]>
Commit | Line | Data |
---|---|---|
629b3d75 MJ |
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 | ||
a945c346 | 5 | Copyright (C) 2005-2024 Free Software Foundation, Inc. |
629b3d75 MJ |
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" | |
a895e6d7 | 55 | #include "alloc-pool.h" |
629b3d75 | 56 | #include "symbol-summary.h" |
629b3d75 MJ |
57 | #include "gomp-constants.h" |
58 | #include "gimple-pretty-print.h" | |
314e6352 ML |
59 | #include "stringpool.h" |
60 | #include "attribs.h" | |
87d6dae3 | 61 | #include "tree-eh.h" |
00f34291 | 62 | #include "opts.h" |
629b3d75 MJ |
63 | |
64 | /* OMP region information. Every parallel and workshare | |
65 | directive is enclosed between two markers, the OMP_* directive | |
66 | and a corresponding GIMPLE_OMP_RETURN statement. */ | |
67 | ||
68 | struct omp_region | |
69 | { | |
70 | /* The enclosing region. */ | |
71 | struct omp_region *outer; | |
72 | ||
73 | /* First child region. */ | |
74 | struct omp_region *inner; | |
75 | ||
76 | /* Next peer region. */ | |
77 | struct omp_region *next; | |
78 | ||
79 | /* Block containing the omp directive as its last stmt. */ | |
80 | basic_block entry; | |
81 | ||
82 | /* Block containing the GIMPLE_OMP_RETURN as its last stmt. */ | |
83 | basic_block exit; | |
84 | ||
85 | /* Block containing the GIMPLE_OMP_CONTINUE as its last stmt. */ | |
86 | basic_block cont; | |
87 | ||
88 | /* If this is a combined parallel+workshare region, this is a list | |
89 | of additional arguments needed by the combined parallel+workshare | |
90 | library call. */ | |
91 | vec<tree, va_gc> *ws_args; | |
92 | ||
93 | /* The code for the omp directive of this region. */ | |
94 | enum gimple_code type; | |
95 | ||
96 | /* Schedule kind, only used for GIMPLE_OMP_FOR type regions. */ | |
97 | enum omp_clause_schedule_kind sched_kind; | |
98 | ||
99 | /* Schedule modifiers. */ | |
100 | unsigned char sched_modifiers; | |
101 | ||
102 | /* True if this is a combined parallel+workshare region. */ | |
103 | bool is_combined_parallel; | |
104 | ||
0b887b75 JJ |
105 | /* Copy of fd.lastprivate_conditional != 0. */ |
106 | bool has_lastprivate_conditional; | |
107 | ||
629b3d75 MJ |
108 | /* The ordered stmt if type is GIMPLE_OMP_ORDERED and it has |
109 | a depend clause. */ | |
110 | gomp_ordered *ord_stmt; | |
111 | }; | |
112 | ||
113 | static struct omp_region *root_omp_region; | |
114 | static bool omp_any_child_fn_dumped; | |
115 | ||
116 | static void expand_omp_build_assign (gimple_stmt_iterator *, tree, tree, | |
117 | bool = false); | |
118 | static gphi *find_phi_with_arg_on_edge (tree, edge); | |
119 | static void expand_omp (struct omp_region *region); | |
120 | ||
121 | /* Return true if REGION is a combined parallel+workshare region. */ | |
122 | ||
123 | static inline bool | |
124 | is_combined_parallel (struct omp_region *region) | |
125 | { | |
126 | return region->is_combined_parallel; | |
127 | } | |
128 | ||
129 | /* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB | |
130 | is the immediate dominator of PAR_ENTRY_BB, return true if there | |
131 | are no data dependencies that would prevent expanding the parallel | |
132 | directive at PAR_ENTRY_BB as a combined parallel+workshare region. | |
133 | ||
134 | When expanding a combined parallel+workshare region, the call to | |
135 | the child function may need additional arguments in the case of | |
136 | GIMPLE_OMP_FOR regions. In some cases, these arguments are | |
137 | computed out of variables passed in from the parent to the child | |
138 | via 'struct .omp_data_s'. For instance: | |
139 | ||
140 | #pragma omp parallel for schedule (guided, i * 4) | |
141 | for (j ...) | |
142 | ||
143 | Is lowered into: | |
144 | ||
01914336 | 145 | # BLOCK 2 (PAR_ENTRY_BB) |
629b3d75 MJ |
146 | .omp_data_o.i = i; |
147 | #pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598) | |
148 | ||
149 | # BLOCK 3 (WS_ENTRY_BB) | |
150 | .omp_data_i = &.omp_data_o; | |
151 | D.1667 = .omp_data_i->i; | |
152 | D.1598 = D.1667 * 4; | |
153 | #pragma omp for schedule (guided, D.1598) | |
154 | ||
155 | When we outline the parallel region, the call to the child function | |
156 | 'bar.omp_fn.0' will need the value D.1598 in its argument list, but | |
157 | that value is computed *after* the call site. So, in principle we | |
158 | cannot do the transformation. | |
159 | ||
160 | To see whether the code in WS_ENTRY_BB blocks the combined | |
161 | parallel+workshare call, we collect all the variables used in the | |
162 | GIMPLE_OMP_FOR header check whether they appear on the LHS of any | |
163 | statement in WS_ENTRY_BB. If so, then we cannot emit the combined | |
164 | call. | |
165 | ||
166 | FIXME. If we had the SSA form built at this point, we could merely | |
167 | hoist the code in block 3 into block 2 and be done with it. But at | |
168 | this point we don't have dataflow information and though we could | |
169 | hack something up here, it is really not worth the aggravation. */ | |
170 | ||
171 | static bool | |
172 | workshare_safe_to_combine_p (basic_block ws_entry_bb) | |
173 | { | |
174 | struct omp_for_data fd; | |
fe8ac82f | 175 | gimple *ws_stmt = last_nondebug_stmt (ws_entry_bb); |
629b3d75 MJ |
176 | |
177 | if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) | |
178 | return true; | |
179 | ||
180 | gcc_assert (gimple_code (ws_stmt) == GIMPLE_OMP_FOR); | |
28567c40 JJ |
181 | if (gimple_omp_for_kind (ws_stmt) != GF_OMP_FOR_KIND_FOR) |
182 | return false; | |
629b3d75 MJ |
183 | |
184 | omp_extract_for_data (as_a <gomp_for *> (ws_stmt), &fd, NULL); | |
185 | ||
186 | if (fd.collapse > 1 && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
187 | return false; | |
188 | if (fd.iter_type != long_integer_type_node) | |
189 | return false; | |
190 | ||
191 | /* FIXME. We give up too easily here. If any of these arguments | |
192 | are not constants, they will likely involve variables that have | |
193 | been mapped into fields of .omp_data_s for sharing with the child | |
194 | function. With appropriate data flow, it would be possible to | |
195 | see through this. */ | |
196 | if (!is_gimple_min_invariant (fd.loop.n1) | |
197 | || !is_gimple_min_invariant (fd.loop.n2) | |
198 | || !is_gimple_min_invariant (fd.loop.step) | |
199 | || (fd.chunk_size && !is_gimple_min_invariant (fd.chunk_size))) | |
200 | return false; | |
201 | ||
202 | return true; | |
203 | } | |
204 | ||
205 | /* Adjust CHUNK_SIZE from SCHEDULE clause, depending on simd modifier | |
206 | presence (SIMD_SCHEDULE). */ | |
207 | ||
208 | static tree | |
209 | omp_adjust_chunk_size (tree chunk_size, bool simd_schedule) | |
210 | { | |
28567c40 | 211 | if (!simd_schedule || integer_zerop (chunk_size)) |
629b3d75 MJ |
212 | return chunk_size; |
213 | ||
9d2f08ab RS |
214 | poly_uint64 vf = omp_max_vf (); |
215 | if (known_eq (vf, 1U)) | |
629b3d75 MJ |
216 | return chunk_size; |
217 | ||
218 | tree type = TREE_TYPE (chunk_size); | |
219 | chunk_size = fold_build2 (PLUS_EXPR, type, chunk_size, | |
220 | build_int_cst (type, vf - 1)); | |
221 | return fold_build2 (BIT_AND_EXPR, type, chunk_size, | |
222 | build_int_cst (type, -vf)); | |
223 | } | |
224 | ||
225 | /* Collect additional arguments needed to emit a combined | |
226 | parallel+workshare call. WS_STMT is the workshare directive being | |
227 | expanded. */ | |
228 | ||
229 | static vec<tree, va_gc> * | |
230 | get_ws_args_for (gimple *par_stmt, gimple *ws_stmt) | |
231 | { | |
232 | tree t; | |
233 | location_t loc = gimple_location (ws_stmt); | |
234 | vec<tree, va_gc> *ws_args; | |
235 | ||
236 | if (gomp_for *for_stmt = dyn_cast <gomp_for *> (ws_stmt)) | |
237 | { | |
238 | struct omp_for_data fd; | |
239 | tree n1, n2; | |
240 | ||
241 | omp_extract_for_data (for_stmt, &fd, NULL); | |
242 | n1 = fd.loop.n1; | |
243 | n2 = fd.loop.n2; | |
244 | ||
245 | if (gimple_omp_for_combined_into_p (for_stmt)) | |
246 | { | |
247 | tree innerc | |
248 | = omp_find_clause (gimple_omp_parallel_clauses (par_stmt), | |
249 | OMP_CLAUSE__LOOPTEMP_); | |
250 | gcc_assert (innerc); | |
251 | n1 = OMP_CLAUSE_DECL (innerc); | |
252 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
253 | OMP_CLAUSE__LOOPTEMP_); | |
254 | gcc_assert (innerc); | |
255 | n2 = OMP_CLAUSE_DECL (innerc); | |
256 | } | |
257 | ||
258 | vec_alloc (ws_args, 3 + (fd.chunk_size != 0)); | |
259 | ||
260 | t = fold_convert_loc (loc, long_integer_type_node, n1); | |
261 | ws_args->quick_push (t); | |
262 | ||
263 | t = fold_convert_loc (loc, long_integer_type_node, n2); | |
264 | ws_args->quick_push (t); | |
265 | ||
266 | t = fold_convert_loc (loc, long_integer_type_node, fd.loop.step); | |
267 | ws_args->quick_push (t); | |
268 | ||
269 | if (fd.chunk_size) | |
270 | { | |
271 | t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size); | |
272 | t = omp_adjust_chunk_size (t, fd.simd_schedule); | |
273 | ws_args->quick_push (t); | |
274 | } | |
275 | ||
276 | return ws_args; | |
277 | } | |
278 | else if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) | |
279 | { | |
280 | /* Number of sections is equal to the number of edges from the | |
281 | GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to | |
282 | the exit of the sections region. */ | |
283 | basic_block bb = single_succ (gimple_bb (ws_stmt)); | |
284 | t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1); | |
285 | vec_alloc (ws_args, 1); | |
286 | ws_args->quick_push (t); | |
287 | return ws_args; | |
288 | } | |
289 | ||
290 | gcc_unreachable (); | |
291 | } | |
292 | ||
293 | /* Discover whether REGION is a combined parallel+workshare region. */ | |
294 | ||
295 | static void | |
296 | determine_parallel_type (struct omp_region *region) | |
297 | { | |
298 | basic_block par_entry_bb, par_exit_bb; | |
299 | basic_block ws_entry_bb, ws_exit_bb; | |
300 | ||
301 | if (region == NULL || region->inner == NULL | |
302 | || region->exit == NULL || region->inner->exit == NULL | |
303 | || region->inner->cont == NULL) | |
304 | return; | |
305 | ||
306 | /* We only support parallel+for and parallel+sections. */ | |
307 | if (region->type != GIMPLE_OMP_PARALLEL | |
308 | || (region->inner->type != GIMPLE_OMP_FOR | |
309 | && region->inner->type != GIMPLE_OMP_SECTIONS)) | |
310 | return; | |
311 | ||
312 | /* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and | |
313 | WS_EXIT_BB -> PAR_EXIT_BB. */ | |
314 | par_entry_bb = region->entry; | |
315 | par_exit_bb = region->exit; | |
316 | ws_entry_bb = region->inner->entry; | |
317 | ws_exit_bb = region->inner->exit; | |
318 | ||
28567c40 JJ |
319 | /* Give up for task reductions on the parallel, while it is implementable, |
320 | adding another big set of APIs or slowing down the normal paths is | |
321 | not acceptable. */ | |
fe8ac82f RB |
322 | tree pclauses |
323 | = gimple_omp_parallel_clauses (last_nondebug_stmt (par_entry_bb)); | |
28567c40 JJ |
324 | if (omp_find_clause (pclauses, OMP_CLAUSE__REDUCTEMP_)) |
325 | return; | |
326 | ||
629b3d75 MJ |
327 | if (single_succ (par_entry_bb) == ws_entry_bb |
328 | && single_succ (ws_exit_bb) == par_exit_bb | |
329 | && workshare_safe_to_combine_p (ws_entry_bb) | |
fe8ac82f | 330 | && (gimple_omp_parallel_combined_p (last_nondebug_stmt (par_entry_bb)) |
629b3d75 MJ |
331 | || (last_and_only_stmt (ws_entry_bb) |
332 | && last_and_only_stmt (par_exit_bb)))) | |
333 | { | |
fe8ac82f RB |
334 | gimple *par_stmt = last_nondebug_stmt (par_entry_bb); |
335 | gimple *ws_stmt = last_nondebug_stmt (ws_entry_bb); | |
629b3d75 MJ |
336 | |
337 | if (region->inner->type == GIMPLE_OMP_FOR) | |
338 | { | |
339 | /* If this is a combined parallel loop, we need to determine | |
340 | whether or not to use the combined library calls. There | |
341 | are two cases where we do not apply the transformation: | |
342 | static loops and any kind of ordered loop. In the first | |
343 | case, we already open code the loop so there is no need | |
344 | to do anything else. In the latter case, the combined | |
345 | parallel loop call would still need extra synchronization | |
346 | to implement ordered semantics, so there would not be any | |
347 | gain in using the combined call. */ | |
348 | tree clauses = gimple_omp_for_clauses (ws_stmt); | |
349 | tree c = omp_find_clause (clauses, OMP_CLAUSE_SCHEDULE); | |
350 | if (c == NULL | |
351 | || ((OMP_CLAUSE_SCHEDULE_KIND (c) & OMP_CLAUSE_SCHEDULE_MASK) | |
352 | == OMP_CLAUSE_SCHEDULE_STATIC) | |
28567c40 | 353 | || omp_find_clause (clauses, OMP_CLAUSE_ORDERED) |
6c7ae8c5 | 354 | || omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_) |
8221c30b JJ |
355 | || ((c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_)) |
356 | && POINTER_TYPE_P (TREE_TYPE (OMP_CLAUSE_DECL (c))))) | |
28567c40 | 357 | return; |
629b3d75 | 358 | } |
28567c40 | 359 | else if (region->inner->type == GIMPLE_OMP_SECTIONS |
8221c30b JJ |
360 | && (omp_find_clause (gimple_omp_sections_clauses (ws_stmt), |
361 | OMP_CLAUSE__REDUCTEMP_) | |
362 | || omp_find_clause (gimple_omp_sections_clauses (ws_stmt), | |
363 | OMP_CLAUSE__CONDTEMP_))) | |
28567c40 | 364 | return; |
629b3d75 MJ |
365 | |
366 | region->is_combined_parallel = true; | |
367 | region->inner->is_combined_parallel = true; | |
368 | region->ws_args = get_ws_args_for (par_stmt, ws_stmt); | |
369 | } | |
370 | } | |
371 | ||
372 | /* Debugging dumps for parallel regions. */ | |
373 | void dump_omp_region (FILE *, struct omp_region *, int); | |
374 | void debug_omp_region (struct omp_region *); | |
375 | void debug_all_omp_regions (void); | |
376 | ||
377 | /* Dump the parallel region tree rooted at REGION. */ | |
378 | ||
379 | void | |
380 | dump_omp_region (FILE *file, struct omp_region *region, int indent) | |
381 | { | |
382 | fprintf (file, "%*sbb %d: %s\n", indent, "", region->entry->index, | |
383 | gimple_code_name[region->type]); | |
384 | ||
385 | if (region->inner) | |
386 | dump_omp_region (file, region->inner, indent + 4); | |
387 | ||
388 | if (region->cont) | |
389 | { | |
390 | fprintf (file, "%*sbb %d: GIMPLE_OMP_CONTINUE\n", indent, "", | |
391 | region->cont->index); | |
392 | } | |
393 | ||
394 | if (region->exit) | |
395 | fprintf (file, "%*sbb %d: GIMPLE_OMP_RETURN\n", indent, "", | |
396 | region->exit->index); | |
397 | else | |
398 | fprintf (file, "%*s[no exit marker]\n", indent, ""); | |
399 | ||
400 | if (region->next) | |
401 | dump_omp_region (file, region->next, indent); | |
402 | } | |
403 | ||
404 | DEBUG_FUNCTION void | |
405 | debug_omp_region (struct omp_region *region) | |
406 | { | |
407 | dump_omp_region (stderr, region, 0); | |
408 | } | |
409 | ||
410 | DEBUG_FUNCTION void | |
411 | debug_all_omp_regions (void) | |
412 | { | |
413 | dump_omp_region (stderr, root_omp_region, 0); | |
414 | } | |
415 | ||
416 | /* Create a new parallel region starting at STMT inside region PARENT. */ | |
417 | ||
418 | static struct omp_region * | |
419 | new_omp_region (basic_block bb, enum gimple_code type, | |
420 | struct omp_region *parent) | |
421 | { | |
422 | struct omp_region *region = XCNEW (struct omp_region); | |
423 | ||
424 | region->outer = parent; | |
425 | region->entry = bb; | |
426 | region->type = type; | |
427 | ||
428 | if (parent) | |
429 | { | |
430 | /* This is a nested region. Add it to the list of inner | |
431 | regions in PARENT. */ | |
432 | region->next = parent->inner; | |
433 | parent->inner = region; | |
434 | } | |
435 | else | |
436 | { | |
437 | /* This is a toplevel region. Add it to the list of toplevel | |
438 | regions in ROOT_OMP_REGION. */ | |
439 | region->next = root_omp_region; | |
440 | root_omp_region = region; | |
441 | } | |
442 | ||
443 | return region; | |
444 | } | |
445 | ||
446 | /* Release the memory associated with the region tree rooted at REGION. */ | |
447 | ||
448 | static void | |
449 | free_omp_region_1 (struct omp_region *region) | |
450 | { | |
451 | struct omp_region *i, *n; | |
452 | ||
453 | for (i = region->inner; i ; i = n) | |
454 | { | |
455 | n = i->next; | |
456 | free_omp_region_1 (i); | |
457 | } | |
458 | ||
459 | free (region); | |
460 | } | |
461 | ||
462 | /* Release the memory for the entire omp region tree. */ | |
463 | ||
464 | void | |
465 | omp_free_regions (void) | |
466 | { | |
467 | struct omp_region *r, *n; | |
468 | for (r = root_omp_region; r ; r = n) | |
469 | { | |
470 | n = r->next; | |
471 | free_omp_region_1 (r); | |
472 | } | |
473 | root_omp_region = NULL; | |
474 | } | |
475 | ||
476 | /* A convenience function to build an empty GIMPLE_COND with just the | |
477 | condition. */ | |
478 | ||
479 | static gcond * | |
480 | gimple_build_cond_empty (tree cond) | |
481 | { | |
482 | enum tree_code pred_code; | |
483 | tree lhs, rhs; | |
484 | ||
485 | gimple_cond_get_ops_from_tree (cond, &pred_code, &lhs, &rhs); | |
486 | return gimple_build_cond (pred_code, lhs, rhs, NULL_TREE, NULL_TREE); | |
487 | } | |
488 | ||
623c6df5 KB |
489 | /* Change DECL_CONTEXT of CHILD_FNDECL to that of the parent function. |
490 | Add CHILD_FNDECL to decl chain of the supercontext of the block | |
491 | ENTRY_BLOCK - this is the block which originally contained the | |
492 | code from which CHILD_FNDECL was created. | |
493 | ||
494 | Together, these actions ensure that the debug info for the outlined | |
495 | function will be emitted with the correct lexical scope. */ | |
496 | ||
497 | static void | |
4ccc4e30 JJ |
498 | adjust_context_and_scope (struct omp_region *region, tree entry_block, |
499 | tree child_fndecl) | |
623c6df5 | 500 | { |
4ccc4e30 JJ |
501 | tree parent_fndecl = NULL_TREE; |
502 | gimple *entry_stmt; | |
503 | /* OMP expansion expands inner regions before outer ones, so if | |
504 | we e.g. have explicit task region nested in parallel region, when | |
505 | expanding the task region current_function_decl will be the original | |
506 | source function, but we actually want to use as context the child | |
507 | function of the parallel. */ | |
508 | for (region = region->outer; | |
509 | region && parent_fndecl == NULL_TREE; region = region->outer) | |
510 | switch (region->type) | |
511 | { | |
512 | case GIMPLE_OMP_PARALLEL: | |
513 | case GIMPLE_OMP_TASK: | |
514 | case GIMPLE_OMP_TEAMS: | |
fe8ac82f | 515 | entry_stmt = last_nondebug_stmt (region->entry); |
4ccc4e30 JJ |
516 | parent_fndecl = gimple_omp_taskreg_child_fn (entry_stmt); |
517 | break; | |
518 | case GIMPLE_OMP_TARGET: | |
fe8ac82f | 519 | entry_stmt = last_nondebug_stmt (region->entry); |
4ccc4e30 JJ |
520 | parent_fndecl |
521 | = gimple_omp_target_child_fn (as_a <gomp_target *> (entry_stmt)); | |
522 | break; | |
523 | default: | |
524 | break; | |
525 | } | |
526 | ||
527 | if (parent_fndecl == NULL_TREE) | |
528 | parent_fndecl = current_function_decl; | |
529 | DECL_CONTEXT (child_fndecl) = parent_fndecl; | |
530 | ||
623c6df5 KB |
531 | if (entry_block != NULL_TREE && TREE_CODE (entry_block) == BLOCK) |
532 | { | |
533 | tree b = BLOCK_SUPERCONTEXT (entry_block); | |
623c6df5 KB |
534 | if (TREE_CODE (b) == BLOCK) |
535 | { | |
623c6df5 KB |
536 | DECL_CHAIN (child_fndecl) = BLOCK_VARS (b); |
537 | BLOCK_VARS (b) = child_fndecl; | |
538 | } | |
539 | } | |
540 | } | |
541 | ||
28567c40 | 542 | /* Build the function calls to GOMP_parallel etc to actually |
629b3d75 MJ |
543 | generate the parallel operation. REGION is the parallel region |
544 | being expanded. BB is the block where to insert the code. WS_ARGS | |
545 | will be set if this is a call to a combined parallel+workshare | |
546 | construct, it contains the list of additional arguments needed by | |
547 | the workshare construct. */ | |
548 | ||
549 | static void | |
550 | expand_parallel_call (struct omp_region *region, basic_block bb, | |
551 | gomp_parallel *entry_stmt, | |
552 | vec<tree, va_gc> *ws_args) | |
553 | { | |
554 | tree t, t1, t2, val, cond, c, clauses, flags; | |
555 | gimple_stmt_iterator gsi; | |
556 | gimple *stmt; | |
557 | enum built_in_function start_ix; | |
558 | int start_ix2; | |
559 | location_t clause_loc; | |
560 | vec<tree, va_gc> *args; | |
561 | ||
562 | clauses = gimple_omp_parallel_clauses (entry_stmt); | |
563 | ||
564 | /* Determine what flavor of GOMP_parallel we will be | |
565 | emitting. */ | |
566 | start_ix = BUILT_IN_GOMP_PARALLEL; | |
28567c40 JJ |
567 | tree rtmp = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); |
568 | if (rtmp) | |
569 | start_ix = BUILT_IN_GOMP_PARALLEL_REDUCTIONS; | |
570 | else if (is_combined_parallel (region)) | |
629b3d75 MJ |
571 | { |
572 | switch (region->inner->type) | |
573 | { | |
574 | case GIMPLE_OMP_FOR: | |
575 | gcc_assert (region->inner->sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); | |
576 | switch (region->inner->sched_kind) | |
577 | { | |
578 | case OMP_CLAUSE_SCHEDULE_RUNTIME: | |
0b887b75 JJ |
579 | /* For lastprivate(conditional:), our implementation |
580 | requires monotonic behavior. */ | |
581 | if (region->inner->has_lastprivate_conditional != 0) | |
582 | start_ix2 = 3; | |
583 | else if ((region->inner->sched_modifiers | |
584 | & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0) | |
28567c40 JJ |
585 | start_ix2 = 6; |
586 | else if ((region->inner->sched_modifiers | |
587 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0) | |
588 | start_ix2 = 7; | |
589 | else | |
590 | start_ix2 = 3; | |
629b3d75 MJ |
591 | break; |
592 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: | |
593 | case OMP_CLAUSE_SCHEDULE_GUIDED: | |
28567c40 | 594 | if ((region->inner->sched_modifiers |
0b887b75 JJ |
595 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 |
596 | && !region->inner->has_lastprivate_conditional) | |
629b3d75 MJ |
597 | { |
598 | start_ix2 = 3 + region->inner->sched_kind; | |
599 | break; | |
600 | } | |
601 | /* FALLTHRU */ | |
602 | default: | |
603 | start_ix2 = region->inner->sched_kind; | |
604 | break; | |
605 | } | |
606 | start_ix2 += (int) BUILT_IN_GOMP_PARALLEL_LOOP_STATIC; | |
607 | start_ix = (enum built_in_function) start_ix2; | |
608 | break; | |
609 | case GIMPLE_OMP_SECTIONS: | |
610 | start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS; | |
611 | break; | |
612 | default: | |
613 | gcc_unreachable (); | |
614 | } | |
615 | } | |
616 | ||
617 | /* By default, the value of NUM_THREADS is zero (selected at run time) | |
618 | and there is no conditional. */ | |
619 | cond = NULL_TREE; | |
620 | val = build_int_cst (unsigned_type_node, 0); | |
621 | flags = build_int_cst (unsigned_type_node, 0); | |
622 | ||
623 | c = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
624 | if (c) | |
625 | cond = OMP_CLAUSE_IF_EXPR (c); | |
626 | ||
627 | c = omp_find_clause (clauses, OMP_CLAUSE_NUM_THREADS); | |
628 | if (c) | |
629 | { | |
630 | val = OMP_CLAUSE_NUM_THREADS_EXPR (c); | |
631 | clause_loc = OMP_CLAUSE_LOCATION (c); | |
632 | } | |
633 | else | |
634 | clause_loc = gimple_location (entry_stmt); | |
635 | ||
636 | c = omp_find_clause (clauses, OMP_CLAUSE_PROC_BIND); | |
637 | if (c) | |
638 | flags = build_int_cst (unsigned_type_node, OMP_CLAUSE_PROC_BIND_KIND (c)); | |
639 | ||
640 | /* Ensure 'val' is of the correct type. */ | |
641 | val = fold_convert_loc (clause_loc, unsigned_type_node, val); | |
642 | ||
643 | /* If we found the clause 'if (cond)', build either | |
644 | (cond != 0) or (cond ? val : 1u). */ | |
645 | if (cond) | |
646 | { | |
647 | cond = gimple_boolify (cond); | |
648 | ||
649 | if (integer_zerop (val)) | |
650 | val = fold_build2_loc (clause_loc, | |
651 | EQ_EXPR, unsigned_type_node, cond, | |
652 | build_int_cst (TREE_TYPE (cond), 0)); | |
653 | else | |
654 | { | |
655 | basic_block cond_bb, then_bb, else_bb; | |
656 | edge e, e_then, e_else; | |
657 | tree tmp_then, tmp_else, tmp_join, tmp_var; | |
658 | ||
659 | tmp_var = create_tmp_var (TREE_TYPE (val)); | |
660 | if (gimple_in_ssa_p (cfun)) | |
661 | { | |
662 | tmp_then = make_ssa_name (tmp_var); | |
663 | tmp_else = make_ssa_name (tmp_var); | |
664 | tmp_join = make_ssa_name (tmp_var); | |
665 | } | |
666 | else | |
667 | { | |
668 | tmp_then = tmp_var; | |
669 | tmp_else = tmp_var; | |
670 | tmp_join = tmp_var; | |
671 | } | |
672 | ||
673 | e = split_block_after_labels (bb); | |
674 | cond_bb = e->src; | |
675 | bb = e->dest; | |
676 | remove_edge (e); | |
677 | ||
678 | then_bb = create_empty_bb (cond_bb); | |
679 | else_bb = create_empty_bb (then_bb); | |
680 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
681 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
682 | ||
683 | stmt = gimple_build_cond_empty (cond); | |
684 | gsi = gsi_start_bb (cond_bb); | |
685 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
686 | ||
687 | gsi = gsi_start_bb (then_bb); | |
688 | expand_omp_build_assign (&gsi, tmp_then, val, true); | |
689 | ||
690 | gsi = gsi_start_bb (else_bb); | |
691 | expand_omp_build_assign (&gsi, tmp_else, | |
692 | build_int_cst (unsigned_type_node, 1), | |
693 | true); | |
694 | ||
695 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
696 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
697 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
698 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
699 | e_then = make_edge (then_bb, bb, EDGE_FALLTHRU); | |
700 | e_else = make_edge (else_bb, bb, EDGE_FALLTHRU); | |
701 | ||
702 | if (gimple_in_ssa_p (cfun)) | |
703 | { | |
704 | gphi *phi = create_phi_node (tmp_join, bb); | |
705 | add_phi_arg (phi, tmp_then, e_then, UNKNOWN_LOCATION); | |
706 | add_phi_arg (phi, tmp_else, e_else, UNKNOWN_LOCATION); | |
707 | } | |
708 | ||
709 | val = tmp_join; | |
710 | } | |
711 | ||
712 | gsi = gsi_start_bb (bb); | |
713 | val = force_gimple_operand_gsi (&gsi, val, true, NULL_TREE, | |
714 | false, GSI_CONTINUE_LINKING); | |
715 | } | |
716 | ||
65f4b875 | 717 | gsi = gsi_last_nondebug_bb (bb); |
629b3d75 MJ |
718 | t = gimple_omp_parallel_data_arg (entry_stmt); |
719 | if (t == NULL) | |
720 | t1 = null_pointer_node; | |
721 | else | |
722 | t1 = build_fold_addr_expr (t); | |
723 | tree child_fndecl = gimple_omp_parallel_child_fn (entry_stmt); | |
724 | t2 = build_fold_addr_expr (child_fndecl); | |
725 | ||
726 | vec_alloc (args, 4 + vec_safe_length (ws_args)); | |
727 | args->quick_push (t2); | |
728 | args->quick_push (t1); | |
729 | args->quick_push (val); | |
730 | if (ws_args) | |
731 | args->splice (*ws_args); | |
732 | args->quick_push (flags); | |
733 | ||
734 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
735 | builtin_decl_explicit (start_ix), args); | |
736 | ||
28567c40 JJ |
737 | if (rtmp) |
738 | { | |
739 | tree type = TREE_TYPE (OMP_CLAUSE_DECL (rtmp)); | |
740 | t = build2 (MODIFY_EXPR, type, OMP_CLAUSE_DECL (rtmp), | |
741 | fold_convert (type, | |
742 | fold_convert (pointer_sized_int_node, t))); | |
743 | } | |
629b3d75 MJ |
744 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
745 | false, GSI_CONTINUE_LINKING); | |
629b3d75 MJ |
746 | } |
747 | ||
629b3d75 MJ |
748 | /* Build the function call to GOMP_task to actually |
749 | generate the task operation. BB is the block where to insert the code. */ | |
750 | ||
751 | static void | |
752 | expand_task_call (struct omp_region *region, basic_block bb, | |
753 | gomp_task *entry_stmt) | |
754 | { | |
755 | tree t1, t2, t3; | |
756 | gimple_stmt_iterator gsi; | |
757 | location_t loc = gimple_location (entry_stmt); | |
758 | ||
759 | tree clauses = gimple_omp_task_clauses (entry_stmt); | |
760 | ||
761 | tree ifc = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
762 | tree untied = omp_find_clause (clauses, OMP_CLAUSE_UNTIED); | |
763 | tree mergeable = omp_find_clause (clauses, OMP_CLAUSE_MERGEABLE); | |
764 | tree depend = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
765 | tree finalc = omp_find_clause (clauses, OMP_CLAUSE_FINAL); | |
766 | tree priority = omp_find_clause (clauses, OMP_CLAUSE_PRIORITY); | |
a6d22fb2 | 767 | tree detach = omp_find_clause (clauses, OMP_CLAUSE_DETACH); |
629b3d75 MJ |
768 | |
769 | unsigned int iflags | |
770 | = (untied ? GOMP_TASK_FLAG_UNTIED : 0) | |
771 | | (mergeable ? GOMP_TASK_FLAG_MERGEABLE : 0) | |
772 | | (depend ? GOMP_TASK_FLAG_DEPEND : 0); | |
773 | ||
774 | bool taskloop_p = gimple_omp_task_taskloop_p (entry_stmt); | |
775 | tree startvar = NULL_TREE, endvar = NULL_TREE, step = NULL_TREE; | |
776 | tree num_tasks = NULL_TREE; | |
777 | bool ull = false; | |
778 | if (taskloop_p) | |
779 | { | |
fe8ac82f | 780 | gimple *g = last_nondebug_stmt (region->outer->entry); |
629b3d75 MJ |
781 | gcc_assert (gimple_code (g) == GIMPLE_OMP_FOR |
782 | && gimple_omp_for_kind (g) == GF_OMP_FOR_KIND_TASKLOOP); | |
783 | struct omp_for_data fd; | |
784 | omp_extract_for_data (as_a <gomp_for *> (g), &fd, NULL); | |
785 | startvar = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
786 | endvar = omp_find_clause (OMP_CLAUSE_CHAIN (startvar), | |
787 | OMP_CLAUSE__LOOPTEMP_); | |
788 | startvar = OMP_CLAUSE_DECL (startvar); | |
789 | endvar = OMP_CLAUSE_DECL (endvar); | |
790 | step = fold_convert_loc (loc, fd.iter_type, fd.loop.step); | |
791 | if (fd.loop.cond_code == LT_EXPR) | |
792 | iflags |= GOMP_TASK_FLAG_UP; | |
793 | tree tclauses = gimple_omp_for_clauses (g); | |
794 | num_tasks = omp_find_clause (tclauses, OMP_CLAUSE_NUM_TASKS); | |
795 | if (num_tasks) | |
3bc75533 JJ |
796 | { |
797 | if (OMP_CLAUSE_NUM_TASKS_STRICT (num_tasks)) | |
798 | iflags |= GOMP_TASK_FLAG_STRICT; | |
799 | num_tasks = OMP_CLAUSE_NUM_TASKS_EXPR (num_tasks); | |
800 | } | |
629b3d75 MJ |
801 | else |
802 | { | |
803 | num_tasks = omp_find_clause (tclauses, OMP_CLAUSE_GRAINSIZE); | |
804 | if (num_tasks) | |
805 | { | |
806 | iflags |= GOMP_TASK_FLAG_GRAINSIZE; | |
3bc75533 JJ |
807 | if (OMP_CLAUSE_GRAINSIZE_STRICT (num_tasks)) |
808 | iflags |= GOMP_TASK_FLAG_STRICT; | |
629b3d75 MJ |
809 | num_tasks = OMP_CLAUSE_GRAINSIZE_EXPR (num_tasks); |
810 | } | |
811 | else | |
812 | num_tasks = integer_zero_node; | |
813 | } | |
814 | num_tasks = fold_convert_loc (loc, long_integer_type_node, num_tasks); | |
815 | if (ifc == NULL_TREE) | |
816 | iflags |= GOMP_TASK_FLAG_IF; | |
817 | if (omp_find_clause (tclauses, OMP_CLAUSE_NOGROUP)) | |
818 | iflags |= GOMP_TASK_FLAG_NOGROUP; | |
819 | ull = fd.iter_type == long_long_unsigned_type_node; | |
28567c40 JJ |
820 | if (omp_find_clause (clauses, OMP_CLAUSE_REDUCTION)) |
821 | iflags |= GOMP_TASK_FLAG_REDUCTION; | |
629b3d75 | 822 | } |
a6d22fb2 KCY |
823 | else |
824 | { | |
825 | if (priority) | |
826 | iflags |= GOMP_TASK_FLAG_PRIORITY; | |
827 | if (detach) | |
828 | iflags |= GOMP_TASK_FLAG_DETACH; | |
829 | } | |
629b3d75 MJ |
830 | |
831 | tree flags = build_int_cst (unsigned_type_node, iflags); | |
832 | ||
833 | tree cond = boolean_true_node; | |
834 | if (ifc) | |
835 | { | |
836 | if (taskloop_p) | |
837 | { | |
838 | tree t = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); | |
839 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, | |
840 | build_int_cst (unsigned_type_node, | |
841 | GOMP_TASK_FLAG_IF), | |
842 | build_int_cst (unsigned_type_node, 0)); | |
843 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, | |
844 | flags, t); | |
845 | } | |
846 | else | |
847 | cond = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); | |
848 | } | |
849 | ||
850 | if (finalc) | |
851 | { | |
852 | tree t = gimple_boolify (OMP_CLAUSE_FINAL_EXPR (finalc)); | |
853 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, | |
854 | build_int_cst (unsigned_type_node, | |
855 | GOMP_TASK_FLAG_FINAL), | |
856 | build_int_cst (unsigned_type_node, 0)); | |
857 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, flags, t); | |
858 | } | |
859 | if (depend) | |
860 | depend = OMP_CLAUSE_DECL (depend); | |
861 | else | |
862 | depend = build_int_cst (ptr_type_node, 0); | |
863 | if (priority) | |
864 | priority = fold_convert (integer_type_node, | |
865 | OMP_CLAUSE_PRIORITY_EXPR (priority)); | |
866 | else | |
867 | priority = integer_zero_node; | |
868 | ||
65f4b875 | 869 | gsi = gsi_last_nondebug_bb (bb); |
a6d22fb2 KCY |
870 | |
871 | detach = (detach | |
872 | ? build_fold_addr_expr (OMP_CLAUSE_DECL (detach)) | |
873 | : null_pointer_node); | |
874 | ||
629b3d75 MJ |
875 | tree t = gimple_omp_task_data_arg (entry_stmt); |
876 | if (t == NULL) | |
877 | t2 = null_pointer_node; | |
878 | else | |
879 | t2 = build_fold_addr_expr_loc (loc, t); | |
880 | t1 = build_fold_addr_expr_loc (loc, gimple_omp_task_child_fn (entry_stmt)); | |
881 | t = gimple_omp_task_copy_fn (entry_stmt); | |
882 | if (t == NULL) | |
883 | t3 = null_pointer_node; | |
884 | else | |
885 | t3 = build_fold_addr_expr_loc (loc, t); | |
886 | ||
887 | if (taskloop_p) | |
888 | t = build_call_expr (ull | |
889 | ? builtin_decl_explicit (BUILT_IN_GOMP_TASKLOOP_ULL) | |
890 | : builtin_decl_explicit (BUILT_IN_GOMP_TASKLOOP), | |
891 | 11, t1, t2, t3, | |
892 | gimple_omp_task_arg_size (entry_stmt), | |
893 | gimple_omp_task_arg_align (entry_stmt), flags, | |
894 | num_tasks, priority, startvar, endvar, step); | |
895 | else | |
896 | t = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASK), | |
a6d22fb2 | 897 | 10, t1, t2, t3, |
629b3d75 MJ |
898 | gimple_omp_task_arg_size (entry_stmt), |
899 | gimple_omp_task_arg_align (entry_stmt), cond, flags, | |
a6d22fb2 | 900 | depend, priority, detach); |
629b3d75 MJ |
901 | |
902 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
903 | false, GSI_CONTINUE_LINKING); | |
904 | } | |
905 | ||
28567c40 JJ |
906 | /* Build the function call to GOMP_taskwait_depend to actually |
907 | generate the taskwait operation. BB is the block where to insert the | |
908 | code. */ | |
909 | ||
910 | static void | |
911 | expand_taskwait_call (basic_block bb, gomp_task *entry_stmt) | |
912 | { | |
913 | tree clauses = gimple_omp_task_clauses (entry_stmt); | |
914 | tree depend = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
915 | if (depend == NULL_TREE) | |
916 | return; | |
917 | ||
918 | depend = OMP_CLAUSE_DECL (depend); | |
919 | ||
b4383691 | 920 | bool nowait = omp_find_clause (clauses, OMP_CLAUSE_NOWAIT) != NULL_TREE; |
28567c40 | 921 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); |
b4383691 JJ |
922 | enum built_in_function f = (nowait |
923 | ? BUILT_IN_GOMP_TASKWAIT_DEPEND_NOWAIT | |
924 | : BUILT_IN_GOMP_TASKWAIT_DEPEND); | |
925 | tree t = build_call_expr (builtin_decl_explicit (f), 1, depend); | |
28567c40 JJ |
926 | |
927 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
928 | false, GSI_CONTINUE_LINKING); | |
929 | } | |
930 | ||
931 | /* Build the function call to GOMP_teams_reg to actually | |
932 | generate the host teams operation. REGION is the teams region | |
933 | being expanded. BB is the block where to insert the code. */ | |
934 | ||
935 | static void | |
936 | expand_teams_call (basic_block bb, gomp_teams *entry_stmt) | |
937 | { | |
938 | tree clauses = gimple_omp_teams_clauses (entry_stmt); | |
939 | tree num_teams = omp_find_clause (clauses, OMP_CLAUSE_NUM_TEAMS); | |
940 | if (num_teams == NULL_TREE) | |
941 | num_teams = build_int_cst (unsigned_type_node, 0); | |
942 | else | |
943 | { | |
48d7327f | 944 | num_teams = OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR (num_teams); |
28567c40 JJ |
945 | num_teams = fold_convert (unsigned_type_node, num_teams); |
946 | } | |
947 | tree thread_limit = omp_find_clause (clauses, OMP_CLAUSE_THREAD_LIMIT); | |
948 | if (thread_limit == NULL_TREE) | |
949 | thread_limit = build_int_cst (unsigned_type_node, 0); | |
950 | else | |
951 | { | |
952 | thread_limit = OMP_CLAUSE_THREAD_LIMIT_EXPR (thread_limit); | |
953 | thread_limit = fold_convert (unsigned_type_node, thread_limit); | |
954 | } | |
955 | ||
956 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); | |
957 | tree t = gimple_omp_teams_data_arg (entry_stmt), t1; | |
958 | if (t == NULL) | |
959 | t1 = null_pointer_node; | |
960 | else | |
961 | t1 = build_fold_addr_expr (t); | |
962 | tree child_fndecl = gimple_omp_teams_child_fn (entry_stmt); | |
963 | tree t2 = build_fold_addr_expr (child_fndecl); | |
964 | ||
28567c40 JJ |
965 | vec<tree, va_gc> *args; |
966 | vec_alloc (args, 5); | |
967 | args->quick_push (t2); | |
968 | args->quick_push (t1); | |
969 | args->quick_push (num_teams); | |
970 | args->quick_push (thread_limit); | |
971 | /* For future extensibility. */ | |
972 | args->quick_push (build_zero_cst (unsigned_type_node)); | |
973 | ||
974 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
975 | builtin_decl_explicit (BUILT_IN_GOMP_TEAMS_REG), | |
976 | args); | |
977 | ||
978 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
979 | false, GSI_CONTINUE_LINKING); | |
980 | } | |
981 | ||
629b3d75 MJ |
982 | /* Chain all the DECLs in LIST by their TREE_CHAIN fields. */ |
983 | ||
984 | static tree | |
985 | vec2chain (vec<tree, va_gc> *v) | |
986 | { | |
987 | tree chain = NULL_TREE, t; | |
988 | unsigned ix; | |
989 | ||
990 | FOR_EACH_VEC_SAFE_ELT_REVERSE (v, ix, t) | |
991 | { | |
992 | DECL_CHAIN (t) = chain; | |
993 | chain = t; | |
994 | } | |
995 | ||
996 | return chain; | |
997 | } | |
998 | ||
999 | /* Remove barriers in REGION->EXIT's block. Note that this is only | |
1000 | valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region | |
1001 | is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that | |
1002 | left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be | |
1003 | removed. */ | |
1004 | ||
1005 | static void | |
1006 | remove_exit_barrier (struct omp_region *region) | |
1007 | { | |
1008 | gimple_stmt_iterator gsi; | |
1009 | basic_block exit_bb; | |
1010 | edge_iterator ei; | |
1011 | edge e; | |
1012 | gimple *stmt; | |
1013 | int any_addressable_vars = -1; | |
1014 | ||
1015 | exit_bb = region->exit; | |
1016 | ||
1017 | /* If the parallel region doesn't return, we don't have REGION->EXIT | |
1018 | block at all. */ | |
1019 | if (! exit_bb) | |
1020 | return; | |
1021 | ||
1022 | /* The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The | |
1023 | workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of | |
1024 | statements that can appear in between are extremely limited -- no | |
1025 | memory operations at all. Here, we allow nothing at all, so the | |
1026 | only thing we allow to precede this GIMPLE_OMP_RETURN is a label. */ | |
65f4b875 | 1027 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 | 1028 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
65f4b875 | 1029 | gsi_prev_nondebug (&gsi); |
629b3d75 MJ |
1030 | if (!gsi_end_p (gsi) && gimple_code (gsi_stmt (gsi)) != GIMPLE_LABEL) |
1031 | return; | |
1032 | ||
1033 | FOR_EACH_EDGE (e, ei, exit_bb->preds) | |
1034 | { | |
65f4b875 | 1035 | gsi = gsi_last_nondebug_bb (e->src); |
629b3d75 MJ |
1036 | if (gsi_end_p (gsi)) |
1037 | continue; | |
1038 | stmt = gsi_stmt (gsi); | |
1039 | if (gimple_code (stmt) == GIMPLE_OMP_RETURN | |
1040 | && !gimple_omp_return_nowait_p (stmt)) | |
1041 | { | |
1042 | /* OpenMP 3.0 tasks unfortunately prevent this optimization | |
1043 | in many cases. If there could be tasks queued, the barrier | |
1044 | might be needed to let the tasks run before some local | |
1045 | variable of the parallel that the task uses as shared | |
1046 | runs out of scope. The task can be spawned either | |
1047 | from within current function (this would be easy to check) | |
1048 | or from some function it calls and gets passed an address | |
1049 | of such a variable. */ | |
1050 | if (any_addressable_vars < 0) | |
1051 | { | |
1052 | gomp_parallel *parallel_stmt | |
fe8ac82f | 1053 | = as_a <gomp_parallel *> (last_nondebug_stmt (region->entry)); |
629b3d75 MJ |
1054 | tree child_fun = gimple_omp_parallel_child_fn (parallel_stmt); |
1055 | tree local_decls, block, decl; | |
1056 | unsigned ix; | |
1057 | ||
1058 | any_addressable_vars = 0; | |
1059 | FOR_EACH_LOCAL_DECL (DECL_STRUCT_FUNCTION (child_fun), ix, decl) | |
1060 | if (TREE_ADDRESSABLE (decl)) | |
1061 | { | |
1062 | any_addressable_vars = 1; | |
1063 | break; | |
1064 | } | |
1065 | for (block = gimple_block (stmt); | |
1066 | !any_addressable_vars | |
1067 | && block | |
1068 | && TREE_CODE (block) == BLOCK; | |
1069 | block = BLOCK_SUPERCONTEXT (block)) | |
1070 | { | |
1071 | for (local_decls = BLOCK_VARS (block); | |
1072 | local_decls; | |
1073 | local_decls = DECL_CHAIN (local_decls)) | |
1074 | if (TREE_ADDRESSABLE (local_decls)) | |
1075 | { | |
1076 | any_addressable_vars = 1; | |
1077 | break; | |
1078 | } | |
1079 | if (block == gimple_block (parallel_stmt)) | |
1080 | break; | |
1081 | } | |
1082 | } | |
1083 | if (!any_addressable_vars) | |
1084 | gimple_omp_return_set_nowait (stmt); | |
1085 | } | |
1086 | } | |
1087 | } | |
1088 | ||
1089 | static void | |
1090 | remove_exit_barriers (struct omp_region *region) | |
1091 | { | |
1092 | if (region->type == GIMPLE_OMP_PARALLEL) | |
1093 | remove_exit_barrier (region); | |
1094 | ||
1095 | if (region->inner) | |
1096 | { | |
1097 | region = region->inner; | |
1098 | remove_exit_barriers (region); | |
1099 | while (region->next) | |
1100 | { | |
1101 | region = region->next; | |
1102 | remove_exit_barriers (region); | |
1103 | } | |
1104 | } | |
1105 | } | |
1106 | ||
1107 | /* Optimize omp_get_thread_num () and omp_get_num_threads () | |
1108 | calls. These can't be declared as const functions, but | |
1109 | within one parallel body they are constant, so they can be | |
1110 | transformed there into __builtin_omp_get_{thread_num,num_threads} () | |
1111 | which are declared const. Similarly for task body, except | |
1112 | that in untied task omp_get_thread_num () can change at any task | |
1113 | scheduling point. */ | |
1114 | ||
1115 | static void | |
1116 | optimize_omp_library_calls (gimple *entry_stmt) | |
1117 | { | |
1118 | basic_block bb; | |
1119 | gimple_stmt_iterator gsi; | |
1120 | tree thr_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
1121 | tree thr_num_id = DECL_ASSEMBLER_NAME (thr_num_tree); | |
1122 | tree num_thr_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
1123 | tree num_thr_id = DECL_ASSEMBLER_NAME (num_thr_tree); | |
1124 | bool untied_task = (gimple_code (entry_stmt) == GIMPLE_OMP_TASK | |
1125 | && omp_find_clause (gimple_omp_task_clauses (entry_stmt), | |
1126 | OMP_CLAUSE_UNTIED) != NULL); | |
1127 | ||
1128 | FOR_EACH_BB_FN (bb, cfun) | |
1129 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1130 | { | |
1131 | gimple *call = gsi_stmt (gsi); | |
1132 | tree decl; | |
1133 | ||
1134 | if (is_gimple_call (call) | |
1135 | && (decl = gimple_call_fndecl (call)) | |
1136 | && DECL_EXTERNAL (decl) | |
1137 | && TREE_PUBLIC (decl) | |
1138 | && DECL_INITIAL (decl) == NULL) | |
1139 | { | |
1140 | tree built_in; | |
1141 | ||
1142 | if (DECL_NAME (decl) == thr_num_id) | |
1143 | { | |
1144 | /* In #pragma omp task untied omp_get_thread_num () can change | |
1145 | during the execution of the task region. */ | |
1146 | if (untied_task) | |
1147 | continue; | |
1148 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
1149 | } | |
1150 | else if (DECL_NAME (decl) == num_thr_id) | |
1151 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
1152 | else | |
1153 | continue; | |
1154 | ||
1155 | if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in) | |
1156 | || gimple_call_num_args (call) != 0) | |
1157 | continue; | |
1158 | ||
1159 | if (flag_exceptions && !TREE_NOTHROW (decl)) | |
1160 | continue; | |
1161 | ||
1162 | if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE | |
1163 | || !types_compatible_p (TREE_TYPE (TREE_TYPE (decl)), | |
1164 | TREE_TYPE (TREE_TYPE (built_in)))) | |
1165 | continue; | |
1166 | ||
1167 | gimple_call_set_fndecl (call, built_in); | |
1168 | } | |
1169 | } | |
1170 | } | |
1171 | ||
1172 | /* Callback for expand_omp_build_assign. Return non-NULL if *tp needs to be | |
1173 | regimplified. */ | |
1174 | ||
1175 | static tree | |
1176 | expand_omp_regimplify_p (tree *tp, int *walk_subtrees, void *) | |
1177 | { | |
1178 | tree t = *tp; | |
1179 | ||
1180 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ | |
1181 | if (VAR_P (t) && DECL_HAS_VALUE_EXPR_P (t)) | |
1182 | return t; | |
1183 | ||
1184 | if (TREE_CODE (t) == ADDR_EXPR) | |
1185 | recompute_tree_invariant_for_addr_expr (t); | |
1186 | ||
1187 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); | |
1188 | return NULL_TREE; | |
1189 | } | |
1190 | ||
1191 | /* Prepend or append TO = FROM assignment before or after *GSI_P. */ | |
1192 | ||
1193 | static void | |
1194 | expand_omp_build_assign (gimple_stmt_iterator *gsi_p, tree to, tree from, | |
1195 | bool after) | |
1196 | { | |
1197 | bool simple_p = DECL_P (to) && TREE_ADDRESSABLE (to); | |
1198 | from = force_gimple_operand_gsi (gsi_p, from, simple_p, NULL_TREE, | |
1199 | !after, after ? GSI_CONTINUE_LINKING | |
1200 | : GSI_SAME_STMT); | |
1201 | gimple *stmt = gimple_build_assign (to, from); | |
1202 | if (after) | |
1203 | gsi_insert_after (gsi_p, stmt, GSI_CONTINUE_LINKING); | |
1204 | else | |
1205 | gsi_insert_before (gsi_p, stmt, GSI_SAME_STMT); | |
1206 | if (walk_tree (&from, expand_omp_regimplify_p, NULL, NULL) | |
1207 | || walk_tree (&to, expand_omp_regimplify_p, NULL, NULL)) | |
1208 | { | |
1209 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1210 | gimple_regimplify_operands (stmt, &gsi); | |
1211 | } | |
1212 | } | |
1213 | ||
47de0b56 JJ |
1214 | /* Prepend or append LHS CODE RHS condition before or after *GSI_P. */ |
1215 | ||
1216 | static gcond * | |
1217 | expand_omp_build_cond (gimple_stmt_iterator *gsi_p, enum tree_code code, | |
1218 | tree lhs, tree rhs, bool after = false) | |
1219 | { | |
1220 | gcond *cond_stmt = gimple_build_cond (code, lhs, rhs, NULL_TREE, NULL_TREE); | |
1221 | if (after) | |
1222 | gsi_insert_after (gsi_p, cond_stmt, GSI_CONTINUE_LINKING); | |
1223 | else | |
1224 | gsi_insert_before (gsi_p, cond_stmt, GSI_SAME_STMT); | |
1225 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
1226 | NULL, NULL) | |
1227 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
1228 | NULL, NULL)) | |
1229 | { | |
1230 | gimple_stmt_iterator gsi = gsi_for_stmt (cond_stmt); | |
1231 | gimple_regimplify_operands (cond_stmt, &gsi); | |
1232 | } | |
1233 | return cond_stmt; | |
1234 | } | |
1235 | ||
629b3d75 MJ |
1236 | /* Expand the OpenMP parallel or task directive starting at REGION. */ |
1237 | ||
1238 | static void | |
1239 | expand_omp_taskreg (struct omp_region *region) | |
1240 | { | |
1241 | basic_block entry_bb, exit_bb, new_bb; | |
1242 | struct function *child_cfun; | |
1243 | tree child_fn, block, t; | |
1244 | gimple_stmt_iterator gsi; | |
1245 | gimple *entry_stmt, *stmt; | |
1246 | edge e; | |
1247 | vec<tree, va_gc> *ws_args; | |
1248 | ||
fe8ac82f | 1249 | entry_stmt = last_nondebug_stmt (region->entry); |
28567c40 JJ |
1250 | if (gimple_code (entry_stmt) == GIMPLE_OMP_TASK |
1251 | && gimple_omp_task_taskwait_p (entry_stmt)) | |
1252 | { | |
1253 | new_bb = region->entry; | |
1254 | gsi = gsi_last_nondebug_bb (region->entry); | |
1255 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK); | |
1256 | gsi_remove (&gsi, true); | |
1257 | expand_taskwait_call (new_bb, as_a <gomp_task *> (entry_stmt)); | |
1258 | return; | |
1259 | } | |
1260 | ||
629b3d75 MJ |
1261 | child_fn = gimple_omp_taskreg_child_fn (entry_stmt); |
1262 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
1263 | ||
1264 | entry_bb = region->entry; | |
1265 | if (gimple_code (entry_stmt) == GIMPLE_OMP_TASK) | |
1266 | exit_bb = region->cont; | |
1267 | else | |
1268 | exit_bb = region->exit; | |
1269 | ||
5e9d6aa4 | 1270 | if (is_combined_parallel (region)) |
629b3d75 MJ |
1271 | ws_args = region->ws_args; |
1272 | else | |
1273 | ws_args = NULL; | |
1274 | ||
1275 | if (child_cfun->cfg) | |
1276 | { | |
1277 | /* Due to inlining, it may happen that we have already outlined | |
1278 | the region, in which case all we need to do is make the | |
1279 | sub-graph unreachable and emit the parallel call. */ | |
1280 | edge entry_succ_e, exit_succ_e; | |
1281 | ||
1282 | entry_succ_e = single_succ_edge (entry_bb); | |
1283 | ||
65f4b875 | 1284 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 | 1285 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_PARALLEL |
28567c40 JJ |
1286 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK |
1287 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TEAMS); | |
629b3d75 MJ |
1288 | gsi_remove (&gsi, true); |
1289 | ||
1290 | new_bb = entry_bb; | |
1291 | if (exit_bb) | |
1292 | { | |
1293 | exit_succ_e = single_succ_edge (exit_bb); | |
1294 | make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU); | |
1295 | } | |
1296 | remove_edge_and_dominated_blocks (entry_succ_e); | |
1297 | } | |
1298 | else | |
1299 | { | |
1300 | unsigned srcidx, dstidx, num; | |
1301 | ||
1302 | /* If the parallel region needs data sent from the parent | |
1303 | function, then the very first statement (except possible | |
1304 | tree profile counter updates) of the parallel body | |
1305 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
1306 | &.OMP_DATA_O is passed as an argument to the child function, | |
1307 | we need to replace it with the argument as seen by the child | |
1308 | function. | |
1309 | ||
1310 | In most cases, this will end up being the identity assignment | |
1311 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had | |
1312 | a function call that has been inlined, the original PARM_DECL | |
1313 | .OMP_DATA_I may have been converted into a different local | |
1314 | variable. In which case, we need to keep the assignment. */ | |
1315 | if (gimple_omp_taskreg_data_arg (entry_stmt)) | |
1316 | { | |
1317 | basic_block entry_succ_bb | |
1318 | = single_succ_p (entry_bb) ? single_succ (entry_bb) | |
1319 | : FALLTHRU_EDGE (entry_bb)->dest; | |
1320 | tree arg; | |
1321 | gimple *parcopy_stmt = NULL; | |
1322 | ||
1323 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
1324 | { | |
1325 | gimple *stmt; | |
1326 | ||
1327 | gcc_assert (!gsi_end_p (gsi)); | |
1328 | stmt = gsi_stmt (gsi); | |
1329 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
1330 | continue; | |
1331 | ||
1332 | if (gimple_num_ops (stmt) == 2) | |
1333 | { | |
1334 | tree arg = gimple_assign_rhs1 (stmt); | |
1335 | ||
1336 | /* We're ignore the subcode because we're | |
1337 | effectively doing a STRIP_NOPS. */ | |
1338 | ||
1339 | if (TREE_CODE (arg) == ADDR_EXPR | |
28567c40 JJ |
1340 | && (TREE_OPERAND (arg, 0) |
1341 | == gimple_omp_taskreg_data_arg (entry_stmt))) | |
629b3d75 MJ |
1342 | { |
1343 | parcopy_stmt = stmt; | |
1344 | break; | |
1345 | } | |
1346 | } | |
1347 | } | |
1348 | ||
1349 | gcc_assert (parcopy_stmt != NULL); | |
1350 | arg = DECL_ARGUMENTS (child_fn); | |
1351 | ||
1352 | if (!gimple_in_ssa_p (cfun)) | |
1353 | { | |
1354 | if (gimple_assign_lhs (parcopy_stmt) == arg) | |
1355 | gsi_remove (&gsi, true); | |
1356 | else | |
1357 | { | |
01914336 | 1358 | /* ?? Is setting the subcode really necessary ?? */ |
629b3d75 MJ |
1359 | gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (arg)); |
1360 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
1361 | } | |
1362 | } | |
1363 | else | |
1364 | { | |
1365 | tree lhs = gimple_assign_lhs (parcopy_stmt); | |
1366 | gcc_assert (SSA_NAME_VAR (lhs) == arg); | |
1367 | /* We'd like to set the rhs to the default def in the child_fn, | |
1368 | but it's too early to create ssa names in the child_fn. | |
1369 | Instead, we set the rhs to the parm. In | |
1370 | move_sese_region_to_fn, we introduce a default def for the | |
1371 | parm, map the parm to it's default def, and once we encounter | |
1372 | this stmt, replace the parm with the default def. */ | |
1373 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
1374 | update_stmt (parcopy_stmt); | |
1375 | } | |
1376 | } | |
1377 | ||
1378 | /* Declare local variables needed in CHILD_CFUN. */ | |
1379 | block = DECL_INITIAL (child_fn); | |
1380 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
1381 | /* The gimplifier could record temporaries in parallel/task block | |
1382 | rather than in containing function's local_decls chain, | |
1383 | which would mean cgraph missed finalizing them. Do it now. */ | |
1384 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
1385 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) | |
1386 | varpool_node::finalize_decl (t); | |
1387 | DECL_SAVED_TREE (child_fn) = NULL; | |
1388 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
1389 | gimple_set_body (child_fn, NULL); | |
1390 | TREE_USED (block) = 1; | |
1391 | ||
1392 | /* Reset DECL_CONTEXT on function arguments. */ | |
1393 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
1394 | DECL_CONTEXT (t) = child_fn; | |
1395 | ||
1396 | /* Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK, | |
1397 | so that it can be moved to the child function. */ | |
65f4b875 | 1398 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
1399 | stmt = gsi_stmt (gsi); |
1400 | gcc_assert (stmt && (gimple_code (stmt) == GIMPLE_OMP_PARALLEL | |
28567c40 JJ |
1401 | || gimple_code (stmt) == GIMPLE_OMP_TASK |
1402 | || gimple_code (stmt) == GIMPLE_OMP_TEAMS)); | |
629b3d75 MJ |
1403 | e = split_block (entry_bb, stmt); |
1404 | gsi_remove (&gsi, true); | |
1405 | entry_bb = e->dest; | |
1406 | edge e2 = NULL; | |
28567c40 | 1407 | if (gimple_code (entry_stmt) != GIMPLE_OMP_TASK) |
629b3d75 MJ |
1408 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; |
1409 | else | |
1410 | { | |
1411 | e2 = make_edge (e->src, BRANCH_EDGE (entry_bb)->dest, EDGE_ABNORMAL); | |
1412 | gcc_assert (e2->dest == region->exit); | |
1413 | remove_edge (BRANCH_EDGE (entry_bb)); | |
1414 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e->src); | |
65f4b875 | 1415 | gsi = gsi_last_nondebug_bb (region->exit); |
629b3d75 MJ |
1416 | gcc_assert (!gsi_end_p (gsi) |
1417 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
1418 | gsi_remove (&gsi, true); | |
1419 | } | |
1420 | ||
1421 | /* Convert GIMPLE_OMP_{RETURN,CONTINUE} into a RETURN_EXPR. */ | |
1422 | if (exit_bb) | |
1423 | { | |
65f4b875 | 1424 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
1425 | gcc_assert (!gsi_end_p (gsi) |
1426 | && (gimple_code (gsi_stmt (gsi)) | |
1427 | == (e2 ? GIMPLE_OMP_CONTINUE : GIMPLE_OMP_RETURN))); | |
1428 | stmt = gimple_build_return (NULL); | |
1429 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
1430 | gsi_remove (&gsi, true); | |
1431 | } | |
1432 | ||
1433 | /* Move the parallel region into CHILD_CFUN. */ | |
1434 | ||
1435 | if (gimple_in_ssa_p (cfun)) | |
1436 | { | |
1437 | init_tree_ssa (child_cfun); | |
1438 | init_ssa_operands (child_cfun); | |
1439 | child_cfun->gimple_df->in_ssa_p = true; | |
1440 | block = NULL_TREE; | |
1441 | } | |
1442 | else | |
1443 | block = gimple_block (entry_stmt); | |
1444 | ||
1445 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
1446 | if (exit_bb) | |
1447 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
1448 | if (e2) | |
1449 | { | |
1450 | basic_block dest_bb = e2->dest; | |
1451 | if (!exit_bb) | |
1452 | make_edge (new_bb, dest_bb, EDGE_FALLTHRU); | |
1453 | remove_edge (e2); | |
1454 | set_immediate_dominator (CDI_DOMINATORS, dest_bb, new_bb); | |
1455 | } | |
1456 | /* When the OMP expansion process cannot guarantee an up-to-date | |
01914336 | 1457 | loop tree arrange for the child function to fixup loops. */ |
629b3d75 MJ |
1458 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) |
1459 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
1460 | ||
1461 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
1462 | num = vec_safe_length (child_cfun->local_decls); | |
1463 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
1464 | { | |
1465 | t = (*child_cfun->local_decls)[srcidx]; | |
1466 | if (DECL_CONTEXT (t) == cfun->decl) | |
1467 | continue; | |
1468 | if (srcidx != dstidx) | |
1469 | (*child_cfun->local_decls)[dstidx] = t; | |
1470 | dstidx++; | |
1471 | } | |
1472 | if (dstidx != num) | |
1473 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
1474 | ||
1475 | /* Inform the callgraph about the new function. */ | |
1476 | child_cfun->curr_properties = cfun->curr_properties; | |
1477 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; | |
1478 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; | |
1479 | cgraph_node *node = cgraph_node::get_create (child_fn); | |
1480 | node->parallelized_function = 1; | |
1481 | cgraph_node::add_new_function (child_fn, true); | |
1482 | ||
1483 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) | |
1484 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); | |
1485 | ||
1486 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
1487 | fixed in a following pass. */ | |
1488 | push_cfun (child_cfun); | |
1489 | if (need_asm) | |
9579db35 | 1490 | assign_assembler_name_if_needed (child_fn); |
629b3d75 MJ |
1491 | |
1492 | if (optimize) | |
1493 | optimize_omp_library_calls (entry_stmt); | |
fc06ae0d | 1494 | update_max_bb_count (); |
629b3d75 MJ |
1495 | cgraph_edge::rebuild_edges (); |
1496 | ||
1497 | /* Some EH regions might become dead, see PR34608. If | |
1498 | pass_cleanup_cfg isn't the first pass to happen with the | |
1499 | new child, these dead EH edges might cause problems. | |
1500 | Clean them up now. */ | |
1501 | if (flag_exceptions) | |
1502 | { | |
1503 | basic_block bb; | |
1504 | bool changed = false; | |
1505 | ||
1506 | FOR_EACH_BB_FN (bb, cfun) | |
1507 | changed |= gimple_purge_dead_eh_edges (bb); | |
1508 | if (changed) | |
1509 | cleanup_tree_cfg (); | |
1510 | } | |
1511 | if (gimple_in_ssa_p (cfun)) | |
1512 | update_ssa (TODO_update_ssa); | |
1513 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
1514 | verify_loop_structure (); | |
1515 | pop_cfun (); | |
1516 | ||
1517 | if (dump_file && !gimple_in_ssa_p (cfun)) | |
1518 | { | |
1519 | omp_any_child_fn_dumped = true; | |
1520 | dump_function_header (dump_file, child_fn, dump_flags); | |
1521 | dump_function_to_file (child_fn, dump_file, dump_flags); | |
1522 | } | |
1523 | } | |
1524 | ||
4ccc4e30 JJ |
1525 | adjust_context_and_scope (region, gimple_block (entry_stmt), child_fn); |
1526 | ||
5e9d6aa4 | 1527 | if (gimple_code (entry_stmt) == GIMPLE_OMP_PARALLEL) |
629b3d75 MJ |
1528 | expand_parallel_call (region, new_bb, |
1529 | as_a <gomp_parallel *> (entry_stmt), ws_args); | |
28567c40 JJ |
1530 | else if (gimple_code (entry_stmt) == GIMPLE_OMP_TEAMS) |
1531 | expand_teams_call (new_bb, as_a <gomp_teams *> (entry_stmt)); | |
629b3d75 MJ |
1532 | else |
1533 | expand_task_call (region, new_bb, as_a <gomp_task *> (entry_stmt)); | |
629b3d75 MJ |
1534 | } |
1535 | ||
1536 | /* Information about members of an OpenACC collapsed loop nest. */ | |
1537 | ||
1538 | struct oacc_collapse | |
1539 | { | |
01914336 | 1540 | tree base; /* Base value. */ |
629b3d75 | 1541 | tree iters; /* Number of steps. */ |
02889d23 CLT |
1542 | tree step; /* Step size. */ |
1543 | tree tile; /* Tile increment (if tiled). */ | |
1544 | tree outer; /* Tile iterator var. */ | |
629b3d75 MJ |
1545 | }; |
1546 | ||
1547 | /* Helper for expand_oacc_for. Determine collapsed loop information. | |
1548 | Fill in COUNTS array. Emit any initialization code before GSI. | |
1549 | Return the calculated outer loop bound of BOUND_TYPE. */ | |
1550 | ||
1551 | static tree | |
1552 | expand_oacc_collapse_init (const struct omp_for_data *fd, | |
1553 | gimple_stmt_iterator *gsi, | |
f324479c TB |
1554 | oacc_collapse *counts, tree diff_type, |
1555 | tree bound_type, location_t loc) | |
629b3d75 | 1556 | { |
02889d23 | 1557 | tree tiling = fd->tiling; |
629b3d75 MJ |
1558 | tree total = build_int_cst (bound_type, 1); |
1559 | int ix; | |
1560 | ||
1561 | gcc_assert (integer_onep (fd->loop.step)); | |
1562 | gcc_assert (integer_zerop (fd->loop.n1)); | |
1563 | ||
02889d23 CLT |
1564 | /* When tiling, the first operand of the tile clause applies to the |
1565 | innermost loop, and we work outwards from there. Seems | |
1566 | backwards, but whatever. */ | |
1567 | for (ix = fd->collapse; ix--;) | |
629b3d75 MJ |
1568 | { |
1569 | const omp_for_data_loop *loop = &fd->loops[ix]; | |
1570 | ||
1571 | tree iter_type = TREE_TYPE (loop->v); | |
629b3d75 MJ |
1572 | tree plus_type = iter_type; |
1573 | ||
ac200799 | 1574 | gcc_assert (loop->cond_code == LT_EXPR || loop->cond_code == GT_EXPR); |
629b3d75 MJ |
1575 | |
1576 | if (POINTER_TYPE_P (iter_type)) | |
1577 | plus_type = sizetype; | |
629b3d75 | 1578 | |
02889d23 CLT |
1579 | if (tiling) |
1580 | { | |
1581 | tree num = build_int_cst (integer_type_node, fd->collapse); | |
1582 | tree loop_no = build_int_cst (integer_type_node, ix); | |
1583 | tree tile = TREE_VALUE (tiling); | |
1584 | gcall *call | |
1585 | = gimple_build_call_internal (IFN_GOACC_TILE, 5, num, loop_no, tile, | |
1586 | /* gwv-outer=*/integer_zero_node, | |
1587 | /* gwv-inner=*/integer_zero_node); | |
1588 | ||
1589 | counts[ix].outer = create_tmp_var (iter_type, ".outer"); | |
1590 | counts[ix].tile = create_tmp_var (diff_type, ".tile"); | |
1591 | gimple_call_set_lhs (call, counts[ix].tile); | |
1592 | gimple_set_location (call, loc); | |
1593 | gsi_insert_before (gsi, call, GSI_SAME_STMT); | |
1594 | ||
1595 | tiling = TREE_CHAIN (tiling); | |
1596 | } | |
1597 | else | |
1598 | { | |
1599 | counts[ix].tile = NULL; | |
1600 | counts[ix].outer = loop->v; | |
1601 | } | |
1602 | ||
629b3d75 MJ |
1603 | tree b = loop->n1; |
1604 | tree e = loop->n2; | |
1605 | tree s = loop->step; | |
1606 | bool up = loop->cond_code == LT_EXPR; | |
1607 | tree dir = build_int_cst (diff_type, up ? +1 : -1); | |
1608 | bool negating; | |
1609 | tree expr; | |
1610 | ||
1611 | b = force_gimple_operand_gsi (gsi, b, true, NULL_TREE, | |
1612 | true, GSI_SAME_STMT); | |
1613 | e = force_gimple_operand_gsi (gsi, e, true, NULL_TREE, | |
1614 | true, GSI_SAME_STMT); | |
1615 | ||
01914336 | 1616 | /* Convert the step, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
1617 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); |
1618 | if (negating) | |
1619 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); | |
1620 | s = fold_convert (diff_type, s); | |
1621 | if (negating) | |
1622 | s = fold_build1 (NEGATE_EXPR, diff_type, s); | |
1623 | s = force_gimple_operand_gsi (gsi, s, true, NULL_TREE, | |
1624 | true, GSI_SAME_STMT); | |
1625 | ||
01914336 | 1626 | /* Determine the range, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
1627 | negating = !up && TYPE_UNSIGNED (iter_type); |
1628 | expr = fold_build2 (MINUS_EXPR, plus_type, | |
1629 | fold_convert (plus_type, negating ? b : e), | |
1630 | fold_convert (plus_type, negating ? e : b)); | |
1631 | expr = fold_convert (diff_type, expr); | |
1632 | if (negating) | |
1633 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); | |
1634 | tree range = force_gimple_operand_gsi | |
1635 | (gsi, expr, true, NULL_TREE, true, GSI_SAME_STMT); | |
1636 | ||
1637 | /* Determine number of iterations. */ | |
1638 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); | |
1639 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); | |
1640 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); | |
1641 | ||
1642 | tree iters = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE, | |
1643 | true, GSI_SAME_STMT); | |
1644 | ||
1645 | counts[ix].base = b; | |
1646 | counts[ix].iters = iters; | |
1647 | counts[ix].step = s; | |
1648 | ||
1649 | total = fold_build2 (MULT_EXPR, bound_type, total, | |
1650 | fold_convert (bound_type, iters)); | |
1651 | } | |
1652 | ||
1653 | return total; | |
1654 | } | |
1655 | ||
02889d23 CLT |
1656 | /* Emit initializers for collapsed loop members. INNER is true if |
1657 | this is for the element loop of a TILE. IVAR is the outer | |
629b3d75 MJ |
1658 | loop iteration variable, from which collapsed loop iteration values |
1659 | are calculated. COUNTS array has been initialized by | |
1660 | expand_oacc_collapse_inits. */ | |
1661 | ||
1662 | static void | |
02889d23 | 1663 | expand_oacc_collapse_vars (const struct omp_for_data *fd, bool inner, |
629b3d75 | 1664 | gimple_stmt_iterator *gsi, |
f324479c TB |
1665 | const oacc_collapse *counts, tree ivar, |
1666 | tree diff_type) | |
629b3d75 MJ |
1667 | { |
1668 | tree ivar_type = TREE_TYPE (ivar); | |
1669 | ||
1670 | /* The most rapidly changing iteration variable is the innermost | |
1671 | one. */ | |
1672 | for (int ix = fd->collapse; ix--;) | |
1673 | { | |
1674 | const omp_for_data_loop *loop = &fd->loops[ix]; | |
1675 | const oacc_collapse *collapse = &counts[ix]; | |
02889d23 CLT |
1676 | tree v = inner ? loop->v : collapse->outer; |
1677 | tree iter_type = TREE_TYPE (v); | |
629b3d75 MJ |
1678 | tree plus_type = iter_type; |
1679 | enum tree_code plus_code = PLUS_EXPR; | |
1680 | tree expr; | |
1681 | ||
1682 | if (POINTER_TYPE_P (iter_type)) | |
1683 | { | |
1684 | plus_code = POINTER_PLUS_EXPR; | |
1685 | plus_type = sizetype; | |
1686 | } | |
1687 | ||
02889d23 CLT |
1688 | expr = ivar; |
1689 | if (ix) | |
1690 | { | |
1691 | tree mod = fold_convert (ivar_type, collapse->iters); | |
1692 | ivar = fold_build2 (TRUNC_DIV_EXPR, ivar_type, expr, mod); | |
1693 | expr = fold_build2 (TRUNC_MOD_EXPR, ivar_type, expr, mod); | |
1694 | ivar = force_gimple_operand_gsi (gsi, ivar, true, NULL_TREE, | |
1695 | true, GSI_SAME_STMT); | |
1696 | } | |
1697 | ||
629b3d75 | 1698 | expr = fold_build2 (MULT_EXPR, diff_type, fold_convert (diff_type, expr), |
f324479c | 1699 | fold_convert (diff_type, collapse->step)); |
02889d23 CLT |
1700 | expr = fold_build2 (plus_code, iter_type, |
1701 | inner ? collapse->outer : collapse->base, | |
629b3d75 MJ |
1702 | fold_convert (plus_type, expr)); |
1703 | expr = force_gimple_operand_gsi (gsi, expr, false, NULL_TREE, | |
1704 | true, GSI_SAME_STMT); | |
02889d23 | 1705 | gassign *ass = gimple_build_assign (v, expr); |
629b3d75 | 1706 | gsi_insert_before (gsi, ass, GSI_SAME_STMT); |
629b3d75 MJ |
1707 | } |
1708 | } | |
1709 | ||
1710 | /* Helper function for expand_omp_{for_*,simd}. If this is the outermost | |
1711 | of the combined collapse > 1 loop constructs, generate code like: | |
1712 | if (__builtin_expect (N32 cond3 N31, 0)) goto ZERO_ITER_BB; | |
1713 | if (cond3 is <) | |
1714 | adj = STEP3 - 1; | |
1715 | else | |
1716 | adj = STEP3 + 1; | |
1717 | count3 = (adj + N32 - N31) / STEP3; | |
1718 | if (__builtin_expect (N22 cond2 N21, 0)) goto ZERO_ITER_BB; | |
1719 | if (cond2 is <) | |
1720 | adj = STEP2 - 1; | |
1721 | else | |
1722 | adj = STEP2 + 1; | |
1723 | count2 = (adj + N22 - N21) / STEP2; | |
1724 | if (__builtin_expect (N12 cond1 N11, 0)) goto ZERO_ITER_BB; | |
1725 | if (cond1 is <) | |
1726 | adj = STEP1 - 1; | |
1727 | else | |
1728 | adj = STEP1 + 1; | |
1729 | count1 = (adj + N12 - N11) / STEP1; | |
1730 | count = count1 * count2 * count3; | |
1731 | Furthermore, if ZERO_ITER_BB is NULL, create a BB which does: | |
1732 | count = 0; | |
1733 | and set ZERO_ITER_BB to that bb. If this isn't the outermost | |
1734 | of the combined loop constructs, just initialize COUNTS array | |
aed3ab25 JJ |
1735 | from the _looptemp_ clauses. For loop nests with non-rectangular |
1736 | loops, do this only for the rectangular loops. Then pick | |
1737 | the loops which reference outer vars in their bound expressions | |
1738 | and the loops which they refer to and for this sub-nest compute | |
c2ebf4f1 JJ |
1739 | number of iterations. For triangular loops use Faulhaber's formula, |
1740 | otherwise as a fallback, compute by iterating the loops. | |
aed3ab25 JJ |
1741 | If e.g. the sub-nest is |
1742 | for (I = N11; I COND1 N12; I += STEP1) | |
1743 | for (J = M21 * I + N21; J COND2 M22 * I + N22; J += STEP2) | |
1744 | for (K = M31 * J + N31; K COND3 M32 * J + N32; K += STEP3) | |
1745 | do: | |
1746 | COUNT = 0; | |
1747 | for (tmpi = N11; tmpi COND1 N12; tmpi += STEP1) | |
1748 | for (tmpj = M21 * tmpi + N21; | |
1749 | tmpj COND2 M22 * tmpi + N22; tmpj += STEP2) | |
1750 | { | |
1751 | int tmpk1 = M31 * tmpj + N31; | |
1752 | int tmpk2 = M32 * tmpj + N32; | |
1753 | if (tmpk1 COND3 tmpk2) | |
1754 | { | |
1755 | if (COND3 is <) | |
1756 | adj = STEP3 - 1; | |
1757 | else | |
1758 | adj = STEP3 + 1; | |
1759 | COUNT += (adj + tmpk2 - tmpk1) / STEP3; | |
1760 | } | |
1761 | } | |
1762 | and finally multiply the counts of the rectangular loops not | |
1763 | in the sub-nest with COUNT. Also, as counts[fd->last_nonrect] | |
1764 | store number of iterations of the loops from fd->first_nonrect | |
1765 | to fd->last_nonrect inclusive, i.e. the above COUNT multiplied | |
1766 | by the counts of rectangular loops not referenced in any non-rectangular | |
1767 | loops sandwitched in between those. */ | |
629b3d75 MJ |
1768 | |
1769 | /* NOTE: It *could* be better to moosh all of the BBs together, | |
1770 | creating one larger BB with all the computation and the unexpected | |
1771 | jump at the end. I.e. | |
1772 | ||
1773 | bool zero3, zero2, zero1, zero; | |
1774 | ||
1775 | zero3 = N32 c3 N31; | |
1776 | count3 = (N32 - N31) /[cl] STEP3; | |
1777 | zero2 = N22 c2 N21; | |
1778 | count2 = (N22 - N21) /[cl] STEP2; | |
1779 | zero1 = N12 c1 N11; | |
1780 | count1 = (N12 - N11) /[cl] STEP1; | |
1781 | zero = zero3 || zero2 || zero1; | |
1782 | count = count1 * count2 * count3; | |
1783 | if (__builtin_expect(zero, false)) goto zero_iter_bb; | |
1784 | ||
1785 | After all, we expect the zero=false, and thus we expect to have to | |
1786 | evaluate all of the comparison expressions, so short-circuiting | |
1787 | oughtn't be a win. Since the condition isn't protecting a | |
1788 | denominator, we're not concerned about divide-by-zero, so we can | |
1789 | fully evaluate count even if a numerator turned out to be wrong. | |
1790 | ||
1791 | It seems like putting this all together would create much better | |
1792 | scheduling opportunities, and less pressure on the chip's branch | |
1793 | predictor. */ | |
1794 | ||
1795 | static void | |
1796 | expand_omp_for_init_counts (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
1797 | basic_block &entry_bb, tree *counts, | |
1798 | basic_block &zero_iter1_bb, int &first_zero_iter1, | |
1799 | basic_block &zero_iter2_bb, int &first_zero_iter2, | |
1800 | basic_block &l2_dom_bb) | |
1801 | { | |
1802 | tree t, type = TREE_TYPE (fd->loop.v); | |
1803 | edge e, ne; | |
1804 | int i; | |
1805 | ||
1806 | /* Collapsed loops need work for expansion into SSA form. */ | |
1807 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
1808 | ||
1809 | if (gimple_omp_for_combined_into_p (fd->for_stmt) | |
1810 | && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
1811 | { | |
1812 | gcc_assert (fd->ordered == 0); | |
1813 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
1814 | isn't supposed to be handled, as the inner loop doesn't | |
1815 | use it. */ | |
1816 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
1817 | OMP_CLAUSE__LOOPTEMP_); | |
1818 | gcc_assert (innerc); | |
1819 | for (i = 0; i < fd->collapse; i++) | |
1820 | { | |
1821 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
1822 | OMP_CLAUSE__LOOPTEMP_); | |
1823 | gcc_assert (innerc); | |
1824 | if (i) | |
1825 | counts[i] = OMP_CLAUSE_DECL (innerc); | |
1826 | else | |
1827 | counts[0] = NULL_TREE; | |
1828 | } | |
14707c89 JJ |
1829 | if (fd->non_rect |
1830 | && fd->last_nonrect == fd->first_nonrect + 1 | |
1831 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) | |
1832 | { | |
1833 | tree c[4]; | |
1834 | for (i = 0; i < 4; i++) | |
1835 | { | |
1836 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
1837 | OMP_CLAUSE__LOOPTEMP_); | |
1838 | gcc_assert (innerc); | |
1839 | c[i] = OMP_CLAUSE_DECL (innerc); | |
1840 | } | |
1841 | counts[0] = c[0]; | |
1842 | fd->first_inner_iterations = c[1]; | |
1843 | fd->factor = c[2]; | |
1844 | fd->adjn1 = c[3]; | |
1845 | } | |
629b3d75 MJ |
1846 | return; |
1847 | } | |
1848 | ||
1849 | for (i = fd->collapse; i < fd->ordered; i++) | |
1850 | { | |
1851 | tree itype = TREE_TYPE (fd->loops[i].v); | |
1852 | counts[i] = NULL_TREE; | |
1853 | t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
1854 | fold_convert (itype, fd->loops[i].n1), | |
1855 | fold_convert (itype, fd->loops[i].n2)); | |
1856 | if (t && integer_zerop (t)) | |
1857 | { | |
1858 | for (i = fd->collapse; i < fd->ordered; i++) | |
1859 | counts[i] = build_int_cst (type, 0); | |
1860 | break; | |
1861 | } | |
1862 | } | |
aed3ab25 | 1863 | bool rect_count_seen = false; |
629b3d75 MJ |
1864 | for (i = 0; i < (fd->ordered ? fd->ordered : fd->collapse); i++) |
1865 | { | |
1866 | tree itype = TREE_TYPE (fd->loops[i].v); | |
1867 | ||
1868 | if (i >= fd->collapse && counts[i]) | |
1869 | continue; | |
aed3ab25 JJ |
1870 | if (fd->non_rect) |
1871 | { | |
1872 | /* Skip loops that use outer iterators in their expressions | |
1873 | during this phase. */ | |
1874 | if (fd->loops[i].m1 || fd->loops[i].m2) | |
1875 | { | |
1876 | counts[i] = build_zero_cst (type); | |
1877 | continue; | |
1878 | } | |
1879 | } | |
629b3d75 MJ |
1880 | if ((SSA_VAR_P (fd->loop.n2) || i >= fd->collapse) |
1881 | && ((t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
1882 | fold_convert (itype, fd->loops[i].n1), | |
1883 | fold_convert (itype, fd->loops[i].n2))) | |
1884 | == NULL_TREE || !integer_onep (t))) | |
1885 | { | |
1886 | gcond *cond_stmt; | |
1887 | tree n1, n2; | |
1888 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
1889 | n1 = force_gimple_operand_gsi (gsi, n1, true, NULL_TREE, | |
1890 | true, GSI_SAME_STMT); | |
1891 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
1892 | n2 = force_gimple_operand_gsi (gsi, n2, true, NULL_TREE, | |
1893 | true, GSI_SAME_STMT); | |
47de0b56 JJ |
1894 | cond_stmt = expand_omp_build_cond (gsi, fd->loops[i].cond_code, |
1895 | n1, n2); | |
629b3d75 MJ |
1896 | e = split_block (entry_bb, cond_stmt); |
1897 | basic_block &zero_iter_bb | |
1898 | = i < fd->collapse ? zero_iter1_bb : zero_iter2_bb; | |
1899 | int &first_zero_iter | |
1900 | = i < fd->collapse ? first_zero_iter1 : first_zero_iter2; | |
1901 | if (zero_iter_bb == NULL) | |
1902 | { | |
1903 | gassign *assign_stmt; | |
1904 | first_zero_iter = i; | |
1905 | zero_iter_bb = create_empty_bb (entry_bb); | |
1906 | add_bb_to_loop (zero_iter_bb, entry_bb->loop_father); | |
1907 | *gsi = gsi_after_labels (zero_iter_bb); | |
1908 | if (i < fd->collapse) | |
1909 | assign_stmt = gimple_build_assign (fd->loop.n2, | |
1910 | build_zero_cst (type)); | |
1911 | else | |
1912 | { | |
1913 | counts[i] = create_tmp_reg (type, ".count"); | |
1914 | assign_stmt | |
1915 | = gimple_build_assign (counts[i], build_zero_cst (type)); | |
1916 | } | |
1917 | gsi_insert_before (gsi, assign_stmt, GSI_SAME_STMT); | |
1918 | set_immediate_dominator (CDI_DOMINATORS, zero_iter_bb, | |
1919 | entry_bb); | |
1920 | } | |
1921 | ne = make_edge (entry_bb, zero_iter_bb, EDGE_FALSE_VALUE); | |
357067f2 | 1922 | ne->probability = profile_probability::very_unlikely (); |
629b3d75 | 1923 | e->flags = EDGE_TRUE_VALUE; |
357067f2 | 1924 | e->probability = ne->probability.invert (); |
629b3d75 MJ |
1925 | if (l2_dom_bb == NULL) |
1926 | l2_dom_bb = entry_bb; | |
1927 | entry_bb = e->dest; | |
65f4b875 | 1928 | *gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
1929 | } |
1930 | ||
1931 | if (POINTER_TYPE_P (itype)) | |
1932 | itype = signed_type_for (itype); | |
1933 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
1934 | ? -1 : 1)); | |
1935 | t = fold_build2 (PLUS_EXPR, itype, | |
1936 | fold_convert (itype, fd->loops[i].step), t); | |
1937 | t = fold_build2 (PLUS_EXPR, itype, t, | |
1938 | fold_convert (itype, fd->loops[i].n2)); | |
1939 | t = fold_build2 (MINUS_EXPR, itype, t, | |
1940 | fold_convert (itype, fd->loops[i].n1)); | |
1941 | /* ?? We could probably use CEIL_DIV_EXPR instead of | |
1942 | TRUNC_DIV_EXPR and adjusting by hand. Unless we can't | |
1943 | generate the same code in the end because generically we | |
1944 | don't know that the values involved must be negative for | |
1945 | GT?? */ | |
1946 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
1947 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
1948 | fold_build1 (NEGATE_EXPR, itype, t), | |
1949 | fold_build1 (NEGATE_EXPR, itype, | |
1950 | fold_convert (itype, | |
1951 | fd->loops[i].step))); | |
1952 | else | |
1953 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
1954 | fold_convert (itype, fd->loops[i].step)); | |
1955 | t = fold_convert (type, t); | |
1956 | if (TREE_CODE (t) == INTEGER_CST) | |
1957 | counts[i] = t; | |
1958 | else | |
1959 | { | |
1960 | if (i < fd->collapse || i != first_zero_iter2) | |
1961 | counts[i] = create_tmp_reg (type, ".count"); | |
1962 | expand_omp_build_assign (gsi, counts[i], t); | |
1963 | } | |
1964 | if (SSA_VAR_P (fd->loop.n2) && i < fd->collapse) | |
1965 | { | |
aed3ab25 JJ |
1966 | if (fd->non_rect && i >= fd->first_nonrect && i <= fd->last_nonrect) |
1967 | continue; | |
1968 | if (!rect_count_seen) | |
1969 | { | |
1970 | t = counts[i]; | |
1971 | rect_count_seen = true; | |
1972 | } | |
629b3d75 MJ |
1973 | else |
1974 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, counts[i]); | |
1975 | expand_omp_build_assign (gsi, fd->loop.n2, t); | |
1976 | } | |
1977 | } | |
aed3ab25 JJ |
1978 | if (fd->non_rect && SSA_VAR_P (fd->loop.n2)) |
1979 | { | |
1980 | gcc_assert (fd->last_nonrect != -1); | |
1981 | ||
aed3ab25 JJ |
1982 | counts[fd->last_nonrect] = create_tmp_reg (type, ".count"); |
1983 | expand_omp_build_assign (gsi, counts[fd->last_nonrect], | |
1984 | build_zero_cst (type)); | |
29e0ad45 JJ |
1985 | for (i = fd->first_nonrect + 1; i < fd->last_nonrect; i++) |
1986 | if (fd->loops[i].m1 | |
1987 | || fd->loops[i].m2 | |
1988 | || fd->loops[i].non_rect_referenced) | |
1989 | break; | |
1990 | if (i == fd->last_nonrect | |
1991 | && fd->loops[i].outer == fd->last_nonrect - fd->first_nonrect | |
2084b5f4 | 1992 | && !POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v)) |
29e0ad45 | 1993 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[i].v))) |
aed3ab25 | 1994 | { |
29e0ad45 JJ |
1995 | int o = fd->first_nonrect; |
1996 | tree itype = TREE_TYPE (fd->loops[o].v); | |
1997 | tree n1o = create_tmp_reg (itype, ".n1o"); | |
1998 | t = fold_convert (itype, unshare_expr (fd->loops[o].n1)); | |
1999 | expand_omp_build_assign (gsi, n1o, t); | |
2000 | tree n2o = create_tmp_reg (itype, ".n2o"); | |
2001 | t = fold_convert (itype, unshare_expr (fd->loops[o].n2)); | |
2002 | expand_omp_build_assign (gsi, n2o, t); | |
2003 | if (fd->loops[i].m1 && fd->loops[i].m2) | |
2004 | t = fold_build2 (MINUS_EXPR, itype, unshare_expr (fd->loops[i].m2), | |
2005 | unshare_expr (fd->loops[i].m1)); | |
2006 | else if (fd->loops[i].m1) | |
46644ec9 JJ |
2007 | t = fold_build1 (NEGATE_EXPR, itype, |
2008 | unshare_expr (fd->loops[i].m1)); | |
29e0ad45 JJ |
2009 | else |
2010 | t = unshare_expr (fd->loops[i].m2); | |
2011 | tree m2minusm1 | |
2012 | = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, | |
2013 | true, GSI_SAME_STMT); | |
aed3ab25 | 2014 | |
29e0ad45 JJ |
2015 | gimple_stmt_iterator gsi2 = *gsi; |
2016 | gsi_prev (&gsi2); | |
2017 | e = split_block (entry_bb, gsi_stmt (gsi2)); | |
2018 | e = split_block (e->dest, (gimple *) NULL); | |
2019 | basic_block bb1 = e->src; | |
2020 | entry_bb = e->dest; | |
2021 | *gsi = gsi_after_labels (entry_bb); | |
aed3ab25 | 2022 | |
29e0ad45 JJ |
2023 | gsi2 = gsi_after_labels (bb1); |
2024 | tree ostep = fold_convert (itype, fd->loops[o].step); | |
2025 | t = build_int_cst (itype, (fd->loops[o].cond_code | |
2026 | == LT_EXPR ? -1 : 1)); | |
2027 | t = fold_build2 (PLUS_EXPR, itype, ostep, t); | |
2028 | t = fold_build2 (PLUS_EXPR, itype, t, n2o); | |
2029 | t = fold_build2 (MINUS_EXPR, itype, t, n1o); | |
2030 | if (TYPE_UNSIGNED (itype) | |
2031 | && fd->loops[o].cond_code == GT_EXPR) | |
2032 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2033 | fold_build1 (NEGATE_EXPR, itype, t), | |
2034 | fold_build1 (NEGATE_EXPR, itype, ostep)); | |
2035 | else | |
2036 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, ostep); | |
2037 | tree outer_niters | |
2038 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2039 | true, GSI_SAME_STMT); | |
2040 | t = fold_build2 (MINUS_EXPR, itype, outer_niters, | |
2041 | build_one_cst (itype)); | |
2042 | t = fold_build2 (MULT_EXPR, itype, t, ostep); | |
2043 | t = fold_build2 (PLUS_EXPR, itype, n1o, t); | |
2044 | tree last = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2045 | true, GSI_SAME_STMT); | |
2046 | tree n1, n2, n1e, n2e; | |
aed3ab25 JJ |
2047 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); |
2048 | if (fd->loops[i].m1) | |
2049 | { | |
2050 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); | |
29e0ad45 | 2051 | n1 = fold_build2 (MULT_EXPR, itype, n1o, n1); |
aed3ab25 JJ |
2052 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); |
2053 | } | |
2054 | else | |
2055 | n1 = t; | |
2056 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, | |
2057 | true, GSI_SAME_STMT); | |
aed3ab25 JJ |
2058 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
2059 | if (fd->loops[i].m2) | |
2060 | { | |
2061 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
29e0ad45 | 2062 | n2 = fold_build2 (MULT_EXPR, itype, n1o, n2); |
aed3ab25 JJ |
2063 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); |
2064 | } | |
2065 | else | |
2066 | n2 = t; | |
2067 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, | |
2068 | true, GSI_SAME_STMT); | |
29e0ad45 JJ |
2069 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); |
2070 | if (fd->loops[i].m1) | |
aed3ab25 | 2071 | { |
29e0ad45 JJ |
2072 | n1e = fold_convert (itype, unshare_expr (fd->loops[i].m1)); |
2073 | n1e = fold_build2 (MULT_EXPR, itype, last, n1e); | |
2074 | n1e = fold_build2 (PLUS_EXPR, itype, n1e, t); | |
2075 | } | |
2076 | else | |
2077 | n1e = t; | |
2078 | n1e = force_gimple_operand_gsi (&gsi2, n1e, true, NULL_TREE, | |
2079 | true, GSI_SAME_STMT); | |
2080 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
2081 | if (fd->loops[i].m2) | |
2082 | { | |
2083 | n2e = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2084 | n2e = fold_build2 (MULT_EXPR, itype, last, n2e); | |
2085 | n2e = fold_build2 (PLUS_EXPR, itype, n2e, t); | |
2086 | } | |
2087 | else | |
2088 | n2e = t; | |
2089 | n2e = force_gimple_operand_gsi (&gsi2, n2e, true, NULL_TREE, | |
2090 | true, GSI_SAME_STMT); | |
2091 | gcond *cond_stmt | |
47de0b56 JJ |
2092 | = expand_omp_build_cond (&gsi2, fd->loops[i].cond_code, |
2093 | n1, n2); | |
29e0ad45 JJ |
2094 | e = split_block (bb1, cond_stmt); |
2095 | e->flags = EDGE_TRUE_VALUE; | |
2096 | e->probability = profile_probability::likely ().guessed (); | |
2097 | basic_block bb2 = e->dest; | |
2098 | gsi2 = gsi_after_labels (bb2); | |
2099 | ||
47de0b56 JJ |
2100 | cond_stmt = expand_omp_build_cond (&gsi2, fd->loops[i].cond_code, |
2101 | n1e, n2e); | |
29e0ad45 JJ |
2102 | e = split_block (bb2, cond_stmt); |
2103 | e->flags = EDGE_TRUE_VALUE; | |
2104 | e->probability = profile_probability::likely ().guessed (); | |
2105 | gsi2 = gsi_after_labels (e->dest); | |
2106 | ||
2107 | tree step = fold_convert (itype, fd->loops[i].step); | |
2108 | t = build_int_cst (itype, (fd->loops[i].cond_code | |
2109 | == LT_EXPR ? -1 : 1)); | |
2110 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
2111 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
2112 | t = fold_build2 (MINUS_EXPR, itype, t, n1); | |
2113 | if (TYPE_UNSIGNED (itype) | |
2114 | && fd->loops[i].cond_code == GT_EXPR) | |
2115 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2116 | fold_build1 (NEGATE_EXPR, itype, t), | |
2117 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2118 | else | |
2119 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2120 | tree first_inner_iterations | |
2121 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2122 | true, GSI_SAME_STMT); | |
2123 | t = fold_build2 (MULT_EXPR, itype, m2minusm1, ostep); | |
2124 | if (TYPE_UNSIGNED (itype) | |
2125 | && fd->loops[i].cond_code == GT_EXPR) | |
2126 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2127 | fold_build1 (NEGATE_EXPR, itype, t), | |
2128 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2129 | else | |
2130 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2131 | tree factor | |
2132 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2133 | true, GSI_SAME_STMT); | |
2134 | t = fold_build2 (MINUS_EXPR, itype, outer_niters, | |
2135 | build_one_cst (itype)); | |
2136 | t = fold_build2 (MULT_EXPR, itype, t, outer_niters); | |
2137 | t = fold_build2 (RSHIFT_EXPR, itype, t, integer_one_node); | |
2138 | t = fold_build2 (MULT_EXPR, itype, factor, t); | |
2139 | t = fold_build2 (PLUS_EXPR, itype, | |
2140 | fold_build2 (MULT_EXPR, itype, outer_niters, | |
2141 | first_inner_iterations), t); | |
2142 | expand_omp_build_assign (&gsi2, counts[fd->last_nonrect], | |
2143 | fold_convert (type, t)); | |
2144 | ||
2145 | basic_block bb3 = create_empty_bb (bb1); | |
2146 | add_bb_to_loop (bb3, bb1->loop_father); | |
2147 | ||
2148 | e = make_edge (bb1, bb3, EDGE_FALSE_VALUE); | |
2149 | e->probability = profile_probability::unlikely ().guessed (); | |
2150 | ||
2151 | gsi2 = gsi_after_labels (bb3); | |
47de0b56 JJ |
2152 | cond_stmt = expand_omp_build_cond (&gsi2, fd->loops[i].cond_code, |
2153 | n1e, n2e); | |
29e0ad45 JJ |
2154 | e = split_block (bb3, cond_stmt); |
2155 | e->flags = EDGE_TRUE_VALUE; | |
2156 | e->probability = profile_probability::likely ().guessed (); | |
2157 | basic_block bb4 = e->dest; | |
2158 | ||
2159 | ne = make_edge (bb3, entry_bb, EDGE_FALSE_VALUE); | |
2160 | ne->probability = e->probability.invert (); | |
2161 | ||
2162 | basic_block bb5 = create_empty_bb (bb2); | |
2163 | add_bb_to_loop (bb5, bb2->loop_father); | |
2164 | ||
2165 | ne = make_edge (bb2, bb5, EDGE_FALSE_VALUE); | |
2166 | ne->probability = profile_probability::unlikely ().guessed (); | |
2167 | ||
2168 | for (int j = 0; j < 2; j++) | |
2169 | { | |
2170 | gsi2 = gsi_after_labels (j ? bb5 : bb4); | |
2171 | t = fold_build2 (MINUS_EXPR, itype, | |
2172 | unshare_expr (fd->loops[i].n1), | |
2173 | unshare_expr (fd->loops[i].n2)); | |
2174 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, m2minusm1); | |
2175 | tree tem | |
2176 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2177 | true, GSI_SAME_STMT); | |
2178 | t = fold_build2 (MINUS_EXPR, itype, tem, n1o); | |
2179 | t = fold_build2 (TRUNC_MOD_EXPR, itype, t, ostep); | |
2180 | t = fold_build2 (MINUS_EXPR, itype, tem, t); | |
2181 | tem = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2182 | true, GSI_SAME_STMT); | |
2183 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
2184 | if (fd->loops[i].m1) | |
2185 | { | |
2186 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); | |
2187 | n1 = fold_build2 (MULT_EXPR, itype, tem, n1); | |
2188 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2189 | } | |
2190 | else | |
2191 | n1 = t; | |
2192 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, | |
2193 | true, GSI_SAME_STMT); | |
2194 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
2195 | if (fd->loops[i].m2) | |
2196 | { | |
2197 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2198 | n2 = fold_build2 (MULT_EXPR, itype, tem, n2); | |
2199 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); | |
2200 | } | |
2201 | else | |
2202 | n2 = t; | |
2203 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, | |
2204 | true, GSI_SAME_STMT); | |
2205 | expand_omp_build_assign (&gsi2, j ? n2o : n1o, tem); | |
2206 | ||
47de0b56 JJ |
2207 | cond_stmt = expand_omp_build_cond (&gsi2, fd->loops[i].cond_code, |
2208 | n1, n2); | |
29e0ad45 JJ |
2209 | e = split_block (gsi_bb (gsi2), cond_stmt); |
2210 | e->flags = j ? EDGE_TRUE_VALUE : EDGE_FALSE_VALUE; | |
2211 | e->probability = profile_probability::unlikely ().guessed (); | |
2212 | ne = make_edge (e->src, bb1, | |
2213 | j ? EDGE_FALSE_VALUE : EDGE_TRUE_VALUE); | |
aed3ab25 JJ |
2214 | ne->probability = e->probability.invert (); |
2215 | gsi2 = gsi_after_labels (e->dest); | |
2216 | ||
29e0ad45 JJ |
2217 | t = fold_build2 (PLUS_EXPR, itype, tem, ostep); |
2218 | expand_omp_build_assign (&gsi2, j ? n2o : n1o, t); | |
aed3ab25 | 2219 | |
29e0ad45 JJ |
2220 | make_edge (e->dest, bb1, EDGE_FALLTHRU); |
2221 | } | |
aed3ab25 | 2222 | |
29e0ad45 JJ |
2223 | set_immediate_dominator (CDI_DOMINATORS, bb3, bb1); |
2224 | set_immediate_dominator (CDI_DOMINATORS, bb5, bb2); | |
2225 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, bb1); | |
325714b4 JJ |
2226 | |
2227 | if (fd->first_nonrect + 1 == fd->last_nonrect) | |
2228 | { | |
2229 | fd->first_inner_iterations = first_inner_iterations; | |
2230 | fd->factor = factor; | |
2231 | fd->adjn1 = n1o; | |
2232 | } | |
29e0ad45 JJ |
2233 | } |
2234 | else | |
2235 | { | |
2236 | /* Fallback implementation. Evaluate the loops with m1/m2 | |
2237 | non-NULL as well as their outer loops at runtime using temporaries | |
2238 | instead of the original iteration variables, and in the | |
2239 | body just bump the counter. */ | |
2240 | gimple_stmt_iterator gsi2 = *gsi; | |
2241 | gsi_prev (&gsi2); | |
2242 | e = split_block (entry_bb, gsi_stmt (gsi2)); | |
2243 | e = split_block (e->dest, (gimple *) NULL); | |
2244 | basic_block cur_bb = e->src; | |
2245 | basic_block next_bb = e->dest; | |
2246 | entry_bb = e->dest; | |
2247 | *gsi = gsi_after_labels (entry_bb); | |
aed3ab25 | 2248 | |
29e0ad45 JJ |
2249 | tree *vs = XALLOCAVEC (tree, fd->last_nonrect); |
2250 | memset (vs, 0, fd->last_nonrect * sizeof (tree)); | |
aed3ab25 | 2251 | |
29e0ad45 | 2252 | for (i = 0; i <= fd->last_nonrect; i++) |
aed3ab25 | 2253 | { |
29e0ad45 JJ |
2254 | if (fd->loops[i].m1 == NULL_TREE |
2255 | && fd->loops[i].m2 == NULL_TREE | |
2256 | && !fd->loops[i].non_rect_referenced) | |
2257 | continue; | |
2258 | ||
2259 | tree itype = TREE_TYPE (fd->loops[i].v); | |
2260 | ||
2261 | gsi2 = gsi_after_labels (cur_bb); | |
2262 | tree n1, n2; | |
2263 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
2084b5f4 JJ |
2264 | if (fd->loops[i].m1 == NULL_TREE) |
2265 | n1 = t; | |
2266 | else if (POINTER_TYPE_P (itype)) | |
2267 | { | |
2268 | gcc_assert (integer_onep (fd->loops[i].m1)); | |
4796d16d | 2269 | t = unshare_expr (fd->loops[i].n1); |
2084b5f4 JJ |
2270 | n1 = fold_build_pointer_plus (vs[i - fd->loops[i].outer], t); |
2271 | } | |
2272 | else | |
29e0ad45 JJ |
2273 | { |
2274 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); | |
2275 | n1 = fold_build2 (MULT_EXPR, itype, | |
2276 | vs[i - fd->loops[i].outer], n1); | |
2277 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2278 | } | |
29e0ad45 JJ |
2279 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, |
2280 | true, GSI_SAME_STMT); | |
2281 | if (i < fd->last_nonrect) | |
2282 | { | |
2283 | vs[i] = create_tmp_reg (itype, ".it"); | |
2284 | expand_omp_build_assign (&gsi2, vs[i], n1); | |
2285 | } | |
2286 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
2084b5f4 JJ |
2287 | if (fd->loops[i].m2 == NULL_TREE) |
2288 | n2 = t; | |
2289 | else if (POINTER_TYPE_P (itype)) | |
2290 | { | |
2291 | gcc_assert (integer_onep (fd->loops[i].m2)); | |
4796d16d | 2292 | t = unshare_expr (fd->loops[i].n2); |
2084b5f4 JJ |
2293 | n2 = fold_build_pointer_plus (vs[i - fd->loops[i].outer], t); |
2294 | } | |
2295 | else | |
29e0ad45 JJ |
2296 | { |
2297 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
2298 | n2 = fold_build2 (MULT_EXPR, itype, | |
2299 | vs[i - fd->loops[i].outer], n2); | |
2300 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); | |
2301 | } | |
29e0ad45 JJ |
2302 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, |
2303 | true, GSI_SAME_STMT); | |
2084b5f4 JJ |
2304 | if (POINTER_TYPE_P (itype)) |
2305 | itype = signed_type_for (itype); | |
29e0ad45 JJ |
2306 | if (i == fd->last_nonrect) |
2307 | { | |
2308 | gcond *cond_stmt | |
47de0b56 JJ |
2309 | = expand_omp_build_cond (&gsi2, fd->loops[i].cond_code, |
2310 | n1, n2); | |
29e0ad45 JJ |
2311 | e = split_block (cur_bb, cond_stmt); |
2312 | e->flags = EDGE_TRUE_VALUE; | |
2313 | ne = make_edge (cur_bb, next_bb, EDGE_FALSE_VALUE); | |
2314 | e->probability = profile_probability::likely ().guessed (); | |
2315 | ne->probability = e->probability.invert (); | |
2316 | gsi2 = gsi_after_labels (e->dest); | |
2317 | ||
2318 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
2319 | ? -1 : 1)); | |
2320 | t = fold_build2 (PLUS_EXPR, itype, | |
2321 | fold_convert (itype, fd->loops[i].step), t); | |
2084b5f4 JJ |
2322 | t = fold_build2 (PLUS_EXPR, itype, t, |
2323 | fold_convert (itype, n2)); | |
2324 | t = fold_build2 (MINUS_EXPR, itype, t, | |
2325 | fold_convert (itype, n1)); | |
29e0ad45 JJ |
2326 | tree step = fold_convert (itype, fd->loops[i].step); |
2327 | if (TYPE_UNSIGNED (itype) | |
2328 | && fd->loops[i].cond_code == GT_EXPR) | |
2329 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2330 | fold_build1 (NEGATE_EXPR, itype, t), | |
2331 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2332 | else | |
2333 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2334 | t = fold_convert (type, t); | |
2335 | t = fold_build2 (PLUS_EXPR, type, | |
2336 | counts[fd->last_nonrect], t); | |
2337 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2338 | true, GSI_SAME_STMT); | |
2339 | expand_omp_build_assign (&gsi2, counts[fd->last_nonrect], t); | |
2340 | e = make_edge (e->dest, next_bb, EDGE_FALLTHRU); | |
2341 | set_immediate_dominator (CDI_DOMINATORS, next_bb, cur_bb); | |
2342 | break; | |
2343 | } | |
fe8ac82f | 2344 | e = split_block (cur_bb, last_nondebug_stmt (cur_bb)); |
29e0ad45 JJ |
2345 | |
2346 | basic_block new_cur_bb = create_empty_bb (cur_bb); | |
2347 | add_bb_to_loop (new_cur_bb, cur_bb->loop_father); | |
2348 | ||
2349 | gsi2 = gsi_after_labels (e->dest); | |
2350 | tree step = fold_convert (itype, | |
2351 | unshare_expr (fd->loops[i].step)); | |
2084b5f4 | 2352 | if (POINTER_TYPE_P (TREE_TYPE (vs[i]))) |
4796d16d | 2353 | t = fold_build_pointer_plus (vs[i], step); |
2084b5f4 JJ |
2354 | else |
2355 | t = fold_build2 (PLUS_EXPR, itype, vs[i], step); | |
29e0ad45 JJ |
2356 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2357 | true, GSI_SAME_STMT); | |
2358 | expand_omp_build_assign (&gsi2, vs[i], t); | |
2359 | ||
fe8ac82f | 2360 | ne = split_block (e->dest, last_nondebug_stmt (e->dest)); |
29e0ad45 JJ |
2361 | gsi2 = gsi_after_labels (ne->dest); |
2362 | ||
47de0b56 | 2363 | expand_omp_build_cond (&gsi2, fd->loops[i].cond_code, vs[i], n2); |
29e0ad45 JJ |
2364 | edge e3, e4; |
2365 | if (next_bb == entry_bb) | |
2366 | { | |
2367 | e3 = find_edge (ne->dest, next_bb); | |
2368 | e3->flags = EDGE_FALSE_VALUE; | |
2369 | } | |
2370 | else | |
2371 | e3 = make_edge (ne->dest, next_bb, EDGE_FALSE_VALUE); | |
2372 | e4 = make_edge (ne->dest, new_cur_bb, EDGE_TRUE_VALUE); | |
2373 | e4->probability = profile_probability::likely ().guessed (); | |
2374 | e3->probability = e4->probability.invert (); | |
2375 | basic_block esrc = e->src; | |
2376 | make_edge (e->src, ne->dest, EDGE_FALLTHRU); | |
2377 | cur_bb = new_cur_bb; | |
2378 | basic_block latch_bb = next_bb; | |
2379 | next_bb = e->dest; | |
2380 | remove_edge (e); | |
2381 | set_immediate_dominator (CDI_DOMINATORS, ne->dest, esrc); | |
2382 | set_immediate_dominator (CDI_DOMINATORS, latch_bb, ne->dest); | |
2383 | set_immediate_dominator (CDI_DOMINATORS, cur_bb, ne->dest); | |
aed3ab25 | 2384 | } |
aed3ab25 JJ |
2385 | } |
2386 | t = NULL_TREE; | |
2387 | for (i = fd->first_nonrect; i < fd->last_nonrect; i++) | |
2388 | if (!fd->loops[i].non_rect_referenced | |
2389 | && fd->loops[i].m1 == NULL_TREE | |
2390 | && fd->loops[i].m2 == NULL_TREE) | |
2391 | { | |
2392 | if (t == NULL_TREE) | |
2393 | t = counts[i]; | |
2394 | else | |
2395 | t = fold_build2 (MULT_EXPR, type, t, counts[i]); | |
2396 | } | |
2397 | if (t) | |
2398 | { | |
2399 | t = fold_build2 (MULT_EXPR, type, counts[fd->last_nonrect], t); | |
2400 | expand_omp_build_assign (gsi, counts[fd->last_nonrect], t); | |
2401 | } | |
2402 | if (!rect_count_seen) | |
2403 | t = counts[fd->last_nonrect]; | |
2404 | else | |
2405 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, | |
2406 | counts[fd->last_nonrect]); | |
2407 | expand_omp_build_assign (gsi, fd->loop.n2, t); | |
2408 | } | |
2409 | else if (fd->non_rect) | |
2410 | { | |
2411 | tree t = fd->loop.n2; | |
2412 | gcc_assert (TREE_CODE (t) == INTEGER_CST); | |
2413 | int non_rect_referenced = 0, non_rect = 0; | |
2414 | for (i = 0; i < fd->collapse; i++) | |
2415 | { | |
5acef69f | 2416 | if ((i < fd->first_nonrect || i > fd->last_nonrect) |
aed3ab25 JJ |
2417 | && !integer_zerop (counts[i])) |
2418 | t = fold_build2 (TRUNC_DIV_EXPR, type, t, counts[i]); | |
2419 | if (fd->loops[i].non_rect_referenced) | |
2420 | non_rect_referenced++; | |
2421 | if (fd->loops[i].m1 || fd->loops[i].m2) | |
2422 | non_rect++; | |
2423 | } | |
2424 | gcc_assert (non_rect == 1 && non_rect_referenced == 1); | |
2425 | counts[fd->last_nonrect] = t; | |
2426 | } | |
629b3d75 MJ |
2427 | } |
2428 | ||
2429 | /* Helper function for expand_omp_{for_*,simd}. Generate code like: | |
2430 | T = V; | |
2431 | V3 = N31 + (T % count3) * STEP3; | |
2432 | T = T / count3; | |
2433 | V2 = N21 + (T % count2) * STEP2; | |
2434 | T = T / count2; | |
2435 | V1 = N11 + T * STEP1; | |
2436 | if this loop doesn't have an inner loop construct combined with it. | |
2437 | If it does have an inner loop construct combined with it and the | |
2438 | iteration count isn't known constant, store values from counts array | |
aed3ab25 JJ |
2439 | into its _looptemp_ temporaries instead. |
2440 | For non-rectangular loops (between fd->first_nonrect and fd->last_nonrect | |
2441 | inclusive), use the count of all those loops together, and either | |
c2ebf4f1 | 2442 | find quadratic etc. equation roots, or as a fallback, do: |
aed3ab25 JJ |
2443 | COUNT = 0; |
2444 | for (tmpi = N11; tmpi COND1 N12; tmpi += STEP1) | |
2445 | for (tmpj = M21 * tmpi + N21; | |
2446 | tmpj COND2 M22 * tmpi + N22; tmpj += STEP2) | |
2447 | { | |
2448 | int tmpk1 = M31 * tmpj + N31; | |
2449 | int tmpk2 = M32 * tmpj + N32; | |
2450 | if (tmpk1 COND3 tmpk2) | |
2451 | { | |
2452 | if (COND3 is <) | |
2453 | adj = STEP3 - 1; | |
2454 | else | |
2455 | adj = STEP3 + 1; | |
2456 | int temp = (adj + tmpk2 - tmpk1) / STEP3; | |
2457 | if (COUNT + temp > T) | |
2458 | { | |
2459 | V1 = tmpi; | |
2460 | V2 = tmpj; | |
2461 | V3 = tmpk1 + (T - COUNT) * STEP3; | |
2462 | goto done; | |
2463 | } | |
2464 | else | |
2465 | COUNT += temp; | |
2466 | } | |
2467 | } | |
2468 | done:; | |
2469 | but for optional innermost or outermost rectangular loops that aren't | |
2470 | referenced by other loop expressions keep doing the division/modulo. */ | |
629b3d75 MJ |
2471 | |
2472 | static void | |
2473 | expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
aed3ab25 JJ |
2474 | tree *counts, tree *nonrect_bounds, |
2475 | gimple *inner_stmt, tree startvar) | |
629b3d75 MJ |
2476 | { |
2477 | int i; | |
2478 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
2479 | { | |
2480 | /* If fd->loop.n2 is constant, then no propagation of the counts | |
2481 | is needed, they are constant. */ | |
2482 | if (TREE_CODE (fd->loop.n2) == INTEGER_CST) | |
2483 | return; | |
2484 | ||
2485 | tree clauses = gimple_code (inner_stmt) != GIMPLE_OMP_FOR | |
2486 | ? gimple_omp_taskreg_clauses (inner_stmt) | |
2487 | : gimple_omp_for_clauses (inner_stmt); | |
2488 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
2489 | isn't supposed to be handled, as the inner loop doesn't | |
2490 | use it. */ | |
2491 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
2492 | gcc_assert (innerc); | |
14707c89 JJ |
2493 | int count = 0; |
2494 | if (fd->non_rect | |
2495 | && fd->last_nonrect == fd->first_nonrect + 1 | |
2496 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) | |
2497 | count = 4; | |
2498 | for (i = 0; i < fd->collapse + count; i++) | |
629b3d75 MJ |
2499 | { |
2500 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
2501 | OMP_CLAUSE__LOOPTEMP_); | |
2502 | gcc_assert (innerc); | |
2503 | if (i) | |
2504 | { | |
2505 | tree tem = OMP_CLAUSE_DECL (innerc); | |
14707c89 JJ |
2506 | tree t; |
2507 | if (i < fd->collapse) | |
2508 | t = counts[i]; | |
2509 | else | |
2510 | switch (i - fd->collapse) | |
2511 | { | |
2512 | case 0: t = counts[0]; break; | |
2513 | case 1: t = fd->first_inner_iterations; break; | |
2514 | case 2: t = fd->factor; break; | |
2515 | case 3: t = fd->adjn1; break; | |
2516 | default: gcc_unreachable (); | |
2517 | } | |
2518 | t = fold_convert (TREE_TYPE (tem), t); | |
629b3d75 MJ |
2519 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, |
2520 | false, GSI_CONTINUE_LINKING); | |
2521 | gassign *stmt = gimple_build_assign (tem, t); | |
2522 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2523 | } | |
2524 | } | |
2525 | return; | |
2526 | } | |
2527 | ||
2528 | tree type = TREE_TYPE (fd->loop.v); | |
2529 | tree tem = create_tmp_reg (type, ".tem"); | |
2530 | gassign *stmt = gimple_build_assign (tem, startvar); | |
2531 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2532 | ||
2533 | for (i = fd->collapse - 1; i >= 0; i--) | |
2534 | { | |
2535 | tree vtype = TREE_TYPE (fd->loops[i].v), itype, t; | |
2536 | itype = vtype; | |
2537 | if (POINTER_TYPE_P (vtype)) | |
2538 | itype = signed_type_for (vtype); | |
aed3ab25 | 2539 | if (i != 0 && (i != fd->last_nonrect || fd->first_nonrect)) |
629b3d75 MJ |
2540 | t = fold_build2 (TRUNC_MOD_EXPR, type, tem, counts[i]); |
2541 | else | |
2542 | t = tem; | |
aed3ab25 JJ |
2543 | if (i == fd->last_nonrect) |
2544 | { | |
aed3ab25 JJ |
2545 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, |
2546 | false, GSI_CONTINUE_LINKING); | |
2547 | tree stopval = t; | |
2548 | tree idx = create_tmp_reg (type, ".count"); | |
2549 | expand_omp_build_assign (gsi, idx, | |
2550 | build_zero_cst (type), true); | |
29e0ad45 | 2551 | basic_block bb_triang = NULL, bb_triang_dom = NULL; |
5acef69f | 2552 | if (fd->first_nonrect + 1 == fd->last_nonrect |
325714b4 | 2553 | && (TREE_CODE (fd->loop.n2) == INTEGER_CST |
14707c89 | 2554 | || fd->first_inner_iterations) |
5acef69f | 2555 | && (optab_handler (sqrt_optab, TYPE_MODE (double_type_node)) |
ba009860 JJ |
2556 | != CODE_FOR_nothing) |
2557 | && !integer_zerop (fd->loop.n2)) | |
5acef69f | 2558 | { |
f418bd4b | 2559 | tree outer_n1 = fd->adjn1 ? fd->adjn1 : fd->loops[i - 1].n1; |
5acef69f | 2560 | tree itype = TREE_TYPE (fd->loops[i].v); |
79c12969 | 2561 | tree first_inner_iterations = fd->first_inner_iterations; |
5acef69f JJ |
2562 | tree factor = fd->factor; |
2563 | gcond *cond_stmt | |
47de0b56 | 2564 | = expand_omp_build_cond (gsi, NE_EXPR, factor, |
1dc65003 TB |
2565 | build_zero_cst (TREE_TYPE (factor)), |
2566 | true); | |
5acef69f JJ |
2567 | edge e = split_block (gsi_bb (*gsi), cond_stmt); |
2568 | basic_block bb0 = e->src; | |
2569 | e->flags = EDGE_TRUE_VALUE; | |
2570 | e->probability = profile_probability::likely (); | |
29e0ad45 | 2571 | bb_triang_dom = bb0; |
5acef69f JJ |
2572 | *gsi = gsi_after_labels (e->dest); |
2573 | tree slltype = long_long_integer_type_node; | |
2574 | tree ulltype = long_long_unsigned_type_node; | |
2575 | tree stopvalull = fold_convert (ulltype, stopval); | |
2576 | stopvalull | |
2577 | = force_gimple_operand_gsi (gsi, stopvalull, true, NULL_TREE, | |
2578 | false, GSI_CONTINUE_LINKING); | |
79c12969 JJ |
2579 | first_inner_iterations |
2580 | = fold_convert (slltype, first_inner_iterations); | |
2581 | first_inner_iterations | |
2582 | = force_gimple_operand_gsi (gsi, first_inner_iterations, true, | |
5acef69f JJ |
2583 | NULL_TREE, false, |
2584 | GSI_CONTINUE_LINKING); | |
2585 | factor = fold_convert (slltype, factor); | |
2586 | factor | |
2587 | = force_gimple_operand_gsi (gsi, factor, true, NULL_TREE, | |
2588 | false, GSI_CONTINUE_LINKING); | |
79c12969 | 2589 | tree first_inner_iterationsd |
5acef69f | 2590 | = fold_build1 (FLOAT_EXPR, double_type_node, |
79c12969 JJ |
2591 | first_inner_iterations); |
2592 | first_inner_iterationsd | |
2593 | = force_gimple_operand_gsi (gsi, first_inner_iterationsd, true, | |
5acef69f JJ |
2594 | NULL_TREE, false, |
2595 | GSI_CONTINUE_LINKING); | |
2596 | tree factord = fold_build1 (FLOAT_EXPR, double_type_node, | |
2597 | factor); | |
2598 | factord = force_gimple_operand_gsi (gsi, factord, true, | |
2599 | NULL_TREE, false, | |
2600 | GSI_CONTINUE_LINKING); | |
2601 | tree stopvald = fold_build1 (FLOAT_EXPR, double_type_node, | |
2602 | stopvalull); | |
2603 | stopvald = force_gimple_operand_gsi (gsi, stopvald, true, | |
2604 | NULL_TREE, false, | |
2605 | GSI_CONTINUE_LINKING); | |
2606 | /* Temporarily disable flag_rounding_math, values will be | |
2607 | decimal numbers divided by 2 and worst case imprecisions | |
2608 | due to too large values ought to be caught later by the | |
2609 | checks for fallback. */ | |
2610 | int save_flag_rounding_math = flag_rounding_math; | |
2611 | flag_rounding_math = 0; | |
2612 | t = fold_build2 (RDIV_EXPR, double_type_node, factord, | |
2613 | build_real (double_type_node, dconst2)); | |
2614 | tree t3 = fold_build2 (MINUS_EXPR, double_type_node, | |
79c12969 | 2615 | first_inner_iterationsd, t); |
5acef69f JJ |
2616 | t3 = force_gimple_operand_gsi (gsi, t3, true, NULL_TREE, false, |
2617 | GSI_CONTINUE_LINKING); | |
2618 | t = fold_build2 (MULT_EXPR, double_type_node, factord, | |
2619 | build_real (double_type_node, dconst2)); | |
2620 | t = fold_build2 (MULT_EXPR, double_type_node, t, stopvald); | |
2621 | t = fold_build2 (PLUS_EXPR, double_type_node, t, | |
2622 | fold_build2 (MULT_EXPR, double_type_node, | |
2623 | t3, t3)); | |
2624 | flag_rounding_math = save_flag_rounding_math; | |
2625 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, | |
2626 | GSI_CONTINUE_LINKING); | |
87d6dae3 JJ |
2627 | if (flag_exceptions |
2628 | && cfun->can_throw_non_call_exceptions | |
2629 | && operation_could_trap_p (LT_EXPR, true, false, NULL_TREE)) | |
2630 | { | |
2631 | tree tem = fold_build2 (LT_EXPR, boolean_type_node, t, | |
2632 | build_zero_cst (double_type_node)); | |
2633 | tem = force_gimple_operand_gsi (gsi, tem, true, NULL_TREE, | |
2634 | false, GSI_CONTINUE_LINKING); | |
2635 | cond_stmt = gimple_build_cond (NE_EXPR, tem, | |
2636 | boolean_false_node, | |
2637 | NULL_TREE, NULL_TREE); | |
2638 | } | |
2639 | else | |
2640 | cond_stmt | |
2641 | = gimple_build_cond (LT_EXPR, t, | |
2642 | build_zero_cst (double_type_node), | |
2643 | NULL_TREE, NULL_TREE); | |
5acef69f JJ |
2644 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); |
2645 | e = split_block (gsi_bb (*gsi), cond_stmt); | |
2646 | basic_block bb1 = e->src; | |
2647 | e->flags = EDGE_FALSE_VALUE; | |
2648 | e->probability = profile_probability::very_likely (); | |
2649 | *gsi = gsi_after_labels (e->dest); | |
2650 | gcall *call = gimple_build_call_internal (IFN_SQRT, 1, t); | |
2651 | tree sqrtr = create_tmp_var (double_type_node); | |
2652 | gimple_call_set_lhs (call, sqrtr); | |
2653 | gsi_insert_after (gsi, call, GSI_CONTINUE_LINKING); | |
2654 | t = fold_build2 (MINUS_EXPR, double_type_node, sqrtr, t3); | |
2655 | t = fold_build2 (RDIV_EXPR, double_type_node, t, factord); | |
2656 | t = fold_build1 (FIX_TRUNC_EXPR, ulltype, t); | |
2657 | tree c = create_tmp_var (ulltype); | |
2658 | tree d = create_tmp_var (ulltype); | |
2659 | expand_omp_build_assign (gsi, c, t, true); | |
2660 | t = fold_build2 (MINUS_EXPR, ulltype, c, | |
2661 | build_one_cst (ulltype)); | |
2662 | t = fold_build2 (MULT_EXPR, ulltype, c, t); | |
2663 | t = fold_build2 (RSHIFT_EXPR, ulltype, t, integer_one_node); | |
325714b4 JJ |
2664 | t = fold_build2 (MULT_EXPR, ulltype, |
2665 | fold_convert (ulltype, fd->factor), t); | |
2666 | tree t2 | |
2667 | = fold_build2 (MULT_EXPR, ulltype, c, | |
2668 | fold_convert (ulltype, | |
2669 | fd->first_inner_iterations)); | |
5acef69f JJ |
2670 | t = fold_build2 (PLUS_EXPR, ulltype, t, t2); |
2671 | expand_omp_build_assign (gsi, d, t, true); | |
325714b4 JJ |
2672 | t = fold_build2 (MULT_EXPR, ulltype, |
2673 | fold_convert (ulltype, fd->factor), c); | |
5acef69f | 2674 | t = fold_build2 (PLUS_EXPR, ulltype, |
325714b4 JJ |
2675 | t, fold_convert (ulltype, |
2676 | fd->first_inner_iterations)); | |
5acef69f JJ |
2677 | t2 = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, |
2678 | GSI_CONTINUE_LINKING); | |
2679 | cond_stmt = gimple_build_cond (GE_EXPR, stopvalull, d, | |
2680 | NULL_TREE, NULL_TREE); | |
2681 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
2682 | e = split_block (gsi_bb (*gsi), cond_stmt); | |
2683 | basic_block bb2 = e->src; | |
2684 | e->flags = EDGE_TRUE_VALUE; | |
2685 | e->probability = profile_probability::very_likely (); | |
2686 | *gsi = gsi_after_labels (e->dest); | |
2687 | t = fold_build2 (PLUS_EXPR, ulltype, d, t2); | |
2688 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, | |
2689 | GSI_CONTINUE_LINKING); | |
2690 | cond_stmt = gimple_build_cond (GE_EXPR, stopvalull, t, | |
2691 | NULL_TREE, NULL_TREE); | |
2692 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
2693 | e = split_block (gsi_bb (*gsi), cond_stmt); | |
2694 | basic_block bb3 = e->src; | |
2695 | e->flags = EDGE_FALSE_VALUE; | |
2696 | e->probability = profile_probability::very_likely (); | |
2697 | *gsi = gsi_after_labels (e->dest); | |
2698 | t = fold_convert (itype, c); | |
2699 | t = fold_build2 (MULT_EXPR, itype, t, fd->loops[i - 1].step); | |
f418bd4b | 2700 | t = fold_build2 (PLUS_EXPR, itype, outer_n1, t); |
5acef69f JJ |
2701 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, |
2702 | GSI_CONTINUE_LINKING); | |
2703 | expand_omp_build_assign (gsi, fd->loops[i - 1].v, t, true); | |
2704 | t2 = fold_build2 (MINUS_EXPR, ulltype, stopvalull, d); | |
2705 | t2 = fold_convert (itype, t2); | |
2706 | t2 = fold_build2 (MULT_EXPR, itype, t2, fd->loops[i].step); | |
2707 | t2 = fold_build2 (PLUS_EXPR, itype, t2, fd->loops[i].n1); | |
2708 | if (fd->loops[i].m1) | |
2709 | { | |
2710 | t = fold_build2 (MULT_EXPR, itype, t, fd->loops[i].m1); | |
2711 | t2 = fold_build2 (PLUS_EXPR, itype, t2, t); | |
2712 | } | |
2713 | expand_omp_build_assign (gsi, fd->loops[i].v, t2, true); | |
2714 | e = split_block (gsi_bb (*gsi), gsi_stmt (*gsi)); | |
2715 | bb_triang = e->src; | |
2716 | *gsi = gsi_after_labels (e->dest); | |
2717 | remove_edge (e); | |
2718 | e = make_edge (bb1, gsi_bb (*gsi), EDGE_TRUE_VALUE); | |
2719 | e->probability = profile_probability::very_unlikely (); | |
2720 | e = make_edge (bb2, gsi_bb (*gsi), EDGE_FALSE_VALUE); | |
2721 | e->probability = profile_probability::very_unlikely (); | |
2722 | e = make_edge (bb3, gsi_bb (*gsi), EDGE_TRUE_VALUE); | |
2723 | e->probability = profile_probability::very_unlikely (); | |
2724 | ||
2725 | basic_block bb4 = create_empty_bb (bb0); | |
2726 | add_bb_to_loop (bb4, bb0->loop_father); | |
2727 | e = make_edge (bb0, bb4, EDGE_FALSE_VALUE); | |
2728 | e->probability = profile_probability::unlikely (); | |
2729 | make_edge (bb4, gsi_bb (*gsi), EDGE_FALLTHRU); | |
2730 | set_immediate_dominator (CDI_DOMINATORS, bb4, bb0); | |
2731 | set_immediate_dominator (CDI_DOMINATORS, gsi_bb (*gsi), bb0); | |
2732 | gimple_stmt_iterator gsi2 = gsi_after_labels (bb4); | |
2733 | t2 = fold_build2 (TRUNC_DIV_EXPR, type, | |
2734 | counts[i], counts[i - 1]); | |
2735 | t2 = force_gimple_operand_gsi (&gsi2, t2, true, NULL_TREE, false, | |
2736 | GSI_CONTINUE_LINKING); | |
2737 | t = fold_build2 (TRUNC_MOD_EXPR, type, stopval, t2); | |
2738 | t2 = fold_build2 (TRUNC_DIV_EXPR, type, stopval, t2); | |
2739 | t = fold_convert (itype, t); | |
2740 | t2 = fold_convert (itype, t2); | |
2741 | t = fold_build2 (MULT_EXPR, itype, t, | |
2742 | fold_convert (itype, fd->loops[i].step)); | |
2743 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); | |
2744 | t2 = fold_build2 (MULT_EXPR, itype, t2, | |
2745 | fold_convert (itype, fd->loops[i - 1].step)); | |
2746 | t2 = fold_build2 (PLUS_EXPR, itype, fd->loops[i - 1].n1, t2); | |
2747 | t2 = force_gimple_operand_gsi (&gsi2, t2, false, NULL_TREE, | |
2748 | false, GSI_CONTINUE_LINKING); | |
2749 | stmt = gimple_build_assign (fd->loops[i - 1].v, t2); | |
2750 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
2751 | if (fd->loops[i].m1) | |
2752 | { | |
2753 | t2 = fold_build2 (MULT_EXPR, itype, fd->loops[i].m1, | |
2754 | fd->loops[i - 1].v); | |
2755 | t = fold_build2 (PLUS_EXPR, itype, t, t2); | |
2756 | } | |
2757 | t = force_gimple_operand_gsi (&gsi2, t, false, NULL_TREE, | |
2758 | false, GSI_CONTINUE_LINKING); | |
2759 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
2760 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
2761 | } | |
2762 | /* Fallback implementation. Evaluate the loops in between | |
2763 | (inclusive) fd->first_nonrect and fd->last_nonrect at | |
2764 | runtime unsing temporaries instead of the original iteration | |
2765 | variables, in the body just bump the counter and compare | |
2766 | with the desired value. */ | |
aed3ab25 JJ |
2767 | gimple_stmt_iterator gsi2 = *gsi; |
2768 | basic_block entry_bb = gsi_bb (gsi2); | |
2769 | edge e = split_block (entry_bb, gsi_stmt (gsi2)); | |
2770 | e = split_block (e->dest, (gimple *) NULL); | |
2771 | basic_block dom_bb = NULL; | |
2772 | basic_block cur_bb = e->src; | |
2773 | basic_block next_bb = e->dest; | |
2774 | entry_bb = e->dest; | |
2775 | *gsi = gsi_after_labels (entry_bb); | |
2776 | ||
2777 | tree *vs = XALLOCAVEC (tree, fd->last_nonrect); | |
2778 | tree n1 = NULL_TREE, n2 = NULL_TREE; | |
2779 | memset (vs, 0, fd->last_nonrect * sizeof (tree)); | |
2780 | ||
2781 | for (int j = fd->first_nonrect; j <= fd->last_nonrect; j++) | |
2782 | { | |
2783 | tree itype = TREE_TYPE (fd->loops[j].v); | |
2784 | bool rect_p = (fd->loops[j].m1 == NULL_TREE | |
2785 | && fd->loops[j].m2 == NULL_TREE | |
2786 | && !fd->loops[j].non_rect_referenced); | |
2787 | gsi2 = gsi_after_labels (cur_bb); | |
2788 | t = fold_convert (itype, unshare_expr (fd->loops[j].n1)); | |
2084b5f4 JJ |
2789 | if (fd->loops[j].m1 == NULL_TREE) |
2790 | n1 = rect_p ? build_zero_cst (type) : t; | |
2791 | else if (POINTER_TYPE_P (itype)) | |
2792 | { | |
2793 | gcc_assert (integer_onep (fd->loops[j].m1)); | |
4796d16d | 2794 | t = unshare_expr (fd->loops[j].n1); |
2084b5f4 JJ |
2795 | n1 = fold_build_pointer_plus (vs[j - fd->loops[j].outer], t); |
2796 | } | |
2797 | else | |
aed3ab25 JJ |
2798 | { |
2799 | n1 = fold_convert (itype, unshare_expr (fd->loops[j].m1)); | |
2800 | n1 = fold_build2 (MULT_EXPR, itype, | |
2801 | vs[j - fd->loops[j].outer], n1); | |
2802 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); | |
2803 | } | |
aed3ab25 JJ |
2804 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, |
2805 | true, GSI_SAME_STMT); | |
2806 | if (j < fd->last_nonrect) | |
2807 | { | |
2808 | vs[j] = create_tmp_reg (rect_p ? type : itype, ".it"); | |
2809 | expand_omp_build_assign (&gsi2, vs[j], n1); | |
2810 | } | |
2811 | t = fold_convert (itype, unshare_expr (fd->loops[j].n2)); | |
2084b5f4 JJ |
2812 | if (fd->loops[j].m2 == NULL_TREE) |
2813 | n2 = rect_p ? counts[j] : t; | |
2814 | else if (POINTER_TYPE_P (itype)) | |
2815 | { | |
2816 | gcc_assert (integer_onep (fd->loops[j].m2)); | |
4796d16d | 2817 | t = unshare_expr (fd->loops[j].n2); |
2084b5f4 JJ |
2818 | n2 = fold_build_pointer_plus (vs[j - fd->loops[j].outer], t); |
2819 | } | |
2820 | else | |
aed3ab25 JJ |
2821 | { |
2822 | n2 = fold_convert (itype, unshare_expr (fd->loops[j].m2)); | |
2823 | n2 = fold_build2 (MULT_EXPR, itype, | |
2824 | vs[j - fd->loops[j].outer], n2); | |
2825 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); | |
2826 | } | |
aed3ab25 JJ |
2827 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, |
2828 | true, GSI_SAME_STMT); | |
2084b5f4 JJ |
2829 | if (POINTER_TYPE_P (itype)) |
2830 | itype = signed_type_for (itype); | |
aed3ab25 JJ |
2831 | if (j == fd->last_nonrect) |
2832 | { | |
2833 | gcond *cond_stmt | |
47de0b56 JJ |
2834 | = expand_omp_build_cond (&gsi2, fd->loops[i].cond_code, |
2835 | n1, n2); | |
aed3ab25 JJ |
2836 | e = split_block (cur_bb, cond_stmt); |
2837 | e->flags = EDGE_TRUE_VALUE; | |
2838 | edge ne = make_edge (cur_bb, next_bb, EDGE_FALSE_VALUE); | |
2839 | e->probability = profile_probability::likely ().guessed (); | |
2840 | ne->probability = e->probability.invert (); | |
2841 | gsi2 = gsi_after_labels (e->dest); | |
2842 | ||
2843 | t = build_int_cst (itype, (fd->loops[j].cond_code == LT_EXPR | |
2844 | ? -1 : 1)); | |
2845 | t = fold_build2 (PLUS_EXPR, itype, | |
2846 | fold_convert (itype, fd->loops[j].step), t); | |
2084b5f4 JJ |
2847 | t = fold_build2 (PLUS_EXPR, itype, t, |
2848 | fold_convert (itype, n2)); | |
2849 | t = fold_build2 (MINUS_EXPR, itype, t, | |
2850 | fold_convert (itype, n1)); | |
aed3ab25 JJ |
2851 | tree step = fold_convert (itype, fd->loops[j].step); |
2852 | if (TYPE_UNSIGNED (itype) | |
2853 | && fd->loops[j].cond_code == GT_EXPR) | |
2854 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
2855 | fold_build1 (NEGATE_EXPR, itype, t), | |
2856 | fold_build1 (NEGATE_EXPR, itype, step)); | |
2857 | else | |
2858 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
2859 | t = fold_convert (type, t); | |
2860 | t = fold_build2 (PLUS_EXPR, type, idx, t); | |
2861 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2862 | true, GSI_SAME_STMT); | |
2863 | e = make_edge (e->dest, next_bb, EDGE_FALLTHRU); | |
2864 | set_immediate_dominator (CDI_DOMINATORS, next_bb, cur_bb); | |
2865 | cond_stmt | |
2866 | = gimple_build_cond (LE_EXPR, t, stopval, NULL_TREE, | |
2867 | NULL_TREE); | |
2868 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2869 | e = split_block (gsi_bb (gsi2), cond_stmt); | |
2870 | e->flags = EDGE_TRUE_VALUE; | |
2871 | e->probability = profile_probability::likely ().guessed (); | |
2872 | ne = make_edge (e->src, entry_bb, EDGE_FALSE_VALUE); | |
2873 | ne->probability = e->probability.invert (); | |
2874 | gsi2 = gsi_after_labels (e->dest); | |
2875 | expand_omp_build_assign (&gsi2, idx, t); | |
2876 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, dom_bb); | |
2877 | break; | |
2878 | } | |
fe8ac82f | 2879 | e = split_block (cur_bb, last_nondebug_stmt (cur_bb)); |
aed3ab25 JJ |
2880 | |
2881 | basic_block new_cur_bb = create_empty_bb (cur_bb); | |
2882 | add_bb_to_loop (new_cur_bb, cur_bb->loop_father); | |
2883 | ||
2884 | gsi2 = gsi_after_labels (e->dest); | |
2885 | if (rect_p) | |
2886 | t = fold_build2 (PLUS_EXPR, type, vs[j], | |
2887 | build_one_cst (type)); | |
2888 | else | |
2889 | { | |
2890 | tree step | |
2891 | = fold_convert (itype, unshare_expr (fd->loops[j].step)); | |
2084b5f4 | 2892 | if (POINTER_TYPE_P (vtype)) |
4796d16d | 2893 | t = fold_build_pointer_plus (vs[j], step); |
2084b5f4 JJ |
2894 | else |
2895 | t = fold_build2 (PLUS_EXPR, itype, vs[j], step); | |
aed3ab25 JJ |
2896 | } |
2897 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2898 | true, GSI_SAME_STMT); | |
2899 | expand_omp_build_assign (&gsi2, vs[j], t); | |
2900 | ||
fe8ac82f | 2901 | edge ne = split_block (e->dest, last_nondebug_stmt (e->dest)); |
aed3ab25 JJ |
2902 | gsi2 = gsi_after_labels (ne->dest); |
2903 | ||
2904 | gcond *cond_stmt; | |
2905 | if (next_bb == entry_bb) | |
2906 | /* No need to actually check the outermost condition. */ | |
2907 | cond_stmt | |
2908 | = gimple_build_cond (EQ_EXPR, boolean_true_node, | |
2909 | boolean_true_node, | |
2910 | NULL_TREE, NULL_TREE); | |
2911 | else | |
2912 | cond_stmt | |
2913 | = gimple_build_cond (rect_p ? LT_EXPR | |
2914 | : fd->loops[j].cond_code, | |
2915 | vs[j], n2, NULL_TREE, NULL_TREE); | |
2916 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); | |
2917 | edge e3, e4; | |
2918 | if (next_bb == entry_bb) | |
2919 | { | |
2920 | e3 = find_edge (ne->dest, next_bb); | |
2921 | e3->flags = EDGE_FALSE_VALUE; | |
2922 | dom_bb = ne->dest; | |
2923 | } | |
2924 | else | |
2925 | e3 = make_edge (ne->dest, next_bb, EDGE_FALSE_VALUE); | |
2926 | e4 = make_edge (ne->dest, new_cur_bb, EDGE_TRUE_VALUE); | |
2927 | e4->probability = profile_probability::likely ().guessed (); | |
2928 | e3->probability = e4->probability.invert (); | |
2929 | basic_block esrc = e->src; | |
2930 | make_edge (e->src, ne->dest, EDGE_FALLTHRU); | |
2931 | cur_bb = new_cur_bb; | |
2932 | basic_block latch_bb = next_bb; | |
2933 | next_bb = e->dest; | |
2934 | remove_edge (e); | |
2935 | set_immediate_dominator (CDI_DOMINATORS, ne->dest, esrc); | |
2936 | set_immediate_dominator (CDI_DOMINATORS, latch_bb, ne->dest); | |
2937 | set_immediate_dominator (CDI_DOMINATORS, cur_bb, ne->dest); | |
2938 | } | |
2939 | for (int j = fd->last_nonrect; j >= fd->first_nonrect; j--) | |
2940 | { | |
2084b5f4 JJ |
2941 | tree vtype = TREE_TYPE (fd->loops[j].v); |
2942 | tree itype = vtype; | |
2943 | if (POINTER_TYPE_P (itype)) | |
2944 | itype = signed_type_for (itype); | |
aed3ab25 JJ |
2945 | bool rect_p = (fd->loops[j].m1 == NULL_TREE |
2946 | && fd->loops[j].m2 == NULL_TREE | |
2947 | && !fd->loops[j].non_rect_referenced); | |
2948 | if (j == fd->last_nonrect) | |
2949 | { | |
2950 | t = fold_build2 (MINUS_EXPR, type, stopval, idx); | |
2951 | t = fold_convert (itype, t); | |
2952 | tree t2 | |
2953 | = fold_convert (itype, unshare_expr (fd->loops[j].step)); | |
2954 | t = fold_build2 (MULT_EXPR, itype, t, t2); | |
2084b5f4 | 2955 | if (POINTER_TYPE_P (vtype)) |
4796d16d | 2956 | t = fold_build_pointer_plus (n1, t); |
2084b5f4 JJ |
2957 | else |
2958 | t = fold_build2 (PLUS_EXPR, itype, n1, t); | |
aed3ab25 JJ |
2959 | } |
2960 | else if (rect_p) | |
2961 | { | |
2962 | t = fold_convert (itype, vs[j]); | |
2963 | t = fold_build2 (MULT_EXPR, itype, t, | |
2964 | fold_convert (itype, fd->loops[j].step)); | |
2965 | if (POINTER_TYPE_P (vtype)) | |
4796d16d | 2966 | t = fold_build_pointer_plus (fd->loops[j].n1, t); |
aed3ab25 JJ |
2967 | else |
2968 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[j].n1, t); | |
2969 | } | |
2970 | else | |
2971 | t = vs[j]; | |
2972 | t = force_gimple_operand_gsi (gsi, t, false, | |
2973 | NULL_TREE, true, | |
2974 | GSI_SAME_STMT); | |
2975 | stmt = gimple_build_assign (fd->loops[j].v, t); | |
2976 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); | |
2977 | } | |
2978 | if (gsi_end_p (*gsi)) | |
2979 | *gsi = gsi_last_bb (gsi_bb (*gsi)); | |
2980 | else | |
2981 | gsi_prev (gsi); | |
5acef69f JJ |
2982 | if (bb_triang) |
2983 | { | |
2984 | e = split_block (gsi_bb (*gsi), gsi_stmt (*gsi)); | |
2985 | make_edge (bb_triang, e->dest, EDGE_FALLTHRU); | |
2986 | *gsi = gsi_after_labels (e->dest); | |
2987 | if (!gsi_end_p (*gsi)) | |
2988 | gsi_insert_before (gsi, gimple_build_nop (), GSI_NEW_STMT); | |
29e0ad45 | 2989 | set_immediate_dominator (CDI_DOMINATORS, e->dest, bb_triang_dom); |
5acef69f | 2990 | } |
aed3ab25 | 2991 | } |
629b3d75 | 2992 | else |
aed3ab25 JJ |
2993 | { |
2994 | t = fold_convert (itype, t); | |
2995 | t = fold_build2 (MULT_EXPR, itype, t, | |
2996 | fold_convert (itype, fd->loops[i].step)); | |
2997 | if (POINTER_TYPE_P (vtype)) | |
2998 | t = fold_build_pointer_plus (fd->loops[i].n1, t); | |
2999 | else | |
3000 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); | |
3001 | t = force_gimple_operand_gsi (gsi, t, | |
3002 | DECL_P (fd->loops[i].v) | |
3003 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
3004 | NULL_TREE, false, | |
3005 | GSI_CONTINUE_LINKING); | |
3006 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
3007 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
3008 | } | |
3009 | if (i != 0 && (i != fd->last_nonrect || fd->first_nonrect)) | |
629b3d75 MJ |
3010 | { |
3011 | t = fold_build2 (TRUNC_DIV_EXPR, type, tem, counts[i]); | |
3012 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
3013 | false, GSI_CONTINUE_LINKING); | |
3014 | stmt = gimple_build_assign (tem, t); | |
3015 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
3016 | } | |
aed3ab25 JJ |
3017 | if (i == fd->last_nonrect) |
3018 | i = fd->first_nonrect; | |
629b3d75 | 3019 | } |
aed3ab25 JJ |
3020 | if (fd->non_rect) |
3021 | for (i = 0; i <= fd->last_nonrect; i++) | |
3022 | if (fd->loops[i].m2) | |
3023 | { | |
3024 | tree itype = TREE_TYPE (fd->loops[i].v); | |
3025 | ||
2084b5f4 JJ |
3026 | tree t; |
3027 | if (POINTER_TYPE_P (itype)) | |
3028 | { | |
3029 | gcc_assert (integer_onep (fd->loops[i].m2)); | |
2084b5f4 | 3030 | t = fold_build_pointer_plus (fd->loops[i - fd->loops[i].outer].v, |
4796d16d | 3031 | unshare_expr (fd->loops[i].n2)); |
2084b5f4 JJ |
3032 | } |
3033 | else | |
3034 | { | |
3035 | t = fold_convert (itype, unshare_expr (fd->loops[i].m2)); | |
3036 | t = fold_build2 (MULT_EXPR, itype, | |
3037 | fd->loops[i - fd->loops[i].outer].v, t); | |
3038 | t = fold_build2 (PLUS_EXPR, itype, t, | |
3039 | fold_convert (itype, | |
3040 | unshare_expr (fd->loops[i].n2))); | |
3041 | } | |
aed3ab25 JJ |
3042 | nonrect_bounds[i] = create_tmp_reg (itype, ".bound"); |
3043 | t = force_gimple_operand_gsi (gsi, t, false, | |
3044 | NULL_TREE, false, | |
3045 | GSI_CONTINUE_LINKING); | |
3046 | stmt = gimple_build_assign (nonrect_bounds[i], t); | |
3047 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
3048 | } | |
629b3d75 MJ |
3049 | } |
3050 | ||
3051 | /* Helper function for expand_omp_for_*. Generate code like: | |
3052 | L10: | |
3053 | V3 += STEP3; | |
3054 | if (V3 cond3 N32) goto BODY_BB; else goto L11; | |
3055 | L11: | |
3056 | V3 = N31; | |
3057 | V2 += STEP2; | |
3058 | if (V2 cond2 N22) goto BODY_BB; else goto L12; | |
3059 | L12: | |
3060 | V2 = N21; | |
3061 | V1 += STEP1; | |
aed3ab25 JJ |
3062 | goto BODY_BB; |
3063 | For non-rectangular loops, use temporaries stored in nonrect_bounds | |
3064 | for the upper bounds if M?2 multiplier is present. Given e.g. | |
3065 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
3066 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
3067 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
3068 | for (V4 = N41 + M41 * V2; V4 cond4 N42 + M42 * V2; V4 += STEP4) | |
3069 | do: | |
3070 | L10: | |
3071 | V4 += STEP4; | |
3072 | if (V4 cond4 NONRECT_BOUND4) goto BODY_BB; else goto L11; | |
3073 | L11: | |
3074 | V4 = N41 + M41 * V2; // This can be left out if the loop | |
3075 | // refers to the immediate parent loop | |
3076 | V3 += STEP3; | |
3077 | if (V3 cond3 N32) goto BODY_BB; else goto L12; | |
3078 | L12: | |
3079 | V3 = N31; | |
3080 | V2 += STEP2; | |
3081 | if (V2 cond2 N22) goto L120; else goto L13; | |
3082 | L120: | |
3083 | V4 = N41 + M41 * V2; | |
3084 | NONRECT_BOUND4 = N42 + M42 * V2; | |
3085 | if (V4 cond4 NONRECT_BOUND4) goto BODY_BB; else goto L12; | |
3086 | L13: | |
3087 | V2 = N21; | |
3088 | V1 += STEP1; | |
3089 | goto L120; */ | |
629b3d75 MJ |
3090 | |
3091 | static basic_block | |
aed3ab25 JJ |
3092 | extract_omp_for_update_vars (struct omp_for_data *fd, tree *nonrect_bounds, |
3093 | basic_block cont_bb, basic_block body_bb) | |
629b3d75 MJ |
3094 | { |
3095 | basic_block last_bb, bb, collapse_bb = NULL; | |
3096 | int i; | |
3097 | gimple_stmt_iterator gsi; | |
3098 | edge e; | |
3099 | tree t; | |
3100 | gimple *stmt; | |
3101 | ||
3102 | last_bb = cont_bb; | |
3103 | for (i = fd->collapse - 1; i >= 0; i--) | |
3104 | { | |
3105 | tree vtype = TREE_TYPE (fd->loops[i].v); | |
3106 | ||
3107 | bb = create_empty_bb (last_bb); | |
3108 | add_bb_to_loop (bb, last_bb->loop_father); | |
3109 | gsi = gsi_start_bb (bb); | |
3110 | ||
3111 | if (i < fd->collapse - 1) | |
3112 | { | |
3113 | e = make_edge (last_bb, bb, EDGE_FALSE_VALUE); | |
9f55aee9 | 3114 | e->probability = profile_probability::guessed_always () / 8; |
629b3d75 | 3115 | |
aed3ab25 JJ |
3116 | struct omp_for_data_loop *l = &fd->loops[i + 1]; |
3117 | if (l->m1 == NULL_TREE || l->outer != 1) | |
3118 | { | |
3119 | t = l->n1; | |
3120 | if (l->m1) | |
3121 | { | |
2084b5f4 JJ |
3122 | if (POINTER_TYPE_P (TREE_TYPE (l->v))) |
3123 | t = fold_build_pointer_plus (fd->loops[i + 1 - l->outer].v, | |
4796d16d | 3124 | t); |
2084b5f4 JJ |
3125 | else |
3126 | { | |
3127 | tree t2 | |
3128 | = fold_build2 (MULT_EXPR, TREE_TYPE (t), | |
3129 | fd->loops[i + 1 - l->outer].v, l->m1); | |
3130 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t2, t); | |
3131 | } | |
aed3ab25 JJ |
3132 | } |
3133 | t = force_gimple_operand_gsi (&gsi, t, | |
3134 | DECL_P (l->v) | |
3135 | && TREE_ADDRESSABLE (l->v), | |
3136 | NULL_TREE, false, | |
3137 | GSI_CONTINUE_LINKING); | |
3138 | stmt = gimple_build_assign (l->v, t); | |
3139 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
3140 | } | |
629b3d75 MJ |
3141 | } |
3142 | else | |
3143 | collapse_bb = bb; | |
3144 | ||
3145 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); | |
3146 | ||
3147 | if (POINTER_TYPE_P (vtype)) | |
3148 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); | |
3149 | else | |
3150 | t = fold_build2 (PLUS_EXPR, vtype, fd->loops[i].v, fd->loops[i].step); | |
3151 | t = force_gimple_operand_gsi (&gsi, t, | |
3152 | DECL_P (fd->loops[i].v) | |
3153 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
3154 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
3155 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
3156 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
3157 | ||
aed3ab25 JJ |
3158 | if (fd->loops[i].non_rect_referenced) |
3159 | { | |
3160 | basic_block update_bb = NULL, prev_bb = NULL; | |
3161 | for (int j = i + 1; j <= fd->last_nonrect; j++) | |
3162 | if (j - fd->loops[j].outer == i) | |
3163 | { | |
3164 | tree n1, n2; | |
3165 | struct omp_for_data_loop *l = &fd->loops[j]; | |
3166 | basic_block this_bb = create_empty_bb (last_bb); | |
3167 | add_bb_to_loop (this_bb, last_bb->loop_father); | |
3168 | gimple_stmt_iterator gsi2 = gsi_start_bb (this_bb); | |
3169 | if (prev_bb) | |
3170 | { | |
3171 | e = make_edge (prev_bb, this_bb, EDGE_TRUE_VALUE); | |
3172 | e->probability | |
3173 | = profile_probability::guessed_always ().apply_scale (7, | |
3174 | 8); | |
3175 | set_immediate_dominator (CDI_DOMINATORS, this_bb, prev_bb); | |
aed3ab25 JJ |
3176 | } |
3177 | if (l->m1) | |
3178 | { | |
2084b5f4 | 3179 | if (POINTER_TYPE_P (TREE_TYPE (l->v))) |
4796d16d | 3180 | t = fold_build_pointer_plus (fd->loops[i].v, l->n1); |
2084b5f4 JJ |
3181 | else |
3182 | { | |
3183 | t = fold_build2 (MULT_EXPR, TREE_TYPE (l->m1), l->m1, | |
3184 | fd->loops[i].v); | |
3185 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (l->v), | |
3186 | t, l->n1); | |
3187 | } | |
aed3ab25 JJ |
3188 | n1 = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3189 | false, | |
3190 | GSI_CONTINUE_LINKING); | |
3191 | stmt = gimple_build_assign (l->v, n1); | |
3192 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
3193 | n1 = l->v; | |
3194 | } | |
3195 | else | |
3196 | n1 = force_gimple_operand_gsi (&gsi2, l->n1, true, | |
3197 | NULL_TREE, false, | |
3198 | GSI_CONTINUE_LINKING); | |
3199 | if (l->m2) | |
3200 | { | |
2084b5f4 | 3201 | if (POINTER_TYPE_P (TREE_TYPE (l->v))) |
4796d16d | 3202 | t = fold_build_pointer_plus (fd->loops[i].v, l->n2); |
2084b5f4 JJ |
3203 | else |
3204 | { | |
3205 | t = fold_build2 (MULT_EXPR, TREE_TYPE (l->m2), l->m2, | |
3206 | fd->loops[i].v); | |
3207 | t = fold_build2 (PLUS_EXPR, | |
3208 | TREE_TYPE (nonrect_bounds[j]), | |
3209 | t, unshare_expr (l->n2)); | |
3210 | } | |
aed3ab25 JJ |
3211 | n2 = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3212 | false, | |
3213 | GSI_CONTINUE_LINKING); | |
3214 | stmt = gimple_build_assign (nonrect_bounds[j], n2); | |
3215 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); | |
3216 | n2 = nonrect_bounds[j]; | |
3217 | } | |
3218 | else | |
3219 | n2 = force_gimple_operand_gsi (&gsi2, unshare_expr (l->n2), | |
3220 | true, NULL_TREE, false, | |
3221 | GSI_CONTINUE_LINKING); | |
3222 | gcond *cond_stmt | |
3223 | = gimple_build_cond (l->cond_code, n1, n2, | |
3224 | NULL_TREE, NULL_TREE); | |
3225 | gsi_insert_after (&gsi2, cond_stmt, GSI_CONTINUE_LINKING); | |
3226 | if (update_bb == NULL) | |
3227 | update_bb = this_bb; | |
3228 | e = make_edge (this_bb, bb, EDGE_FALSE_VALUE); | |
9f55aee9 | 3229 | e->probability = profile_probability::guessed_always () / 8; |
aed3ab25 | 3230 | if (prev_bb == NULL) |
29e0ad45 | 3231 | set_immediate_dominator (CDI_DOMINATORS, this_bb, bb); |
aed3ab25 JJ |
3232 | prev_bb = this_bb; |
3233 | } | |
3234 | e = make_edge (prev_bb, body_bb, EDGE_TRUE_VALUE); | |
3235 | e->probability | |
3236 | = profile_probability::guessed_always ().apply_scale (7, 8); | |
3237 | body_bb = update_bb; | |
3238 | } | |
3239 | ||
629b3d75 MJ |
3240 | if (i > 0) |
3241 | { | |
aed3ab25 JJ |
3242 | if (fd->loops[i].m2) |
3243 | t = nonrect_bounds[i]; | |
3244 | else | |
3245 | t = unshare_expr (fd->loops[i].n2); | |
629b3d75 MJ |
3246 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
3247 | false, GSI_CONTINUE_LINKING); | |
3248 | tree v = fd->loops[i].v; | |
3249 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
3250 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
3251 | false, GSI_CONTINUE_LINKING); | |
3252 | t = fold_build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
3253 | stmt = gimple_build_cond_empty (t); | |
3254 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
d1ffbd43 JJ |
3255 | if (walk_tree (gimple_cond_lhs_ptr (as_a <gcond *> (stmt)), |
3256 | expand_omp_regimplify_p, NULL, NULL) | |
3257 | || walk_tree (gimple_cond_rhs_ptr (as_a <gcond *> (stmt)), | |
3258 | expand_omp_regimplify_p, NULL, NULL)) | |
3259 | gimple_regimplify_operands (stmt, &gsi); | |
629b3d75 | 3260 | e = make_edge (bb, body_bb, EDGE_TRUE_VALUE); |
357067f2 | 3261 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); |
629b3d75 MJ |
3262 | } |
3263 | else | |
3264 | make_edge (bb, body_bb, EDGE_FALLTHRU); | |
aed3ab25 | 3265 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); |
629b3d75 MJ |
3266 | last_bb = bb; |
3267 | } | |
3268 | ||
3269 | return collapse_bb; | |
3270 | } | |
3271 | ||
3272 | /* Expand #pragma omp ordered depend(source). */ | |
3273 | ||
3274 | static void | |
3275 | expand_omp_ordered_source (gimple_stmt_iterator *gsi, struct omp_for_data *fd, | |
3276 | tree *counts, location_t loc) | |
3277 | { | |
3278 | enum built_in_function source_ix | |
3279 | = fd->iter_type == long_integer_type_node | |
3280 | ? BUILT_IN_GOMP_DOACROSS_POST : BUILT_IN_GOMP_DOACROSS_ULL_POST; | |
3281 | gimple *g | |
3282 | = gimple_build_call (builtin_decl_explicit (source_ix), 1, | |
3283 | build_fold_addr_expr (counts[fd->ordered])); | |
3284 | gimple_set_location (g, loc); | |
3285 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
3286 | } | |
3287 | ||
3288 | /* Expand a single depend from #pragma omp ordered depend(sink:...). */ | |
3289 | ||
3290 | static void | |
3291 | expand_omp_ordered_sink (gimple_stmt_iterator *gsi, struct omp_for_data *fd, | |
f25a6767 JJ |
3292 | tree *counts, tree c, location_t loc, |
3293 | basic_block cont_bb) | |
629b3d75 MJ |
3294 | { |
3295 | auto_vec<tree, 10> args; | |
3296 | enum built_in_function sink_ix | |
3297 | = fd->iter_type == long_integer_type_node | |
3298 | ? BUILT_IN_GOMP_DOACROSS_WAIT : BUILT_IN_GOMP_DOACROSS_ULL_WAIT; | |
3299 | tree t, off, coff = NULL_TREE, deps = OMP_CLAUSE_DECL (c), cond = NULL_TREE; | |
3300 | int i; | |
3301 | gimple_stmt_iterator gsi2 = *gsi; | |
3302 | bool warned_step = false; | |
3303 | ||
a651e6d5 JJ |
3304 | if (deps == NULL) |
3305 | { | |
f25a6767 JJ |
3306 | /* Handle doacross(sink: omp_cur_iteration - 1). */ |
3307 | gsi_prev (&gsi2); | |
3308 | edge e1 = split_block (gsi_bb (gsi2), gsi_stmt (gsi2)); | |
3309 | edge e2 = split_block_after_labels (e1->dest); | |
3310 | gsi2 = gsi_after_labels (e1->dest); | |
3311 | *gsi = gsi_last_bb (e1->src); | |
3312 | gimple_stmt_iterator gsi3 = *gsi; | |
3313 | ||
3314 | if (counts[fd->collapse - 1]) | |
3315 | { | |
3316 | gcc_assert (fd->collapse == 1); | |
3317 | t = counts[fd->collapse - 1]; | |
3318 | } | |
3319 | else if (fd->collapse > 1) | |
3320 | t = fd->loop.v; | |
3321 | else | |
3322 | { | |
3323 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), | |
3324 | fd->loops[0].v, fd->loops[0].n1); | |
3325 | t = fold_convert (fd->iter_type, t); | |
3326 | } | |
3327 | ||
3328 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, | |
3329 | false, GSI_CONTINUE_LINKING); | |
3330 | gsi_insert_after (gsi, gimple_build_cond (NE_EXPR, t, | |
3331 | build_zero_cst (TREE_TYPE (t)), | |
3332 | NULL_TREE, NULL_TREE), | |
3333 | GSI_NEW_STMT); | |
3334 | ||
3335 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
3336 | build_minus_one_cst (TREE_TYPE (t))); | |
3337 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
3338 | true, GSI_SAME_STMT); | |
3339 | args.safe_push (t); | |
3340 | for (i = fd->collapse; i < fd->ordered; i++) | |
3341 | { | |
3342 | t = counts[fd->ordered + 2 + (i - fd->collapse)]; | |
3343 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
3344 | build_minus_one_cst (TREE_TYPE (t))); | |
3345 | t = fold_convert (fd->iter_type, t); | |
3346 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
3347 | true, GSI_SAME_STMT); | |
3348 | args.safe_push (t); | |
3349 | } | |
3350 | ||
3351 | gimple *g = gimple_build_call_vec (builtin_decl_explicit (sink_ix), | |
3352 | args); | |
3353 | gimple_set_location (g, loc); | |
3354 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); | |
3355 | ||
3356 | edge e3 = make_edge (e1->src, e2->dest, EDGE_FALSE_VALUE); | |
3357 | e3->probability = profile_probability::guessed_always () / 8; | |
3358 | e1->probability = e3->probability.invert (); | |
3359 | e1->flags = EDGE_TRUE_VALUE; | |
3360 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e1->src); | |
3361 | ||
3362 | if (fd->ordered > fd->collapse && cont_bb) | |
3363 | { | |
3364 | if (counts[fd->ordered + 1] == NULL_TREE) | |
3365 | counts[fd->ordered + 1] | |
3366 | = create_tmp_var (boolean_type_node, ".first"); | |
3367 | ||
3368 | edge e4; | |
3369 | if (gsi_end_p (gsi3)) | |
3370 | e4 = split_block_after_labels (e1->src); | |
3371 | else | |
3372 | { | |
3373 | gsi_prev (&gsi3); | |
3374 | e4 = split_block (gsi_bb (gsi3), gsi_stmt (gsi3)); | |
3375 | } | |
3376 | gsi3 = gsi_last_bb (e4->src); | |
3377 | ||
3378 | gsi_insert_after (&gsi3, | |
3379 | gimple_build_cond (NE_EXPR, | |
3380 | counts[fd->ordered + 1], | |
3381 | boolean_false_node, | |
3382 | NULL_TREE, NULL_TREE), | |
3383 | GSI_NEW_STMT); | |
3384 | ||
3385 | edge e5 = make_edge (e4->src, e2->dest, EDGE_FALSE_VALUE); | |
3386 | e4->probability = profile_probability::guessed_always () / 8; | |
3387 | e5->probability = e4->probability.invert (); | |
3388 | e4->flags = EDGE_TRUE_VALUE; | |
3389 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e4->src); | |
3390 | } | |
3391 | ||
3392 | *gsi = gsi_after_labels (e2->dest); | |
a651e6d5 JJ |
3393 | return; |
3394 | } | |
629b3d75 MJ |
3395 | for (i = 0; i < fd->ordered; i++) |
3396 | { | |
3397 | tree step = NULL_TREE; | |
3398 | off = TREE_PURPOSE (deps); | |
3399 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) | |
3400 | { | |
3401 | step = TREE_OPERAND (off, 1); | |
3402 | off = TREE_OPERAND (off, 0); | |
3403 | } | |
3404 | if (!integer_zerop (off)) | |
3405 | { | |
3406 | gcc_assert (fd->loops[i].cond_code == LT_EXPR | |
3407 | || fd->loops[i].cond_code == GT_EXPR); | |
3408 | bool forward = fd->loops[i].cond_code == LT_EXPR; | |
3409 | if (step) | |
3410 | { | |
3411 | /* Non-simple Fortran DO loops. If step is variable, | |
3412 | we don't know at compile even the direction, so can't | |
3413 | warn. */ | |
3414 | if (TREE_CODE (step) != INTEGER_CST) | |
3415 | break; | |
3416 | forward = tree_int_cst_sgn (step) != -1; | |
3417 | } | |
a651e6d5 | 3418 | if (forward ^ OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
449b6b81 TB |
3419 | warning_at (loc, OPT_Wopenmp, |
3420 | "%qs clause with %<sink%> modifier " | |
3421 | "waiting for lexically later iteration", | |
a651e6d5 JJ |
3422 | OMP_CLAUSE_DOACROSS_DEPEND (c) |
3423 | ? "depend" : "doacross"); | |
629b3d75 MJ |
3424 | break; |
3425 | } | |
3426 | deps = TREE_CHAIN (deps); | |
3427 | } | |
3428 | /* If all offsets corresponding to the collapsed loops are zero, | |
3429 | this depend clause can be ignored. FIXME: but there is still a | |
3430 | flush needed. We need to emit one __sync_synchronize () for it | |
3431 | though (perhaps conditionally)? Solve this together with the | |
3432 | conservative dependence folding optimization. | |
3433 | if (i >= fd->collapse) | |
3434 | return; */ | |
3435 | ||
3436 | deps = OMP_CLAUSE_DECL (c); | |
3437 | gsi_prev (&gsi2); | |
3438 | edge e1 = split_block (gsi_bb (gsi2), gsi_stmt (gsi2)); | |
3439 | edge e2 = split_block_after_labels (e1->dest); | |
3440 | ||
3441 | gsi2 = gsi_after_labels (e1->dest); | |
3442 | *gsi = gsi_last_bb (e1->src); | |
3443 | for (i = 0; i < fd->ordered; i++) | |
3444 | { | |
3445 | tree itype = TREE_TYPE (fd->loops[i].v); | |
3446 | tree step = NULL_TREE; | |
3447 | tree orig_off = NULL_TREE; | |
3448 | if (POINTER_TYPE_P (itype)) | |
3449 | itype = sizetype; | |
3450 | if (i) | |
3451 | deps = TREE_CHAIN (deps); | |
3452 | off = TREE_PURPOSE (deps); | |
3453 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) | |
3454 | { | |
3455 | step = TREE_OPERAND (off, 1); | |
3456 | off = TREE_OPERAND (off, 0); | |
3457 | gcc_assert (fd->loops[i].cond_code == LT_EXPR | |
3458 | && integer_onep (fd->loops[i].step) | |
3459 | && !POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))); | |
3460 | } | |
3461 | tree s = fold_convert_loc (loc, itype, step ? step : fd->loops[i].step); | |
3462 | if (step) | |
3463 | { | |
3464 | off = fold_convert_loc (loc, itype, off); | |
3465 | orig_off = off; | |
3466 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); | |
3467 | } | |
3468 | ||
3469 | if (integer_zerop (off)) | |
3470 | t = boolean_true_node; | |
3471 | else | |
3472 | { | |
3473 | tree a; | |
3474 | tree co = fold_convert_loc (loc, itype, off); | |
3475 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
3476 | { | |
a651e6d5 | 3477 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
629b3d75 MJ |
3478 | co = fold_build1_loc (loc, NEGATE_EXPR, itype, co); |
3479 | a = fold_build2_loc (loc, POINTER_PLUS_EXPR, | |
3480 | TREE_TYPE (fd->loops[i].v), fd->loops[i].v, | |
3481 | co); | |
3482 | } | |
a651e6d5 | 3483 | else if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
629b3d75 MJ |
3484 | a = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), |
3485 | fd->loops[i].v, co); | |
3486 | else | |
3487 | a = fold_build2_loc (loc, PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
3488 | fd->loops[i].v, co); | |
3489 | if (step) | |
3490 | { | |
3491 | tree t1, t2; | |
a651e6d5 | 3492 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
629b3d75 MJ |
3493 | t1 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, |
3494 | fd->loops[i].n1); | |
3495 | else | |
3496 | t1 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
3497 | fd->loops[i].n2); | |
a651e6d5 | 3498 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
629b3d75 MJ |
3499 | t2 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, |
3500 | fd->loops[i].n2); | |
3501 | else | |
3502 | t2 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, | |
3503 | fd->loops[i].n1); | |
3504 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, | |
3505 | step, build_int_cst (TREE_TYPE (step), 0)); | |
3506 | if (TREE_CODE (step) != INTEGER_CST) | |
3507 | { | |
3508 | t1 = unshare_expr (t1); | |
3509 | t1 = force_gimple_operand_gsi (gsi, t1, true, NULL_TREE, | |
3510 | false, GSI_CONTINUE_LINKING); | |
3511 | t2 = unshare_expr (t2); | |
3512 | t2 = force_gimple_operand_gsi (gsi, t2, true, NULL_TREE, | |
3513 | false, GSI_CONTINUE_LINKING); | |
3514 | } | |
3515 | t = fold_build3_loc (loc, COND_EXPR, boolean_type_node, | |
3516 | t, t2, t1); | |
3517 | } | |
3518 | else if (fd->loops[i].cond_code == LT_EXPR) | |
3519 | { | |
a651e6d5 | 3520 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
629b3d75 MJ |
3521 | t = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, |
3522 | fd->loops[i].n1); | |
3523 | else | |
3524 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
3525 | fd->loops[i].n2); | |
3526 | } | |
a651e6d5 | 3527 | else if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
629b3d75 MJ |
3528 | t = fold_build2_loc (loc, GT_EXPR, boolean_type_node, a, |
3529 | fd->loops[i].n2); | |
3530 | else | |
3531 | t = fold_build2_loc (loc, LE_EXPR, boolean_type_node, a, | |
3532 | fd->loops[i].n1); | |
3533 | } | |
3534 | if (cond) | |
3535 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, cond, t); | |
3536 | else | |
3537 | cond = t; | |
3538 | ||
3539 | off = fold_convert_loc (loc, itype, off); | |
3540 | ||
3541 | if (step | |
3542 | || (fd->loops[i].cond_code == LT_EXPR | |
3543 | ? !integer_onep (fd->loops[i].step) | |
3544 | : !integer_minus_onep (fd->loops[i].step))) | |
3545 | { | |
3546 | if (step == NULL_TREE | |
3547 | && TYPE_UNSIGNED (itype) | |
3548 | && fd->loops[i].cond_code == GT_EXPR) | |
3549 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, off, | |
3550 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
3551 | s)); | |
3552 | else | |
3553 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, | |
3554 | orig_off ? orig_off : off, s); | |
3555 | t = fold_build2_loc (loc, EQ_EXPR, boolean_type_node, t, | |
3556 | build_int_cst (itype, 0)); | |
3557 | if (integer_zerop (t) && !warned_step) | |
3558 | { | |
449b6b81 TB |
3559 | warning_at (loc, OPT_Wopenmp, |
3560 | "%qs clause with %<sink%> modifier refers to " | |
3561 | "iteration never in the iteration space", | |
a651e6d5 JJ |
3562 | OMP_CLAUSE_DOACROSS_DEPEND (c) |
3563 | ? "depend" : "doacross"); | |
629b3d75 MJ |
3564 | warned_step = true; |
3565 | } | |
3566 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, | |
3567 | cond, t); | |
3568 | } | |
3569 | ||
3570 | if (i <= fd->collapse - 1 && fd->collapse > 1) | |
3571 | t = fd->loop.v; | |
3572 | else if (counts[i]) | |
3573 | t = counts[i]; | |
3574 | else | |
3575 | { | |
3576 | t = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
3577 | fd->loops[i].v, fd->loops[i].n1); | |
3578 | t = fold_convert_loc (loc, fd->iter_type, t); | |
3579 | } | |
3580 | if (step) | |
3581 | /* We have divided off by step already earlier. */; | |
3582 | else if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
3583 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, | |
3584 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
3585 | s)); | |
3586 | else | |
3587 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); | |
a651e6d5 | 3588 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
629b3d75 MJ |
3589 | off = fold_build1_loc (loc, NEGATE_EXPR, itype, off); |
3590 | off = fold_convert_loc (loc, fd->iter_type, off); | |
3591 | if (i <= fd->collapse - 1 && fd->collapse > 1) | |
3592 | { | |
3593 | if (i) | |
3594 | off = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, coff, | |
3595 | off); | |
3596 | if (i < fd->collapse - 1) | |
3597 | { | |
3598 | coff = fold_build2_loc (loc, MULT_EXPR, fd->iter_type, off, | |
3599 | counts[i]); | |
3600 | continue; | |
3601 | } | |
3602 | } | |
3603 | off = unshare_expr (off); | |
3604 | t = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, t, off); | |
3605 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
3606 | true, GSI_SAME_STMT); | |
3607 | args.safe_push (t); | |
3608 | } | |
3609 | gimple *g = gimple_build_call_vec (builtin_decl_explicit (sink_ix), args); | |
3610 | gimple_set_location (g, loc); | |
3611 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); | |
3612 | ||
3613 | cond = unshare_expr (cond); | |
3614 | cond = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE, false, | |
3615 | GSI_CONTINUE_LINKING); | |
3616 | gsi_insert_after (gsi, gimple_build_cond_empty (cond), GSI_NEW_STMT); | |
3617 | edge e3 = make_edge (e1->src, e2->dest, EDGE_FALSE_VALUE); | |
9f55aee9 | 3618 | e3->probability = profile_probability::guessed_always () / 8; |
357067f2 | 3619 | e1->probability = e3->probability.invert (); |
629b3d75 MJ |
3620 | e1->flags = EDGE_TRUE_VALUE; |
3621 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e1->src); | |
3622 | ||
3623 | *gsi = gsi_after_labels (e2->dest); | |
3624 | } | |
3625 | ||
3626 | /* Expand all #pragma omp ordered depend(source) and | |
3627 | #pragma omp ordered depend(sink:...) constructs in the current | |
3628 | #pragma omp for ordered(n) region. */ | |
3629 | ||
3630 | static void | |
3631 | expand_omp_ordered_source_sink (struct omp_region *region, | |
3632 | struct omp_for_data *fd, tree *counts, | |
3633 | basic_block cont_bb) | |
3634 | { | |
3635 | struct omp_region *inner; | |
3636 | int i; | |
3637 | for (i = fd->collapse - 1; i < fd->ordered; i++) | |
3638 | if (i == fd->collapse - 1 && fd->collapse > 1) | |
3639 | counts[i] = NULL_TREE; | |
3640 | else if (i >= fd->collapse && !cont_bb) | |
3641 | counts[i] = build_zero_cst (fd->iter_type); | |
3642 | else if (!POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v)) | |
3643 | && integer_onep (fd->loops[i].step)) | |
3644 | counts[i] = NULL_TREE; | |
3645 | else | |
3646 | counts[i] = create_tmp_var (fd->iter_type, ".orditer"); | |
3647 | tree atype | |
3648 | = build_array_type_nelts (fd->iter_type, fd->ordered - fd->collapse + 1); | |
3649 | counts[fd->ordered] = create_tmp_var (atype, ".orditera"); | |
3650 | TREE_ADDRESSABLE (counts[fd->ordered]) = 1; | |
f25a6767 | 3651 | counts[fd->ordered + 1] = NULL_TREE; |
629b3d75 MJ |
3652 | |
3653 | for (inner = region->inner; inner; inner = inner->next) | |
3654 | if (inner->type == GIMPLE_OMP_ORDERED) | |
3655 | { | |
3656 | gomp_ordered *ord_stmt = inner->ord_stmt; | |
3657 | gimple_stmt_iterator gsi = gsi_for_stmt (ord_stmt); | |
3658 | location_t loc = gimple_location (ord_stmt); | |
3659 | tree c; | |
3660 | for (c = gimple_omp_ordered_clauses (ord_stmt); | |
3661 | c; c = OMP_CLAUSE_CHAIN (c)) | |
a651e6d5 | 3662 | if (OMP_CLAUSE_DOACROSS_KIND (c) == OMP_CLAUSE_DOACROSS_SOURCE) |
629b3d75 MJ |
3663 | break; |
3664 | if (c) | |
3665 | expand_omp_ordered_source (&gsi, fd, counts, loc); | |
3666 | for (c = gimple_omp_ordered_clauses (ord_stmt); | |
3667 | c; c = OMP_CLAUSE_CHAIN (c)) | |
a651e6d5 | 3668 | if (OMP_CLAUSE_DOACROSS_KIND (c) == OMP_CLAUSE_DOACROSS_SINK) |
f25a6767 | 3669 | expand_omp_ordered_sink (&gsi, fd, counts, c, loc, cont_bb); |
629b3d75 MJ |
3670 | gsi_remove (&gsi, true); |
3671 | } | |
3672 | } | |
3673 | ||
3674 | /* Wrap the body into fd->ordered - fd->collapse loops that aren't | |
3675 | collapsed. */ | |
3676 | ||
3677 | static basic_block | |
3678 | expand_omp_for_ordered_loops (struct omp_for_data *fd, tree *counts, | |
3679 | basic_block cont_bb, basic_block body_bb, | |
713fa5db | 3680 | basic_block l0_bb, bool ordered_lastprivate) |
629b3d75 MJ |
3681 | { |
3682 | if (fd->ordered == fd->collapse) | |
3683 | return cont_bb; | |
3684 | ||
3685 | if (!cont_bb) | |
3686 | { | |
3687 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
3688 | for (int i = fd->collapse; i < fd->ordered; i++) | |
3689 | { | |
3690 | tree type = TREE_TYPE (fd->loops[i].v); | |
3691 | tree n1 = fold_convert (type, fd->loops[i].n1); | |
3692 | expand_omp_build_assign (&gsi, fd->loops[i].v, n1); | |
3693 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3694 | size_int (i - fd->collapse + 1), | |
3695 | NULL_TREE, NULL_TREE); | |
3696 | expand_omp_build_assign (&gsi, aref, build_zero_cst (fd->iter_type)); | |
3697 | } | |
3698 | return NULL; | |
3699 | } | |
3700 | ||
3701 | for (int i = fd->ordered - 1; i >= fd->collapse; i--) | |
3702 | { | |
3703 | tree t, type = TREE_TYPE (fd->loops[i].v); | |
3704 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
f25a6767 JJ |
3705 | if (counts[fd->ordered + 1] && i == fd->collapse) |
3706 | expand_omp_build_assign (&gsi, counts[fd->ordered + 1], | |
3707 | boolean_true_node); | |
629b3d75 MJ |
3708 | expand_omp_build_assign (&gsi, fd->loops[i].v, |
3709 | fold_convert (type, fd->loops[i].n1)); | |
3710 | if (counts[i]) | |
3711 | expand_omp_build_assign (&gsi, counts[i], | |
3712 | build_zero_cst (fd->iter_type)); | |
3713 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3714 | size_int (i - fd->collapse + 1), | |
3715 | NULL_TREE, NULL_TREE); | |
3716 | expand_omp_build_assign (&gsi, aref, build_zero_cst (fd->iter_type)); | |
3717 | if (!gsi_end_p (gsi)) | |
3718 | gsi_prev (&gsi); | |
3719 | else | |
3720 | gsi = gsi_last_bb (body_bb); | |
3721 | edge e1 = split_block (body_bb, gsi_stmt (gsi)); | |
3722 | basic_block new_body = e1->dest; | |
3723 | if (body_bb == cont_bb) | |
3724 | cont_bb = new_body; | |
3725 | edge e2 = NULL; | |
3726 | basic_block new_header; | |
3727 | if (EDGE_COUNT (cont_bb->preds) > 0) | |
3728 | { | |
3729 | gsi = gsi_last_bb (cont_bb); | |
3730 | if (POINTER_TYPE_P (type)) | |
4796d16d | 3731 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); |
629b3d75 MJ |
3732 | else |
3733 | t = fold_build2 (PLUS_EXPR, type, fd->loops[i].v, | |
3734 | fold_convert (type, fd->loops[i].step)); | |
3735 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
3736 | if (counts[i]) | |
3737 | { | |
3738 | t = fold_build2 (PLUS_EXPR, fd->iter_type, counts[i], | |
3739 | build_int_cst (fd->iter_type, 1)); | |
3740 | expand_omp_build_assign (&gsi, counts[i], t); | |
3741 | t = counts[i]; | |
3742 | } | |
3743 | else | |
3744 | { | |
3745 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
3746 | fd->loops[i].v, fd->loops[i].n1); | |
3747 | t = fold_convert (fd->iter_type, t); | |
3748 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
3749 | true, GSI_SAME_STMT); | |
3750 | } | |
3751 | aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3752 | size_int (i - fd->collapse + 1), | |
3753 | NULL_TREE, NULL_TREE); | |
3754 | expand_omp_build_assign (&gsi, aref, t); | |
f25a6767 JJ |
3755 | if (counts[fd->ordered + 1] && i == fd->ordered - 1) |
3756 | expand_omp_build_assign (&gsi, counts[fd->ordered + 1], | |
3757 | boolean_false_node); | |
629b3d75 MJ |
3758 | gsi_prev (&gsi); |
3759 | e2 = split_block (cont_bb, gsi_stmt (gsi)); | |
3760 | new_header = e2->dest; | |
3761 | } | |
3762 | else | |
3763 | new_header = cont_bb; | |
3764 | gsi = gsi_after_labels (new_header); | |
3765 | tree v = force_gimple_operand_gsi (&gsi, fd->loops[i].v, true, NULL_TREE, | |
3766 | true, GSI_SAME_STMT); | |
3767 | tree n2 | |
3768 | = force_gimple_operand_gsi (&gsi, fold_convert (type, fd->loops[i].n2), | |
3769 | true, NULL_TREE, true, GSI_SAME_STMT); | |
3770 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, n2); | |
3771 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_NEW_STMT); | |
3772 | edge e3 = split_block (new_header, gsi_stmt (gsi)); | |
3773 | cont_bb = e3->dest; | |
3774 | remove_edge (e1); | |
3775 | make_edge (body_bb, new_header, EDGE_FALLTHRU); | |
3776 | e3->flags = EDGE_FALSE_VALUE; | |
9f55aee9 | 3777 | e3->probability = profile_probability::guessed_always () / 8; |
629b3d75 | 3778 | e1 = make_edge (new_header, new_body, EDGE_TRUE_VALUE); |
357067f2 | 3779 | e1->probability = e3->probability.invert (); |
629b3d75 MJ |
3780 | |
3781 | set_immediate_dominator (CDI_DOMINATORS, new_header, body_bb); | |
3782 | set_immediate_dominator (CDI_DOMINATORS, new_body, new_header); | |
3783 | ||
3784 | if (e2) | |
3785 | { | |
99b1c316 | 3786 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
3787 | loop->header = new_header; |
3788 | loop->latch = e2->src; | |
713fa5db | 3789 | add_loop (loop, l0_bb->loop_father); |
629b3d75 MJ |
3790 | } |
3791 | } | |
3792 | ||
3793 | /* If there are any lastprivate clauses and it is possible some loops | |
3794 | might have zero iterations, ensure all the decls are initialized, | |
3795 | otherwise we could crash evaluating C++ class iterators with lastprivate | |
3796 | clauses. */ | |
3797 | bool need_inits = false; | |
3798 | for (int i = fd->collapse; ordered_lastprivate && i < fd->ordered; i++) | |
3799 | if (need_inits) | |
3800 | { | |
3801 | tree type = TREE_TYPE (fd->loops[i].v); | |
3802 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
3803 | expand_omp_build_assign (&gsi, fd->loops[i].v, | |
3804 | fold_convert (type, fd->loops[i].n1)); | |
3805 | } | |
3806 | else | |
3807 | { | |
3808 | tree type = TREE_TYPE (fd->loops[i].v); | |
3809 | tree this_cond = fold_build2 (fd->loops[i].cond_code, | |
3810 | boolean_type_node, | |
3811 | fold_convert (type, fd->loops[i].n1), | |
3812 | fold_convert (type, fd->loops[i].n2)); | |
3813 | if (!integer_onep (this_cond)) | |
3814 | need_inits = true; | |
3815 | } | |
3816 | ||
3817 | return cont_bb; | |
3818 | } | |
3819 | ||
3820 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
3821 | loop with any schedule. Given parameters: | |
3822 | ||
3823 | for (V = N1; V cond N2; V += STEP) BODY; | |
3824 | ||
3825 | where COND is "<" or ">", we generate pseudocode | |
3826 | ||
3827 | more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0); | |
3828 | if (more) goto L0; else goto L3; | |
3829 | L0: | |
3830 | V = istart0; | |
3831 | iend = iend0; | |
3832 | L1: | |
3833 | BODY; | |
3834 | V += STEP; | |
3835 | if (V cond iend) goto L1; else goto L2; | |
3836 | L2: | |
3837 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
3838 | L3: | |
3839 | ||
3840 | If this is a combined omp parallel loop, instead of the call to | |
3841 | GOMP_loop_foo_start, we call GOMP_loop_foo_next. | |
3842 | If this is gimple_omp_for_combined_p loop, then instead of assigning | |
3843 | V and iend in L0 we assign the first two _looptemp_ clause decls of the | |
3844 | inner GIMPLE_OMP_FOR and V += STEP; and | |
3845 | if (V cond iend) goto L1; else goto L2; are removed. | |
3846 | ||
3847 | For collapsed loops, given parameters: | |
3848 | collapse(3) | |
3849 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
3850 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
3851 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
3852 | BODY; | |
3853 | ||
3854 | we generate pseudocode | |
3855 | ||
3856 | if (__builtin_expect (N32 cond3 N31, 0)) goto Z0; | |
3857 | if (cond3 is <) | |
3858 | adj = STEP3 - 1; | |
3859 | else | |
3860 | adj = STEP3 + 1; | |
3861 | count3 = (adj + N32 - N31) / STEP3; | |
3862 | if (__builtin_expect (N22 cond2 N21, 0)) goto Z0; | |
3863 | if (cond2 is <) | |
3864 | adj = STEP2 - 1; | |
3865 | else | |
3866 | adj = STEP2 + 1; | |
3867 | count2 = (adj + N22 - N21) / STEP2; | |
3868 | if (__builtin_expect (N12 cond1 N11, 0)) goto Z0; | |
3869 | if (cond1 is <) | |
3870 | adj = STEP1 - 1; | |
3871 | else | |
3872 | adj = STEP1 + 1; | |
3873 | count1 = (adj + N12 - N11) / STEP1; | |
3874 | count = count1 * count2 * count3; | |
3875 | goto Z1; | |
3876 | Z0: | |
3877 | count = 0; | |
3878 | Z1: | |
3879 | more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0); | |
3880 | if (more) goto L0; else goto L3; | |
3881 | L0: | |
3882 | V = istart0; | |
3883 | T = V; | |
3884 | V3 = N31 + (T % count3) * STEP3; | |
3885 | T = T / count3; | |
3886 | V2 = N21 + (T % count2) * STEP2; | |
3887 | T = T / count2; | |
3888 | V1 = N11 + T * STEP1; | |
3889 | iend = iend0; | |
3890 | L1: | |
3891 | BODY; | |
3892 | V += 1; | |
3893 | if (V < iend) goto L10; else goto L2; | |
3894 | L10: | |
3895 | V3 += STEP3; | |
3896 | if (V3 cond3 N32) goto L1; else goto L11; | |
3897 | L11: | |
3898 | V3 = N31; | |
3899 | V2 += STEP2; | |
3900 | if (V2 cond2 N22) goto L1; else goto L12; | |
3901 | L12: | |
3902 | V2 = N21; | |
3903 | V1 += STEP1; | |
3904 | goto L1; | |
3905 | L2: | |
3906 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
3907 | L3: | |
3908 | ||
3909 | */ | |
3910 | ||
3911 | static void | |
3912 | expand_omp_for_generic (struct omp_region *region, | |
3913 | struct omp_for_data *fd, | |
3914 | enum built_in_function start_fn, | |
3915 | enum built_in_function next_fn, | |
28567c40 | 3916 | tree sched_arg, |
629b3d75 MJ |
3917 | gimple *inner_stmt) |
3918 | { | |
3919 | tree type, istart0, iend0, iend; | |
3920 | tree t, vmain, vback, bias = NULL_TREE; | |
3921 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, collapse_bb; | |
3922 | basic_block l2_bb = NULL, l3_bb = NULL; | |
3923 | gimple_stmt_iterator gsi; | |
3924 | gassign *assign_stmt; | |
3925 | bool in_combined_parallel = is_combined_parallel (region); | |
3926 | bool broken_loop = region->cont == NULL; | |
3927 | edge e, ne; | |
3928 | tree *counts = NULL; | |
3929 | int i; | |
3930 | bool ordered_lastprivate = false; | |
3931 | ||
3932 | gcc_assert (!broken_loop || !in_combined_parallel); | |
3933 | gcc_assert (fd->iter_type == long_integer_type_node | |
3934 | || !in_combined_parallel); | |
3935 | ||
3936 | entry_bb = region->entry; | |
3937 | cont_bb = region->cont; | |
3938 | collapse_bb = NULL; | |
3939 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
3940 | gcc_assert (broken_loop | |
3941 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
3942 | l0_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
3943 | l1_bb = single_succ (l0_bb); | |
3944 | if (!broken_loop) | |
3945 | { | |
3946 | l2_bb = create_empty_bb (cont_bb); | |
3947 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb | |
3948 | || (single_succ_edge (BRANCH_EDGE (cont_bb)->dest)->dest | |
3949 | == l1_bb)); | |
3950 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
3951 | } | |
3952 | else | |
3953 | l2_bb = NULL; | |
3954 | l3_bb = BRANCH_EDGE (entry_bb)->dest; | |
3955 | exit_bb = region->exit; | |
3956 | ||
65f4b875 | 3957 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
3958 | |
3959 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
3960 | if (fd->ordered | |
6c7ae8c5 | 3961 | && omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
629b3d75 MJ |
3962 | OMP_CLAUSE_LASTPRIVATE)) |
3963 | ordered_lastprivate = false; | |
28567c40 | 3964 | tree reductions = NULL_TREE; |
6c7ae8c5 JJ |
3965 | tree mem = NULL_TREE, cond_var = NULL_TREE, condtemp = NULL_TREE; |
3966 | tree memv = NULL_TREE; | |
8221c30b JJ |
3967 | if (fd->lastprivate_conditional) |
3968 | { | |
3969 | tree c = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
3970 | OMP_CLAUSE__CONDTEMP_); | |
3971 | if (fd->have_pointer_condtemp) | |
3972 | condtemp = OMP_CLAUSE_DECL (c); | |
3973 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
3974 | cond_var = OMP_CLAUSE_DECL (c); | |
3975 | } | |
28567c40 JJ |
3976 | if (sched_arg) |
3977 | { | |
3978 | if (fd->have_reductemp) | |
3979 | { | |
6c7ae8c5 | 3980 | tree c = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
28567c40 JJ |
3981 | OMP_CLAUSE__REDUCTEMP_); |
3982 | reductions = OMP_CLAUSE_DECL (c); | |
3983 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
3984 | gimple *g = SSA_NAME_DEF_STMT (reductions); | |
3985 | reductions = gimple_assign_rhs1 (g); | |
3986 | OMP_CLAUSE_DECL (c) = reductions; | |
3987 | entry_bb = gimple_bb (g); | |
3988 | edge e = split_block (entry_bb, g); | |
3989 | if (region->entry == entry_bb) | |
3990 | region->entry = e->dest; | |
3991 | gsi = gsi_last_bb (entry_bb); | |
3992 | } | |
3993 | else | |
3994 | reductions = null_pointer_node; | |
8221c30b | 3995 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 3996 | { |
6c7ae8c5 JJ |
3997 | tree type = TREE_TYPE (condtemp); |
3998 | memv = create_tmp_var (type); | |
3999 | TREE_ADDRESSABLE (memv) = 1; | |
4000 | unsigned HOST_WIDE_INT sz | |
4001 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
4002 | sz *= fd->lastprivate_conditional; | |
4003 | expand_omp_build_assign (&gsi, memv, build_int_cst (type, sz), | |
4004 | false); | |
4005 | mem = build_fold_addr_expr (memv); | |
4006 | } | |
4007 | else | |
4008 | mem = null_pointer_node; | |
28567c40 | 4009 | } |
629b3d75 MJ |
4010 | if (fd->collapse > 1 || fd->ordered) |
4011 | { | |
4012 | int first_zero_iter1 = -1, first_zero_iter2 = -1; | |
4013 | basic_block zero_iter1_bb = NULL, zero_iter2_bb = NULL, l2_dom_bb = NULL; | |
4014 | ||
f25a6767 JJ |
4015 | counts = XALLOCAVEC (tree, fd->ordered |
4016 | ? fd->ordered + 2 | |
4017 | + (fd->ordered - fd->collapse) | |
4018 | : fd->collapse); | |
629b3d75 MJ |
4019 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, |
4020 | zero_iter1_bb, first_zero_iter1, | |
4021 | zero_iter2_bb, first_zero_iter2, l2_dom_bb); | |
4022 | ||
4023 | if (zero_iter1_bb) | |
4024 | { | |
4025 | /* Some counts[i] vars might be uninitialized if | |
4026 | some loop has zero iterations. But the body shouldn't | |
4027 | be executed in that case, so just avoid uninit warnings. */ | |
4028 | for (i = first_zero_iter1; | |
4029 | i < (fd->ordered ? fd->ordered : fd->collapse); i++) | |
4030 | if (SSA_VAR_P (counts[i])) | |
e9e2bad7 | 4031 | suppress_warning (counts[i], OPT_Wuninitialized); |
629b3d75 MJ |
4032 | gsi_prev (&gsi); |
4033 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
4034 | entry_bb = e->dest; | |
4035 | make_edge (zero_iter1_bb, entry_bb, EDGE_FALLTHRU); | |
65f4b875 | 4036 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
4037 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, |
4038 | get_immediate_dominator (CDI_DOMINATORS, | |
4039 | zero_iter1_bb)); | |
4040 | } | |
4041 | if (zero_iter2_bb) | |
4042 | { | |
4043 | /* Some counts[i] vars might be uninitialized if | |
4044 | some loop has zero iterations. But the body shouldn't | |
4045 | be executed in that case, so just avoid uninit warnings. */ | |
4046 | for (i = first_zero_iter2; i < fd->ordered; i++) | |
4047 | if (SSA_VAR_P (counts[i])) | |
e9e2bad7 | 4048 | suppress_warning (counts[i], OPT_Wuninitialized); |
629b3d75 MJ |
4049 | if (zero_iter1_bb) |
4050 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); | |
4051 | else | |
4052 | { | |
4053 | gsi_prev (&gsi); | |
4054 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
4055 | entry_bb = e->dest; | |
4056 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); | |
65f4b875 | 4057 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
4058 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, |
4059 | get_immediate_dominator | |
4060 | (CDI_DOMINATORS, zero_iter2_bb)); | |
4061 | } | |
4062 | } | |
4063 | if (fd->collapse == 1) | |
4064 | { | |
4065 | counts[0] = fd->loop.n2; | |
4066 | fd->loop = fd->loops[0]; | |
4067 | } | |
4068 | } | |
4069 | ||
4070 | type = TREE_TYPE (fd->loop.v); | |
4071 | istart0 = create_tmp_var (fd->iter_type, ".istart0"); | |
4072 | iend0 = create_tmp_var (fd->iter_type, ".iend0"); | |
4073 | TREE_ADDRESSABLE (istart0) = 1; | |
4074 | TREE_ADDRESSABLE (iend0) = 1; | |
4075 | ||
4076 | /* See if we need to bias by LLONG_MIN. */ | |
4077 | if (fd->iter_type == long_long_unsigned_type_node | |
c8f10456 | 4078 | && (TREE_CODE (type) == INTEGER_TYPE || TREE_CODE (type) == BITINT_TYPE) |
629b3d75 MJ |
4079 | && !TYPE_UNSIGNED (type) |
4080 | && fd->ordered == 0) | |
4081 | { | |
4082 | tree n1, n2; | |
4083 | ||
4084 | if (fd->loop.cond_code == LT_EXPR) | |
4085 | { | |
4086 | n1 = fd->loop.n1; | |
4087 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
4088 | } | |
4089 | else | |
4090 | { | |
4091 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
4092 | n2 = fd->loop.n1; | |
4093 | } | |
4094 | if (TREE_CODE (n1) != INTEGER_CST | |
4095 | || TREE_CODE (n2) != INTEGER_CST | |
4096 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
4097 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
4098 | } | |
4099 | ||
4100 | gimple_stmt_iterator gsif = gsi; | |
4101 | gsi_prev (&gsif); | |
4102 | ||
4103 | tree arr = NULL_TREE; | |
4104 | if (in_combined_parallel) | |
4105 | { | |
4106 | gcc_assert (fd->ordered == 0); | |
4107 | /* In a combined parallel loop, emit a call to | |
4108 | GOMP_loop_foo_next. */ | |
4109 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, | |
4110 | build_fold_addr_expr (istart0), | |
4111 | build_fold_addr_expr (iend0)); | |
4112 | } | |
4113 | else | |
4114 | { | |
4115 | tree t0, t1, t2, t3, t4; | |
4116 | /* If this is not a combined parallel loop, emit a call to | |
4117 | GOMP_loop_foo_start in ENTRY_BB. */ | |
4118 | t4 = build_fold_addr_expr (iend0); | |
4119 | t3 = build_fold_addr_expr (istart0); | |
4120 | if (fd->ordered) | |
4121 | { | |
4122 | t0 = build_int_cst (unsigned_type_node, | |
4123 | fd->ordered - fd->collapse + 1); | |
4124 | arr = create_tmp_var (build_array_type_nelts (fd->iter_type, | |
4125 | fd->ordered | |
4126 | - fd->collapse + 1), | |
4127 | ".omp_counts"); | |
4128 | DECL_NAMELESS (arr) = 1; | |
4129 | TREE_ADDRESSABLE (arr) = 1; | |
4130 | TREE_STATIC (arr) = 1; | |
4131 | vec<constructor_elt, va_gc> *v; | |
4132 | vec_alloc (v, fd->ordered - fd->collapse + 1); | |
4133 | int idx; | |
4134 | ||
4135 | for (idx = 0; idx < fd->ordered - fd->collapse + 1; idx++) | |
4136 | { | |
4137 | tree c; | |
4138 | if (idx == 0 && fd->collapse > 1) | |
4139 | c = fd->loop.n2; | |
4140 | else | |
4141 | c = counts[idx + fd->collapse - 1]; | |
4142 | tree purpose = size_int (idx); | |
4143 | CONSTRUCTOR_APPEND_ELT (v, purpose, c); | |
4144 | if (TREE_CODE (c) != INTEGER_CST) | |
4145 | TREE_STATIC (arr) = 0; | |
4146 | } | |
4147 | ||
4148 | DECL_INITIAL (arr) = build_constructor (TREE_TYPE (arr), v); | |
4149 | if (!TREE_STATIC (arr)) | |
4150 | force_gimple_operand_gsi (&gsi, build1 (DECL_EXPR, | |
4151 | void_type_node, arr), | |
4152 | true, NULL_TREE, true, GSI_SAME_STMT); | |
4153 | t1 = build_fold_addr_expr (arr); | |
4154 | t2 = NULL_TREE; | |
4155 | } | |
4156 | else | |
4157 | { | |
4158 | t2 = fold_convert (fd->iter_type, fd->loop.step); | |
4159 | t1 = fd->loop.n2; | |
4160 | t0 = fd->loop.n1; | |
4161 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
4162 | { | |
4163 | tree innerc | |
4164 | = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
4165 | OMP_CLAUSE__LOOPTEMP_); | |
4166 | gcc_assert (innerc); | |
4167 | t0 = OMP_CLAUSE_DECL (innerc); | |
4168 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4169 | OMP_CLAUSE__LOOPTEMP_); | |
4170 | gcc_assert (innerc); | |
4171 | t1 = OMP_CLAUSE_DECL (innerc); | |
4172 | } | |
4173 | if (POINTER_TYPE_P (TREE_TYPE (t0)) | |
4174 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
4175 | != TYPE_PRECISION (fd->iter_type)) | |
4176 | { | |
4177 | /* Avoid casting pointers to integer of a different size. */ | |
4178 | tree itype = signed_type_for (type); | |
4179 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); | |
4180 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
4181 | } | |
4182 | else | |
4183 | { | |
4184 | t1 = fold_convert (fd->iter_type, t1); | |
4185 | t0 = fold_convert (fd->iter_type, t0); | |
4186 | } | |
4187 | if (bias) | |
4188 | { | |
4189 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); | |
4190 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
4191 | } | |
4192 | } | |
4193 | if (fd->iter_type == long_integer_type_node || fd->ordered) | |
4194 | { | |
4195 | if (fd->chunk_size) | |
4196 | { | |
4197 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
4198 | t = omp_adjust_chunk_size (t, fd->simd_schedule); | |
28567c40 JJ |
4199 | if (sched_arg) |
4200 | { | |
4201 | if (fd->ordered) | |
4202 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4203 | 8, t0, t1, sched_arg, t, t3, t4, | |
4204 | reductions, mem); | |
4205 | else | |
4206 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4207 | 9, t0, t1, t2, sched_arg, t, t3, t4, | |
4208 | reductions, mem); | |
4209 | } | |
4210 | else if (fd->ordered) | |
629b3d75 MJ |
4211 | t = build_call_expr (builtin_decl_explicit (start_fn), |
4212 | 5, t0, t1, t, t3, t4); | |
4213 | else | |
4214 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4215 | 6, t0, t1, t2, t, t3, t4); | |
4216 | } | |
4217 | else if (fd->ordered) | |
4218 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4219 | 4, t0, t1, t3, t4); | |
4220 | else | |
4221 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4222 | 5, t0, t1, t2, t3, t4); | |
4223 | } | |
4224 | else | |
4225 | { | |
4226 | tree t5; | |
4227 | tree c_bool_type; | |
4228 | tree bfn_decl; | |
4229 | ||
4230 | /* The GOMP_loop_ull_*start functions have additional boolean | |
4231 | argument, true for < loops and false for > loops. | |
4232 | In Fortran, the C bool type can be different from | |
4233 | boolean_type_node. */ | |
4234 | bfn_decl = builtin_decl_explicit (start_fn); | |
4235 | c_bool_type = TREE_TYPE (TREE_TYPE (bfn_decl)); | |
4236 | t5 = build_int_cst (c_bool_type, | |
4237 | fd->loop.cond_code == LT_EXPR ? 1 : 0); | |
4238 | if (fd->chunk_size) | |
4239 | { | |
4240 | tree bfn_decl = builtin_decl_explicit (start_fn); | |
4241 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
4242 | t = omp_adjust_chunk_size (t, fd->simd_schedule); | |
28567c40 JJ |
4243 | if (sched_arg) |
4244 | t = build_call_expr (bfn_decl, 10, t5, t0, t1, t2, sched_arg, | |
4245 | t, t3, t4, reductions, mem); | |
4246 | else | |
4247 | t = build_call_expr (bfn_decl, 7, t5, t0, t1, t2, t, t3, t4); | |
629b3d75 MJ |
4248 | } |
4249 | else | |
4250 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
4251 | 6, t5, t0, t1, t2, t3, t4); | |
4252 | } | |
4253 | } | |
4254 | if (TREE_TYPE (t) != boolean_type_node) | |
4255 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
4256 | t, build_int_cst (TREE_TYPE (t), 0)); | |
4257 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
01914336 | 4258 | true, GSI_SAME_STMT); |
629b3d75 MJ |
4259 | if (arr && !TREE_STATIC (arr)) |
4260 | { | |
25b45c7c | 4261 | tree clobber = build_clobber (TREE_TYPE (arr)); |
629b3d75 MJ |
4262 | gsi_insert_before (&gsi, gimple_build_assign (arr, clobber), |
4263 | GSI_SAME_STMT); | |
4264 | } | |
8221c30b | 4265 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 4266 | expand_omp_build_assign (&gsi, condtemp, memv, false); |
28567c40 JJ |
4267 | if (fd->have_reductemp) |
4268 | { | |
4269 | gimple *g = gsi_stmt (gsi); | |
4270 | gsi_remove (&gsi, true); | |
4271 | release_ssa_name (gimple_assign_lhs (g)); | |
4272 | ||
4273 | entry_bb = region->entry; | |
4274 | gsi = gsi_last_nondebug_bb (entry_bb); | |
4275 | ||
4276 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
4277 | } | |
629b3d75 MJ |
4278 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); |
4279 | ||
4280 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
4281 | gsi_remove (&gsi, true); | |
4282 | ||
4283 | if (gsi_end_p (gsif)) | |
4284 | gsif = gsi_after_labels (gsi_bb (gsif)); | |
4285 | gsi_next (&gsif); | |
4286 | ||
4287 | /* Iteration setup for sequential loop goes in L0_BB. */ | |
4288 | tree startvar = fd->loop.v; | |
4289 | tree endvar = NULL_TREE; | |
4290 | ||
4291 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4292 | { | |
4293 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_FOR | |
4294 | && gimple_omp_for_kind (inner_stmt) | |
4295 | == GF_OMP_FOR_KIND_SIMD); | |
4296 | tree innerc = omp_find_clause (gimple_omp_for_clauses (inner_stmt), | |
4297 | OMP_CLAUSE__LOOPTEMP_); | |
4298 | gcc_assert (innerc); | |
4299 | startvar = OMP_CLAUSE_DECL (innerc); | |
4300 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4301 | OMP_CLAUSE__LOOPTEMP_); | |
4302 | gcc_assert (innerc); | |
4303 | endvar = OMP_CLAUSE_DECL (innerc); | |
4304 | } | |
4305 | ||
4306 | gsi = gsi_start_bb (l0_bb); | |
4307 | t = istart0; | |
4308 | if (fd->ordered && fd->collapse == 1) | |
4309 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, | |
4310 | fold_convert (fd->iter_type, fd->loop.step)); | |
4311 | else if (bias) | |
4312 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); | |
4313 | if (fd->ordered && fd->collapse == 1) | |
4314 | { | |
4315 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4316 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), | |
4317 | fd->loop.n1, fold_convert (sizetype, t)); | |
4318 | else | |
4319 | { | |
4320 | t = fold_convert (TREE_TYPE (startvar), t); | |
4321 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), | |
4322 | fd->loop.n1, t); | |
4323 | } | |
4324 | } | |
4325 | else | |
4326 | { | |
4327 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4328 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
4329 | t = fold_convert (TREE_TYPE (startvar), t); | |
4330 | } | |
4331 | t = force_gimple_operand_gsi (&gsi, t, | |
4332 | DECL_P (startvar) | |
4333 | && TREE_ADDRESSABLE (startvar), | |
4334 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4335 | assign_stmt = gimple_build_assign (startvar, t); | |
4336 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6c7ae8c5 JJ |
4337 | if (cond_var) |
4338 | { | |
4339 | tree itype = TREE_TYPE (cond_var); | |
4340 | /* For lastprivate(conditional:) itervar, we need some iteration | |
4341 | counter that starts at unsigned non-zero and increases. | |
4342 | Prefer as few IVs as possible, so if we can use startvar | |
4343 | itself, use that, or startvar + constant (those would be | |
4344 | incremented with step), and as last resort use the s0 + 1 | |
4345 | incremented by 1. */ | |
4346 | if ((fd->ordered && fd->collapse == 1) | |
4347 | || bias | |
4348 | || POINTER_TYPE_P (type) | |
4349 | || TREE_CODE (fd->loop.n1) != INTEGER_CST | |
4350 | || fd->loop.cond_code != LT_EXPR) | |
4351 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, istart0), | |
4352 | build_int_cst (itype, 1)); | |
4353 | else if (tree_int_cst_sgn (fd->loop.n1) == 1) | |
4354 | t = fold_convert (itype, t); | |
4355 | else | |
4356 | { | |
4357 | tree c = fold_convert (itype, fd->loop.n1); | |
4358 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
4359 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
4360 | } | |
4361 | t = force_gimple_operand_gsi (&gsi, t, false, | |
4362 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4363 | assign_stmt = gimple_build_assign (cond_var, t); | |
4364 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4365 | } | |
629b3d75 MJ |
4366 | |
4367 | t = iend0; | |
4368 | if (fd->ordered && fd->collapse == 1) | |
4369 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, | |
4370 | fold_convert (fd->iter_type, fd->loop.step)); | |
4371 | else if (bias) | |
4372 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); | |
4373 | if (fd->ordered && fd->collapse == 1) | |
4374 | { | |
4375 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4376 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), | |
4377 | fd->loop.n1, fold_convert (sizetype, t)); | |
4378 | else | |
4379 | { | |
4380 | t = fold_convert (TREE_TYPE (startvar), t); | |
4381 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), | |
4382 | fd->loop.n1, t); | |
4383 | } | |
4384 | } | |
4385 | else | |
4386 | { | |
4387 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
4388 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
4389 | t = fold_convert (TREE_TYPE (startvar), t); | |
4390 | } | |
4391 | iend = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4392 | false, GSI_CONTINUE_LINKING); | |
4393 | if (endvar) | |
4394 | { | |
4395 | assign_stmt = gimple_build_assign (endvar, iend); | |
4396 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4397 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (iend))) | |
4398 | assign_stmt = gimple_build_assign (fd->loop.v, iend); | |
4399 | else | |
4400 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, iend); | |
4401 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4402 | } | |
4403 | /* Handle linear clause adjustments. */ | |
4404 | tree itercnt = NULL_TREE; | |
4405 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
4406 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
4407 | c; c = OMP_CLAUSE_CHAIN (c)) | |
4408 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
4409 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
4410 | { | |
4411 | tree d = OMP_CLAUSE_DECL (c); | |
629b3d75 | 4412 | tree t = d, a, dest; |
22e6b327 | 4413 | if (omp_privatize_by_reference (t)) |
629b3d75 MJ |
4414 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); |
4415 | tree type = TREE_TYPE (t); | |
4416 | if (POINTER_TYPE_P (type)) | |
4417 | type = sizetype; | |
4418 | dest = unshare_expr (t); | |
4419 | tree v = create_tmp_var (TREE_TYPE (t), NULL); | |
4420 | expand_omp_build_assign (&gsif, v, t); | |
4421 | if (itercnt == NULL_TREE) | |
4422 | { | |
4423 | itercnt = startvar; | |
4424 | tree n1 = fd->loop.n1; | |
4425 | if (POINTER_TYPE_P (TREE_TYPE (itercnt))) | |
4426 | { | |
4427 | itercnt | |
4428 | = fold_convert (signed_type_for (TREE_TYPE (itercnt)), | |
4429 | itercnt); | |
4430 | n1 = fold_convert (TREE_TYPE (itercnt), n1); | |
4431 | } | |
4432 | itercnt = fold_build2 (MINUS_EXPR, TREE_TYPE (itercnt), | |
4433 | itercnt, n1); | |
4434 | itercnt = fold_build2 (EXACT_DIV_EXPR, TREE_TYPE (itercnt), | |
4435 | itercnt, fd->loop.step); | |
4436 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
4437 | NULL_TREE, false, | |
4438 | GSI_CONTINUE_LINKING); | |
4439 | } | |
4440 | a = fold_build2 (MULT_EXPR, type, | |
4441 | fold_convert (type, itercnt), | |
4442 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
4443 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
4444 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); | |
4445 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4446 | false, GSI_CONTINUE_LINKING); | |
6fcc3cac | 4447 | expand_omp_build_assign (&gsi, dest, t, true); |
629b3d75 MJ |
4448 | } |
4449 | if (fd->collapse > 1) | |
aed3ab25 | 4450 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, inner_stmt, startvar); |
629b3d75 MJ |
4451 | |
4452 | if (fd->ordered) | |
4453 | { | |
4454 | /* Until now, counts array contained number of iterations or | |
f25a6767 | 4455 | variable containing it for ith loop. From now on, we usually need |
629b3d75 MJ |
4456 | those counts only for collapsed loops, and only for the 2nd |
4457 | till the last collapsed one. Move those one element earlier, | |
4458 | we'll use counts[fd->collapse - 1] for the first source/sink | |
4459 | iteration counter and so on and counts[fd->ordered] | |
4460 | as the array holding the current counter values for | |
f25a6767 JJ |
4461 | depend(source). For doacross(sink:omp_cur_iteration - 1) we need |
4462 | the counts from fd->collapse to fd->ordered - 1; make a copy of | |
4463 | those to counts[fd->ordered + 2] and onwards. | |
4464 | counts[fd->ordered + 1] can be a flag whether it is the first | |
4465 | iteration with a new collapsed counter (used only if | |
4466 | fd->ordered > fd->collapse). */ | |
4467 | if (fd->ordered > fd->collapse) | |
4468 | memcpy (counts + fd->ordered + 2, counts + fd->collapse, | |
4469 | (fd->ordered - fd->collapse) * sizeof (counts[0])); | |
629b3d75 MJ |
4470 | if (fd->collapse > 1) |
4471 | memmove (counts, counts + 1, (fd->collapse - 1) * sizeof (counts[0])); | |
4472 | if (broken_loop) | |
4473 | { | |
4474 | int i; | |
4475 | for (i = fd->collapse; i < fd->ordered; i++) | |
4476 | { | |
4477 | tree type = TREE_TYPE (fd->loops[i].v); | |
4478 | tree this_cond | |
4479 | = fold_build2 (fd->loops[i].cond_code, boolean_type_node, | |
4480 | fold_convert (type, fd->loops[i].n1), | |
4481 | fold_convert (type, fd->loops[i].n2)); | |
4482 | if (!integer_onep (this_cond)) | |
4483 | break; | |
4484 | } | |
4485 | if (i < fd->ordered) | |
4486 | { | |
713fa5db JJ |
4487 | if (entry_bb->loop_father != l0_bb->loop_father) |
4488 | { | |
4489 | remove_bb_from_loops (l0_bb); | |
4490 | add_bb_to_loop (l0_bb, entry_bb->loop_father); | |
4491 | gcc_assert (single_succ (l0_bb) == l1_bb); | |
4492 | } | |
629b3d75 MJ |
4493 | cont_bb |
4494 | = create_empty_bb (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb); | |
713fa5db | 4495 | add_bb_to_loop (cont_bb, l0_bb->loop_father); |
629b3d75 MJ |
4496 | gimple_stmt_iterator gsi = gsi_after_labels (cont_bb); |
4497 | gimple *g = gimple_build_omp_continue (fd->loop.v, fd->loop.v); | |
4498 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
4499 | make_edge (cont_bb, l3_bb, EDGE_FALLTHRU); | |
4500 | make_edge (cont_bb, l1_bb, 0); | |
4501 | l2_bb = create_empty_bb (cont_bb); | |
4502 | broken_loop = false; | |
4503 | } | |
4504 | } | |
4505 | expand_omp_ordered_source_sink (region, fd, counts, cont_bb); | |
4506 | cont_bb = expand_omp_for_ordered_loops (fd, counts, cont_bb, l1_bb, | |
713fa5db | 4507 | l0_bb, ordered_lastprivate); |
629b3d75 MJ |
4508 | if (counts[fd->collapse - 1]) |
4509 | { | |
4510 | gcc_assert (fd->collapse == 1); | |
4511 | gsi = gsi_last_bb (l0_bb); | |
4512 | expand_omp_build_assign (&gsi, counts[fd->collapse - 1], | |
4513 | istart0, true); | |
a2a17ae7 JJ |
4514 | if (cont_bb) |
4515 | { | |
4516 | gsi = gsi_last_bb (cont_bb); | |
4517 | t = fold_build2 (PLUS_EXPR, fd->iter_type, | |
4518 | counts[fd->collapse - 1], | |
4519 | build_int_cst (fd->iter_type, 1)); | |
4520 | expand_omp_build_assign (&gsi, counts[fd->collapse - 1], t); | |
4521 | tree aref = build4 (ARRAY_REF, fd->iter_type, | |
4522 | counts[fd->ordered], size_zero_node, | |
4523 | NULL_TREE, NULL_TREE); | |
4524 | expand_omp_build_assign (&gsi, aref, counts[fd->collapse - 1]); | |
4525 | } | |
629b3d75 MJ |
4526 | t = counts[fd->collapse - 1]; |
4527 | } | |
4528 | else if (fd->collapse > 1) | |
4529 | t = fd->loop.v; | |
4530 | else | |
4531 | { | |
4532 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), | |
4533 | fd->loops[0].v, fd->loops[0].n1); | |
4534 | t = fold_convert (fd->iter_type, t); | |
4535 | } | |
4536 | gsi = gsi_last_bb (l0_bb); | |
4537 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
4538 | size_zero_node, NULL_TREE, NULL_TREE); | |
4539 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4540 | false, GSI_CONTINUE_LINKING); | |
4541 | expand_omp_build_assign (&gsi, aref, t, true); | |
4542 | } | |
4543 | ||
4544 | if (!broken_loop) | |
4545 | { | |
4546 | /* Code to control the increment and predicate for the sequential | |
4547 | loop goes in the CONT_BB. */ | |
65f4b875 | 4548 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
4549 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
4550 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
4551 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
4552 | vback = gimple_omp_continue_control_def (cont_stmt); | |
4553 | ||
7855700e JJ |
4554 | if (cond_var) |
4555 | { | |
4556 | tree itype = TREE_TYPE (cond_var); | |
4557 | tree t2; | |
4558 | if ((fd->ordered && fd->collapse == 1) | |
4559 | || bias | |
4560 | || POINTER_TYPE_P (type) | |
4561 | || TREE_CODE (fd->loop.n1) != INTEGER_CST | |
4562 | || fd->loop.cond_code != LT_EXPR) | |
4563 | t2 = build_int_cst (itype, 1); | |
4564 | else | |
4565 | t2 = fold_convert (itype, fd->loop.step); | |
4566 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
4567 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
4568 | NULL_TREE, true, GSI_SAME_STMT); | |
4569 | assign_stmt = gimple_build_assign (cond_var, t2); | |
4570 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4571 | } | |
4572 | ||
629b3d75 MJ |
4573 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
4574 | { | |
4575 | if (POINTER_TYPE_P (type)) | |
4576 | t = fold_build_pointer_plus (vmain, fd->loop.step); | |
4577 | else | |
4578 | t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step); | |
4579 | t = force_gimple_operand_gsi (&gsi, t, | |
4580 | DECL_P (vback) | |
4581 | && TREE_ADDRESSABLE (vback), | |
4582 | NULL_TREE, true, GSI_SAME_STMT); | |
4583 | assign_stmt = gimple_build_assign (vback, t); | |
4584 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4585 | ||
4586 | if (fd->ordered && counts[fd->collapse - 1] == NULL_TREE) | |
4587 | { | |
d1ffbd43 | 4588 | tree tem; |
629b3d75 | 4589 | if (fd->collapse > 1) |
d1ffbd43 | 4590 | tem = fd->loop.v; |
629b3d75 MJ |
4591 | else |
4592 | { | |
d1ffbd43 JJ |
4593 | tem = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), |
4594 | fd->loops[0].v, fd->loops[0].n1); | |
4595 | tem = fold_convert (fd->iter_type, tem); | |
629b3d75 MJ |
4596 | } |
4597 | tree aref = build4 (ARRAY_REF, fd->iter_type, | |
4598 | counts[fd->ordered], size_zero_node, | |
4599 | NULL_TREE, NULL_TREE); | |
d1ffbd43 JJ |
4600 | tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, |
4601 | true, GSI_SAME_STMT); | |
4602 | expand_omp_build_assign (&gsi, aref, tem); | |
629b3d75 MJ |
4603 | } |
4604 | ||
4605 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
4606 | DECL_P (vback) && TREE_ADDRESSABLE (vback) ? t : vback, | |
4607 | iend); | |
4608 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
4609 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
4610 | } | |
4611 | ||
4612 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
4613 | gsi_remove (&gsi, true); | |
4614 | ||
4615 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
aed3ab25 | 4616 | collapse_bb = extract_omp_for_update_vars (fd, NULL, cont_bb, l1_bb); |
629b3d75 MJ |
4617 | |
4618 | /* Emit code to get the next parallel iteration in L2_BB. */ | |
4619 | gsi = gsi_start_bb (l2_bb); | |
4620 | ||
4621 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, | |
4622 | build_fold_addr_expr (istart0), | |
4623 | build_fold_addr_expr (iend0)); | |
4624 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4625 | false, GSI_CONTINUE_LINKING); | |
4626 | if (TREE_TYPE (t) != boolean_type_node) | |
4627 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
4628 | t, build_int_cst (TREE_TYPE (t), 0)); | |
4629 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
4630 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
4631 | } | |
4632 | ||
4633 | /* Add the loop cleanup function. */ | |
65f4b875 | 4634 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
4635 | if (gimple_omp_return_nowait_p (gsi_stmt (gsi))) |
4636 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
4637 | else if (gimple_omp_return_lhs (gsi_stmt (gsi))) | |
4638 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
4639 | else | |
4640 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
4641 | gcall *call_stmt = gimple_build_call (t, 0); | |
629b3d75 MJ |
4642 | if (fd->ordered) |
4643 | { | |
4644 | tree arr = counts[fd->ordered]; | |
25b45c7c | 4645 | tree clobber = build_clobber (TREE_TYPE (arr)); |
629b3d75 MJ |
4646 | gsi_insert_after (&gsi, gimple_build_assign (arr, clobber), |
4647 | GSI_SAME_STMT); | |
4648 | } | |
28567c40 JJ |
4649 | if (gimple_omp_return_lhs (gsi_stmt (gsi))) |
4650 | { | |
4651 | gimple_call_set_lhs (call_stmt, gimple_omp_return_lhs (gsi_stmt (gsi))); | |
4652 | if (fd->have_reductemp) | |
4653 | { | |
4654 | gimple *g = gimple_build_assign (reductions, NOP_EXPR, | |
4655 | gimple_call_lhs (call_stmt)); | |
4656 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
4657 | } | |
4658 | } | |
4659 | gsi_insert_after (&gsi, call_stmt, GSI_SAME_STMT); | |
629b3d75 MJ |
4660 | gsi_remove (&gsi, true); |
4661 | ||
4662 | /* Connect the new blocks. */ | |
4663 | find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE; | |
4664 | find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE; | |
4665 | ||
4666 | if (!broken_loop) | |
4667 | { | |
4668 | gimple_seq phis; | |
4669 | ||
4670 | e = find_edge (cont_bb, l3_bb); | |
4671 | ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE); | |
4672 | ||
4673 | phis = phi_nodes (l3_bb); | |
4674 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4675 | { | |
4676 | gimple *phi = gsi_stmt (gsi); | |
4677 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne), | |
4678 | PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
4679 | } | |
4680 | remove_edge (e); | |
4681 | ||
4682 | make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE); | |
4683 | e = find_edge (cont_bb, l1_bb); | |
4684 | if (e == NULL) | |
4685 | { | |
4686 | e = BRANCH_EDGE (cont_bb); | |
4687 | gcc_assert (single_succ (e->dest) == l1_bb); | |
4688 | } | |
4689 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4690 | { | |
4691 | remove_edge (e); | |
4692 | e = NULL; | |
4693 | } | |
4694 | else if (fd->collapse > 1) | |
4695 | { | |
4696 | remove_edge (e); | |
4697 | e = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
4698 | } | |
4699 | else | |
4700 | e->flags = EDGE_TRUE_VALUE; | |
4701 | if (e) | |
4702 | { | |
357067f2 JH |
4703 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); |
4704 | find_edge (cont_bb, l2_bb)->probability = e->probability.invert (); | |
629b3d75 MJ |
4705 | } |
4706 | else | |
4707 | { | |
4708 | e = find_edge (cont_bb, l2_bb); | |
4709 | e->flags = EDGE_FALLTHRU; | |
4710 | } | |
4711 | make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE); | |
4712 | ||
4713 | if (gimple_in_ssa_p (cfun)) | |
4714 | { | |
4715 | /* Add phis to the outer loop that connect to the phis in the inner, | |
4716 | original loop, and move the loop entry value of the inner phi to | |
4717 | the loop entry value of the outer phi. */ | |
4718 | gphi_iterator psi; | |
4719 | for (psi = gsi_start_phis (l3_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
4720 | { | |
620e594b | 4721 | location_t locus; |
629b3d75 MJ |
4722 | gphi *nphi; |
4723 | gphi *exit_phi = psi.phi (); | |
4724 | ||
164485b5 JJ |
4725 | if (virtual_operand_p (gimple_phi_result (exit_phi))) |
4726 | continue; | |
4727 | ||
629b3d75 MJ |
4728 | edge l2_to_l3 = find_edge (l2_bb, l3_bb); |
4729 | tree exit_res = PHI_ARG_DEF_FROM_EDGE (exit_phi, l2_to_l3); | |
4730 | ||
4731 | basic_block latch = BRANCH_EDGE (cont_bb)->dest; | |
4732 | edge latch_to_l1 = find_edge (latch, l1_bb); | |
4733 | gphi *inner_phi | |
4734 | = find_phi_with_arg_on_edge (exit_res, latch_to_l1); | |
4735 | ||
4736 | tree t = gimple_phi_result (exit_phi); | |
4737 | tree new_res = copy_ssa_name (t, NULL); | |
4738 | nphi = create_phi_node (new_res, l0_bb); | |
4739 | ||
4740 | edge l0_to_l1 = find_edge (l0_bb, l1_bb); | |
4741 | t = PHI_ARG_DEF_FROM_EDGE (inner_phi, l0_to_l1); | |
4742 | locus = gimple_phi_arg_location_from_edge (inner_phi, l0_to_l1); | |
4743 | edge entry_to_l0 = find_edge (entry_bb, l0_bb); | |
4744 | add_phi_arg (nphi, t, entry_to_l0, locus); | |
4745 | ||
4746 | edge l2_to_l0 = find_edge (l2_bb, l0_bb); | |
4747 | add_phi_arg (nphi, exit_res, l2_to_l0, UNKNOWN_LOCATION); | |
4748 | ||
4749 | add_phi_arg (inner_phi, new_res, l0_to_l1, UNKNOWN_LOCATION); | |
164485b5 | 4750 | } |
629b3d75 MJ |
4751 | } |
4752 | ||
4753 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, | |
4754 | recompute_dominator (CDI_DOMINATORS, l2_bb)); | |
4755 | set_immediate_dominator (CDI_DOMINATORS, l3_bb, | |
4756 | recompute_dominator (CDI_DOMINATORS, l3_bb)); | |
4757 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, | |
4758 | recompute_dominator (CDI_DOMINATORS, l0_bb)); | |
4759 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, | |
4760 | recompute_dominator (CDI_DOMINATORS, l1_bb)); | |
4761 | ||
4762 | /* We enter expand_omp_for_generic with a loop. This original loop may | |
4763 | have its own loop struct, or it may be part of an outer loop struct | |
4764 | (which may be the fake loop). */ | |
99b1c316 | 4765 | class loop *outer_loop = entry_bb->loop_father; |
629b3d75 MJ |
4766 | bool orig_loop_has_loop_struct = l1_bb->loop_father != outer_loop; |
4767 | ||
4768 | add_bb_to_loop (l2_bb, outer_loop); | |
4769 | ||
4770 | /* We've added a new loop around the original loop. Allocate the | |
4771 | corresponding loop struct. */ | |
99b1c316 | 4772 | class loop *new_loop = alloc_loop (); |
629b3d75 MJ |
4773 | new_loop->header = l0_bb; |
4774 | new_loop->latch = l2_bb; | |
4775 | add_loop (new_loop, outer_loop); | |
4776 | ||
4777 | /* Allocate a loop structure for the original loop unless we already | |
4778 | had one. */ | |
4779 | if (!orig_loop_has_loop_struct | |
4780 | && !gimple_omp_for_combined_p (fd->for_stmt)) | |
4781 | { | |
99b1c316 | 4782 | class loop *orig_loop = alloc_loop (); |
629b3d75 MJ |
4783 | orig_loop->header = l1_bb; |
4784 | /* The loop may have multiple latches. */ | |
4785 | add_loop (orig_loop, new_loop); | |
4786 | } | |
4787 | } | |
4788 | } | |
4789 | ||
2f6bb511 JJ |
4790 | /* Helper function for expand_omp_for_static_nochunk. If PTR is NULL, |
4791 | compute needed allocation size. If !ALLOC of team allocations, | |
4792 | if ALLOC of thread allocation. SZ is the initial needed size for | |
4793 | other purposes, ALLOC_ALIGN guaranteed alignment of allocation in bytes, | |
4794 | CNT number of elements of each array, for !ALLOC this is | |
4795 | omp_get_num_threads (), for ALLOC number of iterations handled by the | |
4796 | current thread. If PTR is non-NULL, it is the start of the allocation | |
4797 | and this routine shall assign to OMP_CLAUSE_DECL (c) of those _scantemp_ | |
4798 | clauses pointers to the corresponding arrays. */ | |
4799 | ||
4800 | static tree | |
4801 | expand_omp_scantemp_alloc (tree clauses, tree ptr, unsigned HOST_WIDE_INT sz, | |
4802 | unsigned HOST_WIDE_INT alloc_align, tree cnt, | |
4803 | gimple_stmt_iterator *gsi, bool alloc) | |
4804 | { | |
4805 | tree eltsz = NULL_TREE; | |
4806 | unsigned HOST_WIDE_INT preval = 0; | |
4807 | if (ptr && sz) | |
4808 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), | |
4809 | ptr, size_int (sz)); | |
4810 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
4811 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
4812 | && !OMP_CLAUSE__SCANTEMP__CONTROL (c) | |
4813 | && (!OMP_CLAUSE__SCANTEMP__ALLOC (c)) != alloc) | |
4814 | { | |
4815 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); | |
4816 | unsigned HOST_WIDE_INT al = TYPE_ALIGN_UNIT (pointee_type); | |
4817 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) | |
4818 | { | |
4819 | unsigned HOST_WIDE_INT szl | |
4820 | = tree_to_uhwi (TYPE_SIZE_UNIT (pointee_type)); | |
4821 | szl = least_bit_hwi (szl); | |
4822 | if (szl) | |
4823 | al = MIN (al, szl); | |
4824 | } | |
4825 | if (ptr == NULL_TREE) | |
4826 | { | |
4827 | if (eltsz == NULL_TREE) | |
4828 | eltsz = TYPE_SIZE_UNIT (pointee_type); | |
4829 | else | |
4830 | eltsz = size_binop (PLUS_EXPR, eltsz, | |
4831 | TYPE_SIZE_UNIT (pointee_type)); | |
4832 | } | |
4833 | if (preval == 0 && al <= alloc_align) | |
4834 | { | |
4835 | unsigned HOST_WIDE_INT diff = ROUND_UP (sz, al) - sz; | |
4836 | sz += diff; | |
4837 | if (diff && ptr) | |
4838 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), | |
4839 | ptr, size_int (diff)); | |
4840 | } | |
4841 | else if (al > preval) | |
4842 | { | |
4843 | if (ptr) | |
4844 | { | |
4845 | ptr = fold_convert (pointer_sized_int_node, ptr); | |
4846 | ptr = fold_build2 (PLUS_EXPR, pointer_sized_int_node, ptr, | |
4847 | build_int_cst (pointer_sized_int_node, | |
4848 | al - 1)); | |
4849 | ptr = fold_build2 (BIT_AND_EXPR, pointer_sized_int_node, ptr, | |
4850 | build_int_cst (pointer_sized_int_node, | |
4851 | -(HOST_WIDE_INT) al)); | |
4852 | ptr = fold_convert (ptr_type_node, ptr); | |
4853 | } | |
4854 | else | |
4855 | sz += al - 1; | |
4856 | } | |
4857 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) | |
4858 | preval = al; | |
4859 | else | |
4860 | preval = 1; | |
4861 | if (ptr) | |
4862 | { | |
4863 | expand_omp_build_assign (gsi, OMP_CLAUSE_DECL (c), ptr, false); | |
4864 | ptr = OMP_CLAUSE_DECL (c); | |
4865 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), ptr, | |
4866 | size_binop (MULT_EXPR, cnt, | |
4867 | TYPE_SIZE_UNIT (pointee_type))); | |
4868 | } | |
4869 | } | |
4870 | ||
4871 | if (ptr == NULL_TREE) | |
4872 | { | |
4873 | eltsz = size_binop (MULT_EXPR, eltsz, cnt); | |
4874 | if (sz) | |
4875 | eltsz = size_binop (PLUS_EXPR, eltsz, size_int (sz)); | |
4876 | return eltsz; | |
4877 | } | |
4878 | else | |
4879 | return ptr; | |
4880 | } | |
4881 | ||
14707c89 JJ |
4882 | /* Return the last _looptemp_ clause if one has been created for |
4883 | lastprivate on distribute parallel for{, simd} or taskloop. | |
4884 | FD is the loop data and INNERC should be the second _looptemp_ | |
4885 | clause (the one holding the end of the range). | |
4886 | This is followed by collapse - 1 _looptemp_ clauses for the | |
4887 | counts[1] and up, and for triangular loops followed by 4 | |
4888 | further _looptemp_ clauses (one for counts[0], one first_inner_iterations, | |
4889 | one factor and one adjn1). After this there is optionally one | |
4890 | _looptemp_ clause that this function returns. */ | |
4891 | ||
4892 | static tree | |
4893 | find_lastprivate_looptemp (struct omp_for_data *fd, tree innerc) | |
4894 | { | |
4895 | gcc_assert (innerc); | |
4896 | int count = fd->collapse - 1; | |
4897 | if (fd->non_rect | |
4898 | && fd->last_nonrect == fd->first_nonrect + 1 | |
4899 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) | |
4900 | count += 4; | |
4901 | for (int i = 0; i < count; i++) | |
4902 | { | |
4903 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4904 | OMP_CLAUSE__LOOPTEMP_); | |
4905 | gcc_assert (innerc); | |
4906 | } | |
4907 | return omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4908 | OMP_CLAUSE__LOOPTEMP_); | |
4909 | } | |
4910 | ||
629b3d75 MJ |
4911 | /* A subroutine of expand_omp_for. Generate code for a parallel |
4912 | loop with static schedule and no specified chunk size. Given | |
4913 | parameters: | |
4914 | ||
4915 | for (V = N1; V cond N2; V += STEP) BODY; | |
4916 | ||
4917 | where COND is "<" or ">", we generate pseudocode | |
4918 | ||
4919 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; | |
4920 | if (cond is <) | |
4921 | adj = STEP - 1; | |
4922 | else | |
4923 | adj = STEP + 1; | |
4924 | if ((__typeof (V)) -1 > 0 && cond is >) | |
4925 | n = -(adj + N2 - N1) / -STEP; | |
4926 | else | |
4927 | n = (adj + N2 - N1) / STEP; | |
4928 | q = n / nthreads; | |
4929 | tt = n % nthreads; | |
4930 | if (threadid < tt) goto L3; else goto L4; | |
4931 | L3: | |
4932 | tt = 0; | |
4933 | q = q + 1; | |
4934 | L4: | |
4935 | s0 = q * threadid + tt; | |
4936 | e0 = s0 + q; | |
4937 | V = s0 * STEP + N1; | |
4938 | if (s0 >= e0) goto L2; else goto L0; | |
4939 | L0: | |
4940 | e = e0 * STEP + N1; | |
4941 | L1: | |
4942 | BODY; | |
4943 | V += STEP; | |
4944 | if (V cond e) goto L1; | |
4945 | L2: | |
4946 | */ | |
4947 | ||
4948 | static void | |
4949 | expand_omp_for_static_nochunk (struct omp_region *region, | |
4950 | struct omp_for_data *fd, | |
4951 | gimple *inner_stmt) | |
4952 | { | |
2f6bb511 | 4953 | tree n, q, s0, e0, e, t, tt, nthreads = NULL_TREE, threadid; |
629b3d75 MJ |
4954 | tree type, itype, vmain, vback; |
4955 | basic_block entry_bb, second_bb, third_bb, exit_bb, seq_start_bb; | |
4956 | basic_block body_bb, cont_bb, collapse_bb = NULL; | |
2f6bb511 JJ |
4957 | basic_block fin_bb, fourth_bb = NULL, fifth_bb = NULL, sixth_bb = NULL; |
4958 | basic_block exit1_bb = NULL, exit2_bb = NULL, exit3_bb = NULL; | |
6c7ae8c5 | 4959 | gimple_stmt_iterator gsi, gsip; |
629b3d75 MJ |
4960 | edge ep; |
4961 | bool broken_loop = region->cont == NULL; | |
4962 | tree *counts = NULL; | |
4963 | tree n1, n2, step; | |
28567c40 | 4964 | tree reductions = NULL_TREE; |
8221c30b | 4965 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; |
629b3d75 MJ |
4966 | |
4967 | itype = type = TREE_TYPE (fd->loop.v); | |
4968 | if (POINTER_TYPE_P (type)) | |
4969 | itype = signed_type_for (type); | |
4970 | ||
4971 | entry_bb = region->entry; | |
4972 | cont_bb = region->cont; | |
4973 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
4974 | fin_bb = BRANCH_EDGE (entry_bb)->dest; | |
4975 | gcc_assert (broken_loop | |
4976 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
4977 | seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
4978 | body_bb = single_succ (seq_start_bb); | |
4979 | if (!broken_loop) | |
4980 | { | |
4981 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb | |
4982 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); | |
4983 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
4984 | } | |
4985 | exit_bb = region->exit; | |
4986 | ||
4987 | /* Iteration space partitioning goes in ENTRY_BB. */ | |
65f4b875 | 4988 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 | 4989 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
6c7ae8c5 JJ |
4990 | gsip = gsi; |
4991 | gsi_prev (&gsip); | |
629b3d75 MJ |
4992 | |
4993 | if (fd->collapse > 1) | |
4994 | { | |
4995 | int first_zero_iter = -1, dummy = -1; | |
4996 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
4997 | ||
4998 | counts = XALLOCAVEC (tree, fd->collapse); | |
4999 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5000 | fin_bb, first_zero_iter, | |
5001 | dummy_bb, dummy, l2_dom_bb); | |
5002 | t = NULL_TREE; | |
5003 | } | |
5004 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5005 | t = integer_one_node; | |
5006 | else | |
5007 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
5008 | fold_convert (type, fd->loop.n1), | |
5009 | fold_convert (type, fd->loop.n2)); | |
5010 | if (fd->collapse == 1 | |
5011 | && TYPE_UNSIGNED (type) | |
5012 | && (t == NULL_TREE || !integer_onep (t))) | |
5013 | { | |
5014 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); | |
5015 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
5016 | true, GSI_SAME_STMT); | |
5017 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
5018 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
5019 | true, GSI_SAME_STMT); | |
47de0b56 JJ |
5020 | gcond *cond_stmt = expand_omp_build_cond (&gsi, fd->loop.cond_code, |
5021 | n1, n2); | |
629b3d75 MJ |
5022 | ep = split_block (entry_bb, cond_stmt); |
5023 | ep->flags = EDGE_TRUE_VALUE; | |
5024 | entry_bb = ep->dest; | |
357067f2 | 5025 | ep->probability = profile_probability::very_likely (); |
629b3d75 | 5026 | ep = make_edge (ep->src, fin_bb, EDGE_FALSE_VALUE); |
357067f2 | 5027 | ep->probability = profile_probability::very_unlikely (); |
629b3d75 MJ |
5028 | if (gimple_in_ssa_p (cfun)) |
5029 | { | |
5030 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; | |
5031 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); | |
5032 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
5033 | { | |
5034 | gphi *phi = gpi.phi (); | |
5035 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
5036 | ep, UNKNOWN_LOCATION); | |
5037 | } | |
5038 | } | |
5039 | gsi = gsi_last_bb (entry_bb); | |
5040 | } | |
5041 | ||
8221c30b JJ |
5042 | if (fd->lastprivate_conditional) |
5043 | { | |
5044 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5045 | tree c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); | |
5046 | if (fd->have_pointer_condtemp) | |
5047 | condtemp = OMP_CLAUSE_DECL (c); | |
5048 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
5049 | cond_var = OMP_CLAUSE_DECL (c); | |
5050 | } | |
2f6bb511 | 5051 | if (fd->have_reductemp |
2f03073f JJ |
5052 | /* For scan, we don't want to reinitialize condtemp before the |
5053 | second loop. */ | |
5054 | || (fd->have_pointer_condtemp && !fd->have_scantemp) | |
2f6bb511 | 5055 | || fd->have_nonctrl_scantemp) |
28567c40 JJ |
5056 | { |
5057 | tree t1 = build_int_cst (long_integer_type_node, 0); | |
5058 | tree t2 = build_int_cst (long_integer_type_node, 1); | |
5059 | tree t3 = build_int_cstu (long_integer_type_node, | |
5060 | (HOST_WIDE_INT_1U << 31) + 1); | |
5061 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
6c7ae8c5 JJ |
5062 | gimple_stmt_iterator gsi2 = gsi_none (); |
5063 | gimple *g = NULL; | |
5064 | tree mem = null_pointer_node, memv = NULL_TREE; | |
2f6bb511 JJ |
5065 | unsigned HOST_WIDE_INT condtemp_sz = 0; |
5066 | unsigned HOST_WIDE_INT alloc_align = 0; | |
6c7ae8c5 JJ |
5067 | if (fd->have_reductemp) |
5068 | { | |
2f6bb511 | 5069 | gcc_assert (!fd->have_nonctrl_scantemp); |
6c7ae8c5 JJ |
5070 | tree c = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); |
5071 | reductions = OMP_CLAUSE_DECL (c); | |
5072 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
5073 | g = SSA_NAME_DEF_STMT (reductions); | |
5074 | reductions = gimple_assign_rhs1 (g); | |
5075 | OMP_CLAUSE_DECL (c) = reductions; | |
5076 | gsi2 = gsi_for_stmt (g); | |
5077 | } | |
5078 | else | |
5079 | { | |
5080 | if (gsi_end_p (gsip)) | |
5081 | gsi2 = gsi_after_labels (region->entry); | |
5082 | else | |
5083 | gsi2 = gsip; | |
5084 | reductions = null_pointer_node; | |
5085 | } | |
2f6bb511 | 5086 | if (fd->have_pointer_condtemp || fd->have_nonctrl_scantemp) |
6c7ae8c5 | 5087 | { |
2f6bb511 JJ |
5088 | tree type; |
5089 | if (fd->have_pointer_condtemp) | |
5090 | type = TREE_TYPE (condtemp); | |
5091 | else | |
5092 | type = ptr_type_node; | |
6c7ae8c5 JJ |
5093 | memv = create_tmp_var (type); |
5094 | TREE_ADDRESSABLE (memv) = 1; | |
2f6bb511 JJ |
5095 | unsigned HOST_WIDE_INT sz = 0; |
5096 | tree size = NULL_TREE; | |
5097 | if (fd->have_pointer_condtemp) | |
5098 | { | |
5099 | sz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
5100 | sz *= fd->lastprivate_conditional; | |
5101 | condtemp_sz = sz; | |
5102 | } | |
5103 | if (fd->have_nonctrl_scantemp) | |
5104 | { | |
5105 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
5106 | gimple *g = gimple_build_call (nthreads, 0); | |
5107 | nthreads = create_tmp_var (integer_type_node); | |
5108 | gimple_call_set_lhs (g, nthreads); | |
5109 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); | |
5110 | nthreads = fold_convert (sizetype, nthreads); | |
5111 | alloc_align = TYPE_ALIGN_UNIT (long_long_integer_type_node); | |
5112 | size = expand_omp_scantemp_alloc (clauses, NULL_TREE, sz, | |
5113 | alloc_align, nthreads, NULL, | |
5114 | false); | |
5115 | size = fold_convert (type, size); | |
5116 | } | |
5117 | else | |
5118 | size = build_int_cst (type, sz); | |
5119 | expand_omp_build_assign (&gsi2, memv, size, false); | |
6c7ae8c5 JJ |
5120 | mem = build_fold_addr_expr (memv); |
5121 | } | |
28567c40 JJ |
5122 | tree t |
5123 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_LOOP_START), | |
5124 | 9, t1, t2, t2, t3, t1, null_pointer_node, | |
6c7ae8c5 | 5125 | null_pointer_node, reductions, mem); |
28567c40 JJ |
5126 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
5127 | true, GSI_SAME_STMT); | |
8221c30b | 5128 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 5129 | expand_omp_build_assign (&gsi2, condtemp, memv, false); |
2f6bb511 JJ |
5130 | if (fd->have_nonctrl_scantemp) |
5131 | { | |
5132 | tree ptr = fd->have_pointer_condtemp ? condtemp : memv; | |
5133 | expand_omp_scantemp_alloc (clauses, ptr, condtemp_sz, | |
5134 | alloc_align, nthreads, &gsi2, false); | |
5135 | } | |
6c7ae8c5 JJ |
5136 | if (fd->have_reductemp) |
5137 | { | |
5138 | gsi_remove (&gsi2, true); | |
5139 | release_ssa_name (gimple_assign_lhs (g)); | |
5140 | } | |
28567c40 | 5141 | } |
629b3d75 MJ |
5142 | switch (gimple_omp_for_kind (fd->for_stmt)) |
5143 | { | |
5144 | case GF_OMP_FOR_KIND_FOR: | |
5145 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
5146 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
5147 | break; | |
5148 | case GF_OMP_FOR_KIND_DISTRIBUTE: | |
5149 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS); | |
5150 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM); | |
5151 | break; | |
5152 | default: | |
5153 | gcc_unreachable (); | |
5154 | } | |
5155 | nthreads = build_call_expr (nthreads, 0); | |
5156 | nthreads = fold_convert (itype, nthreads); | |
5157 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, | |
5158 | true, GSI_SAME_STMT); | |
5159 | threadid = build_call_expr (threadid, 0); | |
5160 | threadid = fold_convert (itype, threadid); | |
5161 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, | |
5162 | true, GSI_SAME_STMT); | |
5163 | ||
5164 | n1 = fd->loop.n1; | |
5165 | n2 = fd->loop.n2; | |
5166 | step = fd->loop.step; | |
5167 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5168 | { | |
5169 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5170 | OMP_CLAUSE__LOOPTEMP_); | |
5171 | gcc_assert (innerc); | |
5172 | n1 = OMP_CLAUSE_DECL (innerc); | |
5173 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5174 | OMP_CLAUSE__LOOPTEMP_); | |
5175 | gcc_assert (innerc); | |
5176 | n2 = OMP_CLAUSE_DECL (innerc); | |
5177 | } | |
5178 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
5179 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5180 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
5181 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5182 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
5183 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5184 | ||
5185 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
5186 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
5187 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
5188 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
5189 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
5190 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
5191 | fold_build1 (NEGATE_EXPR, itype, t), | |
5192 | fold_build1 (NEGATE_EXPR, itype, step)); | |
5193 | else | |
5194 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
5195 | t = fold_convert (itype, t); | |
5196 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
5197 | ||
5198 | q = create_tmp_reg (itype, "q"); | |
5199 | t = fold_build2 (TRUNC_DIV_EXPR, itype, n, nthreads); | |
5200 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
5201 | gsi_insert_before (&gsi, gimple_build_assign (q, t), GSI_SAME_STMT); | |
5202 | ||
5203 | tt = create_tmp_reg (itype, "tt"); | |
5204 | t = fold_build2 (TRUNC_MOD_EXPR, itype, n, nthreads); | |
5205 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
5206 | gsi_insert_before (&gsi, gimple_build_assign (tt, t), GSI_SAME_STMT); | |
5207 | ||
5208 | t = build2 (LT_EXPR, boolean_type_node, threadid, tt); | |
5209 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
5210 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
5211 | ||
5212 | second_bb = split_block (entry_bb, cond_stmt)->dest; | |
65f4b875 | 5213 | gsi = gsi_last_nondebug_bb (second_bb); |
629b3d75 MJ |
5214 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
5215 | ||
5216 | gsi_insert_before (&gsi, gimple_build_assign (tt, build_int_cst (itype, 0)), | |
5217 | GSI_SAME_STMT); | |
5218 | gassign *assign_stmt | |
5219 | = gimple_build_assign (q, PLUS_EXPR, q, build_int_cst (itype, 1)); | |
5220 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5221 | ||
5222 | third_bb = split_block (second_bb, assign_stmt)->dest; | |
65f4b875 | 5223 | gsi = gsi_last_nondebug_bb (third_bb); |
629b3d75 MJ |
5224 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
5225 | ||
2f6bb511 JJ |
5226 | if (fd->have_nonctrl_scantemp) |
5227 | { | |
5228 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5229 | tree controlp = NULL_TREE, controlb = NULL_TREE; | |
5230 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
5231 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
5232 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) | |
5233 | { | |
5234 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) | |
5235 | controlb = OMP_CLAUSE_DECL (c); | |
5236 | else | |
5237 | controlp = OMP_CLAUSE_DECL (c); | |
5238 | if (controlb && controlp) | |
5239 | break; | |
5240 | } | |
5241 | gcc_assert (controlp && controlb); | |
5242 | tree cnt = create_tmp_var (sizetype); | |
5243 | gimple *g = gimple_build_assign (cnt, NOP_EXPR, q); | |
5244 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5245 | unsigned HOST_WIDE_INT alloc_align = TYPE_ALIGN_UNIT (ptr_type_node); | |
5246 | tree sz = expand_omp_scantemp_alloc (clauses, NULL_TREE, 0, | |
5247 | alloc_align, cnt, NULL, true); | |
5248 | tree size = create_tmp_var (sizetype); | |
5249 | expand_omp_build_assign (&gsi, size, sz, false); | |
5250 | tree cmp = fold_build2 (GT_EXPR, boolean_type_node, | |
5251 | size, size_int (16384)); | |
5252 | expand_omp_build_assign (&gsi, controlb, cmp); | |
5253 | g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, | |
5254 | NULL_TREE, NULL_TREE); | |
5255 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5256 | fourth_bb = split_block (third_bb, g)->dest; | |
5257 | gsi = gsi_last_nondebug_bb (fourth_bb); | |
5258 | /* FIXME: Once we have allocators, this should use allocator. */ | |
5259 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_MALLOC), 1, size); | |
5260 | gimple_call_set_lhs (g, controlp); | |
5261 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5262 | expand_omp_scantemp_alloc (clauses, controlp, 0, alloc_align, cnt, | |
5263 | &gsi, true); | |
5264 | gsi_prev (&gsi); | |
5265 | g = gsi_stmt (gsi); | |
5266 | fifth_bb = split_block (fourth_bb, g)->dest; | |
5267 | gsi = gsi_last_nondebug_bb (fifth_bb); | |
5268 | ||
5269 | g = gimple_build_call (builtin_decl_implicit (BUILT_IN_STACK_SAVE), 0); | |
5270 | gimple_call_set_lhs (g, controlp); | |
5271 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5272 | tree alloca_decl = builtin_decl_explicit (BUILT_IN_ALLOCA_WITH_ALIGN); | |
5273 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
5274 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
5275 | && OMP_CLAUSE__SCANTEMP__ALLOC (c)) | |
5276 | { | |
5277 | tree tmp = create_tmp_var (sizetype); | |
5278 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); | |
5279 | g = gimple_build_assign (tmp, MULT_EXPR, cnt, | |
5280 | TYPE_SIZE_UNIT (pointee_type)); | |
5281 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5282 | g = gimple_build_call (alloca_decl, 2, tmp, | |
5283 | size_int (TYPE_ALIGN (pointee_type))); | |
5284 | gimple_call_set_lhs (g, OMP_CLAUSE_DECL (c)); | |
5285 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5286 | } | |
5287 | ||
5288 | sixth_bb = split_block (fifth_bb, g)->dest; | |
5289 | gsi = gsi_last_nondebug_bb (sixth_bb); | |
5290 | } | |
5291 | ||
629b3d75 MJ |
5292 | t = build2 (MULT_EXPR, itype, q, threadid); |
5293 | t = build2 (PLUS_EXPR, itype, t, tt); | |
5294 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
5295 | ||
5296 | t = fold_build2 (PLUS_EXPR, itype, s0, q); | |
5297 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
5298 | ||
5299 | t = build2 (GE_EXPR, boolean_type_node, s0, e0); | |
5300 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
5301 | ||
5302 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
5303 | gsi_remove (&gsi, true); | |
5304 | ||
5305 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
5306 | gsi = gsi_start_bb (seq_start_bb); | |
5307 | ||
5308 | tree startvar = fd->loop.v; | |
5309 | tree endvar = NULL_TREE; | |
5310 | ||
5311 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5312 | { | |
5313 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
5314 | ? gimple_omp_parallel_clauses (inner_stmt) | |
5315 | : gimple_omp_for_clauses (inner_stmt); | |
5316 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
5317 | gcc_assert (innerc); | |
5318 | startvar = OMP_CLAUSE_DECL (innerc); | |
5319 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5320 | OMP_CLAUSE__LOOPTEMP_); | |
5321 | gcc_assert (innerc); | |
5322 | endvar = OMP_CLAUSE_DECL (innerc); | |
5323 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST | |
5324 | && gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) | |
5325 | { | |
14707c89 | 5326 | innerc = find_lastprivate_looptemp (fd, innerc); |
629b3d75 MJ |
5327 | if (innerc) |
5328 | { | |
5329 | /* If needed (distribute parallel for with lastprivate), | |
5330 | propagate down the total number of iterations. */ | |
5331 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), | |
5332 | fd->loop.n2); | |
5333 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, | |
5334 | GSI_CONTINUE_LINKING); | |
5335 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
5336 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5337 | } | |
5338 | } | |
5339 | } | |
5340 | t = fold_convert (itype, s0); | |
5341 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5342 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
5343 | { |
5344 | t = fold_build_pointer_plus (n1, t); | |
5345 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
5346 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
5347 | t = fold_convert (signed_type_for (type), t); | |
5348 | } | |
629b3d75 MJ |
5349 | else |
5350 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5351 | t = fold_convert (TREE_TYPE (startvar), t); | |
5352 | t = force_gimple_operand_gsi (&gsi, t, | |
5353 | DECL_P (startvar) | |
5354 | && TREE_ADDRESSABLE (startvar), | |
5355 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5356 | assign_stmt = gimple_build_assign (startvar, t); | |
5357 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6c7ae8c5 JJ |
5358 | if (cond_var) |
5359 | { | |
5360 | tree itype = TREE_TYPE (cond_var); | |
5361 | /* For lastprivate(conditional:) itervar, we need some iteration | |
5362 | counter that starts at unsigned non-zero and increases. | |
5363 | Prefer as few IVs as possible, so if we can use startvar | |
5364 | itself, use that, or startvar + constant (those would be | |
5365 | incremented with step), and as last resort use the s0 + 1 | |
5366 | incremented by 1. */ | |
5367 | if (POINTER_TYPE_P (type) | |
5368 | || TREE_CODE (n1) != INTEGER_CST | |
5369 | || fd->loop.cond_code != LT_EXPR) | |
5370 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), | |
5371 | build_int_cst (itype, 1)); | |
5372 | else if (tree_int_cst_sgn (n1) == 1) | |
5373 | t = fold_convert (itype, t); | |
5374 | else | |
5375 | { | |
5376 | tree c = fold_convert (itype, n1); | |
5377 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
5378 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
5379 | } | |
5380 | t = force_gimple_operand_gsi (&gsi, t, false, | |
5381 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5382 | assign_stmt = gimple_build_assign (cond_var, t); | |
5383 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5384 | } | |
629b3d75 MJ |
5385 | |
5386 | t = fold_convert (itype, e0); | |
5387 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5388 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
5389 | { |
5390 | t = fold_build_pointer_plus (n1, t); | |
5391 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
5392 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
5393 | t = fold_convert (signed_type_for (type), t); | |
5394 | } | |
629b3d75 MJ |
5395 | else |
5396 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5397 | t = fold_convert (TREE_TYPE (startvar), t); | |
5398 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5399 | false, GSI_CONTINUE_LINKING); | |
5400 | if (endvar) | |
5401 | { | |
5402 | assign_stmt = gimple_build_assign (endvar, e); | |
5403 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5404 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
5405 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
5406 | else | |
5407 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
5408 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5409 | } | |
5410 | /* Handle linear clause adjustments. */ | |
5411 | tree itercnt = NULL_TREE; | |
aed3ab25 | 5412 | tree *nonrect_bounds = NULL; |
629b3d75 MJ |
5413 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) |
5414 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
5415 | c; c = OMP_CLAUSE_CHAIN (c)) | |
5416 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
5417 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
5418 | { | |
5419 | tree d = OMP_CLAUSE_DECL (c); | |
629b3d75 | 5420 | tree t = d, a, dest; |
22e6b327 | 5421 | if (omp_privatize_by_reference (t)) |
629b3d75 MJ |
5422 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); |
5423 | if (itercnt == NULL_TREE) | |
5424 | { | |
5425 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5426 | { | |
5427 | itercnt = fold_build2 (MINUS_EXPR, itype, | |
5428 | fold_convert (itype, n1), | |
5429 | fold_convert (itype, fd->loop.n1)); | |
5430 | itercnt = fold_build2 (EXACT_DIV_EXPR, itype, itercnt, step); | |
5431 | itercnt = fold_build2 (PLUS_EXPR, itype, itercnt, s0); | |
5432 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
5433 | NULL_TREE, false, | |
5434 | GSI_CONTINUE_LINKING); | |
5435 | } | |
5436 | else | |
5437 | itercnt = s0; | |
5438 | } | |
5439 | tree type = TREE_TYPE (t); | |
5440 | if (POINTER_TYPE_P (type)) | |
5441 | type = sizetype; | |
5442 | a = fold_build2 (MULT_EXPR, type, | |
5443 | fold_convert (type, itercnt), | |
5444 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
5445 | dest = unshare_expr (t); | |
5446 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
5447 | : POINTER_PLUS_EXPR, TREE_TYPE (t), t, a); | |
5448 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5449 | false, GSI_CONTINUE_LINKING); | |
6fcc3cac | 5450 | expand_omp_build_assign (&gsi, dest, t, true); |
629b3d75 MJ |
5451 | } |
5452 | if (fd->collapse > 1) | |
aed3ab25 JJ |
5453 | { |
5454 | if (fd->non_rect) | |
5455 | { | |
5456 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); | |
5457 | memset (nonrect_bounds, 0, sizeof (tree) * (fd->last_nonrect + 1)); | |
5458 | } | |
5459 | expand_omp_for_init_vars (fd, &gsi, counts, nonrect_bounds, inner_stmt, | |
5460 | startvar); | |
5461 | } | |
629b3d75 MJ |
5462 | |
5463 | if (!broken_loop) | |
5464 | { | |
5465 | /* The code controlling the sequential loop replaces the | |
5466 | GIMPLE_OMP_CONTINUE. */ | |
65f4b875 | 5467 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
5468 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
5469 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
5470 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
5471 | vback = gimple_omp_continue_control_def (cont_stmt); | |
5472 | ||
7855700e JJ |
5473 | if (cond_var) |
5474 | { | |
5475 | tree itype = TREE_TYPE (cond_var); | |
5476 | tree t2; | |
5477 | if (POINTER_TYPE_P (type) | |
5478 | || TREE_CODE (n1) != INTEGER_CST | |
5479 | || fd->loop.cond_code != LT_EXPR) | |
5480 | t2 = build_int_cst (itype, 1); | |
5481 | else | |
5482 | t2 = fold_convert (itype, step); | |
5483 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
5484 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
5485 | NULL_TREE, true, GSI_SAME_STMT); | |
5486 | assign_stmt = gimple_build_assign (cond_var, t2); | |
5487 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5488 | } | |
5489 | ||
629b3d75 MJ |
5490 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
5491 | { | |
5492 | if (POINTER_TYPE_P (type)) | |
5493 | t = fold_build_pointer_plus (vmain, step); | |
5494 | else | |
5495 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
5496 | t = force_gimple_operand_gsi (&gsi, t, | |
5497 | DECL_P (vback) | |
5498 | && TREE_ADDRESSABLE (vback), | |
5499 | NULL_TREE, true, GSI_SAME_STMT); | |
5500 | assign_stmt = gimple_build_assign (vback, t); | |
5501 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5502 | ||
5503 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
5504 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
5505 | ? t : vback, e); | |
5506 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
5507 | } | |
5508 | ||
5509 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ | |
5510 | gsi_remove (&gsi, true); | |
5511 | ||
5512 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
aed3ab25 JJ |
5513 | collapse_bb = extract_omp_for_update_vars (fd, nonrect_bounds, |
5514 | cont_bb, body_bb); | |
629b3d75 MJ |
5515 | } |
5516 | ||
5517 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ | |
65f4b875 | 5518 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
5519 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) |
5520 | { | |
5521 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
2f6bb511 JJ |
5522 | if (fd->have_reductemp |
5523 | || ((fd->have_pointer_condtemp || fd->have_scantemp) | |
5524 | && !fd->have_nonctrl_scantemp)) | |
28567c40 JJ |
5525 | { |
5526 | tree fn; | |
5527 | if (t) | |
5528 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
5529 | else | |
5530 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
5531 | gcall *g = gimple_build_call (fn, 0); | |
5532 | if (t) | |
5533 | { | |
5534 | gimple_call_set_lhs (g, t); | |
6c7ae8c5 JJ |
5535 | if (fd->have_reductemp) |
5536 | gsi_insert_after (&gsi, gimple_build_assign (reductions, | |
5537 | NOP_EXPR, t), | |
5538 | GSI_SAME_STMT); | |
28567c40 JJ |
5539 | } |
5540 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
5541 | } | |
5542 | else | |
5543 | gsi_insert_after (&gsi, omp_build_barrier (t), GSI_SAME_STMT); | |
629b3d75 | 5544 | } |
2f03073f JJ |
5545 | else if ((fd->have_pointer_condtemp || fd->have_scantemp) |
5546 | && !fd->have_nonctrl_scantemp) | |
1a39b3d3 JJ |
5547 | { |
5548 | tree fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
5549 | gcall *g = gimple_build_call (fn, 0); | |
5550 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
5551 | } | |
2f6bb511 JJ |
5552 | if (fd->have_scantemp && !fd->have_nonctrl_scantemp) |
5553 | { | |
5554 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5555 | tree controlp = NULL_TREE, controlb = NULL_TREE; | |
5556 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
5557 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
5558 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) | |
5559 | { | |
5560 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) | |
5561 | controlb = OMP_CLAUSE_DECL (c); | |
5562 | else | |
5563 | controlp = OMP_CLAUSE_DECL (c); | |
5564 | if (controlb && controlp) | |
5565 | break; | |
5566 | } | |
5567 | gcc_assert (controlp && controlb); | |
5568 | gimple *g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, | |
5569 | NULL_TREE, NULL_TREE); | |
5570 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5571 | exit1_bb = split_block (exit_bb, g)->dest; | |
5572 | gsi = gsi_after_labels (exit1_bb); | |
5573 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_FREE), 1, | |
5574 | controlp); | |
5575 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5576 | exit2_bb = split_block (exit1_bb, g)->dest; | |
5577 | gsi = gsi_after_labels (exit2_bb); | |
5578 | g = gimple_build_call (builtin_decl_implicit (BUILT_IN_STACK_RESTORE), 1, | |
5579 | controlp); | |
5580 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5581 | exit3_bb = split_block (exit2_bb, g)->dest; | |
5582 | gsi = gsi_after_labels (exit3_bb); | |
5583 | } | |
629b3d75 MJ |
5584 | gsi_remove (&gsi, true); |
5585 | ||
5586 | /* Connect all the blocks. */ | |
5587 | ep = make_edge (entry_bb, third_bb, EDGE_FALSE_VALUE); | |
357067f2 | 5588 | ep->probability = profile_probability::guessed_always ().apply_scale (3, 4); |
629b3d75 MJ |
5589 | ep = find_edge (entry_bb, second_bb); |
5590 | ep->flags = EDGE_TRUE_VALUE; | |
9f55aee9 | 5591 | ep->probability = profile_probability::guessed_always () / 4; |
2f6bb511 JJ |
5592 | if (fourth_bb) |
5593 | { | |
5594 | ep = make_edge (third_bb, fifth_bb, EDGE_FALSE_VALUE); | |
9f55aee9 | 5595 | ep->probability = profile_probability::guessed_always () / 2; |
2f6bb511 JJ |
5596 | ep = find_edge (third_bb, fourth_bb); |
5597 | ep->flags = EDGE_TRUE_VALUE; | |
9f55aee9 | 5598 | ep->probability = profile_probability::guessed_always () / 2; |
2f6bb511 JJ |
5599 | ep = find_edge (fourth_bb, fifth_bb); |
5600 | redirect_edge_and_branch (ep, sixth_bb); | |
5601 | } | |
5602 | else | |
5603 | sixth_bb = third_bb; | |
5604 | find_edge (sixth_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE; | |
5605 | find_edge (sixth_bb, fin_bb)->flags = EDGE_TRUE_VALUE; | |
5606 | if (exit1_bb) | |
5607 | { | |
5608 | ep = make_edge (exit_bb, exit2_bb, EDGE_FALSE_VALUE); | |
9f55aee9 | 5609 | ep->probability = profile_probability::guessed_always () / 2; |
2f6bb511 JJ |
5610 | ep = find_edge (exit_bb, exit1_bb); |
5611 | ep->flags = EDGE_TRUE_VALUE; | |
9f55aee9 | 5612 | ep->probability = profile_probability::guessed_always () / 2; |
2f6bb511 JJ |
5613 | ep = find_edge (exit1_bb, exit2_bb); |
5614 | redirect_edge_and_branch (ep, exit3_bb); | |
5615 | } | |
629b3d75 MJ |
5616 | |
5617 | if (!broken_loop) | |
5618 | { | |
5619 | ep = find_edge (cont_bb, body_bb); | |
5620 | if (ep == NULL) | |
5621 | { | |
5622 | ep = BRANCH_EDGE (cont_bb); | |
5623 | gcc_assert (single_succ (ep->dest) == body_bb); | |
5624 | } | |
5625 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5626 | { | |
5627 | remove_edge (ep); | |
5628 | ep = NULL; | |
5629 | } | |
5630 | else if (fd->collapse > 1) | |
5631 | { | |
5632 | remove_edge (ep); | |
5633 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
5634 | } | |
5635 | else | |
5636 | ep->flags = EDGE_TRUE_VALUE; | |
5637 | find_edge (cont_bb, fin_bb)->flags | |
5638 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
5639 | } | |
5640 | ||
5641 | set_immediate_dominator (CDI_DOMINATORS, second_bb, entry_bb); | |
5642 | set_immediate_dominator (CDI_DOMINATORS, third_bb, entry_bb); | |
2f6bb511 JJ |
5643 | if (fourth_bb) |
5644 | { | |
5645 | set_immediate_dominator (CDI_DOMINATORS, fifth_bb, third_bb); | |
5646 | set_immediate_dominator (CDI_DOMINATORS, sixth_bb, third_bb); | |
5647 | } | |
5648 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, sixth_bb); | |
629b3d75 MJ |
5649 | |
5650 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
5651 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
5652 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
5653 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
2f6bb511 JJ |
5654 | if (exit1_bb) |
5655 | { | |
5656 | set_immediate_dominator (CDI_DOMINATORS, exit2_bb, exit_bb); | |
5657 | set_immediate_dominator (CDI_DOMINATORS, exit3_bb, exit_bb); | |
5658 | } | |
629b3d75 | 5659 | |
99b1c316 | 5660 | class loop *loop = body_bb->loop_father; |
629b3d75 MJ |
5661 | if (loop != entry_bb->loop_father) |
5662 | { | |
5663 | gcc_assert (broken_loop || loop->header == body_bb); | |
5664 | gcc_assert (broken_loop | |
5665 | || loop->latch == region->cont | |
5666 | || single_pred (loop->latch) == region->cont); | |
5667 | return; | |
5668 | } | |
5669 | ||
5670 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) | |
5671 | { | |
5672 | loop = alloc_loop (); | |
5673 | loop->header = body_bb; | |
5674 | if (collapse_bb == NULL) | |
5675 | loop->latch = cont_bb; | |
5676 | add_loop (loop, body_bb->loop_father); | |
5677 | } | |
5678 | } | |
5679 | ||
5680 | /* Return phi in E->DEST with ARG on edge E. */ | |
5681 | ||
5682 | static gphi * | |
5683 | find_phi_with_arg_on_edge (tree arg, edge e) | |
5684 | { | |
5685 | basic_block bb = e->dest; | |
5686 | ||
5687 | for (gphi_iterator gpi = gsi_start_phis (bb); | |
5688 | !gsi_end_p (gpi); | |
5689 | gsi_next (&gpi)) | |
5690 | { | |
5691 | gphi *phi = gpi.phi (); | |
5692 | if (PHI_ARG_DEF_FROM_EDGE (phi, e) == arg) | |
5693 | return phi; | |
5694 | } | |
5695 | ||
5696 | return NULL; | |
5697 | } | |
5698 | ||
5699 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
5700 | loop with static schedule and a specified chunk size. Given | |
5701 | parameters: | |
5702 | ||
5703 | for (V = N1; V cond N2; V += STEP) BODY; | |
5704 | ||
5705 | where COND is "<" or ">", we generate pseudocode | |
5706 | ||
5707 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; | |
5708 | if (cond is <) | |
5709 | adj = STEP - 1; | |
5710 | else | |
5711 | adj = STEP + 1; | |
5712 | if ((__typeof (V)) -1 > 0 && cond is >) | |
5713 | n = -(adj + N2 - N1) / -STEP; | |
5714 | else | |
5715 | n = (adj + N2 - N1) / STEP; | |
5716 | trip = 0; | |
5717 | V = threadid * CHUNK * STEP + N1; -- this extra definition of V is | |
5718 | here so that V is defined | |
5719 | if the loop is not entered | |
5720 | L0: | |
5721 | s0 = (trip * nthreads + threadid) * CHUNK; | |
01914336 | 5722 | e0 = min (s0 + CHUNK, n); |
629b3d75 MJ |
5723 | if (s0 < n) goto L1; else goto L4; |
5724 | L1: | |
5725 | V = s0 * STEP + N1; | |
5726 | e = e0 * STEP + N1; | |
5727 | L2: | |
5728 | BODY; | |
5729 | V += STEP; | |
5730 | if (V cond e) goto L2; else goto L3; | |
5731 | L3: | |
5732 | trip += 1; | |
5733 | goto L0; | |
5734 | L4: | |
5735 | */ | |
5736 | ||
5737 | static void | |
5738 | expand_omp_for_static_chunk (struct omp_region *region, | |
5739 | struct omp_for_data *fd, gimple *inner_stmt) | |
5740 | { | |
5741 | tree n, s0, e0, e, t; | |
5742 | tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid; | |
5743 | tree type, itype, vmain, vback, vextra; | |
5744 | basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb; | |
5745 | basic_block trip_update_bb = NULL, cont_bb, collapse_bb = NULL, fin_bb; | |
6c7ae8c5 | 5746 | gimple_stmt_iterator gsi, gsip; |
629b3d75 MJ |
5747 | edge se; |
5748 | bool broken_loop = region->cont == NULL; | |
5749 | tree *counts = NULL; | |
5750 | tree n1, n2, step; | |
28567c40 | 5751 | tree reductions = NULL_TREE; |
8221c30b | 5752 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; |
629b3d75 MJ |
5753 | |
5754 | itype = type = TREE_TYPE (fd->loop.v); | |
5755 | if (POINTER_TYPE_P (type)) | |
5756 | itype = signed_type_for (type); | |
5757 | ||
5758 | entry_bb = region->entry; | |
fe8ac82f | 5759 | se = split_block (entry_bb, last_nondebug_stmt (entry_bb)); |
629b3d75 MJ |
5760 | entry_bb = se->src; |
5761 | iter_part_bb = se->dest; | |
5762 | cont_bb = region->cont; | |
5763 | gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2); | |
5764 | fin_bb = BRANCH_EDGE (iter_part_bb)->dest; | |
5765 | gcc_assert (broken_loop | |
5766 | || fin_bb == FALLTHRU_EDGE (cont_bb)->dest); | |
5767 | seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb)); | |
5768 | body_bb = single_succ (seq_start_bb); | |
5769 | if (!broken_loop) | |
5770 | { | |
5771 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb | |
5772 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); | |
5773 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
5774 | trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb)); | |
5775 | } | |
5776 | exit_bb = region->exit; | |
5777 | ||
5778 | /* Trip and adjustment setup goes in ENTRY_BB. */ | |
65f4b875 | 5779 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 | 5780 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
6c7ae8c5 JJ |
5781 | gsip = gsi; |
5782 | gsi_prev (&gsip); | |
629b3d75 MJ |
5783 | |
5784 | if (fd->collapse > 1) | |
5785 | { | |
5786 | int first_zero_iter = -1, dummy = -1; | |
5787 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
5788 | ||
5789 | counts = XALLOCAVEC (tree, fd->collapse); | |
5790 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5791 | fin_bb, first_zero_iter, | |
5792 | dummy_bb, dummy, l2_dom_bb); | |
5793 | t = NULL_TREE; | |
5794 | } | |
5795 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5796 | t = integer_one_node; | |
5797 | else | |
5798 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
5799 | fold_convert (type, fd->loop.n1), | |
5800 | fold_convert (type, fd->loop.n2)); | |
5801 | if (fd->collapse == 1 | |
5802 | && TYPE_UNSIGNED (type) | |
5803 | && (t == NULL_TREE || !integer_onep (t))) | |
5804 | { | |
5805 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); | |
5806 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
5807 | true, GSI_SAME_STMT); | |
5808 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
5809 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
5810 | true, GSI_SAME_STMT); | |
47de0b56 JJ |
5811 | gcond *cond_stmt = expand_omp_build_cond (&gsi, fd->loop.cond_code, |
5812 | n1, n2); | |
629b3d75 MJ |
5813 | se = split_block (entry_bb, cond_stmt); |
5814 | se->flags = EDGE_TRUE_VALUE; | |
5815 | entry_bb = se->dest; | |
357067f2 | 5816 | se->probability = profile_probability::very_likely (); |
629b3d75 | 5817 | se = make_edge (se->src, fin_bb, EDGE_FALSE_VALUE); |
357067f2 | 5818 | se->probability = profile_probability::very_unlikely (); |
629b3d75 MJ |
5819 | if (gimple_in_ssa_p (cfun)) |
5820 | { | |
5821 | int dest_idx = find_edge (iter_part_bb, fin_bb)->dest_idx; | |
5822 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); | |
5823 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
5824 | { | |
5825 | gphi *phi = gpi.phi (); | |
5826 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
5827 | se, UNKNOWN_LOCATION); | |
5828 | } | |
5829 | } | |
5830 | gsi = gsi_last_bb (entry_bb); | |
5831 | } | |
5832 | ||
8221c30b JJ |
5833 | if (fd->lastprivate_conditional) |
5834 | { | |
5835 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
5836 | tree c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); | |
5837 | if (fd->have_pointer_condtemp) | |
5838 | condtemp = OMP_CLAUSE_DECL (c); | |
5839 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
5840 | cond_var = OMP_CLAUSE_DECL (c); | |
5841 | } | |
5842 | if (fd->have_reductemp || fd->have_pointer_condtemp) | |
28567c40 JJ |
5843 | { |
5844 | tree t1 = build_int_cst (long_integer_type_node, 0); | |
5845 | tree t2 = build_int_cst (long_integer_type_node, 1); | |
5846 | tree t3 = build_int_cstu (long_integer_type_node, | |
5847 | (HOST_WIDE_INT_1U << 31) + 1); | |
5848 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
6c7ae8c5 JJ |
5849 | gimple_stmt_iterator gsi2 = gsi_none (); |
5850 | gimple *g = NULL; | |
5851 | tree mem = null_pointer_node, memv = NULL_TREE; | |
6c7ae8c5 JJ |
5852 | if (fd->have_reductemp) |
5853 | { | |
5854 | tree c = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
5855 | reductions = OMP_CLAUSE_DECL (c); | |
5856 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
5857 | g = SSA_NAME_DEF_STMT (reductions); | |
5858 | reductions = gimple_assign_rhs1 (g); | |
5859 | OMP_CLAUSE_DECL (c) = reductions; | |
5860 | gsi2 = gsi_for_stmt (g); | |
5861 | } | |
5862 | else | |
5863 | { | |
5864 | if (gsi_end_p (gsip)) | |
5865 | gsi2 = gsi_after_labels (region->entry); | |
5866 | else | |
5867 | gsi2 = gsip; | |
5868 | reductions = null_pointer_node; | |
5869 | } | |
8221c30b | 5870 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 5871 | { |
6c7ae8c5 JJ |
5872 | tree type = TREE_TYPE (condtemp); |
5873 | memv = create_tmp_var (type); | |
5874 | TREE_ADDRESSABLE (memv) = 1; | |
5875 | unsigned HOST_WIDE_INT sz | |
5876 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
5877 | sz *= fd->lastprivate_conditional; | |
5878 | expand_omp_build_assign (&gsi2, memv, build_int_cst (type, sz), | |
5879 | false); | |
5880 | mem = build_fold_addr_expr (memv); | |
5881 | } | |
28567c40 JJ |
5882 | tree t |
5883 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_LOOP_START), | |
5884 | 9, t1, t2, t2, t3, t1, null_pointer_node, | |
6c7ae8c5 | 5885 | null_pointer_node, reductions, mem); |
28567c40 JJ |
5886 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
5887 | true, GSI_SAME_STMT); | |
8221c30b | 5888 | if (fd->have_pointer_condtemp) |
6c7ae8c5 JJ |
5889 | expand_omp_build_assign (&gsi2, condtemp, memv, false); |
5890 | if (fd->have_reductemp) | |
5891 | { | |
5892 | gsi_remove (&gsi2, true); | |
5893 | release_ssa_name (gimple_assign_lhs (g)); | |
5894 | } | |
28567c40 | 5895 | } |
629b3d75 MJ |
5896 | switch (gimple_omp_for_kind (fd->for_stmt)) |
5897 | { | |
5898 | case GF_OMP_FOR_KIND_FOR: | |
5899 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
5900 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
5901 | break; | |
5902 | case GF_OMP_FOR_KIND_DISTRIBUTE: | |
5903 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS); | |
5904 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM); | |
5905 | break; | |
5906 | default: | |
5907 | gcc_unreachable (); | |
5908 | } | |
5909 | nthreads = build_call_expr (nthreads, 0); | |
5910 | nthreads = fold_convert (itype, nthreads); | |
5911 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, | |
5912 | true, GSI_SAME_STMT); | |
5913 | threadid = build_call_expr (threadid, 0); | |
5914 | threadid = fold_convert (itype, threadid); | |
5915 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, | |
5916 | true, GSI_SAME_STMT); | |
5917 | ||
5918 | n1 = fd->loop.n1; | |
5919 | n2 = fd->loop.n2; | |
5920 | step = fd->loop.step; | |
5921 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5922 | { | |
5923 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5924 | OMP_CLAUSE__LOOPTEMP_); | |
5925 | gcc_assert (innerc); | |
5926 | n1 = OMP_CLAUSE_DECL (innerc); | |
5927 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5928 | OMP_CLAUSE__LOOPTEMP_); | |
5929 | gcc_assert (innerc); | |
5930 | n2 = OMP_CLAUSE_DECL (innerc); | |
5931 | } | |
5932 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
5933 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5934 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
5935 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5936 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
5937 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5938 | tree chunk_size = fold_convert (itype, fd->chunk_size); | |
5939 | chunk_size = omp_adjust_chunk_size (chunk_size, fd->simd_schedule); | |
5940 | chunk_size | |
5941 | = force_gimple_operand_gsi (&gsi, chunk_size, true, NULL_TREE, true, | |
5942 | GSI_SAME_STMT); | |
5943 | ||
5944 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
5945 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
5946 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
5947 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
5948 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
5949 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
5950 | fold_build1 (NEGATE_EXPR, itype, t), | |
5951 | fold_build1 (NEGATE_EXPR, itype, step)); | |
5952 | else | |
5953 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
5954 | t = fold_convert (itype, t); | |
5955 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5956 | true, GSI_SAME_STMT); | |
5957 | ||
5958 | trip_var = create_tmp_reg (itype, ".trip"); | |
5959 | if (gimple_in_ssa_p (cfun)) | |
5960 | { | |
5961 | trip_init = make_ssa_name (trip_var); | |
5962 | trip_main = make_ssa_name (trip_var); | |
5963 | trip_back = make_ssa_name (trip_var); | |
5964 | } | |
5965 | else | |
5966 | { | |
5967 | trip_init = trip_var; | |
5968 | trip_main = trip_var; | |
5969 | trip_back = trip_var; | |
5970 | } | |
5971 | ||
5972 | gassign *assign_stmt | |
5973 | = gimple_build_assign (trip_init, build_int_cst (itype, 0)); | |
5974 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5975 | ||
5976 | t = fold_build2 (MULT_EXPR, itype, threadid, chunk_size); | |
5977 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
5978 | if (POINTER_TYPE_P (type)) | |
5979 | t = fold_build_pointer_plus (n1, t); | |
5980 | else | |
5981 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
5982 | vextra = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5983 | true, GSI_SAME_STMT); | |
5984 | ||
5985 | /* Remove the GIMPLE_OMP_FOR. */ | |
5986 | gsi_remove (&gsi, true); | |
5987 | ||
5988 | gimple_stmt_iterator gsif = gsi; | |
5989 | ||
5990 | /* Iteration space partitioning goes in ITER_PART_BB. */ | |
5991 | gsi = gsi_last_bb (iter_part_bb); | |
5992 | ||
5993 | t = fold_build2 (MULT_EXPR, itype, trip_main, nthreads); | |
5994 | t = fold_build2 (PLUS_EXPR, itype, t, threadid); | |
5995 | t = fold_build2 (MULT_EXPR, itype, t, chunk_size); | |
5996 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5997 | false, GSI_CONTINUE_LINKING); | |
5998 | ||
5999 | t = fold_build2 (PLUS_EXPR, itype, s0, chunk_size); | |
6000 | t = fold_build2 (MIN_EXPR, itype, t, n); | |
6001 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6002 | false, GSI_CONTINUE_LINKING); | |
6003 | ||
6004 | t = build2 (LT_EXPR, boolean_type_node, s0, n); | |
6005 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_CONTINUE_LINKING); | |
6006 | ||
6007 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
6008 | gsi = gsi_start_bb (seq_start_bb); | |
6009 | ||
6010 | tree startvar = fd->loop.v; | |
6011 | tree endvar = NULL_TREE; | |
6012 | ||
6013 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6014 | { | |
6015 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
6016 | ? gimple_omp_parallel_clauses (inner_stmt) | |
6017 | : gimple_omp_for_clauses (inner_stmt); | |
6018 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
6019 | gcc_assert (innerc); | |
6020 | startvar = OMP_CLAUSE_DECL (innerc); | |
6021 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
6022 | OMP_CLAUSE__LOOPTEMP_); | |
6023 | gcc_assert (innerc); | |
6024 | endvar = OMP_CLAUSE_DECL (innerc); | |
6025 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST | |
6026 | && gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) | |
6027 | { | |
14707c89 | 6028 | innerc = find_lastprivate_looptemp (fd, innerc); |
629b3d75 MJ |
6029 | if (innerc) |
6030 | { | |
6031 | /* If needed (distribute parallel for with lastprivate), | |
6032 | propagate down the total number of iterations. */ | |
6033 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), | |
6034 | fd->loop.n2); | |
6035 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, | |
6036 | GSI_CONTINUE_LINKING); | |
6037 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
6038 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6039 | } | |
6040 | } | |
6041 | } | |
6042 | ||
6043 | t = fold_convert (itype, s0); | |
6044 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
6045 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
6046 | { |
6047 | t = fold_build_pointer_plus (n1, t); | |
6048 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
6049 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
6050 | t = fold_convert (signed_type_for (type), t); | |
6051 | } | |
629b3d75 MJ |
6052 | else |
6053 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
6054 | t = fold_convert (TREE_TYPE (startvar), t); | |
6055 | t = force_gimple_operand_gsi (&gsi, t, | |
6056 | DECL_P (startvar) | |
6057 | && TREE_ADDRESSABLE (startvar), | |
6058 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
6059 | assign_stmt = gimple_build_assign (startvar, t); | |
6060 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6c7ae8c5 JJ |
6061 | if (cond_var) |
6062 | { | |
6063 | tree itype = TREE_TYPE (cond_var); | |
6064 | /* For lastprivate(conditional:) itervar, we need some iteration | |
6065 | counter that starts at unsigned non-zero and increases. | |
6066 | Prefer as few IVs as possible, so if we can use startvar | |
6067 | itself, use that, or startvar + constant (those would be | |
6068 | incremented with step), and as last resort use the s0 + 1 | |
6069 | incremented by 1. */ | |
6070 | if (POINTER_TYPE_P (type) | |
6071 | || TREE_CODE (n1) != INTEGER_CST | |
6072 | || fd->loop.cond_code != LT_EXPR) | |
6073 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), | |
6074 | build_int_cst (itype, 1)); | |
6075 | else if (tree_int_cst_sgn (n1) == 1) | |
6076 | t = fold_convert (itype, t); | |
6077 | else | |
6078 | { | |
6079 | tree c = fold_convert (itype, n1); | |
6080 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
6081 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
6082 | } | |
6083 | t = force_gimple_operand_gsi (&gsi, t, false, | |
6084 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
6085 | assign_stmt = gimple_build_assign (cond_var, t); | |
6086 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6087 | } | |
629b3d75 MJ |
6088 | |
6089 | t = fold_convert (itype, e0); | |
6090 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
6091 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
6092 | { |
6093 | t = fold_build_pointer_plus (n1, t); | |
6094 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
6095 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
6096 | t = fold_convert (signed_type_for (type), t); | |
6097 | } | |
629b3d75 MJ |
6098 | else |
6099 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
6100 | t = fold_convert (TREE_TYPE (startvar), t); | |
6101 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6102 | false, GSI_CONTINUE_LINKING); | |
6103 | if (endvar) | |
6104 | { | |
6105 | assign_stmt = gimple_build_assign (endvar, e); | |
6106 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6107 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
6108 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
6109 | else | |
6110 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
6111 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6112 | } | |
6113 | /* Handle linear clause adjustments. */ | |
6114 | tree itercnt = NULL_TREE, itercntbias = NULL_TREE; | |
6115 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
6116 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
6117 | c; c = OMP_CLAUSE_CHAIN (c)) | |
6118 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
6119 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
6120 | { | |
6121 | tree d = OMP_CLAUSE_DECL (c); | |
629b3d75 | 6122 | tree t = d, a, dest; |
22e6b327 | 6123 | if (omp_privatize_by_reference (t)) |
629b3d75 MJ |
6124 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); |
6125 | tree type = TREE_TYPE (t); | |
6126 | if (POINTER_TYPE_P (type)) | |
6127 | type = sizetype; | |
6128 | dest = unshare_expr (t); | |
6129 | tree v = create_tmp_var (TREE_TYPE (t), NULL); | |
6130 | expand_omp_build_assign (&gsif, v, t); | |
6131 | if (itercnt == NULL_TREE) | |
6132 | { | |
6133 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6134 | { | |
6135 | itercntbias | |
6136 | = fold_build2 (MINUS_EXPR, itype, fold_convert (itype, n1), | |
6137 | fold_convert (itype, fd->loop.n1)); | |
6138 | itercntbias = fold_build2 (EXACT_DIV_EXPR, itype, | |
6139 | itercntbias, step); | |
6140 | itercntbias | |
6141 | = force_gimple_operand_gsi (&gsif, itercntbias, true, | |
6142 | NULL_TREE, true, | |
6143 | GSI_SAME_STMT); | |
6144 | itercnt = fold_build2 (PLUS_EXPR, itype, itercntbias, s0); | |
6145 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
6146 | NULL_TREE, false, | |
6147 | GSI_CONTINUE_LINKING); | |
6148 | } | |
6149 | else | |
6150 | itercnt = s0; | |
6151 | } | |
6152 | a = fold_build2 (MULT_EXPR, type, | |
6153 | fold_convert (type, itercnt), | |
6154 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
6155 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
6156 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); | |
6157 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6158 | false, GSI_CONTINUE_LINKING); | |
6fcc3cac | 6159 | expand_omp_build_assign (&gsi, dest, t, true); |
629b3d75 MJ |
6160 | } |
6161 | if (fd->collapse > 1) | |
aed3ab25 | 6162 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, inner_stmt, startvar); |
629b3d75 MJ |
6163 | |
6164 | if (!broken_loop) | |
6165 | { | |
6166 | /* The code controlling the sequential loop goes in CONT_BB, | |
6167 | replacing the GIMPLE_OMP_CONTINUE. */ | |
65f4b875 | 6168 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
6169 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
6170 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
6171 | vback = gimple_omp_continue_control_def (cont_stmt); | |
6172 | ||
7855700e JJ |
6173 | if (cond_var) |
6174 | { | |
6175 | tree itype = TREE_TYPE (cond_var); | |
6176 | tree t2; | |
6177 | if (POINTER_TYPE_P (type) | |
6178 | || TREE_CODE (n1) != INTEGER_CST | |
6179 | || fd->loop.cond_code != LT_EXPR) | |
6180 | t2 = build_int_cst (itype, 1); | |
6181 | else | |
6182 | t2 = fold_convert (itype, step); | |
6183 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
6184 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
6185 | NULL_TREE, true, GSI_SAME_STMT); | |
6186 | assign_stmt = gimple_build_assign (cond_var, t2); | |
6187 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
6188 | } | |
6189 | ||
629b3d75 MJ |
6190 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
6191 | { | |
6192 | if (POINTER_TYPE_P (type)) | |
6193 | t = fold_build_pointer_plus (vmain, step); | |
6194 | else | |
6195 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
6196 | if (DECL_P (vback) && TREE_ADDRESSABLE (vback)) | |
6197 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
6198 | true, GSI_SAME_STMT); | |
6199 | assign_stmt = gimple_build_assign (vback, t); | |
6200 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
6201 | ||
6202 | if (tree_int_cst_equal (fd->chunk_size, integer_one_node)) | |
6203 | t = build2 (EQ_EXPR, boolean_type_node, | |
6204 | build_int_cst (itype, 0), | |
6205 | build_int_cst (itype, 1)); | |
6206 | else | |
6207 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
6208 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
6209 | ? t : vback, e); | |
6210 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
6211 | } | |
6212 | ||
6213 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
6214 | gsi_remove (&gsi, true); | |
6215 | ||
6216 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
aed3ab25 | 6217 | collapse_bb = extract_omp_for_update_vars (fd, NULL, cont_bb, body_bb); |
629b3d75 MJ |
6218 | |
6219 | /* Trip update code goes into TRIP_UPDATE_BB. */ | |
6220 | gsi = gsi_start_bb (trip_update_bb); | |
6221 | ||
6222 | t = build_int_cst (itype, 1); | |
6223 | t = build2 (PLUS_EXPR, itype, trip_main, t); | |
6224 | assign_stmt = gimple_build_assign (trip_back, t); | |
6225 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6226 | } | |
6227 | ||
6228 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ | |
65f4b875 | 6229 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
6230 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) |
6231 | { | |
6232 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
8221c30b | 6233 | if (fd->have_reductemp || fd->have_pointer_condtemp) |
28567c40 JJ |
6234 | { |
6235 | tree fn; | |
6236 | if (t) | |
6237 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
6238 | else | |
6239 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
6240 | gcall *g = gimple_build_call (fn, 0); | |
6241 | if (t) | |
6242 | { | |
6243 | gimple_call_set_lhs (g, t); | |
6c7ae8c5 JJ |
6244 | if (fd->have_reductemp) |
6245 | gsi_insert_after (&gsi, gimple_build_assign (reductions, | |
6246 | NOP_EXPR, t), | |
6247 | GSI_SAME_STMT); | |
28567c40 JJ |
6248 | } |
6249 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
6250 | } | |
6251 | else | |
6252 | gsi_insert_after (&gsi, omp_build_barrier (t), GSI_SAME_STMT); | |
629b3d75 | 6253 | } |
1a39b3d3 JJ |
6254 | else if (fd->have_pointer_condtemp) |
6255 | { | |
6256 | tree fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
6257 | gcall *g = gimple_build_call (fn, 0); | |
6258 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
6259 | } | |
629b3d75 MJ |
6260 | gsi_remove (&gsi, true); |
6261 | ||
6262 | /* Connect the new blocks. */ | |
6263 | find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE; | |
6264 | find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE; | |
6265 | ||
6266 | if (!broken_loop) | |
6267 | { | |
6268 | se = find_edge (cont_bb, body_bb); | |
6269 | if (se == NULL) | |
6270 | { | |
6271 | se = BRANCH_EDGE (cont_bb); | |
6272 | gcc_assert (single_succ (se->dest) == body_bb); | |
6273 | } | |
6274 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
6275 | { | |
6276 | remove_edge (se); | |
6277 | se = NULL; | |
6278 | } | |
6279 | else if (fd->collapse > 1) | |
6280 | { | |
6281 | remove_edge (se); | |
6282 | se = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
6283 | } | |
6284 | else | |
6285 | se->flags = EDGE_TRUE_VALUE; | |
6286 | find_edge (cont_bb, trip_update_bb)->flags | |
6287 | = se ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
6288 | ||
01914336 MJ |
6289 | redirect_edge_and_branch (single_succ_edge (trip_update_bb), |
6290 | iter_part_bb); | |
629b3d75 MJ |
6291 | } |
6292 | ||
6293 | if (gimple_in_ssa_p (cfun)) | |
6294 | { | |
6295 | gphi_iterator psi; | |
6296 | gphi *phi; | |
6297 | edge re, ene; | |
6298 | edge_var_map *vm; | |
6299 | size_t i; | |
6300 | ||
6301 | gcc_assert (fd->collapse == 1 && !broken_loop); | |
6302 | ||
6303 | /* When we redirect the edge from trip_update_bb to iter_part_bb, we | |
6304 | remove arguments of the phi nodes in fin_bb. We need to create | |
6305 | appropriate phi nodes in iter_part_bb instead. */ | |
6306 | se = find_edge (iter_part_bb, fin_bb); | |
6307 | re = single_succ_edge (trip_update_bb); | |
6308 | vec<edge_var_map> *head = redirect_edge_var_map_vector (re); | |
6309 | ene = single_succ_edge (entry_bb); | |
6310 | ||
6311 | psi = gsi_start_phis (fin_bb); | |
6312 | for (i = 0; !gsi_end_p (psi) && head->iterate (i, &vm); | |
6313 | gsi_next (&psi), ++i) | |
6314 | { | |
6315 | gphi *nphi; | |
620e594b | 6316 | location_t locus; |
629b3d75 MJ |
6317 | |
6318 | phi = psi.phi (); | |
d83cc5cc TV |
6319 | if (operand_equal_p (gimple_phi_arg_def (phi, 0), |
6320 | redirect_edge_var_map_def (vm), 0)) | |
6321 | continue; | |
6322 | ||
629b3d75 MJ |
6323 | t = gimple_phi_result (phi); |
6324 | gcc_assert (t == redirect_edge_var_map_result (vm)); | |
6325 | ||
6326 | if (!single_pred_p (fin_bb)) | |
6327 | t = copy_ssa_name (t, phi); | |
6328 | ||
6329 | nphi = create_phi_node (t, iter_part_bb); | |
6330 | ||
6331 | t = PHI_ARG_DEF_FROM_EDGE (phi, se); | |
6332 | locus = gimple_phi_arg_location_from_edge (phi, se); | |
6333 | ||
6334 | /* A special case -- fd->loop.v is not yet computed in | |
6335 | iter_part_bb, we need to use vextra instead. */ | |
6336 | if (t == fd->loop.v) | |
6337 | t = vextra; | |
6338 | add_phi_arg (nphi, t, ene, locus); | |
6339 | locus = redirect_edge_var_map_location (vm); | |
6340 | tree back_arg = redirect_edge_var_map_def (vm); | |
6341 | add_phi_arg (nphi, back_arg, re, locus); | |
6342 | edge ce = find_edge (cont_bb, body_bb); | |
6343 | if (ce == NULL) | |
6344 | { | |
6345 | ce = BRANCH_EDGE (cont_bb); | |
6346 | gcc_assert (single_succ (ce->dest) == body_bb); | |
6347 | ce = single_succ_edge (ce->dest); | |
6348 | } | |
6349 | gphi *inner_loop_phi = find_phi_with_arg_on_edge (back_arg, ce); | |
6350 | gcc_assert (inner_loop_phi != NULL); | |
6351 | add_phi_arg (inner_loop_phi, gimple_phi_result (nphi), | |
6352 | find_edge (seq_start_bb, body_bb), locus); | |
6353 | ||
6354 | if (!single_pred_p (fin_bb)) | |
6355 | add_phi_arg (phi, gimple_phi_result (nphi), se, locus); | |
6356 | } | |
6357 | gcc_assert (gsi_end_p (psi) && (head == NULL || i == head->length ())); | |
6358 | redirect_edge_var_map_clear (re); | |
6359 | if (single_pred_p (fin_bb)) | |
6360 | while (1) | |
6361 | { | |
6362 | psi = gsi_start_phis (fin_bb); | |
6363 | if (gsi_end_p (psi)) | |
6364 | break; | |
6365 | remove_phi_node (&psi, false); | |
6366 | } | |
6367 | ||
6368 | /* Make phi node for trip. */ | |
6369 | phi = create_phi_node (trip_main, iter_part_bb); | |
6370 | add_phi_arg (phi, trip_back, single_succ_edge (trip_update_bb), | |
6371 | UNKNOWN_LOCATION); | |
6372 | add_phi_arg (phi, trip_init, single_succ_edge (entry_bb), | |
6373 | UNKNOWN_LOCATION); | |
6374 | } | |
6375 | ||
6376 | if (!broken_loop) | |
6377 | set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb); | |
6378 | set_immediate_dominator (CDI_DOMINATORS, iter_part_bb, | |
6379 | recompute_dominator (CDI_DOMINATORS, iter_part_bb)); | |
6380 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
6381 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
6382 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, | |
6383 | recompute_dominator (CDI_DOMINATORS, seq_start_bb)); | |
6384 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
6385 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
6386 | ||
6387 | if (!broken_loop) | |
6388 | { | |
99b1c316 MS |
6389 | class loop *loop = body_bb->loop_father; |
6390 | class loop *trip_loop = alloc_loop (); | |
629b3d75 MJ |
6391 | trip_loop->header = iter_part_bb; |
6392 | trip_loop->latch = trip_update_bb; | |
6393 | add_loop (trip_loop, iter_part_bb->loop_father); | |
6394 | ||
6395 | if (loop != entry_bb->loop_father) | |
6396 | { | |
6397 | gcc_assert (loop->header == body_bb); | |
6398 | gcc_assert (loop->latch == region->cont | |
6399 | || single_pred (loop->latch) == region->cont); | |
6400 | trip_loop->inner = loop; | |
6401 | return; | |
6402 | } | |
6403 | ||
6404 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
6405 | { | |
6406 | loop = alloc_loop (); | |
6407 | loop->header = body_bb; | |
6408 | if (collapse_bb == NULL) | |
6409 | loop->latch = cont_bb; | |
6410 | add_loop (loop, trip_loop); | |
6411 | } | |
6412 | } | |
6413 | } | |
6414 | ||
629b3d75 MJ |
6415 | /* A subroutine of expand_omp_for. Generate code for a simd non-worksharing |
6416 | loop. Given parameters: | |
6417 | ||
6418 | for (V = N1; V cond N2; V += STEP) BODY; | |
6419 | ||
6420 | where COND is "<" or ">", we generate pseudocode | |
6421 | ||
6422 | V = N1; | |
6423 | goto L1; | |
6424 | L0: | |
6425 | BODY; | |
6426 | V += STEP; | |
6427 | L1: | |
6428 | if (V cond N2) goto L0; else goto L2; | |
6429 | L2: | |
6430 | ||
c2ebf4f1 JJ |
6431 | For collapsed loops, emit the outer loops as scalar |
6432 | and only try to vectorize the innermost loop. */ | |
629b3d75 MJ |
6433 | |
6434 | static void | |
6435 | expand_omp_simd (struct omp_region *region, struct omp_for_data *fd) | |
6436 | { | |
6437 | tree type, t; | |
6438 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, l2_bb, l2_dom_bb; | |
6439 | gimple_stmt_iterator gsi; | |
6440 | gimple *stmt; | |
6441 | gcond *cond_stmt; | |
6442 | bool broken_loop = region->cont == NULL; | |
6443 | edge e, ne; | |
6444 | tree *counts = NULL; | |
6445 | int i; | |
6446 | int safelen_int = INT_MAX; | |
fed2a43c | 6447 | bool dont_vectorize = false; |
629b3d75 MJ |
6448 | tree safelen = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
6449 | OMP_CLAUSE_SAFELEN); | |
6450 | tree simduid = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6451 | OMP_CLAUSE__SIMDUID_); | |
fed2a43c JJ |
6452 | tree ifc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
6453 | OMP_CLAUSE_IF); | |
6454 | tree simdlen = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6455 | OMP_CLAUSE_SIMDLEN); | |
e7393c89 JJ |
6456 | tree condtemp = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
6457 | OMP_CLAUSE__CONDTEMP_); | |
629b3d75 | 6458 | tree n1, n2; |
e7393c89 | 6459 | tree cond_var = condtemp ? OMP_CLAUSE_DECL (condtemp) : NULL_TREE; |
629b3d75 MJ |
6460 | |
6461 | if (safelen) | |
6462 | { | |
9d2f08ab | 6463 | poly_uint64 val; |
629b3d75 | 6464 | safelen = OMP_CLAUSE_SAFELEN_EXPR (safelen); |
9d2f08ab | 6465 | if (!poly_int_tree_p (safelen, &val)) |
629b3d75 | 6466 | safelen_int = 0; |
9d2f08ab RS |
6467 | else |
6468 | safelen_int = MIN (constant_lower_bound (val), INT_MAX); | |
629b3d75 MJ |
6469 | if (safelen_int == 1) |
6470 | safelen_int = 0; | |
6471 | } | |
fed2a43c JJ |
6472 | if ((ifc && integer_zerop (OMP_CLAUSE_IF_EXPR (ifc))) |
6473 | || (simdlen && integer_onep (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)))) | |
6474 | { | |
6475 | safelen_int = 0; | |
6476 | dont_vectorize = true; | |
6477 | } | |
629b3d75 MJ |
6478 | type = TREE_TYPE (fd->loop.v); |
6479 | entry_bb = region->entry; | |
6480 | cont_bb = region->cont; | |
6481 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
6482 | gcc_assert (broken_loop | |
6483 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
6484 | l0_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
6485 | if (!broken_loop) | |
6486 | { | |
6487 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l0_bb); | |
6488 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
fe8ac82f | 6489 | l1_bb = split_block (cont_bb, last_nondebug_stmt (cont_bb))->dest; |
629b3d75 MJ |
6490 | l2_bb = BRANCH_EDGE (entry_bb)->dest; |
6491 | } | |
6492 | else | |
6493 | { | |
6494 | BRANCH_EDGE (entry_bb)->flags &= ~EDGE_ABNORMAL; | |
6495 | l1_bb = split_edge (BRANCH_EDGE (entry_bb)); | |
6496 | l2_bb = single_succ (l1_bb); | |
6497 | } | |
6498 | exit_bb = region->exit; | |
6499 | l2_dom_bb = NULL; | |
6500 | ||
65f4b875 | 6501 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
6502 | |
6503 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
6504 | /* Not needed in SSA form right now. */ | |
6505 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
c2ebf4f1 JJ |
6506 | if (fd->collapse > 1 |
6507 | && (gimple_omp_for_combined_into_p (fd->for_stmt) | |
6508 | || broken_loop)) | |
629b3d75 MJ |
6509 | { |
6510 | int first_zero_iter = -1, dummy = -1; | |
6511 | basic_block zero_iter_bb = l2_bb, dummy_bb = NULL; | |
6512 | ||
6513 | counts = XALLOCAVEC (tree, fd->collapse); | |
6514 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
6515 | zero_iter_bb, first_zero_iter, | |
6516 | dummy_bb, dummy, l2_dom_bb); | |
6517 | } | |
6518 | if (l2_dom_bb == NULL) | |
6519 | l2_dom_bb = l1_bb; | |
6520 | ||
6521 | n1 = fd->loop.n1; | |
6522 | n2 = fd->loop.n2; | |
6523 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6524 | { | |
6525 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
6526 | OMP_CLAUSE__LOOPTEMP_); | |
6527 | gcc_assert (innerc); | |
6528 | n1 = OMP_CLAUSE_DECL (innerc); | |
6529 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
6530 | OMP_CLAUSE__LOOPTEMP_); | |
6531 | gcc_assert (innerc); | |
6532 | n2 = OMP_CLAUSE_DECL (innerc); | |
6533 | } | |
6534 | tree step = fd->loop.step; | |
fc14ff61 | 6535 | tree orig_step = step; /* May be different from step if is_simt. */ |
629b3d75 | 6536 | |
4cea8675 AM |
6537 | bool is_simt = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
6538 | OMP_CLAUSE__SIMT_); | |
629b3d75 MJ |
6539 | if (is_simt) |
6540 | { | |
6541 | cfun->curr_properties &= ~PROP_gimple_lomp_dev; | |
4cea8675 AM |
6542 | is_simt = safelen_int > 1; |
6543 | } | |
6544 | tree simt_lane = NULL_TREE, simt_maxlane = NULL_TREE; | |
6545 | if (is_simt) | |
6546 | { | |
629b3d75 MJ |
6547 | simt_lane = create_tmp_var (unsigned_type_node); |
6548 | gimple *g = gimple_build_call_internal (IFN_GOMP_SIMT_LANE, 0); | |
6549 | gimple_call_set_lhs (g, simt_lane); | |
6550 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
6551 | tree offset = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, | |
6552 | fold_convert (TREE_TYPE (step), simt_lane)); | |
6553 | n1 = fold_convert (type, n1); | |
6554 | if (POINTER_TYPE_P (type)) | |
6555 | n1 = fold_build_pointer_plus (n1, offset); | |
6556 | else | |
6557 | n1 = fold_build2 (PLUS_EXPR, type, n1, fold_convert (type, offset)); | |
6558 | ||
6559 | /* Collapsed loops not handled for SIMT yet: limit to one lane only. */ | |
6560 | if (fd->collapse > 1) | |
6561 | simt_maxlane = build_one_cst (unsigned_type_node); | |
6562 | else if (safelen_int < omp_max_simt_vf ()) | |
6563 | simt_maxlane = build_int_cst (unsigned_type_node, safelen_int); | |
6564 | tree vf | |
6565 | = build_call_expr_internal_loc (UNKNOWN_LOCATION, IFN_GOMP_SIMT_VF, | |
6566 | unsigned_type_node, 0); | |
6567 | if (simt_maxlane) | |
6568 | vf = fold_build2 (MIN_EXPR, unsigned_type_node, vf, simt_maxlane); | |
6569 | vf = fold_convert (TREE_TYPE (step), vf); | |
6570 | step = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, vf); | |
6571 | } | |
6572 | ||
c2ebf4f1 JJ |
6573 | tree n2var = NULL_TREE; |
6574 | tree n2v = NULL_TREE; | |
6575 | tree *nonrect_bounds = NULL; | |
83f565ed | 6576 | tree min_arg1 = NULL_TREE, min_arg2 = NULL_TREE; |
629b3d75 MJ |
6577 | if (fd->collapse > 1) |
6578 | { | |
c2ebf4f1 | 6579 | if (broken_loop || gimple_omp_for_combined_into_p (fd->for_stmt)) |
629b3d75 | 6580 | { |
c2ebf4f1 JJ |
6581 | if (fd->non_rect) |
6582 | { | |
6583 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); | |
6584 | memset (nonrect_bounds, 0, | |
6585 | sizeof (tree) * (fd->last_nonrect + 1)); | |
6586 | } | |
6587 | expand_omp_build_assign (&gsi, fd->loop.v, fold_convert (type, n1)); | |
6588 | gcc_assert (entry_bb == gsi_bb (gsi)); | |
6589 | gcc_assert (fd->for_stmt == gsi_stmt (gsi)); | |
629b3d75 | 6590 | gsi_prev (&gsi); |
c2ebf4f1 JJ |
6591 | entry_bb = split_block (entry_bb, gsi_stmt (gsi))->dest; |
6592 | expand_omp_for_init_vars (fd, &gsi, counts, nonrect_bounds, | |
6593 | NULL, n1); | |
6594 | gsi = gsi_for_stmt (fd->for_stmt); | |
6595 | } | |
6596 | if (broken_loop) | |
6597 | ; | |
6598 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6599 | { | |
6600 | /* Compute in n2var the limit for the first innermost loop, | |
6601 | i.e. fd->loop.v + MIN (n2 - fd->loop.v, cnt) | |
6602 | where cnt is how many iterations would the loop have if | |
6603 | all further iterations were assigned to the current task. */ | |
6604 | n2var = create_tmp_var (type); | |
6605 | i = fd->collapse - 1; | |
6606 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6607 | if (POINTER_TYPE_P (itype)) | |
6608 | itype = signed_type_for (itype); | |
6609 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
6610 | ? -1 : 1)); | |
6611 | t = fold_build2 (PLUS_EXPR, itype, | |
6612 | fold_convert (itype, fd->loops[i].step), t); | |
6613 | t = fold_build2 (PLUS_EXPR, itype, t, | |
6614 | fold_convert (itype, fd->loops[i].n2)); | |
6615 | if (fd->loops[i].m2) | |
6616 | { | |
6617 | tree t2 = fold_convert (itype, | |
6618 | fd->loops[i - fd->loops[i].outer].v); | |
6619 | tree t3 = fold_convert (itype, fd->loops[i].m2); | |
6620 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
6621 | t = fold_build2 (PLUS_EXPR, itype, t, t2); | |
6622 | } | |
6623 | t = fold_build2 (MINUS_EXPR, itype, t, | |
6624 | fold_convert (itype, fd->loops[i].v)); | |
6625 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
6626 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6627 | fold_build1 (NEGATE_EXPR, itype, t), | |
6628 | fold_build1 (NEGATE_EXPR, itype, | |
6629 | fold_convert (itype, | |
6630 | fd->loops[i].step))); | |
6631 | else | |
6632 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
6633 | fold_convert (itype, fd->loops[i].step)); | |
6634 | t = fold_convert (type, t); | |
6635 | tree t2 = fold_build2 (MINUS_EXPR, type, n2, n1); | |
83f565ed JJ |
6636 | min_arg1 = create_tmp_var (type); |
6637 | expand_omp_build_assign (&gsi, min_arg1, t2); | |
6638 | min_arg2 = create_tmp_var (type); | |
6639 | expand_omp_build_assign (&gsi, min_arg2, t); | |
629b3d75 MJ |
6640 | } |
6641 | else | |
c2ebf4f1 JJ |
6642 | { |
6643 | if (TREE_CODE (n2) == INTEGER_CST) | |
6644 | { | |
6645 | /* Indicate for lastprivate handling that at least one iteration | |
6646 | has been performed, without wasting runtime. */ | |
6647 | if (integer_nonzerop (n2)) | |
6648 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6649 | fold_convert (type, n2)); | |
6650 | else | |
6651 | /* Indicate that no iteration has been performed. */ | |
6652 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6653 | build_one_cst (type)); | |
6654 | } | |
6655 | else | |
6656 | { | |
6657 | expand_omp_build_assign (&gsi, fd->loop.v, | |
6658 | build_zero_cst (type)); | |
6659 | expand_omp_build_assign (&gsi, n2, build_one_cst (type)); | |
6660 | } | |
6661 | for (i = 0; i < fd->collapse; i++) | |
6662 | { | |
6663 | t = fold_convert (TREE_TYPE (fd->loops[i].v), fd->loops[i].n1); | |
6664 | if (fd->loops[i].m1) | |
6665 | { | |
6666 | tree t2 | |
6667 | = fold_convert (TREE_TYPE (t), | |
6668 | fd->loops[i - fd->loops[i].outer].v); | |
6669 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i].m1); | |
6670 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
6671 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); | |
6672 | } | |
6673 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6674 | /* For normal non-combined collapsed loops just initialize | |
6675 | the outermost iterator in the entry_bb. */ | |
6676 | if (!broken_loop) | |
6677 | break; | |
6678 | } | |
6679 | } | |
629b3d75 | 6680 | } |
c2ebf4f1 JJ |
6681 | else |
6682 | expand_omp_build_assign (&gsi, fd->loop.v, fold_convert (type, n1)); | |
a29bd4f5 JJ |
6683 | tree altv = NULL_TREE, altn2 = NULL_TREE; |
6684 | if (fd->collapse == 1 | |
6685 | && !broken_loop | |
fc14ff61 | 6686 | && TREE_CODE (orig_step) != INTEGER_CST) |
a29bd4f5 JJ |
6687 | { |
6688 | /* The vectorizer currently punts on loops with non-constant steps | |
6689 | for the main IV (can't compute number of iterations and gives up | |
6690 | because of that). As for OpenMP loops it is always possible to | |
6691 | compute the number of iterations upfront, use an alternate IV | |
6692 | as the loop iterator: | |
6693 | altn2 = n1 < n2 ? (n2 - n1 + step - 1) / step : 0; | |
6694 | for (i = n1, altv = 0; altv < altn2; altv++, i += step) */ | |
6695 | altv = create_tmp_var (unsigned_type_for (TREE_TYPE (fd->loops[0].v))); | |
6696 | expand_omp_build_assign (&gsi, altv, build_zero_cst (TREE_TYPE (altv))); | |
6697 | tree itype = TREE_TYPE (fd->loop.v); | |
6698 | if (POINTER_TYPE_P (itype)) | |
6699 | itype = signed_type_for (itype); | |
6700 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
6701 | t = fold_build2 (PLUS_EXPR, itype, | |
fc14ff61 | 6702 | fold_convert (itype, step), t); |
a29bd4f5 JJ |
6703 | t = fold_build2 (PLUS_EXPR, itype, t, fold_convert (itype, n2)); |
6704 | t = fold_build2 (MINUS_EXPR, itype, t, | |
6705 | fold_convert (itype, fd->loop.v)); | |
6706 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
6707 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
6708 | fold_build1 (NEGATE_EXPR, itype, t), | |
6709 | fold_build1 (NEGATE_EXPR, itype, | |
fc14ff61 | 6710 | fold_convert (itype, step))); |
a29bd4f5 JJ |
6711 | else |
6712 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
fc14ff61 | 6713 | fold_convert (itype, step)); |
a29bd4f5 JJ |
6714 | t = fold_convert (TREE_TYPE (altv), t); |
6715 | altn2 = create_tmp_var (TREE_TYPE (altv)); | |
6716 | expand_omp_build_assign (&gsi, altn2, t); | |
6717 | tree t2 = fold_convert (TREE_TYPE (fd->loop.v), n2); | |
68e00633 | 6718 | t2 = fold_build2 (fd->loop.cond_code, boolean_type_node, fd->loop.v, t2); |
a29bd4f5 JJ |
6719 | t2 = force_gimple_operand_gsi (&gsi, t2, true, NULL_TREE, |
6720 | true, GSI_SAME_STMT); | |
a29bd4f5 JJ |
6721 | gassign *g = gimple_build_assign (altn2, COND_EXPR, t2, altn2, |
6722 | build_zero_cst (TREE_TYPE (altv))); | |
6723 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
6724 | } | |
6725 | else if (fd->collapse > 1 | |
6726 | && !broken_loop | |
6727 | && !gimple_omp_for_combined_into_p (fd->for_stmt) | |
6728 | && TREE_CODE (fd->loops[fd->collapse - 1].step) != INTEGER_CST) | |
6729 | { | |
6730 | altv = create_tmp_var (unsigned_type_for (TREE_TYPE (fd->loops[0].v))); | |
6731 | altn2 = create_tmp_var (TREE_TYPE (altv)); | |
6732 | } | |
e7393c89 JJ |
6733 | if (cond_var) |
6734 | { | |
6735 | if (POINTER_TYPE_P (type) | |
6736 | || TREE_CODE (n1) != INTEGER_CST | |
6737 | || fd->loop.cond_code != LT_EXPR | |
6738 | || tree_int_cst_sgn (n1) != 1) | |
6739 | expand_omp_build_assign (&gsi, cond_var, | |
6740 | build_one_cst (TREE_TYPE (cond_var))); | |
6741 | else | |
6742 | expand_omp_build_assign (&gsi, cond_var, | |
6743 | fold_convert (TREE_TYPE (cond_var), n1)); | |
6744 | } | |
629b3d75 MJ |
6745 | |
6746 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
6747 | gsi_remove (&gsi, true); | |
6748 | ||
6749 | if (!broken_loop) | |
6750 | { | |
6751 | /* Code to control the increment goes in the CONT_BB. */ | |
65f4b875 | 6752 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
6753 | stmt = gsi_stmt (gsi); |
6754 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
6755 | ||
c2ebf4f1 JJ |
6756 | if (fd->collapse == 1 |
6757 | || gimple_omp_for_combined_into_p (fd->for_stmt)) | |
6758 | { | |
6759 | if (POINTER_TYPE_P (type)) | |
6760 | t = fold_build_pointer_plus (fd->loop.v, step); | |
6761 | else | |
6762 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); | |
6763 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
6764 | } | |
6765 | else if (TREE_CODE (n2) != INTEGER_CST) | |
6766 | expand_omp_build_assign (&gsi, fd->loop.v, build_one_cst (type)); | |
a29bd4f5 JJ |
6767 | if (altv) |
6768 | { | |
6769 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (altv), altv, | |
6770 | build_one_cst (TREE_TYPE (altv))); | |
6771 | expand_omp_build_assign (&gsi, altv, t); | |
6772 | } | |
629b3d75 MJ |
6773 | |
6774 | if (fd->collapse > 1) | |
6775 | { | |
6776 | i = fd->collapse - 1; | |
6777 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
4796d16d | 6778 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); |
629b3d75 MJ |
6779 | else |
6780 | { | |
6781 | t = fold_convert (TREE_TYPE (fd->loops[i].v), | |
6782 | fd->loops[i].step); | |
6783 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
6784 | fd->loops[i].v, t); | |
6785 | } | |
6786 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
629b3d75 | 6787 | } |
e7393c89 JJ |
6788 | if (cond_var) |
6789 | { | |
6790 | if (POINTER_TYPE_P (type) | |
6791 | || TREE_CODE (n1) != INTEGER_CST | |
6792 | || fd->loop.cond_code != LT_EXPR | |
6793 | || tree_int_cst_sgn (n1) != 1) | |
6794 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, | |
6795 | build_one_cst (TREE_TYPE (cond_var))); | |
6796 | else | |
6797 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, | |
6798 | fold_convert (TREE_TYPE (cond_var), step)); | |
6799 | expand_omp_build_assign (&gsi, cond_var, t); | |
6800 | } | |
629b3d75 MJ |
6801 | |
6802 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
6803 | gsi_remove (&gsi, true); | |
6804 | } | |
6805 | ||
6806 | /* Emit the condition in L1_BB. */ | |
6807 | gsi = gsi_start_bb (l1_bb); | |
6808 | ||
a29bd4f5 JJ |
6809 | if (altv) |
6810 | t = build2 (LT_EXPR, boolean_type_node, altv, altn2); | |
6811 | else if (fd->collapse > 1 | |
6812 | && !gimple_omp_for_combined_into_p (fd->for_stmt) | |
6813 | && !broken_loop) | |
c2ebf4f1 JJ |
6814 | { |
6815 | i = fd->collapse - 1; | |
6816 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6817 | if (fd->loops[i].m2) | |
6818 | t = n2v = create_tmp_var (itype); | |
6819 | else | |
97d32048 | 6820 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
c2ebf4f1 JJ |
6821 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6822 | false, GSI_CONTINUE_LINKING); | |
6823 | tree v = fd->loops[i].v; | |
6824 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
6825 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
6826 | false, GSI_CONTINUE_LINKING); | |
6827 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
6828 | } | |
6829 | else | |
6830 | { | |
6831 | if (fd->collapse > 1 && !broken_loop) | |
6832 | t = n2var; | |
6833 | else | |
97d32048 | 6834 | t = fold_convert (type, unshare_expr (n2)); |
c2ebf4f1 JJ |
6835 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6836 | false, GSI_CONTINUE_LINKING); | |
6837 | tree v = fd->loop.v; | |
6838 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
6839 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
6840 | false, GSI_CONTINUE_LINKING); | |
6841 | t = build2 (fd->loop.cond_code, boolean_type_node, v, t); | |
6842 | } | |
629b3d75 MJ |
6843 | cond_stmt = gimple_build_cond_empty (t); |
6844 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
6845 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
6846 | NULL, NULL) | |
6847 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
6848 | NULL, NULL)) | |
6849 | { | |
6850 | gsi = gsi_for_stmt (cond_stmt); | |
6851 | gimple_regimplify_operands (cond_stmt, &gsi); | |
6852 | } | |
6853 | ||
6854 | /* Add 'V -= STEP * (SIMT_VF - 1)' after the loop. */ | |
6855 | if (is_simt) | |
6856 | { | |
6857 | gsi = gsi_start_bb (l2_bb); | |
fc14ff61 | 6858 | step = fold_build2 (MINUS_EXPR, TREE_TYPE (step), orig_step, step); |
629b3d75 MJ |
6859 | if (POINTER_TYPE_P (type)) |
6860 | t = fold_build_pointer_plus (fd->loop.v, step); | |
6861 | else | |
6862 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); | |
6863 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
6864 | } | |
6865 | ||
6866 | /* Remove GIMPLE_OMP_RETURN. */ | |
65f4b875 | 6867 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
6868 | gsi_remove (&gsi, true); |
6869 | ||
6870 | /* Connect the new blocks. */ | |
6871 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
6872 | ||
6873 | if (!broken_loop) | |
6874 | { | |
6875 | remove_edge (BRANCH_EDGE (entry_bb)); | |
6876 | make_edge (entry_bb, l1_bb, EDGE_FALLTHRU); | |
6877 | ||
6878 | e = BRANCH_EDGE (l1_bb); | |
6879 | ne = FALLTHRU_EDGE (l1_bb); | |
6880 | e->flags = EDGE_TRUE_VALUE; | |
6881 | } | |
6882 | else | |
6883 | { | |
6884 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
6885 | ||
6886 | ne = single_succ_edge (l1_bb); | |
6887 | e = make_edge (l1_bb, l0_bb, EDGE_TRUE_VALUE); | |
6888 | ||
6889 | } | |
6890 | ne->flags = EDGE_FALSE_VALUE; | |
357067f2 JH |
6891 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); |
6892 | ne->probability = e->probability.invert (); | |
629b3d75 MJ |
6893 | |
6894 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, entry_bb); | |
6895 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, l1_bb); | |
6896 | ||
6897 | if (simt_maxlane) | |
6898 | { | |
6899 | cond_stmt = gimple_build_cond (LT_EXPR, simt_lane, simt_maxlane, | |
6900 | NULL_TREE, NULL_TREE); | |
6901 | gsi = gsi_last_bb (entry_bb); | |
6902 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
6903 | make_edge (entry_bb, l2_bb, EDGE_FALSE_VALUE); | |
6904 | FALLTHRU_EDGE (entry_bb)->flags = EDGE_TRUE_VALUE; | |
357067f2 JH |
6905 | FALLTHRU_EDGE (entry_bb)->probability |
6906 | = profile_probability::guessed_always ().apply_scale (7, 8); | |
c2ebf4f1 | 6907 | BRANCH_EDGE (entry_bb)->probability |
357067f2 | 6908 | = FALLTHRU_EDGE (entry_bb)->probability.invert (); |
629b3d75 MJ |
6909 | l2_dom_bb = entry_bb; |
6910 | } | |
6911 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, l2_dom_bb); | |
6912 | ||
c2ebf4f1 JJ |
6913 | if (!broken_loop && fd->collapse > 1) |
6914 | { | |
6915 | basic_block last_bb = l1_bb; | |
6916 | basic_block init_bb = NULL; | |
6917 | for (i = fd->collapse - 2; i >= 0; i--) | |
6918 | { | |
6919 | tree nextn2v = NULL_TREE; | |
6920 | if (EDGE_SUCC (last_bb, 0)->flags & EDGE_FALSE_VALUE) | |
6921 | e = EDGE_SUCC (last_bb, 0); | |
6922 | else | |
6923 | e = EDGE_SUCC (last_bb, 1); | |
6924 | basic_block bb = split_edge (e); | |
6925 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
4796d16d | 6926 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); |
c2ebf4f1 JJ |
6927 | else |
6928 | { | |
6929 | t = fold_convert (TREE_TYPE (fd->loops[i].v), | |
6930 | fd->loops[i].step); | |
6931 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
6932 | fd->loops[i].v, t); | |
6933 | } | |
6934 | gsi = gsi_after_labels (bb); | |
6935 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
6936 | ||
fe8ac82f | 6937 | bb = split_block (bb, last_nondebug_stmt (bb))->dest; |
c2ebf4f1 JJ |
6938 | gsi = gsi_start_bb (bb); |
6939 | tree itype = TREE_TYPE (fd->loops[i].v); | |
6940 | if (fd->loops[i].m2) | |
6941 | t = nextn2v = create_tmp_var (itype); | |
6942 | else | |
97d32048 | 6943 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
c2ebf4f1 JJ |
6944 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6945 | false, GSI_CONTINUE_LINKING); | |
6946 | tree v = fd->loops[i].v; | |
6947 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
6948 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
6949 | false, GSI_CONTINUE_LINKING); | |
6950 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
6951 | cond_stmt = gimple_build_cond_empty (t); | |
6952 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
6953 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
6954 | expand_omp_regimplify_p, NULL, NULL) | |
6955 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
6956 | expand_omp_regimplify_p, NULL, NULL)) | |
6957 | { | |
6958 | gsi = gsi_for_stmt (cond_stmt); | |
6959 | gimple_regimplify_operands (cond_stmt, &gsi); | |
6960 | } | |
6961 | ne = single_succ_edge (bb); | |
6962 | ne->flags = EDGE_FALSE_VALUE; | |
6963 | ||
6964 | init_bb = create_empty_bb (bb); | |
6965 | set_immediate_dominator (CDI_DOMINATORS, init_bb, bb); | |
6966 | add_bb_to_loop (init_bb, bb->loop_father); | |
6967 | e = make_edge (bb, init_bb, EDGE_TRUE_VALUE); | |
6968 | e->probability | |
6969 | = profile_probability::guessed_always ().apply_scale (7, 8); | |
6970 | ne->probability = e->probability.invert (); | |
6971 | ||
6972 | gsi = gsi_after_labels (init_bb); | |
c2ebf4f1 JJ |
6973 | if (fd->loops[i + 1].m1) |
6974 | { | |
97d32048 | 6975 | tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
c2ebf4f1 JJ |
6976 | fd->loops[i + 1 |
6977 | - fd->loops[i + 1].outer].v); | |
97d32048 JJ |
6978 | if (POINTER_TYPE_P (TREE_TYPE (t2))) |
6979 | t = fold_build_pointer_plus (t2, fd->loops[i + 1].n1); | |
6980 | else | |
6981 | { | |
6982 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), | |
6983 | fd->loops[i + 1].n1); | |
6984 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m1); | |
6985 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
6986 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); | |
6987 | } | |
c2ebf4f1 | 6988 | } |
97d32048 JJ |
6989 | else |
6990 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), | |
6991 | fd->loops[i + 1].n1); | |
c2ebf4f1 JJ |
6992 | expand_omp_build_assign (&gsi, fd->loops[i + 1].v, t); |
6993 | if (fd->loops[i + 1].m2) | |
6994 | { | |
a29bd4f5 | 6995 | if (i + 2 == fd->collapse && (n2var || altv)) |
c2ebf4f1 JJ |
6996 | { |
6997 | gcc_assert (n2v == NULL_TREE); | |
6998 | n2v = create_tmp_var (TREE_TYPE (fd->loops[i + 1].v)); | |
6999 | } | |
97d32048 | 7000 | tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
c2ebf4f1 JJ |
7001 | fd->loops[i + 1 |
7002 | - fd->loops[i + 1].outer].v); | |
97d32048 JJ |
7003 | if (POINTER_TYPE_P (TREE_TYPE (t2))) |
7004 | t = fold_build_pointer_plus (t2, fd->loops[i + 1].n2); | |
7005 | else | |
7006 | { | |
7007 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), | |
7008 | fd->loops[i + 1].n2); | |
7009 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m2); | |
7010 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); | |
7011 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); | |
7012 | } | |
c2ebf4f1 JJ |
7013 | expand_omp_build_assign (&gsi, n2v, t); |
7014 | } | |
7015 | if (i + 2 == fd->collapse && n2var) | |
7016 | { | |
7017 | /* For composite simd, n2 is the first iteration the current | |
7018 | task shouldn't already handle, so we effectively want to use | |
7019 | for (V3 = N31; V < N2 && V3 < N32; V++, V3 += STEP3) | |
7020 | as the vectorized loop. Except the vectorizer will not | |
7021 | vectorize that, so instead compute N2VAR as | |
7022 | N2VAR = V + MIN (N2 - V, COUNTS3) and use | |
7023 | for (V3 = N31; V < N2VAR; V++, V3 += STEP3) | |
7024 | as the loop to vectorize. */ | |
7025 | tree t2 = fold_build2 (MINUS_EXPR, type, n2, fd->loop.v); | |
7026 | if (fd->loops[i + 1].m1 || fd->loops[i + 1].m2) | |
7027 | { | |
97d32048 JJ |
7028 | tree itype = TREE_TYPE (fd->loops[i].v); |
7029 | if (POINTER_TYPE_P (itype)) | |
7030 | itype = signed_type_for (itype); | |
c2ebf4f1 JJ |
7031 | t = build_int_cst (itype, (fd->loops[i + 1].cond_code |
7032 | == LT_EXPR ? -1 : 1)); | |
7033 | t = fold_build2 (PLUS_EXPR, itype, | |
7034 | fold_convert (itype, | |
7035 | fd->loops[i + 1].step), t); | |
97d32048 | 7036 | if (fd->loops[i + 1].m2 == NULL_TREE) |
c2ebf4f1 JJ |
7037 | t = fold_build2 (PLUS_EXPR, itype, t, |
7038 | fold_convert (itype, | |
7039 | fd->loops[i + 1].n2)); | |
97d32048 JJ |
7040 | else if (POINTER_TYPE_P (TREE_TYPE (n2v))) |
7041 | { | |
7042 | t = fold_build_pointer_plus (n2v, t); | |
7043 | t = fold_convert (itype, t); | |
7044 | } | |
7045 | else | |
7046 | t = fold_build2 (PLUS_EXPR, itype, t, n2v); | |
c2ebf4f1 JJ |
7047 | t = fold_build2 (MINUS_EXPR, itype, t, |
7048 | fold_convert (itype, fd->loops[i + 1].v)); | |
7049 | tree step = fold_convert (itype, fd->loops[i + 1].step); | |
7050 | if (TYPE_UNSIGNED (itype) | |
7051 | && fd->loops[i + 1].cond_code == GT_EXPR) | |
7052 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
7053 | fold_build1 (NEGATE_EXPR, itype, t), | |
7054 | fold_build1 (NEGATE_EXPR, itype, step)); | |
7055 | else | |
7056 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
7057 | t = fold_convert (type, t); | |
7058 | } | |
7059 | else | |
7060 | t = counts[i + 1]; | |
83f565ed JJ |
7061 | expand_omp_build_assign (&gsi, min_arg1, t2); |
7062 | expand_omp_build_assign (&gsi, min_arg2, t); | |
fe8ac82f | 7063 | e = split_block (init_bb, last_nondebug_stmt (init_bb)); |
83f565ed JJ |
7064 | gsi = gsi_after_labels (e->dest); |
7065 | init_bb = e->dest; | |
7066 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
7067 | make_edge (entry_bb, init_bb, EDGE_FALLTHRU); | |
7068 | set_immediate_dominator (CDI_DOMINATORS, init_bb, entry_bb); | |
7069 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, init_bb); | |
7070 | t = fold_build2 (MIN_EXPR, type, min_arg1, min_arg2); | |
c2ebf4f1 JJ |
7071 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, t); |
7072 | expand_omp_build_assign (&gsi, n2var, t); | |
7073 | } | |
a29bd4f5 JJ |
7074 | if (i + 2 == fd->collapse && altv) |
7075 | { | |
7076 | /* The vectorizer currently punts on loops with non-constant | |
7077 | steps for the main IV (can't compute number of iterations | |
7078 | and gives up because of that). As for OpenMP loops it is | |
7079 | always possible to compute the number of iterations upfront, | |
7080 | use an alternate IV as the loop iterator. */ | |
7081 | expand_omp_build_assign (&gsi, altv, | |
7082 | build_zero_cst (TREE_TYPE (altv))); | |
7083 | tree itype = TREE_TYPE (fd->loops[i + 1].v); | |
7084 | if (POINTER_TYPE_P (itype)) | |
7085 | itype = signed_type_for (itype); | |
7086 | t = build_int_cst (itype, (fd->loops[i + 1].cond_code == LT_EXPR | |
7087 | ? -1 : 1)); | |
7088 | t = fold_build2 (PLUS_EXPR, itype, | |
7089 | fold_convert (itype, fd->loops[i + 1].step), t); | |
7090 | t = fold_build2 (PLUS_EXPR, itype, t, | |
7091 | fold_convert (itype, | |
7092 | fd->loops[i + 1].m2 | |
7093 | ? n2v : fd->loops[i + 1].n2)); | |
7094 | t = fold_build2 (MINUS_EXPR, itype, t, | |
7095 | fold_convert (itype, fd->loops[i + 1].v)); | |
7096 | tree step = fold_convert (itype, fd->loops[i + 1].step); | |
7097 | if (TYPE_UNSIGNED (itype) | |
7098 | && fd->loops[i + 1].cond_code == GT_EXPR) | |
7099 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
7100 | fold_build1 (NEGATE_EXPR, itype, t), | |
7101 | fold_build1 (NEGATE_EXPR, itype, step)); | |
7102 | else | |
7103 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
7104 | t = fold_convert (TREE_TYPE (altv), t); | |
7105 | expand_omp_build_assign (&gsi, altn2, t); | |
7106 | tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v), | |
7107 | fd->loops[i + 1].m2 | |
7108 | ? n2v : fd->loops[i + 1].n2); | |
a29bd4f5 JJ |
7109 | t2 = fold_build2 (fd->loops[i + 1].cond_code, boolean_type_node, |
7110 | fd->loops[i + 1].v, t2); | |
68e00633 RB |
7111 | t2 = force_gimple_operand_gsi (&gsi, t2, true, NULL_TREE, |
7112 | true, GSI_SAME_STMT); | |
a29bd4f5 JJ |
7113 | gassign *g |
7114 | = gimple_build_assign (altn2, COND_EXPR, t2, altn2, | |
7115 | build_zero_cst (TREE_TYPE (altv))); | |
7116 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
7117 | } | |
c2ebf4f1 JJ |
7118 | n2v = nextn2v; |
7119 | ||
7120 | make_edge (init_bb, last_bb, EDGE_FALLTHRU); | |
7121 | if (!gimple_omp_for_combined_into_p (fd->for_stmt)) | |
7122 | { | |
7123 | e = find_edge (entry_bb, last_bb); | |
7124 | redirect_edge_succ (e, bb); | |
7125 | set_immediate_dominator (CDI_DOMINATORS, bb, entry_bb); | |
7126 | set_immediate_dominator (CDI_DOMINATORS, last_bb, init_bb); | |
7127 | } | |
7128 | ||
7129 | last_bb = bb; | |
7130 | } | |
7131 | } | |
629b3d75 MJ |
7132 | if (!broken_loop) |
7133 | { | |
99b1c316 | 7134 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
7135 | loop->header = l1_bb; |
7136 | loop->latch = cont_bb; | |
7137 | add_loop (loop, l1_bb->loop_father); | |
7138 | loop->safelen = safelen_int; | |
7139 | if (simduid) | |
7140 | { | |
7141 | loop->simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); | |
7142 | cfun->has_simduid_loops = true; | |
7143 | } | |
7144 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize | |
7145 | the loop. */ | |
7146 | if ((flag_tree_loop_vectorize | |
00f34291 | 7147 | || !OPTION_SET_P (flag_tree_loop_vectorize)) |
629b3d75 MJ |
7148 | && flag_tree_loop_optimize |
7149 | && loop->safelen > 1) | |
7150 | { | |
7151 | loop->force_vectorize = true; | |
f63445e5 JJ |
7152 | if (simdlen && tree_fits_uhwi_p (OMP_CLAUSE_SIMDLEN_EXPR (simdlen))) |
7153 | { | |
7154 | unsigned HOST_WIDE_INT v | |
7155 | = tree_to_uhwi (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)); | |
7156 | if (v < INT_MAX && v <= (unsigned HOST_WIDE_INT) loop->safelen) | |
7157 | loop->simdlen = v; | |
7158 | } | |
629b3d75 MJ |
7159 | cfun->has_force_vectorize_loops = true; |
7160 | } | |
fed2a43c JJ |
7161 | else if (dont_vectorize) |
7162 | loop->dont_vectorize = true; | |
629b3d75 MJ |
7163 | } |
7164 | else if (simduid) | |
7165 | cfun->has_simduid_loops = true; | |
7166 | } | |
7167 | ||
7168 | /* Taskloop construct is represented after gimplification with | |
7169 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched | |
7170 | in between them. This routine expands the outer GIMPLE_OMP_FOR, | |
7171 | which should just compute all the needed loop temporaries | |
7172 | for GIMPLE_OMP_TASK. */ | |
7173 | ||
7174 | static void | |
7175 | expand_omp_taskloop_for_outer (struct omp_region *region, | |
7176 | struct omp_for_data *fd, | |
7177 | gimple *inner_stmt) | |
7178 | { | |
7179 | tree type, bias = NULL_TREE; | |
7180 | basic_block entry_bb, cont_bb, exit_bb; | |
7181 | gimple_stmt_iterator gsi; | |
7182 | gassign *assign_stmt; | |
7183 | tree *counts = NULL; | |
7184 | int i; | |
7185 | ||
7186 | gcc_assert (inner_stmt); | |
7187 | gcc_assert (region->cont); | |
7188 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_TASK | |
7189 | && gimple_omp_task_taskloop_p (inner_stmt)); | |
7190 | type = TREE_TYPE (fd->loop.v); | |
7191 | ||
7192 | /* See if we need to bias by LLONG_MIN. */ | |
7193 | if (fd->iter_type == long_long_unsigned_type_node | |
c8f10456 | 7194 | && (TREE_CODE (type) == INTEGER_TYPE || TREE_CODE (type) == BITINT_TYPE) |
629b3d75 MJ |
7195 | && !TYPE_UNSIGNED (type)) |
7196 | { | |
7197 | tree n1, n2; | |
7198 | ||
7199 | if (fd->loop.cond_code == LT_EXPR) | |
7200 | { | |
7201 | n1 = fd->loop.n1; | |
7202 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7203 | } | |
7204 | else | |
7205 | { | |
7206 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7207 | n2 = fd->loop.n1; | |
7208 | } | |
7209 | if (TREE_CODE (n1) != INTEGER_CST | |
7210 | || TREE_CODE (n2) != INTEGER_CST | |
7211 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
7212 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
7213 | } | |
7214 | ||
7215 | entry_bb = region->entry; | |
7216 | cont_bb = region->cont; | |
7217 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
7218 | gcc_assert (BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
7219 | exit_bb = region->exit; | |
7220 | ||
65f4b875 | 7221 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
7222 | gimple *for_stmt = gsi_stmt (gsi); |
7223 | gcc_assert (gimple_code (for_stmt) == GIMPLE_OMP_FOR); | |
7224 | if (fd->collapse > 1) | |
7225 | { | |
7226 | int first_zero_iter = -1, dummy = -1; | |
7227 | basic_block zero_iter_bb = NULL, dummy_bb = NULL, l2_dom_bb = NULL; | |
7228 | ||
7229 | counts = XALLOCAVEC (tree, fd->collapse); | |
7230 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
7231 | zero_iter_bb, first_zero_iter, | |
7232 | dummy_bb, dummy, l2_dom_bb); | |
7233 | ||
7234 | if (zero_iter_bb) | |
7235 | { | |
7236 | /* Some counts[i] vars might be uninitialized if | |
7237 | some loop has zero iterations. But the body shouldn't | |
7238 | be executed in that case, so just avoid uninit warnings. */ | |
7239 | for (i = first_zero_iter; i < fd->collapse; i++) | |
7240 | if (SSA_VAR_P (counts[i])) | |
e9e2bad7 | 7241 | suppress_warning (counts[i], OPT_Wuninitialized); |
629b3d75 MJ |
7242 | gsi_prev (&gsi); |
7243 | edge e = split_block (entry_bb, gsi_stmt (gsi)); | |
7244 | entry_bb = e->dest; | |
7245 | make_edge (zero_iter_bb, entry_bb, EDGE_FALLTHRU); | |
7246 | gsi = gsi_last_bb (entry_bb); | |
7247 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, | |
7248 | get_immediate_dominator (CDI_DOMINATORS, | |
7249 | zero_iter_bb)); | |
7250 | } | |
7251 | } | |
7252 | ||
7253 | tree t0, t1; | |
7254 | t1 = fd->loop.n2; | |
7255 | t0 = fd->loop.n1; | |
7256 | if (POINTER_TYPE_P (TREE_TYPE (t0)) | |
7257 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
7258 | != TYPE_PRECISION (fd->iter_type)) | |
7259 | { | |
7260 | /* Avoid casting pointers to integer of a different size. */ | |
7261 | tree itype = signed_type_for (type); | |
7262 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); | |
7263 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
7264 | } | |
7265 | else | |
7266 | { | |
7267 | t1 = fold_convert (fd->iter_type, t1); | |
7268 | t0 = fold_convert (fd->iter_type, t0); | |
7269 | } | |
7270 | if (bias) | |
7271 | { | |
7272 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); | |
7273 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
7274 | } | |
7275 | ||
7276 | tree innerc = omp_find_clause (gimple_omp_task_clauses (inner_stmt), | |
7277 | OMP_CLAUSE__LOOPTEMP_); | |
7278 | gcc_assert (innerc); | |
7279 | tree startvar = OMP_CLAUSE_DECL (innerc); | |
7280 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), OMP_CLAUSE__LOOPTEMP_); | |
7281 | gcc_assert (innerc); | |
7282 | tree endvar = OMP_CLAUSE_DECL (innerc); | |
7283 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
7284 | { | |
14707c89 | 7285 | innerc = find_lastprivate_looptemp (fd, innerc); |
629b3d75 MJ |
7286 | if (innerc) |
7287 | { | |
7288 | /* If needed (inner taskloop has lastprivate clause), propagate | |
7289 | down the total number of iterations. */ | |
7290 | tree t = force_gimple_operand_gsi (&gsi, fd->loop.n2, false, | |
7291 | NULL_TREE, false, | |
7292 | GSI_CONTINUE_LINKING); | |
7293 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
7294 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7295 | } | |
7296 | } | |
7297 | ||
7298 | t0 = force_gimple_operand_gsi (&gsi, t0, false, NULL_TREE, false, | |
7299 | GSI_CONTINUE_LINKING); | |
7300 | assign_stmt = gimple_build_assign (startvar, t0); | |
7301 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7302 | ||
7303 | t1 = force_gimple_operand_gsi (&gsi, t1, false, NULL_TREE, false, | |
7304 | GSI_CONTINUE_LINKING); | |
7305 | assign_stmt = gimple_build_assign (endvar, t1); | |
7306 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7307 | if (fd->collapse > 1) | |
aed3ab25 | 7308 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, inner_stmt, startvar); |
629b3d75 MJ |
7309 | |
7310 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
7311 | gsi = gsi_for_stmt (for_stmt); | |
7312 | gsi_remove (&gsi, true); | |
7313 | ||
65f4b875 | 7314 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
7315 | gsi_remove (&gsi, true); |
7316 | ||
65f4b875 | 7317 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
7318 | gsi_remove (&gsi, true); |
7319 | ||
357067f2 | 7320 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); |
629b3d75 | 7321 | remove_edge (BRANCH_EDGE (entry_bb)); |
357067f2 | 7322 | FALLTHRU_EDGE (cont_bb)->probability = profile_probability::always (); |
629b3d75 MJ |
7323 | remove_edge (BRANCH_EDGE (cont_bb)); |
7324 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, cont_bb); | |
7325 | set_immediate_dominator (CDI_DOMINATORS, region->entry, | |
7326 | recompute_dominator (CDI_DOMINATORS, region->entry)); | |
7327 | } | |
7328 | ||
7329 | /* Taskloop construct is represented after gimplification with | |
7330 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched | |
7331 | in between them. This routine expands the inner GIMPLE_OMP_FOR. | |
7332 | GOMP_taskloop{,_ull} function arranges for each task to be given just | |
7333 | a single range of iterations. */ | |
7334 | ||
7335 | static void | |
7336 | expand_omp_taskloop_for_inner (struct omp_region *region, | |
7337 | struct omp_for_data *fd, | |
7338 | gimple *inner_stmt) | |
7339 | { | |
7340 | tree e, t, type, itype, vmain, vback, bias = NULL_TREE; | |
7341 | basic_block entry_bb, exit_bb, body_bb, cont_bb, collapse_bb = NULL; | |
7342 | basic_block fin_bb; | |
7343 | gimple_stmt_iterator gsi; | |
7344 | edge ep; | |
7345 | bool broken_loop = region->cont == NULL; | |
7346 | tree *counts = NULL; | |
7347 | tree n1, n2, step; | |
7348 | ||
7349 | itype = type = TREE_TYPE (fd->loop.v); | |
7350 | if (POINTER_TYPE_P (type)) | |
7351 | itype = signed_type_for (type); | |
7352 | ||
7353 | /* See if we need to bias by LLONG_MIN. */ | |
7354 | if (fd->iter_type == long_long_unsigned_type_node | |
c8f10456 | 7355 | && (TREE_CODE (type) == INTEGER_TYPE || TREE_CODE (type) == BITINT_TYPE) |
629b3d75 MJ |
7356 | && !TYPE_UNSIGNED (type)) |
7357 | { | |
7358 | tree n1, n2; | |
7359 | ||
7360 | if (fd->loop.cond_code == LT_EXPR) | |
7361 | { | |
7362 | n1 = fd->loop.n1; | |
7363 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7364 | } | |
7365 | else | |
7366 | { | |
7367 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
7368 | n2 = fd->loop.n1; | |
7369 | } | |
7370 | if (TREE_CODE (n1) != INTEGER_CST | |
7371 | || TREE_CODE (n2) != INTEGER_CST | |
7372 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
7373 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
7374 | } | |
7375 | ||
7376 | entry_bb = region->entry; | |
7377 | cont_bb = region->cont; | |
7378 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
7379 | fin_bb = BRANCH_EDGE (entry_bb)->dest; | |
7380 | gcc_assert (broken_loop | |
7381 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
7382 | body_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
7383 | if (!broken_loop) | |
7384 | { | |
7385 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); | |
7386 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
7387 | } | |
7388 | exit_bb = region->exit; | |
7389 | ||
7390 | /* Iteration space partitioning goes in ENTRY_BB. */ | |
65f4b875 | 7391 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
7392 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
7393 | ||
7394 | if (fd->collapse > 1) | |
7395 | { | |
7396 | int first_zero_iter = -1, dummy = -1; | |
7397 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
7398 | ||
7399 | counts = XALLOCAVEC (tree, fd->collapse); | |
7400 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
7401 | fin_bb, first_zero_iter, | |
7402 | dummy_bb, dummy, l2_dom_bb); | |
7403 | t = NULL_TREE; | |
7404 | } | |
7405 | else | |
7406 | t = integer_one_node; | |
7407 | ||
7408 | step = fd->loop.step; | |
7409 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
7410 | OMP_CLAUSE__LOOPTEMP_); | |
7411 | gcc_assert (innerc); | |
7412 | n1 = OMP_CLAUSE_DECL (innerc); | |
7413 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), OMP_CLAUSE__LOOPTEMP_); | |
7414 | gcc_assert (innerc); | |
7415 | n2 = OMP_CLAUSE_DECL (innerc); | |
7416 | if (bias) | |
7417 | { | |
7418 | n1 = fold_build2 (PLUS_EXPR, fd->iter_type, n1, bias); | |
7419 | n2 = fold_build2 (PLUS_EXPR, fd->iter_type, n2, bias); | |
7420 | } | |
7421 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
7422 | true, NULL_TREE, true, GSI_SAME_STMT); | |
7423 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
7424 | true, NULL_TREE, true, GSI_SAME_STMT); | |
7425 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
7426 | true, NULL_TREE, true, GSI_SAME_STMT); | |
7427 | ||
7428 | tree startvar = fd->loop.v; | |
7429 | tree endvar = NULL_TREE; | |
7430 | ||
7431 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
7432 | { | |
7433 | tree clauses = gimple_omp_for_clauses (inner_stmt); | |
7434 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
7435 | gcc_assert (innerc); | |
7436 | startvar = OMP_CLAUSE_DECL (innerc); | |
7437 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
7438 | OMP_CLAUSE__LOOPTEMP_); | |
7439 | gcc_assert (innerc); | |
7440 | endvar = OMP_CLAUSE_DECL (innerc); | |
7441 | } | |
7442 | t = fold_convert (TREE_TYPE (startvar), n1); | |
7443 | t = force_gimple_operand_gsi (&gsi, t, | |
7444 | DECL_P (startvar) | |
7445 | && TREE_ADDRESSABLE (startvar), | |
7446 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
7447 | gimple *assign_stmt = gimple_build_assign (startvar, t); | |
7448 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7449 | ||
7450 | t = fold_convert (TREE_TYPE (startvar), n2); | |
7451 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
7452 | false, GSI_CONTINUE_LINKING); | |
7453 | if (endvar) | |
7454 | { | |
7455 | assign_stmt = gimple_build_assign (endvar, e); | |
7456 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7457 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
7458 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
7459 | else | |
7460 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
7461 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
7462 | } | |
2e47c8c6 JJ |
7463 | |
7464 | tree *nonrect_bounds = NULL; | |
629b3d75 | 7465 | if (fd->collapse > 1) |
2e47c8c6 JJ |
7466 | { |
7467 | if (fd->non_rect) | |
7468 | { | |
7469 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); | |
7470 | memset (nonrect_bounds, 0, sizeof (tree) * (fd->last_nonrect + 1)); | |
7471 | } | |
7472 | gcc_assert (gsi_bb (gsi) == entry_bb); | |
7473 | expand_omp_for_init_vars (fd, &gsi, counts, nonrect_bounds, inner_stmt, | |
7474 | startvar); | |
7475 | entry_bb = gsi_bb (gsi); | |
7476 | } | |
629b3d75 MJ |
7477 | |
7478 | if (!broken_loop) | |
7479 | { | |
7480 | /* The code controlling the sequential loop replaces the | |
7481 | GIMPLE_OMP_CONTINUE. */ | |
65f4b875 | 7482 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
7483 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
7484 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
7485 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
7486 | vback = gimple_omp_continue_control_def (cont_stmt); | |
7487 | ||
7488 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
7489 | { | |
7490 | if (POINTER_TYPE_P (type)) | |
7491 | t = fold_build_pointer_plus (vmain, step); | |
7492 | else | |
7493 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
7494 | t = force_gimple_operand_gsi (&gsi, t, | |
7495 | DECL_P (vback) | |
7496 | && TREE_ADDRESSABLE (vback), | |
7497 | NULL_TREE, true, GSI_SAME_STMT); | |
7498 | assign_stmt = gimple_build_assign (vback, t); | |
7499 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
7500 | ||
7501 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
7502 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
7503 | ? t : vback, e); | |
7504 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
7505 | } | |
7506 | ||
7507 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ | |
7508 | gsi_remove (&gsi, true); | |
7509 | ||
7510 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
2e47c8c6 JJ |
7511 | collapse_bb = extract_omp_for_update_vars (fd, nonrect_bounds, |
7512 | cont_bb, body_bb); | |
629b3d75 MJ |
7513 | } |
7514 | ||
7515 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
7516 | gsi = gsi_for_stmt (fd->for_stmt); | |
7517 | gsi_remove (&gsi, true); | |
7518 | ||
7519 | /* Remove the GIMPLE_OMP_RETURN statement. */ | |
65f4b875 | 7520 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
7521 | gsi_remove (&gsi, true); |
7522 | ||
357067f2 | 7523 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); |
629b3d75 MJ |
7524 | if (!broken_loop) |
7525 | remove_edge (BRANCH_EDGE (entry_bb)); | |
7526 | else | |
7527 | { | |
7528 | remove_edge_and_dominated_blocks (BRANCH_EDGE (entry_bb)); | |
7529 | region->outer->cont = NULL; | |
7530 | } | |
7531 | ||
7532 | /* Connect all the blocks. */ | |
7533 | if (!broken_loop) | |
7534 | { | |
7535 | ep = find_edge (cont_bb, body_bb); | |
7536 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
7537 | { | |
7538 | remove_edge (ep); | |
7539 | ep = NULL; | |
7540 | } | |
7541 | else if (fd->collapse > 1) | |
7542 | { | |
7543 | remove_edge (ep); | |
7544 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
7545 | } | |
7546 | else | |
7547 | ep->flags = EDGE_TRUE_VALUE; | |
7548 | find_edge (cont_bb, fin_bb)->flags | |
7549 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
7550 | } | |
7551 | ||
7552 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
7553 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
7554 | if (!broken_loop) | |
7555 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
7556 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
7557 | ||
7558 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) | |
7559 | { | |
99b1c316 | 7560 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
7561 | loop->header = body_bb; |
7562 | if (collapse_bb == NULL) | |
7563 | loop->latch = cont_bb; | |
7564 | add_loop (loop, body_bb->loop_father); | |
7565 | } | |
7566 | } | |
7567 | ||
7568 | /* A subroutine of expand_omp_for. Generate code for an OpenACC | |
7569 | partitioned loop. The lowering here is abstracted, in that the | |
7570 | loop parameters are passed through internal functions, which are | |
7571 | further lowered by oacc_device_lower, once we get to the target | |
7572 | compiler. The loop is of the form: | |
7573 | ||
7574 | for (V = B; V LTGT E; V += S) {BODY} | |
7575 | ||
7576 | where LTGT is < or >. We may have a specified chunking size, CHUNKING | |
7577 | (constant 0 for no chunking) and we will have a GWV partitioning | |
7578 | mask, specifying dimensions over which the loop is to be | |
02889d23 CLT |
7579 | partitioned (see note below). We generate code that looks like |
7580 | (this ignores tiling): | |
629b3d75 MJ |
7581 | |
7582 | <entry_bb> [incoming FALL->body, BRANCH->exit] | |
7583 | typedef signedintify (typeof (V)) T; // underlying signed integral type | |
7584 | T range = E - B; | |
7585 | T chunk_no = 0; | |
7586 | T DIR = LTGT == '<' ? +1 : -1; | |
7587 | T chunk_max = GOACC_LOOP_CHUNK (dir, range, S, CHUNK_SIZE, GWV); | |
7588 | T step = GOACC_LOOP_STEP (dir, range, S, CHUNK_SIZE, GWV); | |
7589 | ||
7590 | <head_bb> [created by splitting end of entry_bb] | |
7591 | T offset = GOACC_LOOP_OFFSET (dir, range, S, CHUNK_SIZE, GWV, chunk_no); | |
7592 | T bound = GOACC_LOOP_BOUND (dir, range, S, CHUNK_SIZE, GWV, offset); | |
7593 | if (!(offset LTGT bound)) goto bottom_bb; | |
7594 | ||
7595 | <body_bb> [incoming] | |
7596 | V = B + offset; | |
7597 | {BODY} | |
7598 | ||
7599 | <cont_bb> [incoming, may == body_bb FALL->exit_bb, BRANCH->body_bb] | |
7600 | offset += step; | |
7601 | if (offset LTGT bound) goto body_bb; [*] | |
7602 | ||
7603 | <bottom_bb> [created by splitting start of exit_bb] insert BRANCH->head_bb | |
7604 | chunk_no++; | |
7605 | if (chunk < chunk_max) goto head_bb; | |
7606 | ||
7607 | <exit_bb> [incoming] | |
7608 | V = B + ((range -/+ 1) / S +/- 1) * S [*] | |
7609 | ||
02889d23 | 7610 | [*] Needed if V live at end of loop. */ |
629b3d75 MJ |
7611 | |
7612 | static void | |
7613 | expand_oacc_for (struct omp_region *region, struct omp_for_data *fd) | |
7614 | { | |
8c3aa359 TS |
7615 | bool is_oacc_kernels_parallelized |
7616 | = (lookup_attribute ("oacc kernels parallelized", | |
7617 | DECL_ATTRIBUTES (current_function_decl)) != NULL); | |
7618 | { | |
7619 | bool is_oacc_kernels | |
7620 | = (lookup_attribute ("oacc kernels", | |
7621 | DECL_ATTRIBUTES (current_function_decl)) != NULL); | |
7622 | if (is_oacc_kernels_parallelized) | |
7623 | gcc_checking_assert (is_oacc_kernels); | |
7624 | } | |
7625 | gcc_assert (gimple_in_ssa_p (cfun) == is_oacc_kernels_parallelized); | |
7626 | /* In the following, some of the 'gimple_in_ssa_p (cfun)' conditionals are | |
7627 | for SSA specifics, and some are for 'parloops' OpenACC | |
7628 | 'kernels'-parallelized specifics. */ | |
7629 | ||
629b3d75 MJ |
7630 | tree v = fd->loop.v; |
7631 | enum tree_code cond_code = fd->loop.cond_code; | |
7632 | enum tree_code plus_code = PLUS_EXPR; | |
7633 | ||
7634 | tree chunk_size = integer_minus_one_node; | |
7635 | tree gwv = integer_zero_node; | |
7636 | tree iter_type = TREE_TYPE (v); | |
7637 | tree diff_type = iter_type; | |
7638 | tree plus_type = iter_type; | |
7639 | struct oacc_collapse *counts = NULL; | |
7640 | ||
7641 | gcc_checking_assert (gimple_omp_for_kind (fd->for_stmt) | |
7642 | == GF_OMP_FOR_KIND_OACC_LOOP); | |
7643 | gcc_assert (!gimple_omp_for_combined_into_p (fd->for_stmt)); | |
7644 | gcc_assert (cond_code == LT_EXPR || cond_code == GT_EXPR); | |
7645 | ||
7646 | if (POINTER_TYPE_P (iter_type)) | |
7647 | { | |
7648 | plus_code = POINTER_PLUS_EXPR; | |
7649 | plus_type = sizetype; | |
7650 | } | |
f324479c TB |
7651 | for (int ix = fd->collapse; ix--;) |
7652 | { | |
7653 | tree diff_type2 = TREE_TYPE (fd->loops[ix].step); | |
7654 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (diff_type2)) | |
7655 | diff_type = diff_type2; | |
7656 | } | |
629b3d75 MJ |
7657 | if (POINTER_TYPE_P (diff_type) || TYPE_UNSIGNED (diff_type)) |
7658 | diff_type = signed_type_for (diff_type); | |
f4c222c0 TV |
7659 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (integer_type_node)) |
7660 | diff_type = integer_type_node; | |
629b3d75 MJ |
7661 | |
7662 | basic_block entry_bb = region->entry; /* BB ending in OMP_FOR */ | |
7663 | basic_block exit_bb = region->exit; /* BB ending in OMP_RETURN */ | |
7664 | basic_block cont_bb = region->cont; /* BB ending in OMP_CONTINUE */ | |
7665 | basic_block bottom_bb = NULL; | |
7666 | ||
9c3da8cc JJ |
7667 | /* entry_bb has two successors; the branch edge is to the exit |
7668 | block, fallthrough edge to body. */ | |
629b3d75 MJ |
7669 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2 |
7670 | && BRANCH_EDGE (entry_bb)->dest == exit_bb); | |
7671 | ||
7672 | /* If cont_bb non-NULL, it has 2 successors. The branch successor is | |
7673 | body_bb, or to a block whose only successor is the body_bb. Its | |
7674 | fallthrough successor is the final block (same as the branch | |
7675 | successor of the entry_bb). */ | |
7676 | if (cont_bb) | |
7677 | { | |
7678 | basic_block body_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
7679 | basic_block bed = BRANCH_EDGE (cont_bb)->dest; | |
7680 | ||
7681 | gcc_assert (FALLTHRU_EDGE (cont_bb)->dest == exit_bb); | |
7682 | gcc_assert (bed == body_bb || single_succ_edge (bed)->dest == body_bb); | |
7683 | } | |
7684 | else | |
7685 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
7686 | ||
7687 | /* The exit block only has entry_bb and cont_bb as predecessors. */ | |
7688 | gcc_assert (EDGE_COUNT (exit_bb->preds) == 1 + (cont_bb != NULL)); | |
7689 | ||
7690 | tree chunk_no; | |
7691 | tree chunk_max = NULL_TREE; | |
7692 | tree bound, offset; | |
7693 | tree step = create_tmp_var (diff_type, ".step"); | |
7694 | bool up = cond_code == LT_EXPR; | |
7695 | tree dir = build_int_cst (diff_type, up ? +1 : -1); | |
02889d23 | 7696 | bool chunking = !gimple_in_ssa_p (cfun); |
629b3d75 MJ |
7697 | bool negating; |
7698 | ||
02889d23 CLT |
7699 | /* Tiling vars. */ |
7700 | tree tile_size = NULL_TREE; | |
7701 | tree element_s = NULL_TREE; | |
7702 | tree e_bound = NULL_TREE, e_offset = NULL_TREE, e_step = NULL_TREE; | |
7703 | basic_block elem_body_bb = NULL; | |
7704 | basic_block elem_cont_bb = NULL; | |
7705 | ||
629b3d75 MJ |
7706 | /* SSA instances. */ |
7707 | tree offset_incr = NULL_TREE; | |
7708 | tree offset_init = NULL_TREE; | |
7709 | ||
7710 | gimple_stmt_iterator gsi; | |
7711 | gassign *ass; | |
7712 | gcall *call; | |
7713 | gimple *stmt; | |
7714 | tree expr; | |
7715 | location_t loc; | |
7716 | edge split, be, fte; | |
7717 | ||
7718 | /* Split the end of entry_bb to create head_bb. */ | |
fe8ac82f | 7719 | split = split_block (entry_bb, last_nondebug_stmt (entry_bb)); |
629b3d75 MJ |
7720 | basic_block head_bb = split->dest; |
7721 | entry_bb = split->src; | |
7722 | ||
7723 | /* Chunk setup goes at end of entry_bb, replacing the omp_for. */ | |
65f4b875 | 7724 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
7725 | gomp_for *for_stmt = as_a <gomp_for *> (gsi_stmt (gsi)); |
7726 | loc = gimple_location (for_stmt); | |
7727 | ||
7728 | if (gimple_in_ssa_p (cfun)) | |
7729 | { | |
7730 | offset_init = gimple_omp_for_index (for_stmt, 0); | |
7731 | gcc_assert (integer_zerop (fd->loop.n1)); | |
7732 | /* The SSA parallelizer does gang parallelism. */ | |
7733 | gwv = build_int_cst (integer_type_node, GOMP_DIM_MASK (GOMP_DIM_GANG)); | |
7734 | } | |
7735 | ||
02889d23 | 7736 | if (fd->collapse > 1 || fd->tiling) |
629b3d75 | 7737 | { |
02889d23 | 7738 | gcc_assert (!gimple_in_ssa_p (cfun) && up); |
629b3d75 | 7739 | counts = XALLOCAVEC (struct oacc_collapse, fd->collapse); |
f324479c | 7740 | tree total = expand_oacc_collapse_init (fd, &gsi, counts, diff_type, |
02889d23 | 7741 | TREE_TYPE (fd->loop.n2), loc); |
629b3d75 MJ |
7742 | |
7743 | if (SSA_VAR_P (fd->loop.n2)) | |
7744 | { | |
7745 | total = force_gimple_operand_gsi (&gsi, total, false, NULL_TREE, | |
7746 | true, GSI_SAME_STMT); | |
7747 | ass = gimple_build_assign (fd->loop.n2, total); | |
7748 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7749 | } | |
629b3d75 MJ |
7750 | } |
7751 | ||
7752 | tree b = fd->loop.n1; | |
7753 | tree e = fd->loop.n2; | |
7754 | tree s = fd->loop.step; | |
7755 | ||
7756 | b = force_gimple_operand_gsi (&gsi, b, true, NULL_TREE, true, GSI_SAME_STMT); | |
7757 | e = force_gimple_operand_gsi (&gsi, e, true, NULL_TREE, true, GSI_SAME_STMT); | |
7758 | ||
01914336 | 7759 | /* Convert the step, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
7760 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); |
7761 | if (negating) | |
7762 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); | |
7763 | s = fold_convert (diff_type, s); | |
7764 | if (negating) | |
7765 | s = fold_build1 (NEGATE_EXPR, diff_type, s); | |
7766 | s = force_gimple_operand_gsi (&gsi, s, true, NULL_TREE, true, GSI_SAME_STMT); | |
7767 | ||
7768 | if (!chunking) | |
7769 | chunk_size = integer_zero_node; | |
7770 | expr = fold_convert (diff_type, chunk_size); | |
7771 | chunk_size = force_gimple_operand_gsi (&gsi, expr, true, | |
7772 | NULL_TREE, true, GSI_SAME_STMT); | |
02889d23 CLT |
7773 | |
7774 | if (fd->tiling) | |
7775 | { | |
7776 | /* Determine the tile size and element step, | |
7777 | modify the outer loop step size. */ | |
7778 | tile_size = create_tmp_var (diff_type, ".tile_size"); | |
7779 | expr = build_int_cst (diff_type, 1); | |
7780 | for (int ix = 0; ix < fd->collapse; ix++) | |
7781 | expr = fold_build2 (MULT_EXPR, diff_type, counts[ix].tile, expr); | |
7782 | expr = force_gimple_operand_gsi (&gsi, expr, true, | |
7783 | NULL_TREE, true, GSI_SAME_STMT); | |
7784 | ass = gimple_build_assign (tile_size, expr); | |
7785 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7786 | ||
7787 | element_s = create_tmp_var (diff_type, ".element_s"); | |
7788 | ass = gimple_build_assign (element_s, s); | |
7789 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7790 | ||
7791 | expr = fold_build2 (MULT_EXPR, diff_type, s, tile_size); | |
7792 | s = force_gimple_operand_gsi (&gsi, expr, true, | |
7793 | NULL_TREE, true, GSI_SAME_STMT); | |
7794 | } | |
7795 | ||
01914336 | 7796 | /* Determine the range, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
7797 | negating = !up && TYPE_UNSIGNED (iter_type); |
7798 | expr = fold_build2 (MINUS_EXPR, plus_type, | |
7799 | fold_convert (plus_type, negating ? b : e), | |
7800 | fold_convert (plus_type, negating ? e : b)); | |
7801 | expr = fold_convert (diff_type, expr); | |
7802 | if (negating) | |
7803 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); | |
7804 | tree range = force_gimple_operand_gsi (&gsi, expr, true, | |
7805 | NULL_TREE, true, GSI_SAME_STMT); | |
7806 | ||
7807 | chunk_no = build_int_cst (diff_type, 0); | |
7808 | if (chunking) | |
7809 | { | |
7810 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
7811 | ||
7812 | expr = chunk_no; | |
7813 | chunk_max = create_tmp_var (diff_type, ".chunk_max"); | |
7814 | chunk_no = create_tmp_var (diff_type, ".chunk_no"); | |
7815 | ||
7816 | ass = gimple_build_assign (chunk_no, expr); | |
7817 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7818 | ||
7819 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, | |
7820 | build_int_cst (integer_type_node, | |
7821 | IFN_GOACC_LOOP_CHUNKS), | |
7822 | dir, range, s, chunk_size, gwv); | |
7823 | gimple_call_set_lhs (call, chunk_max); | |
7824 | gimple_set_location (call, loc); | |
7825 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7826 | } | |
7827 | else | |
7828 | chunk_size = chunk_no; | |
7829 | ||
7830 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, | |
7831 | build_int_cst (integer_type_node, | |
7832 | IFN_GOACC_LOOP_STEP), | |
7833 | dir, range, s, chunk_size, gwv); | |
7834 | gimple_call_set_lhs (call, step); | |
7835 | gimple_set_location (call, loc); | |
7836 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7837 | ||
7838 | /* Remove the GIMPLE_OMP_FOR. */ | |
7839 | gsi_remove (&gsi, true); | |
7840 | ||
01914336 | 7841 | /* Fixup edges from head_bb. */ |
629b3d75 MJ |
7842 | be = BRANCH_EDGE (head_bb); |
7843 | fte = FALLTHRU_EDGE (head_bb); | |
7844 | be->flags |= EDGE_FALSE_VALUE; | |
7845 | fte->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
7846 | ||
7847 | basic_block body_bb = fte->dest; | |
7848 | ||
7849 | if (gimple_in_ssa_p (cfun)) | |
7850 | { | |
65f4b875 | 7851 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
7852 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
7853 | ||
7854 | offset = gimple_omp_continue_control_use (cont_stmt); | |
7855 | offset_incr = gimple_omp_continue_control_def (cont_stmt); | |
7856 | } | |
7857 | else | |
7858 | { | |
7859 | offset = create_tmp_var (diff_type, ".offset"); | |
7860 | offset_init = offset_incr = offset; | |
7861 | } | |
7862 | bound = create_tmp_var (TREE_TYPE (offset), ".bound"); | |
7863 | ||
7864 | /* Loop offset & bound go into head_bb. */ | |
7865 | gsi = gsi_start_bb (head_bb); | |
7866 | ||
7867 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, | |
7868 | build_int_cst (integer_type_node, | |
7869 | IFN_GOACC_LOOP_OFFSET), | |
7870 | dir, range, s, | |
7871 | chunk_size, gwv, chunk_no); | |
7872 | gimple_call_set_lhs (call, offset_init); | |
7873 | gimple_set_location (call, loc); | |
7874 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); | |
7875 | ||
7876 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, | |
7877 | build_int_cst (integer_type_node, | |
7878 | IFN_GOACC_LOOP_BOUND), | |
7879 | dir, range, s, | |
7880 | chunk_size, gwv, offset_init); | |
7881 | gimple_call_set_lhs (call, bound); | |
7882 | gimple_set_location (call, loc); | |
7883 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); | |
7884 | ||
7885 | expr = build2 (cond_code, boolean_type_node, offset_init, bound); | |
7886 | gsi_insert_after (&gsi, gimple_build_cond_empty (expr), | |
7887 | GSI_CONTINUE_LINKING); | |
7888 | ||
7889 | /* V assignment goes into body_bb. */ | |
7890 | if (!gimple_in_ssa_p (cfun)) | |
7891 | { | |
7892 | gsi = gsi_start_bb (body_bb); | |
7893 | ||
7894 | expr = build2 (plus_code, iter_type, b, | |
7895 | fold_convert (plus_type, offset)); | |
7896 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7897 | true, GSI_SAME_STMT); | |
7898 | ass = gimple_build_assign (v, expr); | |
7899 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
02889d23 CLT |
7900 | |
7901 | if (fd->collapse > 1 || fd->tiling) | |
f324479c | 7902 | expand_oacc_collapse_vars (fd, false, &gsi, counts, v, diff_type); |
02889d23 CLT |
7903 | |
7904 | if (fd->tiling) | |
7905 | { | |
7906 | /* Determine the range of the element loop -- usually simply | |
7907 | the tile_size, but could be smaller if the final | |
7908 | iteration of the outer loop is a partial tile. */ | |
7909 | tree e_range = create_tmp_var (diff_type, ".e_range"); | |
7910 | ||
7911 | expr = build2 (MIN_EXPR, diff_type, | |
7912 | build2 (MINUS_EXPR, diff_type, bound, offset), | |
7913 | build2 (MULT_EXPR, diff_type, tile_size, | |
7914 | element_s)); | |
7915 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7916 | true, GSI_SAME_STMT); | |
7917 | ass = gimple_build_assign (e_range, expr); | |
7918 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7919 | ||
7920 | /* Determine bound, offset & step of inner loop. */ | |
7921 | e_bound = create_tmp_var (diff_type, ".e_bound"); | |
7922 | e_offset = create_tmp_var (diff_type, ".e_offset"); | |
7923 | e_step = create_tmp_var (diff_type, ".e_step"); | |
7924 | ||
7925 | /* Mark these as element loops. */ | |
7926 | tree t, e_gwv = integer_minus_one_node; | |
7927 | tree chunk = build_int_cst (diff_type, 0); /* Never chunked. */ | |
7928 | ||
7929 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_OFFSET); | |
7930 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, | |
7931 | element_s, chunk, e_gwv, chunk); | |
7932 | gimple_call_set_lhs (call, e_offset); | |
7933 | gimple_set_location (call, loc); | |
7934 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7935 | ||
7936 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_BOUND); | |
7937 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, | |
7938 | element_s, chunk, e_gwv, e_offset); | |
7939 | gimple_call_set_lhs (call, e_bound); | |
7940 | gimple_set_location (call, loc); | |
7941 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7942 | ||
7943 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_STEP); | |
7944 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, t, dir, e_range, | |
7945 | element_s, chunk, e_gwv); | |
7946 | gimple_call_set_lhs (call, e_step); | |
7947 | gimple_set_location (call, loc); | |
7948 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
7949 | ||
7950 | /* Add test and split block. */ | |
7951 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); | |
7952 | stmt = gimple_build_cond_empty (expr); | |
7953 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
7954 | split = split_block (body_bb, stmt); | |
7955 | elem_body_bb = split->dest; | |
7956 | if (cont_bb == body_bb) | |
7957 | cont_bb = elem_body_bb; | |
7958 | body_bb = split->src; | |
7959 | ||
7960 | split->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
7961 | ||
05e0af43 CP |
7962 | /* Add a dummy exit for the tiled block when cont_bb is missing. */ |
7963 | if (cont_bb == NULL) | |
7964 | { | |
7965 | edge e = make_edge (body_bb, exit_bb, EDGE_FALSE_VALUE); | |
7966 | e->probability = profile_probability::even (); | |
7967 | split->probability = profile_probability::even (); | |
7968 | } | |
7969 | ||
02889d23 CLT |
7970 | /* Initialize the user's loop vars. */ |
7971 | gsi = gsi_start_bb (elem_body_bb); | |
f324479c TB |
7972 | expand_oacc_collapse_vars (fd, true, &gsi, counts, e_offset, |
7973 | diff_type); | |
02889d23 | 7974 | } |
629b3d75 MJ |
7975 | } |
7976 | ||
7977 | /* Loop increment goes into cont_bb. If this is not a loop, we | |
7978 | will have spawned threads as if it was, and each one will | |
7979 | execute one iteration. The specification is not explicit about | |
7980 | whether such constructs are ill-formed or not, and they can | |
7981 | occur, especially when noreturn routines are involved. */ | |
7982 | if (cont_bb) | |
7983 | { | |
65f4b875 | 7984 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
7985 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
7986 | loc = gimple_location (cont_stmt); | |
7987 | ||
02889d23 CLT |
7988 | if (fd->tiling) |
7989 | { | |
7990 | /* Insert element loop increment and test. */ | |
7991 | expr = build2 (PLUS_EXPR, diff_type, e_offset, e_step); | |
7992 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
7993 | true, GSI_SAME_STMT); | |
7994 | ass = gimple_build_assign (e_offset, expr); | |
7995 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
7996 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); | |
7997 | ||
7998 | stmt = gimple_build_cond_empty (expr); | |
7999 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
8000 | split = split_block (cont_bb, stmt); | |
8001 | elem_cont_bb = split->src; | |
8002 | cont_bb = split->dest; | |
8003 | ||
8004 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; | |
9fba2943 TV |
8005 | split->probability = profile_probability::unlikely ().guessed (); |
8006 | edge latch_edge | |
8007 | = make_edge (elem_cont_bb, elem_body_bb, EDGE_TRUE_VALUE); | |
8008 | latch_edge->probability = profile_probability::likely ().guessed (); | |
8009 | ||
8010 | edge skip_edge = make_edge (body_bb, cont_bb, EDGE_FALSE_VALUE); | |
8011 | skip_edge->probability = profile_probability::unlikely ().guessed (); | |
8012 | edge loop_entry_edge = EDGE_SUCC (body_bb, 1 - skip_edge->dest_idx); | |
8013 | loop_entry_edge->probability | |
8014 | = profile_probability::likely ().guessed (); | |
02889d23 CLT |
8015 | |
8016 | gsi = gsi_for_stmt (cont_stmt); | |
8017 | } | |
8018 | ||
629b3d75 MJ |
8019 | /* Increment offset. */ |
8020 | if (gimple_in_ssa_p (cfun)) | |
02889d23 CLT |
8021 | expr = build2 (plus_code, iter_type, offset, |
8022 | fold_convert (plus_type, step)); | |
629b3d75 MJ |
8023 | else |
8024 | expr = build2 (PLUS_EXPR, diff_type, offset, step); | |
8025 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
8026 | true, GSI_SAME_STMT); | |
8027 | ass = gimple_build_assign (offset_incr, expr); | |
8028 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
8029 | expr = build2 (cond_code, boolean_type_node, offset_incr, bound); | |
8030 | gsi_insert_before (&gsi, gimple_build_cond_empty (expr), GSI_SAME_STMT); | |
8031 | ||
8032 | /* Remove the GIMPLE_OMP_CONTINUE. */ | |
8033 | gsi_remove (&gsi, true); | |
8034 | ||
01914336 | 8035 | /* Fixup edges from cont_bb. */ |
629b3d75 MJ |
8036 | be = BRANCH_EDGE (cont_bb); |
8037 | fte = FALLTHRU_EDGE (cont_bb); | |
8038 | be->flags |= EDGE_TRUE_VALUE; | |
8039 | fte->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; | |
8040 | ||
8041 | if (chunking) | |
8042 | { | |
8043 | /* Split the beginning of exit_bb to make bottom_bb. We | |
8044 | need to insert a nop at the start, because splitting is | |
01914336 | 8045 | after a stmt, not before. */ |
629b3d75 MJ |
8046 | gsi = gsi_start_bb (exit_bb); |
8047 | stmt = gimple_build_nop (); | |
8048 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
8049 | split = split_block (exit_bb, stmt); | |
8050 | bottom_bb = split->src; | |
8051 | exit_bb = split->dest; | |
8052 | gsi = gsi_last_bb (bottom_bb); | |
8053 | ||
8054 | /* Chunk increment and test goes into bottom_bb. */ | |
8055 | expr = build2 (PLUS_EXPR, diff_type, chunk_no, | |
8056 | build_int_cst (diff_type, 1)); | |
8057 | ass = gimple_build_assign (chunk_no, expr); | |
8058 | gsi_insert_after (&gsi, ass, GSI_CONTINUE_LINKING); | |
8059 | ||
8060 | /* Chunk test at end of bottom_bb. */ | |
8061 | expr = build2 (LT_EXPR, boolean_type_node, chunk_no, chunk_max); | |
8062 | gsi_insert_after (&gsi, gimple_build_cond_empty (expr), | |
8063 | GSI_CONTINUE_LINKING); | |
8064 | ||
01914336 | 8065 | /* Fixup edges from bottom_bb. */ |
629b3d75 | 8066 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; |
f38057cc TV |
8067 | split->probability = profile_probability::unlikely ().guessed (); |
8068 | edge latch_edge = make_edge (bottom_bb, head_bb, EDGE_TRUE_VALUE); | |
8069 | latch_edge->probability = profile_probability::likely ().guessed (); | |
629b3d75 MJ |
8070 | } |
8071 | } | |
8072 | ||
65f4b875 | 8073 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
8074 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
8075 | loc = gimple_location (gsi_stmt (gsi)); | |
8076 | ||
8077 | if (!gimple_in_ssa_p (cfun)) | |
8078 | { | |
8079 | /* Insert the final value of V, in case it is live. This is the | |
8080 | value for the only thread that survives past the join. */ | |
8081 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); | |
8082 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); | |
8083 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); | |
8084 | expr = fold_build2 (MULT_EXPR, diff_type, expr, s); | |
8085 | expr = build2 (plus_code, iter_type, b, fold_convert (plus_type, expr)); | |
8086 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
8087 | true, GSI_SAME_STMT); | |
8088 | ass = gimple_build_assign (v, expr); | |
8089 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
8090 | } | |
8091 | ||
01914336 | 8092 | /* Remove the OMP_RETURN. */ |
629b3d75 MJ |
8093 | gsi_remove (&gsi, true); |
8094 | ||
8095 | if (cont_bb) | |
8096 | { | |
02889d23 | 8097 | /* We now have one, two or three nested loops. Update the loop |
629b3d75 | 8098 | structures. */ |
99b1c316 MS |
8099 | class loop *parent = entry_bb->loop_father; |
8100 | class loop *body = body_bb->loop_father; | |
629b3d75 MJ |
8101 | |
8102 | if (chunking) | |
8103 | { | |
99b1c316 | 8104 | class loop *chunk_loop = alloc_loop (); |
629b3d75 MJ |
8105 | chunk_loop->header = head_bb; |
8106 | chunk_loop->latch = bottom_bb; | |
8107 | add_loop (chunk_loop, parent); | |
8108 | parent = chunk_loop; | |
8109 | } | |
8110 | else if (parent != body) | |
8111 | { | |
8112 | gcc_assert (body->header == body_bb); | |
8113 | gcc_assert (body->latch == cont_bb | |
8114 | || single_pred (body->latch) == cont_bb); | |
8115 | parent = NULL; | |
8116 | } | |
8117 | ||
8118 | if (parent) | |
8119 | { | |
99b1c316 | 8120 | class loop *body_loop = alloc_loop (); |
629b3d75 MJ |
8121 | body_loop->header = body_bb; |
8122 | body_loop->latch = cont_bb; | |
8123 | add_loop (body_loop, parent); | |
02889d23 CLT |
8124 | |
8125 | if (fd->tiling) | |
8126 | { | |
8127 | /* Insert tiling's element loop. */ | |
99b1c316 | 8128 | class loop *inner_loop = alloc_loop (); |
02889d23 CLT |
8129 | inner_loop->header = elem_body_bb; |
8130 | inner_loop->latch = elem_cont_bb; | |
8131 | add_loop (inner_loop, body_loop); | |
8132 | } | |
629b3d75 MJ |
8133 | } |
8134 | } | |
8135 | } | |
8136 | ||
8137 | /* Expand the OMP loop defined by REGION. */ | |
8138 | ||
8139 | static void | |
8140 | expand_omp_for (struct omp_region *region, gimple *inner_stmt) | |
8141 | { | |
8142 | struct omp_for_data fd; | |
8143 | struct omp_for_data_loop *loops; | |
8144 | ||
9d50112a | 8145 | loops = XALLOCAVEC (struct omp_for_data_loop, |
fe8ac82f RB |
8146 | gimple_omp_for_collapse |
8147 | (last_nondebug_stmt (region->entry))); | |
8148 | omp_extract_for_data (as_a <gomp_for *> (last_nondebug_stmt (region->entry)), | |
629b3d75 MJ |
8149 | &fd, loops); |
8150 | region->sched_kind = fd.sched_kind; | |
8151 | region->sched_modifiers = fd.sched_modifiers; | |
0b887b75 | 8152 | region->has_lastprivate_conditional = fd.lastprivate_conditional != 0; |
9d50112a JJ |
8153 | if (fd.non_rect && !gimple_omp_for_combined_into_p (fd.for_stmt)) |
8154 | { | |
8155 | for (int i = fd.first_nonrect; i <= fd.last_nonrect; i++) | |
8156 | if ((loops[i].m1 || loops[i].m2) | |
8157 | && (loops[i].m1 == NULL_TREE | |
8158 | || TREE_CODE (loops[i].m1) == INTEGER_CST) | |
8159 | && (loops[i].m2 == NULL_TREE | |
8160 | || TREE_CODE (loops[i].m2) == INTEGER_CST) | |
8161 | && TREE_CODE (loops[i].step) == INTEGER_CST | |
8162 | && TREE_CODE (loops[i - loops[i].outer].step) == INTEGER_CST) | |
8163 | { | |
8164 | tree t; | |
8165 | tree itype = TREE_TYPE (loops[i].v); | |
8166 | if (loops[i].m1 && loops[i].m2) | |
8167 | t = fold_build2 (MINUS_EXPR, itype, loops[i].m2, loops[i].m1); | |
8168 | else if (loops[i].m1) | |
8169 | t = fold_build1 (NEGATE_EXPR, itype, loops[i].m1); | |
8170 | else | |
8171 | t = loops[i].m2; | |
8172 | t = fold_build2 (MULT_EXPR, itype, t, | |
8173 | fold_convert (itype, | |
8174 | loops[i - loops[i].outer].step)); | |
8175 | if (TYPE_UNSIGNED (itype) && loops[i].cond_code == GT_EXPR) | |
8176 | t = fold_build2 (TRUNC_MOD_EXPR, itype, | |
8177 | fold_build1 (NEGATE_EXPR, itype, t), | |
8178 | fold_build1 (NEGATE_EXPR, itype, | |
8179 | fold_convert (itype, | |
8180 | loops[i].step))); | |
8181 | else | |
8182 | t = fold_build2 (TRUNC_MOD_EXPR, itype, t, | |
8183 | fold_convert (itype, loops[i].step)); | |
8184 | if (integer_nonzerop (t)) | |
8185 | error_at (gimple_location (fd.for_stmt), | |
8186 | "invalid OpenMP non-rectangular loop step; " | |
8187 | "%<(%E - %E) * %E%> is not a multiple of loop %d " | |
8188 | "step %qE", | |
8189 | loops[i].m2 ? loops[i].m2 : integer_zero_node, | |
8190 | loops[i].m1 ? loops[i].m1 : integer_zero_node, | |
8191 | loops[i - loops[i].outer].step, i + 1, | |
8192 | loops[i].step); | |
8193 | } | |
8194 | } | |
629b3d75 MJ |
8195 | |
8196 | gcc_assert (EDGE_COUNT (region->entry->succs) == 2); | |
8197 | BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
8198 | FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
8199 | if (region->cont) | |
8200 | { | |
8201 | gcc_assert (EDGE_COUNT (region->cont->succs) == 2); | |
8202 | BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
8203 | FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
8204 | } | |
8205 | else | |
8206 | /* If there isn't a continue then this is a degerate case where | |
8207 | the introduction of abnormal edges during lowering will prevent | |
8208 | original loops from being detected. Fix that up. */ | |
8209 | loops_state_set (LOOPS_NEED_FIXUP); | |
8210 | ||
dfa6e5b4 | 8211 | if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_SIMD) |
c2ebf4f1 | 8212 | expand_omp_simd (region, &fd); |
629b3d75 MJ |
8213 | else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_OACC_LOOP) |
8214 | { | |
1160ec9a | 8215 | gcc_assert (!inner_stmt && !fd.non_rect); |
629b3d75 MJ |
8216 | expand_oacc_for (region, &fd); |
8217 | } | |
8218 | else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_TASKLOOP) | |
8219 | { | |
8220 | if (gimple_omp_for_combined_into_p (fd.for_stmt)) | |
8221 | expand_omp_taskloop_for_inner (region, &fd, inner_stmt); | |
8222 | else | |
8223 | expand_omp_taskloop_for_outer (region, &fd, inner_stmt); | |
8224 | } | |
8225 | else if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC | |
8226 | && !fd.have_ordered) | |
8227 | { | |
8228 | if (fd.chunk_size == NULL) | |
8229 | expand_omp_for_static_nochunk (region, &fd, inner_stmt); | |
8230 | else | |
8231 | expand_omp_for_static_chunk (region, &fd, inner_stmt); | |
8232 | } | |
8233 | else | |
8234 | { | |
8235 | int fn_index, start_ix, next_ix; | |
28567c40 JJ |
8236 | unsigned HOST_WIDE_INT sched = 0; |
8237 | tree sched_arg = NULL_TREE; | |
629b3d75 MJ |
8238 | |
8239 | gcc_assert (gimple_omp_for_kind (fd.for_stmt) | |
1160ec9a | 8240 | == GF_OMP_FOR_KIND_FOR && !fd.non_rect); |
629b3d75 MJ |
8241 | if (fd.chunk_size == NULL |
8242 | && fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) | |
8243 | fd.chunk_size = integer_zero_node; | |
629b3d75 MJ |
8244 | switch (fd.sched_kind) |
8245 | { | |
8246 | case OMP_CLAUSE_SCHEDULE_RUNTIME: | |
0b887b75 JJ |
8247 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0 |
8248 | && fd.lastprivate_conditional == 0) | |
28567c40 JJ |
8249 | { |
8250 | gcc_assert (!fd.have_ordered); | |
8251 | fn_index = 6; | |
8252 | sched = 4; | |
8253 | } | |
8254 | else if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 | |
0b887b75 JJ |
8255 | && !fd.have_ordered |
8256 | && fd.lastprivate_conditional == 0) | |
28567c40 JJ |
8257 | fn_index = 7; |
8258 | else | |
8259 | { | |
8260 | fn_index = 3; | |
8261 | sched = (HOST_WIDE_INT_1U << 31); | |
8262 | } | |
629b3d75 MJ |
8263 | break; |
8264 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: | |
8265 | case OMP_CLAUSE_SCHEDULE_GUIDED: | |
28567c40 | 8266 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 |
0b887b75 JJ |
8267 | && !fd.have_ordered |
8268 | && fd.lastprivate_conditional == 0) | |
629b3d75 MJ |
8269 | { |
8270 | fn_index = 3 + fd.sched_kind; | |
28567c40 | 8271 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; |
629b3d75 MJ |
8272 | break; |
8273 | } | |
629b3d75 | 8274 | fn_index = fd.sched_kind; |
28567c40 JJ |
8275 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; |
8276 | sched += (HOST_WIDE_INT_1U << 31); | |
629b3d75 | 8277 | break; |
28567c40 JJ |
8278 | case OMP_CLAUSE_SCHEDULE_STATIC: |
8279 | gcc_assert (fd.have_ordered); | |
8280 | fn_index = 0; | |
8281 | sched = (HOST_WIDE_INT_1U << 31) + 1; | |
8282 | break; | |
8283 | default: | |
8284 | gcc_unreachable (); | |
629b3d75 MJ |
8285 | } |
8286 | if (!fd.ordered) | |
28567c40 | 8287 | fn_index += fd.have_ordered * 8; |
629b3d75 MJ |
8288 | if (fd.ordered) |
8289 | start_ix = ((int)BUILT_IN_GOMP_LOOP_DOACROSS_STATIC_START) + fn_index; | |
8290 | else | |
8291 | start_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_START) + fn_index; | |
8292 | next_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_NEXT) + fn_index; | |
8221c30b | 8293 | if (fd.have_reductemp || fd.have_pointer_condtemp) |
28567c40 JJ |
8294 | { |
8295 | if (fd.ordered) | |
8296 | start_ix = (int)BUILT_IN_GOMP_LOOP_DOACROSS_START; | |
8297 | else if (fd.have_ordered) | |
8298 | start_ix = (int)BUILT_IN_GOMP_LOOP_ORDERED_START; | |
8299 | else | |
8300 | start_ix = (int)BUILT_IN_GOMP_LOOP_START; | |
8301 | sched_arg = build_int_cstu (long_integer_type_node, sched); | |
8302 | if (!fd.chunk_size) | |
8303 | fd.chunk_size = integer_zero_node; | |
8304 | } | |
629b3d75 MJ |
8305 | if (fd.iter_type == long_long_unsigned_type_node) |
8306 | { | |
8307 | start_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_START | |
8308 | - (int)BUILT_IN_GOMP_LOOP_STATIC_START); | |
8309 | next_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_NEXT | |
8310 | - (int)BUILT_IN_GOMP_LOOP_STATIC_NEXT); | |
8311 | } | |
8312 | expand_omp_for_generic (region, &fd, (enum built_in_function) start_ix, | |
28567c40 JJ |
8313 | (enum built_in_function) next_ix, sched_arg, |
8314 | inner_stmt); | |
629b3d75 | 8315 | } |
629b3d75 MJ |
8316 | } |
8317 | ||
8318 | /* Expand code for an OpenMP sections directive. In pseudo code, we generate | |
8319 | ||
8320 | v = GOMP_sections_start (n); | |
8321 | L0: | |
8322 | switch (v) | |
8323 | { | |
8324 | case 0: | |
8325 | goto L2; | |
8326 | case 1: | |
8327 | section 1; | |
8328 | goto L1; | |
8329 | case 2: | |
8330 | ... | |
8331 | case n: | |
8332 | ... | |
8333 | default: | |
8334 | abort (); | |
8335 | } | |
8336 | L1: | |
8337 | v = GOMP_sections_next (); | |
8338 | goto L0; | |
8339 | L2: | |
8340 | reduction; | |
8341 | ||
8342 | If this is a combined parallel sections, replace the call to | |
8343 | GOMP_sections_start with call to GOMP_sections_next. */ | |
8344 | ||
8345 | static void | |
8346 | expand_omp_sections (struct omp_region *region) | |
8347 | { | |
8348 | tree t, u, vin = NULL, vmain, vnext, l2; | |
8349 | unsigned len; | |
8350 | basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb; | |
8351 | gimple_stmt_iterator si, switch_si; | |
8352 | gomp_sections *sections_stmt; | |
8353 | gimple *stmt; | |
8354 | gomp_continue *cont; | |
8355 | edge_iterator ei; | |
8356 | edge e; | |
8357 | struct omp_region *inner; | |
8358 | unsigned i, casei; | |
8359 | bool exit_reachable = region->cont != NULL; | |
8360 | ||
8361 | gcc_assert (region->exit != NULL); | |
8362 | entry_bb = region->entry; | |
8363 | l0_bb = single_succ (entry_bb); | |
8364 | l1_bb = region->cont; | |
8365 | l2_bb = region->exit; | |
8366 | if (single_pred_p (l2_bb) && single_pred (l2_bb) == l0_bb) | |
8367 | l2 = gimple_block_label (l2_bb); | |
8368 | else | |
8369 | { | |
8370 | /* This can happen if there are reductions. */ | |
8371 | len = EDGE_COUNT (l0_bb->succs); | |
8372 | gcc_assert (len > 0); | |
8373 | e = EDGE_SUCC (l0_bb, len - 1); | |
65f4b875 | 8374 | si = gsi_last_nondebug_bb (e->dest); |
629b3d75 MJ |
8375 | l2 = NULL_TREE; |
8376 | if (gsi_end_p (si) | |
01914336 | 8377 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) |
629b3d75 MJ |
8378 | l2 = gimple_block_label (e->dest); |
8379 | else | |
8380 | FOR_EACH_EDGE (e, ei, l0_bb->succs) | |
8381 | { | |
65f4b875 | 8382 | si = gsi_last_nondebug_bb (e->dest); |
629b3d75 MJ |
8383 | if (gsi_end_p (si) |
8384 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
8385 | { | |
8386 | l2 = gimple_block_label (e->dest); | |
8387 | break; | |
8388 | } | |
8389 | } | |
8390 | } | |
8391 | if (exit_reachable) | |
8392 | default_bb = create_empty_bb (l1_bb->prev_bb); | |
8393 | else | |
8394 | default_bb = create_empty_bb (l0_bb); | |
8395 | ||
8396 | /* We will build a switch() with enough cases for all the | |
8397 | GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work | |
8398 | and a default case to abort if something goes wrong. */ | |
8399 | len = EDGE_COUNT (l0_bb->succs); | |
8400 | ||
8401 | /* Use vec::quick_push on label_vec throughout, since we know the size | |
8402 | in advance. */ | |
8403 | auto_vec<tree> label_vec (len); | |
8404 | ||
8405 | /* The call to GOMP_sections_start goes in ENTRY_BB, replacing the | |
8406 | GIMPLE_OMP_SECTIONS statement. */ | |
65f4b875 | 8407 | si = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
8408 | sections_stmt = as_a <gomp_sections *> (gsi_stmt (si)); |
8409 | gcc_assert (gimple_code (sections_stmt) == GIMPLE_OMP_SECTIONS); | |
8410 | vin = gimple_omp_sections_control (sections_stmt); | |
28567c40 JJ |
8411 | tree clauses = gimple_omp_sections_clauses (sections_stmt); |
8412 | tree reductmp = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
8e7757ba JJ |
8413 | tree condtmp = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); |
8414 | tree cond_var = NULL_TREE; | |
8415 | if (reductmp || condtmp) | |
8416 | { | |
8417 | tree reductions = null_pointer_node, mem = null_pointer_node; | |
8418 | tree memv = NULL_TREE, condtemp = NULL_TREE; | |
8419 | gimple_stmt_iterator gsi = gsi_none (); | |
8420 | gimple *g = NULL; | |
8421 | if (reductmp) | |
8422 | { | |
8423 | reductions = OMP_CLAUSE_DECL (reductmp); | |
8424 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
8425 | g = SSA_NAME_DEF_STMT (reductions); | |
8426 | reductions = gimple_assign_rhs1 (g); | |
8427 | OMP_CLAUSE_DECL (reductmp) = reductions; | |
8428 | gsi = gsi_for_stmt (g); | |
8429 | } | |
8430 | else | |
8431 | gsi = si; | |
8432 | if (condtmp) | |
8433 | { | |
8434 | condtemp = OMP_CLAUSE_DECL (condtmp); | |
8435 | tree c = omp_find_clause (OMP_CLAUSE_CHAIN (condtmp), | |
8436 | OMP_CLAUSE__CONDTEMP_); | |
8437 | cond_var = OMP_CLAUSE_DECL (c); | |
8438 | tree type = TREE_TYPE (condtemp); | |
8439 | memv = create_tmp_var (type); | |
8440 | TREE_ADDRESSABLE (memv) = 1; | |
8441 | unsigned cnt = 0; | |
8442 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
8443 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE | |
8444 | && OMP_CLAUSE_LASTPRIVATE_CONDITIONAL (c)) | |
8445 | ++cnt; | |
8446 | unsigned HOST_WIDE_INT sz | |
8447 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))) * cnt; | |
8448 | expand_omp_build_assign (&gsi, memv, build_int_cst (type, sz), | |
8449 | false); | |
8450 | mem = build_fold_addr_expr (memv); | |
8451 | } | |
28567c40 JJ |
8452 | t = build_int_cst (unsigned_type_node, len - 1); |
8453 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS2_START); | |
8e7757ba | 8454 | stmt = gimple_build_call (u, 3, t, reductions, mem); |
28567c40 JJ |
8455 | gimple_call_set_lhs (stmt, vin); |
8456 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
8e7757ba JJ |
8457 | if (condtmp) |
8458 | { | |
8459 | expand_omp_build_assign (&gsi, condtemp, memv, false); | |
8460 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), | |
8461 | vin, build_one_cst (TREE_TYPE (cond_var))); | |
8462 | expand_omp_build_assign (&gsi, cond_var, t, false); | |
8463 | } | |
8464 | if (reductmp) | |
8465 | { | |
8466 | gsi_remove (&gsi, true); | |
8467 | release_ssa_name (gimple_assign_lhs (g)); | |
8468 | } | |
28567c40 JJ |
8469 | } |
8470 | else if (!is_combined_parallel (region)) | |
629b3d75 MJ |
8471 | { |
8472 | /* If we are not inside a combined parallel+sections region, | |
8473 | call GOMP_sections_start. */ | |
8474 | t = build_int_cst (unsigned_type_node, len - 1); | |
8475 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_START); | |
8476 | stmt = gimple_build_call (u, 1, t); | |
8477 | } | |
8478 | else | |
8479 | { | |
8480 | /* Otherwise, call GOMP_sections_next. */ | |
8481 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); | |
8482 | stmt = gimple_build_call (u, 0); | |
8483 | } | |
8e7757ba | 8484 | if (!reductmp && !condtmp) |
28567c40 JJ |
8485 | { |
8486 | gimple_call_set_lhs (stmt, vin); | |
8487 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
8488 | } | |
629b3d75 MJ |
8489 | gsi_remove (&si, true); |
8490 | ||
8491 | /* The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in | |
8492 | L0_BB. */ | |
65f4b875 | 8493 | switch_si = gsi_last_nondebug_bb (l0_bb); |
629b3d75 MJ |
8494 | gcc_assert (gimple_code (gsi_stmt (switch_si)) == GIMPLE_OMP_SECTIONS_SWITCH); |
8495 | if (exit_reachable) | |
8496 | { | |
fe8ac82f | 8497 | cont = as_a <gomp_continue *> (last_nondebug_stmt (l1_bb)); |
629b3d75 MJ |
8498 | gcc_assert (gimple_code (cont) == GIMPLE_OMP_CONTINUE); |
8499 | vmain = gimple_omp_continue_control_use (cont); | |
8500 | vnext = gimple_omp_continue_control_def (cont); | |
8501 | } | |
8502 | else | |
8503 | { | |
8504 | vmain = vin; | |
8505 | vnext = NULL_TREE; | |
8506 | } | |
8507 | ||
8508 | t = build_case_label (build_int_cst (unsigned_type_node, 0), NULL, l2); | |
8509 | label_vec.quick_push (t); | |
8510 | i = 1; | |
8511 | ||
8512 | /* Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR. */ | |
8513 | for (inner = region->inner, casei = 1; | |
8514 | inner; | |
8515 | inner = inner->next, i++, casei++) | |
8516 | { | |
8517 | basic_block s_entry_bb, s_exit_bb; | |
8518 | ||
8519 | /* Skip optional reduction region. */ | |
8520 | if (inner->type == GIMPLE_OMP_ATOMIC_LOAD) | |
8521 | { | |
8522 | --i; | |
8523 | --casei; | |
8524 | continue; | |
8525 | } | |
8526 | ||
8527 | s_entry_bb = inner->entry; | |
8528 | s_exit_bb = inner->exit; | |
8529 | ||
8530 | t = gimple_block_label (s_entry_bb); | |
8531 | u = build_int_cst (unsigned_type_node, casei); | |
8532 | u = build_case_label (u, NULL, t); | |
8533 | label_vec.quick_push (u); | |
8534 | ||
65f4b875 | 8535 | si = gsi_last_nondebug_bb (s_entry_bb); |
629b3d75 MJ |
8536 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SECTION); |
8537 | gcc_assert (i < len || gimple_omp_section_last_p (gsi_stmt (si))); | |
8538 | gsi_remove (&si, true); | |
8539 | single_succ_edge (s_entry_bb)->flags = EDGE_FALLTHRU; | |
8540 | ||
8541 | if (s_exit_bb == NULL) | |
8542 | continue; | |
8543 | ||
65f4b875 | 8544 | si = gsi_last_nondebug_bb (s_exit_bb); |
629b3d75 MJ |
8545 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); |
8546 | gsi_remove (&si, true); | |
8547 | ||
8548 | single_succ_edge (s_exit_bb)->flags = EDGE_FALLTHRU; | |
8549 | } | |
8550 | ||
8551 | /* Error handling code goes in DEFAULT_BB. */ | |
8552 | t = gimple_block_label (default_bb); | |
8553 | u = build_case_label (NULL, NULL, t); | |
8554 | make_edge (l0_bb, default_bb, 0); | |
8555 | add_bb_to_loop (default_bb, current_loops->tree_root); | |
8556 | ||
8557 | stmt = gimple_build_switch (vmain, u, label_vec); | |
8558 | gsi_insert_after (&switch_si, stmt, GSI_SAME_STMT); | |
8559 | gsi_remove (&switch_si, true); | |
8560 | ||
8561 | si = gsi_start_bb (default_bb); | |
8562 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_TRAP), 0); | |
8563 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); | |
8564 | ||
8565 | if (exit_reachable) | |
8566 | { | |
8567 | tree bfn_decl; | |
8568 | ||
8569 | /* Code to get the next section goes in L1_BB. */ | |
65f4b875 | 8570 | si = gsi_last_nondebug_bb (l1_bb); |
629b3d75 MJ |
8571 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CONTINUE); |
8572 | ||
8573 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); | |
8574 | stmt = gimple_build_call (bfn_decl, 0); | |
8575 | gimple_call_set_lhs (stmt, vnext); | |
8e7757ba JJ |
8576 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
8577 | if (cond_var) | |
8578 | { | |
8579 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), | |
8580 | vnext, build_one_cst (TREE_TYPE (cond_var))); | |
8581 | expand_omp_build_assign (&si, cond_var, t, false); | |
8582 | } | |
629b3d75 MJ |
8583 | gsi_remove (&si, true); |
8584 | ||
8585 | single_succ_edge (l1_bb)->flags = EDGE_FALLTHRU; | |
8586 | } | |
8587 | ||
8588 | /* Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB. */ | |
65f4b875 | 8589 | si = gsi_last_nondebug_bb (l2_bb); |
629b3d75 MJ |
8590 | if (gimple_omp_return_nowait_p (gsi_stmt (si))) |
8591 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_NOWAIT); | |
8592 | else if (gimple_omp_return_lhs (gsi_stmt (si))) | |
8593 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_CANCEL); | |
8594 | else | |
8595 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END); | |
8596 | stmt = gimple_build_call (t, 0); | |
8597 | if (gimple_omp_return_lhs (gsi_stmt (si))) | |
8598 | gimple_call_set_lhs (stmt, gimple_omp_return_lhs (gsi_stmt (si))); | |
8599 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
8600 | gsi_remove (&si, true); | |
8601 | ||
8602 | set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb); | |
8603 | } | |
8604 | ||
e45483c7 | 8605 | /* Expand code for an OpenMP single or scope directive. We've already expanded |
629b3d75 MJ |
8606 | much of the code, here we simply place the GOMP_barrier call. */ |
8607 | ||
8608 | static void | |
8609 | expand_omp_single (struct omp_region *region) | |
8610 | { | |
8611 | basic_block entry_bb, exit_bb; | |
8612 | gimple_stmt_iterator si; | |
8613 | ||
8614 | entry_bb = region->entry; | |
8615 | exit_bb = region->exit; | |
8616 | ||
65f4b875 | 8617 | si = gsi_last_nondebug_bb (entry_bb); |
c4432b27 JJ |
8618 | enum gimple_code code = gimple_code (gsi_stmt (si)); |
8619 | gcc_assert (code == GIMPLE_OMP_SINGLE || code == GIMPLE_OMP_SCOPE); | |
629b3d75 MJ |
8620 | gsi_remove (&si, true); |
8621 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
8622 | ||
c4432b27 JJ |
8623 | if (exit_bb == NULL) |
8624 | { | |
8625 | gcc_assert (code == GIMPLE_OMP_SCOPE); | |
8626 | return; | |
8627 | } | |
8628 | ||
65f4b875 | 8629 | si = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
8630 | if (!gimple_omp_return_nowait_p (gsi_stmt (si))) |
8631 | { | |
8632 | tree t = gimple_omp_return_lhs (gsi_stmt (si)); | |
8633 | gsi_insert_after (&si, omp_build_barrier (t), GSI_SAME_STMT); | |
8634 | } | |
8635 | gsi_remove (&si, true); | |
8636 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; | |
8637 | } | |
8638 | ||
8639 | /* Generic expansion for OpenMP synchronization directives: master, | |
8640 | ordered and critical. All we need to do here is remove the entry | |
8641 | and exit markers for REGION. */ | |
8642 | ||
8643 | static void | |
8644 | expand_omp_synch (struct omp_region *region) | |
8645 | { | |
8646 | basic_block entry_bb, exit_bb; | |
8647 | gimple_stmt_iterator si; | |
8648 | ||
8649 | entry_bb = region->entry; | |
8650 | exit_bb = region->exit; | |
8651 | ||
65f4b875 | 8652 | si = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
8653 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE |
8654 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASTER | |
d0befed7 | 8655 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASKED |
629b3d75 MJ |
8656 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TASKGROUP |
8657 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ORDERED | |
8658 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CRITICAL | |
8659 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS); | |
28567c40 JJ |
8660 | if (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS |
8661 | && gimple_omp_teams_host (as_a <gomp_teams *> (gsi_stmt (si)))) | |
8662 | { | |
8663 | expand_omp_taskreg (region); | |
8664 | return; | |
8665 | } | |
629b3d75 MJ |
8666 | gsi_remove (&si, true); |
8667 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
8668 | ||
8669 | if (exit_bb) | |
8670 | { | |
65f4b875 | 8671 | si = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
8672 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); |
8673 | gsi_remove (&si, true); | |
8674 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; | |
8675 | } | |
8676 | } | |
8677 | ||
8122fbff JJ |
8678 | /* Translate enum omp_memory_order to enum memmodel for the embedded |
8679 | fail clause in there. */ | |
8680 | ||
8681 | static enum memmodel | |
8682 | omp_memory_order_to_fail_memmodel (enum omp_memory_order mo) | |
8683 | { | |
8684 | switch (mo & OMP_FAIL_MEMORY_ORDER_MASK) | |
8685 | { | |
8686 | case OMP_FAIL_MEMORY_ORDER_UNSPECIFIED: | |
8687 | switch (mo & OMP_MEMORY_ORDER_MASK) | |
8688 | { | |
8689 | case OMP_MEMORY_ORDER_RELAXED: return MEMMODEL_RELAXED; | |
8690 | case OMP_MEMORY_ORDER_ACQUIRE: return MEMMODEL_ACQUIRE; | |
8691 | case OMP_MEMORY_ORDER_RELEASE: return MEMMODEL_RELAXED; | |
8692 | case OMP_MEMORY_ORDER_ACQ_REL: return MEMMODEL_ACQUIRE; | |
8693 | case OMP_MEMORY_ORDER_SEQ_CST: return MEMMODEL_SEQ_CST; | |
8694 | default: break; | |
8695 | } | |
8696 | gcc_unreachable (); | |
8697 | case OMP_FAIL_MEMORY_ORDER_RELAXED: return MEMMODEL_RELAXED; | |
8698 | case OMP_FAIL_MEMORY_ORDER_ACQUIRE: return MEMMODEL_ACQUIRE; | |
8699 | case OMP_FAIL_MEMORY_ORDER_SEQ_CST: return MEMMODEL_SEQ_CST; | |
8700 | default: gcc_unreachable (); | |
8701 | } | |
8702 | } | |
8703 | ||
28567c40 JJ |
8704 | /* Translate enum omp_memory_order to enum memmodel. The two enums |
8705 | are using different numbers so that OMP_MEMORY_ORDER_UNSPECIFIED | |
8122fbff | 8706 | is 0 and omp_memory_order has the fail mode encoded in it too. */ |
28567c40 JJ |
8707 | |
8708 | static enum memmodel | |
8709 | omp_memory_order_to_memmodel (enum omp_memory_order mo) | |
8710 | { | |
8122fbff JJ |
8711 | enum memmodel ret, fail_ret; |
8712 | switch (mo & OMP_MEMORY_ORDER_MASK) | |
8713 | { | |
8714 | case OMP_MEMORY_ORDER_RELAXED: ret = MEMMODEL_RELAXED; break; | |
8715 | case OMP_MEMORY_ORDER_ACQUIRE: ret = MEMMODEL_ACQUIRE; break; | |
8716 | case OMP_MEMORY_ORDER_RELEASE: ret = MEMMODEL_RELEASE; break; | |
8717 | case OMP_MEMORY_ORDER_ACQ_REL: ret = MEMMODEL_ACQ_REL; break; | |
8718 | case OMP_MEMORY_ORDER_SEQ_CST: ret = MEMMODEL_SEQ_CST; break; | |
28567c40 JJ |
8719 | default: gcc_unreachable (); |
8720 | } | |
8122fbff JJ |
8721 | /* If we drop the -Winvalid-memory-model warning for C++17 P0418R2, |
8722 | we can just return ret here unconditionally. Otherwise, work around | |
8723 | it here and make sure fail memmodel is not stronger. */ | |
8724 | if ((mo & OMP_FAIL_MEMORY_ORDER_MASK) == OMP_FAIL_MEMORY_ORDER_UNSPECIFIED) | |
8725 | return ret; | |
8726 | fail_ret = omp_memory_order_to_fail_memmodel (mo); | |
8727 | if (fail_ret > ret) | |
8728 | return fail_ret; | |
8729 | return ret; | |
28567c40 JJ |
8730 | } |
8731 | ||
629b3d75 MJ |
8732 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
8733 | operation as a normal volatile load. */ | |
8734 | ||
8735 | static bool | |
8736 | expand_omp_atomic_load (basic_block load_bb, tree addr, | |
8737 | tree loaded_val, int index) | |
8738 | { | |
8739 | enum built_in_function tmpbase; | |
8740 | gimple_stmt_iterator gsi; | |
8741 | basic_block store_bb; | |
8742 | location_t loc; | |
8743 | gimple *stmt; | |
9aa08cd4 | 8744 | tree decl, type, itype; |
629b3d75 | 8745 | |
65f4b875 | 8746 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
8747 | stmt = gsi_stmt (gsi); |
8748 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
8749 | loc = gimple_location (stmt); | |
8750 | ||
8751 | /* ??? If the target does not implement atomic_load_optab[mode], and mode | |
8752 | is smaller than word size, then expand_atomic_load assumes that the load | |
8753 | is atomic. We could avoid the builtin entirely in this case. */ | |
8754 | ||
8755 | tmpbase = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
8756 | decl = builtin_decl_explicit (tmpbase); | |
8757 | if (decl == NULL_TREE) | |
8758 | return false; | |
8759 | ||
8760 | type = TREE_TYPE (loaded_val); | |
8761 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
8762 | ||
28567c40 | 8763 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt); |
9aa08cd4 RB |
8764 | tree mo = build_int_cst (integer_type_node, |
8765 | omp_memory_order_to_memmodel (omo)); | |
8766 | gcall *call = gimple_build_call (decl, 2, addr, mo); | |
8767 | gimple_set_location (call, loc); | |
8768 | gimple_set_vuse (call, gimple_vuse (stmt)); | |
8769 | gimple *repl; | |
629b3d75 | 8770 | if (!useless_type_conversion_p (type, itype)) |
9aa08cd4 RB |
8771 | { |
8772 | tree lhs = make_ssa_name (itype); | |
8773 | gimple_call_set_lhs (call, lhs); | |
8774 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
8775 | repl = gimple_build_assign (loaded_val, | |
8776 | build1 (VIEW_CONVERT_EXPR, type, lhs)); | |
8777 | gimple_set_location (repl, loc); | |
8778 | } | |
8779 | else | |
8780 | { | |
8781 | gimple_call_set_lhs (call, loaded_val); | |
8782 | repl = call; | |
8783 | } | |
8784 | gsi_replace (&gsi, repl, true); | |
629b3d75 MJ |
8785 | |
8786 | store_bb = single_succ (load_bb); | |
65f4b875 | 8787 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
8788 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); |
8789 | gsi_remove (&gsi, true); | |
8790 | ||
629b3d75 MJ |
8791 | return true; |
8792 | } | |
8793 | ||
8794 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
8795 | operation as a normal volatile store. */ | |
8796 | ||
8797 | static bool | |
8798 | expand_omp_atomic_store (basic_block load_bb, tree addr, | |
8799 | tree loaded_val, tree stored_val, int index) | |
8800 | { | |
8801 | enum built_in_function tmpbase; | |
8802 | gimple_stmt_iterator gsi; | |
8803 | basic_block store_bb = single_succ (load_bb); | |
8804 | location_t loc; | |
8805 | gimple *stmt; | |
9aa08cd4 | 8806 | tree decl, type, itype; |
629b3d75 MJ |
8807 | machine_mode imode; |
8808 | bool exchange; | |
8809 | ||
65f4b875 | 8810 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
8811 | stmt = gsi_stmt (gsi); |
8812 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
8813 | ||
8814 | /* If the load value is needed, then this isn't a store but an exchange. */ | |
8815 | exchange = gimple_omp_atomic_need_value_p (stmt); | |
8816 | ||
65f4b875 | 8817 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
8818 | stmt = gsi_stmt (gsi); |
8819 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_STORE); | |
8820 | loc = gimple_location (stmt); | |
8821 | ||
8822 | /* ??? If the target does not implement atomic_store_optab[mode], and mode | |
8823 | is smaller than word size, then expand_atomic_store assumes that the store | |
8824 | is atomic. We could avoid the builtin entirely in this case. */ | |
8825 | ||
8826 | tmpbase = (exchange ? BUILT_IN_ATOMIC_EXCHANGE_N : BUILT_IN_ATOMIC_STORE_N); | |
8827 | tmpbase = (enum built_in_function) ((int) tmpbase + index + 1); | |
8828 | decl = builtin_decl_explicit (tmpbase); | |
8829 | if (decl == NULL_TREE) | |
8830 | return false; | |
8831 | ||
8832 | type = TREE_TYPE (stored_val); | |
8833 | ||
8834 | /* Dig out the type of the function's second argument. */ | |
8835 | itype = TREE_TYPE (decl); | |
8836 | itype = TYPE_ARG_TYPES (itype); | |
8837 | itype = TREE_CHAIN (itype); | |
8838 | itype = TREE_VALUE (itype); | |
8839 | imode = TYPE_MODE (itype); | |
8840 | ||
8841 | if (exchange && !can_atomic_exchange_p (imode, true)) | |
8842 | return false; | |
8843 | ||
8844 | if (!useless_type_conversion_p (itype, type)) | |
8845 | stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val); | |
28567c40 | 8846 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt); |
9aa08cd4 RB |
8847 | tree mo = build_int_cst (integer_type_node, |
8848 | omp_memory_order_to_memmodel (omo)); | |
8849 | stored_val = force_gimple_operand_gsi (&gsi, stored_val, true, NULL_TREE, | |
8850 | true, GSI_SAME_STMT); | |
8851 | gcall *call = gimple_build_call (decl, 3, addr, stored_val, mo); | |
8852 | gimple_set_location (call, loc); | |
8853 | gimple_set_vuse (call, gimple_vuse (stmt)); | |
8854 | gimple_set_vdef (call, gimple_vdef (stmt)); | |
8855 | ||
8856 | gimple *repl = call; | |
629b3d75 MJ |
8857 | if (exchange) |
8858 | { | |
8859 | if (!useless_type_conversion_p (type, itype)) | |
9aa08cd4 RB |
8860 | { |
8861 | tree lhs = make_ssa_name (itype); | |
8862 | gimple_call_set_lhs (call, lhs); | |
8863 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
8864 | repl = gimple_build_assign (loaded_val, | |
8865 | build1 (VIEW_CONVERT_EXPR, type, lhs)); | |
8866 | gimple_set_location (repl, loc); | |
8867 | } | |
8868 | else | |
8869 | gimple_call_set_lhs (call, loaded_val); | |
629b3d75 | 8870 | } |
9aa08cd4 | 8871 | gsi_replace (&gsi, repl, true); |
629b3d75 MJ |
8872 | |
8873 | /* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */ | |
65f4b875 | 8874 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
8875 | gsi_remove (&gsi, true); |
8876 | ||
629b3d75 MJ |
8877 | return true; |
8878 | } | |
8879 | ||
8880 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
8881 | operation as a __atomic_fetch_op builtin. INDEX is log2 of the | |
8882 | size of the data type, and thus usable to find the index of the builtin | |
8883 | decl. Returns false if the expression is not of the proper form. */ | |
8884 | ||
8885 | static bool | |
8886 | expand_omp_atomic_fetch_op (basic_block load_bb, | |
8887 | tree addr, tree loaded_val, | |
8888 | tree stored_val, int index) | |
8889 | { | |
8890 | enum built_in_function oldbase, newbase, tmpbase; | |
8891 | tree decl, itype, call; | |
8892 | tree lhs, rhs; | |
8893 | basic_block store_bb = single_succ (load_bb); | |
8894 | gimple_stmt_iterator gsi; | |
8895 | gimple *stmt; | |
8896 | location_t loc; | |
8897 | enum tree_code code; | |
8898 | bool need_old, need_new; | |
8899 | machine_mode imode; | |
629b3d75 MJ |
8900 | |
8901 | /* We expect to find the following sequences: | |
8902 | ||
8903 | load_bb: | |
8904 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) | |
8905 | ||
8906 | store_bb: | |
8907 | val = tmp OP something; (or: something OP tmp) | |
8908 | GIMPLE_OMP_STORE (val) | |
8909 | ||
8910 | ???FIXME: Allow a more flexible sequence. | |
8911 | Perhaps use data flow to pick the statements. | |
8912 | ||
8913 | */ | |
8914 | ||
8915 | gsi = gsi_after_labels (store_bb); | |
8916 | stmt = gsi_stmt (gsi); | |
65f4b875 AO |
8917 | if (is_gimple_debug (stmt)) |
8918 | { | |
8919 | gsi_next_nondebug (&gsi); | |
8920 | if (gsi_end_p (gsi)) | |
8921 | return false; | |
8922 | stmt = gsi_stmt (gsi); | |
8923 | } | |
629b3d75 MJ |
8924 | loc = gimple_location (stmt); |
8925 | if (!is_gimple_assign (stmt)) | |
8926 | return false; | |
65f4b875 | 8927 | gsi_next_nondebug (&gsi); |
629b3d75 MJ |
8928 | if (gimple_code (gsi_stmt (gsi)) != GIMPLE_OMP_ATOMIC_STORE) |
8929 | return false; | |
8930 | need_new = gimple_omp_atomic_need_value_p (gsi_stmt (gsi)); | |
fe8ac82f | 8931 | need_old = gimple_omp_atomic_need_value_p (last_nondebug_stmt (load_bb)); |
28567c40 | 8932 | enum omp_memory_order omo |
fe8ac82f | 8933 | = gimple_omp_atomic_memory_order (last_nondebug_stmt (load_bb)); |
28567c40 | 8934 | enum memmodel mo = omp_memory_order_to_memmodel (omo); |
629b3d75 MJ |
8935 | gcc_checking_assert (!need_old || !need_new); |
8936 | ||
8937 | if (!operand_equal_p (gimple_assign_lhs (stmt), stored_val, 0)) | |
8938 | return false; | |
8939 | ||
8940 | /* Check for one of the supported fetch-op operations. */ | |
8941 | code = gimple_assign_rhs_code (stmt); | |
8942 | switch (code) | |
8943 | { | |
8944 | case PLUS_EXPR: | |
8945 | case POINTER_PLUS_EXPR: | |
8946 | oldbase = BUILT_IN_ATOMIC_FETCH_ADD_N; | |
8947 | newbase = BUILT_IN_ATOMIC_ADD_FETCH_N; | |
8948 | break; | |
8949 | case MINUS_EXPR: | |
8950 | oldbase = BUILT_IN_ATOMIC_FETCH_SUB_N; | |
8951 | newbase = BUILT_IN_ATOMIC_SUB_FETCH_N; | |
8952 | break; | |
8953 | case BIT_AND_EXPR: | |
8954 | oldbase = BUILT_IN_ATOMIC_FETCH_AND_N; | |
8955 | newbase = BUILT_IN_ATOMIC_AND_FETCH_N; | |
8956 | break; | |
8957 | case BIT_IOR_EXPR: | |
8958 | oldbase = BUILT_IN_ATOMIC_FETCH_OR_N; | |
8959 | newbase = BUILT_IN_ATOMIC_OR_FETCH_N; | |
8960 | break; | |
8961 | case BIT_XOR_EXPR: | |
8962 | oldbase = BUILT_IN_ATOMIC_FETCH_XOR_N; | |
8963 | newbase = BUILT_IN_ATOMIC_XOR_FETCH_N; | |
8964 | break; | |
8965 | default: | |
8966 | return false; | |
8967 | } | |
8968 | ||
8969 | /* Make sure the expression is of the proper form. */ | |
8970 | if (operand_equal_p (gimple_assign_rhs1 (stmt), loaded_val, 0)) | |
8971 | rhs = gimple_assign_rhs2 (stmt); | |
8972 | else if (commutative_tree_code (gimple_assign_rhs_code (stmt)) | |
8973 | && operand_equal_p (gimple_assign_rhs2 (stmt), loaded_val, 0)) | |
8974 | rhs = gimple_assign_rhs1 (stmt); | |
8975 | else | |
8976 | return false; | |
8977 | ||
8978 | tmpbase = ((enum built_in_function) | |
8979 | ((need_new ? newbase : oldbase) + index + 1)); | |
8980 | decl = builtin_decl_explicit (tmpbase); | |
8981 | if (decl == NULL_TREE) | |
8982 | return false; | |
8983 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
8984 | imode = TYPE_MODE (itype); | |
8985 | ||
8986 | /* We could test all of the various optabs involved, but the fact of the | |
8987 | matter is that (with the exception of i486 vs i586 and xadd) all targets | |
8988 | that support any atomic operaton optab also implements compare-and-swap. | |
e53b6e56 | 8989 | Let optabs.cc take care of expanding any compare-and-swap loop. */ |
dc06356a | 8990 | if (!can_compare_and_swap_p (imode, true) || !can_atomic_load_p (imode)) |
629b3d75 MJ |
8991 | return false; |
8992 | ||
65f4b875 | 8993 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
8994 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_LOAD); |
8995 | ||
8996 | /* OpenMP does not imply any barrier-like semantics on its atomic ops. | |
8997 | It only requires that the operation happen atomically. Thus we can | |
8998 | use the RELAXED memory model. */ | |
8999 | call = build_call_expr_loc (loc, decl, 3, addr, | |
9000 | fold_convert_loc (loc, itype, rhs), | |
28567c40 | 9001 | build_int_cst (NULL, mo)); |
629b3d75 MJ |
9002 | |
9003 | if (need_old || need_new) | |
9004 | { | |
9005 | lhs = need_old ? loaded_val : stored_val; | |
9006 | call = fold_convert_loc (loc, TREE_TYPE (lhs), call); | |
9007 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, lhs, call); | |
9008 | } | |
9009 | else | |
9010 | call = fold_convert_loc (loc, void_type_node, call); | |
9011 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
9012 | gsi_remove (&gsi, true); | |
9013 | ||
65f4b875 | 9014 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
9015 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); |
9016 | gsi_remove (&gsi, true); | |
65f4b875 | 9017 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
9018 | stmt = gsi_stmt (gsi); |
9019 | gsi_remove (&gsi, true); | |
9020 | ||
9021 | if (gimple_in_ssa_p (cfun)) | |
9aa08cd4 | 9022 | release_defs (stmt); |
629b3d75 MJ |
9023 | |
9024 | return true; | |
9025 | } | |
9026 | ||
8122fbff JJ |
9027 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
9028 | compare and exchange as an ATOMIC_COMPARE_EXCHANGE internal function. | |
9029 | Returns false if the expression is not of the proper form. */ | |
9030 | ||
9031 | static bool | |
9032 | expand_omp_atomic_cas (basic_block load_bb, tree addr, | |
9033 | tree loaded_val, tree stored_val, int index) | |
9034 | { | |
9035 | /* We expect to find the following sequences: | |
9036 | ||
9037 | load_bb: | |
9038 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) | |
9039 | ||
9040 | store_bb: | |
9041 | val = tmp == e ? d : tmp; | |
9042 | GIMPLE_OMP_ATOMIC_STORE (val) | |
9043 | ||
9044 | or in store_bb instead: | |
9045 | tmp2 = tmp == e; | |
9046 | val = tmp2 ? d : tmp; | |
9047 | GIMPLE_OMP_ATOMIC_STORE (val) | |
9048 | ||
9049 | or: | |
9050 | tmp3 = VIEW_CONVERT_EXPR<integral_type>(tmp); | |
9051 | val = e == tmp3 ? d : tmp; | |
9052 | GIMPLE_OMP_ATOMIC_STORE (val) | |
9053 | ||
9054 | etc. */ | |
9055 | ||
9056 | ||
9057 | basic_block store_bb = single_succ (load_bb); | |
9058 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (store_bb); | |
9059 | gimple *store_stmt = gsi_stmt (gsi); | |
9060 | if (!store_stmt || gimple_code (store_stmt) != GIMPLE_OMP_ATOMIC_STORE) | |
9061 | return false; | |
9062 | gsi_prev_nondebug (&gsi); | |
9063 | if (gsi_end_p (gsi)) | |
9064 | return false; | |
9065 | gimple *condexpr_stmt = gsi_stmt (gsi); | |
9066 | if (!is_gimple_assign (condexpr_stmt) | |
9067 | || gimple_assign_rhs_code (condexpr_stmt) != COND_EXPR) | |
9068 | return false; | |
9069 | if (!operand_equal_p (gimple_assign_lhs (condexpr_stmt), stored_val, 0)) | |
9070 | return false; | |
9071 | gimple *cond_stmt = NULL; | |
9072 | gimple *vce_stmt = NULL; | |
9073 | gsi_prev_nondebug (&gsi); | |
9074 | if (!gsi_end_p (gsi)) | |
9075 | { | |
9076 | cond_stmt = gsi_stmt (gsi); | |
9077 | if (!is_gimple_assign (cond_stmt)) | |
9078 | return false; | |
9079 | if (gimple_assign_rhs_code (cond_stmt) == EQ_EXPR) | |
9080 | { | |
9081 | gsi_prev_nondebug (&gsi); | |
9082 | if (!gsi_end_p (gsi)) | |
9083 | { | |
9084 | vce_stmt = gsi_stmt (gsi); | |
9085 | if (!is_gimple_assign (vce_stmt) | |
9086 | || gimple_assign_rhs_code (vce_stmt) != VIEW_CONVERT_EXPR) | |
9087 | return false; | |
9088 | } | |
9089 | } | |
9090 | else if (gimple_assign_rhs_code (cond_stmt) == VIEW_CONVERT_EXPR) | |
9091 | std::swap (vce_stmt, cond_stmt); | |
9092 | else | |
9093 | return false; | |
9094 | if (vce_stmt) | |
9095 | { | |
9096 | tree vce_rhs = gimple_assign_rhs1 (vce_stmt); | |
9097 | if (TREE_CODE (vce_rhs) != VIEW_CONVERT_EXPR | |
9098 | || !operand_equal_p (TREE_OPERAND (vce_rhs, 0), loaded_val)) | |
9099 | return false; | |
9100 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vce_rhs)) | |
9101 | || !SCALAR_FLOAT_TYPE_P (TREE_TYPE (loaded_val)) | |
9102 | || !tree_int_cst_equal (TYPE_SIZE (TREE_TYPE (vce_rhs)), | |
9103 | TYPE_SIZE (TREE_TYPE (loaded_val)))) | |
9104 | return false; | |
9105 | gsi_prev_nondebug (&gsi); | |
9106 | if (!gsi_end_p (gsi)) | |
9107 | return false; | |
9108 | } | |
9109 | } | |
9110 | tree cond = gimple_assign_rhs1 (condexpr_stmt); | |
9111 | tree cond_op1, cond_op2; | |
9112 | if (cond_stmt) | |
9113 | { | |
68e00633 | 9114 | /* We should now always get a separate cond_stmt. */ |
8122fbff JJ |
9115 | if (!operand_equal_p (cond, gimple_assign_lhs (cond_stmt))) |
9116 | return false; | |
9117 | cond_op1 = gimple_assign_rhs1 (cond_stmt); | |
9118 | cond_op2 = gimple_assign_rhs2 (cond_stmt); | |
9119 | } | |
9120 | else if (TREE_CODE (cond) != EQ_EXPR && TREE_CODE (cond) != NE_EXPR) | |
9121 | return false; | |
9122 | else | |
9123 | { | |
9124 | cond_op1 = TREE_OPERAND (cond, 0); | |
9125 | cond_op2 = TREE_OPERAND (cond, 1); | |
9126 | } | |
9127 | tree d; | |
9128 | if (TREE_CODE (cond) == NE_EXPR) | |
9129 | { | |
9130 | if (!operand_equal_p (gimple_assign_rhs2 (condexpr_stmt), loaded_val)) | |
9131 | return false; | |
9132 | d = gimple_assign_rhs3 (condexpr_stmt); | |
9133 | } | |
9134 | else if (!operand_equal_p (gimple_assign_rhs3 (condexpr_stmt), loaded_val)) | |
9135 | return false; | |
9136 | else | |
9137 | d = gimple_assign_rhs2 (condexpr_stmt); | |
9138 | tree e = vce_stmt ? gimple_assign_lhs (vce_stmt) : loaded_val; | |
9139 | if (operand_equal_p (e, cond_op1)) | |
9140 | e = cond_op2; | |
9141 | else if (operand_equal_p (e, cond_op2)) | |
9142 | e = cond_op1; | |
9143 | else | |
9144 | return false; | |
9145 | ||
9146 | location_t loc = gimple_location (store_stmt); | |
fe8ac82f | 9147 | gimple *load_stmt = last_nondebug_stmt (load_bb); |
8122fbff JJ |
9148 | bool need_new = gimple_omp_atomic_need_value_p (store_stmt); |
9149 | bool need_old = gimple_omp_atomic_need_value_p (load_stmt); | |
9150 | bool weak = gimple_omp_atomic_weak_p (load_stmt); | |
9151 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (load_stmt); | |
9152 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); | |
9153 | tree fmo = build_int_cst (NULL, omp_memory_order_to_fail_memmodel (omo)); | |
9154 | gcc_checking_assert (!need_old || !need_new); | |
9155 | ||
9156 | enum built_in_function fncode | |
9157 | = (enum built_in_function) ((int) BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N | |
9158 | + index + 1); | |
9159 | tree cmpxchg = builtin_decl_explicit (fncode); | |
9160 | if (cmpxchg == NULL_TREE) | |
9161 | return false; | |
9162 | tree itype = TREE_TYPE (TREE_TYPE (cmpxchg)); | |
9163 | ||
9164 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true) | |
9165 | || !can_atomic_load_p (TYPE_MODE (itype))) | |
9166 | return false; | |
9167 | ||
9168 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); | |
9169 | if (SCALAR_FLOAT_TYPE_P (type) && !vce_stmt) | |
9170 | return false; | |
9171 | ||
9172 | gsi = gsi_for_stmt (store_stmt); | |
9173 | if (!useless_type_conversion_p (itype, TREE_TYPE (e))) | |
9174 | { | |
9175 | tree ne = create_tmp_reg (itype); | |
9176 | gimple *g = gimple_build_assign (ne, NOP_EXPR, e); | |
9177 | gimple_set_location (g, loc); | |
9178 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9179 | e = ne; | |
9180 | } | |
9181 | if (!useless_type_conversion_p (itype, TREE_TYPE (d))) | |
9182 | { | |
9183 | tree nd = create_tmp_reg (itype); | |
9184 | enum tree_code code; | |
9185 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (d))) | |
9186 | { | |
9187 | code = VIEW_CONVERT_EXPR; | |
9188 | d = build1 (VIEW_CONVERT_EXPR, itype, d); | |
9189 | } | |
9190 | else | |
9191 | code = NOP_EXPR; | |
9192 | gimple *g = gimple_build_assign (nd, code, d); | |
9193 | gimple_set_location (g, loc); | |
9194 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9195 | d = nd; | |
9196 | } | |
9197 | ||
9198 | tree ctype = build_complex_type (itype); | |
9199 | int flag = int_size_in_bytes (itype) + (weak ? 256 : 0); | |
9200 | gimple *g | |
9201 | = gimple_build_call_internal (IFN_ATOMIC_COMPARE_EXCHANGE, 6, addr, e, d, | |
9202 | build_int_cst (integer_type_node, flag), | |
9203 | mo, fmo); | |
9204 | tree cres = create_tmp_reg (ctype); | |
9205 | gimple_call_set_lhs (g, cres); | |
9206 | gimple_set_location (g, loc); | |
9207 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9208 | ||
9209 | if (cond_stmt || need_old || need_new) | |
9210 | { | |
9211 | tree im = create_tmp_reg (itype); | |
9212 | g = gimple_build_assign (im, IMAGPART_EXPR, | |
9213 | build1 (IMAGPART_EXPR, itype, cres)); | |
9214 | gimple_set_location (g, loc); | |
9215 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9216 | ||
9217 | tree re = NULL_TREE; | |
9218 | if (need_old || need_new) | |
9219 | { | |
9220 | re = create_tmp_reg (itype); | |
9221 | g = gimple_build_assign (re, REALPART_EXPR, | |
9222 | build1 (REALPART_EXPR, itype, cres)); | |
9223 | gimple_set_location (g, loc); | |
9224 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9225 | } | |
9226 | ||
9227 | if (cond_stmt) | |
9228 | { | |
060173dd | 9229 | g = gimple_build_assign (cond, NOP_EXPR, im); |
8122fbff JJ |
9230 | gimple_set_location (g, loc); |
9231 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9232 | } | |
060173dd RB |
9233 | |
9234 | if (need_new) | |
8122fbff JJ |
9235 | { |
9236 | g = gimple_build_assign (create_tmp_reg (itype), COND_EXPR, | |
060173dd RB |
9237 | cond_stmt |
9238 | ? cond : build2 (NE_EXPR, boolean_type_node, | |
9239 | im, build_zero_cst (itype)), | |
8122fbff JJ |
9240 | d, re); |
9241 | gimple_set_location (g, loc); | |
9242 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9243 | re = gimple_assign_lhs (g); | |
9244 | } | |
9245 | ||
9246 | if (need_old || need_new) | |
9247 | { | |
9248 | tree v = need_old ? loaded_val : stored_val; | |
9249 | enum tree_code code; | |
9250 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (v))) | |
9251 | { | |
9252 | code = VIEW_CONVERT_EXPR; | |
9253 | re = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (v), re); | |
9254 | } | |
9255 | else if (!useless_type_conversion_p (TREE_TYPE (v), itype)) | |
9256 | code = NOP_EXPR; | |
9257 | else | |
9258 | code = TREE_CODE (re); | |
9259 | g = gimple_build_assign (v, code, re); | |
9260 | gimple_set_location (g, loc); | |
9261 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
9262 | } | |
9263 | } | |
9264 | ||
9265 | gsi_remove (&gsi, true); | |
9266 | gsi = gsi_for_stmt (load_stmt); | |
9267 | gsi_remove (&gsi, true); | |
9268 | gsi = gsi_for_stmt (condexpr_stmt); | |
9269 | gsi_remove (&gsi, true); | |
9270 | if (cond_stmt) | |
9271 | { | |
9272 | gsi = gsi_for_stmt (cond_stmt); | |
9273 | gsi_remove (&gsi, true); | |
9274 | } | |
9275 | if (vce_stmt) | |
9276 | { | |
9277 | gsi = gsi_for_stmt (vce_stmt); | |
9278 | gsi_remove (&gsi, true); | |
9279 | } | |
9280 | ||
9281 | return true; | |
9282 | } | |
9283 | ||
629b3d75 MJ |
9284 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: |
9285 | ||
9286 | oldval = *addr; | |
9287 | repeat: | |
01914336 | 9288 | newval = rhs; // with oldval replacing *addr in rhs |
629b3d75 MJ |
9289 | oldval = __sync_val_compare_and_swap (addr, oldval, newval); |
9290 | if (oldval != newval) | |
9291 | goto repeat; | |
9292 | ||
9293 | INDEX is log2 of the size of the data type, and thus usable to find the | |
9294 | index of the builtin decl. */ | |
9295 | ||
9296 | static bool | |
9297 | expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb, | |
9298 | tree addr, tree loaded_val, tree stored_val, | |
9299 | int index) | |
9300 | { | |
9301 | tree loadedi, storedi, initial, new_storedi, old_vali; | |
b4e47472 | 9302 | tree type, itype, cmpxchg, iaddr, atype; |
629b3d75 MJ |
9303 | gimple_stmt_iterator si; |
9304 | basic_block loop_header = single_succ (load_bb); | |
9305 | gimple *phi, *stmt; | |
9306 | edge e; | |
9307 | enum built_in_function fncode; | |
9308 | ||
629b3d75 MJ |
9309 | fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N |
9310 | + index + 1); | |
9311 | cmpxchg = builtin_decl_explicit (fncode); | |
9312 | if (cmpxchg == NULL_TREE) | |
9313 | return false; | |
b4e47472 JJ |
9314 | type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); |
9315 | atype = type; | |
629b3d75 MJ |
9316 | itype = TREE_TYPE (TREE_TYPE (cmpxchg)); |
9317 | ||
dc06356a JJ |
9318 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true) |
9319 | || !can_atomic_load_p (TYPE_MODE (itype))) | |
629b3d75 MJ |
9320 | return false; |
9321 | ||
9322 | /* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */ | |
65f4b875 | 9323 | si = gsi_last_nondebug_bb (load_bb); |
629b3d75 | 9324 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); |
090f0d78 JJ |
9325 | location_t loc = gimple_location (gsi_stmt (si)); |
9326 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (gsi_stmt (si)); | |
8122fbff JJ |
9327 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); |
9328 | tree fmo = build_int_cst (NULL, omp_memory_order_to_fail_memmodel (omo)); | |
629b3d75 MJ |
9329 | |
9330 | /* For floating-point values, we'll need to view-convert them to integers | |
9331 | so that we can perform the atomic compare and swap. Simplify the | |
9332 | following code by always setting up the "i"ntegral variables. */ | |
9333 | if (!INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type)) | |
9334 | { | |
9335 | tree iaddr_val; | |
9336 | ||
9337 | iaddr = create_tmp_reg (build_pointer_type_for_mode (itype, ptr_mode, | |
9338 | true)); | |
b4e47472 | 9339 | atype = itype; |
629b3d75 MJ |
9340 | iaddr_val |
9341 | = force_gimple_operand_gsi (&si, | |
9342 | fold_convert (TREE_TYPE (iaddr), addr), | |
9343 | false, NULL_TREE, true, GSI_SAME_STMT); | |
9344 | stmt = gimple_build_assign (iaddr, iaddr_val); | |
9345 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9346 | loadedi = create_tmp_var (itype); | |
9347 | if (gimple_in_ssa_p (cfun)) | |
9348 | loadedi = make_ssa_name (loadedi); | |
9349 | } | |
9350 | else | |
9351 | { | |
9352 | iaddr = addr; | |
9353 | loadedi = loaded_val; | |
9354 | } | |
9355 | ||
9356 | fncode = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
9357 | tree loaddecl = builtin_decl_explicit (fncode); | |
9358 | if (loaddecl) | |
9359 | initial | |
b4e47472 | 9360 | = fold_convert (atype, |
629b3d75 MJ |
9361 | build_call_expr (loaddecl, 2, iaddr, |
9362 | build_int_cst (NULL_TREE, | |
9363 | MEMMODEL_RELAXED))); | |
9364 | else | |
b4e47472 JJ |
9365 | { |
9366 | tree off | |
9367 | = build_int_cst (build_pointer_type_for_mode (atype, ptr_mode, | |
9368 | true), 0); | |
9369 | initial = build2 (MEM_REF, atype, iaddr, off); | |
9370 | } | |
629b3d75 MJ |
9371 | |
9372 | initial | |
9373 | = force_gimple_operand_gsi (&si, initial, true, NULL_TREE, true, | |
9374 | GSI_SAME_STMT); | |
9375 | ||
9376 | /* Move the value to the LOADEDI temporary. */ | |
9377 | if (gimple_in_ssa_p (cfun)) | |
9378 | { | |
9379 | gcc_assert (gimple_seq_empty_p (phi_nodes (loop_header))); | |
9380 | phi = create_phi_node (loadedi, loop_header); | |
9381 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)), | |
9382 | initial); | |
9383 | } | |
9384 | else | |
9385 | gsi_insert_before (&si, | |
9386 | gimple_build_assign (loadedi, initial), | |
9387 | GSI_SAME_STMT); | |
9388 | if (loadedi != loaded_val) | |
9389 | { | |
9390 | gimple_stmt_iterator gsi2; | |
9391 | tree x; | |
9392 | ||
9393 | x = build1 (VIEW_CONVERT_EXPR, type, loadedi); | |
9394 | gsi2 = gsi_start_bb (loop_header); | |
9395 | if (gimple_in_ssa_p (cfun)) | |
9396 | { | |
9397 | gassign *stmt; | |
9398 | x = force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
9399 | true, GSI_SAME_STMT); | |
9400 | stmt = gimple_build_assign (loaded_val, x); | |
9401 | gsi_insert_before (&gsi2, stmt, GSI_SAME_STMT); | |
9402 | } | |
9403 | else | |
9404 | { | |
9405 | x = build2 (MODIFY_EXPR, TREE_TYPE (loaded_val), loaded_val, x); | |
9406 | force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
9407 | true, GSI_SAME_STMT); | |
9408 | } | |
9409 | } | |
9410 | gsi_remove (&si, true); | |
9411 | ||
65f4b875 | 9412 | si = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
9413 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); |
9414 | ||
9415 | if (iaddr == addr) | |
9416 | storedi = stored_val; | |
9417 | else | |
01914336 MJ |
9418 | storedi |
9419 | = force_gimple_operand_gsi (&si, | |
9420 | build1 (VIEW_CONVERT_EXPR, itype, | |
9421 | stored_val), true, NULL_TREE, true, | |
9422 | GSI_SAME_STMT); | |
629b3d75 MJ |
9423 | |
9424 | /* Build the compare&swap statement. */ | |
090f0d78 JJ |
9425 | tree ctype = build_complex_type (itype); |
9426 | int flag = int_size_in_bytes (itype); | |
9427 | new_storedi = build_call_expr_internal_loc (loc, IFN_ATOMIC_COMPARE_EXCHANGE, | |
9428 | ctype, 6, iaddr, loadedi, | |
9429 | storedi, | |
9430 | build_int_cst (integer_type_node, | |
9431 | flag), | |
9432 | mo, fmo); | |
9433 | new_storedi = build1 (REALPART_EXPR, itype, new_storedi); | |
629b3d75 MJ |
9434 | new_storedi = force_gimple_operand_gsi (&si, |
9435 | fold_convert (TREE_TYPE (loadedi), | |
9436 | new_storedi), | |
9437 | true, NULL_TREE, | |
9438 | true, GSI_SAME_STMT); | |
9439 | ||
9440 | if (gimple_in_ssa_p (cfun)) | |
9441 | old_vali = loadedi; | |
9442 | else | |
9443 | { | |
9444 | old_vali = create_tmp_var (TREE_TYPE (loadedi)); | |
9445 | stmt = gimple_build_assign (old_vali, loadedi); | |
9446 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9447 | ||
9448 | stmt = gimple_build_assign (loadedi, new_storedi); | |
9449 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
9450 | } | |
9451 | ||
9452 | /* Note that we always perform the comparison as an integer, even for | |
9453 | floating point. This allows the atomic operation to properly | |
9454 | succeed even with NaNs and -0.0. */ | |
01914336 MJ |
9455 | tree ne = build2 (NE_EXPR, boolean_type_node, new_storedi, old_vali); |
9456 | stmt = gimple_build_cond_empty (ne); | |
629b3d75 MJ |
9457 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9458 | ||
9459 | /* Update cfg. */ | |
9460 | e = single_succ_edge (store_bb); | |
9461 | e->flags &= ~EDGE_FALLTHRU; | |
9462 | e->flags |= EDGE_FALSE_VALUE; | |
357067f2 JH |
9463 | /* Expect no looping. */ |
9464 | e->probability = profile_probability::guessed_always (); | |
629b3d75 MJ |
9465 | |
9466 | e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE); | |
357067f2 | 9467 | e->probability = profile_probability::guessed_never (); |
629b3d75 MJ |
9468 | |
9469 | /* Copy the new value to loadedi (we already did that before the condition | |
9470 | if we are not in SSA). */ | |
9471 | if (gimple_in_ssa_p (cfun)) | |
9472 | { | |
9473 | phi = gimple_seq_first_stmt (phi_nodes (loop_header)); | |
9474 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_storedi); | |
9475 | } | |
9476 | ||
9477 | /* Remove GIMPLE_OMP_ATOMIC_STORE. */ | |
9aa08cd4 | 9478 | stmt = gsi_stmt (si); |
629b3d75 | 9479 | gsi_remove (&si, true); |
9aa08cd4 RB |
9480 | if (gimple_in_ssa_p (cfun)) |
9481 | release_defs (stmt); | |
629b3d75 | 9482 | |
99b1c316 | 9483 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
9484 | loop->header = loop_header; |
9485 | loop->latch = store_bb; | |
9486 | add_loop (loop, loop_header->loop_father); | |
9487 | ||
629b3d75 MJ |
9488 | return true; |
9489 | } | |
9490 | ||
9491 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
9492 | ||
01914336 MJ |
9493 | GOMP_atomic_start (); |
9494 | *addr = rhs; | |
9495 | GOMP_atomic_end (); | |
629b3d75 MJ |
9496 | |
9497 | The result is not globally atomic, but works so long as all parallel | |
9498 | references are within #pragma omp atomic directives. According to | |
9499 | responses received from omp@openmp.org, appears to be within spec. | |
9500 | Which makes sense, since that's how several other compilers handle | |
9501 | this situation as well. | |
9502 | LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're | |
9503 | expanding. STORED_VAL is the operand of the matching | |
9504 | GIMPLE_OMP_ATOMIC_STORE. | |
9505 | ||
9506 | We replace | |
9507 | GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with | |
9508 | loaded_val = *addr; | |
9509 | ||
9510 | and replace | |
9511 | GIMPLE_OMP_ATOMIC_STORE (stored_val) with | |
9512 | *addr = stored_val; | |
9513 | */ | |
9514 | ||
9515 | static bool | |
9516 | expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb, | |
9517 | tree addr, tree loaded_val, tree stored_val) | |
9518 | { | |
9519 | gimple_stmt_iterator si; | |
9520 | gassign *stmt; | |
9521 | tree t; | |
9522 | ||
65f4b875 | 9523 | si = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
9524 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); |
9525 | ||
9526 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START); | |
9527 | t = build_call_expr (t, 0); | |
9528 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
9529 | ||
b4e47472 JJ |
9530 | tree mem = build_simple_mem_ref (addr); |
9531 | TREE_TYPE (mem) = TREE_TYPE (loaded_val); | |
9532 | TREE_OPERAND (mem, 1) | |
9533 | = fold_convert (build_pointer_type_for_mode (TREE_TYPE (mem), ptr_mode, | |
9534 | true), | |
9535 | TREE_OPERAND (mem, 1)); | |
9536 | stmt = gimple_build_assign (loaded_val, mem); | |
629b3d75 MJ |
9537 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9538 | gsi_remove (&si, true); | |
9539 | ||
65f4b875 | 9540 | si = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
9541 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); |
9542 | ||
b4e47472 | 9543 | stmt = gimple_build_assign (unshare_expr (mem), stored_val); |
9aa08cd4 RB |
9544 | gimple_set_vuse (stmt, gimple_vuse (gsi_stmt (si))); |
9545 | gimple_set_vdef (stmt, gimple_vdef (gsi_stmt (si))); | |
629b3d75 MJ |
9546 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9547 | ||
9548 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END); | |
9549 | t = build_call_expr (t, 0); | |
9550 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
9551 | gsi_remove (&si, true); | |
629b3d75 MJ |
9552 | return true; |
9553 | } | |
9554 | ||
9555 | /* Expand an GIMPLE_OMP_ATOMIC statement. We try to expand | |
01914336 | 9556 | using expand_omp_atomic_fetch_op. If it failed, we try to |
629b3d75 MJ |
9557 | call expand_omp_atomic_pipeline, and if it fails too, the |
9558 | ultimate fallback is wrapping the operation in a mutex | |
9559 | (expand_omp_atomic_mutex). REGION is the atomic region built | |
9560 | by build_omp_regions_1(). */ | |
9561 | ||
9562 | static void | |
9563 | expand_omp_atomic (struct omp_region *region) | |
9564 | { | |
9565 | basic_block load_bb = region->entry, store_bb = region->exit; | |
fe8ac82f RB |
9566 | gomp_atomic_load *load |
9567 | = as_a <gomp_atomic_load *> (last_nondebug_stmt (load_bb)); | |
9568 | gomp_atomic_store *store | |
9569 | = as_a <gomp_atomic_store *> (last_nondebug_stmt (store_bb)); | |
629b3d75 MJ |
9570 | tree loaded_val = gimple_omp_atomic_load_lhs (load); |
9571 | tree addr = gimple_omp_atomic_load_rhs (load); | |
9572 | tree stored_val = gimple_omp_atomic_store_val (store); | |
b4e47472 | 9573 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); |
629b3d75 MJ |
9574 | HOST_WIDE_INT index; |
9575 | ||
9576 | /* Make sure the type is one of the supported sizes. */ | |
9577 | index = tree_to_uhwi (TYPE_SIZE_UNIT (type)); | |
9578 | index = exact_log2 (index); | |
9579 | if (index >= 0 && index <= 4) | |
9580 | { | |
9581 | unsigned int align = TYPE_ALIGN_UNIT (type); | |
9582 | ||
9583 | /* __sync builtins require strict data alignment. */ | |
9584 | if (exact_log2 (align) >= index) | |
9585 | { | |
9586 | /* Atomic load. */ | |
3bd8f481 | 9587 | scalar_mode smode; |
629b3d75 | 9588 | if (loaded_val == stored_val |
3bd8f481 RS |
9589 | && (is_int_mode (TYPE_MODE (type), &smode) |
9590 | || is_float_mode (TYPE_MODE (type), &smode)) | |
9591 | && GET_MODE_BITSIZE (smode) <= BITS_PER_WORD | |
629b3d75 MJ |
9592 | && expand_omp_atomic_load (load_bb, addr, loaded_val, index)) |
9593 | return; | |
9594 | ||
9595 | /* Atomic store. */ | |
3bd8f481 RS |
9596 | if ((is_int_mode (TYPE_MODE (type), &smode) |
9597 | || is_float_mode (TYPE_MODE (type), &smode)) | |
9598 | && GET_MODE_BITSIZE (smode) <= BITS_PER_WORD | |
629b3d75 MJ |
9599 | && store_bb == single_succ (load_bb) |
9600 | && first_stmt (store_bb) == store | |
9601 | && expand_omp_atomic_store (load_bb, addr, loaded_val, | |
9602 | stored_val, index)) | |
9603 | return; | |
9604 | ||
9605 | /* When possible, use specialized atomic update functions. */ | |
9606 | if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type)) | |
9607 | && store_bb == single_succ (load_bb) | |
9608 | && expand_omp_atomic_fetch_op (load_bb, addr, | |
9609 | loaded_val, stored_val, index)) | |
9610 | return; | |
9611 | ||
8122fbff JJ |
9612 | /* When possible, use ATOMIC_COMPARE_EXCHANGE ifn without a loop. */ |
9613 | if (store_bb == single_succ (load_bb) | |
9614 | && !gimple_in_ssa_p (cfun) | |
9615 | && expand_omp_atomic_cas (load_bb, addr, loaded_val, stored_val, | |
9616 | index)) | |
9617 | return; | |
9618 | ||
629b3d75 MJ |
9619 | /* If we don't have specialized __sync builtins, try and implement |
9620 | as a compare and swap loop. */ | |
9621 | if (expand_omp_atomic_pipeline (load_bb, store_bb, addr, | |
9622 | loaded_val, stored_val, index)) | |
9623 | return; | |
9624 | } | |
9625 | } | |
9626 | ||
9627 | /* The ultimate fallback is wrapping the operation in a mutex. */ | |
9628 | expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val); | |
9629 | } | |
9630 | ||
9631 | /* Mark the loops inside the kernels region starting at REGION_ENTRY and ending | |
9632 | at REGION_EXIT. */ | |
9633 | ||
9634 | static void | |
9635 | mark_loops_in_oacc_kernels_region (basic_block region_entry, | |
9636 | basic_block region_exit) | |
9637 | { | |
99b1c316 | 9638 | class loop *outer = region_entry->loop_father; |
629b3d75 MJ |
9639 | gcc_assert (region_exit == NULL || outer == region_exit->loop_father); |
9640 | ||
9641 | /* Don't parallelize the kernels region if it contains more than one outer | |
9642 | loop. */ | |
9643 | unsigned int nr_outer_loops = 0; | |
99b1c316 MS |
9644 | class loop *single_outer = NULL; |
9645 | for (class loop *loop = outer->inner; loop != NULL; loop = loop->next) | |
629b3d75 MJ |
9646 | { |
9647 | gcc_assert (loop_outer (loop) == outer); | |
9648 | ||
9649 | if (!dominated_by_p (CDI_DOMINATORS, loop->header, region_entry)) | |
9650 | continue; | |
9651 | ||
9652 | if (region_exit != NULL | |
9653 | && dominated_by_p (CDI_DOMINATORS, loop->header, region_exit)) | |
9654 | continue; | |
9655 | ||
9656 | nr_outer_loops++; | |
9657 | single_outer = loop; | |
9658 | } | |
9659 | if (nr_outer_loops != 1) | |
9660 | return; | |
9661 | ||
99b1c316 | 9662 | for (class loop *loop = single_outer->inner; |
01914336 MJ |
9663 | loop != NULL; |
9664 | loop = loop->inner) | |
629b3d75 MJ |
9665 | if (loop->next) |
9666 | return; | |
9667 | ||
9668 | /* Mark the loops in the region. */ | |
99b1c316 | 9669 | for (class loop *loop = single_outer; loop != NULL; loop = loop->inner) |
629b3d75 MJ |
9670 | loop->in_oacc_kernels_region = true; |
9671 | } | |
9672 | ||
629b3d75 MJ |
9673 | /* Build target argument identifier from the DEVICE identifier, value |
9674 | identifier ID and whether the element also has a SUBSEQUENT_PARAM. */ | |
9675 | ||
9676 | static tree | |
9677 | get_target_argument_identifier_1 (int device, bool subseqent_param, int id) | |
9678 | { | |
9679 | tree t = build_int_cst (integer_type_node, device); | |
9680 | if (subseqent_param) | |
9681 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
9682 | build_int_cst (integer_type_node, | |
9683 | GOMP_TARGET_ARG_SUBSEQUENT_PARAM)); | |
9684 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
9685 | build_int_cst (integer_type_node, id)); | |
9686 | return t; | |
9687 | } | |
9688 | ||
9689 | /* Like above but return it in type that can be directly stored as an element | |
9690 | of the argument array. */ | |
9691 | ||
9692 | static tree | |
9693 | get_target_argument_identifier (int device, bool subseqent_param, int id) | |
9694 | { | |
9695 | tree t = get_target_argument_identifier_1 (device, subseqent_param, id); | |
9696 | return fold_convert (ptr_type_node, t); | |
9697 | } | |
9698 | ||
9699 | /* Return a target argument consisting of DEVICE identifier, value identifier | |
9700 | ID, and the actual VALUE. */ | |
9701 | ||
9702 | static tree | |
9703 | get_target_argument_value (gimple_stmt_iterator *gsi, int device, int id, | |
9704 | tree value) | |
9705 | { | |
9706 | tree t = fold_build2 (LSHIFT_EXPR, integer_type_node, | |
9707 | fold_convert (integer_type_node, value), | |
9708 | build_int_cst (unsigned_type_node, | |
9709 | GOMP_TARGET_ARG_VALUE_SHIFT)); | |
9710 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
9711 | get_target_argument_identifier_1 (device, false, id)); | |
9712 | t = fold_convert (ptr_type_node, t); | |
9713 | return force_gimple_operand_gsi (gsi, t, true, NULL, true, GSI_SAME_STMT); | |
9714 | } | |
9715 | ||
9716 | /* If VALUE is an integer constant greater than -2^15 and smaller than 2^15, | |
9717 | push one argument to ARGS with both the DEVICE, ID and VALUE embedded in it, | |
9718 | otherwise push an identifier (with DEVICE and ID) and the VALUE in two | |
9719 | arguments. */ | |
9720 | ||
9721 | static void | |
9722 | push_target_argument_according_to_value (gimple_stmt_iterator *gsi, int device, | |
9723 | int id, tree value, vec <tree> *args) | |
9724 | { | |
9725 | if (tree_fits_shwi_p (value) | |
9726 | && tree_to_shwi (value) > -(1 << 15) | |
9727 | && tree_to_shwi (value) < (1 << 15)) | |
9728 | args->quick_push (get_target_argument_value (gsi, device, id, value)); | |
9729 | else | |
9730 | { | |
9731 | args->quick_push (get_target_argument_identifier (device, true, id)); | |
9732 | value = fold_convert (ptr_type_node, value); | |
9733 | value = force_gimple_operand_gsi (gsi, value, true, NULL, true, | |
9734 | GSI_SAME_STMT); | |
9735 | args->quick_push (value); | |
9736 | } | |
9737 | } | |
9738 | ||
01914336 | 9739 | /* Create an array of arguments that is then passed to GOMP_target. */ |
629b3d75 MJ |
9740 | |
9741 | static tree | |
9742 | get_target_arguments (gimple_stmt_iterator *gsi, gomp_target *tgt_stmt) | |
9743 | { | |
9744 | auto_vec <tree, 6> args; | |
9745 | tree clauses = gimple_omp_target_clauses (tgt_stmt); | |
9746 | tree t, c = omp_find_clause (clauses, OMP_CLAUSE_NUM_TEAMS); | |
9747 | if (c) | |
48d7327f | 9748 | t = OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR (c); |
629b3d75 MJ |
9749 | else |
9750 | t = integer_minus_one_node; | |
9751 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, | |
9752 | GOMP_TARGET_ARG_NUM_TEAMS, t, &args); | |
9753 | ||
9754 | c = omp_find_clause (clauses, OMP_CLAUSE_THREAD_LIMIT); | |
9755 | if (c) | |
9756 | t = OMP_CLAUSE_THREAD_LIMIT_EXPR (c); | |
9757 | else | |
9758 | t = integer_minus_one_node; | |
9759 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, | |
9760 | GOMP_TARGET_ARG_THREAD_LIMIT, t, | |
9761 | &args); | |
9762 | ||
629b3d75 MJ |
9763 | /* Produce more, perhaps device specific, arguments here. */ |
9764 | ||
9765 | tree argarray = create_tmp_var (build_array_type_nelts (ptr_type_node, | |
9766 | args.length () + 1), | |
9767 | ".omp_target_args"); | |
9768 | for (unsigned i = 0; i < args.length (); i++) | |
9769 | { | |
9770 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, | |
9771 | build_int_cst (integer_type_node, i), | |
9772 | NULL_TREE, NULL_TREE); | |
9773 | gsi_insert_before (gsi, gimple_build_assign (ref, args[i]), | |
9774 | GSI_SAME_STMT); | |
9775 | } | |
9776 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, | |
9777 | build_int_cst (integer_type_node, args.length ()), | |
9778 | NULL_TREE, NULL_TREE); | |
9779 | gsi_insert_before (gsi, gimple_build_assign (ref, null_pointer_node), | |
9780 | GSI_SAME_STMT); | |
9781 | TREE_ADDRESSABLE (argarray) = 1; | |
9782 | return build_fold_addr_expr (argarray); | |
9783 | } | |
9784 | ||
9785 | /* Expand the GIMPLE_OMP_TARGET starting at REGION. */ | |
9786 | ||
9787 | static void | |
9788 | expand_omp_target (struct omp_region *region) | |
9789 | { | |
9790 | basic_block entry_bb, exit_bb, new_bb; | |
9791 | struct function *child_cfun; | |
d6621a2f | 9792 | tree child_fn, child_fn2, block, t, c; |
629b3d75 MJ |
9793 | gimple_stmt_iterator gsi; |
9794 | gomp_target *entry_stmt; | |
9795 | gimple *stmt; | |
9796 | edge e; | |
cc9b9c0b | 9797 | bool offloaded; |
62aee289 | 9798 | int target_kind; |
629b3d75 | 9799 | |
fe8ac82f | 9800 | entry_stmt = as_a <gomp_target *> (last_nondebug_stmt (region->entry)); |
62aee289 | 9801 | target_kind = gimple_omp_target_kind (entry_stmt); |
629b3d75 MJ |
9802 | new_bb = region->entry; |
9803 | ||
9804 | offloaded = is_gimple_omp_offloaded (entry_stmt); | |
62aee289 | 9805 | switch (target_kind) |
629b3d75 MJ |
9806 | { |
9807 | case GF_OMP_TARGET_KIND_REGION: | |
9808 | case GF_OMP_TARGET_KIND_UPDATE: | |
9809 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
9810 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
9811 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
9812 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
62aee289 | 9813 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
629b3d75 | 9814 | case GF_OMP_TARGET_KIND_OACC_UPDATE: |
7aefef31 AS |
9815 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
9816 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
629b3d75 | 9817 | case GF_OMP_TARGET_KIND_OACC_DECLARE: |
e898ce79 GB |
9818 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
9819 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
629b3d75 MJ |
9820 | case GF_OMP_TARGET_KIND_DATA: |
9821 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
9822 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
e898ce79 | 9823 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: |
629b3d75 MJ |
9824 | break; |
9825 | default: | |
9826 | gcc_unreachable (); | |
9827 | } | |
9828 | ||
d6621a2f TB |
9829 | tree clauses = gimple_omp_target_clauses (entry_stmt); |
9830 | ||
9831 | bool is_ancestor = false; | |
9832 | child_fn = child_fn2 = NULL_TREE; | |
629b3d75 MJ |
9833 | child_cfun = NULL; |
9834 | if (offloaded) | |
9835 | { | |
d6621a2f TB |
9836 | c = omp_find_clause (clauses, OMP_CLAUSE_DEVICE); |
9837 | if (ENABLE_OFFLOADING && c) | |
9838 | is_ancestor = OMP_CLAUSE_DEVICE_ANCESTOR (c); | |
629b3d75 MJ |
9839 | child_fn = gimple_omp_target_child_fn (entry_stmt); |
9840 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
9841 | } | |
9842 | ||
9843 | /* Supported by expand_omp_taskreg, but not here. */ | |
9844 | if (child_cfun != NULL) | |
9845 | gcc_checking_assert (!child_cfun->cfg); | |
9846 | gcc_checking_assert (!gimple_in_ssa_p (cfun)); | |
9847 | ||
9848 | entry_bb = region->entry; | |
9849 | exit_bb = region->exit; | |
9850 | ||
703e4f86 TS |
9851 | if (target_kind == GF_OMP_TARGET_KIND_OACC_KERNELS) |
9852 | mark_loops_in_oacc_kernels_region (region->entry, region->exit); | |
9853 | ||
9854 | /* Going on, all OpenACC compute constructs are mapped to | |
9855 | 'BUILT_IN_GOACC_PARALLEL', and get their compute regions outlined. | |
9856 | To distinguish between them, we attach attributes. */ | |
62aee289 | 9857 | switch (target_kind) |
25651634 | 9858 | { |
703e4f86 TS |
9859 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
9860 | DECL_ATTRIBUTES (child_fn) | |
9861 | = tree_cons (get_identifier ("oacc parallel"), | |
9862 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9863 | break; | |
62aee289 | 9864 | case GF_OMP_TARGET_KIND_OACC_KERNELS: |
25651634 TS |
9865 | DECL_ATTRIBUTES (child_fn) |
9866 | = tree_cons (get_identifier ("oacc kernels"), | |
9867 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
62aee289 MR |
9868 | break; |
9869 | case GF_OMP_TARGET_KIND_OACC_SERIAL: | |
62aee289 MR |
9870 | DECL_ATTRIBUTES (child_fn) |
9871 | = tree_cons (get_identifier ("oacc serial"), | |
9872 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9873 | break; | |
e898ce79 GB |
9874 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
9875 | DECL_ATTRIBUTES (child_fn) | |
9876 | = tree_cons (get_identifier ("oacc parallel_kernels_parallelized"), | |
9877 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9878 | break; | |
9879 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
9880 | DECL_ATTRIBUTES (child_fn) | |
9881 | = tree_cons (get_identifier ("oacc parallel_kernels_gang_single"), | |
9882 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
9883 | break; | |
62aee289 | 9884 | default: |
703e4f86 TS |
9885 | /* Make sure we don't miss any. */ |
9886 | gcc_checking_assert (!(is_gimple_omp_oacc (entry_stmt) | |
9887 | && is_gimple_omp_offloaded (entry_stmt))); | |
62aee289 | 9888 | break; |
25651634 | 9889 | } |
629b3d75 MJ |
9890 | |
9891 | if (offloaded) | |
9892 | { | |
9893 | unsigned srcidx, dstidx, num; | |
9894 | ||
9895 | /* If the offloading region needs data sent from the parent | |
9896 | function, then the very first statement (except possible | |
9897 | tree profile counter updates) of the offloading body | |
9898 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
9899 | &.OMP_DATA_O is passed as an argument to the child function, | |
9900 | we need to replace it with the argument as seen by the child | |
9901 | function. | |
9902 | ||
9903 | In most cases, this will end up being the identity assignment | |
9904 | .OMP_DATA_I = .OMP_DATA_I. However, if the offloading body had | |
9905 | a function call that has been inlined, the original PARM_DECL | |
9906 | .OMP_DATA_I may have been converted into a different local | |
9907 | variable. In which case, we need to keep the assignment. */ | |
9908 | tree data_arg = gimple_omp_target_data_arg (entry_stmt); | |
9909 | if (data_arg) | |
9910 | { | |
9911 | basic_block entry_succ_bb = single_succ (entry_bb); | |
9912 | gimple_stmt_iterator gsi; | |
9913 | tree arg; | |
9914 | gimple *tgtcopy_stmt = NULL; | |
9915 | tree sender = TREE_VEC_ELT (data_arg, 0); | |
9916 | ||
9917 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
9918 | { | |
9919 | gcc_assert (!gsi_end_p (gsi)); | |
9920 | stmt = gsi_stmt (gsi); | |
9921 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
9922 | continue; | |
9923 | ||
9924 | if (gimple_num_ops (stmt) == 2) | |
9925 | { | |
9926 | tree arg = gimple_assign_rhs1 (stmt); | |
9927 | ||
9928 | /* We're ignoring the subcode because we're | |
9929 | effectively doing a STRIP_NOPS. */ | |
9930 | ||
9931 | if (TREE_CODE (arg) == ADDR_EXPR | |
9932 | && TREE_OPERAND (arg, 0) == sender) | |
9933 | { | |
9934 | tgtcopy_stmt = stmt; | |
9935 | break; | |
9936 | } | |
9937 | } | |
9938 | } | |
9939 | ||
9940 | gcc_assert (tgtcopy_stmt != NULL); | |
9941 | arg = DECL_ARGUMENTS (child_fn); | |
9942 | ||
9943 | gcc_assert (gimple_assign_lhs (tgtcopy_stmt) == arg); | |
9944 | gsi_remove (&gsi, true); | |
9945 | } | |
9946 | ||
9947 | /* Declare local variables needed in CHILD_CFUN. */ | |
9948 | block = DECL_INITIAL (child_fn); | |
9949 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
9950 | /* The gimplifier could record temporaries in the offloading block | |
9951 | rather than in containing function's local_decls chain, | |
9952 | which would mean cgraph missed finalizing them. Do it now. */ | |
9953 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
9954 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) | |
9955 | varpool_node::finalize_decl (t); | |
9956 | DECL_SAVED_TREE (child_fn) = NULL; | |
9957 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
9958 | gimple_set_body (child_fn, NULL); | |
9959 | TREE_USED (block) = 1; | |
9960 | ||
9961 | /* Reset DECL_CONTEXT on function arguments. */ | |
9962 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
9963 | DECL_CONTEXT (t) = child_fn; | |
9964 | ||
9965 | /* Split ENTRY_BB at GIMPLE_*, | |
9966 | so that it can be moved to the child function. */ | |
65f4b875 | 9967 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
9968 | stmt = gsi_stmt (gsi); |
9969 | gcc_assert (stmt | |
9970 | && gimple_code (stmt) == gimple_code (entry_stmt)); | |
9971 | e = split_block (entry_bb, stmt); | |
9972 | gsi_remove (&gsi, true); | |
9973 | entry_bb = e->dest; | |
9974 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
9975 | ||
9976 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ | |
9977 | if (exit_bb) | |
9978 | { | |
65f4b875 | 9979 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
9980 | gcc_assert (!gsi_end_p (gsi) |
9981 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
9982 | stmt = gimple_build_return (NULL); | |
9983 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
9984 | gsi_remove (&gsi, true); | |
9985 | } | |
9986 | ||
9987 | /* Move the offloading region into CHILD_CFUN. */ | |
9988 | ||
9989 | block = gimple_block (entry_stmt); | |
9990 | ||
9991 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
9992 | if (exit_bb) | |
9993 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
9994 | /* When the OMP expansion process cannot guarantee an up-to-date | |
9995 | loop tree arrange for the child function to fixup loops. */ | |
9996 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
9997 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
9998 | ||
9999 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
10000 | num = vec_safe_length (child_cfun->local_decls); | |
10001 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
10002 | { | |
10003 | t = (*child_cfun->local_decls)[srcidx]; | |
10004 | if (DECL_CONTEXT (t) == cfun->decl) | |
10005 | continue; | |
10006 | if (srcidx != dstidx) | |
10007 | (*child_cfun->local_decls)[dstidx] = t; | |
10008 | dstidx++; | |
10009 | } | |
10010 | if (dstidx != num) | |
10011 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
10012 | ||
10013 | /* Inform the callgraph about the new function. */ | |
10014 | child_cfun->curr_properties = cfun->curr_properties; | |
10015 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; | |
10016 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; | |
10017 | cgraph_node *node = cgraph_node::get_create (child_fn); | |
10018 | node->parallelized_function = 1; | |
10019 | cgraph_node::add_new_function (child_fn, true); | |
10020 | ||
10021 | /* Add the new function to the offload table. */ | |
10022 | if (ENABLE_OFFLOADING) | |
60bf575c TV |
10023 | { |
10024 | if (in_lto_p) | |
10025 | DECL_PRESERVE_P (child_fn) = 1; | |
d6621a2f TB |
10026 | if (!is_ancestor) |
10027 | vec_safe_push (offload_funcs, child_fn); | |
60bf575c | 10028 | } |
629b3d75 MJ |
10029 | |
10030 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) | |
10031 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); | |
10032 | ||
10033 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
10034 | fixed in a following pass. */ | |
10035 | push_cfun (child_cfun); | |
10036 | if (need_asm) | |
9579db35 | 10037 | assign_assembler_name_if_needed (child_fn); |
629b3d75 MJ |
10038 | cgraph_edge::rebuild_edges (); |
10039 | ||
10040 | /* Some EH regions might become dead, see PR34608. If | |
10041 | pass_cleanup_cfg isn't the first pass to happen with the | |
10042 | new child, these dead EH edges might cause problems. | |
10043 | Clean them up now. */ | |
10044 | if (flag_exceptions) | |
10045 | { | |
10046 | basic_block bb; | |
10047 | bool changed = false; | |
10048 | ||
10049 | FOR_EACH_BB_FN (bb, cfun) | |
10050 | changed |= gimple_purge_dead_eh_edges (bb); | |
10051 | if (changed) | |
10052 | cleanup_tree_cfg (); | |
10053 | } | |
10054 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
10055 | verify_loop_structure (); | |
10056 | pop_cfun (); | |
10057 | ||
10058 | if (dump_file && !gimple_in_ssa_p (cfun)) | |
10059 | { | |
10060 | omp_any_child_fn_dumped = true; | |
10061 | dump_function_header (dump_file, child_fn, dump_flags); | |
10062 | dump_function_to_file (child_fn, dump_file, dump_flags); | |
10063 | } | |
4ccc4e30 JJ |
10064 | |
10065 | adjust_context_and_scope (region, gimple_block (entry_stmt), child_fn); | |
d6621a2f TB |
10066 | |
10067 | /* Handle the case that an inner ancestor:1 target is called by an outer | |
10068 | target region. */ | |
178ac530 | 10069 | if (is_ancestor) |
d6621a2f | 10070 | { |
d6621a2f TB |
10071 | cgraph_node *fn2_node; |
10072 | child_fn2 = build_decl (DECL_SOURCE_LOCATION (child_fn), | |
10073 | FUNCTION_DECL, | |
10074 | clone_function_name (child_fn, "nohost"), | |
10075 | TREE_TYPE (child_fn)); | |
10076 | if (in_lto_p) | |
10077 | DECL_PRESERVE_P (child_fn2) = 1; | |
10078 | TREE_STATIC (child_fn2) = 1; | |
10079 | DECL_ARTIFICIAL (child_fn2) = 1; | |
10080 | DECL_IGNORED_P (child_fn2) = 0; | |
10081 | TREE_PUBLIC (child_fn2) = 0; | |
10082 | DECL_UNINLINABLE (child_fn2) = 1; | |
10083 | DECL_EXTERNAL (child_fn2) = 0; | |
178ac530 | 10084 | DECL_CONTEXT (child_fn2) = DECL_CONTEXT (child_fn); |
d6621a2f TB |
10085 | DECL_INITIAL (child_fn2) = make_node (BLOCK); |
10086 | BLOCK_SUPERCONTEXT (DECL_INITIAL (child_fn2)) = child_fn2; | |
10087 | DECL_ATTRIBUTES (child_fn) | |
10088 | = remove_attribute ("omp target entrypoint", | |
10089 | DECL_ATTRIBUTES (child_fn)); | |
10090 | DECL_ATTRIBUTES (child_fn2) | |
10091 | = tree_cons (get_identifier ("omp target device_ancestor_nohost"), | |
10092 | NULL_TREE, copy_list (DECL_ATTRIBUTES (child_fn))); | |
10093 | DECL_ATTRIBUTES (child_fn) | |
10094 | = tree_cons (get_identifier ("omp target device_ancestor_host"), | |
10095 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
10096 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (child_fn2) | |
10097 | = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (current_function_decl); | |
10098 | DECL_FUNCTION_SPECIFIC_TARGET (child_fn2) | |
10099 | = DECL_FUNCTION_SPECIFIC_TARGET (current_function_decl); | |
10100 | DECL_FUNCTION_VERSIONED (child_fn2) | |
10101 | = DECL_FUNCTION_VERSIONED (current_function_decl); | |
10102 | ||
10103 | fn2_node = cgraph_node::get_create (child_fn2); | |
10104 | fn2_node->offloadable = 1; | |
10105 | fn2_node->force_output = 1; | |
10106 | node->offloadable = 0; | |
10107 | ||
178ac530 TB |
10108 | /* Enable pass_omp_device_lower pass. */ |
10109 | fn2_node = cgraph_node::get (DECL_CONTEXT (child_fn)); | |
10110 | fn2_node->calls_declare_variant_alt = 1; | |
10111 | ||
d6621a2f TB |
10112 | t = build_decl (DECL_SOURCE_LOCATION (child_fn), |
10113 | RESULT_DECL, NULL_TREE, void_type_node); | |
10114 | DECL_ARTIFICIAL (t) = 1; | |
10115 | DECL_IGNORED_P (t) = 1; | |
10116 | DECL_CONTEXT (t) = child_fn2; | |
10117 | DECL_RESULT (child_fn2) = t; | |
10118 | DECL_SAVED_TREE (child_fn2) = build1 (RETURN_EXPR, | |
10119 | void_type_node, NULL); | |
10120 | tree tmp = DECL_ARGUMENTS (child_fn); | |
10121 | t = build_decl (DECL_SOURCE_LOCATION (child_fn), PARM_DECL, | |
10122 | DECL_NAME (tmp), TREE_TYPE (tmp)); | |
10123 | DECL_ARTIFICIAL (t) = 1; | |
10124 | DECL_NAMELESS (t) = 1; | |
10125 | DECL_ARG_TYPE (t) = ptr_type_node; | |
10126 | DECL_CONTEXT (t) = current_function_decl; | |
10127 | TREE_USED (t) = 1; | |
10128 | TREE_READONLY (t) = 1; | |
10129 | DECL_ARGUMENTS (child_fn2) = t; | |
10130 | gcc_assert (TREE_CHAIN (tmp) == NULL_TREE); | |
10131 | ||
10132 | gimplify_function_tree (child_fn2); | |
10133 | cgraph_node::add_new_function (child_fn2, true); | |
10134 | ||
10135 | vec_safe_push (offload_funcs, child_fn2); | |
10136 | if (dump_file && !gimple_in_ssa_p (cfun)) | |
10137 | { | |
10138 | dump_function_header (dump_file, child_fn2, dump_flags); | |
10139 | dump_function_to_file (child_fn2, dump_file, dump_flags); | |
10140 | } | |
10141 | } | |
629b3d75 MJ |
10142 | } |
10143 | ||
10144 | /* Emit a library call to launch the offloading region, or do data | |
10145 | transfers. */ | |
d6621a2f | 10146 | tree t1, t2, t3, t4, depend; |
629b3d75 | 10147 | enum built_in_function start_ix; |
629b3d75 | 10148 | unsigned int flags_i = 0; |
629b3d75 MJ |
10149 | |
10150 | switch (gimple_omp_target_kind (entry_stmt)) | |
10151 | { | |
10152 | case GF_OMP_TARGET_KIND_REGION: | |
10153 | start_ix = BUILT_IN_GOMP_TARGET; | |
10154 | break; | |
10155 | case GF_OMP_TARGET_KIND_DATA: | |
10156 | start_ix = BUILT_IN_GOMP_TARGET_DATA; | |
10157 | break; | |
10158 | case GF_OMP_TARGET_KIND_UPDATE: | |
10159 | start_ix = BUILT_IN_GOMP_TARGET_UPDATE; | |
10160 | break; | |
10161 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
10162 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; | |
10163 | break; | |
10164 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
10165 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; | |
10166 | flags_i |= GOMP_TARGET_FLAG_EXIT_DATA; | |
10167 | break; | |
629b3d75 | 10168 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
62aee289 MR |
10169 | case GF_OMP_TARGET_KIND_OACC_KERNELS: |
10170 | case GF_OMP_TARGET_KIND_OACC_SERIAL: | |
e898ce79 GB |
10171 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
10172 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
629b3d75 MJ |
10173 | start_ix = BUILT_IN_GOACC_PARALLEL; |
10174 | break; | |
10175 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
10176 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
e898ce79 | 10177 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: |
629b3d75 MJ |
10178 | start_ix = BUILT_IN_GOACC_DATA_START; |
10179 | break; | |
10180 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
10181 | start_ix = BUILT_IN_GOACC_UPDATE; | |
10182 | break; | |
7aefef31 AS |
10183 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
10184 | start_ix = BUILT_IN_GOACC_ENTER_DATA; | |
10185 | break; | |
10186 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
10187 | start_ix = BUILT_IN_GOACC_EXIT_DATA; | |
629b3d75 MJ |
10188 | break; |
10189 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
10190 | start_ix = BUILT_IN_GOACC_DECLARE; | |
10191 | break; | |
10192 | default: | |
10193 | gcc_unreachable (); | |
10194 | } | |
10195 | ||
59d5960c TS |
10196 | tree device = NULL_TREE; |
10197 | location_t device_loc = UNKNOWN_LOCATION; | |
10198 | tree goacc_flags = NULL_TREE; | |
1158fe43 JJ |
10199 | bool need_device_adjustment = false; |
10200 | gimple_stmt_iterator adj_gsi; | |
59d5960c | 10201 | if (is_gimple_omp_oacc (entry_stmt)) |
629b3d75 | 10202 | { |
59d5960c TS |
10203 | /* By default, no GOACC_FLAGs are set. */ |
10204 | goacc_flags = integer_zero_node; | |
629b3d75 MJ |
10205 | } |
10206 | else | |
59d5960c TS |
10207 | { |
10208 | c = omp_find_clause (clauses, OMP_CLAUSE_DEVICE); | |
10209 | if (c) | |
10210 | { | |
10211 | device = OMP_CLAUSE_DEVICE_ID (c); | |
1158fe43 JJ |
10212 | /* Ensure 'device' is of the correct type. */ |
10213 | device = fold_convert_loc (device_loc, integer_type_node, device); | |
10214 | if (TREE_CODE (device) == INTEGER_CST) | |
10215 | { | |
10216 | if (wi::to_wide (device) == GOMP_DEVICE_ICV) | |
10217 | device = build_int_cst (integer_type_node, | |
10218 | GOMP_DEVICE_HOST_FALLBACK); | |
10219 | else if (wi::to_wide (device) == GOMP_DEVICE_HOST_FALLBACK) | |
10220 | device = build_int_cst (integer_type_node, | |
10221 | GOMP_DEVICE_HOST_FALLBACK - 1); | |
10222 | } | |
10223 | else | |
10224 | need_device_adjustment = true; | |
59d5960c | 10225 | device_loc = OMP_CLAUSE_LOCATION (c); |
03be3cfe | 10226 | if (OMP_CLAUSE_DEVICE_ANCESTOR (c)) |
d6621a2f TB |
10227 | device = build_int_cst (integer_type_node, |
10228 | GOMP_DEVICE_HOST_FALLBACK); | |
59d5960c TS |
10229 | } |
10230 | else | |
10231 | { | |
10232 | /* By default, the value of DEVICE is GOMP_DEVICE_ICV (let runtime | |
10233 | library choose). */ | |
10234 | device = build_int_cst (integer_type_node, GOMP_DEVICE_ICV); | |
10235 | device_loc = gimple_location (entry_stmt); | |
10236 | } | |
629b3d75 | 10237 | |
59d5960c | 10238 | c = omp_find_clause (clauses, OMP_CLAUSE_NOWAIT); |
7619d334 JJ |
10239 | /* FIXME: in_reduction(...) nowait is unimplemented yet, pretend |
10240 | nowait doesn't appear. */ | |
10241 | if (c && omp_find_clause (clauses, OMP_CLAUSE_IN_REDUCTION)) | |
10242 | c = NULL; | |
59d5960c TS |
10243 | if (c) |
10244 | flags_i |= GOMP_TARGET_FLAG_NOWAIT; | |
10245 | } | |
629b3d75 | 10246 | |
59d5960c TS |
10247 | /* By default, there is no conditional. */ |
10248 | tree cond = NULL_TREE; | |
10249 | c = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
10250 | if (c) | |
10251 | cond = OMP_CLAUSE_IF_EXPR (c); | |
10252 | /* If we found the clause 'if (cond)', build: | |
1158fe43 JJ |
10253 | OpenACC: goacc_flags = (cond ? goacc_flags |
10254 | : goacc_flags | GOACC_FLAG_HOST_FALLBACK) | |
59d5960c | 10255 | OpenMP: device = (cond ? device : GOMP_DEVICE_HOST_FALLBACK) */ |
629b3d75 MJ |
10256 | if (cond) |
10257 | { | |
59d5960c TS |
10258 | tree *tp; |
10259 | if (is_gimple_omp_oacc (entry_stmt)) | |
10260 | tp = &goacc_flags; | |
10261 | else | |
1158fe43 | 10262 | tp = &device; |
59d5960c | 10263 | |
629b3d75 MJ |
10264 | cond = gimple_boolify (cond); |
10265 | ||
10266 | basic_block cond_bb, then_bb, else_bb; | |
10267 | edge e; | |
1158fe43 | 10268 | tree tmp_var = create_tmp_var (TREE_TYPE (*tp)); |
629b3d75 MJ |
10269 | if (offloaded) |
10270 | e = split_block_after_labels (new_bb); | |
10271 | else | |
10272 | { | |
65f4b875 | 10273 | gsi = gsi_last_nondebug_bb (new_bb); |
629b3d75 MJ |
10274 | gsi_prev (&gsi); |
10275 | e = split_block (new_bb, gsi_stmt (gsi)); | |
10276 | } | |
10277 | cond_bb = e->src; | |
10278 | new_bb = e->dest; | |
10279 | remove_edge (e); | |
10280 | ||
10281 | then_bb = create_empty_bb (cond_bb); | |
10282 | else_bb = create_empty_bb (then_bb); | |
10283 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
10284 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
10285 | ||
10286 | stmt = gimple_build_cond_empty (cond); | |
10287 | gsi = gsi_last_bb (cond_bb); | |
10288 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
10289 | ||
10290 | gsi = gsi_start_bb (then_bb); | |
59d5960c | 10291 | stmt = gimple_build_assign (tmp_var, *tp); |
629b3d75 | 10292 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
1158fe43 | 10293 | adj_gsi = gsi; |
629b3d75 MJ |
10294 | |
10295 | gsi = gsi_start_bb (else_bb); | |
59d5960c TS |
10296 | if (is_gimple_omp_oacc (entry_stmt)) |
10297 | stmt = gimple_build_assign (tmp_var, | |
10298 | BIT_IOR_EXPR, | |
10299 | *tp, | |
10300 | build_int_cst (integer_type_node, | |
10301 | GOACC_FLAG_HOST_FALLBACK)); | |
10302 | else | |
10303 | stmt = gimple_build_assign (tmp_var, | |
10304 | build_int_cst (integer_type_node, | |
10305 | GOMP_DEVICE_HOST_FALLBACK)); | |
629b3d75 MJ |
10306 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
10307 | ||
10308 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
10309 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
10310 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
10311 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
10312 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); | |
10313 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); | |
10314 | ||
59d5960c TS |
10315 | *tp = tmp_var; |
10316 | ||
65f4b875 | 10317 | gsi = gsi_last_nondebug_bb (new_bb); |
629b3d75 MJ |
10318 | } |
10319 | else | |
10320 | { | |
65f4b875 | 10321 | gsi = gsi_last_nondebug_bb (new_bb); |
59d5960c TS |
10322 | |
10323 | if (device != NULL_TREE) | |
10324 | device = force_gimple_operand_gsi (&gsi, device, true, NULL_TREE, | |
10325 | true, GSI_SAME_STMT); | |
1158fe43 JJ |
10326 | if (need_device_adjustment) |
10327 | { | |
10328 | tree tmp_var = create_tmp_var (TREE_TYPE (device)); | |
10329 | stmt = gimple_build_assign (tmp_var, device); | |
10330 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
10331 | adj_gsi = gsi_for_stmt (stmt); | |
10332 | device = tmp_var; | |
10333 | } | |
10334 | } | |
10335 | ||
3a359638 CLT |
10336 | if ((c = omp_find_clause (clauses, OMP_CLAUSE_SELF)) != NULL_TREE) |
10337 | { | |
7b2ae64b TS |
10338 | gcc_assert ((is_gimple_omp_oacc (entry_stmt) && offloaded) |
10339 | || (gimple_omp_target_kind (entry_stmt) | |
10340 | == GF_OMP_TARGET_KIND_OACC_DATA_KERNELS)); | |
3a359638 | 10341 | |
7b2ae64b TS |
10342 | edge e; |
10343 | if (offloaded) | |
10344 | e = split_block_after_labels (new_bb); | |
10345 | else | |
10346 | { | |
10347 | gsi = gsi_last_nondebug_bb (new_bb); | |
10348 | gsi_prev (&gsi); | |
10349 | e = split_block (new_bb, gsi_stmt (gsi)); | |
10350 | } | |
3a359638 CLT |
10351 | basic_block cond_bb = e->src; |
10352 | new_bb = e->dest; | |
10353 | remove_edge (e); | |
10354 | ||
10355 | basic_block then_bb = create_empty_bb (cond_bb); | |
10356 | basic_block else_bb = create_empty_bb (then_bb); | |
10357 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
10358 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
10359 | ||
10360 | tree self_cond = gimple_boolify (OMP_CLAUSE_SELF_EXPR (c)); | |
10361 | stmt = gimple_build_cond_empty (self_cond); | |
10362 | gsi = gsi_last_bb (cond_bb); | |
10363 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
10364 | ||
10365 | tree tmp_var = create_tmp_var (TREE_TYPE (goacc_flags)); | |
10366 | stmt = gimple_build_assign (tmp_var, BIT_IOR_EXPR, goacc_flags, | |
10367 | build_int_cst (integer_type_node, | |
10368 | GOACC_FLAG_LOCAL_DEVICE)); | |
10369 | gsi = gsi_start_bb (then_bb); | |
10370 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
10371 | ||
10372 | gsi = gsi_start_bb (else_bb); | |
10373 | stmt = gimple_build_assign (tmp_var, goacc_flags); | |
10374 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
10375 | ||
10376 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
10377 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
10378 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
10379 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
10380 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); | |
10381 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); | |
10382 | ||
10383 | goacc_flags = tmp_var; | |
10384 | gsi = gsi_last_nondebug_bb (new_bb); | |
10385 | } | |
10386 | ||
1158fe43 JJ |
10387 | if (need_device_adjustment) |
10388 | { | |
10389 | tree uns = fold_convert (unsigned_type_node, device); | |
10390 | uns = force_gimple_operand_gsi (&adj_gsi, uns, true, NULL_TREE, | |
10391 | false, GSI_CONTINUE_LINKING); | |
10392 | edge e = split_block (gsi_bb (adj_gsi), gsi_stmt (adj_gsi)); | |
10393 | basic_block cond_bb = e->src; | |
10394 | basic_block else_bb = e->dest; | |
10395 | if (gsi_bb (adj_gsi) == new_bb) | |
10396 | { | |
10397 | new_bb = else_bb; | |
10398 | gsi = gsi_last_nondebug_bb (new_bb); | |
10399 | } | |
10400 | ||
10401 | basic_block then_bb = create_empty_bb (cond_bb); | |
10402 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
10403 | ||
10404 | cond = build2 (GT_EXPR, boolean_type_node, uns, | |
10405 | build_int_cst (unsigned_type_node, | |
10406 | GOMP_DEVICE_HOST_FALLBACK - 1)); | |
10407 | stmt = gimple_build_cond_empty (cond); | |
10408 | adj_gsi = gsi_last_bb (cond_bb); | |
10409 | gsi_insert_after (&adj_gsi, stmt, GSI_CONTINUE_LINKING); | |
10410 | ||
10411 | adj_gsi = gsi_start_bb (then_bb); | |
10412 | tree add = build2 (PLUS_EXPR, integer_type_node, device, | |
10413 | build_int_cst (integer_type_node, -1)); | |
10414 | stmt = gimple_build_assign (device, add); | |
10415 | gsi_insert_after (&adj_gsi, stmt, GSI_CONTINUE_LINKING); | |
10416 | ||
10417 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
10418 | e->flags = EDGE_FALSE_VALUE; | |
10419 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
10420 | make_edge (then_bb, else_bb, EDGE_FALLTHRU); | |
629b3d75 MJ |
10421 | } |
10422 | ||
10423 | t = gimple_omp_target_data_arg (entry_stmt); | |
10424 | if (t == NULL) | |
10425 | { | |
10426 | t1 = size_zero_node; | |
10427 | t2 = build_zero_cst (ptr_type_node); | |
10428 | t3 = t2; | |
10429 | t4 = t2; | |
10430 | } | |
10431 | else | |
10432 | { | |
10433 | t1 = TYPE_MAX_VALUE (TYPE_DOMAIN (TREE_TYPE (TREE_VEC_ELT (t, 1)))); | |
10434 | t1 = size_binop (PLUS_EXPR, t1, size_int (1)); | |
10435 | t2 = build_fold_addr_expr (TREE_VEC_ELT (t, 0)); | |
10436 | t3 = build_fold_addr_expr (TREE_VEC_ELT (t, 1)); | |
10437 | t4 = build_fold_addr_expr (TREE_VEC_ELT (t, 2)); | |
10438 | } | |
10439 | ||
10440 | gimple *g; | |
10441 | bool tagging = false; | |
10442 | /* The maximum number used by any start_ix, without varargs. */ | |
10443 | auto_vec<tree, 11> args; | |
59d5960c TS |
10444 | if (is_gimple_omp_oacc (entry_stmt)) |
10445 | { | |
10446 | tree goacc_flags_m = fold_build1 (GOACC_FLAGS_MARSHAL_OP, | |
10447 | TREE_TYPE (goacc_flags), goacc_flags); | |
10448 | goacc_flags_m = force_gimple_operand_gsi (&gsi, goacc_flags_m, true, | |
10449 | NULL_TREE, true, | |
10450 | GSI_SAME_STMT); | |
10451 | args.quick_push (goacc_flags_m); | |
10452 | } | |
10453 | else | |
10454 | args.quick_push (device); | |
629b3d75 | 10455 | if (offloaded) |
d6621a2f | 10456 | args.quick_push (build_fold_addr_expr (child_fn2 ? child_fn2 : child_fn)); |
629b3d75 MJ |
10457 | args.quick_push (t1); |
10458 | args.quick_push (t2); | |
10459 | args.quick_push (t3); | |
10460 | args.quick_push (t4); | |
10461 | switch (start_ix) | |
10462 | { | |
10463 | case BUILT_IN_GOACC_DATA_START: | |
10464 | case BUILT_IN_GOACC_DECLARE: | |
10465 | case BUILT_IN_GOMP_TARGET_DATA: | |
10466 | break; | |
10467 | case BUILT_IN_GOMP_TARGET: | |
10468 | case BUILT_IN_GOMP_TARGET_UPDATE: | |
10469 | case BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA: | |
10470 | args.quick_push (build_int_cst (unsigned_type_node, flags_i)); | |
10471 | c = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
10472 | if (c) | |
10473 | depend = OMP_CLAUSE_DECL (c); | |
10474 | else | |
10475 | depend = build_int_cst (ptr_type_node, 0); | |
10476 | args.quick_push (depend); | |
10477 | if (start_ix == BUILT_IN_GOMP_TARGET) | |
10478 | args.quick_push (get_target_arguments (&gsi, entry_stmt)); | |
10479 | break; | |
10480 | case BUILT_IN_GOACC_PARALLEL: | |
62aee289 MR |
10481 | if (lookup_attribute ("oacc serial", DECL_ATTRIBUTES (child_fn)) != NULL) |
10482 | { | |
10483 | tree dims = NULL_TREE; | |
10484 | unsigned int ix; | |
10485 | ||
10486 | /* For serial constructs we set all dimensions to 1. */ | |
10487 | for (ix = GOMP_DIM_MAX; ix--;) | |
10488 | dims = tree_cons (NULL_TREE, integer_one_node, dims); | |
10489 | oacc_replace_fn_attrib (child_fn, dims); | |
10490 | } | |
10491 | else | |
10492 | oacc_set_fn_attrib (child_fn, clauses, &args); | |
25651634 | 10493 | tagging = true; |
629b3d75 | 10494 | /* FALLTHRU */ |
7aefef31 AS |
10495 | case BUILT_IN_GOACC_ENTER_DATA: |
10496 | case BUILT_IN_GOACC_EXIT_DATA: | |
629b3d75 MJ |
10497 | case BUILT_IN_GOACC_UPDATE: |
10498 | { | |
10499 | tree t_async = NULL_TREE; | |
10500 | ||
10501 | /* If present, use the value specified by the respective | |
10502 | clause, making sure that is of the correct type. */ | |
10503 | c = omp_find_clause (clauses, OMP_CLAUSE_ASYNC); | |
10504 | if (c) | |
10505 | t_async = fold_convert_loc (OMP_CLAUSE_LOCATION (c), | |
10506 | integer_type_node, | |
10507 | OMP_CLAUSE_ASYNC_EXPR (c)); | |
10508 | else if (!tagging) | |
10509 | /* Default values for t_async. */ | |
10510 | t_async = fold_convert_loc (gimple_location (entry_stmt), | |
10511 | integer_type_node, | |
10512 | build_int_cst (integer_type_node, | |
10513 | GOMP_ASYNC_SYNC)); | |
10514 | if (tagging && t_async) | |
10515 | { | |
10516 | unsigned HOST_WIDE_INT i_async = GOMP_LAUNCH_OP_MAX; | |
10517 | ||
10518 | if (TREE_CODE (t_async) == INTEGER_CST) | |
10519 | { | |
10520 | /* See if we can pack the async arg in to the tag's | |
10521 | operand. */ | |
10522 | i_async = TREE_INT_CST_LOW (t_async); | |
10523 | if (i_async < GOMP_LAUNCH_OP_MAX) | |
10524 | t_async = NULL_TREE; | |
10525 | else | |
10526 | i_async = GOMP_LAUNCH_OP_MAX; | |
10527 | } | |
10528 | args.safe_push (oacc_launch_pack (GOMP_LAUNCH_ASYNC, NULL_TREE, | |
10529 | i_async)); | |
10530 | } | |
10531 | if (t_async) | |
ee9fcee3 AS |
10532 | args.safe_push (force_gimple_operand_gsi (&gsi, t_async, true, |
10533 | NULL_TREE, true, | |
10534 | GSI_SAME_STMT)); | |
629b3d75 MJ |
10535 | |
10536 | /* Save the argument index, and ... */ | |
10537 | unsigned t_wait_idx = args.length (); | |
10538 | unsigned num_waits = 0; | |
10539 | c = omp_find_clause (clauses, OMP_CLAUSE_WAIT); | |
10540 | if (!tagging || c) | |
10541 | /* ... push a placeholder. */ | |
10542 | args.safe_push (integer_zero_node); | |
10543 | ||
10544 | for (; c; c = OMP_CLAUSE_CHAIN (c)) | |
10545 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_WAIT) | |
10546 | { | |
ee9fcee3 AS |
10547 | tree arg = fold_convert_loc (OMP_CLAUSE_LOCATION (c), |
10548 | integer_type_node, | |
10549 | OMP_CLAUSE_WAIT_EXPR (c)); | |
10550 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, true, | |
10551 | GSI_SAME_STMT); | |
10552 | args.safe_push (arg); | |
629b3d75 MJ |
10553 | num_waits++; |
10554 | } | |
10555 | ||
10556 | if (!tagging || num_waits) | |
10557 | { | |
10558 | tree len; | |
10559 | ||
10560 | /* Now that we know the number, update the placeholder. */ | |
10561 | if (tagging) | |
10562 | len = oacc_launch_pack (GOMP_LAUNCH_WAIT, NULL_TREE, num_waits); | |
10563 | else | |
10564 | len = build_int_cst (integer_type_node, num_waits); | |
10565 | len = fold_convert_loc (gimple_location (entry_stmt), | |
10566 | unsigned_type_node, len); | |
10567 | args[t_wait_idx] = len; | |
10568 | } | |
10569 | } | |
10570 | break; | |
10571 | default: | |
10572 | gcc_unreachable (); | |
10573 | } | |
10574 | if (tagging) | |
10575 | /* Push terminal marker - zero. */ | |
10576 | args.safe_push (oacc_launch_pack (0, NULL_TREE, 0)); | |
10577 | ||
d6621a2f TB |
10578 | if (child_fn2) |
10579 | { | |
10580 | g = gimple_build_call_internal (IFN_GOMP_TARGET_REV, 1, | |
10581 | build_fold_addr_expr (child_fn)); | |
10582 | gimple_set_location (g, gimple_location (entry_stmt)); | |
10583 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
10584 | } | |
10585 | ||
629b3d75 MJ |
10586 | g = gimple_build_call_vec (builtin_decl_explicit (start_ix), args); |
10587 | gimple_set_location (g, gimple_location (entry_stmt)); | |
10588 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
10589 | if (!offloaded) | |
10590 | { | |
10591 | g = gsi_stmt (gsi); | |
10592 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_TARGET); | |
10593 | gsi_remove (&gsi, true); | |
10594 | } | |
629b3d75 MJ |
10595 | } |
10596 | ||
629b3d75 MJ |
10597 | /* Expand the parallel region tree rooted at REGION. Expansion |
10598 | proceeds in depth-first order. Innermost regions are expanded | |
10599 | first. This way, parallel regions that require a new function to | |
10600 | be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any | |
10601 | internal dependencies in their body. */ | |
10602 | ||
10603 | static void | |
10604 | expand_omp (struct omp_region *region) | |
10605 | { | |
10606 | omp_any_child_fn_dumped = false; | |
10607 | while (region) | |
10608 | { | |
10609 | location_t saved_location; | |
10610 | gimple *inner_stmt = NULL; | |
10611 | ||
10612 | /* First, determine whether this is a combined parallel+workshare | |
01914336 | 10613 | region. */ |
629b3d75 MJ |
10614 | if (region->type == GIMPLE_OMP_PARALLEL) |
10615 | determine_parallel_type (region); | |
629b3d75 MJ |
10616 | |
10617 | if (region->type == GIMPLE_OMP_FOR | |
fe8ac82f RB |
10618 | && gimple_omp_for_combined_p (last_nondebug_stmt (region->entry))) |
10619 | inner_stmt = last_nondebug_stmt (region->inner->entry); | |
629b3d75 MJ |
10620 | |
10621 | if (region->inner) | |
10622 | expand_omp (region->inner); | |
10623 | ||
10624 | saved_location = input_location; | |
fe8ac82f RB |
10625 | if (gimple_has_location (last_nondebug_stmt (region->entry))) |
10626 | input_location = gimple_location (last_nondebug_stmt (region->entry)); | |
629b3d75 MJ |
10627 | |
10628 | switch (region->type) | |
10629 | { | |
10630 | case GIMPLE_OMP_PARALLEL: | |
10631 | case GIMPLE_OMP_TASK: | |
10632 | expand_omp_taskreg (region); | |
10633 | break; | |
10634 | ||
10635 | case GIMPLE_OMP_FOR: | |
10636 | expand_omp_for (region, inner_stmt); | |
10637 | break; | |
10638 | ||
10639 | case GIMPLE_OMP_SECTIONS: | |
10640 | expand_omp_sections (region); | |
10641 | break; | |
10642 | ||
10643 | case GIMPLE_OMP_SECTION: | |
10644 | /* Individual omp sections are handled together with their | |
10645 | parent GIMPLE_OMP_SECTIONS region. */ | |
10646 | break; | |
10647 | ||
a62c8324 SL |
10648 | case GIMPLE_OMP_STRUCTURED_BLOCK: |
10649 | /* We should have gotten rid of these in gimple lowering. */ | |
10650 | gcc_unreachable (); | |
10651 | ||
629b3d75 | 10652 | case GIMPLE_OMP_SINGLE: |
e45483c7 | 10653 | case GIMPLE_OMP_SCOPE: |
629b3d75 MJ |
10654 | expand_omp_single (region); |
10655 | break; | |
10656 | ||
10657 | case GIMPLE_OMP_ORDERED: | |
10658 | { | |
10659 | gomp_ordered *ord_stmt | |
fe8ac82f | 10660 | = as_a <gomp_ordered *> (last_nondebug_stmt (region->entry)); |
3f585bda | 10661 | if (gimple_omp_ordered_standalone_p (ord_stmt)) |
629b3d75 MJ |
10662 | { |
10663 | /* We'll expand these when expanding corresponding | |
10664 | worksharing region with ordered(n) clause. */ | |
10665 | gcc_assert (region->outer | |
10666 | && region->outer->type == GIMPLE_OMP_FOR); | |
10667 | region->ord_stmt = ord_stmt; | |
10668 | break; | |
10669 | } | |
10670 | } | |
10671 | /* FALLTHRU */ | |
10672 | case GIMPLE_OMP_MASTER: | |
d0befed7 | 10673 | case GIMPLE_OMP_MASKED: |
629b3d75 MJ |
10674 | case GIMPLE_OMP_TASKGROUP: |
10675 | case GIMPLE_OMP_CRITICAL: | |
10676 | case GIMPLE_OMP_TEAMS: | |
10677 | expand_omp_synch (region); | |
10678 | break; | |
10679 | ||
10680 | case GIMPLE_OMP_ATOMIC_LOAD: | |
10681 | expand_omp_atomic (region); | |
10682 | break; | |
10683 | ||
10684 | case GIMPLE_OMP_TARGET: | |
10685 | expand_omp_target (region); | |
10686 | break; | |
10687 | ||
10688 | default: | |
10689 | gcc_unreachable (); | |
10690 | } | |
10691 | ||
10692 | input_location = saved_location; | |
10693 | region = region->next; | |
10694 | } | |
10695 | if (omp_any_child_fn_dumped) | |
10696 | { | |
10697 | if (dump_file) | |
10698 | dump_function_header (dump_file, current_function_decl, dump_flags); | |
10699 | omp_any_child_fn_dumped = false; | |
10700 | } | |
10701 | } | |
10702 | ||
10703 | /* Helper for build_omp_regions. Scan the dominator tree starting at | |
10704 | block BB. PARENT is the region that contains BB. If SINGLE_TREE is | |
10705 | true, the function ends once a single tree is built (otherwise, whole | |
10706 | forest of OMP constructs may be built). */ | |
10707 | ||
10708 | static void | |
10709 | build_omp_regions_1 (basic_block bb, struct omp_region *parent, | |
10710 | bool single_tree) | |
10711 | { | |
10712 | gimple_stmt_iterator gsi; | |
10713 | gimple *stmt; | |
10714 | basic_block son; | |
10715 | ||
65f4b875 | 10716 | gsi = gsi_last_nondebug_bb (bb); |
629b3d75 MJ |
10717 | if (!gsi_end_p (gsi) && is_gimple_omp (gsi_stmt (gsi))) |
10718 | { | |
10719 | struct omp_region *region; | |
10720 | enum gimple_code code; | |
10721 | ||
10722 | stmt = gsi_stmt (gsi); | |
10723 | code = gimple_code (stmt); | |
10724 | if (code == GIMPLE_OMP_RETURN) | |
10725 | { | |
10726 | /* STMT is the return point out of region PARENT. Mark it | |
10727 | as the exit point and make PARENT the immediately | |
10728 | enclosing region. */ | |
10729 | gcc_assert (parent); | |
10730 | region = parent; | |
10731 | region->exit = bb; | |
10732 | parent = parent->outer; | |
10733 | } | |
10734 | else if (code == GIMPLE_OMP_ATOMIC_STORE) | |
10735 | { | |
5764ee3c | 10736 | /* GIMPLE_OMP_ATOMIC_STORE is analogous to |
629b3d75 MJ |
10737 | GIMPLE_OMP_RETURN, but matches with |
10738 | GIMPLE_OMP_ATOMIC_LOAD. */ | |
10739 | gcc_assert (parent); | |
10740 | gcc_assert (parent->type == GIMPLE_OMP_ATOMIC_LOAD); | |
10741 | region = parent; | |
10742 | region->exit = bb; | |
10743 | parent = parent->outer; | |
10744 | } | |
10745 | else if (code == GIMPLE_OMP_CONTINUE) | |
10746 | { | |
10747 | gcc_assert (parent); | |
10748 | parent->cont = bb; | |
10749 | } | |
10750 | else if (code == GIMPLE_OMP_SECTIONS_SWITCH) | |
10751 | { | |
10752 | /* GIMPLE_OMP_SECTIONS_SWITCH is part of | |
10753 | GIMPLE_OMP_SECTIONS, and we do nothing for it. */ | |
10754 | } | |
10755 | else | |
10756 | { | |
10757 | region = new_omp_region (bb, code, parent); | |
10758 | /* Otherwise... */ | |
10759 | if (code == GIMPLE_OMP_TARGET) | |
10760 | { | |
10761 | switch (gimple_omp_target_kind (stmt)) | |
10762 | { | |
10763 | case GF_OMP_TARGET_KIND_REGION: | |
629b3d75 MJ |
10764 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
10765 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
62aee289 | 10766 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
e898ce79 GB |
10767 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
10768 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
629b3d75 MJ |
10769 | break; |
10770 | case GF_OMP_TARGET_KIND_UPDATE: | |
10771 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
10772 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
cc9b9c0b JJ |
10773 | case GF_OMP_TARGET_KIND_DATA: |
10774 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
10775 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
10776 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: | |
629b3d75 | 10777 | case GF_OMP_TARGET_KIND_OACC_UPDATE: |
7aefef31 AS |
10778 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
10779 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
629b3d75 | 10780 | case GF_OMP_TARGET_KIND_OACC_DECLARE: |
ad2aab5c JJ |
10781 | /* ..., other than for those stand-alone directives... |
10782 | To be precise, target data isn't stand-alone, but | |
10783 | gimplifier put the end API call into try finally block | |
10784 | for it, so omp expansion can treat it as such. */ | |
629b3d75 MJ |
10785 | region = NULL; |
10786 | break; | |
10787 | default: | |
10788 | gcc_unreachable (); | |
10789 | } | |
10790 | } | |
10791 | else if (code == GIMPLE_OMP_ORDERED | |
3f585bda | 10792 | && gimple_omp_ordered_standalone_p (stmt)) |
629b3d75 MJ |
10793 | /* #pragma omp ordered depend is also just a stand-alone |
10794 | directive. */ | |
10795 | region = NULL; | |
28567c40 JJ |
10796 | else if (code == GIMPLE_OMP_TASK |
10797 | && gimple_omp_task_taskwait_p (stmt)) | |
10798 | /* #pragma omp taskwait depend(...) is a stand-alone directive. */ | |
10799 | region = NULL; | |
ad2aab5c JJ |
10800 | else if (code == GIMPLE_OMP_TASKGROUP) |
10801 | /* #pragma omp taskgroup isn't a stand-alone directive, but | |
10802 | gimplifier put the end API call into try finall block | |
10803 | for it, so omp expansion can treat it as such. */ | |
10804 | region = NULL; | |
629b3d75 MJ |
10805 | /* ..., this directive becomes the parent for a new region. */ |
10806 | if (region) | |
10807 | parent = region; | |
10808 | } | |
10809 | } | |
10810 | ||
10811 | if (single_tree && !parent) | |
10812 | return; | |
10813 | ||
10814 | for (son = first_dom_son (CDI_DOMINATORS, bb); | |
10815 | son; | |
10816 | son = next_dom_son (CDI_DOMINATORS, son)) | |
10817 | build_omp_regions_1 (son, parent, single_tree); | |
10818 | } | |
10819 | ||
10820 | /* Builds the tree of OMP regions rooted at ROOT, storing it to | |
10821 | root_omp_region. */ | |
10822 | ||
10823 | static void | |
10824 | build_omp_regions_root (basic_block root) | |
10825 | { | |
10826 | gcc_assert (root_omp_region == NULL); | |
10827 | build_omp_regions_1 (root, NULL, true); | |
10828 | gcc_assert (root_omp_region != NULL); | |
10829 | } | |
10830 | ||
10831 | /* Expands omp construct (and its subconstructs) starting in HEAD. */ | |
10832 | ||
10833 | void | |
10834 | omp_expand_local (basic_block head) | |
10835 | { | |
10836 | build_omp_regions_root (head); | |
10837 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
10838 | { | |
10839 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
10840 | dump_omp_region (dump_file, root_omp_region, 0); | |
10841 | fprintf (dump_file, "\n"); | |
10842 | } | |
10843 | ||
10844 | remove_exit_barriers (root_omp_region); | |
10845 | expand_omp (root_omp_region); | |
10846 | ||
10847 | omp_free_regions (); | |
10848 | } | |
10849 | ||
10850 | /* Scan the CFG and build a tree of OMP regions. Return the root of | |
10851 | the OMP region tree. */ | |
10852 | ||
10853 | static void | |
10854 | build_omp_regions (void) | |
10855 | { | |
10856 | gcc_assert (root_omp_region == NULL); | |
10857 | calculate_dominance_info (CDI_DOMINATORS); | |
10858 | build_omp_regions_1 (ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, false); | |
10859 | } | |
10860 | ||
10861 | /* Main entry point for expanding OMP-GIMPLE into runtime calls. */ | |
10862 | ||
10863 | static unsigned int | |
10864 | execute_expand_omp (void) | |
10865 | { | |
10866 | build_omp_regions (); | |
10867 | ||
10868 | if (!root_omp_region) | |
10869 | return 0; | |
10870 | ||
10871 | if (dump_file) | |
10872 | { | |
10873 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
10874 | dump_omp_region (dump_file, root_omp_region, 0); | |
10875 | fprintf (dump_file, "\n"); | |
10876 | } | |
10877 | ||
10878 | remove_exit_barriers (root_omp_region); | |
10879 | ||
10880 | expand_omp (root_omp_region); | |
10881 | ||
629b3d75 MJ |
10882 | omp_free_regions (); |
10883 | ||
c30bbd4d RB |
10884 | return (TODO_cleanup_cfg |
10885 | | (gimple_in_ssa_p (cfun) ? TODO_update_ssa_only_virtuals : 0)); | |
629b3d75 MJ |
10886 | } |
10887 | ||
10888 | /* OMP expansion -- the default pass, run before creation of SSA form. */ | |
10889 | ||
10890 | namespace { | |
10891 | ||
10892 | const pass_data pass_data_expand_omp = | |
10893 | { | |
10894 | GIMPLE_PASS, /* type */ | |
10895 | "ompexp", /* name */ | |
fd2b8c8b | 10896 | OPTGROUP_OMP, /* optinfo_flags */ |
629b3d75 MJ |
10897 | TV_NONE, /* tv_id */ |
10898 | PROP_gimple_any, /* properties_required */ | |
10899 | PROP_gimple_eomp, /* properties_provided */ | |
10900 | 0, /* properties_destroyed */ | |
10901 | 0, /* todo_flags_start */ | |
10902 | 0, /* todo_flags_finish */ | |
10903 | }; | |
10904 | ||
10905 | class pass_expand_omp : public gimple_opt_pass | |
10906 | { | |
10907 | public: | |
10908 | pass_expand_omp (gcc::context *ctxt) | |
10909 | : gimple_opt_pass (pass_data_expand_omp, ctxt) | |
10910 | {} | |
10911 | ||
10912 | /* opt_pass methods: */ | |
725793af | 10913 | unsigned int execute (function *) final override |
629b3d75 | 10914 | { |
5e9d6aa4 | 10915 | bool gate = ((flag_openacc != 0 || flag_openmp != 0 |
629b3d75 MJ |
10916 | || flag_openmp_simd != 0) |
10917 | && !seen_error ()); | |
10918 | ||
10919 | /* This pass always runs, to provide PROP_gimple_eomp. | |
10920 | But often, there is nothing to do. */ | |
10921 | if (!gate) | |
10922 | return 0; | |
10923 | ||
10924 | return execute_expand_omp (); | |
10925 | } | |
10926 | ||
10927 | }; // class pass_expand_omp | |
10928 | ||
10929 | } // anon namespace | |
10930 | ||
10931 | gimple_opt_pass * | |
10932 | make_pass_expand_omp (gcc::context *ctxt) | |
10933 | { | |
10934 | return new pass_expand_omp (ctxt); | |
10935 | } | |
10936 | ||
10937 | namespace { | |
10938 | ||
10939 | const pass_data pass_data_expand_omp_ssa = | |
10940 | { | |
10941 | GIMPLE_PASS, /* type */ | |
10942 | "ompexpssa", /* name */ | |
fd2b8c8b | 10943 | OPTGROUP_OMP, /* optinfo_flags */ |
629b3d75 MJ |
10944 | TV_NONE, /* tv_id */ |
10945 | PROP_cfg | PROP_ssa, /* properties_required */ | |
10946 | PROP_gimple_eomp, /* properties_provided */ | |
10947 | 0, /* properties_destroyed */ | |
10948 | 0, /* todo_flags_start */ | |
10949 | TODO_cleanup_cfg | TODO_rebuild_alias, /* todo_flags_finish */ | |
10950 | }; | |
10951 | ||
10952 | class pass_expand_omp_ssa : public gimple_opt_pass | |
10953 | { | |
10954 | public: | |
10955 | pass_expand_omp_ssa (gcc::context *ctxt) | |
10956 | : gimple_opt_pass (pass_data_expand_omp_ssa, ctxt) | |
10957 | {} | |
10958 | ||
10959 | /* opt_pass methods: */ | |
725793af | 10960 | bool gate (function *fun) final override |
629b3d75 MJ |
10961 | { |
10962 | return !(fun->curr_properties & PROP_gimple_eomp); | |
10963 | } | |
725793af DM |
10964 | unsigned int execute (function *) final override |
10965 | { | |
10966 | return execute_expand_omp (); | |
10967 | } | |
10968 | opt_pass * clone () final override | |
10969 | { | |
10970 | return new pass_expand_omp_ssa (m_ctxt); | |
10971 | } | |
629b3d75 MJ |
10972 | |
10973 | }; // class pass_expand_omp_ssa | |
10974 | ||
10975 | } // anon namespace | |
10976 | ||
10977 | gimple_opt_pass * | |
10978 | make_pass_expand_omp_ssa (gcc::context *ctxt) | |
10979 | { | |
10980 | return new pass_expand_omp_ssa (ctxt); | |
10981 | } | |
10982 | ||
e53b6e56 | 10983 | /* Called from tree-cfg.cc::make_edges to create cfg edges for all relevant |
629b3d75 MJ |
10984 | GIMPLE_* codes. */ |
10985 | ||
10986 | bool | |
10987 | omp_make_gimple_edges (basic_block bb, struct omp_region **region, | |
10988 | int *region_idx) | |
10989 | { | |
fe8ac82f | 10990 | gimple *last = last_nondebug_stmt (bb); |
629b3d75 MJ |
10991 | enum gimple_code code = gimple_code (last); |
10992 | struct omp_region *cur_region = *region; | |
10993 | bool fallthru = false; | |
10994 | ||
10995 | switch (code) | |
10996 | { | |
10997 | case GIMPLE_OMP_PARALLEL: | |
629b3d75 MJ |
10998 | case GIMPLE_OMP_FOR: |
10999 | case GIMPLE_OMP_SINGLE: | |
11000 | case GIMPLE_OMP_TEAMS: | |
11001 | case GIMPLE_OMP_MASTER: | |
d0befed7 | 11002 | case GIMPLE_OMP_MASKED: |
e45483c7 | 11003 | case GIMPLE_OMP_SCOPE: |
629b3d75 MJ |
11004 | case GIMPLE_OMP_CRITICAL: |
11005 | case GIMPLE_OMP_SECTION: | |
629b3d75 MJ |
11006 | cur_region = new_omp_region (bb, code, cur_region); |
11007 | fallthru = true; | |
11008 | break; | |
11009 | ||
ad2aab5c JJ |
11010 | case GIMPLE_OMP_TASKGROUP: |
11011 | cur_region = new_omp_region (bb, code, cur_region); | |
11012 | fallthru = true; | |
11013 | cur_region = cur_region->outer; | |
11014 | break; | |
11015 | ||
28567c40 JJ |
11016 | case GIMPLE_OMP_TASK: |
11017 | cur_region = new_omp_region (bb, code, cur_region); | |
11018 | fallthru = true; | |
11019 | if (gimple_omp_task_taskwait_p (last)) | |
11020 | cur_region = cur_region->outer; | |
11021 | break; | |
11022 | ||
629b3d75 MJ |
11023 | case GIMPLE_OMP_ORDERED: |
11024 | cur_region = new_omp_region (bb, code, cur_region); | |
11025 | fallthru = true; | |
3f585bda | 11026 | if (gimple_omp_ordered_standalone_p (last)) |
629b3d75 MJ |
11027 | cur_region = cur_region->outer; |
11028 | break; | |
11029 | ||
11030 | case GIMPLE_OMP_TARGET: | |
11031 | cur_region = new_omp_region (bb, code, cur_region); | |
11032 | fallthru = true; | |
11033 | switch (gimple_omp_target_kind (last)) | |
11034 | { | |
11035 | case GF_OMP_TARGET_KIND_REGION: | |
629b3d75 MJ |
11036 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
11037 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
62aee289 | 11038 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
e898ce79 GB |
11039 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
11040 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: | |
629b3d75 MJ |
11041 | break; |
11042 | case GF_OMP_TARGET_KIND_UPDATE: | |
11043 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
11044 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
cc9b9c0b JJ |
11045 | case GF_OMP_TARGET_KIND_DATA: |
11046 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
11047 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
11048 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: | |
629b3d75 | 11049 | case GF_OMP_TARGET_KIND_OACC_UPDATE: |
7aefef31 AS |
11050 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
11051 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: | |
629b3d75 MJ |
11052 | case GF_OMP_TARGET_KIND_OACC_DECLARE: |
11053 | cur_region = cur_region->outer; | |
11054 | break; | |
11055 | default: | |
11056 | gcc_unreachable (); | |
11057 | } | |
11058 | break; | |
11059 | ||
11060 | case GIMPLE_OMP_SECTIONS: | |
11061 | cur_region = new_omp_region (bb, code, cur_region); | |
11062 | fallthru = true; | |
11063 | break; | |
11064 | ||
11065 | case GIMPLE_OMP_SECTIONS_SWITCH: | |
11066 | fallthru = false; | |
11067 | break; | |
11068 | ||
11069 | case GIMPLE_OMP_ATOMIC_LOAD: | |
11070 | case GIMPLE_OMP_ATOMIC_STORE: | |
11071 | fallthru = true; | |
11072 | break; | |
11073 | ||
11074 | case GIMPLE_OMP_RETURN: | |
11075 | /* In the case of a GIMPLE_OMP_SECTION, the edge will go | |
11076 | somewhere other than the next block. This will be | |
11077 | created later. */ | |
11078 | cur_region->exit = bb; | |
11079 | if (cur_region->type == GIMPLE_OMP_TASK) | |
11080 | /* Add an edge corresponding to not scheduling the task | |
11081 | immediately. */ | |
11082 | make_edge (cur_region->entry, bb, EDGE_ABNORMAL); | |
11083 | fallthru = cur_region->type != GIMPLE_OMP_SECTION; | |
11084 | cur_region = cur_region->outer; | |
11085 | break; | |
11086 | ||
11087 | case GIMPLE_OMP_CONTINUE: | |
11088 | cur_region->cont = bb; | |
11089 | switch (cur_region->type) | |
11090 | { | |
11091 | case GIMPLE_OMP_FOR: | |
11092 | /* Mark all GIMPLE_OMP_FOR and GIMPLE_OMP_CONTINUE | |
11093 | succs edges as abnormal to prevent splitting | |
11094 | them. */ | |
11095 | single_succ_edge (cur_region->entry)->flags |= EDGE_ABNORMAL; | |
11096 | /* Make the loopback edge. */ | |
11097 | make_edge (bb, single_succ (cur_region->entry), | |
11098 | EDGE_ABNORMAL); | |
11099 | ||
11100 | /* Create an edge from GIMPLE_OMP_FOR to exit, which | |
11101 | corresponds to the case that the body of the loop | |
11102 | is not executed at all. */ | |
11103 | make_edge (cur_region->entry, bb->next_bb, EDGE_ABNORMAL); | |
11104 | make_edge (bb, bb->next_bb, EDGE_FALLTHRU | EDGE_ABNORMAL); | |
11105 | fallthru = false; | |
11106 | break; | |
11107 | ||
11108 | case GIMPLE_OMP_SECTIONS: | |
11109 | /* Wire up the edges into and out of the nested sections. */ | |
11110 | { | |
11111 | basic_block switch_bb = single_succ (cur_region->entry); | |
11112 | ||
11113 | struct omp_region *i; | |
11114 | for (i = cur_region->inner; i ; i = i->next) | |
11115 | { | |
11116 | gcc_assert (i->type == GIMPLE_OMP_SECTION); | |
11117 | make_edge (switch_bb, i->entry, 0); | |
11118 | make_edge (i->exit, bb, EDGE_FALLTHRU); | |
11119 | } | |
11120 | ||
11121 | /* Make the loopback edge to the block with | |
11122 | GIMPLE_OMP_SECTIONS_SWITCH. */ | |
11123 | make_edge (bb, switch_bb, 0); | |
11124 | ||
11125 | /* Make the edge from the switch to exit. */ | |
11126 | make_edge (switch_bb, bb->next_bb, 0); | |
11127 | fallthru = false; | |
11128 | } | |
11129 | break; | |
11130 | ||
11131 | case GIMPLE_OMP_TASK: | |
11132 | fallthru = true; | |
11133 | break; | |
11134 | ||
11135 | default: | |
11136 | gcc_unreachable (); | |
11137 | } | |
11138 | break; | |
11139 | ||
11140 | default: | |
11141 | gcc_unreachable (); | |
11142 | } | |
11143 | ||
11144 | if (*region != cur_region) | |
11145 | { | |
11146 | *region = cur_region; | |
11147 | if (cur_region) | |
11148 | *region_idx = cur_region->entry->index; | |
11149 | else | |
11150 | *region_idx = 0; | |
11151 | } | |
11152 | ||
11153 | return fallthru; | |
11154 | } |