP_COPY; | |
11689 | } | |
11690 | else | |
11691 | { | |
11692 | if (INTEGRAL_TYPE_P (parm_type) || POINTER_TYPE_P (parm_type)) | |
11693 | veclen = node->simdclone->vecsize_int; | |
11694 | else | |
11695 | veclen = node->simdclone->vecsize_float; | |
11696 | veclen /= GET_MODE_BITSIZE (TYPE_MODE (parm_type)); | |
11697 | if (veclen > node->simdclone->simdlen) | |
11698 | veclen = node->simdclone->simdlen; | |
11699 | adj.arg_prefix = "simd"; | |
11700 | adj.type = build_vector_type (parm_type, veclen); | |
11701 | node->simdclone->args[i].vector_type = adj.type; | |
11702 | for (j = veclen; j < node->simdclone->simdlen; j += veclen) | |
11703 | { | |
11704 | adjustments.safe_push (adj); | |
11705 | if (j == veclen) | |
11706 | { | |
11707 | memset (&adj, 0, sizeof (adj)); | |
11708 | adj.op = IPA_PARM_OP_NEW; | |
11709 | adj.arg_prefix = "simd"; | |
11710 | adj.base_index = i; | |
11711 | adj.type = node->simdclone->args[i].vector_type; | |
11712 | } | |
11713 | } | |
11714 | ||
11715 | if (node->definition) | |
11716 | node->simdclone->args[i].simd_array | |
11717 | = create_tmp_simd_array (IDENTIFIER_POINTER (DECL_NAME (parm)), | |
11718 | parm_type, node->simdclone->simdlen); | |
11719 | } | |
11720 | adjustments.safe_push (adj); | |
11721 | } | |
11722 | ||
11723 | if (node->simdclone->inbranch) | |
11724 | { | |
11725 | tree base_type | |
11726 | = simd_clone_compute_base_data_type (node->simdclone->origin, | |
11727 | node->simdclone); | |
11728 | ||
11729 | memset (&adj, 0, sizeof (adj)); | |
11730 | adj.op = IPA_PARM_OP_NEW; | |
11731 | adj.arg_prefix = "mask"; | |
11732 | ||
11733 | adj.base_index = i; | |
11734 | if (INTEGRAL_TYPE_P (base_type) || POINTER_TYPE_P (base_type)) | |
11735 | veclen = node->simdclone->vecsize_int; | |
11736 | else | |
11737 | veclen = node->simdclone->vecsize_float; | |
11738 | veclen /= GET_MODE_BITSIZE (TYPE_MODE (base_type)); | |
11739 | if (veclen > node->simdclone->simdlen) | |
11740 | veclen = node->simdclone->simdlen; | |
11741 | adj.type = build_vector_type (base_type, veclen); | |
11742 | adjustments.safe_push (adj); | |
11743 | ||
11744 | for (j = veclen; j < node->simdclone->simdlen; j += veclen) | |
11745 | adjustments.safe_push (adj); | |
11746 | ||
11747 | /* We have previously allocated one extra entry for the mask. Use | |
11748 | it and fill it. */ | |
11749 | struct cgraph_simd_clone *sc = node->simdclone; | |
11750 | sc->nargs++; | |
11751 | if (node->definition) | |
11752 | { | |
11753 | sc->args[i].orig_arg | |
11754 | = build_decl (UNKNOWN_LOCATION, PARM_DECL, NULL, base_type); | |
11755 | sc->args[i].simd_array | |
11756 | = create_tmp_simd_array ("mask", base_type, sc->simdlen); | |
11757 | } | |
11758 | sc->args[i].orig_type = base_type; | |
11759 | sc->args[i].arg_type = SIMD_CLONE_ARG_TYPE_MASK; | |
11760 | } | |
11761 | ||
11762 | if (node->definition) | |
11763 | ipa_modify_formal_parameters (node->decl, adjustments); | |
11764 | else | |
11765 | { | |
11766 | tree new_arg_types = NULL_TREE, new_reversed; | |
11767 | bool last_parm_void = false; | |
11768 | if (args.length () > 0 && args.last () == void_type_node) | |
11769 | last_parm_void = true; | |
11770 | ||
11771 | gcc_assert (TYPE_ARG_TYPES (TREE_TYPE (node->decl))); | |
11772 | j = adjustments.length (); | |
11773 | for (i = 0; i < j; i++) | |
11774 | { | |
11775 | struct ipa_parm_adjustment *adj = &adjustments[i]; | |
11776 | tree ptype; | |
11777 | if (adj->op == IPA_PARM_OP_COPY) | |
11778 | ptype = args[adj->base_index]; | |
11779 | else | |
11780 | ptype = adj->type; | |
11781 | new_arg_types = tree_cons (NULL_TREE, ptype, new_arg_types); | |
11782 | } | |
11783 | new_reversed = nreverse (new_arg_types); | |
11784 | if (last_parm_void) | |
11785 | { | |
11786 | if (new_reversed) | |
11787 | TREE_CHAIN (new_arg_types) = void_list_node; | |
11788 | else | |
11789 | new_reversed = void_list_node; | |
11790 | } | |
11791 | ||
11792 | tree new_type = build_distinct_type_copy (TREE_TYPE (node->decl)); | |
11793 | TYPE_ARG_TYPES (new_type) = new_reversed; | |
11794 | TREE_TYPE (node->decl) = new_type; | |
11795 | ||
11796 | adjustments.release (); | |
11797 | } | |
11798 | args.release (); | |
11799 | return adjustments; | |
11800 | } | |
11801 | ||
11802 | /* Initialize and copy the function arguments in NODE to their | |
11803 | corresponding local simd arrays. Returns a fresh gimple_seq with | |
11804 | the instruction sequence generated. */ | |
11805 | ||
11806 | static gimple_seq | |
11807 | simd_clone_init_simd_arrays (struct cgraph_node *node, | |
11808 | ipa_parm_adjustment_vec adjustments) | |
11809 | { | |
11810 | gimple_seq seq = NULL; | |
11811 | unsigned i = 0, j = 0, k; | |
11812 | ||
11813 | for (tree arg = DECL_ARGUMENTS (node->decl); | |
11814 | arg; | |
11815 | arg = DECL_CHAIN (arg), i++, j++) | |
11816 | { | |
11817 | if (adjustments[j].op == IPA_PARM_OP_COPY) | |
11818 | continue; | |
11819 | ||
11820 | node->simdclone->args[i].vector_arg = arg; | |
11821 | ||
11822 | tree array = node->simdclone->args[i].simd_array; | |
11823 | if (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg)) == node->simdclone->simdlen) | |
11824 | { | |
11825 | tree ptype = build_pointer_type (TREE_TYPE (TREE_TYPE (array))); | |
11826 | tree ptr = build_fold_addr_expr (array); | |
11827 | tree t = build2 (MEM_REF, TREE_TYPE (arg), ptr, | |
11828 | build_int_cst (ptype, 0)); | |
11829 | t = build2 (MODIFY_EXPR, TREE_TYPE (t), t, arg); | |
11830 | gimplify_and_add (t, &seq); | |
11831 | } | |
11832 | else | |
11833 | { | |
11834 | unsigned int simdlen = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg)); | |
11835 | tree ptype = build_pointer_type (TREE_TYPE (TREE_TYPE (array))); | |
11836 | for (k = 0; k < node->simdclone->simdlen; k += simdlen) | |
11837 | { | |
11838 | tree ptr = build_fold_addr_expr (array); | |
11839 | int elemsize; | |
11840 | if (k) | |
11841 | { | |
11842 | arg = DECL_CHAIN (arg); | |
11843 | j++; | |
11844 | } | |
11845 | elemsize | |
11846 | = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (TREE_TYPE (arg)))); | |
11847 | tree t = build2 (MEM_REF, TREE_TYPE (arg), ptr, | |
11848 | build_int_cst (ptype, k * elemsize)); | |
11849 | t = build2 (MODIFY_EXPR, TREE_TYPE (t), t, arg); | |
11850 | gimplify_and_add (t, &seq); | |
11851 | } | |
11852 | } | |
11853 | } | |
11854 | return seq; | |
11855 | } | |
11856 | ||
11857 | /* Callback info for ipa_simd_modify_stmt_ops below. */ | |
11858 | ||
11859 | struct modify_stmt_info { | |
11860 | ipa_parm_adjustment_vec adjustments; | |
11861 | gimple stmt; | |
11862 | /* True if the parent statement was modified by | |
11863 | ipa_simd_modify_stmt_ops. */ | |
11864 | bool modified; | |
11865 | }; | |
11866 | ||
11867 | /* Callback for walk_gimple_op. | |
11868 | ||
11869 | Adjust operands from a given statement as specified in the | |
11870 | adjustments vector in the callback data. */ | |
11871 | ||
11872 | static tree | |
11873 | ipa_simd_modify_stmt_ops (tree *tp, int *walk_subtrees, void *data) | |
11874 | { | |
11875 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; | |
2d78e89f | 11876 | struct modify_stmt_info *info = (struct modify_stmt_info *) wi->info; |
11877 | tree *orig_tp = tp; | |
11878 | if (TREE_CODE (*tp) == ADDR_EXPR) | |
11879 | tp = &TREE_OPERAND (*tp, 0); | |
11880 | struct ipa_parm_adjustment *cand = NULL; | |
11881 | if (TREE_CODE (*tp) == PARM_DECL) | |
11882 | cand = ipa_get_adjustment_candidate (&tp, NULL, info->adjustments, true); | |
11883 | else | |
d09768a4 | 11884 | { |
d09768a4 | 11885 | if (TYPE_P (*tp)) |
11886 | *walk_subtrees = 0; | |
d09768a4 | 11887 | } |
d09768a4 | 11888 | |
2d78e89f | 11889 | tree repl = NULL_TREE; |
11890 | if (cand) | |
11891 | repl = unshare_expr (cand->new_decl); | |
11892 | else | |
d09768a4 | 11893 | { |
2d78e89f | 11894 | if (tp != orig_tp) |
11895 | { | |
11896 | *walk_subtrees = 0; | |
11897 | bool modified = info->modified; | |
11898 | info->modified = false; | |
11899 | walk_tree (tp, ipa_simd_modify_stmt_ops, wi, wi->pset); | |
11900 | if (!info->modified) | |
11901 | { | |
11902 | info->modified = modified; | |
11903 | return NULL_TREE; | |
11904 | } | |
11905 | info->modified = modified; | |
11906 | repl = *tp; | |
11907 | } | |
11908 | else | |
11909 | return NULL_TREE; | |
d09768a4 | 11910 | } |
2d78e89f | 11911 | |
11912 | if (tp != orig_tp) | |
d09768a4 | 11913 | { |
2d78e89f | 11914 | repl = build_fold_addr_expr (repl); |
9c1dd4d9 | 11915 | gimple stmt; |
11916 | if (is_gimple_debug (info->stmt)) | |
11917 | { | |
11918 | tree vexpr = make_node (DEBUG_EXPR_DECL); | |
11919 | stmt = gimple_build_debug_source_bind (vexpr, repl, NULL); | |
11920 | DECL_ARTIFICIAL (vexpr) = 1; | |
11921 | TREE_TYPE (vexpr) = TREE_TYPE (repl); | |
11922 | DECL_MODE (vexpr) = TYPE_MODE (TREE_TYPE (repl)); | |
11923 | repl = vexpr; | |
11924 | } | |
11925 | else | |
11926 | { | |
f9e245b2 | 11927 | stmt = gimple_build_assign (make_ssa_name (TREE_TYPE (repl)), repl); |
9c1dd4d9 | 11928 | repl = gimple_assign_lhs (stmt); |
11929 | } | |
2d78e89f | 11930 | gimple_stmt_iterator gsi = gsi_for_stmt (info->stmt); |
d09768a4 | 11931 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
2d78e89f | 11932 | *orig_tp = repl; |
d09768a4 | 11933 | } |
2d78e89f | 11934 | else if (!useless_type_conversion_p (TREE_TYPE (*tp), TREE_TYPE (repl))) |
d09768a4 | 11935 | { |
11936 | tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*tp), repl); | |
11937 | *tp = vce; | |
11938 | } | |
11939 | else | |
11940 | *tp = repl; | |
11941 | ||
11942 | info->modified = true; | |
d09768a4 | 11943 | return NULL_TREE; |
11944 | } | |
11945 | ||
11946 | /* Traverse the function body and perform all modifications as | |
11947 | described in ADJUSTMENTS. At function return, ADJUSTMENTS will be | |
11948 | modified such that the replacement/reduction value will now be an | |
11949 | offset into the corresponding simd_array. | |
11950 | ||
11951 | This function will replace all function argument uses with their | |
11952 | corresponding simd array elements, and ajust the return values | |
11953 | accordingly. */ | |
11954 | ||
11955 | static void | |
11956 | ipa_simd_modify_function_body (struct cgraph_node *node, | |
11957 | ipa_parm_adjustment_vec adjustments, | |
11958 | tree retval_array, tree iter) | |
11959 | { | |
11960 | basic_block bb; | |
2d78e89f | 11961 | unsigned int i, j, l; |
d09768a4 | 11962 | |
11963 | /* Re-use the adjustments array, but this time use it to replace | |
11964 | every function argument use to an offset into the corresponding | |
11965 | simd_array. */ | |
11966 | for (i = 0, j = 0; i < node->simdclone->nargs; ++i, ++j) | |
11967 | { | |
11968 | if (!node->simdclone->args[i].vector_arg) | |
11969 | continue; | |
11970 | ||
11971 | tree basetype = TREE_TYPE (node->simdclone->args[i].orig_arg); | |
11972 | tree vectype = TREE_TYPE (node->simdclone->args[i].vector_arg); | |
11973 | adjustments[j].new_decl | |
11974 | = build4 (ARRAY_REF, | |
11975 | basetype, | |
11976 | node->simdclone->args[i].simd_array, | |
11977 | iter, | |
11978 | NULL_TREE, NULL_TREE); | |
11979 | if (adjustments[j].op == IPA_PARM_OP_NONE | |
11980 | && TYPE_VECTOR_SUBPARTS (vectype) < node->simdclone->simdlen) | |
11981 | j += node->simdclone->simdlen / TYPE_VECTOR_SUBPARTS (vectype) - 1; | |
11982 | } | |
11983 | ||
2d78e89f | 11984 | l = adjustments.length (); |
11985 | for (i = 1; i < num_ssa_names; i++) | |
11986 | { | |
11987 | tree name = ssa_name (i); | |
11988 | if (name | |
11989 | && SSA_NAME_VAR (name) | |
11990 | && TREE_CODE (SSA_NAME_VAR (name)) == PARM_DECL) | |
11991 | { | |
11992 | for (j = 0; j < l; j++) | |
11993 | if (SSA_NAME_VAR (name) == adjustments[j].base | |
11994 | && adjustments[j].new_decl) | |
11995 | { | |
11996 | tree base_var; | |
11997 | if (adjustments[j].new_ssa_base == NULL_TREE) | |
11998 | { | |
11999 | base_var | |
12000 | = copy_var_decl (adjustments[j].base, | |
12001 | DECL_NAME (adjustments[j].base), | |
12002 | TREE_TYPE (adjustments[j].base)); | |
12003 | adjustments[j].new_ssa_base = base_var; | |
12004 | } | |
12005 | else | |
12006 | base_var = adjustments[j].new_ssa_base; | |
12007 | if (SSA_NAME_IS_DEFAULT_DEF (name)) | |
12008 | { | |
12009 | bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
12010 | gimple_stmt_iterator gsi = gsi_after_labels (bb); | |
12011 | tree new_decl = unshare_expr (adjustments[j].new_decl); | |
12012 | set_ssa_default_def (cfun, adjustments[j].base, NULL_TREE); | |
12013 | SET_SSA_NAME_VAR_OR_IDENTIFIER (name, base_var); | |
12014 | SSA_NAME_IS_DEFAULT_DEF (name) = 0; | |
12015 | gimple stmt = gimple_build_assign (name, new_decl); | |
12016 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
12017 | } | |
12018 | else | |
12019 | SET_SSA_NAME_VAR_OR_IDENTIFIER (name, base_var); | |
12020 | } | |
12021 | } | |
12022 | } | |
12023 | ||
d09768a4 | 12024 | struct modify_stmt_info info; |
12025 | info.adjustments = adjustments; | |
12026 | ||
12027 | FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl)) | |
12028 | { | |
12029 | gimple_stmt_iterator gsi; | |
12030 | ||
12031 | gsi = gsi_start_bb (bb); | |
12032 | while (!gsi_end_p (gsi)) | |
12033 | { | |
12034 | gimple stmt = gsi_stmt (gsi); | |
12035 | info.stmt = stmt; | |
12036 | struct walk_stmt_info wi; | |
12037 | ||
12038 | memset (&wi, 0, sizeof (wi)); | |
12039 | info.modified = false; | |
12040 | wi.info = &info; | |
12041 | walk_gimple_op (stmt, ipa_simd_modify_stmt_ops, &wi); | |
12042 | ||
1a91d914 | 12043 | if (greturn *return_stmt = dyn_cast <greturn *> (stmt)) |
d09768a4 | 12044 | { |
1a91d914 | 12045 | tree retval = gimple_return_retval (return_stmt); |
d09768a4 | 12046 | if (!retval) |
12047 | { | |
12048 | gsi_remove (&gsi, true); | |
12049 | continue; | |
12050 | } | |
12051 | ||
12052 | /* Replace `return foo' with `retval_array[iter] = foo'. */ | |
12053 | tree ref = build4 (ARRAY_REF, TREE_TYPE (retval), | |
12054 | retval_array, iter, NULL, NULL); | |
12055 | stmt = gimple_build_assign (ref, retval); | |
12056 | gsi_replace (&gsi, stmt, true); | |
12057 | info.modified = true; | |
12058 | } | |
12059 | ||
12060 | if (info.modified) | |
12061 | { | |
12062 | update_stmt (stmt); | |
12063 | if (maybe_clean_eh_stmt (stmt)) | |
12064 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
12065 | } | |
12066 | gsi_next (&gsi); | |
12067 | } | |
12068 | } | |
12069 | } | |
12070 | ||
12071 | /* Adjust the argument types in NODE to their appropriate vector | |
12072 | counterparts. */ | |
12073 | ||
12074 | static void | |
12075 | simd_clone_adjust (struct cgraph_node *node) | |
12076 | { | |
12077 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); | |
12078 | ||
12079 | targetm.simd_clone.adjust (node); | |
12080 | ||
12081 | tree retval = simd_clone_adjust_return_type (node); | |
12082 | ipa_parm_adjustment_vec adjustments | |
12083 | = simd_clone_adjust_argument_types (node); | |
12084 | ||
12085 | push_gimplify_context (); | |
12086 | ||
12087 | gimple_seq seq = simd_clone_init_simd_arrays (node, adjustments); | |
12088 | ||
12089 | /* Adjust all uses of vector arguments accordingly. Adjust all | |
12090 | return values accordingly. */ | |
12091 | tree iter = create_tmp_var (unsigned_type_node, "iter"); | |
f9e245b2 | 12092 | tree iter1 = make_ssa_name (iter); |
12093 | tree iter2 = make_ssa_name (iter); | |
d09768a4 | 12094 | ipa_simd_modify_function_body (node, adjustments, retval, iter1); |
12095 | ||
12096 | /* Initialize the iteration variable. */ | |
12097 | basic_block entry_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
12098 | basic_block body_bb = split_block_after_labels (entry_bb)->dest; | |
12099 | gimple_stmt_iterator gsi = gsi_after_labels (entry_bb); | |
12100 | /* Insert the SIMD array and iv initialization at function | |
12101 | entry. */ | |
12102 | gsi_insert_seq_before (&gsi, seq, GSI_NEW_STMT); | |
12103 | ||
12104 | pop_gimplify_context (NULL); | |
12105 | ||
12106 | /* Create a new BB right before the original exit BB, to hold the | |
12107 | iteration increment and the condition/branch. */ | |
12108 | basic_block orig_exit = EDGE_PRED (EXIT_BLOCK_PTR_FOR_FN (cfun), 0)->src; | |
12109 | basic_block incr_bb = create_empty_bb (orig_exit); | |
343ee723 | 12110 | add_bb_to_loop (incr_bb, body_bb->loop_father); |
d09768a4 | 12111 | /* The succ of orig_exit was EXIT_BLOCK_PTR_FOR_FN (cfun), with an empty |
12112 | flag. Set it now to be a FALLTHRU_EDGE. */ | |
12113 | gcc_assert (EDGE_COUNT (orig_exit->succs) == 1); | |
12114 | EDGE_SUCC (orig_exit, 0)->flags |= EDGE_FALLTHRU; | |
12115 | for (unsigned i = 0; | |
12116 | i < EDGE_COUNT (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds); ++i) | |
12117 | { | |
12118 | edge e = EDGE_PRED (EXIT_BLOCK_PTR_FOR_FN (cfun), i); | |
12119 | redirect_edge_succ (e, incr_bb); | |
12120 | } | |
12121 | edge e = make_edge (incr_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0); | |
12122 | e->probability = REG_BR_PROB_BASE; | |
12123 | gsi = gsi_last_bb (incr_bb); | |
12124 | gimple g = gimple_build_assign_with_ops (PLUS_EXPR, iter2, iter1, | |
12125 | build_int_cst (unsigned_type_node, | |
12126 | 1)); | |
12127 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
12128 | ||
12129 | /* Mostly annotate the loop for the vectorizer (the rest is done below). */ | |
12130 | struct loop *loop = alloc_loop (); | |
4c73695b | 12131 | cfun->has_force_vectorize_loops = true; |
d09768a4 | 12132 | loop->safelen = node->simdclone->simdlen; |
4c73695b | 12133 | loop->force_vectorize = true; |
d09768a4 | 12134 | loop->header = body_bb; |
d09768a4 | 12135 | |
12136 | /* Branch around the body if the mask applies. */ | |
12137 | if (node->simdclone->inbranch) | |
12138 | { | |
12139 | gimple_stmt_iterator gsi = gsi_last_bb (loop->header); | |
12140 | tree mask_array | |
12141 | = node->simdclone->args[node->simdclone->nargs - 1].simd_array; | |
f9e245b2 | 12142 | tree mask = make_ssa_name (TREE_TYPE (TREE_TYPE (mask_array))); |
d09768a4 | 12143 | tree aref = build4 (ARRAY_REF, |
12144 | TREE_TYPE (TREE_TYPE (mask_array)), | |
12145 | mask_array, iter1, | |
12146 | NULL, NULL); | |
12147 | g = gimple_build_assign (mask, aref); | |
12148 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
12149 | int bitsize = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (aref))); | |
12150 | if (!INTEGRAL_TYPE_P (TREE_TYPE (aref))) | |
12151 | { | |
12152 | aref = build1 (VIEW_CONVERT_EXPR, | |
12153 | build_nonstandard_integer_type (bitsize, 0), mask); | |
f9e245b2 | 12154 | mask = make_ssa_name (TREE_TYPE (aref)); |
d09768a4 | 12155 | g = gimple_build_assign (mask, aref); |
12156 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
12157 | } | |
12158 | ||
12159 | g = gimple_build_cond (EQ_EXPR, mask, build_zero_cst (TREE_TYPE (mask)), | |
12160 | NULL, NULL); | |
12161 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
12162 | make_edge (loop->header, incr_bb, EDGE_TRUE_VALUE); | |
12163 | FALLTHRU_EDGE (loop->header)->flags = EDGE_FALSE_VALUE; | |
12164 | } | |
12165 | ||
12166 | /* Generate the condition. */ | |
12167 | g = gimple_build_cond (LT_EXPR, | |
12168 | iter2, | |
12169 | build_int_cst (unsigned_type_node, | |
12170 | node->simdclone->simdlen), | |
12171 | NULL, NULL); | |
12172 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
12173 | e = split_block (incr_bb, gsi_stmt (gsi)); | |
12174 | basic_block latch_bb = e->dest; | |
343ee723 | 12175 | basic_block new_exit_bb; |
d09768a4 | 12176 | new_exit_bb = split_block (latch_bb, NULL)->dest; |
12177 | loop->latch = latch_bb; | |
12178 | ||
12179 | redirect_edge_succ (FALLTHRU_EDGE (latch_bb), body_bb); | |
12180 | ||
12181 | make_edge (incr_bb, new_exit_bb, EDGE_FALSE_VALUE); | |
12182 | /* The successor of incr_bb is already pointing to latch_bb; just | |
12183 | change the flags. | |
12184 | make_edge (incr_bb, latch_bb, EDGE_TRUE_VALUE); */ | |
12185 | FALLTHRU_EDGE (incr_bb)->flags = EDGE_TRUE_VALUE; | |
12186 | ||
1a91d914 | 12187 | gphi *phi = create_phi_node (iter1, body_bb); |
d09768a4 | 12188 | edge preheader_edge = find_edge (entry_bb, body_bb); |
12189 | edge latch_edge = single_succ_edge (latch_bb); | |
12190 | add_phi_arg (phi, build_zero_cst (unsigned_type_node), preheader_edge, | |
12191 | UNKNOWN_LOCATION); | |
12192 | add_phi_arg (phi, iter2, latch_edge, UNKNOWN_LOCATION); | |
12193 | ||
12194 | /* Generate the new return. */ | |
12195 | gsi = gsi_last_bb (new_exit_bb); | |
12196 | if (retval | |
12197 | && TREE_CODE (retval) == VIEW_CONVERT_EXPR | |
12198 | && TREE_CODE (TREE_OPERAND (retval, 0)) == RESULT_DECL) | |
12199 | retval = TREE_OPERAND (retval, 0); | |
12200 | else if (retval) | |
12201 | { | |
12202 | retval = build1 (VIEW_CONVERT_EXPR, | |
12203 | TREE_TYPE (TREE_TYPE (node->decl)), | |
12204 | retval); | |
12205 | retval = force_gimple_operand_gsi (&gsi, retval, true, NULL, | |
12206 | false, GSI_CONTINUE_LINKING); | |
12207 | } | |
12208 | g = gimple_build_return (retval); | |
12209 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
12210 | ||
12211 | /* Handle aligned clauses by replacing default defs of the aligned | |
12212 | uniform args with __builtin_assume_aligned (arg_N(D), alignment) | |
12213 | lhs. Handle linear by adding PHIs. */ | |
12214 | for (unsigned i = 0; i < node->simdclone->nargs; i++) | |
12215 | if (node->simdclone->args[i].alignment | |
12216 | && node->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_UNIFORM | |
12217 | && (node->simdclone->args[i].alignment | |
12218 | & (node->simdclone->args[i].alignment - 1)) == 0 | |
12219 | && TREE_CODE (TREE_TYPE (node->simdclone->args[i].orig_arg)) | |
12220 | == POINTER_TYPE) | |
12221 | { | |
12222 | unsigned int alignment = node->simdclone->args[i].alignment; | |
12223 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
12224 | tree def = ssa_default_def (cfun, orig_arg); | |
6db33a3d | 12225 | if (def && !has_zero_uses (def)) |
d09768a4 | 12226 | { |
12227 | tree fn = builtin_decl_explicit (BUILT_IN_ASSUME_ALIGNED); | |
12228 | gimple_seq seq = NULL; | |
12229 | bool need_cvt = false; | |
1a91d914 | 12230 | gcall *call |
d09768a4 | 12231 | = gimple_build_call (fn, 2, def, size_int (alignment)); |
12232 | g = call; | |
12233 | if (!useless_type_conversion_p (TREE_TYPE (orig_arg), | |
12234 | ptr_type_node)) | |
12235 | need_cvt = true; | |
f9e245b2 | 12236 | tree t = make_ssa_name (need_cvt ? ptr_type_node : orig_arg); |
d09768a4 | 12237 | gimple_call_set_lhs (g, t); |
12238 | gimple_seq_add_stmt_without_update (&seq, g); | |
12239 | if (need_cvt) | |
12240 | { | |
f9e245b2 | 12241 | t = make_ssa_name (orig_arg); |
d09768a4 | 12242 | g = gimple_build_assign_with_ops (NOP_EXPR, t, |
806413d2 | 12243 | gimple_call_lhs (g)); |
d09768a4 | 12244 | gimple_seq_add_stmt_without_update (&seq, g); |
12245 | } | |
12246 | gsi_insert_seq_on_edge_immediate | |
12247 | (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)), seq); | |
12248 | ||
12249 | entry_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
12250 | int freq = compute_call_stmt_bb_frequency (current_function_decl, | |
12251 | entry_bb); | |
415d1b9a | 12252 | node->create_edge (cgraph_node::get_create (fn), |
12253 | call, entry_bb->count, freq); | |
d09768a4 | 12254 | |
12255 | imm_use_iterator iter; | |
12256 | use_operand_p use_p; | |
12257 | gimple use_stmt; | |
12258 | tree repl = gimple_get_lhs (g); | |
12259 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, def) | |
12260 | if (is_gimple_debug (use_stmt) || use_stmt == call) | |
12261 | continue; | |
12262 | else | |
12263 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
12264 | SET_USE (use_p, repl); | |
12265 | } | |
12266 | } | |
12267 | else if (node->simdclone->args[i].arg_type | |
12268 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
12269 | { | |
12270 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
12271 | tree def = ssa_default_def (cfun, orig_arg); | |
12272 | gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
12273 | || POINTER_TYPE_P (TREE_TYPE (orig_arg))); | |
6db33a3d | 12274 | if (def && !has_zero_uses (def)) |
d09768a4 | 12275 | { |
f9e245b2 | 12276 | iter1 = make_ssa_name (orig_arg); |
12277 | iter2 = make_ssa_name (orig_arg); | |
d09768a4 | 12278 | phi = create_phi_node (iter1, body_bb); |
12279 | add_phi_arg (phi, def, preheader_edge, UNKNOWN_LOCATION); | |
12280 | add_phi_arg (phi, iter2, latch_edge, UNKNOWN_LOCATION); | |
12281 | enum tree_code code = INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
12282 | ? PLUS_EXPR : POINTER_PLUS_EXPR; | |
12283 | tree addtype = INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
12284 | ? TREE_TYPE (orig_arg) : sizetype; | |
12285 | tree addcst | |
12286 | = build_int_cst (addtype, node->simdclone->args[i].linear_step); | |
12287 | g = gimple_build_assign_with_ops (code, iter2, iter1, addcst); | |
12288 | gsi = gsi_last_bb (incr_bb); | |
12289 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
12290 | ||
12291 | imm_use_iterator iter; | |
12292 | use_operand_p use_p; | |
12293 | gimple use_stmt; | |
12294 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, def) | |
12295 | if (use_stmt == phi) | |
12296 | continue; | |
12297 | else | |
12298 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
12299 | SET_USE (use_p, iter1); | |
12300 | } | |
12301 | } | |
12302 | ||
12303 | calculate_dominance_info (CDI_DOMINATORS); | |
12304 | add_loop (loop, loop->header->loop_father); | |
12305 | update_ssa (TODO_update_ssa); | |
12306 | ||
12307 | pop_cfun (); | |
12308 | } | |
12309 | ||
12310 | /* If the function in NODE is tagged as an elemental SIMD function, | |
12311 | create the appropriate SIMD clones. */ | |
12312 | ||
12313 | static void | |
12314 | expand_simd_clones (struct cgraph_node *node) | |
12315 | { | |
d09768a4 | 12316 | tree attr = lookup_attribute ("omp declare simd", |
12317 | DECL_ATTRIBUTES (node->decl)); | |
ea0695f9 | 12318 | if (attr == NULL_TREE |
12319 | || node->global.inlined_to | |
12320 | || lookup_attribute ("noclone", DECL_ATTRIBUTES (node->decl))) | |
d09768a4 | 12321 | return; |
ea0695f9 | 12322 | |
d09768a4 | 12323 | /* Ignore |
12324 | #pragma omp declare simd | |
12325 | extern int foo (); | |
12326 | in C, there we don't know the argument types at all. */ | |
12327 | if (!node->definition | |
12328 | && TYPE_ARG_TYPES (TREE_TYPE (node->decl)) == NULL_TREE) | |
12329 | return; | |
12330 | ||
12331 | do | |
12332 | { | |
12333 | /* Start with parsing the "omp declare simd" attribute(s). */ | |
12334 | bool inbranch_clause_specified; | |
12335 | struct cgraph_simd_clone *clone_info | |
12336 | = simd_clone_clauses_extract (node, TREE_VALUE (attr), | |
12337 | &inbranch_clause_specified); | |
12338 | if (clone_info == NULL) | |
12339 | continue; | |
12340 | ||
12341 | int orig_simdlen = clone_info->simdlen; | |
12342 | tree base_type = simd_clone_compute_base_data_type (node, clone_info); | |
12343 | /* The target can return 0 (no simd clones should be created), | |
12344 | 1 (just one ISA of simd clones should be created) or higher | |
12345 | count of ISA variants. In that case, clone_info is initialized | |
12346 | for the first ISA variant. */ | |
12347 | int count | |
12348 | = targetm.simd_clone.compute_vecsize_and_simdlen (node, clone_info, | |
12349 | base_type, 0); | |
12350 | if (count == 0) | |
12351 | continue; | |
12352 | ||
12353 | /* Loop over all COUNT ISA variants, and if !INBRANCH_CLAUSE_SPECIFIED, | |
12354 | also create one inbranch and one !inbranch clone of it. */ | |
12355 | for (int i = 0; i < count * 2; i++) | |
12356 | { | |
12357 | struct cgraph_simd_clone *clone = clone_info; | |
12358 | if (inbranch_clause_specified && (i & 1) != 0) | |
12359 | continue; | |
12360 | ||
12361 | if (i != 0) | |
12362 | { | |
12363 | clone = simd_clone_struct_alloc (clone_info->nargs | |
d09768a4 | 12364 | + ((i & 1) != 0)); |
12365 | simd_clone_struct_copy (clone, clone_info); | |
12366 | /* Undo changes targetm.simd_clone.compute_vecsize_and_simdlen | |
12367 | and simd_clone_adjust_argument_types did to the first | |
12368 | clone's info. */ | |
12369 | clone->nargs -= clone_info->inbranch; | |
12370 | clone->simdlen = orig_simdlen; | |
12371 | /* And call the target hook again to get the right ISA. */ | |
12372 | targetm.simd_clone.compute_vecsize_and_simdlen (node, clone, | |
12373 | base_type, | |
12374 | i / 2); | |
12375 | if ((i & 1) != 0) | |
12376 | clone->inbranch = 1; | |
12377 | } | |
12378 | ||
12379 | /* simd_clone_mangle might fail if such a clone has been created | |
12380 | already. */ | |
12381 | tree id = simd_clone_mangle (node, clone); | |
12382 | if (id == NULL_TREE) | |
12383 | continue; | |
12384 | ||
12385 | /* Only when we are sure we want to create the clone actually | |
12386 | clone the function (or definitions) or create another | |
12387 | extern FUNCTION_DECL (for prototypes without definitions). */ | |
12388 | struct cgraph_node *n = simd_clone_create (node); | |
12389 | if (n == NULL) | |
12390 | continue; | |
12391 | ||
12392 | n->simdclone = clone; | |
12393 | clone->origin = node; | |
12394 | clone->next_clone = NULL; | |
12395 | if (node->simd_clones == NULL) | |
12396 | { | |
12397 | clone->prev_clone = n; | |
12398 | node->simd_clones = n; | |
12399 | } | |
12400 | else | |
12401 | { | |
12402 | clone->prev_clone = node->simd_clones->simdclone->prev_clone; | |
12403 | clone->prev_clone->simdclone->next_clone = n; | |
12404 | node->simd_clones->simdclone->prev_clone = n; | |
12405 | } | |
35ee1c66 | 12406 | symtab->change_decl_assembler_name (n->decl, id); |
d09768a4 | 12407 | /* And finally adjust the return type, parameters and for |
12408 | definitions also function body. */ | |
12409 | if (node->definition) | |
12410 | simd_clone_adjust (n); | |
12411 | else | |
12412 | { | |
12413 | simd_clone_adjust_return_type (n); | |
12414 | simd_clone_adjust_argument_types (n); | |
12415 | } | |
12416 | } | |
12417 | } | |
12418 | while ((attr = lookup_attribute ("omp declare simd", TREE_CHAIN (attr)))); | |
12419 | } | |
12420 | ||
12421 | /* Entry point for IPA simd clone creation pass. */ | |
12422 | ||
12423 | static unsigned int | |
12424 | ipa_omp_simd_clone (void) | |
12425 | { | |
12426 | struct cgraph_node *node; | |
12427 | FOR_EACH_FUNCTION (node) | |
12428 | expand_simd_clones (node); | |
12429 | return 0; | |
12430 | } | |
12431 | ||
12432 | namespace { | |
12433 | ||
12434 | const pass_data pass_data_omp_simd_clone = | |
12435 | { | |
12436 | SIMPLE_IPA_PASS, /* type */ | |
12437 | "simdclone", /* name */ | |
12438 | OPTGROUP_NONE, /* optinfo_flags */ | |
d09768a4 | 12439 | TV_NONE, /* tv_id */ |
12440 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
12441 | 0, /* properties_provided */ | |
12442 | 0, /* properties_destroyed */ | |
12443 | 0, /* todo_flags_start */ | |
12444 | 0, /* todo_flags_finish */ | |
12445 | }; | |
12446 | ||
12447 | class pass_omp_simd_clone : public simple_ipa_opt_pass | |
12448 | { | |
12449 | public: | |
12450 | pass_omp_simd_clone(gcc::context *ctxt) | |
12451 | : simple_ipa_opt_pass(pass_data_omp_simd_clone, ctxt) | |
12452 | {} | |
12453 | ||
12454 | /* opt_pass methods: */ | |
31315c24 | 12455 | virtual bool gate (function *); |
65b0537f | 12456 | virtual unsigned int execute (function *) { return ipa_omp_simd_clone (); } |
d09768a4 | 12457 | }; |
12458 | ||
31315c24 | 12459 | bool |
12460 | pass_omp_simd_clone::gate (function *) | |
12461 | { | |
12462 | return ((flag_openmp || flag_openmp_simd | |
12463 | || flag_cilkplus | |
12464 | || (in_lto_p && !flag_wpa)) | |
12465 | && (targetm.simd_clone.compute_vecsize_and_simdlen != NULL)); | |
12466 | } | |
12467 | ||
d09768a4 | 12468 | } // anon namespace |
12469 | ||
12470 | simple_ipa_opt_pass * | |
12471 | make_pass_omp_simd_clone (gcc::context *ctxt) | |
12472 | { | |
12473 | return new pass_omp_simd_clone (ctxt); | |
12474 | } | |
cbe8bda8 | 12475 | |
dccabdd1 | 12476 | /* Helper function for omp_finish_file routine. Takes decls from V_DECLS and |
12477 | adds their addresses and sizes to constructor-vector V_CTOR. */ | |
12478 | static void | |
12479 | add_decls_addresses_to_decl_constructor (vec<tree, va_gc> *v_decls, | |
12480 | vec<constructor_elt, va_gc> *v_ctor) | |
12481 | { | |
12482 | unsigned len = vec_safe_length (v_decls); | |
12483 | for (unsigned i = 0; i < len; i++) | |
12484 | { | |
12485 | tree it = (*v_decls)[i]; | |
12486 | bool is_function = TREE_CODE (it) != VAR_DECL; | |
12487 | ||
12488 | CONSTRUCTOR_APPEND_ELT (v_ctor, NULL_TREE, build_fold_addr_expr (it)); | |
12489 | if (!is_function) | |
12490 | CONSTRUCTOR_APPEND_ELT (v_ctor, NULL_TREE, | |
12491 | fold_convert (const_ptr_type_node, | |
12492 | DECL_SIZE_UNIT (it))); | |
12493 | } | |
12494 | } | |
12495 | ||
12496 | /* Create new symbols containing (address, size) pairs for global variables, | |
12497 | marked with "omp declare target" attribute, as well as addresses for the | |
12498 | functions, which are outlined target regions. */ | |
12499 | void | |
12500 | omp_finish_file (void) | |
12501 | { | |
12502 | unsigned num_funcs = vec_safe_length (offload_funcs); | |
12503 | unsigned num_vars = vec_safe_length (offload_vars); | |
12504 | ||
12505 | if (num_funcs == 0 && num_vars == 0) | |
12506 | return; | |
12507 | ||
12508 | if (targetm_common.have_named_sections) | |
12509 | { | |
12510 | vec<constructor_elt, va_gc> *v_f, *v_v; | |
12511 | vec_alloc (v_f, num_funcs); | |
12512 | vec_alloc (v_v, num_vars * 2); | |
12513 | ||
12514 | add_decls_addresses_to_decl_constructor (offload_funcs, v_f); | |
12515 | add_decls_addresses_to_decl_constructor (offload_vars, v_v); | |
12516 | ||
12517 | tree vars_decl_type = build_array_type_nelts (pointer_sized_int_node, | |
12518 | num_vars * 2); | |
12519 | tree funcs_decl_type = build_array_type_nelts (pointer_sized_int_node, | |
12520 | num_funcs); | |
12521 | TYPE_ALIGN (vars_decl_type) = TYPE_ALIGN (pointer_sized_int_node); | |
12522 | TYPE_ALIGN (funcs_decl_type) = TYPE_ALIGN (pointer_sized_int_node); | |
12523 | tree ctor_v = build_constructor (vars_decl_type, v_v); | |
12524 | tree ctor_f = build_constructor (funcs_decl_type, v_f); | |
12525 | TREE_CONSTANT (ctor_v) = TREE_CONSTANT (ctor_f) = 1; | |
12526 | TREE_STATIC (ctor_v) = TREE_STATIC (ctor_f) = 1; | |
12527 | tree funcs_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, | |
12528 | get_identifier (".offload_func_table"), | |
12529 | funcs_decl_type); | |
12530 | tree vars_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, | |
12531 | get_identifier (".offload_var_table"), | |
12532 | vars_decl_type); | |
12533 | TREE_STATIC (funcs_decl) = TREE_STATIC (vars_decl) = 1; | |
12534 | /* Do not align tables more than TYPE_ALIGN (pointer_sized_int_node), | |
12535 | otherwise a joint table in a binary will contain padding between | |
12536 | tables from multiple object files. */ | |
12537 | DECL_USER_ALIGN (funcs_decl) = DECL_USER_ALIGN (vars_decl) = 1; | |
12538 | DECL_ALIGN (funcs_decl) = TYPE_ALIGN (funcs_decl_type); | |
12539 | DECL_ALIGN (vars_decl) = TYPE_ALIGN (vars_decl_type); | |
12540 | DECL_INITIAL (funcs_decl) = ctor_f; | |
12541 | DECL_INITIAL (vars_decl) = ctor_v; | |
12542 | set_decl_section_name (funcs_decl, OFFLOAD_FUNC_TABLE_SECTION_NAME); | |
12543 | set_decl_section_name (vars_decl, OFFLOAD_VAR_TABLE_SECTION_NAME); | |
12544 | ||
12545 | varpool_node::finalize_decl (vars_decl); | |
12546 | varpool_node::finalize_decl (funcs_decl); | |
7114ebdd | 12547 | } |
dccabdd1 | 12548 | else |
12549 | { | |
12550 | for (unsigned i = 0; i < num_funcs; i++) | |
12551 | { | |
12552 | tree it = (*offload_funcs)[i]; | |
12553 | targetm.record_offload_symbol (it); | |
12554 | } | |
12555 | for (unsigned i = 0; i < num_vars; i++) | |
12556 | { | |
12557 | tree it = (*offload_vars)[i]; | |
12558 | targetm.record_offload_symbol (it); | |
12559 | } | |
12560 | } | |
12561 | } | |
12562 | ||
1e8e9920 | 12563 | #include "gt-omp-low.h" |