]>
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 | ||
a5544970 | 5 | Copyright (C) 2005-2019 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" | |
55 | #include "symbol-summary.h" | |
629b3d75 MJ |
56 | #include "gomp-constants.h" |
57 | #include "gimple-pretty-print.h" | |
13293add | 58 | #include "hsa-common.h" |
314e6352 ML |
59 | #include "stringpool.h" |
60 | #include "attribs.h" | |
629b3d75 MJ |
61 | |
62 | /* OMP region information. Every parallel and workshare | |
63 | directive is enclosed between two markers, the OMP_* directive | |
64 | and a corresponding GIMPLE_OMP_RETURN statement. */ | |
65 | ||
66 | struct omp_region | |
67 | { | |
68 | /* The enclosing region. */ | |
69 | struct omp_region *outer; | |
70 | ||
71 | /* First child region. */ | |
72 | struct omp_region *inner; | |
73 | ||
74 | /* Next peer region. */ | |
75 | struct omp_region *next; | |
76 | ||
77 | /* Block containing the omp directive as its last stmt. */ | |
78 | basic_block entry; | |
79 | ||
80 | /* Block containing the GIMPLE_OMP_RETURN as its last stmt. */ | |
81 | basic_block exit; | |
82 | ||
83 | /* Block containing the GIMPLE_OMP_CONTINUE as its last stmt. */ | |
84 | basic_block cont; | |
85 | ||
86 | /* If this is a combined parallel+workshare region, this is a list | |
87 | of additional arguments needed by the combined parallel+workshare | |
88 | library call. */ | |
89 | vec<tree, va_gc> *ws_args; | |
90 | ||
91 | /* The code for the omp directive of this region. */ | |
92 | enum gimple_code type; | |
93 | ||
94 | /* Schedule kind, only used for GIMPLE_OMP_FOR type regions. */ | |
95 | enum omp_clause_schedule_kind sched_kind; | |
96 | ||
97 | /* Schedule modifiers. */ | |
98 | unsigned char sched_modifiers; | |
99 | ||
100 | /* True if this is a combined parallel+workshare region. */ | |
101 | bool is_combined_parallel; | |
102 | ||
0b887b75 JJ |
103 | /* Copy of fd.lastprivate_conditional != 0. */ |
104 | bool has_lastprivate_conditional; | |
105 | ||
629b3d75 MJ |
106 | /* The ordered stmt if type is GIMPLE_OMP_ORDERED and it has |
107 | a depend clause. */ | |
108 | gomp_ordered *ord_stmt; | |
109 | }; | |
110 | ||
111 | static struct omp_region *root_omp_region; | |
112 | static bool omp_any_child_fn_dumped; | |
113 | ||
114 | static void expand_omp_build_assign (gimple_stmt_iterator *, tree, tree, | |
115 | bool = false); | |
116 | static gphi *find_phi_with_arg_on_edge (tree, edge); | |
117 | static void expand_omp (struct omp_region *region); | |
118 | ||
119 | /* Return true if REGION is a combined parallel+workshare region. */ | |
120 | ||
121 | static inline bool | |
122 | is_combined_parallel (struct omp_region *region) | |
123 | { | |
124 | return region->is_combined_parallel; | |
125 | } | |
126 | ||
127 | /* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB | |
128 | is the immediate dominator of PAR_ENTRY_BB, return true if there | |
129 | are no data dependencies that would prevent expanding the parallel | |
130 | directive at PAR_ENTRY_BB as a combined parallel+workshare region. | |
131 | ||
132 | When expanding a combined parallel+workshare region, the call to | |
133 | the child function may need additional arguments in the case of | |
134 | GIMPLE_OMP_FOR regions. In some cases, these arguments are | |
135 | computed out of variables passed in from the parent to the child | |
136 | via 'struct .omp_data_s'. For instance: | |
137 | ||
138 | #pragma omp parallel for schedule (guided, i * 4) | |
139 | for (j ...) | |
140 | ||
141 | Is lowered into: | |
142 | ||
01914336 | 143 | # BLOCK 2 (PAR_ENTRY_BB) |
629b3d75 MJ |
144 | .omp_data_o.i = i; |
145 | #pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598) | |
146 | ||
147 | # BLOCK 3 (WS_ENTRY_BB) | |
148 | .omp_data_i = &.omp_data_o; | |
149 | D.1667 = .omp_data_i->i; | |
150 | D.1598 = D.1667 * 4; | |
151 | #pragma omp for schedule (guided, D.1598) | |
152 | ||
153 | When we outline the parallel region, the call to the child function | |
154 | 'bar.omp_fn.0' will need the value D.1598 in its argument list, but | |
155 | that value is computed *after* the call site. So, in principle we | |
156 | cannot do the transformation. | |
157 | ||
158 | To see whether the code in WS_ENTRY_BB blocks the combined | |
159 | parallel+workshare call, we collect all the variables used in the | |
160 | GIMPLE_OMP_FOR header check whether they appear on the LHS of any | |
161 | statement in WS_ENTRY_BB. If so, then we cannot emit the combined | |
162 | call. | |
163 | ||
164 | FIXME. If we had the SSA form built at this point, we could merely | |
165 | hoist the code in block 3 into block 2 and be done with it. But at | |
166 | this point we don't have dataflow information and though we could | |
167 | hack something up here, it is really not worth the aggravation. */ | |
168 | ||
169 | static bool | |
170 | workshare_safe_to_combine_p (basic_block ws_entry_bb) | |
171 | { | |
172 | struct omp_for_data fd; | |
173 | gimple *ws_stmt = last_stmt (ws_entry_bb); | |
174 | ||
175 | if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) | |
176 | return true; | |
177 | ||
178 | gcc_assert (gimple_code (ws_stmt) == GIMPLE_OMP_FOR); | |
28567c40 JJ |
179 | if (gimple_omp_for_kind (ws_stmt) != GF_OMP_FOR_KIND_FOR) |
180 | return false; | |
629b3d75 MJ |
181 | |
182 | omp_extract_for_data (as_a <gomp_for *> (ws_stmt), &fd, NULL); | |
183 | ||
184 | if (fd.collapse > 1 && TREE_CODE (fd.loop.n2) != INTEGER_CST) | |
185 | return false; | |
186 | if (fd.iter_type != long_integer_type_node) | |
187 | return false; | |
188 | ||
189 | /* FIXME. We give up too easily here. If any of these arguments | |
190 | are not constants, they will likely involve variables that have | |
191 | been mapped into fields of .omp_data_s for sharing with the child | |
192 | function. With appropriate data flow, it would be possible to | |
193 | see through this. */ | |
194 | if (!is_gimple_min_invariant (fd.loop.n1) | |
195 | || !is_gimple_min_invariant (fd.loop.n2) | |
196 | || !is_gimple_min_invariant (fd.loop.step) | |
197 | || (fd.chunk_size && !is_gimple_min_invariant (fd.chunk_size))) | |
198 | return false; | |
199 | ||
200 | return true; | |
201 | } | |
202 | ||
203 | /* Adjust CHUNK_SIZE from SCHEDULE clause, depending on simd modifier | |
204 | presence (SIMD_SCHEDULE). */ | |
205 | ||
206 | static tree | |
207 | omp_adjust_chunk_size (tree chunk_size, bool simd_schedule) | |
208 | { | |
28567c40 | 209 | if (!simd_schedule || integer_zerop (chunk_size)) |
629b3d75 MJ |
210 | return chunk_size; |
211 | ||
9d2f08ab RS |
212 | poly_uint64 vf = omp_max_vf (); |
213 | if (known_eq (vf, 1U)) | |
629b3d75 MJ |
214 | return chunk_size; |
215 | ||
216 | tree type = TREE_TYPE (chunk_size); | |
217 | chunk_size = fold_build2 (PLUS_EXPR, type, chunk_size, | |
218 | build_int_cst (type, vf - 1)); | |
219 | return fold_build2 (BIT_AND_EXPR, type, chunk_size, | |
220 | build_int_cst (type, -vf)); | |
221 | } | |
222 | ||
223 | /* Collect additional arguments needed to emit a combined | |
224 | parallel+workshare call. WS_STMT is the workshare directive being | |
225 | expanded. */ | |
226 | ||
227 | static vec<tree, va_gc> * | |
228 | get_ws_args_for (gimple *par_stmt, gimple *ws_stmt) | |
229 | { | |
230 | tree t; | |
231 | location_t loc = gimple_location (ws_stmt); | |
232 | vec<tree, va_gc> *ws_args; | |
233 | ||
234 | if (gomp_for *for_stmt = dyn_cast <gomp_for *> (ws_stmt)) | |
235 | { | |
236 | struct omp_for_data fd; | |
237 | tree n1, n2; | |
238 | ||
239 | omp_extract_for_data (for_stmt, &fd, NULL); | |
240 | n1 = fd.loop.n1; | |
241 | n2 = fd.loop.n2; | |
242 | ||
243 | if (gimple_omp_for_combined_into_p (for_stmt)) | |
244 | { | |
245 | tree innerc | |
246 | = omp_find_clause (gimple_omp_parallel_clauses (par_stmt), | |
247 | OMP_CLAUSE__LOOPTEMP_); | |
248 | gcc_assert (innerc); | |
249 | n1 = OMP_CLAUSE_DECL (innerc); | |
250 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
251 | OMP_CLAUSE__LOOPTEMP_); | |
252 | gcc_assert (innerc); | |
253 | n2 = OMP_CLAUSE_DECL (innerc); | |
254 | } | |
255 | ||
256 | vec_alloc (ws_args, 3 + (fd.chunk_size != 0)); | |
257 | ||
258 | t = fold_convert_loc (loc, long_integer_type_node, n1); | |
259 | ws_args->quick_push (t); | |
260 | ||
261 | t = fold_convert_loc (loc, long_integer_type_node, n2); | |
262 | ws_args->quick_push (t); | |
263 | ||
264 | t = fold_convert_loc (loc, long_integer_type_node, fd.loop.step); | |
265 | ws_args->quick_push (t); | |
266 | ||
267 | if (fd.chunk_size) | |
268 | { | |
269 | t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size); | |
270 | t = omp_adjust_chunk_size (t, fd.simd_schedule); | |
271 | ws_args->quick_push (t); | |
272 | } | |
273 | ||
274 | return ws_args; | |
275 | } | |
276 | else if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS) | |
277 | { | |
278 | /* Number of sections is equal to the number of edges from the | |
279 | GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to | |
280 | the exit of the sections region. */ | |
281 | basic_block bb = single_succ (gimple_bb (ws_stmt)); | |
282 | t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1); | |
283 | vec_alloc (ws_args, 1); | |
284 | ws_args->quick_push (t); | |
285 | return ws_args; | |
286 | } | |
287 | ||
288 | gcc_unreachable (); | |
289 | } | |
290 | ||
291 | /* Discover whether REGION is a combined parallel+workshare region. */ | |
292 | ||
293 | static void | |
294 | determine_parallel_type (struct omp_region *region) | |
295 | { | |
296 | basic_block par_entry_bb, par_exit_bb; | |
297 | basic_block ws_entry_bb, ws_exit_bb; | |
298 | ||
299 | if (region == NULL || region->inner == NULL | |
300 | || region->exit == NULL || region->inner->exit == NULL | |
301 | || region->inner->cont == NULL) | |
302 | return; | |
303 | ||
304 | /* We only support parallel+for and parallel+sections. */ | |
305 | if (region->type != GIMPLE_OMP_PARALLEL | |
306 | || (region->inner->type != GIMPLE_OMP_FOR | |
307 | && region->inner->type != GIMPLE_OMP_SECTIONS)) | |
308 | return; | |
309 | ||
310 | /* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and | |
311 | WS_EXIT_BB -> PAR_EXIT_BB. */ | |
312 | par_entry_bb = region->entry; | |
313 | par_exit_bb = region->exit; | |
314 | ws_entry_bb = region->inner->entry; | |
315 | ws_exit_bb = region->inner->exit; | |
316 | ||
28567c40 JJ |
317 | /* Give up for task reductions on the parallel, while it is implementable, |
318 | adding another big set of APIs or slowing down the normal paths is | |
319 | not acceptable. */ | |
320 | tree pclauses = gimple_omp_parallel_clauses (last_stmt (par_entry_bb)); | |
321 | if (omp_find_clause (pclauses, OMP_CLAUSE__REDUCTEMP_)) | |
322 | return; | |
323 | ||
629b3d75 MJ |
324 | if (single_succ (par_entry_bb) == ws_entry_bb |
325 | && single_succ (ws_exit_bb) == par_exit_bb | |
326 | && workshare_safe_to_combine_p (ws_entry_bb) | |
327 | && (gimple_omp_parallel_combined_p (last_stmt (par_entry_bb)) | |
328 | || (last_and_only_stmt (ws_entry_bb) | |
329 | && last_and_only_stmt (par_exit_bb)))) | |
330 | { | |
331 | gimple *par_stmt = last_stmt (par_entry_bb); | |
332 | gimple *ws_stmt = last_stmt (ws_entry_bb); | |
333 | ||
334 | if (region->inner->type == GIMPLE_OMP_FOR) | |
335 | { | |
336 | /* If this is a combined parallel loop, we need to determine | |
337 | whether or not to use the combined library calls. There | |
338 | are two cases where we do not apply the transformation: | |
339 | static loops and any kind of ordered loop. In the first | |
340 | case, we already open code the loop so there is no need | |
341 | to do anything else. In the latter case, the combined | |
342 | parallel loop call would still need extra synchronization | |
343 | to implement ordered semantics, so there would not be any | |
344 | gain in using the combined call. */ | |
345 | tree clauses = gimple_omp_for_clauses (ws_stmt); | |
346 | tree c = omp_find_clause (clauses, OMP_CLAUSE_SCHEDULE); | |
347 | if (c == NULL | |
348 | || ((OMP_CLAUSE_SCHEDULE_KIND (c) & OMP_CLAUSE_SCHEDULE_MASK) | |
349 | == OMP_CLAUSE_SCHEDULE_STATIC) | |
28567c40 | 350 | || omp_find_clause (clauses, OMP_CLAUSE_ORDERED) |
6c7ae8c5 | 351 | || omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_) |
8221c30b JJ |
352 | || ((c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_)) |
353 | && POINTER_TYPE_P (TREE_TYPE (OMP_CLAUSE_DECL (c))))) | |
28567c40 | 354 | return; |
629b3d75 | 355 | } |
28567c40 | 356 | else if (region->inner->type == GIMPLE_OMP_SECTIONS |
8221c30b JJ |
357 | && (omp_find_clause (gimple_omp_sections_clauses (ws_stmt), |
358 | OMP_CLAUSE__REDUCTEMP_) | |
359 | || omp_find_clause (gimple_omp_sections_clauses (ws_stmt), | |
360 | OMP_CLAUSE__CONDTEMP_))) | |
28567c40 | 361 | return; |
629b3d75 MJ |
362 | |
363 | region->is_combined_parallel = true; | |
364 | region->inner->is_combined_parallel = true; | |
365 | region->ws_args = get_ws_args_for (par_stmt, ws_stmt); | |
366 | } | |
367 | } | |
368 | ||
369 | /* Debugging dumps for parallel regions. */ | |
370 | void dump_omp_region (FILE *, struct omp_region *, int); | |
371 | void debug_omp_region (struct omp_region *); | |
372 | void debug_all_omp_regions (void); | |
373 | ||
374 | /* Dump the parallel region tree rooted at REGION. */ | |
375 | ||
376 | void | |
377 | dump_omp_region (FILE *file, struct omp_region *region, int indent) | |
378 | { | |
379 | fprintf (file, "%*sbb %d: %s\n", indent, "", region->entry->index, | |
380 | gimple_code_name[region->type]); | |
381 | ||
382 | if (region->inner) | |
383 | dump_omp_region (file, region->inner, indent + 4); | |
384 | ||
385 | if (region->cont) | |
386 | { | |
387 | fprintf (file, "%*sbb %d: GIMPLE_OMP_CONTINUE\n", indent, "", | |
388 | region->cont->index); | |
389 | } | |
390 | ||
391 | if (region->exit) | |
392 | fprintf (file, "%*sbb %d: GIMPLE_OMP_RETURN\n", indent, "", | |
393 | region->exit->index); | |
394 | else | |
395 | fprintf (file, "%*s[no exit marker]\n", indent, ""); | |
396 | ||
397 | if (region->next) | |
398 | dump_omp_region (file, region->next, indent); | |
399 | } | |
400 | ||
401 | DEBUG_FUNCTION void | |
402 | debug_omp_region (struct omp_region *region) | |
403 | { | |
404 | dump_omp_region (stderr, region, 0); | |
405 | } | |
406 | ||
407 | DEBUG_FUNCTION void | |
408 | debug_all_omp_regions (void) | |
409 | { | |
410 | dump_omp_region (stderr, root_omp_region, 0); | |
411 | } | |
412 | ||
413 | /* Create a new parallel region starting at STMT inside region PARENT. */ | |
414 | ||
415 | static struct omp_region * | |
416 | new_omp_region (basic_block bb, enum gimple_code type, | |
417 | struct omp_region *parent) | |
418 | { | |
419 | struct omp_region *region = XCNEW (struct omp_region); | |
420 | ||
421 | region->outer = parent; | |
422 | region->entry = bb; | |
423 | region->type = type; | |
424 | ||
425 | if (parent) | |
426 | { | |
427 | /* This is a nested region. Add it to the list of inner | |
428 | regions in PARENT. */ | |
429 | region->next = parent->inner; | |
430 | parent->inner = region; | |
431 | } | |
432 | else | |
433 | { | |
434 | /* This is a toplevel region. Add it to the list of toplevel | |
435 | regions in ROOT_OMP_REGION. */ | |
436 | region->next = root_omp_region; | |
437 | root_omp_region = region; | |
438 | } | |
439 | ||
440 | return region; | |
441 | } | |
442 | ||
443 | /* Release the memory associated with the region tree rooted at REGION. */ | |
444 | ||
445 | static void | |
446 | free_omp_region_1 (struct omp_region *region) | |
447 | { | |
448 | struct omp_region *i, *n; | |
449 | ||
450 | for (i = region->inner; i ; i = n) | |
451 | { | |
452 | n = i->next; | |
453 | free_omp_region_1 (i); | |
454 | } | |
455 | ||
456 | free (region); | |
457 | } | |
458 | ||
459 | /* Release the memory for the entire omp region tree. */ | |
460 | ||
461 | void | |
462 | omp_free_regions (void) | |
463 | { | |
464 | struct omp_region *r, *n; | |
465 | for (r = root_omp_region; r ; r = n) | |
466 | { | |
467 | n = r->next; | |
468 | free_omp_region_1 (r); | |
469 | } | |
470 | root_omp_region = NULL; | |
471 | } | |
472 | ||
473 | /* A convenience function to build an empty GIMPLE_COND with just the | |
474 | condition. */ | |
475 | ||
476 | static gcond * | |
477 | gimple_build_cond_empty (tree cond) | |
478 | { | |
479 | enum tree_code pred_code; | |
480 | tree lhs, rhs; | |
481 | ||
482 | gimple_cond_get_ops_from_tree (cond, &pred_code, &lhs, &rhs); | |
483 | return gimple_build_cond (pred_code, lhs, rhs, NULL_TREE, NULL_TREE); | |
484 | } | |
485 | ||
486 | /* Return true if a parallel REGION is within a declare target function or | |
487 | within a target region and is not a part of a gridified target. */ | |
488 | ||
489 | static bool | |
490 | parallel_needs_hsa_kernel_p (struct omp_region *region) | |
491 | { | |
492 | bool indirect = false; | |
493 | for (region = region->outer; region; region = region->outer) | |
494 | { | |
495 | if (region->type == GIMPLE_OMP_PARALLEL) | |
496 | indirect = true; | |
497 | else if (region->type == GIMPLE_OMP_TARGET) | |
498 | { | |
499 | gomp_target *tgt_stmt | |
500 | = as_a <gomp_target *> (last_stmt (region->entry)); | |
501 | ||
502 | if (omp_find_clause (gimple_omp_target_clauses (tgt_stmt), | |
503 | OMP_CLAUSE__GRIDDIM_)) | |
504 | return indirect; | |
505 | else | |
506 | return true; | |
507 | } | |
508 | } | |
509 | ||
510 | if (lookup_attribute ("omp declare target", | |
511 | DECL_ATTRIBUTES (current_function_decl))) | |
512 | return true; | |
513 | ||
514 | return false; | |
515 | } | |
516 | ||
623c6df5 KB |
517 | /* Change DECL_CONTEXT of CHILD_FNDECL to that of the parent function. |
518 | Add CHILD_FNDECL to decl chain of the supercontext of the block | |
519 | ENTRY_BLOCK - this is the block which originally contained the | |
520 | code from which CHILD_FNDECL was created. | |
521 | ||
522 | Together, these actions ensure that the debug info for the outlined | |
523 | function will be emitted with the correct lexical scope. */ | |
524 | ||
525 | static void | |
4ccc4e30 JJ |
526 | adjust_context_and_scope (struct omp_region *region, tree entry_block, |
527 | tree child_fndecl) | |
623c6df5 | 528 | { |
4ccc4e30 JJ |
529 | tree parent_fndecl = NULL_TREE; |
530 | gimple *entry_stmt; | |
531 | /* OMP expansion expands inner regions before outer ones, so if | |
532 | we e.g. have explicit task region nested in parallel region, when | |
533 | expanding the task region current_function_decl will be the original | |
534 | source function, but we actually want to use as context the child | |
535 | function of the parallel. */ | |
536 | for (region = region->outer; | |
537 | region && parent_fndecl == NULL_TREE; region = region->outer) | |
538 | switch (region->type) | |
539 | { | |
540 | case GIMPLE_OMP_PARALLEL: | |
541 | case GIMPLE_OMP_TASK: | |
542 | case GIMPLE_OMP_TEAMS: | |
543 | entry_stmt = last_stmt (region->entry); | |
544 | parent_fndecl = gimple_omp_taskreg_child_fn (entry_stmt); | |
545 | break; | |
546 | case GIMPLE_OMP_TARGET: | |
547 | entry_stmt = last_stmt (region->entry); | |
548 | parent_fndecl | |
549 | = gimple_omp_target_child_fn (as_a <gomp_target *> (entry_stmt)); | |
550 | break; | |
551 | default: | |
552 | break; | |
553 | } | |
554 | ||
555 | if (parent_fndecl == NULL_TREE) | |
556 | parent_fndecl = current_function_decl; | |
557 | DECL_CONTEXT (child_fndecl) = parent_fndecl; | |
558 | ||
623c6df5 KB |
559 | if (entry_block != NULL_TREE && TREE_CODE (entry_block) == BLOCK) |
560 | { | |
561 | tree b = BLOCK_SUPERCONTEXT (entry_block); | |
623c6df5 KB |
562 | if (TREE_CODE (b) == BLOCK) |
563 | { | |
623c6df5 KB |
564 | DECL_CHAIN (child_fndecl) = BLOCK_VARS (b); |
565 | BLOCK_VARS (b) = child_fndecl; | |
566 | } | |
567 | } | |
568 | } | |
569 | ||
28567c40 | 570 | /* Build the function calls to GOMP_parallel etc to actually |
629b3d75 MJ |
571 | generate the parallel operation. REGION is the parallel region |
572 | being expanded. BB is the block where to insert the code. WS_ARGS | |
573 | will be set if this is a call to a combined parallel+workshare | |
574 | construct, it contains the list of additional arguments needed by | |
575 | the workshare construct. */ | |
576 | ||
577 | static void | |
578 | expand_parallel_call (struct omp_region *region, basic_block bb, | |
579 | gomp_parallel *entry_stmt, | |
580 | vec<tree, va_gc> *ws_args) | |
581 | { | |
582 | tree t, t1, t2, val, cond, c, clauses, flags; | |
583 | gimple_stmt_iterator gsi; | |
584 | gimple *stmt; | |
585 | enum built_in_function start_ix; | |
586 | int start_ix2; | |
587 | location_t clause_loc; | |
588 | vec<tree, va_gc> *args; | |
589 | ||
590 | clauses = gimple_omp_parallel_clauses (entry_stmt); | |
591 | ||
592 | /* Determine what flavor of GOMP_parallel we will be | |
593 | emitting. */ | |
594 | start_ix = BUILT_IN_GOMP_PARALLEL; | |
28567c40 JJ |
595 | tree rtmp = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); |
596 | if (rtmp) | |
597 | start_ix = BUILT_IN_GOMP_PARALLEL_REDUCTIONS; | |
598 | else if (is_combined_parallel (region)) | |
629b3d75 MJ |
599 | { |
600 | switch (region->inner->type) | |
601 | { | |
602 | case GIMPLE_OMP_FOR: | |
603 | gcc_assert (region->inner->sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); | |
604 | switch (region->inner->sched_kind) | |
605 | { | |
606 | case OMP_CLAUSE_SCHEDULE_RUNTIME: | |
0b887b75 JJ |
607 | /* For lastprivate(conditional:), our implementation |
608 | requires monotonic behavior. */ | |
609 | if (region->inner->has_lastprivate_conditional != 0) | |
610 | start_ix2 = 3; | |
611 | else if ((region->inner->sched_modifiers | |
612 | & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0) | |
28567c40 JJ |
613 | start_ix2 = 6; |
614 | else if ((region->inner->sched_modifiers | |
615 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0) | |
616 | start_ix2 = 7; | |
617 | else | |
618 | start_ix2 = 3; | |
629b3d75 MJ |
619 | break; |
620 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: | |
621 | case OMP_CLAUSE_SCHEDULE_GUIDED: | |
28567c40 | 622 | if ((region->inner->sched_modifiers |
0b887b75 JJ |
623 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 |
624 | && !region->inner->has_lastprivate_conditional) | |
629b3d75 MJ |
625 | { |
626 | start_ix2 = 3 + region->inner->sched_kind; | |
627 | break; | |
628 | } | |
629 | /* FALLTHRU */ | |
630 | default: | |
631 | start_ix2 = region->inner->sched_kind; | |
632 | break; | |
633 | } | |
634 | start_ix2 += (int) BUILT_IN_GOMP_PARALLEL_LOOP_STATIC; | |
635 | start_ix = (enum built_in_function) start_ix2; | |
636 | break; | |
637 | case GIMPLE_OMP_SECTIONS: | |
638 | start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS; | |
639 | break; | |
640 | default: | |
641 | gcc_unreachable (); | |
642 | } | |
643 | } | |
644 | ||
645 | /* By default, the value of NUM_THREADS is zero (selected at run time) | |
646 | and there is no conditional. */ | |
647 | cond = NULL_TREE; | |
648 | val = build_int_cst (unsigned_type_node, 0); | |
649 | flags = build_int_cst (unsigned_type_node, 0); | |
650 | ||
651 | c = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
652 | if (c) | |
653 | cond = OMP_CLAUSE_IF_EXPR (c); | |
654 | ||
655 | c = omp_find_clause (clauses, OMP_CLAUSE_NUM_THREADS); | |
656 | if (c) | |
657 | { | |
658 | val = OMP_CLAUSE_NUM_THREADS_EXPR (c); | |
659 | clause_loc = OMP_CLAUSE_LOCATION (c); | |
660 | } | |
661 | else | |
662 | clause_loc = gimple_location (entry_stmt); | |
663 | ||
664 | c = omp_find_clause (clauses, OMP_CLAUSE_PROC_BIND); | |
665 | if (c) | |
666 | flags = build_int_cst (unsigned_type_node, OMP_CLAUSE_PROC_BIND_KIND (c)); | |
667 | ||
668 | /* Ensure 'val' is of the correct type. */ | |
669 | val = fold_convert_loc (clause_loc, unsigned_type_node, val); | |
670 | ||
671 | /* If we found the clause 'if (cond)', build either | |
672 | (cond != 0) or (cond ? val : 1u). */ | |
673 | if (cond) | |
674 | { | |
675 | cond = gimple_boolify (cond); | |
676 | ||
677 | if (integer_zerop (val)) | |
678 | val = fold_build2_loc (clause_loc, | |
679 | EQ_EXPR, unsigned_type_node, cond, | |
680 | build_int_cst (TREE_TYPE (cond), 0)); | |
681 | else | |
682 | { | |
683 | basic_block cond_bb, then_bb, else_bb; | |
684 | edge e, e_then, e_else; | |
685 | tree tmp_then, tmp_else, tmp_join, tmp_var; | |
686 | ||
687 | tmp_var = create_tmp_var (TREE_TYPE (val)); | |
688 | if (gimple_in_ssa_p (cfun)) | |
689 | { | |
690 | tmp_then = make_ssa_name (tmp_var); | |
691 | tmp_else = make_ssa_name (tmp_var); | |
692 | tmp_join = make_ssa_name (tmp_var); | |
693 | } | |
694 | else | |
695 | { | |
696 | tmp_then = tmp_var; | |
697 | tmp_else = tmp_var; | |
698 | tmp_join = tmp_var; | |
699 | } | |
700 | ||
701 | e = split_block_after_labels (bb); | |
702 | cond_bb = e->src; | |
703 | bb = e->dest; | |
704 | remove_edge (e); | |
705 | ||
706 | then_bb = create_empty_bb (cond_bb); | |
707 | else_bb = create_empty_bb (then_bb); | |
708 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
709 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
710 | ||
711 | stmt = gimple_build_cond_empty (cond); | |
712 | gsi = gsi_start_bb (cond_bb); | |
713 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
714 | ||
715 | gsi = gsi_start_bb (then_bb); | |
716 | expand_omp_build_assign (&gsi, tmp_then, val, true); | |
717 | ||
718 | gsi = gsi_start_bb (else_bb); | |
719 | expand_omp_build_assign (&gsi, tmp_else, | |
720 | build_int_cst (unsigned_type_node, 1), | |
721 | true); | |
722 | ||
723 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
724 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
725 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
726 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
727 | e_then = make_edge (then_bb, bb, EDGE_FALLTHRU); | |
728 | e_else = make_edge (else_bb, bb, EDGE_FALLTHRU); | |
729 | ||
730 | if (gimple_in_ssa_p (cfun)) | |
731 | { | |
732 | gphi *phi = create_phi_node (tmp_join, bb); | |
733 | add_phi_arg (phi, tmp_then, e_then, UNKNOWN_LOCATION); | |
734 | add_phi_arg (phi, tmp_else, e_else, UNKNOWN_LOCATION); | |
735 | } | |
736 | ||
737 | val = tmp_join; | |
738 | } | |
739 | ||
740 | gsi = gsi_start_bb (bb); | |
741 | val = force_gimple_operand_gsi (&gsi, val, true, NULL_TREE, | |
742 | false, GSI_CONTINUE_LINKING); | |
743 | } | |
744 | ||
65f4b875 | 745 | gsi = gsi_last_nondebug_bb (bb); |
629b3d75 MJ |
746 | t = gimple_omp_parallel_data_arg (entry_stmt); |
747 | if (t == NULL) | |
748 | t1 = null_pointer_node; | |
749 | else | |
750 | t1 = build_fold_addr_expr (t); | |
751 | tree child_fndecl = gimple_omp_parallel_child_fn (entry_stmt); | |
752 | t2 = build_fold_addr_expr (child_fndecl); | |
753 | ||
754 | vec_alloc (args, 4 + vec_safe_length (ws_args)); | |
755 | args->quick_push (t2); | |
756 | args->quick_push (t1); | |
757 | args->quick_push (val); | |
758 | if (ws_args) | |
759 | args->splice (*ws_args); | |
760 | args->quick_push (flags); | |
761 | ||
762 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
763 | builtin_decl_explicit (start_ix), args); | |
764 | ||
28567c40 JJ |
765 | if (rtmp) |
766 | { | |
767 | tree type = TREE_TYPE (OMP_CLAUSE_DECL (rtmp)); | |
768 | t = build2 (MODIFY_EXPR, type, OMP_CLAUSE_DECL (rtmp), | |
769 | fold_convert (type, | |
770 | fold_convert (pointer_sized_int_node, t))); | |
771 | } | |
629b3d75 MJ |
772 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
773 | false, GSI_CONTINUE_LINKING); | |
774 | ||
775 | if (hsa_gen_requested_p () | |
776 | && parallel_needs_hsa_kernel_p (region)) | |
777 | { | |
778 | cgraph_node *child_cnode = cgraph_node::get (child_fndecl); | |
779 | hsa_register_kernel (child_cnode); | |
780 | } | |
781 | } | |
782 | ||
629b3d75 MJ |
783 | /* Build the function call to GOMP_task to actually |
784 | generate the task operation. BB is the block where to insert the code. */ | |
785 | ||
786 | static void | |
787 | expand_task_call (struct omp_region *region, basic_block bb, | |
788 | gomp_task *entry_stmt) | |
789 | { | |
790 | tree t1, t2, t3; | |
791 | gimple_stmt_iterator gsi; | |
792 | location_t loc = gimple_location (entry_stmt); | |
793 | ||
794 | tree clauses = gimple_omp_task_clauses (entry_stmt); | |
795 | ||
796 | tree ifc = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
797 | tree untied = omp_find_clause (clauses, OMP_CLAUSE_UNTIED); | |
798 | tree mergeable = omp_find_clause (clauses, OMP_CLAUSE_MERGEABLE); | |
799 | tree depend = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
800 | tree finalc = omp_find_clause (clauses, OMP_CLAUSE_FINAL); | |
801 | tree priority = omp_find_clause (clauses, OMP_CLAUSE_PRIORITY); | |
802 | ||
803 | unsigned int iflags | |
804 | = (untied ? GOMP_TASK_FLAG_UNTIED : 0) | |
805 | | (mergeable ? GOMP_TASK_FLAG_MERGEABLE : 0) | |
806 | | (depend ? GOMP_TASK_FLAG_DEPEND : 0); | |
807 | ||
808 | bool taskloop_p = gimple_omp_task_taskloop_p (entry_stmt); | |
809 | tree startvar = NULL_TREE, endvar = NULL_TREE, step = NULL_TREE; | |
810 | tree num_tasks = NULL_TREE; | |
811 | bool ull = false; | |
812 | if (taskloop_p) | |
813 | { | |
814 | gimple *g = last_stmt (region->outer->entry); | |
815 | gcc_assert (gimple_code (g) == GIMPLE_OMP_FOR | |
816 | && gimple_omp_for_kind (g) == GF_OMP_FOR_KIND_TASKLOOP); | |
817 | struct omp_for_data fd; | |
818 | omp_extract_for_data (as_a <gomp_for *> (g), &fd, NULL); | |
819 | startvar = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
820 | endvar = omp_find_clause (OMP_CLAUSE_CHAIN (startvar), | |
821 | OMP_CLAUSE__LOOPTEMP_); | |
822 | startvar = OMP_CLAUSE_DECL (startvar); | |
823 | endvar = OMP_CLAUSE_DECL (endvar); | |
824 | step = fold_convert_loc (loc, fd.iter_type, fd.loop.step); | |
825 | if (fd.loop.cond_code == LT_EXPR) | |
826 | iflags |= GOMP_TASK_FLAG_UP; | |
827 | tree tclauses = gimple_omp_for_clauses (g); | |
828 | num_tasks = omp_find_clause (tclauses, OMP_CLAUSE_NUM_TASKS); | |
829 | if (num_tasks) | |
830 | num_tasks = OMP_CLAUSE_NUM_TASKS_EXPR (num_tasks); | |
831 | else | |
832 | { | |
833 | num_tasks = omp_find_clause (tclauses, OMP_CLAUSE_GRAINSIZE); | |
834 | if (num_tasks) | |
835 | { | |
836 | iflags |= GOMP_TASK_FLAG_GRAINSIZE; | |
837 | num_tasks = OMP_CLAUSE_GRAINSIZE_EXPR (num_tasks); | |
838 | } | |
839 | else | |
840 | num_tasks = integer_zero_node; | |
841 | } | |
842 | num_tasks = fold_convert_loc (loc, long_integer_type_node, num_tasks); | |
843 | if (ifc == NULL_TREE) | |
844 | iflags |= GOMP_TASK_FLAG_IF; | |
845 | if (omp_find_clause (tclauses, OMP_CLAUSE_NOGROUP)) | |
846 | iflags |= GOMP_TASK_FLAG_NOGROUP; | |
847 | ull = fd.iter_type == long_long_unsigned_type_node; | |
28567c40 JJ |
848 | if (omp_find_clause (clauses, OMP_CLAUSE_REDUCTION)) |
849 | iflags |= GOMP_TASK_FLAG_REDUCTION; | |
629b3d75 MJ |
850 | } |
851 | else if (priority) | |
852 | iflags |= GOMP_TASK_FLAG_PRIORITY; | |
853 | ||
854 | tree flags = build_int_cst (unsigned_type_node, iflags); | |
855 | ||
856 | tree cond = boolean_true_node; | |
857 | if (ifc) | |
858 | { | |
859 | if (taskloop_p) | |
860 | { | |
861 | tree t = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); | |
862 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, | |
863 | build_int_cst (unsigned_type_node, | |
864 | GOMP_TASK_FLAG_IF), | |
865 | build_int_cst (unsigned_type_node, 0)); | |
866 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, | |
867 | flags, t); | |
868 | } | |
869 | else | |
870 | cond = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); | |
871 | } | |
872 | ||
873 | if (finalc) | |
874 | { | |
875 | tree t = gimple_boolify (OMP_CLAUSE_FINAL_EXPR (finalc)); | |
876 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, | |
877 | build_int_cst (unsigned_type_node, | |
878 | GOMP_TASK_FLAG_FINAL), | |
879 | build_int_cst (unsigned_type_node, 0)); | |
880 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, flags, t); | |
881 | } | |
882 | if (depend) | |
883 | depend = OMP_CLAUSE_DECL (depend); | |
884 | else | |
885 | depend = build_int_cst (ptr_type_node, 0); | |
886 | if (priority) | |
887 | priority = fold_convert (integer_type_node, | |
888 | OMP_CLAUSE_PRIORITY_EXPR (priority)); | |
889 | else | |
890 | priority = integer_zero_node; | |
891 | ||
65f4b875 | 892 | gsi = gsi_last_nondebug_bb (bb); |
629b3d75 MJ |
893 | tree t = gimple_omp_task_data_arg (entry_stmt); |
894 | if (t == NULL) | |
895 | t2 = null_pointer_node; | |
896 | else | |
897 | t2 = build_fold_addr_expr_loc (loc, t); | |
898 | t1 = build_fold_addr_expr_loc (loc, gimple_omp_task_child_fn (entry_stmt)); | |
899 | t = gimple_omp_task_copy_fn (entry_stmt); | |
900 | if (t == NULL) | |
901 | t3 = null_pointer_node; | |
902 | else | |
903 | t3 = build_fold_addr_expr_loc (loc, t); | |
904 | ||
905 | if (taskloop_p) | |
906 | t = build_call_expr (ull | |
907 | ? builtin_decl_explicit (BUILT_IN_GOMP_TASKLOOP_ULL) | |
908 | : builtin_decl_explicit (BUILT_IN_GOMP_TASKLOOP), | |
909 | 11, t1, t2, t3, | |
910 | gimple_omp_task_arg_size (entry_stmt), | |
911 | gimple_omp_task_arg_align (entry_stmt), flags, | |
912 | num_tasks, priority, startvar, endvar, step); | |
913 | else | |
914 | t = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASK), | |
915 | 9, t1, t2, t3, | |
916 | gimple_omp_task_arg_size (entry_stmt), | |
917 | gimple_omp_task_arg_align (entry_stmt), cond, flags, | |
918 | depend, priority); | |
919 | ||
920 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
921 | false, GSI_CONTINUE_LINKING); | |
922 | } | |
923 | ||
28567c40 JJ |
924 | /* Build the function call to GOMP_taskwait_depend to actually |
925 | generate the taskwait operation. BB is the block where to insert the | |
926 | code. */ | |
927 | ||
928 | static void | |
929 | expand_taskwait_call (basic_block bb, gomp_task *entry_stmt) | |
930 | { | |
931 | tree clauses = gimple_omp_task_clauses (entry_stmt); | |
932 | tree depend = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
933 | if (depend == NULL_TREE) | |
934 | return; | |
935 | ||
936 | depend = OMP_CLAUSE_DECL (depend); | |
937 | ||
938 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); | |
939 | tree t | |
940 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASKWAIT_DEPEND), | |
941 | 1, depend); | |
942 | ||
943 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
944 | false, GSI_CONTINUE_LINKING); | |
945 | } | |
946 | ||
947 | /* Build the function call to GOMP_teams_reg to actually | |
948 | generate the host teams operation. REGION is the teams region | |
949 | being expanded. BB is the block where to insert the code. */ | |
950 | ||
951 | static void | |
952 | expand_teams_call (basic_block bb, gomp_teams *entry_stmt) | |
953 | { | |
954 | tree clauses = gimple_omp_teams_clauses (entry_stmt); | |
955 | tree num_teams = omp_find_clause (clauses, OMP_CLAUSE_NUM_TEAMS); | |
956 | if (num_teams == NULL_TREE) | |
957 | num_teams = build_int_cst (unsigned_type_node, 0); | |
958 | else | |
959 | { | |
960 | num_teams = OMP_CLAUSE_NUM_TEAMS_EXPR (num_teams); | |
961 | num_teams = fold_convert (unsigned_type_node, num_teams); | |
962 | } | |
963 | tree thread_limit = omp_find_clause (clauses, OMP_CLAUSE_THREAD_LIMIT); | |
964 | if (thread_limit == NULL_TREE) | |
965 | thread_limit = build_int_cst (unsigned_type_node, 0); | |
966 | else | |
967 | { | |
968 | thread_limit = OMP_CLAUSE_THREAD_LIMIT_EXPR (thread_limit); | |
969 | thread_limit = fold_convert (unsigned_type_node, thread_limit); | |
970 | } | |
971 | ||
972 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); | |
973 | tree t = gimple_omp_teams_data_arg (entry_stmt), t1; | |
974 | if (t == NULL) | |
975 | t1 = null_pointer_node; | |
976 | else | |
977 | t1 = build_fold_addr_expr (t); | |
978 | tree child_fndecl = gimple_omp_teams_child_fn (entry_stmt); | |
979 | tree t2 = build_fold_addr_expr (child_fndecl); | |
980 | ||
28567c40 JJ |
981 | vec<tree, va_gc> *args; |
982 | vec_alloc (args, 5); | |
983 | args->quick_push (t2); | |
984 | args->quick_push (t1); | |
985 | args->quick_push (num_teams); | |
986 | args->quick_push (thread_limit); | |
987 | /* For future extensibility. */ | |
988 | args->quick_push (build_zero_cst (unsigned_type_node)); | |
989 | ||
990 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, | |
991 | builtin_decl_explicit (BUILT_IN_GOMP_TEAMS_REG), | |
992 | args); | |
993 | ||
994 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
995 | false, GSI_CONTINUE_LINKING); | |
996 | } | |
997 | ||
629b3d75 MJ |
998 | /* Chain all the DECLs in LIST by their TREE_CHAIN fields. */ |
999 | ||
1000 | static tree | |
1001 | vec2chain (vec<tree, va_gc> *v) | |
1002 | { | |
1003 | tree chain = NULL_TREE, t; | |
1004 | unsigned ix; | |
1005 | ||
1006 | FOR_EACH_VEC_SAFE_ELT_REVERSE (v, ix, t) | |
1007 | { | |
1008 | DECL_CHAIN (t) = chain; | |
1009 | chain = t; | |
1010 | } | |
1011 | ||
1012 | return chain; | |
1013 | } | |
1014 | ||
1015 | /* Remove barriers in REGION->EXIT's block. Note that this is only | |
1016 | valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region | |
1017 | is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that | |
1018 | left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be | |
1019 | removed. */ | |
1020 | ||
1021 | static void | |
1022 | remove_exit_barrier (struct omp_region *region) | |
1023 | { | |
1024 | gimple_stmt_iterator gsi; | |
1025 | basic_block exit_bb; | |
1026 | edge_iterator ei; | |
1027 | edge e; | |
1028 | gimple *stmt; | |
1029 | int any_addressable_vars = -1; | |
1030 | ||
1031 | exit_bb = region->exit; | |
1032 | ||
1033 | /* If the parallel region doesn't return, we don't have REGION->EXIT | |
1034 | block at all. */ | |
1035 | if (! exit_bb) | |
1036 | return; | |
1037 | ||
1038 | /* The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The | |
1039 | workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of | |
1040 | statements that can appear in between are extremely limited -- no | |
1041 | memory operations at all. Here, we allow nothing at all, so the | |
1042 | only thing we allow to precede this GIMPLE_OMP_RETURN is a label. */ | |
65f4b875 | 1043 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 | 1044 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
65f4b875 | 1045 | gsi_prev_nondebug (&gsi); |
629b3d75 MJ |
1046 | if (!gsi_end_p (gsi) && gimple_code (gsi_stmt (gsi)) != GIMPLE_LABEL) |
1047 | return; | |
1048 | ||
1049 | FOR_EACH_EDGE (e, ei, exit_bb->preds) | |
1050 | { | |
65f4b875 | 1051 | gsi = gsi_last_nondebug_bb (e->src); |
629b3d75 MJ |
1052 | if (gsi_end_p (gsi)) |
1053 | continue; | |
1054 | stmt = gsi_stmt (gsi); | |
1055 | if (gimple_code (stmt) == GIMPLE_OMP_RETURN | |
1056 | && !gimple_omp_return_nowait_p (stmt)) | |
1057 | { | |
1058 | /* OpenMP 3.0 tasks unfortunately prevent this optimization | |
1059 | in many cases. If there could be tasks queued, the barrier | |
1060 | might be needed to let the tasks run before some local | |
1061 | variable of the parallel that the task uses as shared | |
1062 | runs out of scope. The task can be spawned either | |
1063 | from within current function (this would be easy to check) | |
1064 | or from some function it calls and gets passed an address | |
1065 | of such a variable. */ | |
1066 | if (any_addressable_vars < 0) | |
1067 | { | |
1068 | gomp_parallel *parallel_stmt | |
1069 | = as_a <gomp_parallel *> (last_stmt (region->entry)); | |
1070 | tree child_fun = gimple_omp_parallel_child_fn (parallel_stmt); | |
1071 | tree local_decls, block, decl; | |
1072 | unsigned ix; | |
1073 | ||
1074 | any_addressable_vars = 0; | |
1075 | FOR_EACH_LOCAL_DECL (DECL_STRUCT_FUNCTION (child_fun), ix, decl) | |
1076 | if (TREE_ADDRESSABLE (decl)) | |
1077 | { | |
1078 | any_addressable_vars = 1; | |
1079 | break; | |
1080 | } | |
1081 | for (block = gimple_block (stmt); | |
1082 | !any_addressable_vars | |
1083 | && block | |
1084 | && TREE_CODE (block) == BLOCK; | |
1085 | block = BLOCK_SUPERCONTEXT (block)) | |
1086 | { | |
1087 | for (local_decls = BLOCK_VARS (block); | |
1088 | local_decls; | |
1089 | local_decls = DECL_CHAIN (local_decls)) | |
1090 | if (TREE_ADDRESSABLE (local_decls)) | |
1091 | { | |
1092 | any_addressable_vars = 1; | |
1093 | break; | |
1094 | } | |
1095 | if (block == gimple_block (parallel_stmt)) | |
1096 | break; | |
1097 | } | |
1098 | } | |
1099 | if (!any_addressable_vars) | |
1100 | gimple_omp_return_set_nowait (stmt); | |
1101 | } | |
1102 | } | |
1103 | } | |
1104 | ||
1105 | static void | |
1106 | remove_exit_barriers (struct omp_region *region) | |
1107 | { | |
1108 | if (region->type == GIMPLE_OMP_PARALLEL) | |
1109 | remove_exit_barrier (region); | |
1110 | ||
1111 | if (region->inner) | |
1112 | { | |
1113 | region = region->inner; | |
1114 | remove_exit_barriers (region); | |
1115 | while (region->next) | |
1116 | { | |
1117 | region = region->next; | |
1118 | remove_exit_barriers (region); | |
1119 | } | |
1120 | } | |
1121 | } | |
1122 | ||
1123 | /* Optimize omp_get_thread_num () and omp_get_num_threads () | |
1124 | calls. These can't be declared as const functions, but | |
1125 | within one parallel body they are constant, so they can be | |
1126 | transformed there into __builtin_omp_get_{thread_num,num_threads} () | |
1127 | which are declared const. Similarly for task body, except | |
1128 | that in untied task omp_get_thread_num () can change at any task | |
1129 | scheduling point. */ | |
1130 | ||
1131 | static void | |
1132 | optimize_omp_library_calls (gimple *entry_stmt) | |
1133 | { | |
1134 | basic_block bb; | |
1135 | gimple_stmt_iterator gsi; | |
1136 | tree thr_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
1137 | tree thr_num_id = DECL_ASSEMBLER_NAME (thr_num_tree); | |
1138 | tree num_thr_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
1139 | tree num_thr_id = DECL_ASSEMBLER_NAME (num_thr_tree); | |
1140 | bool untied_task = (gimple_code (entry_stmt) == GIMPLE_OMP_TASK | |
1141 | && omp_find_clause (gimple_omp_task_clauses (entry_stmt), | |
1142 | OMP_CLAUSE_UNTIED) != NULL); | |
1143 | ||
1144 | FOR_EACH_BB_FN (bb, cfun) | |
1145 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1146 | { | |
1147 | gimple *call = gsi_stmt (gsi); | |
1148 | tree decl; | |
1149 | ||
1150 | if (is_gimple_call (call) | |
1151 | && (decl = gimple_call_fndecl (call)) | |
1152 | && DECL_EXTERNAL (decl) | |
1153 | && TREE_PUBLIC (decl) | |
1154 | && DECL_INITIAL (decl) == NULL) | |
1155 | { | |
1156 | tree built_in; | |
1157 | ||
1158 | if (DECL_NAME (decl) == thr_num_id) | |
1159 | { | |
1160 | /* In #pragma omp task untied omp_get_thread_num () can change | |
1161 | during the execution of the task region. */ | |
1162 | if (untied_task) | |
1163 | continue; | |
1164 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
1165 | } | |
1166 | else if (DECL_NAME (decl) == num_thr_id) | |
1167 | built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
1168 | else | |
1169 | continue; | |
1170 | ||
1171 | if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in) | |
1172 | || gimple_call_num_args (call) != 0) | |
1173 | continue; | |
1174 | ||
1175 | if (flag_exceptions && !TREE_NOTHROW (decl)) | |
1176 | continue; | |
1177 | ||
1178 | if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE | |
1179 | || !types_compatible_p (TREE_TYPE (TREE_TYPE (decl)), | |
1180 | TREE_TYPE (TREE_TYPE (built_in)))) | |
1181 | continue; | |
1182 | ||
1183 | gimple_call_set_fndecl (call, built_in); | |
1184 | } | |
1185 | } | |
1186 | } | |
1187 | ||
1188 | /* Callback for expand_omp_build_assign. Return non-NULL if *tp needs to be | |
1189 | regimplified. */ | |
1190 | ||
1191 | static tree | |
1192 | expand_omp_regimplify_p (tree *tp, int *walk_subtrees, void *) | |
1193 | { | |
1194 | tree t = *tp; | |
1195 | ||
1196 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ | |
1197 | if (VAR_P (t) && DECL_HAS_VALUE_EXPR_P (t)) | |
1198 | return t; | |
1199 | ||
1200 | if (TREE_CODE (t) == ADDR_EXPR) | |
1201 | recompute_tree_invariant_for_addr_expr (t); | |
1202 | ||
1203 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); | |
1204 | return NULL_TREE; | |
1205 | } | |
1206 | ||
1207 | /* Prepend or append TO = FROM assignment before or after *GSI_P. */ | |
1208 | ||
1209 | static void | |
1210 | expand_omp_build_assign (gimple_stmt_iterator *gsi_p, tree to, tree from, | |
1211 | bool after) | |
1212 | { | |
1213 | bool simple_p = DECL_P (to) && TREE_ADDRESSABLE (to); | |
1214 | from = force_gimple_operand_gsi (gsi_p, from, simple_p, NULL_TREE, | |
1215 | !after, after ? GSI_CONTINUE_LINKING | |
1216 | : GSI_SAME_STMT); | |
1217 | gimple *stmt = gimple_build_assign (to, from); | |
1218 | if (after) | |
1219 | gsi_insert_after (gsi_p, stmt, GSI_CONTINUE_LINKING); | |
1220 | else | |
1221 | gsi_insert_before (gsi_p, stmt, GSI_SAME_STMT); | |
1222 | if (walk_tree (&from, expand_omp_regimplify_p, NULL, NULL) | |
1223 | || walk_tree (&to, expand_omp_regimplify_p, NULL, NULL)) | |
1224 | { | |
1225 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1226 | gimple_regimplify_operands (stmt, &gsi); | |
1227 | } | |
1228 | } | |
1229 | ||
1230 | /* Expand the OpenMP parallel or task directive starting at REGION. */ | |
1231 | ||
1232 | static void | |
1233 | expand_omp_taskreg (struct omp_region *region) | |
1234 | { | |
1235 | basic_block entry_bb, exit_bb, new_bb; | |
1236 | struct function *child_cfun; | |
1237 | tree child_fn, block, t; | |
1238 | gimple_stmt_iterator gsi; | |
1239 | gimple *entry_stmt, *stmt; | |
1240 | edge e; | |
1241 | vec<tree, va_gc> *ws_args; | |
1242 | ||
1243 | entry_stmt = last_stmt (region->entry); | |
28567c40 JJ |
1244 | if (gimple_code (entry_stmt) == GIMPLE_OMP_TASK |
1245 | && gimple_omp_task_taskwait_p (entry_stmt)) | |
1246 | { | |
1247 | new_bb = region->entry; | |
1248 | gsi = gsi_last_nondebug_bb (region->entry); | |
1249 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK); | |
1250 | gsi_remove (&gsi, true); | |
1251 | expand_taskwait_call (new_bb, as_a <gomp_task *> (entry_stmt)); | |
1252 | return; | |
1253 | } | |
1254 | ||
629b3d75 MJ |
1255 | child_fn = gimple_omp_taskreg_child_fn (entry_stmt); |
1256 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
1257 | ||
1258 | entry_bb = region->entry; | |
1259 | if (gimple_code (entry_stmt) == GIMPLE_OMP_TASK) | |
1260 | exit_bb = region->cont; | |
1261 | else | |
1262 | exit_bb = region->exit; | |
1263 | ||
5e9d6aa4 | 1264 | if (is_combined_parallel (region)) |
629b3d75 MJ |
1265 | ws_args = region->ws_args; |
1266 | else | |
1267 | ws_args = NULL; | |
1268 | ||
1269 | if (child_cfun->cfg) | |
1270 | { | |
1271 | /* Due to inlining, it may happen that we have already outlined | |
1272 | the region, in which case all we need to do is make the | |
1273 | sub-graph unreachable and emit the parallel call. */ | |
1274 | edge entry_succ_e, exit_succ_e; | |
1275 | ||
1276 | entry_succ_e = single_succ_edge (entry_bb); | |
1277 | ||
65f4b875 | 1278 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 | 1279 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_PARALLEL |
28567c40 JJ |
1280 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK |
1281 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TEAMS); | |
629b3d75 MJ |
1282 | gsi_remove (&gsi, true); |
1283 | ||
1284 | new_bb = entry_bb; | |
1285 | if (exit_bb) | |
1286 | { | |
1287 | exit_succ_e = single_succ_edge (exit_bb); | |
1288 | make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU); | |
1289 | } | |
1290 | remove_edge_and_dominated_blocks (entry_succ_e); | |
1291 | } | |
1292 | else | |
1293 | { | |
1294 | unsigned srcidx, dstidx, num; | |
1295 | ||
1296 | /* If the parallel region needs data sent from the parent | |
1297 | function, then the very first statement (except possible | |
1298 | tree profile counter updates) of the parallel body | |
1299 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
1300 | &.OMP_DATA_O is passed as an argument to the child function, | |
1301 | we need to replace it with the argument as seen by the child | |
1302 | function. | |
1303 | ||
1304 | In most cases, this will end up being the identity assignment | |
1305 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had | |
1306 | a function call that has been inlined, the original PARM_DECL | |
1307 | .OMP_DATA_I may have been converted into a different local | |
1308 | variable. In which case, we need to keep the assignment. */ | |
1309 | if (gimple_omp_taskreg_data_arg (entry_stmt)) | |
1310 | { | |
1311 | basic_block entry_succ_bb | |
1312 | = single_succ_p (entry_bb) ? single_succ (entry_bb) | |
1313 | : FALLTHRU_EDGE (entry_bb)->dest; | |
1314 | tree arg; | |
1315 | gimple *parcopy_stmt = NULL; | |
1316 | ||
1317 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
1318 | { | |
1319 | gimple *stmt; | |
1320 | ||
1321 | gcc_assert (!gsi_end_p (gsi)); | |
1322 | stmt = gsi_stmt (gsi); | |
1323 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
1324 | continue; | |
1325 | ||
1326 | if (gimple_num_ops (stmt) == 2) | |
1327 | { | |
1328 | tree arg = gimple_assign_rhs1 (stmt); | |
1329 | ||
1330 | /* We're ignore the subcode because we're | |
1331 | effectively doing a STRIP_NOPS. */ | |
1332 | ||
1333 | if (TREE_CODE (arg) == ADDR_EXPR | |
28567c40 JJ |
1334 | && (TREE_OPERAND (arg, 0) |
1335 | == gimple_omp_taskreg_data_arg (entry_stmt))) | |
629b3d75 MJ |
1336 | { |
1337 | parcopy_stmt = stmt; | |
1338 | break; | |
1339 | } | |
1340 | } | |
1341 | } | |
1342 | ||
1343 | gcc_assert (parcopy_stmt != NULL); | |
1344 | arg = DECL_ARGUMENTS (child_fn); | |
1345 | ||
1346 | if (!gimple_in_ssa_p (cfun)) | |
1347 | { | |
1348 | if (gimple_assign_lhs (parcopy_stmt) == arg) | |
1349 | gsi_remove (&gsi, true); | |
1350 | else | |
1351 | { | |
01914336 | 1352 | /* ?? Is setting the subcode really necessary ?? */ |
629b3d75 MJ |
1353 | gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (arg)); |
1354 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
1355 | } | |
1356 | } | |
1357 | else | |
1358 | { | |
1359 | tree lhs = gimple_assign_lhs (parcopy_stmt); | |
1360 | gcc_assert (SSA_NAME_VAR (lhs) == arg); | |
1361 | /* We'd like to set the rhs to the default def in the child_fn, | |
1362 | but it's too early to create ssa names in the child_fn. | |
1363 | Instead, we set the rhs to the parm. In | |
1364 | move_sese_region_to_fn, we introduce a default def for the | |
1365 | parm, map the parm to it's default def, and once we encounter | |
1366 | this stmt, replace the parm with the default def. */ | |
1367 | gimple_assign_set_rhs1 (parcopy_stmt, arg); | |
1368 | update_stmt (parcopy_stmt); | |
1369 | } | |
1370 | } | |
1371 | ||
1372 | /* Declare local variables needed in CHILD_CFUN. */ | |
1373 | block = DECL_INITIAL (child_fn); | |
1374 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
1375 | /* The gimplifier could record temporaries in parallel/task block | |
1376 | rather than in containing function's local_decls chain, | |
1377 | which would mean cgraph missed finalizing them. Do it now. */ | |
1378 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
1379 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) | |
1380 | varpool_node::finalize_decl (t); | |
1381 | DECL_SAVED_TREE (child_fn) = NULL; | |
1382 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
1383 | gimple_set_body (child_fn, NULL); | |
1384 | TREE_USED (block) = 1; | |
1385 | ||
1386 | /* Reset DECL_CONTEXT on function arguments. */ | |
1387 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
1388 | DECL_CONTEXT (t) = child_fn; | |
1389 | ||
1390 | /* Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK, | |
1391 | so that it can be moved to the child function. */ | |
65f4b875 | 1392 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
1393 | stmt = gsi_stmt (gsi); |
1394 | gcc_assert (stmt && (gimple_code (stmt) == GIMPLE_OMP_PARALLEL | |
28567c40 JJ |
1395 | || gimple_code (stmt) == GIMPLE_OMP_TASK |
1396 | || gimple_code (stmt) == GIMPLE_OMP_TEAMS)); | |
629b3d75 MJ |
1397 | e = split_block (entry_bb, stmt); |
1398 | gsi_remove (&gsi, true); | |
1399 | entry_bb = e->dest; | |
1400 | edge e2 = NULL; | |
28567c40 | 1401 | if (gimple_code (entry_stmt) != GIMPLE_OMP_TASK) |
629b3d75 MJ |
1402 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; |
1403 | else | |
1404 | { | |
1405 | e2 = make_edge (e->src, BRANCH_EDGE (entry_bb)->dest, EDGE_ABNORMAL); | |
1406 | gcc_assert (e2->dest == region->exit); | |
1407 | remove_edge (BRANCH_EDGE (entry_bb)); | |
1408 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e->src); | |
65f4b875 | 1409 | gsi = gsi_last_nondebug_bb (region->exit); |
629b3d75 MJ |
1410 | gcc_assert (!gsi_end_p (gsi) |
1411 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
1412 | gsi_remove (&gsi, true); | |
1413 | } | |
1414 | ||
1415 | /* Convert GIMPLE_OMP_{RETURN,CONTINUE} into a RETURN_EXPR. */ | |
1416 | if (exit_bb) | |
1417 | { | |
65f4b875 | 1418 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
1419 | gcc_assert (!gsi_end_p (gsi) |
1420 | && (gimple_code (gsi_stmt (gsi)) | |
1421 | == (e2 ? GIMPLE_OMP_CONTINUE : GIMPLE_OMP_RETURN))); | |
1422 | stmt = gimple_build_return (NULL); | |
1423 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
1424 | gsi_remove (&gsi, true); | |
1425 | } | |
1426 | ||
1427 | /* Move the parallel region into CHILD_CFUN. */ | |
1428 | ||
1429 | if (gimple_in_ssa_p (cfun)) | |
1430 | { | |
1431 | init_tree_ssa (child_cfun); | |
1432 | init_ssa_operands (child_cfun); | |
1433 | child_cfun->gimple_df->in_ssa_p = true; | |
1434 | block = NULL_TREE; | |
1435 | } | |
1436 | else | |
1437 | block = gimple_block (entry_stmt); | |
1438 | ||
1439 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
1440 | if (exit_bb) | |
1441 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
1442 | if (e2) | |
1443 | { | |
1444 | basic_block dest_bb = e2->dest; | |
1445 | if (!exit_bb) | |
1446 | make_edge (new_bb, dest_bb, EDGE_FALLTHRU); | |
1447 | remove_edge (e2); | |
1448 | set_immediate_dominator (CDI_DOMINATORS, dest_bb, new_bb); | |
1449 | } | |
1450 | /* When the OMP expansion process cannot guarantee an up-to-date | |
01914336 | 1451 | loop tree arrange for the child function to fixup loops. */ |
629b3d75 MJ |
1452 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) |
1453 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
1454 | ||
1455 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
1456 | num = vec_safe_length (child_cfun->local_decls); | |
1457 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
1458 | { | |
1459 | t = (*child_cfun->local_decls)[srcidx]; | |
1460 | if (DECL_CONTEXT (t) == cfun->decl) | |
1461 | continue; | |
1462 | if (srcidx != dstidx) | |
1463 | (*child_cfun->local_decls)[dstidx] = t; | |
1464 | dstidx++; | |
1465 | } | |
1466 | if (dstidx != num) | |
1467 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
1468 | ||
1469 | /* Inform the callgraph about the new function. */ | |
1470 | child_cfun->curr_properties = cfun->curr_properties; | |
1471 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; | |
1472 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; | |
1473 | cgraph_node *node = cgraph_node::get_create (child_fn); | |
1474 | node->parallelized_function = 1; | |
1475 | cgraph_node::add_new_function (child_fn, true); | |
1476 | ||
1477 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) | |
1478 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); | |
1479 | ||
1480 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
1481 | fixed in a following pass. */ | |
1482 | push_cfun (child_cfun); | |
1483 | if (need_asm) | |
9579db35 | 1484 | assign_assembler_name_if_needed (child_fn); |
629b3d75 MJ |
1485 | |
1486 | if (optimize) | |
1487 | optimize_omp_library_calls (entry_stmt); | |
fc06ae0d | 1488 | update_max_bb_count (); |
629b3d75 MJ |
1489 | cgraph_edge::rebuild_edges (); |
1490 | ||
1491 | /* Some EH regions might become dead, see PR34608. If | |
1492 | pass_cleanup_cfg isn't the first pass to happen with the | |
1493 | new child, these dead EH edges might cause problems. | |
1494 | Clean them up now. */ | |
1495 | if (flag_exceptions) | |
1496 | { | |
1497 | basic_block bb; | |
1498 | bool changed = false; | |
1499 | ||
1500 | FOR_EACH_BB_FN (bb, cfun) | |
1501 | changed |= gimple_purge_dead_eh_edges (bb); | |
1502 | if (changed) | |
1503 | cleanup_tree_cfg (); | |
1504 | } | |
1505 | if (gimple_in_ssa_p (cfun)) | |
1506 | update_ssa (TODO_update_ssa); | |
1507 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
1508 | verify_loop_structure (); | |
1509 | pop_cfun (); | |
1510 | ||
1511 | if (dump_file && !gimple_in_ssa_p (cfun)) | |
1512 | { | |
1513 | omp_any_child_fn_dumped = true; | |
1514 | dump_function_header (dump_file, child_fn, dump_flags); | |
1515 | dump_function_to_file (child_fn, dump_file, dump_flags); | |
1516 | } | |
1517 | } | |
1518 | ||
4ccc4e30 JJ |
1519 | adjust_context_and_scope (region, gimple_block (entry_stmt), child_fn); |
1520 | ||
5e9d6aa4 | 1521 | if (gimple_code (entry_stmt) == GIMPLE_OMP_PARALLEL) |
629b3d75 MJ |
1522 | expand_parallel_call (region, new_bb, |
1523 | as_a <gomp_parallel *> (entry_stmt), ws_args); | |
28567c40 JJ |
1524 | else if (gimple_code (entry_stmt) == GIMPLE_OMP_TEAMS) |
1525 | expand_teams_call (new_bb, as_a <gomp_teams *> (entry_stmt)); | |
629b3d75 MJ |
1526 | else |
1527 | expand_task_call (region, new_bb, as_a <gomp_task *> (entry_stmt)); | |
1528 | if (gimple_in_ssa_p (cfun)) | |
1529 | update_ssa (TODO_update_ssa_only_virtuals); | |
1530 | } | |
1531 | ||
1532 | /* Information about members of an OpenACC collapsed loop nest. */ | |
1533 | ||
1534 | struct oacc_collapse | |
1535 | { | |
01914336 | 1536 | tree base; /* Base value. */ |
629b3d75 | 1537 | tree iters; /* Number of steps. */ |
02889d23 CLT |
1538 | tree step; /* Step size. */ |
1539 | tree tile; /* Tile increment (if tiled). */ | |
1540 | tree outer; /* Tile iterator var. */ | |
629b3d75 MJ |
1541 | }; |
1542 | ||
1543 | /* Helper for expand_oacc_for. Determine collapsed loop information. | |
1544 | Fill in COUNTS array. Emit any initialization code before GSI. | |
1545 | Return the calculated outer loop bound of BOUND_TYPE. */ | |
1546 | ||
1547 | static tree | |
1548 | expand_oacc_collapse_init (const struct omp_for_data *fd, | |
1549 | gimple_stmt_iterator *gsi, | |
02889d23 CLT |
1550 | oacc_collapse *counts, tree bound_type, |
1551 | location_t loc) | |
629b3d75 | 1552 | { |
02889d23 | 1553 | tree tiling = fd->tiling; |
629b3d75 MJ |
1554 | tree total = build_int_cst (bound_type, 1); |
1555 | int ix; | |
1556 | ||
1557 | gcc_assert (integer_onep (fd->loop.step)); | |
1558 | gcc_assert (integer_zerop (fd->loop.n1)); | |
1559 | ||
02889d23 CLT |
1560 | /* When tiling, the first operand of the tile clause applies to the |
1561 | innermost loop, and we work outwards from there. Seems | |
1562 | backwards, but whatever. */ | |
1563 | for (ix = fd->collapse; ix--;) | |
629b3d75 MJ |
1564 | { |
1565 | const omp_for_data_loop *loop = &fd->loops[ix]; | |
1566 | ||
1567 | tree iter_type = TREE_TYPE (loop->v); | |
1568 | tree diff_type = iter_type; | |
1569 | tree plus_type = iter_type; | |
1570 | ||
1571 | gcc_assert (loop->cond_code == fd->loop.cond_code); | |
1572 | ||
1573 | if (POINTER_TYPE_P (iter_type)) | |
1574 | plus_type = sizetype; | |
1575 | if (POINTER_TYPE_P (diff_type) || TYPE_UNSIGNED (diff_type)) | |
1576 | diff_type = signed_type_for (diff_type); | |
c31bc4ac TV |
1577 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (integer_type_node)) |
1578 | diff_type = integer_type_node; | |
629b3d75 | 1579 | |
02889d23 CLT |
1580 | if (tiling) |
1581 | { | |
1582 | tree num = build_int_cst (integer_type_node, fd->collapse); | |
1583 | tree loop_no = build_int_cst (integer_type_node, ix); | |
1584 | tree tile = TREE_VALUE (tiling); | |
1585 | gcall *call | |
1586 | = gimple_build_call_internal (IFN_GOACC_TILE, 5, num, loop_no, tile, | |
1587 | /* gwv-outer=*/integer_zero_node, | |
1588 | /* gwv-inner=*/integer_zero_node); | |
1589 | ||
1590 | counts[ix].outer = create_tmp_var (iter_type, ".outer"); | |
1591 | counts[ix].tile = create_tmp_var (diff_type, ".tile"); | |
1592 | gimple_call_set_lhs (call, counts[ix].tile); | |
1593 | gimple_set_location (call, loc); | |
1594 | gsi_insert_before (gsi, call, GSI_SAME_STMT); | |
1595 | ||
1596 | tiling = TREE_CHAIN (tiling); | |
1597 | } | |
1598 | else | |
1599 | { | |
1600 | counts[ix].tile = NULL; | |
1601 | counts[ix].outer = loop->v; | |
1602 | } | |
1603 | ||
629b3d75 MJ |
1604 | tree b = loop->n1; |
1605 | tree e = loop->n2; | |
1606 | tree s = loop->step; | |
1607 | bool up = loop->cond_code == LT_EXPR; | |
1608 | tree dir = build_int_cst (diff_type, up ? +1 : -1); | |
1609 | bool negating; | |
1610 | tree expr; | |
1611 | ||
1612 | b = force_gimple_operand_gsi (gsi, b, true, NULL_TREE, | |
1613 | true, GSI_SAME_STMT); | |
1614 | e = force_gimple_operand_gsi (gsi, e, true, NULL_TREE, | |
1615 | true, GSI_SAME_STMT); | |
1616 | ||
01914336 | 1617 | /* Convert the step, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
1618 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); |
1619 | if (negating) | |
1620 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); | |
1621 | s = fold_convert (diff_type, s); | |
1622 | if (negating) | |
1623 | s = fold_build1 (NEGATE_EXPR, diff_type, s); | |
1624 | s = force_gimple_operand_gsi (gsi, s, true, NULL_TREE, | |
1625 | true, GSI_SAME_STMT); | |
1626 | ||
01914336 | 1627 | /* Determine the range, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
1628 | negating = !up && TYPE_UNSIGNED (iter_type); |
1629 | expr = fold_build2 (MINUS_EXPR, plus_type, | |
1630 | fold_convert (plus_type, negating ? b : e), | |
1631 | fold_convert (plus_type, negating ? e : b)); | |
1632 | expr = fold_convert (diff_type, expr); | |
1633 | if (negating) | |
1634 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); | |
1635 | tree range = force_gimple_operand_gsi | |
1636 | (gsi, expr, true, NULL_TREE, true, GSI_SAME_STMT); | |
1637 | ||
1638 | /* Determine number of iterations. */ | |
1639 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); | |
1640 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); | |
1641 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); | |
1642 | ||
1643 | tree iters = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE, | |
1644 | true, GSI_SAME_STMT); | |
1645 | ||
1646 | counts[ix].base = b; | |
1647 | counts[ix].iters = iters; | |
1648 | counts[ix].step = s; | |
1649 | ||
1650 | total = fold_build2 (MULT_EXPR, bound_type, total, | |
1651 | fold_convert (bound_type, iters)); | |
1652 | } | |
1653 | ||
1654 | return total; | |
1655 | } | |
1656 | ||
02889d23 CLT |
1657 | /* Emit initializers for collapsed loop members. INNER is true if |
1658 | this is for the element loop of a TILE. IVAR is the outer | |
629b3d75 MJ |
1659 | loop iteration variable, from which collapsed loop iteration values |
1660 | are calculated. COUNTS array has been initialized by | |
1661 | expand_oacc_collapse_inits. */ | |
1662 | ||
1663 | static void | |
02889d23 | 1664 | expand_oacc_collapse_vars (const struct omp_for_data *fd, bool inner, |
629b3d75 MJ |
1665 | gimple_stmt_iterator *gsi, |
1666 | const oacc_collapse *counts, tree ivar) | |
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 diff_type = TREE_TYPE (collapse->step); |
1679 | tree plus_type = iter_type; | |
1680 | enum tree_code plus_code = PLUS_EXPR; | |
1681 | tree expr; | |
1682 | ||
1683 | if (POINTER_TYPE_P (iter_type)) | |
1684 | { | |
1685 | plus_code = POINTER_PLUS_EXPR; | |
1686 | plus_type = sizetype; | |
1687 | } | |
1688 | ||
02889d23 CLT |
1689 | expr = ivar; |
1690 | if (ix) | |
1691 | { | |
1692 | tree mod = fold_convert (ivar_type, collapse->iters); | |
1693 | ivar = fold_build2 (TRUNC_DIV_EXPR, ivar_type, expr, mod); | |
1694 | expr = fold_build2 (TRUNC_MOD_EXPR, ivar_type, expr, mod); | |
1695 | ivar = force_gimple_operand_gsi (gsi, ivar, true, NULL_TREE, | |
1696 | true, GSI_SAME_STMT); | |
1697 | } | |
1698 | ||
629b3d75 MJ |
1699 | expr = fold_build2 (MULT_EXPR, diff_type, fold_convert (diff_type, expr), |
1700 | collapse->step); | |
02889d23 CLT |
1701 | expr = fold_build2 (plus_code, iter_type, |
1702 | inner ? collapse->outer : collapse->base, | |
629b3d75 MJ |
1703 | fold_convert (plus_type, expr)); |
1704 | expr = force_gimple_operand_gsi (gsi, expr, false, NULL_TREE, | |
1705 | true, GSI_SAME_STMT); | |
02889d23 | 1706 | gassign *ass = gimple_build_assign (v, expr); |
629b3d75 | 1707 | gsi_insert_before (gsi, ass, GSI_SAME_STMT); |
629b3d75 MJ |
1708 | } |
1709 | } | |
1710 | ||
1711 | /* Helper function for expand_omp_{for_*,simd}. If this is the outermost | |
1712 | of the combined collapse > 1 loop constructs, generate code like: | |
1713 | if (__builtin_expect (N32 cond3 N31, 0)) goto ZERO_ITER_BB; | |
1714 | if (cond3 is <) | |
1715 | adj = STEP3 - 1; | |
1716 | else | |
1717 | adj = STEP3 + 1; | |
1718 | count3 = (adj + N32 - N31) / STEP3; | |
1719 | if (__builtin_expect (N22 cond2 N21, 0)) goto ZERO_ITER_BB; | |
1720 | if (cond2 is <) | |
1721 | adj = STEP2 - 1; | |
1722 | else | |
1723 | adj = STEP2 + 1; | |
1724 | count2 = (adj + N22 - N21) / STEP2; | |
1725 | if (__builtin_expect (N12 cond1 N11, 0)) goto ZERO_ITER_BB; | |
1726 | if (cond1 is <) | |
1727 | adj = STEP1 - 1; | |
1728 | else | |
1729 | adj = STEP1 + 1; | |
1730 | count1 = (adj + N12 - N11) / STEP1; | |
1731 | count = count1 * count2 * count3; | |
1732 | Furthermore, if ZERO_ITER_BB is NULL, create a BB which does: | |
1733 | count = 0; | |
1734 | and set ZERO_ITER_BB to that bb. If this isn't the outermost | |
1735 | of the combined loop constructs, just initialize COUNTS array | |
1736 | from the _looptemp_ clauses. */ | |
1737 | ||
1738 | /* NOTE: It *could* be better to moosh all of the BBs together, | |
1739 | creating one larger BB with all the computation and the unexpected | |
1740 | jump at the end. I.e. | |
1741 | ||
1742 | bool zero3, zero2, zero1, zero; | |
1743 | ||
1744 | zero3 = N32 c3 N31; | |
1745 | count3 = (N32 - N31) /[cl] STEP3; | |
1746 | zero2 = N22 c2 N21; | |
1747 | count2 = (N22 - N21) /[cl] STEP2; | |
1748 | zero1 = N12 c1 N11; | |
1749 | count1 = (N12 - N11) /[cl] STEP1; | |
1750 | zero = zero3 || zero2 || zero1; | |
1751 | count = count1 * count2 * count3; | |
1752 | if (__builtin_expect(zero, false)) goto zero_iter_bb; | |
1753 | ||
1754 | After all, we expect the zero=false, and thus we expect to have to | |
1755 | evaluate all of the comparison expressions, so short-circuiting | |
1756 | oughtn't be a win. Since the condition isn't protecting a | |
1757 | denominator, we're not concerned about divide-by-zero, so we can | |
1758 | fully evaluate count even if a numerator turned out to be wrong. | |
1759 | ||
1760 | It seems like putting this all together would create much better | |
1761 | scheduling opportunities, and less pressure on the chip's branch | |
1762 | predictor. */ | |
1763 | ||
1764 | static void | |
1765 | expand_omp_for_init_counts (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
1766 | basic_block &entry_bb, tree *counts, | |
1767 | basic_block &zero_iter1_bb, int &first_zero_iter1, | |
1768 | basic_block &zero_iter2_bb, int &first_zero_iter2, | |
1769 | basic_block &l2_dom_bb) | |
1770 | { | |
1771 | tree t, type = TREE_TYPE (fd->loop.v); | |
1772 | edge e, ne; | |
1773 | int i; | |
1774 | ||
1775 | /* Collapsed loops need work for expansion into SSA form. */ | |
1776 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
1777 | ||
1778 | if (gimple_omp_for_combined_into_p (fd->for_stmt) | |
1779 | && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
1780 | { | |
1781 | gcc_assert (fd->ordered == 0); | |
1782 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
1783 | isn't supposed to be handled, as the inner loop doesn't | |
1784 | use it. */ | |
1785 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
1786 | OMP_CLAUSE__LOOPTEMP_); | |
1787 | gcc_assert (innerc); | |
1788 | for (i = 0; i < fd->collapse; i++) | |
1789 | { | |
1790 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
1791 | OMP_CLAUSE__LOOPTEMP_); | |
1792 | gcc_assert (innerc); | |
1793 | if (i) | |
1794 | counts[i] = OMP_CLAUSE_DECL (innerc); | |
1795 | else | |
1796 | counts[0] = NULL_TREE; | |
1797 | } | |
1798 | return; | |
1799 | } | |
1800 | ||
1801 | for (i = fd->collapse; i < fd->ordered; i++) | |
1802 | { | |
1803 | tree itype = TREE_TYPE (fd->loops[i].v); | |
1804 | counts[i] = NULL_TREE; | |
1805 | t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
1806 | fold_convert (itype, fd->loops[i].n1), | |
1807 | fold_convert (itype, fd->loops[i].n2)); | |
1808 | if (t && integer_zerop (t)) | |
1809 | { | |
1810 | for (i = fd->collapse; i < fd->ordered; i++) | |
1811 | counts[i] = build_int_cst (type, 0); | |
1812 | break; | |
1813 | } | |
1814 | } | |
1815 | for (i = 0; i < (fd->ordered ? fd->ordered : fd->collapse); i++) | |
1816 | { | |
1817 | tree itype = TREE_TYPE (fd->loops[i].v); | |
1818 | ||
1819 | if (i >= fd->collapse && counts[i]) | |
1820 | continue; | |
1821 | if ((SSA_VAR_P (fd->loop.n2) || i >= fd->collapse) | |
1822 | && ((t = fold_binary (fd->loops[i].cond_code, boolean_type_node, | |
1823 | fold_convert (itype, fd->loops[i].n1), | |
1824 | fold_convert (itype, fd->loops[i].n2))) | |
1825 | == NULL_TREE || !integer_onep (t))) | |
1826 | { | |
1827 | gcond *cond_stmt; | |
1828 | tree n1, n2; | |
1829 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].n1)); | |
1830 | n1 = force_gimple_operand_gsi (gsi, n1, true, NULL_TREE, | |
1831 | true, GSI_SAME_STMT); | |
1832 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].n2)); | |
1833 | n2 = force_gimple_operand_gsi (gsi, n2, true, NULL_TREE, | |
1834 | true, GSI_SAME_STMT); | |
1835 | cond_stmt = gimple_build_cond (fd->loops[i].cond_code, n1, n2, | |
1836 | NULL_TREE, NULL_TREE); | |
1837 | gsi_insert_before (gsi, cond_stmt, GSI_SAME_STMT); | |
1838 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
1839 | expand_omp_regimplify_p, NULL, NULL) | |
1840 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
1841 | expand_omp_regimplify_p, NULL, NULL)) | |
1842 | { | |
1843 | *gsi = gsi_for_stmt (cond_stmt); | |
1844 | gimple_regimplify_operands (cond_stmt, gsi); | |
1845 | } | |
1846 | e = split_block (entry_bb, cond_stmt); | |
1847 | basic_block &zero_iter_bb | |
1848 | = i < fd->collapse ? zero_iter1_bb : zero_iter2_bb; | |
1849 | int &first_zero_iter | |
1850 | = i < fd->collapse ? first_zero_iter1 : first_zero_iter2; | |
1851 | if (zero_iter_bb == NULL) | |
1852 | { | |
1853 | gassign *assign_stmt; | |
1854 | first_zero_iter = i; | |
1855 | zero_iter_bb = create_empty_bb (entry_bb); | |
1856 | add_bb_to_loop (zero_iter_bb, entry_bb->loop_father); | |
1857 | *gsi = gsi_after_labels (zero_iter_bb); | |
1858 | if (i < fd->collapse) | |
1859 | assign_stmt = gimple_build_assign (fd->loop.n2, | |
1860 | build_zero_cst (type)); | |
1861 | else | |
1862 | { | |
1863 | counts[i] = create_tmp_reg (type, ".count"); | |
1864 | assign_stmt | |
1865 | = gimple_build_assign (counts[i], build_zero_cst (type)); | |
1866 | } | |
1867 | gsi_insert_before (gsi, assign_stmt, GSI_SAME_STMT); | |
1868 | set_immediate_dominator (CDI_DOMINATORS, zero_iter_bb, | |
1869 | entry_bb); | |
1870 | } | |
1871 | ne = make_edge (entry_bb, zero_iter_bb, EDGE_FALSE_VALUE); | |
357067f2 | 1872 | ne->probability = profile_probability::very_unlikely (); |
629b3d75 | 1873 | e->flags = EDGE_TRUE_VALUE; |
357067f2 | 1874 | e->probability = ne->probability.invert (); |
629b3d75 MJ |
1875 | if (l2_dom_bb == NULL) |
1876 | l2_dom_bb = entry_bb; | |
1877 | entry_bb = e->dest; | |
65f4b875 | 1878 | *gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
1879 | } |
1880 | ||
1881 | if (POINTER_TYPE_P (itype)) | |
1882 | itype = signed_type_for (itype); | |
1883 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR | |
1884 | ? -1 : 1)); | |
1885 | t = fold_build2 (PLUS_EXPR, itype, | |
1886 | fold_convert (itype, fd->loops[i].step), t); | |
1887 | t = fold_build2 (PLUS_EXPR, itype, t, | |
1888 | fold_convert (itype, fd->loops[i].n2)); | |
1889 | t = fold_build2 (MINUS_EXPR, itype, t, | |
1890 | fold_convert (itype, fd->loops[i].n1)); | |
1891 | /* ?? We could probably use CEIL_DIV_EXPR instead of | |
1892 | TRUNC_DIV_EXPR and adjusting by hand. Unless we can't | |
1893 | generate the same code in the end because generically we | |
1894 | don't know that the values involved must be negative for | |
1895 | GT?? */ | |
1896 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
1897 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
1898 | fold_build1 (NEGATE_EXPR, itype, t), | |
1899 | fold_build1 (NEGATE_EXPR, itype, | |
1900 | fold_convert (itype, | |
1901 | fd->loops[i].step))); | |
1902 | else | |
1903 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, | |
1904 | fold_convert (itype, fd->loops[i].step)); | |
1905 | t = fold_convert (type, t); | |
1906 | if (TREE_CODE (t) == INTEGER_CST) | |
1907 | counts[i] = t; | |
1908 | else | |
1909 | { | |
1910 | if (i < fd->collapse || i != first_zero_iter2) | |
1911 | counts[i] = create_tmp_reg (type, ".count"); | |
1912 | expand_omp_build_assign (gsi, counts[i], t); | |
1913 | } | |
1914 | if (SSA_VAR_P (fd->loop.n2) && i < fd->collapse) | |
1915 | { | |
1916 | if (i == 0) | |
1917 | t = counts[0]; | |
1918 | else | |
1919 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, counts[i]); | |
1920 | expand_omp_build_assign (gsi, fd->loop.n2, t); | |
1921 | } | |
1922 | } | |
1923 | } | |
1924 | ||
1925 | /* Helper function for expand_omp_{for_*,simd}. Generate code like: | |
1926 | T = V; | |
1927 | V3 = N31 + (T % count3) * STEP3; | |
1928 | T = T / count3; | |
1929 | V2 = N21 + (T % count2) * STEP2; | |
1930 | T = T / count2; | |
1931 | V1 = N11 + T * STEP1; | |
1932 | if this loop doesn't have an inner loop construct combined with it. | |
1933 | If it does have an inner loop construct combined with it and the | |
1934 | iteration count isn't known constant, store values from counts array | |
1935 | into its _looptemp_ temporaries instead. */ | |
1936 | ||
1937 | static void | |
1938 | expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi, | |
1939 | tree *counts, gimple *inner_stmt, tree startvar) | |
1940 | { | |
1941 | int i; | |
1942 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
1943 | { | |
1944 | /* If fd->loop.n2 is constant, then no propagation of the counts | |
1945 | is needed, they are constant. */ | |
1946 | if (TREE_CODE (fd->loop.n2) == INTEGER_CST) | |
1947 | return; | |
1948 | ||
1949 | tree clauses = gimple_code (inner_stmt) != GIMPLE_OMP_FOR | |
1950 | ? gimple_omp_taskreg_clauses (inner_stmt) | |
1951 | : gimple_omp_for_clauses (inner_stmt); | |
1952 | /* First two _looptemp_ clauses are for istart/iend, counts[0] | |
1953 | isn't supposed to be handled, as the inner loop doesn't | |
1954 | use it. */ | |
1955 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
1956 | gcc_assert (innerc); | |
1957 | for (i = 0; i < fd->collapse; i++) | |
1958 | { | |
1959 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
1960 | OMP_CLAUSE__LOOPTEMP_); | |
1961 | gcc_assert (innerc); | |
1962 | if (i) | |
1963 | { | |
1964 | tree tem = OMP_CLAUSE_DECL (innerc); | |
1965 | tree t = fold_convert (TREE_TYPE (tem), counts[i]); | |
1966 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
1967 | false, GSI_CONTINUE_LINKING); | |
1968 | gassign *stmt = gimple_build_assign (tem, t); | |
1969 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
1970 | } | |
1971 | } | |
1972 | return; | |
1973 | } | |
1974 | ||
1975 | tree type = TREE_TYPE (fd->loop.v); | |
1976 | tree tem = create_tmp_reg (type, ".tem"); | |
1977 | gassign *stmt = gimple_build_assign (tem, startvar); | |
1978 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
1979 | ||
1980 | for (i = fd->collapse - 1; i >= 0; i--) | |
1981 | { | |
1982 | tree vtype = TREE_TYPE (fd->loops[i].v), itype, t; | |
1983 | itype = vtype; | |
1984 | if (POINTER_TYPE_P (vtype)) | |
1985 | itype = signed_type_for (vtype); | |
1986 | if (i != 0) | |
1987 | t = fold_build2 (TRUNC_MOD_EXPR, type, tem, counts[i]); | |
1988 | else | |
1989 | t = tem; | |
1990 | t = fold_convert (itype, t); | |
1991 | t = fold_build2 (MULT_EXPR, itype, t, | |
1992 | fold_convert (itype, fd->loops[i].step)); | |
1993 | if (POINTER_TYPE_P (vtype)) | |
1994 | t = fold_build_pointer_plus (fd->loops[i].n1, t); | |
1995 | else | |
1996 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); | |
1997 | t = force_gimple_operand_gsi (gsi, t, | |
1998 | DECL_P (fd->loops[i].v) | |
1999 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
2000 | NULL_TREE, false, | |
2001 | GSI_CONTINUE_LINKING); | |
2002 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
2003 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2004 | if (i != 0) | |
2005 | { | |
2006 | t = fold_build2 (TRUNC_DIV_EXPR, type, tem, counts[i]); | |
2007 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, | |
2008 | false, GSI_CONTINUE_LINKING); | |
2009 | stmt = gimple_build_assign (tem, t); | |
2010 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); | |
2011 | } | |
2012 | } | |
2013 | } | |
2014 | ||
2015 | /* Helper function for expand_omp_for_*. Generate code like: | |
2016 | L10: | |
2017 | V3 += STEP3; | |
2018 | if (V3 cond3 N32) goto BODY_BB; else goto L11; | |
2019 | L11: | |
2020 | V3 = N31; | |
2021 | V2 += STEP2; | |
2022 | if (V2 cond2 N22) goto BODY_BB; else goto L12; | |
2023 | L12: | |
2024 | V2 = N21; | |
2025 | V1 += STEP1; | |
2026 | goto BODY_BB; */ | |
2027 | ||
2028 | static basic_block | |
2029 | extract_omp_for_update_vars (struct omp_for_data *fd, basic_block cont_bb, | |
2030 | basic_block body_bb) | |
2031 | { | |
2032 | basic_block last_bb, bb, collapse_bb = NULL; | |
2033 | int i; | |
2034 | gimple_stmt_iterator gsi; | |
2035 | edge e; | |
2036 | tree t; | |
2037 | gimple *stmt; | |
2038 | ||
2039 | last_bb = cont_bb; | |
2040 | for (i = fd->collapse - 1; i >= 0; i--) | |
2041 | { | |
2042 | tree vtype = TREE_TYPE (fd->loops[i].v); | |
2043 | ||
2044 | bb = create_empty_bb (last_bb); | |
2045 | add_bb_to_loop (bb, last_bb->loop_father); | |
2046 | gsi = gsi_start_bb (bb); | |
2047 | ||
2048 | if (i < fd->collapse - 1) | |
2049 | { | |
2050 | e = make_edge (last_bb, bb, EDGE_FALSE_VALUE); | |
357067f2 | 2051 | e->probability = profile_probability::guessed_always ().apply_scale (1, 8); |
629b3d75 MJ |
2052 | |
2053 | t = fd->loops[i + 1].n1; | |
2054 | t = force_gimple_operand_gsi (&gsi, t, | |
2055 | DECL_P (fd->loops[i + 1].v) | |
2056 | && TREE_ADDRESSABLE (fd->loops[i | |
2057 | + 1].v), | |
2058 | NULL_TREE, false, | |
2059 | GSI_CONTINUE_LINKING); | |
2060 | stmt = gimple_build_assign (fd->loops[i + 1].v, t); | |
2061 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2062 | } | |
2063 | else | |
2064 | collapse_bb = bb; | |
2065 | ||
2066 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); | |
2067 | ||
2068 | if (POINTER_TYPE_P (vtype)) | |
2069 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); | |
2070 | else | |
2071 | t = fold_build2 (PLUS_EXPR, vtype, fd->loops[i].v, fd->loops[i].step); | |
2072 | t = force_gimple_operand_gsi (&gsi, t, | |
2073 | DECL_P (fd->loops[i].v) | |
2074 | && TREE_ADDRESSABLE (fd->loops[i].v), | |
2075 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
2076 | stmt = gimple_build_assign (fd->loops[i].v, t); | |
2077 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2078 | ||
2079 | if (i > 0) | |
2080 | { | |
2081 | t = fd->loops[i].n2; | |
2082 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
2083 | false, GSI_CONTINUE_LINKING); | |
2084 | tree v = fd->loops[i].v; | |
2085 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
2086 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
2087 | false, GSI_CONTINUE_LINKING); | |
2088 | t = fold_build2 (fd->loops[i].cond_code, boolean_type_node, v, t); | |
2089 | stmt = gimple_build_cond_empty (t); | |
2090 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
d1ffbd43 JJ |
2091 | if (walk_tree (gimple_cond_lhs_ptr (as_a <gcond *> (stmt)), |
2092 | expand_omp_regimplify_p, NULL, NULL) | |
2093 | || walk_tree (gimple_cond_rhs_ptr (as_a <gcond *> (stmt)), | |
2094 | expand_omp_regimplify_p, NULL, NULL)) | |
2095 | gimple_regimplify_operands (stmt, &gsi); | |
629b3d75 | 2096 | e = make_edge (bb, body_bb, EDGE_TRUE_VALUE); |
357067f2 | 2097 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); |
629b3d75 MJ |
2098 | } |
2099 | else | |
2100 | make_edge (bb, body_bb, EDGE_FALLTHRU); | |
2101 | last_bb = bb; | |
2102 | } | |
2103 | ||
2104 | return collapse_bb; | |
2105 | } | |
2106 | ||
2107 | /* Expand #pragma omp ordered depend(source). */ | |
2108 | ||
2109 | static void | |
2110 | expand_omp_ordered_source (gimple_stmt_iterator *gsi, struct omp_for_data *fd, | |
2111 | tree *counts, location_t loc) | |
2112 | { | |
2113 | enum built_in_function source_ix | |
2114 | = fd->iter_type == long_integer_type_node | |
2115 | ? BUILT_IN_GOMP_DOACROSS_POST : BUILT_IN_GOMP_DOACROSS_ULL_POST; | |
2116 | gimple *g | |
2117 | = gimple_build_call (builtin_decl_explicit (source_ix), 1, | |
2118 | build_fold_addr_expr (counts[fd->ordered])); | |
2119 | gimple_set_location (g, loc); | |
2120 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
2121 | } | |
2122 | ||
2123 | /* Expand a single depend from #pragma omp ordered depend(sink:...). */ | |
2124 | ||
2125 | static void | |
2126 | expand_omp_ordered_sink (gimple_stmt_iterator *gsi, struct omp_for_data *fd, | |
2127 | tree *counts, tree c, location_t loc) | |
2128 | { | |
2129 | auto_vec<tree, 10> args; | |
2130 | enum built_in_function sink_ix | |
2131 | = fd->iter_type == long_integer_type_node | |
2132 | ? BUILT_IN_GOMP_DOACROSS_WAIT : BUILT_IN_GOMP_DOACROSS_ULL_WAIT; | |
2133 | tree t, off, coff = NULL_TREE, deps = OMP_CLAUSE_DECL (c), cond = NULL_TREE; | |
2134 | int i; | |
2135 | gimple_stmt_iterator gsi2 = *gsi; | |
2136 | bool warned_step = false; | |
2137 | ||
2138 | for (i = 0; i < fd->ordered; i++) | |
2139 | { | |
2140 | tree step = NULL_TREE; | |
2141 | off = TREE_PURPOSE (deps); | |
2142 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) | |
2143 | { | |
2144 | step = TREE_OPERAND (off, 1); | |
2145 | off = TREE_OPERAND (off, 0); | |
2146 | } | |
2147 | if (!integer_zerop (off)) | |
2148 | { | |
2149 | gcc_assert (fd->loops[i].cond_code == LT_EXPR | |
2150 | || fd->loops[i].cond_code == GT_EXPR); | |
2151 | bool forward = fd->loops[i].cond_code == LT_EXPR; | |
2152 | if (step) | |
2153 | { | |
2154 | /* Non-simple Fortran DO loops. If step is variable, | |
2155 | we don't know at compile even the direction, so can't | |
2156 | warn. */ | |
2157 | if (TREE_CODE (step) != INTEGER_CST) | |
2158 | break; | |
2159 | forward = tree_int_cst_sgn (step) != -1; | |
2160 | } | |
2161 | if (forward ^ OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
90a0bf4e JJ |
2162 | warning_at (loc, 0, "%<depend%> clause with %<sink%> modifier " |
2163 | "waiting for lexically later iteration"); | |
629b3d75 MJ |
2164 | break; |
2165 | } | |
2166 | deps = TREE_CHAIN (deps); | |
2167 | } | |
2168 | /* If all offsets corresponding to the collapsed loops are zero, | |
2169 | this depend clause can be ignored. FIXME: but there is still a | |
2170 | flush needed. We need to emit one __sync_synchronize () for it | |
2171 | though (perhaps conditionally)? Solve this together with the | |
2172 | conservative dependence folding optimization. | |
2173 | if (i >= fd->collapse) | |
2174 | return; */ | |
2175 | ||
2176 | deps = OMP_CLAUSE_DECL (c); | |
2177 | gsi_prev (&gsi2); | |
2178 | edge e1 = split_block (gsi_bb (gsi2), gsi_stmt (gsi2)); | |
2179 | edge e2 = split_block_after_labels (e1->dest); | |
2180 | ||
2181 | gsi2 = gsi_after_labels (e1->dest); | |
2182 | *gsi = gsi_last_bb (e1->src); | |
2183 | for (i = 0; i < fd->ordered; i++) | |
2184 | { | |
2185 | tree itype = TREE_TYPE (fd->loops[i].v); | |
2186 | tree step = NULL_TREE; | |
2187 | tree orig_off = NULL_TREE; | |
2188 | if (POINTER_TYPE_P (itype)) | |
2189 | itype = sizetype; | |
2190 | if (i) | |
2191 | deps = TREE_CHAIN (deps); | |
2192 | off = TREE_PURPOSE (deps); | |
2193 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) | |
2194 | { | |
2195 | step = TREE_OPERAND (off, 1); | |
2196 | off = TREE_OPERAND (off, 0); | |
2197 | gcc_assert (fd->loops[i].cond_code == LT_EXPR | |
2198 | && integer_onep (fd->loops[i].step) | |
2199 | && !POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))); | |
2200 | } | |
2201 | tree s = fold_convert_loc (loc, itype, step ? step : fd->loops[i].step); | |
2202 | if (step) | |
2203 | { | |
2204 | off = fold_convert_loc (loc, itype, off); | |
2205 | orig_off = off; | |
2206 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); | |
2207 | } | |
2208 | ||
2209 | if (integer_zerop (off)) | |
2210 | t = boolean_true_node; | |
2211 | else | |
2212 | { | |
2213 | tree a; | |
2214 | tree co = fold_convert_loc (loc, itype, off); | |
2215 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
2216 | { | |
2217 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
2218 | co = fold_build1_loc (loc, NEGATE_EXPR, itype, co); | |
2219 | a = fold_build2_loc (loc, POINTER_PLUS_EXPR, | |
2220 | TREE_TYPE (fd->loops[i].v), fd->loops[i].v, | |
2221 | co); | |
2222 | } | |
2223 | else if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
2224 | a = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
2225 | fd->loops[i].v, co); | |
2226 | else | |
2227 | a = fold_build2_loc (loc, PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
2228 | fd->loops[i].v, co); | |
2229 | if (step) | |
2230 | { | |
2231 | tree t1, t2; | |
2232 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
2233 | t1 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, | |
2234 | fd->loops[i].n1); | |
2235 | else | |
2236 | t1 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
2237 | fd->loops[i].n2); | |
2238 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
2239 | t2 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
2240 | fd->loops[i].n2); | |
2241 | else | |
2242 | t2 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, | |
2243 | fd->loops[i].n1); | |
2244 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, | |
2245 | step, build_int_cst (TREE_TYPE (step), 0)); | |
2246 | if (TREE_CODE (step) != INTEGER_CST) | |
2247 | { | |
2248 | t1 = unshare_expr (t1); | |
2249 | t1 = force_gimple_operand_gsi (gsi, t1, true, NULL_TREE, | |
2250 | false, GSI_CONTINUE_LINKING); | |
2251 | t2 = unshare_expr (t2); | |
2252 | t2 = force_gimple_operand_gsi (gsi, t2, true, NULL_TREE, | |
2253 | false, GSI_CONTINUE_LINKING); | |
2254 | } | |
2255 | t = fold_build3_loc (loc, COND_EXPR, boolean_type_node, | |
2256 | t, t2, t1); | |
2257 | } | |
2258 | else if (fd->loops[i].cond_code == LT_EXPR) | |
2259 | { | |
2260 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
2261 | t = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, | |
2262 | fd->loops[i].n1); | |
2263 | else | |
2264 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, | |
2265 | fd->loops[i].n2); | |
2266 | } | |
2267 | else if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
2268 | t = fold_build2_loc (loc, GT_EXPR, boolean_type_node, a, | |
2269 | fd->loops[i].n2); | |
2270 | else | |
2271 | t = fold_build2_loc (loc, LE_EXPR, boolean_type_node, a, | |
2272 | fd->loops[i].n1); | |
2273 | } | |
2274 | if (cond) | |
2275 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, cond, t); | |
2276 | else | |
2277 | cond = t; | |
2278 | ||
2279 | off = fold_convert_loc (loc, itype, off); | |
2280 | ||
2281 | if (step | |
2282 | || (fd->loops[i].cond_code == LT_EXPR | |
2283 | ? !integer_onep (fd->loops[i].step) | |
2284 | : !integer_minus_onep (fd->loops[i].step))) | |
2285 | { | |
2286 | if (step == NULL_TREE | |
2287 | && TYPE_UNSIGNED (itype) | |
2288 | && fd->loops[i].cond_code == GT_EXPR) | |
2289 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, off, | |
2290 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
2291 | s)); | |
2292 | else | |
2293 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, | |
2294 | orig_off ? orig_off : off, s); | |
2295 | t = fold_build2_loc (loc, EQ_EXPR, boolean_type_node, t, | |
2296 | build_int_cst (itype, 0)); | |
2297 | if (integer_zerop (t) && !warned_step) | |
2298 | { | |
90a0bf4e JJ |
2299 | warning_at (loc, 0, "%<depend%> clause with %<sink%> modifier " |
2300 | "refers to iteration never in the iteration " | |
2301 | "space"); | |
629b3d75 MJ |
2302 | warned_step = true; |
2303 | } | |
2304 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, | |
2305 | cond, t); | |
2306 | } | |
2307 | ||
2308 | if (i <= fd->collapse - 1 && fd->collapse > 1) | |
2309 | t = fd->loop.v; | |
2310 | else if (counts[i]) | |
2311 | t = counts[i]; | |
2312 | else | |
2313 | { | |
2314 | t = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
2315 | fd->loops[i].v, fd->loops[i].n1); | |
2316 | t = fold_convert_loc (loc, fd->iter_type, t); | |
2317 | } | |
2318 | if (step) | |
2319 | /* We have divided off by step already earlier. */; | |
2320 | else if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) | |
2321 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, | |
2322 | fold_build1_loc (loc, NEGATE_EXPR, itype, | |
2323 | s)); | |
2324 | else | |
2325 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); | |
2326 | if (OMP_CLAUSE_DEPEND_SINK_NEGATIVE (deps)) | |
2327 | off = fold_build1_loc (loc, NEGATE_EXPR, itype, off); | |
2328 | off = fold_convert_loc (loc, fd->iter_type, off); | |
2329 | if (i <= fd->collapse - 1 && fd->collapse > 1) | |
2330 | { | |
2331 | if (i) | |
2332 | off = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, coff, | |
2333 | off); | |
2334 | if (i < fd->collapse - 1) | |
2335 | { | |
2336 | coff = fold_build2_loc (loc, MULT_EXPR, fd->iter_type, off, | |
2337 | counts[i]); | |
2338 | continue; | |
2339 | } | |
2340 | } | |
2341 | off = unshare_expr (off); | |
2342 | t = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, t, off); | |
2343 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, | |
2344 | true, GSI_SAME_STMT); | |
2345 | args.safe_push (t); | |
2346 | } | |
2347 | gimple *g = gimple_build_call_vec (builtin_decl_explicit (sink_ix), args); | |
2348 | gimple_set_location (g, loc); | |
2349 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); | |
2350 | ||
2351 | cond = unshare_expr (cond); | |
2352 | cond = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE, false, | |
2353 | GSI_CONTINUE_LINKING); | |
2354 | gsi_insert_after (gsi, gimple_build_cond_empty (cond), GSI_NEW_STMT); | |
2355 | edge e3 = make_edge (e1->src, e2->dest, EDGE_FALSE_VALUE); | |
357067f2 JH |
2356 | e3->probability = profile_probability::guessed_always ().apply_scale (1, 8); |
2357 | e1->probability = e3->probability.invert (); | |
629b3d75 MJ |
2358 | e1->flags = EDGE_TRUE_VALUE; |
2359 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e1->src); | |
2360 | ||
2361 | *gsi = gsi_after_labels (e2->dest); | |
2362 | } | |
2363 | ||
2364 | /* Expand all #pragma omp ordered depend(source) and | |
2365 | #pragma omp ordered depend(sink:...) constructs in the current | |
2366 | #pragma omp for ordered(n) region. */ | |
2367 | ||
2368 | static void | |
2369 | expand_omp_ordered_source_sink (struct omp_region *region, | |
2370 | struct omp_for_data *fd, tree *counts, | |
2371 | basic_block cont_bb) | |
2372 | { | |
2373 | struct omp_region *inner; | |
2374 | int i; | |
2375 | for (i = fd->collapse - 1; i < fd->ordered; i++) | |
2376 | if (i == fd->collapse - 1 && fd->collapse > 1) | |
2377 | counts[i] = NULL_TREE; | |
2378 | else if (i >= fd->collapse && !cont_bb) | |
2379 | counts[i] = build_zero_cst (fd->iter_type); | |
2380 | else if (!POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v)) | |
2381 | && integer_onep (fd->loops[i].step)) | |
2382 | counts[i] = NULL_TREE; | |
2383 | else | |
2384 | counts[i] = create_tmp_var (fd->iter_type, ".orditer"); | |
2385 | tree atype | |
2386 | = build_array_type_nelts (fd->iter_type, fd->ordered - fd->collapse + 1); | |
2387 | counts[fd->ordered] = create_tmp_var (atype, ".orditera"); | |
2388 | TREE_ADDRESSABLE (counts[fd->ordered]) = 1; | |
2389 | ||
2390 | for (inner = region->inner; inner; inner = inner->next) | |
2391 | if (inner->type == GIMPLE_OMP_ORDERED) | |
2392 | { | |
2393 | gomp_ordered *ord_stmt = inner->ord_stmt; | |
2394 | gimple_stmt_iterator gsi = gsi_for_stmt (ord_stmt); | |
2395 | location_t loc = gimple_location (ord_stmt); | |
2396 | tree c; | |
2397 | for (c = gimple_omp_ordered_clauses (ord_stmt); | |
2398 | c; c = OMP_CLAUSE_CHAIN (c)) | |
2399 | if (OMP_CLAUSE_DEPEND_KIND (c) == OMP_CLAUSE_DEPEND_SOURCE) | |
2400 | break; | |
2401 | if (c) | |
2402 | expand_omp_ordered_source (&gsi, fd, counts, loc); | |
2403 | for (c = gimple_omp_ordered_clauses (ord_stmt); | |
2404 | c; c = OMP_CLAUSE_CHAIN (c)) | |
2405 | if (OMP_CLAUSE_DEPEND_KIND (c) == OMP_CLAUSE_DEPEND_SINK) | |
2406 | expand_omp_ordered_sink (&gsi, fd, counts, c, loc); | |
2407 | gsi_remove (&gsi, true); | |
2408 | } | |
2409 | } | |
2410 | ||
2411 | /* Wrap the body into fd->ordered - fd->collapse loops that aren't | |
2412 | collapsed. */ | |
2413 | ||
2414 | static basic_block | |
2415 | expand_omp_for_ordered_loops (struct omp_for_data *fd, tree *counts, | |
2416 | basic_block cont_bb, basic_block body_bb, | |
2417 | bool ordered_lastprivate) | |
2418 | { | |
2419 | if (fd->ordered == fd->collapse) | |
2420 | return cont_bb; | |
2421 | ||
2422 | if (!cont_bb) | |
2423 | { | |
2424 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
2425 | for (int i = fd->collapse; i < fd->ordered; i++) | |
2426 | { | |
2427 | tree type = TREE_TYPE (fd->loops[i].v); | |
2428 | tree n1 = fold_convert (type, fd->loops[i].n1); | |
2429 | expand_omp_build_assign (&gsi, fd->loops[i].v, n1); | |
2430 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
2431 | size_int (i - fd->collapse + 1), | |
2432 | NULL_TREE, NULL_TREE); | |
2433 | expand_omp_build_assign (&gsi, aref, build_zero_cst (fd->iter_type)); | |
2434 | } | |
2435 | return NULL; | |
2436 | } | |
2437 | ||
2438 | for (int i = fd->ordered - 1; i >= fd->collapse; i--) | |
2439 | { | |
2440 | tree t, type = TREE_TYPE (fd->loops[i].v); | |
2441 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
2442 | expand_omp_build_assign (&gsi, fd->loops[i].v, | |
2443 | fold_convert (type, fd->loops[i].n1)); | |
2444 | if (counts[i]) | |
2445 | expand_omp_build_assign (&gsi, counts[i], | |
2446 | build_zero_cst (fd->iter_type)); | |
2447 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
2448 | size_int (i - fd->collapse + 1), | |
2449 | NULL_TREE, NULL_TREE); | |
2450 | expand_omp_build_assign (&gsi, aref, build_zero_cst (fd->iter_type)); | |
2451 | if (!gsi_end_p (gsi)) | |
2452 | gsi_prev (&gsi); | |
2453 | else | |
2454 | gsi = gsi_last_bb (body_bb); | |
2455 | edge e1 = split_block (body_bb, gsi_stmt (gsi)); | |
2456 | basic_block new_body = e1->dest; | |
2457 | if (body_bb == cont_bb) | |
2458 | cont_bb = new_body; | |
2459 | edge e2 = NULL; | |
2460 | basic_block new_header; | |
2461 | if (EDGE_COUNT (cont_bb->preds) > 0) | |
2462 | { | |
2463 | gsi = gsi_last_bb (cont_bb); | |
2464 | if (POINTER_TYPE_P (type)) | |
2465 | t = fold_build_pointer_plus (fd->loops[i].v, | |
2466 | fold_convert (sizetype, | |
2467 | fd->loops[i].step)); | |
2468 | else | |
2469 | t = fold_build2 (PLUS_EXPR, type, fd->loops[i].v, | |
2470 | fold_convert (type, fd->loops[i].step)); | |
2471 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
2472 | if (counts[i]) | |
2473 | { | |
2474 | t = fold_build2 (PLUS_EXPR, fd->iter_type, counts[i], | |
2475 | build_int_cst (fd->iter_type, 1)); | |
2476 | expand_omp_build_assign (&gsi, counts[i], t); | |
2477 | t = counts[i]; | |
2478 | } | |
2479 | else | |
2480 | { | |
2481 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
2482 | fd->loops[i].v, fd->loops[i].n1); | |
2483 | t = fold_convert (fd->iter_type, t); | |
2484 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
2485 | true, GSI_SAME_STMT); | |
2486 | } | |
2487 | aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
2488 | size_int (i - fd->collapse + 1), | |
2489 | NULL_TREE, NULL_TREE); | |
2490 | expand_omp_build_assign (&gsi, aref, t); | |
2491 | gsi_prev (&gsi); | |
2492 | e2 = split_block (cont_bb, gsi_stmt (gsi)); | |
2493 | new_header = e2->dest; | |
2494 | } | |
2495 | else | |
2496 | new_header = cont_bb; | |
2497 | gsi = gsi_after_labels (new_header); | |
2498 | tree v = force_gimple_operand_gsi (&gsi, fd->loops[i].v, true, NULL_TREE, | |
2499 | true, GSI_SAME_STMT); | |
2500 | tree n2 | |
2501 | = force_gimple_operand_gsi (&gsi, fold_convert (type, fd->loops[i].n2), | |
2502 | true, NULL_TREE, true, GSI_SAME_STMT); | |
2503 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, n2); | |
2504 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_NEW_STMT); | |
2505 | edge e3 = split_block (new_header, gsi_stmt (gsi)); | |
2506 | cont_bb = e3->dest; | |
2507 | remove_edge (e1); | |
2508 | make_edge (body_bb, new_header, EDGE_FALLTHRU); | |
2509 | e3->flags = EDGE_FALSE_VALUE; | |
357067f2 | 2510 | e3->probability = profile_probability::guessed_always ().apply_scale (1, 8); |
629b3d75 | 2511 | e1 = make_edge (new_header, new_body, EDGE_TRUE_VALUE); |
357067f2 | 2512 | e1->probability = e3->probability.invert (); |
629b3d75 MJ |
2513 | |
2514 | set_immediate_dominator (CDI_DOMINATORS, new_header, body_bb); | |
2515 | set_immediate_dominator (CDI_DOMINATORS, new_body, new_header); | |
2516 | ||
2517 | if (e2) | |
2518 | { | |
99b1c316 | 2519 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
2520 | loop->header = new_header; |
2521 | loop->latch = e2->src; | |
2522 | add_loop (loop, body_bb->loop_father); | |
2523 | } | |
2524 | } | |
2525 | ||
2526 | /* If there are any lastprivate clauses and it is possible some loops | |
2527 | might have zero iterations, ensure all the decls are initialized, | |
2528 | otherwise we could crash evaluating C++ class iterators with lastprivate | |
2529 | clauses. */ | |
2530 | bool need_inits = false; | |
2531 | for (int i = fd->collapse; ordered_lastprivate && i < fd->ordered; i++) | |
2532 | if (need_inits) | |
2533 | { | |
2534 | tree type = TREE_TYPE (fd->loops[i].v); | |
2535 | gimple_stmt_iterator gsi = gsi_after_labels (body_bb); | |
2536 | expand_omp_build_assign (&gsi, fd->loops[i].v, | |
2537 | fold_convert (type, fd->loops[i].n1)); | |
2538 | } | |
2539 | else | |
2540 | { | |
2541 | tree type = TREE_TYPE (fd->loops[i].v); | |
2542 | tree this_cond = fold_build2 (fd->loops[i].cond_code, | |
2543 | boolean_type_node, | |
2544 | fold_convert (type, fd->loops[i].n1), | |
2545 | fold_convert (type, fd->loops[i].n2)); | |
2546 | if (!integer_onep (this_cond)) | |
2547 | need_inits = true; | |
2548 | } | |
2549 | ||
2550 | return cont_bb; | |
2551 | } | |
2552 | ||
2553 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
2554 | loop with any schedule. Given parameters: | |
2555 | ||
2556 | for (V = N1; V cond N2; V += STEP) BODY; | |
2557 | ||
2558 | where COND is "<" or ">", we generate pseudocode | |
2559 | ||
2560 | more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0); | |
2561 | if (more) goto L0; else goto L3; | |
2562 | L0: | |
2563 | V = istart0; | |
2564 | iend = iend0; | |
2565 | L1: | |
2566 | BODY; | |
2567 | V += STEP; | |
2568 | if (V cond iend) goto L1; else goto L2; | |
2569 | L2: | |
2570 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
2571 | L3: | |
2572 | ||
2573 | If this is a combined omp parallel loop, instead of the call to | |
2574 | GOMP_loop_foo_start, we call GOMP_loop_foo_next. | |
2575 | If this is gimple_omp_for_combined_p loop, then instead of assigning | |
2576 | V and iend in L0 we assign the first two _looptemp_ clause decls of the | |
2577 | inner GIMPLE_OMP_FOR and V += STEP; and | |
2578 | if (V cond iend) goto L1; else goto L2; are removed. | |
2579 | ||
2580 | For collapsed loops, given parameters: | |
2581 | collapse(3) | |
2582 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
2583 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
2584 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
2585 | BODY; | |
2586 | ||
2587 | we generate pseudocode | |
2588 | ||
2589 | if (__builtin_expect (N32 cond3 N31, 0)) goto Z0; | |
2590 | if (cond3 is <) | |
2591 | adj = STEP3 - 1; | |
2592 | else | |
2593 | adj = STEP3 + 1; | |
2594 | count3 = (adj + N32 - N31) / STEP3; | |
2595 | if (__builtin_expect (N22 cond2 N21, 0)) goto Z0; | |
2596 | if (cond2 is <) | |
2597 | adj = STEP2 - 1; | |
2598 | else | |
2599 | adj = STEP2 + 1; | |
2600 | count2 = (adj + N22 - N21) / STEP2; | |
2601 | if (__builtin_expect (N12 cond1 N11, 0)) goto Z0; | |
2602 | if (cond1 is <) | |
2603 | adj = STEP1 - 1; | |
2604 | else | |
2605 | adj = STEP1 + 1; | |
2606 | count1 = (adj + N12 - N11) / STEP1; | |
2607 | count = count1 * count2 * count3; | |
2608 | goto Z1; | |
2609 | Z0: | |
2610 | count = 0; | |
2611 | Z1: | |
2612 | more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0); | |
2613 | if (more) goto L0; else goto L3; | |
2614 | L0: | |
2615 | V = istart0; | |
2616 | T = V; | |
2617 | V3 = N31 + (T % count3) * STEP3; | |
2618 | T = T / count3; | |
2619 | V2 = N21 + (T % count2) * STEP2; | |
2620 | T = T / count2; | |
2621 | V1 = N11 + T * STEP1; | |
2622 | iend = iend0; | |
2623 | L1: | |
2624 | BODY; | |
2625 | V += 1; | |
2626 | if (V < iend) goto L10; else goto L2; | |
2627 | L10: | |
2628 | V3 += STEP3; | |
2629 | if (V3 cond3 N32) goto L1; else goto L11; | |
2630 | L11: | |
2631 | V3 = N31; | |
2632 | V2 += STEP2; | |
2633 | if (V2 cond2 N22) goto L1; else goto L12; | |
2634 | L12: | |
2635 | V2 = N21; | |
2636 | V1 += STEP1; | |
2637 | goto L1; | |
2638 | L2: | |
2639 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; | |
2640 | L3: | |
2641 | ||
2642 | */ | |
2643 | ||
2644 | static void | |
2645 | expand_omp_for_generic (struct omp_region *region, | |
2646 | struct omp_for_data *fd, | |
2647 | enum built_in_function start_fn, | |
2648 | enum built_in_function next_fn, | |
28567c40 | 2649 | tree sched_arg, |
629b3d75 MJ |
2650 | gimple *inner_stmt) |
2651 | { | |
2652 | tree type, istart0, iend0, iend; | |
2653 | tree t, vmain, vback, bias = NULL_TREE; | |
2654 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, collapse_bb; | |
2655 | basic_block l2_bb = NULL, l3_bb = NULL; | |
2656 | gimple_stmt_iterator gsi; | |
2657 | gassign *assign_stmt; | |
2658 | bool in_combined_parallel = is_combined_parallel (region); | |
2659 | bool broken_loop = region->cont == NULL; | |
2660 | edge e, ne; | |
2661 | tree *counts = NULL; | |
2662 | int i; | |
2663 | bool ordered_lastprivate = false; | |
2664 | ||
2665 | gcc_assert (!broken_loop || !in_combined_parallel); | |
2666 | gcc_assert (fd->iter_type == long_integer_type_node | |
2667 | || !in_combined_parallel); | |
2668 | ||
2669 | entry_bb = region->entry; | |
2670 | cont_bb = region->cont; | |
2671 | collapse_bb = NULL; | |
2672 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
2673 | gcc_assert (broken_loop | |
2674 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
2675 | l0_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
2676 | l1_bb = single_succ (l0_bb); | |
2677 | if (!broken_loop) | |
2678 | { | |
2679 | l2_bb = create_empty_bb (cont_bb); | |
2680 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb | |
2681 | || (single_succ_edge (BRANCH_EDGE (cont_bb)->dest)->dest | |
2682 | == l1_bb)); | |
2683 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
2684 | } | |
2685 | else | |
2686 | l2_bb = NULL; | |
2687 | l3_bb = BRANCH_EDGE (entry_bb)->dest; | |
2688 | exit_bb = region->exit; | |
2689 | ||
65f4b875 | 2690 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
2691 | |
2692 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
2693 | if (fd->ordered | |
6c7ae8c5 | 2694 | && omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
629b3d75 MJ |
2695 | OMP_CLAUSE_LASTPRIVATE)) |
2696 | ordered_lastprivate = false; | |
28567c40 | 2697 | tree reductions = NULL_TREE; |
6c7ae8c5 JJ |
2698 | tree mem = NULL_TREE, cond_var = NULL_TREE, condtemp = NULL_TREE; |
2699 | tree memv = NULL_TREE; | |
8221c30b JJ |
2700 | if (fd->lastprivate_conditional) |
2701 | { | |
2702 | tree c = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
2703 | OMP_CLAUSE__CONDTEMP_); | |
2704 | if (fd->have_pointer_condtemp) | |
2705 | condtemp = OMP_CLAUSE_DECL (c); | |
2706 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
2707 | cond_var = OMP_CLAUSE_DECL (c); | |
2708 | } | |
28567c40 JJ |
2709 | if (sched_arg) |
2710 | { | |
2711 | if (fd->have_reductemp) | |
2712 | { | |
6c7ae8c5 | 2713 | tree c = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
28567c40 JJ |
2714 | OMP_CLAUSE__REDUCTEMP_); |
2715 | reductions = OMP_CLAUSE_DECL (c); | |
2716 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
2717 | gimple *g = SSA_NAME_DEF_STMT (reductions); | |
2718 | reductions = gimple_assign_rhs1 (g); | |
2719 | OMP_CLAUSE_DECL (c) = reductions; | |
2720 | entry_bb = gimple_bb (g); | |
2721 | edge e = split_block (entry_bb, g); | |
2722 | if (region->entry == entry_bb) | |
2723 | region->entry = e->dest; | |
2724 | gsi = gsi_last_bb (entry_bb); | |
2725 | } | |
2726 | else | |
2727 | reductions = null_pointer_node; | |
8221c30b | 2728 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 2729 | { |
6c7ae8c5 JJ |
2730 | tree type = TREE_TYPE (condtemp); |
2731 | memv = create_tmp_var (type); | |
2732 | TREE_ADDRESSABLE (memv) = 1; | |
2733 | unsigned HOST_WIDE_INT sz | |
2734 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
2735 | sz *= fd->lastprivate_conditional; | |
2736 | expand_omp_build_assign (&gsi, memv, build_int_cst (type, sz), | |
2737 | false); | |
2738 | mem = build_fold_addr_expr (memv); | |
2739 | } | |
2740 | else | |
2741 | mem = null_pointer_node; | |
28567c40 | 2742 | } |
629b3d75 MJ |
2743 | if (fd->collapse > 1 || fd->ordered) |
2744 | { | |
2745 | int first_zero_iter1 = -1, first_zero_iter2 = -1; | |
2746 | basic_block zero_iter1_bb = NULL, zero_iter2_bb = NULL, l2_dom_bb = NULL; | |
2747 | ||
2748 | counts = XALLOCAVEC (tree, fd->ordered ? fd->ordered + 1 : fd->collapse); | |
2749 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
2750 | zero_iter1_bb, first_zero_iter1, | |
2751 | zero_iter2_bb, first_zero_iter2, l2_dom_bb); | |
2752 | ||
2753 | if (zero_iter1_bb) | |
2754 | { | |
2755 | /* Some counts[i] vars might be uninitialized if | |
2756 | some loop has zero iterations. But the body shouldn't | |
2757 | be executed in that case, so just avoid uninit warnings. */ | |
2758 | for (i = first_zero_iter1; | |
2759 | i < (fd->ordered ? fd->ordered : fd->collapse); i++) | |
2760 | if (SSA_VAR_P (counts[i])) | |
2761 | TREE_NO_WARNING (counts[i]) = 1; | |
2762 | gsi_prev (&gsi); | |
2763 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
2764 | entry_bb = e->dest; | |
2765 | make_edge (zero_iter1_bb, entry_bb, EDGE_FALLTHRU); | |
65f4b875 | 2766 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
2767 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, |
2768 | get_immediate_dominator (CDI_DOMINATORS, | |
2769 | zero_iter1_bb)); | |
2770 | } | |
2771 | if (zero_iter2_bb) | |
2772 | { | |
2773 | /* Some counts[i] vars might be uninitialized if | |
2774 | some loop has zero iterations. But the body shouldn't | |
2775 | be executed in that case, so just avoid uninit warnings. */ | |
2776 | for (i = first_zero_iter2; i < fd->ordered; i++) | |
2777 | if (SSA_VAR_P (counts[i])) | |
2778 | TREE_NO_WARNING (counts[i]) = 1; | |
2779 | if (zero_iter1_bb) | |
2780 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); | |
2781 | else | |
2782 | { | |
2783 | gsi_prev (&gsi); | |
2784 | e = split_block (entry_bb, gsi_stmt (gsi)); | |
2785 | entry_bb = e->dest; | |
2786 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); | |
65f4b875 | 2787 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
2788 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, |
2789 | get_immediate_dominator | |
2790 | (CDI_DOMINATORS, zero_iter2_bb)); | |
2791 | } | |
2792 | } | |
2793 | if (fd->collapse == 1) | |
2794 | { | |
2795 | counts[0] = fd->loop.n2; | |
2796 | fd->loop = fd->loops[0]; | |
2797 | } | |
2798 | } | |
2799 | ||
2800 | type = TREE_TYPE (fd->loop.v); | |
2801 | istart0 = create_tmp_var (fd->iter_type, ".istart0"); | |
2802 | iend0 = create_tmp_var (fd->iter_type, ".iend0"); | |
2803 | TREE_ADDRESSABLE (istart0) = 1; | |
2804 | TREE_ADDRESSABLE (iend0) = 1; | |
2805 | ||
2806 | /* See if we need to bias by LLONG_MIN. */ | |
2807 | if (fd->iter_type == long_long_unsigned_type_node | |
2808 | && TREE_CODE (type) == INTEGER_TYPE | |
2809 | && !TYPE_UNSIGNED (type) | |
2810 | && fd->ordered == 0) | |
2811 | { | |
2812 | tree n1, n2; | |
2813 | ||
2814 | if (fd->loop.cond_code == LT_EXPR) | |
2815 | { | |
2816 | n1 = fd->loop.n1; | |
2817 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
2818 | } | |
2819 | else | |
2820 | { | |
2821 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
2822 | n2 = fd->loop.n1; | |
2823 | } | |
2824 | if (TREE_CODE (n1) != INTEGER_CST | |
2825 | || TREE_CODE (n2) != INTEGER_CST | |
2826 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
2827 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
2828 | } | |
2829 | ||
2830 | gimple_stmt_iterator gsif = gsi; | |
2831 | gsi_prev (&gsif); | |
2832 | ||
2833 | tree arr = NULL_TREE; | |
2834 | if (in_combined_parallel) | |
2835 | { | |
2836 | gcc_assert (fd->ordered == 0); | |
2837 | /* In a combined parallel loop, emit a call to | |
2838 | GOMP_loop_foo_next. */ | |
2839 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, | |
2840 | build_fold_addr_expr (istart0), | |
2841 | build_fold_addr_expr (iend0)); | |
2842 | } | |
2843 | else | |
2844 | { | |
2845 | tree t0, t1, t2, t3, t4; | |
2846 | /* If this is not a combined parallel loop, emit a call to | |
2847 | GOMP_loop_foo_start in ENTRY_BB. */ | |
2848 | t4 = build_fold_addr_expr (iend0); | |
2849 | t3 = build_fold_addr_expr (istart0); | |
2850 | if (fd->ordered) | |
2851 | { | |
2852 | t0 = build_int_cst (unsigned_type_node, | |
2853 | fd->ordered - fd->collapse + 1); | |
2854 | arr = create_tmp_var (build_array_type_nelts (fd->iter_type, | |
2855 | fd->ordered | |
2856 | - fd->collapse + 1), | |
2857 | ".omp_counts"); | |
2858 | DECL_NAMELESS (arr) = 1; | |
2859 | TREE_ADDRESSABLE (arr) = 1; | |
2860 | TREE_STATIC (arr) = 1; | |
2861 | vec<constructor_elt, va_gc> *v; | |
2862 | vec_alloc (v, fd->ordered - fd->collapse + 1); | |
2863 | int idx; | |
2864 | ||
2865 | for (idx = 0; idx < fd->ordered - fd->collapse + 1; idx++) | |
2866 | { | |
2867 | tree c; | |
2868 | if (idx == 0 && fd->collapse > 1) | |
2869 | c = fd->loop.n2; | |
2870 | else | |
2871 | c = counts[idx + fd->collapse - 1]; | |
2872 | tree purpose = size_int (idx); | |
2873 | CONSTRUCTOR_APPEND_ELT (v, purpose, c); | |
2874 | if (TREE_CODE (c) != INTEGER_CST) | |
2875 | TREE_STATIC (arr) = 0; | |
2876 | } | |
2877 | ||
2878 | DECL_INITIAL (arr) = build_constructor (TREE_TYPE (arr), v); | |
2879 | if (!TREE_STATIC (arr)) | |
2880 | force_gimple_operand_gsi (&gsi, build1 (DECL_EXPR, | |
2881 | void_type_node, arr), | |
2882 | true, NULL_TREE, true, GSI_SAME_STMT); | |
2883 | t1 = build_fold_addr_expr (arr); | |
2884 | t2 = NULL_TREE; | |
2885 | } | |
2886 | else | |
2887 | { | |
2888 | t2 = fold_convert (fd->iter_type, fd->loop.step); | |
2889 | t1 = fd->loop.n2; | |
2890 | t0 = fd->loop.n1; | |
2891 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
2892 | { | |
2893 | tree innerc | |
2894 | = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
2895 | OMP_CLAUSE__LOOPTEMP_); | |
2896 | gcc_assert (innerc); | |
2897 | t0 = OMP_CLAUSE_DECL (innerc); | |
2898 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
2899 | OMP_CLAUSE__LOOPTEMP_); | |
2900 | gcc_assert (innerc); | |
2901 | t1 = OMP_CLAUSE_DECL (innerc); | |
2902 | } | |
2903 | if (POINTER_TYPE_P (TREE_TYPE (t0)) | |
2904 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
2905 | != TYPE_PRECISION (fd->iter_type)) | |
2906 | { | |
2907 | /* Avoid casting pointers to integer of a different size. */ | |
2908 | tree itype = signed_type_for (type); | |
2909 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); | |
2910 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
2911 | } | |
2912 | else | |
2913 | { | |
2914 | t1 = fold_convert (fd->iter_type, t1); | |
2915 | t0 = fold_convert (fd->iter_type, t0); | |
2916 | } | |
2917 | if (bias) | |
2918 | { | |
2919 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); | |
2920 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
2921 | } | |
2922 | } | |
2923 | if (fd->iter_type == long_integer_type_node || fd->ordered) | |
2924 | { | |
2925 | if (fd->chunk_size) | |
2926 | { | |
2927 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
2928 | t = omp_adjust_chunk_size (t, fd->simd_schedule); | |
28567c40 JJ |
2929 | if (sched_arg) |
2930 | { | |
2931 | if (fd->ordered) | |
2932 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
2933 | 8, t0, t1, sched_arg, t, t3, t4, | |
2934 | reductions, mem); | |
2935 | else | |
2936 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
2937 | 9, t0, t1, t2, sched_arg, t, t3, t4, | |
2938 | reductions, mem); | |
2939 | } | |
2940 | else if (fd->ordered) | |
629b3d75 MJ |
2941 | t = build_call_expr (builtin_decl_explicit (start_fn), |
2942 | 5, t0, t1, t, t3, t4); | |
2943 | else | |
2944 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
2945 | 6, t0, t1, t2, t, t3, t4); | |
2946 | } | |
2947 | else if (fd->ordered) | |
2948 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
2949 | 4, t0, t1, t3, t4); | |
2950 | else | |
2951 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
2952 | 5, t0, t1, t2, t3, t4); | |
2953 | } | |
2954 | else | |
2955 | { | |
2956 | tree t5; | |
2957 | tree c_bool_type; | |
2958 | tree bfn_decl; | |
2959 | ||
2960 | /* The GOMP_loop_ull_*start functions have additional boolean | |
2961 | argument, true for < loops and false for > loops. | |
2962 | In Fortran, the C bool type can be different from | |
2963 | boolean_type_node. */ | |
2964 | bfn_decl = builtin_decl_explicit (start_fn); | |
2965 | c_bool_type = TREE_TYPE (TREE_TYPE (bfn_decl)); | |
2966 | t5 = build_int_cst (c_bool_type, | |
2967 | fd->loop.cond_code == LT_EXPR ? 1 : 0); | |
2968 | if (fd->chunk_size) | |
2969 | { | |
2970 | tree bfn_decl = builtin_decl_explicit (start_fn); | |
2971 | t = fold_convert (fd->iter_type, fd->chunk_size); | |
2972 | t = omp_adjust_chunk_size (t, fd->simd_schedule); | |
28567c40 JJ |
2973 | if (sched_arg) |
2974 | t = build_call_expr (bfn_decl, 10, t5, t0, t1, t2, sched_arg, | |
2975 | t, t3, t4, reductions, mem); | |
2976 | else | |
2977 | t = build_call_expr (bfn_decl, 7, t5, t0, t1, t2, t, t3, t4); | |
629b3d75 MJ |
2978 | } |
2979 | else | |
2980 | t = build_call_expr (builtin_decl_explicit (start_fn), | |
2981 | 6, t5, t0, t1, t2, t3, t4); | |
2982 | } | |
2983 | } | |
2984 | if (TREE_TYPE (t) != boolean_type_node) | |
2985 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
2986 | t, build_int_cst (TREE_TYPE (t), 0)); | |
2987 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
01914336 | 2988 | true, GSI_SAME_STMT); |
629b3d75 MJ |
2989 | if (arr && !TREE_STATIC (arr)) |
2990 | { | |
25b45c7c | 2991 | tree clobber = build_clobber (TREE_TYPE (arr)); |
629b3d75 MJ |
2992 | gsi_insert_before (&gsi, gimple_build_assign (arr, clobber), |
2993 | GSI_SAME_STMT); | |
2994 | } | |
8221c30b | 2995 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 2996 | expand_omp_build_assign (&gsi, condtemp, memv, false); |
28567c40 JJ |
2997 | if (fd->have_reductemp) |
2998 | { | |
2999 | gimple *g = gsi_stmt (gsi); | |
3000 | gsi_remove (&gsi, true); | |
3001 | release_ssa_name (gimple_assign_lhs (g)); | |
3002 | ||
3003 | entry_bb = region->entry; | |
3004 | gsi = gsi_last_nondebug_bb (entry_bb); | |
3005 | ||
3006 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
3007 | } | |
629b3d75 MJ |
3008 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); |
3009 | ||
3010 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
3011 | gsi_remove (&gsi, true); | |
3012 | ||
3013 | if (gsi_end_p (gsif)) | |
3014 | gsif = gsi_after_labels (gsi_bb (gsif)); | |
3015 | gsi_next (&gsif); | |
3016 | ||
3017 | /* Iteration setup for sequential loop goes in L0_BB. */ | |
3018 | tree startvar = fd->loop.v; | |
3019 | tree endvar = NULL_TREE; | |
3020 | ||
3021 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
3022 | { | |
3023 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_FOR | |
3024 | && gimple_omp_for_kind (inner_stmt) | |
3025 | == GF_OMP_FOR_KIND_SIMD); | |
3026 | tree innerc = omp_find_clause (gimple_omp_for_clauses (inner_stmt), | |
3027 | OMP_CLAUSE__LOOPTEMP_); | |
3028 | gcc_assert (innerc); | |
3029 | startvar = OMP_CLAUSE_DECL (innerc); | |
3030 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
3031 | OMP_CLAUSE__LOOPTEMP_); | |
3032 | gcc_assert (innerc); | |
3033 | endvar = OMP_CLAUSE_DECL (innerc); | |
3034 | } | |
3035 | ||
3036 | gsi = gsi_start_bb (l0_bb); | |
3037 | t = istart0; | |
3038 | if (fd->ordered && fd->collapse == 1) | |
3039 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, | |
3040 | fold_convert (fd->iter_type, fd->loop.step)); | |
3041 | else if (bias) | |
3042 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); | |
3043 | if (fd->ordered && fd->collapse == 1) | |
3044 | { | |
3045 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
3046 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), | |
3047 | fd->loop.n1, fold_convert (sizetype, t)); | |
3048 | else | |
3049 | { | |
3050 | t = fold_convert (TREE_TYPE (startvar), t); | |
3051 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), | |
3052 | fd->loop.n1, t); | |
3053 | } | |
3054 | } | |
3055 | else | |
3056 | { | |
3057 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
3058 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
3059 | t = fold_convert (TREE_TYPE (startvar), t); | |
3060 | } | |
3061 | t = force_gimple_operand_gsi (&gsi, t, | |
3062 | DECL_P (startvar) | |
3063 | && TREE_ADDRESSABLE (startvar), | |
3064 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
3065 | assign_stmt = gimple_build_assign (startvar, t); | |
3066 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6c7ae8c5 JJ |
3067 | if (cond_var) |
3068 | { | |
3069 | tree itype = TREE_TYPE (cond_var); | |
3070 | /* For lastprivate(conditional:) itervar, we need some iteration | |
3071 | counter that starts at unsigned non-zero and increases. | |
3072 | Prefer as few IVs as possible, so if we can use startvar | |
3073 | itself, use that, or startvar + constant (those would be | |
3074 | incremented with step), and as last resort use the s0 + 1 | |
3075 | incremented by 1. */ | |
3076 | if ((fd->ordered && fd->collapse == 1) | |
3077 | || bias | |
3078 | || POINTER_TYPE_P (type) | |
3079 | || TREE_CODE (fd->loop.n1) != INTEGER_CST | |
3080 | || fd->loop.cond_code != LT_EXPR) | |
3081 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, istart0), | |
3082 | build_int_cst (itype, 1)); | |
3083 | else if (tree_int_cst_sgn (fd->loop.n1) == 1) | |
3084 | t = fold_convert (itype, t); | |
3085 | else | |
3086 | { | |
3087 | tree c = fold_convert (itype, fd->loop.n1); | |
3088 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
3089 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
3090 | } | |
3091 | t = force_gimple_operand_gsi (&gsi, t, false, | |
3092 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
3093 | assign_stmt = gimple_build_assign (cond_var, t); | |
3094 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
3095 | } | |
629b3d75 MJ |
3096 | |
3097 | t = iend0; | |
3098 | if (fd->ordered && fd->collapse == 1) | |
3099 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, | |
3100 | fold_convert (fd->iter_type, fd->loop.step)); | |
3101 | else if (bias) | |
3102 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); | |
3103 | if (fd->ordered && fd->collapse == 1) | |
3104 | { | |
3105 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
3106 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), | |
3107 | fd->loop.n1, fold_convert (sizetype, t)); | |
3108 | else | |
3109 | { | |
3110 | t = fold_convert (TREE_TYPE (startvar), t); | |
3111 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), | |
3112 | fd->loop.n1, t); | |
3113 | } | |
3114 | } | |
3115 | else | |
3116 | { | |
3117 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) | |
3118 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); | |
3119 | t = fold_convert (TREE_TYPE (startvar), t); | |
3120 | } | |
3121 | iend = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
3122 | false, GSI_CONTINUE_LINKING); | |
3123 | if (endvar) | |
3124 | { | |
3125 | assign_stmt = gimple_build_assign (endvar, iend); | |
3126 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
3127 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (iend))) | |
3128 | assign_stmt = gimple_build_assign (fd->loop.v, iend); | |
3129 | else | |
3130 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, iend); | |
3131 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
3132 | } | |
3133 | /* Handle linear clause adjustments. */ | |
3134 | tree itercnt = NULL_TREE; | |
3135 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
3136 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
3137 | c; c = OMP_CLAUSE_CHAIN (c)) | |
3138 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
3139 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
3140 | { | |
3141 | tree d = OMP_CLAUSE_DECL (c); | |
3142 | bool is_ref = omp_is_reference (d); | |
3143 | tree t = d, a, dest; | |
3144 | if (is_ref) | |
3145 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); | |
3146 | tree type = TREE_TYPE (t); | |
3147 | if (POINTER_TYPE_P (type)) | |
3148 | type = sizetype; | |
3149 | dest = unshare_expr (t); | |
3150 | tree v = create_tmp_var (TREE_TYPE (t), NULL); | |
3151 | expand_omp_build_assign (&gsif, v, t); | |
3152 | if (itercnt == NULL_TREE) | |
3153 | { | |
3154 | itercnt = startvar; | |
3155 | tree n1 = fd->loop.n1; | |
3156 | if (POINTER_TYPE_P (TREE_TYPE (itercnt))) | |
3157 | { | |
3158 | itercnt | |
3159 | = fold_convert (signed_type_for (TREE_TYPE (itercnt)), | |
3160 | itercnt); | |
3161 | n1 = fold_convert (TREE_TYPE (itercnt), n1); | |
3162 | } | |
3163 | itercnt = fold_build2 (MINUS_EXPR, TREE_TYPE (itercnt), | |
3164 | itercnt, n1); | |
3165 | itercnt = fold_build2 (EXACT_DIV_EXPR, TREE_TYPE (itercnt), | |
3166 | itercnt, fd->loop.step); | |
3167 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
3168 | NULL_TREE, false, | |
3169 | GSI_CONTINUE_LINKING); | |
3170 | } | |
3171 | a = fold_build2 (MULT_EXPR, type, | |
3172 | fold_convert (type, itercnt), | |
3173 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
3174 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
3175 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); | |
3176 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
3177 | false, GSI_CONTINUE_LINKING); | |
3178 | assign_stmt = gimple_build_assign (dest, t); | |
3179 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
3180 | } | |
3181 | if (fd->collapse > 1) | |
3182 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); | |
3183 | ||
3184 | if (fd->ordered) | |
3185 | { | |
3186 | /* Until now, counts array contained number of iterations or | |
3187 | variable containing it for ith loop. From now on, we need | |
3188 | those counts only for collapsed loops, and only for the 2nd | |
3189 | till the last collapsed one. Move those one element earlier, | |
3190 | we'll use counts[fd->collapse - 1] for the first source/sink | |
3191 | iteration counter and so on and counts[fd->ordered] | |
3192 | as the array holding the current counter values for | |
3193 | depend(source). */ | |
3194 | if (fd->collapse > 1) | |
3195 | memmove (counts, counts + 1, (fd->collapse - 1) * sizeof (counts[0])); | |
3196 | if (broken_loop) | |
3197 | { | |
3198 | int i; | |
3199 | for (i = fd->collapse; i < fd->ordered; i++) | |
3200 | { | |
3201 | tree type = TREE_TYPE (fd->loops[i].v); | |
3202 | tree this_cond | |
3203 | = fold_build2 (fd->loops[i].cond_code, boolean_type_node, | |
3204 | fold_convert (type, fd->loops[i].n1), | |
3205 | fold_convert (type, fd->loops[i].n2)); | |
3206 | if (!integer_onep (this_cond)) | |
3207 | break; | |
3208 | } | |
3209 | if (i < fd->ordered) | |
3210 | { | |
3211 | cont_bb | |
3212 | = create_empty_bb (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb); | |
3213 | add_bb_to_loop (cont_bb, l1_bb->loop_father); | |
3214 | gimple_stmt_iterator gsi = gsi_after_labels (cont_bb); | |
3215 | gimple *g = gimple_build_omp_continue (fd->loop.v, fd->loop.v); | |
3216 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
3217 | make_edge (cont_bb, l3_bb, EDGE_FALLTHRU); | |
3218 | make_edge (cont_bb, l1_bb, 0); | |
3219 | l2_bb = create_empty_bb (cont_bb); | |
3220 | broken_loop = false; | |
3221 | } | |
3222 | } | |
3223 | expand_omp_ordered_source_sink (region, fd, counts, cont_bb); | |
3224 | cont_bb = expand_omp_for_ordered_loops (fd, counts, cont_bb, l1_bb, | |
3225 | ordered_lastprivate); | |
3226 | if (counts[fd->collapse - 1]) | |
3227 | { | |
3228 | gcc_assert (fd->collapse == 1); | |
3229 | gsi = gsi_last_bb (l0_bb); | |
3230 | expand_omp_build_assign (&gsi, counts[fd->collapse - 1], | |
3231 | istart0, true); | |
3232 | gsi = gsi_last_bb (cont_bb); | |
3233 | t = fold_build2 (PLUS_EXPR, fd->iter_type, counts[fd->collapse - 1], | |
3234 | build_int_cst (fd->iter_type, 1)); | |
3235 | expand_omp_build_assign (&gsi, counts[fd->collapse - 1], t); | |
3236 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3237 | size_zero_node, NULL_TREE, NULL_TREE); | |
3238 | expand_omp_build_assign (&gsi, aref, counts[fd->collapse - 1]); | |
3239 | t = counts[fd->collapse - 1]; | |
3240 | } | |
3241 | else if (fd->collapse > 1) | |
3242 | t = fd->loop.v; | |
3243 | else | |
3244 | { | |
3245 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), | |
3246 | fd->loops[0].v, fd->loops[0].n1); | |
3247 | t = fold_convert (fd->iter_type, t); | |
3248 | } | |
3249 | gsi = gsi_last_bb (l0_bb); | |
3250 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], | |
3251 | size_zero_node, NULL_TREE, NULL_TREE); | |
3252 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
3253 | false, GSI_CONTINUE_LINKING); | |
3254 | expand_omp_build_assign (&gsi, aref, t, true); | |
3255 | } | |
3256 | ||
3257 | if (!broken_loop) | |
3258 | { | |
3259 | /* Code to control the increment and predicate for the sequential | |
3260 | loop goes in the CONT_BB. */ | |
65f4b875 | 3261 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
3262 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
3263 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
3264 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
3265 | vback = gimple_omp_continue_control_def (cont_stmt); | |
3266 | ||
7855700e JJ |
3267 | if (cond_var) |
3268 | { | |
3269 | tree itype = TREE_TYPE (cond_var); | |
3270 | tree t2; | |
3271 | if ((fd->ordered && fd->collapse == 1) | |
3272 | || bias | |
3273 | || POINTER_TYPE_P (type) | |
3274 | || TREE_CODE (fd->loop.n1) != INTEGER_CST | |
3275 | || fd->loop.cond_code != LT_EXPR) | |
3276 | t2 = build_int_cst (itype, 1); | |
3277 | else | |
3278 | t2 = fold_convert (itype, fd->loop.step); | |
3279 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
3280 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
3281 | NULL_TREE, true, GSI_SAME_STMT); | |
3282 | assign_stmt = gimple_build_assign (cond_var, t2); | |
3283 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
3284 | } | |
3285 | ||
629b3d75 MJ |
3286 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
3287 | { | |
3288 | if (POINTER_TYPE_P (type)) | |
3289 | t = fold_build_pointer_plus (vmain, fd->loop.step); | |
3290 | else | |
3291 | t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step); | |
3292 | t = force_gimple_operand_gsi (&gsi, t, | |
3293 | DECL_P (vback) | |
3294 | && TREE_ADDRESSABLE (vback), | |
3295 | NULL_TREE, true, GSI_SAME_STMT); | |
3296 | assign_stmt = gimple_build_assign (vback, t); | |
3297 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
3298 | ||
3299 | if (fd->ordered && counts[fd->collapse - 1] == NULL_TREE) | |
3300 | { | |
d1ffbd43 | 3301 | tree tem; |
629b3d75 | 3302 | if (fd->collapse > 1) |
d1ffbd43 | 3303 | tem = fd->loop.v; |
629b3d75 MJ |
3304 | else |
3305 | { | |
d1ffbd43 JJ |
3306 | tem = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), |
3307 | fd->loops[0].v, fd->loops[0].n1); | |
3308 | tem = fold_convert (fd->iter_type, tem); | |
629b3d75 MJ |
3309 | } |
3310 | tree aref = build4 (ARRAY_REF, fd->iter_type, | |
3311 | counts[fd->ordered], size_zero_node, | |
3312 | NULL_TREE, NULL_TREE); | |
d1ffbd43 JJ |
3313 | tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, |
3314 | true, GSI_SAME_STMT); | |
3315 | expand_omp_build_assign (&gsi, aref, tem); | |
629b3d75 MJ |
3316 | } |
3317 | ||
3318 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
3319 | DECL_P (vback) && TREE_ADDRESSABLE (vback) ? t : vback, | |
3320 | iend); | |
3321 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
3322 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
3323 | } | |
3324 | ||
3325 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
3326 | gsi_remove (&gsi, true); | |
3327 | ||
3328 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
3329 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, l1_bb); | |
3330 | ||
3331 | /* Emit code to get the next parallel iteration in L2_BB. */ | |
3332 | gsi = gsi_start_bb (l2_bb); | |
3333 | ||
3334 | t = build_call_expr (builtin_decl_explicit (next_fn), 2, | |
3335 | build_fold_addr_expr (istart0), | |
3336 | build_fold_addr_expr (iend0)); | |
3337 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
3338 | false, GSI_CONTINUE_LINKING); | |
3339 | if (TREE_TYPE (t) != boolean_type_node) | |
3340 | t = fold_build2 (NE_EXPR, boolean_type_node, | |
3341 | t, build_int_cst (TREE_TYPE (t), 0)); | |
3342 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
3343 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
3344 | } | |
3345 | ||
3346 | /* Add the loop cleanup function. */ | |
65f4b875 | 3347 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
3348 | if (gimple_omp_return_nowait_p (gsi_stmt (gsi))) |
3349 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
3350 | else if (gimple_omp_return_lhs (gsi_stmt (gsi))) | |
3351 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
3352 | else | |
3353 | t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
3354 | gcall *call_stmt = gimple_build_call (t, 0); | |
629b3d75 MJ |
3355 | if (fd->ordered) |
3356 | { | |
3357 | tree arr = counts[fd->ordered]; | |
25b45c7c | 3358 | tree clobber = build_clobber (TREE_TYPE (arr)); |
629b3d75 MJ |
3359 | gsi_insert_after (&gsi, gimple_build_assign (arr, clobber), |
3360 | GSI_SAME_STMT); | |
3361 | } | |
28567c40 JJ |
3362 | if (gimple_omp_return_lhs (gsi_stmt (gsi))) |
3363 | { | |
3364 | gimple_call_set_lhs (call_stmt, gimple_omp_return_lhs (gsi_stmt (gsi))); | |
3365 | if (fd->have_reductemp) | |
3366 | { | |
3367 | gimple *g = gimple_build_assign (reductions, NOP_EXPR, | |
3368 | gimple_call_lhs (call_stmt)); | |
3369 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
3370 | } | |
3371 | } | |
3372 | gsi_insert_after (&gsi, call_stmt, GSI_SAME_STMT); | |
629b3d75 MJ |
3373 | gsi_remove (&gsi, true); |
3374 | ||
3375 | /* Connect the new blocks. */ | |
3376 | find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE; | |
3377 | find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE; | |
3378 | ||
3379 | if (!broken_loop) | |
3380 | { | |
3381 | gimple_seq phis; | |
3382 | ||
3383 | e = find_edge (cont_bb, l3_bb); | |
3384 | ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE); | |
3385 | ||
3386 | phis = phi_nodes (l3_bb); | |
3387 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3388 | { | |
3389 | gimple *phi = gsi_stmt (gsi); | |
3390 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne), | |
3391 | PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
3392 | } | |
3393 | remove_edge (e); | |
3394 | ||
3395 | make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE); | |
3396 | e = find_edge (cont_bb, l1_bb); | |
3397 | if (e == NULL) | |
3398 | { | |
3399 | e = BRANCH_EDGE (cont_bb); | |
3400 | gcc_assert (single_succ (e->dest) == l1_bb); | |
3401 | } | |
3402 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
3403 | { | |
3404 | remove_edge (e); | |
3405 | e = NULL; | |
3406 | } | |
3407 | else if (fd->collapse > 1) | |
3408 | { | |
3409 | remove_edge (e); | |
3410 | e = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
3411 | } | |
3412 | else | |
3413 | e->flags = EDGE_TRUE_VALUE; | |
3414 | if (e) | |
3415 | { | |
357067f2 JH |
3416 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); |
3417 | find_edge (cont_bb, l2_bb)->probability = e->probability.invert (); | |
629b3d75 MJ |
3418 | } |
3419 | else | |
3420 | { | |
3421 | e = find_edge (cont_bb, l2_bb); | |
3422 | e->flags = EDGE_FALLTHRU; | |
3423 | } | |
3424 | make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE); | |
3425 | ||
3426 | if (gimple_in_ssa_p (cfun)) | |
3427 | { | |
3428 | /* Add phis to the outer loop that connect to the phis in the inner, | |
3429 | original loop, and move the loop entry value of the inner phi to | |
3430 | the loop entry value of the outer phi. */ | |
3431 | gphi_iterator psi; | |
3432 | for (psi = gsi_start_phis (l3_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
3433 | { | |
620e594b | 3434 | location_t locus; |
629b3d75 MJ |
3435 | gphi *nphi; |
3436 | gphi *exit_phi = psi.phi (); | |
3437 | ||
164485b5 JJ |
3438 | if (virtual_operand_p (gimple_phi_result (exit_phi))) |
3439 | continue; | |
3440 | ||
629b3d75 MJ |
3441 | edge l2_to_l3 = find_edge (l2_bb, l3_bb); |
3442 | tree exit_res = PHI_ARG_DEF_FROM_EDGE (exit_phi, l2_to_l3); | |
3443 | ||
3444 | basic_block latch = BRANCH_EDGE (cont_bb)->dest; | |
3445 | edge latch_to_l1 = find_edge (latch, l1_bb); | |
3446 | gphi *inner_phi | |
3447 | = find_phi_with_arg_on_edge (exit_res, latch_to_l1); | |
3448 | ||
3449 | tree t = gimple_phi_result (exit_phi); | |
3450 | tree new_res = copy_ssa_name (t, NULL); | |
3451 | nphi = create_phi_node (new_res, l0_bb); | |
3452 | ||
3453 | edge l0_to_l1 = find_edge (l0_bb, l1_bb); | |
3454 | t = PHI_ARG_DEF_FROM_EDGE (inner_phi, l0_to_l1); | |
3455 | locus = gimple_phi_arg_location_from_edge (inner_phi, l0_to_l1); | |
3456 | edge entry_to_l0 = find_edge (entry_bb, l0_bb); | |
3457 | add_phi_arg (nphi, t, entry_to_l0, locus); | |
3458 | ||
3459 | edge l2_to_l0 = find_edge (l2_bb, l0_bb); | |
3460 | add_phi_arg (nphi, exit_res, l2_to_l0, UNKNOWN_LOCATION); | |
3461 | ||
3462 | add_phi_arg (inner_phi, new_res, l0_to_l1, UNKNOWN_LOCATION); | |
164485b5 | 3463 | } |
629b3d75 MJ |
3464 | } |
3465 | ||
3466 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, | |
3467 | recompute_dominator (CDI_DOMINATORS, l2_bb)); | |
3468 | set_immediate_dominator (CDI_DOMINATORS, l3_bb, | |
3469 | recompute_dominator (CDI_DOMINATORS, l3_bb)); | |
3470 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, | |
3471 | recompute_dominator (CDI_DOMINATORS, l0_bb)); | |
3472 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, | |
3473 | recompute_dominator (CDI_DOMINATORS, l1_bb)); | |
3474 | ||
3475 | /* We enter expand_omp_for_generic with a loop. This original loop may | |
3476 | have its own loop struct, or it may be part of an outer loop struct | |
3477 | (which may be the fake loop). */ | |
99b1c316 | 3478 | class loop *outer_loop = entry_bb->loop_father; |
629b3d75 MJ |
3479 | bool orig_loop_has_loop_struct = l1_bb->loop_father != outer_loop; |
3480 | ||
3481 | add_bb_to_loop (l2_bb, outer_loop); | |
3482 | ||
3483 | /* We've added a new loop around the original loop. Allocate the | |
3484 | corresponding loop struct. */ | |
99b1c316 | 3485 | class loop *new_loop = alloc_loop (); |
629b3d75 MJ |
3486 | new_loop->header = l0_bb; |
3487 | new_loop->latch = l2_bb; | |
3488 | add_loop (new_loop, outer_loop); | |
3489 | ||
3490 | /* Allocate a loop structure for the original loop unless we already | |
3491 | had one. */ | |
3492 | if (!orig_loop_has_loop_struct | |
3493 | && !gimple_omp_for_combined_p (fd->for_stmt)) | |
3494 | { | |
99b1c316 | 3495 | class loop *orig_loop = alloc_loop (); |
629b3d75 MJ |
3496 | orig_loop->header = l1_bb; |
3497 | /* The loop may have multiple latches. */ | |
3498 | add_loop (orig_loop, new_loop); | |
3499 | } | |
3500 | } | |
3501 | } | |
3502 | ||
2f6bb511 JJ |
3503 | /* Helper function for expand_omp_for_static_nochunk. If PTR is NULL, |
3504 | compute needed allocation size. If !ALLOC of team allocations, | |
3505 | if ALLOC of thread allocation. SZ is the initial needed size for | |
3506 | other purposes, ALLOC_ALIGN guaranteed alignment of allocation in bytes, | |
3507 | CNT number of elements of each array, for !ALLOC this is | |
3508 | omp_get_num_threads (), for ALLOC number of iterations handled by the | |
3509 | current thread. If PTR is non-NULL, it is the start of the allocation | |
3510 | and this routine shall assign to OMP_CLAUSE_DECL (c) of those _scantemp_ | |
3511 | clauses pointers to the corresponding arrays. */ | |
3512 | ||
3513 | static tree | |
3514 | expand_omp_scantemp_alloc (tree clauses, tree ptr, unsigned HOST_WIDE_INT sz, | |
3515 | unsigned HOST_WIDE_INT alloc_align, tree cnt, | |
3516 | gimple_stmt_iterator *gsi, bool alloc) | |
3517 | { | |
3518 | tree eltsz = NULL_TREE; | |
3519 | unsigned HOST_WIDE_INT preval = 0; | |
3520 | if (ptr && sz) | |
3521 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), | |
3522 | ptr, size_int (sz)); | |
3523 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
3524 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
3525 | && !OMP_CLAUSE__SCANTEMP__CONTROL (c) | |
3526 | && (!OMP_CLAUSE__SCANTEMP__ALLOC (c)) != alloc) | |
3527 | { | |
3528 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); | |
3529 | unsigned HOST_WIDE_INT al = TYPE_ALIGN_UNIT (pointee_type); | |
3530 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) | |
3531 | { | |
3532 | unsigned HOST_WIDE_INT szl | |
3533 | = tree_to_uhwi (TYPE_SIZE_UNIT (pointee_type)); | |
3534 | szl = least_bit_hwi (szl); | |
3535 | if (szl) | |
3536 | al = MIN (al, szl); | |
3537 | } | |
3538 | if (ptr == NULL_TREE) | |
3539 | { | |
3540 | if (eltsz == NULL_TREE) | |
3541 | eltsz = TYPE_SIZE_UNIT (pointee_type); | |
3542 | else | |
3543 | eltsz = size_binop (PLUS_EXPR, eltsz, | |
3544 | TYPE_SIZE_UNIT (pointee_type)); | |
3545 | } | |
3546 | if (preval == 0 && al <= alloc_align) | |
3547 | { | |
3548 | unsigned HOST_WIDE_INT diff = ROUND_UP (sz, al) - sz; | |
3549 | sz += diff; | |
3550 | if (diff && ptr) | |
3551 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), | |
3552 | ptr, size_int (diff)); | |
3553 | } | |
3554 | else if (al > preval) | |
3555 | { | |
3556 | if (ptr) | |
3557 | { | |
3558 | ptr = fold_convert (pointer_sized_int_node, ptr); | |
3559 | ptr = fold_build2 (PLUS_EXPR, pointer_sized_int_node, ptr, | |
3560 | build_int_cst (pointer_sized_int_node, | |
3561 | al - 1)); | |
3562 | ptr = fold_build2 (BIT_AND_EXPR, pointer_sized_int_node, ptr, | |
3563 | build_int_cst (pointer_sized_int_node, | |
3564 | -(HOST_WIDE_INT) al)); | |
3565 | ptr = fold_convert (ptr_type_node, ptr); | |
3566 | } | |
3567 | else | |
3568 | sz += al - 1; | |
3569 | } | |
3570 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) | |
3571 | preval = al; | |
3572 | else | |
3573 | preval = 1; | |
3574 | if (ptr) | |
3575 | { | |
3576 | expand_omp_build_assign (gsi, OMP_CLAUSE_DECL (c), ptr, false); | |
3577 | ptr = OMP_CLAUSE_DECL (c); | |
3578 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), ptr, | |
3579 | size_binop (MULT_EXPR, cnt, | |
3580 | TYPE_SIZE_UNIT (pointee_type))); | |
3581 | } | |
3582 | } | |
3583 | ||
3584 | if (ptr == NULL_TREE) | |
3585 | { | |
3586 | eltsz = size_binop (MULT_EXPR, eltsz, cnt); | |
3587 | if (sz) | |
3588 | eltsz = size_binop (PLUS_EXPR, eltsz, size_int (sz)); | |
3589 | return eltsz; | |
3590 | } | |
3591 | else | |
3592 | return ptr; | |
3593 | } | |
3594 | ||
629b3d75 MJ |
3595 | /* A subroutine of expand_omp_for. Generate code for a parallel |
3596 | loop with static schedule and no specified chunk size. Given | |
3597 | parameters: | |
3598 | ||
3599 | for (V = N1; V cond N2; V += STEP) BODY; | |
3600 | ||
3601 | where COND is "<" or ">", we generate pseudocode | |
3602 | ||
3603 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; | |
3604 | if (cond is <) | |
3605 | adj = STEP - 1; | |
3606 | else | |
3607 | adj = STEP + 1; | |
3608 | if ((__typeof (V)) -1 > 0 && cond is >) | |
3609 | n = -(adj + N2 - N1) / -STEP; | |
3610 | else | |
3611 | n = (adj + N2 - N1) / STEP; | |
3612 | q = n / nthreads; | |
3613 | tt = n % nthreads; | |
3614 | if (threadid < tt) goto L3; else goto L4; | |
3615 | L3: | |
3616 | tt = 0; | |
3617 | q = q + 1; | |
3618 | L4: | |
3619 | s0 = q * threadid + tt; | |
3620 | e0 = s0 + q; | |
3621 | V = s0 * STEP + N1; | |
3622 | if (s0 >= e0) goto L2; else goto L0; | |
3623 | L0: | |
3624 | e = e0 * STEP + N1; | |
3625 | L1: | |
3626 | BODY; | |
3627 | V += STEP; | |
3628 | if (V cond e) goto L1; | |
3629 | L2: | |
3630 | */ | |
3631 | ||
3632 | static void | |
3633 | expand_omp_for_static_nochunk (struct omp_region *region, | |
3634 | struct omp_for_data *fd, | |
3635 | gimple *inner_stmt) | |
3636 | { | |
2f6bb511 | 3637 | tree n, q, s0, e0, e, t, tt, nthreads = NULL_TREE, threadid; |
629b3d75 MJ |
3638 | tree type, itype, vmain, vback; |
3639 | basic_block entry_bb, second_bb, third_bb, exit_bb, seq_start_bb; | |
3640 | basic_block body_bb, cont_bb, collapse_bb = NULL; | |
2f6bb511 JJ |
3641 | basic_block fin_bb, fourth_bb = NULL, fifth_bb = NULL, sixth_bb = NULL; |
3642 | basic_block exit1_bb = NULL, exit2_bb = NULL, exit3_bb = NULL; | |
6c7ae8c5 | 3643 | gimple_stmt_iterator gsi, gsip; |
629b3d75 MJ |
3644 | edge ep; |
3645 | bool broken_loop = region->cont == NULL; | |
3646 | tree *counts = NULL; | |
3647 | tree n1, n2, step; | |
28567c40 | 3648 | tree reductions = NULL_TREE; |
8221c30b | 3649 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; |
629b3d75 MJ |
3650 | |
3651 | itype = type = TREE_TYPE (fd->loop.v); | |
3652 | if (POINTER_TYPE_P (type)) | |
3653 | itype = signed_type_for (type); | |
3654 | ||
3655 | entry_bb = region->entry; | |
3656 | cont_bb = region->cont; | |
3657 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
3658 | fin_bb = BRANCH_EDGE (entry_bb)->dest; | |
3659 | gcc_assert (broken_loop | |
3660 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
3661 | seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb)); | |
3662 | body_bb = single_succ (seq_start_bb); | |
3663 | if (!broken_loop) | |
3664 | { | |
3665 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb | |
3666 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); | |
3667 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
3668 | } | |
3669 | exit_bb = region->exit; | |
3670 | ||
3671 | /* Iteration space partitioning goes in ENTRY_BB. */ | |
65f4b875 | 3672 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 | 3673 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
6c7ae8c5 JJ |
3674 | gsip = gsi; |
3675 | gsi_prev (&gsip); | |
629b3d75 MJ |
3676 | |
3677 | if (fd->collapse > 1) | |
3678 | { | |
3679 | int first_zero_iter = -1, dummy = -1; | |
3680 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
3681 | ||
3682 | counts = XALLOCAVEC (tree, fd->collapse); | |
3683 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
3684 | fin_bb, first_zero_iter, | |
3685 | dummy_bb, dummy, l2_dom_bb); | |
3686 | t = NULL_TREE; | |
3687 | } | |
3688 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
3689 | t = integer_one_node; | |
3690 | else | |
3691 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
3692 | fold_convert (type, fd->loop.n1), | |
3693 | fold_convert (type, fd->loop.n2)); | |
3694 | if (fd->collapse == 1 | |
3695 | && TYPE_UNSIGNED (type) | |
3696 | && (t == NULL_TREE || !integer_onep (t))) | |
3697 | { | |
3698 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); | |
3699 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
3700 | true, GSI_SAME_STMT); | |
3701 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
3702 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
3703 | true, GSI_SAME_STMT); | |
3704 | gcond *cond_stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, | |
6c7ae8c5 | 3705 | NULL_TREE, NULL_TREE); |
629b3d75 MJ |
3706 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); |
3707 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
3708 | expand_omp_regimplify_p, NULL, NULL) | |
3709 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
3710 | expand_omp_regimplify_p, NULL, NULL)) | |
3711 | { | |
3712 | gsi = gsi_for_stmt (cond_stmt); | |
3713 | gimple_regimplify_operands (cond_stmt, &gsi); | |
3714 | } | |
3715 | ep = split_block (entry_bb, cond_stmt); | |
3716 | ep->flags = EDGE_TRUE_VALUE; | |
3717 | entry_bb = ep->dest; | |
357067f2 | 3718 | ep->probability = profile_probability::very_likely (); |
629b3d75 | 3719 | ep = make_edge (ep->src, fin_bb, EDGE_FALSE_VALUE); |
357067f2 | 3720 | ep->probability = profile_probability::very_unlikely (); |
629b3d75 MJ |
3721 | if (gimple_in_ssa_p (cfun)) |
3722 | { | |
3723 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; | |
3724 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); | |
3725 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
3726 | { | |
3727 | gphi *phi = gpi.phi (); | |
3728 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
3729 | ep, UNKNOWN_LOCATION); | |
3730 | } | |
3731 | } | |
3732 | gsi = gsi_last_bb (entry_bb); | |
3733 | } | |
3734 | ||
8221c30b JJ |
3735 | if (fd->lastprivate_conditional) |
3736 | { | |
3737 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
3738 | tree c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); | |
3739 | if (fd->have_pointer_condtemp) | |
3740 | condtemp = OMP_CLAUSE_DECL (c); | |
3741 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
3742 | cond_var = OMP_CLAUSE_DECL (c); | |
3743 | } | |
2f6bb511 | 3744 | if (fd->have_reductemp |
2f03073f JJ |
3745 | /* For scan, we don't want to reinitialize condtemp before the |
3746 | second loop. */ | |
3747 | || (fd->have_pointer_condtemp && !fd->have_scantemp) | |
2f6bb511 | 3748 | || fd->have_nonctrl_scantemp) |
28567c40 JJ |
3749 | { |
3750 | tree t1 = build_int_cst (long_integer_type_node, 0); | |
3751 | tree t2 = build_int_cst (long_integer_type_node, 1); | |
3752 | tree t3 = build_int_cstu (long_integer_type_node, | |
3753 | (HOST_WIDE_INT_1U << 31) + 1); | |
3754 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
6c7ae8c5 JJ |
3755 | gimple_stmt_iterator gsi2 = gsi_none (); |
3756 | gimple *g = NULL; | |
3757 | tree mem = null_pointer_node, memv = NULL_TREE; | |
2f6bb511 JJ |
3758 | unsigned HOST_WIDE_INT condtemp_sz = 0; |
3759 | unsigned HOST_WIDE_INT alloc_align = 0; | |
6c7ae8c5 JJ |
3760 | if (fd->have_reductemp) |
3761 | { | |
2f6bb511 | 3762 | gcc_assert (!fd->have_nonctrl_scantemp); |
6c7ae8c5 JJ |
3763 | tree c = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); |
3764 | reductions = OMP_CLAUSE_DECL (c); | |
3765 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
3766 | g = SSA_NAME_DEF_STMT (reductions); | |
3767 | reductions = gimple_assign_rhs1 (g); | |
3768 | OMP_CLAUSE_DECL (c) = reductions; | |
3769 | gsi2 = gsi_for_stmt (g); | |
3770 | } | |
3771 | else | |
3772 | { | |
3773 | if (gsi_end_p (gsip)) | |
3774 | gsi2 = gsi_after_labels (region->entry); | |
3775 | else | |
3776 | gsi2 = gsip; | |
3777 | reductions = null_pointer_node; | |
3778 | } | |
2f6bb511 | 3779 | if (fd->have_pointer_condtemp || fd->have_nonctrl_scantemp) |
6c7ae8c5 | 3780 | { |
2f6bb511 JJ |
3781 | tree type; |
3782 | if (fd->have_pointer_condtemp) | |
3783 | type = TREE_TYPE (condtemp); | |
3784 | else | |
3785 | type = ptr_type_node; | |
6c7ae8c5 JJ |
3786 | memv = create_tmp_var (type); |
3787 | TREE_ADDRESSABLE (memv) = 1; | |
2f6bb511 JJ |
3788 | unsigned HOST_WIDE_INT sz = 0; |
3789 | tree size = NULL_TREE; | |
3790 | if (fd->have_pointer_condtemp) | |
3791 | { | |
3792 | sz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
3793 | sz *= fd->lastprivate_conditional; | |
3794 | condtemp_sz = sz; | |
3795 | } | |
3796 | if (fd->have_nonctrl_scantemp) | |
3797 | { | |
3798 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
3799 | gimple *g = gimple_build_call (nthreads, 0); | |
3800 | nthreads = create_tmp_var (integer_type_node); | |
3801 | gimple_call_set_lhs (g, nthreads); | |
3802 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); | |
3803 | nthreads = fold_convert (sizetype, nthreads); | |
3804 | alloc_align = TYPE_ALIGN_UNIT (long_long_integer_type_node); | |
3805 | size = expand_omp_scantemp_alloc (clauses, NULL_TREE, sz, | |
3806 | alloc_align, nthreads, NULL, | |
3807 | false); | |
3808 | size = fold_convert (type, size); | |
3809 | } | |
3810 | else | |
3811 | size = build_int_cst (type, sz); | |
3812 | expand_omp_build_assign (&gsi2, memv, size, false); | |
6c7ae8c5 JJ |
3813 | mem = build_fold_addr_expr (memv); |
3814 | } | |
28567c40 JJ |
3815 | tree t |
3816 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_LOOP_START), | |
3817 | 9, t1, t2, t2, t3, t1, null_pointer_node, | |
6c7ae8c5 | 3818 | null_pointer_node, reductions, mem); |
28567c40 JJ |
3819 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3820 | true, GSI_SAME_STMT); | |
8221c30b | 3821 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 3822 | expand_omp_build_assign (&gsi2, condtemp, memv, false); |
2f6bb511 JJ |
3823 | if (fd->have_nonctrl_scantemp) |
3824 | { | |
3825 | tree ptr = fd->have_pointer_condtemp ? condtemp : memv; | |
3826 | expand_omp_scantemp_alloc (clauses, ptr, condtemp_sz, | |
3827 | alloc_align, nthreads, &gsi2, false); | |
3828 | } | |
6c7ae8c5 JJ |
3829 | if (fd->have_reductemp) |
3830 | { | |
3831 | gsi_remove (&gsi2, true); | |
3832 | release_ssa_name (gimple_assign_lhs (g)); | |
3833 | } | |
28567c40 | 3834 | } |
629b3d75 MJ |
3835 | switch (gimple_omp_for_kind (fd->for_stmt)) |
3836 | { | |
3837 | case GF_OMP_FOR_KIND_FOR: | |
3838 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
3839 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
3840 | break; | |
3841 | case GF_OMP_FOR_KIND_DISTRIBUTE: | |
3842 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS); | |
3843 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM); | |
3844 | break; | |
3845 | default: | |
3846 | gcc_unreachable (); | |
3847 | } | |
3848 | nthreads = build_call_expr (nthreads, 0); | |
3849 | nthreads = fold_convert (itype, nthreads); | |
3850 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, | |
3851 | true, GSI_SAME_STMT); | |
3852 | threadid = build_call_expr (threadid, 0); | |
3853 | threadid = fold_convert (itype, threadid); | |
3854 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, | |
3855 | true, GSI_SAME_STMT); | |
3856 | ||
3857 | n1 = fd->loop.n1; | |
3858 | n2 = fd->loop.n2; | |
3859 | step = fd->loop.step; | |
3860 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
3861 | { | |
3862 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
3863 | OMP_CLAUSE__LOOPTEMP_); | |
3864 | gcc_assert (innerc); | |
3865 | n1 = OMP_CLAUSE_DECL (innerc); | |
3866 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
3867 | OMP_CLAUSE__LOOPTEMP_); | |
3868 | gcc_assert (innerc); | |
3869 | n2 = OMP_CLAUSE_DECL (innerc); | |
3870 | } | |
3871 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
3872 | true, NULL_TREE, true, GSI_SAME_STMT); | |
3873 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
3874 | true, NULL_TREE, true, GSI_SAME_STMT); | |
3875 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
3876 | true, NULL_TREE, true, GSI_SAME_STMT); | |
3877 | ||
3878 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
3879 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
3880 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
3881 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
3882 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
3883 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
3884 | fold_build1 (NEGATE_EXPR, itype, t), | |
3885 | fold_build1 (NEGATE_EXPR, itype, step)); | |
3886 | else | |
3887 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
3888 | t = fold_convert (itype, t); | |
3889 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
3890 | ||
3891 | q = create_tmp_reg (itype, "q"); | |
3892 | t = fold_build2 (TRUNC_DIV_EXPR, itype, n, nthreads); | |
3893 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
3894 | gsi_insert_before (&gsi, gimple_build_assign (q, t), GSI_SAME_STMT); | |
3895 | ||
3896 | tt = create_tmp_reg (itype, "tt"); | |
3897 | t = fold_build2 (TRUNC_MOD_EXPR, itype, n, nthreads); | |
3898 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); | |
3899 | gsi_insert_before (&gsi, gimple_build_assign (tt, t), GSI_SAME_STMT); | |
3900 | ||
3901 | t = build2 (LT_EXPR, boolean_type_node, threadid, tt); | |
3902 | gcond *cond_stmt = gimple_build_cond_empty (t); | |
3903 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
3904 | ||
3905 | second_bb = split_block (entry_bb, cond_stmt)->dest; | |
65f4b875 | 3906 | gsi = gsi_last_nondebug_bb (second_bb); |
629b3d75 MJ |
3907 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
3908 | ||
3909 | gsi_insert_before (&gsi, gimple_build_assign (tt, build_int_cst (itype, 0)), | |
3910 | GSI_SAME_STMT); | |
3911 | gassign *assign_stmt | |
3912 | = gimple_build_assign (q, PLUS_EXPR, q, build_int_cst (itype, 1)); | |
3913 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
3914 | ||
3915 | third_bb = split_block (second_bb, assign_stmt)->dest; | |
65f4b875 | 3916 | gsi = gsi_last_nondebug_bb (third_bb); |
629b3d75 MJ |
3917 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
3918 | ||
2f6bb511 JJ |
3919 | if (fd->have_nonctrl_scantemp) |
3920 | { | |
3921 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
3922 | tree controlp = NULL_TREE, controlb = NULL_TREE; | |
3923 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
3924 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
3925 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) | |
3926 | { | |
3927 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) | |
3928 | controlb = OMP_CLAUSE_DECL (c); | |
3929 | else | |
3930 | controlp = OMP_CLAUSE_DECL (c); | |
3931 | if (controlb && controlp) | |
3932 | break; | |
3933 | } | |
3934 | gcc_assert (controlp && controlb); | |
3935 | tree cnt = create_tmp_var (sizetype); | |
3936 | gimple *g = gimple_build_assign (cnt, NOP_EXPR, q); | |
3937 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
3938 | unsigned HOST_WIDE_INT alloc_align = TYPE_ALIGN_UNIT (ptr_type_node); | |
3939 | tree sz = expand_omp_scantemp_alloc (clauses, NULL_TREE, 0, | |
3940 | alloc_align, cnt, NULL, true); | |
3941 | tree size = create_tmp_var (sizetype); | |
3942 | expand_omp_build_assign (&gsi, size, sz, false); | |
3943 | tree cmp = fold_build2 (GT_EXPR, boolean_type_node, | |
3944 | size, size_int (16384)); | |
3945 | expand_omp_build_assign (&gsi, controlb, cmp); | |
3946 | g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, | |
3947 | NULL_TREE, NULL_TREE); | |
3948 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
3949 | fourth_bb = split_block (third_bb, g)->dest; | |
3950 | gsi = gsi_last_nondebug_bb (fourth_bb); | |
3951 | /* FIXME: Once we have allocators, this should use allocator. */ | |
3952 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_MALLOC), 1, size); | |
3953 | gimple_call_set_lhs (g, controlp); | |
3954 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
3955 | expand_omp_scantemp_alloc (clauses, controlp, 0, alloc_align, cnt, | |
3956 | &gsi, true); | |
3957 | gsi_prev (&gsi); | |
3958 | g = gsi_stmt (gsi); | |
3959 | fifth_bb = split_block (fourth_bb, g)->dest; | |
3960 | gsi = gsi_last_nondebug_bb (fifth_bb); | |
3961 | ||
3962 | g = gimple_build_call (builtin_decl_implicit (BUILT_IN_STACK_SAVE), 0); | |
3963 | gimple_call_set_lhs (g, controlp); | |
3964 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
3965 | tree alloca_decl = builtin_decl_explicit (BUILT_IN_ALLOCA_WITH_ALIGN); | |
3966 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
3967 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
3968 | && OMP_CLAUSE__SCANTEMP__ALLOC (c)) | |
3969 | { | |
3970 | tree tmp = create_tmp_var (sizetype); | |
3971 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); | |
3972 | g = gimple_build_assign (tmp, MULT_EXPR, cnt, | |
3973 | TYPE_SIZE_UNIT (pointee_type)); | |
3974 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
3975 | g = gimple_build_call (alloca_decl, 2, tmp, | |
3976 | size_int (TYPE_ALIGN (pointee_type))); | |
3977 | gimple_call_set_lhs (g, OMP_CLAUSE_DECL (c)); | |
3978 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
3979 | } | |
3980 | ||
3981 | sixth_bb = split_block (fifth_bb, g)->dest; | |
3982 | gsi = gsi_last_nondebug_bb (sixth_bb); | |
3983 | } | |
3984 | ||
629b3d75 MJ |
3985 | t = build2 (MULT_EXPR, itype, q, threadid); |
3986 | t = build2 (PLUS_EXPR, itype, t, tt); | |
3987 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
3988 | ||
3989 | t = fold_build2 (PLUS_EXPR, itype, s0, q); | |
3990 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
3991 | ||
3992 | t = build2 (GE_EXPR, boolean_type_node, s0, e0); | |
3993 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
3994 | ||
3995 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
3996 | gsi_remove (&gsi, true); | |
3997 | ||
3998 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
3999 | gsi = gsi_start_bb (seq_start_bb); | |
4000 | ||
4001 | tree startvar = fd->loop.v; | |
4002 | tree endvar = NULL_TREE; | |
4003 | ||
4004 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4005 | { | |
4006 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
4007 | ? gimple_omp_parallel_clauses (inner_stmt) | |
4008 | : gimple_omp_for_clauses (inner_stmt); | |
4009 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
4010 | gcc_assert (innerc); | |
4011 | startvar = OMP_CLAUSE_DECL (innerc); | |
4012 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4013 | OMP_CLAUSE__LOOPTEMP_); | |
4014 | gcc_assert (innerc); | |
4015 | endvar = OMP_CLAUSE_DECL (innerc); | |
4016 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST | |
4017 | && gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) | |
4018 | { | |
4019 | int i; | |
4020 | for (i = 1; i < fd->collapse; i++) | |
4021 | { | |
4022 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4023 | OMP_CLAUSE__LOOPTEMP_); | |
4024 | gcc_assert (innerc); | |
4025 | } | |
4026 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4027 | OMP_CLAUSE__LOOPTEMP_); | |
4028 | if (innerc) | |
4029 | { | |
4030 | /* If needed (distribute parallel for with lastprivate), | |
4031 | propagate down the total number of iterations. */ | |
4032 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), | |
4033 | fd->loop.n2); | |
4034 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, | |
4035 | GSI_CONTINUE_LINKING); | |
4036 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
4037 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4038 | } | |
4039 | } | |
4040 | } | |
4041 | t = fold_convert (itype, s0); | |
4042 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
4043 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
4044 | { |
4045 | t = fold_build_pointer_plus (n1, t); | |
4046 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
4047 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
4048 | t = fold_convert (signed_type_for (type), t); | |
4049 | } | |
629b3d75 MJ |
4050 | else |
4051 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
4052 | t = fold_convert (TREE_TYPE (startvar), t); | |
4053 | t = force_gimple_operand_gsi (&gsi, t, | |
4054 | DECL_P (startvar) | |
4055 | && TREE_ADDRESSABLE (startvar), | |
4056 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4057 | assign_stmt = gimple_build_assign (startvar, t); | |
4058 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6c7ae8c5 JJ |
4059 | if (cond_var) |
4060 | { | |
4061 | tree itype = TREE_TYPE (cond_var); | |
4062 | /* For lastprivate(conditional:) itervar, we need some iteration | |
4063 | counter that starts at unsigned non-zero and increases. | |
4064 | Prefer as few IVs as possible, so if we can use startvar | |
4065 | itself, use that, or startvar + constant (those would be | |
4066 | incremented with step), and as last resort use the s0 + 1 | |
4067 | incremented by 1. */ | |
4068 | if (POINTER_TYPE_P (type) | |
4069 | || TREE_CODE (n1) != INTEGER_CST | |
4070 | || fd->loop.cond_code != LT_EXPR) | |
4071 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), | |
4072 | build_int_cst (itype, 1)); | |
4073 | else if (tree_int_cst_sgn (n1) == 1) | |
4074 | t = fold_convert (itype, t); | |
4075 | else | |
4076 | { | |
4077 | tree c = fold_convert (itype, n1); | |
4078 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
4079 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
4080 | } | |
4081 | t = force_gimple_operand_gsi (&gsi, t, false, | |
4082 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4083 | assign_stmt = gimple_build_assign (cond_var, t); | |
4084 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4085 | } | |
629b3d75 MJ |
4086 | |
4087 | t = fold_convert (itype, e0); | |
4088 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
4089 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
4090 | { |
4091 | t = fold_build_pointer_plus (n1, t); | |
4092 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
4093 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
4094 | t = fold_convert (signed_type_for (type), t); | |
4095 | } | |
629b3d75 MJ |
4096 | else |
4097 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
4098 | t = fold_convert (TREE_TYPE (startvar), t); | |
4099 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4100 | false, GSI_CONTINUE_LINKING); | |
4101 | if (endvar) | |
4102 | { | |
4103 | assign_stmt = gimple_build_assign (endvar, e); | |
4104 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4105 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
4106 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
4107 | else | |
4108 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
4109 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4110 | } | |
4111 | /* Handle linear clause adjustments. */ | |
4112 | tree itercnt = NULL_TREE; | |
4113 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
4114 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
4115 | c; c = OMP_CLAUSE_CHAIN (c)) | |
4116 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
4117 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
4118 | { | |
4119 | tree d = OMP_CLAUSE_DECL (c); | |
4120 | bool is_ref = omp_is_reference (d); | |
4121 | tree t = d, a, dest; | |
4122 | if (is_ref) | |
4123 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); | |
4124 | if (itercnt == NULL_TREE) | |
4125 | { | |
4126 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
4127 | { | |
4128 | itercnt = fold_build2 (MINUS_EXPR, itype, | |
4129 | fold_convert (itype, n1), | |
4130 | fold_convert (itype, fd->loop.n1)); | |
4131 | itercnt = fold_build2 (EXACT_DIV_EXPR, itype, itercnt, step); | |
4132 | itercnt = fold_build2 (PLUS_EXPR, itype, itercnt, s0); | |
4133 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
4134 | NULL_TREE, false, | |
4135 | GSI_CONTINUE_LINKING); | |
4136 | } | |
4137 | else | |
4138 | itercnt = s0; | |
4139 | } | |
4140 | tree type = TREE_TYPE (t); | |
4141 | if (POINTER_TYPE_P (type)) | |
4142 | type = sizetype; | |
4143 | a = fold_build2 (MULT_EXPR, type, | |
4144 | fold_convert (type, itercnt), | |
4145 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
4146 | dest = unshare_expr (t); | |
4147 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
4148 | : POINTER_PLUS_EXPR, TREE_TYPE (t), t, a); | |
4149 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4150 | false, GSI_CONTINUE_LINKING); | |
4151 | assign_stmt = gimple_build_assign (dest, t); | |
4152 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4153 | } | |
4154 | if (fd->collapse > 1) | |
4155 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); | |
4156 | ||
4157 | if (!broken_loop) | |
4158 | { | |
4159 | /* The code controlling the sequential loop replaces the | |
4160 | GIMPLE_OMP_CONTINUE. */ | |
65f4b875 | 4161 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
4162 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
4163 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
4164 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
4165 | vback = gimple_omp_continue_control_def (cont_stmt); | |
4166 | ||
7855700e JJ |
4167 | if (cond_var) |
4168 | { | |
4169 | tree itype = TREE_TYPE (cond_var); | |
4170 | tree t2; | |
4171 | if (POINTER_TYPE_P (type) | |
4172 | || TREE_CODE (n1) != INTEGER_CST | |
4173 | || fd->loop.cond_code != LT_EXPR) | |
4174 | t2 = build_int_cst (itype, 1); | |
4175 | else | |
4176 | t2 = fold_convert (itype, step); | |
4177 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
4178 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
4179 | NULL_TREE, true, GSI_SAME_STMT); | |
4180 | assign_stmt = gimple_build_assign (cond_var, t2); | |
4181 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4182 | } | |
4183 | ||
629b3d75 MJ |
4184 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
4185 | { | |
4186 | if (POINTER_TYPE_P (type)) | |
4187 | t = fold_build_pointer_plus (vmain, step); | |
4188 | else | |
4189 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
4190 | t = force_gimple_operand_gsi (&gsi, t, | |
4191 | DECL_P (vback) | |
4192 | && TREE_ADDRESSABLE (vback), | |
4193 | NULL_TREE, true, GSI_SAME_STMT); | |
4194 | assign_stmt = gimple_build_assign (vback, t); | |
4195 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4196 | ||
4197 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
4198 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
4199 | ? t : vback, e); | |
4200 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
4201 | } | |
4202 | ||
4203 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ | |
4204 | gsi_remove (&gsi, true); | |
4205 | ||
4206 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
4207 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, body_bb); | |
4208 | } | |
4209 | ||
4210 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ | |
65f4b875 | 4211 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
4212 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) |
4213 | { | |
4214 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
2f6bb511 JJ |
4215 | if (fd->have_reductemp |
4216 | || ((fd->have_pointer_condtemp || fd->have_scantemp) | |
4217 | && !fd->have_nonctrl_scantemp)) | |
28567c40 JJ |
4218 | { |
4219 | tree fn; | |
4220 | if (t) | |
4221 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
4222 | else | |
4223 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
4224 | gcall *g = gimple_build_call (fn, 0); | |
4225 | if (t) | |
4226 | { | |
4227 | gimple_call_set_lhs (g, t); | |
6c7ae8c5 JJ |
4228 | if (fd->have_reductemp) |
4229 | gsi_insert_after (&gsi, gimple_build_assign (reductions, | |
4230 | NOP_EXPR, t), | |
4231 | GSI_SAME_STMT); | |
28567c40 JJ |
4232 | } |
4233 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
4234 | } | |
4235 | else | |
4236 | gsi_insert_after (&gsi, omp_build_barrier (t), GSI_SAME_STMT); | |
629b3d75 | 4237 | } |
2f03073f JJ |
4238 | else if ((fd->have_pointer_condtemp || fd->have_scantemp) |
4239 | && !fd->have_nonctrl_scantemp) | |
1a39b3d3 JJ |
4240 | { |
4241 | tree fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
4242 | gcall *g = gimple_build_call (fn, 0); | |
4243 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
4244 | } | |
2f6bb511 JJ |
4245 | if (fd->have_scantemp && !fd->have_nonctrl_scantemp) |
4246 | { | |
4247 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
4248 | tree controlp = NULL_TREE, controlb = NULL_TREE; | |
4249 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
4250 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ | |
4251 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) | |
4252 | { | |
4253 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) | |
4254 | controlb = OMP_CLAUSE_DECL (c); | |
4255 | else | |
4256 | controlp = OMP_CLAUSE_DECL (c); | |
4257 | if (controlb && controlp) | |
4258 | break; | |
4259 | } | |
4260 | gcc_assert (controlp && controlb); | |
4261 | gimple *g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, | |
4262 | NULL_TREE, NULL_TREE); | |
4263 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
4264 | exit1_bb = split_block (exit_bb, g)->dest; | |
4265 | gsi = gsi_after_labels (exit1_bb); | |
4266 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_FREE), 1, | |
4267 | controlp); | |
4268 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
4269 | exit2_bb = split_block (exit1_bb, g)->dest; | |
4270 | gsi = gsi_after_labels (exit2_bb); | |
4271 | g = gimple_build_call (builtin_decl_implicit (BUILT_IN_STACK_RESTORE), 1, | |
4272 | controlp); | |
4273 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
4274 | exit3_bb = split_block (exit2_bb, g)->dest; | |
4275 | gsi = gsi_after_labels (exit3_bb); | |
4276 | } | |
629b3d75 MJ |
4277 | gsi_remove (&gsi, true); |
4278 | ||
4279 | /* Connect all the blocks. */ | |
4280 | ep = make_edge (entry_bb, third_bb, EDGE_FALSE_VALUE); | |
357067f2 | 4281 | ep->probability = profile_probability::guessed_always ().apply_scale (3, 4); |
629b3d75 MJ |
4282 | ep = find_edge (entry_bb, second_bb); |
4283 | ep->flags = EDGE_TRUE_VALUE; | |
357067f2 | 4284 | ep->probability = profile_probability::guessed_always ().apply_scale (1, 4); |
2f6bb511 JJ |
4285 | if (fourth_bb) |
4286 | { | |
4287 | ep = make_edge (third_bb, fifth_bb, EDGE_FALSE_VALUE); | |
4288 | ep->probability | |
4289 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
4290 | ep = find_edge (third_bb, fourth_bb); | |
4291 | ep->flags = EDGE_TRUE_VALUE; | |
4292 | ep->probability | |
4293 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
4294 | ep = find_edge (fourth_bb, fifth_bb); | |
4295 | redirect_edge_and_branch (ep, sixth_bb); | |
4296 | } | |
4297 | else | |
4298 | sixth_bb = third_bb; | |
4299 | find_edge (sixth_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE; | |
4300 | find_edge (sixth_bb, fin_bb)->flags = EDGE_TRUE_VALUE; | |
4301 | if (exit1_bb) | |
4302 | { | |
4303 | ep = make_edge (exit_bb, exit2_bb, EDGE_FALSE_VALUE); | |
4304 | ep->probability | |
4305 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
4306 | ep = find_edge (exit_bb, exit1_bb); | |
4307 | ep->flags = EDGE_TRUE_VALUE; | |
4308 | ep->probability | |
4309 | = profile_probability::guessed_always ().apply_scale (1, 2); | |
4310 | ep = find_edge (exit1_bb, exit2_bb); | |
4311 | redirect_edge_and_branch (ep, exit3_bb); | |
4312 | } | |
629b3d75 MJ |
4313 | |
4314 | if (!broken_loop) | |
4315 | { | |
4316 | ep = find_edge (cont_bb, body_bb); | |
4317 | if (ep == NULL) | |
4318 | { | |
4319 | ep = BRANCH_EDGE (cont_bb); | |
4320 | gcc_assert (single_succ (ep->dest) == body_bb); | |
4321 | } | |
4322 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4323 | { | |
4324 | remove_edge (ep); | |
4325 | ep = NULL; | |
4326 | } | |
4327 | else if (fd->collapse > 1) | |
4328 | { | |
4329 | remove_edge (ep); | |
4330 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
4331 | } | |
4332 | else | |
4333 | ep->flags = EDGE_TRUE_VALUE; | |
4334 | find_edge (cont_bb, fin_bb)->flags | |
4335 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
4336 | } | |
4337 | ||
4338 | set_immediate_dominator (CDI_DOMINATORS, second_bb, entry_bb); | |
4339 | set_immediate_dominator (CDI_DOMINATORS, third_bb, entry_bb); | |
2f6bb511 JJ |
4340 | if (fourth_bb) |
4341 | { | |
4342 | set_immediate_dominator (CDI_DOMINATORS, fifth_bb, third_bb); | |
4343 | set_immediate_dominator (CDI_DOMINATORS, sixth_bb, third_bb); | |
4344 | } | |
4345 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, sixth_bb); | |
629b3d75 MJ |
4346 | |
4347 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
4348 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
4349 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
4350 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
2f6bb511 JJ |
4351 | if (exit1_bb) |
4352 | { | |
4353 | set_immediate_dominator (CDI_DOMINATORS, exit2_bb, exit_bb); | |
4354 | set_immediate_dominator (CDI_DOMINATORS, exit3_bb, exit_bb); | |
4355 | } | |
629b3d75 | 4356 | |
99b1c316 | 4357 | class loop *loop = body_bb->loop_father; |
629b3d75 MJ |
4358 | if (loop != entry_bb->loop_father) |
4359 | { | |
4360 | gcc_assert (broken_loop || loop->header == body_bb); | |
4361 | gcc_assert (broken_loop | |
4362 | || loop->latch == region->cont | |
4363 | || single_pred (loop->latch) == region->cont); | |
4364 | return; | |
4365 | } | |
4366 | ||
4367 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) | |
4368 | { | |
4369 | loop = alloc_loop (); | |
4370 | loop->header = body_bb; | |
4371 | if (collapse_bb == NULL) | |
4372 | loop->latch = cont_bb; | |
4373 | add_loop (loop, body_bb->loop_father); | |
4374 | } | |
4375 | } | |
4376 | ||
4377 | /* Return phi in E->DEST with ARG on edge E. */ | |
4378 | ||
4379 | static gphi * | |
4380 | find_phi_with_arg_on_edge (tree arg, edge e) | |
4381 | { | |
4382 | basic_block bb = e->dest; | |
4383 | ||
4384 | for (gphi_iterator gpi = gsi_start_phis (bb); | |
4385 | !gsi_end_p (gpi); | |
4386 | gsi_next (&gpi)) | |
4387 | { | |
4388 | gphi *phi = gpi.phi (); | |
4389 | if (PHI_ARG_DEF_FROM_EDGE (phi, e) == arg) | |
4390 | return phi; | |
4391 | } | |
4392 | ||
4393 | return NULL; | |
4394 | } | |
4395 | ||
4396 | /* A subroutine of expand_omp_for. Generate code for a parallel | |
4397 | loop with static schedule and a specified chunk size. Given | |
4398 | parameters: | |
4399 | ||
4400 | for (V = N1; V cond N2; V += STEP) BODY; | |
4401 | ||
4402 | where COND is "<" or ">", we generate pseudocode | |
4403 | ||
4404 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; | |
4405 | if (cond is <) | |
4406 | adj = STEP - 1; | |
4407 | else | |
4408 | adj = STEP + 1; | |
4409 | if ((__typeof (V)) -1 > 0 && cond is >) | |
4410 | n = -(adj + N2 - N1) / -STEP; | |
4411 | else | |
4412 | n = (adj + N2 - N1) / STEP; | |
4413 | trip = 0; | |
4414 | V = threadid * CHUNK * STEP + N1; -- this extra definition of V is | |
4415 | here so that V is defined | |
4416 | if the loop is not entered | |
4417 | L0: | |
4418 | s0 = (trip * nthreads + threadid) * CHUNK; | |
01914336 | 4419 | e0 = min (s0 + CHUNK, n); |
629b3d75 MJ |
4420 | if (s0 < n) goto L1; else goto L4; |
4421 | L1: | |
4422 | V = s0 * STEP + N1; | |
4423 | e = e0 * STEP + N1; | |
4424 | L2: | |
4425 | BODY; | |
4426 | V += STEP; | |
4427 | if (V cond e) goto L2; else goto L3; | |
4428 | L3: | |
4429 | trip += 1; | |
4430 | goto L0; | |
4431 | L4: | |
4432 | */ | |
4433 | ||
4434 | static void | |
4435 | expand_omp_for_static_chunk (struct omp_region *region, | |
4436 | struct omp_for_data *fd, gimple *inner_stmt) | |
4437 | { | |
4438 | tree n, s0, e0, e, t; | |
4439 | tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid; | |
4440 | tree type, itype, vmain, vback, vextra; | |
4441 | basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb; | |
4442 | basic_block trip_update_bb = NULL, cont_bb, collapse_bb = NULL, fin_bb; | |
6c7ae8c5 | 4443 | gimple_stmt_iterator gsi, gsip; |
629b3d75 MJ |
4444 | edge se; |
4445 | bool broken_loop = region->cont == NULL; | |
4446 | tree *counts = NULL; | |
4447 | tree n1, n2, step; | |
28567c40 | 4448 | tree reductions = NULL_TREE; |
8221c30b | 4449 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; |
629b3d75 MJ |
4450 | |
4451 | itype = type = TREE_TYPE (fd->loop.v); | |
4452 | if (POINTER_TYPE_P (type)) | |
4453 | itype = signed_type_for (type); | |
4454 | ||
4455 | entry_bb = region->entry; | |
4456 | se = split_block (entry_bb, last_stmt (entry_bb)); | |
4457 | entry_bb = se->src; | |
4458 | iter_part_bb = se->dest; | |
4459 | cont_bb = region->cont; | |
4460 | gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2); | |
4461 | fin_bb = BRANCH_EDGE (iter_part_bb)->dest; | |
4462 | gcc_assert (broken_loop | |
4463 | || fin_bb == FALLTHRU_EDGE (cont_bb)->dest); | |
4464 | seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb)); | |
4465 | body_bb = single_succ (seq_start_bb); | |
4466 | if (!broken_loop) | |
4467 | { | |
4468 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb | |
4469 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); | |
4470 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
4471 | trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb)); | |
4472 | } | |
4473 | exit_bb = region->exit; | |
4474 | ||
4475 | /* Trip and adjustment setup goes in ENTRY_BB. */ | |
65f4b875 | 4476 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 | 4477 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
6c7ae8c5 JJ |
4478 | gsip = gsi; |
4479 | gsi_prev (&gsip); | |
629b3d75 MJ |
4480 | |
4481 | if (fd->collapse > 1) | |
4482 | { | |
4483 | int first_zero_iter = -1, dummy = -1; | |
4484 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
4485 | ||
4486 | counts = XALLOCAVEC (tree, fd->collapse); | |
4487 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
4488 | fin_bb, first_zero_iter, | |
4489 | dummy_bb, dummy, l2_dom_bb); | |
4490 | t = NULL_TREE; | |
4491 | } | |
4492 | else if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
4493 | t = integer_one_node; | |
4494 | else | |
4495 | t = fold_binary (fd->loop.cond_code, boolean_type_node, | |
4496 | fold_convert (type, fd->loop.n1), | |
4497 | fold_convert (type, fd->loop.n2)); | |
4498 | if (fd->collapse == 1 | |
4499 | && TYPE_UNSIGNED (type) | |
4500 | && (t == NULL_TREE || !integer_onep (t))) | |
4501 | { | |
4502 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); | |
4503 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, | |
4504 | true, GSI_SAME_STMT); | |
4505 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); | |
4506 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, | |
4507 | true, GSI_SAME_STMT); | |
4508 | gcond *cond_stmt = gimple_build_cond (fd->loop.cond_code, n1, n2, | |
4509 | NULL_TREE, NULL_TREE); | |
4510 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); | |
4511 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), | |
4512 | expand_omp_regimplify_p, NULL, NULL) | |
4513 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), | |
4514 | expand_omp_regimplify_p, NULL, NULL)) | |
4515 | { | |
4516 | gsi = gsi_for_stmt (cond_stmt); | |
4517 | gimple_regimplify_operands (cond_stmt, &gsi); | |
4518 | } | |
4519 | se = split_block (entry_bb, cond_stmt); | |
4520 | se->flags = EDGE_TRUE_VALUE; | |
4521 | entry_bb = se->dest; | |
357067f2 | 4522 | se->probability = profile_probability::very_likely (); |
629b3d75 | 4523 | se = make_edge (se->src, fin_bb, EDGE_FALSE_VALUE); |
357067f2 | 4524 | se->probability = profile_probability::very_unlikely (); |
629b3d75 MJ |
4525 | if (gimple_in_ssa_p (cfun)) |
4526 | { | |
4527 | int dest_idx = find_edge (iter_part_bb, fin_bb)->dest_idx; | |
4528 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); | |
4529 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
4530 | { | |
4531 | gphi *phi = gpi.phi (); | |
4532 | add_phi_arg (phi, gimple_phi_arg_def (phi, dest_idx), | |
4533 | se, UNKNOWN_LOCATION); | |
4534 | } | |
4535 | } | |
4536 | gsi = gsi_last_bb (entry_bb); | |
4537 | } | |
4538 | ||
8221c30b JJ |
4539 | if (fd->lastprivate_conditional) |
4540 | { | |
4541 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
4542 | tree c = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); | |
4543 | if (fd->have_pointer_condtemp) | |
4544 | condtemp = OMP_CLAUSE_DECL (c); | |
4545 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), OMP_CLAUSE__CONDTEMP_); | |
4546 | cond_var = OMP_CLAUSE_DECL (c); | |
4547 | } | |
4548 | if (fd->have_reductemp || fd->have_pointer_condtemp) | |
28567c40 JJ |
4549 | { |
4550 | tree t1 = build_int_cst (long_integer_type_node, 0); | |
4551 | tree t2 = build_int_cst (long_integer_type_node, 1); | |
4552 | tree t3 = build_int_cstu (long_integer_type_node, | |
4553 | (HOST_WIDE_INT_1U << 31) + 1); | |
4554 | tree clauses = gimple_omp_for_clauses (fd->for_stmt); | |
6c7ae8c5 JJ |
4555 | gimple_stmt_iterator gsi2 = gsi_none (); |
4556 | gimple *g = NULL; | |
4557 | tree mem = null_pointer_node, memv = NULL_TREE; | |
6c7ae8c5 JJ |
4558 | if (fd->have_reductemp) |
4559 | { | |
4560 | tree c = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
4561 | reductions = OMP_CLAUSE_DECL (c); | |
4562 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
4563 | g = SSA_NAME_DEF_STMT (reductions); | |
4564 | reductions = gimple_assign_rhs1 (g); | |
4565 | OMP_CLAUSE_DECL (c) = reductions; | |
4566 | gsi2 = gsi_for_stmt (g); | |
4567 | } | |
4568 | else | |
4569 | { | |
4570 | if (gsi_end_p (gsip)) | |
4571 | gsi2 = gsi_after_labels (region->entry); | |
4572 | else | |
4573 | gsi2 = gsip; | |
4574 | reductions = null_pointer_node; | |
4575 | } | |
8221c30b | 4576 | if (fd->have_pointer_condtemp) |
6c7ae8c5 | 4577 | { |
6c7ae8c5 JJ |
4578 | tree type = TREE_TYPE (condtemp); |
4579 | memv = create_tmp_var (type); | |
4580 | TREE_ADDRESSABLE (memv) = 1; | |
4581 | unsigned HOST_WIDE_INT sz | |
4582 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); | |
4583 | sz *= fd->lastprivate_conditional; | |
4584 | expand_omp_build_assign (&gsi2, memv, build_int_cst (type, sz), | |
4585 | false); | |
4586 | mem = build_fold_addr_expr (memv); | |
4587 | } | |
28567c40 JJ |
4588 | tree t |
4589 | = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_LOOP_START), | |
4590 | 9, t1, t2, t2, t3, t1, null_pointer_node, | |
6c7ae8c5 | 4591 | null_pointer_node, reductions, mem); |
28567c40 JJ |
4592 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
4593 | true, GSI_SAME_STMT); | |
8221c30b | 4594 | if (fd->have_pointer_condtemp) |
6c7ae8c5 JJ |
4595 | expand_omp_build_assign (&gsi2, condtemp, memv, false); |
4596 | if (fd->have_reductemp) | |
4597 | { | |
4598 | gsi_remove (&gsi2, true); | |
4599 | release_ssa_name (gimple_assign_lhs (g)); | |
4600 | } | |
28567c40 | 4601 | } |
629b3d75 MJ |
4602 | switch (gimple_omp_for_kind (fd->for_stmt)) |
4603 | { | |
4604 | case GF_OMP_FOR_KIND_FOR: | |
4605 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
4606 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); | |
4607 | break; | |
4608 | case GF_OMP_FOR_KIND_DISTRIBUTE: | |
4609 | nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS); | |
4610 | threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM); | |
4611 | break; | |
4612 | default: | |
4613 | gcc_unreachable (); | |
4614 | } | |
4615 | nthreads = build_call_expr (nthreads, 0); | |
4616 | nthreads = fold_convert (itype, nthreads); | |
4617 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, | |
4618 | true, GSI_SAME_STMT); | |
4619 | threadid = build_call_expr (threadid, 0); | |
4620 | threadid = fold_convert (itype, threadid); | |
4621 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, | |
4622 | true, GSI_SAME_STMT); | |
4623 | ||
4624 | n1 = fd->loop.n1; | |
4625 | n2 = fd->loop.n2; | |
4626 | step = fd->loop.step; | |
4627 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
4628 | { | |
4629 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
4630 | OMP_CLAUSE__LOOPTEMP_); | |
4631 | gcc_assert (innerc); | |
4632 | n1 = OMP_CLAUSE_DECL (innerc); | |
4633 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4634 | OMP_CLAUSE__LOOPTEMP_); | |
4635 | gcc_assert (innerc); | |
4636 | n2 = OMP_CLAUSE_DECL (innerc); | |
4637 | } | |
4638 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
4639 | true, NULL_TREE, true, GSI_SAME_STMT); | |
4640 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
4641 | true, NULL_TREE, true, GSI_SAME_STMT); | |
4642 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
4643 | true, NULL_TREE, true, GSI_SAME_STMT); | |
4644 | tree chunk_size = fold_convert (itype, fd->chunk_size); | |
4645 | chunk_size = omp_adjust_chunk_size (chunk_size, fd->simd_schedule); | |
4646 | chunk_size | |
4647 | = force_gimple_operand_gsi (&gsi, chunk_size, true, NULL_TREE, true, | |
4648 | GSI_SAME_STMT); | |
4649 | ||
4650 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); | |
4651 | t = fold_build2 (PLUS_EXPR, itype, step, t); | |
4652 | t = fold_build2 (PLUS_EXPR, itype, t, n2); | |
4653 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); | |
4654 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) | |
4655 | t = fold_build2 (TRUNC_DIV_EXPR, itype, | |
4656 | fold_build1 (NEGATE_EXPR, itype, t), | |
4657 | fold_build1 (NEGATE_EXPR, itype, step)); | |
4658 | else | |
4659 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); | |
4660 | t = fold_convert (itype, t); | |
4661 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4662 | true, GSI_SAME_STMT); | |
4663 | ||
4664 | trip_var = create_tmp_reg (itype, ".trip"); | |
4665 | if (gimple_in_ssa_p (cfun)) | |
4666 | { | |
4667 | trip_init = make_ssa_name (trip_var); | |
4668 | trip_main = make_ssa_name (trip_var); | |
4669 | trip_back = make_ssa_name (trip_var); | |
4670 | } | |
4671 | else | |
4672 | { | |
4673 | trip_init = trip_var; | |
4674 | trip_main = trip_var; | |
4675 | trip_back = trip_var; | |
4676 | } | |
4677 | ||
4678 | gassign *assign_stmt | |
4679 | = gimple_build_assign (trip_init, build_int_cst (itype, 0)); | |
4680 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4681 | ||
4682 | t = fold_build2 (MULT_EXPR, itype, threadid, chunk_size); | |
4683 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
4684 | if (POINTER_TYPE_P (type)) | |
4685 | t = fold_build_pointer_plus (n1, t); | |
4686 | else | |
4687 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
4688 | vextra = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4689 | true, GSI_SAME_STMT); | |
4690 | ||
4691 | /* Remove the GIMPLE_OMP_FOR. */ | |
4692 | gsi_remove (&gsi, true); | |
4693 | ||
4694 | gimple_stmt_iterator gsif = gsi; | |
4695 | ||
4696 | /* Iteration space partitioning goes in ITER_PART_BB. */ | |
4697 | gsi = gsi_last_bb (iter_part_bb); | |
4698 | ||
4699 | t = fold_build2 (MULT_EXPR, itype, trip_main, nthreads); | |
4700 | t = fold_build2 (PLUS_EXPR, itype, t, threadid); | |
4701 | t = fold_build2 (MULT_EXPR, itype, t, chunk_size); | |
4702 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4703 | false, GSI_CONTINUE_LINKING); | |
4704 | ||
4705 | t = fold_build2 (PLUS_EXPR, itype, s0, chunk_size); | |
4706 | t = fold_build2 (MIN_EXPR, itype, t, n); | |
4707 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4708 | false, GSI_CONTINUE_LINKING); | |
4709 | ||
4710 | t = build2 (LT_EXPR, boolean_type_node, s0, n); | |
4711 | gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_CONTINUE_LINKING); | |
4712 | ||
4713 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ | |
4714 | gsi = gsi_start_bb (seq_start_bb); | |
4715 | ||
4716 | tree startvar = fd->loop.v; | |
4717 | tree endvar = NULL_TREE; | |
4718 | ||
4719 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4720 | { | |
4721 | tree clauses = gimple_code (inner_stmt) == GIMPLE_OMP_PARALLEL | |
4722 | ? gimple_omp_parallel_clauses (inner_stmt) | |
4723 | : gimple_omp_for_clauses (inner_stmt); | |
4724 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
4725 | gcc_assert (innerc); | |
4726 | startvar = OMP_CLAUSE_DECL (innerc); | |
4727 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4728 | OMP_CLAUSE__LOOPTEMP_); | |
4729 | gcc_assert (innerc); | |
4730 | endvar = OMP_CLAUSE_DECL (innerc); | |
4731 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST | |
4732 | && gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) | |
4733 | { | |
4734 | int i; | |
4735 | for (i = 1; i < fd->collapse; i++) | |
4736 | { | |
4737 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4738 | OMP_CLAUSE__LOOPTEMP_); | |
4739 | gcc_assert (innerc); | |
4740 | } | |
4741 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
4742 | OMP_CLAUSE__LOOPTEMP_); | |
4743 | if (innerc) | |
4744 | { | |
4745 | /* If needed (distribute parallel for with lastprivate), | |
4746 | propagate down the total number of iterations. */ | |
4747 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), | |
4748 | fd->loop.n2); | |
4749 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, | |
4750 | GSI_CONTINUE_LINKING); | |
4751 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
4752 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4753 | } | |
4754 | } | |
4755 | } | |
4756 | ||
4757 | t = fold_convert (itype, s0); | |
4758 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
4759 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
4760 | { |
4761 | t = fold_build_pointer_plus (n1, t); | |
4762 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
4763 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
4764 | t = fold_convert (signed_type_for (type), t); | |
4765 | } | |
629b3d75 MJ |
4766 | else |
4767 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
4768 | t = fold_convert (TREE_TYPE (startvar), t); | |
4769 | t = force_gimple_operand_gsi (&gsi, t, | |
4770 | DECL_P (startvar) | |
4771 | && TREE_ADDRESSABLE (startvar), | |
4772 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4773 | assign_stmt = gimple_build_assign (startvar, t); | |
4774 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
6c7ae8c5 JJ |
4775 | if (cond_var) |
4776 | { | |
4777 | tree itype = TREE_TYPE (cond_var); | |
4778 | /* For lastprivate(conditional:) itervar, we need some iteration | |
4779 | counter that starts at unsigned non-zero and increases. | |
4780 | Prefer as few IVs as possible, so if we can use startvar | |
4781 | itself, use that, or startvar + constant (those would be | |
4782 | incremented with step), and as last resort use the s0 + 1 | |
4783 | incremented by 1. */ | |
4784 | if (POINTER_TYPE_P (type) | |
4785 | || TREE_CODE (n1) != INTEGER_CST | |
4786 | || fd->loop.cond_code != LT_EXPR) | |
4787 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), | |
4788 | build_int_cst (itype, 1)); | |
4789 | else if (tree_int_cst_sgn (n1) == 1) | |
4790 | t = fold_convert (itype, t); | |
4791 | else | |
4792 | { | |
4793 | tree c = fold_convert (itype, n1); | |
4794 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); | |
4795 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); | |
4796 | } | |
4797 | t = force_gimple_operand_gsi (&gsi, t, false, | |
4798 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
4799 | assign_stmt = gimple_build_assign (cond_var, t); | |
4800 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4801 | } | |
629b3d75 MJ |
4802 | |
4803 | t = fold_convert (itype, e0); | |
4804 | t = fold_build2 (MULT_EXPR, itype, t, step); | |
4805 | if (POINTER_TYPE_P (type)) | |
bde84d51 RB |
4806 | { |
4807 | t = fold_build_pointer_plus (n1, t); | |
4808 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) | |
4809 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) | |
4810 | t = fold_convert (signed_type_for (type), t); | |
4811 | } | |
629b3d75 MJ |
4812 | else |
4813 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
4814 | t = fold_convert (TREE_TYPE (startvar), t); | |
4815 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4816 | false, GSI_CONTINUE_LINKING); | |
4817 | if (endvar) | |
4818 | { | |
4819 | assign_stmt = gimple_build_assign (endvar, e); | |
4820 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4821 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
4822 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
4823 | else | |
4824 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
4825 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4826 | } | |
4827 | /* Handle linear clause adjustments. */ | |
4828 | tree itercnt = NULL_TREE, itercntbias = NULL_TREE; | |
4829 | if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_FOR) | |
4830 | for (tree c = gimple_omp_for_clauses (fd->for_stmt); | |
4831 | c; c = OMP_CLAUSE_CHAIN (c)) | |
4832 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR | |
4833 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) | |
4834 | { | |
4835 | tree d = OMP_CLAUSE_DECL (c); | |
4836 | bool is_ref = omp_is_reference (d); | |
4837 | tree t = d, a, dest; | |
4838 | if (is_ref) | |
4839 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); | |
4840 | tree type = TREE_TYPE (t); | |
4841 | if (POINTER_TYPE_P (type)) | |
4842 | type = sizetype; | |
4843 | dest = unshare_expr (t); | |
4844 | tree v = create_tmp_var (TREE_TYPE (t), NULL); | |
4845 | expand_omp_build_assign (&gsif, v, t); | |
4846 | if (itercnt == NULL_TREE) | |
4847 | { | |
4848 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
4849 | { | |
4850 | itercntbias | |
4851 | = fold_build2 (MINUS_EXPR, itype, fold_convert (itype, n1), | |
4852 | fold_convert (itype, fd->loop.n1)); | |
4853 | itercntbias = fold_build2 (EXACT_DIV_EXPR, itype, | |
4854 | itercntbias, step); | |
4855 | itercntbias | |
4856 | = force_gimple_operand_gsi (&gsif, itercntbias, true, | |
4857 | NULL_TREE, true, | |
4858 | GSI_SAME_STMT); | |
4859 | itercnt = fold_build2 (PLUS_EXPR, itype, itercntbias, s0); | |
4860 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, | |
4861 | NULL_TREE, false, | |
4862 | GSI_CONTINUE_LINKING); | |
4863 | } | |
4864 | else | |
4865 | itercnt = s0; | |
4866 | } | |
4867 | a = fold_build2 (MULT_EXPR, type, | |
4868 | fold_convert (type, itercnt), | |
4869 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); | |
4870 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR | |
4871 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); | |
4872 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4873 | false, GSI_CONTINUE_LINKING); | |
4874 | assign_stmt = gimple_build_assign (dest, t); | |
4875 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4876 | } | |
4877 | if (fd->collapse > 1) | |
4878 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); | |
4879 | ||
4880 | if (!broken_loop) | |
4881 | { | |
4882 | /* The code controlling the sequential loop goes in CONT_BB, | |
4883 | replacing the GIMPLE_OMP_CONTINUE. */ | |
65f4b875 | 4884 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
4885 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
4886 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
4887 | vback = gimple_omp_continue_control_def (cont_stmt); | |
4888 | ||
7855700e JJ |
4889 | if (cond_var) |
4890 | { | |
4891 | tree itype = TREE_TYPE (cond_var); | |
4892 | tree t2; | |
4893 | if (POINTER_TYPE_P (type) | |
4894 | || TREE_CODE (n1) != INTEGER_CST | |
4895 | || fd->loop.cond_code != LT_EXPR) | |
4896 | t2 = build_int_cst (itype, 1); | |
4897 | else | |
4898 | t2 = fold_convert (itype, step); | |
4899 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); | |
4900 | t2 = force_gimple_operand_gsi (&gsi, t2, false, | |
4901 | NULL_TREE, true, GSI_SAME_STMT); | |
4902 | assign_stmt = gimple_build_assign (cond_var, t2); | |
4903 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4904 | } | |
4905 | ||
629b3d75 MJ |
4906 | if (!gimple_omp_for_combined_p (fd->for_stmt)) |
4907 | { | |
4908 | if (POINTER_TYPE_P (type)) | |
4909 | t = fold_build_pointer_plus (vmain, step); | |
4910 | else | |
4911 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
4912 | if (DECL_P (vback) && TREE_ADDRESSABLE (vback)) | |
4913 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
4914 | true, GSI_SAME_STMT); | |
4915 | assign_stmt = gimple_build_assign (vback, t); | |
4916 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
4917 | ||
4918 | if (tree_int_cst_equal (fd->chunk_size, integer_one_node)) | |
4919 | t = build2 (EQ_EXPR, boolean_type_node, | |
4920 | build_int_cst (itype, 0), | |
4921 | build_int_cst (itype, 1)); | |
4922 | else | |
4923 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
4924 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
4925 | ? t : vback, e); | |
4926 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
4927 | } | |
4928 | ||
4929 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
4930 | gsi_remove (&gsi, true); | |
4931 | ||
4932 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
4933 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, body_bb); | |
4934 | ||
4935 | /* Trip update code goes into TRIP_UPDATE_BB. */ | |
4936 | gsi = gsi_start_bb (trip_update_bb); | |
4937 | ||
4938 | t = build_int_cst (itype, 1); | |
4939 | t = build2 (PLUS_EXPR, itype, trip_main, t); | |
4940 | assign_stmt = gimple_build_assign (trip_back, t); | |
4941 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
4942 | } | |
4943 | ||
4944 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ | |
65f4b875 | 4945 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
4946 | if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))) |
4947 | { | |
4948 | t = gimple_omp_return_lhs (gsi_stmt (gsi)); | |
8221c30b | 4949 | if (fd->have_reductemp || fd->have_pointer_condtemp) |
28567c40 JJ |
4950 | { |
4951 | tree fn; | |
4952 | if (t) | |
4953 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_CANCEL); | |
4954 | else | |
4955 | fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END); | |
4956 | gcall *g = gimple_build_call (fn, 0); | |
4957 | if (t) | |
4958 | { | |
4959 | gimple_call_set_lhs (g, t); | |
6c7ae8c5 JJ |
4960 | if (fd->have_reductemp) |
4961 | gsi_insert_after (&gsi, gimple_build_assign (reductions, | |
4962 | NOP_EXPR, t), | |
4963 | GSI_SAME_STMT); | |
28567c40 JJ |
4964 | } |
4965 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
4966 | } | |
4967 | else | |
4968 | gsi_insert_after (&gsi, omp_build_barrier (t), GSI_SAME_STMT); | |
629b3d75 | 4969 | } |
1a39b3d3 JJ |
4970 | else if (fd->have_pointer_condtemp) |
4971 | { | |
4972 | tree fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT); | |
4973 | gcall *g = gimple_build_call (fn, 0); | |
4974 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); | |
4975 | } | |
629b3d75 MJ |
4976 | gsi_remove (&gsi, true); |
4977 | ||
4978 | /* Connect the new blocks. */ | |
4979 | find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE; | |
4980 | find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE; | |
4981 | ||
4982 | if (!broken_loop) | |
4983 | { | |
4984 | se = find_edge (cont_bb, body_bb); | |
4985 | if (se == NULL) | |
4986 | { | |
4987 | se = BRANCH_EDGE (cont_bb); | |
4988 | gcc_assert (single_succ (se->dest) == body_bb); | |
4989 | } | |
4990 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
4991 | { | |
4992 | remove_edge (se); | |
4993 | se = NULL; | |
4994 | } | |
4995 | else if (fd->collapse > 1) | |
4996 | { | |
4997 | remove_edge (se); | |
4998 | se = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
4999 | } | |
5000 | else | |
5001 | se->flags = EDGE_TRUE_VALUE; | |
5002 | find_edge (cont_bb, trip_update_bb)->flags | |
5003 | = se ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
5004 | ||
01914336 MJ |
5005 | redirect_edge_and_branch (single_succ_edge (trip_update_bb), |
5006 | iter_part_bb); | |
629b3d75 MJ |
5007 | } |
5008 | ||
5009 | if (gimple_in_ssa_p (cfun)) | |
5010 | { | |
5011 | gphi_iterator psi; | |
5012 | gphi *phi; | |
5013 | edge re, ene; | |
5014 | edge_var_map *vm; | |
5015 | size_t i; | |
5016 | ||
5017 | gcc_assert (fd->collapse == 1 && !broken_loop); | |
5018 | ||
5019 | /* When we redirect the edge from trip_update_bb to iter_part_bb, we | |
5020 | remove arguments of the phi nodes in fin_bb. We need to create | |
5021 | appropriate phi nodes in iter_part_bb instead. */ | |
5022 | se = find_edge (iter_part_bb, fin_bb); | |
5023 | re = single_succ_edge (trip_update_bb); | |
5024 | vec<edge_var_map> *head = redirect_edge_var_map_vector (re); | |
5025 | ene = single_succ_edge (entry_bb); | |
5026 | ||
5027 | psi = gsi_start_phis (fin_bb); | |
5028 | for (i = 0; !gsi_end_p (psi) && head->iterate (i, &vm); | |
5029 | gsi_next (&psi), ++i) | |
5030 | { | |
5031 | gphi *nphi; | |
620e594b | 5032 | location_t locus; |
629b3d75 MJ |
5033 | |
5034 | phi = psi.phi (); | |
d83cc5cc TV |
5035 | if (operand_equal_p (gimple_phi_arg_def (phi, 0), |
5036 | redirect_edge_var_map_def (vm), 0)) | |
5037 | continue; | |
5038 | ||
629b3d75 MJ |
5039 | t = gimple_phi_result (phi); |
5040 | gcc_assert (t == redirect_edge_var_map_result (vm)); | |
5041 | ||
5042 | if (!single_pred_p (fin_bb)) | |
5043 | t = copy_ssa_name (t, phi); | |
5044 | ||
5045 | nphi = create_phi_node (t, iter_part_bb); | |
5046 | ||
5047 | t = PHI_ARG_DEF_FROM_EDGE (phi, se); | |
5048 | locus = gimple_phi_arg_location_from_edge (phi, se); | |
5049 | ||
5050 | /* A special case -- fd->loop.v is not yet computed in | |
5051 | iter_part_bb, we need to use vextra instead. */ | |
5052 | if (t == fd->loop.v) | |
5053 | t = vextra; | |
5054 | add_phi_arg (nphi, t, ene, locus); | |
5055 | locus = redirect_edge_var_map_location (vm); | |
5056 | tree back_arg = redirect_edge_var_map_def (vm); | |
5057 | add_phi_arg (nphi, back_arg, re, locus); | |
5058 | edge ce = find_edge (cont_bb, body_bb); | |
5059 | if (ce == NULL) | |
5060 | { | |
5061 | ce = BRANCH_EDGE (cont_bb); | |
5062 | gcc_assert (single_succ (ce->dest) == body_bb); | |
5063 | ce = single_succ_edge (ce->dest); | |
5064 | } | |
5065 | gphi *inner_loop_phi = find_phi_with_arg_on_edge (back_arg, ce); | |
5066 | gcc_assert (inner_loop_phi != NULL); | |
5067 | add_phi_arg (inner_loop_phi, gimple_phi_result (nphi), | |
5068 | find_edge (seq_start_bb, body_bb), locus); | |
5069 | ||
5070 | if (!single_pred_p (fin_bb)) | |
5071 | add_phi_arg (phi, gimple_phi_result (nphi), se, locus); | |
5072 | } | |
5073 | gcc_assert (gsi_end_p (psi) && (head == NULL || i == head->length ())); | |
5074 | redirect_edge_var_map_clear (re); | |
5075 | if (single_pred_p (fin_bb)) | |
5076 | while (1) | |
5077 | { | |
5078 | psi = gsi_start_phis (fin_bb); | |
5079 | if (gsi_end_p (psi)) | |
5080 | break; | |
5081 | remove_phi_node (&psi, false); | |
5082 | } | |
5083 | ||
5084 | /* Make phi node for trip. */ | |
5085 | phi = create_phi_node (trip_main, iter_part_bb); | |
5086 | add_phi_arg (phi, trip_back, single_succ_edge (trip_update_bb), | |
5087 | UNKNOWN_LOCATION); | |
5088 | add_phi_arg (phi, trip_init, single_succ_edge (entry_bb), | |
5089 | UNKNOWN_LOCATION); | |
5090 | } | |
5091 | ||
5092 | if (!broken_loop) | |
5093 | set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb); | |
5094 | set_immediate_dominator (CDI_DOMINATORS, iter_part_bb, | |
5095 | recompute_dominator (CDI_DOMINATORS, iter_part_bb)); | |
5096 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
5097 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
5098 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, | |
5099 | recompute_dominator (CDI_DOMINATORS, seq_start_bb)); | |
5100 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
5101 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
5102 | ||
5103 | if (!broken_loop) | |
5104 | { | |
99b1c316 MS |
5105 | class loop *loop = body_bb->loop_father; |
5106 | class loop *trip_loop = alloc_loop (); | |
629b3d75 MJ |
5107 | trip_loop->header = iter_part_bb; |
5108 | trip_loop->latch = trip_update_bb; | |
5109 | add_loop (trip_loop, iter_part_bb->loop_father); | |
5110 | ||
5111 | if (loop != entry_bb->loop_father) | |
5112 | { | |
5113 | gcc_assert (loop->header == body_bb); | |
5114 | gcc_assert (loop->latch == region->cont | |
5115 | || single_pred (loop->latch) == region->cont); | |
5116 | trip_loop->inner = loop; | |
5117 | return; | |
5118 | } | |
5119 | ||
5120 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
5121 | { | |
5122 | loop = alloc_loop (); | |
5123 | loop->header = body_bb; | |
5124 | if (collapse_bb == NULL) | |
5125 | loop->latch = cont_bb; | |
5126 | add_loop (loop, trip_loop); | |
5127 | } | |
5128 | } | |
5129 | } | |
5130 | ||
629b3d75 MJ |
5131 | /* A subroutine of expand_omp_for. Generate code for a simd non-worksharing |
5132 | loop. Given parameters: | |
5133 | ||
5134 | for (V = N1; V cond N2; V += STEP) BODY; | |
5135 | ||
5136 | where COND is "<" or ">", we generate pseudocode | |
5137 | ||
5138 | V = N1; | |
5139 | goto L1; | |
5140 | L0: | |
5141 | BODY; | |
5142 | V += STEP; | |
5143 | L1: | |
5144 | if (V cond N2) goto L0; else goto L2; | |
5145 | L2: | |
5146 | ||
5147 | For collapsed loops, given parameters: | |
5148 | collapse(3) | |
5149 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) | |
5150 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) | |
5151 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) | |
5152 | BODY; | |
5153 | ||
5154 | we generate pseudocode | |
5155 | ||
5156 | if (cond3 is <) | |
5157 | adj = STEP3 - 1; | |
5158 | else | |
5159 | adj = STEP3 + 1; | |
5160 | count3 = (adj + N32 - N31) / STEP3; | |
5161 | if (cond2 is <) | |
5162 | adj = STEP2 - 1; | |
5163 | else | |
5164 | adj = STEP2 + 1; | |
5165 | count2 = (adj + N22 - N21) / STEP2; | |
5166 | if (cond1 is <) | |
5167 | adj = STEP1 - 1; | |
5168 | else | |
5169 | adj = STEP1 + 1; | |
5170 | count1 = (adj + N12 - N11) / STEP1; | |
5171 | count = count1 * count2 * count3; | |
5172 | V = 0; | |
5173 | V1 = N11; | |
5174 | V2 = N21; | |
5175 | V3 = N31; | |
5176 | goto L1; | |
5177 | L0: | |
5178 | BODY; | |
5179 | V += 1; | |
5180 | V3 += STEP3; | |
5181 | V2 += (V3 cond3 N32) ? 0 : STEP2; | |
5182 | V3 = (V3 cond3 N32) ? V3 : N31; | |
5183 | V1 += (V2 cond2 N22) ? 0 : STEP1; | |
5184 | V2 = (V2 cond2 N22) ? V2 : N21; | |
5185 | L1: | |
5186 | if (V < count) goto L0; else goto L2; | |
5187 | L2: | |
5188 | ||
5189 | */ | |
5190 | ||
5191 | static void | |
5192 | expand_omp_simd (struct omp_region *region, struct omp_for_data *fd) | |
5193 | { | |
5194 | tree type, t; | |
5195 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, l2_bb, l2_dom_bb; | |
5196 | gimple_stmt_iterator gsi; | |
5197 | gimple *stmt; | |
5198 | gcond *cond_stmt; | |
5199 | bool broken_loop = region->cont == NULL; | |
5200 | edge e, ne; | |
5201 | tree *counts = NULL; | |
5202 | int i; | |
5203 | int safelen_int = INT_MAX; | |
fed2a43c | 5204 | bool dont_vectorize = false; |
629b3d75 MJ |
5205 | tree safelen = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
5206 | OMP_CLAUSE_SAFELEN); | |
5207 | tree simduid = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5208 | OMP_CLAUSE__SIMDUID_); | |
fed2a43c JJ |
5209 | tree ifc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
5210 | OMP_CLAUSE_IF); | |
5211 | tree simdlen = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5212 | OMP_CLAUSE_SIMDLEN); | |
e7393c89 JJ |
5213 | tree condtemp = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
5214 | OMP_CLAUSE__CONDTEMP_); | |
629b3d75 | 5215 | tree n1, n2; |
e7393c89 | 5216 | tree cond_var = condtemp ? OMP_CLAUSE_DECL (condtemp) : NULL_TREE; |
629b3d75 MJ |
5217 | |
5218 | if (safelen) | |
5219 | { | |
9d2f08ab | 5220 | poly_uint64 val; |
629b3d75 | 5221 | safelen = OMP_CLAUSE_SAFELEN_EXPR (safelen); |
9d2f08ab | 5222 | if (!poly_int_tree_p (safelen, &val)) |
629b3d75 | 5223 | safelen_int = 0; |
9d2f08ab RS |
5224 | else |
5225 | safelen_int = MIN (constant_lower_bound (val), INT_MAX); | |
629b3d75 MJ |
5226 | if (safelen_int == 1) |
5227 | safelen_int = 0; | |
5228 | } | |
fed2a43c JJ |
5229 | if ((ifc && integer_zerop (OMP_CLAUSE_IF_EXPR (ifc))) |
5230 | || (simdlen && integer_onep (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)))) | |
5231 | { | |
5232 | safelen_int = 0; | |
5233 | dont_vectorize = true; | |
5234 | } | |
629b3d75 MJ |
5235 | type = TREE_TYPE (fd->loop.v); |
5236 | entry_bb = region->entry; | |
5237 | cont_bb = region->cont; | |
5238 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
5239 | gcc_assert (broken_loop | |
5240 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
5241 | l0_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
5242 | if (!broken_loop) | |
5243 | { | |
5244 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l0_bb); | |
5245 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
5246 | l1_bb = split_block (cont_bb, last_stmt (cont_bb))->dest; | |
5247 | l2_bb = BRANCH_EDGE (entry_bb)->dest; | |
5248 | } | |
5249 | else | |
5250 | { | |
5251 | BRANCH_EDGE (entry_bb)->flags &= ~EDGE_ABNORMAL; | |
5252 | l1_bb = split_edge (BRANCH_EDGE (entry_bb)); | |
5253 | l2_bb = single_succ (l1_bb); | |
5254 | } | |
5255 | exit_bb = region->exit; | |
5256 | l2_dom_bb = NULL; | |
5257 | ||
65f4b875 | 5258 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
5259 | |
5260 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); | |
5261 | /* Not needed in SSA form right now. */ | |
5262 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
5263 | if (fd->collapse > 1) | |
5264 | { | |
5265 | int first_zero_iter = -1, dummy = -1; | |
5266 | basic_block zero_iter_bb = l2_bb, dummy_bb = NULL; | |
5267 | ||
5268 | counts = XALLOCAVEC (tree, fd->collapse); | |
5269 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5270 | zero_iter_bb, first_zero_iter, | |
5271 | dummy_bb, dummy, l2_dom_bb); | |
5272 | } | |
5273 | if (l2_dom_bb == NULL) | |
5274 | l2_dom_bb = l1_bb; | |
5275 | ||
5276 | n1 = fd->loop.n1; | |
5277 | n2 = fd->loop.n2; | |
5278 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5279 | { | |
5280 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5281 | OMP_CLAUSE__LOOPTEMP_); | |
5282 | gcc_assert (innerc); | |
5283 | n1 = OMP_CLAUSE_DECL (innerc); | |
5284 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5285 | OMP_CLAUSE__LOOPTEMP_); | |
5286 | gcc_assert (innerc); | |
5287 | n2 = OMP_CLAUSE_DECL (innerc); | |
5288 | } | |
5289 | tree step = fd->loop.step; | |
5290 | ||
4cea8675 AM |
5291 | bool is_simt = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), |
5292 | OMP_CLAUSE__SIMT_); | |
629b3d75 MJ |
5293 | if (is_simt) |
5294 | { | |
5295 | cfun->curr_properties &= ~PROP_gimple_lomp_dev; | |
4cea8675 AM |
5296 | is_simt = safelen_int > 1; |
5297 | } | |
5298 | tree simt_lane = NULL_TREE, simt_maxlane = NULL_TREE; | |
5299 | if (is_simt) | |
5300 | { | |
629b3d75 MJ |
5301 | simt_lane = create_tmp_var (unsigned_type_node); |
5302 | gimple *g = gimple_build_call_internal (IFN_GOMP_SIMT_LANE, 0); | |
5303 | gimple_call_set_lhs (g, simt_lane); | |
5304 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5305 | tree offset = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, | |
5306 | fold_convert (TREE_TYPE (step), simt_lane)); | |
5307 | n1 = fold_convert (type, n1); | |
5308 | if (POINTER_TYPE_P (type)) | |
5309 | n1 = fold_build_pointer_plus (n1, offset); | |
5310 | else | |
5311 | n1 = fold_build2 (PLUS_EXPR, type, n1, fold_convert (type, offset)); | |
5312 | ||
5313 | /* Collapsed loops not handled for SIMT yet: limit to one lane only. */ | |
5314 | if (fd->collapse > 1) | |
5315 | simt_maxlane = build_one_cst (unsigned_type_node); | |
5316 | else if (safelen_int < omp_max_simt_vf ()) | |
5317 | simt_maxlane = build_int_cst (unsigned_type_node, safelen_int); | |
5318 | tree vf | |
5319 | = build_call_expr_internal_loc (UNKNOWN_LOCATION, IFN_GOMP_SIMT_VF, | |
5320 | unsigned_type_node, 0); | |
5321 | if (simt_maxlane) | |
5322 | vf = fold_build2 (MIN_EXPR, unsigned_type_node, vf, simt_maxlane); | |
5323 | vf = fold_convert (TREE_TYPE (step), vf); | |
5324 | step = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, vf); | |
5325 | } | |
5326 | ||
5327 | expand_omp_build_assign (&gsi, fd->loop.v, fold_convert (type, n1)); | |
5328 | if (fd->collapse > 1) | |
5329 | { | |
5330 | if (gimple_omp_for_combined_into_p (fd->for_stmt)) | |
5331 | { | |
5332 | gsi_prev (&gsi); | |
5333 | expand_omp_for_init_vars (fd, &gsi, counts, NULL, n1); | |
5334 | gsi_next (&gsi); | |
5335 | } | |
5336 | else | |
5337 | for (i = 0; i < fd->collapse; i++) | |
5338 | { | |
5339 | tree itype = TREE_TYPE (fd->loops[i].v); | |
5340 | if (POINTER_TYPE_P (itype)) | |
5341 | itype = signed_type_for (itype); | |
5342 | t = fold_convert (TREE_TYPE (fd->loops[i].v), fd->loops[i].n1); | |
5343 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
5344 | } | |
5345 | } | |
e7393c89 JJ |
5346 | if (cond_var) |
5347 | { | |
5348 | if (POINTER_TYPE_P (type) | |
5349 | || TREE_CODE (n1) != INTEGER_CST | |
5350 | || fd->loop.cond_code != LT_EXPR | |
5351 | || tree_int_cst_sgn (n1) != 1) | |
5352 | expand_omp_build_assign (&gsi, cond_var, | |
5353 | build_one_cst (TREE_TYPE (cond_var))); | |
5354 | else | |
5355 | expand_omp_build_assign (&gsi, cond_var, | |
5356 | fold_convert (TREE_TYPE (cond_var), n1)); | |
5357 | } | |
629b3d75 MJ |
5358 | |
5359 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
5360 | gsi_remove (&gsi, true); | |
5361 | ||
5362 | if (!broken_loop) | |
5363 | { | |
5364 | /* Code to control the increment goes in the CONT_BB. */ | |
65f4b875 | 5365 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
5366 | stmt = gsi_stmt (gsi); |
5367 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); | |
5368 | ||
5369 | if (POINTER_TYPE_P (type)) | |
5370 | t = fold_build_pointer_plus (fd->loop.v, step); | |
5371 | else | |
5372 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); | |
5373 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
5374 | ||
5375 | if (fd->collapse > 1) | |
5376 | { | |
5377 | i = fd->collapse - 1; | |
5378 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) | |
5379 | { | |
5380 | t = fold_convert (sizetype, fd->loops[i].step); | |
5381 | t = fold_build_pointer_plus (fd->loops[i].v, t); | |
5382 | } | |
5383 | else | |
5384 | { | |
5385 | t = fold_convert (TREE_TYPE (fd->loops[i].v), | |
5386 | fd->loops[i].step); | |
5387 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), | |
5388 | fd->loops[i].v, t); | |
5389 | } | |
5390 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); | |
5391 | ||
5392 | for (i = fd->collapse - 1; i > 0; i--) | |
5393 | { | |
5394 | tree itype = TREE_TYPE (fd->loops[i].v); | |
5395 | tree itype2 = TREE_TYPE (fd->loops[i - 1].v); | |
5396 | if (POINTER_TYPE_P (itype2)) | |
5397 | itype2 = signed_type_for (itype2); | |
bcc6842b JJ |
5398 | t = fold_convert (itype2, fd->loops[i - 1].step); |
5399 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, | |
5400 | GSI_SAME_STMT); | |
629b3d75 MJ |
5401 | t = build3 (COND_EXPR, itype2, |
5402 | build2 (fd->loops[i].cond_code, boolean_type_node, | |
5403 | fd->loops[i].v, | |
5404 | fold_convert (itype, fd->loops[i].n2)), | |
bcc6842b | 5405 | build_int_cst (itype2, 0), t); |
629b3d75 MJ |
5406 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i - 1].v))) |
5407 | t = fold_build_pointer_plus (fd->loops[i - 1].v, t); | |
5408 | else | |
5409 | t = fold_build2 (PLUS_EXPR, itype2, fd->loops[i - 1].v, t); | |
5410 | expand_omp_build_assign (&gsi, fd->loops[i - 1].v, t); | |
5411 | ||
bcc6842b JJ |
5412 | t = fold_convert (itype, fd->loops[i].n1); |
5413 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, | |
5414 | GSI_SAME_STMT); | |
629b3d75 MJ |
5415 | t = build3 (COND_EXPR, itype, |
5416 | build2 (fd->loops[i].cond_code, boolean_type_node, | |
5417 | fd->loops[i].v, | |
5418 | fold_convert (itype, fd->loops[i].n2)), | |
bcc6842b | 5419 | fd->loops[i].v, t); |
629b3d75 MJ |
5420 | expand_omp_build_assign (&gsi, fd->loops[i].v, t); |
5421 | } | |
5422 | } | |
e7393c89 JJ |
5423 | if (cond_var) |
5424 | { | |
5425 | if (POINTER_TYPE_P (type) | |
5426 | || TREE_CODE (n1) != INTEGER_CST | |
5427 | || fd->loop.cond_code != LT_EXPR | |
5428 | || tree_int_cst_sgn (n1) != 1) | |
5429 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, | |
5430 | build_one_cst (TREE_TYPE (cond_var))); | |
5431 | else | |
5432 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, | |
5433 | fold_convert (TREE_TYPE (cond_var), step)); | |
5434 | expand_omp_build_assign (&gsi, cond_var, t); | |
5435 | } | |
629b3d75 MJ |
5436 | |
5437 | /* Remove GIMPLE_OMP_CONTINUE. */ | |
5438 | gsi_remove (&gsi, true); | |
5439 | } | |
5440 | ||
5441 | /* Emit the condition in L1_BB. */ | |
5442 | gsi = gsi_start_bb (l1_bb); | |
5443 | ||
5444 | t = fold_convert (type, n2); | |
5445 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5446 | false, GSI_CONTINUE_LINKING); | |
5447 | tree v = fd->loop.v; | |
5448 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) | |
5449 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, | |
5450 | false, GSI_CONTINUE_LINKING); | |
5451 | t = build2 (fd->loop.cond_code, boolean_type_node, v, t); | |
5452 | cond_stmt = gimple_build_cond_empty (t); | |
5453 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); | |
5454 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
5455 | NULL, NULL) | |
5456 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), expand_omp_regimplify_p, | |
5457 | NULL, NULL)) | |
5458 | { | |
5459 | gsi = gsi_for_stmt (cond_stmt); | |
5460 | gimple_regimplify_operands (cond_stmt, &gsi); | |
5461 | } | |
5462 | ||
5463 | /* Add 'V -= STEP * (SIMT_VF - 1)' after the loop. */ | |
5464 | if (is_simt) | |
5465 | { | |
5466 | gsi = gsi_start_bb (l2_bb); | |
5467 | step = fold_build2 (MINUS_EXPR, TREE_TYPE (step), fd->loop.step, step); | |
5468 | if (POINTER_TYPE_P (type)) | |
5469 | t = fold_build_pointer_plus (fd->loop.v, step); | |
5470 | else | |
5471 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); | |
5472 | expand_omp_build_assign (&gsi, fd->loop.v, t); | |
5473 | } | |
5474 | ||
5475 | /* Remove GIMPLE_OMP_RETURN. */ | |
65f4b875 | 5476 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
5477 | gsi_remove (&gsi, true); |
5478 | ||
5479 | /* Connect the new blocks. */ | |
5480 | remove_edge (FALLTHRU_EDGE (entry_bb)); | |
5481 | ||
5482 | if (!broken_loop) | |
5483 | { | |
5484 | remove_edge (BRANCH_EDGE (entry_bb)); | |
5485 | make_edge (entry_bb, l1_bb, EDGE_FALLTHRU); | |
5486 | ||
5487 | e = BRANCH_EDGE (l1_bb); | |
5488 | ne = FALLTHRU_EDGE (l1_bb); | |
5489 | e->flags = EDGE_TRUE_VALUE; | |
5490 | } | |
5491 | else | |
5492 | { | |
5493 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
5494 | ||
5495 | ne = single_succ_edge (l1_bb); | |
5496 | e = make_edge (l1_bb, l0_bb, EDGE_TRUE_VALUE); | |
5497 | ||
5498 | } | |
5499 | ne->flags = EDGE_FALSE_VALUE; | |
357067f2 JH |
5500 | e->probability = profile_probability::guessed_always ().apply_scale (7, 8); |
5501 | ne->probability = e->probability.invert (); | |
629b3d75 MJ |
5502 | |
5503 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, entry_bb); | |
5504 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, l1_bb); | |
5505 | ||
5506 | if (simt_maxlane) | |
5507 | { | |
5508 | cond_stmt = gimple_build_cond (LT_EXPR, simt_lane, simt_maxlane, | |
5509 | NULL_TREE, NULL_TREE); | |
5510 | gsi = gsi_last_bb (entry_bb); | |
5511 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
5512 | make_edge (entry_bb, l2_bb, EDGE_FALSE_VALUE); | |
5513 | FALLTHRU_EDGE (entry_bb)->flags = EDGE_TRUE_VALUE; | |
357067f2 JH |
5514 | FALLTHRU_EDGE (entry_bb)->probability |
5515 | = profile_probability::guessed_always ().apply_scale (7, 8); | |
5516 | BRANCH_EDGE (entry_bb)->probability | |
5517 | = FALLTHRU_EDGE (entry_bb)->probability.invert (); | |
629b3d75 MJ |
5518 | l2_dom_bb = entry_bb; |
5519 | } | |
5520 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, l2_dom_bb); | |
5521 | ||
5522 | if (!broken_loop) | |
5523 | { | |
99b1c316 | 5524 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
5525 | loop->header = l1_bb; |
5526 | loop->latch = cont_bb; | |
5527 | add_loop (loop, l1_bb->loop_father); | |
5528 | loop->safelen = safelen_int; | |
5529 | if (simduid) | |
5530 | { | |
5531 | loop->simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); | |
5532 | cfun->has_simduid_loops = true; | |
5533 | } | |
5534 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize | |
5535 | the loop. */ | |
5536 | if ((flag_tree_loop_vectorize | |
26d476cd | 5537 | || !global_options_set.x_flag_tree_loop_vectorize) |
629b3d75 MJ |
5538 | && flag_tree_loop_optimize |
5539 | && loop->safelen > 1) | |
5540 | { | |
5541 | loop->force_vectorize = true; | |
f63445e5 JJ |
5542 | if (simdlen && tree_fits_uhwi_p (OMP_CLAUSE_SIMDLEN_EXPR (simdlen))) |
5543 | { | |
5544 | unsigned HOST_WIDE_INT v | |
5545 | = tree_to_uhwi (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)); | |
5546 | if (v < INT_MAX && v <= (unsigned HOST_WIDE_INT) loop->safelen) | |
5547 | loop->simdlen = v; | |
5548 | } | |
629b3d75 MJ |
5549 | cfun->has_force_vectorize_loops = true; |
5550 | } | |
fed2a43c JJ |
5551 | else if (dont_vectorize) |
5552 | loop->dont_vectorize = true; | |
629b3d75 MJ |
5553 | } |
5554 | else if (simduid) | |
5555 | cfun->has_simduid_loops = true; | |
5556 | } | |
5557 | ||
5558 | /* Taskloop construct is represented after gimplification with | |
5559 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched | |
5560 | in between them. This routine expands the outer GIMPLE_OMP_FOR, | |
5561 | which should just compute all the needed loop temporaries | |
5562 | for GIMPLE_OMP_TASK. */ | |
5563 | ||
5564 | static void | |
5565 | expand_omp_taskloop_for_outer (struct omp_region *region, | |
5566 | struct omp_for_data *fd, | |
5567 | gimple *inner_stmt) | |
5568 | { | |
5569 | tree type, bias = NULL_TREE; | |
5570 | basic_block entry_bb, cont_bb, exit_bb; | |
5571 | gimple_stmt_iterator gsi; | |
5572 | gassign *assign_stmt; | |
5573 | tree *counts = NULL; | |
5574 | int i; | |
5575 | ||
5576 | gcc_assert (inner_stmt); | |
5577 | gcc_assert (region->cont); | |
5578 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_TASK | |
5579 | && gimple_omp_task_taskloop_p (inner_stmt)); | |
5580 | type = TREE_TYPE (fd->loop.v); | |
5581 | ||
5582 | /* See if we need to bias by LLONG_MIN. */ | |
5583 | if (fd->iter_type == long_long_unsigned_type_node | |
5584 | && TREE_CODE (type) == INTEGER_TYPE | |
5585 | && !TYPE_UNSIGNED (type)) | |
5586 | { | |
5587 | tree n1, n2; | |
5588 | ||
5589 | if (fd->loop.cond_code == LT_EXPR) | |
5590 | { | |
5591 | n1 = fd->loop.n1; | |
5592 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5593 | } | |
5594 | else | |
5595 | { | |
5596 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5597 | n2 = fd->loop.n1; | |
5598 | } | |
5599 | if (TREE_CODE (n1) != INTEGER_CST | |
5600 | || TREE_CODE (n2) != INTEGER_CST | |
5601 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
5602 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
5603 | } | |
5604 | ||
5605 | entry_bb = region->entry; | |
5606 | cont_bb = region->cont; | |
5607 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
5608 | gcc_assert (BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); | |
5609 | exit_bb = region->exit; | |
5610 | ||
65f4b875 | 5611 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
5612 | gimple *for_stmt = gsi_stmt (gsi); |
5613 | gcc_assert (gimple_code (for_stmt) == GIMPLE_OMP_FOR); | |
5614 | if (fd->collapse > 1) | |
5615 | { | |
5616 | int first_zero_iter = -1, dummy = -1; | |
5617 | basic_block zero_iter_bb = NULL, dummy_bb = NULL, l2_dom_bb = NULL; | |
5618 | ||
5619 | counts = XALLOCAVEC (tree, fd->collapse); | |
5620 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5621 | zero_iter_bb, first_zero_iter, | |
5622 | dummy_bb, dummy, l2_dom_bb); | |
5623 | ||
5624 | if (zero_iter_bb) | |
5625 | { | |
5626 | /* Some counts[i] vars might be uninitialized if | |
5627 | some loop has zero iterations. But the body shouldn't | |
5628 | be executed in that case, so just avoid uninit warnings. */ | |
5629 | for (i = first_zero_iter; i < fd->collapse; i++) | |
5630 | if (SSA_VAR_P (counts[i])) | |
5631 | TREE_NO_WARNING (counts[i]) = 1; | |
5632 | gsi_prev (&gsi); | |
5633 | edge e = split_block (entry_bb, gsi_stmt (gsi)); | |
5634 | entry_bb = e->dest; | |
5635 | make_edge (zero_iter_bb, entry_bb, EDGE_FALLTHRU); | |
5636 | gsi = gsi_last_bb (entry_bb); | |
5637 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, | |
5638 | get_immediate_dominator (CDI_DOMINATORS, | |
5639 | zero_iter_bb)); | |
5640 | } | |
5641 | } | |
5642 | ||
5643 | tree t0, t1; | |
5644 | t1 = fd->loop.n2; | |
5645 | t0 = fd->loop.n1; | |
5646 | if (POINTER_TYPE_P (TREE_TYPE (t0)) | |
5647 | && TYPE_PRECISION (TREE_TYPE (t0)) | |
5648 | != TYPE_PRECISION (fd->iter_type)) | |
5649 | { | |
5650 | /* Avoid casting pointers to integer of a different size. */ | |
5651 | tree itype = signed_type_for (type); | |
5652 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); | |
5653 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); | |
5654 | } | |
5655 | else | |
5656 | { | |
5657 | t1 = fold_convert (fd->iter_type, t1); | |
5658 | t0 = fold_convert (fd->iter_type, t0); | |
5659 | } | |
5660 | if (bias) | |
5661 | { | |
5662 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); | |
5663 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); | |
5664 | } | |
5665 | ||
5666 | tree innerc = omp_find_clause (gimple_omp_task_clauses (inner_stmt), | |
5667 | OMP_CLAUSE__LOOPTEMP_); | |
5668 | gcc_assert (innerc); | |
5669 | tree startvar = OMP_CLAUSE_DECL (innerc); | |
5670 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), OMP_CLAUSE__LOOPTEMP_); | |
5671 | gcc_assert (innerc); | |
5672 | tree endvar = OMP_CLAUSE_DECL (innerc); | |
5673 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST) | |
5674 | { | |
5675 | gcc_assert (innerc); | |
5676 | for (i = 1; i < fd->collapse; i++) | |
5677 | { | |
5678 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5679 | OMP_CLAUSE__LOOPTEMP_); | |
5680 | gcc_assert (innerc); | |
5681 | } | |
5682 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5683 | OMP_CLAUSE__LOOPTEMP_); | |
5684 | if (innerc) | |
5685 | { | |
5686 | /* If needed (inner taskloop has lastprivate clause), propagate | |
5687 | down the total number of iterations. */ | |
5688 | tree t = force_gimple_operand_gsi (&gsi, fd->loop.n2, false, | |
5689 | NULL_TREE, false, | |
5690 | GSI_CONTINUE_LINKING); | |
5691 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); | |
5692 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5693 | } | |
5694 | } | |
5695 | ||
5696 | t0 = force_gimple_operand_gsi (&gsi, t0, false, NULL_TREE, false, | |
5697 | GSI_CONTINUE_LINKING); | |
5698 | assign_stmt = gimple_build_assign (startvar, t0); | |
5699 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5700 | ||
5701 | t1 = force_gimple_operand_gsi (&gsi, t1, false, NULL_TREE, false, | |
5702 | GSI_CONTINUE_LINKING); | |
5703 | assign_stmt = gimple_build_assign (endvar, t1); | |
5704 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5705 | if (fd->collapse > 1) | |
5706 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); | |
5707 | ||
5708 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
5709 | gsi = gsi_for_stmt (for_stmt); | |
5710 | gsi_remove (&gsi, true); | |
5711 | ||
65f4b875 | 5712 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
5713 | gsi_remove (&gsi, true); |
5714 | ||
65f4b875 | 5715 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
5716 | gsi_remove (&gsi, true); |
5717 | ||
357067f2 | 5718 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); |
629b3d75 | 5719 | remove_edge (BRANCH_EDGE (entry_bb)); |
357067f2 | 5720 | FALLTHRU_EDGE (cont_bb)->probability = profile_probability::always (); |
629b3d75 MJ |
5721 | remove_edge (BRANCH_EDGE (cont_bb)); |
5722 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, cont_bb); | |
5723 | set_immediate_dominator (CDI_DOMINATORS, region->entry, | |
5724 | recompute_dominator (CDI_DOMINATORS, region->entry)); | |
5725 | } | |
5726 | ||
5727 | /* Taskloop construct is represented after gimplification with | |
5728 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched | |
5729 | in between them. This routine expands the inner GIMPLE_OMP_FOR. | |
5730 | GOMP_taskloop{,_ull} function arranges for each task to be given just | |
5731 | a single range of iterations. */ | |
5732 | ||
5733 | static void | |
5734 | expand_omp_taskloop_for_inner (struct omp_region *region, | |
5735 | struct omp_for_data *fd, | |
5736 | gimple *inner_stmt) | |
5737 | { | |
5738 | tree e, t, type, itype, vmain, vback, bias = NULL_TREE; | |
5739 | basic_block entry_bb, exit_bb, body_bb, cont_bb, collapse_bb = NULL; | |
5740 | basic_block fin_bb; | |
5741 | gimple_stmt_iterator gsi; | |
5742 | edge ep; | |
5743 | bool broken_loop = region->cont == NULL; | |
5744 | tree *counts = NULL; | |
5745 | tree n1, n2, step; | |
5746 | ||
5747 | itype = type = TREE_TYPE (fd->loop.v); | |
5748 | if (POINTER_TYPE_P (type)) | |
5749 | itype = signed_type_for (type); | |
5750 | ||
5751 | /* See if we need to bias by LLONG_MIN. */ | |
5752 | if (fd->iter_type == long_long_unsigned_type_node | |
5753 | && TREE_CODE (type) == INTEGER_TYPE | |
5754 | && !TYPE_UNSIGNED (type)) | |
5755 | { | |
5756 | tree n1, n2; | |
5757 | ||
5758 | if (fd->loop.cond_code == LT_EXPR) | |
5759 | { | |
5760 | n1 = fd->loop.n1; | |
5761 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5762 | } | |
5763 | else | |
5764 | { | |
5765 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); | |
5766 | n2 = fd->loop.n1; | |
5767 | } | |
5768 | if (TREE_CODE (n1) != INTEGER_CST | |
5769 | || TREE_CODE (n2) != INTEGER_CST | |
5770 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) | |
5771 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); | |
5772 | } | |
5773 | ||
5774 | entry_bb = region->entry; | |
5775 | cont_bb = region->cont; | |
5776 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); | |
5777 | fin_bb = BRANCH_EDGE (entry_bb)->dest; | |
5778 | gcc_assert (broken_loop | |
5779 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); | |
5780 | body_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
5781 | if (!broken_loop) | |
5782 | { | |
5783 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); | |
5784 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); | |
5785 | } | |
5786 | exit_bb = region->exit; | |
5787 | ||
5788 | /* Iteration space partitioning goes in ENTRY_BB. */ | |
65f4b875 | 5789 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
5790 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
5791 | ||
5792 | if (fd->collapse > 1) | |
5793 | { | |
5794 | int first_zero_iter = -1, dummy = -1; | |
5795 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; | |
5796 | ||
5797 | counts = XALLOCAVEC (tree, fd->collapse); | |
5798 | expand_omp_for_init_counts (fd, &gsi, entry_bb, counts, | |
5799 | fin_bb, first_zero_iter, | |
5800 | dummy_bb, dummy, l2_dom_bb); | |
5801 | t = NULL_TREE; | |
5802 | } | |
5803 | else | |
5804 | t = integer_one_node; | |
5805 | ||
5806 | step = fd->loop.step; | |
5807 | tree innerc = omp_find_clause (gimple_omp_for_clauses (fd->for_stmt), | |
5808 | OMP_CLAUSE__LOOPTEMP_); | |
5809 | gcc_assert (innerc); | |
5810 | n1 = OMP_CLAUSE_DECL (innerc); | |
5811 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), OMP_CLAUSE__LOOPTEMP_); | |
5812 | gcc_assert (innerc); | |
5813 | n2 = OMP_CLAUSE_DECL (innerc); | |
5814 | if (bias) | |
5815 | { | |
5816 | n1 = fold_build2 (PLUS_EXPR, fd->iter_type, n1, bias); | |
5817 | n2 = fold_build2 (PLUS_EXPR, fd->iter_type, n2, bias); | |
5818 | } | |
5819 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
5820 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5821 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), | |
5822 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5823 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
5824 | true, NULL_TREE, true, GSI_SAME_STMT); | |
5825 | ||
5826 | tree startvar = fd->loop.v; | |
5827 | tree endvar = NULL_TREE; | |
5828 | ||
5829 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5830 | { | |
5831 | tree clauses = gimple_omp_for_clauses (inner_stmt); | |
5832 | tree innerc = omp_find_clause (clauses, OMP_CLAUSE__LOOPTEMP_); | |
5833 | gcc_assert (innerc); | |
5834 | startvar = OMP_CLAUSE_DECL (innerc); | |
5835 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), | |
5836 | OMP_CLAUSE__LOOPTEMP_); | |
5837 | gcc_assert (innerc); | |
5838 | endvar = OMP_CLAUSE_DECL (innerc); | |
5839 | } | |
5840 | t = fold_convert (TREE_TYPE (startvar), n1); | |
5841 | t = force_gimple_operand_gsi (&gsi, t, | |
5842 | DECL_P (startvar) | |
5843 | && TREE_ADDRESSABLE (startvar), | |
5844 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
5845 | gimple *assign_stmt = gimple_build_assign (startvar, t); | |
5846 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5847 | ||
5848 | t = fold_convert (TREE_TYPE (startvar), n2); | |
5849 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, | |
5850 | false, GSI_CONTINUE_LINKING); | |
5851 | if (endvar) | |
5852 | { | |
5853 | assign_stmt = gimple_build_assign (endvar, e); | |
5854 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5855 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) | |
5856 | assign_stmt = gimple_build_assign (fd->loop.v, e); | |
5857 | else | |
5858 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); | |
5859 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); | |
5860 | } | |
5861 | if (fd->collapse > 1) | |
5862 | expand_omp_for_init_vars (fd, &gsi, counts, inner_stmt, startvar); | |
5863 | ||
5864 | if (!broken_loop) | |
5865 | { | |
5866 | /* The code controlling the sequential loop replaces the | |
5867 | GIMPLE_OMP_CONTINUE. */ | |
65f4b875 | 5868 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
5869 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
5870 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); | |
5871 | vmain = gimple_omp_continue_control_use (cont_stmt); | |
5872 | vback = gimple_omp_continue_control_def (cont_stmt); | |
5873 | ||
5874 | if (!gimple_omp_for_combined_p (fd->for_stmt)) | |
5875 | { | |
5876 | if (POINTER_TYPE_P (type)) | |
5877 | t = fold_build_pointer_plus (vmain, step); | |
5878 | else | |
5879 | t = fold_build2 (PLUS_EXPR, type, vmain, step); | |
5880 | t = force_gimple_operand_gsi (&gsi, t, | |
5881 | DECL_P (vback) | |
5882 | && TREE_ADDRESSABLE (vback), | |
5883 | NULL_TREE, true, GSI_SAME_STMT); | |
5884 | assign_stmt = gimple_build_assign (vback, t); | |
5885 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
5886 | ||
5887 | t = build2 (fd->loop.cond_code, boolean_type_node, | |
5888 | DECL_P (vback) && TREE_ADDRESSABLE (vback) | |
5889 | ? t : vback, e); | |
5890 | gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT); | |
5891 | } | |
5892 | ||
5893 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ | |
5894 | gsi_remove (&gsi, true); | |
5895 | ||
5896 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (fd->for_stmt)) | |
5897 | collapse_bb = extract_omp_for_update_vars (fd, cont_bb, body_bb); | |
5898 | } | |
5899 | ||
5900 | /* Remove the GIMPLE_OMP_FOR statement. */ | |
5901 | gsi = gsi_for_stmt (fd->for_stmt); | |
5902 | gsi_remove (&gsi, true); | |
5903 | ||
5904 | /* Remove the GIMPLE_OMP_RETURN statement. */ | |
65f4b875 | 5905 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
5906 | gsi_remove (&gsi, true); |
5907 | ||
357067f2 | 5908 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); |
629b3d75 MJ |
5909 | if (!broken_loop) |
5910 | remove_edge (BRANCH_EDGE (entry_bb)); | |
5911 | else | |
5912 | { | |
5913 | remove_edge_and_dominated_blocks (BRANCH_EDGE (entry_bb)); | |
5914 | region->outer->cont = NULL; | |
5915 | } | |
5916 | ||
5917 | /* Connect all the blocks. */ | |
5918 | if (!broken_loop) | |
5919 | { | |
5920 | ep = find_edge (cont_bb, body_bb); | |
5921 | if (gimple_omp_for_combined_p (fd->for_stmt)) | |
5922 | { | |
5923 | remove_edge (ep); | |
5924 | ep = NULL; | |
5925 | } | |
5926 | else if (fd->collapse > 1) | |
5927 | { | |
5928 | remove_edge (ep); | |
5929 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); | |
5930 | } | |
5931 | else | |
5932 | ep->flags = EDGE_TRUE_VALUE; | |
5933 | find_edge (cont_bb, fin_bb)->flags | |
5934 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; | |
5935 | } | |
5936 | ||
5937 | set_immediate_dominator (CDI_DOMINATORS, body_bb, | |
5938 | recompute_dominator (CDI_DOMINATORS, body_bb)); | |
5939 | if (!broken_loop) | |
5940 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, | |
5941 | recompute_dominator (CDI_DOMINATORS, fin_bb)); | |
5942 | ||
5943 | if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt)) | |
5944 | { | |
99b1c316 | 5945 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
5946 | loop->header = body_bb; |
5947 | if (collapse_bb == NULL) | |
5948 | loop->latch = cont_bb; | |
5949 | add_loop (loop, body_bb->loop_father); | |
5950 | } | |
5951 | } | |
5952 | ||
5953 | /* A subroutine of expand_omp_for. Generate code for an OpenACC | |
5954 | partitioned loop. The lowering here is abstracted, in that the | |
5955 | loop parameters are passed through internal functions, which are | |
5956 | further lowered by oacc_device_lower, once we get to the target | |
5957 | compiler. The loop is of the form: | |
5958 | ||
5959 | for (V = B; V LTGT E; V += S) {BODY} | |
5960 | ||
5961 | where LTGT is < or >. We may have a specified chunking size, CHUNKING | |
5962 | (constant 0 for no chunking) and we will have a GWV partitioning | |
5963 | mask, specifying dimensions over which the loop is to be | |
02889d23 CLT |
5964 | partitioned (see note below). We generate code that looks like |
5965 | (this ignores tiling): | |
629b3d75 MJ |
5966 | |
5967 | <entry_bb> [incoming FALL->body, BRANCH->exit] | |
5968 | typedef signedintify (typeof (V)) T; // underlying signed integral type | |
5969 | T range = E - B; | |
5970 | T chunk_no = 0; | |
5971 | T DIR = LTGT == '<' ? +1 : -1; | |
5972 | T chunk_max = GOACC_LOOP_CHUNK (dir, range, S, CHUNK_SIZE, GWV); | |
5973 | T step = GOACC_LOOP_STEP (dir, range, S, CHUNK_SIZE, GWV); | |
5974 | ||
5975 | <head_bb> [created by splitting end of entry_bb] | |
5976 | T offset = GOACC_LOOP_OFFSET (dir, range, S, CHUNK_SIZE, GWV, chunk_no); | |
5977 | T bound = GOACC_LOOP_BOUND (dir, range, S, CHUNK_SIZE, GWV, offset); | |
5978 | if (!(offset LTGT bound)) goto bottom_bb; | |
5979 | ||
5980 | <body_bb> [incoming] | |
5981 | V = B + offset; | |
5982 | {BODY} | |
5983 | ||
5984 | <cont_bb> [incoming, may == body_bb FALL->exit_bb, BRANCH->body_bb] | |
5985 | offset += step; | |
5986 | if (offset LTGT bound) goto body_bb; [*] | |
5987 | ||
5988 | <bottom_bb> [created by splitting start of exit_bb] insert BRANCH->head_bb | |
5989 | chunk_no++; | |
5990 | if (chunk < chunk_max) goto head_bb; | |
5991 | ||
5992 | <exit_bb> [incoming] | |
5993 | V = B + ((range -/+ 1) / S +/- 1) * S [*] | |
5994 | ||
02889d23 | 5995 | [*] Needed if V live at end of loop. */ |
629b3d75 MJ |
5996 | |
5997 | static void | |
5998 | expand_oacc_for (struct omp_region *region, struct omp_for_data *fd) | |
5999 | { | |
6000 | tree v = fd->loop.v; | |
6001 | enum tree_code cond_code = fd->loop.cond_code; | |
6002 | enum tree_code plus_code = PLUS_EXPR; | |
6003 | ||
6004 | tree chunk_size = integer_minus_one_node; | |
6005 | tree gwv = integer_zero_node; | |
6006 | tree iter_type = TREE_TYPE (v); | |
6007 | tree diff_type = iter_type; | |
6008 | tree plus_type = iter_type; | |
6009 | struct oacc_collapse *counts = NULL; | |
6010 | ||
6011 | gcc_checking_assert (gimple_omp_for_kind (fd->for_stmt) | |
6012 | == GF_OMP_FOR_KIND_OACC_LOOP); | |
6013 | gcc_assert (!gimple_omp_for_combined_into_p (fd->for_stmt)); | |
6014 | gcc_assert (cond_code == LT_EXPR || cond_code == GT_EXPR); | |
6015 | ||
6016 | if (POINTER_TYPE_P (iter_type)) | |
6017 | { | |
6018 | plus_code = POINTER_PLUS_EXPR; | |
6019 | plus_type = sizetype; | |
6020 | } | |
6021 | if (POINTER_TYPE_P (diff_type) || TYPE_UNSIGNED (diff_type)) | |
6022 | diff_type = signed_type_for (diff_type); | |
f4c222c0 TV |
6023 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (integer_type_node)) |
6024 | diff_type = integer_type_node; | |
629b3d75 MJ |
6025 | |
6026 | basic_block entry_bb = region->entry; /* BB ending in OMP_FOR */ | |
6027 | basic_block exit_bb = region->exit; /* BB ending in OMP_RETURN */ | |
6028 | basic_block cont_bb = region->cont; /* BB ending in OMP_CONTINUE */ | |
6029 | basic_block bottom_bb = NULL; | |
6030 | ||
6031 | /* entry_bb has two sucessors; the branch edge is to the exit | |
6032 | block, fallthrough edge to body. */ | |
6033 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2 | |
6034 | && BRANCH_EDGE (entry_bb)->dest == exit_bb); | |
6035 | ||
6036 | /* If cont_bb non-NULL, it has 2 successors. The branch successor is | |
6037 | body_bb, or to a block whose only successor is the body_bb. Its | |
6038 | fallthrough successor is the final block (same as the branch | |
6039 | successor of the entry_bb). */ | |
6040 | if (cont_bb) | |
6041 | { | |
6042 | basic_block body_bb = FALLTHRU_EDGE (entry_bb)->dest; | |
6043 | basic_block bed = BRANCH_EDGE (cont_bb)->dest; | |
6044 | ||
6045 | gcc_assert (FALLTHRU_EDGE (cont_bb)->dest == exit_bb); | |
6046 | gcc_assert (bed == body_bb || single_succ_edge (bed)->dest == body_bb); | |
6047 | } | |
6048 | else | |
6049 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
6050 | ||
6051 | /* The exit block only has entry_bb and cont_bb as predecessors. */ | |
6052 | gcc_assert (EDGE_COUNT (exit_bb->preds) == 1 + (cont_bb != NULL)); | |
6053 | ||
6054 | tree chunk_no; | |
6055 | tree chunk_max = NULL_TREE; | |
6056 | tree bound, offset; | |
6057 | tree step = create_tmp_var (diff_type, ".step"); | |
6058 | bool up = cond_code == LT_EXPR; | |
6059 | tree dir = build_int_cst (diff_type, up ? +1 : -1); | |
02889d23 | 6060 | bool chunking = !gimple_in_ssa_p (cfun); |
629b3d75 MJ |
6061 | bool negating; |
6062 | ||
02889d23 CLT |
6063 | /* Tiling vars. */ |
6064 | tree tile_size = NULL_TREE; | |
6065 | tree element_s = NULL_TREE; | |
6066 | tree e_bound = NULL_TREE, e_offset = NULL_TREE, e_step = NULL_TREE; | |
6067 | basic_block elem_body_bb = NULL; | |
6068 | basic_block elem_cont_bb = NULL; | |
6069 | ||
629b3d75 MJ |
6070 | /* SSA instances. */ |
6071 | tree offset_incr = NULL_TREE; | |
6072 | tree offset_init = NULL_TREE; | |
6073 | ||
6074 | gimple_stmt_iterator gsi; | |
6075 | gassign *ass; | |
6076 | gcall *call; | |
6077 | gimple *stmt; | |
6078 | tree expr; | |
6079 | location_t loc; | |
6080 | edge split, be, fte; | |
6081 | ||
6082 | /* Split the end of entry_bb to create head_bb. */ | |
6083 | split = split_block (entry_bb, last_stmt (entry_bb)); | |
6084 | basic_block head_bb = split->dest; | |
6085 | entry_bb = split->src; | |
6086 | ||
6087 | /* Chunk setup goes at end of entry_bb, replacing the omp_for. */ | |
65f4b875 | 6088 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
6089 | gomp_for *for_stmt = as_a <gomp_for *> (gsi_stmt (gsi)); |
6090 | loc = gimple_location (for_stmt); | |
6091 | ||
6092 | if (gimple_in_ssa_p (cfun)) | |
6093 | { | |
6094 | offset_init = gimple_omp_for_index (for_stmt, 0); | |
6095 | gcc_assert (integer_zerop (fd->loop.n1)); | |
6096 | /* The SSA parallelizer does gang parallelism. */ | |
6097 | gwv = build_int_cst (integer_type_node, GOMP_DIM_MASK (GOMP_DIM_GANG)); | |
6098 | } | |
6099 | ||
02889d23 | 6100 | if (fd->collapse > 1 || fd->tiling) |
629b3d75 | 6101 | { |
02889d23 | 6102 | gcc_assert (!gimple_in_ssa_p (cfun) && up); |
629b3d75 MJ |
6103 | counts = XALLOCAVEC (struct oacc_collapse, fd->collapse); |
6104 | tree total = expand_oacc_collapse_init (fd, &gsi, counts, | |
02889d23 | 6105 | TREE_TYPE (fd->loop.n2), loc); |
629b3d75 MJ |
6106 | |
6107 | if (SSA_VAR_P (fd->loop.n2)) | |
6108 | { | |
6109 | total = force_gimple_operand_gsi (&gsi, total, false, NULL_TREE, | |
6110 | true, GSI_SAME_STMT); | |
6111 | ass = gimple_build_assign (fd->loop.n2, total); | |
6112 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6113 | } | |
629b3d75 MJ |
6114 | } |
6115 | ||
6116 | tree b = fd->loop.n1; | |
6117 | tree e = fd->loop.n2; | |
6118 | tree s = fd->loop.step; | |
6119 | ||
6120 | b = force_gimple_operand_gsi (&gsi, b, true, NULL_TREE, true, GSI_SAME_STMT); | |
6121 | e = force_gimple_operand_gsi (&gsi, e, true, NULL_TREE, true, GSI_SAME_STMT); | |
6122 | ||
01914336 | 6123 | /* Convert the step, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
6124 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); |
6125 | if (negating) | |
6126 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); | |
6127 | s = fold_convert (diff_type, s); | |
6128 | if (negating) | |
6129 | s = fold_build1 (NEGATE_EXPR, diff_type, s); | |
6130 | s = force_gimple_operand_gsi (&gsi, s, true, NULL_TREE, true, GSI_SAME_STMT); | |
6131 | ||
6132 | if (!chunking) | |
6133 | chunk_size = integer_zero_node; | |
6134 | expr = fold_convert (diff_type, chunk_size); | |
6135 | chunk_size = force_gimple_operand_gsi (&gsi, expr, true, | |
6136 | NULL_TREE, true, GSI_SAME_STMT); | |
02889d23 CLT |
6137 | |
6138 | if (fd->tiling) | |
6139 | { | |
6140 | /* Determine the tile size and element step, | |
6141 | modify the outer loop step size. */ | |
6142 | tile_size = create_tmp_var (diff_type, ".tile_size"); | |
6143 | expr = build_int_cst (diff_type, 1); | |
6144 | for (int ix = 0; ix < fd->collapse; ix++) | |
6145 | expr = fold_build2 (MULT_EXPR, diff_type, counts[ix].tile, expr); | |
6146 | expr = force_gimple_operand_gsi (&gsi, expr, true, | |
6147 | NULL_TREE, true, GSI_SAME_STMT); | |
6148 | ass = gimple_build_assign (tile_size, expr); | |
6149 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6150 | ||
6151 | element_s = create_tmp_var (diff_type, ".element_s"); | |
6152 | ass = gimple_build_assign (element_s, s); | |
6153 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6154 | ||
6155 | expr = fold_build2 (MULT_EXPR, diff_type, s, tile_size); | |
6156 | s = force_gimple_operand_gsi (&gsi, expr, true, | |
6157 | NULL_TREE, true, GSI_SAME_STMT); | |
6158 | } | |
6159 | ||
01914336 | 6160 | /* Determine the range, avoiding possible unsigned->signed overflow. */ |
629b3d75 MJ |
6161 | negating = !up && TYPE_UNSIGNED (iter_type); |
6162 | expr = fold_build2 (MINUS_EXPR, plus_type, | |
6163 | fold_convert (plus_type, negating ? b : e), | |
6164 | fold_convert (plus_type, negating ? e : b)); | |
6165 | expr = fold_convert (diff_type, expr); | |
6166 | if (negating) | |
6167 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); | |
6168 | tree range = force_gimple_operand_gsi (&gsi, expr, true, | |
6169 | NULL_TREE, true, GSI_SAME_STMT); | |
6170 | ||
6171 | chunk_no = build_int_cst (diff_type, 0); | |
6172 | if (chunking) | |
6173 | { | |
6174 | gcc_assert (!gimple_in_ssa_p (cfun)); | |
6175 | ||
6176 | expr = chunk_no; | |
6177 | chunk_max = create_tmp_var (diff_type, ".chunk_max"); | |
6178 | chunk_no = create_tmp_var (diff_type, ".chunk_no"); | |
6179 | ||
6180 | ass = gimple_build_assign (chunk_no, expr); | |
6181 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6182 | ||
6183 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, | |
6184 | build_int_cst (integer_type_node, | |
6185 | IFN_GOACC_LOOP_CHUNKS), | |
6186 | dir, range, s, chunk_size, gwv); | |
6187 | gimple_call_set_lhs (call, chunk_max); | |
6188 | gimple_set_location (call, loc); | |
6189 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
6190 | } | |
6191 | else | |
6192 | chunk_size = chunk_no; | |
6193 | ||
6194 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, | |
6195 | build_int_cst (integer_type_node, | |
6196 | IFN_GOACC_LOOP_STEP), | |
6197 | dir, range, s, chunk_size, gwv); | |
6198 | gimple_call_set_lhs (call, step); | |
6199 | gimple_set_location (call, loc); | |
6200 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
6201 | ||
6202 | /* Remove the GIMPLE_OMP_FOR. */ | |
6203 | gsi_remove (&gsi, true); | |
6204 | ||
01914336 | 6205 | /* Fixup edges from head_bb. */ |
629b3d75 MJ |
6206 | be = BRANCH_EDGE (head_bb); |
6207 | fte = FALLTHRU_EDGE (head_bb); | |
6208 | be->flags |= EDGE_FALSE_VALUE; | |
6209 | fte->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
6210 | ||
6211 | basic_block body_bb = fte->dest; | |
6212 | ||
6213 | if (gimple_in_ssa_p (cfun)) | |
6214 | { | |
65f4b875 | 6215 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
6216 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
6217 | ||
6218 | offset = gimple_omp_continue_control_use (cont_stmt); | |
6219 | offset_incr = gimple_omp_continue_control_def (cont_stmt); | |
6220 | } | |
6221 | else | |
6222 | { | |
6223 | offset = create_tmp_var (diff_type, ".offset"); | |
6224 | offset_init = offset_incr = offset; | |
6225 | } | |
6226 | bound = create_tmp_var (TREE_TYPE (offset), ".bound"); | |
6227 | ||
6228 | /* Loop offset & bound go into head_bb. */ | |
6229 | gsi = gsi_start_bb (head_bb); | |
6230 | ||
6231 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, | |
6232 | build_int_cst (integer_type_node, | |
6233 | IFN_GOACC_LOOP_OFFSET), | |
6234 | dir, range, s, | |
6235 | chunk_size, gwv, chunk_no); | |
6236 | gimple_call_set_lhs (call, offset_init); | |
6237 | gimple_set_location (call, loc); | |
6238 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); | |
6239 | ||
6240 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, | |
6241 | build_int_cst (integer_type_node, | |
6242 | IFN_GOACC_LOOP_BOUND), | |
6243 | dir, range, s, | |
6244 | chunk_size, gwv, offset_init); | |
6245 | gimple_call_set_lhs (call, bound); | |
6246 | gimple_set_location (call, loc); | |
6247 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); | |
6248 | ||
6249 | expr = build2 (cond_code, boolean_type_node, offset_init, bound); | |
6250 | gsi_insert_after (&gsi, gimple_build_cond_empty (expr), | |
6251 | GSI_CONTINUE_LINKING); | |
6252 | ||
6253 | /* V assignment goes into body_bb. */ | |
6254 | if (!gimple_in_ssa_p (cfun)) | |
6255 | { | |
6256 | gsi = gsi_start_bb (body_bb); | |
6257 | ||
6258 | expr = build2 (plus_code, iter_type, b, | |
6259 | fold_convert (plus_type, offset)); | |
6260 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
6261 | true, GSI_SAME_STMT); | |
6262 | ass = gimple_build_assign (v, expr); | |
6263 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
02889d23 CLT |
6264 | |
6265 | if (fd->collapse > 1 || fd->tiling) | |
6266 | expand_oacc_collapse_vars (fd, false, &gsi, counts, v); | |
6267 | ||
6268 | if (fd->tiling) | |
6269 | { | |
6270 | /* Determine the range of the element loop -- usually simply | |
6271 | the tile_size, but could be smaller if the final | |
6272 | iteration of the outer loop is a partial tile. */ | |
6273 | tree e_range = create_tmp_var (diff_type, ".e_range"); | |
6274 | ||
6275 | expr = build2 (MIN_EXPR, diff_type, | |
6276 | build2 (MINUS_EXPR, diff_type, bound, offset), | |
6277 | build2 (MULT_EXPR, diff_type, tile_size, | |
6278 | element_s)); | |
6279 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
6280 | true, GSI_SAME_STMT); | |
6281 | ass = gimple_build_assign (e_range, expr); | |
6282 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6283 | ||
6284 | /* Determine bound, offset & step of inner loop. */ | |
6285 | e_bound = create_tmp_var (diff_type, ".e_bound"); | |
6286 | e_offset = create_tmp_var (diff_type, ".e_offset"); | |
6287 | e_step = create_tmp_var (diff_type, ".e_step"); | |
6288 | ||
6289 | /* Mark these as element loops. */ | |
6290 | tree t, e_gwv = integer_minus_one_node; | |
6291 | tree chunk = build_int_cst (diff_type, 0); /* Never chunked. */ | |
6292 | ||
6293 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_OFFSET); | |
6294 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, | |
6295 | element_s, chunk, e_gwv, chunk); | |
6296 | gimple_call_set_lhs (call, e_offset); | |
6297 | gimple_set_location (call, loc); | |
6298 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
6299 | ||
6300 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_BOUND); | |
6301 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, | |
6302 | element_s, chunk, e_gwv, e_offset); | |
6303 | gimple_call_set_lhs (call, e_bound); | |
6304 | gimple_set_location (call, loc); | |
6305 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
6306 | ||
6307 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_STEP); | |
6308 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, t, dir, e_range, | |
6309 | element_s, chunk, e_gwv); | |
6310 | gimple_call_set_lhs (call, e_step); | |
6311 | gimple_set_location (call, loc); | |
6312 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
6313 | ||
6314 | /* Add test and split block. */ | |
6315 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); | |
6316 | stmt = gimple_build_cond_empty (expr); | |
6317 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
6318 | split = split_block (body_bb, stmt); | |
6319 | elem_body_bb = split->dest; | |
6320 | if (cont_bb == body_bb) | |
6321 | cont_bb = elem_body_bb; | |
6322 | body_bb = split->src; | |
6323 | ||
6324 | split->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
6325 | ||
05e0af43 CP |
6326 | /* Add a dummy exit for the tiled block when cont_bb is missing. */ |
6327 | if (cont_bb == NULL) | |
6328 | { | |
6329 | edge e = make_edge (body_bb, exit_bb, EDGE_FALSE_VALUE); | |
6330 | e->probability = profile_probability::even (); | |
6331 | split->probability = profile_probability::even (); | |
6332 | } | |
6333 | ||
02889d23 CLT |
6334 | /* Initialize the user's loop vars. */ |
6335 | gsi = gsi_start_bb (elem_body_bb); | |
6336 | expand_oacc_collapse_vars (fd, true, &gsi, counts, e_offset); | |
6337 | } | |
629b3d75 MJ |
6338 | } |
6339 | ||
6340 | /* Loop increment goes into cont_bb. If this is not a loop, we | |
6341 | will have spawned threads as if it was, and each one will | |
6342 | execute one iteration. The specification is not explicit about | |
6343 | whether such constructs are ill-formed or not, and they can | |
6344 | occur, especially when noreturn routines are involved. */ | |
6345 | if (cont_bb) | |
6346 | { | |
65f4b875 | 6347 | gsi = gsi_last_nondebug_bb (cont_bb); |
629b3d75 MJ |
6348 | gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi)); |
6349 | loc = gimple_location (cont_stmt); | |
6350 | ||
02889d23 CLT |
6351 | if (fd->tiling) |
6352 | { | |
6353 | /* Insert element loop increment and test. */ | |
6354 | expr = build2 (PLUS_EXPR, diff_type, e_offset, e_step); | |
6355 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
6356 | true, GSI_SAME_STMT); | |
6357 | ass = gimple_build_assign (e_offset, expr); | |
6358 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6359 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); | |
6360 | ||
6361 | stmt = gimple_build_cond_empty (expr); | |
6362 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
6363 | split = split_block (cont_bb, stmt); | |
6364 | elem_cont_bb = split->src; | |
6365 | cont_bb = split->dest; | |
6366 | ||
6367 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; | |
9fba2943 TV |
6368 | split->probability = profile_probability::unlikely ().guessed (); |
6369 | edge latch_edge | |
6370 | = make_edge (elem_cont_bb, elem_body_bb, EDGE_TRUE_VALUE); | |
6371 | latch_edge->probability = profile_probability::likely ().guessed (); | |
6372 | ||
6373 | edge skip_edge = make_edge (body_bb, cont_bb, EDGE_FALSE_VALUE); | |
6374 | skip_edge->probability = profile_probability::unlikely ().guessed (); | |
6375 | edge loop_entry_edge = EDGE_SUCC (body_bb, 1 - skip_edge->dest_idx); | |
6376 | loop_entry_edge->probability | |
6377 | = profile_probability::likely ().guessed (); | |
02889d23 CLT |
6378 | |
6379 | gsi = gsi_for_stmt (cont_stmt); | |
6380 | } | |
6381 | ||
629b3d75 MJ |
6382 | /* Increment offset. */ |
6383 | if (gimple_in_ssa_p (cfun)) | |
02889d23 CLT |
6384 | expr = build2 (plus_code, iter_type, offset, |
6385 | fold_convert (plus_type, step)); | |
629b3d75 MJ |
6386 | else |
6387 | expr = build2 (PLUS_EXPR, diff_type, offset, step); | |
6388 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
6389 | true, GSI_SAME_STMT); | |
6390 | ass = gimple_build_assign (offset_incr, expr); | |
6391 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6392 | expr = build2 (cond_code, boolean_type_node, offset_incr, bound); | |
6393 | gsi_insert_before (&gsi, gimple_build_cond_empty (expr), GSI_SAME_STMT); | |
6394 | ||
6395 | /* Remove the GIMPLE_OMP_CONTINUE. */ | |
6396 | gsi_remove (&gsi, true); | |
6397 | ||
01914336 | 6398 | /* Fixup edges from cont_bb. */ |
629b3d75 MJ |
6399 | be = BRANCH_EDGE (cont_bb); |
6400 | fte = FALLTHRU_EDGE (cont_bb); | |
6401 | be->flags |= EDGE_TRUE_VALUE; | |
6402 | fte->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; | |
6403 | ||
6404 | if (chunking) | |
6405 | { | |
6406 | /* Split the beginning of exit_bb to make bottom_bb. We | |
6407 | need to insert a nop at the start, because splitting is | |
01914336 | 6408 | after a stmt, not before. */ |
629b3d75 MJ |
6409 | gsi = gsi_start_bb (exit_bb); |
6410 | stmt = gimple_build_nop (); | |
6411 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
6412 | split = split_block (exit_bb, stmt); | |
6413 | bottom_bb = split->src; | |
6414 | exit_bb = split->dest; | |
6415 | gsi = gsi_last_bb (bottom_bb); | |
6416 | ||
6417 | /* Chunk increment and test goes into bottom_bb. */ | |
6418 | expr = build2 (PLUS_EXPR, diff_type, chunk_no, | |
6419 | build_int_cst (diff_type, 1)); | |
6420 | ass = gimple_build_assign (chunk_no, expr); | |
6421 | gsi_insert_after (&gsi, ass, GSI_CONTINUE_LINKING); | |
6422 | ||
6423 | /* Chunk test at end of bottom_bb. */ | |
6424 | expr = build2 (LT_EXPR, boolean_type_node, chunk_no, chunk_max); | |
6425 | gsi_insert_after (&gsi, gimple_build_cond_empty (expr), | |
6426 | GSI_CONTINUE_LINKING); | |
6427 | ||
01914336 | 6428 | /* Fixup edges from bottom_bb. */ |
629b3d75 | 6429 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; |
f38057cc TV |
6430 | split->probability = profile_probability::unlikely ().guessed (); |
6431 | edge latch_edge = make_edge (bottom_bb, head_bb, EDGE_TRUE_VALUE); | |
6432 | latch_edge->probability = profile_probability::likely ().guessed (); | |
629b3d75 MJ |
6433 | } |
6434 | } | |
6435 | ||
65f4b875 | 6436 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
6437 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
6438 | loc = gimple_location (gsi_stmt (gsi)); | |
6439 | ||
6440 | if (!gimple_in_ssa_p (cfun)) | |
6441 | { | |
6442 | /* Insert the final value of V, in case it is live. This is the | |
6443 | value for the only thread that survives past the join. */ | |
6444 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); | |
6445 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); | |
6446 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); | |
6447 | expr = fold_build2 (MULT_EXPR, diff_type, expr, s); | |
6448 | expr = build2 (plus_code, iter_type, b, fold_convert (plus_type, expr)); | |
6449 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, | |
6450 | true, GSI_SAME_STMT); | |
6451 | ass = gimple_build_assign (v, expr); | |
6452 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); | |
6453 | } | |
6454 | ||
01914336 | 6455 | /* Remove the OMP_RETURN. */ |
629b3d75 MJ |
6456 | gsi_remove (&gsi, true); |
6457 | ||
6458 | if (cont_bb) | |
6459 | { | |
02889d23 | 6460 | /* We now have one, two or three nested loops. Update the loop |
629b3d75 | 6461 | structures. */ |
99b1c316 MS |
6462 | class loop *parent = entry_bb->loop_father; |
6463 | class loop *body = body_bb->loop_father; | |
629b3d75 MJ |
6464 | |
6465 | if (chunking) | |
6466 | { | |
99b1c316 | 6467 | class loop *chunk_loop = alloc_loop (); |
629b3d75 MJ |
6468 | chunk_loop->header = head_bb; |
6469 | chunk_loop->latch = bottom_bb; | |
6470 | add_loop (chunk_loop, parent); | |
6471 | parent = chunk_loop; | |
6472 | } | |
6473 | else if (parent != body) | |
6474 | { | |
6475 | gcc_assert (body->header == body_bb); | |
6476 | gcc_assert (body->latch == cont_bb | |
6477 | || single_pred (body->latch) == cont_bb); | |
6478 | parent = NULL; | |
6479 | } | |
6480 | ||
6481 | if (parent) | |
6482 | { | |
99b1c316 | 6483 | class loop *body_loop = alloc_loop (); |
629b3d75 MJ |
6484 | body_loop->header = body_bb; |
6485 | body_loop->latch = cont_bb; | |
6486 | add_loop (body_loop, parent); | |
02889d23 CLT |
6487 | |
6488 | if (fd->tiling) | |
6489 | { | |
6490 | /* Insert tiling's element loop. */ | |
99b1c316 | 6491 | class loop *inner_loop = alloc_loop (); |
02889d23 CLT |
6492 | inner_loop->header = elem_body_bb; |
6493 | inner_loop->latch = elem_cont_bb; | |
6494 | add_loop (inner_loop, body_loop); | |
6495 | } | |
629b3d75 MJ |
6496 | } |
6497 | } | |
6498 | } | |
6499 | ||
6500 | /* Expand the OMP loop defined by REGION. */ | |
6501 | ||
6502 | static void | |
6503 | expand_omp_for (struct omp_region *region, gimple *inner_stmt) | |
6504 | { | |
6505 | struct omp_for_data fd; | |
6506 | struct omp_for_data_loop *loops; | |
6507 | ||
6508 | loops | |
6509 | = (struct omp_for_data_loop *) | |
6510 | alloca (gimple_omp_for_collapse (last_stmt (region->entry)) | |
6511 | * sizeof (struct omp_for_data_loop)); | |
6512 | omp_extract_for_data (as_a <gomp_for *> (last_stmt (region->entry)), | |
6513 | &fd, loops); | |
6514 | region->sched_kind = fd.sched_kind; | |
6515 | region->sched_modifiers = fd.sched_modifiers; | |
0b887b75 | 6516 | region->has_lastprivate_conditional = fd.lastprivate_conditional != 0; |
629b3d75 MJ |
6517 | |
6518 | gcc_assert (EDGE_COUNT (region->entry->succs) == 2); | |
6519 | BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
6520 | FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; | |
6521 | if (region->cont) | |
6522 | { | |
6523 | gcc_assert (EDGE_COUNT (region->cont->succs) == 2); | |
6524 | BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
6525 | FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; | |
6526 | } | |
6527 | else | |
6528 | /* If there isn't a continue then this is a degerate case where | |
6529 | the introduction of abnormal edges during lowering will prevent | |
6530 | original loops from being detected. Fix that up. */ | |
6531 | loops_state_set (LOOPS_NEED_FIXUP); | |
6532 | ||
dfa6e5b4 | 6533 | if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_SIMD) |
629b3d75 | 6534 | expand_omp_simd (region, &fd); |
629b3d75 MJ |
6535 | else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_OACC_LOOP) |
6536 | { | |
6537 | gcc_assert (!inner_stmt); | |
6538 | expand_oacc_for (region, &fd); | |
6539 | } | |
6540 | else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_TASKLOOP) | |
6541 | { | |
6542 | if (gimple_omp_for_combined_into_p (fd.for_stmt)) | |
6543 | expand_omp_taskloop_for_inner (region, &fd, inner_stmt); | |
6544 | else | |
6545 | expand_omp_taskloop_for_outer (region, &fd, inner_stmt); | |
6546 | } | |
6547 | else if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC | |
6548 | && !fd.have_ordered) | |
6549 | { | |
6550 | if (fd.chunk_size == NULL) | |
6551 | expand_omp_for_static_nochunk (region, &fd, inner_stmt); | |
6552 | else | |
6553 | expand_omp_for_static_chunk (region, &fd, inner_stmt); | |
6554 | } | |
6555 | else | |
6556 | { | |
6557 | int fn_index, start_ix, next_ix; | |
28567c40 JJ |
6558 | unsigned HOST_WIDE_INT sched = 0; |
6559 | tree sched_arg = NULL_TREE; | |
629b3d75 MJ |
6560 | |
6561 | gcc_assert (gimple_omp_for_kind (fd.for_stmt) | |
6562 | == GF_OMP_FOR_KIND_FOR); | |
6563 | if (fd.chunk_size == NULL | |
6564 | && fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) | |
6565 | fd.chunk_size = integer_zero_node; | |
629b3d75 MJ |
6566 | switch (fd.sched_kind) |
6567 | { | |
6568 | case OMP_CLAUSE_SCHEDULE_RUNTIME: | |
0b887b75 JJ |
6569 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0 |
6570 | && fd.lastprivate_conditional == 0) | |
28567c40 JJ |
6571 | { |
6572 | gcc_assert (!fd.have_ordered); | |
6573 | fn_index = 6; | |
6574 | sched = 4; | |
6575 | } | |
6576 | else if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 | |
0b887b75 JJ |
6577 | && !fd.have_ordered |
6578 | && fd.lastprivate_conditional == 0) | |
28567c40 JJ |
6579 | fn_index = 7; |
6580 | else | |
6581 | { | |
6582 | fn_index = 3; | |
6583 | sched = (HOST_WIDE_INT_1U << 31); | |
6584 | } | |
629b3d75 MJ |
6585 | break; |
6586 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: | |
6587 | case OMP_CLAUSE_SCHEDULE_GUIDED: | |
28567c40 | 6588 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 |
0b887b75 JJ |
6589 | && !fd.have_ordered |
6590 | && fd.lastprivate_conditional == 0) | |
629b3d75 MJ |
6591 | { |
6592 | fn_index = 3 + fd.sched_kind; | |
28567c40 | 6593 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; |
629b3d75 MJ |
6594 | break; |
6595 | } | |
629b3d75 | 6596 | fn_index = fd.sched_kind; |
28567c40 JJ |
6597 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; |
6598 | sched += (HOST_WIDE_INT_1U << 31); | |
629b3d75 | 6599 | break; |
28567c40 JJ |
6600 | case OMP_CLAUSE_SCHEDULE_STATIC: |
6601 | gcc_assert (fd.have_ordered); | |
6602 | fn_index = 0; | |
6603 | sched = (HOST_WIDE_INT_1U << 31) + 1; | |
6604 | break; | |
6605 | default: | |
6606 | gcc_unreachable (); | |
629b3d75 MJ |
6607 | } |
6608 | if (!fd.ordered) | |
28567c40 | 6609 | fn_index += fd.have_ordered * 8; |
629b3d75 MJ |
6610 | if (fd.ordered) |
6611 | start_ix = ((int)BUILT_IN_GOMP_LOOP_DOACROSS_STATIC_START) + fn_index; | |
6612 | else | |
6613 | start_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_START) + fn_index; | |
6614 | next_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_NEXT) + fn_index; | |
8221c30b | 6615 | if (fd.have_reductemp || fd.have_pointer_condtemp) |
28567c40 JJ |
6616 | { |
6617 | if (fd.ordered) | |
6618 | start_ix = (int)BUILT_IN_GOMP_LOOP_DOACROSS_START; | |
6619 | else if (fd.have_ordered) | |
6620 | start_ix = (int)BUILT_IN_GOMP_LOOP_ORDERED_START; | |
6621 | else | |
6622 | start_ix = (int)BUILT_IN_GOMP_LOOP_START; | |
6623 | sched_arg = build_int_cstu (long_integer_type_node, sched); | |
6624 | if (!fd.chunk_size) | |
6625 | fd.chunk_size = integer_zero_node; | |
6626 | } | |
629b3d75 MJ |
6627 | if (fd.iter_type == long_long_unsigned_type_node) |
6628 | { | |
6629 | start_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_START | |
6630 | - (int)BUILT_IN_GOMP_LOOP_STATIC_START); | |
6631 | next_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_NEXT | |
6632 | - (int)BUILT_IN_GOMP_LOOP_STATIC_NEXT); | |
6633 | } | |
6634 | expand_omp_for_generic (region, &fd, (enum built_in_function) start_ix, | |
28567c40 JJ |
6635 | (enum built_in_function) next_ix, sched_arg, |
6636 | inner_stmt); | |
629b3d75 MJ |
6637 | } |
6638 | ||
6639 | if (gimple_in_ssa_p (cfun)) | |
6640 | update_ssa (TODO_update_ssa_only_virtuals); | |
6641 | } | |
6642 | ||
6643 | /* Expand code for an OpenMP sections directive. In pseudo code, we generate | |
6644 | ||
6645 | v = GOMP_sections_start (n); | |
6646 | L0: | |
6647 | switch (v) | |
6648 | { | |
6649 | case 0: | |
6650 | goto L2; | |
6651 | case 1: | |
6652 | section 1; | |
6653 | goto L1; | |
6654 | case 2: | |
6655 | ... | |
6656 | case n: | |
6657 | ... | |
6658 | default: | |
6659 | abort (); | |
6660 | } | |
6661 | L1: | |
6662 | v = GOMP_sections_next (); | |
6663 | goto L0; | |
6664 | L2: | |
6665 | reduction; | |
6666 | ||
6667 | If this is a combined parallel sections, replace the call to | |
6668 | GOMP_sections_start with call to GOMP_sections_next. */ | |
6669 | ||
6670 | static void | |
6671 | expand_omp_sections (struct omp_region *region) | |
6672 | { | |
6673 | tree t, u, vin = NULL, vmain, vnext, l2; | |
6674 | unsigned len; | |
6675 | basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb; | |
6676 | gimple_stmt_iterator si, switch_si; | |
6677 | gomp_sections *sections_stmt; | |
6678 | gimple *stmt; | |
6679 | gomp_continue *cont; | |
6680 | edge_iterator ei; | |
6681 | edge e; | |
6682 | struct omp_region *inner; | |
6683 | unsigned i, casei; | |
6684 | bool exit_reachable = region->cont != NULL; | |
6685 | ||
6686 | gcc_assert (region->exit != NULL); | |
6687 | entry_bb = region->entry; | |
6688 | l0_bb = single_succ (entry_bb); | |
6689 | l1_bb = region->cont; | |
6690 | l2_bb = region->exit; | |
6691 | if (single_pred_p (l2_bb) && single_pred (l2_bb) == l0_bb) | |
6692 | l2 = gimple_block_label (l2_bb); | |
6693 | else | |
6694 | { | |
6695 | /* This can happen if there are reductions. */ | |
6696 | len = EDGE_COUNT (l0_bb->succs); | |
6697 | gcc_assert (len > 0); | |
6698 | e = EDGE_SUCC (l0_bb, len - 1); | |
65f4b875 | 6699 | si = gsi_last_nondebug_bb (e->dest); |
629b3d75 MJ |
6700 | l2 = NULL_TREE; |
6701 | if (gsi_end_p (si) | |
01914336 | 6702 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) |
629b3d75 MJ |
6703 | l2 = gimple_block_label (e->dest); |
6704 | else | |
6705 | FOR_EACH_EDGE (e, ei, l0_bb->succs) | |
6706 | { | |
65f4b875 | 6707 | si = gsi_last_nondebug_bb (e->dest); |
629b3d75 MJ |
6708 | if (gsi_end_p (si) |
6709 | || gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION) | |
6710 | { | |
6711 | l2 = gimple_block_label (e->dest); | |
6712 | break; | |
6713 | } | |
6714 | } | |
6715 | } | |
6716 | if (exit_reachable) | |
6717 | default_bb = create_empty_bb (l1_bb->prev_bb); | |
6718 | else | |
6719 | default_bb = create_empty_bb (l0_bb); | |
6720 | ||
6721 | /* We will build a switch() with enough cases for all the | |
6722 | GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work | |
6723 | and a default case to abort if something goes wrong. */ | |
6724 | len = EDGE_COUNT (l0_bb->succs); | |
6725 | ||
6726 | /* Use vec::quick_push on label_vec throughout, since we know the size | |
6727 | in advance. */ | |
6728 | auto_vec<tree> label_vec (len); | |
6729 | ||
6730 | /* The call to GOMP_sections_start goes in ENTRY_BB, replacing the | |
6731 | GIMPLE_OMP_SECTIONS statement. */ | |
65f4b875 | 6732 | si = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
6733 | sections_stmt = as_a <gomp_sections *> (gsi_stmt (si)); |
6734 | gcc_assert (gimple_code (sections_stmt) == GIMPLE_OMP_SECTIONS); | |
6735 | vin = gimple_omp_sections_control (sections_stmt); | |
28567c40 JJ |
6736 | tree clauses = gimple_omp_sections_clauses (sections_stmt); |
6737 | tree reductmp = omp_find_clause (clauses, OMP_CLAUSE__REDUCTEMP_); | |
8e7757ba JJ |
6738 | tree condtmp = omp_find_clause (clauses, OMP_CLAUSE__CONDTEMP_); |
6739 | tree cond_var = NULL_TREE; | |
6740 | if (reductmp || condtmp) | |
6741 | { | |
6742 | tree reductions = null_pointer_node, mem = null_pointer_node; | |
6743 | tree memv = NULL_TREE, condtemp = NULL_TREE; | |
6744 | gimple_stmt_iterator gsi = gsi_none (); | |
6745 | gimple *g = NULL; | |
6746 | if (reductmp) | |
6747 | { | |
6748 | reductions = OMP_CLAUSE_DECL (reductmp); | |
6749 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); | |
6750 | g = SSA_NAME_DEF_STMT (reductions); | |
6751 | reductions = gimple_assign_rhs1 (g); | |
6752 | OMP_CLAUSE_DECL (reductmp) = reductions; | |
6753 | gsi = gsi_for_stmt (g); | |
6754 | } | |
6755 | else | |
6756 | gsi = si; | |
6757 | if (condtmp) | |
6758 | { | |
6759 | condtemp = OMP_CLAUSE_DECL (condtmp); | |
6760 | tree c = omp_find_clause (OMP_CLAUSE_CHAIN (condtmp), | |
6761 | OMP_CLAUSE__CONDTEMP_); | |
6762 | cond_var = OMP_CLAUSE_DECL (c); | |
6763 | tree type = TREE_TYPE (condtemp); | |
6764 | memv = create_tmp_var (type); | |
6765 | TREE_ADDRESSABLE (memv) = 1; | |
6766 | unsigned cnt = 0; | |
6767 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) | |
6768 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE | |
6769 | && OMP_CLAUSE_LASTPRIVATE_CONDITIONAL (c)) | |
6770 | ++cnt; | |
6771 | unsigned HOST_WIDE_INT sz | |
6772 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))) * cnt; | |
6773 | expand_omp_build_assign (&gsi, memv, build_int_cst (type, sz), | |
6774 | false); | |
6775 | mem = build_fold_addr_expr (memv); | |
6776 | } | |
28567c40 JJ |
6777 | t = build_int_cst (unsigned_type_node, len - 1); |
6778 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS2_START); | |
8e7757ba | 6779 | stmt = gimple_build_call (u, 3, t, reductions, mem); |
28567c40 JJ |
6780 | gimple_call_set_lhs (stmt, vin); |
6781 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
8e7757ba JJ |
6782 | if (condtmp) |
6783 | { | |
6784 | expand_omp_build_assign (&gsi, condtemp, memv, false); | |
6785 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), | |
6786 | vin, build_one_cst (TREE_TYPE (cond_var))); | |
6787 | expand_omp_build_assign (&gsi, cond_var, t, false); | |
6788 | } | |
6789 | if (reductmp) | |
6790 | { | |
6791 | gsi_remove (&gsi, true); | |
6792 | release_ssa_name (gimple_assign_lhs (g)); | |
6793 | } | |
28567c40 JJ |
6794 | } |
6795 | else if (!is_combined_parallel (region)) | |
629b3d75 MJ |
6796 | { |
6797 | /* If we are not inside a combined parallel+sections region, | |
6798 | call GOMP_sections_start. */ | |
6799 | t = build_int_cst (unsigned_type_node, len - 1); | |
6800 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_START); | |
6801 | stmt = gimple_build_call (u, 1, t); | |
6802 | } | |
6803 | else | |
6804 | { | |
6805 | /* Otherwise, call GOMP_sections_next. */ | |
6806 | u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); | |
6807 | stmt = gimple_build_call (u, 0); | |
6808 | } | |
8e7757ba | 6809 | if (!reductmp && !condtmp) |
28567c40 JJ |
6810 | { |
6811 | gimple_call_set_lhs (stmt, vin); | |
6812 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
6813 | } | |
629b3d75 MJ |
6814 | gsi_remove (&si, true); |
6815 | ||
6816 | /* The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in | |
6817 | L0_BB. */ | |
65f4b875 | 6818 | switch_si = gsi_last_nondebug_bb (l0_bb); |
629b3d75 MJ |
6819 | gcc_assert (gimple_code (gsi_stmt (switch_si)) == GIMPLE_OMP_SECTIONS_SWITCH); |
6820 | if (exit_reachable) | |
6821 | { | |
6822 | cont = as_a <gomp_continue *> (last_stmt (l1_bb)); | |
6823 | gcc_assert (gimple_code (cont) == GIMPLE_OMP_CONTINUE); | |
6824 | vmain = gimple_omp_continue_control_use (cont); | |
6825 | vnext = gimple_omp_continue_control_def (cont); | |
6826 | } | |
6827 | else | |
6828 | { | |
6829 | vmain = vin; | |
6830 | vnext = NULL_TREE; | |
6831 | } | |
6832 | ||
6833 | t = build_case_label (build_int_cst (unsigned_type_node, 0), NULL, l2); | |
6834 | label_vec.quick_push (t); | |
6835 | i = 1; | |
6836 | ||
6837 | /* Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR. */ | |
6838 | for (inner = region->inner, casei = 1; | |
6839 | inner; | |
6840 | inner = inner->next, i++, casei++) | |
6841 | { | |
6842 | basic_block s_entry_bb, s_exit_bb; | |
6843 | ||
6844 | /* Skip optional reduction region. */ | |
6845 | if (inner->type == GIMPLE_OMP_ATOMIC_LOAD) | |
6846 | { | |
6847 | --i; | |
6848 | --casei; | |
6849 | continue; | |
6850 | } | |
6851 | ||
6852 | s_entry_bb = inner->entry; | |
6853 | s_exit_bb = inner->exit; | |
6854 | ||
6855 | t = gimple_block_label (s_entry_bb); | |
6856 | u = build_int_cst (unsigned_type_node, casei); | |
6857 | u = build_case_label (u, NULL, t); | |
6858 | label_vec.quick_push (u); | |
6859 | ||
65f4b875 | 6860 | si = gsi_last_nondebug_bb (s_entry_bb); |
629b3d75 MJ |
6861 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SECTION); |
6862 | gcc_assert (i < len || gimple_omp_section_last_p (gsi_stmt (si))); | |
6863 | gsi_remove (&si, true); | |
6864 | single_succ_edge (s_entry_bb)->flags = EDGE_FALLTHRU; | |
6865 | ||
6866 | if (s_exit_bb == NULL) | |
6867 | continue; | |
6868 | ||
65f4b875 | 6869 | si = gsi_last_nondebug_bb (s_exit_bb); |
629b3d75 MJ |
6870 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); |
6871 | gsi_remove (&si, true); | |
6872 | ||
6873 | single_succ_edge (s_exit_bb)->flags = EDGE_FALLTHRU; | |
6874 | } | |
6875 | ||
6876 | /* Error handling code goes in DEFAULT_BB. */ | |
6877 | t = gimple_block_label (default_bb); | |
6878 | u = build_case_label (NULL, NULL, t); | |
6879 | make_edge (l0_bb, default_bb, 0); | |
6880 | add_bb_to_loop (default_bb, current_loops->tree_root); | |
6881 | ||
6882 | stmt = gimple_build_switch (vmain, u, label_vec); | |
6883 | gsi_insert_after (&switch_si, stmt, GSI_SAME_STMT); | |
6884 | gsi_remove (&switch_si, true); | |
6885 | ||
6886 | si = gsi_start_bb (default_bb); | |
6887 | stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_TRAP), 0); | |
6888 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); | |
6889 | ||
6890 | if (exit_reachable) | |
6891 | { | |
6892 | tree bfn_decl; | |
6893 | ||
6894 | /* Code to get the next section goes in L1_BB. */ | |
65f4b875 | 6895 | si = gsi_last_nondebug_bb (l1_bb); |
629b3d75 MJ |
6896 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CONTINUE); |
6897 | ||
6898 | bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT); | |
6899 | stmt = gimple_build_call (bfn_decl, 0); | |
6900 | gimple_call_set_lhs (stmt, vnext); | |
8e7757ba JJ |
6901 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
6902 | if (cond_var) | |
6903 | { | |
6904 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), | |
6905 | vnext, build_one_cst (TREE_TYPE (cond_var))); | |
6906 | expand_omp_build_assign (&si, cond_var, t, false); | |
6907 | } | |
629b3d75 MJ |
6908 | gsi_remove (&si, true); |
6909 | ||
6910 | single_succ_edge (l1_bb)->flags = EDGE_FALLTHRU; | |
6911 | } | |
6912 | ||
6913 | /* Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB. */ | |
65f4b875 | 6914 | si = gsi_last_nondebug_bb (l2_bb); |
629b3d75 MJ |
6915 | if (gimple_omp_return_nowait_p (gsi_stmt (si))) |
6916 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_NOWAIT); | |
6917 | else if (gimple_omp_return_lhs (gsi_stmt (si))) | |
6918 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_CANCEL); | |
6919 | else | |
6920 | t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END); | |
6921 | stmt = gimple_build_call (t, 0); | |
6922 | if (gimple_omp_return_lhs (gsi_stmt (si))) | |
6923 | gimple_call_set_lhs (stmt, gimple_omp_return_lhs (gsi_stmt (si))); | |
6924 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); | |
6925 | gsi_remove (&si, true); | |
6926 | ||
6927 | set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb); | |
6928 | } | |
6929 | ||
6930 | /* Expand code for an OpenMP single directive. We've already expanded | |
6931 | much of the code, here we simply place the GOMP_barrier call. */ | |
6932 | ||
6933 | static void | |
6934 | expand_omp_single (struct omp_region *region) | |
6935 | { | |
6936 | basic_block entry_bb, exit_bb; | |
6937 | gimple_stmt_iterator si; | |
6938 | ||
6939 | entry_bb = region->entry; | |
6940 | exit_bb = region->exit; | |
6941 | ||
65f4b875 | 6942 | si = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
6943 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE); |
6944 | gsi_remove (&si, true); | |
6945 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
6946 | ||
65f4b875 | 6947 | si = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
6948 | if (!gimple_omp_return_nowait_p (gsi_stmt (si))) |
6949 | { | |
6950 | tree t = gimple_omp_return_lhs (gsi_stmt (si)); | |
6951 | gsi_insert_after (&si, omp_build_barrier (t), GSI_SAME_STMT); | |
6952 | } | |
6953 | gsi_remove (&si, true); | |
6954 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; | |
6955 | } | |
6956 | ||
6957 | /* Generic expansion for OpenMP synchronization directives: master, | |
6958 | ordered and critical. All we need to do here is remove the entry | |
6959 | and exit markers for REGION. */ | |
6960 | ||
6961 | static void | |
6962 | expand_omp_synch (struct omp_region *region) | |
6963 | { | |
6964 | basic_block entry_bb, exit_bb; | |
6965 | gimple_stmt_iterator si; | |
6966 | ||
6967 | entry_bb = region->entry; | |
6968 | exit_bb = region->exit; | |
6969 | ||
65f4b875 | 6970 | si = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
6971 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE |
6972 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASTER | |
6973 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TASKGROUP | |
6974 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ORDERED | |
6975 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CRITICAL | |
6976 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS); | |
28567c40 JJ |
6977 | if (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS |
6978 | && gimple_omp_teams_host (as_a <gomp_teams *> (gsi_stmt (si)))) | |
6979 | { | |
6980 | expand_omp_taskreg (region); | |
6981 | return; | |
6982 | } | |
629b3d75 MJ |
6983 | gsi_remove (&si, true); |
6984 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
6985 | ||
6986 | if (exit_bb) | |
6987 | { | |
65f4b875 | 6988 | si = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
6989 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); |
6990 | gsi_remove (&si, true); | |
6991 | single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU; | |
6992 | } | |
6993 | } | |
6994 | ||
28567c40 JJ |
6995 | /* Translate enum omp_memory_order to enum memmodel. The two enums |
6996 | are using different numbers so that OMP_MEMORY_ORDER_UNSPECIFIED | |
6997 | is 0. */ | |
6998 | ||
6999 | static enum memmodel | |
7000 | omp_memory_order_to_memmodel (enum omp_memory_order mo) | |
7001 | { | |
7002 | switch (mo) | |
7003 | { | |
7004 | case OMP_MEMORY_ORDER_RELAXED: return MEMMODEL_RELAXED; | |
7005 | case OMP_MEMORY_ORDER_ACQUIRE: return MEMMODEL_ACQUIRE; | |
7006 | case OMP_MEMORY_ORDER_RELEASE: return MEMMODEL_RELEASE; | |
7007 | case OMP_MEMORY_ORDER_ACQ_REL: return MEMMODEL_ACQ_REL; | |
7008 | case OMP_MEMORY_ORDER_SEQ_CST: return MEMMODEL_SEQ_CST; | |
7009 | default: gcc_unreachable (); | |
7010 | } | |
7011 | } | |
7012 | ||
629b3d75 MJ |
7013 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
7014 | operation as a normal volatile load. */ | |
7015 | ||
7016 | static bool | |
7017 | expand_omp_atomic_load (basic_block load_bb, tree addr, | |
7018 | tree loaded_val, int index) | |
7019 | { | |
7020 | enum built_in_function tmpbase; | |
7021 | gimple_stmt_iterator gsi; | |
7022 | basic_block store_bb; | |
7023 | location_t loc; | |
7024 | gimple *stmt; | |
7025 | tree decl, call, type, itype; | |
7026 | ||
65f4b875 | 7027 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
7028 | stmt = gsi_stmt (gsi); |
7029 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
7030 | loc = gimple_location (stmt); | |
7031 | ||
7032 | /* ??? If the target does not implement atomic_load_optab[mode], and mode | |
7033 | is smaller than word size, then expand_atomic_load assumes that the load | |
7034 | is atomic. We could avoid the builtin entirely in this case. */ | |
7035 | ||
7036 | tmpbase = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
7037 | decl = builtin_decl_explicit (tmpbase); | |
7038 | if (decl == NULL_TREE) | |
7039 | return false; | |
7040 | ||
7041 | type = TREE_TYPE (loaded_val); | |
7042 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
7043 | ||
28567c40 JJ |
7044 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt); |
7045 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); | |
7046 | call = build_call_expr_loc (loc, decl, 2, addr, mo); | |
629b3d75 MJ |
7047 | if (!useless_type_conversion_p (type, itype)) |
7048 | call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
7049 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
7050 | ||
7051 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
7052 | gsi_remove (&gsi, true); | |
7053 | ||
7054 | store_bb = single_succ (load_bb); | |
65f4b875 | 7055 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
7056 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); |
7057 | gsi_remove (&gsi, true); | |
7058 | ||
7059 | if (gimple_in_ssa_p (cfun)) | |
7060 | update_ssa (TODO_update_ssa_no_phi); | |
7061 | ||
7062 | return true; | |
7063 | } | |
7064 | ||
7065 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
7066 | operation as a normal volatile store. */ | |
7067 | ||
7068 | static bool | |
7069 | expand_omp_atomic_store (basic_block load_bb, tree addr, | |
7070 | tree loaded_val, tree stored_val, int index) | |
7071 | { | |
7072 | enum built_in_function tmpbase; | |
7073 | gimple_stmt_iterator gsi; | |
7074 | basic_block store_bb = single_succ (load_bb); | |
7075 | location_t loc; | |
7076 | gimple *stmt; | |
7077 | tree decl, call, type, itype; | |
7078 | machine_mode imode; | |
7079 | bool exchange; | |
7080 | ||
65f4b875 | 7081 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
7082 | stmt = gsi_stmt (gsi); |
7083 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); | |
7084 | ||
7085 | /* If the load value is needed, then this isn't a store but an exchange. */ | |
7086 | exchange = gimple_omp_atomic_need_value_p (stmt); | |
7087 | ||
65f4b875 | 7088 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
7089 | stmt = gsi_stmt (gsi); |
7090 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_STORE); | |
7091 | loc = gimple_location (stmt); | |
7092 | ||
7093 | /* ??? If the target does not implement atomic_store_optab[mode], and mode | |
7094 | is smaller than word size, then expand_atomic_store assumes that the store | |
7095 | is atomic. We could avoid the builtin entirely in this case. */ | |
7096 | ||
7097 | tmpbase = (exchange ? BUILT_IN_ATOMIC_EXCHANGE_N : BUILT_IN_ATOMIC_STORE_N); | |
7098 | tmpbase = (enum built_in_function) ((int) tmpbase + index + 1); | |
7099 | decl = builtin_decl_explicit (tmpbase); | |
7100 | if (decl == NULL_TREE) | |
7101 | return false; | |
7102 | ||
7103 | type = TREE_TYPE (stored_val); | |
7104 | ||
7105 | /* Dig out the type of the function's second argument. */ | |
7106 | itype = TREE_TYPE (decl); | |
7107 | itype = TYPE_ARG_TYPES (itype); | |
7108 | itype = TREE_CHAIN (itype); | |
7109 | itype = TREE_VALUE (itype); | |
7110 | imode = TYPE_MODE (itype); | |
7111 | ||
7112 | if (exchange && !can_atomic_exchange_p (imode, true)) | |
7113 | return false; | |
7114 | ||
7115 | if (!useless_type_conversion_p (itype, type)) | |
7116 | stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val); | |
28567c40 JJ |
7117 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (stmt); |
7118 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (omo)); | |
7119 | call = build_call_expr_loc (loc, decl, 3, addr, stored_val, mo); | |
629b3d75 MJ |
7120 | if (exchange) |
7121 | { | |
7122 | if (!useless_type_conversion_p (type, itype)) | |
7123 | call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call); | |
7124 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call); | |
7125 | } | |
7126 | ||
7127 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
7128 | gsi_remove (&gsi, true); | |
7129 | ||
7130 | /* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */ | |
65f4b875 | 7131 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
7132 | gsi_remove (&gsi, true); |
7133 | ||
7134 | if (gimple_in_ssa_p (cfun)) | |
7135 | update_ssa (TODO_update_ssa_no_phi); | |
7136 | ||
7137 | return true; | |
7138 | } | |
7139 | ||
7140 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic | |
7141 | operation as a __atomic_fetch_op builtin. INDEX is log2 of the | |
7142 | size of the data type, and thus usable to find the index of the builtin | |
7143 | decl. Returns false if the expression is not of the proper form. */ | |
7144 | ||
7145 | static bool | |
7146 | expand_omp_atomic_fetch_op (basic_block load_bb, | |
7147 | tree addr, tree loaded_val, | |
7148 | tree stored_val, int index) | |
7149 | { | |
7150 | enum built_in_function oldbase, newbase, tmpbase; | |
7151 | tree decl, itype, call; | |
7152 | tree lhs, rhs; | |
7153 | basic_block store_bb = single_succ (load_bb); | |
7154 | gimple_stmt_iterator gsi; | |
7155 | gimple *stmt; | |
7156 | location_t loc; | |
7157 | enum tree_code code; | |
7158 | bool need_old, need_new; | |
7159 | machine_mode imode; | |
629b3d75 MJ |
7160 | |
7161 | /* We expect to find the following sequences: | |
7162 | ||
7163 | load_bb: | |
7164 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) | |
7165 | ||
7166 | store_bb: | |
7167 | val = tmp OP something; (or: something OP tmp) | |
7168 | GIMPLE_OMP_STORE (val) | |
7169 | ||
7170 | ???FIXME: Allow a more flexible sequence. | |
7171 | Perhaps use data flow to pick the statements. | |
7172 | ||
7173 | */ | |
7174 | ||
7175 | gsi = gsi_after_labels (store_bb); | |
7176 | stmt = gsi_stmt (gsi); | |
65f4b875 AO |
7177 | if (is_gimple_debug (stmt)) |
7178 | { | |
7179 | gsi_next_nondebug (&gsi); | |
7180 | if (gsi_end_p (gsi)) | |
7181 | return false; | |
7182 | stmt = gsi_stmt (gsi); | |
7183 | } | |
629b3d75 MJ |
7184 | loc = gimple_location (stmt); |
7185 | if (!is_gimple_assign (stmt)) | |
7186 | return false; | |
65f4b875 | 7187 | gsi_next_nondebug (&gsi); |
629b3d75 MJ |
7188 | if (gimple_code (gsi_stmt (gsi)) != GIMPLE_OMP_ATOMIC_STORE) |
7189 | return false; | |
7190 | need_new = gimple_omp_atomic_need_value_p (gsi_stmt (gsi)); | |
7191 | need_old = gimple_omp_atomic_need_value_p (last_stmt (load_bb)); | |
28567c40 JJ |
7192 | enum omp_memory_order omo |
7193 | = gimple_omp_atomic_memory_order (last_stmt (load_bb)); | |
7194 | enum memmodel mo = omp_memory_order_to_memmodel (omo); | |
629b3d75 MJ |
7195 | gcc_checking_assert (!need_old || !need_new); |
7196 | ||
7197 | if (!operand_equal_p (gimple_assign_lhs (stmt), stored_val, 0)) | |
7198 | return false; | |
7199 | ||
7200 | /* Check for one of the supported fetch-op operations. */ | |
7201 | code = gimple_assign_rhs_code (stmt); | |
7202 | switch (code) | |
7203 | { | |
7204 | case PLUS_EXPR: | |
7205 | case POINTER_PLUS_EXPR: | |
7206 | oldbase = BUILT_IN_ATOMIC_FETCH_ADD_N; | |
7207 | newbase = BUILT_IN_ATOMIC_ADD_FETCH_N; | |
7208 | break; | |
7209 | case MINUS_EXPR: | |
7210 | oldbase = BUILT_IN_ATOMIC_FETCH_SUB_N; | |
7211 | newbase = BUILT_IN_ATOMIC_SUB_FETCH_N; | |
7212 | break; | |
7213 | case BIT_AND_EXPR: | |
7214 | oldbase = BUILT_IN_ATOMIC_FETCH_AND_N; | |
7215 | newbase = BUILT_IN_ATOMIC_AND_FETCH_N; | |
7216 | break; | |
7217 | case BIT_IOR_EXPR: | |
7218 | oldbase = BUILT_IN_ATOMIC_FETCH_OR_N; | |
7219 | newbase = BUILT_IN_ATOMIC_OR_FETCH_N; | |
7220 | break; | |
7221 | case BIT_XOR_EXPR: | |
7222 | oldbase = BUILT_IN_ATOMIC_FETCH_XOR_N; | |
7223 | newbase = BUILT_IN_ATOMIC_XOR_FETCH_N; | |
7224 | break; | |
7225 | default: | |
7226 | return false; | |
7227 | } | |
7228 | ||
7229 | /* Make sure the expression is of the proper form. */ | |
7230 | if (operand_equal_p (gimple_assign_rhs1 (stmt), loaded_val, 0)) | |
7231 | rhs = gimple_assign_rhs2 (stmt); | |
7232 | else if (commutative_tree_code (gimple_assign_rhs_code (stmt)) | |
7233 | && operand_equal_p (gimple_assign_rhs2 (stmt), loaded_val, 0)) | |
7234 | rhs = gimple_assign_rhs1 (stmt); | |
7235 | else | |
7236 | return false; | |
7237 | ||
7238 | tmpbase = ((enum built_in_function) | |
7239 | ((need_new ? newbase : oldbase) + index + 1)); | |
7240 | decl = builtin_decl_explicit (tmpbase); | |
7241 | if (decl == NULL_TREE) | |
7242 | return false; | |
7243 | itype = TREE_TYPE (TREE_TYPE (decl)); | |
7244 | imode = TYPE_MODE (itype); | |
7245 | ||
7246 | /* We could test all of the various optabs involved, but the fact of the | |
7247 | matter is that (with the exception of i486 vs i586 and xadd) all targets | |
7248 | that support any atomic operaton optab also implements compare-and-swap. | |
7249 | Let optabs.c take care of expanding any compare-and-swap loop. */ | |
dc06356a | 7250 | if (!can_compare_and_swap_p (imode, true) || !can_atomic_load_p (imode)) |
629b3d75 MJ |
7251 | return false; |
7252 | ||
65f4b875 | 7253 | gsi = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
7254 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_LOAD); |
7255 | ||
7256 | /* OpenMP does not imply any barrier-like semantics on its atomic ops. | |
7257 | It only requires that the operation happen atomically. Thus we can | |
7258 | use the RELAXED memory model. */ | |
7259 | call = build_call_expr_loc (loc, decl, 3, addr, | |
7260 | fold_convert_loc (loc, itype, rhs), | |
28567c40 | 7261 | build_int_cst (NULL, mo)); |
629b3d75 MJ |
7262 | |
7263 | if (need_old || need_new) | |
7264 | { | |
7265 | lhs = need_old ? loaded_val : stored_val; | |
7266 | call = fold_convert_loc (loc, TREE_TYPE (lhs), call); | |
7267 | call = build2_loc (loc, MODIFY_EXPR, void_type_node, lhs, call); | |
7268 | } | |
7269 | else | |
7270 | call = fold_convert_loc (loc, void_type_node, call); | |
7271 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); | |
7272 | gsi_remove (&gsi, true); | |
7273 | ||
65f4b875 | 7274 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
7275 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); |
7276 | gsi_remove (&gsi, true); | |
65f4b875 | 7277 | gsi = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
7278 | stmt = gsi_stmt (gsi); |
7279 | gsi_remove (&gsi, true); | |
7280 | ||
7281 | if (gimple_in_ssa_p (cfun)) | |
7282 | { | |
7283 | release_defs (stmt); | |
7284 | update_ssa (TODO_update_ssa_no_phi); | |
7285 | } | |
7286 | ||
7287 | return true; | |
7288 | } | |
7289 | ||
7290 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
7291 | ||
7292 | oldval = *addr; | |
7293 | repeat: | |
01914336 | 7294 | newval = rhs; // with oldval replacing *addr in rhs |
629b3d75 MJ |
7295 | oldval = __sync_val_compare_and_swap (addr, oldval, newval); |
7296 | if (oldval != newval) | |
7297 | goto repeat; | |
7298 | ||
7299 | INDEX is log2 of the size of the data type, and thus usable to find the | |
7300 | index of the builtin decl. */ | |
7301 | ||
7302 | static bool | |
7303 | expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb, | |
7304 | tree addr, tree loaded_val, tree stored_val, | |
7305 | int index) | |
7306 | { | |
7307 | tree loadedi, storedi, initial, new_storedi, old_vali; | |
b4e47472 | 7308 | tree type, itype, cmpxchg, iaddr, atype; |
629b3d75 MJ |
7309 | gimple_stmt_iterator si; |
7310 | basic_block loop_header = single_succ (load_bb); | |
7311 | gimple *phi, *stmt; | |
7312 | edge e; | |
7313 | enum built_in_function fncode; | |
7314 | ||
7315 | /* ??? We need a non-pointer interface to __atomic_compare_exchange in | |
7316 | order to use the RELAXED memory model effectively. */ | |
7317 | fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N | |
7318 | + index + 1); | |
7319 | cmpxchg = builtin_decl_explicit (fncode); | |
7320 | if (cmpxchg == NULL_TREE) | |
7321 | return false; | |
b4e47472 JJ |
7322 | type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); |
7323 | atype = type; | |
629b3d75 MJ |
7324 | itype = TREE_TYPE (TREE_TYPE (cmpxchg)); |
7325 | ||
dc06356a JJ |
7326 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true) |
7327 | || !can_atomic_load_p (TYPE_MODE (itype))) | |
629b3d75 MJ |
7328 | return false; |
7329 | ||
7330 | /* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */ | |
65f4b875 | 7331 | si = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
7332 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); |
7333 | ||
7334 | /* For floating-point values, we'll need to view-convert them to integers | |
7335 | so that we can perform the atomic compare and swap. Simplify the | |
7336 | following code by always setting up the "i"ntegral variables. */ | |
7337 | if (!INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type)) | |
7338 | { | |
7339 | tree iaddr_val; | |
7340 | ||
7341 | iaddr = create_tmp_reg (build_pointer_type_for_mode (itype, ptr_mode, | |
7342 | true)); | |
b4e47472 | 7343 | atype = itype; |
629b3d75 MJ |
7344 | iaddr_val |
7345 | = force_gimple_operand_gsi (&si, | |
7346 | fold_convert (TREE_TYPE (iaddr), addr), | |
7347 | false, NULL_TREE, true, GSI_SAME_STMT); | |
7348 | stmt = gimple_build_assign (iaddr, iaddr_val); | |
7349 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
7350 | loadedi = create_tmp_var (itype); | |
7351 | if (gimple_in_ssa_p (cfun)) | |
7352 | loadedi = make_ssa_name (loadedi); | |
7353 | } | |
7354 | else | |
7355 | { | |
7356 | iaddr = addr; | |
7357 | loadedi = loaded_val; | |
7358 | } | |
7359 | ||
7360 | fncode = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); | |
7361 | tree loaddecl = builtin_decl_explicit (fncode); | |
7362 | if (loaddecl) | |
7363 | initial | |
b4e47472 | 7364 | = fold_convert (atype, |
629b3d75 MJ |
7365 | build_call_expr (loaddecl, 2, iaddr, |
7366 | build_int_cst (NULL_TREE, | |
7367 | MEMMODEL_RELAXED))); | |
7368 | else | |
b4e47472 JJ |
7369 | { |
7370 | tree off | |
7371 | = build_int_cst (build_pointer_type_for_mode (atype, ptr_mode, | |
7372 | true), 0); | |
7373 | initial = build2 (MEM_REF, atype, iaddr, off); | |
7374 | } | |
629b3d75 MJ |
7375 | |
7376 | initial | |
7377 | = force_gimple_operand_gsi (&si, initial, true, NULL_TREE, true, | |
7378 | GSI_SAME_STMT); | |
7379 | ||
7380 | /* Move the value to the LOADEDI temporary. */ | |
7381 | if (gimple_in_ssa_p (cfun)) | |
7382 | { | |
7383 | gcc_assert (gimple_seq_empty_p (phi_nodes (loop_header))); | |
7384 | phi = create_phi_node (loadedi, loop_header); | |
7385 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)), | |
7386 | initial); | |
7387 | } | |
7388 | else | |
7389 | gsi_insert_before (&si, | |
7390 | gimple_build_assign (loadedi, initial), | |
7391 | GSI_SAME_STMT); | |
7392 | if (loadedi != loaded_val) | |
7393 | { | |
7394 | gimple_stmt_iterator gsi2; | |
7395 | tree x; | |
7396 | ||
7397 | x = build1 (VIEW_CONVERT_EXPR, type, loadedi); | |
7398 | gsi2 = gsi_start_bb (loop_header); | |
7399 | if (gimple_in_ssa_p (cfun)) | |
7400 | { | |
7401 | gassign *stmt; | |
7402 | x = force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
7403 | true, GSI_SAME_STMT); | |
7404 | stmt = gimple_build_assign (loaded_val, x); | |
7405 | gsi_insert_before (&gsi2, stmt, GSI_SAME_STMT); | |
7406 | } | |
7407 | else | |
7408 | { | |
7409 | x = build2 (MODIFY_EXPR, TREE_TYPE (loaded_val), loaded_val, x); | |
7410 | force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, | |
7411 | true, GSI_SAME_STMT); | |
7412 | } | |
7413 | } | |
7414 | gsi_remove (&si, true); | |
7415 | ||
65f4b875 | 7416 | si = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
7417 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); |
7418 | ||
7419 | if (iaddr == addr) | |
7420 | storedi = stored_val; | |
7421 | else | |
01914336 MJ |
7422 | storedi |
7423 | = force_gimple_operand_gsi (&si, | |
7424 | build1 (VIEW_CONVERT_EXPR, itype, | |
7425 | stored_val), true, NULL_TREE, true, | |
7426 | GSI_SAME_STMT); | |
629b3d75 MJ |
7427 | |
7428 | /* Build the compare&swap statement. */ | |
7429 | new_storedi = build_call_expr (cmpxchg, 3, iaddr, loadedi, storedi); | |
7430 | new_storedi = force_gimple_operand_gsi (&si, | |
7431 | fold_convert (TREE_TYPE (loadedi), | |
7432 | new_storedi), | |
7433 | true, NULL_TREE, | |
7434 | true, GSI_SAME_STMT); | |
7435 | ||
7436 | if (gimple_in_ssa_p (cfun)) | |
7437 | old_vali = loadedi; | |
7438 | else | |
7439 | { | |
7440 | old_vali = create_tmp_var (TREE_TYPE (loadedi)); | |
7441 | stmt = gimple_build_assign (old_vali, loadedi); | |
7442 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
7443 | ||
7444 | stmt = gimple_build_assign (loadedi, new_storedi); | |
7445 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); | |
7446 | } | |
7447 | ||
7448 | /* Note that we always perform the comparison as an integer, even for | |
7449 | floating point. This allows the atomic operation to properly | |
7450 | succeed even with NaNs and -0.0. */ | |
01914336 MJ |
7451 | tree ne = build2 (NE_EXPR, boolean_type_node, new_storedi, old_vali); |
7452 | stmt = gimple_build_cond_empty (ne); | |
629b3d75 MJ |
7453 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
7454 | ||
7455 | /* Update cfg. */ | |
7456 | e = single_succ_edge (store_bb); | |
7457 | e->flags &= ~EDGE_FALLTHRU; | |
7458 | e->flags |= EDGE_FALSE_VALUE; | |
357067f2 JH |
7459 | /* Expect no looping. */ |
7460 | e->probability = profile_probability::guessed_always (); | |
629b3d75 MJ |
7461 | |
7462 | e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE); | |
357067f2 | 7463 | e->probability = profile_probability::guessed_never (); |
629b3d75 MJ |
7464 | |
7465 | /* Copy the new value to loadedi (we already did that before the condition | |
7466 | if we are not in SSA). */ | |
7467 | if (gimple_in_ssa_p (cfun)) | |
7468 | { | |
7469 | phi = gimple_seq_first_stmt (phi_nodes (loop_header)); | |
7470 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_storedi); | |
7471 | } | |
7472 | ||
7473 | /* Remove GIMPLE_OMP_ATOMIC_STORE. */ | |
7474 | gsi_remove (&si, true); | |
7475 | ||
99b1c316 | 7476 | class loop *loop = alloc_loop (); |
629b3d75 MJ |
7477 | loop->header = loop_header; |
7478 | loop->latch = store_bb; | |
7479 | add_loop (loop, loop_header->loop_father); | |
7480 | ||
7481 | if (gimple_in_ssa_p (cfun)) | |
7482 | update_ssa (TODO_update_ssa_no_phi); | |
7483 | ||
7484 | return true; | |
7485 | } | |
7486 | ||
7487 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: | |
7488 | ||
01914336 MJ |
7489 | GOMP_atomic_start (); |
7490 | *addr = rhs; | |
7491 | GOMP_atomic_end (); | |
629b3d75 MJ |
7492 | |
7493 | The result is not globally atomic, but works so long as all parallel | |
7494 | references are within #pragma omp atomic directives. According to | |
7495 | responses received from omp@openmp.org, appears to be within spec. | |
7496 | Which makes sense, since that's how several other compilers handle | |
7497 | this situation as well. | |
7498 | LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're | |
7499 | expanding. STORED_VAL is the operand of the matching | |
7500 | GIMPLE_OMP_ATOMIC_STORE. | |
7501 | ||
7502 | We replace | |
7503 | GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with | |
7504 | loaded_val = *addr; | |
7505 | ||
7506 | and replace | |
7507 | GIMPLE_OMP_ATOMIC_STORE (stored_val) with | |
7508 | *addr = stored_val; | |
7509 | */ | |
7510 | ||
7511 | static bool | |
7512 | expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb, | |
7513 | tree addr, tree loaded_val, tree stored_val) | |
7514 | { | |
7515 | gimple_stmt_iterator si; | |
7516 | gassign *stmt; | |
7517 | tree t; | |
7518 | ||
65f4b875 | 7519 | si = gsi_last_nondebug_bb (load_bb); |
629b3d75 MJ |
7520 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); |
7521 | ||
7522 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START); | |
7523 | t = build_call_expr (t, 0); | |
7524 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
7525 | ||
b4e47472 JJ |
7526 | tree mem = build_simple_mem_ref (addr); |
7527 | TREE_TYPE (mem) = TREE_TYPE (loaded_val); | |
7528 | TREE_OPERAND (mem, 1) | |
7529 | = fold_convert (build_pointer_type_for_mode (TREE_TYPE (mem), ptr_mode, | |
7530 | true), | |
7531 | TREE_OPERAND (mem, 1)); | |
7532 | stmt = gimple_build_assign (loaded_val, mem); | |
629b3d75 MJ |
7533 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
7534 | gsi_remove (&si, true); | |
7535 | ||
65f4b875 | 7536 | si = gsi_last_nondebug_bb (store_bb); |
629b3d75 MJ |
7537 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); |
7538 | ||
b4e47472 | 7539 | stmt = gimple_build_assign (unshare_expr (mem), stored_val); |
629b3d75 MJ |
7540 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
7541 | ||
7542 | t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END); | |
7543 | t = build_call_expr (t, 0); | |
7544 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); | |
7545 | gsi_remove (&si, true); | |
7546 | ||
7547 | if (gimple_in_ssa_p (cfun)) | |
7548 | update_ssa (TODO_update_ssa_no_phi); | |
7549 | return true; | |
7550 | } | |
7551 | ||
7552 | /* Expand an GIMPLE_OMP_ATOMIC statement. We try to expand | |
01914336 | 7553 | using expand_omp_atomic_fetch_op. If it failed, we try to |
629b3d75 MJ |
7554 | call expand_omp_atomic_pipeline, and if it fails too, the |
7555 | ultimate fallback is wrapping the operation in a mutex | |
7556 | (expand_omp_atomic_mutex). REGION is the atomic region built | |
7557 | by build_omp_regions_1(). */ | |
7558 | ||
7559 | static void | |
7560 | expand_omp_atomic (struct omp_region *region) | |
7561 | { | |
7562 | basic_block load_bb = region->entry, store_bb = region->exit; | |
7563 | gomp_atomic_load *load = as_a <gomp_atomic_load *> (last_stmt (load_bb)); | |
7564 | gomp_atomic_store *store = as_a <gomp_atomic_store *> (last_stmt (store_bb)); | |
7565 | tree loaded_val = gimple_omp_atomic_load_lhs (load); | |
7566 | tree addr = gimple_omp_atomic_load_rhs (load); | |
7567 | tree stored_val = gimple_omp_atomic_store_val (store); | |
b4e47472 | 7568 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); |
629b3d75 MJ |
7569 | HOST_WIDE_INT index; |
7570 | ||
7571 | /* Make sure the type is one of the supported sizes. */ | |
7572 | index = tree_to_uhwi (TYPE_SIZE_UNIT (type)); | |
7573 | index = exact_log2 (index); | |
7574 | if (index >= 0 && index <= 4) | |
7575 | { | |
7576 | unsigned int align = TYPE_ALIGN_UNIT (type); | |
7577 | ||
7578 | /* __sync builtins require strict data alignment. */ | |
7579 | if (exact_log2 (align) >= index) | |
7580 | { | |
7581 | /* Atomic load. */ | |
3bd8f481 | 7582 | scalar_mode smode; |
629b3d75 | 7583 | if (loaded_val == stored_val |
3bd8f481 RS |
7584 | && (is_int_mode (TYPE_MODE (type), &smode) |
7585 | || is_float_mode (TYPE_MODE (type), &smode)) | |
7586 | && GET_MODE_BITSIZE (smode) <= BITS_PER_WORD | |
629b3d75 MJ |
7587 | && expand_omp_atomic_load (load_bb, addr, loaded_val, index)) |
7588 | return; | |
7589 | ||
7590 | /* Atomic store. */ | |
3bd8f481 RS |
7591 | if ((is_int_mode (TYPE_MODE (type), &smode) |
7592 | || is_float_mode (TYPE_MODE (type), &smode)) | |
7593 | && GET_MODE_BITSIZE (smode) <= BITS_PER_WORD | |
629b3d75 MJ |
7594 | && store_bb == single_succ (load_bb) |
7595 | && first_stmt (store_bb) == store | |
7596 | && expand_omp_atomic_store (load_bb, addr, loaded_val, | |
7597 | stored_val, index)) | |
7598 | return; | |
7599 | ||
7600 | /* When possible, use specialized atomic update functions. */ | |
7601 | if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type)) | |
7602 | && store_bb == single_succ (load_bb) | |
7603 | && expand_omp_atomic_fetch_op (load_bb, addr, | |
7604 | loaded_val, stored_val, index)) | |
7605 | return; | |
7606 | ||
7607 | /* If we don't have specialized __sync builtins, try and implement | |
7608 | as a compare and swap loop. */ | |
7609 | if (expand_omp_atomic_pipeline (load_bb, store_bb, addr, | |
7610 | loaded_val, stored_val, index)) | |
7611 | return; | |
7612 | } | |
7613 | } | |
7614 | ||
7615 | /* The ultimate fallback is wrapping the operation in a mutex. */ | |
7616 | expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val); | |
7617 | } | |
7618 | ||
7619 | /* Mark the loops inside the kernels region starting at REGION_ENTRY and ending | |
7620 | at REGION_EXIT. */ | |
7621 | ||
7622 | static void | |
7623 | mark_loops_in_oacc_kernels_region (basic_block region_entry, | |
7624 | basic_block region_exit) | |
7625 | { | |
99b1c316 | 7626 | class loop *outer = region_entry->loop_father; |
629b3d75 MJ |
7627 | gcc_assert (region_exit == NULL || outer == region_exit->loop_father); |
7628 | ||
7629 | /* Don't parallelize the kernels region if it contains more than one outer | |
7630 | loop. */ | |
7631 | unsigned int nr_outer_loops = 0; | |
99b1c316 MS |
7632 | class loop *single_outer = NULL; |
7633 | for (class loop *loop = outer->inner; loop != NULL; loop = loop->next) | |
629b3d75 MJ |
7634 | { |
7635 | gcc_assert (loop_outer (loop) == outer); | |
7636 | ||
7637 | if (!dominated_by_p (CDI_DOMINATORS, loop->header, region_entry)) | |
7638 | continue; | |
7639 | ||
7640 | if (region_exit != NULL | |
7641 | && dominated_by_p (CDI_DOMINATORS, loop->header, region_exit)) | |
7642 | continue; | |
7643 | ||
7644 | nr_outer_loops++; | |
7645 | single_outer = loop; | |
7646 | } | |
7647 | if (nr_outer_loops != 1) | |
7648 | return; | |
7649 | ||
99b1c316 | 7650 | for (class loop *loop = single_outer->inner; |
01914336 MJ |
7651 | loop != NULL; |
7652 | loop = loop->inner) | |
629b3d75 MJ |
7653 | if (loop->next) |
7654 | return; | |
7655 | ||
7656 | /* Mark the loops in the region. */ | |
99b1c316 | 7657 | for (class loop *loop = single_outer; loop != NULL; loop = loop->inner) |
629b3d75 MJ |
7658 | loop->in_oacc_kernels_region = true; |
7659 | } | |
7660 | ||
7661 | /* Types used to pass grid and wortkgroup sizes to kernel invocation. */ | |
7662 | ||
7663 | struct GTY(()) grid_launch_attributes_trees | |
7664 | { | |
7665 | tree kernel_dim_array_type; | |
7666 | tree kernel_lattrs_dimnum_decl; | |
7667 | tree kernel_lattrs_grid_decl; | |
7668 | tree kernel_lattrs_group_decl; | |
7669 | tree kernel_launch_attributes_type; | |
7670 | }; | |
7671 | ||
7672 | static GTY(()) struct grid_launch_attributes_trees *grid_attr_trees; | |
7673 | ||
7674 | /* Create types used to pass kernel launch attributes to target. */ | |
7675 | ||
7676 | static void | |
7677 | grid_create_kernel_launch_attr_types (void) | |
7678 | { | |
7679 | if (grid_attr_trees) | |
7680 | return; | |
7681 | grid_attr_trees = ggc_alloc <grid_launch_attributes_trees> (); | |
7682 | ||
7683 | tree dim_arr_index_type | |
7684 | = build_index_type (build_int_cst (integer_type_node, 2)); | |
7685 | grid_attr_trees->kernel_dim_array_type | |
7686 | = build_array_type (uint32_type_node, dim_arr_index_type); | |
7687 | ||
7688 | grid_attr_trees->kernel_launch_attributes_type = make_node (RECORD_TYPE); | |
7689 | grid_attr_trees->kernel_lattrs_dimnum_decl | |
7690 | = build_decl (BUILTINS_LOCATION, FIELD_DECL, get_identifier ("ndim"), | |
7691 | uint32_type_node); | |
7692 | DECL_CHAIN (grid_attr_trees->kernel_lattrs_dimnum_decl) = NULL_TREE; | |
7693 | ||
7694 | grid_attr_trees->kernel_lattrs_grid_decl | |
7695 | = build_decl (BUILTINS_LOCATION, FIELD_DECL, get_identifier ("grid_size"), | |
7696 | grid_attr_trees->kernel_dim_array_type); | |
7697 | DECL_CHAIN (grid_attr_trees->kernel_lattrs_grid_decl) | |
7698 | = grid_attr_trees->kernel_lattrs_dimnum_decl; | |
7699 | grid_attr_trees->kernel_lattrs_group_decl | |
7700 | = build_decl (BUILTINS_LOCATION, FIELD_DECL, get_identifier ("group_size"), | |
7701 | grid_attr_trees->kernel_dim_array_type); | |
7702 | DECL_CHAIN (grid_attr_trees->kernel_lattrs_group_decl) | |
7703 | = grid_attr_trees->kernel_lattrs_grid_decl; | |
7704 | finish_builtin_struct (grid_attr_trees->kernel_launch_attributes_type, | |
7705 | "__gomp_kernel_launch_attributes", | |
7706 | grid_attr_trees->kernel_lattrs_group_decl, NULL_TREE); | |
7707 | } | |
7708 | ||
7709 | /* Insert before the current statement in GSI a store of VALUE to INDEX of | |
7710 | array (of type kernel_dim_array_type) FLD_DECL of RANGE_VAR. VALUE must be | |
7711 | of type uint32_type_node. */ | |
7712 | ||
7713 | static void | |
7714 | grid_insert_store_range_dim (gimple_stmt_iterator *gsi, tree range_var, | |
7715 | tree fld_decl, int index, tree value) | |
7716 | { | |
7717 | tree ref = build4 (ARRAY_REF, uint32_type_node, | |
7718 | build3 (COMPONENT_REF, | |
7719 | grid_attr_trees->kernel_dim_array_type, | |
7720 | range_var, fld_decl, NULL_TREE), | |
7721 | build_int_cst (integer_type_node, index), | |
7722 | NULL_TREE, NULL_TREE); | |
7723 | gsi_insert_before (gsi, gimple_build_assign (ref, value), GSI_SAME_STMT); | |
7724 | } | |
7725 | ||
7726 | /* Return a tree representation of a pointer to a structure with grid and | |
7727 | work-group size information. Statements filling that information will be | |
7728 | inserted before GSI, TGT_STMT is the target statement which has the | |
7729 | necessary information in it. */ | |
7730 | ||
7731 | static tree | |
7732 | grid_get_kernel_launch_attributes (gimple_stmt_iterator *gsi, | |
7733 | gomp_target *tgt_stmt) | |
7734 | { | |
7735 | grid_create_kernel_launch_attr_types (); | |
7736 | tree lattrs = create_tmp_var (grid_attr_trees->kernel_launch_attributes_type, | |
7737 | "__kernel_launch_attrs"); | |
7738 | ||
7739 | unsigned max_dim = 0; | |
7740 | for (tree clause = gimple_omp_target_clauses (tgt_stmt); | |
7741 | clause; | |
7742 | clause = OMP_CLAUSE_CHAIN (clause)) | |
7743 | { | |
7744 | if (OMP_CLAUSE_CODE (clause) != OMP_CLAUSE__GRIDDIM_) | |
7745 | continue; | |
7746 | ||
7747 | unsigned dim = OMP_CLAUSE__GRIDDIM__DIMENSION (clause); | |
7748 | max_dim = MAX (dim, max_dim); | |
7749 | ||
7750 | grid_insert_store_range_dim (gsi, lattrs, | |
7751 | grid_attr_trees->kernel_lattrs_grid_decl, | |
7752 | dim, OMP_CLAUSE__GRIDDIM__SIZE (clause)); | |
7753 | grid_insert_store_range_dim (gsi, lattrs, | |
7754 | grid_attr_trees->kernel_lattrs_group_decl, | |
7755 | dim, OMP_CLAUSE__GRIDDIM__GROUP (clause)); | |
7756 | } | |
7757 | ||
7758 | tree dimref = build3 (COMPONENT_REF, uint32_type_node, lattrs, | |
7759 | grid_attr_trees->kernel_lattrs_dimnum_decl, NULL_TREE); | |
7760 | gcc_checking_assert (max_dim <= 2); | |
7761 | tree dimensions = build_int_cstu (uint32_type_node, max_dim + 1); | |
7762 | gsi_insert_before (gsi, gimple_build_assign (dimref, dimensions), | |
7763 | GSI_SAME_STMT); | |
7764 | TREE_ADDRESSABLE (lattrs) = 1; | |
7765 | return build_fold_addr_expr (lattrs); | |
7766 | } | |
7767 | ||
7768 | /* Build target argument identifier from the DEVICE identifier, value | |
7769 | identifier ID and whether the element also has a SUBSEQUENT_PARAM. */ | |
7770 | ||
7771 | static tree | |
7772 | get_target_argument_identifier_1 (int device, bool subseqent_param, int id) | |
7773 | { | |
7774 | tree t = build_int_cst (integer_type_node, device); | |
7775 | if (subseqent_param) | |
7776 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
7777 | build_int_cst (integer_type_node, | |
7778 | GOMP_TARGET_ARG_SUBSEQUENT_PARAM)); | |
7779 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
7780 | build_int_cst (integer_type_node, id)); | |
7781 | return t; | |
7782 | } | |
7783 | ||
7784 | /* Like above but return it in type that can be directly stored as an element | |
7785 | of the argument array. */ | |
7786 | ||
7787 | static tree | |
7788 | get_target_argument_identifier (int device, bool subseqent_param, int id) | |
7789 | { | |
7790 | tree t = get_target_argument_identifier_1 (device, subseqent_param, id); | |
7791 | return fold_convert (ptr_type_node, t); | |
7792 | } | |
7793 | ||
7794 | /* Return a target argument consisting of DEVICE identifier, value identifier | |
7795 | ID, and the actual VALUE. */ | |
7796 | ||
7797 | static tree | |
7798 | get_target_argument_value (gimple_stmt_iterator *gsi, int device, int id, | |
7799 | tree value) | |
7800 | { | |
7801 | tree t = fold_build2 (LSHIFT_EXPR, integer_type_node, | |
7802 | fold_convert (integer_type_node, value), | |
7803 | build_int_cst (unsigned_type_node, | |
7804 | GOMP_TARGET_ARG_VALUE_SHIFT)); | |
7805 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, | |
7806 | get_target_argument_identifier_1 (device, false, id)); | |
7807 | t = fold_convert (ptr_type_node, t); | |
7808 | return force_gimple_operand_gsi (gsi, t, true, NULL, true, GSI_SAME_STMT); | |
7809 | } | |
7810 | ||
7811 | /* If VALUE is an integer constant greater than -2^15 and smaller than 2^15, | |
7812 | push one argument to ARGS with both the DEVICE, ID and VALUE embedded in it, | |
7813 | otherwise push an identifier (with DEVICE and ID) and the VALUE in two | |
7814 | arguments. */ | |
7815 | ||
7816 | static void | |
7817 | push_target_argument_according_to_value (gimple_stmt_iterator *gsi, int device, | |
7818 | int id, tree value, vec <tree> *args) | |
7819 | { | |
7820 | if (tree_fits_shwi_p (value) | |
7821 | && tree_to_shwi (value) > -(1 << 15) | |
7822 | && tree_to_shwi (value) < (1 << 15)) | |
7823 | args->quick_push (get_target_argument_value (gsi, device, id, value)); | |
7824 | else | |
7825 | { | |
7826 | args->quick_push (get_target_argument_identifier (device, true, id)); | |
7827 | value = fold_convert (ptr_type_node, value); | |
7828 | value = force_gimple_operand_gsi (gsi, value, true, NULL, true, | |
7829 | GSI_SAME_STMT); | |
7830 | args->quick_push (value); | |
7831 | } | |
7832 | } | |
7833 | ||
01914336 | 7834 | /* Create an array of arguments that is then passed to GOMP_target. */ |
629b3d75 MJ |
7835 | |
7836 | static tree | |
7837 | get_target_arguments (gimple_stmt_iterator *gsi, gomp_target *tgt_stmt) | |
7838 | { | |
7839 | auto_vec <tree, 6> args; | |
7840 | tree clauses = gimple_omp_target_clauses (tgt_stmt); | |
7841 | tree t, c = omp_find_clause (clauses, OMP_CLAUSE_NUM_TEAMS); | |
7842 | if (c) | |
7843 | t = OMP_CLAUSE_NUM_TEAMS_EXPR (c); | |
7844 | else | |
7845 | t = integer_minus_one_node; | |
7846 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, | |
7847 | GOMP_TARGET_ARG_NUM_TEAMS, t, &args); | |
7848 | ||
7849 | c = omp_find_clause (clauses, OMP_CLAUSE_THREAD_LIMIT); | |
7850 | if (c) | |
7851 | t = OMP_CLAUSE_THREAD_LIMIT_EXPR (c); | |
7852 | else | |
7853 | t = integer_minus_one_node; | |
7854 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, | |
7855 | GOMP_TARGET_ARG_THREAD_LIMIT, t, | |
7856 | &args); | |
7857 | ||
7858 | /* Add HSA-specific grid sizes, if available. */ | |
7859 | if (omp_find_clause (gimple_omp_target_clauses (tgt_stmt), | |
7860 | OMP_CLAUSE__GRIDDIM_)) | |
7861 | { | |
01914336 MJ |
7862 | int id = GOMP_TARGET_ARG_HSA_KERNEL_ATTRIBUTES; |
7863 | t = get_target_argument_identifier (GOMP_DEVICE_HSA, true, id); | |
629b3d75 MJ |
7864 | args.quick_push (t); |
7865 | args.quick_push (grid_get_kernel_launch_attributes (gsi, tgt_stmt)); | |
7866 | } | |
7867 | ||
7868 | /* Produce more, perhaps device specific, arguments here. */ | |
7869 | ||
7870 | tree argarray = create_tmp_var (build_array_type_nelts (ptr_type_node, | |
7871 | args.length () + 1), | |
7872 | ".omp_target_args"); | |
7873 | for (unsigned i = 0; i < args.length (); i++) | |
7874 | { | |
7875 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, | |
7876 | build_int_cst (integer_type_node, i), | |
7877 | NULL_TREE, NULL_TREE); | |
7878 | gsi_insert_before (gsi, gimple_build_assign (ref, args[i]), | |
7879 | GSI_SAME_STMT); | |
7880 | } | |
7881 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, | |
7882 | build_int_cst (integer_type_node, args.length ()), | |
7883 | NULL_TREE, NULL_TREE); | |
7884 | gsi_insert_before (gsi, gimple_build_assign (ref, null_pointer_node), | |
7885 | GSI_SAME_STMT); | |
7886 | TREE_ADDRESSABLE (argarray) = 1; | |
7887 | return build_fold_addr_expr (argarray); | |
7888 | } | |
7889 | ||
7890 | /* Expand the GIMPLE_OMP_TARGET starting at REGION. */ | |
7891 | ||
7892 | static void | |
7893 | expand_omp_target (struct omp_region *region) | |
7894 | { | |
7895 | basic_block entry_bb, exit_bb, new_bb; | |
7896 | struct function *child_cfun; | |
7897 | tree child_fn, block, t; | |
7898 | gimple_stmt_iterator gsi; | |
7899 | gomp_target *entry_stmt; | |
7900 | gimple *stmt; | |
7901 | edge e; | |
7902 | bool offloaded, data_region; | |
7903 | ||
7904 | entry_stmt = as_a <gomp_target *> (last_stmt (region->entry)); | |
7905 | new_bb = region->entry; | |
7906 | ||
7907 | offloaded = is_gimple_omp_offloaded (entry_stmt); | |
7908 | switch (gimple_omp_target_kind (entry_stmt)) | |
7909 | { | |
7910 | case GF_OMP_TARGET_KIND_REGION: | |
7911 | case GF_OMP_TARGET_KIND_UPDATE: | |
7912 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
7913 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
7914 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
7915 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
7916 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
7917 | case GF_OMP_TARGET_KIND_OACC_ENTER_EXIT_DATA: | |
7918 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
7919 | data_region = false; | |
7920 | break; | |
7921 | case GF_OMP_TARGET_KIND_DATA: | |
7922 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
7923 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
7924 | data_region = true; | |
7925 | break; | |
7926 | default: | |
7927 | gcc_unreachable (); | |
7928 | } | |
7929 | ||
7930 | child_fn = NULL_TREE; | |
7931 | child_cfun = NULL; | |
7932 | if (offloaded) | |
7933 | { | |
7934 | child_fn = gimple_omp_target_child_fn (entry_stmt); | |
7935 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); | |
7936 | } | |
7937 | ||
7938 | /* Supported by expand_omp_taskreg, but not here. */ | |
7939 | if (child_cfun != NULL) | |
7940 | gcc_checking_assert (!child_cfun->cfg); | |
7941 | gcc_checking_assert (!gimple_in_ssa_p (cfun)); | |
7942 | ||
7943 | entry_bb = region->entry; | |
7944 | exit_bb = region->exit; | |
7945 | ||
7946 | if (gimple_omp_target_kind (entry_stmt) == GF_OMP_TARGET_KIND_OACC_KERNELS) | |
25651634 TS |
7947 | { |
7948 | mark_loops_in_oacc_kernels_region (region->entry, region->exit); | |
7949 | ||
7950 | /* Further down, both OpenACC kernels and OpenACC parallel constructs | |
7951 | will be mappted to BUILT_IN_GOACC_PARALLEL, and to distinguish the | |
7952 | two, there is an "oacc kernels" attribute set for OpenACC kernels. */ | |
7953 | DECL_ATTRIBUTES (child_fn) | |
7954 | = tree_cons (get_identifier ("oacc kernels"), | |
7955 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); | |
7956 | } | |
629b3d75 MJ |
7957 | |
7958 | if (offloaded) | |
7959 | { | |
7960 | unsigned srcidx, dstidx, num; | |
7961 | ||
7962 | /* If the offloading region needs data sent from the parent | |
7963 | function, then the very first statement (except possible | |
7964 | tree profile counter updates) of the offloading body | |
7965 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since | |
7966 | &.OMP_DATA_O is passed as an argument to the child function, | |
7967 | we need to replace it with the argument as seen by the child | |
7968 | function. | |
7969 | ||
7970 | In most cases, this will end up being the identity assignment | |
7971 | .OMP_DATA_I = .OMP_DATA_I. However, if the offloading body had | |
7972 | a function call that has been inlined, the original PARM_DECL | |
7973 | .OMP_DATA_I may have been converted into a different local | |
7974 | variable. In which case, we need to keep the assignment. */ | |
7975 | tree data_arg = gimple_omp_target_data_arg (entry_stmt); | |
7976 | if (data_arg) | |
7977 | { | |
7978 | basic_block entry_succ_bb = single_succ (entry_bb); | |
7979 | gimple_stmt_iterator gsi; | |
7980 | tree arg; | |
7981 | gimple *tgtcopy_stmt = NULL; | |
7982 | tree sender = TREE_VEC_ELT (data_arg, 0); | |
7983 | ||
7984 | for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi)) | |
7985 | { | |
7986 | gcc_assert (!gsi_end_p (gsi)); | |
7987 | stmt = gsi_stmt (gsi); | |
7988 | if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
7989 | continue; | |
7990 | ||
7991 | if (gimple_num_ops (stmt) == 2) | |
7992 | { | |
7993 | tree arg = gimple_assign_rhs1 (stmt); | |
7994 | ||
7995 | /* We're ignoring the subcode because we're | |
7996 | effectively doing a STRIP_NOPS. */ | |
7997 | ||
7998 | if (TREE_CODE (arg) == ADDR_EXPR | |
7999 | && TREE_OPERAND (arg, 0) == sender) | |
8000 | { | |
8001 | tgtcopy_stmt = stmt; | |
8002 | break; | |
8003 | } | |
8004 | } | |
8005 | } | |
8006 | ||
8007 | gcc_assert (tgtcopy_stmt != NULL); | |
8008 | arg = DECL_ARGUMENTS (child_fn); | |
8009 | ||
8010 | gcc_assert (gimple_assign_lhs (tgtcopy_stmt) == arg); | |
8011 | gsi_remove (&gsi, true); | |
8012 | } | |
8013 | ||
8014 | /* Declare local variables needed in CHILD_CFUN. */ | |
8015 | block = DECL_INITIAL (child_fn); | |
8016 | BLOCK_VARS (block) = vec2chain (child_cfun->local_decls); | |
8017 | /* The gimplifier could record temporaries in the offloading block | |
8018 | rather than in containing function's local_decls chain, | |
8019 | which would mean cgraph missed finalizing them. Do it now. */ | |
8020 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) | |
8021 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) | |
8022 | varpool_node::finalize_decl (t); | |
8023 | DECL_SAVED_TREE (child_fn) = NULL; | |
8024 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ | |
8025 | gimple_set_body (child_fn, NULL); | |
8026 | TREE_USED (block) = 1; | |
8027 | ||
8028 | /* Reset DECL_CONTEXT on function arguments. */ | |
8029 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) | |
8030 | DECL_CONTEXT (t) = child_fn; | |
8031 | ||
8032 | /* Split ENTRY_BB at GIMPLE_*, | |
8033 | so that it can be moved to the child function. */ | |
65f4b875 | 8034 | gsi = gsi_last_nondebug_bb (entry_bb); |
629b3d75 MJ |
8035 | stmt = gsi_stmt (gsi); |
8036 | gcc_assert (stmt | |
8037 | && gimple_code (stmt) == gimple_code (entry_stmt)); | |
8038 | e = split_block (entry_bb, stmt); | |
8039 | gsi_remove (&gsi, true); | |
8040 | entry_bb = e->dest; | |
8041 | single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU; | |
8042 | ||
8043 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ | |
8044 | if (exit_bb) | |
8045 | { | |
65f4b875 | 8046 | gsi = gsi_last_nondebug_bb (exit_bb); |
629b3d75 MJ |
8047 | gcc_assert (!gsi_end_p (gsi) |
8048 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
8049 | stmt = gimple_build_return (NULL); | |
8050 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); | |
8051 | gsi_remove (&gsi, true); | |
8052 | } | |
8053 | ||
8054 | /* Move the offloading region into CHILD_CFUN. */ | |
8055 | ||
8056 | block = gimple_block (entry_stmt); | |
8057 | ||
8058 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); | |
8059 | if (exit_bb) | |
8060 | single_succ_edge (new_bb)->flags = EDGE_FALLTHRU; | |
8061 | /* When the OMP expansion process cannot guarantee an up-to-date | |
8062 | loop tree arrange for the child function to fixup loops. */ | |
8063 | if (loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
8064 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; | |
8065 | ||
8066 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ | |
8067 | num = vec_safe_length (child_cfun->local_decls); | |
8068 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) | |
8069 | { | |
8070 | t = (*child_cfun->local_decls)[srcidx]; | |
8071 | if (DECL_CONTEXT (t) == cfun->decl) | |
8072 | continue; | |
8073 | if (srcidx != dstidx) | |
8074 | (*child_cfun->local_decls)[dstidx] = t; | |
8075 | dstidx++; | |
8076 | } | |
8077 | if (dstidx != num) | |
8078 | vec_safe_truncate (child_cfun->local_decls, dstidx); | |
8079 | ||
8080 | /* Inform the callgraph about the new function. */ | |
8081 | child_cfun->curr_properties = cfun->curr_properties; | |
8082 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; | |
8083 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; | |
8084 | cgraph_node *node = cgraph_node::get_create (child_fn); | |
8085 | node->parallelized_function = 1; | |
8086 | cgraph_node::add_new_function (child_fn, true); | |
8087 | ||
8088 | /* Add the new function to the offload table. */ | |
8089 | if (ENABLE_OFFLOADING) | |
60bf575c TV |
8090 | { |
8091 | if (in_lto_p) | |
8092 | DECL_PRESERVE_P (child_fn) = 1; | |
8093 | vec_safe_push (offload_funcs, child_fn); | |
8094 | } | |
629b3d75 MJ |
8095 | |
8096 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) | |
8097 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); | |
8098 | ||
8099 | /* Fix the callgraph edges for child_cfun. Those for cfun will be | |
8100 | fixed in a following pass. */ | |
8101 | push_cfun (child_cfun); | |
8102 | if (need_asm) | |
9579db35 | 8103 | assign_assembler_name_if_needed (child_fn); |
629b3d75 MJ |
8104 | cgraph_edge::rebuild_edges (); |
8105 | ||
8106 | /* Some EH regions might become dead, see PR34608. If | |
8107 | pass_cleanup_cfg isn't the first pass to happen with the | |
8108 | new child, these dead EH edges might cause problems. | |
8109 | Clean them up now. */ | |
8110 | if (flag_exceptions) | |
8111 | { | |
8112 | basic_block bb; | |
8113 | bool changed = false; | |
8114 | ||
8115 | FOR_EACH_BB_FN (bb, cfun) | |
8116 | changed |= gimple_purge_dead_eh_edges (bb); | |
8117 | if (changed) | |
8118 | cleanup_tree_cfg (); | |
8119 | } | |
8120 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
8121 | verify_loop_structure (); | |
8122 | pop_cfun (); | |
8123 | ||
8124 | if (dump_file && !gimple_in_ssa_p (cfun)) | |
8125 | { | |
8126 | omp_any_child_fn_dumped = true; | |
8127 | dump_function_header (dump_file, child_fn, dump_flags); | |
8128 | dump_function_to_file (child_fn, dump_file, dump_flags); | |
8129 | } | |
4ccc4e30 JJ |
8130 | |
8131 | adjust_context_and_scope (region, gimple_block (entry_stmt), child_fn); | |
629b3d75 MJ |
8132 | } |
8133 | ||
8134 | /* Emit a library call to launch the offloading region, or do data | |
8135 | transfers. */ | |
59d5960c | 8136 | tree t1, t2, t3, t4, depend, c, clauses; |
629b3d75 | 8137 | enum built_in_function start_ix; |
629b3d75 | 8138 | unsigned int flags_i = 0; |
629b3d75 MJ |
8139 | |
8140 | switch (gimple_omp_target_kind (entry_stmt)) | |
8141 | { | |
8142 | case GF_OMP_TARGET_KIND_REGION: | |
8143 | start_ix = BUILT_IN_GOMP_TARGET; | |
8144 | break; | |
8145 | case GF_OMP_TARGET_KIND_DATA: | |
8146 | start_ix = BUILT_IN_GOMP_TARGET_DATA; | |
8147 | break; | |
8148 | case GF_OMP_TARGET_KIND_UPDATE: | |
8149 | start_ix = BUILT_IN_GOMP_TARGET_UPDATE; | |
8150 | break; | |
8151 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
8152 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; | |
8153 | break; | |
8154 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
8155 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; | |
8156 | flags_i |= GOMP_TARGET_FLAG_EXIT_DATA; | |
8157 | break; | |
8158 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
629b3d75 MJ |
8159 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
8160 | start_ix = BUILT_IN_GOACC_PARALLEL; | |
8161 | break; | |
8162 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
8163 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
8164 | start_ix = BUILT_IN_GOACC_DATA_START; | |
8165 | break; | |
8166 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
8167 | start_ix = BUILT_IN_GOACC_UPDATE; | |
8168 | break; | |
8169 | case GF_OMP_TARGET_KIND_OACC_ENTER_EXIT_DATA: | |
8170 | start_ix = BUILT_IN_GOACC_ENTER_EXIT_DATA; | |
8171 | break; | |
8172 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
8173 | start_ix = BUILT_IN_GOACC_DECLARE; | |
8174 | break; | |
8175 | default: | |
8176 | gcc_unreachable (); | |
8177 | } | |
8178 | ||
8179 | clauses = gimple_omp_target_clauses (entry_stmt); | |
8180 | ||
59d5960c TS |
8181 | tree device = NULL_TREE; |
8182 | location_t device_loc = UNKNOWN_LOCATION; | |
8183 | tree goacc_flags = NULL_TREE; | |
8184 | if (is_gimple_omp_oacc (entry_stmt)) | |
629b3d75 | 8185 | { |
59d5960c TS |
8186 | /* By default, no GOACC_FLAGs are set. */ |
8187 | goacc_flags = integer_zero_node; | |
629b3d75 MJ |
8188 | } |
8189 | else | |
59d5960c TS |
8190 | { |
8191 | c = omp_find_clause (clauses, OMP_CLAUSE_DEVICE); | |
8192 | if (c) | |
8193 | { | |
8194 | device = OMP_CLAUSE_DEVICE_ID (c); | |
8195 | device_loc = OMP_CLAUSE_LOCATION (c); | |
8196 | } | |
8197 | else | |
8198 | { | |
8199 | /* By default, the value of DEVICE is GOMP_DEVICE_ICV (let runtime | |
8200 | library choose). */ | |
8201 | device = build_int_cst (integer_type_node, GOMP_DEVICE_ICV); | |
8202 | device_loc = gimple_location (entry_stmt); | |
8203 | } | |
629b3d75 | 8204 | |
59d5960c TS |
8205 | c = omp_find_clause (clauses, OMP_CLAUSE_NOWAIT); |
8206 | if (c) | |
8207 | flags_i |= GOMP_TARGET_FLAG_NOWAIT; | |
8208 | } | |
629b3d75 | 8209 | |
59d5960c TS |
8210 | /* By default, there is no conditional. */ |
8211 | tree cond = NULL_TREE; | |
8212 | c = omp_find_clause (clauses, OMP_CLAUSE_IF); | |
8213 | if (c) | |
8214 | cond = OMP_CLAUSE_IF_EXPR (c); | |
8215 | /* If we found the clause 'if (cond)', build: | |
8216 | OpenACC: goacc_flags = (cond ? goacc_flags : flags | GOACC_FLAG_HOST_FALLBACK) | |
8217 | OpenMP: device = (cond ? device : GOMP_DEVICE_HOST_FALLBACK) */ | |
629b3d75 MJ |
8218 | if (cond) |
8219 | { | |
59d5960c TS |
8220 | tree *tp; |
8221 | if (is_gimple_omp_oacc (entry_stmt)) | |
8222 | tp = &goacc_flags; | |
8223 | else | |
8224 | { | |
8225 | /* Ensure 'device' is of the correct type. */ | |
8226 | device = fold_convert_loc (device_loc, integer_type_node, device); | |
8227 | ||
8228 | tp = &device; | |
8229 | } | |
8230 | ||
629b3d75 MJ |
8231 | cond = gimple_boolify (cond); |
8232 | ||
8233 | basic_block cond_bb, then_bb, else_bb; | |
8234 | edge e; | |
8235 | tree tmp_var; | |
8236 | ||
59d5960c | 8237 | tmp_var = create_tmp_var (TREE_TYPE (*tp)); |
629b3d75 MJ |
8238 | if (offloaded) |
8239 | e = split_block_after_labels (new_bb); | |
8240 | else | |
8241 | { | |
65f4b875 | 8242 | gsi = gsi_last_nondebug_bb (new_bb); |
629b3d75 MJ |
8243 | gsi_prev (&gsi); |
8244 | e = split_block (new_bb, gsi_stmt (gsi)); | |
8245 | } | |
8246 | cond_bb = e->src; | |
8247 | new_bb = e->dest; | |
8248 | remove_edge (e); | |
8249 | ||
8250 | then_bb = create_empty_bb (cond_bb); | |
8251 | else_bb = create_empty_bb (then_bb); | |
8252 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
8253 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); | |
8254 | ||
8255 | stmt = gimple_build_cond_empty (cond); | |
8256 | gsi = gsi_last_bb (cond_bb); | |
8257 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
8258 | ||
8259 | gsi = gsi_start_bb (then_bb); | |
59d5960c | 8260 | stmt = gimple_build_assign (tmp_var, *tp); |
629b3d75 MJ |
8261 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
8262 | ||
8263 | gsi = gsi_start_bb (else_bb); | |
59d5960c TS |
8264 | if (is_gimple_omp_oacc (entry_stmt)) |
8265 | stmt = gimple_build_assign (tmp_var, | |
8266 | BIT_IOR_EXPR, | |
8267 | *tp, | |
8268 | build_int_cst (integer_type_node, | |
8269 | GOACC_FLAG_HOST_FALLBACK)); | |
8270 | else | |
8271 | stmt = gimple_build_assign (tmp_var, | |
8272 | build_int_cst (integer_type_node, | |
8273 | GOMP_DEVICE_HOST_FALLBACK)); | |
629b3d75 MJ |
8274 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
8275 | ||
8276 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); | |
8277 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); | |
8278 | add_bb_to_loop (then_bb, cond_bb->loop_father); | |
8279 | add_bb_to_loop (else_bb, cond_bb->loop_father); | |
8280 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); | |
8281 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); | |
8282 | ||
59d5960c TS |
8283 | *tp = tmp_var; |
8284 | ||
65f4b875 | 8285 | gsi = gsi_last_nondebug_bb (new_bb); |
629b3d75 MJ |
8286 | } |
8287 | else | |
8288 | { | |
65f4b875 | 8289 | gsi = gsi_last_nondebug_bb (new_bb); |
59d5960c TS |
8290 | |
8291 | if (device != NULL_TREE) | |
8292 | device = force_gimple_operand_gsi (&gsi, device, true, NULL_TREE, | |
8293 | true, GSI_SAME_STMT); | |
629b3d75 MJ |
8294 | } |
8295 | ||
8296 | t = gimple_omp_target_data_arg (entry_stmt); | |
8297 | if (t == NULL) | |
8298 | { | |
8299 | t1 = size_zero_node; | |
8300 | t2 = build_zero_cst (ptr_type_node); | |
8301 | t3 = t2; | |
8302 | t4 = t2; | |
8303 | } | |
8304 | else | |
8305 | { | |
8306 | t1 = TYPE_MAX_VALUE (TYPE_DOMAIN (TREE_TYPE (TREE_VEC_ELT (t, 1)))); | |
8307 | t1 = size_binop (PLUS_EXPR, t1, size_int (1)); | |
8308 | t2 = build_fold_addr_expr (TREE_VEC_ELT (t, 0)); | |
8309 | t3 = build_fold_addr_expr (TREE_VEC_ELT (t, 1)); | |
8310 | t4 = build_fold_addr_expr (TREE_VEC_ELT (t, 2)); | |
8311 | } | |
8312 | ||
8313 | gimple *g; | |
8314 | bool tagging = false; | |
8315 | /* The maximum number used by any start_ix, without varargs. */ | |
8316 | auto_vec<tree, 11> args; | |
59d5960c TS |
8317 | if (is_gimple_omp_oacc (entry_stmt)) |
8318 | { | |
8319 | tree goacc_flags_m = fold_build1 (GOACC_FLAGS_MARSHAL_OP, | |
8320 | TREE_TYPE (goacc_flags), goacc_flags); | |
8321 | goacc_flags_m = force_gimple_operand_gsi (&gsi, goacc_flags_m, true, | |
8322 | NULL_TREE, true, | |
8323 | GSI_SAME_STMT); | |
8324 | args.quick_push (goacc_flags_m); | |
8325 | } | |
8326 | else | |
8327 | args.quick_push (device); | |
629b3d75 MJ |
8328 | if (offloaded) |
8329 | args.quick_push (build_fold_addr_expr (child_fn)); | |
8330 | args.quick_push (t1); | |
8331 | args.quick_push (t2); | |
8332 | args.quick_push (t3); | |
8333 | args.quick_push (t4); | |
8334 | switch (start_ix) | |
8335 | { | |
8336 | case BUILT_IN_GOACC_DATA_START: | |
8337 | case BUILT_IN_GOACC_DECLARE: | |
8338 | case BUILT_IN_GOMP_TARGET_DATA: | |
8339 | break; | |
8340 | case BUILT_IN_GOMP_TARGET: | |
8341 | case BUILT_IN_GOMP_TARGET_UPDATE: | |
8342 | case BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA: | |
8343 | args.quick_push (build_int_cst (unsigned_type_node, flags_i)); | |
8344 | c = omp_find_clause (clauses, OMP_CLAUSE_DEPEND); | |
8345 | if (c) | |
8346 | depend = OMP_CLAUSE_DECL (c); | |
8347 | else | |
8348 | depend = build_int_cst (ptr_type_node, 0); | |
8349 | args.quick_push (depend); | |
8350 | if (start_ix == BUILT_IN_GOMP_TARGET) | |
8351 | args.quick_push (get_target_arguments (&gsi, entry_stmt)); | |
8352 | break; | |
8353 | case BUILT_IN_GOACC_PARALLEL: | |
25651634 TS |
8354 | oacc_set_fn_attrib (child_fn, clauses, &args); |
8355 | tagging = true; | |
629b3d75 MJ |
8356 | /* FALLTHRU */ |
8357 | case BUILT_IN_GOACC_ENTER_EXIT_DATA: | |
8358 | case BUILT_IN_GOACC_UPDATE: | |
8359 | { | |
8360 | tree t_async = NULL_TREE; | |
8361 | ||
8362 | /* If present, use the value specified by the respective | |
8363 | clause, making sure that is of the correct type. */ | |
8364 | c = omp_find_clause (clauses, OMP_CLAUSE_ASYNC); | |
8365 | if (c) | |
8366 | t_async = fold_convert_loc (OMP_CLAUSE_LOCATION (c), | |
8367 | integer_type_node, | |
8368 | OMP_CLAUSE_ASYNC_EXPR (c)); | |
8369 | else if (!tagging) | |
8370 | /* Default values for t_async. */ | |
8371 | t_async = fold_convert_loc (gimple_location (entry_stmt), | |
8372 | integer_type_node, | |
8373 | build_int_cst (integer_type_node, | |
8374 | GOMP_ASYNC_SYNC)); | |
8375 | if (tagging && t_async) | |
8376 | { | |
8377 | unsigned HOST_WIDE_INT i_async = GOMP_LAUNCH_OP_MAX; | |
8378 | ||
8379 | if (TREE_CODE (t_async) == INTEGER_CST) | |
8380 | { | |
8381 | /* See if we can pack the async arg in to the tag's | |
8382 | operand. */ | |
8383 | i_async = TREE_INT_CST_LOW (t_async); | |
8384 | if (i_async < GOMP_LAUNCH_OP_MAX) | |
8385 | t_async = NULL_TREE; | |
8386 | else | |
8387 | i_async = GOMP_LAUNCH_OP_MAX; | |
8388 | } | |
8389 | args.safe_push (oacc_launch_pack (GOMP_LAUNCH_ASYNC, NULL_TREE, | |
8390 | i_async)); | |
8391 | } | |
8392 | if (t_async) | |
8393 | args.safe_push (t_async); | |
8394 | ||
8395 | /* Save the argument index, and ... */ | |
8396 | unsigned t_wait_idx = args.length (); | |
8397 | unsigned num_waits = 0; | |
8398 | c = omp_find_clause (clauses, OMP_CLAUSE_WAIT); | |
8399 | if (!tagging || c) | |
8400 | /* ... push a placeholder. */ | |
8401 | args.safe_push (integer_zero_node); | |
8402 | ||
8403 | for (; c; c = OMP_CLAUSE_CHAIN (c)) | |
8404 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_WAIT) | |
8405 | { | |
8406 | args.safe_push (fold_convert_loc (OMP_CLAUSE_LOCATION (c), | |
8407 | integer_type_node, | |
8408 | OMP_CLAUSE_WAIT_EXPR (c))); | |
8409 | num_waits++; | |
8410 | } | |
8411 | ||
8412 | if (!tagging || num_waits) | |
8413 | { | |
8414 | tree len; | |
8415 | ||
8416 | /* Now that we know the number, update the placeholder. */ | |
8417 | if (tagging) | |
8418 | len = oacc_launch_pack (GOMP_LAUNCH_WAIT, NULL_TREE, num_waits); | |
8419 | else | |
8420 | len = build_int_cst (integer_type_node, num_waits); | |
8421 | len = fold_convert_loc (gimple_location (entry_stmt), | |
8422 | unsigned_type_node, len); | |
8423 | args[t_wait_idx] = len; | |
8424 | } | |
8425 | } | |
8426 | break; | |
8427 | default: | |
8428 | gcc_unreachable (); | |
8429 | } | |
8430 | if (tagging) | |
8431 | /* Push terminal marker - zero. */ | |
8432 | args.safe_push (oacc_launch_pack (0, NULL_TREE, 0)); | |
8433 | ||
8434 | g = gimple_build_call_vec (builtin_decl_explicit (start_ix), args); | |
8435 | gimple_set_location (g, gimple_location (entry_stmt)); | |
8436 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
8437 | if (!offloaded) | |
8438 | { | |
8439 | g = gsi_stmt (gsi); | |
8440 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_TARGET); | |
8441 | gsi_remove (&gsi, true); | |
8442 | } | |
8443 | if (data_region && region->exit) | |
8444 | { | |
65f4b875 | 8445 | gsi = gsi_last_nondebug_bb (region->exit); |
629b3d75 MJ |
8446 | g = gsi_stmt (gsi); |
8447 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_RETURN); | |
8448 | gsi_remove (&gsi, true); | |
8449 | } | |
8450 | } | |
8451 | ||
8452 | /* Expand KFOR loop as a HSA grifidied kernel, i.e. as a body only with | |
8453 | iteration variable derived from the thread number. INTRA_GROUP means this | |
8454 | is an expansion of a loop iterating over work-items within a separate | |
01914336 | 8455 | iteration over groups. */ |
629b3d75 MJ |
8456 | |
8457 | static void | |
8458 | grid_expand_omp_for_loop (struct omp_region *kfor, bool intra_group) | |
8459 | { | |
8460 | gimple_stmt_iterator gsi; | |
8461 | gomp_for *for_stmt = as_a <gomp_for *> (last_stmt (kfor->entry)); | |
8462 | gcc_checking_assert (gimple_omp_for_kind (for_stmt) | |
8463 | == GF_OMP_FOR_KIND_GRID_LOOP); | |
8464 | size_t collapse = gimple_omp_for_collapse (for_stmt); | |
8465 | struct omp_for_data_loop *loops | |
8466 | = XALLOCAVEC (struct omp_for_data_loop, | |
01914336 | 8467 | gimple_omp_for_collapse (for_stmt)); |
629b3d75 MJ |
8468 | struct omp_for_data fd; |
8469 | ||
8470 | remove_edge (BRANCH_EDGE (kfor->entry)); | |
8471 | basic_block body_bb = FALLTHRU_EDGE (kfor->entry)->dest; | |
8472 | ||
8473 | gcc_assert (kfor->cont); | |
8474 | omp_extract_for_data (for_stmt, &fd, loops); | |
8475 | ||
8476 | gsi = gsi_start_bb (body_bb); | |
8477 | ||
8478 | for (size_t dim = 0; dim < collapse; dim++) | |
8479 | { | |
8480 | tree type, itype; | |
8481 | itype = type = TREE_TYPE (fd.loops[dim].v); | |
8482 | if (POINTER_TYPE_P (type)) | |
8483 | itype = signed_type_for (type); | |
8484 | ||
8485 | tree n1 = fd.loops[dim].n1; | |
8486 | tree step = fd.loops[dim].step; | |
8487 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), | |
8488 | true, NULL_TREE, true, GSI_SAME_STMT); | |
8489 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), | |
8490 | true, NULL_TREE, true, GSI_SAME_STMT); | |
8491 | tree threadid; | |
8492 | if (gimple_omp_for_grid_group_iter (for_stmt)) | |
8493 | { | |
8494 | gcc_checking_assert (!intra_group); | |
8495 | threadid = build_call_expr (builtin_decl_explicit | |
8496 | (BUILT_IN_HSA_WORKGROUPID), 1, | |
8497 | build_int_cstu (unsigned_type_node, dim)); | |
8498 | } | |
8499 | else if (intra_group) | |
8500 | threadid = build_call_expr (builtin_decl_explicit | |
8501 | (BUILT_IN_HSA_WORKITEMID), 1, | |
8502 | build_int_cstu (unsigned_type_node, dim)); | |
8503 | else | |
8504 | threadid = build_call_expr (builtin_decl_explicit | |
8505 | (BUILT_IN_HSA_WORKITEMABSID), 1, | |
8506 | build_int_cstu (unsigned_type_node, dim)); | |
8507 | threadid = fold_convert (itype, threadid); | |
8508 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, | |
8509 | true, GSI_SAME_STMT); | |
8510 | ||
8511 | tree startvar = fd.loops[dim].v; | |
8512 | tree t = fold_build2 (MULT_EXPR, itype, threadid, step); | |
8513 | if (POINTER_TYPE_P (type)) | |
8514 | t = fold_build_pointer_plus (n1, t); | |
8515 | else | |
8516 | t = fold_build2 (PLUS_EXPR, type, t, n1); | |
8517 | t = fold_convert (type, t); | |
8518 | t = force_gimple_operand_gsi (&gsi, t, | |
8519 | DECL_P (startvar) | |
8520 | && TREE_ADDRESSABLE (startvar), | |
8521 | NULL_TREE, true, GSI_SAME_STMT); | |
8522 | gassign *assign_stmt = gimple_build_assign (startvar, t); | |
8523 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
8524 | } | |
01914336 | 8525 | /* Remove the omp for statement. */ |
65f4b875 | 8526 | gsi = gsi_last_nondebug_bb (kfor->entry); |
629b3d75 MJ |
8527 | gsi_remove (&gsi, true); |
8528 | ||
8529 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ | |
65f4b875 | 8530 | gsi = gsi_last_nondebug_bb (kfor->cont); |
629b3d75 MJ |
8531 | gcc_assert (!gsi_end_p (gsi) |
8532 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_CONTINUE); | |
8533 | gsi_remove (&gsi, true); | |
8534 | ||
8535 | /* Replace the GIMPLE_OMP_RETURN with a barrier, if necessary. */ | |
65f4b875 | 8536 | gsi = gsi_last_nondebug_bb (kfor->exit); |
629b3d75 MJ |
8537 | gcc_assert (!gsi_end_p (gsi) |
8538 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
8539 | if (intra_group) | |
8540 | gsi_insert_before (&gsi, omp_build_barrier (NULL_TREE), GSI_SAME_STMT); | |
8541 | gsi_remove (&gsi, true); | |
8542 | ||
8543 | /* Fixup the much simpler CFG. */ | |
8544 | remove_edge (find_edge (kfor->cont, body_bb)); | |
8545 | ||
8546 | if (kfor->cont != body_bb) | |
8547 | set_immediate_dominator (CDI_DOMINATORS, kfor->cont, body_bb); | |
8548 | set_immediate_dominator (CDI_DOMINATORS, kfor->exit, kfor->cont); | |
8549 | } | |
8550 | ||
8551 | /* Structure passed to grid_remap_kernel_arg_accesses so that it can remap | |
8552 | argument_decls. */ | |
8553 | ||
8554 | struct grid_arg_decl_map | |
8555 | { | |
8556 | tree old_arg; | |
8557 | tree new_arg; | |
8558 | }; | |
8559 | ||
8560 | /* Invoked through walk_gimple_op, will remap all PARM_DECLs to the ones | |
8561 | pertaining to kernel function. */ | |
8562 | ||
8563 | static tree | |
8564 | grid_remap_kernel_arg_accesses (tree *tp, int *walk_subtrees, void *data) | |
8565 | { | |
8566 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; | |
8567 | struct grid_arg_decl_map *adm = (struct grid_arg_decl_map *) wi->info; | |
8568 | tree t = *tp; | |
8569 | ||
8570 | if (t == adm->old_arg) | |
8571 | *tp = adm->new_arg; | |
8572 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); | |
8573 | return NULL_TREE; | |
8574 | } | |
8575 | ||
8576 | /* If TARGET region contains a kernel body for loop, remove its region from the | |
01914336 | 8577 | TARGET and expand it in HSA gridified kernel fashion. */ |
629b3d75 MJ |
8578 | |
8579 | static void | |
8580 | grid_expand_target_grid_body (struct omp_region *target) | |
8581 | { | |
8582 | if (!hsa_gen_requested_p ()) | |
8583 | return; | |
8584 | ||
8585 | gomp_target *tgt_stmt = as_a <gomp_target *> (last_stmt (target->entry)); | |
8586 | struct omp_region **pp; | |
8587 | ||
8588 | for (pp = &target->inner; *pp; pp = &(*pp)->next) | |
8589 | if ((*pp)->type == GIMPLE_OMP_GRID_BODY) | |
8590 | break; | |
8591 | ||
8592 | struct omp_region *gpukernel = *pp; | |
8593 | ||
8594 | tree orig_child_fndecl = gimple_omp_target_child_fn (tgt_stmt); | |
8595 | if (!gpukernel) | |
8596 | { | |
8597 | /* HSA cannot handle OACC stuff. */ | |
8598 | if (gimple_omp_target_kind (tgt_stmt) != GF_OMP_TARGET_KIND_REGION) | |
8599 | return; | |
8600 | gcc_checking_assert (orig_child_fndecl); | |
8601 | gcc_assert (!omp_find_clause (gimple_omp_target_clauses (tgt_stmt), | |
8602 | OMP_CLAUSE__GRIDDIM_)); | |
8603 | cgraph_node *n = cgraph_node::get (orig_child_fndecl); | |
8604 | ||
8605 | hsa_register_kernel (n); | |
8606 | return; | |
8607 | } | |
8608 | ||
8609 | gcc_assert (omp_find_clause (gimple_omp_target_clauses (tgt_stmt), | |
8610 | OMP_CLAUSE__GRIDDIM_)); | |
01914336 MJ |
8611 | tree inside_block |
8612 | = gimple_block (first_stmt (single_succ (gpukernel->entry))); | |
629b3d75 MJ |
8613 | *pp = gpukernel->next; |
8614 | for (pp = &gpukernel->inner; *pp; pp = &(*pp)->next) | |
8615 | if ((*pp)->type == GIMPLE_OMP_FOR) | |
8616 | break; | |
8617 | ||
8618 | struct omp_region *kfor = *pp; | |
8619 | gcc_assert (kfor); | |
8620 | gomp_for *for_stmt = as_a <gomp_for *> (last_stmt (kfor->entry)); | |
8621 | gcc_assert (gimple_omp_for_kind (for_stmt) == GF_OMP_FOR_KIND_GRID_LOOP); | |
8622 | *pp = kfor->next; | |
8623 | if (kfor->inner) | |
8624 | { | |
8625 | if (gimple_omp_for_grid_group_iter (for_stmt)) | |
8626 | { | |
8627 | struct omp_region **next_pp; | |
8628 | for (pp = &kfor->inner; *pp; pp = next_pp) | |
8629 | { | |
8630 | next_pp = &(*pp)->next; | |
8631 | if ((*pp)->type != GIMPLE_OMP_FOR) | |
8632 | continue; | |
8633 | gomp_for *inner = as_a <gomp_for *> (last_stmt ((*pp)->entry)); | |
8634 | gcc_assert (gimple_omp_for_kind (inner) | |
8635 | == GF_OMP_FOR_KIND_GRID_LOOP); | |
8636 | grid_expand_omp_for_loop (*pp, true); | |
8637 | *pp = (*pp)->next; | |
8638 | next_pp = pp; | |
8639 | } | |
8640 | } | |
8641 | expand_omp (kfor->inner); | |
8642 | } | |
8643 | if (gpukernel->inner) | |
8644 | expand_omp (gpukernel->inner); | |
8645 | ||
8646 | tree kern_fndecl = copy_node (orig_child_fndecl); | |
7958186b MP |
8647 | DECL_NAME (kern_fndecl) = clone_function_name_numbered (kern_fndecl, |
8648 | "kernel"); | |
629b3d75 MJ |
8649 | SET_DECL_ASSEMBLER_NAME (kern_fndecl, DECL_NAME (kern_fndecl)); |
8650 | tree tgtblock = gimple_block (tgt_stmt); | |
8651 | tree fniniblock = make_node (BLOCK); | |
dc16b007 | 8652 | BLOCK_ABSTRACT_ORIGIN (fniniblock) = BLOCK_ORIGIN (tgtblock); |
629b3d75 MJ |
8653 | BLOCK_SOURCE_LOCATION (fniniblock) = BLOCK_SOURCE_LOCATION (tgtblock); |
8654 | BLOCK_SOURCE_END_LOCATION (fniniblock) = BLOCK_SOURCE_END_LOCATION (tgtblock); | |
8655 | BLOCK_SUPERCONTEXT (fniniblock) = kern_fndecl; | |
8656 | DECL_INITIAL (kern_fndecl) = fniniblock; | |
8657 | push_struct_function (kern_fndecl); | |
8658 | cfun->function_end_locus = gimple_location (tgt_stmt); | |
8659 | init_tree_ssa (cfun); | |
8660 | pop_cfun (); | |
8661 | ||
8662 | tree old_parm_decl = DECL_ARGUMENTS (kern_fndecl); | |
8663 | gcc_assert (!DECL_CHAIN (old_parm_decl)); | |
8664 | tree new_parm_decl = copy_node (DECL_ARGUMENTS (kern_fndecl)); | |
8665 | DECL_CONTEXT (new_parm_decl) = kern_fndecl; | |
8666 | DECL_ARGUMENTS (kern_fndecl) = new_parm_decl; | |
8667 | gcc_assert (VOID_TYPE_P (TREE_TYPE (DECL_RESULT (kern_fndecl)))); | |
8668 | DECL_RESULT (kern_fndecl) = copy_node (DECL_RESULT (kern_fndecl)); | |
8669 | DECL_CONTEXT (DECL_RESULT (kern_fndecl)) = kern_fndecl; | |
8670 | struct function *kern_cfun = DECL_STRUCT_FUNCTION (kern_fndecl); | |
8671 | kern_cfun->curr_properties = cfun->curr_properties; | |
8672 | ||
8673 | grid_expand_omp_for_loop (kfor, false); | |
8674 | ||
01914336 | 8675 | /* Remove the omp for statement. */ |
65f4b875 | 8676 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (gpukernel->entry); |
629b3d75 MJ |
8677 | gsi_remove (&gsi, true); |
8678 | /* Replace the GIMPLE_OMP_RETURN at the end of the kernel region with a real | |
8679 | return. */ | |
65f4b875 | 8680 | gsi = gsi_last_nondebug_bb (gpukernel->exit); |
629b3d75 MJ |
8681 | gcc_assert (!gsi_end_p (gsi) |
8682 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); | |
8683 | gimple *ret_stmt = gimple_build_return (NULL); | |
8684 | gsi_insert_after (&gsi, ret_stmt, GSI_SAME_STMT); | |
8685 | gsi_remove (&gsi, true); | |
8686 | ||
8687 | /* Statements in the first BB in the target construct have been produced by | |
8688 | target lowering and must be copied inside the GPUKERNEL, with the two | |
8689 | exceptions of the first OMP statement and the OMP_DATA assignment | |
8690 | statement. */ | |
8691 | gsi = gsi_start_bb (single_succ (gpukernel->entry)); | |
8692 | tree data_arg = gimple_omp_target_data_arg (tgt_stmt); | |
8693 | tree sender = data_arg ? TREE_VEC_ELT (data_arg, 0) : NULL; | |
8694 | for (gimple_stmt_iterator tsi = gsi_start_bb (single_succ (target->entry)); | |
8695 | !gsi_end_p (tsi); gsi_next (&tsi)) | |
8696 | { | |
8697 | gimple *stmt = gsi_stmt (tsi); | |
8698 | if (is_gimple_omp (stmt)) | |
8699 | break; | |
8700 | if (sender | |
8701 | && is_gimple_assign (stmt) | |
8702 | && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR | |
8703 | && TREE_OPERAND (gimple_assign_rhs1 (stmt), 0) == sender) | |
8704 | continue; | |
8705 | gimple *copy = gimple_copy (stmt); | |
8706 | gsi_insert_before (&gsi, copy, GSI_SAME_STMT); | |
8707 | gimple_set_block (copy, fniniblock); | |
8708 | } | |
8709 | ||
8710 | move_sese_region_to_fn (kern_cfun, single_succ (gpukernel->entry), | |
8711 | gpukernel->exit, inside_block); | |
8712 | ||
8713 | cgraph_node *kcn = cgraph_node::get_create (kern_fndecl); | |
8714 | kcn->mark_force_output (); | |
8715 | cgraph_node *orig_child = cgraph_node::get (orig_child_fndecl); | |
8716 | ||
8717 | hsa_register_kernel (kcn, orig_child); | |
8718 | ||
8719 | cgraph_node::add_new_function (kern_fndecl, true); | |
8720 | push_cfun (kern_cfun); | |
8721 | cgraph_edge::rebuild_edges (); | |
8722 | ||
8723 | /* Re-map any mention of the PARM_DECL of the original function to the | |
8724 | PARM_DECL of the new one. | |
8725 | ||
8726 | TODO: It would be great if lowering produced references into the GPU | |
8727 | kernel decl straight away and we did not have to do this. */ | |
8728 | struct grid_arg_decl_map adm; | |
8729 | adm.old_arg = old_parm_decl; | |
8730 | adm.new_arg = new_parm_decl; | |
8731 | basic_block bb; | |
8732 | FOR_EACH_BB_FN (bb, kern_cfun) | |
8733 | { | |
8734 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
8735 | { | |
8736 | gimple *stmt = gsi_stmt (gsi); | |
8737 | struct walk_stmt_info wi; | |
8738 | memset (&wi, 0, sizeof (wi)); | |
8739 | wi.info = &adm; | |
8740 | walk_gimple_op (stmt, grid_remap_kernel_arg_accesses, &wi); | |
8741 | } | |
8742 | } | |
8743 | pop_cfun (); | |
8744 | ||
8745 | return; | |
8746 | } | |
8747 | ||
8748 | /* Expand the parallel region tree rooted at REGION. Expansion | |
8749 | proceeds in depth-first order. Innermost regions are expanded | |
8750 | first. This way, parallel regions that require a new function to | |
8751 | be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any | |
8752 | internal dependencies in their body. */ | |
8753 | ||
8754 | static void | |
8755 | expand_omp (struct omp_region *region) | |
8756 | { | |
8757 | omp_any_child_fn_dumped = false; | |
8758 | while (region) | |
8759 | { | |
8760 | location_t saved_location; | |
8761 | gimple *inner_stmt = NULL; | |
8762 | ||
8763 | /* First, determine whether this is a combined parallel+workshare | |
01914336 | 8764 | region. */ |
629b3d75 MJ |
8765 | if (region->type == GIMPLE_OMP_PARALLEL) |
8766 | determine_parallel_type (region); | |
8767 | else if (region->type == GIMPLE_OMP_TARGET) | |
8768 | grid_expand_target_grid_body (region); | |
8769 | ||
8770 | if (region->type == GIMPLE_OMP_FOR | |
8771 | && gimple_omp_for_combined_p (last_stmt (region->entry))) | |
8772 | inner_stmt = last_stmt (region->inner->entry); | |
8773 | ||
8774 | if (region->inner) | |
8775 | expand_omp (region->inner); | |
8776 | ||
8777 | saved_location = input_location; | |
8778 | if (gimple_has_location (last_stmt (region->entry))) | |
8779 | input_location = gimple_location (last_stmt (region->entry)); | |
8780 | ||
8781 | switch (region->type) | |
8782 | { | |
8783 | case GIMPLE_OMP_PARALLEL: | |
8784 | case GIMPLE_OMP_TASK: | |
8785 | expand_omp_taskreg (region); | |
8786 | break; | |
8787 | ||
8788 | case GIMPLE_OMP_FOR: | |
8789 | expand_omp_for (region, inner_stmt); | |
8790 | break; | |
8791 | ||
8792 | case GIMPLE_OMP_SECTIONS: | |
8793 | expand_omp_sections (region); | |
8794 | break; | |
8795 | ||
8796 | case GIMPLE_OMP_SECTION: | |
8797 | /* Individual omp sections are handled together with their | |
8798 | parent GIMPLE_OMP_SECTIONS region. */ | |
8799 | break; | |
8800 | ||
8801 | case GIMPLE_OMP_SINGLE: | |
8802 | expand_omp_single (region); | |
8803 | break; | |
8804 | ||
8805 | case GIMPLE_OMP_ORDERED: | |
8806 | { | |
8807 | gomp_ordered *ord_stmt | |
8808 | = as_a <gomp_ordered *> (last_stmt (region->entry)); | |
8809 | if (omp_find_clause (gimple_omp_ordered_clauses (ord_stmt), | |
8810 | OMP_CLAUSE_DEPEND)) | |
8811 | { | |
8812 | /* We'll expand these when expanding corresponding | |
8813 | worksharing region with ordered(n) clause. */ | |
8814 | gcc_assert (region->outer | |
8815 | && region->outer->type == GIMPLE_OMP_FOR); | |
8816 | region->ord_stmt = ord_stmt; | |
8817 | break; | |
8818 | } | |
8819 | } | |
8820 | /* FALLTHRU */ | |
8821 | case GIMPLE_OMP_MASTER: | |
8822 | case GIMPLE_OMP_TASKGROUP: | |
8823 | case GIMPLE_OMP_CRITICAL: | |
8824 | case GIMPLE_OMP_TEAMS: | |
8825 | expand_omp_synch (region); | |
8826 | break; | |
8827 | ||
8828 | case GIMPLE_OMP_ATOMIC_LOAD: | |
8829 | expand_omp_atomic (region); | |
8830 | break; | |
8831 | ||
8832 | case GIMPLE_OMP_TARGET: | |
8833 | expand_omp_target (region); | |
8834 | break; | |
8835 | ||
8836 | default: | |
8837 | gcc_unreachable (); | |
8838 | } | |
8839 | ||
8840 | input_location = saved_location; | |
8841 | region = region->next; | |
8842 | } | |
8843 | if (omp_any_child_fn_dumped) | |
8844 | { | |
8845 | if (dump_file) | |
8846 | dump_function_header (dump_file, current_function_decl, dump_flags); | |
8847 | omp_any_child_fn_dumped = false; | |
8848 | } | |
8849 | } | |
8850 | ||
8851 | /* Helper for build_omp_regions. Scan the dominator tree starting at | |
8852 | block BB. PARENT is the region that contains BB. If SINGLE_TREE is | |
8853 | true, the function ends once a single tree is built (otherwise, whole | |
8854 | forest of OMP constructs may be built). */ | |
8855 | ||
8856 | static void | |
8857 | build_omp_regions_1 (basic_block bb, struct omp_region *parent, | |
8858 | bool single_tree) | |
8859 | { | |
8860 | gimple_stmt_iterator gsi; | |
8861 | gimple *stmt; | |
8862 | basic_block son; | |
8863 | ||
65f4b875 | 8864 | gsi = gsi_last_nondebug_bb (bb); |
629b3d75 MJ |
8865 | if (!gsi_end_p (gsi) && is_gimple_omp (gsi_stmt (gsi))) |
8866 | { | |
8867 | struct omp_region *region; | |
8868 | enum gimple_code code; | |
8869 | ||
8870 | stmt = gsi_stmt (gsi); | |
8871 | code = gimple_code (stmt); | |
8872 | if (code == GIMPLE_OMP_RETURN) | |
8873 | { | |
8874 | /* STMT is the return point out of region PARENT. Mark it | |
8875 | as the exit point and make PARENT the immediately | |
8876 | enclosing region. */ | |
8877 | gcc_assert (parent); | |
8878 | region = parent; | |
8879 | region->exit = bb; | |
8880 | parent = parent->outer; | |
8881 | } | |
8882 | else if (code == GIMPLE_OMP_ATOMIC_STORE) | |
8883 | { | |
5764ee3c | 8884 | /* GIMPLE_OMP_ATOMIC_STORE is analogous to |
629b3d75 MJ |
8885 | GIMPLE_OMP_RETURN, but matches with |
8886 | GIMPLE_OMP_ATOMIC_LOAD. */ | |
8887 | gcc_assert (parent); | |
8888 | gcc_assert (parent->type == GIMPLE_OMP_ATOMIC_LOAD); | |
8889 | region = parent; | |
8890 | region->exit = bb; | |
8891 | parent = parent->outer; | |
8892 | } | |
8893 | else if (code == GIMPLE_OMP_CONTINUE) | |
8894 | { | |
8895 | gcc_assert (parent); | |
8896 | parent->cont = bb; | |
8897 | } | |
8898 | else if (code == GIMPLE_OMP_SECTIONS_SWITCH) | |
8899 | { | |
8900 | /* GIMPLE_OMP_SECTIONS_SWITCH is part of | |
8901 | GIMPLE_OMP_SECTIONS, and we do nothing for it. */ | |
8902 | } | |
8903 | else | |
8904 | { | |
8905 | region = new_omp_region (bb, code, parent); | |
8906 | /* Otherwise... */ | |
8907 | if (code == GIMPLE_OMP_TARGET) | |
8908 | { | |
8909 | switch (gimple_omp_target_kind (stmt)) | |
8910 | { | |
8911 | case GF_OMP_TARGET_KIND_REGION: | |
8912 | case GF_OMP_TARGET_KIND_DATA: | |
8913 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
8914 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
8915 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
8916 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
8917 | break; | |
8918 | case GF_OMP_TARGET_KIND_UPDATE: | |
8919 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
8920 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
8921 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
8922 | case GF_OMP_TARGET_KIND_OACC_ENTER_EXIT_DATA: | |
8923 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
8924 | /* ..., other than for those stand-alone directives... */ | |
8925 | region = NULL; | |
8926 | break; | |
8927 | default: | |
8928 | gcc_unreachable (); | |
8929 | } | |
8930 | } | |
8931 | else if (code == GIMPLE_OMP_ORDERED | |
8932 | && omp_find_clause (gimple_omp_ordered_clauses | |
8933 | (as_a <gomp_ordered *> (stmt)), | |
8934 | OMP_CLAUSE_DEPEND)) | |
8935 | /* #pragma omp ordered depend is also just a stand-alone | |
8936 | directive. */ | |
8937 | region = NULL; | |
28567c40 JJ |
8938 | else if (code == GIMPLE_OMP_TASK |
8939 | && gimple_omp_task_taskwait_p (stmt)) | |
8940 | /* #pragma omp taskwait depend(...) is a stand-alone directive. */ | |
8941 | region = NULL; | |
629b3d75 MJ |
8942 | /* ..., this directive becomes the parent for a new region. */ |
8943 | if (region) | |
8944 | parent = region; | |
8945 | } | |
8946 | } | |
8947 | ||
8948 | if (single_tree && !parent) | |
8949 | return; | |
8950 | ||
8951 | for (son = first_dom_son (CDI_DOMINATORS, bb); | |
8952 | son; | |
8953 | son = next_dom_son (CDI_DOMINATORS, son)) | |
8954 | build_omp_regions_1 (son, parent, single_tree); | |
8955 | } | |
8956 | ||
8957 | /* Builds the tree of OMP regions rooted at ROOT, storing it to | |
8958 | root_omp_region. */ | |
8959 | ||
8960 | static void | |
8961 | build_omp_regions_root (basic_block root) | |
8962 | { | |
8963 | gcc_assert (root_omp_region == NULL); | |
8964 | build_omp_regions_1 (root, NULL, true); | |
8965 | gcc_assert (root_omp_region != NULL); | |
8966 | } | |
8967 | ||
8968 | /* Expands omp construct (and its subconstructs) starting in HEAD. */ | |
8969 | ||
8970 | void | |
8971 | omp_expand_local (basic_block head) | |
8972 | { | |
8973 | build_omp_regions_root (head); | |
8974 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
8975 | { | |
8976 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
8977 | dump_omp_region (dump_file, root_omp_region, 0); | |
8978 | fprintf (dump_file, "\n"); | |
8979 | } | |
8980 | ||
8981 | remove_exit_barriers (root_omp_region); | |
8982 | expand_omp (root_omp_region); | |
8983 | ||
8984 | omp_free_regions (); | |
8985 | } | |
8986 | ||
8987 | /* Scan the CFG and build a tree of OMP regions. Return the root of | |
8988 | the OMP region tree. */ | |
8989 | ||
8990 | static void | |
8991 | build_omp_regions (void) | |
8992 | { | |
8993 | gcc_assert (root_omp_region == NULL); | |
8994 | calculate_dominance_info (CDI_DOMINATORS); | |
8995 | build_omp_regions_1 (ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, false); | |
8996 | } | |
8997 | ||
8998 | /* Main entry point for expanding OMP-GIMPLE into runtime calls. */ | |
8999 | ||
9000 | static unsigned int | |
9001 | execute_expand_omp (void) | |
9002 | { | |
9003 | build_omp_regions (); | |
9004 | ||
9005 | if (!root_omp_region) | |
9006 | return 0; | |
9007 | ||
9008 | if (dump_file) | |
9009 | { | |
9010 | fprintf (dump_file, "\nOMP region tree\n\n"); | |
9011 | dump_omp_region (dump_file, root_omp_region, 0); | |
9012 | fprintf (dump_file, "\n"); | |
9013 | } | |
9014 | ||
9015 | remove_exit_barriers (root_omp_region); | |
9016 | ||
9017 | expand_omp (root_omp_region); | |
9018 | ||
9019 | if (flag_checking && !loops_state_satisfies_p (LOOPS_NEED_FIXUP)) | |
9020 | verify_loop_structure (); | |
9021 | cleanup_tree_cfg (); | |
9022 | ||
9023 | omp_free_regions (); | |
9024 | ||
9025 | return 0; | |
9026 | } | |
9027 | ||
9028 | /* OMP expansion -- the default pass, run before creation of SSA form. */ | |
9029 | ||
9030 | namespace { | |
9031 | ||
9032 | const pass_data pass_data_expand_omp = | |
9033 | { | |
9034 | GIMPLE_PASS, /* type */ | |
9035 | "ompexp", /* name */ | |
fd2b8c8b | 9036 | OPTGROUP_OMP, /* optinfo_flags */ |
629b3d75 MJ |
9037 | TV_NONE, /* tv_id */ |
9038 | PROP_gimple_any, /* properties_required */ | |
9039 | PROP_gimple_eomp, /* properties_provided */ | |
9040 | 0, /* properties_destroyed */ | |
9041 | 0, /* todo_flags_start */ | |
9042 | 0, /* todo_flags_finish */ | |
9043 | }; | |
9044 | ||
9045 | class pass_expand_omp : public gimple_opt_pass | |
9046 | { | |
9047 | public: | |
9048 | pass_expand_omp (gcc::context *ctxt) | |
9049 | : gimple_opt_pass (pass_data_expand_omp, ctxt) | |
9050 | {} | |
9051 | ||
9052 | /* opt_pass methods: */ | |
9053 | virtual unsigned int execute (function *) | |
9054 | { | |
5e9d6aa4 | 9055 | bool gate = ((flag_openacc != 0 || flag_openmp != 0 |
629b3d75 MJ |
9056 | || flag_openmp_simd != 0) |
9057 | && !seen_error ()); | |
9058 | ||
9059 | /* This pass always runs, to provide PROP_gimple_eomp. | |
9060 | But often, there is nothing to do. */ | |
9061 | if (!gate) | |
9062 | return 0; | |
9063 | ||
9064 | return execute_expand_omp (); | |
9065 | } | |
9066 | ||
9067 | }; // class pass_expand_omp | |
9068 | ||
9069 | } // anon namespace | |
9070 | ||
9071 | gimple_opt_pass * | |
9072 | make_pass_expand_omp (gcc::context *ctxt) | |
9073 | { | |
9074 | return new pass_expand_omp (ctxt); | |
9075 | } | |
9076 | ||
9077 | namespace { | |
9078 | ||
9079 | const pass_data pass_data_expand_omp_ssa = | |
9080 | { | |
9081 | GIMPLE_PASS, /* type */ | |
9082 | "ompexpssa", /* name */ | |
fd2b8c8b | 9083 | OPTGROUP_OMP, /* optinfo_flags */ |
629b3d75 MJ |
9084 | TV_NONE, /* tv_id */ |
9085 | PROP_cfg | PROP_ssa, /* properties_required */ | |
9086 | PROP_gimple_eomp, /* properties_provided */ | |
9087 | 0, /* properties_destroyed */ | |
9088 | 0, /* todo_flags_start */ | |
9089 | TODO_cleanup_cfg | TODO_rebuild_alias, /* todo_flags_finish */ | |
9090 | }; | |
9091 | ||
9092 | class pass_expand_omp_ssa : public gimple_opt_pass | |
9093 | { | |
9094 | public: | |
9095 | pass_expand_omp_ssa (gcc::context *ctxt) | |
9096 | : gimple_opt_pass (pass_data_expand_omp_ssa, ctxt) | |
9097 | {} | |
9098 | ||
9099 | /* opt_pass methods: */ | |
9100 | virtual bool gate (function *fun) | |
9101 | { | |
9102 | return !(fun->curr_properties & PROP_gimple_eomp); | |
9103 | } | |
9104 | virtual unsigned int execute (function *) { return execute_expand_omp (); } | |
9105 | opt_pass * clone () { return new pass_expand_omp_ssa (m_ctxt); } | |
9106 | ||
9107 | }; // class pass_expand_omp_ssa | |
9108 | ||
9109 | } // anon namespace | |
9110 | ||
9111 | gimple_opt_pass * | |
9112 | make_pass_expand_omp_ssa (gcc::context *ctxt) | |
9113 | { | |
9114 | return new pass_expand_omp_ssa (ctxt); | |
9115 | } | |
9116 | ||
9117 | /* Called from tree-cfg.c::make_edges to create cfg edges for all relevant | |
9118 | GIMPLE_* codes. */ | |
9119 | ||
9120 | bool | |
9121 | omp_make_gimple_edges (basic_block bb, struct omp_region **region, | |
9122 | int *region_idx) | |
9123 | { | |
9124 | gimple *last = last_stmt (bb); | |
9125 | enum gimple_code code = gimple_code (last); | |
9126 | struct omp_region *cur_region = *region; | |
9127 | bool fallthru = false; | |
9128 | ||
9129 | switch (code) | |
9130 | { | |
9131 | case GIMPLE_OMP_PARALLEL: | |
629b3d75 MJ |
9132 | case GIMPLE_OMP_FOR: |
9133 | case GIMPLE_OMP_SINGLE: | |
9134 | case GIMPLE_OMP_TEAMS: | |
9135 | case GIMPLE_OMP_MASTER: | |
9136 | case GIMPLE_OMP_TASKGROUP: | |
9137 | case GIMPLE_OMP_CRITICAL: | |
9138 | case GIMPLE_OMP_SECTION: | |
9139 | case GIMPLE_OMP_GRID_BODY: | |
9140 | cur_region = new_omp_region (bb, code, cur_region); | |
9141 | fallthru = true; | |
9142 | break; | |
9143 | ||
28567c40 JJ |
9144 | case GIMPLE_OMP_TASK: |
9145 | cur_region = new_omp_region (bb, code, cur_region); | |
9146 | fallthru = true; | |
9147 | if (gimple_omp_task_taskwait_p (last)) | |
9148 | cur_region = cur_region->outer; | |
9149 | break; | |
9150 | ||
629b3d75 MJ |
9151 | case GIMPLE_OMP_ORDERED: |
9152 | cur_region = new_omp_region (bb, code, cur_region); | |
9153 | fallthru = true; | |
9154 | if (omp_find_clause (gimple_omp_ordered_clauses | |
9155 | (as_a <gomp_ordered *> (last)), | |
9156 | OMP_CLAUSE_DEPEND)) | |
9157 | cur_region = cur_region->outer; | |
9158 | break; | |
9159 | ||
9160 | case GIMPLE_OMP_TARGET: | |
9161 | cur_region = new_omp_region (bb, code, cur_region); | |
9162 | fallthru = true; | |
9163 | switch (gimple_omp_target_kind (last)) | |
9164 | { | |
9165 | case GF_OMP_TARGET_KIND_REGION: | |
9166 | case GF_OMP_TARGET_KIND_DATA: | |
9167 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: | |
9168 | case GF_OMP_TARGET_KIND_OACC_KERNELS: | |
9169 | case GF_OMP_TARGET_KIND_OACC_DATA: | |
9170 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: | |
9171 | break; | |
9172 | case GF_OMP_TARGET_KIND_UPDATE: | |
9173 | case GF_OMP_TARGET_KIND_ENTER_DATA: | |
9174 | case GF_OMP_TARGET_KIND_EXIT_DATA: | |
9175 | case GF_OMP_TARGET_KIND_OACC_UPDATE: | |
9176 | case GF_OMP_TARGET_KIND_OACC_ENTER_EXIT_DATA: | |
9177 | case GF_OMP_TARGET_KIND_OACC_DECLARE: | |
9178 | cur_region = cur_region->outer; | |
9179 | break; | |
9180 | default: | |
9181 | gcc_unreachable (); | |
9182 | } | |
9183 | break; | |
9184 | ||
9185 | case GIMPLE_OMP_SECTIONS: | |
9186 | cur_region = new_omp_region (bb, code, cur_region); | |
9187 | fallthru = true; | |
9188 | break; | |
9189 | ||
9190 | case GIMPLE_OMP_SECTIONS_SWITCH: | |
9191 | fallthru = false; | |
9192 | break; | |
9193 | ||
9194 | case GIMPLE_OMP_ATOMIC_LOAD: | |
9195 | case GIMPLE_OMP_ATOMIC_STORE: | |
9196 | fallthru = true; | |
9197 | break; | |
9198 | ||
9199 | case GIMPLE_OMP_RETURN: | |
9200 | /* In the case of a GIMPLE_OMP_SECTION, the edge will go | |
9201 | somewhere other than the next block. This will be | |
9202 | created later. */ | |
9203 | cur_region->exit = bb; | |
9204 | if (cur_region->type == GIMPLE_OMP_TASK) | |
9205 | /* Add an edge corresponding to not scheduling the task | |
9206 | immediately. */ | |
9207 | make_edge (cur_region->entry, bb, EDGE_ABNORMAL); | |
9208 | fallthru = cur_region->type != GIMPLE_OMP_SECTION; | |
9209 | cur_region = cur_region->outer; | |
9210 | break; | |
9211 | ||
9212 | case GIMPLE_OMP_CONTINUE: | |
9213 | cur_region->cont = bb; | |
9214 | switch (cur_region->type) | |
9215 | { | |
9216 | case GIMPLE_OMP_FOR: | |
9217 | /* Mark all GIMPLE_OMP_FOR and GIMPLE_OMP_CONTINUE | |
9218 | succs edges as abnormal to prevent splitting | |
9219 | them. */ | |
9220 | single_succ_edge (cur_region->entry)->flags |= EDGE_ABNORMAL; | |
9221 | /* Make the loopback edge. */ | |
9222 | make_edge (bb, single_succ (cur_region->entry), | |
9223 | EDGE_ABNORMAL); | |
9224 | ||
9225 | /* Create an edge from GIMPLE_OMP_FOR to exit, which | |
9226 | corresponds to the case that the body of the loop | |
9227 | is not executed at all. */ | |
9228 | make_edge (cur_region->entry, bb->next_bb, EDGE_ABNORMAL); | |
9229 | make_edge (bb, bb->next_bb, EDGE_FALLTHRU | EDGE_ABNORMAL); | |
9230 | fallthru = false; | |
9231 | break; | |
9232 | ||
9233 | case GIMPLE_OMP_SECTIONS: | |
9234 | /* Wire up the edges into and out of the nested sections. */ | |
9235 | { | |
9236 | basic_block switch_bb = single_succ (cur_region->entry); | |
9237 | ||
9238 | struct omp_region *i; | |
9239 | for (i = cur_region->inner; i ; i = i->next) | |
9240 | { | |
9241 | gcc_assert (i->type == GIMPLE_OMP_SECTION); | |
9242 | make_edge (switch_bb, i->entry, 0); | |
9243 | make_edge (i->exit, bb, EDGE_FALLTHRU); | |
9244 | } | |
9245 | ||
9246 | /* Make the loopback edge to the block with | |
9247 | GIMPLE_OMP_SECTIONS_SWITCH. */ | |
9248 | make_edge (bb, switch_bb, 0); | |
9249 | ||
9250 | /* Make the edge from the switch to exit. */ | |
9251 | make_edge (switch_bb, bb->next_bb, 0); | |
9252 | fallthru = false; | |
9253 | } | |
9254 | break; | |
9255 | ||
9256 | case GIMPLE_OMP_TASK: | |
9257 | fallthru = true; | |
9258 | break; | |
9259 | ||
9260 | default: | |
9261 | gcc_unreachable (); | |
9262 | } | |
9263 | break; | |
9264 | ||
9265 | default: | |
9266 | gcc_unreachable (); | |
9267 | } | |
9268 | ||
9269 | if (*region != cur_region) | |
9270 | { | |
9271 | *region = cur_region; | |
9272 | if (cur_region) | |
9273 | *region_idx = cur_region->entry->index; | |
9274 | else | |
9275 | *region_idx = 0; | |
9276 | } | |
9277 | ||
9278 | return fallthru; | |
9279 | } | |
9280 | ||
9281 | #include "gt-omp-expand.h" |