]>
Commit | Line | Data |
---|---|---|
1 | /* Expansion pass for OMP directives. Outlines regions of certain OMP | |
2 | directives to separate functions, converts others into explicit calls to the | |
3 | runtime library (libgomp) and so forth | |
4 | ||
5 | Copyright (C) 2005-2021 Free Software Foundation, Inc. | |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
11 | Software Foundation; either version 3, or (at your option) any later | |
12 | version. | |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GCC; see the file COPYING3. If not see | |
21 | <http://www.gnu.org/licenses/>. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "memmodel.h" | |
27 | #include "backend.h" | |
28 | #include "target.h" | |
29 | #include "rtl.h" | |
30 | #include "tree.h" | |
31 | #include "gimple.h" | |
32 | #include "cfghooks.h" | |
33 | #include "tree-pass.h" | |
34 | #include "ssa.h" | |
35 | #include "optabs.h" | |
36 | #include "cgraph.h" | |
37 | #include "pretty-print.h" | |
38 | #include "diagnostic-core.h" | |
39 | #include "fold-const.h" | |
40 | #include "stor-layout.h" | |
41 | #include "cfganal.h" | |
42 | #include "internal-fn.h" | |
43 | #include "gimplify.h" | |
44 | #include "gimple-iterator.h" | |
45 | #include "gimplify-me.h" | |
46 | #include "gimple-walk.h" | |
47 | #include "tree-cfg.h" | |
48 | #include "tree-into-ssa.h" | |
49 | #include "tree-ssa.h" | |
50 | #include "splay-tree.h" | |
51 | #include "cfgloop.h" | |
52 | #include "omp-general.h" | |
53 | #include "omp-offload.h" | |
54 | #include "tree-cfgcleanup.h" | |
55 | #include "alloc-pool.h" | |
56 | #include "symbol-summary.h" | |
57 | #include "gomp-constants.h" | |
58 | #include "gimple-pretty-print.h" | |
59 | #include "stringpool.h" | |
60 | #include "attribs.h" | |
61 | #include "tree-eh.h" | |
62 | ||
63 | /* OMP region information. Every parallel and workshare | |
64 | directive is enclosed between two markers, the OMP_* directive | |
65 | and a corresponding GIMPLE_OMP_RETURN statement. */ | |
66 | ||
67 | struct omp_region | |
68 | { | |
69 | /* The enclosing region. */ | |
70 | struct omp_region *outer; | |
71 | ||
72 | /* First child region. */ | |
73 | struct omp_region *inner; | |
74 | ||
75 | /* Next peer region. */ | |
76 | struct omp_region *next; | |
77 | ||
78 | /* Block containing the omp directive as its last stmt. */ | |
79 | basic_block entry; | |
80 | ||
81 | /* Block containing the GIMPLE_OMP_RETURN as its last stmt. */ | |
82 | basic_block exit; | |
83 | ||
84 | /* Block containing the GIMPLE_OMP_CONTINUE as its last stmt. */ | |
85 | basic_block cont; | |
86 | ||
87 | /* If this is a combined parallel+workshare region, this is a list | |
88 | of additional arguments needed by the combined parallel+workshare | |
89 | library call. */ | |
90 | vec<tree, va_gc> *ws_args; | |
91 | ||
92 | /* The code for the omp directive of this region. */ | |
93 | enum gimple_code type; | |
94 | ||
95 | /* Schedule kind, only used for GIMPLE_OMP_FOR type regions. */ | |
96 | enum omp_clause_schedule_kind sched_kind; | |
97 | ||
98 | /* Schedule modifiers. */ | |
99 | unsigned char sched_modifiers; | |
100 | ||
101 | /* True if this is a combined parallel+workshare region. */ | |
102 | bool is_combined_parallel; | |
103 | ||
104 | /* Copy of fd.lastprivate_conditional != 0. */ | |
105 | bool has_lastprivate_conditional; | |
106 | ||
107 | /* The ordered stmt if type is GIMPLE_OMP_ORDERED and it has | |
108 | a depend clause. */ | |
109 | gomp_ordered *ord_stmt; | |
110 | }; | |
111 | ||
112 | static struct omp_region *root_omp_region; | |
113 | static bool omp_any_child_fn_dumped; | |
114 | ||
115 | static void expand_omp_build_assign (gimple_stmt_iterator *, tree, tree, | |
116 | bool = false); | |
117 | static gphi *find_phi_with_arg_on_edge (tree, edge); | |
118 | static void expand_omp (struct omp_region *region); | |
119 | ||
120 | /* Return true if REGION is a combined parallel+workshare region. */ | |
121 | ||
122 | static inline bool | |
123 | is_combined_parallel (struct omp_region *region) | |
124 | { | |
125 | return region->is_combined_parallel; | |
126 | } | |
127 | ||
128 | /* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB | |
129 | is the immediate dominator of PAR_ENTRY_BB, return true if there | |
130 | are no data dependencies that would prevent expanding the parallel | |
131 | directive at PAR_ENTRY_BB as a combined parallel+workshare region. | |
132 | ||
133 | When expanding a combined parallel+workshare region, the call to | |
134 | the child function may need additional arguments in the case of | |
135 | GIMPLE_OMP_FOR regions. In some cases, these arguments are | |
136 | computed out of variables passed in from the parent to the child | |
137 | via 'struct .omp_data_s'. For instance: | |
138 | ||
139 | #pragma omp parallel for schedule (guided, i * 4) | |
140 | for (j ...) | |
141 | ||
142 | Is lowered into: | |
143 | ||
144 | # BLOCK 2 (PAR_ENTRY_BB) | |
145 | .omp_data_o.i = i; | |
146 | #pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598) | |
147 | ||
148 | # BLOCK 3 (WS_ENTRY_BB) | |
149 | .omp_data_i = &.omp_data_o; | |
150 | D.1667 = .omp_data_i->i; | |
151 | D.1598 = D.1667 * 4; | |
152 | #pragma omp for schedule (guided, D.1598) | |
153 | ||
154 | When we outline the parallel region, the call to the child function | |
155 | 'bar.omp_fn.0' will need the value D.1598 in its argument list, but | |
156 | that value is computed *after* the call site. So, in principle we | |
157 | cannot do the transformation. | |
158 | ||
159 | To see whether the code in WS_ENTRY_BB blocks the combined | |
160 | parallel+workshare call, we collect all the variables used in the | |
161 | GIMPLE_OMP_FOR header check whether they appear on the LHS of any | |
162 | statement in WS_ENTRY_BB. If so, then we cannot emit the combined | |
163 | call. | |
164 | ||
165 | FIXME. If we had the SSA form built at this point, we could merely | |
166 | hoist the code in block 3 into block 2 and be done with it. But at | |
167 | this point we don't have dataflow information and though we could | |
168 | hack something up here, it is really not worth the aggravation. */ | |
169 | ||
170 | static bool | |
171 | workshare_safe_to_combine_p (basic_block ws_entry_bb) | |
172 | { | |
173 | struct omp_for_data fd; | |
174 | gimple *ws_stmt = last_stmt (ws_entry_bb); | |
175 | ||
176 | if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) | |
177 | return true; | |
178 | ||
179 | gcc_assert (gimple_code (ws_stmt) == GIMPLE_OMP_FOR); | |
180 | if (gimple_omp_for_kind (ws_stmt) != GF_OMP_FOR_KIND_FOR) | |
181 | return false; | |
182 | ||
183 | omp_extract_for_data (as_a <gomp_for *> (ws_stmt), &fd, NULL); | |
184 | ||
185 | if (fd.collapse > 1 && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
186 | return false; | |
187 | if (fd.iter_type != long_integer_type_node) | |
188 | return false; | |
189 | ||
190 | /* FIXME. We give up too easily here. If any of these arguments | |
191 | are not constants, they will likely involve variables that have | |
192 | been mapped into fields of .omp_data_s for sharing with the child | |
193 | function. With appropriate data flow, it would be possible to | |
194 | see through this. */ | |
195 | if (!is_gimple_min_invariant (fd.loop.n1) | |
196 | || !is_gimple_min_invariant (fd.loop.n2) | |
197 | || !is_gimple_min_invariant (fd.loop.step) | |
198 | || (fd.chunk_size && !is_gimple_min_invariant (fd.chunk_size))) | |
199 | return false; | |
200 | ||
201 | return true; | |
202 | } | |
203 | ||
204 | /* Adjust CHUNK_SIZE from SCHEDULE clause, depending on simd modifier | |
205 | presence (SIMD_SCHEDULE). */ | |
206 | ||
207 | static tree | |
208 | omp_adjust_chunk_size (tree chunk_size, bool simd_schedule) | |
209 | { | |
210 | if (!simd_schedule || integer_zerop (chunk_size)) | |
211 | return chunk_size; | |
212 | ||
213 | poly_uint64 vf = omp_max_vf (); | |
214 | if (known_eq (vf, 1U)) | |
215 | return chunk_size; | |
216 | ||
217 | tree type = TREE_TYPE (chunk_size); | |
218 | chunk_size = fold_build2 (PLUS_EXPR, type, chunk_size, | |
219 | build_int_cst (type, vf - 1)); | |
220 | return fold_build2 (BIT_AND_EXPR, type, chunk_size, | |
221 | build_int_cst (type, -vf)); | |
222 | } | |
223 | ||
224 | /* Collect additional arguments needed to emit a combined | |
225 | parallel+workshare call. WS_STMT is the workshare directive being | |
226 | expanded. */ | |
227 | ||
228 | static vec<tree, va_gc> * | |
229 | get_ws_args_for (gimple *par_stmt, gimple *ws_stmt) | |
230 | { | |
231 | tree t; | |
232 | location_t loc = gimple_location (ws_stmt); | |
233 | vec<tree, va_gc> *ws_args; | |
234 | ||
235 | if (gomp_for *for_stmt = dyn_cast <gomp_for *> (ws_stmt)) | |
236 | { | |
237 | struct omp_for_data fd; | |
238 | tree n1, n2; | |
239 | ||
240 | omp_extract_for_data (for_stmt, &fd, NULL); | |
241 | n1 = fd.loop.n1; | |
242 | n2 = fd.loop.n2; | |
243 | ||
244 | if (gimple_omp_for_combined_into_p (for_stmt)) | |
245 | { | |
246 | tree innerc | |
247 | = omp_find_clause (gimple_omp_parallel_clauses (par_stmt), | |
248 | OMP_CLAUSE__LOOPTEMP_); | |
249 | gcc_assert (innerc); | |
250 | n1 = OMP_CLAUSE_DECL (innerc); | |
251 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
252 | OMP_CLAUSE__LOOPTEMP_); | |
253 | gcc_assert (innerc); | |
254 | n2 = OMP_CLAUSE_DECL (innerc); | |
255 | } | |
256 | ||
257 | vec_alloc (ws_args, 3 + (fd.chunk_size != 0)); | |
258 | ||
259 | t = fold_convert_loc (loc, long_integer_type_node, n1); | |
260 | ws_args->quick_push (t); | |
261 | ||
262 | t = fold_convert_loc (loc, long_integer_type_node, n2); | |
263 | ws_args->quick_push (t); | |
264 | ||
265 | t = fold_convert_loc (loc, long_integer_type_node, fd.loop.step); | |
266 | ws_args->quick_push (t); | |
267 | ||
268 | if (fd.chunk_size) | |
269 | { | |
270 | t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size); | |
271 | t = omp_adjust_chunk_size (t, fd.simd_schedule); | |
272 | ws_args->quick_push (t); | |
273 | } | |
274 | ||
275 | return ws_args; | |
276 | } | |
277 | else if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) | |
278 | { | |
279 | /* Number of sections is equal to the number of edges from the | |
280 | GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to | |
281 | the exit of the sections region. */ | |
282 | basic_block bb = single_succ (gimple_bb (ws_stmt)); | |
283 | t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1); | |
284 | vec_alloc (ws_args, 1); | |
285 | ws_args->quick_push (t); | |
286 | return ws_args; | |
287 | } | |
288 | ||
289 | gcc_unreachable (); | |
290 | } | |
291 | ||
292 | /* Discover whether REGION is a combined parallel+workshare region. */ | |
293 | ||
294 | static void | |
295 | determine_parallel_type (struct omp_region *region) | |
296 | { | |
297 | basic_block par_entry_bb, par_exit_bb; | |
298 | basic_block ws_entry_bb, ws_exit_bb; | |
299 | ||
300 | if (region == NULL || region->inner == NULL | |
301 | || region->exit == NULL || region->inner->exit == NULL | |
302 | || region->inner->cont == NULL) | |
303 | return; | |
304 | ||
305 | /* We only support parallel+for and parallel+sections. */ | |
306 | if (region->type != GIMPLE_OMP_PARALLEL | |
307 | || (region->inner->type != GIMPLE_OMP_FOR | |
308 | && region->inner->type != GIMPLE_OMP_SECTIONS)) | |
309 | return; | |
310 | ||
311 | /* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and | |
312 | WS_EXIT_BB -> PAR_EXIT_BB. */ | |
313 | par_entry_bb = region->entry; | |
314 | par_exit_bb = region->exit; | |
315 | ws_entry_bb = region->inner->entry; | |
316 | ws_exit_bb = region->inner->exit; | |
317 | ||
318 | /* Give up for task reductions on the parallel, while it is implementable, | |
319 | adding another big set of APIs or slowing down the normal paths is | |
320 | not acceptable. */ | |
321 | tree pclauses = gimple_omp_parallel_clauses (last_stmt (par_entry_bb)); | |
322 | if (omp_find_clause (pclauses, OMP_CLAUSE__REDUCTEMP_)) | |
323 | return; | |
324 | ||
325 | if (single_succ (par_entry_bb) == ws_entry_bb | |
326 | && single_succ (ws_exit_bb) == par_exit_bb | |
327 | && workshare_safe_to_combine_p (ws_entry_bb) | |
328 | && (gimple_omp_parallel_combined_p (last_stmt (par_entry_bb)) | |
329 | || (last_and_only_stmt (ws_entry_bb) | |
330 | && last_and_only_stmt (par_exit_bb)))) | |
331 | { | |
332 | gimple *par_stmt = last_stmt (par_entry_bb); | |
333 | gimple *ws_stmt = last_stmt (ws_entry_bb); | |
334 | ||
335 | if (region->inner->type == GIMPLE_OMP_FOR) | |
336 | { | |
337 | /* If this is a combined parallel loop, we need to determine | |
338 | whether or not to use the combined library calls. There | |
339 | are two cases where we do not apply the transformation: | |
340 | static loops and any kind of ordered loop. In the first | |
341 | case, we already open code the loop so there is no need | |
342 | to do anything else. In the latter case, the combined | |
343 | parallel loop call would still need extra synchronization | |
344 | to implement ordered semantics, so there would not be any | |
345 | gain in using the combined call. */ | |
346 | tree clauses = gimple_omp_for_clauses (ws_stmt); | |
347 | tree c = omp_find_clause (clauses, OMP_CLAUSE_SCHEDULE); | |
348 | if (c == NULL | |
349 | || ((OMP_CLAUSE_SCHEDULE_KIND (c) & OMP_CLAUSE_SCHEDULE_MASK) | |
350 | == OMP_CLAUSE_SCHEDULE_STATIC) | |
351 | || omp_find_clause (clauses, OMP_CLAUSE_ORDERED) | |
352 | || omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_) | |
353 | || ((c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_)) | |
354 | && POINTER_TYPE_P (TREE_TYPE (OMP_CLAUSE_DECL (c))))) | |
355 | return; | |
356 | } | |
357 | else if (region->inner->type == GIMPLE_OMP_SECTIONS | |
358 | && (omp_find_clause (gimple_omp_sections_clauses (ws_stmt), | |
359 | OMP_CLAUSE__REDUCTEMP_) | |
360 | || omp_find_clause (gimple_omp_sections_clauses (ws_stmt), | |
361 | OMP_CLAUSE__CONDTEMP_))) | |
362 | return; | |
363 | ||
364 | region->is_combined_parallel = true; | |
365 | region->inner->is_combined_parallel = true; | |
366 | region->ws_args = get_ws_args_for (par_stmt, ws_stmt); | |
367 | } | |
368 | } | |
369 | ||
370 | /* Debugging dumps for parallel regions. */ | |
371 | void dump_omp_region (FILE *, struct omp_region *, int); | |
372 | void debug_omp_region (struct omp_region *); | |
373 | void debug_all_omp_regions (void); | |
374 | ||
375 | /* Dump the parallel region tree rooted at REGION. */ | |
376 | ||
377 | void | |
378 | dump_omp_region (FILE *file, struct omp_region *region, int indent) | |
379 | { | |
380 | fprintf (file, "%*sbb %d: %s\n", indent, "", region->entry->index, | |
381 | gimple_code_name[region->type]); | |
382 | ||
383 | if (region->inner) | |
384 | dump_omp_region (file, region->inner, indent + 4); | |
385 | ||
386 | if (region->cont) | |
387 | { | |
388 | fprintf (file, "%*sbb %d: GIMPLE_OMP_CONTINUE\n", indent, "", | |
389 | region->cont->index); | |
390 | } | |
391 | ||
392 | if (region->exit) | |
393 | fprintf (file, "%*sbb %d: GIMPLE_OMP_RETURN\n", indent, "", | |
394 | region->exit->index); | |
395 | else | |
396 | fprintf (file, "%*s[no exit marker]\n", indent, ""); | |
397 | ||
398 | if (region->next) | |
399 | dump_omp_region (file, region->next, indent); | |
400 | } | |
401 | ||
402 | DEBUG_FUNCTION void | |
403 | debug_omp_region (struct omp_region *region) | |
404 | { | |
405 | dump_omp_region (stderr, region, 0); | |
406 | } | |
407 | ||
408 | DEBUG_FUNCTION void | |
409 | debug_all_omp_regions (void) | |
410 | { | |
411 | dump_omp_region (stderr, root_omp_region, 0); | |
412 | } | |
413 | ||
414 | /* Create a new parallel region starting at STMT inside region PARENT. */ | |
415 | ||
416 | static struct omp_region * | |
417 | new_omp_region (basic_block bb, enum gimple_code type, | |
418 | struct omp_region *parent) | |
419 | { | |
420 | struct omp_region *region = XCNEW (struct omp_region); | |
421 | ||
422 | region->outer = parent; | |
423 | region->entry = bb; | |
424 | region->type = type; | |
425 | ||
426 | if (parent) | |
427 | { | |
428 | /* This is a nested region. Add it to the list of inner | |
429 | regions in PARENT. */ | |
430 | region->next = parent->inner; | |
431 | parent->inner = region; | |
432 | } | |
433 | else | |
434 | { | |
435 | /* This is a toplevel region. Add it to the list of toplevel | |
436 | regions in ROOT_OMP_REGION. */ | |
437 | region->next = root_omp_region; | |
438 | root_omp_region = region; | |
439 | } | |
440 | ||
441 | return region; | |
442 | } | |
443 | ||
444 | /* Release the memory associated with the region tree rooted at REGION. */ | |
445 | ||
446 | static void | |
447 | free_omp_region_1 (struct omp_region *region) | |
448 | { | |
449 | struct omp_region *i, *n; | |
450 | ||
451 | for (i = region->inner; i ; i = n) | |
452 | { | |
453 | n = i->next; | |
454 | free_omp_region_1 (i); | |
455 | } | |
456 | ||
457 | free (region); | |
458 | } | |
459 | ||
460 | /* Release the memory for the entire omp region tree. */ | |
461 | ||
462 | void | |
463 | omp_free_regions (void) | |
464 | { | |
465 | struct omp_region *r, *n; | |
466 | for (r = root_omp_region; r ; r = n) | |
467 | { | |
468 | n = r->next; | |
469 | free_omp_region_1 (r); | |
470 | } | |
471 | root_omp_region = NULL; | |
472 | } | |
473 | ||
474 | /* A convenience function to build an empty GIMPLE_COND with just the | |
475 | condition. */ | |
476 | ||
477 | static gcond * | |
478 | gimple_build_cond_empty (tree cond) | |
479 | { | |
480 | enum tree_code pred_code; | |
481 | tree lhs, rhs; | |
482 | ||
483 | gimple_cond_get_ops_from_tree (cond, &pred_code, &lhs, &rhs); | |
484 | return gimple_build_cond (pred_code, lhs, rhs, NULL_TREE, NULL_TREE); | |
485 | } | |
486 | ||
487 | /* Change DECL_CONTEXT of CHILD_FNDECL to that of the parent function. | |
488 | Add CHILD_FNDECL to decl chain of the supercontext of the block | |
489 | ENTRY_BLOCK - this is the block which originally contained the | |
490 | code from which CHILD_FNDECL was created. | |
491 | ||
492 | Together, these actions ensure that the debug info for the outlined | |
493 | function will be emitted with the correct lexical scope. */ | |
494 | ||
495 | static void | |
496 | adjust_context_and_scope (struct omp_region *region, tree entry_block, | |
497 | tree child_fndecl) | |
498 | { | |
499 | tree parent_fndecl = NULL_TREE; | |
500 | gimple *entry_stmt; | |
501 | /* OMP expansion expands inner regions before outer ones, so if | |
502 | we e.g. have explicit task region nested in parallel region, when | |
503 | expanding the task region current_function_decl will be the original | |
504 | source function, but we actually want to use as context the child | |
505 | function of the parallel. */ | |
506 | for (region = region->outer; | |
507 | region && parent_fndecl == NULL_TREE; region = region->outer) | |
508 | switch (region->type) | |
509 | { | |
510 | case GIMPLE_OMP_PARALLEL: | |
511 | case GIMPLE_OMP_TASK: | |
512 | case GIMPLE_OMP_TEAMS: | |
513 | entry_stmt = last_stmt (region->entry); | |
514 | parent_fndecl = gimple_omp_taskreg_child_fn (entry_stmt); | |
515 | break; | |
516 | case GIMPLE_OMP_TARGET: | |
517 | entry_stmt = last_stmt (region->entry); | |
518 | parent_fndecl | |
519 | = gimple_omp_target_child_fn (as_a <gomp_target *> (entry_stmt)); | |
520 | break; | |
521 | default: | |
522 | break; | |
523 | } | |
524 | ||
525 | if (parent_fndecl == NULL_TREE) | |
526 | parent_fndecl = current_function_decl; | |
527 | DECL_CONTEXT (child_fndecl) = parent_fndecl; | |
528 | ||
529 | if (entry_block != NULL_TREE && TREE_CODE (entry_block) == BLOCK) | |
530 | { | |
531 | tree b = BLOCK_SUPERCONTEXT (entry_block); | |
532 | if (TREE_CODE (b) == BLOCK) | |
533 | { | |
534 | DECL_CHAIN (child_fndecl) = BLOCK_VARS (b); | |
535 | BLOCK_VARS (b) = child_fndecl; | |
536 | } | |
537 | } | |
538 | } | |
539 | ||
540 | /* Build the function calls to GOMP_parallel etc to actually | |
541 | generate the parallel operation. REGION is the parallel region | |
542 | being expanded. BB is the block where to insert the code. WS_ARGS | |
543 | will be set if this is a call to a combined parallel+workshare | |
544 | construct, it contains the list of additional arguments needed by | |
545 | the workshare construct. */ | |
546 | ||
547 | static void | |
548 | expand_parallel_call (struct omp_region *region, basic_block bb, | |
549 | gomp_parallel *entry_stmt, | |
550 | vec<tree, va_gc> *ws_args) | |
551 | { | |
552 | tree t, t1, t2, val, cond, c, clauses, flags; | |
553 | gimple_stmt_iterator gsi; | |
554 | gimple *stmt; | |
555 | enum built_in_function start_ix; | |
556 | int start_ix2; | |
557 | location_t clause_loc; | |
558 | vec<tree, va_gc> *args; | |
559 | ||
560 | clauses = gimple_omp_parallel_clauses (entry_stmt); | |
561 | ||
562 | /* Determine what flavor of GOMP_parallel we will be | |
563 | emitting. */ | |
564 | start_ix = BUILT_IN_GOMP_PARALLEL; | |
565 | tree rtmp = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
566 | if (rtmp) | |
567 | start_ix = BUILT_IN_GOMP_PARALLEL_REDUCTIONS; | |
568 | else if (is_combined_parallel (region)) | |
569 | { | |
570 | switch (region->inner->type) | |
571 | { | |
572 | case GIMPLE_OMP_FOR: | |
573 | gcc_assert (region->inner->sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); | |
574 | switch (region->inner->sched_kind) | |
575 | { | |
576 | case OMP_CLAUSE_SCHEDULE_RUNTIME: | |
577 | /* For lastprivate(conditional:), our implementation | |
578 | requires monotonic behavior. */ | |
579 | if (region->inner->has_lastprivate_conditional != 0) | |
580 | start_ix2 = 3; | |
581 | else if ((region->inner->sched_modifiers | |
582 | & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0) | |
583 | start_ix2 = 6; | |
584 | else if ((region->inner->sched_modifiers | |
585 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0) | |
586 | start_ix2 = 7; | |
587 | else | |
588 | start_ix2 = 3; | |
589 | break; | |
590 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: | |
591 | case OMP_CLAUSE_SCHEDULE_GUIDED: | |
592 | if ((region->inner->sched_modifiers | |
593 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 | |
594 | && !region->inner->has_lastprivate_conditional) | |
595 | { | |
596 | start_ix2 = 3 + region->inner->sched_kind; | |
597 | break; | |
598 | } | |
599 | /* FALLTHRU */ | |
600 | default: | |
601 | start_ix2 = region->inner->sched_kind; | |
602 | break; | |
603 | } | |
604 | start_ix2 += (int) BUILT_IN_GOMP_PARALLEL_LOOP_STATIC; | |
605 | start_ix = (enum built_in_function) start_ix2; | |
606 | break; | |
607 | case GIMPLE_OMP_SECTIONS: | |
608 | start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS; | |
609 | break; | |
610 | default: | |
611 | gcc_unreachable (); | |
612 | } | |
613 | } | |
614 | ||
615 | /* By default, the value of NUM_THREADS is zero (selected at run time) | |
616 | and there is no conditional. */ | |
617 | cond = NULL_TREE; | |
618 | val = build_int_cst (unsigned_type_node, 0); | |
619 | flags = build_int_cst (unsigned_type_node, 0); | |
620 | ||
621 | c = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
622 | if (c) | |
623 | cond = OMP_CLAUSE_IF_EXPR (c); | |
624 | ||
625 | c = omp_find_clause (clauses, OMP_CLAUSE_NUM_THREADS); | |
626 | if (c) | |
627 | { | |
628 | val = OMP_CLAUSE_NUM_THREADS_EXPR (c); | |
629 | clause_loc = OMP_CLAUSE_LOCATION (c); | |
630 | } | |
631 | else | |
632 | clause_loc = gimple_location (entry_stmt); | |
633 | ||
634 | c = omp_find_clause (clauses, OMP_CLAUSE_PROC_BIND); | |
635 | if (c) | |
636 | flags = build_int_cst (unsigned_type_node, OMP_CLAUSE_PROC_BIND_KIND (c)); | |
637 | ||
638 | /* Ensure 'val' is of the correct type. */ | |
639 | val = fold_convert_loc (clause_loc, unsigned_type_node, val); | |
640 | ||
641 | /* If we found the clause 'if (cond)', build either | |
642 | (cond != 0) or (cond ? val : 1u). */ | |
643 | if (cond) | |
644 | { | |
645 | cond = gimple_boolify (cond); | |
646 | ||
647 | if (integer_zerop (val)) | |
648 | val = fold_build2_loc (clause_loc, | |
649 | EQ_EXPR, unsigned_type_node, cond, | |
650 | build_int_cst (TREE_TYPE (cond), 0)); | |
651 | else | |
652 | { | |
653 | basic_block cond_bb, then_bb, else_bb; | |
654 | edge e, e_then, e_else; | |
655 | tree tmp_then, tmp_else, tmp_join, tmp_var; | |
656 | ||
657 | tmp_var = create_tmp_var (TREE_TYPE (val)); | |
658 | if (gimple_in_ssa_p (cfun)) | |
659 | { | |
660 | tmp_then = make_ssa_name (tmp_var); | |
661 | tmp_else = make_ssa_name (tmp_var); | |
662 | tmp_join = make_ssa_name (tmp_var); | |
663 | } | |
664 | else | |
665 | { | |
666 | tmp_then = tmp_var; | |
667 | tmp_else = tmp_var; | |
668 | tmp_join = tmp_var; | |
669 | } | |
670 | ||
671 | e = split_block_after_labels (bb); | |
672 | cond_bb = e->src; | |
673 | bb = e->dest; | |
674 | remove_edge (e); | |
675 | ||
676 | then_bb = create_empty_bb (cond_bb); | |
677 | else_bb = create_empty_bb (then_bb); | |
678 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
679 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
680 | ||
681 | stmt = gimple_build_cond_empty (cond); | |
682 | gsi = gsi_start_bb (cond_bb); | |
683 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
684 | ||
685 | gsi = gsi_start_bb (then_bb); | |
686 | expand_omp_build_assign (&gsi, tmp_then, val, true); | |
687 | ||
688 | gsi = gsi_start_bb (else_bb); | |
689 | expand_omp_build_assign (&gsi, tmp_else, | |
690 | build_int_cst (unsigned_type_node, 1), | |
691 | true); | |
692 | ||
693 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
694 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
695 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
696 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
697 | e_then = make_edge (then_bb, bb, EDGE_FALLTHRU); | |
698 | e_else = make_edge (else_bb, bb, EDGE_FALLTHRU); | |
699 | ||
700 | if (gimple_in_ssa_p (cfun)) | |
701 | { | |
702 | gphi *phi = create_phi_node (tmp_join, bb); | |
703 | add_phi_arg (phi, tmp_then, e_then, UNKNOWN_LOCATION); | |
704 | add_phi_arg (phi, tmp_else, e_else, UNKNOWN_LOCATION); | |
705 | } | |
706 | ||
707 | val = tmp_join; | |
708 | } | |
709 | ||
710 | gsi = gsi_start_bb (bb); | |
711 | val = force_gimple_operand_gsi (&gsi, val, true, NULL_TREE, | |
712 | false, GSI_CONTINUE_LINKING); | |
713 | } | |
714 | ||
715 | gsi = gsi_last_nondebug_bb (bb); | |
716 | t = gimple_omp_parallel_data_arg (entry_stmt); | |
717 | if (t == NULL) | |
718 | t1 = null_pointer_node; | |
719 | else | |
720 | t1 = build_fold_addr_expr (t); | |
721 | tree child_fndecl = gimple_omp_parallel_child_fn (entry_stmt); | |
722 | t2 = build_fold_addr_expr (child_fndecl); | |
723 | ||
724 | vec_alloc (args, 4 + vec_safe_length (ws_args)); | |
725 | args->quick_push (t2); | |
726 | args->quick_push (t1); | |
727 | args->quick_push (val); | |
728 | if (ws_args) | |
729 | args->splice (*ws_args); | |
730 | args->quick_push (flags); | |
731 | ||
732 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
733 | builtin_decl_explicit (start_ix), args); | |
734 | ||
735 | if (rtmp) | |
736 | { | |
737 | tree type = TREE_TYPE (OMP_CLAUSE_DECL (rtmp)); | |
738 | t = build2 (MODIFY_EXPR, type, OMP_CLAUSE_DECL (rtmp), | |
739 | fold_convert (type, | |
740 | fold_convert (pointer_sized_int_node, t))); | |
741 | } | |
742 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
743 | false, GSI_CONTINUE_LINKING); | |
744 | } | |
745 | ||
746 | /* Build the function call to GOMP_task to actually | |
747 | generate the task operation. BB is the block where to insert the code. */ | |
748 | ||
749 | static void | |
750 | expand_task_call (struct omp_region *region, basic_block bb, | |
751 | gomp_task *entry_stmt) | |
752 | { | |
753 | tree t1, t2, t3; | |
754 | gimple_stmt_iterator gsi; | |
755 | location_t loc = gimple_location (entry_stmt); | |
756 | ||
757 | tree clauses = gimple_omp_task_clauses (entry_stmt); | |
758 | ||
759 | tree ifc = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
760 | tree untied = omp_find_clause (clauses, OMP_CLAUSE_UNTIED); | |
761 | tree mergeable = omp_find_clause (clauses, OMP_CLAUSE_MERGEABLE); | |
762 | tree depend = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
763 | tree finalc = omp_find_clause (clauses, OMP_CLAUSE_FINAL); | |
764 | tree priority = omp_find_clause (clauses, OMP_CLAUSE_PRIORITY); | |
765 | tree detach = omp_find_clause (clauses, OMP_CLAUSE_DETACH); | |
766 | ||
767 | unsigned int iflags | |
768 | = (untied ? GOMP_TASK_FLAG_UNTIED : 0) | |
769 | | (mergeable ? GOMP_TASK_FLAG_MERGEABLE : 0) | |
770 | | (depend ? GOMP_TASK_FLAG_DEPEND : 0); | |
771 | ||
772 | bool taskloop_p = gimple_omp_task_taskloop_p (entry_stmt); | |
773 | tree startvar = NULL_TREE, endvar = NULL_TREE, step = NULL_TREE; | |
774 | tree num_tasks = NULL_TREE; | |
775 | bool ull = false; | |
776 | if (taskloop_p) | |
777 | { | |
778 | gimple *g = last_stmt (region->outer->entry); | |
779 | gcc_assert (gimple_code (g) == GIMPLE_OMP_FOR | |
780 | && gimple_omp_for_kind (g) == GF_OMP_FOR_KIND_TASKLOOP); | |
781 | struct omp_for_data fd; | |
782 | omp_extract_for_data (as_a <gomp_for *> (g), &fd, NULL); | |
783 | startvar = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
784 | endvar = omp_find_clause (OMP_CLAUSE_CHAIN (startvar), | |
785 | OMP_CLAUSE__LOOPTEMP_); | |
786 | startvar = OMP_CLAUSE_DECL (startvar); | |
787 | endvar = OMP_CLAUSE_DECL (endvar); | |
788 | step = fold_convert_loc (loc, fd.iter_type, fd.loop.step); | |
789 | if (fd.loop.cond_code == LT_EXPR) | |
790 | iflags |= GOMP_TASK_FLAG_UP; | |
791 | tree tclauses = gimple_omp_for_clauses (g); | |
792 | num_tasks = omp_find_clause (tclauses, OMP_CLAUSE_NUM_TASKS); | |
793 | if (num_tasks) | |
794 | { | |
795 | if (OMP_CLAUSE_NUM_TASKS_STRICT (num_tasks)) | |
796 | iflags |= GOMP_TASK_FLAG_STRICT; | |
797 | num_tasks = OMP_CLAUSE_NUM_TASKS_EXPR (num_tasks); | |
798 | } | |
799 | else | |
800 | { | |
801 | num_tasks = omp_find_clause (tclauses, OMP_CLAUSE_GRAINSIZE); | |
802 | if (num_tasks) | |
803 | { | |
804 | iflags |= GOMP_TASK_FLAG_GRAINSIZE; | |
805 | if (OMP_CLAUSE_GRAINSIZE_STRICT (num_tasks)) | |
806 | iflags |= GOMP_TASK_FLAG_STRICT; | |
807 | num_tasks = OMP_CLAUSE_GRAINSIZE_EXPR (num_tasks); | |
808 | } | |
809 | else | |
810 | num_tasks = integer_zero_node; | |
811 | } | |
812 | num_tasks = fold_convert_loc (loc, long_integer_type_node, num_tasks); | |
813 | if (ifc == NULL_TREE) | |
814 | iflags |= GOMP_TASK_FLAG_IF; | |
815 | if (omp_find_clause (tclauses, OMP_CLAUSE_NOGROUP)) | |
816 | iflags |= GOMP_TASK_FLAG_NOGROUP; | |
817 | ull = fd.iter_type == long_long_unsigned_type_node; | |
818 | if (omp_find_clause (clauses, OMP_CLAUSE_REDUCTION)) | |
819 | iflags |= GOMP_TASK_FLAG_REDUCTION; | |
820 | } | |
821 | else | |
822 | { | |
823 | if (priority) | |
824 | iflags |= GOMP_TASK_FLAG_PRIORITY; | |
825 | if (detach) | |
826 | iflags |= GOMP_TASK_FLAG_DETACH; | |
827 | } | |
828 | ||
829 | tree flags = build_int_cst (unsigned_type_node, iflags); | |
830 | ||
831 | tree cond = boolean_true_node; | |
832 | if (ifc) | |
833 | { | |
834 | if (taskloop_p) | |
835 | { | |
836 | tree t = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); | |
837 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, | |
838 | build_int_cst (unsigned_type_node, | |
839 | GOMP_TASK_FLAG_IF), | |
840 | build_int_cst (unsigned_type_node, 0)); | |
841 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, | |
842 | flags, t); | |
843 | } | |
844 | else | |
845 | cond = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); | |
846 | } | |
847 | ||
848 | if (finalc) | |
849 | { | |
850 | tree t = gimple_boolify (OMP_CLAUSE_FINAL_EXPR (finalc)); | |
851 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, | |
852 | build_int_cst (unsigned_type_node, | |
853 | GOMP_TASK_FLAG_FINAL), | |
854 | build_int_cst (unsigned_type_node, 0)); | |
855 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, flags, t); | |
856 | } | |
857 | if (depend) | |
858 | depend = OMP_CLAUSE_DECL (depend); | |
859 | else | |
860 | depend = build_int_cst (ptr_type_node, 0); | |
861 | if (priority) | |
862 | priority = fold_convert (integer_type_node, | |
863 | OMP_CLAUSE_PRIORITY_EXPR (priority)); | |
864 | else | |
865 | priority = integer_zero_node; | |
866 | ||
867 | gsi = gsi_last_nondebug_bb (bb); | |
868 | ||
869 | detach = (detach | |
870 | ? build_fold_addr_expr (OMP_CLAUSE_DECL (detach)) | |
871 | : null_pointer_node); | |
872 | ||
873 | tree t = gimple_omp_task_data_arg (entry_stmt); | |
874 | if (t == NULL) | |
875 | t2 = null_pointer_node; | |
876 | else | |
877 | t2 = build_fold_addr_expr_loc (loc, t); | |
878 | t1 = build_fold_addr_expr_loc (loc, gimple_omp_task_child_fn (entry_stmt)); | |
879 | t = gimple_omp_task_copy_fn (entry_stmt); | |
880 | if (t == NULL) | |
881 | t3 = null_pointer_node; | |
882 | else | |
883 | t3 = build_fold_addr_expr_loc (loc, t); | |
884 | ||
885 | if (taskloop_p) | |
886 | t = build_call_expr (ull | |
887 | ? builtin_decl_explicit (BUILT_IN_GOMP_TASKLOOP_ULL) | |
888 | : builtin_decl_explicit (BUILT_IN_GOMP_TASKLOOP), | |
889 | 11, t1, t2, t3, | |
890 | gimple_omp_task_arg_size (entry_stmt), | |
891 | gimple_omp_task_arg_align (entry_stmt), flags, | |
892 | num_tasks, priority, startvar, endvar, step); | |
893 | else | |
894 | t = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASK), | |
895 | 10, t1, t2, t3, | |
896 | gimple_omp_task_arg_size (entry_stmt), | |
897 | gimple_omp_task_arg_align (entry_stmt), cond, flags, | |
898 | depend, priority, detach); | |
899 | ||
900 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
901 | false, GSI_CONTINUE_LINKING); | |
902 | } | |
903 | ||
904 | /* Build the function call to GOMP_taskwait_depend to actually | |
905 | generate the taskwait operation. BB is the block where to insert the | |
906 | code. */ | |
907 | ||
908 | static void | |
909 | expand_taskwait_call (basic_block bb, gomp_task *entry_stmt) | |
910 | { | |
911 | tree clauses = gimple_omp_task_clauses (entry_stmt); | |
912 | tree depend = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
913 | if (depend == NULL_TREE) | |
914 | return; | |
915 | ||
916 | depend = OMP_CLAUSE_DECL (depend); | |
917 | ||
918 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); | |
919 | tree t | |
920 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASKWAIT_DEPEND), | |
921 | 1, depend); | |
922 | ||
923 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
924 | false, GSI_CONTINUE_LINKING); | |
925 | } | |
926 | ||
927 | /* Build the function call to GOMP_teams_reg to actually | |
928 | generate the host teams operation. REGION is the teams region | |
929 | being expanded. BB is the block where to insert the code. */ | |
930 | ||
931 | static void | |
932 | expand_teams_call (basic_block bb, gomp_teams *entry_stmt) | |
933 | { | |
934 | tree clauses = gimple_omp_teams_clauses (entry_stmt); | |
935 | tree num_teams = omp_find_clause (clauses, OMP_CLAUSE_NUM_TEAMS); | |
936 | if (num_teams == NULL_TREE) | |
937 | num_teams = build_int_cst (unsigned_type_node, 0); | |
938 | else | |
939 | { | |
940 | num_teams = OMP_CLAUSE_NUM_TEAMS_EXPR (num_teams); | |
941 | num_teams = fold_convert (unsigned_type_node, num_teams); | |
942 | } | |
943 | tree thread_limit = omp_find_clause (clauses, OMP_CLAUSE_THREAD_LIMIT); | |
944 | if (thread_limit == NULL_TREE) | |
945 | thread_limit = build_int_cst (unsigned_type_node, 0); | |
946 | else | |
947 | { | |
948 | thread_limit = OMP_CLAUSE_THREAD_LIMIT_EXPR (thread_limit); | |
949 | thread_limit = fold_convert (unsigned_type_node, thread_limit); | |
950 | } | |
951 | ||
952 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); | |
953 | tree t = gimple_omp_teams_data_arg (entry_stmt), t1; | |
954 | if (t == NULL) | |
955 | t1 = null_pointer_node; | |
956 | else | |
957 | t1 = build_fold_addr_expr (t); | |
958 | tree child_fndecl = gimple_omp_teams_child_fn (entry_stmt); | |
959 | tree t2 = build_fold_addr_expr (child_fndecl); | |
960 | ||
961 | vec<tree, va_gc> *args; | |
962 | vec_alloc (args, 5); | |
963 | args->quick_push (t2); | |
964 | args->quick_push (t1); | |
965 | args->quick_push (num_teams); | |
966 | args->quick_push (thread_limit); | |
967 | /* For future extensibility. */ | |
968 | args->quick_push (build_zero_cst (unsigned_type_node)); | |
969 | ||
970 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
971 | builtin_decl_explicit (BUILT_IN_GOMP_TEAMS_REG), | |
972 | args); | |
973 | ||
974 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
975 | false, GSI_CONTINUE_LINKING); | |
976 | } | |
977 | ||
978 | /* Chain all the DECLs in LIST by their TREE_CHAIN fields. */ | |
979 | ||
980 | static tree | |
981 | vec2chain (vec<tree, va_gc> *v) | |
982 | { | |
983 | tree chain = NULL_TREE, t; | |
984 | unsigned ix; | |
985 | ||
986 | FOR_EACH_VEC_SAFE_ELT_REVERSE (v, ix, t) | |
987 | { | |
988 | DECL_CHAIN (t) = chain; | |
989 | chain = t; | |
990 | } | |
991 | ||
992 | return chain; | |
993 | } | |
994 | ||
995 | /* Remove barriers in REGION->EXIT's block. Note that this is only | |
996 | valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region | |
997 | is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that | |
998 | left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be | |
999 | removed. */ | |
1000 | ||
1001 | static void | |
1002 | remove_exit_barrier (struct omp_region *region) | |
1003 | { | |
1004 | gimple_stmt_iterator gsi; | |
1005 | basic_block exit_bb; | |
1006 | edge_iterator ei; | |
1007 | edge e; | |
1008 | gimple *stmt; | |
1009 | int any_addressable_vars = -1; | |
1010 | ||
1011 | exit_bb = region->exit; | |
1012 | ||
1013 | /* If the parallel region doesn't return, we don't have REGION->EXIT | |
1014 | block at all. */ | |
1015 | if (! exit_bb) | |
1016 | return; | |
1017 | ||
1018 | /* The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The | |
1019 | workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of | |
1020 | statements that can appear in between are extremely limited -- no | |
1021 | memory operations at all. Here, we allow nothing at all, so the | |
1022 | only thing we allow to precede this GIMPLE_OMP_RETURN is a label. */ | |
1023 | gsi = gsi_last_nondebug_bb (exit_bb); | |
1024 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
1025 | gsi_prev_nondebug (&gsi); | |
1026 | if (!gsi_end_p (gsi) && gimple_code (gsi_stmt (gsi)) != GIMPLE_LABEL) | |
1027 | return; | |
1028 | ||
1029 | FOR_EACH_EDGE (e, ei, exit_bb->preds) | |
1030 | { | |
1031 | gsi = gsi_last_nondebug_bb (e->src); | |
1032 | if (gsi_end_p (gsi)) | |
1033 | continue; | |
1034 | stmt = gsi_stmt (gsi); | |
1035 | if (gimple_code (stmt) == GIMPLE_OMP_RETURN | |
1036 | && !gimple_omp_return_nowait_p (stmt)) | |
1037 | { | |
1038 | /* OpenMP 3.0 tasks unfortunately prevent this optimization | |
1039 | in many cases. If there could be tasks queued, the barrier | |
1040 | might be needed to let the tasks run before some local | |
1041 | variable of the parallel that the task uses as shared | |
1042 | runs out of scope. The task can be spawned either | |
1043 | from within current function (this would be easy to check) | |
1044 | or from some function it calls and gets passed an address | |
1045 | of such a variable. */ | |
1046 | if (any_addressable_vars < 0) | |
1047 | { | |
1048 | gomp_parallel *parallel_stmt | |
1049 | = as_a <gomp_parallel *> (last_stmt (region->entry)); | |
1050 | tree child_fun = gimple_omp_parallel_child_fn (parallel_stmt); | |
1051 | tree local_decls, block, decl; | |
1052 | unsigned ix; | |
1053 | ||
1054 | any_addressable_vars = 0; | |
1055 | FOR_EACH_LOCAL_DECL (DECL_STRUCT_FUNCTION (child_fun), ix, decl) | |
1056 | if (TREE_ADDRESSABLE (decl)) | |
1057 | { | |
1058 | any_addressable_vars = 1; | |
1059 | break; | |
1060 | } | |
1061 | for (block = gimple_block (stmt); | |
1062 | !any_addressable_vars | |
1063 | && block | |
1064 | && TREE_CODE (block) == BLOCK; | |
1065 | block = BLOCK_SUPERCONTEXT (block)) | |
1066 | { | |
1067 | for (local_decls = BLOCK_VARS (block); | |
1068 | local_decls; | |
1069 | local_decls = DECL_CHAIN (local_decls)) | |
1070 | if (TREE_ADDRESSABLE (local_decls)) | |
1071 | { | |
1072 | any_addressable_vars = 1; | |
1073 | break; | |
1074 | } | |
1075 | if (block == gimple_block (parallel_stmt)) | |
1076 | break; | |
1077 | } | |
1078 | } | |
1079 | if (!any_addressable_vars) | |
1080 | gimple_omp_return_set_nowait (stmt); | |
1081 | } | |
1082 | } | |
1083 | } | |
1084 | ||
1085 | static void | |
1086 | remove_exit_barriers (struct omp_region *region) | |
1087 | { | |
1088 | if (region->type == GIMPLE_OMP_PARALLEL) | |
1089 | remove_exit_barrier (region); | |
1090 | ||
1091 | if (region->inner) | |
1092 | { | |
1093 | region = region->inner; | |
1094 | remove_exit_barriers (region); | |
1095 | while (region->next) | |
1096 | { | |
1097 | region = region->next; | |
1098 | remove_exit_barriers (region); | |
1099 | } | |
1100 | } | |
1101 | } | |
1102 | ||
1103 | /* Optimize omp_get_thread_num () and omp_get_num_threads () | |
1104 | calls. These can't be declared as const functions, but | |
1105 | within one parallel body they are constant, so they can be | |
1106 | transformed there into __builtin_omp_get_{thread_num,num_threads} () | |
1107 | which are declared const. Similarly for task body, except | |
1108 | that in untied task omp_get_thread_num () can change at any task | |
1109 | scheduling point. */ | |
1110 | ||
1111 | static void | |
1112 | optimize_omp_library_calls (gimple *entry_stmt) | |
1113 | { | |
1114 | basic_block bb; | |
1115 | gimple_stmt_iterator gsi; | |
1116 | tree thr_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
1117 | tree thr_num_id = DECL_ASSEMBLER_NAME (thr_num_tree); | |
1118 | tree num_thr_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
1119 | tree num_thr_id = DECL_ASSEMBLER_NAME (num_thr_tree); | |
1120 | bool untied_task = (gimple_code (entry_stmt) == GIMPLE_OMP_TASK | |
1121 | && omp_find_clause (gimple_omp_task_clauses (entry_stmt), | |
1122 | OMP_CLAUSE_UNTIED) != NULL); | |
1123 | ||
1124 | FOR_EACH_BB_FN (bb, cfun) | |
1125 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1126 | { | |
1127 | gimple *call = gsi_stmt (gsi); | |
1128 | tree decl; | |
1129 | ||
1130 | if (is_gimple_call (call) | |
1131 | && (decl = gimple_call_fndecl (call)) | |
1132 | && DECL_EXTERNAL (decl) | |
1133 | && TREE_PUBLIC (decl) | |
1134 | && DECL_INITIAL (decl) == NULL) | |
1135 | { | |
1136 | tree built_in; | |
1137 | ||
1138 | if (DECL_NAME (decl) == thr_num_id) | |
1139 | { | |
1140 | /* In #pragma omp task untied omp_get_thread_num () can change | |
1141 | during the execution of the task region. */ | |
1142 | if (untied_task) | |
1143 | continue; | |
1144 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
1145 | } | |
1146 | else if (DECL_NAME (decl) == num_thr_id) | |
1147 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
1148 | else | |
1149 | continue; | |
1150 | ||
1151 | if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in) | |
1152 | || gimple_call_num_args (call) != 0) | |
1153 | continue; | |
1154 | ||
1155 | if (flag_exceptions && !TREE_NOTHROW (decl)) | |
1156 | continue; | |
1157 | ||
1158 | if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE | |
1159 | || !types_compatible_p (TREE_TYPE (TREE_TYPE (decl)), | |
1160 | TREE_TYPE (TREE_TYPE (built_in)))) | |
1161 | continue; | |
1162 | ||
1163 | gimple_call_set_fndecl (call, built_in); | |
1164 | } | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | /* Callback for expand_omp_build_assign. Return non-NULL if *tp needs to be | |
1169 | regimplified. */ | |
1170 | ||
1171 | static tree | |
1172 | expand_omp_regimplify_p (tree *tp, int *walk_subtrees, void *) | |
1173 | { | |
1174 | tree t = *tp; | |
1175 | ||
1176 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ | |
1177 | if (VAR_P (t) && DECL_HAS_VALUE_EXPR_P (t)) | |
1178 | return t; | |
1179 | ||
1180 | if (TREE_CODE (t) == ADDR_EXPR) | |
1181 | recompute_tree_invariant_for_addr_expr (t); | |
1182 | ||
1183 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); | |
1184 | return NULL_TREE; | |
1185 | } | |
1186 | ||
1187 | /* Prepend or append TO = FROM assignment before or after *GSI_P. */ | |
1188 | ||
1189 | static void | |
1190 | expand_omp_build_assign (gimple_stmt_iterator *gsi_p, tree to, tree from, | |
1191 | bool after) | |
1192 | { | |
1193 | bool simple_p = DECL_P (to) && TREE_ADDRESSABLE (to); | |
1194 | from = force_gimple_operand_gsi (gsi_p, from, simple_p, NULL_TREE, | |
1195 | !after, after ? GSI_CONTINUE_LINKING | |
1196 | : GSI_SAME_STMT); | |
1197 | gimple *stmt = gimple_build_assign (to, from); | |
1198 | if (after) | |
1199 | gsi_insert_after (gsi_p, stmt, GSI_CONTINUE_LINKING); | |
1200 | else | |
1201 | gsi_insert_before (gsi_p, stmt, GSI_SAME_STMT); | |
1202 | if (walk_tree (&from, expand_omp_regimplify_p, NULL, NULL) | |
1203 | || walk_tree (&to, expand_omp_regimplify_p, NULL, NULL)) | |
1204 | { | |
1205 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1206 | gimple_regimplify_operands (stmt, &gsi); | |
1207 | } | |
1208 | } | |
1209 | ||
1210 | /* Expand the OpenMP parallel or task directive starting at REGION. */ | |
1211 | ||
1212 | static void | |
1213 | expand_omp_taskreg (struct omp_region *region) | |
1214 | { | |
1215 | basic_block entry_bb, exit_bb, new_bb; | |
1216 | struct function *child_cfun; | |
1217 | tree child_fn, block, t; | |
1218 | gimple_stmt_iterator gsi; | |
1219 | gimple *entry_stmt, *stmt; | |
1220 | edge e; | |
1221 | vec<tree, va_gc> *ws_args; | |
1222 | ||
1223 | entry_stmt = last_stmt (region->entry); | |
1224 | if (gimple_code (entry_stmt) == GIMPLE_OMP_TASK | |
1225 | && gimple_omp_task_taskwait_p (entry_stmt)) | |
1226 | { | |
1227 | new_bb = region->entry; | |
1228 | gsi = gsi_last_nondebug_bb (region->entry); | |
1229 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK); | |
1230 | gsi_remove (&gsi, true); | |
1231 | expand_taskwait_call (new_bb, as_a <gomp_task *> (entry_stmt)); | |
1232 | return; | |
1233 | } | |
1234 | ||
1235 | child_fn = gimple_omp_taskreg_child_fn (entry_stmt); | |
1236 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
1237 | ||
1238 | entry_bb = region->entry; | |
1239 | if (gimple_code (entry_stmt) == GIMPLE_OMP_TASK) | |
1240 | exit_bb = region->cont; | |
1241 | else | |
1242 | exit_bb = region->exit; | |
1243 | ||
1244 | if (is_combined_parallel (region)) | |
1245 | ws_args = region->ws_args; | |
1246 | else | |
1247 | ws_args = NULL; | |
1248 | ||
1249 | if (child_cfun->cfg) | |
1250 | { | |
1251 | /* Due to inlining, it may happen that we have already outlined | |
1252 | the region, in which case all we need to do is make the | |
1253 | sub-graph unreachable and emit the parallel call. */ | |
1254 | edge entry_succ_e, exit_succ_e; | |
1255 | ||
1256 | entry_succ_e = single_succ_edge (entry_bb); | |
1257 | ||
1258 | gsi = gsi_last_nondebug_bb (entry_bb); | |
1259 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_PARALLEL | |
1260 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK | |
1261 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TEAMS); | |
1262 | gsi_remove (&gsi, true); | |
1263 | ||
1264 | new_bb = entry_bb; | |
1265 | if (exit_bb) | |
1266 | { | |
1267 | exit_succ_e = single_succ_edge (exit_bb); | |
1268 | make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU); | |
1269 | } | |
1270 | remove_edge_and_dominated_blocks (entry_succ_e); | |
1271 | } | |
1272 | else | |
1273 | { | |
1274 | unsigned srcidx, dstidx, num; | |
1275 | ||
1276 | /* If the parallel region needs data sent from the parent | |
1277 | function, then the very first statement (except possible | |
1278 | tree profile counter updates) of the parallel body | |
1279 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
1280 | &.OMP_DATA_O is passed as an argument to the child function, | |
1281 | we need to replace it with the argument as seen by the child | |
1282 | function. | |
1283 | ||
1284 | In most cases, this will end up being the identity assignment | |
1285 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had | |
1286 | a function call that has been inlined, the original PARM_DECL | |
1287 | .OMP_DATA_I may have been converted into a different local | |
1288 | variable. In which case, we need to keep the assignment. */ | |
1289 | if (gimple_omp_taskreg_data_arg (entry_stmt)) | |
1290 | { | |
1291 | basic_block entry_succ_bb | |
1292 | = single_succ_p (entry_bb) ? single_succ (entry_bb) | |
1293 | : FALLTHRU_EDGE (entry_bb)->dest; | |
1294 | tree arg; | |
1295 | gimple *parcopy_stmt = NULL; | |
1296 | ||
1297 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
1298 | { | |
1299 | gimple *stmt; | |
1300 | ||
1301 | gcc_assert (!gsi_end_p (gsi)); | |
1302 | stmt = gsi_stmt (gsi); | |
1303 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
1304 | continue; | |
1305 | ||
1306 | if (gimple_num_ops (stmt) == 2) | |
1307 | { | |
1308 | tree arg = gimple_assign_rhs1 (stmt); | |
1309 | ||
1310 | /* We're ignore the subcode because we're | |
1311 | effectively doing a STRIP_NOPS. */ | |
1312 | ||
1313 | if (TREE_CODE (arg) == ADDR_EXPR | |
1314 | && (TREE_OPERAND (arg, 0) | |
1315 | == gimple_omp_taskreg_data_arg (entry_stmt))) | |
1316 | { | |
1317 | parcopy_stmt = stmt; | |
1318 | break; | |
1319 | } | |
1320 | } | |
1321 | } | |
1322 | ||
1323 | gcc_assert (parcopy_stmt != NULL); | |
1324 | arg = DECL_ARGUMENTS (child_fn); | |
1325 | ||
1326 | if (!gimple_in_ssa_p (cfun)) | |
1327 | { | |
1328 | if (gimple_assign_lhs (parcopy_stmt) == arg) | |
1329 | gsi_remove (&gsi, true); | |
1330 | else | |
1331 | { | |
1332 | /* ?? Is setting the subcode really necessary ?? */ | |
1333 | gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (arg)); | |
1334 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
1335 | } | |
1336 | } | |
1337 | else | |
1338 | { | |
1339 | tree lhs = gimple_assign_lhs (parcopy_stmt); | |
1340 | gcc_assert (SSA_NAME_VAR (lhs) == arg); | |
1341 | /* We'd like to set the rhs to the default def in the child_fn, | |
1342 | but it's too early to create ssa names in the child_fn. | |
1343 | Instead, we set the rhs to the parm. In | |
1344 | move_sese_region_to_fn, we introduce a default def for the | |
1345 | parm, map the parm to it's default def, and once we encounter | |
1346 | this stmt, replace the parm with the default def. */ | |
1347 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
1348 | update_stmt (parcopy_stmt); | |
1349 | } | |
1350 | } | |
1351 | ||
1352 | /* Declare local variables needed in CHILD_CFUN. */ | |
1353 | block = DECL_INITIAL (child_fn); | |
1354 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
1355 | /* The gimplifier could record temporaries in parallel/task block | |
1356 | rather than in containing function's local_decls chain, | |
1357 | which would mean cgraph missed finalizing them. Do it now. */ | |
1358 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
1359 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) | |
1360 | varpool_node::finalize_decl (t); | |
1361 | DECL_SAVED_TREE (child_fn) = NULL; | |
1362 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
1363 | gimple_set_body (child_fn, NULL); | |
1364 | TREE_USED (block) = 1; | |
1365 | ||
1366 | /* Reset DECL_CONTEXT on function arguments. */ | |
1367 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
1368 | DECL_CONTEXT (t) = child_fn; | |
1369 | ||
1370 | /* Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK, | |
1371 | so that it can be moved to the child function. */ | |
1372 | gsi = gsi_last_nondebug_bb (entry_bb); | |
1373 | stmt = gsi_stmt (gsi); | |
1374 | gcc_assert (stmt && (gimple_code (stmt) == GIMPLE_OMP_PARALLEL | |
1375 | || gimple_code (stmt) == GIMPLE_OMP_TASK | |
1376 | || gimple_code (stmt) == GIMPLE_OMP_TEAMS)); | |
1377 | e = split_block (entry_bb, stmt); | |
1378 | gsi_remove (&gsi, true); | |
1379 | entry_bb = e->dest; | |
1380 | edge e2 = NULL; | |
1381 | if (gimple_code (entry_stmt) != GIMPLE_OMP_TASK) | |
1382 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
1383 | else | |
1384 | { | |
1385 | e2 = make_edge (e->src, BRANCH_EDGE (entry_bb)->dest, EDGE_ABNORMAL); | |
1386 | gcc_assert (e2->dest == region->exit); | |
1387 | remove_edge (BRANCH_EDGE (entry_bb)); | |
1388 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e->src); | |
1389 | gsi = gsi_last_nondebug_bb (region->exit); | |
1390 | gcc_assert (!gsi_end_p (gsi) | |
1391 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
1392 | gsi_remove (&gsi, true); | |
1393 | } | |
1394 | ||
1395 | /* Convert GIMPLE_OMP_{RETURN,CONTINUE} into a RETURN_EXPR. */ | |
1396 | if (exit_bb) | |
1397 | { | |
1398 | gsi = gsi_last_nondebug_bb (exit_bb); | |
1399 | gcc_assert (!gsi_end_p (gsi) | |
1400 | && (gimple_code (gsi_stmt (gsi)) | |
1401 | == (e2 ? GIMPLE_OMP_CONTINUE : GIMPLE_OMP_RETURN))); | |
1402 | stmt = gimple_build_return (NULL); | |
1403 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
1404 | gsi_remove (&gsi, true); | |
1405 | } | |
1406 | ||
1407 | /* Move the parallel region into CHILD_CFUN. */ | |
1408 | ||
1409 | if (gimple_in_ssa_p (cfun)) | |
1410 | { | |
1411 | init_tree_ssa (child_cfun); | |
1412 | init_ssa_operands (child_cfun); | |
1413 | child_cfun->gimple_df->in_ssa_p = true; | |
1414 | block = NULL_TREE; | |
1415 | } | |
1416 | else | |
1417 | block = gimple_block (entry_stmt); | |
1418 | ||
1419 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
1420 | if (exit_bb) | |
1421 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
1422 | if (e2) | |
1423 | { | |
1424 | basic_block dest_bb = e2->dest; | |
1425 | if (!exit_bb) | |
1426 | make_edge (new_bb, dest_bb, EDGE_FALLTHRU); | |
1427 | remove_edge (e2); | |
1428 | set_immediate_dominator (CDI_DOMINATORS, dest_bb, new_bb); | |
1429 | } | |
1430 | /* When the OMP expansion process cannot guarantee an up-to-date | |
1431 | loop tree arrange for the child function to fixup loops. */ | |
1432 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
1433 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
1434 | ||
1435 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
1436 | num = vec_safe_length (child_cfun->local_decls); | |
1437 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
1438 | { | |
1439 | t = (*child_cfun->local_decls)[srcidx]; | |
1440 | if (DECL_CONTEXT (t) == cfun->decl) | |
1441 | continue; | |
1442 | if (srcidx != dstidx) | |
1443 | (*child_cfun->local_decls)[dstidx] = t; | |
1444 | dstidx++; | |
1445 | } | |
1446 | if (dstidx != num) | |
1447 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
1448 | ||
1449 | /* Inform the callgraph about the new function. */ | |
1450 | child_cfun->curr_properties = cfun->curr_properties; | |
1451 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; | |
1452 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; | |
1453 | cgraph_node *node = cgraph_node::get_create (child_fn); | |
1454 | node->parallelized_function = 1; | |
1455 | cgraph_node::add_new_function (child_fn, true); | |
1456 | ||
1457 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) | |
1458 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); | |
1459 | ||
1460 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
1461 | fixed in a following pass. */ | |
1462 | push_cfun (child_cfun); | |
1463 | if (need_asm) | |
1464 | assign_assembler_name_if_needed (child_fn); | |
1465 | ||
1466 | if (optimize) | |
1467 | optimize_omp_library_calls (entry_stmt); | |
1468 | update_max_bb_count (); | |
1469 | cgraph_edge::rebuild_edges (); | |
1470 | ||
1471 | /* Some EH regions might become dead, see PR34608. If | |
1472 | pass_cleanup_cfg isn't the first pass to happen with the | |
1473 | new child, these dead EH edges might cause problems. | |
1474 | Clean them up now. */ | |
1475 | if (flag_exceptions) | |
1476 | { | |
1477 | basic_block bb; | |
1478 | bool changed = false; | |
1479 | ||
1480 | FOR_EACH_BB_FN (bb, cfun) | |
1481 | changed |= gimple_purge_dead_eh_edges (bb); | |
1482 | if (changed) | |
1483 | cleanup_tree_cfg (); | |
1484 | } | |
1485 | if (gimple_in_ssa_p (cfun)) | |
1486 | update_ssa (TODO_update_ssa); | |
1487 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
1488 | verify_loop_structure (); | |
1489 | pop_cfun (); | |
1490 | ||
1491 | if (dump_file && !gimple_in_ssa_p (cfun)) | |
1492 | { | |
1493 | omp_any_child_fn_dumped = true; | |
1494 | dump_function_header (dump_file, child_fn, dump_flags); | |
1495 | dump_function_to_file (child_fn, dump_file, dump_flags); | |
1496 | } | |
1497 | } | |
1498 | ||
1499 | adjust_context_and_scope (region, gimple_block (entry_stmt), child_fn); | |
1500 | ||
1501 | if (gimple_code (entry_stmt) == GIMPLE_OMP_PARALLEL) | |
1502 | expand_parallel_call (region, new_bb, | |
1503 | as_a <gomp_parallel *> (entry_stmt), ws_args); | |
1504 | else if (gimple_code (entry_stmt) == GIMPLE_OMP_TEAMS) | |
1505 | expand_teams_call (new_bb, as_a <gomp_teams *> (entry_stmt)); | |
1506 | else | |
1507 | expand_task_call (region, new_bb, as_a <gomp_task *> (entry_stmt)); | |
1508 | if (gimple_in_ssa_p (cfun)) | |
1509 | update_ssa (TODO_update_ssa_only_virtuals); | |
1510 | } | |
1511 | ||
1512 | /* Information about members of an OpenACC collapsed loop nest. */ | |
1513 | ||
1514 | struct oacc_collapse | |
1515 | { | |
1516 | tree base; /* Base value. */ | |
1517 | tree iters; /* Number of steps. */ | |
1518 | tree step; /* Step size. */ | |
1519 | tree tile; /* Tile increment (if tiled). */ | |
1520 | tree outer; /* Tile iterator var. */ | |
1521 | }; | |
1522 | ||
1523 | /* Helper for expand_oacc_for. Determine collapsed loop information. | |
1524 | Fill in COUNTS array. Emit any initialization code before GSI. | |
1525 | Return the calculated outer loop bound of BOUND_TYPE. */ | |
1526 | ||
1527 | static tree | |
1528 | expand_oacc_collapse_init (const struct omp_for_data *fd, | |
1529 | gimple_stmt_iterator *gsi, | |
1530 | oacc_collapse *counts, tree diff_type, | |
1531 | tree bound_type, location_t loc) | |
1532 | { | |
1533 | tree tiling = fd->tiling; | |
1534 | tree total = build_int_cst (bound_type, 1); | |
1535 | int ix; | |
1536 | ||
1537 | gcc_assert (integer_onep (fd->loop.step)); | |
1538 | gcc_assert (integer_zerop (fd->loop.n1)); | |
1539 | ||
1540 | /* When tiling, the first operand of the tile clause applies to the | |
1541 | innermost loop, and we work outwards from there. Seems | |
1542 | backwards, but whatever. */ | |
1543 | for (ix = fd->collapse; ix--;) | |
1544 | { | |
1545 | const omp_for_data_loop *loop = &fd->loops[ix]; | |
1546 | ||
1547 | tree iter_type = TREE_TYPE (loop->v); | |
1548 | tree plus_type = iter_type; | |
1549 | ||
1550 | gcc_assert (loop->cond_code == LT_EXPR || loop->cond_code == GT_EXPR); | |
1551 | ||
1552 | if (POINTER_TYPE_P (iter_type)) | |
1553 | plus_type = sizetype; | |
1554 | ||
1555 | if (tiling) | |
1556 | { | |
1557 | tree num = build_int_cst (integer_type_node, fd->collapse); | |
1558 | tree loop_no = build_int_cst (integer_type_node, ix); | |
1559 | tree tile = TREE_VALUE (tiling); | |
1560 | gcall *call | |
1561 | = gimple_build_call_internal (IFN_GOACC_TILE, 5, num, loop_no, tile, | |
1562 | /* gwv-outer=*/integer_zero_node, | |
1563 | /* gwv-inner=*/integer_zero_node); | |
1564 | ||
1565 | counts[ix].outer = create_tmp_var (iter_type, ".outer"); | |
1566 | counts[ix].tile = create_tmp_var (diff_type, ".tile"); | |
1567 | gimple_call_set_lhs (call, counts[ix].tile); | |
1568 | gimple_set_location (call, loc); | |
1569 | gsi_insert_before (gsi, call, GSI_SAME_STMT); | |
1570 | ||
1571 | tiling = TREE_CHAIN (tiling); | |
1572 | } | |
1573 | else | |
1574 | { | |
1575 | counts[ix].tile = NULL; | |
1576 | counts[ix].outer = loop->v; | |
1577 | } | |
1578 | ||
1579 | tree b = loop->n1; | |
1580 | tree e = loop->n2; | |
1581 | tree s = loop->step; | |
1582 | bool up = loop->cond_code == LT_EXPR; | |
1583 | tree dir = build_int_cst (diff_type, up ? +1 : -1); | |
1584 | bool negating; | |
1585 | tree expr; | |
1586 | ||
1587 | b = force_gimple_operand_gsi (gsi, b, true, NULL_TREE, | |
1588 | true, GSI_SAME_STMT); | |
1589 | e = force_gimple_operand_gsi (gsi, e, true, NULL_TREE, | |
1590 | true, GSI_SAME_STMT); | |
1591 | ||
1592 | /* Convert the step, avoiding possible unsigned->signed overflow. */ | |
1593 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); | |
1594 | if (negating) | |
1595 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); | |
1596 | s = fold_convert (diff_type, s); | |
1597 | if (negating) | |
1598 | s = fold_build1 (NEGATE_EXPR, diff_type, s); | |
1599 | s = force_gimple_operand_gsi (gsi, s, true, NULL_TREE, | |
1600 | true, GSI_SAME_STMT); | |
1601 | ||
1602 | /* Determine the range, avoiding possible unsigned->signed overflow. */ | |
1603 | negating = !up && TYPE_UNSIGNED (iter_type); | |
1604 | expr = fold_build2 (MINUS_EXPR, plus_type, | |
1605 | fold_convert (plus_type, negating ? b : e), | |
1606 | fold_convert (plus_type, negating ? e : b)); | |
1607 | expr = fold_convert (diff_type, expr); | |
1608 | if (negating) | |
1609 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); | |
1610 | tree range = force_gimple_operand_gsi | |
1611 | (gsi, expr, true, NULL_TREE, true, GSI_SAME_STMT); | |
1612 | ||
1613 | /* Determine number of iterations. */ | |
1614 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); | |
1615 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); | |
1616 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); | |
1617 | ||
1618 | tree iters = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE, | |
1619 | true, GSI_SAME_STMT); | |
1620 | ||
1621 | counts[ix].base = b; | |
1622 | counts[ix].iters = iters; | |
1623 | counts[ix].step = s; | |
1624 | ||
1625 | total = fold_build2 (MULT_EXPR, bound_type, total, | |
1626 | fold_convert (bound_type, iters)); | |
1627 | } | |
1628 | ||
1629 | return total; | |
1630 | } | |
1631 | ||
1632 | /* Emit initializers for collapsed loop members. INNER is true if | |
1633 | this is for the element loop of a TILE. IVAR is the outer | |
1634 | loop iteration variable, from which collapsed loop iteration values | |
1635 | are calculated. COUNTS array has been initialized by | |
1636 | expand_oacc_collapse_inits. */ | |
1637 | ||
1638 | static void | |
1639 | expand_oacc_collapse_vars (const struct omp_for_data *fd, bool inner, | |
1640 | gimple_stmt_iterator *gsi, | |
1641 | const oacc_collapse *counts, tree ivar, | |
1642 | tree diff_type) | |
1643 | { | |
1644 | tree ivar_type = TREE_TYPE (ivar); | |
1645 | ||
1646 | /* The most rapidly changing iteration variable is the innermost | |
1647 | one. */ | |
1648 | for (int ix = fd->collapse; ix--;) | |
1649 | { | |
1650 | const omp_for_data_loop *loop = &fd->loops[ix]; | |
1651 | const oacc_collapse *collapse = &counts[ix]; | |
1652 | tree v = inner ? loop->v : collapse->outer; | |
1653 | tree iter_type = TREE_TYPE (v); | |
1654 | tree plus_type = iter_type; | |
1655 | enum tree_code plus_code = PLUS_EXPR; | |
1656 | tree expr; | |
1657 | ||
1658 | if (POINTER_TYPE_P (iter_type)) | |
1659 | { | |
1660 | plus_code = POINTER_PLUS_EXPR; | |
1661 | plus_type = sizetype; | |
1662 | } | |
1663 | ||
1664 | expr = ivar; | |
1665 | if (ix) | |
1666 | { | |
1667 | tree mod = fold_convert (ivar_type, collapse->iters); | |
1668 | ivar = fold_build2 (TRUNC_DIV_EXPR, ivar_type, expr, mod); | |
1669 | expr = fold_build2 (TRUNC_MOD_EXPR, ivar_type, expr, mod); | |
1670 | ivar = force_gimple_operand_gsi (gsi, ivar, true, NULL_TREE, | |
1671 | true, GSI_SAME_STMT); | |
1672 | } | |
1673 | ||
1674 | expr = fold_build2 (MULT_EXPR, diff_type, fold_convert (diff_type, expr), | |
1675 | fold_convert (diff_type, collapse->step)); | |
1676 | expr = fold_build2 (plus_code, iter_type, | |
1677 | inner ? collapse->outer : collapse->base, | |
1678 | fold_convert (plus_type, expr)); | |
1679 | expr = force_gimple_operand_gsi (gsi, expr, false, NULL_TREE, | |
1680 | true, GSI_SAME_STMT); | |
1681 | gassign *ass = gimple_build_assign (v, expr); | |
1682 | gsi_insert_before (gsi, ass, GSI_SAME_STMT); | |
1683 | } | |
1684 | } | |
1685 | ||
1686 | /* Helper function for expand_omp_{for_*,simd}. If this is the outermost | |
1687 | of the combined collapse > 1 loop constructs, generate code like: | |
1688 | if (__builtin_expect (N32 cond3 N31, 0)) goto ZERO_ITER_BB; | |
1689 | if (cond3 is <) | |
1690 | adj = STEP3 - 1; | |
1691 | else | |
1692 | adj = STEP3 + 1; | |
1693 | count3 = (adj + N32 - N31) / STEP3; | |
1694 | if (__builtin_expect (N22 cond2 N21, 0)) goto ZERO_ITER_BB; | |
1695 | if (cond2 is <) | |
1696 | adj = STEP2 - 1; | |
1697 | else | |
1698 | adj = STEP2 + 1; | |
1699 | count2 = (adj + N22 - N21) / STEP2; | |
1700 | if (__builtin_expect (N12 cond1 N11, 0)) goto ZERO_ITER_BB; | |
1701 | if (cond1 is <) | |
1702 | adj = STEP1 - 1; | |
1703 | else | |
1704 | adj = STEP1 + 1; | |
1705 | count1 = (adj + N12 - N11) / STEP1; | |
1706 | count = count1 * count2 * count3; | |
1707 | Furthermore, if ZERO_ITER_BB is NULL, create a BB which does: | |
1708 | count = 0; | |
1709 | and set ZERO_ITER_BB to that bb. If this isn't the outermost | |
1710 | of the combined loop constructs, just initialize COUNTS array | |
1711 | from the _looptemp_ clauses. For loop nests with non-rectangular | |
1712 | loops, do this only for the rectangular loops. Then pick | |
1713 | the loops which reference outer vars in their bound expressions | |
1714 | and the loops which they refer to and for this sub-nest compute | |
1715 | number of iterations. For triangular loops use Faulhaber's formula, | |
1716 | otherwise as a fallback, compute by iterating the loops. | |
1717 | If e.g. the sub-nest is | |
1718 | for (I = N11; I COND1 N12; I += STEP1) | |
1719 | for (J = M21 * I + N21; J COND2 M22 * I + N22; J += STEP2) | |
1720 | for (K = M31 * J + N31; K COND3 M32 * J + N32; K += STEP3) | |
1721 | do: | |
1722 | COUNT = 0; | |
1723 | for (tmpi = N11; tmpi COND1 N12; tmpi += STEP1) | |
1724 | for (tmpj = M21 * tmpi + N21; | |
1725 | tmpj COND2 M22 * tmpi + N22; tmpj += STEP2) | |
1726 | { | |
1727 | int tmpk1 = M31 * tmpj + N31; | |
1728 | int tmpk2 = M32 * tmpj + N32; | |
1729 | if (tmpk1 COND3 tmpk2) | |
1730 | { | |
1731 | if (COND3 is <) | |
1732 | adj = STEP3 - 1; | |
1733 | else | |
1734 | adj = STEP3 + 1; | |
1735 | COUNT += (adj + tmpk2 - tmpk1) / STEP3; | |
1736 | } | |
1737 | } | |
1738 | and finally multiply the counts of the rectangular loops not | |
1739 | in the sub-nest with COUNT. Also, as counts[fd->last_nonrect] | |
1740 | store number of iterations of the loops from fd->first_nonrect | |
1741 | to fd->last_nonrect inclusive, i.e. the above COUNT multiplied | |
1742 | by the counts of rectangular loops not referenced in any non-rectangular | |
1743 | loops sandwitched in between those. */ | |
1744 | ||
1745 | /* NOTE: It *could* be better to moosh all of the BBs together, | |
1746 | creating one larger BB with all the computation and the unexpected | |
1747 | jump at the end. I.e. | |
1748 | ||
1749 | bool zero3, zero2, zero1, zero; | |
1750 | ||
1751 | zero3 = N32 c3 N31; | |
1752 | count3 = (N32 - N31) /[cl] STEP3; | |
1753 | zero2 = N22 c2 N21; | |
1754 | count2 = (N22 - N21) /[cl] STEP2; | |
1755 | zero1 = N12 c1 N11; | |
1756 | count1 = (N12 - N11) /[cl] STEP1; | |
1757 | zero = zero3 || zero2 || zero1; | |
1758 | count = count1 * count2 * count3; | |
1759 | if (__builtin_expect(zero, false)) goto zero_iter_bb; | |
1760 | ||
1761 | After all, we expect the zero=false, and thus we expect to have to | |
1762 | evaluate all of the comparison expressions, so short-circuiting | |
1763 | oughtn't be a win. Since the condition isn't protecting a | |
1764 | denominator, we're not concerned about divide-by-zero, so we can | |
1765 | fully evaluate count even if a numerator turned out to be wrong. | |
1766 | ||
1767 | It seems like putting this all together would create much better | |
1768 | scheduling opportunities, and less pressure on the chip's branch | |
1769 | predictor. */ | |
1770 | ||
1771 | static void | |
1772 | expand_omp_for_init_counts (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
1773 | basic_block &entry_bb, tree *counts, | |
1774 | basic_block &zero_iter1_bb, int &first_zero_iter1, | |
1775 | basic_block &zero_iter2_bb, int &first_zero_iter2, | |
1776 | basic_block &l2_dom_bb) | |
1777 | { | |
1778 | tree t, type = TREE_TYPE (fd->loop.v); | |
1779 | edge e, ne; | |
1780 | int i; | |
1781 | ||
1782 | /* Collapsed loops need work for expansion into SSA form. */ | |
1783 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
1784 | ||
1785 | if (gimple_omp_for_combined_into_p (fd->for_stmt) | |
1786 | && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
1787 | { | |
1788 | gcc_assert (fd->ordered == 0); | |
1789 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
1790 | isn't supposed to be handled, as the inner loop doesn't | |
1791 | use it. */ | |
1792 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
1793 | OMP_CLAUSE__LOOPTEMP_); | |
1794 | gcc_assert (innerc); | |
1795 | for (i = 0; i < fd->collapse; i++) | |
1796 | { | |
1797 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
1798 | OMP_CLAUSE__LOOPTEMP_); | |
1799 | gcc_assert (innerc); | |
1800 | if (i) | |
1801 | counts[i] = OMP_CLAUSE_DECL (innerc); | |
1802 | else | |
1803 | counts[0] = NULL_TREE; | |
1804 | } | |
1805 | if (fd->non_rect | |
1806 | && fd->last_nonrect == fd->first_nonrect + 1 | |
1807 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) | |
1808 | { | |
1809 | tree c[4]; | |
1810 | for (i = 0; i < 4; i++) | |
1811 | { | |
1812 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
1813 | OMP_CLAUSE__LOOPTEMP_); | |
1814 | gcc_assert (innerc); | |
1815 | c[i] = OMP_CLAUSE_DECL (innerc); | |
1816 | } | |
1817 | counts[0] = c[0]; | |
1818 | fd->first_inner_iterations = c[1]; | |
1819 | fd->factor = c[2]; | |
1820 | fd->adjn1 = c[3]; | |
1821 | } | |
1822 | return; | |
1823 | } | |
1824 | ||
1825 | for (i = fd->collapse; i < fd->ordered; i++) | |
1826 | { | |
1827 | tree itype = TREE_TYPE (fd->loops[i].v); | |
1828 | counts[i] = NULL_TREE; | |
1829 | t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
1830 | fold_convert (itype, fd->loops[i].n1), | |
1831 | fold_convert (itype, fd->loops[i].n2)); | |
1832 | if (t && integer_zerop (t)) | |
1833 | { | |
1834 | for (i = fd->collapse; i < fd->ordered; i++) | |
1835 | counts[i] = build_int_cst (type, 0); | |
1836 | break; | |
1837 | } | |
1838 | } | |
1839 | bool rect_count_seen = false; | |
1840 | for (i = 0; i < (fd->ordered ? fd->ordered : fd->collapse); i++) | |
1841 | { | |
1842 | tree itype = TREE_TYPE (fd->loops[i].v); | |
1843 | ||
1844 | if (i >= fd->collapse && counts[i]) | |
1845 | continue; | |
1846 | if (fd->non_rect) | |
1847 | { | |
1848 | /* Skip loops that use outer iterators in their expressions | |
1849 | during this phase. */ | |
1850 | if (fd->loops[i].m1 || fd->loops[i].m2) | |
1851 | { | |
1852 | counts[i] = build_zero_cst (type); | |
1853 | continue; | |
1854 | } | |
1855 | } | |
1856 | if ((SSA_VAR_P (fd->loop.n2) || i >= fd->collapse) | |
1857 | && ((t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
1858 | fold_convert (itype, fd->loops[i].n1), | |
1859 | fold_convert (itype, fd->loops[i].n2))) | |
1860 | == NULL_TREE || !integer_onep (t))) | |
1861 | { | |
1862 | gcond *cond_stmt; | |
1863 | tree n1, n2; | |
1864 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
1865 | n1 = force_gimple_operand_gsi (gsi, n1, true, NULL_TREE, | |
1866 | true, GSI_SAME_STMT); | |
1867 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
1868 | n2 = force_gimple_operand_gsi (gsi, n2, true, NULL_TREE, | |
1869 | true, GSI_SAME_STMT); | |
1870 | cond_stmt = gimple_build_cond (fd->loops[i].cond_code, n1, n2, | |
1871 | NULL_TREE, NULL_TREE); | |
1872 | gsi_insert_before (gsi, cond_stmt, GSI_SAME_STMT); | |
1873 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
1874 | expand_omp_regimplify_p, NULL, NULL) | |
1875 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
1876 | expand_omp_regimplify_p, NULL, NULL)) | |
1877 | { | |
1878 | *gsi = gsi_for_stmt (cond_stmt); | |
1879 | gimple_regimplify_operands (cond_stmt, gsi); | |
1880 | } | |
1881 | e = split_block (entry_bb, cond_stmt); | |
1882 | basic_block &zero_iter_bb | |
1883 | = i < fd->collapse ? zero_iter1_bb : zero_iter2_bb; | |
1884 | int &first_zero_iter | |
1885 | = i < fd->collapse ? first_zero_iter1 : first_zero_iter2; | |
1886 | if (zero_iter_bb == NULL) | |
1887 | { | |
1888 | gassign *assign_stmt; | |
1889 | first_zero_iter = i; | |
1890 | zero_iter_bb = create_empty_bb (entry_bb); | |
1891 | add_bb_to_loop (zero_iter_bb, entry_bb->loop_father); | |
1892 | *gsi = gsi_after_labels (zero_iter_bb); | |
1893 | if (i < fd->collapse) | |
1894 | assign_stmt = gimple_build_assign (fd->loop.n2, | |
1895 | build_zero_cst (type)); | |
1896 | else | |
1897 | { | |
1898 | counts[i] = create_tmp_reg (type, ".count"); | |
1899 | assign_stmt | |
1900 | = gimple_build_assign (counts[i], build_zero_cst (type)); | |
1901 | } | |
1902 | gsi_insert_before (gsi, assign_stmt, GSI_SAME_STMT); | |
1903 | set_immediate_dominator (CDI_DOMINATORS, zero_iter_bb, | |
1904 | entry_bb); | |
1905 | } | |
1906 | ne = make_edge (entry_bb, zero_iter_bb, EDGE_FALSE_VALUE); | |
1907 | ne->probability = profile_probability::very_unlikely (); | |
1908 | e->flags = EDGE_TRUE_VALUE; | |
1909 | e->probability = ne->probability.invert (); | |
1910 | if (l2_dom_bb == NULL) | |
1911 | l2_dom_bb = entry_bb; | |
1912 | entry_bb = e->dest; | |
1913 | *gsi = gsi_last_nondebug_bb (entry_bb); | |
1914 | } | |
1915 | ||
1916 | if (POINTER_TYPE_P (itype)) | |
1917 | itype = signed_type_for (itype); | |
1918 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
1919 | ? -1 : 1)); | |
1920 | t = fold_build2 (PLUS_EXPR, itype, | |
1921 | fold_convert (itype, fd->loops[i].step), t); | |
1922 | t = fold_build2 (PLUS_EXPR, itype, t, | |
1923 | fold_convert (itype, fd->loops[i].n2)); | |
1924 | t = fold_build2 (MINUS_EXPR, itype, t, | |
1925 | fold_convert (itype, fd->loops[i].n1)); | |
1926 | /* ?? We could probably use CEIL_DIV_EXPR instead of | |
1927 | TRUNC_DIV_EXPR and adjusting by hand. Unless we can't | |
1928 | generate the same code in the end because generically we | |
1929 | don't know that the values involved must be negative for | |
1930 | GT?? */ | |
1931 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
1932 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
1933 | fold_build1 (NEGATE_EXPR, itype, t), | |
1934 | fold_build1 (NEGATE_EXPR, itype, | |
1935 | fold_convert (itype, | |
1936 | fd->loops[i].step))); | |
1937 | else | |
1938 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
1939 | fold_convert (itype, fd->loops[i].step)); | |
1940 | t = fold_convert (type, t); | |
1941 | if (TREE_CODE (t) == INTEGER_CST) | |
1942 | counts[i] = t; | |
1943 | else | |
1944 | { | |
1945 | if (i < fd->collapse || i != first_zero_iter2) | |
1946 | counts[i] = create_tmp_reg (type, ".count"); | |
1947 | expand_omp_build_assign (gsi, counts[i], t); | |
1948 | } | |
1949 | if (SSA_VAR_P (fd->loop.n2) && i < fd->collapse) | |
1950 | { | |
1951 | if (fd->non_rect && i >= fd->first_nonrect && i <= fd->last_nonrect) | |
1952 | continue; | |
1953 | if (!rect_count_seen) | |
1954 | { | |
1955 | t = counts[i]; | |
1956 | rect_count_seen = true; | |
1957 | } | |
1958 | else | |
1959 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, counts[i]); | |
1960 | expand_omp_build_assign (gsi, fd->loop.n2, t); | |
1961 | } | |
1962 | } | |
1963 | if (fd->non_rect && SSA_VAR_P (fd->loop.n2)) | |
1964 | { | |
1965 | gcc_assert (fd->last_nonrect != -1); | |
1966 | ||
1967 | counts[fd->last_nonrect] = create_tmp_reg (type, ".count"); | |
1968 | expand_omp_build_assign (gsi, counts[fd->last_nonrect], | |
1969 | build_zero_cst (type)); | |
1970 | for (i = fd->first_nonrect + 1; i < fd->last_nonrect; i++) | |
1971 | if (fd->loops[i].m1 | |
1972 | || fd->loops[i].m2 | |
1973 | || fd->loops[i].non_rect_referenced) | |
1974 | break; | |
1975 | if (i == fd->last_nonrect | |
1976 | && fd->loops[i].outer == fd->last_nonrect - fd->first_nonrect | |
1977 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[i].v))) | |
1978 | { | |
1979 | int o = fd->first_nonrect; | |
1980 | tree itype = TREE_TYPE (fd->loops[o].v); | |
1981 | tree n1o = create_tmp_reg (itype, ".n1o"); | |
1982 | t = fold_convert (itype, unshare_expr (fd->loops[o].n1)); | |
1983 | expand_omp_build_assign (gsi, n1o, t); | |
1984 | tree n2o = create_tmp_reg (itype, ".n2o"); | |
1985 | t = fold_convert (itype, unshare_expr (fd->loops[o].n2)); | |
1986 | expand_omp_build_assign (gsi, n2o, t); | |
1987 | if (fd->loops[i].m1 && fd->loops[i].m2) | |
1988 | t = fold_build2 (MINUS_EXPR, itype, unshare_expr (fd->loops[i].m2), | |
1989 | unshare_expr (fd->loops[i].m1)); | |
1990 | else if (fd->loops[i].m1) | |
1991 | t = fold_unary (NEGATE_EXPR, itype, | |
1992 | unshare_expr (fd->loops[i].m1)); | |
1993 | else | |
1994 | t = unshare_expr (fd->loops[i].m2); | |
1995 | tree m2minusm1 | |
1996 | = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, | |
1997 | true, GSI_SAME_STMT); | |
1998 | ||
1999 | gimple_stmt_iterator gsi2 = *gsi; | |
2000 | gsi_prev (&gsi2); | |
2001 | e = split_block (entry_bb, gsi_stmt (gsi2)); | |
2002 | e = split_block (e->dest, (gimple *) NULL); | |
2003 | basic_block bb1 = e->src; | |
2004 | entry_bb = e->dest; | |
2005 | *gsi = gsi_after_labels (entry_bb); | |
2006 | ||
2007 | gsi2 = gsi_after_labels (bb1); | |
2008 | tree ostep = fold_convert (itype, fd->loops[o].step); | |
2009 | t = build_int_cst (itype, (fd->loops[o].cond_code | |
2010 | == LT_EXPR ? -1 : 1)); | |
2011 | t = fold_build2 (PLUS_EXPR, itype, ostep, t); | |
2012 | t = fold_build2 (PLUS_EXPR, itype, t, n2o); | |
2013 | t = fold_build2 (MINUS_EXPR, itype, t, n1o); | |
2014 | if (TYPE_UNSIGNED (itype) | |
2015 | && fd->loops[o].cond_code == GT_EXPR) | |
2016 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2017 | fold_build1 (NEGATE_EXPR, itype, t), | |
2018 | fold_build1 (NEGATE_EXPR, itype, ostep)); | |
2019 | else | |
2020 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, ostep); | |
2021 | tree outer_niters | |
2022 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2023 | true, GSI_SAME_STMT); | |
2024 | t = fold_build2 (MINUS_EXPR, itype, outer_niters, | |
2025 | build_one_cst (itype)); | |
2026 | t = fold_build2 (MULT_EXPR, itype, t, ostep); | |
2027 | t = fold_build2 (PLUS_EXPR, itype, n1o, t); | |
2028 | tree last = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2029 | true, GSI_SAME_STMT); | |
2030 | tree n1, n2, n1e, n2e; | |
2031 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
2032 | if (fd->loops[i].m1) | |
2033 | { | |
2034 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); | |
2035 | n1 = fold_build2 (MULT_EXPR, itype, n1o, n1); | |
2036 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2037 | } | |
2038 | else | |
2039 | n1 = t; | |
2040 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, | |
2041 | true, GSI_SAME_STMT); | |
2042 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
2043 | if (fd->loops[i].m2) | |
2044 | { | |
2045 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2046 | n2 = fold_build2 (MULT_EXPR, itype, n1o, n2); | |
2047 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); | |
2048 | } | |
2049 | else | |
2050 | n2 = t; | |
2051 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, | |
2052 | true, GSI_SAME_STMT); | |
2053 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
2054 | if (fd->loops[i].m1) | |
2055 | { | |
2056 | n1e = fold_convert (itype, unshare_expr (fd->loops[i].m1)); | |
2057 | n1e = fold_build2 (MULT_EXPR, itype, last, n1e); | |
2058 | n1e = fold_build2 (PLUS_EXPR, itype, n1e, t); | |
2059 | } | |
2060 | else | |
2061 | n1e = t; | |
2062 | n1e = force_gimple_operand_gsi (&gsi2, n1e, true, NULL_TREE, | |
2063 | true, GSI_SAME_STMT); | |
2064 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
2065 | if (fd->loops[i].m2) | |
2066 | { | |
2067 | n2e = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2068 | n2e = fold_build2 (MULT_EXPR, itype, last, n2e); | |
2069 | n2e = fold_build2 (PLUS_EXPR, itype, n2e, t); | |
2070 | } | |
2071 | else | |
2072 | n2e = t; | |
2073 | n2e = force_gimple_operand_gsi (&gsi2, n2e, true, NULL_TREE, | |
2074 | true, GSI_SAME_STMT); | |
2075 | gcond *cond_stmt | |
2076 | = gimple_build_cond (fd->loops[i].cond_code, n1, n2, | |
2077 | NULL_TREE, NULL_TREE); | |
2078 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2079 | e = split_block (bb1, cond_stmt); | |
2080 | e->flags = EDGE_TRUE_VALUE; | |
2081 | e->probability = profile_probability::likely ().guessed (); | |
2082 | basic_block bb2 = e->dest; | |
2083 | gsi2 = gsi_after_labels (bb2); | |
2084 | ||
2085 | cond_stmt = gimple_build_cond (fd->loops[i].cond_code, n1e, n2e, | |
2086 | NULL_TREE, NULL_TREE); | |
2087 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2088 | e = split_block (bb2, cond_stmt); | |
2089 | e->flags = EDGE_TRUE_VALUE; | |
2090 | e->probability = profile_probability::likely ().guessed (); | |
2091 | gsi2 = gsi_after_labels (e->dest); | |
2092 | ||
2093 | tree step = fold_convert (itype, fd->loops[i].step); | |
2094 | t = build_int_cst (itype, (fd->loops[i].cond_code | |
2095 | == LT_EXPR ? -1 : 1)); | |
2096 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
2097 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
2098 | t = fold_build2 (MINUS_EXPR, itype, t, n1); | |
2099 | if (TYPE_UNSIGNED (itype) | |
2100 | && fd->loops[i].cond_code == GT_EXPR) | |
2101 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2102 | fold_build1 (NEGATE_EXPR, itype, t), | |
2103 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2104 | else | |
2105 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2106 | tree first_inner_iterations | |
2107 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2108 | true, GSI_SAME_STMT); | |
2109 | t = fold_build2 (MULT_EXPR, itype, m2minusm1, ostep); | |
2110 | if (TYPE_UNSIGNED (itype) | |
2111 | && fd->loops[i].cond_code == GT_EXPR) | |
2112 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2113 | fold_build1 (NEGATE_EXPR, itype, t), | |
2114 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2115 | else | |
2116 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2117 | tree factor | |
2118 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2119 | true, GSI_SAME_STMT); | |
2120 | t = fold_build2 (MINUS_EXPR, itype, outer_niters, | |
2121 | build_one_cst (itype)); | |
2122 | t = fold_build2 (MULT_EXPR, itype, t, outer_niters); | |
2123 | t = fold_build2 (RSHIFT_EXPR, itype, t, integer_one_node); | |
2124 | t = fold_build2 (MULT_EXPR, itype, factor, t); | |
2125 | t = fold_build2 (PLUS_EXPR, itype, | |
2126 | fold_build2 (MULT_EXPR, itype, outer_niters, | |
2127 | first_inner_iterations), t); | |
2128 | expand_omp_build_assign (&gsi2, counts[fd->last_nonrect], | |
2129 | fold_convert (type, t)); | |
2130 | ||
2131 | basic_block bb3 = create_empty_bb (bb1); | |
2132 | add_bb_to_loop (bb3, bb1->loop_father); | |
2133 | ||
2134 | e = make_edge (bb1, bb3, EDGE_FALSE_VALUE); | |
2135 | e->probability = profile_probability::unlikely ().guessed (); | |
2136 | ||
2137 | gsi2 = gsi_after_labels (bb3); | |
2138 | cond_stmt = gimple_build_cond (fd->loops[i].cond_code, n1e, n2e, | |
2139 | NULL_TREE, NULL_TREE); | |
2140 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2141 | e = split_block (bb3, cond_stmt); | |
2142 | e->flags = EDGE_TRUE_VALUE; | |
2143 | e->probability = profile_probability::likely ().guessed (); | |
2144 | basic_block bb4 = e->dest; | |
2145 | ||
2146 | ne = make_edge (bb3, entry_bb, EDGE_FALSE_VALUE); | |
2147 | ne->probability = e->probability.invert (); | |
2148 | ||
2149 | basic_block bb5 = create_empty_bb (bb2); | |
2150 | add_bb_to_loop (bb5, bb2->loop_father); | |
2151 | ||
2152 | ne = make_edge (bb2, bb5, EDGE_FALSE_VALUE); | |
2153 | ne->probability = profile_probability::unlikely ().guessed (); | |
2154 | ||
2155 | for (int j = 0; j < 2; j++) | |
2156 | { | |
2157 | gsi2 = gsi_after_labels (j ? bb5 : bb4); | |
2158 | t = fold_build2 (MINUS_EXPR, itype, | |
2159 | unshare_expr (fd->loops[i].n1), | |
2160 | unshare_expr (fd->loops[i].n2)); | |
2161 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, m2minusm1); | |
2162 | tree tem | |
2163 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2164 | true, GSI_SAME_STMT); | |
2165 | t = fold_build2 (MINUS_EXPR, itype, tem, n1o); | |
2166 | t = fold_build2 (TRUNC_MOD_EXPR, itype, t, ostep); | |
2167 | t = fold_build2 (MINUS_EXPR, itype, tem, t); | |
2168 | tem = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2169 | true, GSI_SAME_STMT); | |
2170 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
2171 | if (fd->loops[i].m1) | |
2172 | { | |
2173 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); | |
2174 | n1 = fold_build2 (MULT_EXPR, itype, tem, n1); | |
2175 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2176 | } | |
2177 | else | |
2178 | n1 = t; | |
2179 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, | |
2180 | true, GSI_SAME_STMT); | |
2181 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
2182 | if (fd->loops[i].m2) | |
2183 | { | |
2184 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2185 | n2 = fold_build2 (MULT_EXPR, itype, tem, n2); | |
2186 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); | |
2187 | } | |
2188 | else | |
2189 | n2 = t; | |
2190 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, | |
2191 | true, GSI_SAME_STMT); | |
2192 | expand_omp_build_assign (&gsi2, j ? n2o : n1o, tem); | |
2193 | ||
2194 | cond_stmt = gimple_build_cond (fd->loops[i].cond_code, n1, n2, | |
2195 | NULL_TREE, NULL_TREE); | |
2196 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2197 | e = split_block (gsi_bb (gsi2), cond_stmt); | |
2198 | e->flags = j ? EDGE_TRUE_VALUE : EDGE_FALSE_VALUE; | |
2199 | e->probability = profile_probability::unlikely ().guessed (); | |
2200 | ne = make_edge (e->src, bb1, | |
2201 | j ? EDGE_FALSE_VALUE : EDGE_TRUE_VALUE); | |
2202 | ne->probability = e->probability.invert (); | |
2203 | gsi2 = gsi_after_labels (e->dest); | |
2204 | ||
2205 | t = fold_build2 (PLUS_EXPR, itype, tem, ostep); | |
2206 | expand_omp_build_assign (&gsi2, j ? n2o : n1o, t); | |
2207 | ||
2208 | make_edge (e->dest, bb1, EDGE_FALLTHRU); | |
2209 | } | |
2210 | ||
2211 | set_immediate_dominator (CDI_DOMINATORS, bb3, bb1); | |
2212 | set_immediate_dominator (CDI_DOMINATORS, bb5, bb2); | |
2213 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, bb1); | |
2214 | ||
2215 | if (fd->first_nonrect + 1 == fd->last_nonrect) | |
2216 | { | |
2217 | fd->first_inner_iterations = first_inner_iterations; | |
2218 | fd->factor = factor; | |
2219 | fd->adjn1 = n1o; | |
2220 | } | |
2221 | } | |
2222 | else | |
2223 | { | |
2224 | /* Fallback implementation. Evaluate the loops with m1/m2 | |
2225 | non-NULL as well as their outer loops at runtime using temporaries | |
2226 | instead of the original iteration variables, and in the | |
2227 | body just bump the counter. */ | |
2228 | gimple_stmt_iterator gsi2 = *gsi; | |
2229 | gsi_prev (&gsi2); | |
2230 | e = split_block (entry_bb, gsi_stmt (gsi2)); | |
2231 | e = split_block (e->dest, (gimple *) NULL); | |
2232 | basic_block cur_bb = e->src; | |
2233 | basic_block next_bb = e->dest; | |
2234 | entry_bb = e->dest; | |
2235 | *gsi = gsi_after_labels (entry_bb); | |
2236 | ||
2237 | tree *vs = XALLOCAVEC (tree, fd->last_nonrect); | |
2238 | memset (vs, 0, fd->last_nonrect * sizeof (tree)); | |
2239 | ||
2240 | for (i = 0; i <= fd->last_nonrect; i++) | |
2241 | { | |
2242 | if (fd->loops[i].m1 == NULL_TREE | |
2243 | && fd->loops[i].m2 == NULL_TREE | |
2244 | && !fd->loops[i].non_rect_referenced) | |
2245 | continue; | |
2246 | ||
2247 | tree itype = TREE_TYPE (fd->loops[i].v); | |
2248 | ||
2249 | gsi2 = gsi_after_labels (cur_bb); | |
2250 | tree n1, n2; | |
2251 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
2252 | if (fd->loops[i].m1) | |
2253 | { | |
2254 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); | |
2255 | n1 = fold_build2 (MULT_EXPR, itype, | |
2256 | vs[i - fd->loops[i].outer], n1); | |
2257 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2258 | } | |
2259 | else | |
2260 | n1 = t; | |
2261 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, | |
2262 | true, GSI_SAME_STMT); | |
2263 | if (i < fd->last_nonrect) | |
2264 | { | |
2265 | vs[i] = create_tmp_reg (itype, ".it"); | |
2266 | expand_omp_build_assign (&gsi2, vs[i], n1); | |
2267 | } | |
2268 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
2269 | if (fd->loops[i].m2) | |
2270 | { | |
2271 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2272 | n2 = fold_build2 (MULT_EXPR, itype, | |
2273 | vs[i - fd->loops[i].outer], n2); | |
2274 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); | |
2275 | } | |
2276 | else | |
2277 | n2 = t; | |
2278 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, | |
2279 | true, GSI_SAME_STMT); | |
2280 | if (i == fd->last_nonrect) | |
2281 | { | |
2282 | gcond *cond_stmt | |
2283 | = gimple_build_cond (fd->loops[i].cond_code, n1, n2, | |
2284 | NULL_TREE, NULL_TREE); | |
2285 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2286 | e = split_block (cur_bb, cond_stmt); | |
2287 | e->flags = EDGE_TRUE_VALUE; | |
2288 | ne = make_edge (cur_bb, next_bb, EDGE_FALSE_VALUE); | |
2289 | e->probability = profile_probability::likely ().guessed (); | |
2290 | ne->probability = e->probability.invert (); | |
2291 | gsi2 = gsi_after_labels (e->dest); | |
2292 | ||
2293 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
2294 | ? -1 : 1)); | |
2295 | t = fold_build2 (PLUS_EXPR, itype, | |
2296 | fold_convert (itype, fd->loops[i].step), t); | |
2297 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
2298 | t = fold_build2 (MINUS_EXPR, itype, t, n1); | |
2299 | tree step = fold_convert (itype, fd->loops[i].step); | |
2300 | if (TYPE_UNSIGNED (itype) | |
2301 | && fd->loops[i].cond_code == GT_EXPR) | |
2302 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2303 | fold_build1 (NEGATE_EXPR, itype, t), | |
2304 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2305 | else | |
2306 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2307 | t = fold_convert (type, t); | |
2308 | t = fold_build2 (PLUS_EXPR, type, | |
2309 | counts[fd->last_nonrect], t); | |
2310 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2311 | true, GSI_SAME_STMT); | |
2312 | expand_omp_build_assign (&gsi2, counts[fd->last_nonrect], t); | |
2313 | e = make_edge (e->dest, next_bb, EDGE_FALLTHRU); | |
2314 | set_immediate_dominator (CDI_DOMINATORS, next_bb, cur_bb); | |
2315 | break; | |
2316 | } | |
2317 | e = split_block (cur_bb, last_stmt (cur_bb)); | |
2318 | ||
2319 | basic_block new_cur_bb = create_empty_bb (cur_bb); | |
2320 | add_bb_to_loop (new_cur_bb, cur_bb->loop_father); | |
2321 | ||
2322 | gsi2 = gsi_after_labels (e->dest); | |
2323 | tree step = fold_convert (itype, | |
2324 | unshare_expr (fd->loops[i].step)); | |
2325 | t = fold_build2 (PLUS_EXPR, itype, vs[i], step); | |
2326 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2327 | true, GSI_SAME_STMT); | |
2328 | expand_omp_build_assign (&gsi2, vs[i], t); | |
2329 | ||
2330 | ne = split_block (e->dest, last_stmt (e->dest)); | |
2331 | gsi2 = gsi_after_labels (ne->dest); | |
2332 | ||
2333 | gcond *cond_stmt | |
2334 | = gimple_build_cond (fd->loops[i].cond_code, vs[i], n2, | |
2335 | NULL_TREE, NULL_TREE); | |
2336 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2337 | edge e3, e4; | |
2338 | if (next_bb == entry_bb) | |
2339 | { | |
2340 | e3 = find_edge (ne->dest, next_bb); | |
2341 | e3->flags = EDGE_FALSE_VALUE; | |
2342 | } | |
2343 | else | |
2344 | e3 = make_edge (ne->dest, next_bb, EDGE_FALSE_VALUE); | |
2345 | e4 = make_edge (ne->dest, new_cur_bb, EDGE_TRUE_VALUE); | |
2346 | e4->probability = profile_probability::likely ().guessed (); | |
2347 | e3->probability = e4->probability.invert (); | |
2348 | basic_block esrc = e->src; | |
2349 | make_edge (e->src, ne->dest, EDGE_FALLTHRU); | |
2350 | cur_bb = new_cur_bb; | |
2351 | basic_block latch_bb = next_bb; | |
2352 | next_bb = e->dest; | |
2353 | remove_edge (e); | |
2354 | set_immediate_dominator (CDI_DOMINATORS, ne->dest, esrc); | |
2355 | set_immediate_dominator (CDI_DOMINATORS, latch_bb, ne->dest); | |
2356 | set_immediate_dominator (CDI_DOMINATORS, cur_bb, ne->dest); | |
2357 | } | |
2358 | } | |
2359 | t = NULL_TREE; | |
2360 | for (i = fd->first_nonrect; i < fd->last_nonrect; i++) | |
2361 | if (!fd->loops[i].non_rect_referenced | |
2362 | && fd->loops[i].m1 == NULL_TREE | |
2363 | && fd->loops[i].m2 == NULL_TREE) | |
2364 | { | |
2365 | if (t == NULL_TREE) | |
2366 | t = counts[i]; | |
2367 | else | |
2368 | t = fold_build2 (MULT_EXPR, type, t, counts[i]); | |
2369 | } | |
2370 | if (t) | |
2371 | { | |
2372 | t = fold_build2 (MULT_EXPR, type, counts[fd->last_nonrect], t); | |
2373 | expand_omp_build_assign (gsi, counts[fd->last_nonrect], t); | |
2374 | } | |
2375 | if (!rect_count_seen) | |
2376 | t = counts[fd->last_nonrect]; | |
2377 | else | |
2378 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, | |
2379 | counts[fd->last_nonrect]); | |
2380 | expand_omp_build_assign (gsi, fd->loop.n2, t); | |
2381 | } | |
2382 | else if (fd->non_rect) | |
2383 | { | |
2384 | tree t = fd->loop.n2; | |
2385 | gcc_assert (TREE_CODE (t) == INTEGER_CST); | |
2386 | int non_rect_referenced = 0, non_rect = 0; | |
2387 | for (i = 0; i < fd->collapse; i++) | |
2388 | { | |
2389 | if ((i < fd->first_nonrect || i > fd->last_nonrect) | |
2390 | && !integer_zerop (counts[i])) | |
2391 | t = fold_build2 (TRUNC_DIV_EXPR, type, t, counts[i]); | |
2392 | if (fd->loops[i].non_rect_referenced) | |
2393 | non_rect_referenced++; | |
2394 | if (fd->loops[i].m1 || fd->loops[i].m2) | |
2395 | non_rect++; | |
2396 | } | |
2397 | gcc_assert (non_rect == 1 && non_rect_referenced == 1); | |
2398 | counts[fd->last_nonrect] = t; | |
2399 | } | |
2400 | } | |
2401 | ||
2402 | /* Helper function for expand_omp_{for_*,simd}. Generate code like: | |
2403 | T = V; | |
2404 | V3 = N31 + (T % count3) * STEP3; | |
2405 | T = T / count3; | |
2406 | V2 = N21 + (T % count2) * STEP2; | |
2407 | T = T / count2; | |
2408 | V1 = N11 + T * STEP1; | |
2409 | if this loop doesn't have an inner loop construct combined with it. | |
2410 | If it does have an inner loop construct combined with it and the | |
2411 | iteration count isn't known constant, store values from counts array | |
2412 | into its _looptemp_ temporaries instead. | |
2413 | For non-rectangular loops (between fd->first_nonrect and fd->last_nonrect | |
2414 | inclusive), use the count of all those loops together, and either | |
2415 | find quadratic etc. equation roots, or as a fallback, do: | |
2416 | COUNT = 0; | |
2417 | for (tmpi = N11; tmpi COND1 N12; tmpi += STEP1) | |
2418 | for (tmpj = M21 * tmpi + N21; | |
2419 | tmpj COND2 M22 * tmpi + N22; tmpj += STEP2) | |
2420 | { | |
2421 | int tmpk1 = M31 * tmpj + N31; | |
2422 | int tmpk2 = M32 * tmpj + N32; | |
2423 | if (tmpk1 COND3 tmpk2) | |
2424 | { | |
2425 | if (COND3 is <) | |
2426 | adj = STEP3 - 1; | |
2427 | else | |
2428 | adj = STEP3 + 1; | |
2429 | int temp = (adj + tmpk2 - tmpk1) / STEP3; | |
2430 | if (COUNT + temp > T) | |
2431 | { | |
2432 | V1 = tmpi; | |
2433 | V2 = tmpj; | |
2434 | V3 = tmpk1 + (T - COUNT) * STEP3; | |
2435 | goto done; | |
2436 | } | |
2437 | else | |
2438 | COUNT += temp; | |
2439 | } | |
2440 | } | |
2441 | done:; | |
2442 | but for optional innermost or outermost rectangular loops that aren't | |
2443 | referenced by other loop expressions keep doing the division/modulo. */ | |
2444 | ||
2445 | static void | |
2446 | expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
2447 | tree *counts, tree *nonrect_bounds, | |
2448 | gimple *inner_stmt, tree startvar) | |
2449 | { | |
2450 | int i; | |
2451 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
2452 | { | |
2453 | /* If fd->loop.n2 is constant, then no propagation of the counts | |
2454 | is needed, they are constant. */ | |
2455 | if (TREE_CODE (fd->loop.n2) == INTEGER_CST) | |
2456 | return; | |
2457 | ||
2458 | tree clauses = gimple_code (inner_stmt) != GIMPLE_OMP_FOR | |
2459 | ? gimple_omp_taskreg_clauses (inner_stmt) | |
2460 | : gimple_omp_for_clauses (inner_stmt); | |
2461 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
2462 | isn't supposed to be handled, as the inner loop doesn't | |
2463 | use it. */ | |
2464 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
2465 | gcc_assert (innerc); | |
2466 | int count = 0; | |
2467 | if (fd->non_rect | |
2468 | && fd->last_nonrect == fd->first_nonrect + 1 | |
2469 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) | |
2470 | count = 4; | |
2471 | for (i = 0; i < fd->collapse + count; i++) | |
2472 | { | |
2473 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
2474 | OMP_CLAUSE__LOOPTEMP_); | |
2475 | gcc_assert (innerc); | |
2476 | if (i) | |
2477 | { | |
2478 | tree tem = OMP_CLAUSE_DECL (innerc); | |
2479 | tree t; | |
2480 | if (i < fd->collapse) | |
2481 | t = counts[i]; | |
2482 | else | |
2483 | switch (i - fd->collapse) | |
2484 | { | |
2485 | case 0: t = counts[0]; break; | |
2486 | case 1: t = fd->first_inner_iterations; break; | |
2487 | case 2: t = fd->factor; break; | |
2488 | case 3: t = fd->adjn1; break; | |
2489 | default: gcc_unreachable (); | |
2490 | } | |
2491 | t = fold_convert (TREE_TYPE (tem), t); | |
2492 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
2493 | false, GSI_CONTINUE_LINKING); | |
2494 | gassign *stmt = gimple_build_assign (tem, t); | |
2495 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2496 | } | |
2497 | } | |
2498 | return; | |
2499 | } | |
2500 | ||
2501 | tree type = TREE_TYPE (fd->loop.v); | |
2502 | tree tem = create_tmp_reg (type, ".tem"); | |
2503 | gassign *stmt = gimple_build_assign (tem, startvar); | |
2504 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2505 | ||
2506 | for (i = fd->collapse - 1; i >= 0; i--) | |
2507 | { | |
2508 | tree vtype = TREE_TYPE (fd->loops[i].v), itype, t; | |
2509 | itype = vtype; | |
2510 | if (POINTER_TYPE_P (vtype)) | |
2511 | itype = signed_type_for (vtype); | |
2512 | if (i != 0 && (i != fd->last_nonrect || fd->first_nonrect)) | |
2513 | t = fold_build2 (TRUNC_MOD_EXPR, type, tem, counts[i]); | |
2514 | else | |
2515 | t = tem; | |
2516 | if (i == fd->last_nonrect) | |
2517 | { | |
2518 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, | |
2519 | false, GSI_CONTINUE_LINKING); | |
2520 | tree stopval = t; | |
2521 | tree idx = create_tmp_reg (type, ".count"); | |
2522 | expand_omp_build_assign (gsi, idx, | |
2523 | build_zero_cst (type), true); | |
2524 | basic_block bb_triang = NULL, bb_triang_dom = NULL; | |
2525 | if (fd->first_nonrect + 1 == fd->last_nonrect | |
2526 | && (TREE_CODE (fd->loop.n2) == INTEGER_CST | |
2527 | || fd->first_inner_iterations) | |
2528 | && (optab_handler (sqrt_optab, TYPE_MODE (double_type_node)) | |
2529 | != CODE_FOR_nothing) | |
2530 | && !integer_zerop (fd->loop.n2)) | |
2531 | { | |
2532 | tree outer_n1 = fd->adjn1 ? fd->adjn1 : fd->loops[i - 1].n1; | |
2533 | tree itype = TREE_TYPE (fd->loops[i].v); | |
2534 | tree first_inner_iterations = fd->first_inner_iterations; | |
2535 | tree factor = fd->factor; | |
2536 | gcond *cond_stmt | |
2537 | = gimple_build_cond (NE_EXPR, factor, | |
2538 | build_zero_cst (TREE_TYPE (factor)), | |
2539 | NULL_TREE, NULL_TREE); | |
2540 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
2541 | edge e = split_block (gsi_bb (*gsi), cond_stmt); | |
2542 | basic_block bb0 = e->src; | |
2543 | e->flags = EDGE_TRUE_VALUE; | |
2544 | e->probability = profile_probability::likely (); | |
2545 | bb_triang_dom = bb0; | |
2546 | *gsi = gsi_after_labels (e->dest); | |
2547 | tree slltype = long_long_integer_type_node; | |
2548 | tree ulltype = long_long_unsigned_type_node; | |
2549 | tree stopvalull = fold_convert (ulltype, stopval); | |
2550 | stopvalull | |
2551 | = force_gimple_operand_gsi (gsi, stopvalull, true, NULL_TREE, | |
2552 | false, GSI_CONTINUE_LINKING); | |
2553 | first_inner_iterations | |
2554 | = fold_convert (slltype, first_inner_iterations); | |
2555 | first_inner_iterations | |
2556 | = force_gimple_operand_gsi (gsi, first_inner_iterations, true, | |
2557 | NULL_TREE, false, | |
2558 | GSI_CONTINUE_LINKING); | |
2559 | factor = fold_convert (slltype, factor); | |
2560 | factor | |
2561 | = force_gimple_operand_gsi (gsi, factor, true, NULL_TREE, | |
2562 | false, GSI_CONTINUE_LINKING); | |
2563 | tree first_inner_iterationsd | |
2564 | = fold_build1 (FLOAT_EXPR, double_type_node, | |
2565 | first_inner_iterations); | |
2566 | first_inner_iterationsd | |
2567 | = force_gimple_operand_gsi (gsi, first_inner_iterationsd, true, | |
2568 | NULL_TREE, false, | |
2569 | GSI_CONTINUE_LINKING); | |
2570 | tree factord = fold_build1 (FLOAT_EXPR, double_type_node, | |
2571 | factor); | |
2572 | factord = force_gimple_operand_gsi (gsi, factord, true, | |
2573 | NULL_TREE, false, | |
2574 | GSI_CONTINUE_LINKING); | |
2575 | tree stopvald = fold_build1 (FLOAT_EXPR, double_type_node, | |
2576 | stopvalull); | |
2577 | stopvald = force_gimple_operand_gsi (gsi, stopvald, true, | |
2578 | NULL_TREE, false, | |
2579 | GSI_CONTINUE_LINKING); | |
2580 | /* Temporarily disable flag_rounding_math, values will be | |
2581 | decimal numbers divided by 2 and worst case imprecisions | |
2582 | due to too large values ought to be caught later by the | |
2583 | checks for fallback. */ | |
2584 | int save_flag_rounding_math = flag_rounding_math; | |
2585 | flag_rounding_math = 0; | |
2586 | t = fold_build2 (RDIV_EXPR, double_type_node, factord, | |
2587 | build_real (double_type_node, dconst2)); | |
2588 | tree t3 = fold_build2 (MINUS_EXPR, double_type_node, | |
2589 | first_inner_iterationsd, t); | |
2590 | t3 = force_gimple_operand_gsi (gsi, t3, true, NULL_TREE, false, | |
2591 | GSI_CONTINUE_LINKING); | |
2592 | t = fold_build2 (MULT_EXPR, double_type_node, factord, | |
2593 | build_real (double_type_node, dconst2)); | |
2594 | t = fold_build2 (MULT_EXPR, double_type_node, t, stopvald); | |
2595 | t = fold_build2 (PLUS_EXPR, double_type_node, t, | |
2596 | fold_build2 (MULT_EXPR, double_type_node, | |
2597 | t3, t3)); | |
2598 | flag_rounding_math = save_flag_rounding_math; | |
2599 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, | |
2600 | GSI_CONTINUE_LINKING); | |
2601 | if (flag_exceptions | |
2602 | && cfun->can_throw_non_call_exceptions | |
2603 | && operation_could_trap_p (LT_EXPR, true, false, NULL_TREE)) | |
2604 | { | |
2605 | tree tem = fold_build2 (LT_EXPR, boolean_type_node, t, | |
2606 | build_zero_cst (double_type_node)); | |
2607 | tem = force_gimple_operand_gsi (gsi, tem, true, NULL_TREE, | |
2608 | false, GSI_CONTINUE_LINKING); | |
2609 | cond_stmt = gimple_build_cond (NE_EXPR, tem, | |
2610 | boolean_false_node, | |
2611 | NULL_TREE, NULL_TREE); | |
2612 | } | |
2613 | else | |
2614 | cond_stmt | |
2615 | = gimple_build_cond (LT_EXPR, t, | |
2616 | build_zero_cst (double_type_node), | |
2617 | NULL_TREE, NULL_TREE); | |
2618 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
2619 | e = split_block (gsi_bb (*gsi), cond_stmt); | |
2620 | basic_block bb1 = e->src; | |
2621 | e->flags = EDGE_FALSE_VALUE; | |
2622 | e->probability = profile_probability::very_likely (); | |
2623 | *gsi = gsi_after_labels (e->dest); | |
2624 | gcall *call = gimple_build_call_internal (IFN_SQRT, 1, t); | |
2625 | tree sqrtr = create_tmp_var (double_type_node); | |
2626 | gimple_call_set_lhs (call, sqrtr); | |
2627 | gsi_insert_after (gsi, call, GSI_CONTINUE_LINKING); | |
2628 | t = fold_build2 (MINUS_EXPR, double_type_node, sqrtr, t3); | |
2629 | t = fold_build2 (RDIV_EXPR, double_type_node, t, factord); | |
2630 | t = fold_build1 (FIX_TRUNC_EXPR, ulltype, t); | |
2631 | tree c = create_tmp_var (ulltype); | |
2632 | tree d = create_tmp_var (ulltype); | |
2633 | expand_omp_build_assign (gsi, c, t, true); | |
2634 | t = fold_build2 (MINUS_EXPR, ulltype, c, | |
2635 | build_one_cst (ulltype)); | |
2636 | t = fold_build2 (MULT_EXPR, ulltype, c, t); | |
2637 | t = fold_build2 (RSHIFT_EXPR, ulltype, t, integer_one_node); | |
2638 | t = fold_build2 (MULT_EXPR, ulltype, | |
2639 | fold_convert (ulltype, fd->factor), t); | |
2640 | tree t2 | |
2641 | = fold_build2 (MULT_EXPR, ulltype, c, | |
2642 | fold_convert (ulltype, | |
2643 | fd->first_inner_iterations)); | |
2644 | t = fold_build2 (PLUS_EXPR, ulltype, t, t2); | |
2645 | expand_omp_build_assign (gsi, d, t, true); | |
2646 | t = fold_build2 (MULT_EXPR, ulltype, | |
2647 | fold_convert (ulltype, fd->factor), c); | |
2648 | t = fold_build2 (PLUS_EXPR, ulltype, | |
2649 | t, fold_convert (ulltype, | |
2650 | fd->first_inner_iterations)); | |
2651 | t2 = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, | |
2652 | GSI_CONTINUE_LINKING); | |
2653 | cond_stmt = gimple_build_cond (GE_EXPR, stopvalull, d, | |
2654 | NULL_TREE, NULL_TREE); | |
2655 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
2656 | e = split_block (gsi_bb (*gsi), cond_stmt); | |
2657 | basic_block bb2 = e->src; | |
2658 | e->flags = EDGE_TRUE_VALUE; | |
2659 | e->probability = profile_probability::very_likely (); | |
2660 | *gsi = gsi_after_labels (e->dest); | |
2661 | t = fold_build2 (PLUS_EXPR, ulltype, d, t2); | |
2662 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, | |
2663 | GSI_CONTINUE_LINKING); | |
2664 | cond_stmt = gimple_build_cond (GE_EXPR, stopvalull, t, | |
2665 | NULL_TREE, NULL_TREE); | |
2666 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
2667 | e = split_block (gsi_bb (*gsi), cond_stmt); | |
2668 | basic_block bb3 = e->src; | |
2669 | e->flags = EDGE_FALSE_VALUE; | |
2670 | e->probability = profile_probability::very_likely (); | |
2671 | *gsi = gsi_after_labels (e->dest); | |
2672 | t = fold_convert (itype, c); | |
2673 | t = fold_build2 (MULT_EXPR, itype, t, fd->loops[i - 1].step); | |
2674 | t = fold_build2 (PLUS_EXPR, itype, outer_n1, t); | |
2675 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, | |
2676 | GSI_CONTINUE_LINKING); | |
2677 | expand_omp_build_assign (gsi, fd->loops[i - 1].v, t, true); | |
2678 | t2 = fold_build2 (MINUS_EXPR, ulltype, stopvalull, d); | |
2679 | t2 = fold_convert (itype, t2); | |
2680 | t2 = fold_build2 (MULT_EXPR, itype, t2, fd->loops[i].step); | |
2681 | t2 = fold_build2 (PLUS_EXPR, itype, t2, fd->loops[i].n1); | |
2682 | if (fd->loops[i].m1) | |
2683 | { | |
2684 | t = fold_build2 (MULT_EXPR, itype, t, fd->loops[i].m1); | |
2685 | t2 = fold_build2 (PLUS_EXPR, itype, t2, t); | |
2686 | } | |
2687 | expand_omp_build_assign (gsi, fd->loops[i].v, t2, true); | |
2688 | e = split_block (gsi_bb (*gsi), gsi_stmt (*gsi)); | |
2689 | bb_triang = e->src; | |
2690 | *gsi = gsi_after_labels (e->dest); | |
2691 | remove_edge (e); | |
2692 | e = make_edge (bb1, gsi_bb (*gsi), EDGE_TRUE_VALUE); | |
2693 | e->probability = profile_probability::very_unlikely (); | |
2694 | e = make_edge (bb2, gsi_bb (*gsi), EDGE_FALSE_VALUE); | |
2695 | e->probability = profile_probability::very_unlikely (); | |
2696 | e = make_edge (bb3, gsi_bb (*gsi), EDGE_TRUE_VALUE); | |
2697 | e->probability = profile_probability::very_unlikely (); | |
2698 | ||
2699 | basic_block bb4 = create_empty_bb (bb0); | |
2700 | add_bb_to_loop (bb4, bb0->loop_father); | |
2701 | e = make_edge (bb0, bb4, EDGE_FALSE_VALUE); | |
2702 | e->probability = profile_probability::unlikely (); | |
2703 | make_edge (bb4, gsi_bb (*gsi), EDGE_FALLTHRU); | |
2704 | set_immediate_dominator (CDI_DOMINATORS, bb4, bb0); | |
2705 | set_immediate_dominator (CDI_DOMINATORS, gsi_bb (*gsi), bb0); | |
2706 | gimple_stmt_iterator gsi2 = gsi_after_labels (bb4); | |
2707 | t2 = fold_build2 (TRUNC_DIV_EXPR, type, | |
2708 | counts[i], counts[i - 1]); | |
2709 | t2 = force_gimple_operand_gsi (&gsi2, t2, true, NULL_TREE, false, | |
2710 | GSI_CONTINUE_LINKING); | |
2711 | t = fold_build2 (TRUNC_MOD_EXPR, type, stopval, t2); | |
2712 | t2 = fold_build2 (TRUNC_DIV_EXPR, type, stopval, t2); | |
2713 | t = fold_convert (itype, t); | |
2714 | t2 = fold_convert (itype, t2); | |
2715 | t = fold_build2 (MULT_EXPR, itype, t, | |
2716 | fold_convert (itype, fd->loops[i].step)); | |
2717 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); | |
2718 | t2 = fold_build2 (MULT_EXPR, itype, t2, | |
2719 | fold_convert (itype, fd->loops[i - 1].step)); | |
2720 | t2 = fold_build2 (PLUS_EXPR, itype, fd->loops[i - 1].n1, t2); | |
2721 | t2 = force_gimple_operand_gsi (&gsi2, t2, false, NULL_TREE, | |
2722 | false, GSI_CONTINUE_LINKING); | |
2723 | stmt = gimple_build_assign (fd->loops[i - 1].v, t2); | |
2724 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
2725 | if (fd->loops[i].m1) | |
2726 | { | |
2727 | t2 = fold_build2 (MULT_EXPR, itype, fd->loops[i].m1, | |
2728 | fd->loops[i - 1].v); | |
2729 | t = fold_build2 (PLUS_EXPR, itype, t, t2); | |
2730 | } | |
2731 | t = force_gimple_operand_gsi (&gsi2, t, false, NULL_TREE, | |
2732 | false, GSI_CONTINUE_LINKING); | |
2733 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
2734 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
2735 | } | |
2736 | /* Fallback implementation. Evaluate the loops in between | |
2737 | (inclusive) fd->first_nonrect and fd->last_nonrect at | |
2738 | runtime unsing temporaries instead of the original iteration | |
2739 | variables, in the body just bump the counter and compare | |
2740 | with the desired value. */ | |
2741 | gimple_stmt_iterator gsi2 = *gsi; | |
2742 | basic_block entry_bb = gsi_bb (gsi2); | |
2743 | edge e = split_block (entry_bb, gsi_stmt (gsi2)); | |
2744 | e = split_block (e->dest, (gimple *) NULL); | |
2745 | basic_block dom_bb = NULL; | |
2746 | basic_block cur_bb = e->src; | |
2747 | basic_block next_bb = e->dest; | |
2748 | entry_bb = e->dest; | |
2749 | *gsi = gsi_after_labels (entry_bb); | |
2750 | ||
2751 | tree *vs = XALLOCAVEC (tree, fd->last_nonrect); | |
2752 | tree n1 = NULL_TREE, n2 = NULL_TREE; | |
2753 | memset (vs, 0, fd->last_nonrect * sizeof (tree)); | |
2754 | ||
2755 | for (int j = fd->first_nonrect; j <= fd->last_nonrect; j++) | |
2756 | { | |
2757 | tree itype = TREE_TYPE (fd->loops[j].v); | |
2758 | bool rect_p = (fd->loops[j].m1 == NULL_TREE | |
2759 | && fd->loops[j].m2 == NULL_TREE | |
2760 | && !fd->loops[j].non_rect_referenced); | |
2761 | gsi2 = gsi_after_labels (cur_bb); | |
2762 | t = fold_convert (itype, unshare_expr (fd->loops[j].n1)); | |
2763 | if (fd->loops[j].m1) | |
2764 | { | |
2765 | n1 = fold_convert (itype, unshare_expr (fd->loops[j].m1)); | |
2766 | n1 = fold_build2 (MULT_EXPR, itype, | |
2767 | vs[j - fd->loops[j].outer], n1); | |
2768 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2769 | } | |
2770 | else if (rect_p) | |
2771 | n1 = build_zero_cst (type); | |
2772 | else | |
2773 | n1 = t; | |
2774 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, | |
2775 | true, GSI_SAME_STMT); | |
2776 | if (j < fd->last_nonrect) | |
2777 | { | |
2778 | vs[j] = create_tmp_reg (rect_p ? type : itype, ".it"); | |
2779 | expand_omp_build_assign (&gsi2, vs[j], n1); | |
2780 | } | |
2781 | t = fold_convert (itype, unshare_expr (fd->loops[j].n2)); | |
2782 | if (fd->loops[j].m2) | |
2783 | { | |
2784 | n2 = fold_convert (itype, unshare_expr (fd->loops[j].m2)); | |
2785 | n2 = fold_build2 (MULT_EXPR, itype, | |
2786 | vs[j - fd->loops[j].outer], n2); | |
2787 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); | |
2788 | } | |
2789 | else if (rect_p) | |
2790 | n2 = counts[j]; | |
2791 | else | |
2792 | n2 = t; | |
2793 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, | |
2794 | true, GSI_SAME_STMT); | |
2795 | if (j == fd->last_nonrect) | |
2796 | { | |
2797 | gcond *cond_stmt | |
2798 | = gimple_build_cond (fd->loops[j].cond_code, n1, n2, | |
2799 | NULL_TREE, NULL_TREE); | |
2800 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2801 | e = split_block (cur_bb, cond_stmt); | |
2802 | e->flags = EDGE_TRUE_VALUE; | |
2803 | edge ne = make_edge (cur_bb, next_bb, EDGE_FALSE_VALUE); | |
2804 | e->probability = profile_probability::likely ().guessed (); | |
2805 | ne->probability = e->probability.invert (); | |
2806 | gsi2 = gsi_after_labels (e->dest); | |
2807 | ||
2808 | t = build_int_cst (itype, (fd->loops[j].cond_code == LT_EXPR | |
2809 | ? -1 : 1)); | |
2810 | t = fold_build2 (PLUS_EXPR, itype, | |
2811 | fold_convert (itype, fd->loops[j].step), t); | |
2812 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
2813 | t = fold_build2 (MINUS_EXPR, itype, t, n1); | |
2814 | tree step = fold_convert (itype, fd->loops[j].step); | |
2815 | if (TYPE_UNSIGNED (itype) | |
2816 | && fd->loops[j].cond_code == GT_EXPR) | |
2817 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2818 | fold_build1 (NEGATE_EXPR, itype, t), | |
2819 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2820 | else | |
2821 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2822 | t = fold_convert (type, t); | |
2823 | t = fold_build2 (PLUS_EXPR, type, idx, t); | |
2824 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2825 | true, GSI_SAME_STMT); | |
2826 | e = make_edge (e->dest, next_bb, EDGE_FALLTHRU); | |
2827 | set_immediate_dominator (CDI_DOMINATORS, next_bb, cur_bb); | |
2828 | cond_stmt | |
2829 | = gimple_build_cond (LE_EXPR, t, stopval, NULL_TREE, | |
2830 | NULL_TREE); | |
2831 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2832 | e = split_block (gsi_bb (gsi2), cond_stmt); | |
2833 | e->flags = EDGE_TRUE_VALUE; | |
2834 | e->probability = profile_probability::likely ().guessed (); | |
2835 | ne = make_edge (e->src, entry_bb, EDGE_FALSE_VALUE); | |
2836 | ne->probability = e->probability.invert (); | |
2837 | gsi2 = gsi_after_labels (e->dest); | |
2838 | expand_omp_build_assign (&gsi2, idx, t); | |
2839 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, dom_bb); | |
2840 | break; | |
2841 | } | |
2842 | e = split_block (cur_bb, last_stmt (cur_bb)); | |
2843 | ||
2844 | basic_block new_cur_bb = create_empty_bb (cur_bb); | |
2845 | add_bb_to_loop (new_cur_bb, cur_bb->loop_father); | |
2846 | ||
2847 | gsi2 = gsi_after_labels (e->dest); | |
2848 | if (rect_p) | |
2849 | t = fold_build2 (PLUS_EXPR, type, vs[j], | |
2850 | build_one_cst (type)); | |
2851 | else | |
2852 | { | |
2853 | tree step | |
2854 | = fold_convert (itype, unshare_expr (fd->loops[j].step)); | |
2855 | t = fold_build2 (PLUS_EXPR, itype, vs[j], step); | |
2856 | } | |
2857 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2858 | true, GSI_SAME_STMT); | |
2859 | expand_omp_build_assign (&gsi2, vs[j], t); | |
2860 | ||
2861 | edge ne = split_block (e->dest, last_stmt (e->dest)); | |
2862 | gsi2 = gsi_after_labels (ne->dest); | |
2863 | ||
2864 | gcond *cond_stmt; | |
2865 | if (next_bb == entry_bb) | |
2866 | /* No need to actually check the outermost condition. */ | |
2867 | cond_stmt | |
2868 | = gimple_build_cond (EQ_EXPR, boolean_true_node, | |
2869 | boolean_true_node, | |
2870 | NULL_TREE, NULL_TREE); | |
2871 | else | |
2872 | cond_stmt | |
2873 | = gimple_build_cond (rect_p ? LT_EXPR | |
2874 | : fd->loops[j].cond_code, | |
2875 | vs[j], n2, NULL_TREE, NULL_TREE); | |
2876 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2877 | edge e3, e4; | |
2878 | if (next_bb == entry_bb) | |
2879 | { | |
2880 | e3 = find_edge (ne->dest, next_bb); | |
2881 | e3->flags = EDGE_FALSE_VALUE; | |
2882 | dom_bb = ne->dest; | |
2883 | } | |
2884 | else | |
2885 | e3 = make_edge (ne->dest, next_bb, EDGE_FALSE_VALUE); | |
2886 | e4 = make_edge (ne->dest, new_cur_bb, EDGE_TRUE_VALUE); | |
2887 | e4->probability = profile_probability::likely ().guessed (); | |
2888 | e3->probability = e4->probability.invert (); | |
2889 | basic_block esrc = e->src; | |
2890 | make_edge (e->src, ne->dest, EDGE_FALLTHRU); | |
2891 | cur_bb = new_cur_bb; | |
2892 | basic_block latch_bb = next_bb; | |
2893 | next_bb = e->dest; | |
2894 | remove_edge (e); | |
2895 | set_immediate_dominator (CDI_DOMINATORS, ne->dest, esrc); | |
2896 | set_immediate_dominator (CDI_DOMINATORS, latch_bb, ne->dest); | |
2897 | set_immediate_dominator (CDI_DOMINATORS, cur_bb, ne->dest); | |
2898 | } | |
2899 | for (int j = fd->last_nonrect; j >= fd->first_nonrect; j--) | |
2900 | { | |
2901 | tree itype = TREE_TYPE (fd->loops[j].v); | |
2902 | bool rect_p = (fd->loops[j].m1 == NULL_TREE | |
2903 | && fd->loops[j].m2 == NULL_TREE | |
2904 | && !fd->loops[j].non_rect_referenced); | |
2905 | if (j == fd->last_nonrect) | |
2906 | { | |
2907 | t = fold_build2 (MINUS_EXPR, type, stopval, idx); | |
2908 | t = fold_convert (itype, t); | |
2909 | tree t2 | |
2910 | = fold_convert (itype, unshare_expr (fd->loops[j].step)); | |
2911 | t = fold_build2 (MULT_EXPR, itype, t, t2); | |
2912 | t = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2913 | } | |
2914 | else if (rect_p) | |
2915 | { | |
2916 | t = fold_convert (itype, vs[j]); | |
2917 | t = fold_build2 (MULT_EXPR, itype, t, | |
2918 | fold_convert (itype, fd->loops[j].step)); | |
2919 | if (POINTER_TYPE_P (vtype)) | |
2920 | t = fold_build_pointer_plus (fd->loops[j].n1, t); | |
2921 | else | |
2922 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[j].n1, t); | |
2923 | } | |
2924 | else | |
2925 | t = vs[j]; | |
2926 | t = force_gimple_operand_gsi (gsi, t, false, | |
2927 | NULL_TREE, true, | |
2928 | GSI_SAME_STMT); | |
2929 | stmt = gimple_build_assign (fd->loops[j].v, t); | |
2930 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
2931 | } | |
2932 | if (gsi_end_p (*gsi)) | |
2933 | *gsi = gsi_last_bb (gsi_bb (*gsi)); | |
2934 | else | |
2935 | gsi_prev (gsi); | |
2936 | if (bb_triang) | |
2937 | { | |
2938 | e = split_block (gsi_bb (*gsi), gsi_stmt (*gsi)); | |
2939 | make_edge (bb_triang, e->dest, EDGE_FALLTHRU); | |
2940 | *gsi = gsi_after_labels (e->dest); | |
2941 | if (!gsi_end_p (*gsi)) | |
2942 | gsi_insert_before (gsi, gimple_build_nop (), GSI_NEW_STMT); | |
2943 | set_immediate_dominator (CDI_DOMINATORS, e->dest, bb_triang_dom); | |
2944 | } | |
2945 | } | |
2946 | else | |
2947 | { | |
2948 | t = fold_convert (itype, t); | |
2949 | t = fold_build2 (MULT_EXPR, itype, t, | |
2950 | fold_convert (itype, fd->loops[i].step)); | |
2951 | if (POINTER_TYPE_P (vtype)) | |
2952 | t = fold_build_pointer_plus (fd->loops[i].n1, t); | |
2953 | else | |
2954 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); | |
2955 | t = force_gimple_operand_gsi (gsi, t, | |
2956 | DECL_P (fd->loops[i].v) | |
2957 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
2958 | NULL_TREE, false, | |
2959 | GSI_CONTINUE_LINKING); | |
2960 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
2961 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2962 | } | |
2963 | if (i != 0 && (i != fd->last_nonrect || fd->first_nonrect)) | |
2964 | { | |
2965 | t = fold_build2 (TRUNC_DIV_EXPR, type, tem, counts[i]); | |
2966 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
2967 | false, GSI_CONTINUE_LINKING); | |
2968 | stmt = gimple_build_assign (tem, t); | |
2969 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2970 | } | |
2971 | if (i == fd->last_nonrect) | |
2972 | i = fd->first_nonrect; | |
2973 | } | |
2974 | if (fd->non_rect) | |
2975 | for (i = 0; i <= fd->last_nonrect; i++) | |
2976 | if (fd->loops[i].m2) | |
2977 | { | |
2978 | tree itype = TREE_TYPE (fd->loops[i].v); | |
2979 | ||
2980 | tree t = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2981 | t = fold_build2 (MULT_EXPR, itype, | |
2982 | fd->loops[i - fd->loops[i].outer].v, t); | |
2983 | t = fold_build2 (PLUS_EXPR, itype, t, | |
2984 | fold_convert (itype, | |
2985 | unshare_expr (fd->loops[i].n2))); | |
2986 | nonrect_bounds[i] = create_tmp_reg (itype, ".bound"); | |
2987 | t = force_gimple_operand_gsi (gsi, t, false, | |
2988 | NULL_TREE, false, | |
2989 | GSI_CONTINUE_LINKING); | |
2990 | stmt = gimple_build_assign (nonrect_bounds[i], t); | |
2991 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2992 | } | |
2993 | } | |
2994 | ||
2995 | /* Helper function for expand_omp_for_*. Generate code like: | |
2996 | L10: | |
2997 | V3 += STEP3; | |
2998 | if (V3 cond3 N32) goto BODY_BB; else goto L11; | |
2999 | L11: | |
3000 | V3 = N31; | |
3001 | V2 += STEP2; | |
3002 | if (V2 cond2 N22) goto BODY_BB; else goto L12; | |
3003 | L12: | |
3004 | V2 = N21; | |
3005 | V1 += STEP1; | |
3006 | goto BODY_BB; | |
3007 | For non-rectangular loops, use temporaries stored in nonrect_bounds | |
3008 | for the upper bounds if M?2 multiplier is present. Given e.g. | |
3009 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
3010 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
3011 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
3012 | for (V4 = N41 + M41 * V2; V4 cond4 N42 + M42 * V2; V4 += STEP4) | |
3013 | do: | |
3014 | L10: | |
3015 | V4 += STEP4; | |
3016 | if (V4 cond4 NONRECT_BOUND4) goto BODY_BB; else goto L11; | |
3017 | L11: | |
3018 | V4 = N41 + M41 * V2; // This can be left out if the loop | |
3019 | // refers to the immediate parent loop | |
3020 | V3 += STEP3; | |
3021 | if (V3 cond3 N32) goto BODY_BB; else goto L12; | |
3022 | L12: | |
3023 | V3 = N31; | |
3024 | V2 += STEP2; | |
3025 | if (V2 cond2 N22) goto L120; else goto L13; | |
3026 | L120: | |
3027 | V4 = N41 + M41 * V2; | |
3028 | NONRECT_BOUND4 = N42 + M42 * V2; | |
3029 | if (V4 cond4 NONRECT_BOUND4) goto BODY_BB; else goto L12; | |
3030 | L13: | |
3031 | V2 = N21; | |
3032 | V1 += STEP1; | |
3033 | goto L120; */ | |
3034 | ||
3035 | static basic_block | |
3036 | extract_omp_for_update_vars (struct omp_for_data *fd, tree *nonrect_bounds, | |
3037 | basic_block cont_bb, basic_block body_bb) | |
3038 | { | |
3039 | basic_block last_bb, bb, collapse_bb = NULL; | |
3040 | int i; | |
3041 | gimple_stmt_iterator gsi; | |
3042 | edge e; | |
3043 | tree t; | |
3044 | gimple *stmt; | |
3045 | ||
3046 | last_bb = cont_bb; | |
3047 | for (i = fd->collapse - 1; i >= 0; i--) | |
3048 | { | |
3049 | tree vtype = TREE_TYPE (fd->loops[i].v); | |
3050 | ||
3051 | bb = create_empty_bb (last_bb); | |
3052 | add_bb_to_loop (bb, last_bb->loop_father); | |
3053 | gsi = gsi_start_bb (bb); | |
3054 | ||
3055 | if (i < fd->collapse - 1) | |
3056 | { | |
3057 | e = make_edge (last_bb, bb, EDGE_FALSE_VALUE); | |
3058 | e->probability | |
3059 | = profile_probability::guessed_always ().apply_scale (1, 8); | |
3060 | ||
3061 | struct omp_for_data_loop *l = &fd->loops[i + 1]; | |
3062 | if (l->m1 == NULL_TREE || l->outer != 1) | |
3063 | { | |
3064 | t = l->n1; | |
3065 | if (l->m1) | |
3066 | { | |
3067 | tree t2 | |
3068 | = fold_build2 (MULT_EXPR, TREE_TYPE (t), | |
3069 | fd->loops[i + 1 - l->outer].v, l->m1); | |
3070 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t2, t); | |
3071 | } | |
3072 | t = force_gimple_operand_gsi (&gsi, t, | |
3073 | DECL_P (l->v) | |
3074 | && TREE_ADDRESSABLE (l->v), | |
3075 | NULL_TREE, false, | |
3076 | GSI_CONTINUE_LINKING); | |
3077 | stmt = gimple_build_assign (l->v, t); | |
3078 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
3079 | } | |
3080 | } | |
3081 | else | |
3082 | collapse_bb = bb; | |
3083 | ||
3084 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); | |
3085 | ||
3086 | if (POINTER_TYPE_P (vtype)) | |
3087 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); | |
3088 | else | |
3089 | t = fold_build2 (PLUS_EXPR, vtype, fd->loops[i].v, fd->loops[i].step); | |
3090 | t = force_gimple_operand_gsi (&gsi, t, | |
3091 | DECL_P (fd->loops[i].v) | |
3092 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
3093 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
3094 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
3095 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
3096 | ||
3097 | if (fd->loops[i].non_rect_referenced) | |
3098 | { | |
3099 | basic_block update_bb = NULL, prev_bb = NULL; | |
3100 | for (int j = i + 1; j <= fd->last_nonrect; j++) | |
3101 | if (j - fd->loops[j].outer == i) | |
3102 | { | |
3103 | tree n1, n2; | |
3104 | struct omp_for_data_loop *l = &fd->loops[j]; | |
3105 | basic_block this_bb = create_empty_bb (last_bb); | |
3106 | add_bb_to_loop (this_bb, last_bb->loop_father); | |
3107 | gimple_stmt_iterator gsi2 = gsi_start_bb (this_bb); | |
3108 | if (prev_bb) | |
3109 | { | |
3110 | e = make_edge (prev_bb, this_bb, EDGE_TRUE_VALUE); | |
3111 | e->probability | |
3112 | = profile_probability::guessed_always ().apply_scale (7, | |
3113 | 8); | |
3114 | set_immediate_dominator (CDI_DOMINATORS, this_bb, prev_bb); | |
3115 | } | |
3116 | if (l->m1) | |
3117 | { | |
3118 | t = fold_build2 (MULT_EXPR, TREE_TYPE (l->m1), l->m1, | |
3119 | fd->loops[i].v); | |
3120 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (l->v), t, l->n1); | |
3121 | n1 = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
3122 | false, | |
3123 | GSI_CONTINUE_LINKING); | |
3124 | stmt = gimple_build_assign (l->v, n1); | |
3125 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
3126 | n1 = l->v; | |
3127 | } | |
3128 | else | |
3129 | n1 = force_gimple_operand_gsi (&gsi2, l->n1, true, | |
3130 | NULL_TREE, false, | |
3131 | GSI_CONTINUE_LINKING); | |
3132 | if (l->m2) | |
3133 | { | |
3134 | t = fold_build2 (MULT_EXPR, TREE_TYPE (l->m2), l->m2, | |
3135 | fd->loops[i].v); | |
3136 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (nonrect_bounds[j]), | |
3137 | t, unshare_expr (l->n2)); | |
3138 | n2 = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
3139 | false, | |
3140 | GSI_CONTINUE_LINKING); | |
3141 | stmt = gimple_build_assign (nonrect_bounds[j], n2); | |
3142 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
3143 | n2 = nonrect_bounds[j]; | |
3144 | } | |
3145 | else | |
3146 | n2 = force_gimple_operand_gsi (&gsi2, unshare_expr (l->n2), | |
3147 | true, NULL_TREE, false, | |
3148 | GSI_CONTINUE_LINKING); | |
3149 | gcond *cond_stmt | |
3150 | = gimple_build_cond (l->cond_code, n1, n2, | |
3151 | NULL_TREE, NULL_TREE); | |
3152 | gsi_insert_after (&gsi2, cond_stmt, GSI_CONTINUE_LINKING); | |
3153 | if (update_bb == NULL) | |
3154 | update_bb = this_bb; | |
3155 | e = make_edge (this_bb, bb, EDGE_FALSE_VALUE); | |
3156 | e->probability | |
3157 | = profile_probability::guessed_always ().apply_scale (1, 8); | |
3158 | if (prev_bb == NULL) | |
3159 | set_immediate_dominator (CDI_DOMINATORS, this_bb, bb); | |
3160 | prev_bb = this_bb; | |
3161 | } | |
3162 | e = make_edge (prev_bb, body_bb, EDGE_TRUE_VALUE); | |
3163 | e->probability | |
3164 | = profile_probability::guessed_always ().apply_scale (7, 8); | |
3165 | body_bb = update_bb; | |
3166 | } | |
3167 | ||
3168 | if (i > 0) | |
3169 | { | |
3170 | if (fd->loops[i].m2) | |
3171 | t = nonrect_bounds[i]; | |
3172 | else | |
3173 | t = unshare_expr (fd->loops[i].n2); | |
3174 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
3175 | false, GSI_CONTINUE_LINKING); | |
3176 | tree v = fd->loops[i].v; | |
3177 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
3178 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
3179 | false, GSI_CONTINUE_LINKING); | |
3180 | t = fold_build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
3181 | stmt = gimple_build_cond_empty (t); | |
3182 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
3183 | if (walk_tree (gimple_cond_lhs_ptr (as_a <gcond *> (stmt)), | |
3184 | expand_omp_regimplify_p, NULL, NULL) | |
3185 | || walk_tree (gimple_cond_rhs_ptr (as_a <gcond *> (stmt)), | |
3186 | expand_omp_regimplify_p, NULL, NULL)) | |
3187 | gimple_regimplify_operands (stmt, &gsi); | |
3188 | e = make_edge (bb, body_bb, EDGE_TRUE_VALUE); | |
3189 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); | |
3190 | } | |
3191 | else | |
3192 | make_edge (bb, body_bb, EDGE_FALLTHRU); | |
3193 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); | |
3194 | last_bb = bb; | |
3195 | } | |
3196 | ||
3197 | return collapse_bb; | |
3198 | } | |
3199 | ||
3200 | /* Expand #pragma omp ordered depend(source). */ | |
3201 | ||
3202 | static void | |
3203 | expand_omp_ordered_source (gimple_stmt_iterator *gsi, struct omp_for_data *fd, | |
3204 | tree *counts, location_t loc) | |
3205 | { | |
3206 | enum built_in_function source_ix | |
3207 | = fd->iter_type == long_integer_type_node | |
3208 | ? BUILT_IN_GOMP_DOACROSS_POST : BUILT_IN_GOMP_DOACROSS_ULL_POST; | |
3209 | gimple *g | |
3210 | = gimple_build_call (builtin_decl_explicit (source_ix), 1, | |
3211 | build_fold_addr_expr (counts[fd->ordered])); | |
3212 | gimple_set_location (g, loc); | |
3213 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
3214 | } | |
3215 | ||
3216 | /* Expand a single depend from #pragma omp ordered depend(sink:...). */ | |
3217 | ||
3218 | static void | |
3219 | expand_omp_ordered_sink (gimple_stmt_iterator *gsi, struct omp_for_data *fd, | |
3220 | tree *counts, tree c, location_t loc) | |
3221 | { | |
3222 | auto_vec<tree, 10> args; | |
3223 | enum built_in_function sink_ix | |
3224 | = fd->iter_type == long_integer_type_node | |
3225 | ? BUILT_IN_GOMP_DOACROSS_WAIT : BUILT_IN_GOMP_DOACROSS_ULL_WAIT; | |
3226 | tree t, off, coff = NULL_TREE, deps = OMP_CLAUSE_DECL (c), cond = NULL_TREE; | |
3227 | int i; | |
3228 | gimple_stmt_iterator gsi2 = *gsi; | |
3229 | bool warned_step = false; | |
3230 | ||
3231 | for (i = 0; i < fd->ordered; i++) | |
3232 | { | |
3233 | tree step = NULL_TREE; | |
3234 | off = TREE_PURPOSE (deps); | |
3235 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) | |
3236 | { | |
3237 | step = TREE_OPERAND (off, 1); | |
3238 | off = TREE_OPERAND (off, 0); | |
3239 | } | |
3240 | if (!integer_zerop (off)) | |
3241 | { | |
3242 | gcc_assert (fd->loops[i].cond_code == LT_EXPR | |
3243 | || fd->loops[i].cond_code == GT_EXPR); | |
3244 | bool forward = fd->loops[i].cond_code == LT_EXPR; | |
3245 | if (step) | |
3246 | { | |
3247 | /* Non-simple Fortran DO loops. If step is variable, | |
3248 | we don't know at compile even the direction, so can't | |
3249 | warn. */ | |
3250 | if (TREE_CODE (step) != INTEGER_CST) | |
3251 | break; | |
3252 | forward = tree_int_cst_sgn (step) != -1; | |
3253 | } | |
3254 | if (forward ^ OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3255 | warning_at (loc, 0, "%<depend%> clause with %<sink%> modifier " | |
3256 | "waiting for lexically later iteration"); | |
3257 | break; | |
3258 | } | |
3259 | deps = TREE_CHAIN (deps); | |
3260 | } | |
3261 | /* If all offsets corresponding to the collapsed loops are zero, | |
3262 | this depend clause can be ignored. FIXME: but there is still a | |
3263 | flush needed. We need to emit one __sync_synchronize () for it | |
3264 | though (perhaps conditionally)? Solve this together with the | |
3265 | conservative dependence folding optimization. | |
3266 | if (i >= fd->collapse) | |
3267 | return; */ | |
3268 | ||
3269 | deps = OMP_CLAUSE_DECL (c); | |
3270 | gsi_prev (&gsi2); | |
3271 | edge e1 = split_block (gsi_bb (gsi2), gsi_stmt (gsi2)); | |
3272 | edge e2 = split_block_after_labels (e1->dest); | |
3273 | ||
3274 | gsi2 = gsi_after_labels (e1->dest); | |
3275 | *gsi = gsi_last_bb (e1->src); | |
3276 | for (i = 0; i < fd->ordered; i++) | |
3277 | { | |
3278 | tree itype = TREE_TYPE (fd->loops[i].v); | |
3279 | tree step = NULL_TREE; | |
3280 | tree orig_off = NULL_TREE; | |
3281 | if (POINTER_TYPE_P (itype)) | |
3282 | itype = sizetype; | |
3283 | if (i) | |
3284 | deps = TREE_CHAIN (deps); | |
3285 | off = TREE_PURPOSE (deps); | |
3286 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) | |
3287 | { | |
3288 | step = TREE_OPERAND (off, 1); | |
3289 | off = TREE_OPERAND (off, 0); | |
3290 | gcc_assert (fd->loops[i].cond_code == LT_EXPR | |
3291 | && integer_onep (fd->loops[i].step) | |
3292 | && !POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))); | |
3293 | } | |
3294 | tree s = fold_convert_loc (loc, itype, step ? step : fd->loops[i].step); | |
3295 | if (step) | |
3296 | { | |
3297 | off = fold_convert_loc (loc, itype, off); | |
3298 | orig_off = off; | |
3299 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); | |
3300 | } | |
3301 | ||
3302 | if (integer_zerop (off)) | |
3303 | t = boolean_true_node; | |
3304 | else | |
3305 | { | |
3306 | tree a; | |
3307 | tree co = fold_convert_loc (loc, itype, off); | |
3308 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
3309 | { | |
3310 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3311 | co = fold_build1_loc (loc, NEGATE_EXPR, itype, co); | |
3312 | a = fold_build2_loc (loc, POINTER_PLUS_EXPR, | |
3313 | TREE_TYPE (fd->loops[i].v), fd->loops[i].v, | |
3314 | co); | |
3315 | } | |
3316 | else if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3317 | a = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
3318 | fd->loops[i].v, co); | |
3319 | else | |
3320 | a = fold_build2_loc (loc, PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
3321 | fd->loops[i].v, co); | |
3322 | if (step) | |
3323 | { | |
3324 | tree t1, t2; | |
3325 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3326 | t1 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, | |
3327 | fd->loops[i].n1); | |
3328 | else | |
3329 | t1 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
3330 | fd->loops[i].n2); | |
3331 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3332 | t2 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
3333 | fd->loops[i].n2); | |
3334 | else | |
3335 | t2 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, | |
3336 | fd->loops[i].n1); | |
3337 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, | |
3338 | step, build_int_cst (TREE_TYPE (step), 0)); | |
3339 | if (TREE_CODE (step) != INTEGER_CST) | |
3340 | { | |
3341 | t1 = unshare_expr (t1); | |
3342 | t1 = force_gimple_operand_gsi (gsi, t1, true, NULL_TREE, | |
3343 | false, GSI_CONTINUE_LINKING); | |
3344 | t2 = unshare_expr (t2); | |
3345 | t2 = force_gimple_operand_gsi (gsi, t2, true, NULL_TREE, | |
3346 | false, GSI_CONTINUE_LINKING); | |
3347 | } | |
3348 | t = fold_build3_loc (loc, COND_EXPR, boolean_type_node, | |
3349 | t, t2, t1); | |
3350 | } | |
3351 | else if (fd->loops[i].cond_code == LT_EXPR) | |
3352 | { | |
3353 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3354 | t = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, | |
3355 | fd->loops[i].n1); | |
3356 | else | |
3357 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
3358 | fd->loops[i].n2); | |
3359 | } | |
3360 | else if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3361 | t = fold_build2_loc (loc, GT_EXPR, boolean_type_node, a, | |
3362 | fd->loops[i].n2); | |
3363 | else | |
3364 | t = fold_build2_loc (loc, LE_EXPR, boolean_type_node, a, | |
3365 | fd->loops[i].n1); | |
3366 | } | |
3367 | if (cond) | |
3368 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, cond, t); | |
3369 | else | |
3370 | cond = t; | |
3371 | ||
3372 | off = fold_convert_loc (loc, itype, off); | |
3373 | ||
3374 | if (step | |
3375 | || (fd->loops[i].cond_code == LT_EXPR | |
3376 | ? !integer_onep (fd->loops[i].step) | |
3377 | : !integer_minus_onep (fd->loops[i].step))) | |
3378 | { | |
3379 | if (step == NULL_TREE | |
3380 | && TYPE_UNSIGNED (itype) | |
3381 | && fd->loops[i].cond_code == GT_EXPR) | |
3382 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, off, | |
3383 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
3384 | s)); | |
3385 | else | |
3386 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, | |
3387 | orig_off ? orig_off : off, s); | |
3388 | t = fold_build2_loc (loc, EQ_EXPR, boolean_type_node, t, | |
3389 | build_int_cst (itype, 0)); | |
3390 | if (integer_zerop (t) && !warned_step) | |
3391 | { | |
3392 | warning_at (loc, 0, "%<depend%> clause with %<sink%> modifier " | |
3393 | "refers to iteration never in the iteration " | |
3394 | "space"); | |
3395 | warned_step = true; | |
3396 | } | |
3397 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, | |
3398 | cond, t); | |
3399 | } | |
3400 | ||
3401 | if (i <= fd->collapse - 1 && fd->collapse > 1) | |
3402 | t = fd->loop.v; | |
3403 | else if (counts[i]) | |
3404 | t = counts[i]; | |
3405 | else | |
3406 | { | |
3407 | t = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
3408 | fd->loops[i].v, fd->loops[i].n1); | |
3409 | t = fold_convert_loc (loc, fd->iter_type, t); | |
3410 | } | |
3411 | if (step) | |
3412 | /* We have divided off by step already earlier. */; | |
3413 | else if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
3414 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, | |
3415 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
3416 | s)); | |
3417 | else | |
3418 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); | |
3419 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
3420 | off = fold_build1_loc (loc, NEGATE_EXPR, itype, off); | |
3421 | off = fold_convert_loc (loc, fd->iter_type, off); | |
3422 | if (i <= fd->collapse - 1 && fd->collapse > 1) | |
3423 | { | |
3424 | if (i) | |
3425 | off = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, coff, | |
3426 | off); | |
3427 | if (i < fd->collapse - 1) | |
3428 | { | |
3429 | coff = fold_build2_loc (loc, MULT_EXPR, fd->iter_type, off, | |
3430 | counts[i]); | |
3431 | continue; | |
3432 | } | |
3433 | } | |
3434 | off = unshare_expr (off); | |
3435 | t = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, t, off); | |
3436 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
3437 | true, GSI_SAME_STMT); | |
3438 | args.safe_push (t); | |
3439 | } | |
3440 | gimple *g = gimple_build_call_vec (builtin_decl_explicit (sink_ix), args); | |
3441 | gimple_set_location (g, loc); | |
3442 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); | |
3443 | ||
3444 | cond = unshare_expr (cond); | |
3445 | cond = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE, false, | |
3446 | GSI_CONTINUE_LINKING); | |
3447 | gsi_insert_after (gsi, gimple_build_cond_empty (cond), GSI_NEW_STMT); | |
3448 | edge e3 = make_edge (e1->src, e2->dest, EDGE_FALSE_VALUE); | |
3449 | e3->probability = profile_probability::guessed_always ().apply_scale (1, 8); | |
3450 | e1->probability = e3->probability.invert (); | |
3451 | e1->flags = EDGE_TRUE_VALUE; | |
3452 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e1->src); | |
3453 | ||
3454 | *gsi = gsi_after_labels (e2->dest); | |
3455 | } | |
3456 | ||
3457 | /* Expand all #pragma omp ordered depend(source) and | |
3458 | #pragma omp ordered depend(sink:...) constructs in the current | |
3459 | #pragma omp for ordered(n) region. */ | |
3460 | ||
3461 | static void | |
3462 | expand_omp_ordered_source_sink (struct omp_region *region, | |
3463 | struct omp_for_data *fd, tree *counts, | |
3464 | basic_block cont_bb) | |
3465 | { | |
3466 | struct omp_region *inner; | |
3467 | int i; | |
3468 | for (i = fd->collapse - 1; i < fd->ordered; i++) | |
3469 | if (i == fd->collapse - 1 && fd->collapse > 1) | |
3470 | counts[i] = NULL_TREE; | |
3471 | else if (i >= fd->collapse && !cont_bb) | |
3472 | counts[i] = build_zero_cst (fd->iter_type); | |
3473 | else if (!POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v)) | |
3474 | && integer_onep (fd->loops[i].step)) | |
3475 | counts[i] = NULL_TREE; | |
3476 | else | |
3477 | counts[i] = create_tmp_var (fd->iter_type, ".orditer"); | |
3478 | tree atype | |
3479 | = build_array_type_nelts (fd->iter_type, fd->ordered - fd->collapse + 1); | |
3480 | counts[fd->ordered] = create_tmp_var (atype, ".orditera"); | |
3481 | TREE_ADDRESSABLE (counts[fd->ordered]) = 1; | |
3482 | ||
3483 | for (inner = region->inner; inner; inner = inner->next) | |
3484 | if (inner->type == GIMPLE_OMP_ORDERED) | |
3485 | { | |
3486 | gomp_ordered *ord_stmt = inner->ord_stmt; | |
3487 | gimple_stmt_iterator gsi = gsi_for_stmt (ord_stmt); | |
3488 | location_t loc = gimple_location (ord_stmt); | |
3489 | tree c; | |
3490 | for (c = gimple_omp_ordered_clauses (ord_stmt); | |
3491 | c; c = OMP_CLAUSE_CHAIN (c)) | |
3492 | if (OMP_CLAUSE_DEPEND_KIND (c) == OMP_CLAUSE_DEPEND_SOURCE) | |
3493 | break; | |
3494 | if (c) | |
3495 | expand_omp_ordered_source (&gsi, fd, counts, loc); | |
3496 | for (c = gimple_omp_ordered_clauses (ord_stmt); | |
3497 | c; c = OMP_CLAUSE_CHAIN (c)) | |
3498 | if (OMP_CLAUSE_DEPEND_KIND (c) == OMP_CLAUSE_DEPEND_SINK) | |
3499 | expand_omp_ordered_sink (&gsi, fd, counts, c, loc); | |
3500 | gsi_remove (&gsi, true); | |
3501 | } | |
3502 | } | |
3503 | ||
3504 | /* Wrap the body into fd->ordered - fd->collapse loops that aren't | |
3505 | collapsed. */ | |
3506 | ||
3507 | static basic_block | |
3508 | expand_omp_for_ordered_loops (struct omp_for_data *fd, tree *counts, | |
3509 | basic_block cont_bb, basic_block body_bb, | |
3510 | bool ordered_lastprivate) | |
3511 | { | |
3512 | if (fd->ordered == fd->collapse) | |
3513 | return cont_bb; | |
3514 | ||
3515 | if (!cont_bb) | |
3516 | { | |
3517 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
3518 | for (int i = fd->collapse; i < fd->ordered; i++) | |
3519 | { | |
3520 | tree type = TREE_TYPE (fd->loops[i].v); | |
3521 | tree n1 = fold_convert (type, fd->loops[i].n1); | |
3522 | expand_omp_build_assign (&gsi, fd->loops[i].v, n1); | |
3523 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3524 | size_int (i - fd->collapse + 1), | |
3525 | NULL_TREE, NULL_TREE); | |
3526 | expand_omp_build_assign (&gsi, aref, build_zero_cst (fd->iter_type)); | |
3527 | } | |
3528 | return NULL; | |
3529 | } | |
3530 | ||
3531 | for (int i = fd->ordered - 1; i >= fd->collapse; i--) | |
3532 | { | |
3533 | tree t, type = TREE_TYPE (fd->loops[i].v); | |
3534 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
3535 | expand_omp_build_assign (&gsi, fd->loops[i].v, | |
3536 | fold_convert (type, fd->loops[i].n1)); | |
3537 | if (counts[i]) | |
3538 | expand_omp_build_assign (&gsi, counts[i], | |
3539 | build_zero_cst (fd->iter_type)); | |
3540 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3541 | size_int (i - fd->collapse + 1), | |
3542 | NULL_TREE, NULL_TREE); | |
3543 | expand_omp_build_assign (&gsi, aref, build_zero_cst (fd->iter_type)); | |
3544 | if (!gsi_end_p (gsi)) | |
3545 | gsi_prev (&gsi); | |
3546 | else | |
3547 | gsi = gsi_last_bb (body_bb); | |
3548 | edge e1 = split_block (body_bb, gsi_stmt (gsi)); | |
3549 | basic_block new_body = e1->dest; | |
3550 | if (body_bb == cont_bb) | |
3551 | cont_bb = new_body; | |
3552 | edge e2 = NULL; | |
3553 | basic_block new_header; | |
3554 | if (EDGE_COUNT (cont_bb->preds) > 0) | |
3555 | { | |
3556 | gsi = gsi_last_bb (cont_bb); | |
3557 | if (POINTER_TYPE_P (type)) | |
3558 | t = fold_build_pointer_plus (fd->loops[i].v, | |
3559 | fold_convert (sizetype, | |
3560 | fd->loops[i].step)); | |
3561 | else | |
3562 | t = fold_build2 (PLUS_EXPR, type, fd->loops[i].v, | |
3563 | fold_convert (type, fd->loops[i].step)); | |
3564 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
3565 | if (counts[i]) | |
3566 | { | |
3567 | t = fold_build2 (PLUS_EXPR, fd->iter_type, counts[i], | |
3568 | build_int_cst (fd->iter_type, 1)); | |
3569 | expand_omp_build_assign (&gsi, counts[i], t); | |
3570 | t = counts[i]; | |
3571 | } | |
3572 | else | |
3573 | { | |
3574 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
3575 | fd->loops[i].v, fd->loops[i].n1); | |
3576 | t = fold_convert (fd->iter_type, t); | |
3577 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
3578 | true, GSI_SAME_STMT); | |
3579 | } | |
3580 | aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3581 | size_int (i - fd->collapse + 1), | |
3582 | NULL_TREE, NULL_TREE); | |
3583 | expand_omp_build_assign (&gsi, aref, t); | |
3584 | gsi_prev (&gsi); | |
3585 | e2 = split_block (cont_bb, gsi_stmt (gsi)); | |
3586 | new_header = e2->dest; | |
3587 | } | |
3588 | else | |
3589 | new_header = cont_bb; | |
3590 | gsi = gsi_after_labels (new_header); | |
3591 | tree v = force_gimple_operand_gsi (&gsi, fd->loops[i].v, true, NULL_TREE, | |
3592 | true, GSI_SAME_STMT); | |
3593 | tree n2 | |
3594 | = force_gimple_operand_gsi (&gsi, fold_convert (type, fd->loops[i].n2), | |
3595 | true, NULL_TREE, true, GSI_SAME_STMT); | |
3596 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, n2); | |
3597 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_NEW_STMT); | |
3598 | edge e3 = split_block (new_header, gsi_stmt (gsi)); | |
3599 | cont_bb = e3->dest; | |
3600 | remove_edge (e1); | |
3601 | make_edge (body_bb, new_header, EDGE_FALLTHRU); | |
3602 | e3->flags = EDGE_FALSE_VALUE; | |
3603 | e3->probability = profile_probability::guessed_always ().apply_scale (1, 8); | |
3604 | e1 = make_edge (new_header, new_body, EDGE_TRUE_VALUE); | |
3605 | e1->probability = e3->probability.invert (); | |
3606 | ||
3607 | set_immediate_dominator (CDI_DOMINATORS, new_header, body_bb); | |
3608 | set_immediate_dominator (CDI_DOMINATORS, new_body, new_header); | |
3609 | ||
3610 | if (e2) | |
3611 | { | |
3612 | class loop *loop = alloc_loop (); | |
3613 | loop->header = new_header; | |
3614 | loop->latch = e2->src; | |
3615 | add_loop (loop, body_bb->loop_father); | |
3616 | } | |
3617 | } | |
3618 | ||
3619 | /* If there are any lastprivate clauses and it is possible some loops | |
3620 | might have zero iterations, ensure all the decls are initialized, | |
3621 | otherwise we could crash evaluating C++ class iterators with lastprivate | |
3622 | clauses. */ | |
3623 | bool need_inits = false; | |
3624 | for (int i = fd->collapse; ordered_lastprivate && i < fd->ordered; i++) | |
3625 | if (need_inits) | |
3626 | { | |
3627 | tree type = TREE_TYPE (fd->loops[i].v); | |
3628 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
3629 | expand_omp_build_assign (&gsi, fd->loops[i].v, | |
3630 | fold_convert (type, fd->loops[i].n1)); | |
3631 | } | |
3632 | else | |
3633 | { | |
3634 | tree type = TREE_TYPE (fd->loops[i].v); | |
3635 | tree this_cond = fold_build2 (fd->loops[i].cond_code, | |
3636 | boolean_type_node, | |
3637 | fold_convert (type, fd->loops[i].n1), | |
3638 | fold_convert (type, fd->loops[i].n2)); | |
3639 | if (!integer_onep (this_cond)) | |
3640 | need_inits = true; | |
3641 | } | |
3642 | ||
3643 | return cont_bb; | |
3644 | } | |
3645 | ||
3646 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
3647 | loop with any schedule. Given parameters: | |
3648 | ||
3649 | for (V = N1; V cond N2; V += STEP) BODY; | |
3650 | ||
3651 | where COND is "<" or ">", we generate pseudocode | |
3652 | ||
3653 | more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0); | |
3654 | if (more) goto L0; else goto L3; | |
3655 | L0: | |
3656 | V = istart0; | |
3657 | iend = iend0; | |
3658 | L1: | |
3659 | BODY; | |
3660 | V += STEP; | |
3661 | if (V cond iend) goto L1; else goto L2; | |
3662 | L2: | |
3663 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
3664 | L3: | |
3665 | ||
3666 | If this is a combined omp parallel loop, instead of the call to | |
3667 | GOMP_loop_foo_start, we call GOMP_loop_foo_next. | |
3668 | If this is gimple_omp_for_combined_p loop, then instead of assigning | |
3669 | V and iend in L0 we assign the first two _looptemp_ clause decls of the | |
3670 | inner GIMPLE_OMP_FOR and V += STEP; and | |
3671 | if (V cond iend) goto L1; else goto L2; are removed. | |
3672 | ||
3673 | For collapsed loops, given parameters: | |
3674 | collapse(3) | |
3675 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
3676 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
3677 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
3678 | BODY; | |
3679 | ||
3680 | we generate pseudocode | |
3681 | ||
3682 | if (__builtin_expect (N32 cond3 N31, 0)) goto Z0; | |
3683 | if (cond3 is <) | |
3684 | adj = STEP3 - 1; | |
3685 | else | |
3686 | adj = STEP3 + 1; | |
3687 | count3 = (adj + N32 - N31) / STEP3; | |
3688 | if (__builtin_expect (N22 cond2 N21, 0)) goto Z0; | |
3689 | if (cond2 is <) | |
3690 | adj = STEP2 - 1; | |
3691 | else | |
3692 | adj = STEP2 + 1; | |
3693 | count2 = (adj + N22 - N21) / STEP2; | |
3694 | if (__builtin_expect (N12 cond1 N11, 0)) goto Z0; | |
3695 | if (cond1 is <) | |
3696 | adj = STEP1 - 1; | |
3697 | else | |
3698 | adj = STEP1 + 1; | |
3699 | count1 = (adj + N12 - N11) / STEP1; | |
3700 | count = count1 * count2 * count3; | |
3701 | goto Z1; | |
3702 | Z0: | |
3703 | count = 0; | |
3704 | Z1: | |
3705 | more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0); | |
3706 | if (more) goto L0; else goto L3; | |
3707 | L0: | |
3708 | V = istart0; | |
3709 | T = V; | |
3710 | V3 = N31 + (T % count3) * STEP3; | |
3711 | T = T / count3; | |
3712 | V2 = N21 + (T % count2) * STEP2; | |
3713 | T = T / count2; | |
3714 | V1 = N11 + T * STEP1; | |
3715 | iend = iend0; | |
3716 | L1: | |
3717 | BODY; | |
3718 | V += 1; | |
3719 | if (V < iend) goto L10; else goto L2; | |
3720 | L10: | |
3721 | V3 += STEP3; | |
3722 | if (V3 cond3 N32) goto L1; else goto L11; | |
3723 | L11: | |
3724 | V3 = N31; | |
3725 | V2 += STEP2; | |
3726 | if (V2 cond2 N22) goto L1; else goto L12; | |
3727 | L12: | |
3728 | V2 = N21; | |
3729 | V1 += STEP1; | |
3730 | goto L1; | |
3731 | L2: | |
3732 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
3733 | L3: | |
3734 | ||
3735 | */ | |
3736 | ||
3737 | static void | |
3738 | expand_omp_for_generic (struct omp_region *region, | |
3739 | struct omp_for_data *fd, | |
3740 | enum built_in_function start_fn, | |
3741 | enum built_in_function next_fn, | |
3742 | tree sched_arg, | |
3743 | gimple *inner_stmt) | |
3744 | { | |
3745 | tree type, istart0, iend0, iend; | |
3746 | tree t, vmain, vback, bias = NULL_TREE; | |
3747 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, collapse_bb; | |
3748 | basic_block l2_bb = NULL, l3_bb = NULL; | |
3749 | gimple_stmt_iterator gsi; | |
3750 | gassign *assign_stmt; | |
3751 | bool in_combined_parallel = is_combined_parallel (region); | |
3752 | bool broken_loop = region->cont == NULL; | |
3753 | edge e, ne; | |
3754 | tree *counts = NULL; | |
3755 | int i; | |
3756 | bool ordered_lastprivate = false; | |
3757 | ||
3758 | gcc_assert (!broken_loop || !in_combined_parallel); | |
3759 | gcc_assert (fd->iter_type == long_integer_type_node | |
3760 | || !in_combined_parallel); | |
3761 | ||
3762 | entry_bb = region->entry; | |
3763 | cont_bb = region->cont; | |
3764 | collapse_bb = NULL; | |
3765 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
3766 | gcc_assert (broken_loop | |
3767 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
3768 | l0_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
3769 | l1_bb = single_succ (l0_bb); | |
3770 | if (!broken_loop) | |
3771 | { | |
3772 | l2_bb = create_empty_bb (cont_bb); | |
3773 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb | |
3774 | || (single_succ_edge (BRANCH_EDGE (cont_bb)->dest)->dest | |
3775 | == l1_bb)); | |
3776 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
3777 | } | |
3778 | else | |
3779 | l2_bb = NULL; | |
3780 | l3_bb = BRANCH_EDGE (entry_bb)->dest; | |
3781 | exit_bb = region->exit; | |
3782 | ||
3783 | gsi = gsi_last_nondebug_bb (entry_bb); | |
3784 | ||
3785 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
3786 | if (fd->ordered | |
3787 | && omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
3788 | OMP_CLAUSE_LASTPRIVATE)) | |
3789 | ordered_lastprivate = false; | |
3790 | tree reductions = NULL_TREE; | |
3791 | tree mem = NULL_TREE, cond_var = NULL_TREE, condtemp = NULL_TREE; | |
3792 | tree memv = NULL_TREE; | |
3793 | if (fd->lastprivate_conditional) | |
3794 | { | |
3795 | tree c = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
3796 | OMP_CLAUSE__CONDTEMP_); | |
3797 | if (fd->have_pointer_condtemp) | |
3798 | condtemp = OMP_CLAUSE_DECL (c); | |
3799 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
3800 | cond_var = OMP_CLAUSE_DECL (c); | |
3801 | } | |
3802 | if (sched_arg) | |
3803 | { | |
3804 | if (fd->have_reductemp) | |
3805 | { | |
3806 | tree c = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
3807 | OMP_CLAUSE__REDUCTEMP_); | |
3808 | reductions = OMP_CLAUSE_DECL (c); | |
3809 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
3810 | gimple *g = SSA_NAME_DEF_STMT (reductions); | |
3811 | reductions = gimple_assign_rhs1 (g); | |
3812 | OMP_CLAUSE_DECL (c) = reductions; | |
3813 | entry_bb = gimple_bb (g); | |
3814 | edge e = split_block (entry_bb, g); | |
3815 | if (region->entry == entry_bb) | |
3816 | region->entry = e->dest; | |
3817 | gsi = gsi_last_bb (entry_bb); | |
3818 | } | |
3819 | else | |
3820 | reductions = null_pointer_node; | |
3821 | if (fd->have_pointer_condtemp) | |
3822 | { | |
3823 | tree type = TREE_TYPE (condtemp); | |
3824 | memv = create_tmp_var (type); | |
3825 | TREE_ADDRESSABLE (memv) = 1; | |
3826 | unsigned HOST_WIDE_INT sz | |
3827 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
3828 | sz *= fd->lastprivate_conditional; | |
3829 | expand_omp_build_assign (&gsi, memv, build_int_cst (type, sz), | |
3830 | false); | |
3831 | mem = build_fold_addr_expr (memv); | |
3832 | } | |
3833 | else | |
3834 | mem = null_pointer_node; | |
3835 | } | |
3836 | if (fd->collapse > 1 || fd->ordered) | |
3837 | { | |
3838 | int first_zero_iter1 = -1, first_zero_iter2 = -1; | |
3839 | basic_block zero_iter1_bb = NULL, zero_iter2_bb = NULL, l2_dom_bb = NULL; | |
3840 | ||
3841 | counts = XALLOCAVEC (tree, fd->ordered ? fd->ordered + 1 : fd->collapse); | |
3842 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
3843 | zero_iter1_bb, first_zero_iter1, | |
3844 | zero_iter2_bb, first_zero_iter2, l2_dom_bb); | |
3845 | ||
3846 | if (zero_iter1_bb) | |
3847 | { | |
3848 | /* Some counts[i] vars might be uninitialized if | |
3849 | some loop has zero iterations. But the body shouldn't | |
3850 | be executed in that case, so just avoid uninit warnings. */ | |
3851 | for (i = first_zero_iter1; | |
3852 | i < (fd->ordered ? fd->ordered : fd->collapse); i++) | |
3853 | if (SSA_VAR_P (counts[i])) | |
3854 | suppress_warning (counts[i], OPT_Wuninitialized); | |
3855 | gsi_prev (&gsi); | |
3856 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
3857 | entry_bb = e->dest; | |
3858 | make_edge (zero_iter1_bb, entry_bb, EDGE_FALLTHRU); | |
3859 | gsi = gsi_last_nondebug_bb (entry_bb); | |
3860 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, | |
3861 | get_immediate_dominator (CDI_DOMINATORS, | |
3862 | zero_iter1_bb)); | |
3863 | } | |
3864 | if (zero_iter2_bb) | |
3865 | { | |
3866 | /* Some counts[i] vars might be uninitialized if | |
3867 | some loop has zero iterations. But the body shouldn't | |
3868 | be executed in that case, so just avoid uninit warnings. */ | |
3869 | for (i = first_zero_iter2; i < fd->ordered; i++) | |
3870 | if (SSA_VAR_P (counts[i])) | |
3871 | suppress_warning (counts[i], OPT_Wuninitialized); | |
3872 | if (zero_iter1_bb) | |
3873 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); | |
3874 | else | |
3875 | { | |
3876 | gsi_prev (&gsi); | |
3877 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
3878 | entry_bb = e->dest; | |
3879 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); | |
3880 | gsi = gsi_last_nondebug_bb (entry_bb); | |
3881 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, | |
3882 | get_immediate_dominator | |
3883 | (CDI_DOMINATORS, zero_iter2_bb)); | |
3884 | } | |
3885 | } | |
3886 | if (fd->collapse == 1) | |
3887 | { | |
3888 | counts[0] = fd->loop.n2; | |
3889 | fd->loop = fd->loops[0]; | |
3890 | } | |
3891 | } | |
3892 | ||
3893 | type = TREE_TYPE (fd->loop.v); | |
3894 | istart0 = create_tmp_var (fd->iter_type, ".istart0"); | |
3895 | iend0 = create_tmp_var (fd->iter_type, ".iend0"); | |
3896 | TREE_ADDRESSABLE (istart0) = 1; | |
3897 | TREE_ADDRESSABLE (iend0) = 1; | |
3898 | ||
3899 | /* See if we need to bias by LLONG_MIN. */ | |
3900 | if (fd->iter_type == long_long_unsigned_type_node | |
3901 | && TREE_CODE (type) == INTEGER_TYPE | |
3902 | && !TYPE_UNSIGNED (type) | |
3903 | && fd->ordered == 0) | |
3904 | { | |
3905 | tree n1, n2; | |
3906 | ||
3907 | if (fd->loop.cond_code == LT_EXPR) | |
3908 | { | |
3909 | n1 = fd->loop.n1; | |
3910 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
3911 | } | |
3912 | else | |
3913 | { | |
3914 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
3915 | n2 = fd->loop.n1; | |
3916 | } | |
3917 | if (TREE_CODE (n1) != INTEGER_CST | |
3918 | || TREE_CODE (n2) != INTEGER_CST | |
3919 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
3920 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
3921 | } | |
3922 | ||
3923 | gimple_stmt_iterator gsif = gsi; | |
3924 | gsi_prev (&gsif); | |
3925 | ||
3926 | tree arr = NULL_TREE; | |
3927 | if (in_combined_parallel) | |
3928 | { | |
3929 | gcc_assert (fd->ordered == 0); | |
3930 | /* In a combined parallel loop, emit a call to | |
3931 | GOMP_loop_foo_next. */ | |
3932 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, | |
3933 | build_fold_addr_expr (istart0), | |
3934 | build_fold_addr_expr (iend0)); | |
3935 | } | |
3936 | else | |
3937 | { | |
3938 | tree t0, t1, t2, t3, t4; | |
3939 | /* If this is not a combined parallel loop, emit a call to | |
3940 | GOMP_loop_foo_start in ENTRY_BB. */ | |
3941 | t4 = build_fold_addr_expr (iend0); | |
3942 | t3 = build_fold_addr_expr (istart0); | |
3943 | if (fd->ordered) | |
3944 | { | |
3945 | t0 = build_int_cst (unsigned_type_node, | |
3946 | fd->ordered - fd->collapse + 1); | |
3947 | arr = create_tmp_var (build_array_type_nelts (fd->iter_type, | |
3948 | fd->ordered | |
3949 | - fd->collapse + 1), | |
3950 | ".omp_counts"); | |
3951 | DECL_NAMELESS (arr) = 1; | |
3952 | TREE_ADDRESSABLE (arr) = 1; | |
3953 | TREE_STATIC (arr) = 1; | |
3954 | vec<constructor_elt, va_gc> *v; | |
3955 | vec_alloc (v, fd->ordered - fd->collapse + 1); | |
3956 | int idx; | |
3957 | ||
3958 | for (idx = 0; idx < fd->ordered - fd->collapse + 1; idx++) | |
3959 | { | |
3960 | tree c; | |
3961 | if (idx == 0 && fd->collapse > 1) | |
3962 | c = fd->loop.n2; | |
3963 | else | |
3964 | c = counts[idx + fd->collapse - 1]; | |
3965 | tree purpose = size_int (idx); | |
3966 | CONSTRUCTOR_APPEND_ELT (v, purpose, c); | |
3967 | if (TREE_CODE (c) != INTEGER_CST) | |
3968 | TREE_STATIC (arr) = 0; | |
3969 | } | |
3970 | ||
3971 | DECL_INITIAL (arr) = build_constructor (TREE_TYPE (arr), v); | |
3972 | if (!TREE_STATIC (arr)) | |
3973 | force_gimple_operand_gsi (&gsi, build1 (DECL_EXPR, | |
3974 | void_type_node, arr), | |
3975 | true, NULL_TREE, true, GSI_SAME_STMT); | |
3976 | t1 = build_fold_addr_expr (arr); | |
3977 | t2 = NULL_TREE; | |
3978 | } | |
3979 | else | |
3980 | { | |
3981 | t2 = fold_convert (fd->iter_type, fd->loop.step); | |
3982 | t1 = fd->loop.n2; | |
3983 | t0 = fd->loop.n1; | |
3984 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
3985 | { | |
3986 | tree innerc | |
3987 | = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
3988 | OMP_CLAUSE__LOOPTEMP_); | |
3989 | gcc_assert (innerc); | |
3990 | t0 = OMP_CLAUSE_DECL (innerc); | |
3991 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
3992 | OMP_CLAUSE__LOOPTEMP_); | |
3993 | gcc_assert (innerc); | |
3994 | t1 = OMP_CLAUSE_DECL (innerc); | |
3995 | } | |
3996 | if (POINTER_TYPE_P (TREE_TYPE (t0)) | |
3997 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
3998 | != TYPE_PRECISION (fd->iter_type)) | |
3999 | { | |
4000 | /* Avoid casting pointers to integer of a different size. */ | |
4001 | tree itype = signed_type_for (type); | |
4002 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); | |
4003 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
4004 | } | |
4005 | else | |
4006 | { | |
4007 | t1 = fold_convert (fd->iter_type, t1); | |
4008 | t0 = fold_convert (fd->iter_type, t0); | |
4009 | } | |
4010 | if (bias) | |
4011 | { | |
4012 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); | |
4013 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
4014 | } | |
4015 | } | |
4016 | if (fd->iter_type == long_integer_type_node || fd->ordered) | |
4017 | { | |
4018 | if (fd->chunk_size) | |
4019 | { | |
4020 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
4021 | t = omp_adjust_chunk_size (t, fd->simd_schedule); | |
4022 | if (sched_arg) | |
4023 | { | |
4024 | if (fd->ordered) | |
4025 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4026 | 8, t0, t1, sched_arg, t, t3, t4, | |
4027 | reductions, mem); | |
4028 | else | |
4029 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4030 | 9, t0, t1, t2, sched_arg, t, t3, t4, | |
4031 | reductions, mem); | |
4032 | } | |
4033 | else if (fd->ordered) | |
4034 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4035 | 5, t0, t1, t, t3, t4); | |
4036 | else | |
4037 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4038 | 6, t0, t1, t2, t, t3, t4); | |
4039 | } | |
4040 | else if (fd->ordered) | |
4041 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4042 | 4, t0, t1, t3, t4); | |
4043 | else | |
4044 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4045 | 5, t0, t1, t2, t3, t4); | |
4046 | } | |
4047 | else | |
4048 | { | |
4049 | tree t5; | |
4050 | tree c_bool_type; | |
4051 | tree bfn_decl; | |
4052 | ||
4053 | /* The GOMP_loop_ull_*start functions have additional boolean | |
4054 | argument, true for < loops and false for > loops. | |
4055 | In Fortran, the C bool type can be different from | |
4056 | boolean_type_node. */ | |
4057 | bfn_decl = builtin_decl_explicit (start_fn); | |
4058 | c_bool_type = TREE_TYPE (TREE_TYPE (bfn_decl)); | |
4059 | t5 = build_int_cst (c_bool_type, | |
4060 | fd->loop.cond_code == LT_EXPR ? 1 : 0); | |
4061 | if (fd->chunk_size) | |
4062 | { | |
4063 | tree bfn_decl = builtin_decl_explicit (start_fn); | |
4064 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
4065 | t = omp_adjust_chunk_size (t, fd->simd_schedule); | |
4066 | if (sched_arg) | |
4067 | t = build_call_expr (bfn_decl, 10, t5, t0, t1, t2, sched_arg, | |
4068 | t, t3, t4, reductions, mem); | |
4069 | else | |
4070 | t = build_call_expr (bfn_decl, 7, t5, t0, t1, t2, t, t3, t4); | |
4071 | } | |
4072 | else | |
4073 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4074 | 6, t5, t0, t1, t2, t3, t4); | |
4075 | } | |
4076 | } | |
4077 | if (TREE_TYPE (t) != boolean_type_node) | |
4078 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
4079 | t, build_int_cst (TREE_TYPE (t), 0)); | |
4080 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4081 | true, GSI_SAME_STMT); | |
4082 | if (arr && !TREE_STATIC (arr)) | |
4083 | { | |
4084 | tree clobber = build_clobber (TREE_TYPE (arr)); | |
4085 | gsi_insert_before (&gsi, gimple_build_assign (arr, clobber), | |
4086 | GSI_SAME_STMT); | |
4087 | } | |
4088 | if (fd->have_pointer_condtemp) | |
4089 | expand_omp_build_assign (&gsi, condtemp, memv, false); | |
4090 | if (fd->have_reductemp) | |
4091 | { | |
4092 | gimple *g = gsi_stmt (gsi); | |
4093 | gsi_remove (&gsi, true); | |
4094 | release_ssa_name (gimple_assign_lhs (g)); | |
4095 | ||
4096 | entry_bb = region->entry; | |
4097 | gsi = gsi_last_nondebug_bb (entry_bb); | |
4098 | ||
4099 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
4100 | } | |
4101 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
4102 | ||
4103 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
4104 | gsi_remove (&gsi, true); | |
4105 | ||
4106 | if (gsi_end_p (gsif)) | |
4107 | gsif = gsi_after_labels (gsi_bb (gsif)); | |
4108 | gsi_next (&gsif); | |
4109 | ||
4110 | /* Iteration setup for sequential loop goes in L0_BB. */ | |
4111 | tree startvar = fd->loop.v; | |
4112 | tree endvar = NULL_TREE; | |
4113 | ||
4114 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4115 | { | |
4116 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_FOR | |
4117 | && gimple_omp_for_kind (inner_stmt) | |
4118 | == GF_OMP_FOR_KIND_SIMD); | |
4119 | tree innerc = omp_find_clause (gimple_omp_for_clauses (inner_stmt), | |
4120 | OMP_CLAUSE__LOOPTEMP_); | |
4121 | gcc_assert (innerc); | |
4122 | startvar = OMP_CLAUSE_DECL (innerc); | |
4123 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4124 | OMP_CLAUSE__LOOPTEMP_); | |
4125 | gcc_assert (innerc); | |
4126 | endvar = OMP_CLAUSE_DECL (innerc); | |
4127 | } | |
4128 | ||
4129 | gsi = gsi_start_bb (l0_bb); | |
4130 | t = istart0; | |
4131 | if (fd->ordered && fd->collapse == 1) | |
4132 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, | |
4133 | fold_convert (fd->iter_type, fd->loop.step)); | |
4134 | else if (bias) | |
4135 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); | |
4136 | if (fd->ordered && fd->collapse == 1) | |
4137 | { | |
4138 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4139 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), | |
4140 | fd->loop.n1, fold_convert (sizetype, t)); | |
4141 | else | |
4142 | { | |
4143 | t = fold_convert (TREE_TYPE (startvar), t); | |
4144 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), | |
4145 | fd->loop.n1, t); | |
4146 | } | |
4147 | } | |
4148 | else | |
4149 | { | |
4150 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4151 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
4152 | t = fold_convert (TREE_TYPE (startvar), t); | |
4153 | } | |
4154 | t = force_gimple_operand_gsi (&gsi, t, | |
4155 | DECL_P (startvar) | |
4156 | && TREE_ADDRESSABLE (startvar), | |
4157 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4158 | assign_stmt = gimple_build_assign (startvar, t); | |
4159 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4160 | if (cond_var) | |
4161 | { | |
4162 | tree itype = TREE_TYPE (cond_var); | |
4163 | /* For lastprivate(conditional:) itervar, we need some iteration | |
4164 | counter that starts at unsigned non-zero and increases. | |
4165 | Prefer as few IVs as possible, so if we can use startvar | |
4166 | itself, use that, or startvar + constant (those would be | |
4167 | incremented with step), and as last resort use the s0 + 1 | |
4168 | incremented by 1. */ | |
4169 | if ((fd->ordered && fd->collapse == 1) | |
4170 | || bias | |
4171 | || POINTER_TYPE_P (type) | |
4172 | || TREE_CODE (fd->loop.n1) != INTEGER_CST | |
4173 | || fd->loop.cond_code != LT_EXPR) | |
4174 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, istart0), | |
4175 | build_int_cst (itype, 1)); | |
4176 | else if (tree_int_cst_sgn (fd->loop.n1) == 1) | |
4177 | t = fold_convert (itype, t); | |
4178 | else | |
4179 | { | |
4180 | tree c = fold_convert (itype, fd->loop.n1); | |
4181 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
4182 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
4183 | } | |
4184 | t = force_gimple_operand_gsi (&gsi, t, false, | |
4185 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4186 | assign_stmt = gimple_build_assign (cond_var, t); | |
4187 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4188 | } | |
4189 | ||
4190 | t = iend0; | |
4191 | if (fd->ordered && fd->collapse == 1) | |
4192 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, | |
4193 | fold_convert (fd->iter_type, fd->loop.step)); | |
4194 | else if (bias) | |
4195 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); | |
4196 | if (fd->ordered && fd->collapse == 1) | |
4197 | { | |
4198 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4199 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), | |
4200 | fd->loop.n1, fold_convert (sizetype, t)); | |
4201 | else | |
4202 | { | |
4203 | t = fold_convert (TREE_TYPE (startvar), t); | |
4204 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), | |
4205 | fd->loop.n1, t); | |
4206 | } | |
4207 | } | |
4208 | else | |
4209 | { | |
4210 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4211 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
4212 | t = fold_convert (TREE_TYPE (startvar), t); | |
4213 | } | |
4214 | iend = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4215 | false, GSI_CONTINUE_LINKING); | |
4216 | if (endvar) | |
4217 | { | |
4218 | assign_stmt = gimple_build_assign (endvar, iend); | |
4219 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4220 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (iend))) | |
4221 | assign_stmt = gimple_build_assign (fd->loop.v, iend); | |
4222 | else | |
4223 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, iend); | |
4224 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4225 | } | |
4226 | /* Handle linear clause adjustments. */ | |
4227 | tree itercnt = NULL_TREE; | |
4228 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
4229 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
4230 | c; c = OMP_CLAUSE_CHAIN (c)) | |
4231 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
4232 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
4233 | { | |
4234 | tree d = OMP_CLAUSE_DECL (c); | |
4235 | tree t = d, a, dest; | |
4236 | if (omp_privatize_by_reference (t)) | |
4237 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); | |
4238 | tree type = TREE_TYPE (t); | |
4239 | if (POINTER_TYPE_P (type)) | |
4240 | type = sizetype; | |
4241 | dest = unshare_expr (t); | |
4242 | tree v = create_tmp_var (TREE_TYPE (t), NULL); | |
4243 | expand_omp_build_assign (&gsif, v, t); | |
4244 | if (itercnt == NULL_TREE) | |
4245 | { | |
4246 | itercnt = startvar; | |
4247 | tree n1 = fd->loop.n1; | |
4248 | if (POINTER_TYPE_P (TREE_TYPE (itercnt))) | |
4249 | { | |
4250 | itercnt | |
4251 | = fold_convert (signed_type_for (TREE_TYPE (itercnt)), | |
4252 | itercnt); | |
4253 | n1 = fold_convert (TREE_TYPE (itercnt), n1); | |
4254 | } | |
4255 | itercnt = fold_build2 (MINUS_EXPR, TREE_TYPE (itercnt), | |
4256 | itercnt, n1); | |
4257 | itercnt = fold_build2 (EXACT_DIV_EXPR, TREE_TYPE (itercnt), | |
4258 | itercnt, fd->loop.step); | |
4259 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
4260 | NULL_TREE, false, | |
4261 | GSI_CONTINUE_LINKING); | |
4262 | } | |
4263 | a = fold_build2 (MULT_EXPR, type, | |
4264 | fold_convert (type, itercnt), | |
4265 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
4266 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
4267 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); | |
4268 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4269 | false, GSI_CONTINUE_LINKING); | |
4270 | expand_omp_build_assign (&gsi, dest, t, true); | |
4271 | } | |
4272 | if (fd->collapse > 1) | |
4273 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, inner_stmt, startvar); | |
4274 | ||
4275 | if (fd->ordered) | |
4276 | { | |
4277 | /* Until now, counts array contained number of iterations or | |
4278 | variable containing it for ith loop. From now on, we need | |
4279 | those counts only for collapsed loops, and only for the 2nd | |
4280 | till the last collapsed one. Move those one element earlier, | |
4281 | we'll use counts[fd->collapse - 1] for the first source/sink | |
4282 | iteration counter and so on and counts[fd->ordered] | |
4283 | as the array holding the current counter values for | |
4284 | depend(source). */ | |
4285 | if (fd->collapse > 1) | |
4286 | memmove (counts, counts + 1, (fd->collapse - 1) * sizeof (counts[0])); | |
4287 | if (broken_loop) | |
4288 | { | |
4289 | int i; | |
4290 | for (i = fd->collapse; i < fd->ordered; i++) | |
4291 | { | |
4292 | tree type = TREE_TYPE (fd->loops[i].v); | |
4293 | tree this_cond | |
4294 | = fold_build2 (fd->loops[i].cond_code, boolean_type_node, | |
4295 | fold_convert (type, fd->loops[i].n1), | |
4296 | fold_convert (type, fd->loops[i].n2)); | |
4297 | if (!integer_onep (this_cond)) | |
4298 | break; | |
4299 | } | |
4300 | if (i < fd->ordered) | |
4301 | { | |
4302 | cont_bb | |
4303 | = create_empty_bb (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb); | |
4304 | add_bb_to_loop (cont_bb, l1_bb->loop_father); | |
4305 | gimple_stmt_iterator gsi = gsi_after_labels (cont_bb); | |
4306 | gimple *g = gimple_build_omp_continue (fd->loop.v, fd->loop.v); | |
4307 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
4308 | make_edge (cont_bb, l3_bb, EDGE_FALLTHRU); | |
4309 | make_edge (cont_bb, l1_bb, 0); | |
4310 | l2_bb = create_empty_bb (cont_bb); | |
4311 | broken_loop = false; | |
4312 | } | |
4313 | } | |
4314 | expand_omp_ordered_source_sink (region, fd, counts, cont_bb); | |
4315 | cont_bb = expand_omp_for_ordered_loops (fd, counts, cont_bb, l1_bb, | |
4316 | ordered_lastprivate); | |
4317 | if (counts[fd->collapse - 1]) | |
4318 | { | |
4319 | gcc_assert (fd->collapse == 1); | |
4320 | gsi = gsi_last_bb (l0_bb); | |
4321 | expand_omp_build_assign (&gsi, counts[fd->collapse - 1], | |
4322 | istart0, true); | |
4323 | if (cont_bb) | |
4324 | { | |
4325 | gsi = gsi_last_bb (cont_bb); | |
4326 | t = fold_build2 (PLUS_EXPR, fd->iter_type, | |
4327 | counts[fd->collapse - 1], | |
4328 | build_int_cst (fd->iter_type, 1)); | |
4329 | expand_omp_build_assign (&gsi, counts[fd->collapse - 1], t); | |
4330 | tree aref = build4 (ARRAY_REF, fd->iter_type, | |
4331 | counts[fd->ordered], size_zero_node, | |
4332 | NULL_TREE, NULL_TREE); | |
4333 | expand_omp_build_assign (&gsi, aref, counts[fd->collapse - 1]); | |
4334 | } | |
4335 | t = counts[fd->collapse - 1]; | |
4336 | } | |
4337 | else if (fd->collapse > 1) | |
4338 | t = fd->loop.v; | |
4339 | else | |
4340 | { | |
4341 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), | |
4342 | fd->loops[0].v, fd->loops[0].n1); | |
4343 | t = fold_convert (fd->iter_type, t); | |
4344 | } | |
4345 | gsi = gsi_last_bb (l0_bb); | |
4346 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
4347 | size_zero_node, NULL_TREE, NULL_TREE); | |
4348 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4349 | false, GSI_CONTINUE_LINKING); | |
4350 | expand_omp_build_assign (&gsi, aref, t, true); | |
4351 | } | |
4352 | ||
4353 | if (!broken_loop) | |
4354 | { | |
4355 | /* Code to control the increment and predicate for the sequential | |
4356 | loop goes in the CONT_BB. */ | |
4357 | gsi = gsi_last_nondebug_bb (cont_bb); | |
4358 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); | |
4359 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
4360 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
4361 | vback = gimple_omp_continue_control_def (cont_stmt); | |
4362 | ||
4363 | if (cond_var) | |
4364 | { | |
4365 | tree itype = TREE_TYPE (cond_var); | |
4366 | tree t2; | |
4367 | if ((fd->ordered && fd->collapse == 1) | |
4368 | || bias | |
4369 | || POINTER_TYPE_P (type) | |
4370 | || TREE_CODE (fd->loop.n1) != INTEGER_CST | |
4371 | || fd->loop.cond_code != LT_EXPR) | |
4372 | t2 = build_int_cst (itype, 1); | |
4373 | else | |
4374 | t2 = fold_convert (itype, fd->loop.step); | |
4375 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
4376 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
4377 | NULL_TREE, true, GSI_SAME_STMT); | |
4378 | assign_stmt = gimple_build_assign (cond_var, t2); | |
4379 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4380 | } | |
4381 | ||
4382 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
4383 | { | |
4384 | if (POINTER_TYPE_P (type)) | |
4385 | t = fold_build_pointer_plus (vmain, fd->loop.step); | |
4386 | else | |
4387 | t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step); | |
4388 | t = force_gimple_operand_gsi (&gsi, t, | |
4389 | DECL_P (vback) | |
4390 | && TREE_ADDRESSABLE (vback), | |
4391 | NULL_TREE, true, GSI_SAME_STMT); | |
4392 | assign_stmt = gimple_build_assign (vback, t); | |
4393 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4394 | ||
4395 | if (fd->ordered && counts[fd->collapse - 1] == NULL_TREE) | |
4396 | { | |
4397 | tree tem; | |
4398 | if (fd->collapse > 1) | |
4399 | tem = fd->loop.v; | |
4400 | else | |
4401 | { | |
4402 | tem = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), | |
4403 | fd->loops[0].v, fd->loops[0].n1); | |
4404 | tem = fold_convert (fd->iter_type, tem); | |
4405 | } | |
4406 | tree aref = build4 (ARRAY_REF, fd->iter_type, | |
4407 | counts[fd->ordered], size_zero_node, | |
4408 | NULL_TREE, NULL_TREE); | |
4409 | tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, | |
4410 | true, GSI_SAME_STMT); | |
4411 | expand_omp_build_assign (&gsi, aref, tem); | |
4412 | } | |
4413 | ||
4414 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
4415 | DECL_P (vback) && TREE_ADDRESSABLE (vback) ? t : vback, | |
4416 | iend); | |
4417 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
4418 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
4419 | } | |
4420 | ||
4421 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
4422 | gsi_remove (&gsi, true); | |
4423 | ||
4424 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
4425 | collapse_bb = extract_omp_for_update_vars (fd, NULL, cont_bb, l1_bb); | |
4426 | ||
4427 | /* Emit code to get the next parallel iteration in L2_BB. */ | |
4428 | gsi = gsi_start_bb (l2_bb); | |
4429 | ||
4430 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, | |
4431 | build_fold_addr_expr (istart0), | |
4432 | build_fold_addr_expr (iend0)); | |
4433 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4434 | false, GSI_CONTINUE_LINKING); | |
4435 | if (TREE_TYPE (t) != boolean_type_node) | |
4436 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
4437 | t, build_int_cst (TREE_TYPE (t), 0)); | |
4438 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
4439 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
4440 | } | |
4441 | ||
4442 | /* Add the loop cleanup function. */ | |
4443 | gsi = gsi_last_nondebug_bb (exit_bb); | |
4444 | if (gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
4445 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
4446 | else if (gimple_omp_return_lhs (gsi_stmt (gsi))) | |
4447 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
4448 | else | |
4449 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
4450 | gcall *call_stmt = gimple_build_call (t, 0); | |
4451 | if (fd->ordered) | |
4452 | { | |
4453 | tree arr = counts[fd->ordered]; | |
4454 | tree clobber = build_clobber (TREE_TYPE (arr)); | |
4455 | gsi_insert_after (&gsi, gimple_build_assign (arr, clobber), | |
4456 | GSI_SAME_STMT); | |
4457 | } | |
4458 | if (gimple_omp_return_lhs (gsi_stmt (gsi))) | |
4459 | { | |
4460 | gimple_call_set_lhs (call_stmt, gimple_omp_return_lhs (gsi_stmt (gsi))); | |
4461 | if (fd->have_reductemp) | |
4462 | { | |
4463 | gimple *g = gimple_build_assign (reductions, NOP_EXPR, | |
4464 | gimple_call_lhs (call_stmt)); | |
4465 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
4466 | } | |
4467 | } | |
4468 | gsi_insert_after (&gsi, call_stmt, GSI_SAME_STMT); | |
4469 | gsi_remove (&gsi, true); | |
4470 | ||
4471 | /* Connect the new blocks. */ | |
4472 | find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE; | |
4473 | find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE; | |
4474 | ||
4475 | if (!broken_loop) | |
4476 | { | |
4477 | gimple_seq phis; | |
4478 | ||
4479 | e = find_edge (cont_bb, l3_bb); | |
4480 | ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE); | |
4481 | ||
4482 | phis = phi_nodes (l3_bb); | |
4483 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4484 | { | |
4485 | gimple *phi = gsi_stmt (gsi); | |
4486 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne), | |
4487 | PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
4488 | } | |
4489 | remove_edge (e); | |
4490 | ||
4491 | make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE); | |
4492 | e = find_edge (cont_bb, l1_bb); | |
4493 | if (e == NULL) | |
4494 | { | |
4495 | e = BRANCH_EDGE (cont_bb); | |
4496 | gcc_assert (single_succ (e->dest) == l1_bb); | |
4497 | } | |
4498 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4499 | { | |
4500 | remove_edge (e); | |
4501 | e = NULL; | |
4502 | } | |
4503 | else if (fd->collapse > 1) | |
4504 | { | |
4505 | remove_edge (e); | |
4506 | e = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
4507 | } | |
4508 | else | |
4509 | e->flags = EDGE_TRUE_VALUE; | |
4510 | if (e) | |
4511 | { | |
4512 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); | |
4513 | find_edge (cont_bb, l2_bb)->probability = e->probability.invert (); | |
4514 | } | |
4515 | else | |
4516 | { | |
4517 | e = find_edge (cont_bb, l2_bb); | |
4518 | e->flags = EDGE_FALLTHRU; | |
4519 | } | |
4520 | make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE); | |
4521 | ||
4522 | if (gimple_in_ssa_p (cfun)) | |
4523 | { | |
4524 | /* Add phis to the outer loop that connect to the phis in the inner, | |
4525 | original loop, and move the loop entry value of the inner phi to | |
4526 | the loop entry value of the outer phi. */ | |
4527 | gphi_iterator psi; | |
4528 | for (psi = gsi_start_phis (l3_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
4529 | { | |
4530 | location_t locus; | |
4531 | gphi *nphi; | |
4532 | gphi *exit_phi = psi.phi (); | |
4533 | ||
4534 | if (virtual_operand_p (gimple_phi_result (exit_phi))) | |
4535 | continue; | |
4536 | ||
4537 | edge l2_to_l3 = find_edge (l2_bb, l3_bb); | |
4538 | tree exit_res = PHI_ARG_DEF_FROM_EDGE (exit_phi, l2_to_l3); | |
4539 | ||
4540 | basic_block latch = BRANCH_EDGE (cont_bb)->dest; | |
4541 | edge latch_to_l1 = find_edge (latch, l1_bb); | |
4542 | gphi *inner_phi | |
4543 | = find_phi_with_arg_on_edge (exit_res, latch_to_l1); | |
4544 | ||
4545 | tree t = gimple_phi_result (exit_phi); | |
4546 | tree new_res = copy_ssa_name (t, NULL); | |
4547 | nphi = create_phi_node (new_res, l0_bb); | |
4548 | ||
4549 | edge l0_to_l1 = find_edge (l0_bb, l1_bb); | |
4550 | t = PHI_ARG_DEF_FROM_EDGE (inner_phi, l0_to_l1); | |
4551 | locus = gimple_phi_arg_location_from_edge (inner_phi, l0_to_l1); | |
4552 | edge entry_to_l0 = find_edge (entry_bb, l0_bb); | |
4553 | add_phi_arg (nphi, t, entry_to_l0, locus); | |
4554 | ||
4555 | edge l2_to_l0 = find_edge (l2_bb, l0_bb); | |
4556 | add_phi_arg (nphi, exit_res, l2_to_l0, UNKNOWN_LOCATION); | |
4557 | ||
4558 | add_phi_arg (inner_phi, new_res, l0_to_l1, UNKNOWN_LOCATION); | |
4559 | } | |
4560 | } | |
4561 | ||
4562 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, | |
4563 | recompute_dominator (CDI_DOMINATORS, l2_bb)); | |
4564 | set_immediate_dominator (CDI_DOMINATORS, l3_bb, | |
4565 | recompute_dominator (CDI_DOMINATORS, l3_bb)); | |
4566 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, | |
4567 | recompute_dominator (CDI_DOMINATORS, l0_bb)); | |
4568 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, | |
4569 | recompute_dominator (CDI_DOMINATORS, l1_bb)); | |
4570 | ||
4571 | /* We enter expand_omp_for_generic with a loop. This original loop may | |
4572 | have its own loop struct, or it may be part of an outer loop struct | |
4573 | (which may be the fake loop). */ | |
4574 | class loop *outer_loop = entry_bb->loop_father; | |
4575 | bool orig_loop_has_loop_struct = l1_bb->loop_father != outer_loop; | |
4576 | ||
4577 | add_bb_to_loop (l2_bb, outer_loop); | |
4578 | ||
4579 | /* We've added a new loop around the original loop. Allocate the | |
4580 | corresponding loop struct. */ | |
4581 | class loop *new_loop = alloc_loop (); | |
4582 | new_loop->header = l0_bb; | |
4583 | new_loop->latch = l2_bb; | |
4584 | add_loop (new_loop, outer_loop); | |
4585 | ||
4586 | /* Allocate a loop structure for the original loop unless we already | |
4587 | had one. */ | |
4588 | if (!orig_loop_has_loop_struct | |
4589 | && !gimple_omp_for_combined_p (fd->for_stmt)) | |
4590 | { | |
4591 | class loop *orig_loop = alloc_loop (); | |
4592 | orig_loop->header = l1_bb; | |
4593 | /* The loop may have multiple latches. */ | |
4594 | add_loop (orig_loop, new_loop); | |
4595 | } | |
4596 | } | |
4597 | } | |
4598 | ||
4599 | /* Helper function for expand_omp_for_static_nochunk. If PTR is NULL, | |
4600 | compute needed allocation size. If !ALLOC of team allocations, | |
4601 | if ALLOC of thread allocation. SZ is the initial needed size for | |
4602 | other purposes, ALLOC_ALIGN guaranteed alignment of allocation in bytes, | |
4603 | CNT number of elements of each array, for !ALLOC this is | |
4604 | omp_get_num_threads (), for ALLOC number of iterations handled by the | |
4605 | current thread. If PTR is non-NULL, it is the start of the allocation | |
4606 | and this routine shall assign to OMP_CLAUSE_DECL (c) of those _scantemp_ | |
4607 | clauses pointers to the corresponding arrays. */ | |
4608 | ||
4609 | static tree | |
4610 | expand_omp_scantemp_alloc (tree clauses, tree ptr, unsigned HOST_WIDE_INT sz, | |
4611 | unsigned HOST_WIDE_INT alloc_align, tree cnt, | |
4612 | gimple_stmt_iterator *gsi, bool alloc) | |
4613 | { | |
4614 | tree eltsz = NULL_TREE; | |
4615 | unsigned HOST_WIDE_INT preval = 0; | |
4616 | if (ptr && sz) | |
4617 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), | |
4618 | ptr, size_int (sz)); | |
4619 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
4620 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
4621 | && !OMP_CLAUSE__SCANTEMP__CONTROL (c) | |
4622 | && (!OMP_CLAUSE__SCANTEMP__ALLOC (c)) != alloc) | |
4623 | { | |
4624 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); | |
4625 | unsigned HOST_WIDE_INT al = TYPE_ALIGN_UNIT (pointee_type); | |
4626 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) | |
4627 | { | |
4628 | unsigned HOST_WIDE_INT szl | |
4629 | = tree_to_uhwi (TYPE_SIZE_UNIT (pointee_type)); | |
4630 | szl = least_bit_hwi (szl); | |
4631 | if (szl) | |
4632 | al = MIN (al, szl); | |
4633 | } | |
4634 | if (ptr == NULL_TREE) | |
4635 | { | |
4636 | if (eltsz == NULL_TREE) | |
4637 | eltsz = TYPE_SIZE_UNIT (pointee_type); | |
4638 | else | |
4639 | eltsz = size_binop (PLUS_EXPR, eltsz, | |
4640 | TYPE_SIZE_UNIT (pointee_type)); | |
4641 | } | |
4642 | if (preval == 0 && al <= alloc_align) | |
4643 | { | |
4644 | unsigned HOST_WIDE_INT diff = ROUND_UP (sz, al) - sz; | |
4645 | sz += diff; | |
4646 | if (diff && ptr) | |
4647 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), | |
4648 | ptr, size_int (diff)); | |
4649 | } | |
4650 | else if (al > preval) | |
4651 | { | |
4652 | if (ptr) | |
4653 | { | |
4654 | ptr = fold_convert (pointer_sized_int_node, ptr); | |
4655 | ptr = fold_build2 (PLUS_EXPR, pointer_sized_int_node, ptr, | |
4656 | build_int_cst (pointer_sized_int_node, | |
4657 | al - 1)); | |
4658 | ptr = fold_build2 (BIT_AND_EXPR, pointer_sized_int_node, ptr, | |
4659 | build_int_cst (pointer_sized_int_node, | |
4660 | -(HOST_WIDE_INT) al)); | |
4661 | ptr = fold_convert (ptr_type_node, ptr); | |
4662 | } | |
4663 | else | |
4664 | sz += al - 1; | |
4665 | } | |
4666 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) | |
4667 | preval = al; | |
4668 | else | |
4669 | preval = 1; | |
4670 | if (ptr) | |
4671 | { | |
4672 | expand_omp_build_assign (gsi, OMP_CLAUSE_DECL (c), ptr, false); | |
4673 | ptr = OMP_CLAUSE_DECL (c); | |
4674 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), ptr, | |
4675 | size_binop (MULT_EXPR, cnt, | |
4676 | TYPE_SIZE_UNIT (pointee_type))); | |
4677 | } | |
4678 | } | |
4679 | ||
4680 | if (ptr == NULL_TREE) | |
4681 | { | |
4682 | eltsz = size_binop (MULT_EXPR, eltsz, cnt); | |
4683 | if (sz) | |
4684 | eltsz = size_binop (PLUS_EXPR, eltsz, size_int (sz)); | |
4685 | return eltsz; | |
4686 | } | |
4687 | else | |
4688 | return ptr; | |
4689 | } | |
4690 | ||
4691 | /* Return the last _looptemp_ clause if one has been created for | |
4692 | lastprivate on distribute parallel for{, simd} or taskloop. | |
4693 | FD is the loop data and INNERC should be the second _looptemp_ | |
4694 | clause (the one holding the end of the range). | |
4695 | This is followed by collapse - 1 _looptemp_ clauses for the | |
4696 | counts[1] and up, and for triangular loops followed by 4 | |
4697 | further _looptemp_ clauses (one for counts[0], one first_inner_iterations, | |
4698 | one factor and one adjn1). After this there is optionally one | |
4699 | _looptemp_ clause that this function returns. */ | |
4700 | ||
4701 | static tree | |
4702 | find_lastprivate_looptemp (struct omp_for_data *fd, tree innerc) | |
4703 | { | |
4704 | gcc_assert (innerc); | |
4705 | int count = fd->collapse - 1; | |
4706 | if (fd->non_rect | |
4707 | && fd->last_nonrect == fd->first_nonrect + 1 | |
4708 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) | |
4709 | count += 4; | |
4710 | for (int i = 0; i < count; i++) | |
4711 | { | |
4712 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4713 | OMP_CLAUSE__LOOPTEMP_); | |
4714 | gcc_assert (innerc); | |
4715 | } | |
4716 | return omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4717 | OMP_CLAUSE__LOOPTEMP_); | |
4718 | } | |
4719 | ||
4720 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
4721 | loop with static schedule and no specified chunk size. Given | |
4722 | parameters: | |
4723 | ||
4724 | for (V = N1; V cond N2; V += STEP) BODY; | |
4725 | ||
4726 | where COND is "<" or ">", we generate pseudocode | |
4727 | ||
4728 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; | |
4729 | if (cond is <) | |
4730 | adj = STEP - 1; | |
4731 | else | |
4732 | adj = STEP + 1; | |
4733 | if ((__typeof (V)) -1 > 0 && cond is >) | |
4734 | n = -(adj + N2 - N1) / -STEP; | |
4735 | else | |
4736 | n = (adj + N2 - N1) / STEP; | |
4737 | q = n / nthreads; | |
4738 | tt = n % nthreads; | |
4739 | if (threadid < tt) goto L3; else goto L4; | |
4740 | L3: | |
4741 | tt = 0; | |
4742 | q = q + 1; | |
4743 | L4: | |
4744 | s0 = q * threadid + tt; | |
4745 | e0 = s0 + q; | |
4746 | V = s0 * STEP + N1; | |
4747 | if (s0 >= e0) goto L2; else goto L0; | |
4748 | L0: | |
4749 | e = e0 * STEP + N1; | |
4750 | L1: | |
4751 | BODY; | |
4752 | V += STEP; | |
4753 | if (V cond e) goto L1; | |
4754 | L2: | |
4755 | */ | |
4756 | ||
4757 | static void | |
4758 | expand_omp_for_static_nochunk (struct omp_region *region, | |
4759 | struct omp_for_data *fd, | |
4760 | gimple *inner_stmt) | |
4761 | { | |
4762 | tree n, q, s0, e0, e, t, tt, nthreads = NULL_TREE, threadid; | |
4763 | tree type, itype, vmain, vback; | |
4764 | basic_block entry_bb, second_bb, third_bb, exit_bb, seq_start_bb; | |
4765 | basic_block body_bb, cont_bb, collapse_bb = NULL; | |
4766 | basic_block fin_bb, fourth_bb = NULL, fifth_bb = NULL, sixth_bb = NULL; | |
4767 | basic_block exit1_bb = NULL, exit2_bb = NULL, exit3_bb = NULL; | |
4768 | gimple_stmt_iterator gsi, gsip; | |
4769 | edge ep; | |
4770 | bool broken_loop = region->cont == NULL; | |
4771 | tree *counts = NULL; | |
4772 | tree n1, n2, step; | |
4773 | tree reductions = NULL_TREE; | |
4774 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; | |
4775 | ||
4776 | itype = type = TREE_TYPE (fd->loop.v); | |
4777 | if (POINTER_TYPE_P (type)) | |
4778 | itype = signed_type_for (type); | |
4779 | ||
4780 | entry_bb = region->entry; | |
4781 | cont_bb = region->cont; | |
4782 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
4783 | fin_bb = BRANCH_EDGE (entry_bb)->dest; | |
4784 | gcc_assert (broken_loop | |
4785 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
4786 | seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
4787 | body_bb = single_succ (seq_start_bb); | |
4788 | if (!broken_loop) | |
4789 | { | |
4790 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb | |
4791 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); | |
4792 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
4793 | } | |
4794 | exit_bb = region->exit; | |
4795 | ||
4796 | /* Iteration space partitioning goes in ENTRY_BB. */ | |
4797 | gsi = gsi_last_nondebug_bb (entry_bb); | |
4798 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
4799 | gsip = gsi; | |
4800 | gsi_prev (&gsip); | |
4801 | ||
4802 | if (fd->collapse > 1) | |
4803 | { | |
4804 | int first_zero_iter = -1, dummy = -1; | |
4805 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
4806 | ||
4807 | counts = XALLOCAVEC (tree, fd->collapse); | |
4808 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
4809 | fin_bb, first_zero_iter, | |
4810 | dummy_bb, dummy, l2_dom_bb); | |
4811 | t = NULL_TREE; | |
4812 | } | |
4813 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
4814 | t = integer_one_node; | |
4815 | else | |
4816 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
4817 | fold_convert (type, fd->loop.n1), | |
4818 | fold_convert (type, fd->loop.n2)); | |
4819 | if (fd->collapse == 1 | |
4820 | && TYPE_UNSIGNED (type) | |
4821 | && (t == NULL_TREE || !integer_onep (t))) | |
4822 | { | |
4823 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); | |
4824 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
4825 | true, GSI_SAME_STMT); | |
4826 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
4827 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
4828 | true, GSI_SAME_STMT); | |
4829 | gcond *cond_stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, | |
4830 | NULL_TREE, NULL_TREE); | |
4831 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
4832 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
4833 | expand_omp_regimplify_p, NULL, NULL) | |
4834 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
4835 | expand_omp_regimplify_p, NULL, NULL)) | |
4836 | { | |
4837 | gsi = gsi_for_stmt (cond_stmt); | |
4838 | gimple_regimplify_operands (cond_stmt, &gsi); | |
4839 | } | |
4840 | ep = split_block (entry_bb, cond_stmt); | |
4841 | ep->flags = EDGE_TRUE_VALUE; | |
4842 | entry_bb = ep->dest; | |
4843 | ep->probability = profile_probability::very_likely (); | |
4844 | ep = make_edge (ep->src, fin_bb, EDGE_FALSE_VALUE); | |
4845 | ep->probability = profile_probability::very_unlikely (); | |
4846 | if (gimple_in_ssa_p (cfun)) | |
4847 | { | |
4848 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; | |
4849 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); | |
4850 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
4851 | { | |
4852 | gphi *phi = gpi.phi (); | |
4853 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
4854 | ep, UNKNOWN_LOCATION); | |
4855 | } | |
4856 | } | |
4857 | gsi = gsi_last_bb (entry_bb); | |
4858 | } | |
4859 | ||
4860 | if (fd->lastprivate_conditional) | |
4861 | { | |
4862 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
4863 | tree c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); | |
4864 | if (fd->have_pointer_condtemp) | |
4865 | condtemp = OMP_CLAUSE_DECL (c); | |
4866 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
4867 | cond_var = OMP_CLAUSE_DECL (c); | |
4868 | } | |
4869 | if (fd->have_reductemp | |
4870 | /* For scan, we don't want to reinitialize condtemp before the | |
4871 | second loop. */ | |
4872 | || (fd->have_pointer_condtemp && !fd->have_scantemp) | |
4873 | || fd->have_nonctrl_scantemp) | |
4874 | { | |
4875 | tree t1 = build_int_cst (long_integer_type_node, 0); | |
4876 | tree t2 = build_int_cst (long_integer_type_node, 1); | |
4877 | tree t3 = build_int_cstu (long_integer_type_node, | |
4878 | (HOST_WIDE_INT_1U << 31) + 1); | |
4879 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
4880 | gimple_stmt_iterator gsi2 = gsi_none (); | |
4881 | gimple *g = NULL; | |
4882 | tree mem = null_pointer_node, memv = NULL_TREE; | |
4883 | unsigned HOST_WIDE_INT condtemp_sz = 0; | |
4884 | unsigned HOST_WIDE_INT alloc_align = 0; | |
4885 | if (fd->have_reductemp) | |
4886 | { | |
4887 | gcc_assert (!fd->have_nonctrl_scantemp); | |
4888 | tree c = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
4889 | reductions = OMP_CLAUSE_DECL (c); | |
4890 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
4891 | g = SSA_NAME_DEF_STMT (reductions); | |
4892 | reductions = gimple_assign_rhs1 (g); | |
4893 | OMP_CLAUSE_DECL (c) = reductions; | |
4894 | gsi2 = gsi_for_stmt (g); | |
4895 | } | |
4896 | else | |
4897 | { | |
4898 | if (gsi_end_p (gsip)) | |
4899 | gsi2 = gsi_after_labels (region->entry); | |
4900 | else | |
4901 | gsi2 = gsip; | |
4902 | reductions = null_pointer_node; | |
4903 | } | |
4904 | if (fd->have_pointer_condtemp || fd->have_nonctrl_scantemp) | |
4905 | { | |
4906 | tree type; | |
4907 | if (fd->have_pointer_condtemp) | |
4908 | type = TREE_TYPE (condtemp); | |
4909 | else | |
4910 | type = ptr_type_node; | |
4911 | memv = create_tmp_var (type); | |
4912 | TREE_ADDRESSABLE (memv) = 1; | |
4913 | unsigned HOST_WIDE_INT sz = 0; | |
4914 | tree size = NULL_TREE; | |
4915 | if (fd->have_pointer_condtemp) | |
4916 | { | |
4917 | sz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
4918 | sz *= fd->lastprivate_conditional; | |
4919 | condtemp_sz = sz; | |
4920 | } | |
4921 | if (fd->have_nonctrl_scantemp) | |
4922 | { | |
4923 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
4924 | gimple *g = gimple_build_call (nthreads, 0); | |
4925 | nthreads = create_tmp_var (integer_type_node); | |
4926 | gimple_call_set_lhs (g, nthreads); | |
4927 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); | |
4928 | nthreads = fold_convert (sizetype, nthreads); | |
4929 | alloc_align = TYPE_ALIGN_UNIT (long_long_integer_type_node); | |
4930 | size = expand_omp_scantemp_alloc (clauses, NULL_TREE, sz, | |
4931 | alloc_align, nthreads, NULL, | |
4932 | false); | |
4933 | size = fold_convert (type, size); | |
4934 | } | |
4935 | else | |
4936 | size = build_int_cst (type, sz); | |
4937 | expand_omp_build_assign (&gsi2, memv, size, false); | |
4938 | mem = build_fold_addr_expr (memv); | |
4939 | } | |
4940 | tree t | |
4941 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_LOOP_START), | |
4942 | 9, t1, t2, t2, t3, t1, null_pointer_node, | |
4943 | null_pointer_node, reductions, mem); | |
4944 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
4945 | true, GSI_SAME_STMT); | |
4946 | if (fd->have_pointer_condtemp) | |
4947 | expand_omp_build_assign (&gsi2, condtemp, memv, false); | |
4948 | if (fd->have_nonctrl_scantemp) | |
4949 | { | |
4950 | tree ptr = fd->have_pointer_condtemp ? condtemp : memv; | |
4951 | expand_omp_scantemp_alloc (clauses, ptr, condtemp_sz, | |
4952 | alloc_align, nthreads, &gsi2, false); | |
4953 | } | |
4954 | if (fd->have_reductemp) | |
4955 | { | |
4956 | gsi_remove (&gsi2, true); | |
4957 | release_ssa_name (gimple_assign_lhs (g)); | |
4958 | } | |
4959 | } | |
4960 | switch (gimple_omp_for_kind (fd->for_stmt)) | |
4961 | { | |
4962 | case GF_OMP_FOR_KIND_FOR: | |
4963 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
4964 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
4965 | break; | |
4966 | case GF_OMP_FOR_KIND_DISTRIBUTE: | |
4967 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS); | |
4968 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM); | |
4969 | break; | |
4970 | default: | |
4971 | gcc_unreachable (); | |
4972 | } | |
4973 | nthreads = build_call_expr (nthreads, 0); | |
4974 | nthreads = fold_convert (itype, nthreads); | |
4975 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, | |
4976 | true, GSI_SAME_STMT); | |
4977 | threadid = build_call_expr (threadid, 0); | |
4978 | threadid = fold_convert (itype, threadid); | |
4979 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, | |
4980 | true, GSI_SAME_STMT); | |
4981 | ||
4982 | n1 = fd->loop.n1; | |
4983 | n2 = fd->loop.n2; | |
4984 | step = fd->loop.step; | |
4985 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
4986 | { | |
4987 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
4988 | OMP_CLAUSE__LOOPTEMP_); | |
4989 | gcc_assert (innerc); | |
4990 | n1 = OMP_CLAUSE_DECL (innerc); | |
4991 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4992 | OMP_CLAUSE__LOOPTEMP_); | |
4993 | gcc_assert (innerc); | |
4994 | n2 = OMP_CLAUSE_DECL (innerc); | |
4995 | } | |
4996 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
4997 | true, NULL_TREE, true, GSI_SAME_STMT); | |
4998 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
4999 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5000 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
5001 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5002 | ||
5003 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
5004 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
5005 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
5006 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
5007 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
5008 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
5009 | fold_build1 (NEGATE_EXPR, itype, t), | |
5010 | fold_build1 (NEGATE_EXPR, itype, step)); | |
5011 | else | |
5012 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
5013 | t = fold_convert (itype, t); | |
5014 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
5015 | ||
5016 | q = create_tmp_reg (itype, "q"); | |
5017 | t = fold_build2 (TRUNC_DIV_EXPR, itype, n, nthreads); | |
5018 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
5019 | gsi_insert_before (&gsi, gimple_build_assign (q, t), GSI_SAME_STMT); | |
5020 | ||
5021 | tt = create_tmp_reg (itype, "tt"); | |
5022 | t = fold_build2 (TRUNC_MOD_EXPR, itype, n, nthreads); | |
5023 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
5024 | gsi_insert_before (&gsi, gimple_build_assign (tt, t), GSI_SAME_STMT); | |
5025 | ||
5026 | t = build2 (LT_EXPR, boolean_type_node, threadid, tt); | |
5027 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
5028 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
5029 | ||
5030 | second_bb = split_block (entry_bb, cond_stmt)->dest; | |
5031 | gsi = gsi_last_nondebug_bb (second_bb); | |
5032 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
5033 | ||
5034 | gsi_insert_before (&gsi, gimple_build_assign (tt, build_int_cst (itype, 0)), | |
5035 | GSI_SAME_STMT); | |
5036 | gassign *assign_stmt | |
5037 | = gimple_build_assign (q, PLUS_EXPR, q, build_int_cst (itype, 1)); | |
5038 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5039 | ||
5040 | third_bb = split_block (second_bb, assign_stmt)->dest; | |
5041 | gsi = gsi_last_nondebug_bb (third_bb); | |
5042 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
5043 | ||
5044 | if (fd->have_nonctrl_scantemp) | |
5045 | { | |
5046 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5047 | tree controlp = NULL_TREE, controlb = NULL_TREE; | |
5048 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
5049 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
5050 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) | |
5051 | { | |
5052 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) | |
5053 | controlb = OMP_CLAUSE_DECL (c); | |
5054 | else | |
5055 | controlp = OMP_CLAUSE_DECL (c); | |
5056 | if (controlb && controlp) | |
5057 | break; | |
5058 | } | |
5059 | gcc_assert (controlp && controlb); | |
5060 | tree cnt = create_tmp_var (sizetype); | |
5061 | gimple *g = gimple_build_assign (cnt, NOP_EXPR, q); | |
5062 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5063 | unsigned HOST_WIDE_INT alloc_align = TYPE_ALIGN_UNIT (ptr_type_node); | |
5064 | tree sz = expand_omp_scantemp_alloc (clauses, NULL_TREE, 0, | |
5065 | alloc_align, cnt, NULL, true); | |
5066 | tree size = create_tmp_var (sizetype); | |
5067 | expand_omp_build_assign (&gsi, size, sz, false); | |
5068 | tree cmp = fold_build2 (GT_EXPR, boolean_type_node, | |
5069 | size, size_int (16384)); | |
5070 | expand_omp_build_assign (&gsi, controlb, cmp); | |
5071 | g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, | |
5072 | NULL_TREE, NULL_TREE); | |
5073 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5074 | fourth_bb = split_block (third_bb, g)->dest; | |
5075 | gsi = gsi_last_nondebug_bb (fourth_bb); | |
5076 | /* FIXME: Once we have allocators, this should use allocator. */ | |
5077 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_MALLOC), 1, size); | |
5078 | gimple_call_set_lhs (g, controlp); | |
5079 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5080 | expand_omp_scantemp_alloc (clauses, controlp, 0, alloc_align, cnt, | |
5081 | &gsi, true); | |
5082 | gsi_prev (&gsi); | |
5083 | g = gsi_stmt (gsi); | |
5084 | fifth_bb = split_block (fourth_bb, g)->dest; | |
5085 | gsi = gsi_last_nondebug_bb (fifth_bb); | |
5086 | ||
5087 | g = gimple_build_call (builtin_decl_implicit (BUILT_IN_STACK_SAVE), 0); | |
5088 | gimple_call_set_lhs (g, controlp); | |
5089 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5090 | tree alloca_decl = builtin_decl_explicit (BUILT_IN_ALLOCA_WITH_ALIGN); | |
5091 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
5092 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
5093 | && OMP_CLAUSE__SCANTEMP__ALLOC (c)) | |
5094 | { | |
5095 | tree tmp = create_tmp_var (sizetype); | |
5096 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); | |
5097 | g = gimple_build_assign (tmp, MULT_EXPR, cnt, | |
5098 | TYPE_SIZE_UNIT (pointee_type)); | |
5099 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5100 | g = gimple_build_call (alloca_decl, 2, tmp, | |
5101 | size_int (TYPE_ALIGN (pointee_type))); | |
5102 | gimple_call_set_lhs (g, OMP_CLAUSE_DECL (c)); | |
5103 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5104 | } | |
5105 | ||
5106 | sixth_bb = split_block (fifth_bb, g)->dest; | |
5107 | gsi = gsi_last_nondebug_bb (sixth_bb); | |
5108 | } | |
5109 | ||
5110 | t = build2 (MULT_EXPR, itype, q, threadid); | |
5111 | t = build2 (PLUS_EXPR, itype, t, tt); | |
5112 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
5113 | ||
5114 | t = fold_build2 (PLUS_EXPR, itype, s0, q); | |
5115 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
5116 | ||
5117 | t = build2 (GE_EXPR, boolean_type_node, s0, e0); | |
5118 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
5119 | ||
5120 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
5121 | gsi_remove (&gsi, true); | |
5122 | ||
5123 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
5124 | gsi = gsi_start_bb (seq_start_bb); | |
5125 | ||
5126 | tree startvar = fd->loop.v; | |
5127 | tree endvar = NULL_TREE; | |
5128 | ||
5129 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5130 | { | |
5131 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
5132 | ? gimple_omp_parallel_clauses (inner_stmt) | |
5133 | : gimple_omp_for_clauses (inner_stmt); | |
5134 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
5135 | gcc_assert (innerc); | |
5136 | startvar = OMP_CLAUSE_DECL (innerc); | |
5137 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5138 | OMP_CLAUSE__LOOPTEMP_); | |
5139 | gcc_assert (innerc); | |
5140 | endvar = OMP_CLAUSE_DECL (innerc); | |
5141 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST | |
5142 | && gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) | |
5143 | { | |
5144 | innerc = find_lastprivate_looptemp (fd, innerc); | |
5145 | if (innerc) | |
5146 | { | |
5147 | /* If needed (distribute parallel for with lastprivate), | |
5148 | propagate down the total number of iterations. */ | |
5149 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), | |
5150 | fd->loop.n2); | |
5151 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, | |
5152 | GSI_CONTINUE_LINKING); | |
5153 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
5154 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5155 | } | |
5156 | } | |
5157 | } | |
5158 | t = fold_convert (itype, s0); | |
5159 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5160 | if (POINTER_TYPE_P (type)) | |
5161 | { | |
5162 | t = fold_build_pointer_plus (n1, t); | |
5163 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
5164 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
5165 | t = fold_convert (signed_type_for (type), t); | |
5166 | } | |
5167 | else | |
5168 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5169 | t = fold_convert (TREE_TYPE (startvar), t); | |
5170 | t = force_gimple_operand_gsi (&gsi, t, | |
5171 | DECL_P (startvar) | |
5172 | && TREE_ADDRESSABLE (startvar), | |
5173 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5174 | assign_stmt = gimple_build_assign (startvar, t); | |
5175 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5176 | if (cond_var) | |
5177 | { | |
5178 | tree itype = TREE_TYPE (cond_var); | |
5179 | /* For lastprivate(conditional:) itervar, we need some iteration | |
5180 | counter that starts at unsigned non-zero and increases. | |
5181 | Prefer as few IVs as possible, so if we can use startvar | |
5182 | itself, use that, or startvar + constant (those would be | |
5183 | incremented with step), and as last resort use the s0 + 1 | |
5184 | incremented by 1. */ | |
5185 | if (POINTER_TYPE_P (type) | |
5186 | || TREE_CODE (n1) != INTEGER_CST | |
5187 | || fd->loop.cond_code != LT_EXPR) | |
5188 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), | |
5189 | build_int_cst (itype, 1)); | |
5190 | else if (tree_int_cst_sgn (n1) == 1) | |
5191 | t = fold_convert (itype, t); | |
5192 | else | |
5193 | { | |
5194 | tree c = fold_convert (itype, n1); | |
5195 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
5196 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
5197 | } | |
5198 | t = force_gimple_operand_gsi (&gsi, t, false, | |
5199 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5200 | assign_stmt = gimple_build_assign (cond_var, t); | |
5201 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5202 | } | |
5203 | ||
5204 | t = fold_convert (itype, e0); | |
5205 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5206 | if (POINTER_TYPE_P (type)) | |
5207 | { | |
5208 | t = fold_build_pointer_plus (n1, t); | |
5209 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
5210 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
5211 | t = fold_convert (signed_type_for (type), t); | |
5212 | } | |
5213 | else | |
5214 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5215 | t = fold_convert (TREE_TYPE (startvar), t); | |
5216 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5217 | false, GSI_CONTINUE_LINKING); | |
5218 | if (endvar) | |
5219 | { | |
5220 | assign_stmt = gimple_build_assign (endvar, e); | |
5221 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5222 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
5223 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
5224 | else | |
5225 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
5226 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5227 | } | |
5228 | /* Handle linear clause adjustments. */ | |
5229 | tree itercnt = NULL_TREE; | |
5230 | tree *nonrect_bounds = NULL; | |
5231 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
5232 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
5233 | c; c = OMP_CLAUSE_CHAIN (c)) | |
5234 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
5235 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
5236 | { | |
5237 | tree d = OMP_CLAUSE_DECL (c); | |
5238 | tree t = d, a, dest; | |
5239 | if (omp_privatize_by_reference (t)) | |
5240 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); | |
5241 | if (itercnt == NULL_TREE) | |
5242 | { | |
5243 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5244 | { | |
5245 | itercnt = fold_build2 (MINUS_EXPR, itype, | |
5246 | fold_convert (itype, n1), | |
5247 | fold_convert (itype, fd->loop.n1)); | |
5248 | itercnt = fold_build2 (EXACT_DIV_EXPR, itype, itercnt, step); | |
5249 | itercnt = fold_build2 (PLUS_EXPR, itype, itercnt, s0); | |
5250 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
5251 | NULL_TREE, false, | |
5252 | GSI_CONTINUE_LINKING); | |
5253 | } | |
5254 | else | |
5255 | itercnt = s0; | |
5256 | } | |
5257 | tree type = TREE_TYPE (t); | |
5258 | if (POINTER_TYPE_P (type)) | |
5259 | type = sizetype; | |
5260 | a = fold_build2 (MULT_EXPR, type, | |
5261 | fold_convert (type, itercnt), | |
5262 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
5263 | dest = unshare_expr (t); | |
5264 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
5265 | : POINTER_PLUS_EXPR, TREE_TYPE (t), t, a); | |
5266 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5267 | false, GSI_CONTINUE_LINKING); | |
5268 | expand_omp_build_assign (&gsi, dest, t, true); | |
5269 | } | |
5270 | if (fd->collapse > 1) | |
5271 | { | |
5272 | if (fd->non_rect) | |
5273 | { | |
5274 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); | |
5275 | memset (nonrect_bounds, 0, sizeof (tree) * (fd->last_nonrect + 1)); | |
5276 | } | |
5277 | expand_omp_for_init_vars (fd, &gsi, counts, nonrect_bounds, inner_stmt, | |
5278 | startvar); | |
5279 | } | |
5280 | ||
5281 | if (!broken_loop) | |
5282 | { | |
5283 | /* The code controlling the sequential loop replaces the | |
5284 | GIMPLE_OMP_CONTINUE. */ | |
5285 | gsi = gsi_last_nondebug_bb (cont_bb); | |
5286 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); | |
5287 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
5288 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
5289 | vback = gimple_omp_continue_control_def (cont_stmt); | |
5290 | ||
5291 | if (cond_var) | |
5292 | { | |
5293 | tree itype = TREE_TYPE (cond_var); | |
5294 | tree t2; | |
5295 | if (POINTER_TYPE_P (type) | |
5296 | || TREE_CODE (n1) != INTEGER_CST | |
5297 | || fd->loop.cond_code != LT_EXPR) | |
5298 | t2 = build_int_cst (itype, 1); | |
5299 | else | |
5300 | t2 = fold_convert (itype, step); | |
5301 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
5302 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
5303 | NULL_TREE, true, GSI_SAME_STMT); | |
5304 | assign_stmt = gimple_build_assign (cond_var, t2); | |
5305 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5306 | } | |
5307 | ||
5308 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
5309 | { | |
5310 | if (POINTER_TYPE_P (type)) | |
5311 | t = fold_build_pointer_plus (vmain, step); | |
5312 | else | |
5313 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
5314 | t = force_gimple_operand_gsi (&gsi, t, | |
5315 | DECL_P (vback) | |
5316 | && TREE_ADDRESSABLE (vback), | |
5317 | NULL_TREE, true, GSI_SAME_STMT); | |
5318 | assign_stmt = gimple_build_assign (vback, t); | |
5319 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5320 | ||
5321 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
5322 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
5323 | ? t : vback, e); | |
5324 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
5325 | } | |
5326 | ||
5327 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ | |
5328 | gsi_remove (&gsi, true); | |
5329 | ||
5330 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
5331 | collapse_bb = extract_omp_for_update_vars (fd, nonrect_bounds, | |
5332 | cont_bb, body_bb); | |
5333 | } | |
5334 | ||
5335 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ | |
5336 | gsi = gsi_last_nondebug_bb (exit_bb); | |
5337 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
5338 | { | |
5339 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
5340 | if (fd->have_reductemp | |
5341 | || ((fd->have_pointer_condtemp || fd->have_scantemp) | |
5342 | && !fd->have_nonctrl_scantemp)) | |
5343 | { | |
5344 | tree fn; | |
5345 | if (t) | |
5346 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
5347 | else | |
5348 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
5349 | gcall *g = gimple_build_call (fn, 0); | |
5350 | if (t) | |
5351 | { | |
5352 | gimple_call_set_lhs (g, t); | |
5353 | if (fd->have_reductemp) | |
5354 | gsi_insert_after (&gsi, gimple_build_assign (reductions, | |
5355 | NOP_EXPR, t), | |
5356 | GSI_SAME_STMT); | |
5357 | } | |
5358 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
5359 | } | |
5360 | else | |
5361 | gsi_insert_after (&gsi, omp_build_barrier (t), GSI_SAME_STMT); | |
5362 | } | |
5363 | else if ((fd->have_pointer_condtemp || fd->have_scantemp) | |
5364 | && !fd->have_nonctrl_scantemp) | |
5365 | { | |
5366 | tree fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
5367 | gcall *g = gimple_build_call (fn, 0); | |
5368 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
5369 | } | |
5370 | if (fd->have_scantemp && !fd->have_nonctrl_scantemp) | |
5371 | { | |
5372 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5373 | tree controlp = NULL_TREE, controlb = NULL_TREE; | |
5374 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
5375 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
5376 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) | |
5377 | { | |
5378 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) | |
5379 | controlb = OMP_CLAUSE_DECL (c); | |
5380 | else | |
5381 | controlp = OMP_CLAUSE_DECL (c); | |
5382 | if (controlb && controlp) | |
5383 | break; | |
5384 | } | |
5385 | gcc_assert (controlp && controlb); | |
5386 | gimple *g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, | |
5387 | NULL_TREE, NULL_TREE); | |
5388 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5389 | exit1_bb = split_block (exit_bb, g)->dest; | |
5390 | gsi = gsi_after_labels (exit1_bb); | |
5391 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_FREE), 1, | |
5392 | controlp); | |
5393 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5394 | exit2_bb = split_block (exit1_bb, g)->dest; | |
5395 | gsi = gsi_after_labels (exit2_bb); | |
5396 | g = gimple_build_call (builtin_decl_implicit (BUILT_IN_STACK_RESTORE), 1, | |
5397 | controlp); | |
5398 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5399 | exit3_bb = split_block (exit2_bb, g)->dest; | |
5400 | gsi = gsi_after_labels (exit3_bb); | |
5401 | } | |
5402 | gsi_remove (&gsi, true); | |
5403 | ||
5404 | /* Connect all the blocks. */ | |
5405 | ep = make_edge (entry_bb, third_bb, EDGE_FALSE_VALUE); | |
5406 | ep->probability = profile_probability::guessed_always ().apply_scale (3, 4); | |
5407 | ep = find_edge (entry_bb, second_bb); | |
5408 | ep->flags = EDGE_TRUE_VALUE; | |
5409 | ep->probability = profile_probability::guessed_always ().apply_scale (1, 4); | |
5410 | if (fourth_bb) | |
5411 | { | |
5412 | ep = make_edge (third_bb, fifth_bb, EDGE_FALSE_VALUE); | |
5413 | ep->probability | |
5414 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
5415 | ep = find_edge (third_bb, fourth_bb); | |
5416 | ep->flags = EDGE_TRUE_VALUE; | |
5417 | ep->probability | |
5418 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
5419 | ep = find_edge (fourth_bb, fifth_bb); | |
5420 | redirect_edge_and_branch (ep, sixth_bb); | |
5421 | } | |
5422 | else | |
5423 | sixth_bb = third_bb; | |
5424 | find_edge (sixth_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE; | |
5425 | find_edge (sixth_bb, fin_bb)->flags = EDGE_TRUE_VALUE; | |
5426 | if (exit1_bb) | |
5427 | { | |
5428 | ep = make_edge (exit_bb, exit2_bb, EDGE_FALSE_VALUE); | |
5429 | ep->probability | |
5430 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
5431 | ep = find_edge (exit_bb, exit1_bb); | |
5432 | ep->flags = EDGE_TRUE_VALUE; | |
5433 | ep->probability | |
5434 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
5435 | ep = find_edge (exit1_bb, exit2_bb); | |
5436 | redirect_edge_and_branch (ep, exit3_bb); | |
5437 | } | |
5438 | ||
5439 | if (!broken_loop) | |
5440 | { | |
5441 | ep = find_edge (cont_bb, body_bb); | |
5442 | if (ep == NULL) | |
5443 | { | |
5444 | ep = BRANCH_EDGE (cont_bb); | |
5445 | gcc_assert (single_succ (ep->dest) == body_bb); | |
5446 | } | |
5447 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5448 | { | |
5449 | remove_edge (ep); | |
5450 | ep = NULL; | |
5451 | } | |
5452 | else if (fd->collapse > 1) | |
5453 | { | |
5454 | remove_edge (ep); | |
5455 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
5456 | } | |
5457 | else | |
5458 | ep->flags = EDGE_TRUE_VALUE; | |
5459 | find_edge (cont_bb, fin_bb)->flags | |
5460 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
5461 | } | |
5462 | ||
5463 | set_immediate_dominator (CDI_DOMINATORS, second_bb, entry_bb); | |
5464 | set_immediate_dominator (CDI_DOMINATORS, third_bb, entry_bb); | |
5465 | if (fourth_bb) | |
5466 | { | |
5467 | set_immediate_dominator (CDI_DOMINATORS, fifth_bb, third_bb); | |
5468 | set_immediate_dominator (CDI_DOMINATORS, sixth_bb, third_bb); | |
5469 | } | |
5470 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, sixth_bb); | |
5471 | ||
5472 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
5473 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
5474 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
5475 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
5476 | if (exit1_bb) | |
5477 | { | |
5478 | set_immediate_dominator (CDI_DOMINATORS, exit2_bb, exit_bb); | |
5479 | set_immediate_dominator (CDI_DOMINATORS, exit3_bb, exit_bb); | |
5480 | } | |
5481 | ||
5482 | class loop *loop = body_bb->loop_father; | |
5483 | if (loop != entry_bb->loop_father) | |
5484 | { | |
5485 | gcc_assert (broken_loop || loop->header == body_bb); | |
5486 | gcc_assert (broken_loop | |
5487 | || loop->latch == region->cont | |
5488 | || single_pred (loop->latch) == region->cont); | |
5489 | return; | |
5490 | } | |
5491 | ||
5492 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) | |
5493 | { | |
5494 | loop = alloc_loop (); | |
5495 | loop->header = body_bb; | |
5496 | if (collapse_bb == NULL) | |
5497 | loop->latch = cont_bb; | |
5498 | add_loop (loop, body_bb->loop_father); | |
5499 | } | |
5500 | } | |
5501 | ||
5502 | /* Return phi in E->DEST with ARG on edge E. */ | |
5503 | ||
5504 | static gphi * | |
5505 | find_phi_with_arg_on_edge (tree arg, edge e) | |
5506 | { | |
5507 | basic_block bb = e->dest; | |
5508 | ||
5509 | for (gphi_iterator gpi = gsi_start_phis (bb); | |
5510 | !gsi_end_p (gpi); | |
5511 | gsi_next (&gpi)) | |
5512 | { | |
5513 | gphi *phi = gpi.phi (); | |
5514 | if (PHI_ARG_DEF_FROM_EDGE (phi, e) == arg) | |
5515 | return phi; | |
5516 | } | |
5517 | ||
5518 | return NULL; | |
5519 | } | |
5520 | ||
5521 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
5522 | loop with static schedule and a specified chunk size. Given | |
5523 | parameters: | |
5524 | ||
5525 | for (V = N1; V cond N2; V += STEP) BODY; | |
5526 | ||
5527 | where COND is "<" or ">", we generate pseudocode | |
5528 | ||
5529 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; | |
5530 | if (cond is <) | |
5531 | adj = STEP - 1; | |
5532 | else | |
5533 | adj = STEP + 1; | |
5534 | if ((__typeof (V)) -1 > 0 && cond is >) | |
5535 | n = -(adj + N2 - N1) / -STEP; | |
5536 | else | |
5537 | n = (adj + N2 - N1) / STEP; | |
5538 | trip = 0; | |
5539 | V = threadid * CHUNK * STEP + N1; -- this extra definition of V is | |
5540 | here so that V is defined | |
5541 | if the loop is not entered | |
5542 | L0: | |
5543 | s0 = (trip * nthreads + threadid) * CHUNK; | |
5544 | e0 = min (s0 + CHUNK, n); | |
5545 | if (s0 < n) goto L1; else goto L4; | |
5546 | L1: | |
5547 | V = s0 * STEP + N1; | |
5548 | e = e0 * STEP + N1; | |
5549 | L2: | |
5550 | BODY; | |
5551 | V += STEP; | |
5552 | if (V cond e) goto L2; else goto L3; | |
5553 | L3: | |
5554 | trip += 1; | |
5555 | goto L0; | |
5556 | L4: | |
5557 | */ | |
5558 | ||
5559 | static void | |
5560 | expand_omp_for_static_chunk (struct omp_region *region, | |
5561 | struct omp_for_data *fd, gimple *inner_stmt) | |
5562 | { | |
5563 | tree n, s0, e0, e, t; | |
5564 | tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid; | |
5565 | tree type, itype, vmain, vback, vextra; | |
5566 | basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb; | |
5567 | basic_block trip_update_bb = NULL, cont_bb, collapse_bb = NULL, fin_bb; | |
5568 | gimple_stmt_iterator gsi, gsip; | |
5569 | edge se; | |
5570 | bool broken_loop = region->cont == NULL; | |
5571 | tree *counts = NULL; | |
5572 | tree n1, n2, step; | |
5573 | tree reductions = NULL_TREE; | |
5574 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; | |
5575 | ||
5576 | itype = type = TREE_TYPE (fd->loop.v); | |
5577 | if (POINTER_TYPE_P (type)) | |
5578 | itype = signed_type_for (type); | |
5579 | ||
5580 | entry_bb = region->entry; | |
5581 | se = split_block (entry_bb, last_stmt (entry_bb)); | |
5582 | entry_bb = se->src; | |
5583 | iter_part_bb = se->dest; | |
5584 | cont_bb = region->cont; | |
5585 | gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2); | |
5586 | fin_bb = BRANCH_EDGE (iter_part_bb)->dest; | |
5587 | gcc_assert (broken_loop | |
5588 | || fin_bb == FALLTHRU_EDGE (cont_bb)->dest); | |
5589 | seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb)); | |
5590 | body_bb = single_succ (seq_start_bb); | |
5591 | if (!broken_loop) | |
5592 | { | |
5593 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb | |
5594 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); | |
5595 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
5596 | trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb)); | |
5597 | } | |
5598 | exit_bb = region->exit; | |
5599 | ||
5600 | /* Trip and adjustment setup goes in ENTRY_BB. */ | |
5601 | gsi = gsi_last_nondebug_bb (entry_bb); | |
5602 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
5603 | gsip = gsi; | |
5604 | gsi_prev (&gsip); | |
5605 | ||
5606 | if (fd->collapse > 1) | |
5607 | { | |
5608 | int first_zero_iter = -1, dummy = -1; | |
5609 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
5610 | ||
5611 | counts = XALLOCAVEC (tree, fd->collapse); | |
5612 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5613 | fin_bb, first_zero_iter, | |
5614 | dummy_bb, dummy, l2_dom_bb); | |
5615 | t = NULL_TREE; | |
5616 | } | |
5617 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5618 | t = integer_one_node; | |
5619 | else | |
5620 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
5621 | fold_convert (type, fd->loop.n1), | |
5622 | fold_convert (type, fd->loop.n2)); | |
5623 | if (fd->collapse == 1 | |
5624 | && TYPE_UNSIGNED (type) | |
5625 | && (t == NULL_TREE || !integer_onep (t))) | |
5626 | { | |
5627 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); | |
5628 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
5629 | true, GSI_SAME_STMT); | |
5630 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
5631 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
5632 | true, GSI_SAME_STMT); | |
5633 | gcond *cond_stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, | |
5634 | NULL_TREE, NULL_TREE); | |
5635 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
5636 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
5637 | expand_omp_regimplify_p, NULL, NULL) | |
5638 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
5639 | expand_omp_regimplify_p, NULL, NULL)) | |
5640 | { | |
5641 | gsi = gsi_for_stmt (cond_stmt); | |
5642 | gimple_regimplify_operands (cond_stmt, &gsi); | |
5643 | } | |
5644 | se = split_block (entry_bb, cond_stmt); | |
5645 | se->flags = EDGE_TRUE_VALUE; | |
5646 | entry_bb = se->dest; | |
5647 | se->probability = profile_probability::very_likely (); | |
5648 | se = make_edge (se->src, fin_bb, EDGE_FALSE_VALUE); | |
5649 | se->probability = profile_probability::very_unlikely (); | |
5650 | if (gimple_in_ssa_p (cfun)) | |
5651 | { | |
5652 | int dest_idx = find_edge (iter_part_bb, fin_bb)->dest_idx; | |
5653 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); | |
5654 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
5655 | { | |
5656 | gphi *phi = gpi.phi (); | |
5657 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
5658 | se, UNKNOWN_LOCATION); | |
5659 | } | |
5660 | } | |
5661 | gsi = gsi_last_bb (entry_bb); | |
5662 | } | |
5663 | ||
5664 | if (fd->lastprivate_conditional) | |
5665 | { | |
5666 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5667 | tree c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); | |
5668 | if (fd->have_pointer_condtemp) | |
5669 | condtemp = OMP_CLAUSE_DECL (c); | |
5670 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
5671 | cond_var = OMP_CLAUSE_DECL (c); | |
5672 | } | |
5673 | if (fd->have_reductemp || fd->have_pointer_condtemp) | |
5674 | { | |
5675 | tree t1 = build_int_cst (long_integer_type_node, 0); | |
5676 | tree t2 = build_int_cst (long_integer_type_node, 1); | |
5677 | tree t3 = build_int_cstu (long_integer_type_node, | |
5678 | (HOST_WIDE_INT_1U << 31) + 1); | |
5679 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5680 | gimple_stmt_iterator gsi2 = gsi_none (); | |
5681 | gimple *g = NULL; | |
5682 | tree mem = null_pointer_node, memv = NULL_TREE; | |
5683 | if (fd->have_reductemp) | |
5684 | { | |
5685 | tree c = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
5686 | reductions = OMP_CLAUSE_DECL (c); | |
5687 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
5688 | g = SSA_NAME_DEF_STMT (reductions); | |
5689 | reductions = gimple_assign_rhs1 (g); | |
5690 | OMP_CLAUSE_DECL (c) = reductions; | |
5691 | gsi2 = gsi_for_stmt (g); | |
5692 | } | |
5693 | else | |
5694 | { | |
5695 | if (gsi_end_p (gsip)) | |
5696 | gsi2 = gsi_after_labels (region->entry); | |
5697 | else | |
5698 | gsi2 = gsip; | |
5699 | reductions = null_pointer_node; | |
5700 | } | |
5701 | if (fd->have_pointer_condtemp) | |
5702 | { | |
5703 | tree type = TREE_TYPE (condtemp); | |
5704 | memv = create_tmp_var (type); | |
5705 | TREE_ADDRESSABLE (memv) = 1; | |
5706 | unsigned HOST_WIDE_INT sz | |
5707 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
5708 | sz *= fd->lastprivate_conditional; | |
5709 | expand_omp_build_assign (&gsi2, memv, build_int_cst (type, sz), | |
5710 | false); | |
5711 | mem = build_fold_addr_expr (memv); | |
5712 | } | |
5713 | tree t | |
5714 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_LOOP_START), | |
5715 | 9, t1, t2, t2, t3, t1, null_pointer_node, | |
5716 | null_pointer_node, reductions, mem); | |
5717 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
5718 | true, GSI_SAME_STMT); | |
5719 | if (fd->have_pointer_condtemp) | |
5720 | expand_omp_build_assign (&gsi2, condtemp, memv, false); | |
5721 | if (fd->have_reductemp) | |
5722 | { | |
5723 | gsi_remove (&gsi2, true); | |
5724 | release_ssa_name (gimple_assign_lhs (g)); | |
5725 | } | |
5726 | } | |
5727 | switch (gimple_omp_for_kind (fd->for_stmt)) | |
5728 | { | |
5729 | case GF_OMP_FOR_KIND_FOR: | |
5730 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
5731 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
5732 | break; | |
5733 | case GF_OMP_FOR_KIND_DISTRIBUTE: | |
5734 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS); | |
5735 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM); | |
5736 | break; | |
5737 | default: | |
5738 | gcc_unreachable (); | |
5739 | } | |
5740 | nthreads = build_call_expr (nthreads, 0); | |
5741 | nthreads = fold_convert (itype, nthreads); | |
5742 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, | |
5743 | true, GSI_SAME_STMT); | |
5744 | threadid = build_call_expr (threadid, 0); | |
5745 | threadid = fold_convert (itype, threadid); | |
5746 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, | |
5747 | true, GSI_SAME_STMT); | |
5748 | ||
5749 | n1 = fd->loop.n1; | |
5750 | n2 = fd->loop.n2; | |
5751 | step = fd->loop.step; | |
5752 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5753 | { | |
5754 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5755 | OMP_CLAUSE__LOOPTEMP_); | |
5756 | gcc_assert (innerc); | |
5757 | n1 = OMP_CLAUSE_DECL (innerc); | |
5758 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5759 | OMP_CLAUSE__LOOPTEMP_); | |
5760 | gcc_assert (innerc); | |
5761 | n2 = OMP_CLAUSE_DECL (innerc); | |
5762 | } | |
5763 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
5764 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5765 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
5766 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5767 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
5768 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5769 | tree chunk_size = fold_convert (itype, fd->chunk_size); | |
5770 | chunk_size = omp_adjust_chunk_size (chunk_size, fd->simd_schedule); | |
5771 | chunk_size | |
5772 | = force_gimple_operand_gsi (&gsi, chunk_size, true, NULL_TREE, true, | |
5773 | GSI_SAME_STMT); | |
5774 | ||
5775 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
5776 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
5777 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
5778 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
5779 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
5780 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
5781 | fold_build1 (NEGATE_EXPR, itype, t), | |
5782 | fold_build1 (NEGATE_EXPR, itype, step)); | |
5783 | else | |
5784 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
5785 | t = fold_convert (itype, t); | |
5786 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5787 | true, GSI_SAME_STMT); | |
5788 | ||
5789 | trip_var = create_tmp_reg (itype, ".trip"); | |
5790 | if (gimple_in_ssa_p (cfun)) | |
5791 | { | |
5792 | trip_init = make_ssa_name (trip_var); | |
5793 | trip_main = make_ssa_name (trip_var); | |
5794 | trip_back = make_ssa_name (trip_var); | |
5795 | } | |
5796 | else | |
5797 | { | |
5798 | trip_init = trip_var; | |
5799 | trip_main = trip_var; | |
5800 | trip_back = trip_var; | |
5801 | } | |
5802 | ||
5803 | gassign *assign_stmt | |
5804 | = gimple_build_assign (trip_init, build_int_cst (itype, 0)); | |
5805 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5806 | ||
5807 | t = fold_build2 (MULT_EXPR, itype, threadid, chunk_size); | |
5808 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5809 | if (POINTER_TYPE_P (type)) | |
5810 | t = fold_build_pointer_plus (n1, t); | |
5811 | else | |
5812 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5813 | vextra = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5814 | true, GSI_SAME_STMT); | |
5815 | ||
5816 | /* Remove the GIMPLE_OMP_FOR. */ | |
5817 | gsi_remove (&gsi, true); | |
5818 | ||
5819 | gimple_stmt_iterator gsif = gsi; | |
5820 | ||
5821 | /* Iteration space partitioning goes in ITER_PART_BB. */ | |
5822 | gsi = gsi_last_bb (iter_part_bb); | |
5823 | ||
5824 | t = fold_build2 (MULT_EXPR, itype, trip_main, nthreads); | |
5825 | t = fold_build2 (PLUS_EXPR, itype, t, threadid); | |
5826 | t = fold_build2 (MULT_EXPR, itype, t, chunk_size); | |
5827 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5828 | false, GSI_CONTINUE_LINKING); | |
5829 | ||
5830 | t = fold_build2 (PLUS_EXPR, itype, s0, chunk_size); | |
5831 | t = fold_build2 (MIN_EXPR, itype, t, n); | |
5832 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5833 | false, GSI_CONTINUE_LINKING); | |
5834 | ||
5835 | t = build2 (LT_EXPR, boolean_type_node, s0, n); | |
5836 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_CONTINUE_LINKING); | |
5837 | ||
5838 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
5839 | gsi = gsi_start_bb (seq_start_bb); | |
5840 | ||
5841 | tree startvar = fd->loop.v; | |
5842 | tree endvar = NULL_TREE; | |
5843 | ||
5844 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5845 | { | |
5846 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
5847 | ? gimple_omp_parallel_clauses (inner_stmt) | |
5848 | : gimple_omp_for_clauses (inner_stmt); | |
5849 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
5850 | gcc_assert (innerc); | |
5851 | startvar = OMP_CLAUSE_DECL (innerc); | |
5852 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5853 | OMP_CLAUSE__LOOPTEMP_); | |
5854 | gcc_assert (innerc); | |
5855 | endvar = OMP_CLAUSE_DECL (innerc); | |
5856 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST | |
5857 | && gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) | |
5858 | { | |
5859 | innerc = find_lastprivate_looptemp (fd, innerc); | |
5860 | if (innerc) | |
5861 | { | |
5862 | /* If needed (distribute parallel for with lastprivate), | |
5863 | propagate down the total number of iterations. */ | |
5864 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), | |
5865 | fd->loop.n2); | |
5866 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, | |
5867 | GSI_CONTINUE_LINKING); | |
5868 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
5869 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5870 | } | |
5871 | } | |
5872 | } | |
5873 | ||
5874 | t = fold_convert (itype, s0); | |
5875 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5876 | if (POINTER_TYPE_P (type)) | |
5877 | { | |
5878 | t = fold_build_pointer_plus (n1, t); | |
5879 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
5880 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
5881 | t = fold_convert (signed_type_for (type), t); | |
5882 | } | |
5883 | else | |
5884 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5885 | t = fold_convert (TREE_TYPE (startvar), t); | |
5886 | t = force_gimple_operand_gsi (&gsi, t, | |
5887 | DECL_P (startvar) | |
5888 | && TREE_ADDRESSABLE (startvar), | |
5889 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5890 | assign_stmt = gimple_build_assign (startvar, t); | |
5891 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5892 | if (cond_var) | |
5893 | { | |
5894 | tree itype = TREE_TYPE (cond_var); | |
5895 | /* For lastprivate(conditional:) itervar, we need some iteration | |
5896 | counter that starts at unsigned non-zero and increases. | |
5897 | Prefer as few IVs as possible, so if we can use startvar | |
5898 | itself, use that, or startvar + constant (those would be | |
5899 | incremented with step), and as last resort use the s0 + 1 | |
5900 | incremented by 1. */ | |
5901 | if (POINTER_TYPE_P (type) | |
5902 | || TREE_CODE (n1) != INTEGER_CST | |
5903 | || fd->loop.cond_code != LT_EXPR) | |
5904 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), | |
5905 | build_int_cst (itype, 1)); | |
5906 | else if (tree_int_cst_sgn (n1) == 1) | |
5907 | t = fold_convert (itype, t); | |
5908 | else | |
5909 | { | |
5910 | tree c = fold_convert (itype, n1); | |
5911 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
5912 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
5913 | } | |
5914 | t = force_gimple_operand_gsi (&gsi, t, false, | |
5915 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5916 | assign_stmt = gimple_build_assign (cond_var, t); | |
5917 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5918 | } | |
5919 | ||
5920 | t = fold_convert (itype, e0); | |
5921 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5922 | if (POINTER_TYPE_P (type)) | |
5923 | { | |
5924 | t = fold_build_pointer_plus (n1, t); | |
5925 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
5926 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
5927 | t = fold_convert (signed_type_for (type), t); | |
5928 | } | |
5929 | else | |
5930 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5931 | t = fold_convert (TREE_TYPE (startvar), t); | |
5932 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5933 | false, GSI_CONTINUE_LINKING); | |
5934 | if (endvar) | |
5935 | { | |
5936 | assign_stmt = gimple_build_assign (endvar, e); | |
5937 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5938 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
5939 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
5940 | else | |
5941 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
5942 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5943 | } | |
5944 | /* Handle linear clause adjustments. */ | |
5945 | tree itercnt = NULL_TREE, itercntbias = NULL_TREE; | |
5946 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
5947 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
5948 | c; c = OMP_CLAUSE_CHAIN (c)) | |
5949 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
5950 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
5951 | { | |
5952 | tree d = OMP_CLAUSE_DECL (c); | |
5953 | tree t = d, a, dest; | |
5954 | if (omp_privatize_by_reference (t)) | |
5955 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); | |
5956 | tree type = TREE_TYPE (t); | |
5957 | if (POINTER_TYPE_P (type)) | |
5958 | type = sizetype; | |
5959 | dest = unshare_expr (t); | |
5960 | tree v = create_tmp_var (TREE_TYPE (t), NULL); | |
5961 | expand_omp_build_assign (&gsif, v, t); | |
5962 | if (itercnt == NULL_TREE) | |
5963 | { | |
5964 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5965 | { | |
5966 | itercntbias | |
5967 | = fold_build2 (MINUS_EXPR, itype, fold_convert (itype, n1), | |
5968 | fold_convert (itype, fd->loop.n1)); | |
5969 | itercntbias = fold_build2 (EXACT_DIV_EXPR, itype, | |
5970 | itercntbias, step); | |
5971 | itercntbias | |
5972 | = force_gimple_operand_gsi (&gsif, itercntbias, true, | |
5973 | NULL_TREE, true, | |
5974 | GSI_SAME_STMT); | |
5975 | itercnt = fold_build2 (PLUS_EXPR, itype, itercntbias, s0); | |
5976 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
5977 | NULL_TREE, false, | |
5978 | GSI_CONTINUE_LINKING); | |
5979 | } | |
5980 | else | |
5981 | itercnt = s0; | |
5982 | } | |
5983 | a = fold_build2 (MULT_EXPR, type, | |
5984 | fold_convert (type, itercnt), | |
5985 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
5986 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
5987 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); | |
5988 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5989 | false, GSI_CONTINUE_LINKING); | |
5990 | expand_omp_build_assign (&gsi, dest, t, true); | |
5991 | } | |
5992 | if (fd->collapse > 1) | |
5993 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, inner_stmt, startvar); | |
5994 | ||
5995 | if (!broken_loop) | |
5996 | { | |
5997 | /* The code controlling the sequential loop goes in CONT_BB, | |
5998 | replacing the GIMPLE_OMP_CONTINUE. */ | |
5999 | gsi = gsi_last_nondebug_bb (cont_bb); | |
6000 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); | |
6001 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
6002 | vback = gimple_omp_continue_control_def (cont_stmt); | |
6003 | ||
6004 | if (cond_var) | |
6005 | { | |
6006 | tree itype = TREE_TYPE (cond_var); | |
6007 | tree t2; | |
6008 | if (POINTER_TYPE_P (type) | |
6009 | || TREE_CODE (n1) != INTEGER_CST | |
6010 | || fd->loop.cond_code != LT_EXPR) | |
6011 | t2 = build_int_cst (itype, 1); | |
6012 | else | |
6013 | t2 = fold_convert (itype, step); | |
6014 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
6015 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
6016 | NULL_TREE, true, GSI_SAME_STMT); | |
6017 | assign_stmt = gimple_build_assign (cond_var, t2); | |
6018 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
6019 | } | |
6020 | ||
6021 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
6022 | { | |
6023 | if (POINTER_TYPE_P (type)) | |
6024 | t = fold_build_pointer_plus (vmain, step); | |
6025 | else | |
6026 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
6027 | if (DECL_P (vback) && TREE_ADDRESSABLE (vback)) | |
6028 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6029 | true, GSI_SAME_STMT); | |
6030 | assign_stmt = gimple_build_assign (vback, t); | |
6031 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
6032 | ||
6033 | if (tree_int_cst_equal (fd->chunk_size, integer_one_node)) | |
6034 | t = build2 (EQ_EXPR, boolean_type_node, | |
6035 | build_int_cst (itype, 0), | |
6036 | build_int_cst (itype, 1)); | |
6037 | else | |
6038 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
6039 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
6040 | ? t : vback, e); | |
6041 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
6042 | } | |
6043 | ||
6044 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
6045 | gsi_remove (&gsi, true); | |
6046 | ||
6047 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
6048 | collapse_bb = extract_omp_for_update_vars (fd, NULL, cont_bb, body_bb); | |
6049 | ||
6050 | /* Trip update code goes into TRIP_UPDATE_BB. */ | |
6051 | gsi = gsi_start_bb (trip_update_bb); | |
6052 | ||
6053 | t = build_int_cst (itype, 1); | |
6054 | t = build2 (PLUS_EXPR, itype, trip_main, t); | |
6055 | assign_stmt = gimple_build_assign (trip_back, t); | |
6056 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6057 | } | |
6058 | ||
6059 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ | |
6060 | gsi = gsi_last_nondebug_bb (exit_bb); | |
6061 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) | |
6062 | { | |
6063 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
6064 | if (fd->have_reductemp || fd->have_pointer_condtemp) | |
6065 | { | |
6066 | tree fn; | |
6067 | if (t) | |
6068 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
6069 | else | |
6070 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
6071 | gcall *g = gimple_build_call (fn, 0); | |
6072 | if (t) | |
6073 | { | |
6074 | gimple_call_set_lhs (g, t); | |
6075 | if (fd->have_reductemp) | |
6076 | gsi_insert_after (&gsi, gimple_build_assign (reductions, | |
6077 | NOP_EXPR, t), | |
6078 | GSI_SAME_STMT); | |
6079 | } | |
6080 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
6081 | } | |
6082 | else | |
6083 | gsi_insert_after (&gsi, omp_build_barrier (t), GSI_SAME_STMT); | |
6084 | } | |
6085 | else if (fd->have_pointer_condtemp) | |
6086 | { | |
6087 | tree fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
6088 | gcall *g = gimple_build_call (fn, 0); | |
6089 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
6090 | } | |
6091 | gsi_remove (&gsi, true); | |
6092 | ||
6093 | /* Connect the new blocks. */ | |
6094 | find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE; | |
6095 | find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE; | |
6096 | ||
6097 | if (!broken_loop) | |
6098 | { | |
6099 | se = find_edge (cont_bb, body_bb); | |
6100 | if (se == NULL) | |
6101 | { | |
6102 | se = BRANCH_EDGE (cont_bb); | |
6103 | gcc_assert (single_succ (se->dest) == body_bb); | |
6104 | } | |
6105 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6106 | { | |
6107 | remove_edge (se); | |
6108 | se = NULL; | |
6109 | } | |
6110 | else if (fd->collapse > 1) | |
6111 | { | |
6112 | remove_edge (se); | |
6113 | se = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
6114 | } | |
6115 | else | |
6116 | se->flags = EDGE_TRUE_VALUE; | |
6117 | find_edge (cont_bb, trip_update_bb)->flags | |
6118 | = se ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
6119 | ||
6120 | redirect_edge_and_branch (single_succ_edge (trip_update_bb), | |
6121 | iter_part_bb); | |
6122 | } | |
6123 | ||
6124 | if (gimple_in_ssa_p (cfun)) | |
6125 | { | |
6126 | gphi_iterator psi; | |
6127 | gphi *phi; | |
6128 | edge re, ene; | |
6129 | edge_var_map *vm; | |
6130 | size_t i; | |
6131 | ||
6132 | gcc_assert (fd->collapse == 1 && !broken_loop); | |
6133 | ||
6134 | /* When we redirect the edge from trip_update_bb to iter_part_bb, we | |
6135 | remove arguments of the phi nodes in fin_bb. We need to create | |
6136 | appropriate phi nodes in iter_part_bb instead. */ | |
6137 | se = find_edge (iter_part_bb, fin_bb); | |
6138 | re = single_succ_edge (trip_update_bb); | |
6139 | vec<edge_var_map> *head = redirect_edge_var_map_vector (re); | |
6140 | ene = single_succ_edge (entry_bb); | |
6141 | ||
6142 | psi = gsi_start_phis (fin_bb); | |
6143 | for (i = 0; !gsi_end_p (psi) && head->iterate (i, &vm); | |
6144 | gsi_next (&psi), ++i) | |
6145 | { | |
6146 | gphi *nphi; | |
6147 | location_t locus; | |
6148 | ||
6149 | phi = psi.phi (); | |
6150 | if (operand_equal_p (gimple_phi_arg_def (phi, 0), | |
6151 | redirect_edge_var_map_def (vm), 0)) | |
6152 | continue; | |
6153 | ||
6154 | t = gimple_phi_result (phi); | |
6155 | gcc_assert (t == redirect_edge_var_map_result (vm)); | |
6156 | ||
6157 | if (!single_pred_p (fin_bb)) | |
6158 | t = copy_ssa_name (t, phi); | |
6159 | ||
6160 | nphi = create_phi_node (t, iter_part_bb); | |
6161 | ||
6162 | t = PHI_ARG_DEF_FROM_EDGE (phi, se); | |
6163 | locus = gimple_phi_arg_location_from_edge (phi, se); | |
6164 | ||
6165 | /* A special case -- fd->loop.v is not yet computed in | |
6166 | iter_part_bb, we need to use vextra instead. */ | |
6167 | if (t == fd->loop.v) | |
6168 | t = vextra; | |
6169 | add_phi_arg (nphi, t, ene, locus); | |
6170 | locus = redirect_edge_var_map_location (vm); | |
6171 | tree back_arg = redirect_edge_var_map_def (vm); | |
6172 | add_phi_arg (nphi, back_arg, re, locus); | |
6173 | edge ce = find_edge (cont_bb, body_bb); | |
6174 | if (ce == NULL) | |
6175 | { | |
6176 | ce = BRANCH_EDGE (cont_bb); | |
6177 | gcc_assert (single_succ (ce->dest) == body_bb); | |
6178 | ce = single_succ_edge (ce->dest); | |
6179 | } | |
6180 | gphi *inner_loop_phi = find_phi_with_arg_on_edge (back_arg, ce); | |
6181 | gcc_assert (inner_loop_phi != NULL); | |
6182 | add_phi_arg (inner_loop_phi, gimple_phi_result (nphi), | |
6183 | find_edge (seq_start_bb, body_bb), locus); | |
6184 | ||
6185 | if (!single_pred_p (fin_bb)) | |
6186 | add_phi_arg (phi, gimple_phi_result (nphi), se, locus); | |
6187 | } | |
6188 | gcc_assert (gsi_end_p (psi) && (head == NULL || i == head->length ())); | |
6189 | redirect_edge_var_map_clear (re); | |
6190 | if (single_pred_p (fin_bb)) | |
6191 | while (1) | |
6192 | { | |
6193 | psi = gsi_start_phis (fin_bb); | |
6194 | if (gsi_end_p (psi)) | |
6195 | break; | |
6196 | remove_phi_node (&psi, false); | |
6197 | } | |
6198 | ||
6199 | /* Make phi node for trip. */ | |
6200 | phi = create_phi_node (trip_main, iter_part_bb); | |
6201 | add_phi_arg (phi, trip_back, single_succ_edge (trip_update_bb), | |
6202 | UNKNOWN_LOCATION); | |
6203 | add_phi_arg (phi, trip_init, single_succ_edge (entry_bb), | |
6204 | UNKNOWN_LOCATION); | |
6205 | } | |
6206 | ||
6207 | if (!broken_loop) | |
6208 | set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb); | |
6209 | set_immediate_dominator (CDI_DOMINATORS, iter_part_bb, | |
6210 | recompute_dominator (CDI_DOMINATORS, iter_part_bb)); | |
6211 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
6212 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
6213 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, | |
6214 | recompute_dominator (CDI_DOMINATORS, seq_start_bb)); | |
6215 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
6216 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
6217 | ||
6218 | if (!broken_loop) | |
6219 | { | |
6220 | class loop *loop = body_bb->loop_father; | |
6221 | class loop *trip_loop = alloc_loop (); | |
6222 | trip_loop->header = iter_part_bb; | |
6223 | trip_loop->latch = trip_update_bb; | |
6224 | add_loop (trip_loop, iter_part_bb->loop_father); | |
6225 | ||
6226 | if (loop != entry_bb->loop_father) | |
6227 | { | |
6228 | gcc_assert (loop->header == body_bb); | |
6229 | gcc_assert (loop->latch == region->cont | |
6230 | || single_pred (loop->latch) == region->cont); | |
6231 | trip_loop->inner = loop; | |
6232 | return; | |
6233 | } | |
6234 | ||
6235 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
6236 | { | |
6237 | loop = alloc_loop (); | |
6238 | loop->header = body_bb; | |
6239 | if (collapse_bb == NULL) | |
6240 | loop->latch = cont_bb; | |
6241 | add_loop (loop, trip_loop); | |
6242 | } | |
6243 | } | |
6244 | } | |
6245 | ||
6246 | /* A subroutine of expand_omp_for. Generate code for a simd non-worksharing | |
6247 | loop. Given parameters: | |
6248 | ||
6249 | for (V = N1; V cond N2; V += STEP) BODY; | |
6250 | ||
6251 | where COND is "<" or ">", we generate pseudocode | |
6252 | ||
6253 | V = N1; | |
6254 | goto L1; | |
6255 | L0: | |
6256 | BODY; | |
6257 | V += STEP; | |
6258 | L1: | |
6259 | if (V cond N2) goto L0; else goto L2; | |
6260 | L2: | |
6261 | ||
6262 | For collapsed loops, emit the outer loops as scalar | |
6263 | and only try to vectorize the innermost loop. */ | |
6264 | ||
6265 | static void | |
6266 | expand_omp_simd (struct omp_region *region, struct omp_for_data *fd) | |
6267 | { | |
6268 | tree type, t; | |
6269 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, l2_bb, l2_dom_bb; | |
6270 | gimple_stmt_iterator gsi; | |
6271 | gimple *stmt; | |
6272 | gcond *cond_stmt; | |
6273 | bool broken_loop = region->cont == NULL; | |
6274 | edge e, ne; | |
6275 | tree *counts = NULL; | |
6276 | int i; | |
6277 | int safelen_int = INT_MAX; | |
6278 | bool dont_vectorize = false; | |
6279 | tree safelen = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6280 | OMP_CLAUSE_SAFELEN); | |
6281 | tree simduid = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6282 | OMP_CLAUSE__SIMDUID_); | |
6283 | tree ifc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6284 | OMP_CLAUSE_IF); | |
6285 | tree simdlen = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6286 | OMP_CLAUSE_SIMDLEN); | |
6287 | tree condtemp = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6288 | OMP_CLAUSE__CONDTEMP_); | |
6289 | tree n1, n2; | |
6290 | tree cond_var = condtemp ? OMP_CLAUSE_DECL (condtemp) : NULL_TREE; | |
6291 | ||
6292 | if (safelen) | |
6293 | { | |
6294 | poly_uint64 val; | |
6295 | safelen = OMP_CLAUSE_SAFELEN_EXPR (safelen); | |
6296 | if (!poly_int_tree_p (safelen, &val)) | |
6297 | safelen_int = 0; | |
6298 | else | |
6299 | safelen_int = MIN (constant_lower_bound (val), INT_MAX); | |
6300 | if (safelen_int == 1) | |
6301 | safelen_int = 0; | |
6302 | } | |
6303 | if ((ifc && integer_zerop (OMP_CLAUSE_IF_EXPR (ifc))) | |
6304 | || (simdlen && integer_onep (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)))) | |
6305 | { | |
6306 | safelen_int = 0; | |
6307 | dont_vectorize = true; | |
6308 | } | |
6309 | type = TREE_TYPE (fd->loop.v); | |
6310 | entry_bb = region->entry; | |
6311 | cont_bb = region->cont; | |
6312 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
6313 | gcc_assert (broken_loop | |
6314 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
6315 | l0_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
6316 | if (!broken_loop) | |
6317 | { | |
6318 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l0_bb); | |
6319 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
6320 | l1_bb = split_block (cont_bb, last_stmt (cont_bb))->dest; | |
6321 | l2_bb = BRANCH_EDGE (entry_bb)->dest; | |
6322 | } | |
6323 | else | |
6324 | { | |
6325 | BRANCH_EDGE (entry_bb)->flags &= ~EDGE_ABNORMAL; | |
6326 | l1_bb = split_edge (BRANCH_EDGE (entry_bb)); | |
6327 | l2_bb = single_succ (l1_bb); | |
6328 | } | |
6329 | exit_bb = region->exit; | |
6330 | l2_dom_bb = NULL; | |
6331 | ||
6332 | gsi = gsi_last_nondebug_bb (entry_bb); | |
6333 | ||
6334 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
6335 | /* Not needed in SSA form right now. */ | |
6336 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
6337 | if (fd->collapse > 1 | |
6338 | && (gimple_omp_for_combined_into_p (fd->for_stmt) | |
6339 | || broken_loop)) | |
6340 | { | |
6341 | int first_zero_iter = -1, dummy = -1; | |
6342 | basic_block zero_iter_bb = l2_bb, dummy_bb = NULL; | |
6343 | ||
6344 | counts = XALLOCAVEC (tree, fd->collapse); | |
6345 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
6346 | zero_iter_bb, first_zero_iter, | |
6347 | dummy_bb, dummy, l2_dom_bb); | |
6348 | } | |
6349 | if (l2_dom_bb == NULL) | |
6350 | l2_dom_bb = l1_bb; | |
6351 | ||
6352 | n1 = fd->loop.n1; | |
6353 | n2 = fd->loop.n2; | |
6354 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6355 | { | |
6356 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6357 | OMP_CLAUSE__LOOPTEMP_); | |
6358 | gcc_assert (innerc); | |
6359 | n1 = OMP_CLAUSE_DECL (innerc); | |
6360 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
6361 | OMP_CLAUSE__LOOPTEMP_); | |
6362 | gcc_assert (innerc); | |
6363 | n2 = OMP_CLAUSE_DECL (innerc); | |
6364 | } | |
6365 | tree step = fd->loop.step; | |
6366 | tree orig_step = step; /* May be different from step if is_simt. */ | |
6367 | ||
6368 | bool is_simt = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6369 | OMP_CLAUSE__SIMT_); | |
6370 | if (is_simt) | |
6371 | { | |
6372 | cfun->curr_properties &= ~PROP_gimple_lomp_dev; | |
6373 | is_simt = safelen_int > 1; | |
6374 | } | |
6375 | tree simt_lane = NULL_TREE, simt_maxlane = NULL_TREE; | |
6376 | if (is_simt) | |
6377 | { | |
6378 | simt_lane = create_tmp_var (unsigned_type_node); | |
6379 | gimple *g = gimple_build_call_internal (IFN_GOMP_SIMT_LANE, 0); | |
6380 | gimple_call_set_lhs (g, simt_lane); | |
6381 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
6382 | tree offset = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, | |
6383 | fold_convert (TREE_TYPE (step), simt_lane)); | |
6384 | n1 = fold_convert (type, n1); | |
6385 | if (POINTER_TYPE_P (type)) | |
6386 | n1 = fold_build_pointer_plus (n1, offset); | |
6387 | else | |
6388 | n1 = fold_build2 (PLUS_EXPR, type, n1, fold_convert (type, offset)); | |
6389 | ||
6390 | /* Collapsed loops not handled for SIMT yet: limit to one lane only. */ | |
6391 | if (fd->collapse > 1) | |
6392 | simt_maxlane = build_one_cst (unsigned_type_node); | |
6393 | else if (safelen_int < omp_max_simt_vf ()) | |
6394 | simt_maxlane = build_int_cst (unsigned_type_node, safelen_int); | |
6395 | tree vf | |
6396 | = build_call_expr_internal_loc (UNKNOWN_LOCATION, IFN_GOMP_SIMT_VF, | |
6397 | unsigned_type_node, 0); | |
6398 | if (simt_maxlane) | |
6399 | vf = fold_build2 (MIN_EXPR, unsigned_type_node, vf, simt_maxlane); | |
6400 | vf = fold_convert (TREE_TYPE (step), vf); | |
6401 | step = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, vf); | |
6402 | } | |
6403 | ||
6404 | tree n2var = NULL_TREE; | |
6405 | tree n2v = NULL_TREE; | |
6406 | tree *nonrect_bounds = NULL; | |
6407 | tree min_arg1 = NULL_TREE, min_arg2 = NULL_TREE; | |
6408 | if (fd->collapse > 1) | |
6409 | { | |
6410 | if (broken_loop || gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6411 | { | |
6412 | if (fd->non_rect) | |
6413 | { | |
6414 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); | |
6415 | memset (nonrect_bounds, 0, | |
6416 | sizeof (tree) * (fd->last_nonrect + 1)); | |
6417 | } | |
6418 | expand_omp_build_assign (&gsi, fd->loop.v, fold_convert (type, n1)); | |
6419 | gcc_assert (entry_bb == gsi_bb (gsi)); | |
6420 | gcc_assert (fd->for_stmt == gsi_stmt (gsi)); | |
6421 | gsi_prev (&gsi); | |
6422 | entry_bb = split_block (entry_bb, gsi_stmt (gsi))->dest; | |
6423 | expand_omp_for_init_vars (fd, &gsi, counts, nonrect_bounds, | |
6424 | NULL, n1); | |
6425 | gsi = gsi_for_stmt (fd->for_stmt); | |
6426 | } | |
6427 | if (broken_loop) | |
6428 | ; | |
6429 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6430 | { | |
6431 | /* Compute in n2var the limit for the first innermost loop, | |
6432 | i.e. fd->loop.v + MIN (n2 - fd->loop.v, cnt) | |
6433 | where cnt is how many iterations would the loop have if | |
6434 | all further iterations were assigned to the current task. */ | |
6435 | n2var = create_tmp_var (type); | |
6436 | i = fd->collapse - 1; | |
6437 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6438 | if (POINTER_TYPE_P (itype)) | |
6439 | itype = signed_type_for (itype); | |
6440 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
6441 | ? -1 : 1)); | |
6442 | t = fold_build2 (PLUS_EXPR, itype, | |
6443 | fold_convert (itype, fd->loops[i].step), t); | |
6444 | t = fold_build2 (PLUS_EXPR, itype, t, | |
6445 | fold_convert (itype, fd->loops[i].n2)); | |
6446 | if (fd->loops[i].m2) | |
6447 | { | |
6448 | tree t2 = fold_convert (itype, | |
6449 | fd->loops[i - fd->loops[i].outer].v); | |
6450 | tree t3 = fold_convert (itype, fd->loops[i].m2); | |
6451 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
6452 | t = fold_build2 (PLUS_EXPR, itype, t, t2); | |
6453 | } | |
6454 | t = fold_build2 (MINUS_EXPR, itype, t, | |
6455 | fold_convert (itype, fd->loops[i].v)); | |
6456 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
6457 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6458 | fold_build1 (NEGATE_EXPR, itype, t), | |
6459 | fold_build1 (NEGATE_EXPR, itype, | |
6460 | fold_convert (itype, | |
6461 | fd->loops[i].step))); | |
6462 | else | |
6463 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
6464 | fold_convert (itype, fd->loops[i].step)); | |
6465 | t = fold_convert (type, t); | |
6466 | tree t2 = fold_build2 (MINUS_EXPR, type, n2, n1); | |
6467 | min_arg1 = create_tmp_var (type); | |
6468 | expand_omp_build_assign (&gsi, min_arg1, t2); | |
6469 | min_arg2 = create_tmp_var (type); | |
6470 | expand_omp_build_assign (&gsi, min_arg2, t); | |
6471 | } | |
6472 | else | |
6473 | { | |
6474 | if (TREE_CODE (n2) == INTEGER_CST) | |
6475 | { | |
6476 | /* Indicate for lastprivate handling that at least one iteration | |
6477 | has been performed, without wasting runtime. */ | |
6478 | if (integer_nonzerop (n2)) | |
6479 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6480 | fold_convert (type, n2)); | |
6481 | else | |
6482 | /* Indicate that no iteration has been performed. */ | |
6483 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6484 | build_one_cst (type)); | |
6485 | } | |
6486 | else | |
6487 | { | |
6488 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6489 | build_zero_cst (type)); | |
6490 | expand_omp_build_assign (&gsi, n2, build_one_cst (type)); | |
6491 | } | |
6492 | for (i = 0; i < fd->collapse; i++) | |
6493 | { | |
6494 | t = fold_convert (TREE_TYPE (fd->loops[i].v), fd->loops[i].n1); | |
6495 | if (fd->loops[i].m1) | |
6496 | { | |
6497 | tree t2 | |
6498 | = fold_convert (TREE_TYPE (t), | |
6499 | fd->loops[i - fd->loops[i].outer].v); | |
6500 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i].m1); | |
6501 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
6502 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); | |
6503 | } | |
6504 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6505 | /* For normal non-combined collapsed loops just initialize | |
6506 | the outermost iterator in the entry_bb. */ | |
6507 | if (!broken_loop) | |
6508 | break; | |
6509 | } | |
6510 | } | |
6511 | } | |
6512 | else | |
6513 | expand_omp_build_assign (&gsi, fd->loop.v, fold_convert (type, n1)); | |
6514 | tree altv = NULL_TREE, altn2 = NULL_TREE; | |
6515 | if (fd->collapse == 1 | |
6516 | && !broken_loop | |
6517 | && TREE_CODE (orig_step) != INTEGER_CST) | |
6518 | { | |
6519 | /* The vectorizer currently punts on loops with non-constant steps | |
6520 | for the main IV (can't compute number of iterations and gives up | |
6521 | because of that). As for OpenMP loops it is always possible to | |
6522 | compute the number of iterations upfront, use an alternate IV | |
6523 | as the loop iterator: | |
6524 | altn2 = n1 < n2 ? (n2 - n1 + step - 1) / step : 0; | |
6525 | for (i = n1, altv = 0; altv < altn2; altv++, i += step) */ | |
6526 | altv = create_tmp_var (unsigned_type_for (TREE_TYPE (fd->loops[0].v))); | |
6527 | expand_omp_build_assign (&gsi, altv, build_zero_cst (TREE_TYPE (altv))); | |
6528 | tree itype = TREE_TYPE (fd->loop.v); | |
6529 | if (POINTER_TYPE_P (itype)) | |
6530 | itype = signed_type_for (itype); | |
6531 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
6532 | t = fold_build2 (PLUS_EXPR, itype, | |
6533 | fold_convert (itype, step), t); | |
6534 | t = fold_build2 (PLUS_EXPR, itype, t, fold_convert (itype, n2)); | |
6535 | t = fold_build2 (MINUS_EXPR, itype, t, | |
6536 | fold_convert (itype, fd->loop.v)); | |
6537 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
6538 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6539 | fold_build1 (NEGATE_EXPR, itype, t), | |
6540 | fold_build1 (NEGATE_EXPR, itype, | |
6541 | fold_convert (itype, step))); | |
6542 | else | |
6543 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
6544 | fold_convert (itype, step)); | |
6545 | t = fold_convert (TREE_TYPE (altv), t); | |
6546 | altn2 = create_tmp_var (TREE_TYPE (altv)); | |
6547 | expand_omp_build_assign (&gsi, altn2, t); | |
6548 | tree t2 = fold_convert (TREE_TYPE (fd->loop.v), n2); | |
6549 | t2 = force_gimple_operand_gsi (&gsi, t2, true, NULL_TREE, | |
6550 | true, GSI_SAME_STMT); | |
6551 | t2 = fold_build2 (fd->loop.cond_code, boolean_type_node, fd->loop.v, t2); | |
6552 | gassign *g = gimple_build_assign (altn2, COND_EXPR, t2, altn2, | |
6553 | build_zero_cst (TREE_TYPE (altv))); | |
6554 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
6555 | } | |
6556 | else if (fd->collapse > 1 | |
6557 | && !broken_loop | |
6558 | && !gimple_omp_for_combined_into_p (fd->for_stmt) | |
6559 | && TREE_CODE (fd->loops[fd->collapse - 1].step) != INTEGER_CST) | |
6560 | { | |
6561 | altv = create_tmp_var (unsigned_type_for (TREE_TYPE (fd->loops[0].v))); | |
6562 | altn2 = create_tmp_var (TREE_TYPE (altv)); | |
6563 | } | |
6564 | if (cond_var) | |
6565 | { | |
6566 | if (POINTER_TYPE_P (type) | |
6567 | || TREE_CODE (n1) != INTEGER_CST | |
6568 | || fd->loop.cond_code != LT_EXPR | |
6569 | || tree_int_cst_sgn (n1) != 1) | |
6570 | expand_omp_build_assign (&gsi, cond_var, | |
6571 | build_one_cst (TREE_TYPE (cond_var))); | |
6572 | else | |
6573 | expand_omp_build_assign (&gsi, cond_var, | |
6574 | fold_convert (TREE_TYPE (cond_var), n1)); | |
6575 | } | |
6576 | ||
6577 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
6578 | gsi_remove (&gsi, true); | |
6579 | ||
6580 | if (!broken_loop) | |
6581 | { | |
6582 | /* Code to control the increment goes in the CONT_BB. */ | |
6583 | gsi = gsi_last_nondebug_bb (cont_bb); | |
6584 | stmt = gsi_stmt (gsi); | |
6585 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
6586 | ||
6587 | if (fd->collapse == 1 | |
6588 | || gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6589 | { | |
6590 | if (POINTER_TYPE_P (type)) | |
6591 | t = fold_build_pointer_plus (fd->loop.v, step); | |
6592 | else | |
6593 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); | |
6594 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
6595 | } | |
6596 | else if (TREE_CODE (n2) != INTEGER_CST) | |
6597 | expand_omp_build_assign (&gsi, fd->loop.v, build_one_cst (type)); | |
6598 | if (altv) | |
6599 | { | |
6600 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (altv), altv, | |
6601 | build_one_cst (TREE_TYPE (altv))); | |
6602 | expand_omp_build_assign (&gsi, altv, t); | |
6603 | } | |
6604 | ||
6605 | if (fd->collapse > 1) | |
6606 | { | |
6607 | i = fd->collapse - 1; | |
6608 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
6609 | { | |
6610 | t = fold_convert (sizetype, fd->loops[i].step); | |
6611 | t = fold_build_pointer_plus (fd->loops[i].v, t); | |
6612 | } | |
6613 | else | |
6614 | { | |
6615 | t = fold_convert (TREE_TYPE (fd->loops[i].v), | |
6616 | fd->loops[i].step); | |
6617 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
6618 | fd->loops[i].v, t); | |
6619 | } | |
6620 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6621 | } | |
6622 | if (cond_var) | |
6623 | { | |
6624 | if (POINTER_TYPE_P (type) | |
6625 | || TREE_CODE (n1) != INTEGER_CST | |
6626 | || fd->loop.cond_code != LT_EXPR | |
6627 | || tree_int_cst_sgn (n1) != 1) | |
6628 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, | |
6629 | build_one_cst (TREE_TYPE (cond_var))); | |
6630 | else | |
6631 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, | |
6632 | fold_convert (TREE_TYPE (cond_var), step)); | |
6633 | expand_omp_build_assign (&gsi, cond_var, t); | |
6634 | } | |
6635 | ||
6636 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
6637 | gsi_remove (&gsi, true); | |
6638 | } | |
6639 | ||
6640 | /* Emit the condition in L1_BB. */ | |
6641 | gsi = gsi_start_bb (l1_bb); | |
6642 | ||
6643 | if (altv) | |
6644 | t = build2 (LT_EXPR, boolean_type_node, altv, altn2); | |
6645 | else if (fd->collapse > 1 | |
6646 | && !gimple_omp_for_combined_into_p (fd->for_stmt) | |
6647 | && !broken_loop) | |
6648 | { | |
6649 | i = fd->collapse - 1; | |
6650 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6651 | if (fd->loops[i].m2) | |
6652 | t = n2v = create_tmp_var (itype); | |
6653 | else | |
6654 | t = fold_convert (itype, fd->loops[i].n2); | |
6655 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6656 | false, GSI_CONTINUE_LINKING); | |
6657 | tree v = fd->loops[i].v; | |
6658 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
6659 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
6660 | false, GSI_CONTINUE_LINKING); | |
6661 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
6662 | } | |
6663 | else | |
6664 | { | |
6665 | if (fd->collapse > 1 && !broken_loop) | |
6666 | t = n2var; | |
6667 | else | |
6668 | t = fold_convert (type, n2); | |
6669 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6670 | false, GSI_CONTINUE_LINKING); | |
6671 | tree v = fd->loop.v; | |
6672 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
6673 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
6674 | false, GSI_CONTINUE_LINKING); | |
6675 | t = build2 (fd->loop.cond_code, boolean_type_node, v, t); | |
6676 | } | |
6677 | cond_stmt = gimple_build_cond_empty (t); | |
6678 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
6679 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
6680 | NULL, NULL) | |
6681 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
6682 | NULL, NULL)) | |
6683 | { | |
6684 | gsi = gsi_for_stmt (cond_stmt); | |
6685 | gimple_regimplify_operands (cond_stmt, &gsi); | |
6686 | } | |
6687 | ||
6688 | /* Add 'V -= STEP * (SIMT_VF - 1)' after the loop. */ | |
6689 | if (is_simt) | |
6690 | { | |
6691 | gsi = gsi_start_bb (l2_bb); | |
6692 | step = fold_build2 (MINUS_EXPR, TREE_TYPE (step), orig_step, step); | |
6693 | if (POINTER_TYPE_P (type)) | |
6694 | t = fold_build_pointer_plus (fd->loop.v, step); | |
6695 | else | |
6696 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); | |
6697 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
6698 | } | |
6699 | ||
6700 | /* Remove GIMPLE_OMP_RETURN. */ | |
6701 | gsi = gsi_last_nondebug_bb (exit_bb); | |
6702 | gsi_remove (&gsi, true); | |
6703 | ||
6704 | /* Connect the new blocks. */ | |
6705 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
6706 | ||
6707 | if (!broken_loop) | |
6708 | { | |
6709 | remove_edge (BRANCH_EDGE (entry_bb)); | |
6710 | make_edge (entry_bb, l1_bb, EDGE_FALLTHRU); | |
6711 | ||
6712 | e = BRANCH_EDGE (l1_bb); | |
6713 | ne = FALLTHRU_EDGE (l1_bb); | |
6714 | e->flags = EDGE_TRUE_VALUE; | |
6715 | } | |
6716 | else | |
6717 | { | |
6718 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
6719 | ||
6720 | ne = single_succ_edge (l1_bb); | |
6721 | e = make_edge (l1_bb, l0_bb, EDGE_TRUE_VALUE); | |
6722 | ||
6723 | } | |
6724 | ne->flags = EDGE_FALSE_VALUE; | |
6725 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); | |
6726 | ne->probability = e->probability.invert (); | |
6727 | ||
6728 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, entry_bb); | |
6729 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, l1_bb); | |
6730 | ||
6731 | if (simt_maxlane) | |
6732 | { | |
6733 | cond_stmt = gimple_build_cond (LT_EXPR, simt_lane, simt_maxlane, | |
6734 | NULL_TREE, NULL_TREE); | |
6735 | gsi = gsi_last_bb (entry_bb); | |
6736 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
6737 | make_edge (entry_bb, l2_bb, EDGE_FALSE_VALUE); | |
6738 | FALLTHRU_EDGE (entry_bb)->flags = EDGE_TRUE_VALUE; | |
6739 | FALLTHRU_EDGE (entry_bb)->probability | |
6740 | = profile_probability::guessed_always ().apply_scale (7, 8); | |
6741 | BRANCH_EDGE (entry_bb)->probability | |
6742 | = FALLTHRU_EDGE (entry_bb)->probability.invert (); | |
6743 | l2_dom_bb = entry_bb; | |
6744 | } | |
6745 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, l2_dom_bb); | |
6746 | ||
6747 | if (!broken_loop && fd->collapse > 1) | |
6748 | { | |
6749 | basic_block last_bb = l1_bb; | |
6750 | basic_block init_bb = NULL; | |
6751 | for (i = fd->collapse - 2; i >= 0; i--) | |
6752 | { | |
6753 | tree nextn2v = NULL_TREE; | |
6754 | if (EDGE_SUCC (last_bb, 0)->flags & EDGE_FALSE_VALUE) | |
6755 | e = EDGE_SUCC (last_bb, 0); | |
6756 | else | |
6757 | e = EDGE_SUCC (last_bb, 1); | |
6758 | basic_block bb = split_edge (e); | |
6759 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
6760 | { | |
6761 | t = fold_convert (sizetype, fd->loops[i].step); | |
6762 | t = fold_build_pointer_plus (fd->loops[i].v, t); | |
6763 | } | |
6764 | else | |
6765 | { | |
6766 | t = fold_convert (TREE_TYPE (fd->loops[i].v), | |
6767 | fd->loops[i].step); | |
6768 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
6769 | fd->loops[i].v, t); | |
6770 | } | |
6771 | gsi = gsi_after_labels (bb); | |
6772 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6773 | ||
6774 | bb = split_block (bb, last_stmt (bb))->dest; | |
6775 | gsi = gsi_start_bb (bb); | |
6776 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6777 | if (fd->loops[i].m2) | |
6778 | t = nextn2v = create_tmp_var (itype); | |
6779 | else | |
6780 | t = fold_convert (itype, fd->loops[i].n2); | |
6781 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6782 | false, GSI_CONTINUE_LINKING); | |
6783 | tree v = fd->loops[i].v; | |
6784 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
6785 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
6786 | false, GSI_CONTINUE_LINKING); | |
6787 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
6788 | cond_stmt = gimple_build_cond_empty (t); | |
6789 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
6790 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
6791 | expand_omp_regimplify_p, NULL, NULL) | |
6792 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
6793 | expand_omp_regimplify_p, NULL, NULL)) | |
6794 | { | |
6795 | gsi = gsi_for_stmt (cond_stmt); | |
6796 | gimple_regimplify_operands (cond_stmt, &gsi); | |
6797 | } | |
6798 | ne = single_succ_edge (bb); | |
6799 | ne->flags = EDGE_FALSE_VALUE; | |
6800 | ||
6801 | init_bb = create_empty_bb (bb); | |
6802 | set_immediate_dominator (CDI_DOMINATORS, init_bb, bb); | |
6803 | add_bb_to_loop (init_bb, bb->loop_father); | |
6804 | e = make_edge (bb, init_bb, EDGE_TRUE_VALUE); | |
6805 | e->probability | |
6806 | = profile_probability::guessed_always ().apply_scale (7, 8); | |
6807 | ne->probability = e->probability.invert (); | |
6808 | ||
6809 | gsi = gsi_after_labels (init_bb); | |
6810 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), | |
6811 | fd->loops[i + 1].n1); | |
6812 | if (fd->loops[i + 1].m1) | |
6813 | { | |
6814 | tree t2 = fold_convert (TREE_TYPE (t), | |
6815 | fd->loops[i + 1 | |
6816 | - fd->loops[i + 1].outer].v); | |
6817 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m1); | |
6818 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
6819 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); | |
6820 | } | |
6821 | expand_omp_build_assign (&gsi, fd->loops[i + 1].v, t); | |
6822 | if (fd->loops[i + 1].m2) | |
6823 | { | |
6824 | if (i + 2 == fd->collapse && (n2var || altv)) | |
6825 | { | |
6826 | gcc_assert (n2v == NULL_TREE); | |
6827 | n2v = create_tmp_var (TREE_TYPE (fd->loops[i + 1].v)); | |
6828 | } | |
6829 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), | |
6830 | fd->loops[i + 1].n2); | |
6831 | tree t2 = fold_convert (TREE_TYPE (t), | |
6832 | fd->loops[i + 1 | |
6833 | - fd->loops[i + 1].outer].v); | |
6834 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m2); | |
6835 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
6836 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); | |
6837 | expand_omp_build_assign (&gsi, n2v, t); | |
6838 | } | |
6839 | if (i + 2 == fd->collapse && n2var) | |
6840 | { | |
6841 | /* For composite simd, n2 is the first iteration the current | |
6842 | task shouldn't already handle, so we effectively want to use | |
6843 | for (V3 = N31; V < N2 && V3 < N32; V++, V3 += STEP3) | |
6844 | as the vectorized loop. Except the vectorizer will not | |
6845 | vectorize that, so instead compute N2VAR as | |
6846 | N2VAR = V + MIN (N2 - V, COUNTS3) and use | |
6847 | for (V3 = N31; V < N2VAR; V++, V3 += STEP3) | |
6848 | as the loop to vectorize. */ | |
6849 | tree t2 = fold_build2 (MINUS_EXPR, type, n2, fd->loop.v); | |
6850 | if (fd->loops[i + 1].m1 || fd->loops[i + 1].m2) | |
6851 | { | |
6852 | t = build_int_cst (itype, (fd->loops[i + 1].cond_code | |
6853 | == LT_EXPR ? -1 : 1)); | |
6854 | t = fold_build2 (PLUS_EXPR, itype, | |
6855 | fold_convert (itype, | |
6856 | fd->loops[i + 1].step), t); | |
6857 | if (fd->loops[i + 1].m2) | |
6858 | t = fold_build2 (PLUS_EXPR, itype, t, n2v); | |
6859 | else | |
6860 | t = fold_build2 (PLUS_EXPR, itype, t, | |
6861 | fold_convert (itype, | |
6862 | fd->loops[i + 1].n2)); | |
6863 | t = fold_build2 (MINUS_EXPR, itype, t, | |
6864 | fold_convert (itype, fd->loops[i + 1].v)); | |
6865 | tree step = fold_convert (itype, fd->loops[i + 1].step); | |
6866 | if (TYPE_UNSIGNED (itype) | |
6867 | && fd->loops[i + 1].cond_code == GT_EXPR) | |
6868 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6869 | fold_build1 (NEGATE_EXPR, itype, t), | |
6870 | fold_build1 (NEGATE_EXPR, itype, step)); | |
6871 | else | |
6872 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
6873 | t = fold_convert (type, t); | |
6874 | } | |
6875 | else | |
6876 | t = counts[i + 1]; | |
6877 | expand_omp_build_assign (&gsi, min_arg1, t2); | |
6878 | expand_omp_build_assign (&gsi, min_arg2, t); | |
6879 | e = split_block (init_bb, last_stmt (init_bb)); | |
6880 | gsi = gsi_after_labels (e->dest); | |
6881 | init_bb = e->dest; | |
6882 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
6883 | make_edge (entry_bb, init_bb, EDGE_FALLTHRU); | |
6884 | set_immediate_dominator (CDI_DOMINATORS, init_bb, entry_bb); | |
6885 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, init_bb); | |
6886 | t = fold_build2 (MIN_EXPR, type, min_arg1, min_arg2); | |
6887 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, t); | |
6888 | expand_omp_build_assign (&gsi, n2var, t); | |
6889 | } | |
6890 | if (i + 2 == fd->collapse && altv) | |
6891 | { | |
6892 | /* The vectorizer currently punts on loops with non-constant | |
6893 | steps for the main IV (can't compute number of iterations | |
6894 | and gives up because of that). As for OpenMP loops it is | |
6895 | always possible to compute the number of iterations upfront, | |
6896 | use an alternate IV as the loop iterator. */ | |
6897 | expand_omp_build_assign (&gsi, altv, | |
6898 | build_zero_cst (TREE_TYPE (altv))); | |
6899 | tree itype = TREE_TYPE (fd->loops[i + 1].v); | |
6900 | if (POINTER_TYPE_P (itype)) | |
6901 | itype = signed_type_for (itype); | |
6902 | t = build_int_cst (itype, (fd->loops[i + 1].cond_code == LT_EXPR | |
6903 | ? -1 : 1)); | |
6904 | t = fold_build2 (PLUS_EXPR, itype, | |
6905 | fold_convert (itype, fd->loops[i + 1].step), t); | |
6906 | t = fold_build2 (PLUS_EXPR, itype, t, | |
6907 | fold_convert (itype, | |
6908 | fd->loops[i + 1].m2 | |
6909 | ? n2v : fd->loops[i + 1].n2)); | |
6910 | t = fold_build2 (MINUS_EXPR, itype, t, | |
6911 | fold_convert (itype, fd->loops[i + 1].v)); | |
6912 | tree step = fold_convert (itype, fd->loops[i + 1].step); | |
6913 | if (TYPE_UNSIGNED (itype) | |
6914 | && fd->loops[i + 1].cond_code == GT_EXPR) | |
6915 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6916 | fold_build1 (NEGATE_EXPR, itype, t), | |
6917 | fold_build1 (NEGATE_EXPR, itype, step)); | |
6918 | else | |
6919 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
6920 | t = fold_convert (TREE_TYPE (altv), t); | |
6921 | expand_omp_build_assign (&gsi, altn2, t); | |
6922 | tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v), | |
6923 | fd->loops[i + 1].m2 | |
6924 | ? n2v : fd->loops[i + 1].n2); | |
6925 | t2 = force_gimple_operand_gsi (&gsi, t2, true, NULL_TREE, | |
6926 | true, GSI_SAME_STMT); | |
6927 | t2 = fold_build2 (fd->loops[i + 1].cond_code, boolean_type_node, | |
6928 | fd->loops[i + 1].v, t2); | |
6929 | gassign *g | |
6930 | = gimple_build_assign (altn2, COND_EXPR, t2, altn2, | |
6931 | build_zero_cst (TREE_TYPE (altv))); | |
6932 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
6933 | } | |
6934 | n2v = nextn2v; | |
6935 | ||
6936 | make_edge (init_bb, last_bb, EDGE_FALLTHRU); | |
6937 | if (!gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6938 | { | |
6939 | e = find_edge (entry_bb, last_bb); | |
6940 | redirect_edge_succ (e, bb); | |
6941 | set_immediate_dominator (CDI_DOMINATORS, bb, entry_bb); | |
6942 | set_immediate_dominator (CDI_DOMINATORS, last_bb, init_bb); | |
6943 | } | |
6944 | ||
6945 | last_bb = bb; | |
6946 | } | |
6947 | } | |
6948 | if (!broken_loop) | |
6949 | { | |
6950 | class loop *loop = alloc_loop (); | |
6951 | loop->header = l1_bb; | |
6952 | loop->latch = cont_bb; | |
6953 | add_loop (loop, l1_bb->loop_father); | |
6954 | loop->safelen = safelen_int; | |
6955 | if (simduid) | |
6956 | { | |
6957 | loop->simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); | |
6958 | cfun->has_simduid_loops = true; | |
6959 | } | |
6960 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize | |
6961 | the loop. */ | |
6962 | if ((flag_tree_loop_vectorize | |
6963 | || !global_options_set.x_flag_tree_loop_vectorize) | |
6964 | && flag_tree_loop_optimize | |
6965 | && loop->safelen > 1) | |
6966 | { | |
6967 | loop->force_vectorize = true; | |
6968 | if (simdlen && tree_fits_uhwi_p (OMP_CLAUSE_SIMDLEN_EXPR (simdlen))) | |
6969 | { | |
6970 | unsigned HOST_WIDE_INT v | |
6971 | = tree_to_uhwi (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)); | |
6972 | if (v < INT_MAX && v <= (unsigned HOST_WIDE_INT) loop->safelen) | |
6973 | loop->simdlen = v; | |
6974 | } | |
6975 | cfun->has_force_vectorize_loops = true; | |
6976 | } | |
6977 | else if (dont_vectorize) | |
6978 | loop->dont_vectorize = true; | |
6979 | } | |
6980 | else if (simduid) | |
6981 | cfun->has_simduid_loops = true; | |
6982 | } | |
6983 | ||
6984 | /* Taskloop construct is represented after gimplification with | |
6985 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched | |
6986 | in between them. This routine expands the outer GIMPLE_OMP_FOR, | |
6987 | which should just compute all the needed loop temporaries | |
6988 | for GIMPLE_OMP_TASK. */ | |
6989 | ||
6990 | static void | |
6991 | expand_omp_taskloop_for_outer (struct omp_region *region, | |
6992 | struct omp_for_data *fd, | |
6993 | gimple *inner_stmt) | |
6994 | { | |
6995 | tree type, bias = NULL_TREE; | |
6996 | basic_block entry_bb, cont_bb, exit_bb; | |
6997 | gimple_stmt_iterator gsi; | |
6998 | gassign *assign_stmt; | |
6999 | tree *counts = NULL; | |
7000 | int i; | |
7001 | ||
7002 | gcc_assert (inner_stmt); | |
7003 | gcc_assert (region->cont); | |
7004 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_TASK | |
7005 | && gimple_omp_task_taskloop_p (inner_stmt)); | |
7006 | type = TREE_TYPE (fd->loop.v); | |
7007 | ||
7008 | /* See if we need to bias by LLONG_MIN. */ | |
7009 | if (fd->iter_type == long_long_unsigned_type_node | |
7010 | && TREE_CODE (type) == INTEGER_TYPE | |
7011 | && !TYPE_UNSIGNED (type)) | |
7012 | { | |
7013 | tree n1, n2; | |
7014 | ||
7015 | if (fd->loop.cond_code == LT_EXPR) | |
7016 | { | |
7017 | n1 = fd->loop.n1; | |
7018 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7019 | } | |
7020 | else | |
7021 | { | |
7022 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7023 | n2 = fd->loop.n1; | |
7024 | } | |
7025 | if (TREE_CODE (n1) != INTEGER_CST | |
7026 | || TREE_CODE (n2) != INTEGER_CST | |
7027 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
7028 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
7029 | } | |
7030 | ||
7031 | entry_bb = region->entry; | |
7032 | cont_bb = region->cont; | |
7033 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
7034 | gcc_assert (BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
7035 | exit_bb = region->exit; | |
7036 | ||
7037 | gsi = gsi_last_nondebug_bb (entry_bb); | |
7038 | gimple *for_stmt = gsi_stmt (gsi); | |
7039 | gcc_assert (gimple_code (for_stmt) == GIMPLE_OMP_FOR); | |
7040 | if (fd->collapse > 1) | |
7041 | { | |
7042 | int first_zero_iter = -1, dummy = -1; | |
7043 | basic_block zero_iter_bb = NULL, dummy_bb = NULL, l2_dom_bb = NULL; | |
7044 | ||
7045 | counts = XALLOCAVEC (tree, fd->collapse); | |
7046 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
7047 | zero_iter_bb, first_zero_iter, | |
7048 | dummy_bb, dummy, l2_dom_bb); | |
7049 | ||
7050 | if (zero_iter_bb) | |
7051 | { | |
7052 | /* Some counts[i] vars might be uninitialized if | |
7053 | some loop has zero iterations. But the body shouldn't | |
7054 | be executed in that case, so just avoid uninit warnings. */ | |
7055 | for (i = first_zero_iter; i < fd->collapse; i++) | |
7056 | if (SSA_VAR_P (counts[i])) | |
7057 | suppress_warning (counts[i], OPT_Wuninitialized); | |
7058 | gsi_prev (&gsi); | |
7059 | edge e = split_block (entry_bb, gsi_stmt (gsi)); | |
7060 | entry_bb = e->dest; | |
7061 | make_edge (zero_iter_bb, entry_bb, EDGE_FALLTHRU); | |
7062 | gsi = gsi_last_bb (entry_bb); | |
7063 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, | |
7064 | get_immediate_dominator (CDI_DOMINATORS, | |
7065 | zero_iter_bb)); | |
7066 | } | |
7067 | } | |
7068 | ||
7069 | tree t0, t1; | |
7070 | t1 = fd->loop.n2; | |
7071 | t0 = fd->loop.n1; | |
7072 | if (POINTER_TYPE_P (TREE_TYPE (t0)) | |
7073 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
7074 | != TYPE_PRECISION (fd->iter_type)) | |
7075 | { | |
7076 | /* Avoid casting pointers to integer of a different size. */ | |
7077 | tree itype = signed_type_for (type); | |
7078 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); | |
7079 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
7080 | } | |
7081 | else | |
7082 | { | |
7083 | t1 = fold_convert (fd->iter_type, t1); | |
7084 | t0 = fold_convert (fd->iter_type, t0); | |
7085 | } | |
7086 | if (bias) | |
7087 | { | |
7088 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); | |
7089 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
7090 | } | |
7091 | ||
7092 | tree innerc = omp_find_clause (gimple_omp_task_clauses (inner_stmt), | |
7093 | OMP_CLAUSE__LOOPTEMP_); | |
7094 | gcc_assert (innerc); | |
7095 | tree startvar = OMP_CLAUSE_DECL (innerc); | |
7096 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), OMP_CLAUSE__LOOPTEMP_); | |
7097 | gcc_assert (innerc); | |
7098 | tree endvar = OMP_CLAUSE_DECL (innerc); | |
7099 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
7100 | { | |
7101 | innerc = find_lastprivate_looptemp (fd, innerc); | |
7102 | if (innerc) | |
7103 | { | |
7104 | /* If needed (inner taskloop has lastprivate clause), propagate | |
7105 | down the total number of iterations. */ | |
7106 | tree t = force_gimple_operand_gsi (&gsi, fd->loop.n2, false, | |
7107 | NULL_TREE, false, | |
7108 | GSI_CONTINUE_LINKING); | |
7109 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
7110 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7111 | } | |
7112 | } | |
7113 | ||
7114 | t0 = force_gimple_operand_gsi (&gsi, t0, false, NULL_TREE, false, | |
7115 | GSI_CONTINUE_LINKING); | |
7116 | assign_stmt = gimple_build_assign (startvar, t0); | |
7117 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7118 | ||
7119 | t1 = force_gimple_operand_gsi (&gsi, t1, false, NULL_TREE, false, | |
7120 | GSI_CONTINUE_LINKING); | |
7121 | assign_stmt = gimple_build_assign (endvar, t1); | |
7122 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7123 | if (fd->collapse > 1) | |
7124 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, inner_stmt, startvar); | |
7125 | ||
7126 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
7127 | gsi = gsi_for_stmt (for_stmt); | |
7128 | gsi_remove (&gsi, true); | |
7129 | ||
7130 | gsi = gsi_last_nondebug_bb (cont_bb); | |
7131 | gsi_remove (&gsi, true); | |
7132 | ||
7133 | gsi = gsi_last_nondebug_bb (exit_bb); | |
7134 | gsi_remove (&gsi, true); | |
7135 | ||
7136 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); | |
7137 | remove_edge (BRANCH_EDGE (entry_bb)); | |
7138 | FALLTHRU_EDGE (cont_bb)->probability = profile_probability::always (); | |
7139 | remove_edge (BRANCH_EDGE (cont_bb)); | |
7140 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, cont_bb); | |
7141 | set_immediate_dominator (CDI_DOMINATORS, region->entry, | |
7142 | recompute_dominator (CDI_DOMINATORS, region->entry)); | |
7143 | } | |
7144 | ||
7145 | /* Taskloop construct is represented after gimplification with | |
7146 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched | |
7147 | in between them. This routine expands the inner GIMPLE_OMP_FOR. | |
7148 | GOMP_taskloop{,_ull} function arranges for each task to be given just | |
7149 | a single range of iterations. */ | |
7150 | ||
7151 | static void | |
7152 | expand_omp_taskloop_for_inner (struct omp_region *region, | |
7153 | struct omp_for_data *fd, | |
7154 | gimple *inner_stmt) | |
7155 | { | |
7156 | tree e, t, type, itype, vmain, vback, bias = NULL_TREE; | |
7157 | basic_block entry_bb, exit_bb, body_bb, cont_bb, collapse_bb = NULL; | |
7158 | basic_block fin_bb; | |
7159 | gimple_stmt_iterator gsi; | |
7160 | edge ep; | |
7161 | bool broken_loop = region->cont == NULL; | |
7162 | tree *counts = NULL; | |
7163 | tree n1, n2, step; | |
7164 | ||
7165 | itype = type = TREE_TYPE (fd->loop.v); | |
7166 | if (POINTER_TYPE_P (type)) | |
7167 | itype = signed_type_for (type); | |
7168 | ||
7169 | /* See if we need to bias by LLONG_MIN. */ | |
7170 | if (fd->iter_type == long_long_unsigned_type_node | |
7171 | && TREE_CODE (type) == INTEGER_TYPE | |
7172 | && !TYPE_UNSIGNED (type)) | |
7173 | { | |
7174 | tree n1, n2; | |
7175 | ||
7176 | if (fd->loop.cond_code == LT_EXPR) | |
7177 | { | |
7178 | n1 = fd->loop.n1; | |
7179 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7180 | } | |
7181 | else | |
7182 | { | |
7183 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7184 | n2 = fd->loop.n1; | |
7185 | } | |
7186 | if (TREE_CODE (n1) != INTEGER_CST | |
7187 | || TREE_CODE (n2) != INTEGER_CST | |
7188 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
7189 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
7190 | } | |
7191 | ||
7192 | entry_bb = region->entry; | |
7193 | cont_bb = region->cont; | |
7194 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
7195 | fin_bb = BRANCH_EDGE (entry_bb)->dest; | |
7196 | gcc_assert (broken_loop | |
7197 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
7198 | body_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
7199 | if (!broken_loop) | |
7200 | { | |
7201 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); | |
7202 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
7203 | } | |
7204 | exit_bb = region->exit; | |
7205 | ||
7206 | /* Iteration space partitioning goes in ENTRY_BB. */ | |
7207 | gsi = gsi_last_nondebug_bb (entry_bb); | |
7208 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
7209 | ||
7210 | if (fd->collapse > 1) | |
7211 | { | |
7212 | int first_zero_iter = -1, dummy = -1; | |
7213 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
7214 | ||
7215 | counts = XALLOCAVEC (tree, fd->collapse); | |
7216 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
7217 | fin_bb, first_zero_iter, | |
7218 | dummy_bb, dummy, l2_dom_bb); | |
7219 | t = NULL_TREE; | |
7220 | } | |
7221 | else | |
7222 | t = integer_one_node; | |
7223 | ||
7224 | step = fd->loop.step; | |
7225 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
7226 | OMP_CLAUSE__LOOPTEMP_); | |
7227 | gcc_assert (innerc); | |
7228 | n1 = OMP_CLAUSE_DECL (innerc); | |
7229 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), OMP_CLAUSE__LOOPTEMP_); | |
7230 | gcc_assert (innerc); | |
7231 | n2 = OMP_CLAUSE_DECL (innerc); | |
7232 | if (bias) | |
7233 | { | |
7234 | n1 = fold_build2 (PLUS_EXPR, fd->iter_type, n1, bias); | |
7235 | n2 = fold_build2 (PLUS_EXPR, fd->iter_type, n2, bias); | |
7236 | } | |
7237 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
7238 | true, NULL_TREE, true, GSI_SAME_STMT); | |
7239 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
7240 | true, NULL_TREE, true, GSI_SAME_STMT); | |
7241 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
7242 | true, NULL_TREE, true, GSI_SAME_STMT); | |
7243 | ||
7244 | tree startvar = fd->loop.v; | |
7245 | tree endvar = NULL_TREE; | |
7246 | ||
7247 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
7248 | { | |
7249 | tree clauses = gimple_omp_for_clauses (inner_stmt); | |
7250 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
7251 | gcc_assert (innerc); | |
7252 | startvar = OMP_CLAUSE_DECL (innerc); | |
7253 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
7254 | OMP_CLAUSE__LOOPTEMP_); | |
7255 | gcc_assert (innerc); | |
7256 | endvar = OMP_CLAUSE_DECL (innerc); | |
7257 | } | |
7258 | t = fold_convert (TREE_TYPE (startvar), n1); | |
7259 | t = force_gimple_operand_gsi (&gsi, t, | |
7260 | DECL_P (startvar) | |
7261 | && TREE_ADDRESSABLE (startvar), | |
7262 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
7263 | gimple *assign_stmt = gimple_build_assign (startvar, t); | |
7264 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7265 | ||
7266 | t = fold_convert (TREE_TYPE (startvar), n2); | |
7267 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
7268 | false, GSI_CONTINUE_LINKING); | |
7269 | if (endvar) | |
7270 | { | |
7271 | assign_stmt = gimple_build_assign (endvar, e); | |
7272 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7273 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
7274 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
7275 | else | |
7276 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
7277 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7278 | } | |
7279 | ||
7280 | tree *nonrect_bounds = NULL; | |
7281 | if (fd->collapse > 1) | |
7282 | { | |
7283 | if (fd->non_rect) | |
7284 | { | |
7285 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); | |
7286 | memset (nonrect_bounds, 0, sizeof (tree) * (fd->last_nonrect + 1)); | |
7287 | } | |
7288 | gcc_assert (gsi_bb (gsi) == entry_bb); | |
7289 | expand_omp_for_init_vars (fd, &gsi, counts, nonrect_bounds, inner_stmt, | |
7290 | startvar); | |
7291 | entry_bb = gsi_bb (gsi); | |
7292 | } | |
7293 | ||
7294 | if (!broken_loop) | |
7295 | { | |
7296 | /* The code controlling the sequential loop replaces the | |
7297 | GIMPLE_OMP_CONTINUE. */ | |
7298 | gsi = gsi_last_nondebug_bb (cont_bb); | |
7299 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); | |
7300 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
7301 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
7302 | vback = gimple_omp_continue_control_def (cont_stmt); | |
7303 | ||
7304 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
7305 | { | |
7306 | if (POINTER_TYPE_P (type)) | |
7307 | t = fold_build_pointer_plus (vmain, step); | |
7308 | else | |
7309 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
7310 | t = force_gimple_operand_gsi (&gsi, t, | |
7311 | DECL_P (vback) | |
7312 | && TREE_ADDRESSABLE (vback), | |
7313 | NULL_TREE, true, GSI_SAME_STMT); | |
7314 | assign_stmt = gimple_build_assign (vback, t); | |
7315 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
7316 | ||
7317 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
7318 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
7319 | ? t : vback, e); | |
7320 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
7321 | } | |
7322 | ||
7323 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ | |
7324 | gsi_remove (&gsi, true); | |
7325 | ||
7326 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
7327 | collapse_bb = extract_omp_for_update_vars (fd, nonrect_bounds, | |
7328 | cont_bb, body_bb); | |
7329 | } | |
7330 | ||
7331 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
7332 | gsi = gsi_for_stmt (fd->for_stmt); | |
7333 | gsi_remove (&gsi, true); | |
7334 | ||
7335 | /* Remove the GIMPLE_OMP_RETURN statement. */ | |
7336 | gsi = gsi_last_nondebug_bb (exit_bb); | |
7337 | gsi_remove (&gsi, true); | |
7338 | ||
7339 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); | |
7340 | if (!broken_loop) | |
7341 | remove_edge (BRANCH_EDGE (entry_bb)); | |
7342 | else | |
7343 | { | |
7344 | remove_edge_and_dominated_blocks (BRANCH_EDGE (entry_bb)); | |
7345 | region->outer->cont = NULL; | |
7346 | } | |
7347 | ||
7348 | /* Connect all the blocks. */ | |
7349 | if (!broken_loop) | |
7350 | { | |
7351 | ep = find_edge (cont_bb, body_bb); | |
7352 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
7353 | { | |
7354 | remove_edge (ep); | |
7355 | ep = NULL; | |
7356 | } | |
7357 | else if (fd->collapse > 1) | |
7358 | { | |
7359 | remove_edge (ep); | |
7360 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
7361 | } | |
7362 | else | |
7363 | ep->flags = EDGE_TRUE_VALUE; | |
7364 | find_edge (cont_bb, fin_bb)->flags | |
7365 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
7366 | } | |
7367 | ||
7368 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
7369 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
7370 | if (!broken_loop) | |
7371 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
7372 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
7373 | ||
7374 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) | |
7375 | { | |
7376 | class loop *loop = alloc_loop (); | |
7377 | loop->header = body_bb; | |
7378 | if (collapse_bb == NULL) | |
7379 | loop->latch = cont_bb; | |
7380 | add_loop (loop, body_bb->loop_father); | |
7381 | } | |
7382 | } | |
7383 | ||
7384 | /* A subroutine of expand_omp_for. Generate code for an OpenACC | |
7385 | partitioned loop. The lowering here is abstracted, in that the | |
7386 | loop parameters are passed through internal functions, which are | |
7387 | further lowered by oacc_device_lower, once we get to the target | |
7388 | compiler. The loop is of the form: | |
7389 | ||
7390 | for (V = B; V LTGT E; V += S) {BODY} | |
7391 | ||
7392 | where LTGT is < or >. We may have a specified chunking size, CHUNKING | |
7393 | (constant 0 for no chunking) and we will have a GWV partitioning | |
7394 | mask, specifying dimensions over which the loop is to be | |
7395 | partitioned (see note below). We generate code that looks like | |
7396 | (this ignores tiling): | |
7397 | ||
7398 | <entry_bb> [incoming FALL->body, BRANCH->exit] | |
7399 | typedef signedintify (typeof (V)) T; // underlying signed integral type | |
7400 | T range = E - B; | |
7401 | T chunk_no = 0; | |
7402 | T DIR = LTGT == '<' ? +1 : -1; | |
7403 | T chunk_max = GOACC_LOOP_CHUNK (dir, range, S, CHUNK_SIZE, GWV); | |
7404 | T step = GOACC_LOOP_STEP (dir, range, S, CHUNK_SIZE, GWV); | |
7405 | ||
7406 | <head_bb> [created by splitting end of entry_bb] | |
7407 | T offset = GOACC_LOOP_OFFSET (dir, range, S, CHUNK_SIZE, GWV, chunk_no); | |
7408 | T bound = GOACC_LOOP_BOUND (dir, range, S, CHUNK_SIZE, GWV, offset); | |
7409 | if (!(offset LTGT bound)) goto bottom_bb; | |
7410 | ||
7411 | <body_bb> [incoming] | |
7412 | V = B + offset; | |
7413 | {BODY} | |
7414 | ||
7415 | <cont_bb> [incoming, may == body_bb FALL->exit_bb, BRANCH->body_bb] | |
7416 | offset += step; | |
7417 | if (offset LTGT bound) goto body_bb; [*] | |
7418 | ||
7419 | <bottom_bb> [created by splitting start of exit_bb] insert BRANCH->head_bb | |
7420 | chunk_no++; | |
7421 | if (chunk < chunk_max) goto head_bb; | |
7422 | ||
7423 | <exit_bb> [incoming] | |
7424 | V = B + ((range -/+ 1) / S +/- 1) * S [*] | |
7425 | ||
7426 | [*] Needed if V live at end of loop. */ | |
7427 | ||
7428 | static void | |
7429 | expand_oacc_for (struct omp_region *region, struct omp_for_data *fd) | |
7430 | { | |
7431 | bool is_oacc_kernels_parallelized | |
7432 | = (lookup_attribute ("oacc kernels parallelized", | |
7433 | DECL_ATTRIBUTES (current_function_decl)) != NULL); | |
7434 | { | |
7435 | bool is_oacc_kernels | |
7436 | = (lookup_attribute ("oacc kernels", | |
7437 | DECL_ATTRIBUTES (current_function_decl)) != NULL); | |
7438 | if (is_oacc_kernels_parallelized) | |
7439 | gcc_checking_assert (is_oacc_kernels); | |
7440 | } | |
7441 | gcc_assert (gimple_in_ssa_p (cfun) == is_oacc_kernels_parallelized); | |
7442 | /* In the following, some of the 'gimple_in_ssa_p (cfun)' conditionals are | |
7443 | for SSA specifics, and some are for 'parloops' OpenACC | |
7444 | 'kernels'-parallelized specifics. */ | |
7445 | ||
7446 | tree v = fd->loop.v; | |
7447 | enum tree_code cond_code = fd->loop.cond_code; | |
7448 | enum tree_code plus_code = PLUS_EXPR; | |
7449 | ||
7450 | tree chunk_size = integer_minus_one_node; | |
7451 | tree gwv = integer_zero_node; | |
7452 | tree iter_type = TREE_TYPE (v); | |
7453 | tree diff_type = iter_type; | |
7454 | tree plus_type = iter_type; | |
7455 | struct oacc_collapse *counts = NULL; | |
7456 | ||
7457 | gcc_checking_assert (gimple_omp_for_kind (fd->for_stmt) | |
7458 | == GF_OMP_FOR_KIND_OACC_LOOP); | |
7459 | gcc_assert (!gimple_omp_for_combined_into_p (fd->for_stmt)); | |
7460 | gcc_assert (cond_code == LT_EXPR || cond_code == GT_EXPR); | |
7461 | ||
7462 | if (POINTER_TYPE_P (iter_type)) | |
7463 | { | |
7464 | plus_code = POINTER_PLUS_EXPR; | |
7465 | plus_type = sizetype; | |
7466 | } | |
7467 | for (int ix = fd->collapse; ix--;) | |
7468 | { | |
7469 | tree diff_type2 = TREE_TYPE (fd->loops[ix].step); | |
7470 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (diff_type2)) | |
7471 | diff_type = diff_type2; | |
7472 | } | |
7473 | if (POINTER_TYPE_P (diff_type) || TYPE_UNSIGNED (diff_type)) | |
7474 | diff_type = signed_type_for (diff_type); | |
7475 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (integer_type_node)) | |
7476 | diff_type = integer_type_node; | |
7477 | ||
7478 | basic_block entry_bb = region->entry; /* BB ending in OMP_FOR */ | |
7479 | basic_block exit_bb = region->exit; /* BB ending in OMP_RETURN */ | |
7480 | basic_block cont_bb = region->cont; /* BB ending in OMP_CONTINUE */ | |
7481 | basic_block bottom_bb = NULL; | |
7482 | ||
7483 | /* entry_bb has two successors; the branch edge is to the exit | |
7484 | block, fallthrough edge to body. */ | |
7485 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2 | |
7486 | && BRANCH_EDGE (entry_bb)->dest == exit_bb); | |
7487 | ||
7488 | /* If cont_bb non-NULL, it has 2 successors. The branch successor is | |
7489 | body_bb, or to a block whose only successor is the body_bb. Its | |
7490 | fallthrough successor is the final block (same as the branch | |
7491 | successor of the entry_bb). */ | |
7492 | if (cont_bb) | |
7493 | { | |
7494 | basic_block body_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
7495 | basic_block bed = BRANCH_EDGE (cont_bb)->dest; | |
7496 | ||
7497 | gcc_assert (FALLTHRU_EDGE (cont_bb)->dest == exit_bb); | |
7498 | gcc_assert (bed == body_bb || single_succ_edge (bed)->dest == body_bb); | |
7499 | } | |
7500 | else | |
7501 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
7502 | ||
7503 | /* The exit block only has entry_bb and cont_bb as predecessors. */ | |
7504 | gcc_assert (EDGE_COUNT (exit_bb->preds) == 1 + (cont_bb != NULL)); | |
7505 | ||
7506 | tree chunk_no; | |
7507 | tree chunk_max = NULL_TREE; | |
7508 | tree bound, offset; | |
7509 | tree step = create_tmp_var (diff_type, ".step"); | |
7510 | bool up = cond_code == LT_EXPR; | |
7511 | tree dir = build_int_cst (diff_type, up ? +1 : -1); | |
7512 | bool chunking = !gimple_in_ssa_p (cfun); | |
7513 | bool negating; | |
7514 | ||
7515 | /* Tiling vars. */ | |
7516 | tree tile_size = NULL_TREE; | |
7517 | tree element_s = NULL_TREE; | |
7518 | tree e_bound = NULL_TREE, e_offset = NULL_TREE, e_step = NULL_TREE; | |
7519 | basic_block elem_body_bb = NULL; | |
7520 | basic_block elem_cont_bb = NULL; | |
7521 | ||
7522 | /* SSA instances. */ | |
7523 | tree offset_incr = NULL_TREE; | |
7524 | tree offset_init = NULL_TREE; | |
7525 | ||
7526 | gimple_stmt_iterator gsi; | |
7527 | gassign *ass; | |
7528 | gcall *call; | |
7529 | gimple *stmt; | |
7530 | tree expr; | |
7531 | location_t loc; | |
7532 | edge split, be, fte; | |
7533 | ||
7534 | /* Split the end of entry_bb to create head_bb. */ | |
7535 | split = split_block (entry_bb, last_stmt (entry_bb)); | |
7536 | basic_block head_bb = split->dest; | |
7537 | entry_bb = split->src; | |
7538 | ||
7539 | /* Chunk setup goes at end of entry_bb, replacing the omp_for. */ | |
7540 | gsi = gsi_last_nondebug_bb (entry_bb); | |
7541 | gomp_for *for_stmt = as_a <gomp_for *> (gsi_stmt (gsi)); | |
7542 | loc = gimple_location (for_stmt); | |
7543 | ||
7544 | if (gimple_in_ssa_p (cfun)) | |
7545 | { | |
7546 | offset_init = gimple_omp_for_index (for_stmt, 0); | |
7547 | gcc_assert (integer_zerop (fd->loop.n1)); | |
7548 | /* The SSA parallelizer does gang parallelism. */ | |
7549 | gwv = build_int_cst (integer_type_node, GOMP_DIM_MASK (GOMP_DIM_GANG)); | |
7550 | } | |
7551 | ||
7552 | if (fd->collapse > 1 || fd->tiling) | |
7553 | { | |
7554 | gcc_assert (!gimple_in_ssa_p (cfun) && up); | |
7555 | counts = XALLOCAVEC (struct oacc_collapse, fd->collapse); | |
7556 | tree total = expand_oacc_collapse_init (fd, &gsi, counts, diff_type, | |
7557 | TREE_TYPE (fd->loop.n2), loc); | |
7558 | ||
7559 | if (SSA_VAR_P (fd->loop.n2)) | |
7560 | { | |
7561 | total = force_gimple_operand_gsi (&gsi, total, false, NULL_TREE, | |
7562 | true, GSI_SAME_STMT); | |
7563 | ass = gimple_build_assign (fd->loop.n2, total); | |
7564 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7565 | } | |
7566 | } | |
7567 | ||
7568 | tree b = fd->loop.n1; | |
7569 | tree e = fd->loop.n2; | |
7570 | tree s = fd->loop.step; | |
7571 | ||
7572 | b = force_gimple_operand_gsi (&gsi, b, true, NULL_TREE, true, GSI_SAME_STMT); | |
7573 | e = force_gimple_operand_gsi (&gsi, e, true, NULL_TREE, true, GSI_SAME_STMT); | |
7574 | ||
7575 | /* Convert the step, avoiding possible unsigned->signed overflow. */ | |
7576 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); | |
7577 | if (negating) | |
7578 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); | |
7579 | s = fold_convert (diff_type, s); | |
7580 | if (negating) | |
7581 | s = fold_build1 (NEGATE_EXPR, diff_type, s); | |
7582 | s = force_gimple_operand_gsi (&gsi, s, true, NULL_TREE, true, GSI_SAME_STMT); | |
7583 | ||
7584 | if (!chunking) | |
7585 | chunk_size = integer_zero_node; | |
7586 | expr = fold_convert (diff_type, chunk_size); | |
7587 | chunk_size = force_gimple_operand_gsi (&gsi, expr, true, | |
7588 | NULL_TREE, true, GSI_SAME_STMT); | |
7589 | ||
7590 | if (fd->tiling) | |
7591 | { | |
7592 | /* Determine the tile size and element step, | |
7593 | modify the outer loop step size. */ | |
7594 | tile_size = create_tmp_var (diff_type, ".tile_size"); | |
7595 | expr = build_int_cst (diff_type, 1); | |
7596 | for (int ix = 0; ix < fd->collapse; ix++) | |
7597 | expr = fold_build2 (MULT_EXPR, diff_type, counts[ix].tile, expr); | |
7598 | expr = force_gimple_operand_gsi (&gsi, expr, true, | |
7599 | NULL_TREE, true, GSI_SAME_STMT); | |
7600 | ass = gimple_build_assign (tile_size, expr); | |
7601 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7602 | ||
7603 | element_s = create_tmp_var (diff_type, ".element_s"); | |
7604 | ass = gimple_build_assign (element_s, s); | |
7605 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7606 | ||
7607 | expr = fold_build2 (MULT_EXPR, diff_type, s, tile_size); | |
7608 | s = force_gimple_operand_gsi (&gsi, expr, true, | |
7609 | NULL_TREE, true, GSI_SAME_STMT); | |
7610 | } | |
7611 | ||
7612 | /* Determine the range, avoiding possible unsigned->signed overflow. */ | |
7613 | negating = !up && TYPE_UNSIGNED (iter_type); | |
7614 | expr = fold_build2 (MINUS_EXPR, plus_type, | |
7615 | fold_convert (plus_type, negating ? b : e), | |
7616 | fold_convert (plus_type, negating ? e : b)); | |
7617 | expr = fold_convert (diff_type, expr); | |
7618 | if (negating) | |
7619 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); | |
7620 | tree range = force_gimple_operand_gsi (&gsi, expr, true, | |
7621 | NULL_TREE, true, GSI_SAME_STMT); | |
7622 | ||
7623 | chunk_no = build_int_cst (diff_type, 0); | |
7624 | if (chunking) | |
7625 | { | |
7626 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
7627 | ||
7628 | expr = chunk_no; | |
7629 | chunk_max = create_tmp_var (diff_type, ".chunk_max"); | |
7630 | chunk_no = create_tmp_var (diff_type, ".chunk_no"); | |
7631 | ||
7632 | ass = gimple_build_assign (chunk_no, expr); | |
7633 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7634 | ||
7635 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, | |
7636 | build_int_cst (integer_type_node, | |
7637 | IFN_GOACC_LOOP_CHUNKS), | |
7638 | dir, range, s, chunk_size, gwv); | |
7639 | gimple_call_set_lhs (call, chunk_max); | |
7640 | gimple_set_location (call, loc); | |
7641 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7642 | } | |
7643 | else | |
7644 | chunk_size = chunk_no; | |
7645 | ||
7646 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, | |
7647 | build_int_cst (integer_type_node, | |
7648 | IFN_GOACC_LOOP_STEP), | |
7649 | dir, range, s, chunk_size, gwv); | |
7650 | gimple_call_set_lhs (call, step); | |
7651 | gimple_set_location (call, loc); | |
7652 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7653 | ||
7654 | /* Remove the GIMPLE_OMP_FOR. */ | |
7655 | gsi_remove (&gsi, true); | |
7656 | ||
7657 | /* Fixup edges from head_bb. */ | |
7658 | be = BRANCH_EDGE (head_bb); | |
7659 | fte = FALLTHRU_EDGE (head_bb); | |
7660 | be->flags |= EDGE_FALSE_VALUE; | |
7661 | fte->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
7662 | ||
7663 | basic_block body_bb = fte->dest; | |
7664 | ||
7665 | if (gimple_in_ssa_p (cfun)) | |
7666 | { | |
7667 | gsi = gsi_last_nondebug_bb (cont_bb); | |
7668 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); | |
7669 | ||
7670 | offset = gimple_omp_continue_control_use (cont_stmt); | |
7671 | offset_incr = gimple_omp_continue_control_def (cont_stmt); | |
7672 | } | |
7673 | else | |
7674 | { | |
7675 | offset = create_tmp_var (diff_type, ".offset"); | |
7676 | offset_init = offset_incr = offset; | |
7677 | } | |
7678 | bound = create_tmp_var (TREE_TYPE (offset), ".bound"); | |
7679 | ||
7680 | /* Loop offset & bound go into head_bb. */ | |
7681 | gsi = gsi_start_bb (head_bb); | |
7682 | ||
7683 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, | |
7684 | build_int_cst (integer_type_node, | |
7685 | IFN_GOACC_LOOP_OFFSET), | |
7686 | dir, range, s, | |
7687 | chunk_size, gwv, chunk_no); | |
7688 | gimple_call_set_lhs (call, offset_init); | |
7689 | gimple_set_location (call, loc); | |
7690 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); | |
7691 | ||
7692 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, | |
7693 | build_int_cst (integer_type_node, | |
7694 | IFN_GOACC_LOOP_BOUND), | |
7695 | dir, range, s, | |
7696 | chunk_size, gwv, offset_init); | |
7697 | gimple_call_set_lhs (call, bound); | |
7698 | gimple_set_location (call, loc); | |
7699 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); | |
7700 | ||
7701 | expr = build2 (cond_code, boolean_type_node, offset_init, bound); | |
7702 | gsi_insert_after (&gsi, gimple_build_cond_empty (expr), | |
7703 | GSI_CONTINUE_LINKING); | |
7704 | ||
7705 | /* V assignment goes into body_bb. */ | |
7706 | if (!gimple_in_ssa_p (cfun)) | |
7707 | { | |
7708 | gsi = gsi_start_bb (body_bb); | |
7709 | ||
7710 | expr = build2 (plus_code, iter_type, b, | |
7711 | fold_convert (plus_type, offset)); | |
7712 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7713 | true, GSI_SAME_STMT); | |
7714 | ass = gimple_build_assign (v, expr); | |
7715 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7716 | ||
7717 | if (fd->collapse > 1 || fd->tiling) | |
7718 | expand_oacc_collapse_vars (fd, false, &gsi, counts, v, diff_type); | |
7719 | ||
7720 | if (fd->tiling) | |
7721 | { | |
7722 | /* Determine the range of the element loop -- usually simply | |
7723 | the tile_size, but could be smaller if the final | |
7724 | iteration of the outer loop is a partial tile. */ | |
7725 | tree e_range = create_tmp_var (diff_type, ".e_range"); | |
7726 | ||
7727 | expr = build2 (MIN_EXPR, diff_type, | |
7728 | build2 (MINUS_EXPR, diff_type, bound, offset), | |
7729 | build2 (MULT_EXPR, diff_type, tile_size, | |
7730 | element_s)); | |
7731 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7732 | true, GSI_SAME_STMT); | |
7733 | ass = gimple_build_assign (e_range, expr); | |
7734 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7735 | ||
7736 | /* Determine bound, offset & step of inner loop. */ | |
7737 | e_bound = create_tmp_var (diff_type, ".e_bound"); | |
7738 | e_offset = create_tmp_var (diff_type, ".e_offset"); | |
7739 | e_step = create_tmp_var (diff_type, ".e_step"); | |
7740 | ||
7741 | /* Mark these as element loops. */ | |
7742 | tree t, e_gwv = integer_minus_one_node; | |
7743 | tree chunk = build_int_cst (diff_type, 0); /* Never chunked. */ | |
7744 | ||
7745 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_OFFSET); | |
7746 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, | |
7747 | element_s, chunk, e_gwv, chunk); | |
7748 | gimple_call_set_lhs (call, e_offset); | |
7749 | gimple_set_location (call, loc); | |
7750 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7751 | ||
7752 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_BOUND); | |
7753 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, | |
7754 | element_s, chunk, e_gwv, e_offset); | |
7755 | gimple_call_set_lhs (call, e_bound); | |
7756 | gimple_set_location (call, loc); | |
7757 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7758 | ||
7759 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_STEP); | |
7760 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, t, dir, e_range, | |
7761 | element_s, chunk, e_gwv); | |
7762 | gimple_call_set_lhs (call, e_step); | |
7763 | gimple_set_location (call, loc); | |
7764 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7765 | ||
7766 | /* Add test and split block. */ | |
7767 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); | |
7768 | stmt = gimple_build_cond_empty (expr); | |
7769 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
7770 | split = split_block (body_bb, stmt); | |
7771 | elem_body_bb = split->dest; | |
7772 | if (cont_bb == body_bb) | |
7773 | cont_bb = elem_body_bb; | |
7774 | body_bb = split->src; | |
7775 | ||
7776 | split->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
7777 | ||
7778 | /* Add a dummy exit for the tiled block when cont_bb is missing. */ | |
7779 | if (cont_bb == NULL) | |
7780 | { | |
7781 | edge e = make_edge (body_bb, exit_bb, EDGE_FALSE_VALUE); | |
7782 | e->probability = profile_probability::even (); | |
7783 | split->probability = profile_probability::even (); | |
7784 | } | |
7785 | ||
7786 | /* Initialize the user's loop vars. */ | |
7787 | gsi = gsi_start_bb (elem_body_bb); | |
7788 | expand_oacc_collapse_vars (fd, true, &gsi, counts, e_offset, | |
7789 | diff_type); | |
7790 | } | |
7791 | } | |
7792 | ||
7793 | /* Loop increment goes into cont_bb. If this is not a loop, we | |
7794 | will have spawned threads as if it was, and each one will | |
7795 | execute one iteration. The specification is not explicit about | |
7796 | whether such constructs are ill-formed or not, and they can | |
7797 | occur, especially when noreturn routines are involved. */ | |
7798 | if (cont_bb) | |
7799 | { | |
7800 | gsi = gsi_last_nondebug_bb (cont_bb); | |
7801 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); | |
7802 | loc = gimple_location (cont_stmt); | |
7803 | ||
7804 | if (fd->tiling) | |
7805 | { | |
7806 | /* Insert element loop increment and test. */ | |
7807 | expr = build2 (PLUS_EXPR, diff_type, e_offset, e_step); | |
7808 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7809 | true, GSI_SAME_STMT); | |
7810 | ass = gimple_build_assign (e_offset, expr); | |
7811 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7812 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); | |
7813 | ||
7814 | stmt = gimple_build_cond_empty (expr); | |
7815 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
7816 | split = split_block (cont_bb, stmt); | |
7817 | elem_cont_bb = split->src; | |
7818 | cont_bb = split->dest; | |
7819 | ||
7820 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; | |
7821 | split->probability = profile_probability::unlikely ().guessed (); | |
7822 | edge latch_edge | |
7823 | = make_edge (elem_cont_bb, elem_body_bb, EDGE_TRUE_VALUE); | |
7824 | latch_edge->probability = profile_probability::likely ().guessed (); | |
7825 | ||
7826 | edge skip_edge = make_edge (body_bb, cont_bb, EDGE_FALSE_VALUE); | |
7827 | skip_edge->probability = profile_probability::unlikely ().guessed (); | |
7828 | edge loop_entry_edge = EDGE_SUCC (body_bb, 1 - skip_edge->dest_idx); | |
7829 | loop_entry_edge->probability | |
7830 | = profile_probability::likely ().guessed (); | |
7831 | ||
7832 | gsi = gsi_for_stmt (cont_stmt); | |
7833 | } | |
7834 | ||
7835 | /* Increment offset. */ | |
7836 | if (gimple_in_ssa_p (cfun)) | |
7837 | expr = build2 (plus_code, iter_type, offset, | |
7838 | fold_convert (plus_type, step)); | |
7839 | else | |
7840 | expr = build2 (PLUS_EXPR, diff_type, offset, step); | |
7841 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7842 | true, GSI_SAME_STMT); | |
7843 | ass = gimple_build_assign (offset_incr, expr); | |
7844 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7845 | expr = build2 (cond_code, boolean_type_node, offset_incr, bound); | |
7846 | gsi_insert_before (&gsi, gimple_build_cond_empty (expr), GSI_SAME_STMT); | |
7847 | ||
7848 | /* Remove the GIMPLE_OMP_CONTINUE. */ | |
7849 | gsi_remove (&gsi, true); | |
7850 | ||
7851 | /* Fixup edges from cont_bb. */ | |
7852 | be = BRANCH_EDGE (cont_bb); | |
7853 | fte = FALLTHRU_EDGE (cont_bb); | |
7854 | be->flags |= EDGE_TRUE_VALUE; | |
7855 | fte->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; | |
7856 | ||
7857 | if (chunking) | |
7858 | { | |
7859 | /* Split the beginning of exit_bb to make bottom_bb. We | |
7860 | need to insert a nop at the start, because splitting is | |
7861 | after a stmt, not before. */ | |
7862 | gsi = gsi_start_bb (exit_bb); | |
7863 | stmt = gimple_build_nop (); | |
7864 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
7865 | split = split_block (exit_bb, stmt); | |
7866 | bottom_bb = split->src; | |
7867 | exit_bb = split->dest; | |
7868 | gsi = gsi_last_bb (bottom_bb); | |
7869 | ||
7870 | /* Chunk increment and test goes into bottom_bb. */ | |
7871 | expr = build2 (PLUS_EXPR, diff_type, chunk_no, | |
7872 | build_int_cst (diff_type, 1)); | |
7873 | ass = gimple_build_assign (chunk_no, expr); | |
7874 | gsi_insert_after (&gsi, ass, GSI_CONTINUE_LINKING); | |
7875 | ||
7876 | /* Chunk test at end of bottom_bb. */ | |
7877 | expr = build2 (LT_EXPR, boolean_type_node, chunk_no, chunk_max); | |
7878 | gsi_insert_after (&gsi, gimple_build_cond_empty (expr), | |
7879 | GSI_CONTINUE_LINKING); | |
7880 | ||
7881 | /* Fixup edges from bottom_bb. */ | |
7882 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; | |
7883 | split->probability = profile_probability::unlikely ().guessed (); | |
7884 | edge latch_edge = make_edge (bottom_bb, head_bb, EDGE_TRUE_VALUE); | |
7885 | latch_edge->probability = profile_probability::likely ().guessed (); | |
7886 | } | |
7887 | } | |
7888 | ||
7889 | gsi = gsi_last_nondebug_bb (exit_bb); | |
7890 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
7891 | loc = gimple_location (gsi_stmt (gsi)); | |
7892 | ||
7893 | if (!gimple_in_ssa_p (cfun)) | |
7894 | { | |
7895 | /* Insert the final value of V, in case it is live. This is the | |
7896 | value for the only thread that survives past the join. */ | |
7897 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); | |
7898 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); | |
7899 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); | |
7900 | expr = fold_build2 (MULT_EXPR, diff_type, expr, s); | |
7901 | expr = build2 (plus_code, iter_type, b, fold_convert (plus_type, expr)); | |
7902 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7903 | true, GSI_SAME_STMT); | |
7904 | ass = gimple_build_assign (v, expr); | |
7905 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7906 | } | |
7907 | ||
7908 | /* Remove the OMP_RETURN. */ | |
7909 | gsi_remove (&gsi, true); | |
7910 | ||
7911 | if (cont_bb) | |
7912 | { | |
7913 | /* We now have one, two or three nested loops. Update the loop | |
7914 | structures. */ | |
7915 | class loop *parent = entry_bb->loop_father; | |
7916 | class loop *body = body_bb->loop_father; | |
7917 | ||
7918 | if (chunking) | |
7919 | { | |
7920 | class loop *chunk_loop = alloc_loop (); | |
7921 | chunk_loop->header = head_bb; | |
7922 | chunk_loop->latch = bottom_bb; | |
7923 | add_loop (chunk_loop, parent); | |
7924 | parent = chunk_loop; | |
7925 | } | |
7926 | else if (parent != body) | |
7927 | { | |
7928 | gcc_assert (body->header == body_bb); | |
7929 | gcc_assert (body->latch == cont_bb | |
7930 | || single_pred (body->latch) == cont_bb); | |
7931 | parent = NULL; | |
7932 | } | |
7933 | ||
7934 | if (parent) | |
7935 | { | |
7936 | class loop *body_loop = alloc_loop (); | |
7937 | body_loop->header = body_bb; | |
7938 | body_loop->latch = cont_bb; | |
7939 | add_loop (body_loop, parent); | |
7940 | ||
7941 | if (fd->tiling) | |
7942 | { | |
7943 | /* Insert tiling's element loop. */ | |
7944 | class loop *inner_loop = alloc_loop (); | |
7945 | inner_loop->header = elem_body_bb; | |
7946 | inner_loop->latch = elem_cont_bb; | |
7947 | add_loop (inner_loop, body_loop); | |
7948 | } | |
7949 | } | |
7950 | } | |
7951 | } | |
7952 | ||
7953 | /* Expand the OMP loop defined by REGION. */ | |
7954 | ||
7955 | static void | |
7956 | expand_omp_for (struct omp_region *region, gimple *inner_stmt) | |
7957 | { | |
7958 | struct omp_for_data fd; | |
7959 | struct omp_for_data_loop *loops; | |
7960 | ||
7961 | loops = XALLOCAVEC (struct omp_for_data_loop, | |
7962 | gimple_omp_for_collapse (last_stmt (region->entry))); | |
7963 | omp_extract_for_data (as_a <gomp_for *> (last_stmt (region->entry)), | |
7964 | &fd, loops); | |
7965 | region->sched_kind = fd.sched_kind; | |
7966 | region->sched_modifiers = fd.sched_modifiers; | |
7967 | region->has_lastprivate_conditional = fd.lastprivate_conditional != 0; | |
7968 | if (fd.non_rect && !gimple_omp_for_combined_into_p (fd.for_stmt)) | |
7969 | { | |
7970 | for (int i = fd.first_nonrect; i <= fd.last_nonrect; i++) | |
7971 | if ((loops[i].m1 || loops[i].m2) | |
7972 | && (loops[i].m1 == NULL_TREE | |
7973 | || TREE_CODE (loops[i].m1) == INTEGER_CST) | |
7974 | && (loops[i].m2 == NULL_TREE | |
7975 | || TREE_CODE (loops[i].m2) == INTEGER_CST) | |
7976 | && TREE_CODE (loops[i].step) == INTEGER_CST | |
7977 | && TREE_CODE (loops[i - loops[i].outer].step) == INTEGER_CST) | |
7978 | { | |
7979 | tree t; | |
7980 | tree itype = TREE_TYPE (loops[i].v); | |
7981 | if (loops[i].m1 && loops[i].m2) | |
7982 | t = fold_build2 (MINUS_EXPR, itype, loops[i].m2, loops[i].m1); | |
7983 | else if (loops[i].m1) | |
7984 | t = fold_build1 (NEGATE_EXPR, itype, loops[i].m1); | |
7985 | else | |
7986 | t = loops[i].m2; | |
7987 | t = fold_build2 (MULT_EXPR, itype, t, | |
7988 | fold_convert (itype, | |
7989 | loops[i - loops[i].outer].step)); | |
7990 | if (TYPE_UNSIGNED (itype) && loops[i].cond_code == GT_EXPR) | |
7991 | t = fold_build2 (TRUNC_MOD_EXPR, itype, | |
7992 | fold_build1 (NEGATE_EXPR, itype, t), | |
7993 | fold_build1 (NEGATE_EXPR, itype, | |
7994 | fold_convert (itype, | |
7995 | loops[i].step))); | |
7996 | else | |
7997 | t = fold_build2 (TRUNC_MOD_EXPR, itype, t, | |
7998 | fold_convert (itype, loops[i].step)); | |
7999 | if (integer_nonzerop (t)) | |
8000 | error_at (gimple_location (fd.for_stmt), | |
8001 | "invalid OpenMP non-rectangular loop step; " | |
8002 | "%<(%E - %E) * %E%> is not a multiple of loop %d " | |
8003 | "step %qE", | |
8004 | loops[i].m2 ? loops[i].m2 : integer_zero_node, | |
8005 | loops[i].m1 ? loops[i].m1 : integer_zero_node, | |
8006 | loops[i - loops[i].outer].step, i + 1, | |
8007 | loops[i].step); | |
8008 | } | |
8009 | } | |
8010 | ||
8011 | gcc_assert (EDGE_COUNT (region->entry->succs) == 2); | |
8012 | BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
8013 | FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
8014 | if (region->cont) | |
8015 | { | |
8016 | gcc_assert (EDGE_COUNT (region->cont->succs) == 2); | |
8017 | BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
8018 | FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
8019 | } | |
8020 | else | |
8021 | /* If there isn't a continue then this is a degerate case where | |
8022 | the introduction of abnormal edges during lowering will prevent | |
8023 | original loops from being detected. Fix that up. */ | |
8024 | loops_state_set (LOOPS_NEED_FIXUP); | |
8025 | ||
8026 | if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_SIMD) | |
8027 | expand_omp_simd (region, &fd); | |
8028 | else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_OACC_LOOP) | |
8029 | { | |
8030 | gcc_assert (!inner_stmt && !fd.non_rect); | |
8031 | expand_oacc_for (region, &fd); | |
8032 | } | |
8033 | else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_TASKLOOP) | |
8034 | { | |
8035 | if (gimple_omp_for_combined_into_p (fd.for_stmt)) | |
8036 | expand_omp_taskloop_for_inner (region, &fd, inner_stmt); | |
8037 | else | |
8038 | expand_omp_taskloop_for_outer (region, &fd, inner_stmt); | |
8039 | } | |
8040 | else if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC | |
8041 | && !fd.have_ordered) | |
8042 | { | |
8043 | if (fd.chunk_size == NULL) | |
8044 | expand_omp_for_static_nochunk (region, &fd, inner_stmt); | |
8045 | else | |
8046 | expand_omp_for_static_chunk (region, &fd, inner_stmt); | |
8047 | } | |
8048 | else | |
8049 | { | |
8050 | int fn_index, start_ix, next_ix; | |
8051 | unsigned HOST_WIDE_INT sched = 0; | |
8052 | tree sched_arg = NULL_TREE; | |
8053 | ||
8054 | gcc_assert (gimple_omp_for_kind (fd.for_stmt) | |
8055 | == GF_OMP_FOR_KIND_FOR && !fd.non_rect); | |
8056 | if (fd.chunk_size == NULL | |
8057 | && fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) | |
8058 | fd.chunk_size = integer_zero_node; | |
8059 | switch (fd.sched_kind) | |
8060 | { | |
8061 | case OMP_CLAUSE_SCHEDULE_RUNTIME: | |
8062 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0 | |
8063 | && fd.lastprivate_conditional == 0) | |
8064 | { | |
8065 | gcc_assert (!fd.have_ordered); | |
8066 | fn_index = 6; | |
8067 | sched = 4; | |
8068 | } | |
8069 | else if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 | |
8070 | && !fd.have_ordered | |
8071 | && fd.lastprivate_conditional == 0) | |
8072 | fn_index = 7; | |
8073 | else | |
8074 | { | |
8075 | fn_index = 3; | |
8076 | sched = (HOST_WIDE_INT_1U << 31); | |
8077 | } | |
8078 | break; | |
8079 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: | |
8080 | case OMP_CLAUSE_SCHEDULE_GUIDED: | |
8081 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 | |
8082 | && !fd.have_ordered | |
8083 | && fd.lastprivate_conditional == 0) | |
8084 | { | |
8085 | fn_index = 3 + fd.sched_kind; | |
8086 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; | |
8087 | break; | |
8088 | } | |
8089 | fn_index = fd.sched_kind; | |
8090 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; | |
8091 | sched += (HOST_WIDE_INT_1U << 31); | |
8092 | break; | |
8093 | case OMP_CLAUSE_SCHEDULE_STATIC: | |
8094 | gcc_assert (fd.have_ordered); | |
8095 | fn_index = 0; | |
8096 | sched = (HOST_WIDE_INT_1U << 31) + 1; | |
8097 | break; | |
8098 | default: | |
8099 | gcc_unreachable (); | |
8100 | } | |
8101 | if (!fd.ordered) | |
8102 | fn_index += fd.have_ordered * 8; | |
8103 | if (fd.ordered) | |
8104 | start_ix = ((int)BUILT_IN_GOMP_LOOP_DOACROSS_STATIC_START) + fn_index; | |
8105 | else | |
8106 | start_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_START) + fn_index; | |
8107 | next_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_NEXT) + fn_index; | |
8108 | if (fd.have_reductemp || fd.have_pointer_condtemp) | |
8109 | { | |
8110 | if (fd.ordered) | |
8111 | start_ix = (int)BUILT_IN_GOMP_LOOP_DOACROSS_START; | |
8112 | else if (fd.have_ordered) | |
8113 | start_ix = (int)BUILT_IN_GOMP_LOOP_ORDERED_START; | |
8114 | else | |
8115 | start_ix = (int)BUILT_IN_GOMP_LOOP_START; | |
8116 | sched_arg = build_int_cstu (long_integer_type_node, sched); | |
8117 | if (!fd.chunk_size) | |
8118 | fd.chunk_size = integer_zero_node; | |
8119 | } | |
8120 | if (fd.iter_type == long_long_unsigned_type_node) | |
8121 | { | |
8122 | start_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_START | |
8123 | - (int)BUILT_IN_GOMP_LOOP_STATIC_START); | |
8124 | next_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_NEXT | |
8125 | - (int)BUILT_IN_GOMP_LOOP_STATIC_NEXT); | |
8126 | } | |
8127 | expand_omp_for_generic (region, &fd, (enum built_in_function) start_ix, | |
8128 | (enum built_in_function) next_ix, sched_arg, | |
8129 | inner_stmt); | |
8130 | } | |
8131 | ||
8132 | if (gimple_in_ssa_p (cfun)) | |
8133 | update_ssa (TODO_update_ssa_only_virtuals); | |
8134 | } | |
8135 | ||
8136 | /* Expand code for an OpenMP sections directive. In pseudo code, we generate | |
8137 | ||
8138 | v = GOMP_sections_start (n); | |
8139 | L0: | |
8140 | switch (v) | |
8141 | { | |
8142 | case 0: | |
8143 | goto L2; | |
8144 | case 1: | |
8145 | section 1; | |
8146 | goto L1; | |
8147 | case 2: | |
8148 | ... | |
8149 | case n: | |
8150 | ... | |
8151 | default: | |
8152 | abort (); | |
8153 | } | |
8154 | L1: | |
8155 | v = GOMP_sections_next (); | |
8156 | goto L0; | |
8157 | L2: | |
8158 | reduction; | |
8159 | ||
8160 | If this is a combined parallel sections, replace the call to | |
8161 | GOMP_sections_start with call to GOMP_sections_next. */ | |
8162 | ||
8163 | static void | |
8164 | expand_omp_sections (struct omp_region *region) | |
8165 | { | |
8166 | tree t, u, vin = NULL, vmain, vnext, l2; | |
8167 | unsigned len; | |
8168 | basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb; | |
8169 | gimple_stmt_iterator si, switch_si; | |
8170 | gomp_sections *sections_stmt; | |
8171 | gimple *stmt; | |
8172 | gomp_continue *cont; | |
8173 | edge_iterator ei; | |
8174 | edge e; | |
8175 | struct omp_region *inner; | |
8176 | unsigned i, casei; | |
8177 | bool exit_reachable = region->cont != NULL; | |
8178 | ||
8179 | gcc_assert (region->exit != NULL); | |
8180 | entry_bb = region->entry; | |
8181 | l0_bb = single_succ (entry_bb); | |
8182 | l1_bb = region->cont; | |
8183 | l2_bb = region->exit; | |
8184 | if (single_pred_p (l2_bb) && single_pred (l2_bb) == l0_bb) | |
8185 | l2 = gimple_block_label (l2_bb); | |
8186 | else | |
8187 | { | |
8188 | /* This can happen if there are reductions. */ | |
8189 | len = EDGE_COUNT (l0_bb->succs); | |
8190 | gcc_assert (len > 0); | |
8191 | e = EDGE_SUCC (l0_bb, len - 1); | |
8192 | si = gsi_last_nondebug_bb (e->dest); | |
8193 | l2 = NULL_TREE; | |
8194 | if (gsi_end_p (si) | |
8195 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
8196 | l2 = gimple_block_label (e->dest); | |
8197 | else | |
8198 | FOR_EACH_EDGE (e, ei, l0_bb->succs) | |
8199 | { | |
8200 | si = gsi_last_nondebug_bb (e->dest); | |
8201 | if (gsi_end_p (si) | |
8202 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
8203 | { | |
8204 | l2 = gimple_block_label (e->dest); | |
8205 | break; | |
8206 | } | |
8207 | } | |
8208 | } | |
8209 | if (exit_reachable) | |
8210 | default_bb = create_empty_bb (l1_bb->prev_bb); | |
8211 | else | |
8212 | default_bb = create_empty_bb (l0_bb); | |
8213 | ||
8214 | /* We will build a switch() with enough cases for all the | |
8215 | GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work | |
8216 | and a default case to abort if something goes wrong. */ | |
8217 | len = EDGE_COUNT (l0_bb->succs); | |
8218 | ||
8219 | /* Use vec::quick_push on label_vec throughout, since we know the size | |
8220 | in advance. */ | |
8221 | auto_vec<tree> label_vec (len); | |
8222 | ||
8223 | /* The call to GOMP_sections_start goes in ENTRY_BB, replacing the | |
8224 | GIMPLE_OMP_SECTIONS statement. */ | |
8225 | si = gsi_last_nondebug_bb (entry_bb); | |
8226 | sections_stmt = as_a <gomp_sections *> (gsi_stmt (si)); | |
8227 | gcc_assert (gimple_code (sections_stmt) == GIMPLE_OMP_SECTIONS); | |
8228 | vin = gimple_omp_sections_control (sections_stmt); | |
8229 | tree clauses = gimple_omp_sections_clauses (sections_stmt); | |
8230 | tree reductmp = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
8231 | tree condtmp = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); | |
8232 | tree cond_var = NULL_TREE; | |
8233 | if (reductmp || condtmp) | |
8234 | { | |
8235 | tree reductions = null_pointer_node, mem = null_pointer_node; | |
8236 | tree memv = NULL_TREE, condtemp = NULL_TREE; | |
8237 | gimple_stmt_iterator gsi = gsi_none (); | |
8238 | gimple *g = NULL; | |
8239 | if (reductmp) | |
8240 | { | |
8241 | reductions = OMP_CLAUSE_DECL (reductmp); | |
8242 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
8243 | g = SSA_NAME_DEF_STMT (reductions); | |
8244 | reductions = gimple_assign_rhs1 (g); | |
8245 | OMP_CLAUSE_DECL (reductmp) = reductions; | |
8246 | gsi = gsi_for_stmt (g); | |
8247 | } | |
8248 | else | |
8249 | gsi = si; | |
8250 | if (condtmp) | |
8251 | { | |
8252 | condtemp = OMP_CLAUSE_DECL (condtmp); | |
8253 | tree c = omp_find_clause (OMP_CLAUSE_CHAIN (condtmp), | |
8254 | OMP_CLAUSE__CONDTEMP_); | |
8255 | cond_var = OMP_CLAUSE_DECL (c); | |
8256 | tree type = TREE_TYPE (condtemp); | |
8257 | memv = create_tmp_var (type); | |
8258 | TREE_ADDRESSABLE (memv) = 1; | |
8259 | unsigned cnt = 0; | |
8260 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
8261 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE | |
8262 | && OMP_CLAUSE_LASTPRIVATE_CONDITIONAL (c)) | |
8263 | ++cnt; | |
8264 | unsigned HOST_WIDE_INT sz | |
8265 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))) * cnt; | |
8266 | expand_omp_build_assign (&gsi, memv, build_int_cst (type, sz), | |
8267 | false); | |
8268 | mem = build_fold_addr_expr (memv); | |
8269 | } | |
8270 | t = build_int_cst (unsigned_type_node, len - 1); | |
8271 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS2_START); | |
8272 | stmt = gimple_build_call (u, 3, t, reductions, mem); | |
8273 | gimple_call_set_lhs (stmt, vin); | |
8274 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
8275 | if (condtmp) | |
8276 | { | |
8277 | expand_omp_build_assign (&gsi, condtemp, memv, false); | |
8278 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), | |
8279 | vin, build_one_cst (TREE_TYPE (cond_var))); | |
8280 | expand_omp_build_assign (&gsi, cond_var, t, false); | |
8281 | } | |
8282 | if (reductmp) | |
8283 | { | |
8284 | gsi_remove (&gsi, true); | |
8285 | release_ssa_name (gimple_assign_lhs (g)); | |
8286 | } | |
8287 | } | |
8288 | else if (!is_combined_parallel (region)) | |
8289 | { | |
8290 | /* If we are not inside a combined parallel+sections region, | |
8291 | call GOMP_sections_start. */ | |
8292 | t = build_int_cst (unsigned_type_node, len - 1); | |
8293 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_START); | |
8294 | stmt = gimple_build_call (u, 1, t); | |
8295 | } | |
8296 | else | |
8297 | { | |
8298 | /* Otherwise, call GOMP_sections_next. */ | |
8299 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); | |
8300 | stmt = gimple_build_call (u, 0); | |
8301 | } | |
8302 | if (!reductmp && !condtmp) | |
8303 | { | |
8304 | gimple_call_set_lhs (stmt, vin); | |
8305 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
8306 | } | |
8307 | gsi_remove (&si, true); | |
8308 | ||
8309 | /* The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in | |
8310 | L0_BB. */ | |
8311 | switch_si = gsi_last_nondebug_bb (l0_bb); | |
8312 | gcc_assert (gimple_code (gsi_stmt (switch_si)) == GIMPLE_OMP_SECTIONS_SWITCH); | |
8313 | if (exit_reachable) | |
8314 | { | |
8315 | cont = as_a <gomp_continue *> (last_stmt (l1_bb)); | |
8316 | gcc_assert (gimple_code (cont) == GIMPLE_OMP_CONTINUE); | |
8317 | vmain = gimple_omp_continue_control_use (cont); | |
8318 | vnext = gimple_omp_continue_control_def (cont); | |
8319 | } | |
8320 | else | |
8321 | { | |
8322 | vmain = vin; | |
8323 | vnext = NULL_TREE; | |
8324 | } | |
8325 | ||
8326 | t = build_case_label (build_int_cst (unsigned_type_node, 0), NULL, l2); | |
8327 | label_vec.quick_push (t); | |
8328 | i = 1; | |
8329 | ||
8330 | /* Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR. */ | |
8331 | for (inner = region->inner, casei = 1; | |
8332 | inner; | |
8333 | inner = inner->next, i++, casei++) | |
8334 | { | |
8335 | basic_block s_entry_bb, s_exit_bb; | |
8336 | ||
8337 | /* Skip optional reduction region. */ | |
8338 | if (inner->type == GIMPLE_OMP_ATOMIC_LOAD) | |
8339 | { | |
8340 | --i; | |
8341 | --casei; | |
8342 | continue; | |
8343 | } | |
8344 | ||
8345 | s_entry_bb = inner->entry; | |
8346 | s_exit_bb = inner->exit; | |
8347 | ||
8348 | t = gimple_block_label (s_entry_bb); | |
8349 | u = build_int_cst (unsigned_type_node, casei); | |
8350 | u = build_case_label (u, NULL, t); | |
8351 | label_vec.quick_push (u); | |
8352 | ||
8353 | si = gsi_last_nondebug_bb (s_entry_bb); | |
8354 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SECTION); | |
8355 | gcc_assert (i < len || gimple_omp_section_last_p (gsi_stmt (si))); | |
8356 | gsi_remove (&si, true); | |
8357 | single_succ_edge (s_entry_bb)->flags = EDGE_FALLTHRU; | |
8358 | ||
8359 | if (s_exit_bb == NULL) | |
8360 | continue; | |
8361 | ||
8362 | si = gsi_last_nondebug_bb (s_exit_bb); | |
8363 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); | |
8364 | gsi_remove (&si, true); | |
8365 | ||
8366 | single_succ_edge (s_exit_bb)->flags = EDGE_FALLTHRU; | |
8367 | } | |
8368 | ||
8369 | /* Error handling code goes in DEFAULT_BB. */ | |
8370 | t = gimple_block_label (default_bb); | |
8371 | u = build_case_label (NULL, NULL, t); | |
8372 | make_edge (l0_bb, default_bb, 0); | |
8373 | add_bb_to_loop (default_bb, current_loops->tree_root); | |
8374 | ||
8375 | stmt = gimple_build_switch (vmain, u, label_vec); | |
8376 | gsi_insert_after (&switch_si, stmt, GSI_SAME_STMT); | |
8377 | gsi_remove (&switch_si, true); | |
8378 | ||
8379 | si = gsi_start_bb (default_bb); | |
8380 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_TRAP), 0); | |
8381 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); | |
8382 | ||
8383 | if (exit_reachable) | |
8384 | { | |
8385 | tree bfn_decl; | |
8386 | ||
8387 | /* Code to get the next section goes in L1_BB. */ | |
8388 | si = gsi_last_nondebug_bb (l1_bb); | |
8389 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CONTINUE); | |
8390 | ||
8391 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); | |
8392 | stmt = gimple_build_call (bfn_decl, 0); | |
8393 | gimple_call_set_lhs (stmt, vnext); | |
8394 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
8395 | if (cond_var) | |
8396 | { | |
8397 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), | |
8398 | vnext, build_one_cst (TREE_TYPE (cond_var))); | |
8399 | expand_omp_build_assign (&si, cond_var, t, false); | |
8400 | } | |
8401 | gsi_remove (&si, true); | |
8402 | ||
8403 | single_succ_edge (l1_bb)->flags = EDGE_FALLTHRU; | |
8404 | } | |
8405 | ||
8406 | /* Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB. */ | |
8407 | si = gsi_last_nondebug_bb (l2_bb); | |
8408 | if (gimple_omp_return_nowait_p (gsi_stmt (si))) | |
8409 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_NOWAIT); | |
8410 | else if (gimple_omp_return_lhs (gsi_stmt (si))) | |
8411 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_CANCEL); | |
8412 | else | |
8413 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END); | |
8414 | stmt = gimple_build_call (t, 0); | |
8415 | if (gimple_omp_return_lhs (gsi_stmt (si))) | |
8416 | gimple_call_set_lhs (stmt, gimple_omp_return_lhs (gsi_stmt (si))); | |
8417 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
8418 | gsi_remove (&si, true); | |
8419 | ||
8420 | set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb); | |
8421 | } | |
8422 | ||
8423 | /* Expand code for an OpenMP single or scope directive. We've already expanded | |
8424 | much of the code, here we simply place the GOMP_barrier call. */ | |
8425 | ||
8426 | static void | |
8427 | expand_omp_single (struct omp_region *region) | |
8428 | { | |
8429 | basic_block entry_bb, exit_bb; | |
8430 | gimple_stmt_iterator si; | |
8431 | ||
8432 | entry_bb = region->entry; | |
8433 | exit_bb = region->exit; | |
8434 | ||
8435 | si = gsi_last_nondebug_bb (entry_bb); | |
8436 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE | |
8437 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SCOPE); | |
8438 | gsi_remove (&si, true); | |
8439 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
8440 | ||
8441 | si = gsi_last_nondebug_bb (exit_bb); | |
8442 | if (!gimple_omp_return_nowait_p (gsi_stmt (si))) | |
8443 | { | |
8444 | tree t = gimple_omp_return_lhs (gsi_stmt (si)); | |
8445 | gsi_insert_after (&si, omp_build_barrier (t), GSI_SAME_STMT); | |
8446 | } | |
8447 | gsi_remove (&si, true); | |
8448 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; | |
8449 | } | |
8450 | ||
8451 | /* Generic expansion for OpenMP synchronization directives: master, | |
8452 | ordered and critical. All we need to do here is remove the entry | |
8453 | and exit markers for REGION. */ | |
8454 | ||
8455 | static void | |
8456 | expand_omp_synch (struct omp_region *region) | |
8457 | { | |
8458 | basic_block entry_bb, exit_bb; | |
8459 | gimple_stmt_iterator si; | |
8460 | ||
8461 | entry_bb = region->entry; | |
8462 | exit_bb = region->exit; | |
8463 | ||
8464 | si = gsi_last_nondebug_bb (entry_bb); | |
8465 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE | |
8466 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASTER | |
8467 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASKED | |
8468 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TASKGROUP | |
8469 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ORDERED | |
8470 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CRITICAL | |
8471 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS); | |
8472 | if (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS | |
8473 | && gimple_omp_teams_host (as_a <gomp_teams *> (gsi_stmt (si)))) | |
8474 | { | |
8475 | expand_omp_taskreg (region); | |
8476 | return; | |
8477 | } | |
8478 | gsi_remove (&si, true); | |
8479 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
8480 | ||
8481 | if (exit_bb) | |
8482 | { | |
8483 | si = gsi_last_nondebug_bb (exit_bb); | |
8484 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); | |
8485 | gsi_remove (&si, true); | |
8486 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; | |
8487 | } | |
8488 | } | |
8489 | ||
8490 | /* Translate enum omp_memory_order to enum memmodel for the embedded | |
8491 | fail clause in there. */ | |
8492 | ||
8493 | static enum memmodel | |
8494 | omp_memory_order_to_fail_memmodel (enum omp_memory_order mo) | |
8495 | { | |
8496 | switch (mo & OMP_FAIL_MEMORY_ORDER_MASK) | |
8497 | { | |
8498 | case OMP_FAIL_MEMORY_ORDER_UNSPECIFIED: | |
8499 | switch (mo & OMP_MEMORY_ORDER_MASK) | |
8500 | { | |
8501 | case OMP_MEMORY_ORDER_RELAXED: return MEMMODEL_RELAXED; | |
8502 | case OMP_MEMORY_ORDER_ACQUIRE: return MEMMODEL_ACQUIRE; | |
8503 | case OMP_MEMORY_ORDER_RELEASE: return MEMMODEL_RELAXED; | |
8504 | case OMP_MEMORY_ORDER_ACQ_REL: return MEMMODEL_ACQUIRE; | |
8505 | case OMP_MEMORY_ORDER_SEQ_CST: return MEMMODEL_SEQ_CST; | |
8506 | default: break; | |
8507 | } | |
8508 | gcc_unreachable (); | |
8509 | case OMP_FAIL_MEMORY_ORDER_RELAXED: return MEMMODEL_RELAXED; | |
8510 | case OMP_FAIL_MEMORY_ORDER_ACQUIRE: return MEMMODEL_ACQUIRE; | |
8511 | case OMP_FAIL_MEMORY_ORDER_SEQ_CST: return MEMMODEL_SEQ_CST; | |
8512 | default: gcc_unreachable (); | |
8513 | } | |
8514 | } | |
8515 | ||
8516 | /* Translate enum omp_memory_order to enum memmodel. The two enums | |
8517 | are using different numbers so that OMP_MEMORY_ORDER_UNSPECIFIED | |
8518 | is 0 and omp_memory_order has the fail mode encoded in it too. */ | |
8519 | ||
8520 | static enum memmodel | |
8521 | omp_memory_order_to_memmodel (enum omp_memory_order mo) | |
8522 | { | |
8523 | enum memmodel ret, fail_ret; | |
8524 | switch (mo & OMP_MEMORY_ORDER_MASK) | |
8525 | { | |
8526 | case OMP_MEMORY_ORDER_RELAXED: ret = MEMMODEL_RELAXED; break; | |
8527 | case OMP_MEMORY_ORDER_ACQUIRE: ret = MEMMODEL_ACQUIRE; break; | |
8528 | case OMP_MEMORY_ORDER_RELEASE: ret = MEMMODEL_RELEASE; break; | |
8529 | case OMP_MEMORY_ORDER_ACQ_REL: ret = MEMMODEL_ACQ_REL; break; | |
8530 | case OMP_MEMORY_ORDER_SEQ_CST: ret = MEMMODEL_SEQ_CST; break; | |
8531 | default: gcc_unreachable (); | |
8532 | } | |
8533 | /* If we drop the -Winvalid-memory-model warning for C++17 P0418R2, | |
8534 | we can just return ret here unconditionally. Otherwise, work around | |
8535 | it here and make sure fail memmodel is not stronger. */ | |
8536 | if ((mo & OMP_FAIL_MEMORY_ORDER_MASK) == OMP_FAIL_MEMORY_ORDER_UNSPECIFIED) | |
8537 | return ret; | |
8538 | fail_ret = omp_memory_order_to_fail_memmodel (mo); | |
8539 | if (fail_ret > ret) | |
8540 | return fail_ret; | |
8541 | return ret; | |
8542 | } | |
8543 | ||
8544 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
8545 | operation as a normal volatile load. */ | |
8546 | ||
8547 | static bool | |
8548 | expand_omp_atomic_load (basic_block load_bb, tree addr, | |
8549 | tree loaded_val, int index) | |
8550 | { | |
8551 | enum built_in_function tmpbase; | |
8552 | gimple_stmt_iterator gsi; | |
8553 | basic_block store_bb; | |
8554 | location_t loc; | |
8555 | gimple *stmt; | |
8556 | tree decl, call, type, itype; | |
8557 | ||
8558 | gsi = gsi_last_nondebug_bb (load_bb); | |
8559 | stmt = gsi_stmt (gsi); | |
8560 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
8561 | loc = gimple_location (stmt); | |
8562 | ||
8563 | /* ??? If the target does not implement atomic_load_optab[mode], and mode | |
8564 | is smaller than word size, then expand_atomic_load assumes that the load | |
8565 | is atomic. We could avoid the builtin entirely in this case. */ | |
8566 | ||
8567 | tmpbase = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
8568 | decl = builtin_decl_explicit (tmpbase); | |
8569 | if (decl == NULL_TREE) | |
8570 | return false; | |
8571 | ||
8572 | type = TREE_TYPE (loaded_val); | |
8573 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
8574 | ||
8575 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt); | |
8576 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); | |
8577 | call = build_call_expr_loc (loc, decl, 2, addr, mo); | |
8578 | if (!useless_type_conversion_p (type, itype)) | |
8579 | call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
8580 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
8581 | ||
8582 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
8583 | gsi_remove (&gsi, true); | |
8584 | ||
8585 | store_bb = single_succ (load_bb); | |
8586 | gsi = gsi_last_nondebug_bb (store_bb); | |
8587 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); | |
8588 | gsi_remove (&gsi, true); | |
8589 | ||
8590 | if (gimple_in_ssa_p (cfun)) | |
8591 | update_ssa (TODO_update_ssa_no_phi); | |
8592 | ||
8593 | return true; | |
8594 | } | |
8595 | ||
8596 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
8597 | operation as a normal volatile store. */ | |
8598 | ||
8599 | static bool | |
8600 | expand_omp_atomic_store (basic_block load_bb, tree addr, | |
8601 | tree loaded_val, tree stored_val, int index) | |
8602 | { | |
8603 | enum built_in_function tmpbase; | |
8604 | gimple_stmt_iterator gsi; | |
8605 | basic_block store_bb = single_succ (load_bb); | |
8606 | location_t loc; | |
8607 | gimple *stmt; | |
8608 | tree decl, call, type, itype; | |
8609 | machine_mode imode; | |
8610 | bool exchange; | |
8611 | ||
8612 | gsi = gsi_last_nondebug_bb (load_bb); | |
8613 | stmt = gsi_stmt (gsi); | |
8614 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
8615 | ||
8616 | /* If the load value is needed, then this isn't a store but an exchange. */ | |
8617 | exchange = gimple_omp_atomic_need_value_p (stmt); | |
8618 | ||
8619 | gsi = gsi_last_nondebug_bb (store_bb); | |
8620 | stmt = gsi_stmt (gsi); | |
8621 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_STORE); | |
8622 | loc = gimple_location (stmt); | |
8623 | ||
8624 | /* ??? If the target does not implement atomic_store_optab[mode], and mode | |
8625 | is smaller than word size, then expand_atomic_store assumes that the store | |
8626 | is atomic. We could avoid the builtin entirely in this case. */ | |
8627 | ||
8628 | tmpbase = (exchange ? BUILT_IN_ATOMIC_EXCHANGE_N : BUILT_IN_ATOMIC_STORE_N); | |
8629 | tmpbase = (enum built_in_function) ((int) tmpbase + index + 1); | |
8630 | decl = builtin_decl_explicit (tmpbase); | |
8631 | if (decl == NULL_TREE) | |
8632 | return false; | |
8633 | ||
8634 | type = TREE_TYPE (stored_val); | |
8635 | ||
8636 | /* Dig out the type of the function's second argument. */ | |
8637 | itype = TREE_TYPE (decl); | |
8638 | itype = TYPE_ARG_TYPES (itype); | |
8639 | itype = TREE_CHAIN (itype); | |
8640 | itype = TREE_VALUE (itype); | |
8641 | imode = TYPE_MODE (itype); | |
8642 | ||
8643 | if (exchange && !can_atomic_exchange_p (imode, true)) | |
8644 | return false; | |
8645 | ||
8646 | if (!useless_type_conversion_p (itype, type)) | |
8647 | stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val); | |
8648 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt); | |
8649 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); | |
8650 | call = build_call_expr_loc (loc, decl, 3, addr, stored_val, mo); | |
8651 | if (exchange) | |
8652 | { | |
8653 | if (!useless_type_conversion_p (type, itype)) | |
8654 | call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
8655 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
8656 | } | |
8657 | ||
8658 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
8659 | gsi_remove (&gsi, true); | |
8660 | ||
8661 | /* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */ | |
8662 | gsi = gsi_last_nondebug_bb (load_bb); | |
8663 | gsi_remove (&gsi, true); | |
8664 | ||
8665 | if (gimple_in_ssa_p (cfun)) | |
8666 | update_ssa (TODO_update_ssa_no_phi); | |
8667 | ||
8668 | return true; | |
8669 | } | |
8670 | ||
8671 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
8672 | operation as a __atomic_fetch_op builtin. INDEX is log2 of the | |
8673 | size of the data type, and thus usable to find the index of the builtin | |
8674 | decl. Returns false if the expression is not of the proper form. */ | |
8675 | ||
8676 | static bool | |
8677 | expand_omp_atomic_fetch_op (basic_block load_bb, | |
8678 | tree addr, tree loaded_val, | |
8679 | tree stored_val, int index) | |
8680 | { | |
8681 | enum built_in_function oldbase, newbase, tmpbase; | |
8682 | tree decl, itype, call; | |
8683 | tree lhs, rhs; | |
8684 | basic_block store_bb = single_succ (load_bb); | |
8685 | gimple_stmt_iterator gsi; | |
8686 | gimple *stmt; | |
8687 | location_t loc; | |
8688 | enum tree_code code; | |
8689 | bool need_old, need_new; | |
8690 | machine_mode imode; | |
8691 | ||
8692 | /* We expect to find the following sequences: | |
8693 | ||
8694 | load_bb: | |
8695 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) | |
8696 | ||
8697 | store_bb: | |
8698 | val = tmp OP something; (or: something OP tmp) | |
8699 | GIMPLE_OMP_STORE (val) | |
8700 | ||
8701 | ???FIXME: Allow a more flexible sequence. | |
8702 | Perhaps use data flow to pick the statements. | |
8703 | ||
8704 | */ | |
8705 | ||
8706 | gsi = gsi_after_labels (store_bb); | |
8707 | stmt = gsi_stmt (gsi); | |
8708 | if (is_gimple_debug (stmt)) | |
8709 | { | |
8710 | gsi_next_nondebug (&gsi); | |
8711 | if (gsi_end_p (gsi)) | |
8712 | return false; | |
8713 | stmt = gsi_stmt (gsi); | |
8714 | } | |
8715 | loc = gimple_location (stmt); | |
8716 | if (!is_gimple_assign (stmt)) | |
8717 | return false; | |
8718 | gsi_next_nondebug (&gsi); | |
8719 | if (gimple_code (gsi_stmt (gsi)) != GIMPLE_OMP_ATOMIC_STORE) | |
8720 | return false; | |
8721 | need_new = gimple_omp_atomic_need_value_p (gsi_stmt (gsi)); | |
8722 | need_old = gimple_omp_atomic_need_value_p (last_stmt (load_bb)); | |
8723 | enum omp_memory_order omo | |
8724 | = gimple_omp_atomic_memory_order (last_stmt (load_bb)); | |
8725 | enum memmodel mo = omp_memory_order_to_memmodel (omo); | |
8726 | gcc_checking_assert (!need_old || !need_new); | |
8727 | ||
8728 | if (!operand_equal_p (gimple_assign_lhs (stmt), stored_val, 0)) | |
8729 | return false; | |
8730 | ||
8731 | /* Check for one of the supported fetch-op operations. */ | |
8732 | code = gimple_assign_rhs_code (stmt); | |
8733 | switch (code) | |
8734 | { | |
8735 | case PLUS_EXPR: | |
8736 | case POINTER_PLUS_EXPR: | |
8737 | oldbase = BUILT_IN_ATOMIC_FETCH_ADD_N; | |
8738 | newbase = BUILT_IN_ATOMIC_ADD_FETCH_N; | |
8739 | break; | |
8740 | case MINUS_EXPR: | |
8741 | oldbase = BUILT_IN_ATOMIC_FETCH_SUB_N; | |
8742 | newbase = BUILT_IN_ATOMIC_SUB_FETCH_N; | |
8743 | break; | |
8744 | case BIT_AND_EXPR: | |
8745 | oldbase = BUILT_IN_ATOMIC_FETCH_AND_N; | |
8746 | newbase = BUILT_IN_ATOMIC_AND_FETCH_N; | |
8747 | break; | |
8748 | case BIT_IOR_EXPR: | |
8749 | oldbase = BUILT_IN_ATOMIC_FETCH_OR_N; | |
8750 | newbase = BUILT_IN_ATOMIC_OR_FETCH_N; | |
8751 | break; | |
8752 | case BIT_XOR_EXPR: | |
8753 | oldbase = BUILT_IN_ATOMIC_FETCH_XOR_N; | |
8754 | newbase = BUILT_IN_ATOMIC_XOR_FETCH_N; | |
8755 | break; | |
8756 | default: | |
8757 | return false; | |
8758 | } | |
8759 | ||
8760 | /* Make sure the expression is of the proper form. */ | |
8761 | if (operand_equal_p (gimple_assign_rhs1 (stmt), loaded_val, 0)) | |
8762 | rhs = gimple_assign_rhs2 (stmt); | |
8763 | else if (commutative_tree_code (gimple_assign_rhs_code (stmt)) | |
8764 | && operand_equal_p (gimple_assign_rhs2 (stmt), loaded_val, 0)) | |
8765 | rhs = gimple_assign_rhs1 (stmt); | |
8766 | else | |
8767 | return false; | |
8768 | ||
8769 | tmpbase = ((enum built_in_function) | |
8770 | ((need_new ? newbase : oldbase) + index + 1)); | |
8771 | decl = builtin_decl_explicit (tmpbase); | |
8772 | if (decl == NULL_TREE) | |
8773 | return false; | |
8774 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
8775 | imode = TYPE_MODE (itype); | |
8776 | ||
8777 | /* We could test all of the various optabs involved, but the fact of the | |
8778 | matter is that (with the exception of i486 vs i586 and xadd) all targets | |
8779 | that support any atomic operaton optab also implements compare-and-swap. | |
8780 | Let optabs.c take care of expanding any compare-and-swap loop. */ | |
8781 | if (!can_compare_and_swap_p (imode, true) || !can_atomic_load_p (imode)) | |
8782 | return false; | |
8783 | ||
8784 | gsi = gsi_last_nondebug_bb (load_bb); | |
8785 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_LOAD); | |
8786 | ||
8787 | /* OpenMP does not imply any barrier-like semantics on its atomic ops. | |
8788 | It only requires that the operation happen atomically. Thus we can | |
8789 | use the RELAXED memory model. */ | |
8790 | call = build_call_expr_loc (loc, decl, 3, addr, | |
8791 | fold_convert_loc (loc, itype, rhs), | |
8792 | build_int_cst (NULL, mo)); | |
8793 | ||
8794 | if (need_old || need_new) | |
8795 | { | |
8796 | lhs = need_old ? loaded_val : stored_val; | |
8797 | call = fold_convert_loc (loc, TREE_TYPE (lhs), call); | |
8798 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, lhs, call); | |
8799 | } | |
8800 | else | |
8801 | call = fold_convert_loc (loc, void_type_node, call); | |
8802 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
8803 | gsi_remove (&gsi, true); | |
8804 | ||
8805 | gsi = gsi_last_nondebug_bb (store_bb); | |
8806 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); | |
8807 | gsi_remove (&gsi, true); | |
8808 | gsi = gsi_last_nondebug_bb (store_bb); | |
8809 | stmt = gsi_stmt (gsi); | |
8810 | gsi_remove (&gsi, true); | |
8811 | ||
8812 | if (gimple_in_ssa_p (cfun)) | |
8813 | { | |
8814 | release_defs (stmt); | |
8815 | update_ssa (TODO_update_ssa_no_phi); | |
8816 | } | |
8817 | ||
8818 | return true; | |
8819 | } | |
8820 | ||
8821 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
8822 | compare and exchange as an ATOMIC_COMPARE_EXCHANGE internal function. | |
8823 | Returns false if the expression is not of the proper form. */ | |
8824 | ||
8825 | static bool | |
8826 | expand_omp_atomic_cas (basic_block load_bb, tree addr, | |
8827 | tree loaded_val, tree stored_val, int index) | |
8828 | { | |
8829 | /* We expect to find the following sequences: | |
8830 | ||
8831 | load_bb: | |
8832 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) | |
8833 | ||
8834 | store_bb: | |
8835 | val = tmp == e ? d : tmp; | |
8836 | GIMPLE_OMP_ATOMIC_STORE (val) | |
8837 | ||
8838 | or in store_bb instead: | |
8839 | tmp2 = tmp == e; | |
8840 | val = tmp2 ? d : tmp; | |
8841 | GIMPLE_OMP_ATOMIC_STORE (val) | |
8842 | ||
8843 | or: | |
8844 | tmp3 = VIEW_CONVERT_EXPR<integral_type>(tmp); | |
8845 | val = e == tmp3 ? d : tmp; | |
8846 | GIMPLE_OMP_ATOMIC_STORE (val) | |
8847 | ||
8848 | etc. */ | |
8849 | ||
8850 | ||
8851 | basic_block store_bb = single_succ (load_bb); | |
8852 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (store_bb); | |
8853 | gimple *store_stmt = gsi_stmt (gsi); | |
8854 | if (!store_stmt || gimple_code (store_stmt) != GIMPLE_OMP_ATOMIC_STORE) | |
8855 | return false; | |
8856 | gsi_prev_nondebug (&gsi); | |
8857 | if (gsi_end_p (gsi)) | |
8858 | return false; | |
8859 | gimple *condexpr_stmt = gsi_stmt (gsi); | |
8860 | if (!is_gimple_assign (condexpr_stmt) | |
8861 | || gimple_assign_rhs_code (condexpr_stmt) != COND_EXPR) | |
8862 | return false; | |
8863 | if (!operand_equal_p (gimple_assign_lhs (condexpr_stmt), stored_val, 0)) | |
8864 | return false; | |
8865 | gimple *cond_stmt = NULL; | |
8866 | gimple *vce_stmt = NULL; | |
8867 | gsi_prev_nondebug (&gsi); | |
8868 | if (!gsi_end_p (gsi)) | |
8869 | { | |
8870 | cond_stmt = gsi_stmt (gsi); | |
8871 | if (!is_gimple_assign (cond_stmt)) | |
8872 | return false; | |
8873 | if (gimple_assign_rhs_code (cond_stmt) == EQ_EXPR) | |
8874 | { | |
8875 | gsi_prev_nondebug (&gsi); | |
8876 | if (!gsi_end_p (gsi)) | |
8877 | { | |
8878 | vce_stmt = gsi_stmt (gsi); | |
8879 | if (!is_gimple_assign (vce_stmt) | |
8880 | || gimple_assign_rhs_code (vce_stmt) != VIEW_CONVERT_EXPR) | |
8881 | return false; | |
8882 | } | |
8883 | } | |
8884 | else if (gimple_assign_rhs_code (cond_stmt) == VIEW_CONVERT_EXPR) | |
8885 | std::swap (vce_stmt, cond_stmt); | |
8886 | else | |
8887 | return false; | |
8888 | if (vce_stmt) | |
8889 | { | |
8890 | tree vce_rhs = gimple_assign_rhs1 (vce_stmt); | |
8891 | if (TREE_CODE (vce_rhs) != VIEW_CONVERT_EXPR | |
8892 | || !operand_equal_p (TREE_OPERAND (vce_rhs, 0), loaded_val)) | |
8893 | return false; | |
8894 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vce_rhs)) | |
8895 | || !SCALAR_FLOAT_TYPE_P (TREE_TYPE (loaded_val)) | |
8896 | || !tree_int_cst_equal (TYPE_SIZE (TREE_TYPE (vce_rhs)), | |
8897 | TYPE_SIZE (TREE_TYPE (loaded_val)))) | |
8898 | return false; | |
8899 | gsi_prev_nondebug (&gsi); | |
8900 | if (!gsi_end_p (gsi)) | |
8901 | return false; | |
8902 | } | |
8903 | } | |
8904 | tree cond = gimple_assign_rhs1 (condexpr_stmt); | |
8905 | tree cond_op1, cond_op2; | |
8906 | if (cond_stmt) | |
8907 | { | |
8908 | if (!operand_equal_p (cond, gimple_assign_lhs (cond_stmt))) | |
8909 | return false; | |
8910 | cond_op1 = gimple_assign_rhs1 (cond_stmt); | |
8911 | cond_op2 = gimple_assign_rhs2 (cond_stmt); | |
8912 | } | |
8913 | else if (TREE_CODE (cond) != EQ_EXPR && TREE_CODE (cond) != NE_EXPR) | |
8914 | return false; | |
8915 | else | |
8916 | { | |
8917 | cond_op1 = TREE_OPERAND (cond, 0); | |
8918 | cond_op2 = TREE_OPERAND (cond, 1); | |
8919 | } | |
8920 | tree d; | |
8921 | if (TREE_CODE (cond) == NE_EXPR) | |
8922 | { | |
8923 | if (!operand_equal_p (gimple_assign_rhs2 (condexpr_stmt), loaded_val)) | |
8924 | return false; | |
8925 | d = gimple_assign_rhs3 (condexpr_stmt); | |
8926 | } | |
8927 | else if (!operand_equal_p (gimple_assign_rhs3 (condexpr_stmt), loaded_val)) | |
8928 | return false; | |
8929 | else | |
8930 | d = gimple_assign_rhs2 (condexpr_stmt); | |
8931 | tree e = vce_stmt ? gimple_assign_lhs (vce_stmt) : loaded_val; | |
8932 | if (operand_equal_p (e, cond_op1)) | |
8933 | e = cond_op2; | |
8934 | else if (operand_equal_p (e, cond_op2)) | |
8935 | e = cond_op1; | |
8936 | else | |
8937 | return false; | |
8938 | ||
8939 | location_t loc = gimple_location (store_stmt); | |
8940 | gimple *load_stmt = last_stmt (load_bb); | |
8941 | bool need_new = gimple_omp_atomic_need_value_p (store_stmt); | |
8942 | bool need_old = gimple_omp_atomic_need_value_p (load_stmt); | |
8943 | bool weak = gimple_omp_atomic_weak_p (load_stmt); | |
8944 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (load_stmt); | |
8945 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); | |
8946 | tree fmo = build_int_cst (NULL, omp_memory_order_to_fail_memmodel (omo)); | |
8947 | gcc_checking_assert (!need_old || !need_new); | |
8948 | ||
8949 | enum built_in_function fncode | |
8950 | = (enum built_in_function) ((int) BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N | |
8951 | + index + 1); | |
8952 | tree cmpxchg = builtin_decl_explicit (fncode); | |
8953 | if (cmpxchg == NULL_TREE) | |
8954 | return false; | |
8955 | tree itype = TREE_TYPE (TREE_TYPE (cmpxchg)); | |
8956 | ||
8957 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true) | |
8958 | || !can_atomic_load_p (TYPE_MODE (itype))) | |
8959 | return false; | |
8960 | ||
8961 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); | |
8962 | if (SCALAR_FLOAT_TYPE_P (type) && !vce_stmt) | |
8963 | return false; | |
8964 | ||
8965 | gsi = gsi_for_stmt (store_stmt); | |
8966 | if (!useless_type_conversion_p (itype, TREE_TYPE (e))) | |
8967 | { | |
8968 | tree ne = create_tmp_reg (itype); | |
8969 | gimple *g = gimple_build_assign (ne, NOP_EXPR, e); | |
8970 | gimple_set_location (g, loc); | |
8971 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
8972 | e = ne; | |
8973 | } | |
8974 | if (!useless_type_conversion_p (itype, TREE_TYPE (d))) | |
8975 | { | |
8976 | tree nd = create_tmp_reg (itype); | |
8977 | enum tree_code code; | |
8978 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (d))) | |
8979 | { | |
8980 | code = VIEW_CONVERT_EXPR; | |
8981 | d = build1 (VIEW_CONVERT_EXPR, itype, d); | |
8982 | } | |
8983 | else | |
8984 | code = NOP_EXPR; | |
8985 | gimple *g = gimple_build_assign (nd, code, d); | |
8986 | gimple_set_location (g, loc); | |
8987 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
8988 | d = nd; | |
8989 | } | |
8990 | ||
8991 | tree ctype = build_complex_type (itype); | |
8992 | int flag = int_size_in_bytes (itype) + (weak ? 256 : 0); | |
8993 | gimple *g | |
8994 | = gimple_build_call_internal (IFN_ATOMIC_COMPARE_EXCHANGE, 6, addr, e, d, | |
8995 | build_int_cst (integer_type_node, flag), | |
8996 | mo, fmo); | |
8997 | tree cres = create_tmp_reg (ctype); | |
8998 | gimple_call_set_lhs (g, cres); | |
8999 | gimple_set_location (g, loc); | |
9000 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9001 | ||
9002 | if (cond_stmt || need_old || need_new) | |
9003 | { | |
9004 | tree im = create_tmp_reg (itype); | |
9005 | g = gimple_build_assign (im, IMAGPART_EXPR, | |
9006 | build1 (IMAGPART_EXPR, itype, cres)); | |
9007 | gimple_set_location (g, loc); | |
9008 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9009 | ||
9010 | tree re = NULL_TREE; | |
9011 | if (need_old || need_new) | |
9012 | { | |
9013 | re = create_tmp_reg (itype); | |
9014 | g = gimple_build_assign (re, REALPART_EXPR, | |
9015 | build1 (REALPART_EXPR, itype, cres)); | |
9016 | gimple_set_location (g, loc); | |
9017 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9018 | } | |
9019 | ||
9020 | if (cond_stmt) | |
9021 | { | |
9022 | g = gimple_build_assign (gimple_assign_lhs (cond_stmt), | |
9023 | NOP_EXPR, im); | |
9024 | gimple_set_location (g, loc); | |
9025 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9026 | } | |
9027 | else if (need_new) | |
9028 | { | |
9029 | g = gimple_build_assign (create_tmp_reg (itype), COND_EXPR, | |
9030 | build2 (NE_EXPR, boolean_type_node, | |
9031 | im, build_zero_cst (itype)), | |
9032 | d, re); | |
9033 | gimple_set_location (g, loc); | |
9034 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9035 | re = gimple_assign_lhs (g); | |
9036 | } | |
9037 | ||
9038 | if (need_old || need_new) | |
9039 | { | |
9040 | tree v = need_old ? loaded_val : stored_val; | |
9041 | enum tree_code code; | |
9042 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (v))) | |
9043 | { | |
9044 | code = VIEW_CONVERT_EXPR; | |
9045 | re = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (v), re); | |
9046 | } | |
9047 | else if (!useless_type_conversion_p (TREE_TYPE (v), itype)) | |
9048 | code = NOP_EXPR; | |
9049 | else | |
9050 | code = TREE_CODE (re); | |
9051 | g = gimple_build_assign (v, code, re); | |
9052 | gimple_set_location (g, loc); | |
9053 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9054 | } | |
9055 | } | |
9056 | ||
9057 | gsi_remove (&gsi, true); | |
9058 | gsi = gsi_for_stmt (load_stmt); | |
9059 | gsi_remove (&gsi, true); | |
9060 | gsi = gsi_for_stmt (condexpr_stmt); | |
9061 | gsi_remove (&gsi, true); | |
9062 | if (cond_stmt) | |
9063 | { | |
9064 | gsi = gsi_for_stmt (cond_stmt); | |
9065 | gsi_remove (&gsi, true); | |
9066 | } | |
9067 | if (vce_stmt) | |
9068 | { | |
9069 | gsi = gsi_for_stmt (vce_stmt); | |
9070 | gsi_remove (&gsi, true); | |
9071 | } | |
9072 | ||
9073 | return true; | |
9074 | } | |
9075 | ||
9076 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
9077 | ||
9078 | oldval = *addr; | |
9079 | repeat: | |
9080 | newval = rhs; // with oldval replacing *addr in rhs | |
9081 | oldval = __sync_val_compare_and_swap (addr, oldval, newval); | |
9082 | if (oldval != newval) | |
9083 | goto repeat; | |
9084 | ||
9085 | INDEX is log2 of the size of the data type, and thus usable to find the | |
9086 | index of the builtin decl. */ | |
9087 | ||
9088 | static bool | |
9089 | expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb, | |
9090 | tree addr, tree loaded_val, tree stored_val, | |
9091 | int index) | |
9092 | { | |
9093 | tree loadedi, storedi, initial, new_storedi, old_vali; | |
9094 | tree type, itype, cmpxchg, iaddr, atype; | |
9095 | gimple_stmt_iterator si; | |
9096 | basic_block loop_header = single_succ (load_bb); | |
9097 | gimple *phi, *stmt; | |
9098 | edge e; | |
9099 | enum built_in_function fncode; | |
9100 | ||
9101 | fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N | |
9102 | + index + 1); | |
9103 | cmpxchg = builtin_decl_explicit (fncode); | |
9104 | if (cmpxchg == NULL_TREE) | |
9105 | return false; | |
9106 | type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); | |
9107 | atype = type; | |
9108 | itype = TREE_TYPE (TREE_TYPE (cmpxchg)); | |
9109 | ||
9110 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true) | |
9111 | || !can_atomic_load_p (TYPE_MODE (itype))) | |
9112 | return false; | |
9113 | ||
9114 | /* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */ | |
9115 | si = gsi_last_nondebug_bb (load_bb); | |
9116 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); | |
9117 | location_t loc = gimple_location (gsi_stmt (si)); | |
9118 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (gsi_stmt (si)); | |
9119 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); | |
9120 | tree fmo = build_int_cst (NULL, omp_memory_order_to_fail_memmodel (omo)); | |
9121 | ||
9122 | /* For floating-point values, we'll need to view-convert them to integers | |
9123 | so that we can perform the atomic compare and swap. Simplify the | |
9124 | following code by always setting up the "i"ntegral variables. */ | |
9125 | if (!INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type)) | |
9126 | { | |
9127 | tree iaddr_val; | |
9128 | ||
9129 | iaddr = create_tmp_reg (build_pointer_type_for_mode (itype, ptr_mode, | |
9130 | true)); | |
9131 | atype = itype; | |
9132 | iaddr_val | |
9133 | = force_gimple_operand_gsi (&si, | |
9134 | fold_convert (TREE_TYPE (iaddr), addr), | |
9135 | false, NULL_TREE, true, GSI_SAME_STMT); | |
9136 | stmt = gimple_build_assign (iaddr, iaddr_val); | |
9137 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9138 | loadedi = create_tmp_var (itype); | |
9139 | if (gimple_in_ssa_p (cfun)) | |
9140 | loadedi = make_ssa_name (loadedi); | |
9141 | } | |
9142 | else | |
9143 | { | |
9144 | iaddr = addr; | |
9145 | loadedi = loaded_val; | |
9146 | } | |
9147 | ||
9148 | fncode = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
9149 | tree loaddecl = builtin_decl_explicit (fncode); | |
9150 | if (loaddecl) | |
9151 | initial | |
9152 | = fold_convert (atype, | |
9153 | build_call_expr (loaddecl, 2, iaddr, | |
9154 | build_int_cst (NULL_TREE, | |
9155 | MEMMODEL_RELAXED))); | |
9156 | else | |
9157 | { | |
9158 | tree off | |
9159 | = build_int_cst (build_pointer_type_for_mode (atype, ptr_mode, | |
9160 | true), 0); | |
9161 | initial = build2 (MEM_REF, atype, iaddr, off); | |
9162 | } | |
9163 | ||
9164 | initial | |
9165 | = force_gimple_operand_gsi (&si, initial, true, NULL_TREE, true, | |
9166 | GSI_SAME_STMT); | |
9167 | ||
9168 | /* Move the value to the LOADEDI temporary. */ | |
9169 | if (gimple_in_ssa_p (cfun)) | |
9170 | { | |
9171 | gcc_assert (gimple_seq_empty_p (phi_nodes (loop_header))); | |
9172 | phi = create_phi_node (loadedi, loop_header); | |
9173 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)), | |
9174 | initial); | |
9175 | } | |
9176 | else | |
9177 | gsi_insert_before (&si, | |
9178 | gimple_build_assign (loadedi, initial), | |
9179 | GSI_SAME_STMT); | |
9180 | if (loadedi != loaded_val) | |
9181 | { | |
9182 | gimple_stmt_iterator gsi2; | |
9183 | tree x; | |
9184 | ||
9185 | x = build1 (VIEW_CONVERT_EXPR, type, loadedi); | |
9186 | gsi2 = gsi_start_bb (loop_header); | |
9187 | if (gimple_in_ssa_p (cfun)) | |
9188 | { | |
9189 | gassign *stmt; | |
9190 | x = force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
9191 | true, GSI_SAME_STMT); | |
9192 | stmt = gimple_build_assign (loaded_val, x); | |
9193 | gsi_insert_before (&gsi2, stmt, GSI_SAME_STMT); | |
9194 | } | |
9195 | else | |
9196 | { | |
9197 | x = build2 (MODIFY_EXPR, TREE_TYPE (loaded_val), loaded_val, x); | |
9198 | force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
9199 | true, GSI_SAME_STMT); | |
9200 | } | |
9201 | } | |
9202 | gsi_remove (&si, true); | |
9203 | ||
9204 | si = gsi_last_nondebug_bb (store_bb); | |
9205 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); | |
9206 | ||
9207 | if (iaddr == addr) | |
9208 | storedi = stored_val; | |
9209 | else | |
9210 | storedi | |
9211 | = force_gimple_operand_gsi (&si, | |
9212 | build1 (VIEW_CONVERT_EXPR, itype, | |
9213 | stored_val), true, NULL_TREE, true, | |
9214 | GSI_SAME_STMT); | |
9215 | ||
9216 | /* Build the compare&swap statement. */ | |
9217 | tree ctype = build_complex_type (itype); | |
9218 | int flag = int_size_in_bytes (itype); | |
9219 | new_storedi = build_call_expr_internal_loc (loc, IFN_ATOMIC_COMPARE_EXCHANGE, | |
9220 | ctype, 6, iaddr, loadedi, | |
9221 | storedi, | |
9222 | build_int_cst (integer_type_node, | |
9223 | flag), | |
9224 | mo, fmo); | |
9225 | new_storedi = build1 (REALPART_EXPR, itype, new_storedi); | |
9226 | new_storedi = force_gimple_operand_gsi (&si, | |
9227 | fold_convert (TREE_TYPE (loadedi), | |
9228 | new_storedi), | |
9229 | true, NULL_TREE, | |
9230 | true, GSI_SAME_STMT); | |
9231 | ||
9232 | if (gimple_in_ssa_p (cfun)) | |
9233 | old_vali = loadedi; | |
9234 | else | |
9235 | { | |
9236 | old_vali = create_tmp_var (TREE_TYPE (loadedi)); | |
9237 | stmt = gimple_build_assign (old_vali, loadedi); | |
9238 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9239 | ||
9240 | stmt = gimple_build_assign (loadedi, new_storedi); | |
9241 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9242 | } | |
9243 | ||
9244 | /* Note that we always perform the comparison as an integer, even for | |
9245 | floating point. This allows the atomic operation to properly | |
9246 | succeed even with NaNs and -0.0. */ | |
9247 | tree ne = build2 (NE_EXPR, boolean_type_node, new_storedi, old_vali); | |
9248 | stmt = gimple_build_cond_empty (ne); | |
9249 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9250 | ||
9251 | /* Update cfg. */ | |
9252 | e = single_succ_edge (store_bb); | |
9253 | e->flags &= ~EDGE_FALLTHRU; | |
9254 | e->flags |= EDGE_FALSE_VALUE; | |
9255 | /* Expect no looping. */ | |
9256 | e->probability = profile_probability::guessed_always (); | |
9257 | ||
9258 | e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE); | |
9259 | e->probability = profile_probability::guessed_never (); | |
9260 | ||
9261 | /* Copy the new value to loadedi (we already did that before the condition | |
9262 | if we are not in SSA). */ | |
9263 | if (gimple_in_ssa_p (cfun)) | |
9264 | { | |
9265 | phi = gimple_seq_first_stmt (phi_nodes (loop_header)); | |
9266 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_storedi); | |
9267 | } | |
9268 | ||
9269 | /* Remove GIMPLE_OMP_ATOMIC_STORE. */ | |
9270 | gsi_remove (&si, true); | |
9271 | ||
9272 | class loop *loop = alloc_loop (); | |
9273 | loop->header = loop_header; | |
9274 | loop->latch = store_bb; | |
9275 | add_loop (loop, loop_header->loop_father); | |
9276 | ||
9277 | if (gimple_in_ssa_p (cfun)) | |
9278 | update_ssa (TODO_update_ssa_no_phi); | |
9279 | ||
9280 | return true; | |
9281 | } | |
9282 | ||
9283 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
9284 | ||
9285 | GOMP_atomic_start (); | |
9286 | *addr = rhs; | |
9287 | GOMP_atomic_end (); | |
9288 | ||
9289 | The result is not globally atomic, but works so long as all parallel | |
9290 | references are within #pragma omp atomic directives. According to | |
9291 | responses received from omp@openmp.org, appears to be within spec. | |
9292 | Which makes sense, since that's how several other compilers handle | |
9293 | this situation as well. | |
9294 | LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're | |
9295 | expanding. STORED_VAL is the operand of the matching | |
9296 | GIMPLE_OMP_ATOMIC_STORE. | |
9297 | ||
9298 | We replace | |
9299 | GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with | |
9300 | loaded_val = *addr; | |
9301 | ||
9302 | and replace | |
9303 | GIMPLE_OMP_ATOMIC_STORE (stored_val) with | |
9304 | *addr = stored_val; | |
9305 | */ | |
9306 | ||
9307 | static bool | |
9308 | expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb, | |
9309 | tree addr, tree loaded_val, tree stored_val) | |
9310 | { | |
9311 | gimple_stmt_iterator si; | |
9312 | gassign *stmt; | |
9313 | tree t; | |
9314 | ||
9315 | si = gsi_last_nondebug_bb (load_bb); | |
9316 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); | |
9317 | ||
9318 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START); | |
9319 | t = build_call_expr (t, 0); | |
9320 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
9321 | ||
9322 | tree mem = build_simple_mem_ref (addr); | |
9323 | TREE_TYPE (mem) = TREE_TYPE (loaded_val); | |
9324 | TREE_OPERAND (mem, 1) | |
9325 | = fold_convert (build_pointer_type_for_mode (TREE_TYPE (mem), ptr_mode, | |
9326 | true), | |
9327 | TREE_OPERAND (mem, 1)); | |
9328 | stmt = gimple_build_assign (loaded_val, mem); | |
9329 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9330 | gsi_remove (&si, true); | |
9331 | ||
9332 | si = gsi_last_nondebug_bb (store_bb); | |
9333 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); | |
9334 | ||
9335 | stmt = gimple_build_assign (unshare_expr (mem), stored_val); | |
9336 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9337 | ||
9338 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END); | |
9339 | t = build_call_expr (t, 0); | |
9340 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
9341 | gsi_remove (&si, true); | |
9342 | ||
9343 | if (gimple_in_ssa_p (cfun)) | |
9344 | update_ssa (TODO_update_ssa_no_phi); | |
9345 | return true; | |
9346 | } | |
9347 | ||
9348 | /* Expand an GIMPLE_OMP_ATOMIC statement. We try to expand | |
9349 | using expand_omp_atomic_fetch_op. If it failed, we try to | |
9350 | call expand_omp_atomic_pipeline, and if it fails too, the | |
9351 | ultimate fallback is wrapping the operation in a mutex | |
9352 | (expand_omp_atomic_mutex). REGION is the atomic region built | |
9353 | by build_omp_regions_1(). */ | |
9354 | ||
9355 | static void | |
9356 | expand_omp_atomic (struct omp_region *region) | |
9357 | { | |
9358 | basic_block load_bb = region->entry, store_bb = region->exit; | |
9359 | gomp_atomic_load *load = as_a <gomp_atomic_load *> (last_stmt (load_bb)); | |
9360 | gomp_atomic_store *store = as_a <gomp_atomic_store *> (last_stmt (store_bb)); | |
9361 | tree loaded_val = gimple_omp_atomic_load_lhs (load); | |
9362 | tree addr = gimple_omp_atomic_load_rhs (load); | |
9363 | tree stored_val = gimple_omp_atomic_store_val (store); | |
9364 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); | |
9365 | HOST_WIDE_INT index; | |
9366 | ||
9367 | /* Make sure the type is one of the supported sizes. */ | |
9368 | index = tree_to_uhwi (TYPE_SIZE_UNIT (type)); | |
9369 | index = exact_log2 (index); | |
9370 | if (index >= 0 && index <= 4) | |
9371 | { | |
9372 | unsigned int align = TYPE_ALIGN_UNIT (type); | |
9373 | ||
9374 | /* __sync builtins require strict data alignment. */ | |
9375 | if (exact_log2 (align) >= index) | |
9376 | { | |
9377 | /* Atomic load. */ | |
9378 | scalar_mode smode; | |
9379 | if (loaded_val == stored_val | |
9380 | && (is_int_mode (TYPE_MODE (type), &smode) | |
9381 | || is_float_mode (TYPE_MODE (type), &smode)) | |
9382 | && GET_MODE_BITSIZE (smode) <= BITS_PER_WORD | |
9383 | && expand_omp_atomic_load (load_bb, addr, loaded_val, index)) | |
9384 | return; | |
9385 | ||
9386 | /* Atomic store. */ | |
9387 | if ((is_int_mode (TYPE_MODE (type), &smode) | |
9388 | || is_float_mode (TYPE_MODE (type), &smode)) | |
9389 | && GET_MODE_BITSIZE (smode) <= BITS_PER_WORD | |
9390 | && store_bb == single_succ (load_bb) | |
9391 | && first_stmt (store_bb) == store | |
9392 | && expand_omp_atomic_store (load_bb, addr, loaded_val, | |
9393 | stored_val, index)) | |
9394 | return; | |
9395 | ||
9396 | /* When possible, use specialized atomic update functions. */ | |
9397 | if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type)) | |
9398 | && store_bb == single_succ (load_bb) | |
9399 | && expand_omp_atomic_fetch_op (load_bb, addr, | |
9400 | loaded_val, stored_val, index)) | |
9401 | return; | |
9402 | ||
9403 | /* When possible, use ATOMIC_COMPARE_EXCHANGE ifn without a loop. */ | |
9404 | if (store_bb == single_succ (load_bb) | |
9405 | && !gimple_in_ssa_p (cfun) | |
9406 | && expand_omp_atomic_cas (load_bb, addr, loaded_val, stored_val, | |
9407 | index)) | |
9408 | return; | |
9409 | ||
9410 | /* If we don't have specialized __sync builtins, try and implement | |
9411 | as a compare and swap loop. */ | |
9412 | if (expand_omp_atomic_pipeline (load_bb, store_bb, addr, | |
9413 | loaded_val, stored_val, index)) | |
9414 | return; | |
9415 | } | |
9416 | } | |
9417 | ||
9418 | /* The ultimate fallback is wrapping the operation in a mutex. */ | |
9419 | expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val); | |
9420 | } | |
9421 | ||
9422 | /* Mark the loops inside the kernels region starting at REGION_ENTRY and ending | |
9423 | at REGION_EXIT. */ | |
9424 | ||
9425 | static void | |
9426 | mark_loops_in_oacc_kernels_region (basic_block region_entry, | |
9427 | basic_block region_exit) | |
9428 | { | |
9429 | class loop *outer = region_entry->loop_father; | |
9430 | gcc_assert (region_exit == NULL || outer == region_exit->loop_father); | |
9431 | ||
9432 | /* Don't parallelize the kernels region if it contains more than one outer | |
9433 | loop. */ | |
9434 | unsigned int nr_outer_loops = 0; | |
9435 | class loop *single_outer = NULL; | |
9436 | for (class loop *loop = outer->inner; loop != NULL; loop = loop->next) | |
9437 | { | |
9438 | gcc_assert (loop_outer (loop) == outer); | |
9439 | ||
9440 | if (!dominated_by_p (CDI_DOMINATORS, loop->header, region_entry)) | |
9441 | continue; | |
9442 | ||
9443 | if (region_exit != NULL | |
9444 | && dominated_by_p (CDI_DOMINATORS, loop->header, region_exit)) | |
9445 | continue; | |
9446 | ||
9447 | nr_outer_loops++; | |
9448 | single_outer = loop; | |
9449 | } | |
9450 | if (nr_outer_loops != 1) | |
9451 | return; | |
9452 | ||
9453 | for (class loop *loop = single_outer->inner; | |
9454 | loop != NULL; | |
9455 | loop = loop->inner) | |
9456 | if (loop->next) | |
9457 | return; | |
9458 | ||
9459 | /* Mark the loops in the region. */ | |
9460 | for (class loop *loop = single_outer; loop != NULL; loop = loop->inner) | |
9461 | loop->in_oacc_kernels_region = true; | |
9462 | } | |
9463 | ||
9464 | /* Build target argument identifier from the DEVICE identifier, value | |
9465 | identifier ID and whether the element also has a SUBSEQUENT_PARAM. */ | |
9466 | ||
9467 | static tree | |
9468 | get_target_argument_identifier_1 (int device, bool subseqent_param, int id) | |
9469 | { | |
9470 | tree t = build_int_cst (integer_type_node, device); | |
9471 | if (subseqent_param) | |
9472 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
9473 | build_int_cst (integer_type_node, | |
9474 | GOMP_TARGET_ARG_SUBSEQUENT_PARAM)); | |
9475 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
9476 | build_int_cst (integer_type_node, id)); | |
9477 | return t; | |
9478 | } | |
9479 | ||
9480 | /* Like above but return it in type that can be directly stored as an element | |
9481 | of the argument array. */ | |
9482 | ||
9483 | static tree | |
9484 | get_target_argument_identifier (int device, bool subseqent_param, int id) | |
9485 | { | |
9486 | tree t = get_target_argument_identifier_1 (device, subseqent_param, id); | |
9487 | return fold_convert (ptr_type_node, t); | |
9488 | } | |
9489 | ||
9490 | /* Return a target argument consisting of DEVICE identifier, value identifier | |
9491 | ID, and the actual VALUE. */ | |
9492 | ||
9493 | static tree | |
9494 | get_target_argument_value (gimple_stmt_iterator *gsi, int device, int id, | |
9495 | tree value) | |
9496 | { | |
9497 | tree t = fold_build2 (LSHIFT_EXPR, integer_type_node, | |
9498 | fold_convert (integer_type_node, value), | |
9499 | build_int_cst (unsigned_type_node, | |
9500 | GOMP_TARGET_ARG_VALUE_SHIFT)); | |
9501 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
9502 | get_target_argument_identifier_1 (device, false, id)); | |
9503 | t = fold_convert (ptr_type_node, t); | |
9504 | return force_gimple_operand_gsi (gsi, t, true, NULL, true, GSI_SAME_STMT); | |
9505 | } | |
9506 | ||
9507 | /* If VALUE is an integer constant greater than -2^15 and smaller than 2^15, | |
9508 | push one argument to ARGS with both the DEVICE, ID and VALUE embedded in it, | |
9509 | otherwise push an identifier (with DEVICE and ID) and the VALUE in two | |
9510 | arguments. */ | |
9511 | ||
9512 | static void | |
9513 | push_target_argument_according_to_value (gimple_stmt_iterator *gsi, int device, | |
9514 | int id, tree value, vec <tree> *args) | |
9515 | { | |
9516 | if (tree_fits_shwi_p (value) | |
9517 | && tree_to_shwi (value) > -(1 << 15) | |
9518 | && tree_to_shwi (value) < (1 << 15)) | |
9519 | args->quick_push (get_target_argument_value (gsi, device, id, value)); | |
9520 | else | |
9521 | { | |
9522 | args->quick_push (get_target_argument_identifier (device, true, id)); | |
9523 | value = fold_convert (ptr_type_node, value); | |
9524 | value = force_gimple_operand_gsi (gsi, value, true, NULL, true, | |
9525 | GSI_SAME_STMT); | |
9526 | args->quick_push (value); | |
9527 | } | |
9528 | } | |
9529 | ||
9530 | /* Create an array of arguments that is then passed to GOMP_target. */ | |
9531 | ||
9532 | static tree | |
9533 | get_target_arguments (gimple_stmt_iterator *gsi, gomp_target *tgt_stmt) | |
9534 | { | |
9535 | auto_vec <tree, 6> args; | |
9536 | tree clauses = gimple_omp_target_clauses (tgt_stmt); | |
9537 | tree t, c = omp_find_clause (clauses, OMP_CLAUSE_NUM_TEAMS); | |
9538 | if (c) | |
9539 | t = OMP_CLAUSE_NUM_TEAMS_EXPR (c); | |
9540 | else | |
9541 | t = integer_minus_one_node; | |
9542 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, | |
9543 | GOMP_TARGET_ARG_NUM_TEAMS, t, &args); | |
9544 | ||
9545 | c = omp_find_clause (clauses, OMP_CLAUSE_THREAD_LIMIT); | |
9546 | if (c) | |
9547 | t = OMP_CLAUSE_THREAD_LIMIT_EXPR (c); | |
9548 | else | |
9549 | t = integer_minus_one_node; | |
9550 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, | |
9551 | GOMP_TARGET_ARG_THREAD_LIMIT, t, | |
9552 | &args); | |
9553 | ||
9554 | /* Produce more, perhaps device specific, arguments here. */ | |
9555 | ||
9556 | tree argarray = create_tmp_var (build_array_type_nelts (ptr_type_node, | |
9557 | args.length () + 1), | |
9558 | ".omp_target_args"); | |
9559 | for (unsigned i = 0; i < args.length (); i++) | |
9560 | { | |
9561 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, | |
9562 | build_int_cst (integer_type_node, i), | |
9563 | NULL_TREE, NULL_TREE); | |
9564 | gsi_insert_before (gsi, gimple_build_assign (ref, args[i]), | |
9565 | GSI_SAME_STMT); | |
9566 | } | |
9567 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, | |
9568 | build_int_cst (integer_type_node, args.length ()), | |
9569 | NULL_TREE, NULL_TREE); | |
9570 | gsi_insert_before (gsi, gimple_build_assign (ref, null_pointer_node), | |
9571 | GSI_SAME_STMT); | |
9572 | TREE_ADDRESSABLE (argarray) = 1; | |
9573 | return build_fold_addr_expr (argarray); | |
9574 | } | |
9575 | ||
9576 | /* Expand the GIMPLE_OMP_TARGET starting at REGION. */ | |
9577 | ||
9578 | static void | |
9579 | expand_omp_target (struct omp_region *region) | |
9580 | { | |
9581 | basic_block entry_bb, exit_bb, new_bb; | |
9582 | struct function *child_cfun; | |
9583 | tree child_fn, block, t; | |
9584 | gimple_stmt_iterator gsi; | |
9585 | gomp_target *entry_stmt; | |
9586 | gimple *stmt; | |
9587 | edge e; | |
9588 | bool offloaded; | |
9589 | int target_kind; | |
9590 | ||
9591 | entry_stmt = as_a <gomp_target *> (last_stmt (region->entry)); | |
9592 | target_kind = gimple_omp_target_kind (entry_stmt); | |
9593 | new_bb = region->entry; | |
9594 | ||
9595 | offloaded = is_gimple_omp_offloaded (entry_stmt); | |
9596 | switch (target_kind) | |
9597 | { | |
9598 | case GF_OMP_TARGET_KIND_REGION: | |
9599 | case GF_OMP_TARGET_KIND_UPDATE: | |
9600 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
9601 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
9602 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
9603 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
9604 | case GF_OMP_TARGET_KIND_OACC_SERIAL: | |
9605 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
9606 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: | |
9607 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
9608 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
9609 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: | |
9610 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
9611 | case GF_OMP_TARGET_KIND_DATA: | |
9612 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
9613 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
9614 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: | |
9615 | break; | |
9616 | default: | |
9617 | gcc_unreachable (); | |
9618 | } | |
9619 | ||
9620 | child_fn = NULL_TREE; | |
9621 | child_cfun = NULL; | |
9622 | if (offloaded) | |
9623 | { | |
9624 | child_fn = gimple_omp_target_child_fn (entry_stmt); | |
9625 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
9626 | } | |
9627 | ||
9628 | /* Supported by expand_omp_taskreg, but not here. */ | |
9629 | if (child_cfun != NULL) | |
9630 | gcc_checking_assert (!child_cfun->cfg); | |
9631 | gcc_checking_assert (!gimple_in_ssa_p (cfun)); | |
9632 | ||
9633 | entry_bb = region->entry; | |
9634 | exit_bb = region->exit; | |
9635 | ||
9636 | if (target_kind == GF_OMP_TARGET_KIND_OACC_KERNELS) | |
9637 | mark_loops_in_oacc_kernels_region (region->entry, region->exit); | |
9638 | ||
9639 | /* Going on, all OpenACC compute constructs are mapped to | |
9640 | 'BUILT_IN_GOACC_PARALLEL', and get their compute regions outlined. | |
9641 | To distinguish between them, we attach attributes. */ | |
9642 | switch (target_kind) | |
9643 | { | |
9644 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
9645 | DECL_ATTRIBUTES (child_fn) | |
9646 | = tree_cons (get_identifier ("oacc parallel"), | |
9647 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9648 | break; | |
9649 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
9650 | DECL_ATTRIBUTES (child_fn) | |
9651 | = tree_cons (get_identifier ("oacc kernels"), | |
9652 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9653 | break; | |
9654 | case GF_OMP_TARGET_KIND_OACC_SERIAL: | |
9655 | DECL_ATTRIBUTES (child_fn) | |
9656 | = tree_cons (get_identifier ("oacc serial"), | |
9657 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9658 | break; | |
9659 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: | |
9660 | DECL_ATTRIBUTES (child_fn) | |
9661 | = tree_cons (get_identifier ("oacc parallel_kernels_parallelized"), | |
9662 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9663 | break; | |
9664 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
9665 | DECL_ATTRIBUTES (child_fn) | |
9666 | = tree_cons (get_identifier ("oacc parallel_kernels_gang_single"), | |
9667 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9668 | break; | |
9669 | default: | |
9670 | /* Make sure we don't miss any. */ | |
9671 | gcc_checking_assert (!(is_gimple_omp_oacc (entry_stmt) | |
9672 | && is_gimple_omp_offloaded (entry_stmt))); | |
9673 | break; | |
9674 | } | |
9675 | ||
9676 | if (offloaded) | |
9677 | { | |
9678 | unsigned srcidx, dstidx, num; | |
9679 | ||
9680 | /* If the offloading region needs data sent from the parent | |
9681 | function, then the very first statement (except possible | |
9682 | tree profile counter updates) of the offloading body | |
9683 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
9684 | &.OMP_DATA_O is passed as an argument to the child function, | |
9685 | we need to replace it with the argument as seen by the child | |
9686 | function. | |
9687 | ||
9688 | In most cases, this will end up being the identity assignment | |
9689 | .OMP_DATA_I = .OMP_DATA_I. However, if the offloading body had | |
9690 | a function call that has been inlined, the original PARM_DECL | |
9691 | .OMP_DATA_I may have been converted into a different local | |
9692 | variable. In which case, we need to keep the assignment. */ | |
9693 | tree data_arg = gimple_omp_target_data_arg (entry_stmt); | |
9694 | if (data_arg) | |
9695 | { | |
9696 | basic_block entry_succ_bb = single_succ (entry_bb); | |
9697 | gimple_stmt_iterator gsi; | |
9698 | tree arg; | |
9699 | gimple *tgtcopy_stmt = NULL; | |
9700 | tree sender = TREE_VEC_ELT (data_arg, 0); | |
9701 | ||
9702 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
9703 | { | |
9704 | gcc_assert (!gsi_end_p (gsi)); | |
9705 | stmt = gsi_stmt (gsi); | |
9706 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
9707 | continue; | |
9708 | ||
9709 | if (gimple_num_ops (stmt) == 2) | |
9710 | { | |
9711 | tree arg = gimple_assign_rhs1 (stmt); | |
9712 | ||
9713 | /* We're ignoring the subcode because we're | |
9714 | effectively doing a STRIP_NOPS. */ | |
9715 | ||
9716 | if (TREE_CODE (arg) == ADDR_EXPR | |
9717 | && TREE_OPERAND (arg, 0) == sender) | |
9718 | { | |
9719 | tgtcopy_stmt = stmt; | |
9720 | break; | |
9721 | } | |
9722 | } | |
9723 | } | |
9724 | ||
9725 | gcc_assert (tgtcopy_stmt != NULL); | |
9726 | arg = DECL_ARGUMENTS (child_fn); | |
9727 | ||
9728 | gcc_assert (gimple_assign_lhs (tgtcopy_stmt) == arg); | |
9729 | gsi_remove (&gsi, true); | |
9730 | } | |
9731 | ||
9732 | /* Declare local variables needed in CHILD_CFUN. */ | |
9733 | block = DECL_INITIAL (child_fn); | |
9734 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
9735 | /* The gimplifier could record temporaries in the offloading block | |
9736 | rather than in containing function's local_decls chain, | |
9737 | which would mean cgraph missed finalizing them. Do it now. */ | |
9738 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
9739 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) | |
9740 | varpool_node::finalize_decl (t); | |
9741 | DECL_SAVED_TREE (child_fn) = NULL; | |
9742 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
9743 | gimple_set_body (child_fn, NULL); | |
9744 | TREE_USED (block) = 1; | |
9745 | ||
9746 | /* Reset DECL_CONTEXT on function arguments. */ | |
9747 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
9748 | DECL_CONTEXT (t) = child_fn; | |
9749 | ||
9750 | /* Split ENTRY_BB at GIMPLE_*, | |
9751 | so that it can be moved to the child function. */ | |
9752 | gsi = gsi_last_nondebug_bb (entry_bb); | |
9753 | stmt = gsi_stmt (gsi); | |
9754 | gcc_assert (stmt | |
9755 | && gimple_code (stmt) == gimple_code (entry_stmt)); | |
9756 | e = split_block (entry_bb, stmt); | |
9757 | gsi_remove (&gsi, true); | |
9758 | entry_bb = e->dest; | |
9759 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
9760 | ||
9761 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ | |
9762 | if (exit_bb) | |
9763 | { | |
9764 | gsi = gsi_last_nondebug_bb (exit_bb); | |
9765 | gcc_assert (!gsi_end_p (gsi) | |
9766 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
9767 | stmt = gimple_build_return (NULL); | |
9768 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
9769 | gsi_remove (&gsi, true); | |
9770 | } | |
9771 | ||
9772 | /* Move the offloading region into CHILD_CFUN. */ | |
9773 | ||
9774 | block = gimple_block (entry_stmt); | |
9775 | ||
9776 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
9777 | if (exit_bb) | |
9778 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
9779 | /* When the OMP expansion process cannot guarantee an up-to-date | |
9780 | loop tree arrange for the child function to fixup loops. */ | |
9781 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
9782 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
9783 | ||
9784 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
9785 | num = vec_safe_length (child_cfun->local_decls); | |
9786 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
9787 | { | |
9788 | t = (*child_cfun->local_decls)[srcidx]; | |
9789 | if (DECL_CONTEXT (t) == cfun->decl) | |
9790 | continue; | |
9791 | if (srcidx != dstidx) | |
9792 | (*child_cfun->local_decls)[dstidx] = t; | |
9793 | dstidx++; | |
9794 | } | |
9795 | if (dstidx != num) | |
9796 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
9797 | ||
9798 | /* Inform the callgraph about the new function. */ | |
9799 | child_cfun->curr_properties = cfun->curr_properties; | |
9800 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; | |
9801 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; | |
9802 | cgraph_node *node = cgraph_node::get_create (child_fn); | |
9803 | node->parallelized_function = 1; | |
9804 | cgraph_node::add_new_function (child_fn, true); | |
9805 | ||
9806 | /* Add the new function to the offload table. */ | |
9807 | if (ENABLE_OFFLOADING) | |
9808 | { | |
9809 | if (in_lto_p) | |
9810 | DECL_PRESERVE_P (child_fn) = 1; | |
9811 | vec_safe_push (offload_funcs, child_fn); | |
9812 | } | |
9813 | ||
9814 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) | |
9815 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); | |
9816 | ||
9817 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
9818 | fixed in a following pass. */ | |
9819 | push_cfun (child_cfun); | |
9820 | if (need_asm) | |
9821 | assign_assembler_name_if_needed (child_fn); | |
9822 | cgraph_edge::rebuild_edges (); | |
9823 | ||
9824 | /* Some EH regions might become dead, see PR34608. If | |
9825 | pass_cleanup_cfg isn't the first pass to happen with the | |
9826 | new child, these dead EH edges might cause problems. | |
9827 | Clean them up now. */ | |
9828 | if (flag_exceptions) | |
9829 | { | |
9830 | basic_block bb; | |
9831 | bool changed = false; | |
9832 | ||
9833 | FOR_EACH_BB_FN (bb, cfun) | |
9834 | changed |= gimple_purge_dead_eh_edges (bb); | |
9835 | if (changed) | |
9836 | cleanup_tree_cfg (); | |
9837 | } | |
9838 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
9839 | verify_loop_structure (); | |
9840 | pop_cfun (); | |
9841 | ||
9842 | if (dump_file && !gimple_in_ssa_p (cfun)) | |
9843 | { | |
9844 | omp_any_child_fn_dumped = true; | |
9845 | dump_function_header (dump_file, child_fn, dump_flags); | |
9846 | dump_function_to_file (child_fn, dump_file, dump_flags); | |
9847 | } | |
9848 | ||
9849 | adjust_context_and_scope (region, gimple_block (entry_stmt), child_fn); | |
9850 | } | |
9851 | ||
9852 | /* Emit a library call to launch the offloading region, or do data | |
9853 | transfers. */ | |
9854 | tree t1, t2, t3, t4, depend, c, clauses; | |
9855 | enum built_in_function start_ix; | |
9856 | unsigned int flags_i = 0; | |
9857 | ||
9858 | switch (gimple_omp_target_kind (entry_stmt)) | |
9859 | { | |
9860 | case GF_OMP_TARGET_KIND_REGION: | |
9861 | start_ix = BUILT_IN_GOMP_TARGET; | |
9862 | break; | |
9863 | case GF_OMP_TARGET_KIND_DATA: | |
9864 | start_ix = BUILT_IN_GOMP_TARGET_DATA; | |
9865 | break; | |
9866 | case GF_OMP_TARGET_KIND_UPDATE: | |
9867 | start_ix = BUILT_IN_GOMP_TARGET_UPDATE; | |
9868 | break; | |
9869 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
9870 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; | |
9871 | break; | |
9872 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
9873 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; | |
9874 | flags_i |= GOMP_TARGET_FLAG_EXIT_DATA; | |
9875 | break; | |
9876 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
9877 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
9878 | case GF_OMP_TARGET_KIND_OACC_SERIAL: | |
9879 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: | |
9880 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
9881 | start_ix = BUILT_IN_GOACC_PARALLEL; | |
9882 | break; | |
9883 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
9884 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
9885 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: | |
9886 | start_ix = BUILT_IN_GOACC_DATA_START; | |
9887 | break; | |
9888 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
9889 | start_ix = BUILT_IN_GOACC_UPDATE; | |
9890 | break; | |
9891 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: | |
9892 | start_ix = BUILT_IN_GOACC_ENTER_DATA; | |
9893 | break; | |
9894 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
9895 | start_ix = BUILT_IN_GOACC_EXIT_DATA; | |
9896 | break; | |
9897 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
9898 | start_ix = BUILT_IN_GOACC_DECLARE; | |
9899 | break; | |
9900 | default: | |
9901 | gcc_unreachable (); | |
9902 | } | |
9903 | ||
9904 | clauses = gimple_omp_target_clauses (entry_stmt); | |
9905 | ||
9906 | tree device = NULL_TREE; | |
9907 | location_t device_loc = UNKNOWN_LOCATION; | |
9908 | tree goacc_flags = NULL_TREE; | |
9909 | if (is_gimple_omp_oacc (entry_stmt)) | |
9910 | { | |
9911 | /* By default, no GOACC_FLAGs are set. */ | |
9912 | goacc_flags = integer_zero_node; | |
9913 | } | |
9914 | else | |
9915 | { | |
9916 | c = omp_find_clause (clauses, OMP_CLAUSE_DEVICE); | |
9917 | if (c) | |
9918 | { | |
9919 | device = OMP_CLAUSE_DEVICE_ID (c); | |
9920 | device_loc = OMP_CLAUSE_LOCATION (c); | |
9921 | if (OMP_CLAUSE_DEVICE_ANCESTOR (c)) | |
9922 | sorry_at (device_loc, "%<ancestor%> not yet supported"); | |
9923 | } | |
9924 | else | |
9925 | { | |
9926 | /* By default, the value of DEVICE is GOMP_DEVICE_ICV (let runtime | |
9927 | library choose). */ | |
9928 | device = build_int_cst (integer_type_node, GOMP_DEVICE_ICV); | |
9929 | device_loc = gimple_location (entry_stmt); | |
9930 | } | |
9931 | ||
9932 | c = omp_find_clause (clauses, OMP_CLAUSE_NOWAIT); | |
9933 | /* FIXME: in_reduction(...) nowait is unimplemented yet, pretend | |
9934 | nowait doesn't appear. */ | |
9935 | if (c && omp_find_clause (clauses, OMP_CLAUSE_IN_REDUCTION)) | |
9936 | c = NULL; | |
9937 | if (c) | |
9938 | flags_i |= GOMP_TARGET_FLAG_NOWAIT; | |
9939 | } | |
9940 | ||
9941 | /* By default, there is no conditional. */ | |
9942 | tree cond = NULL_TREE; | |
9943 | c = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
9944 | if (c) | |
9945 | cond = OMP_CLAUSE_IF_EXPR (c); | |
9946 | /* If we found the clause 'if (cond)', build: | |
9947 | OpenACC: goacc_flags = (cond ? goacc_flags : flags | GOACC_FLAG_HOST_FALLBACK) | |
9948 | OpenMP: device = (cond ? device : GOMP_DEVICE_HOST_FALLBACK) */ | |
9949 | if (cond) | |
9950 | { | |
9951 | tree *tp; | |
9952 | if (is_gimple_omp_oacc (entry_stmt)) | |
9953 | tp = &goacc_flags; | |
9954 | else | |
9955 | { | |
9956 | /* Ensure 'device' is of the correct type. */ | |
9957 | device = fold_convert_loc (device_loc, integer_type_node, device); | |
9958 | ||
9959 | tp = &device; | |
9960 | } | |
9961 | ||
9962 | cond = gimple_boolify (cond); | |
9963 | ||
9964 | basic_block cond_bb, then_bb, else_bb; | |
9965 | edge e; | |
9966 | tree tmp_var; | |
9967 | ||
9968 | tmp_var = create_tmp_var (TREE_TYPE (*tp)); | |
9969 | if (offloaded) | |
9970 | e = split_block_after_labels (new_bb); | |
9971 | else | |
9972 | { | |
9973 | gsi = gsi_last_nondebug_bb (new_bb); | |
9974 | gsi_prev (&gsi); | |
9975 | e = split_block (new_bb, gsi_stmt (gsi)); | |
9976 | } | |
9977 | cond_bb = e->src; | |
9978 | new_bb = e->dest; | |
9979 | remove_edge (e); | |
9980 | ||
9981 | then_bb = create_empty_bb (cond_bb); | |
9982 | else_bb = create_empty_bb (then_bb); | |
9983 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
9984 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
9985 | ||
9986 | stmt = gimple_build_cond_empty (cond); | |
9987 | gsi = gsi_last_bb (cond_bb); | |
9988 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
9989 | ||
9990 | gsi = gsi_start_bb (then_bb); | |
9991 | stmt = gimple_build_assign (tmp_var, *tp); | |
9992 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
9993 | ||
9994 | gsi = gsi_start_bb (else_bb); | |
9995 | if (is_gimple_omp_oacc (entry_stmt)) | |
9996 | stmt = gimple_build_assign (tmp_var, | |
9997 | BIT_IOR_EXPR, | |
9998 | *tp, | |
9999 | build_int_cst (integer_type_node, | |
10000 | GOACC_FLAG_HOST_FALLBACK)); | |
10001 | else | |
10002 | stmt = gimple_build_assign (tmp_var, | |
10003 | build_int_cst (integer_type_node, | |
10004 | GOMP_DEVICE_HOST_FALLBACK)); | |
10005 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
10006 | ||
10007 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
10008 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
10009 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
10010 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
10011 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); | |
10012 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); | |
10013 | ||
10014 | *tp = tmp_var; | |
10015 | ||
10016 | gsi = gsi_last_nondebug_bb (new_bb); | |
10017 | } | |
10018 | else | |
10019 | { | |
10020 | gsi = gsi_last_nondebug_bb (new_bb); | |
10021 | ||
10022 | if (device != NULL_TREE) | |
10023 | device = force_gimple_operand_gsi (&gsi, device, true, NULL_TREE, | |
10024 | true, GSI_SAME_STMT); | |
10025 | } | |
10026 | ||
10027 | t = gimple_omp_target_data_arg (entry_stmt); | |
10028 | if (t == NULL) | |
10029 | { | |
10030 | t1 = size_zero_node; | |
10031 | t2 = build_zero_cst (ptr_type_node); | |
10032 | t3 = t2; | |
10033 | t4 = t2; | |
10034 | } | |
10035 | else | |
10036 | { | |
10037 | t1 = TYPE_MAX_VALUE (TYPE_DOMAIN (TREE_TYPE (TREE_VEC_ELT (t, 1)))); | |
10038 | t1 = size_binop (PLUS_EXPR, t1, size_int (1)); | |
10039 | t2 = build_fold_addr_expr (TREE_VEC_ELT (t, 0)); | |
10040 | t3 = build_fold_addr_expr (TREE_VEC_ELT (t, 1)); | |
10041 | t4 = build_fold_addr_expr (TREE_VEC_ELT (t, 2)); | |
10042 | } | |
10043 | ||
10044 | gimple *g; | |
10045 | bool tagging = false; | |
10046 | /* The maximum number used by any start_ix, without varargs. */ | |
10047 | auto_vec<tree, 11> args; | |
10048 | if (is_gimple_omp_oacc (entry_stmt)) | |
10049 | { | |
10050 | tree goacc_flags_m = fold_build1 (GOACC_FLAGS_MARSHAL_OP, | |
10051 | TREE_TYPE (goacc_flags), goacc_flags); | |
10052 | goacc_flags_m = force_gimple_operand_gsi (&gsi, goacc_flags_m, true, | |
10053 | NULL_TREE, true, | |
10054 | GSI_SAME_STMT); | |
10055 | args.quick_push (goacc_flags_m); | |
10056 | } | |
10057 | else | |
10058 | args.quick_push (device); | |
10059 | if (offloaded) | |
10060 | args.quick_push (build_fold_addr_expr (child_fn)); | |
10061 | args.quick_push (t1); | |
10062 | args.quick_push (t2); | |
10063 | args.quick_push (t3); | |
10064 | args.quick_push (t4); | |
10065 | switch (start_ix) | |
10066 | { | |
10067 | case BUILT_IN_GOACC_DATA_START: | |
10068 | case BUILT_IN_GOACC_DECLARE: | |
10069 | case BUILT_IN_GOMP_TARGET_DATA: | |
10070 | break; | |
10071 | case BUILT_IN_GOMP_TARGET: | |
10072 | case BUILT_IN_GOMP_TARGET_UPDATE: | |
10073 | case BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA: | |
10074 | args.quick_push (build_int_cst (unsigned_type_node, flags_i)); | |
10075 | c = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
10076 | if (c) | |
10077 | depend = OMP_CLAUSE_DECL (c); | |
10078 | else | |
10079 | depend = build_int_cst (ptr_type_node, 0); | |
10080 | args.quick_push (depend); | |
10081 | if (start_ix == BUILT_IN_GOMP_TARGET) | |
10082 | args.quick_push (get_target_arguments (&gsi, entry_stmt)); | |
10083 | break; | |
10084 | case BUILT_IN_GOACC_PARALLEL: | |
10085 | if (lookup_attribute ("oacc serial", DECL_ATTRIBUTES (child_fn)) != NULL) | |
10086 | { | |
10087 | tree dims = NULL_TREE; | |
10088 | unsigned int ix; | |
10089 | ||
10090 | /* For serial constructs we set all dimensions to 1. */ | |
10091 | for (ix = GOMP_DIM_MAX; ix--;) | |
10092 | dims = tree_cons (NULL_TREE, integer_one_node, dims); | |
10093 | oacc_replace_fn_attrib (child_fn, dims); | |
10094 | } | |
10095 | else | |
10096 | oacc_set_fn_attrib (child_fn, clauses, &args); | |
10097 | tagging = true; | |
10098 | /* FALLTHRU */ | |
10099 | case BUILT_IN_GOACC_ENTER_DATA: | |
10100 | case BUILT_IN_GOACC_EXIT_DATA: | |
10101 | case BUILT_IN_GOACC_UPDATE: | |
10102 | { | |
10103 | tree t_async = NULL_TREE; | |
10104 | ||
10105 | /* If present, use the value specified by the respective | |
10106 | clause, making sure that is of the correct type. */ | |
10107 | c = omp_find_clause (clauses, OMP_CLAUSE_ASYNC); | |
10108 | if (c) | |
10109 | t_async = fold_convert_loc (OMP_CLAUSE_LOCATION (c), | |
10110 | integer_type_node, | |
10111 | OMP_CLAUSE_ASYNC_EXPR (c)); | |
10112 | else if (!tagging) | |
10113 | /* Default values for t_async. */ | |
10114 | t_async = fold_convert_loc (gimple_location (entry_stmt), | |
10115 | integer_type_node, | |
10116 | build_int_cst (integer_type_node, | |
10117 | GOMP_ASYNC_SYNC)); | |
10118 | if (tagging && t_async) | |
10119 | { | |
10120 | unsigned HOST_WIDE_INT i_async = GOMP_LAUNCH_OP_MAX; | |
10121 | ||
10122 | if (TREE_CODE (t_async) == INTEGER_CST) | |
10123 | { | |
10124 | /* See if we can pack the async arg in to the tag's | |
10125 | operand. */ | |
10126 | i_async = TREE_INT_CST_LOW (t_async); | |
10127 | if (i_async < GOMP_LAUNCH_OP_MAX) | |
10128 | t_async = NULL_TREE; | |
10129 | else | |
10130 | i_async = GOMP_LAUNCH_OP_MAX; | |
10131 | } | |
10132 | args.safe_push (oacc_launch_pack (GOMP_LAUNCH_ASYNC, NULL_TREE, | |
10133 | i_async)); | |
10134 | } | |
10135 | if (t_async) | |
10136 | args.safe_push (force_gimple_operand_gsi (&gsi, t_async, true, | |
10137 | NULL_TREE, true, | |
10138 | GSI_SAME_STMT)); | |
10139 | ||
10140 | /* Save the argument index, and ... */ | |
10141 | unsigned t_wait_idx = args.length (); | |
10142 | unsigned num_waits = 0; | |
10143 | c = omp_find_clause (clauses, OMP_CLAUSE_WAIT); | |
10144 | if (!tagging || c) | |
10145 | /* ... push a placeholder. */ | |
10146 | args.safe_push (integer_zero_node); | |
10147 | ||
10148 | for (; c; c = OMP_CLAUSE_CHAIN (c)) | |
10149 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_WAIT) | |
10150 | { | |
10151 | tree arg = fold_convert_loc (OMP_CLAUSE_LOCATION (c), | |
10152 | integer_type_node, | |
10153 | OMP_CLAUSE_WAIT_EXPR (c)); | |
10154 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, true, | |
10155 | GSI_SAME_STMT); | |
10156 | args.safe_push (arg); | |
10157 | num_waits++; | |
10158 | } | |
10159 | ||
10160 | if (!tagging || num_waits) | |
10161 | { | |
10162 | tree len; | |
10163 | ||
10164 | /* Now that we know the number, update the placeholder. */ | |
10165 | if (tagging) | |
10166 | len = oacc_launch_pack (GOMP_LAUNCH_WAIT, NULL_TREE, num_waits); | |
10167 | else | |
10168 | len = build_int_cst (integer_type_node, num_waits); | |
10169 | len = fold_convert_loc (gimple_location (entry_stmt), | |
10170 | unsigned_type_node, len); | |
10171 | args[t_wait_idx] = len; | |
10172 | } | |
10173 | } | |
10174 | break; | |
10175 | default: | |
10176 | gcc_unreachable (); | |
10177 | } | |
10178 | if (tagging) | |
10179 | /* Push terminal marker - zero. */ | |
10180 | args.safe_push (oacc_launch_pack (0, NULL_TREE, 0)); | |
10181 | ||
10182 | g = gimple_build_call_vec (builtin_decl_explicit (start_ix), args); | |
10183 | gimple_set_location (g, gimple_location (entry_stmt)); | |
10184 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
10185 | if (!offloaded) | |
10186 | { | |
10187 | g = gsi_stmt (gsi); | |
10188 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_TARGET); | |
10189 | gsi_remove (&gsi, true); | |
10190 | } | |
10191 | } | |
10192 | ||
10193 | /* Expand the parallel region tree rooted at REGION. Expansion | |
10194 | proceeds in depth-first order. Innermost regions are expanded | |
10195 | first. This way, parallel regions that require a new function to | |
10196 | be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any | |
10197 | internal dependencies in their body. */ | |
10198 | ||
10199 | static void | |
10200 | expand_omp (struct omp_region *region) | |
10201 | { | |
10202 | omp_any_child_fn_dumped = false; | |
10203 | while (region) | |
10204 | { | |
10205 | location_t saved_location; | |
10206 | gimple *inner_stmt = NULL; | |
10207 | ||
10208 | /* First, determine whether this is a combined parallel+workshare | |
10209 | region. */ | |
10210 | if (region->type == GIMPLE_OMP_PARALLEL) | |
10211 | determine_parallel_type (region); | |
10212 | ||
10213 | if (region->type == GIMPLE_OMP_FOR | |
10214 | && gimple_omp_for_combined_p (last_stmt (region->entry))) | |
10215 | inner_stmt = last_stmt (region->inner->entry); | |
10216 | ||
10217 | if (region->inner) | |
10218 | expand_omp (region->inner); | |
10219 | ||
10220 | saved_location = input_location; | |
10221 | if (gimple_has_location (last_stmt (region->entry))) | |
10222 | input_location = gimple_location (last_stmt (region->entry)); | |
10223 | ||
10224 | switch (region->type) | |
10225 | { | |
10226 | case GIMPLE_OMP_PARALLEL: | |
10227 | case GIMPLE_OMP_TASK: | |
10228 | expand_omp_taskreg (region); | |
10229 | break; | |
10230 | ||
10231 | case GIMPLE_OMP_FOR: | |
10232 | expand_omp_for (region, inner_stmt); | |
10233 | break; | |
10234 | ||
10235 | case GIMPLE_OMP_SECTIONS: | |
10236 | expand_omp_sections (region); | |
10237 | break; | |
10238 | ||
10239 | case GIMPLE_OMP_SECTION: | |
10240 | /* Individual omp sections are handled together with their | |
10241 | parent GIMPLE_OMP_SECTIONS region. */ | |
10242 | break; | |
10243 | ||
10244 | case GIMPLE_OMP_SINGLE: | |
10245 | case GIMPLE_OMP_SCOPE: | |
10246 | expand_omp_single (region); | |
10247 | break; | |
10248 | ||
10249 | case GIMPLE_OMP_ORDERED: | |
10250 | { | |
10251 | gomp_ordered *ord_stmt | |
10252 | = as_a <gomp_ordered *> (last_stmt (region->entry)); | |
10253 | if (omp_find_clause (gimple_omp_ordered_clauses (ord_stmt), | |
10254 | OMP_CLAUSE_DEPEND)) | |
10255 | { | |
10256 | /* We'll expand these when expanding corresponding | |
10257 | worksharing region with ordered(n) clause. */ | |
10258 | gcc_assert (region->outer | |
10259 | && region->outer->type == GIMPLE_OMP_FOR); | |
10260 | region->ord_stmt = ord_stmt; | |
10261 | break; | |
10262 | } | |
10263 | } | |
10264 | /* FALLTHRU */ | |
10265 | case GIMPLE_OMP_MASTER: | |
10266 | case GIMPLE_OMP_MASKED: | |
10267 | case GIMPLE_OMP_TASKGROUP: | |
10268 | case GIMPLE_OMP_CRITICAL: | |
10269 | case GIMPLE_OMP_TEAMS: | |
10270 | expand_omp_synch (region); | |
10271 | break; | |
10272 | ||
10273 | case GIMPLE_OMP_ATOMIC_LOAD: | |
10274 | expand_omp_atomic (region); | |
10275 | break; | |
10276 | ||
10277 | case GIMPLE_OMP_TARGET: | |
10278 | expand_omp_target (region); | |
10279 | break; | |
10280 | ||
10281 | default: | |
10282 | gcc_unreachable (); | |
10283 | } | |
10284 | ||
10285 | input_location = saved_location; | |
10286 | region = region->next; | |
10287 | } | |
10288 | if (omp_any_child_fn_dumped) | |
10289 | { | |
10290 | if (dump_file) | |
10291 | dump_function_header (dump_file, current_function_decl, dump_flags); | |
10292 | omp_any_child_fn_dumped = false; | |
10293 | } | |
10294 | } | |
10295 | ||
10296 | /* Helper for build_omp_regions. Scan the dominator tree starting at | |
10297 | block BB. PARENT is the region that contains BB. If SINGLE_TREE is | |
10298 | true, the function ends once a single tree is built (otherwise, whole | |
10299 | forest of OMP constructs may be built). */ | |
10300 | ||
10301 | static void | |
10302 | build_omp_regions_1 (basic_block bb, struct omp_region *parent, | |
10303 | bool single_tree) | |
10304 | { | |
10305 | gimple_stmt_iterator gsi; | |
10306 | gimple *stmt; | |
10307 | basic_block son; | |
10308 | ||
10309 | gsi = gsi_last_nondebug_bb (bb); | |
10310 | if (!gsi_end_p (gsi) && is_gimple_omp (gsi_stmt (gsi))) | |
10311 | { | |
10312 | struct omp_region *region; | |
10313 | enum gimple_code code; | |
10314 | ||
10315 | stmt = gsi_stmt (gsi); | |
10316 | code = gimple_code (stmt); | |
10317 | if (code == GIMPLE_OMP_RETURN) | |
10318 | { | |
10319 | /* STMT is the return point out of region PARENT. Mark it | |
10320 | as the exit point and make PARENT the immediately | |
10321 | enclosing region. */ | |
10322 | gcc_assert (parent); | |
10323 | region = parent; | |
10324 | region->exit = bb; | |
10325 | parent = parent->outer; | |
10326 | } | |
10327 | else if (code == GIMPLE_OMP_ATOMIC_STORE) | |
10328 | { | |
10329 | /* GIMPLE_OMP_ATOMIC_STORE is analogous to | |
10330 | GIMPLE_OMP_RETURN, but matches with | |
10331 | GIMPLE_OMP_ATOMIC_LOAD. */ | |
10332 | gcc_assert (parent); | |
10333 | gcc_assert (parent->type == GIMPLE_OMP_ATOMIC_LOAD); | |
10334 | region = parent; | |
10335 | region->exit = bb; | |
10336 | parent = parent->outer; | |
10337 | } | |
10338 | else if (code == GIMPLE_OMP_CONTINUE) | |
10339 | { | |
10340 | gcc_assert (parent); | |
10341 | parent->cont = bb; | |
10342 | } | |
10343 | else if (code == GIMPLE_OMP_SECTIONS_SWITCH) | |
10344 | { | |
10345 | /* GIMPLE_OMP_SECTIONS_SWITCH is part of | |
10346 | GIMPLE_OMP_SECTIONS, and we do nothing for it. */ | |
10347 | } | |
10348 | else | |
10349 | { | |
10350 | region = new_omp_region (bb, code, parent); | |
10351 | /* Otherwise... */ | |
10352 | if (code == GIMPLE_OMP_TARGET) | |
10353 | { | |
10354 | switch (gimple_omp_target_kind (stmt)) | |
10355 | { | |
10356 | case GF_OMP_TARGET_KIND_REGION: | |
10357 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
10358 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
10359 | case GF_OMP_TARGET_KIND_OACC_SERIAL: | |
10360 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: | |
10361 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
10362 | break; | |
10363 | case GF_OMP_TARGET_KIND_UPDATE: | |
10364 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
10365 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
10366 | case GF_OMP_TARGET_KIND_DATA: | |
10367 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
10368 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
10369 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: | |
10370 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
10371 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: | |
10372 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
10373 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
10374 | /* ..., other than for those stand-alone directives... */ | |
10375 | region = NULL; | |
10376 | break; | |
10377 | default: | |
10378 | gcc_unreachable (); | |
10379 | } | |
10380 | } | |
10381 | else if (code == GIMPLE_OMP_ORDERED | |
10382 | && omp_find_clause (gimple_omp_ordered_clauses | |
10383 | (as_a <gomp_ordered *> (stmt)), | |
10384 | OMP_CLAUSE_DEPEND)) | |
10385 | /* #pragma omp ordered depend is also just a stand-alone | |
10386 | directive. */ | |
10387 | region = NULL; | |
10388 | else if (code == GIMPLE_OMP_TASK | |
10389 | && gimple_omp_task_taskwait_p (stmt)) | |
10390 | /* #pragma omp taskwait depend(...) is a stand-alone directive. */ | |
10391 | region = NULL; | |
10392 | /* ..., this directive becomes the parent for a new region. */ | |
10393 | if (region) | |
10394 | parent = region; | |
10395 | } | |
10396 | } | |
10397 | ||
10398 | if (single_tree && !parent) | |
10399 | return; | |
10400 | ||
10401 | for (son = first_dom_son (CDI_DOMINATORS, bb); | |
10402 | son; | |
10403 | son = next_dom_son (CDI_DOMINATORS, son)) | |
10404 | build_omp_regions_1 (son, parent, single_tree); | |
10405 | } | |
10406 | ||
10407 | /* Builds the tree of OMP regions rooted at ROOT, storing it to | |
10408 | root_omp_region. */ | |
10409 | ||
10410 | static void | |
10411 | build_omp_regions_root (basic_block root) | |
10412 | { | |
10413 | gcc_assert (root_omp_region == NULL); | |
10414 | build_omp_regions_1 (root, NULL, true); | |
10415 | gcc_assert (root_omp_region != NULL); | |
10416 | } | |
10417 | ||
10418 | /* Expands omp construct (and its subconstructs) starting in HEAD. */ | |
10419 | ||
10420 | void | |
10421 | omp_expand_local (basic_block head) | |
10422 | { | |
10423 | build_omp_regions_root (head); | |
10424 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
10425 | { | |
10426 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
10427 | dump_omp_region (dump_file, root_omp_region, 0); | |
10428 | fprintf (dump_file, "\n"); | |
10429 | } | |
10430 | ||
10431 | remove_exit_barriers (root_omp_region); | |
10432 | expand_omp (root_omp_region); | |
10433 | ||
10434 | omp_free_regions (); | |
10435 | } | |
10436 | ||
10437 | /* Scan the CFG and build a tree of OMP regions. Return the root of | |
10438 | the OMP region tree. */ | |
10439 | ||
10440 | static void | |
10441 | build_omp_regions (void) | |
10442 | { | |
10443 | gcc_assert (root_omp_region == NULL); | |
10444 | calculate_dominance_info (CDI_DOMINATORS); | |
10445 | build_omp_regions_1 (ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, false); | |
10446 | } | |
10447 | ||
10448 | /* Main entry point for expanding OMP-GIMPLE into runtime calls. */ | |
10449 | ||
10450 | static unsigned int | |
10451 | execute_expand_omp (void) | |
10452 | { | |
10453 | build_omp_regions (); | |
10454 | ||
10455 | if (!root_omp_region) | |
10456 | return 0; | |
10457 | ||
10458 | if (dump_file) | |
10459 | { | |
10460 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
10461 | dump_omp_region (dump_file, root_omp_region, 0); | |
10462 | fprintf (dump_file, "\n"); | |
10463 | } | |
10464 | ||
10465 | remove_exit_barriers (root_omp_region); | |
10466 | ||
10467 | expand_omp (root_omp_region); | |
10468 | ||
10469 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
10470 | verify_loop_structure (); | |
10471 | cleanup_tree_cfg (); | |
10472 | ||
10473 | omp_free_regions (); | |
10474 | ||
10475 | return 0; | |
10476 | } | |
10477 | ||
10478 | /* OMP expansion -- the default pass, run before creation of SSA form. */ | |
10479 | ||
10480 | namespace { | |
10481 | ||
10482 | const pass_data pass_data_expand_omp = | |
10483 | { | |
10484 | GIMPLE_PASS, /* type */ | |
10485 | "ompexp", /* name */ | |
10486 | OPTGROUP_OMP, /* optinfo_flags */ | |
10487 | TV_NONE, /* tv_id */ | |
10488 | PROP_gimple_any, /* properties_required */ | |
10489 | PROP_gimple_eomp, /* properties_provided */ | |
10490 | 0, /* properties_destroyed */ | |
10491 | 0, /* todo_flags_start */ | |
10492 | 0, /* todo_flags_finish */ | |
10493 | }; | |
10494 | ||
10495 | class pass_expand_omp : public gimple_opt_pass | |
10496 | { | |
10497 | public: | |
10498 | pass_expand_omp (gcc::context *ctxt) | |
10499 | : gimple_opt_pass (pass_data_expand_omp, ctxt) | |
10500 | {} | |
10501 | ||
10502 | /* opt_pass methods: */ | |
10503 | virtual unsigned int execute (function *) | |
10504 | { | |
10505 | bool gate = ((flag_openacc != 0 || flag_openmp != 0 | |
10506 | || flag_openmp_simd != 0) | |
10507 | && !seen_error ()); | |
10508 | ||
10509 | /* This pass always runs, to provide PROP_gimple_eomp. | |
10510 | But often, there is nothing to do. */ | |
10511 | if (!gate) | |
10512 | return 0; | |
10513 | ||
10514 | return execute_expand_omp (); | |
10515 | } | |
10516 | ||
10517 | }; // class pass_expand_omp | |
10518 | ||
10519 | } // anon namespace | |
10520 | ||
10521 | gimple_opt_pass * | |
10522 | make_pass_expand_omp (gcc::context *ctxt) | |
10523 | { | |
10524 | return new pass_expand_omp (ctxt); | |
10525 | } | |
10526 | ||
10527 | namespace { | |
10528 | ||
10529 | const pass_data pass_data_expand_omp_ssa = | |
10530 | { | |
10531 | GIMPLE_PASS, /* type */ | |
10532 | "ompexpssa", /* name */ | |
10533 | OPTGROUP_OMP, /* optinfo_flags */ | |
10534 | TV_NONE, /* tv_id */ | |
10535 | PROP_cfg | PROP_ssa, /* properties_required */ | |
10536 | PROP_gimple_eomp, /* properties_provided */ | |
10537 | 0, /* properties_destroyed */ | |
10538 | 0, /* todo_flags_start */ | |
10539 | TODO_cleanup_cfg | TODO_rebuild_alias, /* todo_flags_finish */ | |
10540 | }; | |
10541 | ||
10542 | class pass_expand_omp_ssa : public gimple_opt_pass | |
10543 | { | |
10544 | public: | |
10545 | pass_expand_omp_ssa (gcc::context *ctxt) | |
10546 | : gimple_opt_pass (pass_data_expand_omp_ssa, ctxt) | |
10547 | {} | |
10548 | ||
10549 | /* opt_pass methods: */ | |
10550 | virtual bool gate (function *fun) | |
10551 | { | |
10552 | return !(fun->curr_properties & PROP_gimple_eomp); | |
10553 | } | |
10554 | virtual unsigned int execute (function *) { return execute_expand_omp (); } | |
10555 | opt_pass * clone () { return new pass_expand_omp_ssa (m_ctxt); } | |
10556 | ||
10557 | }; // class pass_expand_omp_ssa | |
10558 | ||
10559 | } // anon namespace | |
10560 | ||
10561 | gimple_opt_pass * | |
10562 | make_pass_expand_omp_ssa (gcc::context *ctxt) | |
10563 | { | |
10564 | return new pass_expand_omp_ssa (ctxt); | |
10565 | } | |
10566 | ||
10567 | /* Called from tree-cfg.c::make_edges to create cfg edges for all relevant | |
10568 | GIMPLE_* codes. */ | |
10569 | ||
10570 | bool | |
10571 | omp_make_gimple_edges (basic_block bb, struct omp_region **region, | |
10572 | int *region_idx) | |
10573 | { | |
10574 | gimple *last = last_stmt (bb); | |
10575 | enum gimple_code code = gimple_code (last); | |
10576 | struct omp_region *cur_region = *region; | |
10577 | bool fallthru = false; | |
10578 | ||
10579 | switch (code) | |
10580 | { | |
10581 | case GIMPLE_OMP_PARALLEL: | |
10582 | case GIMPLE_OMP_FOR: | |
10583 | case GIMPLE_OMP_SINGLE: | |
10584 | case GIMPLE_OMP_TEAMS: | |
10585 | case GIMPLE_OMP_MASTER: | |
10586 | case GIMPLE_OMP_MASKED: | |
10587 | case GIMPLE_OMP_SCOPE: | |
10588 | case GIMPLE_OMP_TASKGROUP: | |
10589 | case GIMPLE_OMP_CRITICAL: | |
10590 | case GIMPLE_OMP_SECTION: | |
10591 | cur_region = new_omp_region (bb, code, cur_region); | |
10592 | fallthru = true; | |
10593 | break; | |
10594 | ||
10595 | case GIMPLE_OMP_TASK: | |
10596 | cur_region = new_omp_region (bb, code, cur_region); | |
10597 | fallthru = true; | |
10598 | if (gimple_omp_task_taskwait_p (last)) | |
10599 | cur_region = cur_region->outer; | |
10600 | break; | |
10601 | ||
10602 | case GIMPLE_OMP_ORDERED: | |
10603 | cur_region = new_omp_region (bb, code, cur_region); | |
10604 | fallthru = true; | |
10605 | if (omp_find_clause (gimple_omp_ordered_clauses | |
10606 | (as_a <gomp_ordered *> (last)), | |
10607 | OMP_CLAUSE_DEPEND)) | |
10608 | cur_region = cur_region->outer; | |
10609 | break; | |
10610 | ||
10611 | case GIMPLE_OMP_TARGET: | |
10612 | cur_region = new_omp_region (bb, code, cur_region); | |
10613 | fallthru = true; | |
10614 | switch (gimple_omp_target_kind (last)) | |
10615 | { | |
10616 | case GF_OMP_TARGET_KIND_REGION: | |
10617 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
10618 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
10619 | case GF_OMP_TARGET_KIND_OACC_SERIAL: | |
10620 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: | |
10621 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
10622 | break; | |
10623 | case GF_OMP_TARGET_KIND_UPDATE: | |
10624 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
10625 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
10626 | case GF_OMP_TARGET_KIND_DATA: | |
10627 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
10628 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
10629 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: | |
10630 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
10631 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: | |
10632 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
10633 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
10634 | cur_region = cur_region->outer; | |
10635 | break; | |
10636 | default: | |
10637 | gcc_unreachable (); | |
10638 | } | |
10639 | break; | |
10640 | ||
10641 | case GIMPLE_OMP_SECTIONS: | |
10642 | cur_region = new_omp_region (bb, code, cur_region); | |
10643 | fallthru = true; | |
10644 | break; | |
10645 | ||
10646 | case GIMPLE_OMP_SECTIONS_SWITCH: | |
10647 | fallthru = false; | |
10648 | break; | |
10649 | ||
10650 | case GIMPLE_OMP_ATOMIC_LOAD: | |
10651 | case GIMPLE_OMP_ATOMIC_STORE: | |
10652 | fallthru = true; | |
10653 | break; | |
10654 | ||
10655 | case GIMPLE_OMP_RETURN: | |
10656 | /* In the case of a GIMPLE_OMP_SECTION, the edge will go | |
10657 | somewhere other than the next block. This will be | |
10658 | created later. */ | |
10659 | cur_region->exit = bb; | |
10660 | if (cur_region->type == GIMPLE_OMP_TASK) | |
10661 | /* Add an edge corresponding to not scheduling the task | |
10662 | immediately. */ | |
10663 | make_edge (cur_region->entry, bb, EDGE_ABNORMAL); | |
10664 | fallthru = cur_region->type != GIMPLE_OMP_SECTION; | |
10665 | cur_region = cur_region->outer; | |
10666 | break; | |
10667 | ||
10668 | case GIMPLE_OMP_CONTINUE: | |
10669 | cur_region->cont = bb; | |
10670 | switch (cur_region->type) | |
10671 | { | |
10672 | case GIMPLE_OMP_FOR: | |
10673 | /* Mark all GIMPLE_OMP_FOR and GIMPLE_OMP_CONTINUE | |
10674 | succs edges as abnormal to prevent splitting | |
10675 | them. */ | |
10676 | single_succ_edge (cur_region->entry)->flags |= EDGE_ABNORMAL; | |
10677 | /* Make the loopback edge. */ | |
10678 | make_edge (bb, single_succ (cur_region->entry), | |
10679 | EDGE_ABNORMAL); | |
10680 | ||
10681 | /* Create an edge from GIMPLE_OMP_FOR to exit, which | |
10682 | corresponds to the case that the body of the loop | |
10683 | is not executed at all. */ | |
10684 | make_edge (cur_region->entry, bb->next_bb, EDGE_ABNORMAL); | |
10685 | make_edge (bb, bb->next_bb, EDGE_FALLTHRU | EDGE_ABNORMAL); | |
10686 | fallthru = false; | |
10687 | break; | |
10688 | ||
10689 | case GIMPLE_OMP_SECTIONS: | |
10690 | /* Wire up the edges into and out of the nested sections. */ | |
10691 | { | |
10692 | basic_block switch_bb = single_succ (cur_region->entry); | |
10693 | ||
10694 | struct omp_region *i; | |
10695 | for (i = cur_region->inner; i ; i = i->next) | |
10696 | { | |
10697 | gcc_assert (i->type == GIMPLE_OMP_SECTION); | |
10698 | make_edge (switch_bb, i->entry, 0); | |
10699 | make_edge (i->exit, bb, EDGE_FALLTHRU); | |
10700 | } | |
10701 | ||
10702 | /* Make the loopback edge to the block with | |
10703 | GIMPLE_OMP_SECTIONS_SWITCH. */ | |
10704 | make_edge (bb, switch_bb, 0); | |
10705 | ||
10706 | /* Make the edge from the switch to exit. */ | |
10707 | make_edge (switch_bb, bb->next_bb, 0); | |
10708 | fallthru = false; | |
10709 | } | |
10710 | break; | |
10711 | ||
10712 | case GIMPLE_OMP_TASK: | |
10713 | fallthru = true; | |
10714 | break; | |
10715 | ||
10716 | default: | |
10717 | gcc_unreachable (); | |
10718 | } | |
10719 | break; | |
10720 | ||
10721 | default: | |
10722 | gcc_unreachable (); | |
10723 | } | |
10724 | ||
10725 | if (*region != cur_region) | |
10726 | { | |
10727 | *region = cur_region; | |
10728 | if (cur_region) | |
10729 | *region_idx = cur_region->entry->index; | |
10730 | else | |
10731 | *region_idx = 0; | |
10732 | } | |
10733 | ||
10734 | return fallthru; | |
10735 | } |