1 | /* Expansion pass for OMP directives. Outlines regions of certain OMP |
2 | directives to separate functions, converts others into explicit calls to the |
3 | runtime library (libgomp) and so forth |
4 | |
5 | Copyright (C) 2005-2023 Free Software Foundation, Inc. |
6 | |
7 | This file is part of GCC. |
8 | |
9 | GCC is free software; you can redistribute it and/or modify it under |
10 | the terms of the GNU General Public License as published by the Free |
11 | Software Foundation; either version 3, or (at your option) any later |
12 | version. |
13 | |
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
17 | for more details. |
18 | |
19 | You should have received a copy of the GNU General Public License |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ |
22 | |
23 | #include "config.h" |
24 | #include "system.h" |
25 | #include "coretypes.h" |
26 | #include "memmodel.h" |
27 | #include "backend.h" |
28 | #include "target.h" |
29 | #include "rtl.h" |
30 | #include "tree.h" |
31 | #include "gimple.h" |
32 | #include "cfghooks.h" |
33 | #include "tree-pass.h" |
34 | #include "ssa.h" |
35 | #include "optabs.h" |
36 | #include "cgraph.h" |
37 | #include "pretty-print.h" |
38 | #include "diagnostic-core.h" |
39 | #include "fold-const.h" |
40 | #include "stor-layout.h" |
41 | #include "cfganal.h" |
42 | #include "internal-fn.h" |
43 | #include "gimplify.h" |
44 | #include "gimple-iterator.h" |
45 | #include "gimplify-me.h" |
46 | #include "gimple-walk.h" |
47 | #include "tree-cfg.h" |
48 | #include "tree-into-ssa.h" |
49 | #include "tree-ssa.h" |
50 | #include "splay-tree.h" |
51 | #include "cfgloop.h" |
52 | #include "omp-general.h" |
53 | #include "omp-offload.h" |
54 | #include "tree-cfgcleanup.h" |
55 | #include "alloc-pool.h" |
56 | #include "symbol-summary.h" |
57 | #include "gomp-constants.h" |
58 | #include "gimple-pretty-print.h" |
59 | #include "stringpool.h" |
60 | #include "attribs.h" |
61 | #include "tree-eh.h" |
62 | #include "opts.h" |
63 | |
64 | /* OMP region information. Every parallel and workshare |
65 | directive is enclosed between two markers, the OMP_* directive |
66 | and a corresponding GIMPLE_OMP_RETURN statement. */ |
67 | |
68 | struct omp_region |
69 | { |
70 | /* The enclosing region. */ |
71 | struct omp_region *outer; |
72 | |
73 | /* First child region. */ |
74 | struct omp_region *inner; |
75 | |
76 | /* Next peer region. */ |
77 | struct omp_region *next; |
78 | |
79 | /* Block containing the omp directive as its last stmt. */ |
80 | basic_block entry; |
81 | |
82 | /* Block containing the GIMPLE_OMP_RETURN as its last stmt. */ |
83 | basic_block exit; |
84 | |
85 | /* Block containing the GIMPLE_OMP_CONTINUE as its last stmt. */ |
86 | basic_block cont; |
87 | |
88 | /* If this is a combined parallel+workshare region, this is a list |
89 | of additional arguments needed by the combined parallel+workshare |
90 | library call. */ |
91 | vec<tree, va_gc> *ws_args; |
92 | |
93 | /* The code for the omp directive of this region. */ |
94 | enum gimple_code type; |
95 | |
96 | /* Schedule kind, only used for GIMPLE_OMP_FOR type regions. */ |
97 | enum omp_clause_schedule_kind sched_kind; |
98 | |
99 | /* Schedule modifiers. */ |
100 | unsigned char sched_modifiers; |
101 | |
102 | /* True if this is a combined parallel+workshare region. */ |
103 | bool is_combined_parallel; |
104 | |
105 | /* Copy of fd.lastprivate_conditional != 0. */ |
106 | bool has_lastprivate_conditional; |
107 | |
108 | /* The ordered stmt if type is GIMPLE_OMP_ORDERED and it has |
109 | a depend clause. */ |
110 | gomp_ordered *ord_stmt; |
111 | }; |
112 | |
113 | static struct omp_region *root_omp_region; |
114 | static bool omp_any_child_fn_dumped; |
115 | |
116 | static void expand_omp_build_assign (gimple_stmt_iterator *, tree, tree, |
117 | bool = false); |
118 | static gphi *find_phi_with_arg_on_edge (tree, edge); |
119 | static void expand_omp (struct omp_region *region); |
120 | |
121 | /* Return true if REGION is a combined parallel+workshare region. */ |
122 | |
123 | static inline bool |
124 | is_combined_parallel (struct omp_region *region) |
125 | { |
126 | return region->is_combined_parallel; |
127 | } |
128 | |
129 | /* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB |
130 | is the immediate dominator of PAR_ENTRY_BB, return true if there |
131 | are no data dependencies that would prevent expanding the parallel |
132 | directive at PAR_ENTRY_BB as a combined parallel+workshare region. |
133 | |
134 | When expanding a combined parallel+workshare region, the call to |
135 | the child function may need additional arguments in the case of |
136 | GIMPLE_OMP_FOR regions. In some cases, these arguments are |
137 | computed out of variables passed in from the parent to the child |
138 | via 'struct .omp_data_s'. For instance: |
139 | |
140 | #pragma omp parallel for schedule (guided, i * 4) |
141 | for (j ...) |
142 | |
143 | Is lowered into: |
144 | |
145 | # BLOCK 2 (PAR_ENTRY_BB) |
146 | .omp_data_o.i = i; |
147 | #pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598) |
148 | |
149 | # BLOCK 3 (WS_ENTRY_BB) |
150 | .omp_data_i = &.omp_data_o; |
151 | D.1667 = .omp_data_i->i; |
152 | D.1598 = D.1667 * 4; |
153 | #pragma omp for schedule (guided, D.1598) |
154 | |
155 | When we outline the parallel region, the call to the child function |
156 | 'bar.omp_fn.0' will need the value D.1598 in its argument list, but |
157 | that value is computed *after* the call site. So, in principle we |
158 | cannot do the transformation. |
159 | |
160 | To see whether the code in WS_ENTRY_BB blocks the combined |
161 | parallel+workshare call, we collect all the variables used in the |
162 | GIMPLE_OMP_FOR header check whether they appear on the LHS of any |
163 | statement in WS_ENTRY_BB. If so, then we cannot emit the combined |
164 | call. |
165 | |
166 | FIXME. If we had the SSA form built at this point, we could merely |
167 | hoist the code in block 3 into block 2 and be done with it. But at |
168 | this point we don't have dataflow information and though we could |
169 | hack something up here, it is really not worth the aggravation. */ |
170 | |
171 | static bool |
172 | workshare_safe_to_combine_p (basic_block ws_entry_bb) |
173 | { |
174 | struct omp_for_data fd; |
175 | gimple *ws_stmt = last_nondebug_stmt (ws_entry_bb); |
176 | |
177 | if (gimple_code (g: ws_stmt) == GIMPLE_OMP_SECTIONS) |
178 | return true; |
179 | |
180 | gcc_assert (gimple_code (ws_stmt) == GIMPLE_OMP_FOR); |
181 | if (gimple_omp_for_kind (g: ws_stmt) != GF_OMP_FOR_KIND_FOR) |
182 | return false; |
183 | |
184 | omp_extract_for_data (for_stmt: as_a <gomp_for *> (p: ws_stmt), fd: &fd, NULL); |
185 | |
186 | if (fd.collapse > 1 && TREE_CODE (fd.loop.n2) != INTEGER_CST) |
187 | return false; |
188 | if (fd.iter_type != long_integer_type_node) |
189 | return false; |
190 | |
191 | /* FIXME. We give up too easily here. If any of these arguments |
192 | are not constants, they will likely involve variables that have |
193 | been mapped into fields of .omp_data_s for sharing with the child |
194 | function. With appropriate data flow, it would be possible to |
195 | see through this. */ |
196 | if (!is_gimple_min_invariant (fd.loop.n1) |
197 | || !is_gimple_min_invariant (fd.loop.n2) |
198 | || !is_gimple_min_invariant (fd.loop.step) |
199 | || (fd.chunk_size && !is_gimple_min_invariant (fd.chunk_size))) |
200 | return false; |
201 | |
202 | return true; |
203 | } |
204 | |
205 | /* Adjust CHUNK_SIZE from SCHEDULE clause, depending on simd modifier |
206 | presence (SIMD_SCHEDULE). */ |
207 | |
208 | static tree |
209 | omp_adjust_chunk_size (tree chunk_size, bool simd_schedule) |
210 | { |
211 | if (!simd_schedule || integer_zerop (chunk_size)) |
212 | return chunk_size; |
213 | |
214 | poly_uint64 vf = omp_max_vf (); |
215 | if (known_eq (vf, 1U)) |
216 | return chunk_size; |
217 | |
218 | tree type = TREE_TYPE (chunk_size); |
219 | chunk_size = fold_build2 (PLUS_EXPR, type, chunk_size, |
220 | build_int_cst (type, vf - 1)); |
221 | return fold_build2 (BIT_AND_EXPR, type, chunk_size, |
222 | build_int_cst (type, -vf)); |
223 | } |
224 | |
225 | /* Collect additional arguments needed to emit a combined |
226 | parallel+workshare call. WS_STMT is the workshare directive being |
227 | expanded. */ |
228 | |
229 | static vec<tree, va_gc> * |
230 | get_ws_args_for (gimple *par_stmt, gimple *ws_stmt) |
231 | { |
232 | tree t; |
233 | location_t loc = gimple_location (g: ws_stmt); |
234 | vec<tree, va_gc> *ws_args; |
235 | |
236 | if (gomp_for *for_stmt = dyn_cast <gomp_for *> (p: ws_stmt)) |
237 | { |
238 | struct omp_for_data fd; |
239 | tree n1, n2; |
240 | |
241 | omp_extract_for_data (for_stmt, fd: &fd, NULL); |
242 | n1 = fd.loop.n1; |
243 | n2 = fd.loop.n2; |
244 | |
245 | if (gimple_omp_for_combined_into_p (g: for_stmt)) |
246 | { |
247 | tree innerc |
248 | = omp_find_clause (clauses: gimple_omp_parallel_clauses (gs: par_stmt), |
249 | kind: OMP_CLAUSE__LOOPTEMP_); |
250 | gcc_assert (innerc); |
251 | n1 = OMP_CLAUSE_DECL (innerc); |
252 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
253 | kind: OMP_CLAUSE__LOOPTEMP_); |
254 | gcc_assert (innerc); |
255 | n2 = OMP_CLAUSE_DECL (innerc); |
256 | } |
257 | |
258 | vec_alloc (v&: ws_args, nelems: 3 + (fd.chunk_size != 0)); |
259 | |
260 | t = fold_convert_loc (loc, long_integer_type_node, n1); |
261 | ws_args->quick_push (obj: t); |
262 | |
263 | t = fold_convert_loc (loc, long_integer_type_node, n2); |
264 | ws_args->quick_push (obj: t); |
265 | |
266 | t = fold_convert_loc (loc, long_integer_type_node, fd.loop.step); |
267 | ws_args->quick_push (obj: t); |
268 | |
269 | if (fd.chunk_size) |
270 | { |
271 | t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size); |
272 | t = omp_adjust_chunk_size (chunk_size: t, simd_schedule: fd.simd_schedule); |
273 | ws_args->quick_push (obj: t); |
274 | } |
275 | |
276 | return ws_args; |
277 | } |
278 | else if (gimple_code (g: ws_stmt) == GIMPLE_OMP_SECTIONS) |
279 | { |
280 | /* Number of sections is equal to the number of edges from the |
281 | GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to |
282 | the exit of the sections region. */ |
283 | basic_block bb = single_succ (bb: gimple_bb (g: ws_stmt)); |
284 | t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1); |
285 | vec_alloc (v&: ws_args, nelems: 1); |
286 | ws_args->quick_push (obj: t); |
287 | return ws_args; |
288 | } |
289 | |
290 | gcc_unreachable (); |
291 | } |
292 | |
293 | /* Discover whether REGION is a combined parallel+workshare region. */ |
294 | |
295 | static void |
296 | determine_parallel_type (struct omp_region *region) |
297 | { |
298 | basic_block par_entry_bb, par_exit_bb; |
299 | basic_block ws_entry_bb, ws_exit_bb; |
300 | |
301 | if (region == NULL || region->inner == NULL |
302 | || region->exit == NULL || region->inner->exit == NULL |
303 | || region->inner->cont == NULL) |
304 | return; |
305 | |
306 | /* We only support parallel+for and parallel+sections. */ |
307 | if (region->type != GIMPLE_OMP_PARALLEL |
308 | || (region->inner->type != GIMPLE_OMP_FOR |
309 | && region->inner->type != GIMPLE_OMP_SECTIONS)) |
310 | return; |
311 | |
312 | /* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and |
313 | WS_EXIT_BB -> PAR_EXIT_BB. */ |
314 | par_entry_bb = region->entry; |
315 | par_exit_bb = region->exit; |
316 | ws_entry_bb = region->inner->entry; |
317 | ws_exit_bb = region->inner->exit; |
318 | |
319 | /* Give up for task reductions on the parallel, while it is implementable, |
320 | adding another big set of APIs or slowing down the normal paths is |
321 | not acceptable. */ |
322 | tree pclauses |
323 | = gimple_omp_parallel_clauses (gs: last_nondebug_stmt (par_entry_bb)); |
324 | if (omp_find_clause (clauses: pclauses, kind: OMP_CLAUSE__REDUCTEMP_)) |
325 | return; |
326 | |
327 | if (single_succ (bb: par_entry_bb) == ws_entry_bb |
328 | && single_succ (bb: ws_exit_bb) == par_exit_bb |
329 | && workshare_safe_to_combine_p (ws_entry_bb) |
330 | && (gimple_omp_parallel_combined_p (g: last_nondebug_stmt (par_entry_bb)) |
331 | || (last_and_only_stmt (ws_entry_bb) |
332 | && last_and_only_stmt (par_exit_bb)))) |
333 | { |
334 | gimple *par_stmt = last_nondebug_stmt (par_entry_bb); |
335 | gimple *ws_stmt = last_nondebug_stmt (ws_entry_bb); |
336 | |
337 | if (region->inner->type == GIMPLE_OMP_FOR) |
338 | { |
339 | /* If this is a combined parallel loop, we need to determine |
340 | whether or not to use the combined library calls. There |
341 | are two cases where we do not apply the transformation: |
342 | static loops and any kind of ordered loop. In the first |
343 | case, we already open code the loop so there is no need |
344 | to do anything else. In the latter case, the combined |
345 | parallel loop call would still need extra synchronization |
346 | to implement ordered semantics, so there would not be any |
347 | gain in using the combined call. */ |
348 | tree clauses = gimple_omp_for_clauses (gs: ws_stmt); |
349 | tree c = omp_find_clause (clauses, kind: OMP_CLAUSE_SCHEDULE); |
350 | if (c == NULL |
351 | || ((OMP_CLAUSE_SCHEDULE_KIND (c) & OMP_CLAUSE_SCHEDULE_MASK) |
352 | == OMP_CLAUSE_SCHEDULE_STATIC) |
353 | || omp_find_clause (clauses, kind: OMP_CLAUSE_ORDERED) |
354 | || omp_find_clause (clauses, kind: OMP_CLAUSE__REDUCTEMP_) |
355 | || ((c = omp_find_clause (clauses, kind: OMP_CLAUSE__CONDTEMP_)) |
356 | && POINTER_TYPE_P (TREE_TYPE (OMP_CLAUSE_DECL (c))))) |
357 | return; |
358 | } |
359 | else if (region->inner->type == GIMPLE_OMP_SECTIONS |
360 | && (omp_find_clause (clauses: gimple_omp_sections_clauses (gs: ws_stmt), |
361 | kind: OMP_CLAUSE__REDUCTEMP_) |
362 | || omp_find_clause (clauses: gimple_omp_sections_clauses (gs: ws_stmt), |
363 | kind: OMP_CLAUSE__CONDTEMP_))) |
364 | return; |
365 | |
366 | region->is_combined_parallel = true; |
367 | region->inner->is_combined_parallel = true; |
368 | region->ws_args = get_ws_args_for (par_stmt, ws_stmt); |
369 | } |
370 | } |
371 | |
372 | /* Debugging dumps for parallel regions. */ |
373 | void dump_omp_region (FILE *, struct omp_region *, int); |
374 | void debug_omp_region (struct omp_region *); |
375 | void debug_all_omp_regions (void); |
376 | |
377 | /* Dump the parallel region tree rooted at REGION. */ |
378 | |
379 | void |
380 | dump_omp_region (FILE *file, struct omp_region *region, int indent) |
381 | { |
382 | fprintf (stream: file, format: "%*sbb %d: %s\n" , indent, "" , region->entry->index, |
383 | gimple_code_name[region->type]); |
384 | |
385 | if (region->inner) |
386 | dump_omp_region (file, region: region->inner, indent: indent + 4); |
387 | |
388 | if (region->cont) |
389 | { |
390 | fprintf (stream: file, format: "%*sbb %d: GIMPLE_OMP_CONTINUE\n" , indent, "" , |
391 | region->cont->index); |
392 | } |
393 | |
394 | if (region->exit) |
395 | fprintf (stream: file, format: "%*sbb %d: GIMPLE_OMP_RETURN\n" , indent, "" , |
396 | region->exit->index); |
397 | else |
398 | fprintf (stream: file, format: "%*s[no exit marker]\n" , indent, "" ); |
399 | |
400 | if (region->next) |
401 | dump_omp_region (file, region: region->next, indent); |
402 | } |
403 | |
404 | DEBUG_FUNCTION void |
405 | debug_omp_region (struct omp_region *region) |
406 | { |
407 | dump_omp_region (stderr, region, indent: 0); |
408 | } |
409 | |
410 | DEBUG_FUNCTION void |
411 | debug_all_omp_regions (void) |
412 | { |
413 | dump_omp_region (stderr, region: root_omp_region, indent: 0); |
414 | } |
415 | |
416 | /* Create a new parallel region starting at STMT inside region PARENT. */ |
417 | |
418 | static struct omp_region * |
419 | new_omp_region (basic_block bb, enum gimple_code type, |
420 | struct omp_region *parent) |
421 | { |
422 | struct omp_region *region = XCNEW (struct omp_region); |
423 | |
424 | region->outer = parent; |
425 | region->entry = bb; |
426 | region->type = type; |
427 | |
428 | if (parent) |
429 | { |
430 | /* This is a nested region. Add it to the list of inner |
431 | regions in PARENT. */ |
432 | region->next = parent->inner; |
433 | parent->inner = region; |
434 | } |
435 | else |
436 | { |
437 | /* This is a toplevel region. Add it to the list of toplevel |
438 | regions in ROOT_OMP_REGION. */ |
439 | region->next = root_omp_region; |
440 | root_omp_region = region; |
441 | } |
442 | |
443 | return region; |
444 | } |
445 | |
446 | /* Release the memory associated with the region tree rooted at REGION. */ |
447 | |
448 | static void |
449 | free_omp_region_1 (struct omp_region *region) |
450 | { |
451 | struct omp_region *i, *n; |
452 | |
453 | for (i = region->inner; i ; i = n) |
454 | { |
455 | n = i->next; |
456 | free_omp_region_1 (region: i); |
457 | } |
458 | |
459 | free (ptr: region); |
460 | } |
461 | |
462 | /* Release the memory for the entire omp region tree. */ |
463 | |
464 | void |
465 | omp_free_regions (void) |
466 | { |
467 | struct omp_region *r, *n; |
468 | for (r = root_omp_region; r ; r = n) |
469 | { |
470 | n = r->next; |
471 | free_omp_region_1 (region: r); |
472 | } |
473 | root_omp_region = NULL; |
474 | } |
475 | |
476 | /* A convenience function to build an empty GIMPLE_COND with just the |
477 | condition. */ |
478 | |
479 | static gcond * |
480 | gimple_build_cond_empty (tree cond) |
481 | { |
482 | enum tree_code pred_code; |
483 | tree lhs, rhs; |
484 | |
485 | gimple_cond_get_ops_from_tree (cond, &pred_code, &lhs, &rhs); |
486 | return gimple_build_cond (pred_code, lhs, rhs, NULL_TREE, NULL_TREE); |
487 | } |
488 | |
489 | /* Change DECL_CONTEXT of CHILD_FNDECL to that of the parent function. |
490 | Add CHILD_FNDECL to decl chain of the supercontext of the block |
491 | ENTRY_BLOCK - this is the block which originally contained the |
492 | code from which CHILD_FNDECL was created. |
493 | |
494 | Together, these actions ensure that the debug info for the outlined |
495 | function will be emitted with the correct lexical scope. */ |
496 | |
497 | static void |
498 | adjust_context_and_scope (struct omp_region *region, tree entry_block, |
499 | tree child_fndecl) |
500 | { |
501 | tree parent_fndecl = NULL_TREE; |
502 | gimple *entry_stmt; |
503 | /* OMP expansion expands inner regions before outer ones, so if |
504 | we e.g. have explicit task region nested in parallel region, when |
505 | expanding the task region current_function_decl will be the original |
506 | source function, but we actually want to use as context the child |
507 | function of the parallel. */ |
508 | for (region = region->outer; |
509 | region && parent_fndecl == NULL_TREE; region = region->outer) |
510 | switch (region->type) |
511 | { |
512 | case GIMPLE_OMP_PARALLEL: |
513 | case GIMPLE_OMP_TASK: |
514 | case GIMPLE_OMP_TEAMS: |
515 | entry_stmt = last_nondebug_stmt (region->entry); |
516 | parent_fndecl = gimple_omp_taskreg_child_fn (gs: entry_stmt); |
517 | break; |
518 | case GIMPLE_OMP_TARGET: |
519 | entry_stmt = last_nondebug_stmt (region->entry); |
520 | parent_fndecl |
521 | = gimple_omp_target_child_fn (omp_target_stmt: as_a <gomp_target *> (p: entry_stmt)); |
522 | break; |
523 | default: |
524 | break; |
525 | } |
526 | |
527 | if (parent_fndecl == NULL_TREE) |
528 | parent_fndecl = current_function_decl; |
529 | DECL_CONTEXT (child_fndecl) = parent_fndecl; |
530 | |
531 | if (entry_block != NULL_TREE && TREE_CODE (entry_block) == BLOCK) |
532 | { |
533 | tree b = BLOCK_SUPERCONTEXT (entry_block); |
534 | if (TREE_CODE (b) == BLOCK) |
535 | { |
536 | DECL_CHAIN (child_fndecl) = BLOCK_VARS (b); |
537 | BLOCK_VARS (b) = child_fndecl; |
538 | } |
539 | } |
540 | } |
541 | |
542 | /* Build the function calls to GOMP_parallel etc to actually |
543 | generate the parallel operation. REGION is the parallel region |
544 | being expanded. BB is the block where to insert the code. WS_ARGS |
545 | will be set if this is a call to a combined parallel+workshare |
546 | construct, it contains the list of additional arguments needed by |
547 | the workshare construct. */ |
548 | |
549 | static void |
550 | expand_parallel_call (struct omp_region *region, basic_block bb, |
551 | gomp_parallel *entry_stmt, |
552 | vec<tree, va_gc> *ws_args) |
553 | { |
554 | tree t, t1, t2, val, cond, c, clauses, flags; |
555 | gimple_stmt_iterator gsi; |
556 | gimple *stmt; |
557 | enum built_in_function start_ix; |
558 | int start_ix2; |
559 | location_t clause_loc; |
560 | vec<tree, va_gc> *args; |
561 | |
562 | clauses = gimple_omp_parallel_clauses (gs: entry_stmt); |
563 | |
564 | /* Determine what flavor of GOMP_parallel we will be |
565 | emitting. */ |
566 | start_ix = BUILT_IN_GOMP_PARALLEL; |
567 | tree rtmp = omp_find_clause (clauses, kind: OMP_CLAUSE__REDUCTEMP_); |
568 | if (rtmp) |
569 | start_ix = BUILT_IN_GOMP_PARALLEL_REDUCTIONS; |
570 | else if (is_combined_parallel (region)) |
571 | { |
572 | switch (region->inner->type) |
573 | { |
574 | case GIMPLE_OMP_FOR: |
575 | gcc_assert (region->inner->sched_kind != OMP_CLAUSE_SCHEDULE_AUTO); |
576 | switch (region->inner->sched_kind) |
577 | { |
578 | case OMP_CLAUSE_SCHEDULE_RUNTIME: |
579 | /* For lastprivate(conditional:), our implementation |
580 | requires monotonic behavior. */ |
581 | if (region->inner->has_lastprivate_conditional != 0) |
582 | start_ix2 = 3; |
583 | else if ((region->inner->sched_modifiers |
584 | & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0) |
585 | start_ix2 = 6; |
586 | else if ((region->inner->sched_modifiers |
587 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0) |
588 | start_ix2 = 7; |
589 | else |
590 | start_ix2 = 3; |
591 | break; |
592 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: |
593 | case OMP_CLAUSE_SCHEDULE_GUIDED: |
594 | if ((region->inner->sched_modifiers |
595 | & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 |
596 | && !region->inner->has_lastprivate_conditional) |
597 | { |
598 | start_ix2 = 3 + region->inner->sched_kind; |
599 | break; |
600 | } |
601 | /* FALLTHRU */ |
602 | default: |
603 | start_ix2 = region->inner->sched_kind; |
604 | break; |
605 | } |
606 | start_ix2 += (int) BUILT_IN_GOMP_PARALLEL_LOOP_STATIC; |
607 | start_ix = (enum built_in_function) start_ix2; |
608 | break; |
609 | case GIMPLE_OMP_SECTIONS: |
610 | start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS; |
611 | break; |
612 | default: |
613 | gcc_unreachable (); |
614 | } |
615 | } |
616 | |
617 | /* By default, the value of NUM_THREADS is zero (selected at run time) |
618 | and there is no conditional. */ |
619 | cond = NULL_TREE; |
620 | val = build_int_cst (unsigned_type_node, 0); |
621 | flags = build_int_cst (unsigned_type_node, 0); |
622 | |
623 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_IF); |
624 | if (c) |
625 | cond = OMP_CLAUSE_IF_EXPR (c); |
626 | |
627 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_NUM_THREADS); |
628 | if (c) |
629 | { |
630 | val = OMP_CLAUSE_NUM_THREADS_EXPR (c); |
631 | clause_loc = OMP_CLAUSE_LOCATION (c); |
632 | } |
633 | else |
634 | clause_loc = gimple_location (g: entry_stmt); |
635 | |
636 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_PROC_BIND); |
637 | if (c) |
638 | flags = build_int_cst (unsigned_type_node, OMP_CLAUSE_PROC_BIND_KIND (c)); |
639 | |
640 | /* Ensure 'val' is of the correct type. */ |
641 | val = fold_convert_loc (clause_loc, unsigned_type_node, val); |
642 | |
643 | /* If we found the clause 'if (cond)', build either |
644 | (cond != 0) or (cond ? val : 1u). */ |
645 | if (cond) |
646 | { |
647 | cond = gimple_boolify (cond); |
648 | |
649 | if (integer_zerop (val)) |
650 | val = fold_build2_loc (clause_loc, |
651 | EQ_EXPR, unsigned_type_node, cond, |
652 | build_int_cst (TREE_TYPE (cond), 0)); |
653 | else |
654 | { |
655 | basic_block cond_bb, then_bb, else_bb; |
656 | edge e, e_then, e_else; |
657 | tree tmp_then, tmp_else, tmp_join, tmp_var; |
658 | |
659 | tmp_var = create_tmp_var (TREE_TYPE (val)); |
660 | if (gimple_in_ssa_p (cfun)) |
661 | { |
662 | tmp_then = make_ssa_name (var: tmp_var); |
663 | tmp_else = make_ssa_name (var: tmp_var); |
664 | tmp_join = make_ssa_name (var: tmp_var); |
665 | } |
666 | else |
667 | { |
668 | tmp_then = tmp_var; |
669 | tmp_else = tmp_var; |
670 | tmp_join = tmp_var; |
671 | } |
672 | |
673 | e = split_block_after_labels (bb); |
674 | cond_bb = e->src; |
675 | bb = e->dest; |
676 | remove_edge (e); |
677 | |
678 | then_bb = create_empty_bb (cond_bb); |
679 | else_bb = create_empty_bb (then_bb); |
680 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
681 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); |
682 | |
683 | stmt = gimple_build_cond_empty (cond); |
684 | gsi = gsi_start_bb (bb: cond_bb); |
685 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
686 | |
687 | gsi = gsi_start_bb (bb: then_bb); |
688 | expand_omp_build_assign (&gsi, tmp_then, val, true); |
689 | |
690 | gsi = gsi_start_bb (bb: else_bb); |
691 | expand_omp_build_assign (&gsi, tmp_else, |
692 | build_int_cst (unsigned_type_node, 1), |
693 | true); |
694 | |
695 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); |
696 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); |
697 | add_bb_to_loop (then_bb, cond_bb->loop_father); |
698 | add_bb_to_loop (else_bb, cond_bb->loop_father); |
699 | e_then = make_edge (then_bb, bb, EDGE_FALLTHRU); |
700 | e_else = make_edge (else_bb, bb, EDGE_FALLTHRU); |
701 | |
702 | if (gimple_in_ssa_p (cfun)) |
703 | { |
704 | gphi *phi = create_phi_node (tmp_join, bb); |
705 | add_phi_arg (phi, tmp_then, e_then, UNKNOWN_LOCATION); |
706 | add_phi_arg (phi, tmp_else, e_else, UNKNOWN_LOCATION); |
707 | } |
708 | |
709 | val = tmp_join; |
710 | } |
711 | |
712 | gsi = gsi_start_bb (bb); |
713 | val = force_gimple_operand_gsi (&gsi, val, true, NULL_TREE, |
714 | false, GSI_CONTINUE_LINKING); |
715 | } |
716 | |
717 | gsi = gsi_last_nondebug_bb (bb); |
718 | t = gimple_omp_parallel_data_arg (omp_parallel_stmt: entry_stmt); |
719 | if (t == NULL) |
720 | t1 = null_pointer_node; |
721 | else |
722 | t1 = build_fold_addr_expr (t); |
723 | tree child_fndecl = gimple_omp_parallel_child_fn (omp_parallel_stmt: entry_stmt); |
724 | t2 = build_fold_addr_expr (child_fndecl); |
725 | |
726 | vec_alloc (v&: args, nelems: 4 + vec_safe_length (v: ws_args)); |
727 | args->quick_push (obj: t2); |
728 | args->quick_push (obj: t1); |
729 | args->quick_push (obj: val); |
730 | if (ws_args) |
731 | args->splice (src: *ws_args); |
732 | args->quick_push (obj: flags); |
733 | |
734 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, |
735 | builtin_decl_explicit (fncode: start_ix), args); |
736 | |
737 | if (rtmp) |
738 | { |
739 | tree type = TREE_TYPE (OMP_CLAUSE_DECL (rtmp)); |
740 | t = build2 (MODIFY_EXPR, type, OMP_CLAUSE_DECL (rtmp), |
741 | fold_convert (type, |
742 | fold_convert (pointer_sized_int_node, t))); |
743 | } |
744 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
745 | false, GSI_CONTINUE_LINKING); |
746 | } |
747 | |
748 | /* Build the function call to GOMP_task to actually |
749 | generate the task operation. BB is the block where to insert the code. */ |
750 | |
751 | static void |
752 | expand_task_call (struct omp_region *region, basic_block bb, |
753 | gomp_task *entry_stmt) |
754 | { |
755 | tree t1, t2, t3; |
756 | gimple_stmt_iterator gsi; |
757 | location_t loc = gimple_location (g: entry_stmt); |
758 | |
759 | tree clauses = gimple_omp_task_clauses (gs: entry_stmt); |
760 | |
761 | tree ifc = omp_find_clause (clauses, kind: OMP_CLAUSE_IF); |
762 | tree untied = omp_find_clause (clauses, kind: OMP_CLAUSE_UNTIED); |
763 | tree mergeable = omp_find_clause (clauses, kind: OMP_CLAUSE_MERGEABLE); |
764 | tree depend = omp_find_clause (clauses, kind: OMP_CLAUSE_DEPEND); |
765 | tree finalc = omp_find_clause (clauses, kind: OMP_CLAUSE_FINAL); |
766 | tree priority = omp_find_clause (clauses, kind: OMP_CLAUSE_PRIORITY); |
767 | tree detach = omp_find_clause (clauses, kind: OMP_CLAUSE_DETACH); |
768 | |
769 | unsigned int iflags |
770 | = (untied ? GOMP_TASK_FLAG_UNTIED : 0) |
771 | | (mergeable ? GOMP_TASK_FLAG_MERGEABLE : 0) |
772 | | (depend ? GOMP_TASK_FLAG_DEPEND : 0); |
773 | |
774 | bool taskloop_p = gimple_omp_task_taskloop_p (g: entry_stmt); |
775 | tree startvar = NULL_TREE, endvar = NULL_TREE, step = NULL_TREE; |
776 | tree num_tasks = NULL_TREE; |
777 | bool ull = false; |
778 | if (taskloop_p) |
779 | { |
780 | gimple *g = last_nondebug_stmt (region->outer->entry); |
781 | gcc_assert (gimple_code (g) == GIMPLE_OMP_FOR |
782 | && gimple_omp_for_kind (g) == GF_OMP_FOR_KIND_TASKLOOP); |
783 | struct omp_for_data fd; |
784 | omp_extract_for_data (for_stmt: as_a <gomp_for *> (p: g), fd: &fd, NULL); |
785 | startvar = omp_find_clause (clauses, kind: OMP_CLAUSE__LOOPTEMP_); |
786 | endvar = omp_find_clause (OMP_CLAUSE_CHAIN (startvar), |
787 | kind: OMP_CLAUSE__LOOPTEMP_); |
788 | startvar = OMP_CLAUSE_DECL (startvar); |
789 | endvar = OMP_CLAUSE_DECL (endvar); |
790 | step = fold_convert_loc (loc, fd.iter_type, fd.loop.step); |
791 | if (fd.loop.cond_code == LT_EXPR) |
792 | iflags |= GOMP_TASK_FLAG_UP; |
793 | tree tclauses = gimple_omp_for_clauses (gs: g); |
794 | num_tasks = omp_find_clause (clauses: tclauses, kind: OMP_CLAUSE_NUM_TASKS); |
795 | if (num_tasks) |
796 | { |
797 | if (OMP_CLAUSE_NUM_TASKS_STRICT (num_tasks)) |
798 | iflags |= GOMP_TASK_FLAG_STRICT; |
799 | num_tasks = OMP_CLAUSE_NUM_TASKS_EXPR (num_tasks); |
800 | } |
801 | else |
802 | { |
803 | num_tasks = omp_find_clause (clauses: tclauses, kind: OMP_CLAUSE_GRAINSIZE); |
804 | if (num_tasks) |
805 | { |
806 | iflags |= GOMP_TASK_FLAG_GRAINSIZE; |
807 | if (OMP_CLAUSE_GRAINSIZE_STRICT (num_tasks)) |
808 | iflags |= GOMP_TASK_FLAG_STRICT; |
809 | num_tasks = OMP_CLAUSE_GRAINSIZE_EXPR (num_tasks); |
810 | } |
811 | else |
812 | num_tasks = integer_zero_node; |
813 | } |
814 | num_tasks = fold_convert_loc (loc, long_integer_type_node, num_tasks); |
815 | if (ifc == NULL_TREE) |
816 | iflags |= GOMP_TASK_FLAG_IF; |
817 | if (omp_find_clause (clauses: tclauses, kind: OMP_CLAUSE_NOGROUP)) |
818 | iflags |= GOMP_TASK_FLAG_NOGROUP; |
819 | ull = fd.iter_type == long_long_unsigned_type_node; |
820 | if (omp_find_clause (clauses, kind: OMP_CLAUSE_REDUCTION)) |
821 | iflags |= GOMP_TASK_FLAG_REDUCTION; |
822 | } |
823 | else |
824 | { |
825 | if (priority) |
826 | iflags |= GOMP_TASK_FLAG_PRIORITY; |
827 | if (detach) |
828 | iflags |= GOMP_TASK_FLAG_DETACH; |
829 | } |
830 | |
831 | tree flags = build_int_cst (unsigned_type_node, iflags); |
832 | |
833 | tree cond = boolean_true_node; |
834 | if (ifc) |
835 | { |
836 | if (taskloop_p) |
837 | { |
838 | tree t = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); |
839 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, |
840 | build_int_cst (unsigned_type_node, |
841 | GOMP_TASK_FLAG_IF), |
842 | build_int_cst (unsigned_type_node, 0)); |
843 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, |
844 | flags, t); |
845 | } |
846 | else |
847 | cond = gimple_boolify (OMP_CLAUSE_IF_EXPR (ifc)); |
848 | } |
849 | |
850 | if (finalc) |
851 | { |
852 | tree t = gimple_boolify (OMP_CLAUSE_FINAL_EXPR (finalc)); |
853 | t = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, t, |
854 | build_int_cst (unsigned_type_node, |
855 | GOMP_TASK_FLAG_FINAL), |
856 | build_int_cst (unsigned_type_node, 0)); |
857 | flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, flags, t); |
858 | } |
859 | if (depend) |
860 | depend = OMP_CLAUSE_DECL (depend); |
861 | else |
862 | depend = build_int_cst (ptr_type_node, 0); |
863 | if (priority) |
864 | priority = fold_convert (integer_type_node, |
865 | OMP_CLAUSE_PRIORITY_EXPR (priority)); |
866 | else |
867 | priority = integer_zero_node; |
868 | |
869 | gsi = gsi_last_nondebug_bb (bb); |
870 | |
871 | detach = (detach |
872 | ? build_fold_addr_expr (OMP_CLAUSE_DECL (detach)) |
873 | : null_pointer_node); |
874 | |
875 | tree t = gimple_omp_task_data_arg (gs: entry_stmt); |
876 | if (t == NULL) |
877 | t2 = null_pointer_node; |
878 | else |
879 | t2 = build_fold_addr_expr_loc (loc, t); |
880 | t1 = build_fold_addr_expr_loc (loc, gimple_omp_task_child_fn (gs: entry_stmt)); |
881 | t = gimple_omp_task_copy_fn (gs: entry_stmt); |
882 | if (t == NULL) |
883 | t3 = null_pointer_node; |
884 | else |
885 | t3 = build_fold_addr_expr_loc (loc, t); |
886 | |
887 | if (taskloop_p) |
888 | t = build_call_expr (ull |
889 | ? builtin_decl_explicit (fncode: BUILT_IN_GOMP_TASKLOOP_ULL) |
890 | : builtin_decl_explicit (fncode: BUILT_IN_GOMP_TASKLOOP), |
891 | 11, t1, t2, t3, |
892 | gimple_omp_task_arg_size (gs: entry_stmt), |
893 | gimple_omp_task_arg_align (gs: entry_stmt), flags, |
894 | num_tasks, priority, startvar, endvar, step); |
895 | else |
896 | t = build_call_expr (builtin_decl_explicit (fncode: BUILT_IN_GOMP_TASK), |
897 | 10, t1, t2, t3, |
898 | gimple_omp_task_arg_size (gs: entry_stmt), |
899 | gimple_omp_task_arg_align (gs: entry_stmt), cond, flags, |
900 | depend, priority, detach); |
901 | |
902 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
903 | false, GSI_CONTINUE_LINKING); |
904 | } |
905 | |
906 | /* Build the function call to GOMP_taskwait_depend to actually |
907 | generate the taskwait operation. BB is the block where to insert the |
908 | code. */ |
909 | |
910 | static void |
911 | expand_taskwait_call (basic_block bb, gomp_task *entry_stmt) |
912 | { |
913 | tree clauses = gimple_omp_task_clauses (gs: entry_stmt); |
914 | tree depend = omp_find_clause (clauses, kind: OMP_CLAUSE_DEPEND); |
915 | if (depend == NULL_TREE) |
916 | return; |
917 | |
918 | depend = OMP_CLAUSE_DECL (depend); |
919 | |
920 | bool nowait = omp_find_clause (clauses, kind: OMP_CLAUSE_NOWAIT) != NULL_TREE; |
921 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); |
922 | enum built_in_function f = (nowait |
923 | ? BUILT_IN_GOMP_TASKWAIT_DEPEND_NOWAIT |
924 | : BUILT_IN_GOMP_TASKWAIT_DEPEND); |
925 | tree t = build_call_expr (builtin_decl_explicit (fncode: f), 1, depend); |
926 | |
927 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
928 | false, GSI_CONTINUE_LINKING); |
929 | } |
930 | |
931 | /* Build the function call to GOMP_teams_reg to actually |
932 | generate the host teams operation. REGION is the teams region |
933 | being expanded. BB is the block where to insert the code. */ |
934 | |
935 | static void |
936 | expand_teams_call (basic_block bb, gomp_teams *entry_stmt) |
937 | { |
938 | tree clauses = gimple_omp_teams_clauses (gs: entry_stmt); |
939 | tree num_teams = omp_find_clause (clauses, kind: OMP_CLAUSE_NUM_TEAMS); |
940 | if (num_teams == NULL_TREE) |
941 | num_teams = build_int_cst (unsigned_type_node, 0); |
942 | else |
943 | { |
944 | num_teams = OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR (num_teams); |
945 | num_teams = fold_convert (unsigned_type_node, num_teams); |
946 | } |
947 | tree thread_limit = omp_find_clause (clauses, kind: OMP_CLAUSE_THREAD_LIMIT); |
948 | if (thread_limit == NULL_TREE) |
949 | thread_limit = build_int_cst (unsigned_type_node, 0); |
950 | else |
951 | { |
952 | thread_limit = OMP_CLAUSE_THREAD_LIMIT_EXPR (thread_limit); |
953 | thread_limit = fold_convert (unsigned_type_node, thread_limit); |
954 | } |
955 | |
956 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb); |
957 | tree t = gimple_omp_teams_data_arg (omp_teams_stmt: entry_stmt), t1; |
958 | if (t == NULL) |
959 | t1 = null_pointer_node; |
960 | else |
961 | t1 = build_fold_addr_expr (t); |
962 | tree child_fndecl = gimple_omp_teams_child_fn (omp_teams_stmt: entry_stmt); |
963 | tree t2 = build_fold_addr_expr (child_fndecl); |
964 | |
965 | vec<tree, va_gc> *args; |
966 | vec_alloc (v&: args, nelems: 5); |
967 | args->quick_push (obj: t2); |
968 | args->quick_push (obj: t1); |
969 | args->quick_push (obj: num_teams); |
970 | args->quick_push (obj: thread_limit); |
971 | /* For future extensibility. */ |
972 | args->quick_push (obj: build_zero_cst (unsigned_type_node)); |
973 | |
974 | t = build_call_expr_loc_vec (UNKNOWN_LOCATION, |
975 | builtin_decl_explicit (fncode: BUILT_IN_GOMP_TEAMS_REG), |
976 | args); |
977 | |
978 | force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
979 | false, GSI_CONTINUE_LINKING); |
980 | } |
981 | |
982 | /* Chain all the DECLs in LIST by their TREE_CHAIN fields. */ |
983 | |
984 | static tree |
985 | vec2chain (vec<tree, va_gc> *v) |
986 | { |
987 | tree chain = NULL_TREE, t; |
988 | unsigned ix; |
989 | |
990 | FOR_EACH_VEC_SAFE_ELT_REVERSE (v, ix, t) |
991 | { |
992 | DECL_CHAIN (t) = chain; |
993 | chain = t; |
994 | } |
995 | |
996 | return chain; |
997 | } |
998 | |
999 | /* Remove barriers in REGION->EXIT's block. Note that this is only |
1000 | valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region |
1001 | is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that |
1002 | left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be |
1003 | removed. */ |
1004 | |
1005 | static void |
1006 | remove_exit_barrier (struct omp_region *region) |
1007 | { |
1008 | gimple_stmt_iterator gsi; |
1009 | basic_block exit_bb; |
1010 | edge_iterator ei; |
1011 | edge e; |
1012 | gimple *stmt; |
1013 | int any_addressable_vars = -1; |
1014 | |
1015 | exit_bb = region->exit; |
1016 | |
1017 | /* If the parallel region doesn't return, we don't have REGION->EXIT |
1018 | block at all. */ |
1019 | if (! exit_bb) |
1020 | return; |
1021 | |
1022 | /* The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The |
1023 | workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of |
1024 | statements that can appear in between are extremely limited -- no |
1025 | memory operations at all. Here, we allow nothing at all, so the |
1026 | only thing we allow to precede this GIMPLE_OMP_RETURN is a label. */ |
1027 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
1028 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
1029 | gsi_prev_nondebug (i: &gsi); |
1030 | if (!gsi_end_p (i: gsi) && gimple_code (g: gsi_stmt (i: gsi)) != GIMPLE_LABEL) |
1031 | return; |
1032 | |
1033 | FOR_EACH_EDGE (e, ei, exit_bb->preds) |
1034 | { |
1035 | gsi = gsi_last_nondebug_bb (bb: e->src); |
1036 | if (gsi_end_p (i: gsi)) |
1037 | continue; |
1038 | stmt = gsi_stmt (i: gsi); |
1039 | if (gimple_code (g: stmt) == GIMPLE_OMP_RETURN |
1040 | && !gimple_omp_return_nowait_p (g: stmt)) |
1041 | { |
1042 | /* OpenMP 3.0 tasks unfortunately prevent this optimization |
1043 | in many cases. If there could be tasks queued, the barrier |
1044 | might be needed to let the tasks run before some local |
1045 | variable of the parallel that the task uses as shared |
1046 | runs out of scope. The task can be spawned either |
1047 | from within current function (this would be easy to check) |
1048 | or from some function it calls and gets passed an address |
1049 | of such a variable. */ |
1050 | if (any_addressable_vars < 0) |
1051 | { |
1052 | gomp_parallel *parallel_stmt |
1053 | = as_a <gomp_parallel *> (p: last_nondebug_stmt (region->entry)); |
1054 | tree child_fun = gimple_omp_parallel_child_fn (omp_parallel_stmt: parallel_stmt); |
1055 | tree local_decls, block, decl; |
1056 | unsigned ix; |
1057 | |
1058 | any_addressable_vars = 0; |
1059 | FOR_EACH_LOCAL_DECL (DECL_STRUCT_FUNCTION (child_fun), ix, decl) |
1060 | if (TREE_ADDRESSABLE (decl)) |
1061 | { |
1062 | any_addressable_vars = 1; |
1063 | break; |
1064 | } |
1065 | for (block = gimple_block (g: stmt); |
1066 | !any_addressable_vars |
1067 | && block |
1068 | && TREE_CODE (block) == BLOCK; |
1069 | block = BLOCK_SUPERCONTEXT (block)) |
1070 | { |
1071 | for (local_decls = BLOCK_VARS (block); |
1072 | local_decls; |
1073 | local_decls = DECL_CHAIN (local_decls)) |
1074 | if (TREE_ADDRESSABLE (local_decls)) |
1075 | { |
1076 | any_addressable_vars = 1; |
1077 | break; |
1078 | } |
1079 | if (block == gimple_block (g: parallel_stmt)) |
1080 | break; |
1081 | } |
1082 | } |
1083 | if (!any_addressable_vars) |
1084 | gimple_omp_return_set_nowait (s: stmt); |
1085 | } |
1086 | } |
1087 | } |
1088 | |
1089 | static void |
1090 | remove_exit_barriers (struct omp_region *region) |
1091 | { |
1092 | if (region->type == GIMPLE_OMP_PARALLEL) |
1093 | remove_exit_barrier (region); |
1094 | |
1095 | if (region->inner) |
1096 | { |
1097 | region = region->inner; |
1098 | remove_exit_barriers (region); |
1099 | while (region->next) |
1100 | { |
1101 | region = region->next; |
1102 | remove_exit_barriers (region); |
1103 | } |
1104 | } |
1105 | } |
1106 | |
1107 | /* Optimize omp_get_thread_num () and omp_get_num_threads () |
1108 | calls. These can't be declared as const functions, but |
1109 | within one parallel body they are constant, so they can be |
1110 | transformed there into __builtin_omp_get_{thread_num,num_threads} () |
1111 | which are declared const. Similarly for task body, except |
1112 | that in untied task omp_get_thread_num () can change at any task |
1113 | scheduling point. */ |
1114 | |
1115 | static void |
1116 | optimize_omp_library_calls (gimple *entry_stmt) |
1117 | { |
1118 | basic_block bb; |
1119 | gimple_stmt_iterator gsi; |
1120 | tree thr_num_tree = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_THREAD_NUM); |
1121 | tree thr_num_id = DECL_ASSEMBLER_NAME (thr_num_tree); |
1122 | tree num_thr_tree = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_NUM_THREADS); |
1123 | tree num_thr_id = DECL_ASSEMBLER_NAME (num_thr_tree); |
1124 | bool untied_task = (gimple_code (g: entry_stmt) == GIMPLE_OMP_TASK |
1125 | && omp_find_clause (clauses: gimple_omp_task_clauses (gs: entry_stmt), |
1126 | kind: OMP_CLAUSE_UNTIED) != NULL); |
1127 | |
1128 | FOR_EACH_BB_FN (bb, cfun) |
1129 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
1130 | { |
1131 | gimple *call = gsi_stmt (i: gsi); |
1132 | tree decl; |
1133 | |
1134 | if (is_gimple_call (gs: call) |
1135 | && (decl = gimple_call_fndecl (gs: call)) |
1136 | && DECL_EXTERNAL (decl) |
1137 | && TREE_PUBLIC (decl) |
1138 | && DECL_INITIAL (decl) == NULL) |
1139 | { |
1140 | tree built_in; |
1141 | |
1142 | if (DECL_NAME (decl) == thr_num_id) |
1143 | { |
1144 | /* In #pragma omp task untied omp_get_thread_num () can change |
1145 | during the execution of the task region. */ |
1146 | if (untied_task) |
1147 | continue; |
1148 | built_in = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_THREAD_NUM); |
1149 | } |
1150 | else if (DECL_NAME (decl) == num_thr_id) |
1151 | built_in = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_NUM_THREADS); |
1152 | else |
1153 | continue; |
1154 | |
1155 | if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in) |
1156 | || gimple_call_num_args (gs: call) != 0) |
1157 | continue; |
1158 | |
1159 | if (flag_exceptions && !TREE_NOTHROW (decl)) |
1160 | continue; |
1161 | |
1162 | if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE |
1163 | || !types_compatible_p (TREE_TYPE (TREE_TYPE (decl)), |
1164 | TREE_TYPE (TREE_TYPE (built_in)))) |
1165 | continue; |
1166 | |
1167 | gimple_call_set_fndecl (gs: call, decl: built_in); |
1168 | } |
1169 | } |
1170 | } |
1171 | |
1172 | /* Callback for expand_omp_build_assign. Return non-NULL if *tp needs to be |
1173 | regimplified. */ |
1174 | |
1175 | static tree |
1176 | expand_omp_regimplify_p (tree *tp, int *walk_subtrees, void *) |
1177 | { |
1178 | tree t = *tp; |
1179 | |
1180 | /* Any variable with DECL_VALUE_EXPR needs to be regimplified. */ |
1181 | if (VAR_P (t) && DECL_HAS_VALUE_EXPR_P (t)) |
1182 | return t; |
1183 | |
1184 | if (TREE_CODE (t) == ADDR_EXPR) |
1185 | recompute_tree_invariant_for_addr_expr (t); |
1186 | |
1187 | *walk_subtrees = !TYPE_P (t) && !DECL_P (t); |
1188 | return NULL_TREE; |
1189 | } |
1190 | |
1191 | /* Prepend or append TO = FROM assignment before or after *GSI_P. */ |
1192 | |
1193 | static void |
1194 | expand_omp_build_assign (gimple_stmt_iterator *gsi_p, tree to, tree from, |
1195 | bool after) |
1196 | { |
1197 | bool simple_p = DECL_P (to) && TREE_ADDRESSABLE (to); |
1198 | from = force_gimple_operand_gsi (gsi_p, from, simple_p, NULL_TREE, |
1199 | !after, after ? GSI_CONTINUE_LINKING |
1200 | : GSI_SAME_STMT); |
1201 | gimple *stmt = gimple_build_assign (to, from); |
1202 | if (after) |
1203 | gsi_insert_after (gsi_p, stmt, GSI_CONTINUE_LINKING); |
1204 | else |
1205 | gsi_insert_before (gsi_p, stmt, GSI_SAME_STMT); |
1206 | if (walk_tree (&from, expand_omp_regimplify_p, NULL, NULL) |
1207 | || walk_tree (&to, expand_omp_regimplify_p, NULL, NULL)) |
1208 | { |
1209 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
1210 | gimple_regimplify_operands (stmt, &gsi); |
1211 | } |
1212 | } |
1213 | |
1214 | /* Prepend or append LHS CODE RHS condition before or after *GSI_P. */ |
1215 | |
1216 | static gcond * |
1217 | expand_omp_build_cond (gimple_stmt_iterator *gsi_p, enum tree_code code, |
1218 | tree lhs, tree rhs, bool after = false) |
1219 | { |
1220 | gcond *cond_stmt = gimple_build_cond (code, lhs, rhs, NULL_TREE, NULL_TREE); |
1221 | if (after) |
1222 | gsi_insert_after (gsi_p, cond_stmt, GSI_CONTINUE_LINKING); |
1223 | else |
1224 | gsi_insert_before (gsi_p, cond_stmt, GSI_SAME_STMT); |
1225 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), expand_omp_regimplify_p, |
1226 | NULL, NULL) |
1227 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), expand_omp_regimplify_p, |
1228 | NULL, NULL)) |
1229 | { |
1230 | gimple_stmt_iterator gsi = gsi_for_stmt (cond_stmt); |
1231 | gimple_regimplify_operands (cond_stmt, &gsi); |
1232 | } |
1233 | return cond_stmt; |
1234 | } |
1235 | |
1236 | /* Expand the OpenMP parallel or task directive starting at REGION. */ |
1237 | |
1238 | static void |
1239 | expand_omp_taskreg (struct omp_region *region) |
1240 | { |
1241 | basic_block entry_bb, exit_bb, new_bb; |
1242 | struct function *child_cfun; |
1243 | tree child_fn, block, t; |
1244 | gimple_stmt_iterator gsi; |
1245 | gimple *entry_stmt, *stmt; |
1246 | edge e; |
1247 | vec<tree, va_gc> *ws_args; |
1248 | |
1249 | entry_stmt = last_nondebug_stmt (region->entry); |
1250 | if (gimple_code (g: entry_stmt) == GIMPLE_OMP_TASK |
1251 | && gimple_omp_task_taskwait_p (g: entry_stmt)) |
1252 | { |
1253 | new_bb = region->entry; |
1254 | gsi = gsi_last_nondebug_bb (bb: region->entry); |
1255 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK); |
1256 | gsi_remove (&gsi, true); |
1257 | expand_taskwait_call (bb: new_bb, entry_stmt: as_a <gomp_task *> (p: entry_stmt)); |
1258 | return; |
1259 | } |
1260 | |
1261 | child_fn = gimple_omp_taskreg_child_fn (gs: entry_stmt); |
1262 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); |
1263 | |
1264 | entry_bb = region->entry; |
1265 | if (gimple_code (g: entry_stmt) == GIMPLE_OMP_TASK) |
1266 | exit_bb = region->cont; |
1267 | else |
1268 | exit_bb = region->exit; |
1269 | |
1270 | if (is_combined_parallel (region)) |
1271 | ws_args = region->ws_args; |
1272 | else |
1273 | ws_args = NULL; |
1274 | |
1275 | if (child_cfun->cfg) |
1276 | { |
1277 | /* Due to inlining, it may happen that we have already outlined |
1278 | the region, in which case all we need to do is make the |
1279 | sub-graph unreachable and emit the parallel call. */ |
1280 | edge entry_succ_e, exit_succ_e; |
1281 | |
1282 | entry_succ_e = single_succ_edge (bb: entry_bb); |
1283 | |
1284 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
1285 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_PARALLEL |
1286 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK |
1287 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TEAMS); |
1288 | gsi_remove (&gsi, true); |
1289 | |
1290 | new_bb = entry_bb; |
1291 | if (exit_bb) |
1292 | { |
1293 | exit_succ_e = single_succ_edge (bb: exit_bb); |
1294 | make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU); |
1295 | } |
1296 | remove_edge_and_dominated_blocks (entry_succ_e); |
1297 | } |
1298 | else |
1299 | { |
1300 | unsigned srcidx, dstidx, num; |
1301 | |
1302 | /* If the parallel region needs data sent from the parent |
1303 | function, then the very first statement (except possible |
1304 | tree profile counter updates) of the parallel body |
1305 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since |
1306 | &.OMP_DATA_O is passed as an argument to the child function, |
1307 | we need to replace it with the argument as seen by the child |
1308 | function. |
1309 | |
1310 | In most cases, this will end up being the identity assignment |
1311 | .OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had |
1312 | a function call that has been inlined, the original PARM_DECL |
1313 | .OMP_DATA_I may have been converted into a different local |
1314 | variable. In which case, we need to keep the assignment. */ |
1315 | if (gimple_omp_taskreg_data_arg (gs: entry_stmt)) |
1316 | { |
1317 | basic_block entry_succ_bb |
1318 | = single_succ_p (bb: entry_bb) ? single_succ (bb: entry_bb) |
1319 | : FALLTHRU_EDGE (entry_bb)->dest; |
1320 | tree arg; |
1321 | gimple *parcopy_stmt = NULL; |
1322 | |
1323 | for (gsi = gsi_start_bb (bb: entry_succ_bb); ; gsi_next (i: &gsi)) |
1324 | { |
1325 | gimple *stmt; |
1326 | |
1327 | gcc_assert (!gsi_end_p (gsi)); |
1328 | stmt = gsi_stmt (i: gsi); |
1329 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
1330 | continue; |
1331 | |
1332 | if (gimple_num_ops (gs: stmt) == 2) |
1333 | { |
1334 | tree arg = gimple_assign_rhs1 (gs: stmt); |
1335 | |
1336 | /* We're ignore the subcode because we're |
1337 | effectively doing a STRIP_NOPS. */ |
1338 | |
1339 | if (TREE_CODE (arg) == ADDR_EXPR |
1340 | && (TREE_OPERAND (arg, 0) |
1341 | == gimple_omp_taskreg_data_arg (gs: entry_stmt))) |
1342 | { |
1343 | parcopy_stmt = stmt; |
1344 | break; |
1345 | } |
1346 | } |
1347 | } |
1348 | |
1349 | gcc_assert (parcopy_stmt != NULL); |
1350 | arg = DECL_ARGUMENTS (child_fn); |
1351 | |
1352 | if (!gimple_in_ssa_p (cfun)) |
1353 | { |
1354 | if (gimple_assign_lhs (gs: parcopy_stmt) == arg) |
1355 | gsi_remove (&gsi, true); |
1356 | else |
1357 | { |
1358 | /* ?? Is setting the subcode really necessary ?? */ |
1359 | gimple_omp_set_subcode (s: parcopy_stmt, TREE_CODE (arg)); |
1360 | gimple_assign_set_rhs1 (gs: parcopy_stmt, rhs: arg); |
1361 | } |
1362 | } |
1363 | else |
1364 | { |
1365 | tree lhs = gimple_assign_lhs (gs: parcopy_stmt); |
1366 | gcc_assert (SSA_NAME_VAR (lhs) == arg); |
1367 | /* We'd like to set the rhs to the default def in the child_fn, |
1368 | but it's too early to create ssa names in the child_fn. |
1369 | Instead, we set the rhs to the parm. In |
1370 | move_sese_region_to_fn, we introduce a default def for the |
1371 | parm, map the parm to it's default def, and once we encounter |
1372 | this stmt, replace the parm with the default def. */ |
1373 | gimple_assign_set_rhs1 (gs: parcopy_stmt, rhs: arg); |
1374 | update_stmt (s: parcopy_stmt); |
1375 | } |
1376 | } |
1377 | |
1378 | /* Declare local variables needed in CHILD_CFUN. */ |
1379 | block = DECL_INITIAL (child_fn); |
1380 | BLOCK_VARS (block) = vec2chain (v: child_cfun->local_decls); |
1381 | /* The gimplifier could record temporaries in parallel/task block |
1382 | rather than in containing function's local_decls chain, |
1383 | which would mean cgraph missed finalizing them. Do it now. */ |
1384 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) |
1385 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) |
1386 | varpool_node::finalize_decl (decl: t); |
1387 | DECL_SAVED_TREE (child_fn) = NULL; |
1388 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ |
1389 | gimple_set_body (child_fn, NULL); |
1390 | TREE_USED (block) = 1; |
1391 | |
1392 | /* Reset DECL_CONTEXT on function arguments. */ |
1393 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) |
1394 | DECL_CONTEXT (t) = child_fn; |
1395 | |
1396 | /* Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK, |
1397 | so that it can be moved to the child function. */ |
1398 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
1399 | stmt = gsi_stmt (i: gsi); |
1400 | gcc_assert (stmt && (gimple_code (stmt) == GIMPLE_OMP_PARALLEL |
1401 | || gimple_code (stmt) == GIMPLE_OMP_TASK |
1402 | || gimple_code (stmt) == GIMPLE_OMP_TEAMS)); |
1403 | e = split_block (entry_bb, stmt); |
1404 | gsi_remove (&gsi, true); |
1405 | entry_bb = e->dest; |
1406 | edge e2 = NULL; |
1407 | if (gimple_code (g: entry_stmt) != GIMPLE_OMP_TASK) |
1408 | single_succ_edge (bb: entry_bb)->flags = EDGE_FALLTHRU; |
1409 | else |
1410 | { |
1411 | e2 = make_edge (e->src, BRANCH_EDGE (entry_bb)->dest, EDGE_ABNORMAL); |
1412 | gcc_assert (e2->dest == region->exit); |
1413 | remove_edge (BRANCH_EDGE (entry_bb)); |
1414 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e->src); |
1415 | gsi = gsi_last_nondebug_bb (bb: region->exit); |
1416 | gcc_assert (!gsi_end_p (gsi) |
1417 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
1418 | gsi_remove (&gsi, true); |
1419 | } |
1420 | |
1421 | /* Convert GIMPLE_OMP_{RETURN,CONTINUE} into a RETURN_EXPR. */ |
1422 | if (exit_bb) |
1423 | { |
1424 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
1425 | gcc_assert (!gsi_end_p (gsi) |
1426 | && (gimple_code (gsi_stmt (gsi)) |
1427 | == (e2 ? GIMPLE_OMP_CONTINUE : GIMPLE_OMP_RETURN))); |
1428 | stmt = gimple_build_return (NULL); |
1429 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); |
1430 | gsi_remove (&gsi, true); |
1431 | } |
1432 | |
1433 | /* Move the parallel region into CHILD_CFUN. */ |
1434 | |
1435 | if (gimple_in_ssa_p (cfun)) |
1436 | { |
1437 | init_tree_ssa (child_cfun); |
1438 | init_ssa_operands (fn: child_cfun); |
1439 | child_cfun->gimple_df->in_ssa_p = true; |
1440 | block = NULL_TREE; |
1441 | } |
1442 | else |
1443 | block = gimple_block (g: entry_stmt); |
1444 | |
1445 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); |
1446 | if (exit_bb) |
1447 | single_succ_edge (bb: new_bb)->flags = EDGE_FALLTHRU; |
1448 | if (e2) |
1449 | { |
1450 | basic_block dest_bb = e2->dest; |
1451 | if (!exit_bb) |
1452 | make_edge (new_bb, dest_bb, EDGE_FALLTHRU); |
1453 | remove_edge (e2); |
1454 | set_immediate_dominator (CDI_DOMINATORS, dest_bb, new_bb); |
1455 | } |
1456 | /* When the OMP expansion process cannot guarantee an up-to-date |
1457 | loop tree arrange for the child function to fixup loops. */ |
1458 | if (loops_state_satisfies_p (flags: LOOPS_NEED_FIXUP)) |
1459 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; |
1460 | |
1461 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ |
1462 | num = vec_safe_length (v: child_cfun->local_decls); |
1463 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) |
1464 | { |
1465 | t = (*child_cfun->local_decls)[srcidx]; |
1466 | if (DECL_CONTEXT (t) == cfun->decl) |
1467 | continue; |
1468 | if (srcidx != dstidx) |
1469 | (*child_cfun->local_decls)[dstidx] = t; |
1470 | dstidx++; |
1471 | } |
1472 | if (dstidx != num) |
1473 | vec_safe_truncate (v: child_cfun->local_decls, size: dstidx); |
1474 | |
1475 | /* Inform the callgraph about the new function. */ |
1476 | child_cfun->curr_properties = cfun->curr_properties; |
1477 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; |
1478 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; |
1479 | cgraph_node *node = cgraph_node::get_create (child_fn); |
1480 | node->parallelized_function = 1; |
1481 | cgraph_node::add_new_function (fndecl: child_fn, lowered: true); |
1482 | |
1483 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) |
1484 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); |
1485 | |
1486 | /* Fix the callgraph edges for child_cfun. Those for cfun will be |
1487 | fixed in a following pass. */ |
1488 | push_cfun (new_cfun: child_cfun); |
1489 | if (need_asm) |
1490 | assign_assembler_name_if_needed (child_fn); |
1491 | |
1492 | if (optimize) |
1493 | optimize_omp_library_calls (entry_stmt); |
1494 | update_max_bb_count (); |
1495 | cgraph_edge::rebuild_edges (); |
1496 | |
1497 | /* Some EH regions might become dead, see PR34608. If |
1498 | pass_cleanup_cfg isn't the first pass to happen with the |
1499 | new child, these dead EH edges might cause problems. |
1500 | Clean them up now. */ |
1501 | if (flag_exceptions) |
1502 | { |
1503 | basic_block bb; |
1504 | bool changed = false; |
1505 | |
1506 | FOR_EACH_BB_FN (bb, cfun) |
1507 | changed |= gimple_purge_dead_eh_edges (bb); |
1508 | if (changed) |
1509 | cleanup_tree_cfg (); |
1510 | } |
1511 | if (gimple_in_ssa_p (cfun)) |
1512 | update_ssa (TODO_update_ssa); |
1513 | if (flag_checking && !loops_state_satisfies_p (flags: LOOPS_NEED_FIXUP)) |
1514 | verify_loop_structure (); |
1515 | pop_cfun (); |
1516 | |
1517 | if (dump_file && !gimple_in_ssa_p (cfun)) |
1518 | { |
1519 | omp_any_child_fn_dumped = true; |
1520 | dump_function_header (dump_file, child_fn, dump_flags); |
1521 | dump_function_to_file (child_fn, dump_file, dump_flags); |
1522 | } |
1523 | } |
1524 | |
1525 | adjust_context_and_scope (region, entry_block: gimple_block (g: entry_stmt), child_fndecl: child_fn); |
1526 | |
1527 | if (gimple_code (g: entry_stmt) == GIMPLE_OMP_PARALLEL) |
1528 | expand_parallel_call (region, bb: new_bb, |
1529 | entry_stmt: as_a <gomp_parallel *> (p: entry_stmt), ws_args); |
1530 | else if (gimple_code (g: entry_stmt) == GIMPLE_OMP_TEAMS) |
1531 | expand_teams_call (bb: new_bb, entry_stmt: as_a <gomp_teams *> (p: entry_stmt)); |
1532 | else |
1533 | expand_task_call (region, bb: new_bb, entry_stmt: as_a <gomp_task *> (p: entry_stmt)); |
1534 | } |
1535 | |
1536 | /* Information about members of an OpenACC collapsed loop nest. */ |
1537 | |
1538 | struct oacc_collapse |
1539 | { |
1540 | tree base; /* Base value. */ |
1541 | tree iters; /* Number of steps. */ |
1542 | tree step; /* Step size. */ |
1543 | tree tile; /* Tile increment (if tiled). */ |
1544 | tree outer; /* Tile iterator var. */ |
1545 | }; |
1546 | |
1547 | /* Helper for expand_oacc_for. Determine collapsed loop information. |
1548 | Fill in COUNTS array. Emit any initialization code before GSI. |
1549 | Return the calculated outer loop bound of BOUND_TYPE. */ |
1550 | |
1551 | static tree |
1552 | expand_oacc_collapse_init (const struct omp_for_data *fd, |
1553 | gimple_stmt_iterator *gsi, |
1554 | oacc_collapse *counts, tree diff_type, |
1555 | tree bound_type, location_t loc) |
1556 | { |
1557 | tree tiling = fd->tiling; |
1558 | tree total = build_int_cst (bound_type, 1); |
1559 | int ix; |
1560 | |
1561 | gcc_assert (integer_onep (fd->loop.step)); |
1562 | gcc_assert (integer_zerop (fd->loop.n1)); |
1563 | |
1564 | /* When tiling, the first operand of the tile clause applies to the |
1565 | innermost loop, and we work outwards from there. Seems |
1566 | backwards, but whatever. */ |
1567 | for (ix = fd->collapse; ix--;) |
1568 | { |
1569 | const omp_for_data_loop *loop = &fd->loops[ix]; |
1570 | |
1571 | tree iter_type = TREE_TYPE (loop->v); |
1572 | tree plus_type = iter_type; |
1573 | |
1574 | gcc_assert (loop->cond_code == LT_EXPR || loop->cond_code == GT_EXPR); |
1575 | |
1576 | if (POINTER_TYPE_P (iter_type)) |
1577 | plus_type = sizetype; |
1578 | |
1579 | if (tiling) |
1580 | { |
1581 | tree num = build_int_cst (integer_type_node, fd->collapse); |
1582 | tree loop_no = build_int_cst (integer_type_node, ix); |
1583 | tree tile = TREE_VALUE (tiling); |
1584 | gcall *call |
1585 | = gimple_build_call_internal (IFN_GOACC_TILE, 5, num, loop_no, tile, |
1586 | /* gwv-outer=*/integer_zero_node, |
1587 | /* gwv-inner=*/integer_zero_node); |
1588 | |
1589 | counts[ix].outer = create_tmp_var (iter_type, ".outer" ); |
1590 | counts[ix].tile = create_tmp_var (diff_type, ".tile" ); |
1591 | gimple_call_set_lhs (gs: call, lhs: counts[ix].tile); |
1592 | gimple_set_location (g: call, location: loc); |
1593 | gsi_insert_before (gsi, call, GSI_SAME_STMT); |
1594 | |
1595 | tiling = TREE_CHAIN (tiling); |
1596 | } |
1597 | else |
1598 | { |
1599 | counts[ix].tile = NULL; |
1600 | counts[ix].outer = loop->v; |
1601 | } |
1602 | |
1603 | tree b = loop->n1; |
1604 | tree e = loop->n2; |
1605 | tree s = loop->step; |
1606 | bool up = loop->cond_code == LT_EXPR; |
1607 | tree dir = build_int_cst (diff_type, up ? +1 : -1); |
1608 | bool negating; |
1609 | tree expr; |
1610 | |
1611 | b = force_gimple_operand_gsi (gsi, b, true, NULL_TREE, |
1612 | true, GSI_SAME_STMT); |
1613 | e = force_gimple_operand_gsi (gsi, e, true, NULL_TREE, |
1614 | true, GSI_SAME_STMT); |
1615 | |
1616 | /* Convert the step, avoiding possible unsigned->signed overflow. */ |
1617 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); |
1618 | if (negating) |
1619 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); |
1620 | s = fold_convert (diff_type, s); |
1621 | if (negating) |
1622 | s = fold_build1 (NEGATE_EXPR, diff_type, s); |
1623 | s = force_gimple_operand_gsi (gsi, s, true, NULL_TREE, |
1624 | true, GSI_SAME_STMT); |
1625 | |
1626 | /* Determine the range, avoiding possible unsigned->signed overflow. */ |
1627 | negating = !up && TYPE_UNSIGNED (iter_type); |
1628 | expr = fold_build2 (MINUS_EXPR, plus_type, |
1629 | fold_convert (plus_type, negating ? b : e), |
1630 | fold_convert (plus_type, negating ? e : b)); |
1631 | expr = fold_convert (diff_type, expr); |
1632 | if (negating) |
1633 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); |
1634 | tree range = force_gimple_operand_gsi |
1635 | (gsi, expr, true, NULL_TREE, true, GSI_SAME_STMT); |
1636 | |
1637 | /* Determine number of iterations. */ |
1638 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); |
1639 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); |
1640 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); |
1641 | |
1642 | tree iters = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE, |
1643 | true, GSI_SAME_STMT); |
1644 | |
1645 | counts[ix].base = b; |
1646 | counts[ix].iters = iters; |
1647 | counts[ix].step = s; |
1648 | |
1649 | total = fold_build2 (MULT_EXPR, bound_type, total, |
1650 | fold_convert (bound_type, iters)); |
1651 | } |
1652 | |
1653 | return total; |
1654 | } |
1655 | |
1656 | /* Emit initializers for collapsed loop members. INNER is true if |
1657 | this is for the element loop of a TILE. IVAR is the outer |
1658 | loop iteration variable, from which collapsed loop iteration values |
1659 | are calculated. COUNTS array has been initialized by |
1660 | expand_oacc_collapse_inits. */ |
1661 | |
1662 | static void |
1663 | expand_oacc_collapse_vars (const struct omp_for_data *fd, bool inner, |
1664 | gimple_stmt_iterator *gsi, |
1665 | const oacc_collapse *counts, tree ivar, |
1666 | tree diff_type) |
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]; |
1676 | tree v = inner ? loop->v : collapse->outer; |
1677 | tree iter_type = TREE_TYPE (v); |
1678 | tree plus_type = iter_type; |
1679 | enum tree_code plus_code = PLUS_EXPR; |
1680 | tree expr; |
1681 | |
1682 | if (POINTER_TYPE_P (iter_type)) |
1683 | { |
1684 | plus_code = POINTER_PLUS_EXPR; |
1685 | plus_type = sizetype; |
1686 | } |
1687 | |
1688 | expr = ivar; |
1689 | if (ix) |
1690 | { |
1691 | tree mod = fold_convert (ivar_type, collapse->iters); |
1692 | ivar = fold_build2 (TRUNC_DIV_EXPR, ivar_type, expr, mod); |
1693 | expr = fold_build2 (TRUNC_MOD_EXPR, ivar_type, expr, mod); |
1694 | ivar = force_gimple_operand_gsi (gsi, ivar, true, NULL_TREE, |
1695 | true, GSI_SAME_STMT); |
1696 | } |
1697 | |
1698 | expr = fold_build2 (MULT_EXPR, diff_type, fold_convert (diff_type, expr), |
1699 | fold_convert (diff_type, collapse->step)); |
1700 | expr = fold_build2 (plus_code, iter_type, |
1701 | inner ? collapse->outer : collapse->base, |
1702 | fold_convert (plus_type, expr)); |
1703 | expr = force_gimple_operand_gsi (gsi, expr, false, NULL_TREE, |
1704 | true, GSI_SAME_STMT); |
1705 | gassign *ass = gimple_build_assign (v, expr); |
1706 | gsi_insert_before (gsi, ass, GSI_SAME_STMT); |
1707 | } |
1708 | } |
1709 | |
1710 | /* Helper function for expand_omp_{for_*,simd}. If this is the outermost |
1711 | of the combined collapse > 1 loop constructs, generate code like: |
1712 | if (__builtin_expect (N32 cond3 N31, 0)) goto ZERO_ITER_BB; |
1713 | if (cond3 is <) |
1714 | adj = STEP3 - 1; |
1715 | else |
1716 | adj = STEP3 + 1; |
1717 | count3 = (adj + N32 - N31) / STEP3; |
1718 | if (__builtin_expect (N22 cond2 N21, 0)) goto ZERO_ITER_BB; |
1719 | if (cond2 is <) |
1720 | adj = STEP2 - 1; |
1721 | else |
1722 | adj = STEP2 + 1; |
1723 | count2 = (adj + N22 - N21) / STEP2; |
1724 | if (__builtin_expect (N12 cond1 N11, 0)) goto ZERO_ITER_BB; |
1725 | if (cond1 is <) |
1726 | adj = STEP1 - 1; |
1727 | else |
1728 | adj = STEP1 + 1; |
1729 | count1 = (adj + N12 - N11) / STEP1; |
1730 | count = count1 * count2 * count3; |
1731 | Furthermore, if ZERO_ITER_BB is NULL, create a BB which does: |
1732 | count = 0; |
1733 | and set ZERO_ITER_BB to that bb. If this isn't the outermost |
1734 | of the combined loop constructs, just initialize COUNTS array |
1735 | from the _looptemp_ clauses. For loop nests with non-rectangular |
1736 | loops, do this only for the rectangular loops. Then pick |
1737 | the loops which reference outer vars in their bound expressions |
1738 | and the loops which they refer to and for this sub-nest compute |
1739 | number of iterations. For triangular loops use Faulhaber's formula, |
1740 | otherwise as a fallback, compute by iterating the loops. |
1741 | If e.g. the sub-nest is |
1742 | for (I = N11; I COND1 N12; I += STEP1) |
1743 | for (J = M21 * I + N21; J COND2 M22 * I + N22; J += STEP2) |
1744 | for (K = M31 * J + N31; K COND3 M32 * J + N32; K += STEP3) |
1745 | do: |
1746 | COUNT = 0; |
1747 | for (tmpi = N11; tmpi COND1 N12; tmpi += STEP1) |
1748 | for (tmpj = M21 * tmpi + N21; |
1749 | tmpj COND2 M22 * tmpi + N22; tmpj += STEP2) |
1750 | { |
1751 | int tmpk1 = M31 * tmpj + N31; |
1752 | int tmpk2 = M32 * tmpj + N32; |
1753 | if (tmpk1 COND3 tmpk2) |
1754 | { |
1755 | if (COND3 is <) |
1756 | adj = STEP3 - 1; |
1757 | else |
1758 | adj = STEP3 + 1; |
1759 | COUNT += (adj + tmpk2 - tmpk1) / STEP3; |
1760 | } |
1761 | } |
1762 | and finally multiply the counts of the rectangular loops not |
1763 | in the sub-nest with COUNT. Also, as counts[fd->last_nonrect] |
1764 | store number of iterations of the loops from fd->first_nonrect |
1765 | to fd->last_nonrect inclusive, i.e. the above COUNT multiplied |
1766 | by the counts of rectangular loops not referenced in any non-rectangular |
1767 | loops sandwitched in between those. */ |
1768 | |
1769 | /* NOTE: It *could* be better to moosh all of the BBs together, |
1770 | creating one larger BB with all the computation and the unexpected |
1771 | jump at the end. I.e. |
1772 | |
1773 | bool zero3, zero2, zero1, zero; |
1774 | |
1775 | zero3 = N32 c3 N31; |
1776 | count3 = (N32 - N31) /[cl] STEP3; |
1777 | zero2 = N22 c2 N21; |
1778 | count2 = (N22 - N21) /[cl] STEP2; |
1779 | zero1 = N12 c1 N11; |
1780 | count1 = (N12 - N11) /[cl] STEP1; |
1781 | zero = zero3 || zero2 || zero1; |
1782 | count = count1 * count2 * count3; |
1783 | if (__builtin_expect(zero, false)) goto zero_iter_bb; |
1784 | |
1785 | After all, we expect the zero=false, and thus we expect to have to |
1786 | evaluate all of the comparison expressions, so short-circuiting |
1787 | oughtn't be a win. Since the condition isn't protecting a |
1788 | denominator, we're not concerned about divide-by-zero, so we can |
1789 | fully evaluate count even if a numerator turned out to be wrong. |
1790 | |
1791 | It seems like putting this all together would create much better |
1792 | scheduling opportunities, and less pressure on the chip's branch |
1793 | predictor. */ |
1794 | |
1795 | static void |
1796 | expand_omp_for_init_counts (struct omp_for_data *fd, gimple_stmt_iterator *gsi, |
1797 | basic_block &entry_bb, tree *counts, |
1798 | basic_block &zero_iter1_bb, int &first_zero_iter1, |
1799 | basic_block &zero_iter2_bb, int &first_zero_iter2, |
1800 | basic_block &l2_dom_bb) |
1801 | { |
1802 | tree t, type = TREE_TYPE (fd->loop.v); |
1803 | edge e, ne; |
1804 | int i; |
1805 | |
1806 | /* Collapsed loops need work for expansion into SSA form. */ |
1807 | gcc_assert (!gimple_in_ssa_p (cfun)); |
1808 | |
1809 | if (gimple_omp_for_combined_into_p (g: fd->for_stmt) |
1810 | && TREE_CODE (fd->loop.n2) != INTEGER_CST) |
1811 | { |
1812 | gcc_assert (fd->ordered == 0); |
1813 | /* First two _looptemp_ clauses are for istart/iend, counts[0] |
1814 | isn't supposed to be handled, as the inner loop doesn't |
1815 | use it. */ |
1816 | tree innerc = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
1817 | kind: OMP_CLAUSE__LOOPTEMP_); |
1818 | gcc_assert (innerc); |
1819 | for (i = 0; i < fd->collapse; i++) |
1820 | { |
1821 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
1822 | kind: OMP_CLAUSE__LOOPTEMP_); |
1823 | gcc_assert (innerc); |
1824 | if (i) |
1825 | counts[i] = OMP_CLAUSE_DECL (innerc); |
1826 | else |
1827 | counts[0] = NULL_TREE; |
1828 | } |
1829 | if (fd->non_rect |
1830 | && fd->last_nonrect == fd->first_nonrect + 1 |
1831 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) |
1832 | { |
1833 | tree c[4]; |
1834 | for (i = 0; i < 4; i++) |
1835 | { |
1836 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
1837 | kind: OMP_CLAUSE__LOOPTEMP_); |
1838 | gcc_assert (innerc); |
1839 | c[i] = OMP_CLAUSE_DECL (innerc); |
1840 | } |
1841 | counts[0] = c[0]; |
1842 | fd->first_inner_iterations = c[1]; |
1843 | fd->factor = c[2]; |
1844 | fd->adjn1 = c[3]; |
1845 | } |
1846 | return; |
1847 | } |
1848 | |
1849 | for (i = fd->collapse; i < fd->ordered; i++) |
1850 | { |
1851 | tree itype = TREE_TYPE (fd->loops[i].v); |
1852 | counts[i] = NULL_TREE; |
1853 | t = fold_binary (fd->loops[i].cond_code, boolean_type_node, |
1854 | fold_convert (itype, fd->loops[i].n1), |
1855 | fold_convert (itype, fd->loops[i].n2)); |
1856 | if (t && integer_zerop (t)) |
1857 | { |
1858 | for (i = fd->collapse; i < fd->ordered; i++) |
1859 | counts[i] = build_int_cst (type, 0); |
1860 | break; |
1861 | } |
1862 | } |
1863 | bool rect_count_seen = false; |
1864 | for (i = 0; i < (fd->ordered ? fd->ordered : fd->collapse); i++) |
1865 | { |
1866 | tree itype = TREE_TYPE (fd->loops[i].v); |
1867 | |
1868 | if (i >= fd->collapse && counts[i]) |
1869 | continue; |
1870 | if (fd->non_rect) |
1871 | { |
1872 | /* Skip loops that use outer iterators in their expressions |
1873 | during this phase. */ |
1874 | if (fd->loops[i].m1 || fd->loops[i].m2) |
1875 | { |
1876 | counts[i] = build_zero_cst (type); |
1877 | continue; |
1878 | } |
1879 | } |
1880 | if ((SSA_VAR_P (fd->loop.n2) || i >= fd->collapse) |
1881 | && ((t = fold_binary (fd->loops[i].cond_code, boolean_type_node, |
1882 | fold_convert (itype, fd->loops[i].n1), |
1883 | fold_convert (itype, fd->loops[i].n2))) |
1884 | == NULL_TREE || !integer_onep (t))) |
1885 | { |
1886 | gcond *cond_stmt; |
1887 | tree n1, n2; |
1888 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].n1)); |
1889 | n1 = force_gimple_operand_gsi (gsi, n1, true, NULL_TREE, |
1890 | true, GSI_SAME_STMT); |
1891 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
1892 | n2 = force_gimple_operand_gsi (gsi, n2, true, NULL_TREE, |
1893 | true, GSI_SAME_STMT); |
1894 | cond_stmt = expand_omp_build_cond (gsi_p: gsi, code: fd->loops[i].cond_code, |
1895 | lhs: n1, rhs: n2); |
1896 | e = split_block (entry_bb, cond_stmt); |
1897 | basic_block &zero_iter_bb |
1898 | = i < fd->collapse ? zero_iter1_bb : zero_iter2_bb; |
1899 | int &first_zero_iter |
1900 | = i < fd->collapse ? first_zero_iter1 : first_zero_iter2; |
1901 | if (zero_iter_bb == NULL) |
1902 | { |
1903 | gassign *assign_stmt; |
1904 | first_zero_iter = i; |
1905 | zero_iter_bb = create_empty_bb (entry_bb); |
1906 | add_bb_to_loop (zero_iter_bb, entry_bb->loop_father); |
1907 | *gsi = gsi_after_labels (bb: zero_iter_bb); |
1908 | if (i < fd->collapse) |
1909 | assign_stmt = gimple_build_assign (fd->loop.n2, |
1910 | build_zero_cst (type)); |
1911 | else |
1912 | { |
1913 | counts[i] = create_tmp_reg (type, ".count" ); |
1914 | assign_stmt |
1915 | = gimple_build_assign (counts[i], build_zero_cst (type)); |
1916 | } |
1917 | gsi_insert_before (gsi, assign_stmt, GSI_SAME_STMT); |
1918 | set_immediate_dominator (CDI_DOMINATORS, zero_iter_bb, |
1919 | entry_bb); |
1920 | } |
1921 | ne = make_edge (entry_bb, zero_iter_bb, EDGE_FALSE_VALUE); |
1922 | ne->probability = profile_probability::very_unlikely (); |
1923 | e->flags = EDGE_TRUE_VALUE; |
1924 | e->probability = ne->probability.invert (); |
1925 | if (l2_dom_bb == NULL) |
1926 | l2_dom_bb = entry_bb; |
1927 | entry_bb = e->dest; |
1928 | *gsi = gsi_last_nondebug_bb (bb: entry_bb); |
1929 | } |
1930 | |
1931 | if (POINTER_TYPE_P (itype)) |
1932 | itype = signed_type_for (itype); |
1933 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR |
1934 | ? -1 : 1)); |
1935 | t = fold_build2 (PLUS_EXPR, itype, |
1936 | fold_convert (itype, fd->loops[i].step), t); |
1937 | t = fold_build2 (PLUS_EXPR, itype, t, |
1938 | fold_convert (itype, fd->loops[i].n2)); |
1939 | t = fold_build2 (MINUS_EXPR, itype, t, |
1940 | fold_convert (itype, fd->loops[i].n1)); |
1941 | /* ?? We could probably use CEIL_DIV_EXPR instead of |
1942 | TRUNC_DIV_EXPR and adjusting by hand. Unless we can't |
1943 | generate the same code in the end because generically we |
1944 | don't know that the values involved must be negative for |
1945 | GT?? */ |
1946 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) |
1947 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
1948 | fold_build1 (NEGATE_EXPR, itype, t), |
1949 | fold_build1 (NEGATE_EXPR, itype, |
1950 | fold_convert (itype, |
1951 | fd->loops[i].step))); |
1952 | else |
1953 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, |
1954 | fold_convert (itype, fd->loops[i].step)); |
1955 | t = fold_convert (type, t); |
1956 | if (TREE_CODE (t) == INTEGER_CST) |
1957 | counts[i] = t; |
1958 | else |
1959 | { |
1960 | if (i < fd->collapse || i != first_zero_iter2) |
1961 | counts[i] = create_tmp_reg (type, ".count" ); |
1962 | expand_omp_build_assign (gsi_p: gsi, to: counts[i], from: t); |
1963 | } |
1964 | if (SSA_VAR_P (fd->loop.n2) && i < fd->collapse) |
1965 | { |
1966 | if (fd->non_rect && i >= fd->first_nonrect && i <= fd->last_nonrect) |
1967 | continue; |
1968 | if (!rect_count_seen) |
1969 | { |
1970 | t = counts[i]; |
1971 | rect_count_seen = true; |
1972 | } |
1973 | else |
1974 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, counts[i]); |
1975 | expand_omp_build_assign (gsi_p: gsi, to: fd->loop.n2, from: t); |
1976 | } |
1977 | } |
1978 | if (fd->non_rect && SSA_VAR_P (fd->loop.n2)) |
1979 | { |
1980 | gcc_assert (fd->last_nonrect != -1); |
1981 | |
1982 | counts[fd->last_nonrect] = create_tmp_reg (type, ".count" ); |
1983 | expand_omp_build_assign (gsi_p: gsi, to: counts[fd->last_nonrect], |
1984 | from: build_zero_cst (type)); |
1985 | for (i = fd->first_nonrect + 1; i < fd->last_nonrect; i++) |
1986 | if (fd->loops[i].m1 |
1987 | || fd->loops[i].m2 |
1988 | || fd->loops[i].non_rect_referenced) |
1989 | break; |
1990 | if (i == fd->last_nonrect |
1991 | && fd->loops[i].outer == fd->last_nonrect - fd->first_nonrect |
1992 | && !POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v)) |
1993 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[i].v))) |
1994 | { |
1995 | int o = fd->first_nonrect; |
1996 | tree itype = TREE_TYPE (fd->loops[o].v); |
1997 | tree n1o = create_tmp_reg (itype, ".n1o" ); |
1998 | t = fold_convert (itype, unshare_expr (fd->loops[o].n1)); |
1999 | expand_omp_build_assign (gsi_p: gsi, to: n1o, from: t); |
2000 | tree n2o = create_tmp_reg (itype, ".n2o" ); |
2001 | t = fold_convert (itype, unshare_expr (fd->loops[o].n2)); |
2002 | expand_omp_build_assign (gsi_p: gsi, to: n2o, from: t); |
2003 | if (fd->loops[i].m1 && fd->loops[i].m2) |
2004 | t = fold_build2 (MINUS_EXPR, itype, unshare_expr (fd->loops[i].m2), |
2005 | unshare_expr (fd->loops[i].m1)); |
2006 | else if (fd->loops[i].m1) |
2007 | t = fold_build1 (NEGATE_EXPR, itype, |
2008 | unshare_expr (fd->loops[i].m1)); |
2009 | else |
2010 | t = unshare_expr (fd->loops[i].m2); |
2011 | tree m2minusm1 |
2012 | = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, |
2013 | true, GSI_SAME_STMT); |
2014 | |
2015 | gimple_stmt_iterator gsi2 = *gsi; |
2016 | gsi_prev (i: &gsi2); |
2017 | e = split_block (entry_bb, gsi_stmt (i: gsi2)); |
2018 | e = split_block (e->dest, (gimple *) NULL); |
2019 | basic_block bb1 = e->src; |
2020 | entry_bb = e->dest; |
2021 | *gsi = gsi_after_labels (bb: entry_bb); |
2022 | |
2023 | gsi2 = gsi_after_labels (bb: bb1); |
2024 | tree ostep = fold_convert (itype, fd->loops[o].step); |
2025 | t = build_int_cst (itype, (fd->loops[o].cond_code |
2026 | == LT_EXPR ? -1 : 1)); |
2027 | t = fold_build2 (PLUS_EXPR, itype, ostep, t); |
2028 | t = fold_build2 (PLUS_EXPR, itype, t, n2o); |
2029 | t = fold_build2 (MINUS_EXPR, itype, t, n1o); |
2030 | if (TYPE_UNSIGNED (itype) |
2031 | && fd->loops[o].cond_code == GT_EXPR) |
2032 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
2033 | fold_build1 (NEGATE_EXPR, itype, t), |
2034 | fold_build1 (NEGATE_EXPR, itype, ostep)); |
2035 | else |
2036 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, ostep); |
2037 | tree outer_niters |
2038 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2039 | true, GSI_SAME_STMT); |
2040 | t = fold_build2 (MINUS_EXPR, itype, outer_niters, |
2041 | build_one_cst (itype)); |
2042 | t = fold_build2 (MULT_EXPR, itype, t, ostep); |
2043 | t = fold_build2 (PLUS_EXPR, itype, n1o, t); |
2044 | tree last = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2045 | true, GSI_SAME_STMT); |
2046 | tree n1, n2, n1e, n2e; |
2047 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); |
2048 | if (fd->loops[i].m1) |
2049 | { |
2050 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); |
2051 | n1 = fold_build2 (MULT_EXPR, itype, n1o, n1); |
2052 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); |
2053 | } |
2054 | else |
2055 | n1 = t; |
2056 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, |
2057 | true, GSI_SAME_STMT); |
2058 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
2059 | if (fd->loops[i].m2) |
2060 | { |
2061 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); |
2062 | n2 = fold_build2 (MULT_EXPR, itype, n1o, n2); |
2063 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); |
2064 | } |
2065 | else |
2066 | n2 = t; |
2067 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, |
2068 | true, GSI_SAME_STMT); |
2069 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); |
2070 | if (fd->loops[i].m1) |
2071 | { |
2072 | n1e = fold_convert (itype, unshare_expr (fd->loops[i].m1)); |
2073 | n1e = fold_build2 (MULT_EXPR, itype, last, n1e); |
2074 | n1e = fold_build2 (PLUS_EXPR, itype, n1e, t); |
2075 | } |
2076 | else |
2077 | n1e = t; |
2078 | n1e = force_gimple_operand_gsi (&gsi2, n1e, true, NULL_TREE, |
2079 | true, GSI_SAME_STMT); |
2080 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
2081 | if (fd->loops[i].m2) |
2082 | { |
2083 | n2e = fold_convert (itype, unshare_expr (fd->loops[i].m2)); |
2084 | n2e = fold_build2 (MULT_EXPR, itype, last, n2e); |
2085 | n2e = fold_build2 (PLUS_EXPR, itype, n2e, t); |
2086 | } |
2087 | else |
2088 | n2e = t; |
2089 | n2e = force_gimple_operand_gsi (&gsi2, n2e, true, NULL_TREE, |
2090 | true, GSI_SAME_STMT); |
2091 | gcond *cond_stmt |
2092 | = expand_omp_build_cond (gsi_p: &gsi2, code: fd->loops[i].cond_code, |
2093 | lhs: n1, rhs: n2); |
2094 | e = split_block (bb1, cond_stmt); |
2095 | e->flags = EDGE_TRUE_VALUE; |
2096 | e->probability = profile_probability::likely ().guessed (); |
2097 | basic_block bb2 = e->dest; |
2098 | gsi2 = gsi_after_labels (bb: bb2); |
2099 | |
2100 | cond_stmt = expand_omp_build_cond (gsi_p: &gsi2, code: fd->loops[i].cond_code, |
2101 | lhs: n1e, rhs: n2e); |
2102 | e = split_block (bb2, cond_stmt); |
2103 | e->flags = EDGE_TRUE_VALUE; |
2104 | e->probability = profile_probability::likely ().guessed (); |
2105 | gsi2 = gsi_after_labels (bb: e->dest); |
2106 | |
2107 | tree step = fold_convert (itype, fd->loops[i].step); |
2108 | t = build_int_cst (itype, (fd->loops[i].cond_code |
2109 | == LT_EXPR ? -1 : 1)); |
2110 | t = fold_build2 (PLUS_EXPR, itype, step, t); |
2111 | t = fold_build2 (PLUS_EXPR, itype, t, n2); |
2112 | t = fold_build2 (MINUS_EXPR, itype, t, n1); |
2113 | if (TYPE_UNSIGNED (itype) |
2114 | && fd->loops[i].cond_code == GT_EXPR) |
2115 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
2116 | fold_build1 (NEGATE_EXPR, itype, t), |
2117 | fold_build1 (NEGATE_EXPR, itype, step)); |
2118 | else |
2119 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
2120 | tree first_inner_iterations |
2121 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2122 | true, GSI_SAME_STMT); |
2123 | t = fold_build2 (MULT_EXPR, itype, m2minusm1, ostep); |
2124 | if (TYPE_UNSIGNED (itype) |
2125 | && fd->loops[i].cond_code == GT_EXPR) |
2126 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
2127 | fold_build1 (NEGATE_EXPR, itype, t), |
2128 | fold_build1 (NEGATE_EXPR, itype, step)); |
2129 | else |
2130 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
2131 | tree factor |
2132 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2133 | true, GSI_SAME_STMT); |
2134 | t = fold_build2 (MINUS_EXPR, itype, outer_niters, |
2135 | build_one_cst (itype)); |
2136 | t = fold_build2 (MULT_EXPR, itype, t, outer_niters); |
2137 | t = fold_build2 (RSHIFT_EXPR, itype, t, integer_one_node); |
2138 | t = fold_build2 (MULT_EXPR, itype, factor, t); |
2139 | t = fold_build2 (PLUS_EXPR, itype, |
2140 | fold_build2 (MULT_EXPR, itype, outer_niters, |
2141 | first_inner_iterations), t); |
2142 | expand_omp_build_assign (gsi_p: &gsi2, to: counts[fd->last_nonrect], |
2143 | fold_convert (type, t)); |
2144 | |
2145 | basic_block bb3 = create_empty_bb (bb1); |
2146 | add_bb_to_loop (bb3, bb1->loop_father); |
2147 | |
2148 | e = make_edge (bb1, bb3, EDGE_FALSE_VALUE); |
2149 | e->probability = profile_probability::unlikely ().guessed (); |
2150 | |
2151 | gsi2 = gsi_after_labels (bb: bb3); |
2152 | cond_stmt = expand_omp_build_cond (gsi_p: &gsi2, code: fd->loops[i].cond_code, |
2153 | lhs: n1e, rhs: n2e); |
2154 | e = split_block (bb3, cond_stmt); |
2155 | e->flags = EDGE_TRUE_VALUE; |
2156 | e->probability = profile_probability::likely ().guessed (); |
2157 | basic_block bb4 = e->dest; |
2158 | |
2159 | ne = make_edge (bb3, entry_bb, EDGE_FALSE_VALUE); |
2160 | ne->probability = e->probability.invert (); |
2161 | |
2162 | basic_block bb5 = create_empty_bb (bb2); |
2163 | add_bb_to_loop (bb5, bb2->loop_father); |
2164 | |
2165 | ne = make_edge (bb2, bb5, EDGE_FALSE_VALUE); |
2166 | ne->probability = profile_probability::unlikely ().guessed (); |
2167 | |
2168 | for (int j = 0; j < 2; j++) |
2169 | { |
2170 | gsi2 = gsi_after_labels (bb: j ? bb5 : bb4); |
2171 | t = fold_build2 (MINUS_EXPR, itype, |
2172 | unshare_expr (fd->loops[i].n1), |
2173 | unshare_expr (fd->loops[i].n2)); |
2174 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, m2minusm1); |
2175 | tree tem |
2176 | = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2177 | true, GSI_SAME_STMT); |
2178 | t = fold_build2 (MINUS_EXPR, itype, tem, n1o); |
2179 | t = fold_build2 (TRUNC_MOD_EXPR, itype, t, ostep); |
2180 | t = fold_build2 (MINUS_EXPR, itype, tem, t); |
2181 | tem = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2182 | true, GSI_SAME_STMT); |
2183 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); |
2184 | if (fd->loops[i].m1) |
2185 | { |
2186 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); |
2187 | n1 = fold_build2 (MULT_EXPR, itype, tem, n1); |
2188 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); |
2189 | } |
2190 | else |
2191 | n1 = t; |
2192 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, |
2193 | true, GSI_SAME_STMT); |
2194 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
2195 | if (fd->loops[i].m2) |
2196 | { |
2197 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); |
2198 | n2 = fold_build2 (MULT_EXPR, itype, tem, n2); |
2199 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); |
2200 | } |
2201 | else |
2202 | n2 = t; |
2203 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, |
2204 | true, GSI_SAME_STMT); |
2205 | expand_omp_build_assign (gsi_p: &gsi2, to: j ? n2o : n1o, from: tem); |
2206 | |
2207 | cond_stmt = expand_omp_build_cond (gsi_p: &gsi2, code: fd->loops[i].cond_code, |
2208 | lhs: n1, rhs: n2); |
2209 | e = split_block (gsi_bb (i: gsi2), cond_stmt); |
2210 | e->flags = j ? EDGE_TRUE_VALUE : EDGE_FALSE_VALUE; |
2211 | e->probability = profile_probability::unlikely ().guessed (); |
2212 | ne = make_edge (e->src, bb1, |
2213 | j ? EDGE_FALSE_VALUE : EDGE_TRUE_VALUE); |
2214 | ne->probability = e->probability.invert (); |
2215 | gsi2 = gsi_after_labels (bb: e->dest); |
2216 | |
2217 | t = fold_build2 (PLUS_EXPR, itype, tem, ostep); |
2218 | expand_omp_build_assign (gsi_p: &gsi2, to: j ? n2o : n1o, from: t); |
2219 | |
2220 | make_edge (e->dest, bb1, EDGE_FALLTHRU); |
2221 | } |
2222 | |
2223 | set_immediate_dominator (CDI_DOMINATORS, bb3, bb1); |
2224 | set_immediate_dominator (CDI_DOMINATORS, bb5, bb2); |
2225 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, bb1); |
2226 | |
2227 | if (fd->first_nonrect + 1 == fd->last_nonrect) |
2228 | { |
2229 | fd->first_inner_iterations = first_inner_iterations; |
2230 | fd->factor = factor; |
2231 | fd->adjn1 = n1o; |
2232 | } |
2233 | } |
2234 | else |
2235 | { |
2236 | /* Fallback implementation. Evaluate the loops with m1/m2 |
2237 | non-NULL as well as their outer loops at runtime using temporaries |
2238 | instead of the original iteration variables, and in the |
2239 | body just bump the counter. */ |
2240 | gimple_stmt_iterator gsi2 = *gsi; |
2241 | gsi_prev (i: &gsi2); |
2242 | e = split_block (entry_bb, gsi_stmt (i: gsi2)); |
2243 | e = split_block (e->dest, (gimple *) NULL); |
2244 | basic_block cur_bb = e->src; |
2245 | basic_block next_bb = e->dest; |
2246 | entry_bb = e->dest; |
2247 | *gsi = gsi_after_labels (bb: entry_bb); |
2248 | |
2249 | tree *vs = XALLOCAVEC (tree, fd->last_nonrect); |
2250 | memset (s: vs, c: 0, n: fd->last_nonrect * sizeof (tree)); |
2251 | |
2252 | for (i = 0; i <= fd->last_nonrect; i++) |
2253 | { |
2254 | if (fd->loops[i].m1 == NULL_TREE |
2255 | && fd->loops[i].m2 == NULL_TREE |
2256 | && !fd->loops[i].non_rect_referenced) |
2257 | continue; |
2258 | |
2259 | tree itype = TREE_TYPE (fd->loops[i].v); |
2260 | |
2261 | gsi2 = gsi_after_labels (bb: cur_bb); |
2262 | tree n1, n2; |
2263 | t = fold_convert (itype, unshare_expr (fd->loops[i].n1)); |
2264 | if (fd->loops[i].m1 == NULL_TREE) |
2265 | n1 = t; |
2266 | else if (POINTER_TYPE_P (itype)) |
2267 | { |
2268 | gcc_assert (integer_onep (fd->loops[i].m1)); |
2269 | t = unshare_expr (fd->loops[i].n1); |
2270 | n1 = fold_build_pointer_plus (vs[i - fd->loops[i].outer], t); |
2271 | } |
2272 | else |
2273 | { |
2274 | n1 = fold_convert (itype, unshare_expr (fd->loops[i].m1)); |
2275 | n1 = fold_build2 (MULT_EXPR, itype, |
2276 | vs[i - fd->loops[i].outer], n1); |
2277 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); |
2278 | } |
2279 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, |
2280 | true, GSI_SAME_STMT); |
2281 | if (i < fd->last_nonrect) |
2282 | { |
2283 | vs[i] = create_tmp_reg (itype, ".it" ); |
2284 | expand_omp_build_assign (gsi_p: &gsi2, to: vs[i], from: n1); |
2285 | } |
2286 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
2287 | if (fd->loops[i].m2 == NULL_TREE) |
2288 | n2 = t; |
2289 | else if (POINTER_TYPE_P (itype)) |
2290 | { |
2291 | gcc_assert (integer_onep (fd->loops[i].m2)); |
2292 | t = unshare_expr (fd->loops[i].n2); |
2293 | n2 = fold_build_pointer_plus (vs[i - fd->loops[i].outer], t); |
2294 | } |
2295 | else |
2296 | { |
2297 | n2 = fold_convert (itype, unshare_expr (fd->loops[i].m2)); |
2298 | n2 = fold_build2 (MULT_EXPR, itype, |
2299 | vs[i - fd->loops[i].outer], n2); |
2300 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); |
2301 | } |
2302 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, |
2303 | true, GSI_SAME_STMT); |
2304 | if (POINTER_TYPE_P (itype)) |
2305 | itype = signed_type_for (itype); |
2306 | if (i == fd->last_nonrect) |
2307 | { |
2308 | gcond *cond_stmt |
2309 | = expand_omp_build_cond (gsi_p: &gsi2, code: fd->loops[i].cond_code, |
2310 | lhs: n1, rhs: n2); |
2311 | e = split_block (cur_bb, cond_stmt); |
2312 | e->flags = EDGE_TRUE_VALUE; |
2313 | ne = make_edge (cur_bb, next_bb, EDGE_FALSE_VALUE); |
2314 | e->probability = profile_probability::likely ().guessed (); |
2315 | ne->probability = e->probability.invert (); |
2316 | gsi2 = gsi_after_labels (bb: e->dest); |
2317 | |
2318 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR |
2319 | ? -1 : 1)); |
2320 | t = fold_build2 (PLUS_EXPR, itype, |
2321 | fold_convert (itype, fd->loops[i].step), t); |
2322 | t = fold_build2 (PLUS_EXPR, itype, t, |
2323 | fold_convert (itype, n2)); |
2324 | t = fold_build2 (MINUS_EXPR, itype, t, |
2325 | fold_convert (itype, n1)); |
2326 | tree step = fold_convert (itype, fd->loops[i].step); |
2327 | if (TYPE_UNSIGNED (itype) |
2328 | && fd->loops[i].cond_code == GT_EXPR) |
2329 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
2330 | fold_build1 (NEGATE_EXPR, itype, t), |
2331 | fold_build1 (NEGATE_EXPR, itype, step)); |
2332 | else |
2333 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
2334 | t = fold_convert (type, t); |
2335 | t = fold_build2 (PLUS_EXPR, type, |
2336 | counts[fd->last_nonrect], t); |
2337 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2338 | true, GSI_SAME_STMT); |
2339 | expand_omp_build_assign (gsi_p: &gsi2, to: counts[fd->last_nonrect], from: t); |
2340 | e = make_edge (e->dest, next_bb, EDGE_FALLTHRU); |
2341 | set_immediate_dominator (CDI_DOMINATORS, next_bb, cur_bb); |
2342 | break; |
2343 | } |
2344 | e = split_block (cur_bb, last_nondebug_stmt (cur_bb)); |
2345 | |
2346 | basic_block new_cur_bb = create_empty_bb (cur_bb); |
2347 | add_bb_to_loop (new_cur_bb, cur_bb->loop_father); |
2348 | |
2349 | gsi2 = gsi_after_labels (bb: e->dest); |
2350 | tree step = fold_convert (itype, |
2351 | unshare_expr (fd->loops[i].step)); |
2352 | if (POINTER_TYPE_P (TREE_TYPE (vs[i]))) |
2353 | t = fold_build_pointer_plus (vs[i], step); |
2354 | else |
2355 | t = fold_build2 (PLUS_EXPR, itype, vs[i], step); |
2356 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2357 | true, GSI_SAME_STMT); |
2358 | expand_omp_build_assign (gsi_p: &gsi2, to: vs[i], from: t); |
2359 | |
2360 | ne = split_block (e->dest, last_nondebug_stmt (e->dest)); |
2361 | gsi2 = gsi_after_labels (bb: ne->dest); |
2362 | |
2363 | expand_omp_build_cond (gsi_p: &gsi2, code: fd->loops[i].cond_code, lhs: vs[i], rhs: n2); |
2364 | edge e3, e4; |
2365 | if (next_bb == entry_bb) |
2366 | { |
2367 | e3 = find_edge (ne->dest, next_bb); |
2368 | e3->flags = EDGE_FALSE_VALUE; |
2369 | } |
2370 | else |
2371 | e3 = make_edge (ne->dest, next_bb, EDGE_FALSE_VALUE); |
2372 | e4 = make_edge (ne->dest, new_cur_bb, EDGE_TRUE_VALUE); |
2373 | e4->probability = profile_probability::likely ().guessed (); |
2374 | e3->probability = e4->probability.invert (); |
2375 | basic_block esrc = e->src; |
2376 | make_edge (e->src, ne->dest, EDGE_FALLTHRU); |
2377 | cur_bb = new_cur_bb; |
2378 | basic_block latch_bb = next_bb; |
2379 | next_bb = e->dest; |
2380 | remove_edge (e); |
2381 | set_immediate_dominator (CDI_DOMINATORS, ne->dest, esrc); |
2382 | set_immediate_dominator (CDI_DOMINATORS, latch_bb, ne->dest); |
2383 | set_immediate_dominator (CDI_DOMINATORS, cur_bb, ne->dest); |
2384 | } |
2385 | } |
2386 | t = NULL_TREE; |
2387 | for (i = fd->first_nonrect; i < fd->last_nonrect; i++) |
2388 | if (!fd->loops[i].non_rect_referenced |
2389 | && fd->loops[i].m1 == NULL_TREE |
2390 | && fd->loops[i].m2 == NULL_TREE) |
2391 | { |
2392 | if (t == NULL_TREE) |
2393 | t = counts[i]; |
2394 | else |
2395 | t = fold_build2 (MULT_EXPR, type, t, counts[i]); |
2396 | } |
2397 | if (t) |
2398 | { |
2399 | t = fold_build2 (MULT_EXPR, type, counts[fd->last_nonrect], t); |
2400 | expand_omp_build_assign (gsi_p: gsi, to: counts[fd->last_nonrect], from: t); |
2401 | } |
2402 | if (!rect_count_seen) |
2403 | t = counts[fd->last_nonrect]; |
2404 | else |
2405 | t = fold_build2 (MULT_EXPR, type, fd->loop.n2, |
2406 | counts[fd->last_nonrect]); |
2407 | expand_omp_build_assign (gsi_p: gsi, to: fd->loop.n2, from: t); |
2408 | } |
2409 | else if (fd->non_rect) |
2410 | { |
2411 | tree t = fd->loop.n2; |
2412 | gcc_assert (TREE_CODE (t) == INTEGER_CST); |
2413 | int non_rect_referenced = 0, non_rect = 0; |
2414 | for (i = 0; i < fd->collapse; i++) |
2415 | { |
2416 | if ((i < fd->first_nonrect || i > fd->last_nonrect) |
2417 | && !integer_zerop (counts[i])) |
2418 | t = fold_build2 (TRUNC_DIV_EXPR, type, t, counts[i]); |
2419 | if (fd->loops[i].non_rect_referenced) |
2420 | non_rect_referenced++; |
2421 | if (fd->loops[i].m1 || fd->loops[i].m2) |
2422 | non_rect++; |
2423 | } |
2424 | gcc_assert (non_rect == 1 && non_rect_referenced == 1); |
2425 | counts[fd->last_nonrect] = t; |
2426 | } |
2427 | } |
2428 | |
2429 | /* Helper function for expand_omp_{for_*,simd}. Generate code like: |
2430 | T = V; |
2431 | V3 = N31 + (T % count3) * STEP3; |
2432 | T = T / count3; |
2433 | V2 = N21 + (T % count2) * STEP2; |
2434 | T = T / count2; |
2435 | V1 = N11 + T * STEP1; |
2436 | if this loop doesn't have an inner loop construct combined with it. |
2437 | If it does have an inner loop construct combined with it and the |
2438 | iteration count isn't known constant, store values from counts array |
2439 | into its _looptemp_ temporaries instead. |
2440 | For non-rectangular loops (between fd->first_nonrect and fd->last_nonrect |
2441 | inclusive), use the count of all those loops together, and either |
2442 | find quadratic etc. equation roots, or as a fallback, do: |
2443 | COUNT = 0; |
2444 | for (tmpi = N11; tmpi COND1 N12; tmpi += STEP1) |
2445 | for (tmpj = M21 * tmpi + N21; |
2446 | tmpj COND2 M22 * tmpi + N22; tmpj += STEP2) |
2447 | { |
2448 | int tmpk1 = M31 * tmpj + N31; |
2449 | int tmpk2 = M32 * tmpj + N32; |
2450 | if (tmpk1 COND3 tmpk2) |
2451 | { |
2452 | if (COND3 is <) |
2453 | adj = STEP3 - 1; |
2454 | else |
2455 | adj = STEP3 + 1; |
2456 | int temp = (adj + tmpk2 - tmpk1) / STEP3; |
2457 | if (COUNT + temp > T) |
2458 | { |
2459 | V1 = tmpi; |
2460 | V2 = tmpj; |
2461 | V3 = tmpk1 + (T - COUNT) * STEP3; |
2462 | goto done; |
2463 | } |
2464 | else |
2465 | COUNT += temp; |
2466 | } |
2467 | } |
2468 | done:; |
2469 | but for optional innermost or outermost rectangular loops that aren't |
2470 | referenced by other loop expressions keep doing the division/modulo. */ |
2471 | |
2472 | static void |
2473 | expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi, |
2474 | tree *counts, tree *nonrect_bounds, |
2475 | gimple *inner_stmt, tree startvar) |
2476 | { |
2477 | int i; |
2478 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
2479 | { |
2480 | /* If fd->loop.n2 is constant, then no propagation of the counts |
2481 | is needed, they are constant. */ |
2482 | if (TREE_CODE (fd->loop.n2) == INTEGER_CST) |
2483 | return; |
2484 | |
2485 | tree clauses = gimple_code (g: inner_stmt) != GIMPLE_OMP_FOR |
2486 | ? gimple_omp_taskreg_clauses (gs: inner_stmt) |
2487 | : gimple_omp_for_clauses (gs: inner_stmt); |
2488 | /* First two _looptemp_ clauses are for istart/iend, counts[0] |
2489 | isn't supposed to be handled, as the inner loop doesn't |
2490 | use it. */ |
2491 | tree innerc = omp_find_clause (clauses, kind: OMP_CLAUSE__LOOPTEMP_); |
2492 | gcc_assert (innerc); |
2493 | int count = 0; |
2494 | if (fd->non_rect |
2495 | && fd->last_nonrect == fd->first_nonrect + 1 |
2496 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) |
2497 | count = 4; |
2498 | for (i = 0; i < fd->collapse + count; i++) |
2499 | { |
2500 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
2501 | kind: OMP_CLAUSE__LOOPTEMP_); |
2502 | gcc_assert (innerc); |
2503 | if (i) |
2504 | { |
2505 | tree tem = OMP_CLAUSE_DECL (innerc); |
2506 | tree t; |
2507 | if (i < fd->collapse) |
2508 | t = counts[i]; |
2509 | else |
2510 | switch (i - fd->collapse) |
2511 | { |
2512 | case 0: t = counts[0]; break; |
2513 | case 1: t = fd->first_inner_iterations; break; |
2514 | case 2: t = fd->factor; break; |
2515 | case 3: t = fd->adjn1; break; |
2516 | default: gcc_unreachable (); |
2517 | } |
2518 | t = fold_convert (TREE_TYPE (tem), t); |
2519 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, |
2520 | false, GSI_CONTINUE_LINKING); |
2521 | gassign *stmt = gimple_build_assign (tem, t); |
2522 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); |
2523 | } |
2524 | } |
2525 | return; |
2526 | } |
2527 | |
2528 | tree type = TREE_TYPE (fd->loop.v); |
2529 | tree tem = create_tmp_reg (type, ".tem" ); |
2530 | gassign *stmt = gimple_build_assign (tem, startvar); |
2531 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); |
2532 | |
2533 | for (i = fd->collapse - 1; i >= 0; i--) |
2534 | { |
2535 | tree vtype = TREE_TYPE (fd->loops[i].v), itype, t; |
2536 | itype = vtype; |
2537 | if (POINTER_TYPE_P (vtype)) |
2538 | itype = signed_type_for (vtype); |
2539 | if (i != 0 && (i != fd->last_nonrect || fd->first_nonrect)) |
2540 | t = fold_build2 (TRUNC_MOD_EXPR, type, tem, counts[i]); |
2541 | else |
2542 | t = tem; |
2543 | if (i == fd->last_nonrect) |
2544 | { |
2545 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, |
2546 | false, GSI_CONTINUE_LINKING); |
2547 | tree stopval = t; |
2548 | tree idx = create_tmp_reg (type, ".count" ); |
2549 | expand_omp_build_assign (gsi_p: gsi, to: idx, |
2550 | from: build_zero_cst (type), after: true); |
2551 | basic_block bb_triang = NULL, bb_triang_dom = NULL; |
2552 | if (fd->first_nonrect + 1 == fd->last_nonrect |
2553 | && (TREE_CODE (fd->loop.n2) == INTEGER_CST |
2554 | || fd->first_inner_iterations) |
2555 | && (optab_handler (op: sqrt_optab, TYPE_MODE (double_type_node)) |
2556 | != CODE_FOR_nothing) |
2557 | && !integer_zerop (fd->loop.n2)) |
2558 | { |
2559 | tree outer_n1 = fd->adjn1 ? fd->adjn1 : fd->loops[i - 1].n1; |
2560 | tree itype = TREE_TYPE (fd->loops[i].v); |
2561 | tree first_inner_iterations = fd->first_inner_iterations; |
2562 | tree factor = fd->factor; |
2563 | gcond *cond_stmt |
2564 | = expand_omp_build_cond (gsi_p: gsi, code: NE_EXPR, lhs: factor, |
2565 | rhs: build_zero_cst (TREE_TYPE (factor)), |
2566 | after: true); |
2567 | edge e = split_block (gsi_bb (i: *gsi), cond_stmt); |
2568 | basic_block bb0 = e->src; |
2569 | e->flags = EDGE_TRUE_VALUE; |
2570 | e->probability = profile_probability::likely (); |
2571 | bb_triang_dom = bb0; |
2572 | *gsi = gsi_after_labels (bb: e->dest); |
2573 | tree slltype = long_long_integer_type_node; |
2574 | tree ulltype = long_long_unsigned_type_node; |
2575 | tree stopvalull = fold_convert (ulltype, stopval); |
2576 | stopvalull |
2577 | = force_gimple_operand_gsi (gsi, stopvalull, true, NULL_TREE, |
2578 | false, GSI_CONTINUE_LINKING); |
2579 | first_inner_iterations |
2580 | = fold_convert (slltype, first_inner_iterations); |
2581 | first_inner_iterations |
2582 | = force_gimple_operand_gsi (gsi, first_inner_iterations, true, |
2583 | NULL_TREE, false, |
2584 | GSI_CONTINUE_LINKING); |
2585 | factor = fold_convert (slltype, factor); |
2586 | factor |
2587 | = force_gimple_operand_gsi (gsi, factor, true, NULL_TREE, |
2588 | false, GSI_CONTINUE_LINKING); |
2589 | tree first_inner_iterationsd |
2590 | = fold_build1 (FLOAT_EXPR, double_type_node, |
2591 | first_inner_iterations); |
2592 | first_inner_iterationsd |
2593 | = force_gimple_operand_gsi (gsi, first_inner_iterationsd, true, |
2594 | NULL_TREE, false, |
2595 | GSI_CONTINUE_LINKING); |
2596 | tree factord = fold_build1 (FLOAT_EXPR, double_type_node, |
2597 | factor); |
2598 | factord = force_gimple_operand_gsi (gsi, factord, true, |
2599 | NULL_TREE, false, |
2600 | GSI_CONTINUE_LINKING); |
2601 | tree stopvald = fold_build1 (FLOAT_EXPR, double_type_node, |
2602 | stopvalull); |
2603 | stopvald = force_gimple_operand_gsi (gsi, stopvald, true, |
2604 | NULL_TREE, false, |
2605 | GSI_CONTINUE_LINKING); |
2606 | /* Temporarily disable flag_rounding_math, values will be |
2607 | decimal numbers divided by 2 and worst case imprecisions |
2608 | due to too large values ought to be caught later by the |
2609 | checks for fallback. */ |
2610 | int save_flag_rounding_math = flag_rounding_math; |
2611 | flag_rounding_math = 0; |
2612 | t = fold_build2 (RDIV_EXPR, double_type_node, factord, |
2613 | build_real (double_type_node, dconst2)); |
2614 | tree t3 = fold_build2 (MINUS_EXPR, double_type_node, |
2615 | first_inner_iterationsd, t); |
2616 | t3 = force_gimple_operand_gsi (gsi, t3, true, NULL_TREE, false, |
2617 | GSI_CONTINUE_LINKING); |
2618 | t = fold_build2 (MULT_EXPR, double_type_node, factord, |
2619 | build_real (double_type_node, dconst2)); |
2620 | t = fold_build2 (MULT_EXPR, double_type_node, t, stopvald); |
2621 | t = fold_build2 (PLUS_EXPR, double_type_node, t, |
2622 | fold_build2 (MULT_EXPR, double_type_node, |
2623 | t3, t3)); |
2624 | flag_rounding_math = save_flag_rounding_math; |
2625 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, |
2626 | GSI_CONTINUE_LINKING); |
2627 | if (flag_exceptions |
2628 | && cfun->can_throw_non_call_exceptions |
2629 | && operation_could_trap_p (LT_EXPR, true, false, NULL_TREE)) |
2630 | { |
2631 | tree tem = fold_build2 (LT_EXPR, boolean_type_node, t, |
2632 | build_zero_cst (double_type_node)); |
2633 | tem = force_gimple_operand_gsi (gsi, tem, true, NULL_TREE, |
2634 | false, GSI_CONTINUE_LINKING); |
2635 | cond_stmt = gimple_build_cond (NE_EXPR, tem, |
2636 | boolean_false_node, |
2637 | NULL_TREE, NULL_TREE); |
2638 | } |
2639 | else |
2640 | cond_stmt |
2641 | = gimple_build_cond (LT_EXPR, t, |
2642 | build_zero_cst (double_type_node), |
2643 | NULL_TREE, NULL_TREE); |
2644 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); |
2645 | e = split_block (gsi_bb (i: *gsi), cond_stmt); |
2646 | basic_block bb1 = e->src; |
2647 | e->flags = EDGE_FALSE_VALUE; |
2648 | e->probability = profile_probability::very_likely (); |
2649 | *gsi = gsi_after_labels (bb: e->dest); |
2650 | gcall *call = gimple_build_call_internal (IFN_SQRT, 1, t); |
2651 | tree sqrtr = create_tmp_var (double_type_node); |
2652 | gimple_call_set_lhs (gs: call, lhs: sqrtr); |
2653 | gsi_insert_after (gsi, call, GSI_CONTINUE_LINKING); |
2654 | t = fold_build2 (MINUS_EXPR, double_type_node, sqrtr, t3); |
2655 | t = fold_build2 (RDIV_EXPR, double_type_node, t, factord); |
2656 | t = fold_build1 (FIX_TRUNC_EXPR, ulltype, t); |
2657 | tree c = create_tmp_var (ulltype); |
2658 | tree d = create_tmp_var (ulltype); |
2659 | expand_omp_build_assign (gsi_p: gsi, to: c, from: t, after: true); |
2660 | t = fold_build2 (MINUS_EXPR, ulltype, c, |
2661 | build_one_cst (ulltype)); |
2662 | t = fold_build2 (MULT_EXPR, ulltype, c, t); |
2663 | t = fold_build2 (RSHIFT_EXPR, ulltype, t, integer_one_node); |
2664 | t = fold_build2 (MULT_EXPR, ulltype, |
2665 | fold_convert (ulltype, fd->factor), t); |
2666 | tree t2 |
2667 | = fold_build2 (MULT_EXPR, ulltype, c, |
2668 | fold_convert (ulltype, |
2669 | fd->first_inner_iterations)); |
2670 | t = fold_build2 (PLUS_EXPR, ulltype, t, t2); |
2671 | expand_omp_build_assign (gsi_p: gsi, to: d, from: t, after: true); |
2672 | t = fold_build2 (MULT_EXPR, ulltype, |
2673 | fold_convert (ulltype, fd->factor), c); |
2674 | t = fold_build2 (PLUS_EXPR, ulltype, |
2675 | t, fold_convert (ulltype, |
2676 | fd->first_inner_iterations)); |
2677 | t2 = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, |
2678 | GSI_CONTINUE_LINKING); |
2679 | cond_stmt = gimple_build_cond (GE_EXPR, stopvalull, d, |
2680 | NULL_TREE, NULL_TREE); |
2681 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); |
2682 | e = split_block (gsi_bb (i: *gsi), cond_stmt); |
2683 | basic_block bb2 = e->src; |
2684 | e->flags = EDGE_TRUE_VALUE; |
2685 | e->probability = profile_probability::very_likely (); |
2686 | *gsi = gsi_after_labels (bb: e->dest); |
2687 | t = fold_build2 (PLUS_EXPR, ulltype, d, t2); |
2688 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, |
2689 | GSI_CONTINUE_LINKING); |
2690 | cond_stmt = gimple_build_cond (GE_EXPR, stopvalull, t, |
2691 | NULL_TREE, NULL_TREE); |
2692 | gsi_insert_after (gsi, cond_stmt, GSI_CONTINUE_LINKING); |
2693 | e = split_block (gsi_bb (i: *gsi), cond_stmt); |
2694 | basic_block bb3 = e->src; |
2695 | e->flags = EDGE_FALSE_VALUE; |
2696 | e->probability = profile_probability::very_likely (); |
2697 | *gsi = gsi_after_labels (bb: e->dest); |
2698 | t = fold_convert (itype, c); |
2699 | t = fold_build2 (MULT_EXPR, itype, t, fd->loops[i - 1].step); |
2700 | t = fold_build2 (PLUS_EXPR, itype, outer_n1, t); |
2701 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false, |
2702 | GSI_CONTINUE_LINKING); |
2703 | expand_omp_build_assign (gsi_p: gsi, to: fd->loops[i - 1].v, from: t, after: true); |
2704 | t2 = fold_build2 (MINUS_EXPR, ulltype, stopvalull, d); |
2705 | t2 = fold_convert (itype, t2); |
2706 | t2 = fold_build2 (MULT_EXPR, itype, t2, fd->loops[i].step); |
2707 | t2 = fold_build2 (PLUS_EXPR, itype, t2, fd->loops[i].n1); |
2708 | if (fd->loops[i].m1) |
2709 | { |
2710 | t = fold_build2 (MULT_EXPR, itype, t, fd->loops[i].m1); |
2711 | t2 = fold_build2 (PLUS_EXPR, itype, t2, t); |
2712 | } |
2713 | expand_omp_build_assign (gsi_p: gsi, to: fd->loops[i].v, from: t2, after: true); |
2714 | e = split_block (gsi_bb (i: *gsi), gsi_stmt (i: *gsi)); |
2715 | bb_triang = e->src; |
2716 | *gsi = gsi_after_labels (bb: e->dest); |
2717 | remove_edge (e); |
2718 | e = make_edge (bb1, gsi_bb (i: *gsi), EDGE_TRUE_VALUE); |
2719 | e->probability = profile_probability::very_unlikely (); |
2720 | e = make_edge (bb2, gsi_bb (i: *gsi), EDGE_FALSE_VALUE); |
2721 | e->probability = profile_probability::very_unlikely (); |
2722 | e = make_edge (bb3, gsi_bb (i: *gsi), EDGE_TRUE_VALUE); |
2723 | e->probability = profile_probability::very_unlikely (); |
2724 | |
2725 | basic_block bb4 = create_empty_bb (bb0); |
2726 | add_bb_to_loop (bb4, bb0->loop_father); |
2727 | e = make_edge (bb0, bb4, EDGE_FALSE_VALUE); |
2728 | e->probability = profile_probability::unlikely (); |
2729 | make_edge (bb4, gsi_bb (i: *gsi), EDGE_FALLTHRU); |
2730 | set_immediate_dominator (CDI_DOMINATORS, bb4, bb0); |
2731 | set_immediate_dominator (CDI_DOMINATORS, gsi_bb (i: *gsi), bb0); |
2732 | gimple_stmt_iterator gsi2 = gsi_after_labels (bb: bb4); |
2733 | t2 = fold_build2 (TRUNC_DIV_EXPR, type, |
2734 | counts[i], counts[i - 1]); |
2735 | t2 = force_gimple_operand_gsi (&gsi2, t2, true, NULL_TREE, false, |
2736 | GSI_CONTINUE_LINKING); |
2737 | t = fold_build2 (TRUNC_MOD_EXPR, type, stopval, t2); |
2738 | t2 = fold_build2 (TRUNC_DIV_EXPR, type, stopval, t2); |
2739 | t = fold_convert (itype, t); |
2740 | t2 = fold_convert (itype, t2); |
2741 | t = fold_build2 (MULT_EXPR, itype, t, |
2742 | fold_convert (itype, fd->loops[i].step)); |
2743 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); |
2744 | t2 = fold_build2 (MULT_EXPR, itype, t2, |
2745 | fold_convert (itype, fd->loops[i - 1].step)); |
2746 | t2 = fold_build2 (PLUS_EXPR, itype, fd->loops[i - 1].n1, t2); |
2747 | t2 = force_gimple_operand_gsi (&gsi2, t2, false, NULL_TREE, |
2748 | false, GSI_CONTINUE_LINKING); |
2749 | stmt = gimple_build_assign (fd->loops[i - 1].v, t2); |
2750 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); |
2751 | if (fd->loops[i].m1) |
2752 | { |
2753 | t2 = fold_build2 (MULT_EXPR, itype, fd->loops[i].m1, |
2754 | fd->loops[i - 1].v); |
2755 | t = fold_build2 (PLUS_EXPR, itype, t, t2); |
2756 | } |
2757 | t = force_gimple_operand_gsi (&gsi2, t, false, NULL_TREE, |
2758 | false, GSI_CONTINUE_LINKING); |
2759 | stmt = gimple_build_assign (fd->loops[i].v, t); |
2760 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); |
2761 | } |
2762 | /* Fallback implementation. Evaluate the loops in between |
2763 | (inclusive) fd->first_nonrect and fd->last_nonrect at |
2764 | runtime unsing temporaries instead of the original iteration |
2765 | variables, in the body just bump the counter and compare |
2766 | with the desired value. */ |
2767 | gimple_stmt_iterator gsi2 = *gsi; |
2768 | basic_block entry_bb = gsi_bb (i: gsi2); |
2769 | edge e = split_block (entry_bb, gsi_stmt (i: gsi2)); |
2770 | e = split_block (e->dest, (gimple *) NULL); |
2771 | basic_block dom_bb = NULL; |
2772 | basic_block cur_bb = e->src; |
2773 | basic_block next_bb = e->dest; |
2774 | entry_bb = e->dest; |
2775 | *gsi = gsi_after_labels (bb: entry_bb); |
2776 | |
2777 | tree *vs = XALLOCAVEC (tree, fd->last_nonrect); |
2778 | tree n1 = NULL_TREE, n2 = NULL_TREE; |
2779 | memset (s: vs, c: 0, n: fd->last_nonrect * sizeof (tree)); |
2780 | |
2781 | for (int j = fd->first_nonrect; j <= fd->last_nonrect; j++) |
2782 | { |
2783 | tree itype = TREE_TYPE (fd->loops[j].v); |
2784 | bool rect_p = (fd->loops[j].m1 == NULL_TREE |
2785 | && fd->loops[j].m2 == NULL_TREE |
2786 | && !fd->loops[j].non_rect_referenced); |
2787 | gsi2 = gsi_after_labels (bb: cur_bb); |
2788 | t = fold_convert (itype, unshare_expr (fd->loops[j].n1)); |
2789 | if (fd->loops[j].m1 == NULL_TREE) |
2790 | n1 = rect_p ? build_zero_cst (type) : t; |
2791 | else if (POINTER_TYPE_P (itype)) |
2792 | { |
2793 | gcc_assert (integer_onep (fd->loops[j].m1)); |
2794 | t = unshare_expr (fd->loops[j].n1); |
2795 | n1 = fold_build_pointer_plus (vs[j - fd->loops[j].outer], t); |
2796 | } |
2797 | else |
2798 | { |
2799 | n1 = fold_convert (itype, unshare_expr (fd->loops[j].m1)); |
2800 | n1 = fold_build2 (MULT_EXPR, itype, |
2801 | vs[j - fd->loops[j].outer], n1); |
2802 | n1 = fold_build2 (PLUS_EXPR, itype, n1, t); |
2803 | } |
2804 | n1 = force_gimple_operand_gsi (&gsi2, n1, true, NULL_TREE, |
2805 | true, GSI_SAME_STMT); |
2806 | if (j < fd->last_nonrect) |
2807 | { |
2808 | vs[j] = create_tmp_reg (rect_p ? type : itype, ".it" ); |
2809 | expand_omp_build_assign (gsi_p: &gsi2, to: vs[j], from: n1); |
2810 | } |
2811 | t = fold_convert (itype, unshare_expr (fd->loops[j].n2)); |
2812 | if (fd->loops[j].m2 == NULL_TREE) |
2813 | n2 = rect_p ? counts[j] : t; |
2814 | else if (POINTER_TYPE_P (itype)) |
2815 | { |
2816 | gcc_assert (integer_onep (fd->loops[j].m2)); |
2817 | t = unshare_expr (fd->loops[j].n2); |
2818 | n2 = fold_build_pointer_plus (vs[j - fd->loops[j].outer], t); |
2819 | } |
2820 | else |
2821 | { |
2822 | n2 = fold_convert (itype, unshare_expr (fd->loops[j].m2)); |
2823 | n2 = fold_build2 (MULT_EXPR, itype, |
2824 | vs[j - fd->loops[j].outer], n2); |
2825 | n2 = fold_build2 (PLUS_EXPR, itype, n2, t); |
2826 | } |
2827 | n2 = force_gimple_operand_gsi (&gsi2, n2, true, NULL_TREE, |
2828 | true, GSI_SAME_STMT); |
2829 | if (POINTER_TYPE_P (itype)) |
2830 | itype = signed_type_for (itype); |
2831 | if (j == fd->last_nonrect) |
2832 | { |
2833 | gcond *cond_stmt |
2834 | = expand_omp_build_cond (gsi_p: &gsi2, code: fd->loops[i].cond_code, |
2835 | lhs: n1, rhs: n2); |
2836 | e = split_block (cur_bb, cond_stmt); |
2837 | e->flags = EDGE_TRUE_VALUE; |
2838 | edge ne = make_edge (cur_bb, next_bb, EDGE_FALSE_VALUE); |
2839 | e->probability = profile_probability::likely ().guessed (); |
2840 | ne->probability = e->probability.invert (); |
2841 | gsi2 = gsi_after_labels (bb: e->dest); |
2842 | |
2843 | t = build_int_cst (itype, (fd->loops[j].cond_code == LT_EXPR |
2844 | ? -1 : 1)); |
2845 | t = fold_build2 (PLUS_EXPR, itype, |
2846 | fold_convert (itype, fd->loops[j].step), t); |
2847 | t = fold_build2 (PLUS_EXPR, itype, t, |
2848 | fold_convert (itype, n2)); |
2849 | t = fold_build2 (MINUS_EXPR, itype, t, |
2850 | fold_convert (itype, n1)); |
2851 | tree step = fold_convert (itype, fd->loops[j].step); |
2852 | if (TYPE_UNSIGNED (itype) |
2853 | && fd->loops[j].cond_code == GT_EXPR) |
2854 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
2855 | fold_build1 (NEGATE_EXPR, itype, t), |
2856 | fold_build1 (NEGATE_EXPR, itype, step)); |
2857 | else |
2858 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
2859 | t = fold_convert (type, t); |
2860 | t = fold_build2 (PLUS_EXPR, type, idx, t); |
2861 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2862 | true, GSI_SAME_STMT); |
2863 | e = make_edge (e->dest, next_bb, EDGE_FALLTHRU); |
2864 | set_immediate_dominator (CDI_DOMINATORS, next_bb, cur_bb); |
2865 | cond_stmt |
2866 | = gimple_build_cond (LE_EXPR, t, stopval, NULL_TREE, |
2867 | NULL_TREE); |
2868 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); |
2869 | e = split_block (gsi_bb (i: gsi2), cond_stmt); |
2870 | e->flags = EDGE_TRUE_VALUE; |
2871 | e->probability = profile_probability::likely ().guessed (); |
2872 | ne = make_edge (e->src, entry_bb, EDGE_FALSE_VALUE); |
2873 | ne->probability = e->probability.invert (); |
2874 | gsi2 = gsi_after_labels (bb: e->dest); |
2875 | expand_omp_build_assign (gsi_p: &gsi2, to: idx, from: t); |
2876 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, dom_bb); |
2877 | break; |
2878 | } |
2879 | e = split_block (cur_bb, last_nondebug_stmt (cur_bb)); |
2880 | |
2881 | basic_block new_cur_bb = create_empty_bb (cur_bb); |
2882 | add_bb_to_loop (new_cur_bb, cur_bb->loop_father); |
2883 | |
2884 | gsi2 = gsi_after_labels (bb: e->dest); |
2885 | if (rect_p) |
2886 | t = fold_build2 (PLUS_EXPR, type, vs[j], |
2887 | build_one_cst (type)); |
2888 | else |
2889 | { |
2890 | tree step |
2891 | = fold_convert (itype, unshare_expr (fd->loops[j].step)); |
2892 | if (POINTER_TYPE_P (vtype)) |
2893 | t = fold_build_pointer_plus (vs[j], step); |
2894 | else |
2895 | t = fold_build2 (PLUS_EXPR, itype, vs[j], step); |
2896 | } |
2897 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
2898 | true, GSI_SAME_STMT); |
2899 | expand_omp_build_assign (gsi_p: &gsi2, to: vs[j], from: t); |
2900 | |
2901 | edge ne = split_block (e->dest, last_nondebug_stmt (e->dest)); |
2902 | gsi2 = gsi_after_labels (bb: ne->dest); |
2903 | |
2904 | gcond *cond_stmt; |
2905 | if (next_bb == entry_bb) |
2906 | /* No need to actually check the outermost condition. */ |
2907 | cond_stmt |
2908 | = gimple_build_cond (EQ_EXPR, boolean_true_node, |
2909 | boolean_true_node, |
2910 | NULL_TREE, NULL_TREE); |
2911 | else |
2912 | cond_stmt |
2913 | = gimple_build_cond (rect_p ? LT_EXPR |
2914 | : fd->loops[j].cond_code, |
2915 | vs[j], n2, NULL_TREE, NULL_TREE); |
2916 | gsi_insert_before (&gsi2, cond_stmt, GSI_SAME_STMT); |
2917 | edge e3, e4; |
2918 | if (next_bb == entry_bb) |
2919 | { |
2920 | e3 = find_edge (ne->dest, next_bb); |
2921 | e3->flags = EDGE_FALSE_VALUE; |
2922 | dom_bb = ne->dest; |
2923 | } |
2924 | else |
2925 | e3 = make_edge (ne->dest, next_bb, EDGE_FALSE_VALUE); |
2926 | e4 = make_edge (ne->dest, new_cur_bb, EDGE_TRUE_VALUE); |
2927 | e4->probability = profile_probability::likely ().guessed (); |
2928 | e3->probability = e4->probability.invert (); |
2929 | basic_block esrc = e->src; |
2930 | make_edge (e->src, ne->dest, EDGE_FALLTHRU); |
2931 | cur_bb = new_cur_bb; |
2932 | basic_block latch_bb = next_bb; |
2933 | next_bb = e->dest; |
2934 | remove_edge (e); |
2935 | set_immediate_dominator (CDI_DOMINATORS, ne->dest, esrc); |
2936 | set_immediate_dominator (CDI_DOMINATORS, latch_bb, ne->dest); |
2937 | set_immediate_dominator (CDI_DOMINATORS, cur_bb, ne->dest); |
2938 | } |
2939 | for (int j = fd->last_nonrect; j >= fd->first_nonrect; j--) |
2940 | { |
2941 | tree vtype = TREE_TYPE (fd->loops[j].v); |
2942 | tree itype = vtype; |
2943 | if (POINTER_TYPE_P (itype)) |
2944 | itype = signed_type_for (itype); |
2945 | bool rect_p = (fd->loops[j].m1 == NULL_TREE |
2946 | && fd->loops[j].m2 == NULL_TREE |
2947 | && !fd->loops[j].non_rect_referenced); |
2948 | if (j == fd->last_nonrect) |
2949 | { |
2950 | t = fold_build2 (MINUS_EXPR, type, stopval, idx); |
2951 | t = fold_convert (itype, t); |
2952 | tree t2 |
2953 | = fold_convert (itype, unshare_expr (fd->loops[j].step)); |
2954 | t = fold_build2 (MULT_EXPR, itype, t, t2); |
2955 | if (POINTER_TYPE_P (vtype)) |
2956 | t = fold_build_pointer_plus (n1, t); |
2957 | else |
2958 | t = fold_build2 (PLUS_EXPR, itype, n1, t); |
2959 | } |
2960 | else if (rect_p) |
2961 | { |
2962 | t = fold_convert (itype, vs[j]); |
2963 | t = fold_build2 (MULT_EXPR, itype, t, |
2964 | fold_convert (itype, fd->loops[j].step)); |
2965 | if (POINTER_TYPE_P (vtype)) |
2966 | t = fold_build_pointer_plus (fd->loops[j].n1, t); |
2967 | else |
2968 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[j].n1, t); |
2969 | } |
2970 | else |
2971 | t = vs[j]; |
2972 | t = force_gimple_operand_gsi (gsi, t, false, |
2973 | NULL_TREE, true, |
2974 | GSI_SAME_STMT); |
2975 | stmt = gimple_build_assign (fd->loops[j].v, t); |
2976 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); |
2977 | } |
2978 | if (gsi_end_p (i: *gsi)) |
2979 | *gsi = gsi_last_bb (bb: gsi_bb (i: *gsi)); |
2980 | else |
2981 | gsi_prev (i: gsi); |
2982 | if (bb_triang) |
2983 | { |
2984 | e = split_block (gsi_bb (i: *gsi), gsi_stmt (i: *gsi)); |
2985 | make_edge (bb_triang, e->dest, EDGE_FALLTHRU); |
2986 | *gsi = gsi_after_labels (bb: e->dest); |
2987 | if (!gsi_end_p (i: *gsi)) |
2988 | gsi_insert_before (gsi, gimple_build_nop (), GSI_NEW_STMT); |
2989 | set_immediate_dominator (CDI_DOMINATORS, e->dest, bb_triang_dom); |
2990 | } |
2991 | } |
2992 | else |
2993 | { |
2994 | t = fold_convert (itype, t); |
2995 | t = fold_build2 (MULT_EXPR, itype, t, |
2996 | fold_convert (itype, fd->loops[i].step)); |
2997 | if (POINTER_TYPE_P (vtype)) |
2998 | t = fold_build_pointer_plus (fd->loops[i].n1, t); |
2999 | else |
3000 | t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t); |
3001 | t = force_gimple_operand_gsi (gsi, t, |
3002 | DECL_P (fd->loops[i].v) |
3003 | && TREE_ADDRESSABLE (fd->loops[i].v), |
3004 | NULL_TREE, false, |
3005 | GSI_CONTINUE_LINKING); |
3006 | stmt = gimple_build_assign (fd->loops[i].v, t); |
3007 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); |
3008 | } |
3009 | if (i != 0 && (i != fd->last_nonrect || fd->first_nonrect)) |
3010 | { |
3011 | t = fold_build2 (TRUNC_DIV_EXPR, type, tem, counts[i]); |
3012 | t = force_gimple_operand_gsi (gsi, t, false, NULL_TREE, |
3013 | false, GSI_CONTINUE_LINKING); |
3014 | stmt = gimple_build_assign (tem, t); |
3015 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); |
3016 | } |
3017 | if (i == fd->last_nonrect) |
3018 | i = fd->first_nonrect; |
3019 | } |
3020 | if (fd->non_rect) |
3021 | for (i = 0; i <= fd->last_nonrect; i++) |
3022 | if (fd->loops[i].m2) |
3023 | { |
3024 | tree itype = TREE_TYPE (fd->loops[i].v); |
3025 | |
3026 | tree t; |
3027 | if (POINTER_TYPE_P (itype)) |
3028 | { |
3029 | gcc_assert (integer_onep (fd->loops[i].m2)); |
3030 | t = fold_build_pointer_plus (fd->loops[i - fd->loops[i].outer].v, |
3031 | unshare_expr (fd->loops[i].n2)); |
3032 | } |
3033 | else |
3034 | { |
3035 | t = fold_convert (itype, unshare_expr (fd->loops[i].m2)); |
3036 | t = fold_build2 (MULT_EXPR, itype, |
3037 | fd->loops[i - fd->loops[i].outer].v, t); |
3038 | t = fold_build2 (PLUS_EXPR, itype, t, |
3039 | fold_convert (itype, |
3040 | unshare_expr (fd->loops[i].n2))); |
3041 | } |
3042 | nonrect_bounds[i] = create_tmp_reg (itype, ".bound" ); |
3043 | t = force_gimple_operand_gsi (gsi, t, false, |
3044 | NULL_TREE, false, |
3045 | GSI_CONTINUE_LINKING); |
3046 | stmt = gimple_build_assign (nonrect_bounds[i], t); |
3047 | gsi_insert_after (gsi, stmt, GSI_CONTINUE_LINKING); |
3048 | } |
3049 | } |
3050 | |
3051 | /* Helper function for expand_omp_for_*. Generate code like: |
3052 | L10: |
3053 | V3 += STEP3; |
3054 | if (V3 cond3 N32) goto BODY_BB; else goto L11; |
3055 | L11: |
3056 | V3 = N31; |
3057 | V2 += STEP2; |
3058 | if (V2 cond2 N22) goto BODY_BB; else goto L12; |
3059 | L12: |
3060 | V2 = N21; |
3061 | V1 += STEP1; |
3062 | goto BODY_BB; |
3063 | For non-rectangular loops, use temporaries stored in nonrect_bounds |
3064 | for the upper bounds if M?2 multiplier is present. Given e.g. |
3065 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) |
3066 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) |
3067 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) |
3068 | for (V4 = N41 + M41 * V2; V4 cond4 N42 + M42 * V2; V4 += STEP4) |
3069 | do: |
3070 | L10: |
3071 | V4 += STEP4; |
3072 | if (V4 cond4 NONRECT_BOUND4) goto BODY_BB; else goto L11; |
3073 | L11: |
3074 | V4 = N41 + M41 * V2; // This can be left out if the loop |
3075 | // refers to the immediate parent loop |
3076 | V3 += STEP3; |
3077 | if (V3 cond3 N32) goto BODY_BB; else goto L12; |
3078 | L12: |
3079 | V3 = N31; |
3080 | V2 += STEP2; |
3081 | if (V2 cond2 N22) goto L120; else goto L13; |
3082 | L120: |
3083 | V4 = N41 + M41 * V2; |
3084 | NONRECT_BOUND4 = N42 + M42 * V2; |
3085 | if (V4 cond4 NONRECT_BOUND4) goto BODY_BB; else goto L12; |
3086 | L13: |
3087 | V2 = N21; |
3088 | V1 += STEP1; |
3089 | goto L120; */ |
3090 | |
3091 | static basic_block |
3092 | (struct omp_for_data *fd, tree *nonrect_bounds, |
3093 | basic_block cont_bb, basic_block body_bb) |
3094 | { |
3095 | basic_block last_bb, bb, collapse_bb = NULL; |
3096 | int i; |
3097 | gimple_stmt_iterator gsi; |
3098 | edge e; |
3099 | tree t; |
3100 | gimple *stmt; |
3101 | |
3102 | last_bb = cont_bb; |
3103 | for (i = fd->collapse - 1; i >= 0; i--) |
3104 | { |
3105 | tree vtype = TREE_TYPE (fd->loops[i].v); |
3106 | |
3107 | bb = create_empty_bb (last_bb); |
3108 | add_bb_to_loop (bb, last_bb->loop_father); |
3109 | gsi = gsi_start_bb (bb); |
3110 | |
3111 | if (i < fd->collapse - 1) |
3112 | { |
3113 | e = make_edge (last_bb, bb, EDGE_FALSE_VALUE); |
3114 | e->probability = profile_probability::guessed_always () / 8; |
3115 | |
3116 | struct omp_for_data_loop *l = &fd->loops[i + 1]; |
3117 | if (l->m1 == NULL_TREE || l->outer != 1) |
3118 | { |
3119 | t = l->n1; |
3120 | if (l->m1) |
3121 | { |
3122 | if (POINTER_TYPE_P (TREE_TYPE (l->v))) |
3123 | t = fold_build_pointer_plus (fd->loops[i + 1 - l->outer].v, |
3124 | t); |
3125 | else |
3126 | { |
3127 | tree t2 |
3128 | = fold_build2 (MULT_EXPR, TREE_TYPE (t), |
3129 | fd->loops[i + 1 - l->outer].v, l->m1); |
3130 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t2, t); |
3131 | } |
3132 | } |
3133 | t = force_gimple_operand_gsi (&gsi, t, |
3134 | DECL_P (l->v) |
3135 | && TREE_ADDRESSABLE (l->v), |
3136 | NULL_TREE, false, |
3137 | GSI_CONTINUE_LINKING); |
3138 | stmt = gimple_build_assign (l->v, t); |
3139 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
3140 | } |
3141 | } |
3142 | else |
3143 | collapse_bb = bb; |
3144 | |
3145 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); |
3146 | |
3147 | if (POINTER_TYPE_P (vtype)) |
3148 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); |
3149 | else |
3150 | t = fold_build2 (PLUS_EXPR, vtype, fd->loops[i].v, fd->loops[i].step); |
3151 | t = force_gimple_operand_gsi (&gsi, t, |
3152 | DECL_P (fd->loops[i].v) |
3153 | && TREE_ADDRESSABLE (fd->loops[i].v), |
3154 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
3155 | stmt = gimple_build_assign (fd->loops[i].v, t); |
3156 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
3157 | |
3158 | if (fd->loops[i].non_rect_referenced) |
3159 | { |
3160 | basic_block update_bb = NULL, prev_bb = NULL; |
3161 | for (int j = i + 1; j <= fd->last_nonrect; j++) |
3162 | if (j - fd->loops[j].outer == i) |
3163 | { |
3164 | tree n1, n2; |
3165 | struct omp_for_data_loop *l = &fd->loops[j]; |
3166 | basic_block this_bb = create_empty_bb (last_bb); |
3167 | add_bb_to_loop (this_bb, last_bb->loop_father); |
3168 | gimple_stmt_iterator gsi2 = gsi_start_bb (bb: this_bb); |
3169 | if (prev_bb) |
3170 | { |
3171 | e = make_edge (prev_bb, this_bb, EDGE_TRUE_VALUE); |
3172 | e->probability |
3173 | = profile_probability::guessed_always ().apply_scale (num: 7, |
3174 | den: 8); |
3175 | set_immediate_dominator (CDI_DOMINATORS, this_bb, prev_bb); |
3176 | } |
3177 | if (l->m1) |
3178 | { |
3179 | if (POINTER_TYPE_P (TREE_TYPE (l->v))) |
3180 | t = fold_build_pointer_plus (fd->loops[i].v, l->n1); |
3181 | else |
3182 | { |
3183 | t = fold_build2 (MULT_EXPR, TREE_TYPE (l->m1), l->m1, |
3184 | fd->loops[i].v); |
3185 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (l->v), |
3186 | t, l->n1); |
3187 | } |
3188 | n1 = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3189 | false, |
3190 | GSI_CONTINUE_LINKING); |
3191 | stmt = gimple_build_assign (l->v, n1); |
3192 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); |
3193 | n1 = l->v; |
3194 | } |
3195 | else |
3196 | n1 = force_gimple_operand_gsi (&gsi2, l->n1, true, |
3197 | NULL_TREE, false, |
3198 | GSI_CONTINUE_LINKING); |
3199 | if (l->m2) |
3200 | { |
3201 | if (POINTER_TYPE_P (TREE_TYPE (l->v))) |
3202 | t = fold_build_pointer_plus (fd->loops[i].v, l->n2); |
3203 | else |
3204 | { |
3205 | t = fold_build2 (MULT_EXPR, TREE_TYPE (l->m2), l->m2, |
3206 | fd->loops[i].v); |
3207 | t = fold_build2 (PLUS_EXPR, |
3208 | TREE_TYPE (nonrect_bounds[j]), |
3209 | t, unshare_expr (l->n2)); |
3210 | } |
3211 | n2 = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3212 | false, |
3213 | GSI_CONTINUE_LINKING); |
3214 | stmt = gimple_build_assign (nonrect_bounds[j], n2); |
3215 | gsi_insert_after (&gsi2, stmt, GSI_CONTINUE_LINKING); |
3216 | n2 = nonrect_bounds[j]; |
3217 | } |
3218 | else |
3219 | n2 = force_gimple_operand_gsi (&gsi2, unshare_expr (l->n2), |
3220 | true, NULL_TREE, false, |
3221 | GSI_CONTINUE_LINKING); |
3222 | gcond *cond_stmt |
3223 | = gimple_build_cond (l->cond_code, n1, n2, |
3224 | NULL_TREE, NULL_TREE); |
3225 | gsi_insert_after (&gsi2, cond_stmt, GSI_CONTINUE_LINKING); |
3226 | if (update_bb == NULL) |
3227 | update_bb = this_bb; |
3228 | e = make_edge (this_bb, bb, EDGE_FALSE_VALUE); |
3229 | e->probability = profile_probability::guessed_always () / 8; |
3230 | if (prev_bb == NULL) |
3231 | set_immediate_dominator (CDI_DOMINATORS, this_bb, bb); |
3232 | prev_bb = this_bb; |
3233 | } |
3234 | e = make_edge (prev_bb, body_bb, EDGE_TRUE_VALUE); |
3235 | e->probability |
3236 | = profile_probability::guessed_always ().apply_scale (num: 7, den: 8); |
3237 | body_bb = update_bb; |
3238 | } |
3239 | |
3240 | if (i > 0) |
3241 | { |
3242 | if (fd->loops[i].m2) |
3243 | t = nonrect_bounds[i]; |
3244 | else |
3245 | t = unshare_expr (fd->loops[i].n2); |
3246 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
3247 | false, GSI_CONTINUE_LINKING); |
3248 | tree v = fd->loops[i].v; |
3249 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) |
3250 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, |
3251 | false, GSI_CONTINUE_LINKING); |
3252 | t = fold_build2 (fd->loops[i].cond_code, boolean_type_node, v, t); |
3253 | stmt = gimple_build_cond_empty (cond: t); |
3254 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
3255 | if (walk_tree (gimple_cond_lhs_ptr (as_a <gcond *> (stmt)), |
3256 | expand_omp_regimplify_p, NULL, NULL) |
3257 | || walk_tree (gimple_cond_rhs_ptr (as_a <gcond *> (stmt)), |
3258 | expand_omp_regimplify_p, NULL, NULL)) |
3259 | gimple_regimplify_operands (stmt, &gsi); |
3260 | e = make_edge (bb, body_bb, EDGE_TRUE_VALUE); |
3261 | e->probability = profile_probability::guessed_always ().apply_scale (num: 7, den: 8); |
3262 | } |
3263 | else |
3264 | make_edge (bb, body_bb, EDGE_FALLTHRU); |
3265 | set_immediate_dominator (CDI_DOMINATORS, bb, last_bb); |
3266 | last_bb = bb; |
3267 | } |
3268 | |
3269 | return collapse_bb; |
3270 | } |
3271 | |
3272 | /* Expand #pragma omp ordered depend(source). */ |
3273 | |
3274 | static void |
3275 | expand_omp_ordered_source (gimple_stmt_iterator *gsi, struct omp_for_data *fd, |
3276 | tree *counts, location_t loc) |
3277 | { |
3278 | enum built_in_function source_ix |
3279 | = fd->iter_type == long_integer_type_node |
3280 | ? BUILT_IN_GOMP_DOACROSS_POST : BUILT_IN_GOMP_DOACROSS_ULL_POST; |
3281 | gimple *g |
3282 | = gimple_build_call (builtin_decl_explicit (fncode: source_ix), 1, |
3283 | build_fold_addr_expr (counts[fd->ordered])); |
3284 | gimple_set_location (g, location: loc); |
3285 | gsi_insert_before (gsi, g, GSI_SAME_STMT); |
3286 | } |
3287 | |
3288 | /* Expand a single depend from #pragma omp ordered depend(sink:...). */ |
3289 | |
3290 | static void |
3291 | expand_omp_ordered_sink (gimple_stmt_iterator *gsi, struct omp_for_data *fd, |
3292 | tree *counts, tree c, location_t loc, |
3293 | basic_block cont_bb) |
3294 | { |
3295 | auto_vec<tree, 10> args; |
3296 | enum built_in_function sink_ix |
3297 | = fd->iter_type == long_integer_type_node |
3298 | ? BUILT_IN_GOMP_DOACROSS_WAIT : BUILT_IN_GOMP_DOACROSS_ULL_WAIT; |
3299 | tree t, off, coff = NULL_TREE, deps = OMP_CLAUSE_DECL (c), cond = NULL_TREE; |
3300 | int i; |
3301 | gimple_stmt_iterator gsi2 = *gsi; |
3302 | bool warned_step = false; |
3303 | |
3304 | if (deps == NULL) |
3305 | { |
3306 | /* Handle doacross(sink: omp_cur_iteration - 1). */ |
3307 | gsi_prev (i: &gsi2); |
3308 | edge e1 = split_block (gsi_bb (i: gsi2), gsi_stmt (i: gsi2)); |
3309 | edge e2 = split_block_after_labels (e1->dest); |
3310 | gsi2 = gsi_after_labels (bb: e1->dest); |
3311 | *gsi = gsi_last_bb (bb: e1->src); |
3312 | gimple_stmt_iterator gsi3 = *gsi; |
3313 | |
3314 | if (counts[fd->collapse - 1]) |
3315 | { |
3316 | gcc_assert (fd->collapse == 1); |
3317 | t = counts[fd->collapse - 1]; |
3318 | } |
3319 | else if (fd->collapse > 1) |
3320 | t = fd->loop.v; |
3321 | else |
3322 | { |
3323 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), |
3324 | fd->loops[0].v, fd->loops[0].n1); |
3325 | t = fold_convert (fd->iter_type, t); |
3326 | } |
3327 | |
3328 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, |
3329 | false, GSI_CONTINUE_LINKING); |
3330 | gsi_insert_after (gsi, gimple_build_cond (NE_EXPR, t, |
3331 | build_zero_cst (TREE_TYPE (t)), |
3332 | NULL_TREE, NULL_TREE), |
3333 | GSI_NEW_STMT); |
3334 | |
3335 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, |
3336 | build_minus_one_cst (TREE_TYPE (t))); |
3337 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3338 | true, GSI_SAME_STMT); |
3339 | args.safe_push (obj: t); |
3340 | for (i = fd->collapse; i < fd->ordered; i++) |
3341 | { |
3342 | t = counts[fd->ordered + 2 + (i - fd->collapse)]; |
3343 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, |
3344 | build_minus_one_cst (TREE_TYPE (t))); |
3345 | t = fold_convert (fd->iter_type, t); |
3346 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3347 | true, GSI_SAME_STMT); |
3348 | args.safe_push (obj: t); |
3349 | } |
3350 | |
3351 | gimple *g = gimple_build_call_vec (builtin_decl_explicit (fncode: sink_ix), |
3352 | args); |
3353 | gimple_set_location (g, location: loc); |
3354 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); |
3355 | |
3356 | edge e3 = make_edge (e1->src, e2->dest, EDGE_FALSE_VALUE); |
3357 | e3->probability = profile_probability::guessed_always () / 8; |
3358 | e1->probability = e3->probability.invert (); |
3359 | e1->flags = EDGE_TRUE_VALUE; |
3360 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e1->src); |
3361 | |
3362 | if (fd->ordered > fd->collapse && cont_bb) |
3363 | { |
3364 | if (counts[fd->ordered + 1] == NULL_TREE) |
3365 | counts[fd->ordered + 1] |
3366 | = create_tmp_var (boolean_type_node, ".first" ); |
3367 | |
3368 | edge e4; |
3369 | if (gsi_end_p (i: gsi3)) |
3370 | e4 = split_block_after_labels (e1->src); |
3371 | else |
3372 | { |
3373 | gsi_prev (i: &gsi3); |
3374 | e4 = split_block (gsi_bb (i: gsi3), gsi_stmt (i: gsi3)); |
3375 | } |
3376 | gsi3 = gsi_last_bb (bb: e4->src); |
3377 | |
3378 | gsi_insert_after (&gsi3, |
3379 | gimple_build_cond (NE_EXPR, |
3380 | counts[fd->ordered + 1], |
3381 | boolean_false_node, |
3382 | NULL_TREE, NULL_TREE), |
3383 | GSI_NEW_STMT); |
3384 | |
3385 | edge e5 = make_edge (e4->src, e2->dest, EDGE_FALSE_VALUE); |
3386 | e4->probability = profile_probability::guessed_always () / 8; |
3387 | e5->probability = e4->probability.invert (); |
3388 | e4->flags = EDGE_TRUE_VALUE; |
3389 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e4->src); |
3390 | } |
3391 | |
3392 | *gsi = gsi_after_labels (bb: e2->dest); |
3393 | return; |
3394 | } |
3395 | for (i = 0; i < fd->ordered; i++) |
3396 | { |
3397 | tree step = NULL_TREE; |
3398 | off = TREE_PURPOSE (deps); |
3399 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) |
3400 | { |
3401 | step = TREE_OPERAND (off, 1); |
3402 | off = TREE_OPERAND (off, 0); |
3403 | } |
3404 | if (!integer_zerop (off)) |
3405 | { |
3406 | gcc_assert (fd->loops[i].cond_code == LT_EXPR |
3407 | || fd->loops[i].cond_code == GT_EXPR); |
3408 | bool forward = fd->loops[i].cond_code == LT_EXPR; |
3409 | if (step) |
3410 | { |
3411 | /* Non-simple Fortran DO loops. If step is variable, |
3412 | we don't know at compile even the direction, so can't |
3413 | warn. */ |
3414 | if (TREE_CODE (step) != INTEGER_CST) |
3415 | break; |
3416 | forward = tree_int_cst_sgn (step) != -1; |
3417 | } |
3418 | if (forward ^ OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3419 | warning_at (loc, 0, "%qs clause with %<sink%> modifier " |
3420 | "waiting for lexically later iteration" , |
3421 | OMP_CLAUSE_DOACROSS_DEPEND (c) |
3422 | ? "depend" : "doacross" ); |
3423 | break; |
3424 | } |
3425 | deps = TREE_CHAIN (deps); |
3426 | } |
3427 | /* If all offsets corresponding to the collapsed loops are zero, |
3428 | this depend clause can be ignored. FIXME: but there is still a |
3429 | flush needed. We need to emit one __sync_synchronize () for it |
3430 | though (perhaps conditionally)? Solve this together with the |
3431 | conservative dependence folding optimization. |
3432 | if (i >= fd->collapse) |
3433 | return; */ |
3434 | |
3435 | deps = OMP_CLAUSE_DECL (c); |
3436 | gsi_prev (i: &gsi2); |
3437 | edge e1 = split_block (gsi_bb (i: gsi2), gsi_stmt (i: gsi2)); |
3438 | edge e2 = split_block_after_labels (e1->dest); |
3439 | |
3440 | gsi2 = gsi_after_labels (bb: e1->dest); |
3441 | *gsi = gsi_last_bb (bb: e1->src); |
3442 | for (i = 0; i < fd->ordered; i++) |
3443 | { |
3444 | tree itype = TREE_TYPE (fd->loops[i].v); |
3445 | tree step = NULL_TREE; |
3446 | tree orig_off = NULL_TREE; |
3447 | if (POINTER_TYPE_P (itype)) |
3448 | itype = sizetype; |
3449 | if (i) |
3450 | deps = TREE_CHAIN (deps); |
3451 | off = TREE_PURPOSE (deps); |
3452 | if (TREE_CODE (off) == TRUNC_DIV_EXPR) |
3453 | { |
3454 | step = TREE_OPERAND (off, 1); |
3455 | off = TREE_OPERAND (off, 0); |
3456 | gcc_assert (fd->loops[i].cond_code == LT_EXPR |
3457 | && integer_onep (fd->loops[i].step) |
3458 | && !POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))); |
3459 | } |
3460 | tree s = fold_convert_loc (loc, itype, step ? step : fd->loops[i].step); |
3461 | if (step) |
3462 | { |
3463 | off = fold_convert_loc (loc, itype, off); |
3464 | orig_off = off; |
3465 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); |
3466 | } |
3467 | |
3468 | if (integer_zerop (off)) |
3469 | t = boolean_true_node; |
3470 | else |
3471 | { |
3472 | tree a; |
3473 | tree co = fold_convert_loc (loc, itype, off); |
3474 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) |
3475 | { |
3476 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3477 | co = fold_build1_loc (loc, NEGATE_EXPR, itype, co); |
3478 | a = fold_build2_loc (loc, POINTER_PLUS_EXPR, |
3479 | TREE_TYPE (fd->loops[i].v), fd->loops[i].v, |
3480 | co); |
3481 | } |
3482 | else if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3483 | a = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), |
3484 | fd->loops[i].v, co); |
3485 | else |
3486 | a = fold_build2_loc (loc, PLUS_EXPR, TREE_TYPE (fd->loops[i].v), |
3487 | fd->loops[i].v, co); |
3488 | if (step) |
3489 | { |
3490 | tree t1, t2; |
3491 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3492 | t1 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, |
3493 | fd->loops[i].n1); |
3494 | else |
3495 | t1 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, |
3496 | fd->loops[i].n2); |
3497 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3498 | t2 = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, |
3499 | fd->loops[i].n2); |
3500 | else |
3501 | t2 = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, |
3502 | fd->loops[i].n1); |
3503 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, |
3504 | step, build_int_cst (TREE_TYPE (step), 0)); |
3505 | if (TREE_CODE (step) != INTEGER_CST) |
3506 | { |
3507 | t1 = unshare_expr (t1); |
3508 | t1 = force_gimple_operand_gsi (gsi, t1, true, NULL_TREE, |
3509 | false, GSI_CONTINUE_LINKING); |
3510 | t2 = unshare_expr (t2); |
3511 | t2 = force_gimple_operand_gsi (gsi, t2, true, NULL_TREE, |
3512 | false, GSI_CONTINUE_LINKING); |
3513 | } |
3514 | t = fold_build3_loc (loc, COND_EXPR, boolean_type_node, |
3515 | t, t2, t1); |
3516 | } |
3517 | else if (fd->loops[i].cond_code == LT_EXPR) |
3518 | { |
3519 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3520 | t = fold_build2_loc (loc, GE_EXPR, boolean_type_node, a, |
3521 | fd->loops[i].n1); |
3522 | else |
3523 | t = fold_build2_loc (loc, LT_EXPR, boolean_type_node, a, |
3524 | fd->loops[i].n2); |
3525 | } |
3526 | else if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3527 | t = fold_build2_loc (loc, GT_EXPR, boolean_type_node, a, |
3528 | fd->loops[i].n2); |
3529 | else |
3530 | t = fold_build2_loc (loc, LE_EXPR, boolean_type_node, a, |
3531 | fd->loops[i].n1); |
3532 | } |
3533 | if (cond) |
3534 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, cond, t); |
3535 | else |
3536 | cond = t; |
3537 | |
3538 | off = fold_convert_loc (loc, itype, off); |
3539 | |
3540 | if (step |
3541 | || (fd->loops[i].cond_code == LT_EXPR |
3542 | ? !integer_onep (fd->loops[i].step) |
3543 | : !integer_minus_onep (fd->loops[i].step))) |
3544 | { |
3545 | if (step == NULL_TREE |
3546 | && TYPE_UNSIGNED (itype) |
3547 | && fd->loops[i].cond_code == GT_EXPR) |
3548 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, off, |
3549 | fold_build1_loc (loc, NEGATE_EXPR, itype, |
3550 | s)); |
3551 | else |
3552 | t = fold_build2_loc (loc, TRUNC_MOD_EXPR, itype, |
3553 | orig_off ? orig_off : off, s); |
3554 | t = fold_build2_loc (loc, EQ_EXPR, boolean_type_node, t, |
3555 | build_int_cst (itype, 0)); |
3556 | if (integer_zerop (t) && !warned_step) |
3557 | { |
3558 | warning_at (loc, 0, "%qs clause with %<sink%> modifier " |
3559 | "refers to iteration never in the iteration " |
3560 | "space" , |
3561 | OMP_CLAUSE_DOACROSS_DEPEND (c) |
3562 | ? "depend" : "doacross" ); |
3563 | warned_step = true; |
3564 | } |
3565 | cond = fold_build2_loc (loc, BIT_AND_EXPR, boolean_type_node, |
3566 | cond, t); |
3567 | } |
3568 | |
3569 | if (i <= fd->collapse - 1 && fd->collapse > 1) |
3570 | t = fd->loop.v; |
3571 | else if (counts[i]) |
3572 | t = counts[i]; |
3573 | else |
3574 | { |
3575 | t = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (fd->loops[i].v), |
3576 | fd->loops[i].v, fd->loops[i].n1); |
3577 | t = fold_convert_loc (loc, fd->iter_type, t); |
3578 | } |
3579 | if (step) |
3580 | /* We have divided off by step already earlier. */; |
3581 | else if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) |
3582 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, |
3583 | fold_build1_loc (loc, NEGATE_EXPR, itype, |
3584 | s)); |
3585 | else |
3586 | off = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, off, s); |
3587 | if (OMP_CLAUSE_DOACROSS_SINK_NEGATIVE (deps)) |
3588 | off = fold_build1_loc (loc, NEGATE_EXPR, itype, off); |
3589 | off = fold_convert_loc (loc, fd->iter_type, off); |
3590 | if (i <= fd->collapse - 1 && fd->collapse > 1) |
3591 | { |
3592 | if (i) |
3593 | off = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, coff, |
3594 | off); |
3595 | if (i < fd->collapse - 1) |
3596 | { |
3597 | coff = fold_build2_loc (loc, MULT_EXPR, fd->iter_type, off, |
3598 | counts[i]); |
3599 | continue; |
3600 | } |
3601 | } |
3602 | off = unshare_expr (off); |
3603 | t = fold_build2_loc (loc, PLUS_EXPR, fd->iter_type, t, off); |
3604 | t = force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
3605 | true, GSI_SAME_STMT); |
3606 | args.safe_push (obj: t); |
3607 | } |
3608 | gimple *g = gimple_build_call_vec (builtin_decl_explicit (fncode: sink_ix), args); |
3609 | gimple_set_location (g, location: loc); |
3610 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); |
3611 | |
3612 | cond = unshare_expr (cond); |
3613 | cond = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE, false, |
3614 | GSI_CONTINUE_LINKING); |
3615 | gsi_insert_after (gsi, gimple_build_cond_empty (cond), GSI_NEW_STMT); |
3616 | edge e3 = make_edge (e1->src, e2->dest, EDGE_FALSE_VALUE); |
3617 | e3->probability = profile_probability::guessed_always () / 8; |
3618 | e1->probability = e3->probability.invert (); |
3619 | e1->flags = EDGE_TRUE_VALUE; |
3620 | set_immediate_dominator (CDI_DOMINATORS, e2->dest, e1->src); |
3621 | |
3622 | *gsi = gsi_after_labels (bb: e2->dest); |
3623 | } |
3624 | |
3625 | /* Expand all #pragma omp ordered depend(source) and |
3626 | #pragma omp ordered depend(sink:...) constructs in the current |
3627 | #pragma omp for ordered(n) region. */ |
3628 | |
3629 | static void |
3630 | expand_omp_ordered_source_sink (struct omp_region *region, |
3631 | struct omp_for_data *fd, tree *counts, |
3632 | basic_block cont_bb) |
3633 | { |
3634 | struct omp_region *inner; |
3635 | int i; |
3636 | for (i = fd->collapse - 1; i < fd->ordered; i++) |
3637 | if (i == fd->collapse - 1 && fd->collapse > 1) |
3638 | counts[i] = NULL_TREE; |
3639 | else if (i >= fd->collapse && !cont_bb) |
3640 | counts[i] = build_zero_cst (fd->iter_type); |
3641 | else if (!POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v)) |
3642 | && integer_onep (fd->loops[i].step)) |
3643 | counts[i] = NULL_TREE; |
3644 | else |
3645 | counts[i] = create_tmp_var (fd->iter_type, ".orditer" ); |
3646 | tree atype |
3647 | = build_array_type_nelts (fd->iter_type, fd->ordered - fd->collapse + 1); |
3648 | counts[fd->ordered] = create_tmp_var (atype, ".orditera" ); |
3649 | TREE_ADDRESSABLE (counts[fd->ordered]) = 1; |
3650 | counts[fd->ordered + 1] = NULL_TREE; |
3651 | |
3652 | for (inner = region->inner; inner; inner = inner->next) |
3653 | if (inner->type == GIMPLE_OMP_ORDERED) |
3654 | { |
3655 | gomp_ordered *ord_stmt = inner->ord_stmt; |
3656 | gimple_stmt_iterator gsi = gsi_for_stmt (ord_stmt); |
3657 | location_t loc = gimple_location (g: ord_stmt); |
3658 | tree c; |
3659 | for (c = gimple_omp_ordered_clauses (ord_stmt); |
3660 | c; c = OMP_CLAUSE_CHAIN (c)) |
3661 | if (OMP_CLAUSE_DOACROSS_KIND (c) == OMP_CLAUSE_DOACROSS_SOURCE) |
3662 | break; |
3663 | if (c) |
3664 | expand_omp_ordered_source (gsi: &gsi, fd, counts, loc); |
3665 | for (c = gimple_omp_ordered_clauses (ord_stmt); |
3666 | c; c = OMP_CLAUSE_CHAIN (c)) |
3667 | if (OMP_CLAUSE_DOACROSS_KIND (c) == OMP_CLAUSE_DOACROSS_SINK) |
3668 | expand_omp_ordered_sink (gsi: &gsi, fd, counts, c, loc, cont_bb); |
3669 | gsi_remove (&gsi, true); |
3670 | } |
3671 | } |
3672 | |
3673 | /* Wrap the body into fd->ordered - fd->collapse loops that aren't |
3674 | collapsed. */ |
3675 | |
3676 | static basic_block |
3677 | expand_omp_for_ordered_loops (struct omp_for_data *fd, tree *counts, |
3678 | basic_block cont_bb, basic_block body_bb, |
3679 | basic_block l0_bb, bool ordered_lastprivate) |
3680 | { |
3681 | if (fd->ordered == fd->collapse) |
3682 | return cont_bb; |
3683 | |
3684 | if (!cont_bb) |
3685 | { |
3686 | gimple_stmt_iterator gsi = gsi_after_labels (bb: body_bb); |
3687 | for (int i = fd->collapse; i < fd->ordered; i++) |
3688 | { |
3689 | tree type = TREE_TYPE (fd->loops[i].v); |
3690 | tree n1 = fold_convert (type, fd->loops[i].n1); |
3691 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i].v, from: n1); |
3692 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], |
3693 | size_int (i - fd->collapse + 1), |
3694 | NULL_TREE, NULL_TREE); |
3695 | expand_omp_build_assign (gsi_p: &gsi, to: aref, from: build_zero_cst (fd->iter_type)); |
3696 | } |
3697 | return NULL; |
3698 | } |
3699 | |
3700 | for (int i = fd->ordered - 1; i >= fd->collapse; i--) |
3701 | { |
3702 | tree t, type = TREE_TYPE (fd->loops[i].v); |
3703 | gimple_stmt_iterator gsi = gsi_after_labels (bb: body_bb); |
3704 | if (counts[fd->ordered + 1] && i == fd->collapse) |
3705 | expand_omp_build_assign (gsi_p: &gsi, to: counts[fd->ordered + 1], |
3706 | boolean_true_node); |
3707 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i].v, |
3708 | fold_convert (type, fd->loops[i].n1)); |
3709 | if (counts[i]) |
3710 | expand_omp_build_assign (gsi_p: &gsi, to: counts[i], |
3711 | from: build_zero_cst (fd->iter_type)); |
3712 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], |
3713 | size_int (i - fd->collapse + 1), |
3714 | NULL_TREE, NULL_TREE); |
3715 | expand_omp_build_assign (gsi_p: &gsi, to: aref, from: build_zero_cst (fd->iter_type)); |
3716 | if (!gsi_end_p (i: gsi)) |
3717 | gsi_prev (i: &gsi); |
3718 | else |
3719 | gsi = gsi_last_bb (bb: body_bb); |
3720 | edge e1 = split_block (body_bb, gsi_stmt (i: gsi)); |
3721 | basic_block new_body = e1->dest; |
3722 | if (body_bb == cont_bb) |
3723 | cont_bb = new_body; |
3724 | edge e2 = NULL; |
3725 | basic_block ; |
3726 | if (EDGE_COUNT (cont_bb->preds) > 0) |
3727 | { |
3728 | gsi = gsi_last_bb (bb: cont_bb); |
3729 | if (POINTER_TYPE_P (type)) |
3730 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); |
3731 | else |
3732 | t = fold_build2 (PLUS_EXPR, type, fd->loops[i].v, |
3733 | fold_convert (type, fd->loops[i].step)); |
3734 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i].v, from: t); |
3735 | if (counts[i]) |
3736 | { |
3737 | t = fold_build2 (PLUS_EXPR, fd->iter_type, counts[i], |
3738 | build_int_cst (fd->iter_type, 1)); |
3739 | expand_omp_build_assign (gsi_p: &gsi, to: counts[i], from: t); |
3740 | t = counts[i]; |
3741 | } |
3742 | else |
3743 | { |
3744 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[i].v), |
3745 | fd->loops[i].v, fd->loops[i].n1); |
3746 | t = fold_convert (fd->iter_type, t); |
3747 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
3748 | true, GSI_SAME_STMT); |
3749 | } |
3750 | aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], |
3751 | size_int (i - fd->collapse + 1), |
3752 | NULL_TREE, NULL_TREE); |
3753 | expand_omp_build_assign (gsi_p: &gsi, to: aref, from: t); |
3754 | if (counts[fd->ordered + 1] && i == fd->ordered - 1) |
3755 | expand_omp_build_assign (gsi_p: &gsi, to: counts[fd->ordered + 1], |
3756 | boolean_false_node); |
3757 | gsi_prev (i: &gsi); |
3758 | e2 = split_block (cont_bb, gsi_stmt (i: gsi)); |
3759 | new_header = e2->dest; |
3760 | } |
3761 | else |
3762 | new_header = cont_bb; |
3763 | gsi = gsi_after_labels (bb: new_header); |
3764 | tree v = force_gimple_operand_gsi (&gsi, fd->loops[i].v, true, NULL_TREE, |
3765 | true, GSI_SAME_STMT); |
3766 | tree n2 |
3767 | = force_gimple_operand_gsi (&gsi, fold_convert (type, fd->loops[i].n2), |
3768 | true, NULL_TREE, true, GSI_SAME_STMT); |
3769 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, n2); |
3770 | gsi_insert_before (&gsi, gimple_build_cond_empty (cond: t), GSI_NEW_STMT); |
3771 | edge e3 = split_block (new_header, gsi_stmt (i: gsi)); |
3772 | cont_bb = e3->dest; |
3773 | remove_edge (e1); |
3774 | make_edge (body_bb, new_header, EDGE_FALLTHRU); |
3775 | e3->flags = EDGE_FALSE_VALUE; |
3776 | e3->probability = profile_probability::guessed_always () / 8; |
3777 | e1 = make_edge (new_header, new_body, EDGE_TRUE_VALUE); |
3778 | e1->probability = e3->probability.invert (); |
3779 | |
3780 | set_immediate_dominator (CDI_DOMINATORS, new_header, body_bb); |
3781 | set_immediate_dominator (CDI_DOMINATORS, new_body, new_header); |
3782 | |
3783 | if (e2) |
3784 | { |
3785 | class loop *loop = alloc_loop (); |
3786 | loop->header = new_header; |
3787 | loop->latch = e2->src; |
3788 | add_loop (loop, l0_bb->loop_father); |
3789 | } |
3790 | } |
3791 | |
3792 | /* If there are any lastprivate clauses and it is possible some loops |
3793 | might have zero iterations, ensure all the decls are initialized, |
3794 | otherwise we could crash evaluating C++ class iterators with lastprivate |
3795 | clauses. */ |
3796 | bool need_inits = false; |
3797 | for (int i = fd->collapse; ordered_lastprivate && i < fd->ordered; i++) |
3798 | if (need_inits) |
3799 | { |
3800 | tree type = TREE_TYPE (fd->loops[i].v); |
3801 | gimple_stmt_iterator gsi = gsi_after_labels (bb: body_bb); |
3802 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i].v, |
3803 | fold_convert (type, fd->loops[i].n1)); |
3804 | } |
3805 | else |
3806 | { |
3807 | tree type = TREE_TYPE (fd->loops[i].v); |
3808 | tree this_cond = fold_build2 (fd->loops[i].cond_code, |
3809 | boolean_type_node, |
3810 | fold_convert (type, fd->loops[i].n1), |
3811 | fold_convert (type, fd->loops[i].n2)); |
3812 | if (!integer_onep (this_cond)) |
3813 | need_inits = true; |
3814 | } |
3815 | |
3816 | return cont_bb; |
3817 | } |
3818 | |
3819 | /* A subroutine of expand_omp_for. Generate code for a parallel |
3820 | loop with any schedule. Given parameters: |
3821 | |
3822 | for (V = N1; V cond N2; V += STEP) BODY; |
3823 | |
3824 | where COND is "<" or ">", we generate pseudocode |
3825 | |
3826 | more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0); |
3827 | if (more) goto L0; else goto L3; |
3828 | L0: |
3829 | V = istart0; |
3830 | iend = iend0; |
3831 | L1: |
3832 | BODY; |
3833 | V += STEP; |
3834 | if (V cond iend) goto L1; else goto L2; |
3835 | L2: |
3836 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; |
3837 | L3: |
3838 | |
3839 | If this is a combined omp parallel loop, instead of the call to |
3840 | GOMP_loop_foo_start, we call GOMP_loop_foo_next. |
3841 | If this is gimple_omp_for_combined_p loop, then instead of assigning |
3842 | V and iend in L0 we assign the first two _looptemp_ clause decls of the |
3843 | inner GIMPLE_OMP_FOR and V += STEP; and |
3844 | if (V cond iend) goto L1; else goto L2; are removed. |
3845 | |
3846 | For collapsed loops, given parameters: |
3847 | collapse(3) |
3848 | for (V1 = N11; V1 cond1 N12; V1 += STEP1) |
3849 | for (V2 = N21; V2 cond2 N22; V2 += STEP2) |
3850 | for (V3 = N31; V3 cond3 N32; V3 += STEP3) |
3851 | BODY; |
3852 | |
3853 | we generate pseudocode |
3854 | |
3855 | if (__builtin_expect (N32 cond3 N31, 0)) goto Z0; |
3856 | if (cond3 is <) |
3857 | adj = STEP3 - 1; |
3858 | else |
3859 | adj = STEP3 + 1; |
3860 | count3 = (adj + N32 - N31) / STEP3; |
3861 | if (__builtin_expect (N22 cond2 N21, 0)) goto Z0; |
3862 | if (cond2 is <) |
3863 | adj = STEP2 - 1; |
3864 | else |
3865 | adj = STEP2 + 1; |
3866 | count2 = (adj + N22 - N21) / STEP2; |
3867 | if (__builtin_expect (N12 cond1 N11, 0)) goto Z0; |
3868 | if (cond1 is <) |
3869 | adj = STEP1 - 1; |
3870 | else |
3871 | adj = STEP1 + 1; |
3872 | count1 = (adj + N12 - N11) / STEP1; |
3873 | count = count1 * count2 * count3; |
3874 | goto Z1; |
3875 | Z0: |
3876 | count = 0; |
3877 | Z1: |
3878 | more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0); |
3879 | if (more) goto L0; else goto L3; |
3880 | L0: |
3881 | V = istart0; |
3882 | T = V; |
3883 | V3 = N31 + (T % count3) * STEP3; |
3884 | T = T / count3; |
3885 | V2 = N21 + (T % count2) * STEP2; |
3886 | T = T / count2; |
3887 | V1 = N11 + T * STEP1; |
3888 | iend = iend0; |
3889 | L1: |
3890 | BODY; |
3891 | V += 1; |
3892 | if (V < iend) goto L10; else goto L2; |
3893 | L10: |
3894 | V3 += STEP3; |
3895 | if (V3 cond3 N32) goto L1; else goto L11; |
3896 | L11: |
3897 | V3 = N31; |
3898 | V2 += STEP2; |
3899 | if (V2 cond2 N22) goto L1; else goto L12; |
3900 | L12: |
3901 | V2 = N21; |
3902 | V1 += STEP1; |
3903 | goto L1; |
3904 | L2: |
3905 | if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3; |
3906 | L3: |
3907 | |
3908 | */ |
3909 | |
3910 | static void |
3911 | expand_omp_for_generic (struct omp_region *region, |
3912 | struct omp_for_data *fd, |
3913 | enum built_in_function start_fn, |
3914 | enum built_in_function next_fn, |
3915 | tree sched_arg, |
3916 | gimple *inner_stmt) |
3917 | { |
3918 | tree type, istart0, iend0, iend; |
3919 | tree t, vmain, vback, bias = NULL_TREE; |
3920 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, collapse_bb; |
3921 | basic_block l2_bb = NULL, l3_bb = NULL; |
3922 | gimple_stmt_iterator gsi; |
3923 | gassign *assign_stmt; |
3924 | bool in_combined_parallel = is_combined_parallel (region); |
3925 | bool broken_loop = region->cont == NULL; |
3926 | edge e, ne; |
3927 | tree *counts = NULL; |
3928 | int i; |
3929 | bool ordered_lastprivate = false; |
3930 | |
3931 | gcc_assert (!broken_loop || !in_combined_parallel); |
3932 | gcc_assert (fd->iter_type == long_integer_type_node |
3933 | || !in_combined_parallel); |
3934 | |
3935 | entry_bb = region->entry; |
3936 | cont_bb = region->cont; |
3937 | collapse_bb = NULL; |
3938 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
3939 | gcc_assert (broken_loop |
3940 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); |
3941 | l0_bb = split_edge (FALLTHRU_EDGE (entry_bb)); |
3942 | l1_bb = single_succ (bb: l0_bb); |
3943 | if (!broken_loop) |
3944 | { |
3945 | l2_bb = create_empty_bb (cont_bb); |
3946 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb |
3947 | || (single_succ_edge (BRANCH_EDGE (cont_bb)->dest)->dest |
3948 | == l1_bb)); |
3949 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); |
3950 | } |
3951 | else |
3952 | l2_bb = NULL; |
3953 | l3_bb = BRANCH_EDGE (entry_bb)->dest; |
3954 | exit_bb = region->exit; |
3955 | |
3956 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
3957 | |
3958 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
3959 | if (fd->ordered |
3960 | && omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
3961 | kind: OMP_CLAUSE_LASTPRIVATE)) |
3962 | ordered_lastprivate = false; |
3963 | tree reductions = NULL_TREE; |
3964 | tree mem = NULL_TREE, cond_var = NULL_TREE, condtemp = NULL_TREE; |
3965 | tree memv = NULL_TREE; |
3966 | if (fd->lastprivate_conditional) |
3967 | { |
3968 | tree c = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
3969 | kind: OMP_CLAUSE__CONDTEMP_); |
3970 | if (fd->have_pointer_condtemp) |
3971 | condtemp = OMP_CLAUSE_DECL (c); |
3972 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), kind: OMP_CLAUSE__CONDTEMP_); |
3973 | cond_var = OMP_CLAUSE_DECL (c); |
3974 | } |
3975 | if (sched_arg) |
3976 | { |
3977 | if (fd->have_reductemp) |
3978 | { |
3979 | tree c = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
3980 | kind: OMP_CLAUSE__REDUCTEMP_); |
3981 | reductions = OMP_CLAUSE_DECL (c); |
3982 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); |
3983 | gimple *g = SSA_NAME_DEF_STMT (reductions); |
3984 | reductions = gimple_assign_rhs1 (gs: g); |
3985 | OMP_CLAUSE_DECL (c) = reductions; |
3986 | entry_bb = gimple_bb (g); |
3987 | edge e = split_block (entry_bb, g); |
3988 | if (region->entry == entry_bb) |
3989 | region->entry = e->dest; |
3990 | gsi = gsi_last_bb (bb: entry_bb); |
3991 | } |
3992 | else |
3993 | reductions = null_pointer_node; |
3994 | if (fd->have_pointer_condtemp) |
3995 | { |
3996 | tree type = TREE_TYPE (condtemp); |
3997 | memv = create_tmp_var (type); |
3998 | TREE_ADDRESSABLE (memv) = 1; |
3999 | unsigned HOST_WIDE_INT sz |
4000 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); |
4001 | sz *= fd->lastprivate_conditional; |
4002 | expand_omp_build_assign (gsi_p: &gsi, to: memv, from: build_int_cst (type, sz), |
4003 | after: false); |
4004 | mem = build_fold_addr_expr (memv); |
4005 | } |
4006 | else |
4007 | mem = null_pointer_node; |
4008 | } |
4009 | if (fd->collapse > 1 || fd->ordered) |
4010 | { |
4011 | int first_zero_iter1 = -1, first_zero_iter2 = -1; |
4012 | basic_block zero_iter1_bb = NULL, zero_iter2_bb = NULL, l2_dom_bb = NULL; |
4013 | |
4014 | counts = XALLOCAVEC (tree, fd->ordered |
4015 | ? fd->ordered + 2 |
4016 | + (fd->ordered - fd->collapse) |
4017 | : fd->collapse); |
4018 | expand_omp_for_init_counts (fd, gsi: &gsi, entry_bb, counts, |
4019 | zero_iter1_bb, first_zero_iter1, |
4020 | zero_iter2_bb, first_zero_iter2, l2_dom_bb); |
4021 | |
4022 | if (zero_iter1_bb) |
4023 | { |
4024 | /* Some counts[i] vars might be uninitialized if |
4025 | some loop has zero iterations. But the body shouldn't |
4026 | be executed in that case, so just avoid uninit warnings. */ |
4027 | for (i = first_zero_iter1; |
4028 | i < (fd->ordered ? fd->ordered : fd->collapse); i++) |
4029 | if (SSA_VAR_P (counts[i])) |
4030 | suppress_warning (counts[i], OPT_Wuninitialized); |
4031 | gsi_prev (i: &gsi); |
4032 | e = split_block (entry_bb, gsi_stmt (i: gsi)); |
4033 | entry_bb = e->dest; |
4034 | make_edge (zero_iter1_bb, entry_bb, EDGE_FALLTHRU); |
4035 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
4036 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, |
4037 | get_immediate_dominator (CDI_DOMINATORS, |
4038 | zero_iter1_bb)); |
4039 | } |
4040 | if (zero_iter2_bb) |
4041 | { |
4042 | /* Some counts[i] vars might be uninitialized if |
4043 | some loop has zero iterations. But the body shouldn't |
4044 | be executed in that case, so just avoid uninit warnings. */ |
4045 | for (i = first_zero_iter2; i < fd->ordered; i++) |
4046 | if (SSA_VAR_P (counts[i])) |
4047 | suppress_warning (counts[i], OPT_Wuninitialized); |
4048 | if (zero_iter1_bb) |
4049 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); |
4050 | else |
4051 | { |
4052 | gsi_prev (i: &gsi); |
4053 | e = split_block (entry_bb, gsi_stmt (i: gsi)); |
4054 | entry_bb = e->dest; |
4055 | make_edge (zero_iter2_bb, entry_bb, EDGE_FALLTHRU); |
4056 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
4057 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, |
4058 | get_immediate_dominator |
4059 | (CDI_DOMINATORS, zero_iter2_bb)); |
4060 | } |
4061 | } |
4062 | if (fd->collapse == 1) |
4063 | { |
4064 | counts[0] = fd->loop.n2; |
4065 | fd->loop = fd->loops[0]; |
4066 | } |
4067 | } |
4068 | |
4069 | type = TREE_TYPE (fd->loop.v); |
4070 | istart0 = create_tmp_var (fd->iter_type, ".istart0" ); |
4071 | iend0 = create_tmp_var (fd->iter_type, ".iend0" ); |
4072 | TREE_ADDRESSABLE (istart0) = 1; |
4073 | TREE_ADDRESSABLE (iend0) = 1; |
4074 | |
4075 | /* See if we need to bias by LLONG_MIN. */ |
4076 | if (fd->iter_type == long_long_unsigned_type_node |
4077 | && TREE_CODE (type) == INTEGER_TYPE |
4078 | && !TYPE_UNSIGNED (type) |
4079 | && fd->ordered == 0) |
4080 | { |
4081 | tree n1, n2; |
4082 | |
4083 | if (fd->loop.cond_code == LT_EXPR) |
4084 | { |
4085 | n1 = fd->loop.n1; |
4086 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); |
4087 | } |
4088 | else |
4089 | { |
4090 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); |
4091 | n2 = fd->loop.n1; |
4092 | } |
4093 | if (TREE_CODE (n1) != INTEGER_CST |
4094 | || TREE_CODE (n2) != INTEGER_CST |
4095 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) |
4096 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); |
4097 | } |
4098 | |
4099 | gimple_stmt_iterator gsif = gsi; |
4100 | gsi_prev (i: &gsif); |
4101 | |
4102 | tree arr = NULL_TREE; |
4103 | if (in_combined_parallel) |
4104 | { |
4105 | gcc_assert (fd->ordered == 0); |
4106 | /* In a combined parallel loop, emit a call to |
4107 | GOMP_loop_foo_next. */ |
4108 | t = build_call_expr (builtin_decl_explicit (fncode: next_fn), 2, |
4109 | build_fold_addr_expr (istart0), |
4110 | build_fold_addr_expr (iend0)); |
4111 | } |
4112 | else |
4113 | { |
4114 | tree t0, t1, t2, t3, t4; |
4115 | /* If this is not a combined parallel loop, emit a call to |
4116 | GOMP_loop_foo_start in ENTRY_BB. */ |
4117 | t4 = build_fold_addr_expr (iend0); |
4118 | t3 = build_fold_addr_expr (istart0); |
4119 | if (fd->ordered) |
4120 | { |
4121 | t0 = build_int_cst (unsigned_type_node, |
4122 | fd->ordered - fd->collapse + 1); |
4123 | arr = create_tmp_var (build_array_type_nelts (fd->iter_type, |
4124 | fd->ordered |
4125 | - fd->collapse + 1), |
4126 | ".omp_counts" ); |
4127 | DECL_NAMELESS (arr) = 1; |
4128 | TREE_ADDRESSABLE (arr) = 1; |
4129 | TREE_STATIC (arr) = 1; |
4130 | vec<constructor_elt, va_gc> *v; |
4131 | vec_alloc (v, nelems: fd->ordered - fd->collapse + 1); |
4132 | int idx; |
4133 | |
4134 | for (idx = 0; idx < fd->ordered - fd->collapse + 1; idx++) |
4135 | { |
4136 | tree c; |
4137 | if (idx == 0 && fd->collapse > 1) |
4138 | c = fd->loop.n2; |
4139 | else |
4140 | c = counts[idx + fd->collapse - 1]; |
4141 | tree purpose = size_int (idx); |
4142 | CONSTRUCTOR_APPEND_ELT (v, purpose, c); |
4143 | if (TREE_CODE (c) != INTEGER_CST) |
4144 | TREE_STATIC (arr) = 0; |
4145 | } |
4146 | |
4147 | DECL_INITIAL (arr) = build_constructor (TREE_TYPE (arr), v); |
4148 | if (!TREE_STATIC (arr)) |
4149 | force_gimple_operand_gsi (&gsi, build1 (DECL_EXPR, |
4150 | void_type_node, arr), |
4151 | true, NULL_TREE, true, GSI_SAME_STMT); |
4152 | t1 = build_fold_addr_expr (arr); |
4153 | t2 = NULL_TREE; |
4154 | } |
4155 | else |
4156 | { |
4157 | t2 = fold_convert (fd->iter_type, fd->loop.step); |
4158 | t1 = fd->loop.n2; |
4159 | t0 = fd->loop.n1; |
4160 | if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
4161 | { |
4162 | tree innerc |
4163 | = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
4164 | kind: OMP_CLAUSE__LOOPTEMP_); |
4165 | gcc_assert (innerc); |
4166 | t0 = OMP_CLAUSE_DECL (innerc); |
4167 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
4168 | kind: OMP_CLAUSE__LOOPTEMP_); |
4169 | gcc_assert (innerc); |
4170 | t1 = OMP_CLAUSE_DECL (innerc); |
4171 | } |
4172 | if (POINTER_TYPE_P (TREE_TYPE (t0)) |
4173 | && TYPE_PRECISION (TREE_TYPE (t0)) |
4174 | != TYPE_PRECISION (fd->iter_type)) |
4175 | { |
4176 | /* Avoid casting pointers to integer of a different size. */ |
4177 | tree itype = signed_type_for (type); |
4178 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); |
4179 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); |
4180 | } |
4181 | else |
4182 | { |
4183 | t1 = fold_convert (fd->iter_type, t1); |
4184 | t0 = fold_convert (fd->iter_type, t0); |
4185 | } |
4186 | if (bias) |
4187 | { |
4188 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); |
4189 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); |
4190 | } |
4191 | } |
4192 | if (fd->iter_type == long_integer_type_node || fd->ordered) |
4193 | { |
4194 | if (fd->chunk_size) |
4195 | { |
4196 | t = fold_convert (fd->iter_type, fd->chunk_size); |
4197 | t = omp_adjust_chunk_size (chunk_size: t, simd_schedule: fd->simd_schedule); |
4198 | if (sched_arg) |
4199 | { |
4200 | if (fd->ordered) |
4201 | t = build_call_expr (builtin_decl_explicit (fncode: start_fn), |
4202 | 8, t0, t1, sched_arg, t, t3, t4, |
4203 | reductions, mem); |
4204 | else |
4205 | t = build_call_expr (builtin_decl_explicit (fncode: start_fn), |
4206 | 9, t0, t1, t2, sched_arg, t, t3, t4, |
4207 | reductions, mem); |
4208 | } |
4209 | else if (fd->ordered) |
4210 | t = build_call_expr (builtin_decl_explicit (fncode: start_fn), |
4211 | 5, t0, t1, t, t3, t4); |
4212 | else |
4213 | t = build_call_expr (builtin_decl_explicit (fncode: start_fn), |
4214 | 6, t0, t1, t2, t, t3, t4); |
4215 | } |
4216 | else if (fd->ordered) |
4217 | t = build_call_expr (builtin_decl_explicit (fncode: start_fn), |
4218 | 4, t0, t1, t3, t4); |
4219 | else |
4220 | t = build_call_expr (builtin_decl_explicit (fncode: start_fn), |
4221 | 5, t0, t1, t2, t3, t4); |
4222 | } |
4223 | else |
4224 | { |
4225 | tree t5; |
4226 | tree c_bool_type; |
4227 | tree bfn_decl; |
4228 | |
4229 | /* The GOMP_loop_ull_*start functions have additional boolean |
4230 | argument, true for < loops and false for > loops. |
4231 | In Fortran, the C bool type can be different from |
4232 | boolean_type_node. */ |
4233 | bfn_decl = builtin_decl_explicit (fncode: start_fn); |
4234 | c_bool_type = TREE_TYPE (TREE_TYPE (bfn_decl)); |
4235 | t5 = build_int_cst (c_bool_type, |
4236 | fd->loop.cond_code == LT_EXPR ? 1 : 0); |
4237 | if (fd->chunk_size) |
4238 | { |
4239 | tree bfn_decl = builtin_decl_explicit (fncode: start_fn); |
4240 | t = fold_convert (fd->iter_type, fd->chunk_size); |
4241 | t = omp_adjust_chunk_size (chunk_size: t, simd_schedule: fd->simd_schedule); |
4242 | if (sched_arg) |
4243 | t = build_call_expr (bfn_decl, 10, t5, t0, t1, t2, sched_arg, |
4244 | t, t3, t4, reductions, mem); |
4245 | else |
4246 | t = build_call_expr (bfn_decl, 7, t5, t0, t1, t2, t, t3, t4); |
4247 | } |
4248 | else |
4249 | t = build_call_expr (builtin_decl_explicit (fncode: start_fn), |
4250 | 6, t5, t0, t1, t2, t3, t4); |
4251 | } |
4252 | } |
4253 | if (TREE_TYPE (t) != boolean_type_node) |
4254 | t = fold_build2 (NE_EXPR, boolean_type_node, |
4255 | t, build_int_cst (TREE_TYPE (t), 0)); |
4256 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4257 | true, GSI_SAME_STMT); |
4258 | if (arr && !TREE_STATIC (arr)) |
4259 | { |
4260 | tree clobber = build_clobber (TREE_TYPE (arr)); |
4261 | gsi_insert_before (&gsi, gimple_build_assign (arr, clobber), |
4262 | GSI_SAME_STMT); |
4263 | } |
4264 | if (fd->have_pointer_condtemp) |
4265 | expand_omp_build_assign (gsi_p: &gsi, to: condtemp, from: memv, after: false); |
4266 | if (fd->have_reductemp) |
4267 | { |
4268 | gimple *g = gsi_stmt (i: gsi); |
4269 | gsi_remove (&gsi, true); |
4270 | release_ssa_name (name: gimple_assign_lhs (gs: g)); |
4271 | |
4272 | entry_bb = region->entry; |
4273 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
4274 | |
4275 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
4276 | } |
4277 | gsi_insert_after (&gsi, gimple_build_cond_empty (cond: t), GSI_SAME_STMT); |
4278 | |
4279 | /* Remove the GIMPLE_OMP_FOR statement. */ |
4280 | gsi_remove (&gsi, true); |
4281 | |
4282 | if (gsi_end_p (i: gsif)) |
4283 | gsif = gsi_after_labels (bb: gsi_bb (i: gsif)); |
4284 | gsi_next (i: &gsif); |
4285 | |
4286 | /* Iteration setup for sequential loop goes in L0_BB. */ |
4287 | tree startvar = fd->loop.v; |
4288 | tree endvar = NULL_TREE; |
4289 | |
4290 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
4291 | { |
4292 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_FOR |
4293 | && gimple_omp_for_kind (inner_stmt) |
4294 | == GF_OMP_FOR_KIND_SIMD); |
4295 | tree innerc = omp_find_clause (clauses: gimple_omp_for_clauses (gs: inner_stmt), |
4296 | kind: OMP_CLAUSE__LOOPTEMP_); |
4297 | gcc_assert (innerc); |
4298 | startvar = OMP_CLAUSE_DECL (innerc); |
4299 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
4300 | kind: OMP_CLAUSE__LOOPTEMP_); |
4301 | gcc_assert (innerc); |
4302 | endvar = OMP_CLAUSE_DECL (innerc); |
4303 | } |
4304 | |
4305 | gsi = gsi_start_bb (bb: l0_bb); |
4306 | t = istart0; |
4307 | if (fd->ordered && fd->collapse == 1) |
4308 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, |
4309 | fold_convert (fd->iter_type, fd->loop.step)); |
4310 | else if (bias) |
4311 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); |
4312 | if (fd->ordered && fd->collapse == 1) |
4313 | { |
4314 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
4315 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), |
4316 | fd->loop.n1, fold_convert (sizetype, t)); |
4317 | else |
4318 | { |
4319 | t = fold_convert (TREE_TYPE (startvar), t); |
4320 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), |
4321 | fd->loop.n1, t); |
4322 | } |
4323 | } |
4324 | else |
4325 | { |
4326 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
4327 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); |
4328 | t = fold_convert (TREE_TYPE (startvar), t); |
4329 | } |
4330 | t = force_gimple_operand_gsi (&gsi, t, |
4331 | DECL_P (startvar) |
4332 | && TREE_ADDRESSABLE (startvar), |
4333 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
4334 | assign_stmt = gimple_build_assign (startvar, t); |
4335 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
4336 | if (cond_var) |
4337 | { |
4338 | tree itype = TREE_TYPE (cond_var); |
4339 | /* For lastprivate(conditional:) itervar, we need some iteration |
4340 | counter that starts at unsigned non-zero and increases. |
4341 | Prefer as few IVs as possible, so if we can use startvar |
4342 | itself, use that, or startvar + constant (those would be |
4343 | incremented with step), and as last resort use the s0 + 1 |
4344 | incremented by 1. */ |
4345 | if ((fd->ordered && fd->collapse == 1) |
4346 | || bias |
4347 | || POINTER_TYPE_P (type) |
4348 | || TREE_CODE (fd->loop.n1) != INTEGER_CST |
4349 | || fd->loop.cond_code != LT_EXPR) |
4350 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, istart0), |
4351 | build_int_cst (itype, 1)); |
4352 | else if (tree_int_cst_sgn (fd->loop.n1) == 1) |
4353 | t = fold_convert (itype, t); |
4354 | else |
4355 | { |
4356 | tree c = fold_convert (itype, fd->loop.n1); |
4357 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); |
4358 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); |
4359 | } |
4360 | t = force_gimple_operand_gsi (&gsi, t, false, |
4361 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
4362 | assign_stmt = gimple_build_assign (cond_var, t); |
4363 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
4364 | } |
4365 | |
4366 | t = iend0; |
4367 | if (fd->ordered && fd->collapse == 1) |
4368 | t = fold_build2 (MULT_EXPR, fd->iter_type, t, |
4369 | fold_convert (fd->iter_type, fd->loop.step)); |
4370 | else if (bias) |
4371 | t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias); |
4372 | if (fd->ordered && fd->collapse == 1) |
4373 | { |
4374 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
4375 | t = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (startvar), |
4376 | fd->loop.n1, fold_convert (sizetype, t)); |
4377 | else |
4378 | { |
4379 | t = fold_convert (TREE_TYPE (startvar), t); |
4380 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (startvar), |
4381 | fd->loop.n1, t); |
4382 | } |
4383 | } |
4384 | else |
4385 | { |
4386 | if (POINTER_TYPE_P (TREE_TYPE (startvar))) |
4387 | t = fold_convert (signed_type_for (TREE_TYPE (startvar)), t); |
4388 | t = fold_convert (TREE_TYPE (startvar), t); |
4389 | } |
4390 | iend = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4391 | false, GSI_CONTINUE_LINKING); |
4392 | if (endvar) |
4393 | { |
4394 | assign_stmt = gimple_build_assign (endvar, iend); |
4395 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
4396 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (iend))) |
4397 | assign_stmt = gimple_build_assign (fd->loop.v, iend); |
4398 | else |
4399 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, iend); |
4400 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
4401 | } |
4402 | /* Handle linear clause adjustments. */ |
4403 | tree itercnt = NULL_TREE; |
4404 | if (gimple_omp_for_kind (g: fd->for_stmt) == GF_OMP_FOR_KIND_FOR) |
4405 | for (tree c = gimple_omp_for_clauses (gs: fd->for_stmt); |
4406 | c; c = OMP_CLAUSE_CHAIN (c)) |
4407 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
4408 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) |
4409 | { |
4410 | tree d = OMP_CLAUSE_DECL (c); |
4411 | tree t = d, a, dest; |
4412 | if (omp_privatize_by_reference (decl: t)) |
4413 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); |
4414 | tree type = TREE_TYPE (t); |
4415 | if (POINTER_TYPE_P (type)) |
4416 | type = sizetype; |
4417 | dest = unshare_expr (t); |
4418 | tree v = create_tmp_var (TREE_TYPE (t), NULL); |
4419 | expand_omp_build_assign (gsi_p: &gsif, to: v, from: t); |
4420 | if (itercnt == NULL_TREE) |
4421 | { |
4422 | itercnt = startvar; |
4423 | tree n1 = fd->loop.n1; |
4424 | if (POINTER_TYPE_P (TREE_TYPE (itercnt))) |
4425 | { |
4426 | itercnt |
4427 | = fold_convert (signed_type_for (TREE_TYPE (itercnt)), |
4428 | itercnt); |
4429 | n1 = fold_convert (TREE_TYPE (itercnt), n1); |
4430 | } |
4431 | itercnt = fold_build2 (MINUS_EXPR, TREE_TYPE (itercnt), |
4432 | itercnt, n1); |
4433 | itercnt = fold_build2 (EXACT_DIV_EXPR, TREE_TYPE (itercnt), |
4434 | itercnt, fd->loop.step); |
4435 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, |
4436 | NULL_TREE, false, |
4437 | GSI_CONTINUE_LINKING); |
4438 | } |
4439 | a = fold_build2 (MULT_EXPR, type, |
4440 | fold_convert (type, itercnt), |
4441 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); |
4442 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR |
4443 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); |
4444 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4445 | false, GSI_CONTINUE_LINKING); |
4446 | expand_omp_build_assign (gsi_p: &gsi, to: dest, from: t, after: true); |
4447 | } |
4448 | if (fd->collapse > 1) |
4449 | expand_omp_for_init_vars (fd, gsi: &gsi, counts, NULL, inner_stmt, startvar); |
4450 | |
4451 | if (fd->ordered) |
4452 | { |
4453 | /* Until now, counts array contained number of iterations or |
4454 | variable containing it for ith loop. From now on, we usually need |
4455 | those counts only for collapsed loops, and only for the 2nd |
4456 | till the last collapsed one. Move those one element earlier, |
4457 | we'll use counts[fd->collapse - 1] for the first source/sink |
4458 | iteration counter and so on and counts[fd->ordered] |
4459 | as the array holding the current counter values for |
4460 | depend(source). For doacross(sink:omp_cur_iteration - 1) we need |
4461 | the counts from fd->collapse to fd->ordered - 1; make a copy of |
4462 | those to counts[fd->ordered + 2] and onwards. |
4463 | counts[fd->ordered + 1] can be a flag whether it is the first |
4464 | iteration with a new collapsed counter (used only if |
4465 | fd->ordered > fd->collapse). */ |
4466 | if (fd->ordered > fd->collapse) |
4467 | memcpy (dest: counts + fd->ordered + 2, src: counts + fd->collapse, |
4468 | n: (fd->ordered - fd->collapse) * sizeof (counts[0])); |
4469 | if (fd->collapse > 1) |
4470 | memmove (dest: counts, src: counts + 1, n: (fd->collapse - 1) * sizeof (counts[0])); |
4471 | if (broken_loop) |
4472 | { |
4473 | int i; |
4474 | for (i = fd->collapse; i < fd->ordered; i++) |
4475 | { |
4476 | tree type = TREE_TYPE (fd->loops[i].v); |
4477 | tree this_cond |
4478 | = fold_build2 (fd->loops[i].cond_code, boolean_type_node, |
4479 | fold_convert (type, fd->loops[i].n1), |
4480 | fold_convert (type, fd->loops[i].n2)); |
4481 | if (!integer_onep (this_cond)) |
4482 | break; |
4483 | } |
4484 | if (i < fd->ordered) |
4485 | { |
4486 | if (entry_bb->loop_father != l0_bb->loop_father) |
4487 | { |
4488 | remove_bb_from_loops (l0_bb); |
4489 | add_bb_to_loop (l0_bb, entry_bb->loop_father); |
4490 | gcc_assert (single_succ (l0_bb) == l1_bb); |
4491 | } |
4492 | cont_bb |
4493 | = create_empty_bb (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb); |
4494 | add_bb_to_loop (cont_bb, l0_bb->loop_father); |
4495 | gimple_stmt_iterator gsi = gsi_after_labels (bb: cont_bb); |
4496 | gimple *g = gimple_build_omp_continue (fd->loop.v, fd->loop.v); |
4497 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
4498 | make_edge (cont_bb, l3_bb, EDGE_FALLTHRU); |
4499 | make_edge (cont_bb, l1_bb, 0); |
4500 | l2_bb = create_empty_bb (cont_bb); |
4501 | broken_loop = false; |
4502 | } |
4503 | } |
4504 | expand_omp_ordered_source_sink (region, fd, counts, cont_bb); |
4505 | cont_bb = expand_omp_for_ordered_loops (fd, counts, cont_bb, body_bb: l1_bb, |
4506 | l0_bb, ordered_lastprivate); |
4507 | if (counts[fd->collapse - 1]) |
4508 | { |
4509 | gcc_assert (fd->collapse == 1); |
4510 | gsi = gsi_last_bb (bb: l0_bb); |
4511 | expand_omp_build_assign (gsi_p: &gsi, to: counts[fd->collapse - 1], |
4512 | from: istart0, after: true); |
4513 | if (cont_bb) |
4514 | { |
4515 | gsi = gsi_last_bb (bb: cont_bb); |
4516 | t = fold_build2 (PLUS_EXPR, fd->iter_type, |
4517 | counts[fd->collapse - 1], |
4518 | build_int_cst (fd->iter_type, 1)); |
4519 | expand_omp_build_assign (gsi_p: &gsi, to: counts[fd->collapse - 1], from: t); |
4520 | tree aref = build4 (ARRAY_REF, fd->iter_type, |
4521 | counts[fd->ordered], size_zero_node, |
4522 | NULL_TREE, NULL_TREE); |
4523 | expand_omp_build_assign (gsi_p: &gsi, to: aref, from: counts[fd->collapse - 1]); |
4524 | } |
4525 | t = counts[fd->collapse - 1]; |
4526 | } |
4527 | else if (fd->collapse > 1) |
4528 | t = fd->loop.v; |
4529 | else |
4530 | { |
4531 | t = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), |
4532 | fd->loops[0].v, fd->loops[0].n1); |
4533 | t = fold_convert (fd->iter_type, t); |
4534 | } |
4535 | gsi = gsi_last_bb (bb: l0_bb); |
4536 | tree aref = build4 (ARRAY_REF, fd->iter_type, counts[fd->ordered], |
4537 | size_zero_node, NULL_TREE, NULL_TREE); |
4538 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4539 | false, GSI_CONTINUE_LINKING); |
4540 | expand_omp_build_assign (gsi_p: &gsi, to: aref, from: t, after: true); |
4541 | } |
4542 | |
4543 | if (!broken_loop) |
4544 | { |
4545 | /* Code to control the increment and predicate for the sequential |
4546 | loop goes in the CONT_BB. */ |
4547 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
4548 | gomp_continue *cont_stmt = as_a <gomp_continue *> (p: gsi_stmt (i: gsi)); |
4549 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); |
4550 | vmain = gimple_omp_continue_control_use (cont_stmt); |
4551 | vback = gimple_omp_continue_control_def (cont_stmt); |
4552 | |
4553 | if (cond_var) |
4554 | { |
4555 | tree itype = TREE_TYPE (cond_var); |
4556 | tree t2; |
4557 | if ((fd->ordered && fd->collapse == 1) |
4558 | || bias |
4559 | || POINTER_TYPE_P (type) |
4560 | || TREE_CODE (fd->loop.n1) != INTEGER_CST |
4561 | || fd->loop.cond_code != LT_EXPR) |
4562 | t2 = build_int_cst (itype, 1); |
4563 | else |
4564 | t2 = fold_convert (itype, fd->loop.step); |
4565 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); |
4566 | t2 = force_gimple_operand_gsi (&gsi, t2, false, |
4567 | NULL_TREE, true, GSI_SAME_STMT); |
4568 | assign_stmt = gimple_build_assign (cond_var, t2); |
4569 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
4570 | } |
4571 | |
4572 | if (!gimple_omp_for_combined_p (g: fd->for_stmt)) |
4573 | { |
4574 | if (POINTER_TYPE_P (type)) |
4575 | t = fold_build_pointer_plus (vmain, fd->loop.step); |
4576 | else |
4577 | t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step); |
4578 | t = force_gimple_operand_gsi (&gsi, t, |
4579 | DECL_P (vback) |
4580 | && TREE_ADDRESSABLE (vback), |
4581 | NULL_TREE, true, GSI_SAME_STMT); |
4582 | assign_stmt = gimple_build_assign (vback, t); |
4583 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
4584 | |
4585 | if (fd->ordered && counts[fd->collapse - 1] == NULL_TREE) |
4586 | { |
4587 | tree tem; |
4588 | if (fd->collapse > 1) |
4589 | tem = fd->loop.v; |
4590 | else |
4591 | { |
4592 | tem = fold_build2 (MINUS_EXPR, TREE_TYPE (fd->loops[0].v), |
4593 | fd->loops[0].v, fd->loops[0].n1); |
4594 | tem = fold_convert (fd->iter_type, tem); |
4595 | } |
4596 | tree aref = build4 (ARRAY_REF, fd->iter_type, |
4597 | counts[fd->ordered], size_zero_node, |
4598 | NULL_TREE, NULL_TREE); |
4599 | tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, |
4600 | true, GSI_SAME_STMT); |
4601 | expand_omp_build_assign (gsi_p: &gsi, to: aref, from: tem); |
4602 | } |
4603 | |
4604 | t = build2 (fd->loop.cond_code, boolean_type_node, |
4605 | DECL_P (vback) && TREE_ADDRESSABLE (vback) ? t : vback, |
4606 | iend); |
4607 | gcond *cond_stmt = gimple_build_cond_empty (cond: t); |
4608 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); |
4609 | } |
4610 | |
4611 | /* Remove GIMPLE_OMP_CONTINUE. */ |
4612 | gsi_remove (&gsi, true); |
4613 | |
4614 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (g: fd->for_stmt)) |
4615 | collapse_bb = extract_omp_for_update_vars (fd, NULL, cont_bb, body_bb: l1_bb); |
4616 | |
4617 | /* Emit code to get the next parallel iteration in L2_BB. */ |
4618 | gsi = gsi_start_bb (bb: l2_bb); |
4619 | |
4620 | t = build_call_expr (builtin_decl_explicit (fncode: next_fn), 2, |
4621 | build_fold_addr_expr (istart0), |
4622 | build_fold_addr_expr (iend0)); |
4623 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
4624 | false, GSI_CONTINUE_LINKING); |
4625 | if (TREE_TYPE (t) != boolean_type_node) |
4626 | t = fold_build2 (NE_EXPR, boolean_type_node, |
4627 | t, build_int_cst (TREE_TYPE (t), 0)); |
4628 | gcond *cond_stmt = gimple_build_cond_empty (cond: t); |
4629 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); |
4630 | } |
4631 | |
4632 | /* Add the loop cleanup function. */ |
4633 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
4634 | if (gimple_omp_return_nowait_p (g: gsi_stmt (i: gsi))) |
4635 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END_NOWAIT); |
4636 | else if (gimple_omp_return_lhs (g: gsi_stmt (i: gsi))) |
4637 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END_CANCEL); |
4638 | else |
4639 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END); |
4640 | gcall *call_stmt = gimple_build_call (t, 0); |
4641 | if (fd->ordered) |
4642 | { |
4643 | tree arr = counts[fd->ordered]; |
4644 | tree clobber = build_clobber (TREE_TYPE (arr)); |
4645 | gsi_insert_after (&gsi, gimple_build_assign (arr, clobber), |
4646 | GSI_SAME_STMT); |
4647 | } |
4648 | if (gimple_omp_return_lhs (g: gsi_stmt (i: gsi))) |
4649 | { |
4650 | gimple_call_set_lhs (gs: call_stmt, lhs: gimple_omp_return_lhs (g: gsi_stmt (i: gsi))); |
4651 | if (fd->have_reductemp) |
4652 | { |
4653 | gimple *g = gimple_build_assign (reductions, NOP_EXPR, |
4654 | gimple_call_lhs (gs: call_stmt)); |
4655 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); |
4656 | } |
4657 | } |
4658 | gsi_insert_after (&gsi, call_stmt, GSI_SAME_STMT); |
4659 | gsi_remove (&gsi, true); |
4660 | |
4661 | /* Connect the new blocks. */ |
4662 | find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE; |
4663 | find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE; |
4664 | |
4665 | if (!broken_loop) |
4666 | { |
4667 | gimple_seq phis; |
4668 | |
4669 | e = find_edge (cont_bb, l3_bb); |
4670 | ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE); |
4671 | |
4672 | phis = phi_nodes (bb: l3_bb); |
4673 | for (gsi = gsi_start (seq&: phis); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
4674 | { |
4675 | gimple *phi = gsi_stmt (i: gsi); |
4676 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne), |
4677 | PHI_ARG_DEF_FROM_EDGE (phi, e)); |
4678 | } |
4679 | remove_edge (e); |
4680 | |
4681 | make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE); |
4682 | e = find_edge (cont_bb, l1_bb); |
4683 | if (e == NULL) |
4684 | { |
4685 | e = BRANCH_EDGE (cont_bb); |
4686 | gcc_assert (single_succ (e->dest) == l1_bb); |
4687 | } |
4688 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
4689 | { |
4690 | remove_edge (e); |
4691 | e = NULL; |
4692 | } |
4693 | else if (fd->collapse > 1) |
4694 | { |
4695 | remove_edge (e); |
4696 | e = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); |
4697 | } |
4698 | else |
4699 | e->flags = EDGE_TRUE_VALUE; |
4700 | if (e) |
4701 | { |
4702 | e->probability = profile_probability::guessed_always ().apply_scale (num: 7, den: 8); |
4703 | find_edge (cont_bb, l2_bb)->probability = e->probability.invert (); |
4704 | } |
4705 | else |
4706 | { |
4707 | e = find_edge (cont_bb, l2_bb); |
4708 | e->flags = EDGE_FALLTHRU; |
4709 | } |
4710 | make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE); |
4711 | |
4712 | if (gimple_in_ssa_p (cfun)) |
4713 | { |
4714 | /* Add phis to the outer loop that connect to the phis in the inner, |
4715 | original loop, and move the loop entry value of the inner phi to |
4716 | the loop entry value of the outer phi. */ |
4717 | gphi_iterator psi; |
4718 | for (psi = gsi_start_phis (l3_bb); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
4719 | { |
4720 | location_t locus; |
4721 | gphi *nphi; |
4722 | gphi *exit_phi = psi.phi (); |
4723 | |
4724 | if (virtual_operand_p (op: gimple_phi_result (gs: exit_phi))) |
4725 | continue; |
4726 | |
4727 | edge l2_to_l3 = find_edge (l2_bb, l3_bb); |
4728 | tree exit_res = PHI_ARG_DEF_FROM_EDGE (exit_phi, l2_to_l3); |
4729 | |
4730 | basic_block latch = BRANCH_EDGE (cont_bb)->dest; |
4731 | edge latch_to_l1 = find_edge (latch, l1_bb); |
4732 | gphi *inner_phi |
4733 | = find_phi_with_arg_on_edge (exit_res, latch_to_l1); |
4734 | |
4735 | tree t = gimple_phi_result (gs: exit_phi); |
4736 | tree new_res = copy_ssa_name (var: t, NULL); |
4737 | nphi = create_phi_node (new_res, l0_bb); |
4738 | |
4739 | edge l0_to_l1 = find_edge (l0_bb, l1_bb); |
4740 | t = PHI_ARG_DEF_FROM_EDGE (inner_phi, l0_to_l1); |
4741 | locus = gimple_phi_arg_location_from_edge (phi: inner_phi, e: l0_to_l1); |
4742 | edge entry_to_l0 = find_edge (entry_bb, l0_bb); |
4743 | add_phi_arg (nphi, t, entry_to_l0, locus); |
4744 | |
4745 | edge l2_to_l0 = find_edge (l2_bb, l0_bb); |
4746 | add_phi_arg (nphi, exit_res, l2_to_l0, UNKNOWN_LOCATION); |
4747 | |
4748 | add_phi_arg (inner_phi, new_res, l0_to_l1, UNKNOWN_LOCATION); |
4749 | } |
4750 | } |
4751 | |
4752 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, |
4753 | recompute_dominator (CDI_DOMINATORS, l2_bb)); |
4754 | set_immediate_dominator (CDI_DOMINATORS, l3_bb, |
4755 | recompute_dominator (CDI_DOMINATORS, l3_bb)); |
4756 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, |
4757 | recompute_dominator (CDI_DOMINATORS, l0_bb)); |
4758 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, |
4759 | recompute_dominator (CDI_DOMINATORS, l1_bb)); |
4760 | |
4761 | /* We enter expand_omp_for_generic with a loop. This original loop may |
4762 | have its own loop struct, or it may be part of an outer loop struct |
4763 | (which may be the fake loop). */ |
4764 | class loop *outer_loop = entry_bb->loop_father; |
4765 | bool orig_loop_has_loop_struct = l1_bb->loop_father != outer_loop; |
4766 | |
4767 | add_bb_to_loop (l2_bb, outer_loop); |
4768 | |
4769 | /* We've added a new loop around the original loop. Allocate the |
4770 | corresponding loop struct. */ |
4771 | class loop *new_loop = alloc_loop (); |
4772 | new_loop->header = l0_bb; |
4773 | new_loop->latch = l2_bb; |
4774 | add_loop (new_loop, outer_loop); |
4775 | |
4776 | /* Allocate a loop structure for the original loop unless we already |
4777 | had one. */ |
4778 | if (!orig_loop_has_loop_struct |
4779 | && !gimple_omp_for_combined_p (g: fd->for_stmt)) |
4780 | { |
4781 | class loop *orig_loop = alloc_loop (); |
4782 | orig_loop->header = l1_bb; |
4783 | /* The loop may have multiple latches. */ |
4784 | add_loop (orig_loop, new_loop); |
4785 | } |
4786 | } |
4787 | } |
4788 | |
4789 | /* Helper function for expand_omp_for_static_nochunk. If PTR is NULL, |
4790 | compute needed allocation size. If !ALLOC of team allocations, |
4791 | if ALLOC of thread allocation. SZ is the initial needed size for |
4792 | other purposes, ALLOC_ALIGN guaranteed alignment of allocation in bytes, |
4793 | CNT number of elements of each array, for !ALLOC this is |
4794 | omp_get_num_threads (), for ALLOC number of iterations handled by the |
4795 | current thread. If PTR is non-NULL, it is the start of the allocation |
4796 | and this routine shall assign to OMP_CLAUSE_DECL (c) of those _scantemp_ |
4797 | clauses pointers to the corresponding arrays. */ |
4798 | |
4799 | static tree |
4800 | expand_omp_scantemp_alloc (tree clauses, tree ptr, unsigned HOST_WIDE_INT sz, |
4801 | unsigned HOST_WIDE_INT alloc_align, tree cnt, |
4802 | gimple_stmt_iterator *gsi, bool alloc) |
4803 | { |
4804 | tree eltsz = NULL_TREE; |
4805 | unsigned HOST_WIDE_INT preval = 0; |
4806 | if (ptr && sz) |
4807 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), |
4808 | ptr, size_int (sz)); |
4809 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) |
4810 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ |
4811 | && !OMP_CLAUSE__SCANTEMP__CONTROL (c) |
4812 | && (!OMP_CLAUSE__SCANTEMP__ALLOC (c)) != alloc) |
4813 | { |
4814 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); |
4815 | unsigned HOST_WIDE_INT al = TYPE_ALIGN_UNIT (pointee_type); |
4816 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) |
4817 | { |
4818 | unsigned HOST_WIDE_INT szl |
4819 | = tree_to_uhwi (TYPE_SIZE_UNIT (pointee_type)); |
4820 | szl = least_bit_hwi (x: szl); |
4821 | if (szl) |
4822 | al = MIN (al, szl); |
4823 | } |
4824 | if (ptr == NULL_TREE) |
4825 | { |
4826 | if (eltsz == NULL_TREE) |
4827 | eltsz = TYPE_SIZE_UNIT (pointee_type); |
4828 | else |
4829 | eltsz = size_binop (PLUS_EXPR, eltsz, |
4830 | TYPE_SIZE_UNIT (pointee_type)); |
4831 | } |
4832 | if (preval == 0 && al <= alloc_align) |
4833 | { |
4834 | unsigned HOST_WIDE_INT diff = ROUND_UP (sz, al) - sz; |
4835 | sz += diff; |
4836 | if (diff && ptr) |
4837 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), |
4838 | ptr, size_int (diff)); |
4839 | } |
4840 | else if (al > preval) |
4841 | { |
4842 | if (ptr) |
4843 | { |
4844 | ptr = fold_convert (pointer_sized_int_node, ptr); |
4845 | ptr = fold_build2 (PLUS_EXPR, pointer_sized_int_node, ptr, |
4846 | build_int_cst (pointer_sized_int_node, |
4847 | al - 1)); |
4848 | ptr = fold_build2 (BIT_AND_EXPR, pointer_sized_int_node, ptr, |
4849 | build_int_cst (pointer_sized_int_node, |
4850 | -(HOST_WIDE_INT) al)); |
4851 | ptr = fold_convert (ptr_type_node, ptr); |
4852 | } |
4853 | else |
4854 | sz += al - 1; |
4855 | } |
4856 | if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (pointee_type))) |
4857 | preval = al; |
4858 | else |
4859 | preval = 1; |
4860 | if (ptr) |
4861 | { |
4862 | expand_omp_build_assign (gsi_p: gsi, OMP_CLAUSE_DECL (c), from: ptr, after: false); |
4863 | ptr = OMP_CLAUSE_DECL (c); |
4864 | ptr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (ptr), ptr, |
4865 | size_binop (MULT_EXPR, cnt, |
4866 | TYPE_SIZE_UNIT (pointee_type))); |
4867 | } |
4868 | } |
4869 | |
4870 | if (ptr == NULL_TREE) |
4871 | { |
4872 | eltsz = size_binop (MULT_EXPR, eltsz, cnt); |
4873 | if (sz) |
4874 | eltsz = size_binop (PLUS_EXPR, eltsz, size_int (sz)); |
4875 | return eltsz; |
4876 | } |
4877 | else |
4878 | return ptr; |
4879 | } |
4880 | |
4881 | /* Return the last _looptemp_ clause if one has been created for |
4882 | lastprivate on distribute parallel for{, simd} or taskloop. |
4883 | FD is the loop data and INNERC should be the second _looptemp_ |
4884 | clause (the one holding the end of the range). |
4885 | This is followed by collapse - 1 _looptemp_ clauses for the |
4886 | counts[1] and up, and for triangular loops followed by 4 |
4887 | further _looptemp_ clauses (one for counts[0], one first_inner_iterations, |
4888 | one factor and one adjn1). After this there is optionally one |
4889 | _looptemp_ clause that this function returns. */ |
4890 | |
4891 | static tree |
4892 | find_lastprivate_looptemp (struct omp_for_data *fd, tree innerc) |
4893 | { |
4894 | gcc_assert (innerc); |
4895 | int count = fd->collapse - 1; |
4896 | if (fd->non_rect |
4897 | && fd->last_nonrect == fd->first_nonrect + 1 |
4898 | && !TYPE_UNSIGNED (TREE_TYPE (fd->loops[fd->last_nonrect].v))) |
4899 | count += 4; |
4900 | for (int i = 0; i < count; i++) |
4901 | { |
4902 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
4903 | kind: OMP_CLAUSE__LOOPTEMP_); |
4904 | gcc_assert (innerc); |
4905 | } |
4906 | return omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
4907 | kind: OMP_CLAUSE__LOOPTEMP_); |
4908 | } |
4909 | |
4910 | /* A subroutine of expand_omp_for. Generate code for a parallel |
4911 | loop with static schedule and no specified chunk size. Given |
4912 | parameters: |
4913 | |
4914 | for (V = N1; V cond N2; V += STEP) BODY; |
4915 | |
4916 | where COND is "<" or ">", we generate pseudocode |
4917 | |
4918 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; |
4919 | if (cond is <) |
4920 | adj = STEP - 1; |
4921 | else |
4922 | adj = STEP + 1; |
4923 | if ((__typeof (V)) -1 > 0 && cond is >) |
4924 | n = -(adj + N2 - N1) / -STEP; |
4925 | else |
4926 | n = (adj + N2 - N1) / STEP; |
4927 | q = n / nthreads; |
4928 | tt = n % nthreads; |
4929 | if (threadid < tt) goto L3; else goto L4; |
4930 | L3: |
4931 | tt = 0; |
4932 | q = q + 1; |
4933 | L4: |
4934 | s0 = q * threadid + tt; |
4935 | e0 = s0 + q; |
4936 | V = s0 * STEP + N1; |
4937 | if (s0 >= e0) goto L2; else goto L0; |
4938 | L0: |
4939 | e = e0 * STEP + N1; |
4940 | L1: |
4941 | BODY; |
4942 | V += STEP; |
4943 | if (V cond e) goto L1; |
4944 | L2: |
4945 | */ |
4946 | |
4947 | static void |
4948 | expand_omp_for_static_nochunk (struct omp_region *region, |
4949 | struct omp_for_data *fd, |
4950 | gimple *inner_stmt) |
4951 | { |
4952 | tree n, q, s0, e0, e, t, tt, nthreads = NULL_TREE, threadid; |
4953 | tree type, itype, vmain, vback; |
4954 | basic_block entry_bb, second_bb, third_bb, exit_bb, seq_start_bb; |
4955 | basic_block body_bb, cont_bb, collapse_bb = NULL; |
4956 | basic_block fin_bb, fourth_bb = NULL, fifth_bb = NULL, sixth_bb = NULL; |
4957 | basic_block exit1_bb = NULL, exit2_bb = NULL, exit3_bb = NULL; |
4958 | gimple_stmt_iterator gsi, gsip; |
4959 | edge ep; |
4960 | bool broken_loop = region->cont == NULL; |
4961 | tree *counts = NULL; |
4962 | tree n1, n2, step; |
4963 | tree reductions = NULL_TREE; |
4964 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; |
4965 | |
4966 | itype = type = TREE_TYPE (fd->loop.v); |
4967 | if (POINTER_TYPE_P (type)) |
4968 | itype = signed_type_for (type); |
4969 | |
4970 | entry_bb = region->entry; |
4971 | cont_bb = region->cont; |
4972 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
4973 | fin_bb = BRANCH_EDGE (entry_bb)->dest; |
4974 | gcc_assert (broken_loop |
4975 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); |
4976 | seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb)); |
4977 | body_bb = single_succ (bb: seq_start_bb); |
4978 | if (!broken_loop) |
4979 | { |
4980 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb |
4981 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); |
4982 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); |
4983 | } |
4984 | exit_bb = region->exit; |
4985 | |
4986 | /* Iteration space partitioning goes in ENTRY_BB. */ |
4987 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
4988 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
4989 | gsip = gsi; |
4990 | gsi_prev (i: &gsip); |
4991 | |
4992 | if (fd->collapse > 1) |
4993 | { |
4994 | int first_zero_iter = -1, dummy = -1; |
4995 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; |
4996 | |
4997 | counts = XALLOCAVEC (tree, fd->collapse); |
4998 | expand_omp_for_init_counts (fd, gsi: &gsi, entry_bb, counts, |
4999 | zero_iter1_bb&: fin_bb, first_zero_iter1&: first_zero_iter, |
5000 | zero_iter2_bb&: dummy_bb, first_zero_iter2&: dummy, l2_dom_bb); |
5001 | t = NULL_TREE; |
5002 | } |
5003 | else if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
5004 | t = integer_one_node; |
5005 | else |
5006 | t = fold_binary (fd->loop.cond_code, boolean_type_node, |
5007 | fold_convert (type, fd->loop.n1), |
5008 | fold_convert (type, fd->loop.n2)); |
5009 | if (fd->collapse == 1 |
5010 | && TYPE_UNSIGNED (type) |
5011 | && (t == NULL_TREE || !integer_onep (t))) |
5012 | { |
5013 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); |
5014 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, |
5015 | true, GSI_SAME_STMT); |
5016 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); |
5017 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, |
5018 | true, GSI_SAME_STMT); |
5019 | gcond *cond_stmt = expand_omp_build_cond (gsi_p: &gsi, code: fd->loop.cond_code, |
5020 | lhs: n1, rhs: n2); |
5021 | ep = split_block (entry_bb, cond_stmt); |
5022 | ep->flags = EDGE_TRUE_VALUE; |
5023 | entry_bb = ep->dest; |
5024 | ep->probability = profile_probability::very_likely (); |
5025 | ep = make_edge (ep->src, fin_bb, EDGE_FALSE_VALUE); |
5026 | ep->probability = profile_probability::very_unlikely (); |
5027 | if (gimple_in_ssa_p (cfun)) |
5028 | { |
5029 | int dest_idx = find_edge (entry_bb, fin_bb)->dest_idx; |
5030 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); |
5031 | !gsi_end_p (i: gpi); gsi_next (i: &gpi)) |
5032 | { |
5033 | gphi *phi = gpi.phi (); |
5034 | add_phi_arg (phi, gimple_phi_arg_def (gs: phi, index: dest_idx), |
5035 | ep, UNKNOWN_LOCATION); |
5036 | } |
5037 | } |
5038 | gsi = gsi_last_bb (bb: entry_bb); |
5039 | } |
5040 | |
5041 | if (fd->lastprivate_conditional) |
5042 | { |
5043 | tree clauses = gimple_omp_for_clauses (gs: fd->for_stmt); |
5044 | tree c = omp_find_clause (clauses, kind: OMP_CLAUSE__CONDTEMP_); |
5045 | if (fd->have_pointer_condtemp) |
5046 | condtemp = OMP_CLAUSE_DECL (c); |
5047 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), kind: OMP_CLAUSE__CONDTEMP_); |
5048 | cond_var = OMP_CLAUSE_DECL (c); |
5049 | } |
5050 | if (fd->have_reductemp |
5051 | /* For scan, we don't want to reinitialize condtemp before the |
5052 | second loop. */ |
5053 | || (fd->have_pointer_condtemp && !fd->have_scantemp) |
5054 | || fd->have_nonctrl_scantemp) |
5055 | { |
5056 | tree t1 = build_int_cst (long_integer_type_node, 0); |
5057 | tree t2 = build_int_cst (long_integer_type_node, 1); |
5058 | tree t3 = build_int_cstu (long_integer_type_node, |
5059 | (HOST_WIDE_INT_1U << 31) + 1); |
5060 | tree clauses = gimple_omp_for_clauses (gs: fd->for_stmt); |
5061 | gimple_stmt_iterator gsi2 = gsi_none (); |
5062 | gimple *g = NULL; |
5063 | tree mem = null_pointer_node, memv = NULL_TREE; |
5064 | unsigned HOST_WIDE_INT condtemp_sz = 0; |
5065 | unsigned HOST_WIDE_INT alloc_align = 0; |
5066 | if (fd->have_reductemp) |
5067 | { |
5068 | gcc_assert (!fd->have_nonctrl_scantemp); |
5069 | tree c = omp_find_clause (clauses, kind: OMP_CLAUSE__REDUCTEMP_); |
5070 | reductions = OMP_CLAUSE_DECL (c); |
5071 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); |
5072 | g = SSA_NAME_DEF_STMT (reductions); |
5073 | reductions = gimple_assign_rhs1 (gs: g); |
5074 | OMP_CLAUSE_DECL (c) = reductions; |
5075 | gsi2 = gsi_for_stmt (g); |
5076 | } |
5077 | else |
5078 | { |
5079 | if (gsi_end_p (i: gsip)) |
5080 | gsi2 = gsi_after_labels (bb: region->entry); |
5081 | else |
5082 | gsi2 = gsip; |
5083 | reductions = null_pointer_node; |
5084 | } |
5085 | if (fd->have_pointer_condtemp || fd->have_nonctrl_scantemp) |
5086 | { |
5087 | tree type; |
5088 | if (fd->have_pointer_condtemp) |
5089 | type = TREE_TYPE (condtemp); |
5090 | else |
5091 | type = ptr_type_node; |
5092 | memv = create_tmp_var (type); |
5093 | TREE_ADDRESSABLE (memv) = 1; |
5094 | unsigned HOST_WIDE_INT sz = 0; |
5095 | tree size = NULL_TREE; |
5096 | if (fd->have_pointer_condtemp) |
5097 | { |
5098 | sz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); |
5099 | sz *= fd->lastprivate_conditional; |
5100 | condtemp_sz = sz; |
5101 | } |
5102 | if (fd->have_nonctrl_scantemp) |
5103 | { |
5104 | nthreads = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_NUM_THREADS); |
5105 | gimple *g = gimple_build_call (nthreads, 0); |
5106 | nthreads = create_tmp_var (integer_type_node); |
5107 | gimple_call_set_lhs (gs: g, lhs: nthreads); |
5108 | gsi_insert_before (&gsi2, g, GSI_SAME_STMT); |
5109 | nthreads = fold_convert (sizetype, nthreads); |
5110 | alloc_align = TYPE_ALIGN_UNIT (long_long_integer_type_node); |
5111 | size = expand_omp_scantemp_alloc (clauses, NULL_TREE, sz, |
5112 | alloc_align, cnt: nthreads, NULL, |
5113 | alloc: false); |
5114 | size = fold_convert (type, size); |
5115 | } |
5116 | else |
5117 | size = build_int_cst (type, sz); |
5118 | expand_omp_build_assign (gsi_p: &gsi2, to: memv, from: size, after: false); |
5119 | mem = build_fold_addr_expr (memv); |
5120 | } |
5121 | tree t |
5122 | = build_call_expr (builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_START), |
5123 | 9, t1, t2, t2, t3, t1, null_pointer_node, |
5124 | null_pointer_node, reductions, mem); |
5125 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
5126 | true, GSI_SAME_STMT); |
5127 | if (fd->have_pointer_condtemp) |
5128 | expand_omp_build_assign (gsi_p: &gsi2, to: condtemp, from: memv, after: false); |
5129 | if (fd->have_nonctrl_scantemp) |
5130 | { |
5131 | tree ptr = fd->have_pointer_condtemp ? condtemp : memv; |
5132 | expand_omp_scantemp_alloc (clauses, ptr, sz: condtemp_sz, |
5133 | alloc_align, cnt: nthreads, gsi: &gsi2, alloc: false); |
5134 | } |
5135 | if (fd->have_reductemp) |
5136 | { |
5137 | gsi_remove (&gsi2, true); |
5138 | release_ssa_name (name: gimple_assign_lhs (gs: g)); |
5139 | } |
5140 | } |
5141 | switch (gimple_omp_for_kind (g: fd->for_stmt)) |
5142 | { |
5143 | case GF_OMP_FOR_KIND_FOR: |
5144 | nthreads = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_NUM_THREADS); |
5145 | threadid = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_THREAD_NUM); |
5146 | break; |
5147 | case GF_OMP_FOR_KIND_DISTRIBUTE: |
5148 | nthreads = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_NUM_TEAMS); |
5149 | threadid = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_TEAM_NUM); |
5150 | break; |
5151 | default: |
5152 | gcc_unreachable (); |
5153 | } |
5154 | nthreads = build_call_expr (nthreads, 0); |
5155 | nthreads = fold_convert (itype, nthreads); |
5156 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, |
5157 | true, GSI_SAME_STMT); |
5158 | threadid = build_call_expr (threadid, 0); |
5159 | threadid = fold_convert (itype, threadid); |
5160 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, |
5161 | true, GSI_SAME_STMT); |
5162 | |
5163 | n1 = fd->loop.n1; |
5164 | n2 = fd->loop.n2; |
5165 | step = fd->loop.step; |
5166 | if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
5167 | { |
5168 | tree innerc = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
5169 | kind: OMP_CLAUSE__LOOPTEMP_); |
5170 | gcc_assert (innerc); |
5171 | n1 = OMP_CLAUSE_DECL (innerc); |
5172 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
5173 | kind: OMP_CLAUSE__LOOPTEMP_); |
5174 | gcc_assert (innerc); |
5175 | n2 = OMP_CLAUSE_DECL (innerc); |
5176 | } |
5177 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), |
5178 | true, NULL_TREE, true, GSI_SAME_STMT); |
5179 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), |
5180 | true, NULL_TREE, true, GSI_SAME_STMT); |
5181 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), |
5182 | true, NULL_TREE, true, GSI_SAME_STMT); |
5183 | |
5184 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); |
5185 | t = fold_build2 (PLUS_EXPR, itype, step, t); |
5186 | t = fold_build2 (PLUS_EXPR, itype, t, n2); |
5187 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); |
5188 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) |
5189 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
5190 | fold_build1 (NEGATE_EXPR, itype, t), |
5191 | fold_build1 (NEGATE_EXPR, itype, step)); |
5192 | else |
5193 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
5194 | t = fold_convert (itype, t); |
5195 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
5196 | |
5197 | q = create_tmp_reg (itype, "q" ); |
5198 | t = fold_build2 (TRUNC_DIV_EXPR, itype, n, nthreads); |
5199 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); |
5200 | gsi_insert_before (&gsi, gimple_build_assign (q, t), GSI_SAME_STMT); |
5201 | |
5202 | tt = create_tmp_reg (itype, "tt" ); |
5203 | t = fold_build2 (TRUNC_MOD_EXPR, itype, n, nthreads); |
5204 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT); |
5205 | gsi_insert_before (&gsi, gimple_build_assign (tt, t), GSI_SAME_STMT); |
5206 | |
5207 | t = build2 (LT_EXPR, boolean_type_node, threadid, tt); |
5208 | gcond *cond_stmt = gimple_build_cond_empty (cond: t); |
5209 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); |
5210 | |
5211 | second_bb = split_block (entry_bb, cond_stmt)->dest; |
5212 | gsi = gsi_last_nondebug_bb (bb: second_bb); |
5213 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
5214 | |
5215 | gsi_insert_before (&gsi, gimple_build_assign (tt, build_int_cst (itype, 0)), |
5216 | GSI_SAME_STMT); |
5217 | gassign *assign_stmt |
5218 | = gimple_build_assign (q, PLUS_EXPR, q, build_int_cst (itype, 1)); |
5219 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
5220 | |
5221 | third_bb = split_block (second_bb, assign_stmt)->dest; |
5222 | gsi = gsi_last_nondebug_bb (bb: third_bb); |
5223 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
5224 | |
5225 | if (fd->have_nonctrl_scantemp) |
5226 | { |
5227 | tree clauses = gimple_omp_for_clauses (gs: fd->for_stmt); |
5228 | tree controlp = NULL_TREE, controlb = NULL_TREE; |
5229 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) |
5230 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ |
5231 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) |
5232 | { |
5233 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) |
5234 | controlb = OMP_CLAUSE_DECL (c); |
5235 | else |
5236 | controlp = OMP_CLAUSE_DECL (c); |
5237 | if (controlb && controlp) |
5238 | break; |
5239 | } |
5240 | gcc_assert (controlp && controlb); |
5241 | tree cnt = create_tmp_var (sizetype); |
5242 | gimple *g = gimple_build_assign (cnt, NOP_EXPR, q); |
5243 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5244 | unsigned HOST_WIDE_INT alloc_align = TYPE_ALIGN_UNIT (ptr_type_node); |
5245 | tree sz = expand_omp_scantemp_alloc (clauses, NULL_TREE, sz: 0, |
5246 | alloc_align, cnt, NULL, alloc: true); |
5247 | tree size = create_tmp_var (sizetype); |
5248 | expand_omp_build_assign (gsi_p: &gsi, to: size, from: sz, after: false); |
5249 | tree cmp = fold_build2 (GT_EXPR, boolean_type_node, |
5250 | size, size_int (16384)); |
5251 | expand_omp_build_assign (gsi_p: &gsi, to: controlb, from: cmp); |
5252 | g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, |
5253 | NULL_TREE, NULL_TREE); |
5254 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5255 | fourth_bb = split_block (third_bb, g)->dest; |
5256 | gsi = gsi_last_nondebug_bb (bb: fourth_bb); |
5257 | /* FIXME: Once we have allocators, this should use allocator. */ |
5258 | g = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_MALLOC), 1, size); |
5259 | gimple_call_set_lhs (gs: g, lhs: controlp); |
5260 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5261 | expand_omp_scantemp_alloc (clauses, ptr: controlp, sz: 0, alloc_align, cnt, |
5262 | gsi: &gsi, alloc: true); |
5263 | gsi_prev (i: &gsi); |
5264 | g = gsi_stmt (i: gsi); |
5265 | fifth_bb = split_block (fourth_bb, g)->dest; |
5266 | gsi = gsi_last_nondebug_bb (bb: fifth_bb); |
5267 | |
5268 | g = gimple_build_call (builtin_decl_implicit (fncode: BUILT_IN_STACK_SAVE), 0); |
5269 | gimple_call_set_lhs (gs: g, lhs: controlp); |
5270 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5271 | tree alloca_decl = builtin_decl_explicit (fncode: BUILT_IN_ALLOCA_WITH_ALIGN); |
5272 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) |
5273 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ |
5274 | && OMP_CLAUSE__SCANTEMP__ALLOC (c)) |
5275 | { |
5276 | tree tmp = create_tmp_var (sizetype); |
5277 | tree pointee_type = TREE_TYPE (TREE_TYPE (OMP_CLAUSE_DECL (c))); |
5278 | g = gimple_build_assign (tmp, MULT_EXPR, cnt, |
5279 | TYPE_SIZE_UNIT (pointee_type)); |
5280 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5281 | g = gimple_build_call (alloca_decl, 2, tmp, |
5282 | size_int (TYPE_ALIGN (pointee_type))); |
5283 | gimple_call_set_lhs (gs: g, OMP_CLAUSE_DECL (c)); |
5284 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5285 | } |
5286 | |
5287 | sixth_bb = split_block (fifth_bb, g)->dest; |
5288 | gsi = gsi_last_nondebug_bb (bb: sixth_bb); |
5289 | } |
5290 | |
5291 | t = build2 (MULT_EXPR, itype, q, threadid); |
5292 | t = build2 (PLUS_EXPR, itype, t, tt); |
5293 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
5294 | |
5295 | t = fold_build2 (PLUS_EXPR, itype, s0, q); |
5296 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT); |
5297 | |
5298 | t = build2 (GE_EXPR, boolean_type_node, s0, e0); |
5299 | gsi_insert_before (&gsi, gimple_build_cond_empty (cond: t), GSI_SAME_STMT); |
5300 | |
5301 | /* Remove the GIMPLE_OMP_FOR statement. */ |
5302 | gsi_remove (&gsi, true); |
5303 | |
5304 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ |
5305 | gsi = gsi_start_bb (bb: seq_start_bb); |
5306 | |
5307 | tree startvar = fd->loop.v; |
5308 | tree endvar = NULL_TREE; |
5309 | |
5310 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
5311 | { |
5312 | tree clauses = gimple_code (g: inner_stmt) == GIMPLE_OMP_PARALLEL |
5313 | ? gimple_omp_parallel_clauses (gs: inner_stmt) |
5314 | : gimple_omp_for_clauses (gs: inner_stmt); |
5315 | tree innerc = omp_find_clause (clauses, kind: OMP_CLAUSE__LOOPTEMP_); |
5316 | gcc_assert (innerc); |
5317 | startvar = OMP_CLAUSE_DECL (innerc); |
5318 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
5319 | kind: OMP_CLAUSE__LOOPTEMP_); |
5320 | gcc_assert (innerc); |
5321 | endvar = OMP_CLAUSE_DECL (innerc); |
5322 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST |
5323 | && gimple_omp_for_kind (g: fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
5324 | { |
5325 | innerc = find_lastprivate_looptemp (fd, innerc); |
5326 | if (innerc) |
5327 | { |
5328 | /* If needed (distribute parallel for with lastprivate), |
5329 | propagate down the total number of iterations. */ |
5330 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), |
5331 | fd->loop.n2); |
5332 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, |
5333 | GSI_CONTINUE_LINKING); |
5334 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); |
5335 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
5336 | } |
5337 | } |
5338 | } |
5339 | t = fold_convert (itype, s0); |
5340 | t = fold_build2 (MULT_EXPR, itype, t, step); |
5341 | if (POINTER_TYPE_P (type)) |
5342 | { |
5343 | t = fold_build_pointer_plus (n1, t); |
5344 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) |
5345 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) |
5346 | t = fold_convert (signed_type_for (type), t); |
5347 | } |
5348 | else |
5349 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
5350 | t = fold_convert (TREE_TYPE (startvar), t); |
5351 | t = force_gimple_operand_gsi (&gsi, t, |
5352 | DECL_P (startvar) |
5353 | && TREE_ADDRESSABLE (startvar), |
5354 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
5355 | assign_stmt = gimple_build_assign (startvar, t); |
5356 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
5357 | if (cond_var) |
5358 | { |
5359 | tree itype = TREE_TYPE (cond_var); |
5360 | /* For lastprivate(conditional:) itervar, we need some iteration |
5361 | counter that starts at unsigned non-zero and increases. |
5362 | Prefer as few IVs as possible, so if we can use startvar |
5363 | itself, use that, or startvar + constant (those would be |
5364 | incremented with step), and as last resort use the s0 + 1 |
5365 | incremented by 1. */ |
5366 | if (POINTER_TYPE_P (type) |
5367 | || TREE_CODE (n1) != INTEGER_CST |
5368 | || fd->loop.cond_code != LT_EXPR) |
5369 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), |
5370 | build_int_cst (itype, 1)); |
5371 | else if (tree_int_cst_sgn (n1) == 1) |
5372 | t = fold_convert (itype, t); |
5373 | else |
5374 | { |
5375 | tree c = fold_convert (itype, n1); |
5376 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); |
5377 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); |
5378 | } |
5379 | t = force_gimple_operand_gsi (&gsi, t, false, |
5380 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
5381 | assign_stmt = gimple_build_assign (cond_var, t); |
5382 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
5383 | } |
5384 | |
5385 | t = fold_convert (itype, e0); |
5386 | t = fold_build2 (MULT_EXPR, itype, t, step); |
5387 | if (POINTER_TYPE_P (type)) |
5388 | { |
5389 | t = fold_build_pointer_plus (n1, t); |
5390 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) |
5391 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) |
5392 | t = fold_convert (signed_type_for (type), t); |
5393 | } |
5394 | else |
5395 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
5396 | t = fold_convert (TREE_TYPE (startvar), t); |
5397 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5398 | false, GSI_CONTINUE_LINKING); |
5399 | if (endvar) |
5400 | { |
5401 | assign_stmt = gimple_build_assign (endvar, e); |
5402 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
5403 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) |
5404 | assign_stmt = gimple_build_assign (fd->loop.v, e); |
5405 | else |
5406 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); |
5407 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
5408 | } |
5409 | /* Handle linear clause adjustments. */ |
5410 | tree itercnt = NULL_TREE; |
5411 | tree *nonrect_bounds = NULL; |
5412 | if (gimple_omp_for_kind (g: fd->for_stmt) == GF_OMP_FOR_KIND_FOR) |
5413 | for (tree c = gimple_omp_for_clauses (gs: fd->for_stmt); |
5414 | c; c = OMP_CLAUSE_CHAIN (c)) |
5415 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
5416 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) |
5417 | { |
5418 | tree d = OMP_CLAUSE_DECL (c); |
5419 | tree t = d, a, dest; |
5420 | if (omp_privatize_by_reference (decl: t)) |
5421 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); |
5422 | if (itercnt == NULL_TREE) |
5423 | { |
5424 | if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
5425 | { |
5426 | itercnt = fold_build2 (MINUS_EXPR, itype, |
5427 | fold_convert (itype, n1), |
5428 | fold_convert (itype, fd->loop.n1)); |
5429 | itercnt = fold_build2 (EXACT_DIV_EXPR, itype, itercnt, step); |
5430 | itercnt = fold_build2 (PLUS_EXPR, itype, itercnt, s0); |
5431 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, |
5432 | NULL_TREE, false, |
5433 | GSI_CONTINUE_LINKING); |
5434 | } |
5435 | else |
5436 | itercnt = s0; |
5437 | } |
5438 | tree type = TREE_TYPE (t); |
5439 | if (POINTER_TYPE_P (type)) |
5440 | type = sizetype; |
5441 | a = fold_build2 (MULT_EXPR, type, |
5442 | fold_convert (type, itercnt), |
5443 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); |
5444 | dest = unshare_expr (t); |
5445 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR |
5446 | : POINTER_PLUS_EXPR, TREE_TYPE (t), t, a); |
5447 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5448 | false, GSI_CONTINUE_LINKING); |
5449 | expand_omp_build_assign (gsi_p: &gsi, to: dest, from: t, after: true); |
5450 | } |
5451 | if (fd->collapse > 1) |
5452 | { |
5453 | if (fd->non_rect) |
5454 | { |
5455 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); |
5456 | memset (s: nonrect_bounds, c: 0, n: sizeof (tree) * (fd->last_nonrect + 1)); |
5457 | } |
5458 | expand_omp_for_init_vars (fd, gsi: &gsi, counts, nonrect_bounds, inner_stmt, |
5459 | startvar); |
5460 | } |
5461 | |
5462 | if (!broken_loop) |
5463 | { |
5464 | /* The code controlling the sequential loop replaces the |
5465 | GIMPLE_OMP_CONTINUE. */ |
5466 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
5467 | gomp_continue *cont_stmt = as_a <gomp_continue *> (p: gsi_stmt (i: gsi)); |
5468 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); |
5469 | vmain = gimple_omp_continue_control_use (cont_stmt); |
5470 | vback = gimple_omp_continue_control_def (cont_stmt); |
5471 | |
5472 | if (cond_var) |
5473 | { |
5474 | tree itype = TREE_TYPE (cond_var); |
5475 | tree t2; |
5476 | if (POINTER_TYPE_P (type) |
5477 | || TREE_CODE (n1) != INTEGER_CST |
5478 | || fd->loop.cond_code != LT_EXPR) |
5479 | t2 = build_int_cst (itype, 1); |
5480 | else |
5481 | t2 = fold_convert (itype, step); |
5482 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); |
5483 | t2 = force_gimple_operand_gsi (&gsi, t2, false, |
5484 | NULL_TREE, true, GSI_SAME_STMT); |
5485 | assign_stmt = gimple_build_assign (cond_var, t2); |
5486 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
5487 | } |
5488 | |
5489 | if (!gimple_omp_for_combined_p (g: fd->for_stmt)) |
5490 | { |
5491 | if (POINTER_TYPE_P (type)) |
5492 | t = fold_build_pointer_plus (vmain, step); |
5493 | else |
5494 | t = fold_build2 (PLUS_EXPR, type, vmain, step); |
5495 | t = force_gimple_operand_gsi (&gsi, t, |
5496 | DECL_P (vback) |
5497 | && TREE_ADDRESSABLE (vback), |
5498 | NULL_TREE, true, GSI_SAME_STMT); |
5499 | assign_stmt = gimple_build_assign (vback, t); |
5500 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
5501 | |
5502 | t = build2 (fd->loop.cond_code, boolean_type_node, |
5503 | DECL_P (vback) && TREE_ADDRESSABLE (vback) |
5504 | ? t : vback, e); |
5505 | gsi_insert_before (&gsi, gimple_build_cond_empty (cond: t), GSI_SAME_STMT); |
5506 | } |
5507 | |
5508 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ |
5509 | gsi_remove (&gsi, true); |
5510 | |
5511 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (g: fd->for_stmt)) |
5512 | collapse_bb = extract_omp_for_update_vars (fd, nonrect_bounds, |
5513 | cont_bb, body_bb); |
5514 | } |
5515 | |
5516 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ |
5517 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
5518 | if (!gimple_omp_return_nowait_p (g: gsi_stmt (i: gsi))) |
5519 | { |
5520 | t = gimple_omp_return_lhs (g: gsi_stmt (i: gsi)); |
5521 | if (fd->have_reductemp |
5522 | || ((fd->have_pointer_condtemp || fd->have_scantemp) |
5523 | && !fd->have_nonctrl_scantemp)) |
5524 | { |
5525 | tree fn; |
5526 | if (t) |
5527 | fn = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END_CANCEL); |
5528 | else |
5529 | fn = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END); |
5530 | gcall *g = gimple_build_call (fn, 0); |
5531 | if (t) |
5532 | { |
5533 | gimple_call_set_lhs (gs: g, lhs: t); |
5534 | if (fd->have_reductemp) |
5535 | gsi_insert_after (&gsi, gimple_build_assign (reductions, |
5536 | NOP_EXPR, t), |
5537 | GSI_SAME_STMT); |
5538 | } |
5539 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); |
5540 | } |
5541 | else |
5542 | gsi_insert_after (&gsi, omp_build_barrier (lhs: t), GSI_SAME_STMT); |
5543 | } |
5544 | else if ((fd->have_pointer_condtemp || fd->have_scantemp) |
5545 | && !fd->have_nonctrl_scantemp) |
5546 | { |
5547 | tree fn = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END_NOWAIT); |
5548 | gcall *g = gimple_build_call (fn, 0); |
5549 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); |
5550 | } |
5551 | if (fd->have_scantemp && !fd->have_nonctrl_scantemp) |
5552 | { |
5553 | tree clauses = gimple_omp_for_clauses (gs: fd->for_stmt); |
5554 | tree controlp = NULL_TREE, controlb = NULL_TREE; |
5555 | for (tree c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) |
5556 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE__SCANTEMP_ |
5557 | && OMP_CLAUSE__SCANTEMP__CONTROL (c)) |
5558 | { |
5559 | if (TREE_TYPE (OMP_CLAUSE_DECL (c)) == boolean_type_node) |
5560 | controlb = OMP_CLAUSE_DECL (c); |
5561 | else |
5562 | controlp = OMP_CLAUSE_DECL (c); |
5563 | if (controlb && controlp) |
5564 | break; |
5565 | } |
5566 | gcc_assert (controlp && controlb); |
5567 | gimple *g = gimple_build_cond (NE_EXPR, controlb, boolean_false_node, |
5568 | NULL_TREE, NULL_TREE); |
5569 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5570 | exit1_bb = split_block (exit_bb, g)->dest; |
5571 | gsi = gsi_after_labels (bb: exit1_bb); |
5572 | g = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_FREE), 1, |
5573 | controlp); |
5574 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5575 | exit2_bb = split_block (exit1_bb, g)->dest; |
5576 | gsi = gsi_after_labels (bb: exit2_bb); |
5577 | g = gimple_build_call (builtin_decl_implicit (fncode: BUILT_IN_STACK_RESTORE), 1, |
5578 | controlp); |
5579 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5580 | exit3_bb = split_block (exit2_bb, g)->dest; |
5581 | gsi = gsi_after_labels (bb: exit3_bb); |
5582 | } |
5583 | gsi_remove (&gsi, true); |
5584 | |
5585 | /* Connect all the blocks. */ |
5586 | ep = make_edge (entry_bb, third_bb, EDGE_FALSE_VALUE); |
5587 | ep->probability = profile_probability::guessed_always ().apply_scale (num: 3, den: 4); |
5588 | ep = find_edge (entry_bb, second_bb); |
5589 | ep->flags = EDGE_TRUE_VALUE; |
5590 | ep->probability = profile_probability::guessed_always () / 4; |
5591 | if (fourth_bb) |
5592 | { |
5593 | ep = make_edge (third_bb, fifth_bb, EDGE_FALSE_VALUE); |
5594 | ep->probability = profile_probability::guessed_always () / 2; |
5595 | ep = find_edge (third_bb, fourth_bb); |
5596 | ep->flags = EDGE_TRUE_VALUE; |
5597 | ep->probability = profile_probability::guessed_always () / 2; |
5598 | ep = find_edge (fourth_bb, fifth_bb); |
5599 | redirect_edge_and_branch (ep, sixth_bb); |
5600 | } |
5601 | else |
5602 | sixth_bb = third_bb; |
5603 | find_edge (sixth_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE; |
5604 | find_edge (sixth_bb, fin_bb)->flags = EDGE_TRUE_VALUE; |
5605 | if (exit1_bb) |
5606 | { |
5607 | ep = make_edge (exit_bb, exit2_bb, EDGE_FALSE_VALUE); |
5608 | ep->probability = profile_probability::guessed_always () / 2; |
5609 | ep = find_edge (exit_bb, exit1_bb); |
5610 | ep->flags = EDGE_TRUE_VALUE; |
5611 | ep->probability = profile_probability::guessed_always () / 2; |
5612 | ep = find_edge (exit1_bb, exit2_bb); |
5613 | redirect_edge_and_branch (ep, exit3_bb); |
5614 | } |
5615 | |
5616 | if (!broken_loop) |
5617 | { |
5618 | ep = find_edge (cont_bb, body_bb); |
5619 | if (ep == NULL) |
5620 | { |
5621 | ep = BRANCH_EDGE (cont_bb); |
5622 | gcc_assert (single_succ (ep->dest) == body_bb); |
5623 | } |
5624 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
5625 | { |
5626 | remove_edge (ep); |
5627 | ep = NULL; |
5628 | } |
5629 | else if (fd->collapse > 1) |
5630 | { |
5631 | remove_edge (ep); |
5632 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); |
5633 | } |
5634 | else |
5635 | ep->flags = EDGE_TRUE_VALUE; |
5636 | find_edge (cont_bb, fin_bb)->flags |
5637 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; |
5638 | } |
5639 | |
5640 | set_immediate_dominator (CDI_DOMINATORS, second_bb, entry_bb); |
5641 | set_immediate_dominator (CDI_DOMINATORS, third_bb, entry_bb); |
5642 | if (fourth_bb) |
5643 | { |
5644 | set_immediate_dominator (CDI_DOMINATORS, fifth_bb, third_bb); |
5645 | set_immediate_dominator (CDI_DOMINATORS, sixth_bb, third_bb); |
5646 | } |
5647 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, sixth_bb); |
5648 | |
5649 | set_immediate_dominator (CDI_DOMINATORS, body_bb, |
5650 | recompute_dominator (CDI_DOMINATORS, body_bb)); |
5651 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, |
5652 | recompute_dominator (CDI_DOMINATORS, fin_bb)); |
5653 | if (exit1_bb) |
5654 | { |
5655 | set_immediate_dominator (CDI_DOMINATORS, exit2_bb, exit_bb); |
5656 | set_immediate_dominator (CDI_DOMINATORS, exit3_bb, exit_bb); |
5657 | } |
5658 | |
5659 | class loop *loop = body_bb->loop_father; |
5660 | if (loop != entry_bb->loop_father) |
5661 | { |
5662 | gcc_assert (broken_loop || loop->header == body_bb); |
5663 | gcc_assert (broken_loop |
5664 | || loop->latch == region->cont |
5665 | || single_pred (loop->latch) == region->cont); |
5666 | return; |
5667 | } |
5668 | |
5669 | if (!broken_loop && !gimple_omp_for_combined_p (g: fd->for_stmt)) |
5670 | { |
5671 | loop = alloc_loop (); |
5672 | loop->header = body_bb; |
5673 | if (collapse_bb == NULL) |
5674 | loop->latch = cont_bb; |
5675 | add_loop (loop, body_bb->loop_father); |
5676 | } |
5677 | } |
5678 | |
5679 | /* Return phi in E->DEST with ARG on edge E. */ |
5680 | |
5681 | static gphi * |
5682 | find_phi_with_arg_on_edge (tree arg, edge e) |
5683 | { |
5684 | basic_block bb = e->dest; |
5685 | |
5686 | for (gphi_iterator gpi = gsi_start_phis (bb); |
5687 | !gsi_end_p (i: gpi); |
5688 | gsi_next (i: &gpi)) |
5689 | { |
5690 | gphi *phi = gpi.phi (); |
5691 | if (PHI_ARG_DEF_FROM_EDGE (phi, e) == arg) |
5692 | return phi; |
5693 | } |
5694 | |
5695 | return NULL; |
5696 | } |
5697 | |
5698 | /* A subroutine of expand_omp_for. Generate code for a parallel |
5699 | loop with static schedule and a specified chunk size. Given |
5700 | parameters: |
5701 | |
5702 | for (V = N1; V cond N2; V += STEP) BODY; |
5703 | |
5704 | where COND is "<" or ">", we generate pseudocode |
5705 | |
5706 | if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; |
5707 | if (cond is <) |
5708 | adj = STEP - 1; |
5709 | else |
5710 | adj = STEP + 1; |
5711 | if ((__typeof (V)) -1 > 0 && cond is >) |
5712 | n = -(adj + N2 - N1) / -STEP; |
5713 | else |
5714 | n = (adj + N2 - N1) / STEP; |
5715 | trip = 0; |
5716 | V = threadid * CHUNK * STEP + N1; -- this extra definition of V is |
5717 | here so that V is defined |
5718 | if the loop is not entered |
5719 | L0: |
5720 | s0 = (trip * nthreads + threadid) * CHUNK; |
5721 | e0 = min (s0 + CHUNK, n); |
5722 | if (s0 < n) goto L1; else goto L4; |
5723 | L1: |
5724 | V = s0 * STEP + N1; |
5725 | e = e0 * STEP + N1; |
5726 | L2: |
5727 | BODY; |
5728 | V += STEP; |
5729 | if (V cond e) goto L2; else goto L3; |
5730 | L3: |
5731 | trip += 1; |
5732 | goto L0; |
5733 | L4: |
5734 | */ |
5735 | |
5736 | static void |
5737 | expand_omp_for_static_chunk (struct omp_region *region, |
5738 | struct omp_for_data *fd, gimple *inner_stmt) |
5739 | { |
5740 | tree n, s0, e0, e, t; |
5741 | tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid; |
5742 | tree type, itype, vmain, vback, ; |
5743 | basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb; |
5744 | basic_block trip_update_bb = NULL, cont_bb, collapse_bb = NULL, fin_bb; |
5745 | gimple_stmt_iterator gsi, gsip; |
5746 | edge se; |
5747 | bool broken_loop = region->cont == NULL; |
5748 | tree *counts = NULL; |
5749 | tree n1, n2, step; |
5750 | tree reductions = NULL_TREE; |
5751 | tree cond_var = NULL_TREE, condtemp = NULL_TREE; |
5752 | |
5753 | itype = type = TREE_TYPE (fd->loop.v); |
5754 | if (POINTER_TYPE_P (type)) |
5755 | itype = signed_type_for (type); |
5756 | |
5757 | entry_bb = region->entry; |
5758 | se = split_block (entry_bb, last_nondebug_stmt (entry_bb)); |
5759 | entry_bb = se->src; |
5760 | iter_part_bb = se->dest; |
5761 | cont_bb = region->cont; |
5762 | gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2); |
5763 | fin_bb = BRANCH_EDGE (iter_part_bb)->dest; |
5764 | gcc_assert (broken_loop |
5765 | || fin_bb == FALLTHRU_EDGE (cont_bb)->dest); |
5766 | seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb)); |
5767 | body_bb = single_succ (bb: seq_start_bb); |
5768 | if (!broken_loop) |
5769 | { |
5770 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb |
5771 | || single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb); |
5772 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); |
5773 | trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb)); |
5774 | } |
5775 | exit_bb = region->exit; |
5776 | |
5777 | /* Trip and adjustment setup goes in ENTRY_BB. */ |
5778 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
5779 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
5780 | gsip = gsi; |
5781 | gsi_prev (i: &gsip); |
5782 | |
5783 | if (fd->collapse > 1) |
5784 | { |
5785 | int first_zero_iter = -1, dummy = -1; |
5786 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; |
5787 | |
5788 | counts = XALLOCAVEC (tree, fd->collapse); |
5789 | expand_omp_for_init_counts (fd, gsi: &gsi, entry_bb, counts, |
5790 | zero_iter1_bb&: fin_bb, first_zero_iter1&: first_zero_iter, |
5791 | zero_iter2_bb&: dummy_bb, first_zero_iter2&: dummy, l2_dom_bb); |
5792 | t = NULL_TREE; |
5793 | } |
5794 | else if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
5795 | t = integer_one_node; |
5796 | else |
5797 | t = fold_binary (fd->loop.cond_code, boolean_type_node, |
5798 | fold_convert (type, fd->loop.n1), |
5799 | fold_convert (type, fd->loop.n2)); |
5800 | if (fd->collapse == 1 |
5801 | && TYPE_UNSIGNED (type) |
5802 | && (t == NULL_TREE || !integer_onep (t))) |
5803 | { |
5804 | n1 = fold_convert (type, unshare_expr (fd->loop.n1)); |
5805 | n1 = force_gimple_operand_gsi (&gsi, n1, true, NULL_TREE, |
5806 | true, GSI_SAME_STMT); |
5807 | n2 = fold_convert (type, unshare_expr (fd->loop.n2)); |
5808 | n2 = force_gimple_operand_gsi (&gsi, n2, true, NULL_TREE, |
5809 | true, GSI_SAME_STMT); |
5810 | gcond *cond_stmt = expand_omp_build_cond (gsi_p: &gsi, code: fd->loop.cond_code, |
5811 | lhs: n1, rhs: n2); |
5812 | se = split_block (entry_bb, cond_stmt); |
5813 | se->flags = EDGE_TRUE_VALUE; |
5814 | entry_bb = se->dest; |
5815 | se->probability = profile_probability::very_likely (); |
5816 | se = make_edge (se->src, fin_bb, EDGE_FALSE_VALUE); |
5817 | se->probability = profile_probability::very_unlikely (); |
5818 | if (gimple_in_ssa_p (cfun)) |
5819 | { |
5820 | int dest_idx = find_edge (iter_part_bb, fin_bb)->dest_idx; |
5821 | for (gphi_iterator gpi = gsi_start_phis (fin_bb); |
5822 | !gsi_end_p (i: gpi); gsi_next (i: &gpi)) |
5823 | { |
5824 | gphi *phi = gpi.phi (); |
5825 | add_phi_arg (phi, gimple_phi_arg_def (gs: phi, index: dest_idx), |
5826 | se, UNKNOWN_LOCATION); |
5827 | } |
5828 | } |
5829 | gsi = gsi_last_bb (bb: entry_bb); |
5830 | } |
5831 | |
5832 | if (fd->lastprivate_conditional) |
5833 | { |
5834 | tree clauses = gimple_omp_for_clauses (gs: fd->for_stmt); |
5835 | tree c = omp_find_clause (clauses, kind: OMP_CLAUSE__CONDTEMP_); |
5836 | if (fd->have_pointer_condtemp) |
5837 | condtemp = OMP_CLAUSE_DECL (c); |
5838 | c = omp_find_clause (OMP_CLAUSE_CHAIN (c), kind: OMP_CLAUSE__CONDTEMP_); |
5839 | cond_var = OMP_CLAUSE_DECL (c); |
5840 | } |
5841 | if (fd->have_reductemp || fd->have_pointer_condtemp) |
5842 | { |
5843 | tree t1 = build_int_cst (long_integer_type_node, 0); |
5844 | tree t2 = build_int_cst (long_integer_type_node, 1); |
5845 | tree t3 = build_int_cstu (long_integer_type_node, |
5846 | (HOST_WIDE_INT_1U << 31) + 1); |
5847 | tree clauses = gimple_omp_for_clauses (gs: fd->for_stmt); |
5848 | gimple_stmt_iterator gsi2 = gsi_none (); |
5849 | gimple *g = NULL; |
5850 | tree mem = null_pointer_node, memv = NULL_TREE; |
5851 | if (fd->have_reductemp) |
5852 | { |
5853 | tree c = omp_find_clause (clauses, kind: OMP_CLAUSE__REDUCTEMP_); |
5854 | reductions = OMP_CLAUSE_DECL (c); |
5855 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); |
5856 | g = SSA_NAME_DEF_STMT (reductions); |
5857 | reductions = gimple_assign_rhs1 (gs: g); |
5858 | OMP_CLAUSE_DECL (c) = reductions; |
5859 | gsi2 = gsi_for_stmt (g); |
5860 | } |
5861 | else |
5862 | { |
5863 | if (gsi_end_p (i: gsip)) |
5864 | gsi2 = gsi_after_labels (bb: region->entry); |
5865 | else |
5866 | gsi2 = gsip; |
5867 | reductions = null_pointer_node; |
5868 | } |
5869 | if (fd->have_pointer_condtemp) |
5870 | { |
5871 | tree type = TREE_TYPE (condtemp); |
5872 | memv = create_tmp_var (type); |
5873 | TREE_ADDRESSABLE (memv) = 1; |
5874 | unsigned HOST_WIDE_INT sz |
5875 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))); |
5876 | sz *= fd->lastprivate_conditional; |
5877 | expand_omp_build_assign (gsi_p: &gsi2, to: memv, from: build_int_cst (type, sz), |
5878 | after: false); |
5879 | mem = build_fold_addr_expr (memv); |
5880 | } |
5881 | tree t |
5882 | = build_call_expr (builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_START), |
5883 | 9, t1, t2, t2, t3, t1, null_pointer_node, |
5884 | null_pointer_node, reductions, mem); |
5885 | force_gimple_operand_gsi (&gsi2, t, true, NULL_TREE, |
5886 | true, GSI_SAME_STMT); |
5887 | if (fd->have_pointer_condtemp) |
5888 | expand_omp_build_assign (gsi_p: &gsi2, to: condtemp, from: memv, after: false); |
5889 | if (fd->have_reductemp) |
5890 | { |
5891 | gsi_remove (&gsi2, true); |
5892 | release_ssa_name (name: gimple_assign_lhs (gs: g)); |
5893 | } |
5894 | } |
5895 | switch (gimple_omp_for_kind (g: fd->for_stmt)) |
5896 | { |
5897 | case GF_OMP_FOR_KIND_FOR: |
5898 | nthreads = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_NUM_THREADS); |
5899 | threadid = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_THREAD_NUM); |
5900 | break; |
5901 | case GF_OMP_FOR_KIND_DISTRIBUTE: |
5902 | nthreads = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_NUM_TEAMS); |
5903 | threadid = builtin_decl_explicit (fncode: BUILT_IN_OMP_GET_TEAM_NUM); |
5904 | break; |
5905 | default: |
5906 | gcc_unreachable (); |
5907 | } |
5908 | nthreads = build_call_expr (nthreads, 0); |
5909 | nthreads = fold_convert (itype, nthreads); |
5910 | nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE, |
5911 | true, GSI_SAME_STMT); |
5912 | threadid = build_call_expr (threadid, 0); |
5913 | threadid = fold_convert (itype, threadid); |
5914 | threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE, |
5915 | true, GSI_SAME_STMT); |
5916 | |
5917 | n1 = fd->loop.n1; |
5918 | n2 = fd->loop.n2; |
5919 | step = fd->loop.step; |
5920 | if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
5921 | { |
5922 | tree innerc = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
5923 | kind: OMP_CLAUSE__LOOPTEMP_); |
5924 | gcc_assert (innerc); |
5925 | n1 = OMP_CLAUSE_DECL (innerc); |
5926 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
5927 | kind: OMP_CLAUSE__LOOPTEMP_); |
5928 | gcc_assert (innerc); |
5929 | n2 = OMP_CLAUSE_DECL (innerc); |
5930 | } |
5931 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), |
5932 | true, NULL_TREE, true, GSI_SAME_STMT); |
5933 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), |
5934 | true, NULL_TREE, true, GSI_SAME_STMT); |
5935 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), |
5936 | true, NULL_TREE, true, GSI_SAME_STMT); |
5937 | tree chunk_size = fold_convert (itype, fd->chunk_size); |
5938 | chunk_size = omp_adjust_chunk_size (chunk_size, simd_schedule: fd->simd_schedule); |
5939 | chunk_size |
5940 | = force_gimple_operand_gsi (&gsi, chunk_size, true, NULL_TREE, true, |
5941 | GSI_SAME_STMT); |
5942 | |
5943 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); |
5944 | t = fold_build2 (PLUS_EXPR, itype, step, t); |
5945 | t = fold_build2 (PLUS_EXPR, itype, t, n2); |
5946 | t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1)); |
5947 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) |
5948 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
5949 | fold_build1 (NEGATE_EXPR, itype, t), |
5950 | fold_build1 (NEGATE_EXPR, itype, step)); |
5951 | else |
5952 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
5953 | t = fold_convert (itype, t); |
5954 | n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5955 | true, GSI_SAME_STMT); |
5956 | |
5957 | trip_var = create_tmp_reg (itype, ".trip" ); |
5958 | if (gimple_in_ssa_p (cfun)) |
5959 | { |
5960 | trip_init = make_ssa_name (var: trip_var); |
5961 | trip_main = make_ssa_name (var: trip_var); |
5962 | trip_back = make_ssa_name (var: trip_var); |
5963 | } |
5964 | else |
5965 | { |
5966 | trip_init = trip_var; |
5967 | trip_main = trip_var; |
5968 | trip_back = trip_var; |
5969 | } |
5970 | |
5971 | gassign *assign_stmt |
5972 | = gimple_build_assign (trip_init, build_int_cst (itype, 0)); |
5973 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
5974 | |
5975 | t = fold_build2 (MULT_EXPR, itype, threadid, chunk_size); |
5976 | t = fold_build2 (MULT_EXPR, itype, t, step); |
5977 | if (POINTER_TYPE_P (type)) |
5978 | t = fold_build_pointer_plus (n1, t); |
5979 | else |
5980 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
5981 | vextra = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5982 | true, GSI_SAME_STMT); |
5983 | |
5984 | /* Remove the GIMPLE_OMP_FOR. */ |
5985 | gsi_remove (&gsi, true); |
5986 | |
5987 | gimple_stmt_iterator gsif = gsi; |
5988 | |
5989 | /* Iteration space partitioning goes in ITER_PART_BB. */ |
5990 | gsi = gsi_last_bb (bb: iter_part_bb); |
5991 | |
5992 | t = fold_build2 (MULT_EXPR, itype, trip_main, nthreads); |
5993 | t = fold_build2 (PLUS_EXPR, itype, t, threadid); |
5994 | t = fold_build2 (MULT_EXPR, itype, t, chunk_size); |
5995 | s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
5996 | false, GSI_CONTINUE_LINKING); |
5997 | |
5998 | t = fold_build2 (PLUS_EXPR, itype, s0, chunk_size); |
5999 | t = fold_build2 (MIN_EXPR, itype, t, n); |
6000 | e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6001 | false, GSI_CONTINUE_LINKING); |
6002 | |
6003 | t = build2 (LT_EXPR, boolean_type_node, s0, n); |
6004 | gsi_insert_after (&gsi, gimple_build_cond_empty (cond: t), GSI_CONTINUE_LINKING); |
6005 | |
6006 | /* Setup code for sequential iteration goes in SEQ_START_BB. */ |
6007 | gsi = gsi_start_bb (bb: seq_start_bb); |
6008 | |
6009 | tree startvar = fd->loop.v; |
6010 | tree endvar = NULL_TREE; |
6011 | |
6012 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
6013 | { |
6014 | tree clauses = gimple_code (g: inner_stmt) == GIMPLE_OMP_PARALLEL |
6015 | ? gimple_omp_parallel_clauses (gs: inner_stmt) |
6016 | : gimple_omp_for_clauses (gs: inner_stmt); |
6017 | tree innerc = omp_find_clause (clauses, kind: OMP_CLAUSE__LOOPTEMP_); |
6018 | gcc_assert (innerc); |
6019 | startvar = OMP_CLAUSE_DECL (innerc); |
6020 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
6021 | kind: OMP_CLAUSE__LOOPTEMP_); |
6022 | gcc_assert (innerc); |
6023 | endvar = OMP_CLAUSE_DECL (innerc); |
6024 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST |
6025 | && gimple_omp_for_kind (g: fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE) |
6026 | { |
6027 | innerc = find_lastprivate_looptemp (fd, innerc); |
6028 | if (innerc) |
6029 | { |
6030 | /* If needed (distribute parallel for with lastprivate), |
6031 | propagate down the total number of iterations. */ |
6032 | tree t = fold_convert (TREE_TYPE (OMP_CLAUSE_DECL (innerc)), |
6033 | fd->loop.n2); |
6034 | t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, false, |
6035 | GSI_CONTINUE_LINKING); |
6036 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); |
6037 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
6038 | } |
6039 | } |
6040 | } |
6041 | |
6042 | t = fold_convert (itype, s0); |
6043 | t = fold_build2 (MULT_EXPR, itype, t, step); |
6044 | if (POINTER_TYPE_P (type)) |
6045 | { |
6046 | t = fold_build_pointer_plus (n1, t); |
6047 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) |
6048 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) |
6049 | t = fold_convert (signed_type_for (type), t); |
6050 | } |
6051 | else |
6052 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6053 | t = fold_convert (TREE_TYPE (startvar), t); |
6054 | t = force_gimple_operand_gsi (&gsi, t, |
6055 | DECL_P (startvar) |
6056 | && TREE_ADDRESSABLE (startvar), |
6057 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
6058 | assign_stmt = gimple_build_assign (startvar, t); |
6059 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
6060 | if (cond_var) |
6061 | { |
6062 | tree itype = TREE_TYPE (cond_var); |
6063 | /* For lastprivate(conditional:) itervar, we need some iteration |
6064 | counter that starts at unsigned non-zero and increases. |
6065 | Prefer as few IVs as possible, so if we can use startvar |
6066 | itself, use that, or startvar + constant (those would be |
6067 | incremented with step), and as last resort use the s0 + 1 |
6068 | incremented by 1. */ |
6069 | if (POINTER_TYPE_P (type) |
6070 | || TREE_CODE (n1) != INTEGER_CST |
6071 | || fd->loop.cond_code != LT_EXPR) |
6072 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, s0), |
6073 | build_int_cst (itype, 1)); |
6074 | else if (tree_int_cst_sgn (n1) == 1) |
6075 | t = fold_convert (itype, t); |
6076 | else |
6077 | { |
6078 | tree c = fold_convert (itype, n1); |
6079 | c = fold_build2 (MINUS_EXPR, itype, build_int_cst (itype, 1), c); |
6080 | t = fold_build2 (PLUS_EXPR, itype, fold_convert (itype, t), c); |
6081 | } |
6082 | t = force_gimple_operand_gsi (&gsi, t, false, |
6083 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
6084 | assign_stmt = gimple_build_assign (cond_var, t); |
6085 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
6086 | } |
6087 | |
6088 | t = fold_convert (itype, e0); |
6089 | t = fold_build2 (MULT_EXPR, itype, t, step); |
6090 | if (POINTER_TYPE_P (type)) |
6091 | { |
6092 | t = fold_build_pointer_plus (n1, t); |
6093 | if (!POINTER_TYPE_P (TREE_TYPE (startvar)) |
6094 | && TYPE_PRECISION (TREE_TYPE (startvar)) > TYPE_PRECISION (type)) |
6095 | t = fold_convert (signed_type_for (type), t); |
6096 | } |
6097 | else |
6098 | t = fold_build2 (PLUS_EXPR, type, t, n1); |
6099 | t = fold_convert (TREE_TYPE (startvar), t); |
6100 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6101 | false, GSI_CONTINUE_LINKING); |
6102 | if (endvar) |
6103 | { |
6104 | assign_stmt = gimple_build_assign (endvar, e); |
6105 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
6106 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) |
6107 | assign_stmt = gimple_build_assign (fd->loop.v, e); |
6108 | else |
6109 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); |
6110 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
6111 | } |
6112 | /* Handle linear clause adjustments. */ |
6113 | tree itercnt = NULL_TREE, itercntbias = NULL_TREE; |
6114 | if (gimple_omp_for_kind (g: fd->for_stmt) == GF_OMP_FOR_KIND_FOR) |
6115 | for (tree c = gimple_omp_for_clauses (gs: fd->for_stmt); |
6116 | c; c = OMP_CLAUSE_CHAIN (c)) |
6117 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR |
6118 | && !OMP_CLAUSE_LINEAR_NO_COPYIN (c)) |
6119 | { |
6120 | tree d = OMP_CLAUSE_DECL (c); |
6121 | tree t = d, a, dest; |
6122 | if (omp_privatize_by_reference (decl: t)) |
6123 | t = build_simple_mem_ref_loc (OMP_CLAUSE_LOCATION (c), t); |
6124 | tree type = TREE_TYPE (t); |
6125 | if (POINTER_TYPE_P (type)) |
6126 | type = sizetype; |
6127 | dest = unshare_expr (t); |
6128 | tree v = create_tmp_var (TREE_TYPE (t), NULL); |
6129 | expand_omp_build_assign (gsi_p: &gsif, to: v, from: t); |
6130 | if (itercnt == NULL_TREE) |
6131 | { |
6132 | if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
6133 | { |
6134 | itercntbias |
6135 | = fold_build2 (MINUS_EXPR, itype, fold_convert (itype, n1), |
6136 | fold_convert (itype, fd->loop.n1)); |
6137 | itercntbias = fold_build2 (EXACT_DIV_EXPR, itype, |
6138 | itercntbias, step); |
6139 | itercntbias |
6140 | = force_gimple_operand_gsi (&gsif, itercntbias, true, |
6141 | NULL_TREE, true, |
6142 | GSI_SAME_STMT); |
6143 | itercnt = fold_build2 (PLUS_EXPR, itype, itercntbias, s0); |
6144 | itercnt = force_gimple_operand_gsi (&gsi, itercnt, true, |
6145 | NULL_TREE, false, |
6146 | GSI_CONTINUE_LINKING); |
6147 | } |
6148 | else |
6149 | itercnt = s0; |
6150 | } |
6151 | a = fold_build2 (MULT_EXPR, type, |
6152 | fold_convert (type, itercnt), |
6153 | fold_convert (type, OMP_CLAUSE_LINEAR_STEP (c))); |
6154 | t = fold_build2 (type == TREE_TYPE (t) ? PLUS_EXPR |
6155 | : POINTER_PLUS_EXPR, TREE_TYPE (t), v, a); |
6156 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6157 | false, GSI_CONTINUE_LINKING); |
6158 | expand_omp_build_assign (gsi_p: &gsi, to: dest, from: t, after: true); |
6159 | } |
6160 | if (fd->collapse > 1) |
6161 | expand_omp_for_init_vars (fd, gsi: &gsi, counts, NULL, inner_stmt, startvar); |
6162 | |
6163 | if (!broken_loop) |
6164 | { |
6165 | /* The code controlling the sequential loop goes in CONT_BB, |
6166 | replacing the GIMPLE_OMP_CONTINUE. */ |
6167 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
6168 | gomp_continue *cont_stmt = as_a <gomp_continue *> (p: gsi_stmt (i: gsi)); |
6169 | vmain = gimple_omp_continue_control_use (cont_stmt); |
6170 | vback = gimple_omp_continue_control_def (cont_stmt); |
6171 | |
6172 | if (cond_var) |
6173 | { |
6174 | tree itype = TREE_TYPE (cond_var); |
6175 | tree t2; |
6176 | if (POINTER_TYPE_P (type) |
6177 | || TREE_CODE (n1) != INTEGER_CST |
6178 | || fd->loop.cond_code != LT_EXPR) |
6179 | t2 = build_int_cst (itype, 1); |
6180 | else |
6181 | t2 = fold_convert (itype, step); |
6182 | t2 = fold_build2 (PLUS_EXPR, itype, cond_var, t2); |
6183 | t2 = force_gimple_operand_gsi (&gsi, t2, false, |
6184 | NULL_TREE, true, GSI_SAME_STMT); |
6185 | assign_stmt = gimple_build_assign (cond_var, t2); |
6186 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
6187 | } |
6188 | |
6189 | if (!gimple_omp_for_combined_p (g: fd->for_stmt)) |
6190 | { |
6191 | if (POINTER_TYPE_P (type)) |
6192 | t = fold_build_pointer_plus (vmain, step); |
6193 | else |
6194 | t = fold_build2 (PLUS_EXPR, type, vmain, step); |
6195 | if (DECL_P (vback) && TREE_ADDRESSABLE (vback)) |
6196 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6197 | true, GSI_SAME_STMT); |
6198 | assign_stmt = gimple_build_assign (vback, t); |
6199 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
6200 | |
6201 | if (tree_int_cst_equal (fd->chunk_size, integer_one_node)) |
6202 | t = build2 (EQ_EXPR, boolean_type_node, |
6203 | build_int_cst (itype, 0), |
6204 | build_int_cst (itype, 1)); |
6205 | else |
6206 | t = build2 (fd->loop.cond_code, boolean_type_node, |
6207 | DECL_P (vback) && TREE_ADDRESSABLE (vback) |
6208 | ? t : vback, e); |
6209 | gsi_insert_before (&gsi, gimple_build_cond_empty (cond: t), GSI_SAME_STMT); |
6210 | } |
6211 | |
6212 | /* Remove GIMPLE_OMP_CONTINUE. */ |
6213 | gsi_remove (&gsi, true); |
6214 | |
6215 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (g: fd->for_stmt)) |
6216 | collapse_bb = extract_omp_for_update_vars (fd, NULL, cont_bb, body_bb); |
6217 | |
6218 | /* Trip update code goes into TRIP_UPDATE_BB. */ |
6219 | gsi = gsi_start_bb (bb: trip_update_bb); |
6220 | |
6221 | t = build_int_cst (itype, 1); |
6222 | t = build2 (PLUS_EXPR, itype, trip_main, t); |
6223 | assign_stmt = gimple_build_assign (trip_back, t); |
6224 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
6225 | } |
6226 | |
6227 | /* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */ |
6228 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
6229 | if (!gimple_omp_return_nowait_p (g: gsi_stmt (i: gsi))) |
6230 | { |
6231 | t = gimple_omp_return_lhs (g: gsi_stmt (i: gsi)); |
6232 | if (fd->have_reductemp || fd->have_pointer_condtemp) |
6233 | { |
6234 | tree fn; |
6235 | if (t) |
6236 | fn = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END_CANCEL); |
6237 | else |
6238 | fn = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END); |
6239 | gcall *g = gimple_build_call (fn, 0); |
6240 | if (t) |
6241 | { |
6242 | gimple_call_set_lhs (gs: g, lhs: t); |
6243 | if (fd->have_reductemp) |
6244 | gsi_insert_after (&gsi, gimple_build_assign (reductions, |
6245 | NOP_EXPR, t), |
6246 | GSI_SAME_STMT); |
6247 | } |
6248 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); |
6249 | } |
6250 | else |
6251 | gsi_insert_after (&gsi, omp_build_barrier (lhs: t), GSI_SAME_STMT); |
6252 | } |
6253 | else if (fd->have_pointer_condtemp) |
6254 | { |
6255 | tree fn = builtin_decl_explicit (fncode: BUILT_IN_GOMP_LOOP_END_NOWAIT); |
6256 | gcall *g = gimple_build_call (fn, 0); |
6257 | gsi_insert_after (&gsi, g, GSI_SAME_STMT); |
6258 | } |
6259 | gsi_remove (&gsi, true); |
6260 | |
6261 | /* Connect the new blocks. */ |
6262 | find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE; |
6263 | find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE; |
6264 | |
6265 | if (!broken_loop) |
6266 | { |
6267 | se = find_edge (cont_bb, body_bb); |
6268 | if (se == NULL) |
6269 | { |
6270 | se = BRANCH_EDGE (cont_bb); |
6271 | gcc_assert (single_succ (se->dest) == body_bb); |
6272 | } |
6273 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
6274 | { |
6275 | remove_edge (se); |
6276 | se = NULL; |
6277 | } |
6278 | else if (fd->collapse > 1) |
6279 | { |
6280 | remove_edge (se); |
6281 | se = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); |
6282 | } |
6283 | else |
6284 | se->flags = EDGE_TRUE_VALUE; |
6285 | find_edge (cont_bb, trip_update_bb)->flags |
6286 | = se ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; |
6287 | |
6288 | redirect_edge_and_branch (single_succ_edge (bb: trip_update_bb), |
6289 | iter_part_bb); |
6290 | } |
6291 | |
6292 | if (gimple_in_ssa_p (cfun)) |
6293 | { |
6294 | gphi_iterator psi; |
6295 | gphi *phi; |
6296 | edge re, ene; |
6297 | edge_var_map *vm; |
6298 | size_t i; |
6299 | |
6300 | gcc_assert (fd->collapse == 1 && !broken_loop); |
6301 | |
6302 | /* When we redirect the edge from trip_update_bb to iter_part_bb, we |
6303 | remove arguments of the phi nodes in fin_bb. We need to create |
6304 | appropriate phi nodes in iter_part_bb instead. */ |
6305 | se = find_edge (iter_part_bb, fin_bb); |
6306 | re = single_succ_edge (bb: trip_update_bb); |
6307 | vec<edge_var_map> *head = redirect_edge_var_map_vector (re); |
6308 | ene = single_succ_edge (bb: entry_bb); |
6309 | |
6310 | psi = gsi_start_phis (fin_bb); |
6311 | for (i = 0; !gsi_end_p (i: psi) && head->iterate (ix: i, ptr: &vm); |
6312 | gsi_next (i: &psi), ++i) |
6313 | { |
6314 | gphi *nphi; |
6315 | location_t locus; |
6316 | |
6317 | phi = psi.phi (); |
6318 | if (operand_equal_p (gimple_phi_arg_def (gs: phi, index: 0), |
6319 | redirect_edge_var_map_def (v: vm), flags: 0)) |
6320 | continue; |
6321 | |
6322 | t = gimple_phi_result (gs: phi); |
6323 | gcc_assert (t == redirect_edge_var_map_result (vm)); |
6324 | |
6325 | if (!single_pred_p (bb: fin_bb)) |
6326 | t = copy_ssa_name (var: t, stmt: phi); |
6327 | |
6328 | nphi = create_phi_node (t, iter_part_bb); |
6329 | |
6330 | t = PHI_ARG_DEF_FROM_EDGE (phi, se); |
6331 | locus = gimple_phi_arg_location_from_edge (phi, e: se); |
6332 | |
6333 | /* A special case -- fd->loop.v is not yet computed in |
6334 | iter_part_bb, we need to use vextra instead. */ |
6335 | if (t == fd->loop.v) |
6336 | t = vextra; |
6337 | add_phi_arg (nphi, t, ene, locus); |
6338 | locus = redirect_edge_var_map_location (v: vm); |
6339 | tree back_arg = redirect_edge_var_map_def (v: vm); |
6340 | add_phi_arg (nphi, back_arg, re, locus); |
6341 | edge ce = find_edge (cont_bb, body_bb); |
6342 | if (ce == NULL) |
6343 | { |
6344 | ce = BRANCH_EDGE (cont_bb); |
6345 | gcc_assert (single_succ (ce->dest) == body_bb); |
6346 | ce = single_succ_edge (bb: ce->dest); |
6347 | } |
6348 | gphi *inner_loop_phi = find_phi_with_arg_on_edge (arg: back_arg, e: ce); |
6349 | gcc_assert (inner_loop_phi != NULL); |
6350 | add_phi_arg (inner_loop_phi, gimple_phi_result (gs: nphi), |
6351 | find_edge (seq_start_bb, body_bb), locus); |
6352 | |
6353 | if (!single_pred_p (bb: fin_bb)) |
6354 | add_phi_arg (phi, gimple_phi_result (gs: nphi), se, locus); |
6355 | } |
6356 | gcc_assert (gsi_end_p (psi) && (head == NULL || i == head->length ())); |
6357 | redirect_edge_var_map_clear (re); |
6358 | if (single_pred_p (bb: fin_bb)) |
6359 | while (1) |
6360 | { |
6361 | psi = gsi_start_phis (fin_bb); |
6362 | if (gsi_end_p (i: psi)) |
6363 | break; |
6364 | remove_phi_node (&psi, false); |
6365 | } |
6366 | |
6367 | /* Make phi node for trip. */ |
6368 | phi = create_phi_node (trip_main, iter_part_bb); |
6369 | add_phi_arg (phi, trip_back, single_succ_edge (bb: trip_update_bb), |
6370 | UNKNOWN_LOCATION); |
6371 | add_phi_arg (phi, trip_init, single_succ_edge (bb: entry_bb), |
6372 | UNKNOWN_LOCATION); |
6373 | } |
6374 | |
6375 | if (!broken_loop) |
6376 | set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb); |
6377 | set_immediate_dominator (CDI_DOMINATORS, iter_part_bb, |
6378 | recompute_dominator (CDI_DOMINATORS, iter_part_bb)); |
6379 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, |
6380 | recompute_dominator (CDI_DOMINATORS, fin_bb)); |
6381 | set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, |
6382 | recompute_dominator (CDI_DOMINATORS, seq_start_bb)); |
6383 | set_immediate_dominator (CDI_DOMINATORS, body_bb, |
6384 | recompute_dominator (CDI_DOMINATORS, body_bb)); |
6385 | |
6386 | if (!broken_loop) |
6387 | { |
6388 | class loop *loop = body_bb->loop_father; |
6389 | class loop *trip_loop = alloc_loop (); |
6390 | trip_loop->header = iter_part_bb; |
6391 | trip_loop->latch = trip_update_bb; |
6392 | add_loop (trip_loop, iter_part_bb->loop_father); |
6393 | |
6394 | if (loop != entry_bb->loop_father) |
6395 | { |
6396 | gcc_assert (loop->header == body_bb); |
6397 | gcc_assert (loop->latch == region->cont |
6398 | || single_pred (loop->latch) == region->cont); |
6399 | trip_loop->inner = loop; |
6400 | return; |
6401 | } |
6402 | |
6403 | if (!gimple_omp_for_combined_p (g: fd->for_stmt)) |
6404 | { |
6405 | loop = alloc_loop (); |
6406 | loop->header = body_bb; |
6407 | if (collapse_bb == NULL) |
6408 | loop->latch = cont_bb; |
6409 | add_loop (loop, trip_loop); |
6410 | } |
6411 | } |
6412 | } |
6413 | |
6414 | /* A subroutine of expand_omp_for. Generate code for a simd non-worksharing |
6415 | loop. Given parameters: |
6416 | |
6417 | for (V = N1; V cond N2; V += STEP) BODY; |
6418 | |
6419 | where COND is "<" or ">", we generate pseudocode |
6420 | |
6421 | V = N1; |
6422 | goto L1; |
6423 | L0: |
6424 | BODY; |
6425 | V += STEP; |
6426 | L1: |
6427 | if (V cond N2) goto L0; else goto L2; |
6428 | L2: |
6429 | |
6430 | For collapsed loops, emit the outer loops as scalar |
6431 | and only try to vectorize the innermost loop. */ |
6432 | |
6433 | static void |
6434 | expand_omp_simd (struct omp_region *region, struct omp_for_data *fd) |
6435 | { |
6436 | tree type, t; |
6437 | basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, l2_bb, l2_dom_bb; |
6438 | gimple_stmt_iterator gsi; |
6439 | gimple *stmt; |
6440 | gcond *cond_stmt; |
6441 | bool broken_loop = region->cont == NULL; |
6442 | edge e, ne; |
6443 | tree *counts = NULL; |
6444 | int i; |
6445 | int safelen_int = INT_MAX; |
6446 | bool dont_vectorize = false; |
6447 | tree safelen = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
6448 | kind: OMP_CLAUSE_SAFELEN); |
6449 | tree simduid = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
6450 | kind: OMP_CLAUSE__SIMDUID_); |
6451 | tree ifc = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
6452 | kind: OMP_CLAUSE_IF); |
6453 | tree simdlen = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
6454 | kind: OMP_CLAUSE_SIMDLEN); |
6455 | tree condtemp = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
6456 | kind: OMP_CLAUSE__CONDTEMP_); |
6457 | tree n1, n2; |
6458 | tree cond_var = condtemp ? OMP_CLAUSE_DECL (condtemp) : NULL_TREE; |
6459 | |
6460 | if (safelen) |
6461 | { |
6462 | poly_uint64 val; |
6463 | safelen = OMP_CLAUSE_SAFELEN_EXPR (safelen); |
6464 | if (!poly_int_tree_p (t: safelen, value: &val)) |
6465 | safelen_int = 0; |
6466 | else |
6467 | safelen_int = MIN (constant_lower_bound (val), INT_MAX); |
6468 | if (safelen_int == 1) |
6469 | safelen_int = 0; |
6470 | } |
6471 | if ((ifc && integer_zerop (OMP_CLAUSE_IF_EXPR (ifc))) |
6472 | || (simdlen && integer_onep (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)))) |
6473 | { |
6474 | safelen_int = 0; |
6475 | dont_vectorize = true; |
6476 | } |
6477 | type = TREE_TYPE (fd->loop.v); |
6478 | entry_bb = region->entry; |
6479 | cont_bb = region->cont; |
6480 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
6481 | gcc_assert (broken_loop |
6482 | || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); |
6483 | l0_bb = FALLTHRU_EDGE (entry_bb)->dest; |
6484 | if (!broken_loop) |
6485 | { |
6486 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == l0_bb); |
6487 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); |
6488 | l1_bb = split_block (cont_bb, last_nondebug_stmt (cont_bb))->dest; |
6489 | l2_bb = BRANCH_EDGE (entry_bb)->dest; |
6490 | } |
6491 | else |
6492 | { |
6493 | BRANCH_EDGE (entry_bb)->flags &= ~EDGE_ABNORMAL; |
6494 | l1_bb = split_edge (BRANCH_EDGE (entry_bb)); |
6495 | l2_bb = single_succ (bb: l1_bb); |
6496 | } |
6497 | exit_bb = region->exit; |
6498 | l2_dom_bb = NULL; |
6499 | |
6500 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
6501 | |
6502 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
6503 | /* Not needed in SSA form right now. */ |
6504 | gcc_assert (!gimple_in_ssa_p (cfun)); |
6505 | if (fd->collapse > 1 |
6506 | && (gimple_omp_for_combined_into_p (g: fd->for_stmt) |
6507 | || broken_loop)) |
6508 | { |
6509 | int first_zero_iter = -1, dummy = -1; |
6510 | basic_block zero_iter_bb = l2_bb, dummy_bb = NULL; |
6511 | |
6512 | counts = XALLOCAVEC (tree, fd->collapse); |
6513 | expand_omp_for_init_counts (fd, gsi: &gsi, entry_bb, counts, |
6514 | zero_iter1_bb&: zero_iter_bb, first_zero_iter1&: first_zero_iter, |
6515 | zero_iter2_bb&: dummy_bb, first_zero_iter2&: dummy, l2_dom_bb); |
6516 | } |
6517 | if (l2_dom_bb == NULL) |
6518 | l2_dom_bb = l1_bb; |
6519 | |
6520 | n1 = fd->loop.n1; |
6521 | n2 = fd->loop.n2; |
6522 | if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
6523 | { |
6524 | tree innerc = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
6525 | kind: OMP_CLAUSE__LOOPTEMP_); |
6526 | gcc_assert (innerc); |
6527 | n1 = OMP_CLAUSE_DECL (innerc); |
6528 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
6529 | kind: OMP_CLAUSE__LOOPTEMP_); |
6530 | gcc_assert (innerc); |
6531 | n2 = OMP_CLAUSE_DECL (innerc); |
6532 | } |
6533 | tree step = fd->loop.step; |
6534 | tree orig_step = step; /* May be different from step if is_simt. */ |
6535 | |
6536 | bool is_simt = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
6537 | kind: OMP_CLAUSE__SIMT_); |
6538 | if (is_simt) |
6539 | { |
6540 | cfun->curr_properties &= ~PROP_gimple_lomp_dev; |
6541 | is_simt = safelen_int > 1; |
6542 | } |
6543 | tree simt_lane = NULL_TREE, simt_maxlane = NULL_TREE; |
6544 | if (is_simt) |
6545 | { |
6546 | simt_lane = create_tmp_var (unsigned_type_node); |
6547 | gimple *g = gimple_build_call_internal (IFN_GOMP_SIMT_LANE, 0); |
6548 | gimple_call_set_lhs (gs: g, lhs: simt_lane); |
6549 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
6550 | tree offset = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, |
6551 | fold_convert (TREE_TYPE (step), simt_lane)); |
6552 | n1 = fold_convert (type, n1); |
6553 | if (POINTER_TYPE_P (type)) |
6554 | n1 = fold_build_pointer_plus (n1, offset); |
6555 | else |
6556 | n1 = fold_build2 (PLUS_EXPR, type, n1, fold_convert (type, offset)); |
6557 | |
6558 | /* Collapsed loops not handled for SIMT yet: limit to one lane only. */ |
6559 | if (fd->collapse > 1) |
6560 | simt_maxlane = build_one_cst (unsigned_type_node); |
6561 | else if (safelen_int < omp_max_simt_vf ()) |
6562 | simt_maxlane = build_int_cst (unsigned_type_node, safelen_int); |
6563 | tree vf |
6564 | = build_call_expr_internal_loc (UNKNOWN_LOCATION, IFN_GOMP_SIMT_VF, |
6565 | unsigned_type_node, 0); |
6566 | if (simt_maxlane) |
6567 | vf = fold_build2 (MIN_EXPR, unsigned_type_node, vf, simt_maxlane); |
6568 | vf = fold_convert (TREE_TYPE (step), vf); |
6569 | step = fold_build2 (MULT_EXPR, TREE_TYPE (step), step, vf); |
6570 | } |
6571 | |
6572 | tree n2var = NULL_TREE; |
6573 | tree n2v = NULL_TREE; |
6574 | tree *nonrect_bounds = NULL; |
6575 | tree min_arg1 = NULL_TREE, min_arg2 = NULL_TREE; |
6576 | if (fd->collapse > 1) |
6577 | { |
6578 | if (broken_loop || gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
6579 | { |
6580 | if (fd->non_rect) |
6581 | { |
6582 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); |
6583 | memset (s: nonrect_bounds, c: 0, |
6584 | n: sizeof (tree) * (fd->last_nonrect + 1)); |
6585 | } |
6586 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, fold_convert (type, n1)); |
6587 | gcc_assert (entry_bb == gsi_bb (gsi)); |
6588 | gcc_assert (fd->for_stmt == gsi_stmt (gsi)); |
6589 | gsi_prev (i: &gsi); |
6590 | entry_bb = split_block (entry_bb, gsi_stmt (i: gsi))->dest; |
6591 | expand_omp_for_init_vars (fd, gsi: &gsi, counts, nonrect_bounds, |
6592 | NULL, startvar: n1); |
6593 | gsi = gsi_for_stmt (fd->for_stmt); |
6594 | } |
6595 | if (broken_loop) |
6596 | ; |
6597 | else if (gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
6598 | { |
6599 | /* Compute in n2var the limit for the first innermost loop, |
6600 | i.e. fd->loop.v + MIN (n2 - fd->loop.v, cnt) |
6601 | where cnt is how many iterations would the loop have if |
6602 | all further iterations were assigned to the current task. */ |
6603 | n2var = create_tmp_var (type); |
6604 | i = fd->collapse - 1; |
6605 | tree itype = TREE_TYPE (fd->loops[i].v); |
6606 | if (POINTER_TYPE_P (itype)) |
6607 | itype = signed_type_for (itype); |
6608 | t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR |
6609 | ? -1 : 1)); |
6610 | t = fold_build2 (PLUS_EXPR, itype, |
6611 | fold_convert (itype, fd->loops[i].step), t); |
6612 | t = fold_build2 (PLUS_EXPR, itype, t, |
6613 | fold_convert (itype, fd->loops[i].n2)); |
6614 | if (fd->loops[i].m2) |
6615 | { |
6616 | tree t2 = fold_convert (itype, |
6617 | fd->loops[i - fd->loops[i].outer].v); |
6618 | tree t3 = fold_convert (itype, fd->loops[i].m2); |
6619 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); |
6620 | t = fold_build2 (PLUS_EXPR, itype, t, t2); |
6621 | } |
6622 | t = fold_build2 (MINUS_EXPR, itype, t, |
6623 | fold_convert (itype, fd->loops[i].v)); |
6624 | if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR) |
6625 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
6626 | fold_build1 (NEGATE_EXPR, itype, t), |
6627 | fold_build1 (NEGATE_EXPR, itype, |
6628 | fold_convert (itype, |
6629 | fd->loops[i].step))); |
6630 | else |
6631 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, |
6632 | fold_convert (itype, fd->loops[i].step)); |
6633 | t = fold_convert (type, t); |
6634 | tree t2 = fold_build2 (MINUS_EXPR, type, n2, n1); |
6635 | min_arg1 = create_tmp_var (type); |
6636 | expand_omp_build_assign (gsi_p: &gsi, to: min_arg1, from: t2); |
6637 | min_arg2 = create_tmp_var (type); |
6638 | expand_omp_build_assign (gsi_p: &gsi, to: min_arg2, from: t); |
6639 | } |
6640 | else |
6641 | { |
6642 | if (TREE_CODE (n2) == INTEGER_CST) |
6643 | { |
6644 | /* Indicate for lastprivate handling that at least one iteration |
6645 | has been performed, without wasting runtime. */ |
6646 | if (integer_nonzerop (n2)) |
6647 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, |
6648 | fold_convert (type, n2)); |
6649 | else |
6650 | /* Indicate that no iteration has been performed. */ |
6651 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, |
6652 | from: build_one_cst (type)); |
6653 | } |
6654 | else |
6655 | { |
6656 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, |
6657 | from: build_zero_cst (type)); |
6658 | expand_omp_build_assign (gsi_p: &gsi, to: n2, from: build_one_cst (type)); |
6659 | } |
6660 | for (i = 0; i < fd->collapse; i++) |
6661 | { |
6662 | t = fold_convert (TREE_TYPE (fd->loops[i].v), fd->loops[i].n1); |
6663 | if (fd->loops[i].m1) |
6664 | { |
6665 | tree t2 |
6666 | = fold_convert (TREE_TYPE (t), |
6667 | fd->loops[i - fd->loops[i].outer].v); |
6668 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i].m1); |
6669 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); |
6670 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); |
6671 | } |
6672 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i].v, from: t); |
6673 | /* For normal non-combined collapsed loops just initialize |
6674 | the outermost iterator in the entry_bb. */ |
6675 | if (!broken_loop) |
6676 | break; |
6677 | } |
6678 | } |
6679 | } |
6680 | else |
6681 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, fold_convert (type, n1)); |
6682 | tree altv = NULL_TREE, altn2 = NULL_TREE; |
6683 | if (fd->collapse == 1 |
6684 | && !broken_loop |
6685 | && TREE_CODE (orig_step) != INTEGER_CST) |
6686 | { |
6687 | /* The vectorizer currently punts on loops with non-constant steps |
6688 | for the main IV (can't compute number of iterations and gives up |
6689 | because of that). As for OpenMP loops it is always possible to |
6690 | compute the number of iterations upfront, use an alternate IV |
6691 | as the loop iterator: |
6692 | altn2 = n1 < n2 ? (n2 - n1 + step - 1) / step : 0; |
6693 | for (i = n1, altv = 0; altv < altn2; altv++, i += step) */ |
6694 | altv = create_tmp_var (unsigned_type_for (TREE_TYPE (fd->loops[0].v))); |
6695 | expand_omp_build_assign (gsi_p: &gsi, to: altv, from: build_zero_cst (TREE_TYPE (altv))); |
6696 | tree itype = TREE_TYPE (fd->loop.v); |
6697 | if (POINTER_TYPE_P (itype)) |
6698 | itype = signed_type_for (itype); |
6699 | t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1)); |
6700 | t = fold_build2 (PLUS_EXPR, itype, |
6701 | fold_convert (itype, step), t); |
6702 | t = fold_build2 (PLUS_EXPR, itype, t, fold_convert (itype, n2)); |
6703 | t = fold_build2 (MINUS_EXPR, itype, t, |
6704 | fold_convert (itype, fd->loop.v)); |
6705 | if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR) |
6706 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
6707 | fold_build1 (NEGATE_EXPR, itype, t), |
6708 | fold_build1 (NEGATE_EXPR, itype, |
6709 | fold_convert (itype, step))); |
6710 | else |
6711 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, |
6712 | fold_convert (itype, step)); |
6713 | t = fold_convert (TREE_TYPE (altv), t); |
6714 | altn2 = create_tmp_var (TREE_TYPE (altv)); |
6715 | expand_omp_build_assign (gsi_p: &gsi, to: altn2, from: t); |
6716 | tree t2 = fold_convert (TREE_TYPE (fd->loop.v), n2); |
6717 | t2 = fold_build2 (fd->loop.cond_code, boolean_type_node, fd->loop.v, t2); |
6718 | t2 = force_gimple_operand_gsi (&gsi, t2, true, NULL_TREE, |
6719 | true, GSI_SAME_STMT); |
6720 | gassign *g = gimple_build_assign (altn2, COND_EXPR, t2, altn2, |
6721 | build_zero_cst (TREE_TYPE (altv))); |
6722 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
6723 | } |
6724 | else if (fd->collapse > 1 |
6725 | && !broken_loop |
6726 | && !gimple_omp_for_combined_into_p (g: fd->for_stmt) |
6727 | && TREE_CODE (fd->loops[fd->collapse - 1].step) != INTEGER_CST) |
6728 | { |
6729 | altv = create_tmp_var (unsigned_type_for (TREE_TYPE (fd->loops[0].v))); |
6730 | altn2 = create_tmp_var (TREE_TYPE (altv)); |
6731 | } |
6732 | if (cond_var) |
6733 | { |
6734 | if (POINTER_TYPE_P (type) |
6735 | || TREE_CODE (n1) != INTEGER_CST |
6736 | || fd->loop.cond_code != LT_EXPR |
6737 | || tree_int_cst_sgn (n1) != 1) |
6738 | expand_omp_build_assign (gsi_p: &gsi, to: cond_var, |
6739 | from: build_one_cst (TREE_TYPE (cond_var))); |
6740 | else |
6741 | expand_omp_build_assign (gsi_p: &gsi, to: cond_var, |
6742 | fold_convert (TREE_TYPE (cond_var), n1)); |
6743 | } |
6744 | |
6745 | /* Remove the GIMPLE_OMP_FOR statement. */ |
6746 | gsi_remove (&gsi, true); |
6747 | |
6748 | if (!broken_loop) |
6749 | { |
6750 | /* Code to control the increment goes in the CONT_BB. */ |
6751 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
6752 | stmt = gsi_stmt (i: gsi); |
6753 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE); |
6754 | |
6755 | if (fd->collapse == 1 |
6756 | || gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
6757 | { |
6758 | if (POINTER_TYPE_P (type)) |
6759 | t = fold_build_pointer_plus (fd->loop.v, step); |
6760 | else |
6761 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); |
6762 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, from: t); |
6763 | } |
6764 | else if (TREE_CODE (n2) != INTEGER_CST) |
6765 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, from: build_one_cst (type)); |
6766 | if (altv) |
6767 | { |
6768 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (altv), altv, |
6769 | build_one_cst (TREE_TYPE (altv))); |
6770 | expand_omp_build_assign (gsi_p: &gsi, to: altv, from: t); |
6771 | } |
6772 | |
6773 | if (fd->collapse > 1) |
6774 | { |
6775 | i = fd->collapse - 1; |
6776 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) |
6777 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); |
6778 | else |
6779 | { |
6780 | t = fold_convert (TREE_TYPE (fd->loops[i].v), |
6781 | fd->loops[i].step); |
6782 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), |
6783 | fd->loops[i].v, t); |
6784 | } |
6785 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i].v, from: t); |
6786 | } |
6787 | if (cond_var) |
6788 | { |
6789 | if (POINTER_TYPE_P (type) |
6790 | || TREE_CODE (n1) != INTEGER_CST |
6791 | || fd->loop.cond_code != LT_EXPR |
6792 | || tree_int_cst_sgn (n1) != 1) |
6793 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, |
6794 | build_one_cst (TREE_TYPE (cond_var))); |
6795 | else |
6796 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (cond_var), cond_var, |
6797 | fold_convert (TREE_TYPE (cond_var), step)); |
6798 | expand_omp_build_assign (gsi_p: &gsi, to: cond_var, from: t); |
6799 | } |
6800 | |
6801 | /* Remove GIMPLE_OMP_CONTINUE. */ |
6802 | gsi_remove (&gsi, true); |
6803 | } |
6804 | |
6805 | /* Emit the condition in L1_BB. */ |
6806 | gsi = gsi_start_bb (bb: l1_bb); |
6807 | |
6808 | if (altv) |
6809 | t = build2 (LT_EXPR, boolean_type_node, altv, altn2); |
6810 | else if (fd->collapse > 1 |
6811 | && !gimple_omp_for_combined_into_p (g: fd->for_stmt) |
6812 | && !broken_loop) |
6813 | { |
6814 | i = fd->collapse - 1; |
6815 | tree itype = TREE_TYPE (fd->loops[i].v); |
6816 | if (fd->loops[i].m2) |
6817 | t = n2v = create_tmp_var (itype); |
6818 | else |
6819 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
6820 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6821 | false, GSI_CONTINUE_LINKING); |
6822 | tree v = fd->loops[i].v; |
6823 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) |
6824 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, |
6825 | false, GSI_CONTINUE_LINKING); |
6826 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, t); |
6827 | } |
6828 | else |
6829 | { |
6830 | if (fd->collapse > 1 && !broken_loop) |
6831 | t = n2var; |
6832 | else |
6833 | t = fold_convert (type, unshare_expr (n2)); |
6834 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6835 | false, GSI_CONTINUE_LINKING); |
6836 | tree v = fd->loop.v; |
6837 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) |
6838 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, |
6839 | false, GSI_CONTINUE_LINKING); |
6840 | t = build2 (fd->loop.cond_code, boolean_type_node, v, t); |
6841 | } |
6842 | cond_stmt = gimple_build_cond_empty (cond: t); |
6843 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); |
6844 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), expand_omp_regimplify_p, |
6845 | NULL, NULL) |
6846 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), expand_omp_regimplify_p, |
6847 | NULL, NULL)) |
6848 | { |
6849 | gsi = gsi_for_stmt (cond_stmt); |
6850 | gimple_regimplify_operands (cond_stmt, &gsi); |
6851 | } |
6852 | |
6853 | /* Add 'V -= STEP * (SIMT_VF - 1)' after the loop. */ |
6854 | if (is_simt) |
6855 | { |
6856 | gsi = gsi_start_bb (bb: l2_bb); |
6857 | step = fold_build2 (MINUS_EXPR, TREE_TYPE (step), orig_step, step); |
6858 | if (POINTER_TYPE_P (type)) |
6859 | t = fold_build_pointer_plus (fd->loop.v, step); |
6860 | else |
6861 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, step); |
6862 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loop.v, from: t); |
6863 | } |
6864 | |
6865 | /* Remove GIMPLE_OMP_RETURN. */ |
6866 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
6867 | gsi_remove (&gsi, true); |
6868 | |
6869 | /* Connect the new blocks. */ |
6870 | remove_edge (FALLTHRU_EDGE (entry_bb)); |
6871 | |
6872 | if (!broken_loop) |
6873 | { |
6874 | remove_edge (BRANCH_EDGE (entry_bb)); |
6875 | make_edge (entry_bb, l1_bb, EDGE_FALLTHRU); |
6876 | |
6877 | e = BRANCH_EDGE (l1_bb); |
6878 | ne = FALLTHRU_EDGE (l1_bb); |
6879 | e->flags = EDGE_TRUE_VALUE; |
6880 | } |
6881 | else |
6882 | { |
6883 | single_succ_edge (bb: entry_bb)->flags = EDGE_FALLTHRU; |
6884 | |
6885 | ne = single_succ_edge (bb: l1_bb); |
6886 | e = make_edge (l1_bb, l0_bb, EDGE_TRUE_VALUE); |
6887 | |
6888 | } |
6889 | ne->flags = EDGE_FALSE_VALUE; |
6890 | e->probability = profile_probability::guessed_always ().apply_scale (num: 7, den: 8); |
6891 | ne->probability = e->probability.invert (); |
6892 | |
6893 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, entry_bb); |
6894 | set_immediate_dominator (CDI_DOMINATORS, l0_bb, l1_bb); |
6895 | |
6896 | if (simt_maxlane) |
6897 | { |
6898 | cond_stmt = gimple_build_cond (LT_EXPR, simt_lane, simt_maxlane, |
6899 | NULL_TREE, NULL_TREE); |
6900 | gsi = gsi_last_bb (bb: entry_bb); |
6901 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); |
6902 | make_edge (entry_bb, l2_bb, EDGE_FALSE_VALUE); |
6903 | FALLTHRU_EDGE (entry_bb)->flags = EDGE_TRUE_VALUE; |
6904 | FALLTHRU_EDGE (entry_bb)->probability |
6905 | = profile_probability::guessed_always ().apply_scale (num: 7, den: 8); |
6906 | BRANCH_EDGE (entry_bb)->probability |
6907 | = FALLTHRU_EDGE (entry_bb)->probability.invert (); |
6908 | l2_dom_bb = entry_bb; |
6909 | } |
6910 | set_immediate_dominator (CDI_DOMINATORS, l2_bb, l2_dom_bb); |
6911 | |
6912 | if (!broken_loop && fd->collapse > 1) |
6913 | { |
6914 | basic_block last_bb = l1_bb; |
6915 | basic_block init_bb = NULL; |
6916 | for (i = fd->collapse - 2; i >= 0; i--) |
6917 | { |
6918 | tree nextn2v = NULL_TREE; |
6919 | if (EDGE_SUCC (last_bb, 0)->flags & EDGE_FALSE_VALUE) |
6920 | e = EDGE_SUCC (last_bb, 0); |
6921 | else |
6922 | e = EDGE_SUCC (last_bb, 1); |
6923 | basic_block bb = split_edge (e); |
6924 | if (POINTER_TYPE_P (TREE_TYPE (fd->loops[i].v))) |
6925 | t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step); |
6926 | else |
6927 | { |
6928 | t = fold_convert (TREE_TYPE (fd->loops[i].v), |
6929 | fd->loops[i].step); |
6930 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (fd->loops[i].v), |
6931 | fd->loops[i].v, t); |
6932 | } |
6933 | gsi = gsi_after_labels (bb); |
6934 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i].v, from: t); |
6935 | |
6936 | bb = split_block (bb, last_nondebug_stmt (bb))->dest; |
6937 | gsi = gsi_start_bb (bb); |
6938 | tree itype = TREE_TYPE (fd->loops[i].v); |
6939 | if (fd->loops[i].m2) |
6940 | t = nextn2v = create_tmp_var (itype); |
6941 | else |
6942 | t = fold_convert (itype, unshare_expr (fd->loops[i].n2)); |
6943 | t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
6944 | false, GSI_CONTINUE_LINKING); |
6945 | tree v = fd->loops[i].v; |
6946 | if (DECL_P (v) && TREE_ADDRESSABLE (v)) |
6947 | v = force_gimple_operand_gsi (&gsi, v, true, NULL_TREE, |
6948 | false, GSI_CONTINUE_LINKING); |
6949 | t = build2 (fd->loops[i].cond_code, boolean_type_node, v, t); |
6950 | cond_stmt = gimple_build_cond_empty (cond: t); |
6951 | gsi_insert_after (&gsi, cond_stmt, GSI_CONTINUE_LINKING); |
6952 | if (walk_tree (gimple_cond_lhs_ptr (cond_stmt), |
6953 | expand_omp_regimplify_p, NULL, NULL) |
6954 | || walk_tree (gimple_cond_rhs_ptr (cond_stmt), |
6955 | expand_omp_regimplify_p, NULL, NULL)) |
6956 | { |
6957 | gsi = gsi_for_stmt (cond_stmt); |
6958 | gimple_regimplify_operands (cond_stmt, &gsi); |
6959 | } |
6960 | ne = single_succ_edge (bb); |
6961 | ne->flags = EDGE_FALSE_VALUE; |
6962 | |
6963 | init_bb = create_empty_bb (bb); |
6964 | set_immediate_dominator (CDI_DOMINATORS, init_bb, bb); |
6965 | add_bb_to_loop (init_bb, bb->loop_father); |
6966 | e = make_edge (bb, init_bb, EDGE_TRUE_VALUE); |
6967 | e->probability |
6968 | = profile_probability::guessed_always ().apply_scale (num: 7, den: 8); |
6969 | ne->probability = e->probability.invert (); |
6970 | |
6971 | gsi = gsi_after_labels (bb: init_bb); |
6972 | if (fd->loops[i + 1].m1) |
6973 | { |
6974 | tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
6975 | fd->loops[i + 1 |
6976 | - fd->loops[i + 1].outer].v); |
6977 | if (POINTER_TYPE_P (TREE_TYPE (t2))) |
6978 | t = fold_build_pointer_plus (t2, fd->loops[i + 1].n1); |
6979 | else |
6980 | { |
6981 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
6982 | fd->loops[i + 1].n1); |
6983 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m1); |
6984 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); |
6985 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); |
6986 | } |
6987 | } |
6988 | else |
6989 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
6990 | fd->loops[i + 1].n1); |
6991 | expand_omp_build_assign (gsi_p: &gsi, to: fd->loops[i + 1].v, from: t); |
6992 | if (fd->loops[i + 1].m2) |
6993 | { |
6994 | if (i + 2 == fd->collapse && (n2var || altv)) |
6995 | { |
6996 | gcc_assert (n2v == NULL_TREE); |
6997 | n2v = create_tmp_var (TREE_TYPE (fd->loops[i + 1].v)); |
6998 | } |
6999 | tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
7000 | fd->loops[i + 1 |
7001 | - fd->loops[i + 1].outer].v); |
7002 | if (POINTER_TYPE_P (TREE_TYPE (t2))) |
7003 | t = fold_build_pointer_plus (t2, fd->loops[i + 1].n2); |
7004 | else |
7005 | { |
7006 | t = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
7007 | fd->loops[i + 1].n2); |
7008 | tree t3 = fold_convert (TREE_TYPE (t), fd->loops[i + 1].m2); |
7009 | t2 = fold_build2 (MULT_EXPR, TREE_TYPE (t), t2, t3); |
7010 | t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2); |
7011 | } |
7012 | expand_omp_build_assign (gsi_p: &gsi, to: n2v, from: t); |
7013 | } |
7014 | if (i + 2 == fd->collapse && n2var) |
7015 | { |
7016 | /* For composite simd, n2 is the first iteration the current |
7017 | task shouldn't already handle, so we effectively want to use |
7018 | for (V3 = N31; V < N2 && V3 < N32; V++, V3 += STEP3) |
7019 | as the vectorized loop. Except the vectorizer will not |
7020 | vectorize that, so instead compute N2VAR as |
7021 | N2VAR = V + MIN (N2 - V, COUNTS3) and use |
7022 | for (V3 = N31; V < N2VAR; V++, V3 += STEP3) |
7023 | as the loop to vectorize. */ |
7024 | tree t2 = fold_build2 (MINUS_EXPR, type, n2, fd->loop.v); |
7025 | if (fd->loops[i + 1].m1 || fd->loops[i + 1].m2) |
7026 | { |
7027 | tree itype = TREE_TYPE (fd->loops[i].v); |
7028 | if (POINTER_TYPE_P (itype)) |
7029 | itype = signed_type_for (itype); |
7030 | t = build_int_cst (itype, (fd->loops[i + 1].cond_code |
7031 | == LT_EXPR ? -1 : 1)); |
7032 | t = fold_build2 (PLUS_EXPR, itype, |
7033 | fold_convert (itype, |
7034 | fd->loops[i + 1].step), t); |
7035 | if (fd->loops[i + 1].m2 == NULL_TREE) |
7036 | t = fold_build2 (PLUS_EXPR, itype, t, |
7037 | fold_convert (itype, |
7038 | fd->loops[i + 1].n2)); |
7039 | else if (POINTER_TYPE_P (TREE_TYPE (n2v))) |
7040 | { |
7041 | t = fold_build_pointer_plus (n2v, t); |
7042 | t = fold_convert (itype, t); |
7043 | } |
7044 | else |
7045 | t = fold_build2 (PLUS_EXPR, itype, t, n2v); |
7046 | t = fold_build2 (MINUS_EXPR, itype, t, |
7047 | fold_convert (itype, fd->loops[i + 1].v)); |
7048 | tree step = fold_convert (itype, fd->loops[i + 1].step); |
7049 | if (TYPE_UNSIGNED (itype) |
7050 | && fd->loops[i + 1].cond_code == GT_EXPR) |
7051 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
7052 | fold_build1 (NEGATE_EXPR, itype, t), |
7053 | fold_build1 (NEGATE_EXPR, itype, step)); |
7054 | else |
7055 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
7056 | t = fold_convert (type, t); |
7057 | } |
7058 | else |
7059 | t = counts[i + 1]; |
7060 | expand_omp_build_assign (gsi_p: &gsi, to: min_arg1, from: t2); |
7061 | expand_omp_build_assign (gsi_p: &gsi, to: min_arg2, from: t); |
7062 | e = split_block (init_bb, last_nondebug_stmt (init_bb)); |
7063 | gsi = gsi_after_labels (bb: e->dest); |
7064 | init_bb = e->dest; |
7065 | remove_edge (FALLTHRU_EDGE (entry_bb)); |
7066 | make_edge (entry_bb, init_bb, EDGE_FALLTHRU); |
7067 | set_immediate_dominator (CDI_DOMINATORS, init_bb, entry_bb); |
7068 | set_immediate_dominator (CDI_DOMINATORS, l1_bb, init_bb); |
7069 | t = fold_build2 (MIN_EXPR, type, min_arg1, min_arg2); |
7070 | t = fold_build2 (PLUS_EXPR, type, fd->loop.v, t); |
7071 | expand_omp_build_assign (gsi_p: &gsi, to: n2var, from: t); |
7072 | } |
7073 | if (i + 2 == fd->collapse && altv) |
7074 | { |
7075 | /* The vectorizer currently punts on loops with non-constant |
7076 | steps for the main IV (can't compute number of iterations |
7077 | and gives up because of that). As for OpenMP loops it is |
7078 | always possible to compute the number of iterations upfront, |
7079 | use an alternate IV as the loop iterator. */ |
7080 | expand_omp_build_assign (gsi_p: &gsi, to: altv, |
7081 | from: build_zero_cst (TREE_TYPE (altv))); |
7082 | tree itype = TREE_TYPE (fd->loops[i + 1].v); |
7083 | if (POINTER_TYPE_P (itype)) |
7084 | itype = signed_type_for (itype); |
7085 | t = build_int_cst (itype, (fd->loops[i + 1].cond_code == LT_EXPR |
7086 | ? -1 : 1)); |
7087 | t = fold_build2 (PLUS_EXPR, itype, |
7088 | fold_convert (itype, fd->loops[i + 1].step), t); |
7089 | t = fold_build2 (PLUS_EXPR, itype, t, |
7090 | fold_convert (itype, |
7091 | fd->loops[i + 1].m2 |
7092 | ? n2v : fd->loops[i + 1].n2)); |
7093 | t = fold_build2 (MINUS_EXPR, itype, t, |
7094 | fold_convert (itype, fd->loops[i + 1].v)); |
7095 | tree step = fold_convert (itype, fd->loops[i + 1].step); |
7096 | if (TYPE_UNSIGNED (itype) |
7097 | && fd->loops[i + 1].cond_code == GT_EXPR) |
7098 | t = fold_build2 (TRUNC_DIV_EXPR, itype, |
7099 | fold_build1 (NEGATE_EXPR, itype, t), |
7100 | fold_build1 (NEGATE_EXPR, itype, step)); |
7101 | else |
7102 | t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step); |
7103 | t = fold_convert (TREE_TYPE (altv), t); |
7104 | expand_omp_build_assign (gsi_p: &gsi, to: altn2, from: t); |
7105 | tree t2 = fold_convert (TREE_TYPE (fd->loops[i + 1].v), |
7106 | fd->loops[i + 1].m2 |
7107 | ? n2v : fd->loops[i + 1].n2); |
7108 | t2 = fold_build2 (fd->loops[i + 1].cond_code, boolean_type_node, |
7109 | fd->loops[i + 1].v, t2); |
7110 | t2 = force_gimple_operand_gsi (&gsi, t2, true, NULL_TREE, |
7111 | true, GSI_SAME_STMT); |
7112 | gassign *g |
7113 | = gimple_build_assign (altn2, COND_EXPR, t2, altn2, |
7114 | build_zero_cst (TREE_TYPE (altv))); |
7115 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
7116 | } |
7117 | n2v = nextn2v; |
7118 | |
7119 | make_edge (init_bb, last_bb, EDGE_FALLTHRU); |
7120 | if (!gimple_omp_for_combined_into_p (g: fd->for_stmt)) |
7121 | { |
7122 | e = find_edge (entry_bb, last_bb); |
7123 | redirect_edge_succ (e, bb); |
7124 | set_immediate_dominator (CDI_DOMINATORS, bb, entry_bb); |
7125 | set_immediate_dominator (CDI_DOMINATORS, last_bb, init_bb); |
7126 | } |
7127 | |
7128 | last_bb = bb; |
7129 | } |
7130 | } |
7131 | if (!broken_loop) |
7132 | { |
7133 | class loop *loop = alloc_loop (); |
7134 | loop->header = l1_bb; |
7135 | loop->latch = cont_bb; |
7136 | add_loop (loop, l1_bb->loop_father); |
7137 | loop->safelen = safelen_int; |
7138 | if (simduid) |
7139 | { |
7140 | loop->simduid = OMP_CLAUSE__SIMDUID__DECL (simduid); |
7141 | cfun->has_simduid_loops = true; |
7142 | } |
7143 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize |
7144 | the loop. */ |
7145 | if ((flag_tree_loop_vectorize |
7146 | || !OPTION_SET_P (flag_tree_loop_vectorize)) |
7147 | && flag_tree_loop_optimize |
7148 | && loop->safelen > 1) |
7149 | { |
7150 | loop->force_vectorize = true; |
7151 | if (simdlen && tree_fits_uhwi_p (OMP_CLAUSE_SIMDLEN_EXPR (simdlen))) |
7152 | { |
7153 | unsigned HOST_WIDE_INT v |
7154 | = tree_to_uhwi (OMP_CLAUSE_SIMDLEN_EXPR (simdlen)); |
7155 | if (v < INT_MAX && v <= (unsigned HOST_WIDE_INT) loop->safelen) |
7156 | loop->simdlen = v; |
7157 | } |
7158 | cfun->has_force_vectorize_loops = true; |
7159 | } |
7160 | else if (dont_vectorize) |
7161 | loop->dont_vectorize = true; |
7162 | } |
7163 | else if (simduid) |
7164 | cfun->has_simduid_loops = true; |
7165 | } |
7166 | |
7167 | /* Taskloop construct is represented after gimplification with |
7168 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched |
7169 | in between them. This routine expands the outer GIMPLE_OMP_FOR, |
7170 | which should just compute all the needed loop temporaries |
7171 | for GIMPLE_OMP_TASK. */ |
7172 | |
7173 | static void |
7174 | expand_omp_taskloop_for_outer (struct omp_region *region, |
7175 | struct omp_for_data *fd, |
7176 | gimple *inner_stmt) |
7177 | { |
7178 | tree type, bias = NULL_TREE; |
7179 | basic_block entry_bb, cont_bb, exit_bb; |
7180 | gimple_stmt_iterator gsi; |
7181 | gassign *assign_stmt; |
7182 | tree *counts = NULL; |
7183 | int i; |
7184 | |
7185 | gcc_assert (inner_stmt); |
7186 | gcc_assert (region->cont); |
7187 | gcc_assert (gimple_code (inner_stmt) == GIMPLE_OMP_TASK |
7188 | && gimple_omp_task_taskloop_p (inner_stmt)); |
7189 | type = TREE_TYPE (fd->loop.v); |
7190 | |
7191 | /* See if we need to bias by LLONG_MIN. */ |
7192 | if (fd->iter_type == long_long_unsigned_type_node |
7193 | && TREE_CODE (type) == INTEGER_TYPE |
7194 | && !TYPE_UNSIGNED (type)) |
7195 | { |
7196 | tree n1, n2; |
7197 | |
7198 | if (fd->loop.cond_code == LT_EXPR) |
7199 | { |
7200 | n1 = fd->loop.n1; |
7201 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); |
7202 | } |
7203 | else |
7204 | { |
7205 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); |
7206 | n2 = fd->loop.n1; |
7207 | } |
7208 | if (TREE_CODE (n1) != INTEGER_CST |
7209 | || TREE_CODE (n2) != INTEGER_CST |
7210 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) |
7211 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); |
7212 | } |
7213 | |
7214 | entry_bb = region->entry; |
7215 | cont_bb = region->cont; |
7216 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
7217 | gcc_assert (BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest); |
7218 | exit_bb = region->exit; |
7219 | |
7220 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
7221 | gimple *for_stmt = gsi_stmt (i: gsi); |
7222 | gcc_assert (gimple_code (for_stmt) == GIMPLE_OMP_FOR); |
7223 | if (fd->collapse > 1) |
7224 | { |
7225 | int first_zero_iter = -1, dummy = -1; |
7226 | basic_block zero_iter_bb = NULL, dummy_bb = NULL, l2_dom_bb = NULL; |
7227 | |
7228 | counts = XALLOCAVEC (tree, fd->collapse); |
7229 | expand_omp_for_init_counts (fd, gsi: &gsi, entry_bb, counts, |
7230 | zero_iter1_bb&: zero_iter_bb, first_zero_iter1&: first_zero_iter, |
7231 | zero_iter2_bb&: dummy_bb, first_zero_iter2&: dummy, l2_dom_bb); |
7232 | |
7233 | if (zero_iter_bb) |
7234 | { |
7235 | /* Some counts[i] vars might be uninitialized if |
7236 | some loop has zero iterations. But the body shouldn't |
7237 | be executed in that case, so just avoid uninit warnings. */ |
7238 | for (i = first_zero_iter; i < fd->collapse; i++) |
7239 | if (SSA_VAR_P (counts[i])) |
7240 | suppress_warning (counts[i], OPT_Wuninitialized); |
7241 | gsi_prev (i: &gsi); |
7242 | edge e = split_block (entry_bb, gsi_stmt (i: gsi)); |
7243 | entry_bb = e->dest; |
7244 | make_edge (zero_iter_bb, entry_bb, EDGE_FALLTHRU); |
7245 | gsi = gsi_last_bb (bb: entry_bb); |
7246 | set_immediate_dominator (CDI_DOMINATORS, entry_bb, |
7247 | get_immediate_dominator (CDI_DOMINATORS, |
7248 | zero_iter_bb)); |
7249 | } |
7250 | } |
7251 | |
7252 | tree t0, t1; |
7253 | t1 = fd->loop.n2; |
7254 | t0 = fd->loop.n1; |
7255 | if (POINTER_TYPE_P (TREE_TYPE (t0)) |
7256 | && TYPE_PRECISION (TREE_TYPE (t0)) |
7257 | != TYPE_PRECISION (fd->iter_type)) |
7258 | { |
7259 | /* Avoid casting pointers to integer of a different size. */ |
7260 | tree itype = signed_type_for (type); |
7261 | t1 = fold_convert (fd->iter_type, fold_convert (itype, t1)); |
7262 | t0 = fold_convert (fd->iter_type, fold_convert (itype, t0)); |
7263 | } |
7264 | else |
7265 | { |
7266 | t1 = fold_convert (fd->iter_type, t1); |
7267 | t0 = fold_convert (fd->iter_type, t0); |
7268 | } |
7269 | if (bias) |
7270 | { |
7271 | t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias); |
7272 | t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias); |
7273 | } |
7274 | |
7275 | tree innerc = omp_find_clause (clauses: gimple_omp_task_clauses (gs: inner_stmt), |
7276 | kind: OMP_CLAUSE__LOOPTEMP_); |
7277 | gcc_assert (innerc); |
7278 | tree startvar = OMP_CLAUSE_DECL (innerc); |
7279 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), kind: OMP_CLAUSE__LOOPTEMP_); |
7280 | gcc_assert (innerc); |
7281 | tree endvar = OMP_CLAUSE_DECL (innerc); |
7282 | if (fd->collapse > 1 && TREE_CODE (fd->loop.n2) != INTEGER_CST) |
7283 | { |
7284 | innerc = find_lastprivate_looptemp (fd, innerc); |
7285 | if (innerc) |
7286 | { |
7287 | /* If needed (inner taskloop has lastprivate clause), propagate |
7288 | down the total number of iterations. */ |
7289 | tree t = force_gimple_operand_gsi (&gsi, fd->loop.n2, false, |
7290 | NULL_TREE, false, |
7291 | GSI_CONTINUE_LINKING); |
7292 | assign_stmt = gimple_build_assign (OMP_CLAUSE_DECL (innerc), t); |
7293 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
7294 | } |
7295 | } |
7296 | |
7297 | t0 = force_gimple_operand_gsi (&gsi, t0, false, NULL_TREE, false, |
7298 | GSI_CONTINUE_LINKING); |
7299 | assign_stmt = gimple_build_assign (startvar, t0); |
7300 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
7301 | |
7302 | t1 = force_gimple_operand_gsi (&gsi, t1, false, NULL_TREE, false, |
7303 | GSI_CONTINUE_LINKING); |
7304 | assign_stmt = gimple_build_assign (endvar, t1); |
7305 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
7306 | if (fd->collapse > 1) |
7307 | expand_omp_for_init_vars (fd, gsi: &gsi, counts, NULL, inner_stmt, startvar); |
7308 | |
7309 | /* Remove the GIMPLE_OMP_FOR statement. */ |
7310 | gsi = gsi_for_stmt (for_stmt); |
7311 | gsi_remove (&gsi, true); |
7312 | |
7313 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
7314 | gsi_remove (&gsi, true); |
7315 | |
7316 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
7317 | gsi_remove (&gsi, true); |
7318 | |
7319 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); |
7320 | remove_edge (BRANCH_EDGE (entry_bb)); |
7321 | FALLTHRU_EDGE (cont_bb)->probability = profile_probability::always (); |
7322 | remove_edge (BRANCH_EDGE (cont_bb)); |
7323 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, cont_bb); |
7324 | set_immediate_dominator (CDI_DOMINATORS, region->entry, |
7325 | recompute_dominator (CDI_DOMINATORS, region->entry)); |
7326 | } |
7327 | |
7328 | /* Taskloop construct is represented after gimplification with |
7329 | two GIMPLE_OMP_FOR constructs with GIMPLE_OMP_TASK sandwiched |
7330 | in between them. This routine expands the inner GIMPLE_OMP_FOR. |
7331 | GOMP_taskloop{,_ull} function arranges for each task to be given just |
7332 | a single range of iterations. */ |
7333 | |
7334 | static void |
7335 | expand_omp_taskloop_for_inner (struct omp_region *region, |
7336 | struct omp_for_data *fd, |
7337 | gimple *inner_stmt) |
7338 | { |
7339 | tree e, t, type, itype, vmain, vback, bias = NULL_TREE; |
7340 | basic_block entry_bb, exit_bb, body_bb, cont_bb, collapse_bb = NULL; |
7341 | basic_block fin_bb; |
7342 | gimple_stmt_iterator gsi; |
7343 | edge ep; |
7344 | bool broken_loop = region->cont == NULL; |
7345 | tree *counts = NULL; |
7346 | tree n1, n2, step; |
7347 | |
7348 | itype = type = TREE_TYPE (fd->loop.v); |
7349 | if (POINTER_TYPE_P (type)) |
7350 | itype = signed_type_for (type); |
7351 | |
7352 | /* See if we need to bias by LLONG_MIN. */ |
7353 | if (fd->iter_type == long_long_unsigned_type_node |
7354 | && TREE_CODE (type) == INTEGER_TYPE |
7355 | && !TYPE_UNSIGNED (type)) |
7356 | { |
7357 | tree n1, n2; |
7358 | |
7359 | if (fd->loop.cond_code == LT_EXPR) |
7360 | { |
7361 | n1 = fd->loop.n1; |
7362 | n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step); |
7363 | } |
7364 | else |
7365 | { |
7366 | n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step); |
7367 | n2 = fd->loop.n1; |
7368 | } |
7369 | if (TREE_CODE (n1) != INTEGER_CST |
7370 | || TREE_CODE (n2) != INTEGER_CST |
7371 | || ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0))) |
7372 | bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type)); |
7373 | } |
7374 | |
7375 | entry_bb = region->entry; |
7376 | cont_bb = region->cont; |
7377 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2); |
7378 | fin_bb = BRANCH_EDGE (entry_bb)->dest; |
7379 | gcc_assert (broken_loop |
7380 | || (fin_bb == FALLTHRU_EDGE (cont_bb)->dest)); |
7381 | body_bb = FALLTHRU_EDGE (entry_bb)->dest; |
7382 | if (!broken_loop) |
7383 | { |
7384 | gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb); |
7385 | gcc_assert (EDGE_COUNT (cont_bb->succs) == 2); |
7386 | } |
7387 | exit_bb = region->exit; |
7388 | |
7389 | /* Iteration space partitioning goes in ENTRY_BB. */ |
7390 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
7391 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR); |
7392 | |
7393 | if (fd->collapse > 1) |
7394 | { |
7395 | int first_zero_iter = -1, dummy = -1; |
7396 | basic_block l2_dom_bb = NULL, dummy_bb = NULL; |
7397 | |
7398 | counts = XALLOCAVEC (tree, fd->collapse); |
7399 | expand_omp_for_init_counts (fd, gsi: &gsi, entry_bb, counts, |
7400 | zero_iter1_bb&: fin_bb, first_zero_iter1&: first_zero_iter, |
7401 | zero_iter2_bb&: dummy_bb, first_zero_iter2&: dummy, l2_dom_bb); |
7402 | t = NULL_TREE; |
7403 | } |
7404 | else |
7405 | t = integer_one_node; |
7406 | |
7407 | step = fd->loop.step; |
7408 | tree innerc = omp_find_clause (clauses: gimple_omp_for_clauses (gs: fd->for_stmt), |
7409 | kind: OMP_CLAUSE__LOOPTEMP_); |
7410 | gcc_assert (innerc); |
7411 | n1 = OMP_CLAUSE_DECL (innerc); |
7412 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), kind: OMP_CLAUSE__LOOPTEMP_); |
7413 | gcc_assert (innerc); |
7414 | n2 = OMP_CLAUSE_DECL (innerc); |
7415 | if (bias) |
7416 | { |
7417 | n1 = fold_build2 (PLUS_EXPR, fd->iter_type, n1, bias); |
7418 | n2 = fold_build2 (PLUS_EXPR, fd->iter_type, n2, bias); |
7419 | } |
7420 | n1 = force_gimple_operand_gsi (&gsi, fold_convert (type, n1), |
7421 | true, NULL_TREE, true, GSI_SAME_STMT); |
7422 | n2 = force_gimple_operand_gsi (&gsi, fold_convert (itype, n2), |
7423 | true, NULL_TREE, true, GSI_SAME_STMT); |
7424 | step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step), |
7425 | true, NULL_TREE, true, GSI_SAME_STMT); |
7426 | |
7427 | tree startvar = fd->loop.v; |
7428 | tree endvar = NULL_TREE; |
7429 | |
7430 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
7431 | { |
7432 | tree clauses = gimple_omp_for_clauses (gs: inner_stmt); |
7433 | tree innerc = omp_find_clause (clauses, kind: OMP_CLAUSE__LOOPTEMP_); |
7434 | gcc_assert (innerc); |
7435 | startvar = OMP_CLAUSE_DECL (innerc); |
7436 | innerc = omp_find_clause (OMP_CLAUSE_CHAIN (innerc), |
7437 | kind: OMP_CLAUSE__LOOPTEMP_); |
7438 | gcc_assert (innerc); |
7439 | endvar = OMP_CLAUSE_DECL (innerc); |
7440 | } |
7441 | t = fold_convert (TREE_TYPE (startvar), n1); |
7442 | t = force_gimple_operand_gsi (&gsi, t, |
7443 | DECL_P (startvar) |
7444 | && TREE_ADDRESSABLE (startvar), |
7445 | NULL_TREE, false, GSI_CONTINUE_LINKING); |
7446 | gimple *assign_stmt = gimple_build_assign (startvar, t); |
7447 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
7448 | |
7449 | t = fold_convert (TREE_TYPE (startvar), n2); |
7450 | e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, |
7451 | false, GSI_CONTINUE_LINKING); |
7452 | if (endvar) |
7453 | { |
7454 | assign_stmt = gimple_build_assign (endvar, e); |
7455 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
7456 | if (useless_type_conversion_p (TREE_TYPE (fd->loop.v), TREE_TYPE (e))) |
7457 | assign_stmt = gimple_build_assign (fd->loop.v, e); |
7458 | else |
7459 | assign_stmt = gimple_build_assign (fd->loop.v, NOP_EXPR, e); |
7460 | gsi_insert_after (&gsi, assign_stmt, GSI_CONTINUE_LINKING); |
7461 | } |
7462 | |
7463 | tree *nonrect_bounds = NULL; |
7464 | if (fd->collapse > 1) |
7465 | { |
7466 | if (fd->non_rect) |
7467 | { |
7468 | nonrect_bounds = XALLOCAVEC (tree, fd->last_nonrect + 1); |
7469 | memset (s: nonrect_bounds, c: 0, n: sizeof (tree) * (fd->last_nonrect + 1)); |
7470 | } |
7471 | gcc_assert (gsi_bb (gsi) == entry_bb); |
7472 | expand_omp_for_init_vars (fd, gsi: &gsi, counts, nonrect_bounds, inner_stmt, |
7473 | startvar); |
7474 | entry_bb = gsi_bb (i: gsi); |
7475 | } |
7476 | |
7477 | if (!broken_loop) |
7478 | { |
7479 | /* The code controlling the sequential loop replaces the |
7480 | GIMPLE_OMP_CONTINUE. */ |
7481 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
7482 | gomp_continue *cont_stmt = as_a <gomp_continue *> (p: gsi_stmt (i: gsi)); |
7483 | gcc_assert (gimple_code (cont_stmt) == GIMPLE_OMP_CONTINUE); |
7484 | vmain = gimple_omp_continue_control_use (cont_stmt); |
7485 | vback = gimple_omp_continue_control_def (cont_stmt); |
7486 | |
7487 | if (!gimple_omp_for_combined_p (g: fd->for_stmt)) |
7488 | { |
7489 | if (POINTER_TYPE_P (type)) |
7490 | t = fold_build_pointer_plus (vmain, step); |
7491 | else |
7492 | t = fold_build2 (PLUS_EXPR, type, vmain, step); |
7493 | t = force_gimple_operand_gsi (&gsi, t, |
7494 | DECL_P (vback) |
7495 | && TREE_ADDRESSABLE (vback), |
7496 | NULL_TREE, true, GSI_SAME_STMT); |
7497 | assign_stmt = gimple_build_assign (vback, t); |
7498 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
7499 | |
7500 | t = build2 (fd->loop.cond_code, boolean_type_node, |
7501 | DECL_P (vback) && TREE_ADDRESSABLE (vback) |
7502 | ? t : vback, e); |
7503 | gsi_insert_before (&gsi, gimple_build_cond_empty (cond: t), GSI_SAME_STMT); |
7504 | } |
7505 | |
7506 | /* Remove the GIMPLE_OMP_CONTINUE statement. */ |
7507 | gsi_remove (&gsi, true); |
7508 | |
7509 | if (fd->collapse > 1 && !gimple_omp_for_combined_p (g: fd->for_stmt)) |
7510 | collapse_bb = extract_omp_for_update_vars (fd, nonrect_bounds, |
7511 | cont_bb, body_bb); |
7512 | } |
7513 | |
7514 | /* Remove the GIMPLE_OMP_FOR statement. */ |
7515 | gsi = gsi_for_stmt (fd->for_stmt); |
7516 | gsi_remove (&gsi, true); |
7517 | |
7518 | /* Remove the GIMPLE_OMP_RETURN statement. */ |
7519 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
7520 | gsi_remove (&gsi, true); |
7521 | |
7522 | FALLTHRU_EDGE (entry_bb)->probability = profile_probability::always (); |
7523 | if (!broken_loop) |
7524 | remove_edge (BRANCH_EDGE (entry_bb)); |
7525 | else |
7526 | { |
7527 | remove_edge_and_dominated_blocks (BRANCH_EDGE (entry_bb)); |
7528 | region->outer->cont = NULL; |
7529 | } |
7530 | |
7531 | /* Connect all the blocks. */ |
7532 | if (!broken_loop) |
7533 | { |
7534 | ep = find_edge (cont_bb, body_bb); |
7535 | if (gimple_omp_for_combined_p (g: fd->for_stmt)) |
7536 | { |
7537 | remove_edge (ep); |
7538 | ep = NULL; |
7539 | } |
7540 | else if (fd->collapse > 1) |
7541 | { |
7542 | remove_edge (ep); |
7543 | ep = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE); |
7544 | } |
7545 | else |
7546 | ep->flags = EDGE_TRUE_VALUE; |
7547 | find_edge (cont_bb, fin_bb)->flags |
7548 | = ep ? EDGE_FALSE_VALUE : EDGE_FALLTHRU; |
7549 | } |
7550 | |
7551 | set_immediate_dominator (CDI_DOMINATORS, body_bb, |
7552 | recompute_dominator (CDI_DOMINATORS, body_bb)); |
7553 | if (!broken_loop) |
7554 | set_immediate_dominator (CDI_DOMINATORS, fin_bb, |
7555 | recompute_dominator (CDI_DOMINATORS, fin_bb)); |
7556 | |
7557 | if (!broken_loop && !gimple_omp_for_combined_p (g: fd->for_stmt)) |
7558 | { |
7559 | class loop *loop = alloc_loop (); |
7560 | loop->header = body_bb; |
7561 | if (collapse_bb == NULL) |
7562 | loop->latch = cont_bb; |
7563 | add_loop (loop, body_bb->loop_father); |
7564 | } |
7565 | } |
7566 | |
7567 | /* A subroutine of expand_omp_for. Generate code for an OpenACC |
7568 | partitioned loop. The lowering here is abstracted, in that the |
7569 | loop parameters are passed through internal functions, which are |
7570 | further lowered by oacc_device_lower, once we get to the target |
7571 | compiler. The loop is of the form: |
7572 | |
7573 | for (V = B; V LTGT E; V += S) {BODY} |
7574 | |
7575 | where LTGT is < or >. We may have a specified chunking size, CHUNKING |
7576 | (constant 0 for no chunking) and we will have a GWV partitioning |
7577 | mask, specifying dimensions over which the loop is to be |
7578 | partitioned (see note below). We generate code that looks like |
7579 | (this ignores tiling): |
7580 | |
7581 | <entry_bb> [incoming FALL->body, BRANCH->exit] |
7582 | typedef signedintify (typeof (V)) T; // underlying signed integral type |
7583 | T range = E - B; |
7584 | T chunk_no = 0; |
7585 | T DIR = LTGT == '<' ? +1 : -1; |
7586 | T chunk_max = GOACC_LOOP_CHUNK (dir, range, S, CHUNK_SIZE, GWV); |
7587 | T step = GOACC_LOOP_STEP (dir, range, S, CHUNK_SIZE, GWV); |
7588 | |
7589 | <head_bb> [created by splitting end of entry_bb] |
7590 | T offset = GOACC_LOOP_OFFSET (dir, range, S, CHUNK_SIZE, GWV, chunk_no); |
7591 | T bound = GOACC_LOOP_BOUND (dir, range, S, CHUNK_SIZE, GWV, offset); |
7592 | if (!(offset LTGT bound)) goto bottom_bb; |
7593 | |
7594 | <body_bb> [incoming] |
7595 | V = B + offset; |
7596 | {BODY} |
7597 | |
7598 | <cont_bb> [incoming, may == body_bb FALL->exit_bb, BRANCH->body_bb] |
7599 | offset += step; |
7600 | if (offset LTGT bound) goto body_bb; [*] |
7601 | |
7602 | <bottom_bb> [created by splitting start of exit_bb] insert BRANCH->head_bb |
7603 | chunk_no++; |
7604 | if (chunk < chunk_max) goto head_bb; |
7605 | |
7606 | <exit_bb> [incoming] |
7607 | V = B + ((range -/+ 1) / S +/- 1) * S [*] |
7608 | |
7609 | [*] Needed if V live at end of loop. */ |
7610 | |
7611 | static void |
7612 | expand_oacc_for (struct omp_region *region, struct omp_for_data *fd) |
7613 | { |
7614 | bool is_oacc_kernels_parallelized |
7615 | = (lookup_attribute (attr_name: "oacc kernels parallelized" , |
7616 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
7617 | { |
7618 | bool is_oacc_kernels |
7619 | = (lookup_attribute (attr_name: "oacc kernels" , |
7620 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
7621 | if (is_oacc_kernels_parallelized) |
7622 | gcc_checking_assert (is_oacc_kernels); |
7623 | } |
7624 | gcc_assert (gimple_in_ssa_p (cfun) == is_oacc_kernels_parallelized); |
7625 | /* In the following, some of the 'gimple_in_ssa_p (cfun)' conditionals are |
7626 | for SSA specifics, and some are for 'parloops' OpenACC |
7627 | 'kernels'-parallelized specifics. */ |
7628 | |
7629 | tree v = fd->loop.v; |
7630 | enum tree_code cond_code = fd->loop.cond_code; |
7631 | enum tree_code plus_code = PLUS_EXPR; |
7632 | |
7633 | tree chunk_size = integer_minus_one_node; |
7634 | tree gwv = integer_zero_node; |
7635 | tree iter_type = TREE_TYPE (v); |
7636 | tree diff_type = iter_type; |
7637 | tree plus_type = iter_type; |
7638 | struct oacc_collapse *counts = NULL; |
7639 | |
7640 | gcc_checking_assert (gimple_omp_for_kind (fd->for_stmt) |
7641 | == GF_OMP_FOR_KIND_OACC_LOOP); |
7642 | gcc_assert (!gimple_omp_for_combined_into_p (fd->for_stmt)); |
7643 | gcc_assert (cond_code == LT_EXPR || cond_code == GT_EXPR); |
7644 | |
7645 | if (POINTER_TYPE_P (iter_type)) |
7646 | { |
7647 | plus_code = POINTER_PLUS_EXPR; |
7648 | plus_type = sizetype; |
7649 | } |
7650 | for (int ix = fd->collapse; ix--;) |
7651 | { |
7652 | tree diff_type2 = TREE_TYPE (fd->loops[ix].step); |
7653 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (diff_type2)) |
7654 | diff_type = diff_type2; |
7655 | } |
7656 | if (POINTER_TYPE_P (diff_type) || TYPE_UNSIGNED (diff_type)) |
7657 | diff_type = signed_type_for (diff_type); |
7658 | if (TYPE_PRECISION (diff_type) < TYPE_PRECISION (integer_type_node)) |
7659 | diff_type = integer_type_node; |
7660 | |
7661 | basic_block entry_bb = region->entry; /* BB ending in OMP_FOR */ |
7662 | basic_block exit_bb = region->exit; /* BB ending in OMP_RETURN */ |
7663 | basic_block cont_bb = region->cont; /* BB ending in OMP_CONTINUE */ |
7664 | basic_block bottom_bb = NULL; |
7665 | |
7666 | /* entry_bb has two successors; the branch edge is to the exit |
7667 | block, fallthrough edge to body. */ |
7668 | gcc_assert (EDGE_COUNT (entry_bb->succs) == 2 |
7669 | && BRANCH_EDGE (entry_bb)->dest == exit_bb); |
7670 | |
7671 | /* If cont_bb non-NULL, it has 2 successors. The branch successor is |
7672 | body_bb, or to a block whose only successor is the body_bb. Its |
7673 | fallthrough successor is the final block (same as the branch |
7674 | successor of the entry_bb). */ |
7675 | if (cont_bb) |
7676 | { |
7677 | basic_block body_bb = FALLTHRU_EDGE (entry_bb)->dest; |
7678 | basic_block bed = BRANCH_EDGE (cont_bb)->dest; |
7679 | |
7680 | gcc_assert (FALLTHRU_EDGE (cont_bb)->dest == exit_bb); |
7681 | gcc_assert (bed == body_bb || single_succ_edge (bed)->dest == body_bb); |
7682 | } |
7683 | else |
7684 | gcc_assert (!gimple_in_ssa_p (cfun)); |
7685 | |
7686 | /* The exit block only has entry_bb and cont_bb as predecessors. */ |
7687 | gcc_assert (EDGE_COUNT (exit_bb->preds) == 1 + (cont_bb != NULL)); |
7688 | |
7689 | tree chunk_no; |
7690 | tree chunk_max = NULL_TREE; |
7691 | tree bound, offset; |
7692 | tree step = create_tmp_var (diff_type, ".step" ); |
7693 | bool up = cond_code == LT_EXPR; |
7694 | tree dir = build_int_cst (diff_type, up ? +1 : -1); |
7695 | bool chunking = !gimple_in_ssa_p (cfun); |
7696 | bool negating; |
7697 | |
7698 | /* Tiling vars. */ |
7699 | tree tile_size = NULL_TREE; |
7700 | tree element_s = NULL_TREE; |
7701 | tree e_bound = NULL_TREE, e_offset = NULL_TREE, e_step = NULL_TREE; |
7702 | basic_block elem_body_bb = NULL; |
7703 | basic_block elem_cont_bb = NULL; |
7704 | |
7705 | /* SSA instances. */ |
7706 | tree offset_incr = NULL_TREE; |
7707 | tree offset_init = NULL_TREE; |
7708 | |
7709 | gimple_stmt_iterator gsi; |
7710 | gassign *ass; |
7711 | gcall *call; |
7712 | gimple *stmt; |
7713 | tree expr; |
7714 | location_t loc; |
7715 | edge split, be, fte; |
7716 | |
7717 | /* Split the end of entry_bb to create head_bb. */ |
7718 | split = split_block (entry_bb, last_nondebug_stmt (entry_bb)); |
7719 | basic_block head_bb = split->dest; |
7720 | entry_bb = split->src; |
7721 | |
7722 | /* Chunk setup goes at end of entry_bb, replacing the omp_for. */ |
7723 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
7724 | gomp_for *for_stmt = as_a <gomp_for *> (p: gsi_stmt (i: gsi)); |
7725 | loc = gimple_location (g: for_stmt); |
7726 | |
7727 | if (gimple_in_ssa_p (cfun)) |
7728 | { |
7729 | offset_init = gimple_omp_for_index (gs: for_stmt, i: 0); |
7730 | gcc_assert (integer_zerop (fd->loop.n1)); |
7731 | /* The SSA parallelizer does gang parallelism. */ |
7732 | gwv = build_int_cst (integer_type_node, GOMP_DIM_MASK (GOMP_DIM_GANG)); |
7733 | } |
7734 | |
7735 | if (fd->collapse > 1 || fd->tiling) |
7736 | { |
7737 | gcc_assert (!gimple_in_ssa_p (cfun) && up); |
7738 | counts = XALLOCAVEC (struct oacc_collapse, fd->collapse); |
7739 | tree total = expand_oacc_collapse_init (fd, gsi: &gsi, counts, diff_type, |
7740 | TREE_TYPE (fd->loop.n2), loc); |
7741 | |
7742 | if (SSA_VAR_P (fd->loop.n2)) |
7743 | { |
7744 | total = force_gimple_operand_gsi (&gsi, total, false, NULL_TREE, |
7745 | true, GSI_SAME_STMT); |
7746 | ass = gimple_build_assign (fd->loop.n2, total); |
7747 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
7748 | } |
7749 | } |
7750 | |
7751 | tree b = fd->loop.n1; |
7752 | tree e = fd->loop.n2; |
7753 | tree s = fd->loop.step; |
7754 | |
7755 | b = force_gimple_operand_gsi (&gsi, b, true, NULL_TREE, true, GSI_SAME_STMT); |
7756 | e = force_gimple_operand_gsi (&gsi, e, true, NULL_TREE, true, GSI_SAME_STMT); |
7757 | |
7758 | /* Convert the step, avoiding possible unsigned->signed overflow. */ |
7759 | negating = !up && TYPE_UNSIGNED (TREE_TYPE (s)); |
7760 | if (negating) |
7761 | s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s); |
7762 | s = fold_convert (diff_type, s); |
7763 | if (negating) |
7764 | s = fold_build1 (NEGATE_EXPR, diff_type, s); |
7765 | s = force_gimple_operand_gsi (&gsi, s, true, NULL_TREE, true, GSI_SAME_STMT); |
7766 | |
7767 | if (!chunking) |
7768 | chunk_size = integer_zero_node; |
7769 | expr = fold_convert (diff_type, chunk_size); |
7770 | chunk_size = force_gimple_operand_gsi (&gsi, expr, true, |
7771 | NULL_TREE, true, GSI_SAME_STMT); |
7772 | |
7773 | if (fd->tiling) |
7774 | { |
7775 | /* Determine the tile size and element step, |
7776 | modify the outer loop step size. */ |
7777 | tile_size = create_tmp_var (diff_type, ".tile_size" ); |
7778 | expr = build_int_cst (diff_type, 1); |
7779 | for (int ix = 0; ix < fd->collapse; ix++) |
7780 | expr = fold_build2 (MULT_EXPR, diff_type, counts[ix].tile, expr); |
7781 | expr = force_gimple_operand_gsi (&gsi, expr, true, |
7782 | NULL_TREE, true, GSI_SAME_STMT); |
7783 | ass = gimple_build_assign (tile_size, expr); |
7784 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
7785 | |
7786 | element_s = create_tmp_var (diff_type, ".element_s" ); |
7787 | ass = gimple_build_assign (element_s, s); |
7788 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
7789 | |
7790 | expr = fold_build2 (MULT_EXPR, diff_type, s, tile_size); |
7791 | s = force_gimple_operand_gsi (&gsi, expr, true, |
7792 | NULL_TREE, true, GSI_SAME_STMT); |
7793 | } |
7794 | |
7795 | /* Determine the range, avoiding possible unsigned->signed overflow. */ |
7796 | negating = !up && TYPE_UNSIGNED (iter_type); |
7797 | expr = fold_build2 (MINUS_EXPR, plus_type, |
7798 | fold_convert (plus_type, negating ? b : e), |
7799 | fold_convert (plus_type, negating ? e : b)); |
7800 | expr = fold_convert (diff_type, expr); |
7801 | if (negating) |
7802 | expr = fold_build1 (NEGATE_EXPR, diff_type, expr); |
7803 | tree range = force_gimple_operand_gsi (&gsi, expr, true, |
7804 | NULL_TREE, true, GSI_SAME_STMT); |
7805 | |
7806 | chunk_no = build_int_cst (diff_type, 0); |
7807 | if (chunking) |
7808 | { |
7809 | gcc_assert (!gimple_in_ssa_p (cfun)); |
7810 | |
7811 | expr = chunk_no; |
7812 | chunk_max = create_tmp_var (diff_type, ".chunk_max" ); |
7813 | chunk_no = create_tmp_var (diff_type, ".chunk_no" ); |
7814 | |
7815 | ass = gimple_build_assign (chunk_no, expr); |
7816 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
7817 | |
7818 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, |
7819 | build_int_cst (integer_type_node, |
7820 | IFN_GOACC_LOOP_CHUNKS), |
7821 | dir, range, s, chunk_size, gwv); |
7822 | gimple_call_set_lhs (gs: call, lhs: chunk_max); |
7823 | gimple_set_location (g: call, location: loc); |
7824 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
7825 | } |
7826 | else |
7827 | chunk_size = chunk_no; |
7828 | |
7829 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, |
7830 | build_int_cst (integer_type_node, |
7831 | IFN_GOACC_LOOP_STEP), |
7832 | dir, range, s, chunk_size, gwv); |
7833 | gimple_call_set_lhs (gs: call, lhs: step); |
7834 | gimple_set_location (g: call, location: loc); |
7835 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
7836 | |
7837 | /* Remove the GIMPLE_OMP_FOR. */ |
7838 | gsi_remove (&gsi, true); |
7839 | |
7840 | /* Fixup edges from head_bb. */ |
7841 | be = BRANCH_EDGE (head_bb); |
7842 | fte = FALLTHRU_EDGE (head_bb); |
7843 | be->flags |= EDGE_FALSE_VALUE; |
7844 | fte->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; |
7845 | |
7846 | basic_block body_bb = fte->dest; |
7847 | |
7848 | if (gimple_in_ssa_p (cfun)) |
7849 | { |
7850 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
7851 | gomp_continue *cont_stmt = as_a <gomp_continue *> (p: gsi_stmt (i: gsi)); |
7852 | |
7853 | offset = gimple_omp_continue_control_use (cont_stmt); |
7854 | offset_incr = gimple_omp_continue_control_def (cont_stmt); |
7855 | } |
7856 | else |
7857 | { |
7858 | offset = create_tmp_var (diff_type, ".offset" ); |
7859 | offset_init = offset_incr = offset; |
7860 | } |
7861 | bound = create_tmp_var (TREE_TYPE (offset), ".bound" ); |
7862 | |
7863 | /* Loop offset & bound go into head_bb. */ |
7864 | gsi = gsi_start_bb (bb: head_bb); |
7865 | |
7866 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, |
7867 | build_int_cst (integer_type_node, |
7868 | IFN_GOACC_LOOP_OFFSET), |
7869 | dir, range, s, |
7870 | chunk_size, gwv, chunk_no); |
7871 | gimple_call_set_lhs (gs: call, lhs: offset_init); |
7872 | gimple_set_location (g: call, location: loc); |
7873 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); |
7874 | |
7875 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, |
7876 | build_int_cst (integer_type_node, |
7877 | IFN_GOACC_LOOP_BOUND), |
7878 | dir, range, s, |
7879 | chunk_size, gwv, offset_init); |
7880 | gimple_call_set_lhs (gs: call, lhs: bound); |
7881 | gimple_set_location (g: call, location: loc); |
7882 | gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING); |
7883 | |
7884 | expr = build2 (cond_code, boolean_type_node, offset_init, bound); |
7885 | gsi_insert_after (&gsi, gimple_build_cond_empty (cond: expr), |
7886 | GSI_CONTINUE_LINKING); |
7887 | |
7888 | /* V assignment goes into body_bb. */ |
7889 | if (!gimple_in_ssa_p (cfun)) |
7890 | { |
7891 | gsi = gsi_start_bb (bb: body_bb); |
7892 | |
7893 | expr = build2 (plus_code, iter_type, b, |
7894 | fold_convert (plus_type, offset)); |
7895 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, |
7896 | true, GSI_SAME_STMT); |
7897 | ass = gimple_build_assign (v, expr); |
7898 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
7899 | |
7900 | if (fd->collapse > 1 || fd->tiling) |
7901 | expand_oacc_collapse_vars (fd, inner: false, gsi: &gsi, counts, ivar: v, diff_type); |
7902 | |
7903 | if (fd->tiling) |
7904 | { |
7905 | /* Determine the range of the element loop -- usually simply |
7906 | the tile_size, but could be smaller if the final |
7907 | iteration of the outer loop is a partial tile. */ |
7908 | tree e_range = create_tmp_var (diff_type, ".e_range" ); |
7909 | |
7910 | expr = build2 (MIN_EXPR, diff_type, |
7911 | build2 (MINUS_EXPR, diff_type, bound, offset), |
7912 | build2 (MULT_EXPR, diff_type, tile_size, |
7913 | element_s)); |
7914 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, |
7915 | true, GSI_SAME_STMT); |
7916 | ass = gimple_build_assign (e_range, expr); |
7917 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
7918 | |
7919 | /* Determine bound, offset & step of inner loop. */ |
7920 | e_bound = create_tmp_var (diff_type, ".e_bound" ); |
7921 | e_offset = create_tmp_var (diff_type, ".e_offset" ); |
7922 | e_step = create_tmp_var (diff_type, ".e_step" ); |
7923 | |
7924 | /* Mark these as element loops. */ |
7925 | tree t, e_gwv = integer_minus_one_node; |
7926 | tree chunk = build_int_cst (diff_type, 0); /* Never chunked. */ |
7927 | |
7928 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_OFFSET); |
7929 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, |
7930 | element_s, chunk, e_gwv, chunk); |
7931 | gimple_call_set_lhs (gs: call, lhs: e_offset); |
7932 | gimple_set_location (g: call, location: loc); |
7933 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
7934 | |
7935 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_BOUND); |
7936 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 7, t, dir, e_range, |
7937 | element_s, chunk, e_gwv, e_offset); |
7938 | gimple_call_set_lhs (gs: call, lhs: e_bound); |
7939 | gimple_set_location (g: call, location: loc); |
7940 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
7941 | |
7942 | t = build_int_cst (integer_type_node, IFN_GOACC_LOOP_STEP); |
7943 | call = gimple_build_call_internal (IFN_GOACC_LOOP, 6, t, dir, e_range, |
7944 | element_s, chunk, e_gwv); |
7945 | gimple_call_set_lhs (gs: call, lhs: e_step); |
7946 | gimple_set_location (g: call, location: loc); |
7947 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
7948 | |
7949 | /* Add test and split block. */ |
7950 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); |
7951 | stmt = gimple_build_cond_empty (cond: expr); |
7952 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
7953 | split = split_block (body_bb, stmt); |
7954 | elem_body_bb = split->dest; |
7955 | if (cont_bb == body_bb) |
7956 | cont_bb = elem_body_bb; |
7957 | body_bb = split->src; |
7958 | |
7959 | split->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; |
7960 | |
7961 | /* Add a dummy exit for the tiled block when cont_bb is missing. */ |
7962 | if (cont_bb == NULL) |
7963 | { |
7964 | edge e = make_edge (body_bb, exit_bb, EDGE_FALSE_VALUE); |
7965 | e->probability = profile_probability::even (); |
7966 | split->probability = profile_probability::even (); |
7967 | } |
7968 | |
7969 | /* Initialize the user's loop vars. */ |
7970 | gsi = gsi_start_bb (bb: elem_body_bb); |
7971 | expand_oacc_collapse_vars (fd, inner: true, gsi: &gsi, counts, ivar: e_offset, |
7972 | diff_type); |
7973 | } |
7974 | } |
7975 | |
7976 | /* Loop increment goes into cont_bb. If this is not a loop, we |
7977 | will have spawned threads as if it was, and each one will |
7978 | execute one iteration. The specification is not explicit about |
7979 | whether such constructs are ill-formed or not, and they can |
7980 | occur, especially when noreturn routines are involved. */ |
7981 | if (cont_bb) |
7982 | { |
7983 | gsi = gsi_last_nondebug_bb (bb: cont_bb); |
7984 | gomp_continue *cont_stmt = as_a <gomp_continue *> (p: gsi_stmt (i: gsi)); |
7985 | loc = gimple_location (g: cont_stmt); |
7986 | |
7987 | if (fd->tiling) |
7988 | { |
7989 | /* Insert element loop increment and test. */ |
7990 | expr = build2 (PLUS_EXPR, diff_type, e_offset, e_step); |
7991 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, |
7992 | true, GSI_SAME_STMT); |
7993 | ass = gimple_build_assign (e_offset, expr); |
7994 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
7995 | expr = build2 (cond_code, boolean_type_node, e_offset, e_bound); |
7996 | |
7997 | stmt = gimple_build_cond_empty (cond: expr); |
7998 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
7999 | split = split_block (cont_bb, stmt); |
8000 | elem_cont_bb = split->src; |
8001 | cont_bb = split->dest; |
8002 | |
8003 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; |
8004 | split->probability = profile_probability::unlikely ().guessed (); |
8005 | edge latch_edge |
8006 | = make_edge (elem_cont_bb, elem_body_bb, EDGE_TRUE_VALUE); |
8007 | latch_edge->probability = profile_probability::likely ().guessed (); |
8008 | |
8009 | edge skip_edge = make_edge (body_bb, cont_bb, EDGE_FALSE_VALUE); |
8010 | skip_edge->probability = profile_probability::unlikely ().guessed (); |
8011 | edge loop_entry_edge = EDGE_SUCC (body_bb, 1 - skip_edge->dest_idx); |
8012 | loop_entry_edge->probability |
8013 | = profile_probability::likely ().guessed (); |
8014 | |
8015 | gsi = gsi_for_stmt (cont_stmt); |
8016 | } |
8017 | |
8018 | /* Increment offset. */ |
8019 | if (gimple_in_ssa_p (cfun)) |
8020 | expr = build2 (plus_code, iter_type, offset, |
8021 | fold_convert (plus_type, step)); |
8022 | else |
8023 | expr = build2 (PLUS_EXPR, diff_type, offset, step); |
8024 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, |
8025 | true, GSI_SAME_STMT); |
8026 | ass = gimple_build_assign (offset_incr, expr); |
8027 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
8028 | expr = build2 (cond_code, boolean_type_node, offset_incr, bound); |
8029 | gsi_insert_before (&gsi, gimple_build_cond_empty (cond: expr), GSI_SAME_STMT); |
8030 | |
8031 | /* Remove the GIMPLE_OMP_CONTINUE. */ |
8032 | gsi_remove (&gsi, true); |
8033 | |
8034 | /* Fixup edges from cont_bb. */ |
8035 | be = BRANCH_EDGE (cont_bb); |
8036 | fte = FALLTHRU_EDGE (cont_bb); |
8037 | be->flags |= EDGE_TRUE_VALUE; |
8038 | fte->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; |
8039 | |
8040 | if (chunking) |
8041 | { |
8042 | /* Split the beginning of exit_bb to make bottom_bb. We |
8043 | need to insert a nop at the start, because splitting is |
8044 | after a stmt, not before. */ |
8045 | gsi = gsi_start_bb (bb: exit_bb); |
8046 | stmt = gimple_build_nop (); |
8047 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
8048 | split = split_block (exit_bb, stmt); |
8049 | bottom_bb = split->src; |
8050 | exit_bb = split->dest; |
8051 | gsi = gsi_last_bb (bb: bottom_bb); |
8052 | |
8053 | /* Chunk increment and test goes into bottom_bb. */ |
8054 | expr = build2 (PLUS_EXPR, diff_type, chunk_no, |
8055 | build_int_cst (diff_type, 1)); |
8056 | ass = gimple_build_assign (chunk_no, expr); |
8057 | gsi_insert_after (&gsi, ass, GSI_CONTINUE_LINKING); |
8058 | |
8059 | /* Chunk test at end of bottom_bb. */ |
8060 | expr = build2 (LT_EXPR, boolean_type_node, chunk_no, chunk_max); |
8061 | gsi_insert_after (&gsi, gimple_build_cond_empty (cond: expr), |
8062 | GSI_CONTINUE_LINKING); |
8063 | |
8064 | /* Fixup edges from bottom_bb. */ |
8065 | split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE; |
8066 | split->probability = profile_probability::unlikely ().guessed (); |
8067 | edge latch_edge = make_edge (bottom_bb, head_bb, EDGE_TRUE_VALUE); |
8068 | latch_edge->probability = profile_probability::likely ().guessed (); |
8069 | } |
8070 | } |
8071 | |
8072 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
8073 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
8074 | loc = gimple_location (g: gsi_stmt (i: gsi)); |
8075 | |
8076 | if (!gimple_in_ssa_p (cfun)) |
8077 | { |
8078 | /* Insert the final value of V, in case it is live. This is the |
8079 | value for the only thread that survives past the join. */ |
8080 | expr = fold_build2 (MINUS_EXPR, diff_type, range, dir); |
8081 | expr = fold_build2 (PLUS_EXPR, diff_type, expr, s); |
8082 | expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s); |
8083 | expr = fold_build2 (MULT_EXPR, diff_type, expr, s); |
8084 | expr = build2 (plus_code, iter_type, b, fold_convert (plus_type, expr)); |
8085 | expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE, |
8086 | true, GSI_SAME_STMT); |
8087 | ass = gimple_build_assign (v, expr); |
8088 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
8089 | } |
8090 | |
8091 | /* Remove the OMP_RETURN. */ |
8092 | gsi_remove (&gsi, true); |
8093 | |
8094 | if (cont_bb) |
8095 | { |
8096 | /* We now have one, two or three nested loops. Update the loop |
8097 | structures. */ |
8098 | class loop *parent = entry_bb->loop_father; |
8099 | class loop *body = body_bb->loop_father; |
8100 | |
8101 | if (chunking) |
8102 | { |
8103 | class loop *chunk_loop = alloc_loop (); |
8104 | chunk_loop->header = head_bb; |
8105 | chunk_loop->latch = bottom_bb; |
8106 | add_loop (chunk_loop, parent); |
8107 | parent = chunk_loop; |
8108 | } |
8109 | else if (parent != body) |
8110 | { |
8111 | gcc_assert (body->header == body_bb); |
8112 | gcc_assert (body->latch == cont_bb |
8113 | || single_pred (body->latch) == cont_bb); |
8114 | parent = NULL; |
8115 | } |
8116 | |
8117 | if (parent) |
8118 | { |
8119 | class loop *body_loop = alloc_loop (); |
8120 | body_loop->header = body_bb; |
8121 | body_loop->latch = cont_bb; |
8122 | add_loop (body_loop, parent); |
8123 | |
8124 | if (fd->tiling) |
8125 | { |
8126 | /* Insert tiling's element loop. */ |
8127 | class loop *inner_loop = alloc_loop (); |
8128 | inner_loop->header = elem_body_bb; |
8129 | inner_loop->latch = elem_cont_bb; |
8130 | add_loop (inner_loop, body_loop); |
8131 | } |
8132 | } |
8133 | } |
8134 | } |
8135 | |
8136 | /* Expand the OMP loop defined by REGION. */ |
8137 | |
8138 | static void |
8139 | expand_omp_for (struct omp_region *region, gimple *inner_stmt) |
8140 | { |
8141 | struct omp_for_data fd; |
8142 | struct omp_for_data_loop *loops; |
8143 | |
8144 | loops = XALLOCAVEC (struct omp_for_data_loop, |
8145 | gimple_omp_for_collapse |
8146 | (last_nondebug_stmt (region->entry))); |
8147 | omp_extract_for_data (for_stmt: as_a <gomp_for *> (p: last_nondebug_stmt (region->entry)), |
8148 | fd: &fd, loops); |
8149 | region->sched_kind = fd.sched_kind; |
8150 | region->sched_modifiers = fd.sched_modifiers; |
8151 | region->has_lastprivate_conditional = fd.lastprivate_conditional != 0; |
8152 | if (fd.non_rect && !gimple_omp_for_combined_into_p (g: fd.for_stmt)) |
8153 | { |
8154 | for (int i = fd.first_nonrect; i <= fd.last_nonrect; i++) |
8155 | if ((loops[i].m1 || loops[i].m2) |
8156 | && (loops[i].m1 == NULL_TREE |
8157 | || TREE_CODE (loops[i].m1) == INTEGER_CST) |
8158 | && (loops[i].m2 == NULL_TREE |
8159 | || TREE_CODE (loops[i].m2) == INTEGER_CST) |
8160 | && TREE_CODE (loops[i].step) == INTEGER_CST |
8161 | && TREE_CODE (loops[i - loops[i].outer].step) == INTEGER_CST) |
8162 | { |
8163 | tree t; |
8164 | tree itype = TREE_TYPE (loops[i].v); |
8165 | if (loops[i].m1 && loops[i].m2) |
8166 | t = fold_build2 (MINUS_EXPR, itype, loops[i].m2, loops[i].m1); |
8167 | else if (loops[i].m1) |
8168 | t = fold_build1 (NEGATE_EXPR, itype, loops[i].m1); |
8169 | else |
8170 | t = loops[i].m2; |
8171 | t = fold_build2 (MULT_EXPR, itype, t, |
8172 | fold_convert (itype, |
8173 | loops[i - loops[i].outer].step)); |
8174 | if (TYPE_UNSIGNED (itype) && loops[i].cond_code == GT_EXPR) |
8175 | t = fold_build2 (TRUNC_MOD_EXPR, itype, |
8176 | fold_build1 (NEGATE_EXPR, itype, t), |
8177 | fold_build1 (NEGATE_EXPR, itype, |
8178 | fold_convert (itype, |
8179 | loops[i].step))); |
8180 | else |
8181 | t = fold_build2 (TRUNC_MOD_EXPR, itype, t, |
8182 | fold_convert (itype, loops[i].step)); |
8183 | if (integer_nonzerop (t)) |
8184 | error_at (gimple_location (g: fd.for_stmt), |
8185 | "invalid OpenMP non-rectangular loop step; " |
8186 | "%<(%E - %E) * %E%> is not a multiple of loop %d " |
8187 | "step %qE" , |
8188 | loops[i].m2 ? loops[i].m2 : integer_zero_node, |
8189 | loops[i].m1 ? loops[i].m1 : integer_zero_node, |
8190 | loops[i - loops[i].outer].step, i + 1, |
8191 | loops[i].step); |
8192 | } |
8193 | } |
8194 | |
8195 | gcc_assert (EDGE_COUNT (region->entry->succs) == 2); |
8196 | BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; |
8197 | FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL; |
8198 | if (region->cont) |
8199 | { |
8200 | gcc_assert (EDGE_COUNT (region->cont->succs) == 2); |
8201 | BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; |
8202 | FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL; |
8203 | } |
8204 | else |
8205 | /* If there isn't a continue then this is a degerate case where |
8206 | the introduction of abnormal edges during lowering will prevent |
8207 | original loops from being detected. Fix that up. */ |
8208 | loops_state_set (flags: LOOPS_NEED_FIXUP); |
8209 | |
8210 | if (gimple_omp_for_kind (g: fd.for_stmt) == GF_OMP_FOR_KIND_SIMD) |
8211 | expand_omp_simd (region, fd: &fd); |
8212 | else if (gimple_omp_for_kind (g: fd.for_stmt) == GF_OMP_FOR_KIND_OACC_LOOP) |
8213 | { |
8214 | gcc_assert (!inner_stmt && !fd.non_rect); |
8215 | expand_oacc_for (region, fd: &fd); |
8216 | } |
8217 | else if (gimple_omp_for_kind (g: fd.for_stmt) == GF_OMP_FOR_KIND_TASKLOOP) |
8218 | { |
8219 | if (gimple_omp_for_combined_into_p (g: fd.for_stmt)) |
8220 | expand_omp_taskloop_for_inner (region, fd: &fd, inner_stmt); |
8221 | else |
8222 | expand_omp_taskloop_for_outer (region, fd: &fd, inner_stmt); |
8223 | } |
8224 | else if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC |
8225 | && !fd.have_ordered) |
8226 | { |
8227 | if (fd.chunk_size == NULL) |
8228 | expand_omp_for_static_nochunk (region, fd: &fd, inner_stmt); |
8229 | else |
8230 | expand_omp_for_static_chunk (region, fd: &fd, inner_stmt); |
8231 | } |
8232 | else |
8233 | { |
8234 | int fn_index, start_ix, next_ix; |
8235 | unsigned HOST_WIDE_INT sched = 0; |
8236 | tree sched_arg = NULL_TREE; |
8237 | |
8238 | gcc_assert (gimple_omp_for_kind (fd.for_stmt) |
8239 | == GF_OMP_FOR_KIND_FOR && !fd.non_rect); |
8240 | if (fd.chunk_size == NULL |
8241 | && fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC) |
8242 | fd.chunk_size = integer_zero_node; |
8243 | switch (fd.sched_kind) |
8244 | { |
8245 | case OMP_CLAUSE_SCHEDULE_RUNTIME: |
8246 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_NONMONOTONIC) != 0 |
8247 | && fd.lastprivate_conditional == 0) |
8248 | { |
8249 | gcc_assert (!fd.have_ordered); |
8250 | fn_index = 6; |
8251 | sched = 4; |
8252 | } |
8253 | else if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 |
8254 | && !fd.have_ordered |
8255 | && fd.lastprivate_conditional == 0) |
8256 | fn_index = 7; |
8257 | else |
8258 | { |
8259 | fn_index = 3; |
8260 | sched = (HOST_WIDE_INT_1U << 31); |
8261 | } |
8262 | break; |
8263 | case OMP_CLAUSE_SCHEDULE_DYNAMIC: |
8264 | case OMP_CLAUSE_SCHEDULE_GUIDED: |
8265 | if ((fd.sched_modifiers & OMP_CLAUSE_SCHEDULE_MONOTONIC) == 0 |
8266 | && !fd.have_ordered |
8267 | && fd.lastprivate_conditional == 0) |
8268 | { |
8269 | fn_index = 3 + fd.sched_kind; |
8270 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; |
8271 | break; |
8272 | } |
8273 | fn_index = fd.sched_kind; |
8274 | sched = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_GUIDED) + 2; |
8275 | sched += (HOST_WIDE_INT_1U << 31); |
8276 | break; |
8277 | case OMP_CLAUSE_SCHEDULE_STATIC: |
8278 | gcc_assert (fd.have_ordered); |
8279 | fn_index = 0; |
8280 | sched = (HOST_WIDE_INT_1U << 31) + 1; |
8281 | break; |
8282 | default: |
8283 | gcc_unreachable (); |
8284 | } |
8285 | if (!fd.ordered) |
8286 | fn_index += fd.have_ordered * 8; |
8287 | if (fd.ordered) |
8288 | start_ix = ((int)BUILT_IN_GOMP_LOOP_DOACROSS_STATIC_START) + fn_index; |
8289 | else |
8290 | start_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_START) + fn_index; |
8291 | next_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_NEXT) + fn_index; |
8292 | if (fd.have_reductemp || fd.have_pointer_condtemp) |
8293 | { |
8294 | if (fd.ordered) |
8295 | start_ix = (int)BUILT_IN_GOMP_LOOP_DOACROSS_START; |
8296 | else if (fd.have_ordered) |
8297 | start_ix = (int)BUILT_IN_GOMP_LOOP_ORDERED_START; |
8298 | else |
8299 | start_ix = (int)BUILT_IN_GOMP_LOOP_START; |
8300 | sched_arg = build_int_cstu (long_integer_type_node, sched); |
8301 | if (!fd.chunk_size) |
8302 | fd.chunk_size = integer_zero_node; |
8303 | } |
8304 | if (fd.iter_type == long_long_unsigned_type_node) |
8305 | { |
8306 | start_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_START |
8307 | - (int)BUILT_IN_GOMP_LOOP_STATIC_START); |
8308 | next_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_NEXT |
8309 | - (int)BUILT_IN_GOMP_LOOP_STATIC_NEXT); |
8310 | } |
8311 | expand_omp_for_generic (region, fd: &fd, start_fn: (enum built_in_function) start_ix, |
8312 | next_fn: (enum built_in_function) next_ix, sched_arg, |
8313 | inner_stmt); |
8314 | } |
8315 | } |
8316 | |
8317 | /* Expand code for an OpenMP sections directive. In pseudo code, we generate |
8318 | |
8319 | v = GOMP_sections_start (n); |
8320 | L0: |
8321 | switch (v) |
8322 | { |
8323 | case 0: |
8324 | goto L2; |
8325 | case 1: |
8326 | section 1; |
8327 | goto L1; |
8328 | case 2: |
8329 | ... |
8330 | case n: |
8331 | ... |
8332 | default: |
8333 | abort (); |
8334 | } |
8335 | L1: |
8336 | v = GOMP_sections_next (); |
8337 | goto L0; |
8338 | L2: |
8339 | reduction; |
8340 | |
8341 | If this is a combined parallel sections, replace the call to |
8342 | GOMP_sections_start with call to GOMP_sections_next. */ |
8343 | |
8344 | static void |
8345 | expand_omp_sections (struct omp_region *region) |
8346 | { |
8347 | tree t, u, vin = NULL, vmain, vnext, l2; |
8348 | unsigned len; |
8349 | basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb; |
8350 | gimple_stmt_iterator si, switch_si; |
8351 | gomp_sections *sections_stmt; |
8352 | gimple *stmt; |
8353 | gomp_continue *cont; |
8354 | edge_iterator ei; |
8355 | edge e; |
8356 | struct omp_region *inner; |
8357 | unsigned i, casei; |
8358 | bool exit_reachable = region->cont != NULL; |
8359 | |
8360 | gcc_assert (region->exit != NULL); |
8361 | entry_bb = region->entry; |
8362 | l0_bb = single_succ (bb: entry_bb); |
8363 | l1_bb = region->cont; |
8364 | l2_bb = region->exit; |
8365 | if (single_pred_p (bb: l2_bb) && single_pred (bb: l2_bb) == l0_bb) |
8366 | l2 = gimple_block_label (l2_bb); |
8367 | else |
8368 | { |
8369 | /* This can happen if there are reductions. */ |
8370 | len = EDGE_COUNT (l0_bb->succs); |
8371 | gcc_assert (len > 0); |
8372 | e = EDGE_SUCC (l0_bb, len - 1); |
8373 | si = gsi_last_nondebug_bb (bb: e->dest); |
8374 | l2 = NULL_TREE; |
8375 | if (gsi_end_p (i: si) |
8376 | || gimple_code (g: gsi_stmt (i: si)) != GIMPLE_OMP_SECTION) |
8377 | l2 = gimple_block_label (e->dest); |
8378 | else |
8379 | FOR_EACH_EDGE (e, ei, l0_bb->succs) |
8380 | { |
8381 | si = gsi_last_nondebug_bb (bb: e->dest); |
8382 | if (gsi_end_p (i: si) |
8383 | || gimple_code (g: gsi_stmt (i: si)) != GIMPLE_OMP_SECTION) |
8384 | { |
8385 | l2 = gimple_block_label (e->dest); |
8386 | break; |
8387 | } |
8388 | } |
8389 | } |
8390 | if (exit_reachable) |
8391 | default_bb = create_empty_bb (l1_bb->prev_bb); |
8392 | else |
8393 | default_bb = create_empty_bb (l0_bb); |
8394 | |
8395 | /* We will build a switch() with enough cases for all the |
8396 | GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work |
8397 | and a default case to abort if something goes wrong. */ |
8398 | len = EDGE_COUNT (l0_bb->succs); |
8399 | |
8400 | /* Use vec::quick_push on label_vec throughout, since we know the size |
8401 | in advance. */ |
8402 | auto_vec<tree> label_vec (len); |
8403 | |
8404 | /* The call to GOMP_sections_start goes in ENTRY_BB, replacing the |
8405 | GIMPLE_OMP_SECTIONS statement. */ |
8406 | si = gsi_last_nondebug_bb (bb: entry_bb); |
8407 | sections_stmt = as_a <gomp_sections *> (p: gsi_stmt (i: si)); |
8408 | gcc_assert (gimple_code (sections_stmt) == GIMPLE_OMP_SECTIONS); |
8409 | vin = gimple_omp_sections_control (gs: sections_stmt); |
8410 | tree clauses = gimple_omp_sections_clauses (gs: sections_stmt); |
8411 | tree reductmp = omp_find_clause (clauses, kind: OMP_CLAUSE__REDUCTEMP_); |
8412 | tree condtmp = omp_find_clause (clauses, kind: OMP_CLAUSE__CONDTEMP_); |
8413 | tree cond_var = NULL_TREE; |
8414 | if (reductmp || condtmp) |
8415 | { |
8416 | tree reductions = null_pointer_node, mem = null_pointer_node; |
8417 | tree memv = NULL_TREE, condtemp = NULL_TREE; |
8418 | gimple_stmt_iterator gsi = gsi_none (); |
8419 | gimple *g = NULL; |
8420 | if (reductmp) |
8421 | { |
8422 | reductions = OMP_CLAUSE_DECL (reductmp); |
8423 | gcc_assert (TREE_CODE (reductions) == SSA_NAME); |
8424 | g = SSA_NAME_DEF_STMT (reductions); |
8425 | reductions = gimple_assign_rhs1 (gs: g); |
8426 | OMP_CLAUSE_DECL (reductmp) = reductions; |
8427 | gsi = gsi_for_stmt (g); |
8428 | } |
8429 | else |
8430 | gsi = si; |
8431 | if (condtmp) |
8432 | { |
8433 | condtemp = OMP_CLAUSE_DECL (condtmp); |
8434 | tree c = omp_find_clause (OMP_CLAUSE_CHAIN (condtmp), |
8435 | kind: OMP_CLAUSE__CONDTEMP_); |
8436 | cond_var = OMP_CLAUSE_DECL (c); |
8437 | tree type = TREE_TYPE (condtemp); |
8438 | memv = create_tmp_var (type); |
8439 | TREE_ADDRESSABLE (memv) = 1; |
8440 | unsigned cnt = 0; |
8441 | for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c)) |
8442 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE |
8443 | && OMP_CLAUSE_LASTPRIVATE_CONDITIONAL (c)) |
8444 | ++cnt; |
8445 | unsigned HOST_WIDE_INT sz |
8446 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type))) * cnt; |
8447 | expand_omp_build_assign (gsi_p: &gsi, to: memv, from: build_int_cst (type, sz), |
8448 | after: false); |
8449 | mem = build_fold_addr_expr (memv); |
8450 | } |
8451 | t = build_int_cst (unsigned_type_node, len - 1); |
8452 | u = builtin_decl_explicit (fncode: BUILT_IN_GOMP_SECTIONS2_START); |
8453 | stmt = gimple_build_call (u, 3, t, reductions, mem); |
8454 | gimple_call_set_lhs (gs: stmt, lhs: vin); |
8455 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
8456 | if (condtmp) |
8457 | { |
8458 | expand_omp_build_assign (gsi_p: &gsi, to: condtemp, from: memv, after: false); |
8459 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), |
8460 | vin, build_one_cst (TREE_TYPE (cond_var))); |
8461 | expand_omp_build_assign (gsi_p: &gsi, to: cond_var, from: t, after: false); |
8462 | } |
8463 | if (reductmp) |
8464 | { |
8465 | gsi_remove (&gsi, true); |
8466 | release_ssa_name (name: gimple_assign_lhs (gs: g)); |
8467 | } |
8468 | } |
8469 | else if (!is_combined_parallel (region)) |
8470 | { |
8471 | /* If we are not inside a combined parallel+sections region, |
8472 | call GOMP_sections_start. */ |
8473 | t = build_int_cst (unsigned_type_node, len - 1); |
8474 | u = builtin_decl_explicit (fncode: BUILT_IN_GOMP_SECTIONS_START); |
8475 | stmt = gimple_build_call (u, 1, t); |
8476 | } |
8477 | else |
8478 | { |
8479 | /* Otherwise, call GOMP_sections_next. */ |
8480 | u = builtin_decl_explicit (fncode: BUILT_IN_GOMP_SECTIONS_NEXT); |
8481 | stmt = gimple_build_call (u, 0); |
8482 | } |
8483 | if (!reductmp && !condtmp) |
8484 | { |
8485 | gimple_call_set_lhs (gs: stmt, lhs: vin); |
8486 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); |
8487 | } |
8488 | gsi_remove (&si, true); |
8489 | |
8490 | /* The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in |
8491 | L0_BB. */ |
8492 | switch_si = gsi_last_nondebug_bb (bb: l0_bb); |
8493 | gcc_assert (gimple_code (gsi_stmt (switch_si)) == GIMPLE_OMP_SECTIONS_SWITCH); |
8494 | if (exit_reachable) |
8495 | { |
8496 | cont = as_a <gomp_continue *> (p: last_nondebug_stmt (l1_bb)); |
8497 | gcc_assert (gimple_code (cont) == GIMPLE_OMP_CONTINUE); |
8498 | vmain = gimple_omp_continue_control_use (cont_stmt: cont); |
8499 | vnext = gimple_omp_continue_control_def (cont_stmt: cont); |
8500 | } |
8501 | else |
8502 | { |
8503 | vmain = vin; |
8504 | vnext = NULL_TREE; |
8505 | } |
8506 | |
8507 | t = build_case_label (build_int_cst (unsigned_type_node, 0), NULL, l2); |
8508 | label_vec.quick_push (obj: t); |
8509 | i = 1; |
8510 | |
8511 | /* Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR. */ |
8512 | for (inner = region->inner, casei = 1; |
8513 | inner; |
8514 | inner = inner->next, i++, casei++) |
8515 | { |
8516 | basic_block s_entry_bb, s_exit_bb; |
8517 | |
8518 | /* Skip optional reduction region. */ |
8519 | if (inner->type == GIMPLE_OMP_ATOMIC_LOAD) |
8520 | { |
8521 | --i; |
8522 | --casei; |
8523 | continue; |
8524 | } |
8525 | |
8526 | s_entry_bb = inner->entry; |
8527 | s_exit_bb = inner->exit; |
8528 | |
8529 | t = gimple_block_label (s_entry_bb); |
8530 | u = build_int_cst (unsigned_type_node, casei); |
8531 | u = build_case_label (u, NULL, t); |
8532 | label_vec.quick_push (obj: u); |
8533 | |
8534 | si = gsi_last_nondebug_bb (bb: s_entry_bb); |
8535 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SECTION); |
8536 | gcc_assert (i < len || gimple_omp_section_last_p (gsi_stmt (si))); |
8537 | gsi_remove (&si, true); |
8538 | single_succ_edge (bb: s_entry_bb)->flags = EDGE_FALLTHRU; |
8539 | |
8540 | if (s_exit_bb == NULL) |
8541 | continue; |
8542 | |
8543 | si = gsi_last_nondebug_bb (bb: s_exit_bb); |
8544 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); |
8545 | gsi_remove (&si, true); |
8546 | |
8547 | single_succ_edge (bb: s_exit_bb)->flags = EDGE_FALLTHRU; |
8548 | } |
8549 | |
8550 | /* Error handling code goes in DEFAULT_BB. */ |
8551 | t = gimple_block_label (default_bb); |
8552 | u = build_case_label (NULL, NULL, t); |
8553 | make_edge (l0_bb, default_bb, 0); |
8554 | add_bb_to_loop (default_bb, current_loops->tree_root); |
8555 | |
8556 | stmt = gimple_build_switch (vmain, u, label_vec); |
8557 | gsi_insert_after (&switch_si, stmt, GSI_SAME_STMT); |
8558 | gsi_remove (&switch_si, true); |
8559 | |
8560 | si = gsi_start_bb (bb: default_bb); |
8561 | stmt = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_TRAP), 0); |
8562 | gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING); |
8563 | |
8564 | if (exit_reachable) |
8565 | { |
8566 | tree bfn_decl; |
8567 | |
8568 | /* Code to get the next section goes in L1_BB. */ |
8569 | si = gsi_last_nondebug_bb (bb: l1_bb); |
8570 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CONTINUE); |
8571 | |
8572 | bfn_decl = builtin_decl_explicit (fncode: BUILT_IN_GOMP_SECTIONS_NEXT); |
8573 | stmt = gimple_build_call (bfn_decl, 0); |
8574 | gimple_call_set_lhs (gs: stmt, lhs: vnext); |
8575 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
8576 | if (cond_var) |
8577 | { |
8578 | tree t = build2 (PLUS_EXPR, TREE_TYPE (cond_var), |
8579 | vnext, build_one_cst (TREE_TYPE (cond_var))); |
8580 | expand_omp_build_assign (gsi_p: &si, to: cond_var, from: t, after: false); |
8581 | } |
8582 | gsi_remove (&si, true); |
8583 | |
8584 | single_succ_edge (bb: l1_bb)->flags = EDGE_FALLTHRU; |
8585 | } |
8586 | |
8587 | /* Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB. */ |
8588 | si = gsi_last_nondebug_bb (bb: l2_bb); |
8589 | if (gimple_omp_return_nowait_p (g: gsi_stmt (i: si))) |
8590 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_SECTIONS_END_NOWAIT); |
8591 | else if (gimple_omp_return_lhs (g: gsi_stmt (i: si))) |
8592 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_SECTIONS_END_CANCEL); |
8593 | else |
8594 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_SECTIONS_END); |
8595 | stmt = gimple_build_call (t, 0); |
8596 | if (gimple_omp_return_lhs (g: gsi_stmt (i: si))) |
8597 | gimple_call_set_lhs (gs: stmt, lhs: gimple_omp_return_lhs (g: gsi_stmt (i: si))); |
8598 | gsi_insert_after (&si, stmt, GSI_SAME_STMT); |
8599 | gsi_remove (&si, true); |
8600 | |
8601 | set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb); |
8602 | } |
8603 | |
8604 | /* Expand code for an OpenMP single or scope directive. We've already expanded |
8605 | much of the code, here we simply place the GOMP_barrier call. */ |
8606 | |
8607 | static void |
8608 | expand_omp_single (struct omp_region *region) |
8609 | { |
8610 | basic_block entry_bb, exit_bb; |
8611 | gimple_stmt_iterator si; |
8612 | |
8613 | entry_bb = region->entry; |
8614 | exit_bb = region->exit; |
8615 | |
8616 | si = gsi_last_nondebug_bb (bb: entry_bb); |
8617 | enum gimple_code code = gimple_code (g: gsi_stmt (i: si)); |
8618 | gcc_assert (code == GIMPLE_OMP_SINGLE || code == GIMPLE_OMP_SCOPE); |
8619 | gsi_remove (&si, true); |
8620 | single_succ_edge (bb: entry_bb)->flags = EDGE_FALLTHRU; |
8621 | |
8622 | if (exit_bb == NULL) |
8623 | { |
8624 | gcc_assert (code == GIMPLE_OMP_SCOPE); |
8625 | return; |
8626 | } |
8627 | |
8628 | si = gsi_last_nondebug_bb (bb: exit_bb); |
8629 | if (!gimple_omp_return_nowait_p (g: gsi_stmt (i: si))) |
8630 | { |
8631 | tree t = gimple_omp_return_lhs (g: gsi_stmt (i: si)); |
8632 | gsi_insert_after (&si, omp_build_barrier (lhs: t), GSI_SAME_STMT); |
8633 | } |
8634 | gsi_remove (&si, true); |
8635 | single_succ_edge (bb: exit_bb)->flags = EDGE_FALLTHRU; |
8636 | } |
8637 | |
8638 | /* Generic expansion for OpenMP synchronization directives: master, |
8639 | ordered and critical. All we need to do here is remove the entry |
8640 | and exit markers for REGION. */ |
8641 | |
8642 | static void |
8643 | expand_omp_synch (struct omp_region *region) |
8644 | { |
8645 | basic_block entry_bb, exit_bb; |
8646 | gimple_stmt_iterator si; |
8647 | |
8648 | entry_bb = region->entry; |
8649 | exit_bb = region->exit; |
8650 | |
8651 | si = gsi_last_nondebug_bb (bb: entry_bb); |
8652 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE |
8653 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASTER |
8654 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASKED |
8655 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TASKGROUP |
8656 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ORDERED |
8657 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CRITICAL |
8658 | || gimple_code (gsi_stmt (si)) == GIMPLE_OMP_TEAMS); |
8659 | if (gimple_code (g: gsi_stmt (i: si)) == GIMPLE_OMP_TEAMS |
8660 | && gimple_omp_teams_host (omp_teams_stmt: as_a <gomp_teams *> (p: gsi_stmt (i: si)))) |
8661 | { |
8662 | expand_omp_taskreg (region); |
8663 | return; |
8664 | } |
8665 | gsi_remove (&si, true); |
8666 | single_succ_edge (bb: entry_bb)->flags = EDGE_FALLTHRU; |
8667 | |
8668 | if (exit_bb) |
8669 | { |
8670 | si = gsi_last_nondebug_bb (bb: exit_bb); |
8671 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN); |
8672 | gsi_remove (&si, true); |
8673 | single_succ_edge (bb: exit_bb)->flags = EDGE_FALLTHRU; |
8674 | } |
8675 | } |
8676 | |
8677 | /* Translate enum omp_memory_order to enum memmodel for the embedded |
8678 | fail clause in there. */ |
8679 | |
8680 | static enum memmodel |
8681 | omp_memory_order_to_fail_memmodel (enum omp_memory_order mo) |
8682 | { |
8683 | switch (mo & OMP_FAIL_MEMORY_ORDER_MASK) |
8684 | { |
8685 | case OMP_FAIL_MEMORY_ORDER_UNSPECIFIED: |
8686 | switch (mo & OMP_MEMORY_ORDER_MASK) |
8687 | { |
8688 | case OMP_MEMORY_ORDER_RELAXED: return MEMMODEL_RELAXED; |
8689 | case OMP_MEMORY_ORDER_ACQUIRE: return MEMMODEL_ACQUIRE; |
8690 | case OMP_MEMORY_ORDER_RELEASE: return MEMMODEL_RELAXED; |
8691 | case OMP_MEMORY_ORDER_ACQ_REL: return MEMMODEL_ACQUIRE; |
8692 | case OMP_MEMORY_ORDER_SEQ_CST: return MEMMODEL_SEQ_CST; |
8693 | default: break; |
8694 | } |
8695 | gcc_unreachable (); |
8696 | case OMP_FAIL_MEMORY_ORDER_RELAXED: return MEMMODEL_RELAXED; |
8697 | case OMP_FAIL_MEMORY_ORDER_ACQUIRE: return MEMMODEL_ACQUIRE; |
8698 | case OMP_FAIL_MEMORY_ORDER_SEQ_CST: return MEMMODEL_SEQ_CST; |
8699 | default: gcc_unreachable (); |
8700 | } |
8701 | } |
8702 | |
8703 | /* Translate enum omp_memory_order to enum memmodel. The two enums |
8704 | are using different numbers so that OMP_MEMORY_ORDER_UNSPECIFIED |
8705 | is 0 and omp_memory_order has the fail mode encoded in it too. */ |
8706 | |
8707 | static enum memmodel |
8708 | omp_memory_order_to_memmodel (enum omp_memory_order mo) |
8709 | { |
8710 | enum memmodel ret, fail_ret; |
8711 | switch (mo & OMP_MEMORY_ORDER_MASK) |
8712 | { |
8713 | case OMP_MEMORY_ORDER_RELAXED: ret = MEMMODEL_RELAXED; break; |
8714 | case OMP_MEMORY_ORDER_ACQUIRE: ret = MEMMODEL_ACQUIRE; break; |
8715 | case OMP_MEMORY_ORDER_RELEASE: ret = MEMMODEL_RELEASE; break; |
8716 | case OMP_MEMORY_ORDER_ACQ_REL: ret = MEMMODEL_ACQ_REL; break; |
8717 | case OMP_MEMORY_ORDER_SEQ_CST: ret = MEMMODEL_SEQ_CST; break; |
8718 | default: gcc_unreachable (); |
8719 | } |
8720 | /* If we drop the -Winvalid-memory-model warning for C++17 P0418R2, |
8721 | we can just return ret here unconditionally. Otherwise, work around |
8722 | it here and make sure fail memmodel is not stronger. */ |
8723 | if ((mo & OMP_FAIL_MEMORY_ORDER_MASK) == OMP_FAIL_MEMORY_ORDER_UNSPECIFIED) |
8724 | return ret; |
8725 | fail_ret = omp_memory_order_to_fail_memmodel (mo); |
8726 | if (fail_ret > ret) |
8727 | return fail_ret; |
8728 | return ret; |
8729 | } |
8730 | |
8731 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
8732 | operation as a normal volatile load. */ |
8733 | |
8734 | static bool |
8735 | expand_omp_atomic_load (basic_block load_bb, tree addr, |
8736 | tree loaded_val, int index) |
8737 | { |
8738 | enum built_in_function tmpbase; |
8739 | gimple_stmt_iterator gsi; |
8740 | basic_block store_bb; |
8741 | location_t loc; |
8742 | gimple *stmt; |
8743 | tree decl, type, itype; |
8744 | |
8745 | gsi = gsi_last_nondebug_bb (bb: load_bb); |
8746 | stmt = gsi_stmt (i: gsi); |
8747 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); |
8748 | loc = gimple_location (g: stmt); |
8749 | |
8750 | /* ??? If the target does not implement atomic_load_optab[mode], and mode |
8751 | is smaller than word size, then expand_atomic_load assumes that the load |
8752 | is atomic. We could avoid the builtin entirely in this case. */ |
8753 | |
8754 | tmpbase = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); |
8755 | decl = builtin_decl_explicit (fncode: tmpbase); |
8756 | if (decl == NULL_TREE) |
8757 | return false; |
8758 | |
8759 | type = TREE_TYPE (loaded_val); |
8760 | itype = TREE_TYPE (TREE_TYPE (decl)); |
8761 | |
8762 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (g: stmt); |
8763 | tree mo = build_int_cst (integer_type_node, |
8764 | omp_memory_order_to_memmodel (mo: omo)); |
8765 | gcall *call = gimple_build_call (decl, 2, addr, mo); |
8766 | gimple_set_location (g: call, location: loc); |
8767 | gimple_set_vuse (g: call, vuse: gimple_vuse (g: stmt)); |
8768 | gimple *repl; |
8769 | if (!useless_type_conversion_p (type, itype)) |
8770 | { |
8771 | tree lhs = make_ssa_name (var: itype); |
8772 | gimple_call_set_lhs (gs: call, lhs); |
8773 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
8774 | repl = gimple_build_assign (loaded_val, |
8775 | build1 (VIEW_CONVERT_EXPR, type, lhs)); |
8776 | gimple_set_location (g: repl, location: loc); |
8777 | } |
8778 | else |
8779 | { |
8780 | gimple_call_set_lhs (gs: call, lhs: loaded_val); |
8781 | repl = call; |
8782 | } |
8783 | gsi_replace (&gsi, repl, true); |
8784 | |
8785 | store_bb = single_succ (bb: load_bb); |
8786 | gsi = gsi_last_nondebug_bb (bb: store_bb); |
8787 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); |
8788 | gsi_remove (&gsi, true); |
8789 | |
8790 | return true; |
8791 | } |
8792 | |
8793 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
8794 | operation as a normal volatile store. */ |
8795 | |
8796 | static bool |
8797 | expand_omp_atomic_store (basic_block load_bb, tree addr, |
8798 | tree loaded_val, tree stored_val, int index) |
8799 | { |
8800 | enum built_in_function tmpbase; |
8801 | gimple_stmt_iterator gsi; |
8802 | basic_block store_bb = single_succ (bb: load_bb); |
8803 | location_t loc; |
8804 | gimple *stmt; |
8805 | tree decl, type, itype; |
8806 | machine_mode imode; |
8807 | bool exchange; |
8808 | |
8809 | gsi = gsi_last_nondebug_bb (bb: load_bb); |
8810 | stmt = gsi_stmt (i: gsi); |
8811 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD); |
8812 | |
8813 | /* If the load value is needed, then this isn't a store but an exchange. */ |
8814 | exchange = gimple_omp_atomic_need_value_p (g: stmt); |
8815 | |
8816 | gsi = gsi_last_nondebug_bb (bb: store_bb); |
8817 | stmt = gsi_stmt (i: gsi); |
8818 | gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_STORE); |
8819 | loc = gimple_location (g: stmt); |
8820 | |
8821 | /* ??? If the target does not implement atomic_store_optab[mode], and mode |
8822 | is smaller than word size, then expand_atomic_store assumes that the store |
8823 | is atomic. We could avoid the builtin entirely in this case. */ |
8824 | |
8825 | tmpbase = (exchange ? BUILT_IN_ATOMIC_EXCHANGE_N : BUILT_IN_ATOMIC_STORE_N); |
8826 | tmpbase = (enum built_in_function) ((int) tmpbase + index + 1); |
8827 | decl = builtin_decl_explicit (fncode: tmpbase); |
8828 | if (decl == NULL_TREE) |
8829 | return false; |
8830 | |
8831 | type = TREE_TYPE (stored_val); |
8832 | |
8833 | /* Dig out the type of the function's second argument. */ |
8834 | itype = TREE_TYPE (decl); |
8835 | itype = TYPE_ARG_TYPES (itype); |
8836 | itype = TREE_CHAIN (itype); |
8837 | itype = TREE_VALUE (itype); |
8838 | imode = TYPE_MODE (itype); |
8839 | |
8840 | if (exchange && !can_atomic_exchange_p (imode, true)) |
8841 | return false; |
8842 | |
8843 | if (!useless_type_conversion_p (itype, type)) |
8844 | stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val); |
8845 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (g: stmt); |
8846 | tree mo = build_int_cst (integer_type_node, |
8847 | omp_memory_order_to_memmodel (mo: omo)); |
8848 | stored_val = force_gimple_operand_gsi (&gsi, stored_val, true, NULL_TREE, |
8849 | true, GSI_SAME_STMT); |
8850 | gcall *call = gimple_build_call (decl, 3, addr, stored_val, mo); |
8851 | gimple_set_location (g: call, location: loc); |
8852 | gimple_set_vuse (g: call, vuse: gimple_vuse (g: stmt)); |
8853 | gimple_set_vdef (g: call, vdef: gimple_vdef (g: stmt)); |
8854 | |
8855 | gimple *repl = call; |
8856 | if (exchange) |
8857 | { |
8858 | if (!useless_type_conversion_p (type, itype)) |
8859 | { |
8860 | tree lhs = make_ssa_name (var: itype); |
8861 | gimple_call_set_lhs (gs: call, lhs); |
8862 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
8863 | repl = gimple_build_assign (loaded_val, |
8864 | build1 (VIEW_CONVERT_EXPR, type, lhs)); |
8865 | gimple_set_location (g: repl, location: loc); |
8866 | } |
8867 | else |
8868 | gimple_call_set_lhs (gs: call, lhs: loaded_val); |
8869 | } |
8870 | gsi_replace (&gsi, repl, true); |
8871 | |
8872 | /* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */ |
8873 | gsi = gsi_last_nondebug_bb (bb: load_bb); |
8874 | gsi_remove (&gsi, true); |
8875 | |
8876 | return true; |
8877 | } |
8878 | |
8879 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
8880 | operation as a __atomic_fetch_op builtin. INDEX is log2 of the |
8881 | size of the data type, and thus usable to find the index of the builtin |
8882 | decl. Returns false if the expression is not of the proper form. */ |
8883 | |
8884 | static bool |
8885 | expand_omp_atomic_fetch_op (basic_block load_bb, |
8886 | tree addr, tree loaded_val, |
8887 | tree stored_val, int index) |
8888 | { |
8889 | enum built_in_function oldbase, newbase, tmpbase; |
8890 | tree decl, itype, call; |
8891 | tree lhs, rhs; |
8892 | basic_block store_bb = single_succ (bb: load_bb); |
8893 | gimple_stmt_iterator gsi; |
8894 | gimple *stmt; |
8895 | location_t loc; |
8896 | enum tree_code code; |
8897 | bool need_old, need_new; |
8898 | machine_mode imode; |
8899 | |
8900 | /* We expect to find the following sequences: |
8901 | |
8902 | load_bb: |
8903 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) |
8904 | |
8905 | store_bb: |
8906 | val = tmp OP something; (or: something OP tmp) |
8907 | GIMPLE_OMP_STORE (val) |
8908 | |
8909 | ???FIXME: Allow a more flexible sequence. |
8910 | Perhaps use data flow to pick the statements. |
8911 | |
8912 | */ |
8913 | |
8914 | gsi = gsi_after_labels (bb: store_bb); |
8915 | stmt = gsi_stmt (i: gsi); |
8916 | if (is_gimple_debug (gs: stmt)) |
8917 | { |
8918 | gsi_next_nondebug (i: &gsi); |
8919 | if (gsi_end_p (i: gsi)) |
8920 | return false; |
8921 | stmt = gsi_stmt (i: gsi); |
8922 | } |
8923 | loc = gimple_location (g: stmt); |
8924 | if (!is_gimple_assign (gs: stmt)) |
8925 | return false; |
8926 | gsi_next_nondebug (i: &gsi); |
8927 | if (gimple_code (g: gsi_stmt (i: gsi)) != GIMPLE_OMP_ATOMIC_STORE) |
8928 | return false; |
8929 | need_new = gimple_omp_atomic_need_value_p (g: gsi_stmt (i: gsi)); |
8930 | need_old = gimple_omp_atomic_need_value_p (g: last_nondebug_stmt (load_bb)); |
8931 | enum omp_memory_order omo |
8932 | = gimple_omp_atomic_memory_order (g: last_nondebug_stmt (load_bb)); |
8933 | enum memmodel mo = omp_memory_order_to_memmodel (mo: omo); |
8934 | gcc_checking_assert (!need_old || !need_new); |
8935 | |
8936 | if (!operand_equal_p (gimple_assign_lhs (gs: stmt), stored_val, flags: 0)) |
8937 | return false; |
8938 | |
8939 | /* Check for one of the supported fetch-op operations. */ |
8940 | code = gimple_assign_rhs_code (gs: stmt); |
8941 | switch (code) |
8942 | { |
8943 | case PLUS_EXPR: |
8944 | case POINTER_PLUS_EXPR: |
8945 | oldbase = BUILT_IN_ATOMIC_FETCH_ADD_N; |
8946 | newbase = BUILT_IN_ATOMIC_ADD_FETCH_N; |
8947 | break; |
8948 | case MINUS_EXPR: |
8949 | oldbase = BUILT_IN_ATOMIC_FETCH_SUB_N; |
8950 | newbase = BUILT_IN_ATOMIC_SUB_FETCH_N; |
8951 | break; |
8952 | case BIT_AND_EXPR: |
8953 | oldbase = BUILT_IN_ATOMIC_FETCH_AND_N; |
8954 | newbase = BUILT_IN_ATOMIC_AND_FETCH_N; |
8955 | break; |
8956 | case BIT_IOR_EXPR: |
8957 | oldbase = BUILT_IN_ATOMIC_FETCH_OR_N; |
8958 | newbase = BUILT_IN_ATOMIC_OR_FETCH_N; |
8959 | break; |
8960 | case BIT_XOR_EXPR: |
8961 | oldbase = BUILT_IN_ATOMIC_FETCH_XOR_N; |
8962 | newbase = BUILT_IN_ATOMIC_XOR_FETCH_N; |
8963 | break; |
8964 | default: |
8965 | return false; |
8966 | } |
8967 | |
8968 | /* Make sure the expression is of the proper form. */ |
8969 | if (operand_equal_p (gimple_assign_rhs1 (gs: stmt), loaded_val, flags: 0)) |
8970 | rhs = gimple_assign_rhs2 (gs: stmt); |
8971 | else if (commutative_tree_code (gimple_assign_rhs_code (gs: stmt)) |
8972 | && operand_equal_p (gimple_assign_rhs2 (gs: stmt), loaded_val, flags: 0)) |
8973 | rhs = gimple_assign_rhs1 (gs: stmt); |
8974 | else |
8975 | return false; |
8976 | |
8977 | tmpbase = ((enum built_in_function) |
8978 | ((need_new ? newbase : oldbase) + index + 1)); |
8979 | decl = builtin_decl_explicit (fncode: tmpbase); |
8980 | if (decl == NULL_TREE) |
8981 | return false; |
8982 | itype = TREE_TYPE (TREE_TYPE (decl)); |
8983 | imode = TYPE_MODE (itype); |
8984 | |
8985 | /* We could test all of the various optabs involved, but the fact of the |
8986 | matter is that (with the exception of i486 vs i586 and xadd) all targets |
8987 | that support any atomic operaton optab also implements compare-and-swap. |
8988 | Let optabs.cc take care of expanding any compare-and-swap loop. */ |
8989 | if (!can_compare_and_swap_p (imode, true) || !can_atomic_load_p (imode)) |
8990 | return false; |
8991 | |
8992 | gsi = gsi_last_nondebug_bb (bb: load_bb); |
8993 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_LOAD); |
8994 | |
8995 | /* OpenMP does not imply any barrier-like semantics on its atomic ops. |
8996 | It only requires that the operation happen atomically. Thus we can |
8997 | use the RELAXED memory model. */ |
8998 | call = build_call_expr_loc (loc, decl, 3, addr, |
8999 | fold_convert_loc (loc, itype, rhs), |
9000 | build_int_cst (NULL, mo)); |
9001 | |
9002 | if (need_old || need_new) |
9003 | { |
9004 | lhs = need_old ? loaded_val : stored_val; |
9005 | call = fold_convert_loc (loc, TREE_TYPE (lhs), call); |
9006 | call = build2_loc (loc, code: MODIFY_EXPR, void_type_node, arg0: lhs, arg1: call); |
9007 | } |
9008 | else |
9009 | call = fold_convert_loc (loc, void_type_node, call); |
9010 | force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT); |
9011 | gsi_remove (&gsi, true); |
9012 | |
9013 | gsi = gsi_last_nondebug_bb (bb: store_bb); |
9014 | gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE); |
9015 | gsi_remove (&gsi, true); |
9016 | gsi = gsi_last_nondebug_bb (bb: store_bb); |
9017 | stmt = gsi_stmt (i: gsi); |
9018 | gsi_remove (&gsi, true); |
9019 | |
9020 | if (gimple_in_ssa_p (cfun)) |
9021 | release_defs (stmt); |
9022 | |
9023 | return true; |
9024 | } |
9025 | |
9026 | /* A subroutine of expand_omp_atomic. Attempt to implement the atomic |
9027 | compare and exchange as an ATOMIC_COMPARE_EXCHANGE internal function. |
9028 | Returns false if the expression is not of the proper form. */ |
9029 | |
9030 | static bool |
9031 | expand_omp_atomic_cas (basic_block load_bb, tree addr, |
9032 | tree loaded_val, tree stored_val, int index) |
9033 | { |
9034 | /* We expect to find the following sequences: |
9035 | |
9036 | load_bb: |
9037 | GIMPLE_OMP_ATOMIC_LOAD (tmp, mem) |
9038 | |
9039 | store_bb: |
9040 | val = tmp == e ? d : tmp; |
9041 | GIMPLE_OMP_ATOMIC_STORE (val) |
9042 | |
9043 | or in store_bb instead: |
9044 | tmp2 = tmp == e; |
9045 | val = tmp2 ? d : tmp; |
9046 | GIMPLE_OMP_ATOMIC_STORE (val) |
9047 | |
9048 | or: |
9049 | tmp3 = VIEW_CONVERT_EXPR<integral_type>(tmp); |
9050 | val = e == tmp3 ? d : tmp; |
9051 | GIMPLE_OMP_ATOMIC_STORE (val) |
9052 | |
9053 | etc. */ |
9054 | |
9055 | |
9056 | basic_block store_bb = single_succ (bb: load_bb); |
9057 | gimple_stmt_iterator gsi = gsi_last_nondebug_bb (bb: store_bb); |
9058 | gimple *store_stmt = gsi_stmt (i: gsi); |
9059 | if (!store_stmt || gimple_code (g: store_stmt) != GIMPLE_OMP_ATOMIC_STORE) |
9060 | return false; |
9061 | gsi_prev_nondebug (i: &gsi); |
9062 | if (gsi_end_p (i: gsi)) |
9063 | return false; |
9064 | gimple *condexpr_stmt = gsi_stmt (i: gsi); |
9065 | if (!is_gimple_assign (gs: condexpr_stmt) |
9066 | || gimple_assign_rhs_code (gs: condexpr_stmt) != COND_EXPR) |
9067 | return false; |
9068 | if (!operand_equal_p (gimple_assign_lhs (gs: condexpr_stmt), stored_val, flags: 0)) |
9069 | return false; |
9070 | gimple *cond_stmt = NULL; |
9071 | gimple *vce_stmt = NULL; |
9072 | gsi_prev_nondebug (i: &gsi); |
9073 | if (!gsi_end_p (i: gsi)) |
9074 | { |
9075 | cond_stmt = gsi_stmt (i: gsi); |
9076 | if (!is_gimple_assign (gs: cond_stmt)) |
9077 | return false; |
9078 | if (gimple_assign_rhs_code (gs: cond_stmt) == EQ_EXPR) |
9079 | { |
9080 | gsi_prev_nondebug (i: &gsi); |
9081 | if (!gsi_end_p (i: gsi)) |
9082 | { |
9083 | vce_stmt = gsi_stmt (i: gsi); |
9084 | if (!is_gimple_assign (gs: vce_stmt) |
9085 | || gimple_assign_rhs_code (gs: vce_stmt) != VIEW_CONVERT_EXPR) |
9086 | return false; |
9087 | } |
9088 | } |
9089 | else if (gimple_assign_rhs_code (gs: cond_stmt) == VIEW_CONVERT_EXPR) |
9090 | std::swap (a&: vce_stmt, b&: cond_stmt); |
9091 | else |
9092 | return false; |
9093 | if (vce_stmt) |
9094 | { |
9095 | tree vce_rhs = gimple_assign_rhs1 (gs: vce_stmt); |
9096 | if (TREE_CODE (vce_rhs) != VIEW_CONVERT_EXPR |
9097 | || !operand_equal_p (TREE_OPERAND (vce_rhs, 0), loaded_val)) |
9098 | return false; |
9099 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vce_rhs)) |
9100 | || !SCALAR_FLOAT_TYPE_P (TREE_TYPE (loaded_val)) |
9101 | || !tree_int_cst_equal (TYPE_SIZE (TREE_TYPE (vce_rhs)), |
9102 | TYPE_SIZE (TREE_TYPE (loaded_val)))) |
9103 | return false; |
9104 | gsi_prev_nondebug (i: &gsi); |
9105 | if (!gsi_end_p (i: gsi)) |
9106 | return false; |
9107 | } |
9108 | } |
9109 | tree cond = gimple_assign_rhs1 (gs: condexpr_stmt); |
9110 | tree cond_op1, cond_op2; |
9111 | if (cond_stmt) |
9112 | { |
9113 | /* We should now always get a separate cond_stmt. */ |
9114 | if (!operand_equal_p (cond, gimple_assign_lhs (gs: cond_stmt))) |
9115 | return false; |
9116 | cond_op1 = gimple_assign_rhs1 (gs: cond_stmt); |
9117 | cond_op2 = gimple_assign_rhs2 (gs: cond_stmt); |
9118 | } |
9119 | else if (TREE_CODE (cond) != EQ_EXPR && TREE_CODE (cond) != NE_EXPR) |
9120 | return false; |
9121 | else |
9122 | { |
9123 | cond_op1 = TREE_OPERAND (cond, 0); |
9124 | cond_op2 = TREE_OPERAND (cond, 1); |
9125 | } |
9126 | tree d; |
9127 | if (TREE_CODE (cond) == NE_EXPR) |
9128 | { |
9129 | if (!operand_equal_p (gimple_assign_rhs2 (gs: condexpr_stmt), loaded_val)) |
9130 | return false; |
9131 | d = gimple_assign_rhs3 (gs: condexpr_stmt); |
9132 | } |
9133 | else if (!operand_equal_p (gimple_assign_rhs3 (gs: condexpr_stmt), loaded_val)) |
9134 | return false; |
9135 | else |
9136 | d = gimple_assign_rhs2 (gs: condexpr_stmt); |
9137 | tree e = vce_stmt ? gimple_assign_lhs (gs: vce_stmt) : loaded_val; |
9138 | if (operand_equal_p (e, cond_op1)) |
9139 | e = cond_op2; |
9140 | else if (operand_equal_p (e, cond_op2)) |
9141 | e = cond_op1; |
9142 | else |
9143 | return false; |
9144 | |
9145 | location_t loc = gimple_location (g: store_stmt); |
9146 | gimple *load_stmt = last_nondebug_stmt (load_bb); |
9147 | bool need_new = gimple_omp_atomic_need_value_p (g: store_stmt); |
9148 | bool need_old = gimple_omp_atomic_need_value_p (g: load_stmt); |
9149 | bool weak = gimple_omp_atomic_weak_p (g: load_stmt); |
9150 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (g: load_stmt); |
9151 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (mo: omo)); |
9152 | tree fmo = build_int_cst (NULL, omp_memory_order_to_fail_memmodel (mo: omo)); |
9153 | gcc_checking_assert (!need_old || !need_new); |
9154 | |
9155 | enum built_in_function fncode |
9156 | = (enum built_in_function) ((int) BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N |
9157 | + index + 1); |
9158 | tree cmpxchg = builtin_decl_explicit (fncode); |
9159 | if (cmpxchg == NULL_TREE) |
9160 | return false; |
9161 | tree itype = TREE_TYPE (TREE_TYPE (cmpxchg)); |
9162 | |
9163 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true) |
9164 | || !can_atomic_load_p (TYPE_MODE (itype))) |
9165 | return false; |
9166 | |
9167 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); |
9168 | if (SCALAR_FLOAT_TYPE_P (type) && !vce_stmt) |
9169 | return false; |
9170 | |
9171 | gsi = gsi_for_stmt (store_stmt); |
9172 | if (!useless_type_conversion_p (itype, TREE_TYPE (e))) |
9173 | { |
9174 | tree ne = create_tmp_reg (itype); |
9175 | gimple *g = gimple_build_assign (ne, NOP_EXPR, e); |
9176 | gimple_set_location (g, location: loc); |
9177 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9178 | e = ne; |
9179 | } |
9180 | if (!useless_type_conversion_p (itype, TREE_TYPE (d))) |
9181 | { |
9182 | tree nd = create_tmp_reg (itype); |
9183 | enum tree_code code; |
9184 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (d))) |
9185 | { |
9186 | code = VIEW_CONVERT_EXPR; |
9187 | d = build1 (VIEW_CONVERT_EXPR, itype, d); |
9188 | } |
9189 | else |
9190 | code = NOP_EXPR; |
9191 | gimple *g = gimple_build_assign (nd, code, d); |
9192 | gimple_set_location (g, location: loc); |
9193 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9194 | d = nd; |
9195 | } |
9196 | |
9197 | tree ctype = build_complex_type (itype); |
9198 | int flag = int_size_in_bytes (itype) + (weak ? 256 : 0); |
9199 | gimple *g |
9200 | = gimple_build_call_internal (IFN_ATOMIC_COMPARE_EXCHANGE, 6, addr, e, d, |
9201 | build_int_cst (integer_type_node, flag), |
9202 | mo, fmo); |
9203 | tree cres = create_tmp_reg (ctype); |
9204 | gimple_call_set_lhs (gs: g, lhs: cres); |
9205 | gimple_set_location (g, location: loc); |
9206 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9207 | |
9208 | if (cond_stmt || need_old || need_new) |
9209 | { |
9210 | tree im = create_tmp_reg (itype); |
9211 | g = gimple_build_assign (im, IMAGPART_EXPR, |
9212 | build1 (IMAGPART_EXPR, itype, cres)); |
9213 | gimple_set_location (g, location: loc); |
9214 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9215 | |
9216 | tree re = NULL_TREE; |
9217 | if (need_old || need_new) |
9218 | { |
9219 | re = create_tmp_reg (itype); |
9220 | g = gimple_build_assign (re, REALPART_EXPR, |
9221 | build1 (REALPART_EXPR, itype, cres)); |
9222 | gimple_set_location (g, location: loc); |
9223 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9224 | } |
9225 | |
9226 | if (cond_stmt) |
9227 | { |
9228 | g = gimple_build_assign (cond, NOP_EXPR, im); |
9229 | gimple_set_location (g, location: loc); |
9230 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9231 | } |
9232 | |
9233 | if (need_new) |
9234 | { |
9235 | g = gimple_build_assign (create_tmp_reg (itype), COND_EXPR, |
9236 | cond_stmt |
9237 | ? cond : build2 (NE_EXPR, boolean_type_node, |
9238 | im, build_zero_cst (itype)), |
9239 | d, re); |
9240 | gimple_set_location (g, location: loc); |
9241 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9242 | re = gimple_assign_lhs (gs: g); |
9243 | } |
9244 | |
9245 | if (need_old || need_new) |
9246 | { |
9247 | tree v = need_old ? loaded_val : stored_val; |
9248 | enum tree_code code; |
9249 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (v))) |
9250 | { |
9251 | code = VIEW_CONVERT_EXPR; |
9252 | re = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (v), re); |
9253 | } |
9254 | else if (!useless_type_conversion_p (TREE_TYPE (v), itype)) |
9255 | code = NOP_EXPR; |
9256 | else |
9257 | code = TREE_CODE (re); |
9258 | g = gimple_build_assign (v, code, re); |
9259 | gimple_set_location (g, location: loc); |
9260 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
9261 | } |
9262 | } |
9263 | |
9264 | gsi_remove (&gsi, true); |
9265 | gsi = gsi_for_stmt (load_stmt); |
9266 | gsi_remove (&gsi, true); |
9267 | gsi = gsi_for_stmt (condexpr_stmt); |
9268 | gsi_remove (&gsi, true); |
9269 | if (cond_stmt) |
9270 | { |
9271 | gsi = gsi_for_stmt (cond_stmt); |
9272 | gsi_remove (&gsi, true); |
9273 | } |
9274 | if (vce_stmt) |
9275 | { |
9276 | gsi = gsi_for_stmt (vce_stmt); |
9277 | gsi_remove (&gsi, true); |
9278 | } |
9279 | |
9280 | return true; |
9281 | } |
9282 | |
9283 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: |
9284 | |
9285 | oldval = *addr; |
9286 | repeat: |
9287 | newval = rhs; // with oldval replacing *addr in rhs |
9288 | oldval = __sync_val_compare_and_swap (addr, oldval, newval); |
9289 | if (oldval != newval) |
9290 | goto repeat; |
9291 | |
9292 | INDEX is log2 of the size of the data type, and thus usable to find the |
9293 | index of the builtin decl. */ |
9294 | |
9295 | static bool |
9296 | expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb, |
9297 | tree addr, tree loaded_val, tree stored_val, |
9298 | int index) |
9299 | { |
9300 | tree loadedi, storedi, initial, new_storedi, old_vali; |
9301 | tree type, itype, cmpxchg, iaddr, atype; |
9302 | gimple_stmt_iterator si; |
9303 | basic_block = single_succ (bb: load_bb); |
9304 | gimple *phi, *stmt; |
9305 | edge e; |
9306 | enum built_in_function fncode; |
9307 | |
9308 | fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N |
9309 | + index + 1); |
9310 | cmpxchg = builtin_decl_explicit (fncode); |
9311 | if (cmpxchg == NULL_TREE) |
9312 | return false; |
9313 | type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); |
9314 | atype = type; |
9315 | itype = TREE_TYPE (TREE_TYPE (cmpxchg)); |
9316 | |
9317 | if (!can_compare_and_swap_p (TYPE_MODE (itype), true) |
9318 | || !can_atomic_load_p (TYPE_MODE (itype))) |
9319 | return false; |
9320 | |
9321 | /* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */ |
9322 | si = gsi_last_nondebug_bb (bb: load_bb); |
9323 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); |
9324 | location_t loc = gimple_location (g: gsi_stmt (i: si)); |
9325 | enum omp_memory_order omo = gimple_omp_atomic_memory_order (g: gsi_stmt (i: si)); |
9326 | tree mo = build_int_cst (NULL, omp_memory_order_to_memmodel (mo: omo)); |
9327 | tree fmo = build_int_cst (NULL, omp_memory_order_to_fail_memmodel (mo: omo)); |
9328 | |
9329 | /* For floating-point values, we'll need to view-convert them to integers |
9330 | so that we can perform the atomic compare and swap. Simplify the |
9331 | following code by always setting up the "i"ntegral variables. */ |
9332 | if (!INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type)) |
9333 | { |
9334 | tree iaddr_val; |
9335 | |
9336 | iaddr = create_tmp_reg (build_pointer_type_for_mode (itype, ptr_mode, |
9337 | true)); |
9338 | atype = itype; |
9339 | iaddr_val |
9340 | = force_gimple_operand_gsi (&si, |
9341 | fold_convert (TREE_TYPE (iaddr), addr), |
9342 | false, NULL_TREE, true, GSI_SAME_STMT); |
9343 | stmt = gimple_build_assign (iaddr, iaddr_val); |
9344 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9345 | loadedi = create_tmp_var (itype); |
9346 | if (gimple_in_ssa_p (cfun)) |
9347 | loadedi = make_ssa_name (var: loadedi); |
9348 | } |
9349 | else |
9350 | { |
9351 | iaddr = addr; |
9352 | loadedi = loaded_val; |
9353 | } |
9354 | |
9355 | fncode = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1); |
9356 | tree loaddecl = builtin_decl_explicit (fncode); |
9357 | if (loaddecl) |
9358 | initial |
9359 | = fold_convert (atype, |
9360 | build_call_expr (loaddecl, 2, iaddr, |
9361 | build_int_cst (NULL_TREE, |
9362 | MEMMODEL_RELAXED))); |
9363 | else |
9364 | { |
9365 | tree off |
9366 | = build_int_cst (build_pointer_type_for_mode (atype, ptr_mode, |
9367 | true), 0); |
9368 | initial = build2 (MEM_REF, atype, iaddr, off); |
9369 | } |
9370 | |
9371 | initial |
9372 | = force_gimple_operand_gsi (&si, initial, true, NULL_TREE, true, |
9373 | GSI_SAME_STMT); |
9374 | |
9375 | /* Move the value to the LOADEDI temporary. */ |
9376 | if (gimple_in_ssa_p (cfun)) |
9377 | { |
9378 | gcc_assert (gimple_seq_empty_p (phi_nodes (loop_header))); |
9379 | phi = create_phi_node (loadedi, loop_header); |
9380 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)), |
9381 | initial); |
9382 | } |
9383 | else |
9384 | gsi_insert_before (&si, |
9385 | gimple_build_assign (loadedi, initial), |
9386 | GSI_SAME_STMT); |
9387 | if (loadedi != loaded_val) |
9388 | { |
9389 | gimple_stmt_iterator gsi2; |
9390 | tree x; |
9391 | |
9392 | x = build1 (VIEW_CONVERT_EXPR, type, loadedi); |
9393 | gsi2 = gsi_start_bb (bb: loop_header); |
9394 | if (gimple_in_ssa_p (cfun)) |
9395 | { |
9396 | gassign *stmt; |
9397 | x = force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, |
9398 | true, GSI_SAME_STMT); |
9399 | stmt = gimple_build_assign (loaded_val, x); |
9400 | gsi_insert_before (&gsi2, stmt, GSI_SAME_STMT); |
9401 | } |
9402 | else |
9403 | { |
9404 | x = build2 (MODIFY_EXPR, TREE_TYPE (loaded_val), loaded_val, x); |
9405 | force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE, |
9406 | true, GSI_SAME_STMT); |
9407 | } |
9408 | } |
9409 | gsi_remove (&si, true); |
9410 | |
9411 | si = gsi_last_nondebug_bb (bb: store_bb); |
9412 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); |
9413 | |
9414 | if (iaddr == addr) |
9415 | storedi = stored_val; |
9416 | else |
9417 | storedi |
9418 | = force_gimple_operand_gsi (&si, |
9419 | build1 (VIEW_CONVERT_EXPR, itype, |
9420 | stored_val), true, NULL_TREE, true, |
9421 | GSI_SAME_STMT); |
9422 | |
9423 | /* Build the compare&swap statement. */ |
9424 | tree ctype = build_complex_type (itype); |
9425 | int flag = int_size_in_bytes (itype); |
9426 | new_storedi = build_call_expr_internal_loc (loc, IFN_ATOMIC_COMPARE_EXCHANGE, |
9427 | ctype, 6, iaddr, loadedi, |
9428 | storedi, |
9429 | build_int_cst (integer_type_node, |
9430 | flag), |
9431 | mo, fmo); |
9432 | new_storedi = build1 (REALPART_EXPR, itype, new_storedi); |
9433 | new_storedi = force_gimple_operand_gsi (&si, |
9434 | fold_convert (TREE_TYPE (loadedi), |
9435 | new_storedi), |
9436 | true, NULL_TREE, |
9437 | true, GSI_SAME_STMT); |
9438 | |
9439 | if (gimple_in_ssa_p (cfun)) |
9440 | old_vali = loadedi; |
9441 | else |
9442 | { |
9443 | old_vali = create_tmp_var (TREE_TYPE (loadedi)); |
9444 | stmt = gimple_build_assign (old_vali, loadedi); |
9445 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9446 | |
9447 | stmt = gimple_build_assign (loadedi, new_storedi); |
9448 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9449 | } |
9450 | |
9451 | /* Note that we always perform the comparison as an integer, even for |
9452 | floating point. This allows the atomic operation to properly |
9453 | succeed even with NaNs and -0.0. */ |
9454 | tree ne = build2 (NE_EXPR, boolean_type_node, new_storedi, old_vali); |
9455 | stmt = gimple_build_cond_empty (cond: ne); |
9456 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9457 | |
9458 | /* Update cfg. */ |
9459 | e = single_succ_edge (bb: store_bb); |
9460 | e->flags &= ~EDGE_FALLTHRU; |
9461 | e->flags |= EDGE_FALSE_VALUE; |
9462 | /* Expect no looping. */ |
9463 | e->probability = profile_probability::guessed_always (); |
9464 | |
9465 | e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE); |
9466 | e->probability = profile_probability::guessed_never (); |
9467 | |
9468 | /* Copy the new value to loadedi (we already did that before the condition |
9469 | if we are not in SSA). */ |
9470 | if (gimple_in_ssa_p (cfun)) |
9471 | { |
9472 | phi = gimple_seq_first_stmt (s: phi_nodes (bb: loop_header)); |
9473 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_storedi); |
9474 | } |
9475 | |
9476 | /* Remove GIMPLE_OMP_ATOMIC_STORE. */ |
9477 | stmt = gsi_stmt (i: si); |
9478 | gsi_remove (&si, true); |
9479 | if (gimple_in_ssa_p (cfun)) |
9480 | release_defs (stmt); |
9481 | |
9482 | class loop *loop = alloc_loop (); |
9483 | loop->header = loop_header; |
9484 | loop->latch = store_bb; |
9485 | add_loop (loop, loop_header->loop_father); |
9486 | |
9487 | return true; |
9488 | } |
9489 | |
9490 | /* A subroutine of expand_omp_atomic. Implement the atomic operation as: |
9491 | |
9492 | GOMP_atomic_start (); |
9493 | *addr = rhs; |
9494 | GOMP_atomic_end (); |
9495 | |
9496 | The result is not globally atomic, but works so long as all parallel |
9497 | references are within #pragma omp atomic directives. According to |
9498 | responses received from omp@openmp.org, appears to be within spec. |
9499 | Which makes sense, since that's how several other compilers handle |
9500 | this situation as well. |
9501 | LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're |
9502 | expanding. STORED_VAL is the operand of the matching |
9503 | GIMPLE_OMP_ATOMIC_STORE. |
9504 | |
9505 | We replace |
9506 | GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with |
9507 | loaded_val = *addr; |
9508 | |
9509 | and replace |
9510 | GIMPLE_OMP_ATOMIC_STORE (stored_val) with |
9511 | *addr = stored_val; |
9512 | */ |
9513 | |
9514 | static bool |
9515 | expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb, |
9516 | tree addr, tree loaded_val, tree stored_val) |
9517 | { |
9518 | gimple_stmt_iterator si; |
9519 | gassign *stmt; |
9520 | tree t; |
9521 | |
9522 | si = gsi_last_nondebug_bb (bb: load_bb); |
9523 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD); |
9524 | |
9525 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_ATOMIC_START); |
9526 | t = build_call_expr (t, 0); |
9527 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); |
9528 | |
9529 | tree mem = build_simple_mem_ref (addr); |
9530 | TREE_TYPE (mem) = TREE_TYPE (loaded_val); |
9531 | TREE_OPERAND (mem, 1) |
9532 | = fold_convert (build_pointer_type_for_mode (TREE_TYPE (mem), ptr_mode, |
9533 | true), |
9534 | TREE_OPERAND (mem, 1)); |
9535 | stmt = gimple_build_assign (loaded_val, mem); |
9536 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9537 | gsi_remove (&si, true); |
9538 | |
9539 | si = gsi_last_nondebug_bb (bb: store_bb); |
9540 | gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE); |
9541 | |
9542 | stmt = gimple_build_assign (unshare_expr (mem), stored_val); |
9543 | gimple_set_vuse (g: stmt, vuse: gimple_vuse (g: gsi_stmt (i: si))); |
9544 | gimple_set_vdef (g: stmt, vdef: gimple_vdef (g: gsi_stmt (i: si))); |
9545 | gsi_insert_before (&si, stmt, GSI_SAME_STMT); |
9546 | |
9547 | t = builtin_decl_explicit (fncode: BUILT_IN_GOMP_ATOMIC_END); |
9548 | t = build_call_expr (t, 0); |
9549 | force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT); |
9550 | gsi_remove (&si, true); |
9551 | return true; |
9552 | } |
9553 | |
9554 | /* Expand an GIMPLE_OMP_ATOMIC statement. We try to expand |
9555 | using expand_omp_atomic_fetch_op. If it failed, we try to |
9556 | call expand_omp_atomic_pipeline, and if it fails too, the |
9557 | ultimate fallback is wrapping the operation in a mutex |
9558 | (expand_omp_atomic_mutex). REGION is the atomic region built |
9559 | by build_omp_regions_1(). */ |
9560 | |
9561 | static void |
9562 | expand_omp_atomic (struct omp_region *region) |
9563 | { |
9564 | basic_block load_bb = region->entry, store_bb = region->exit; |
9565 | gomp_atomic_load *load |
9566 | = as_a <gomp_atomic_load *> (p: last_nondebug_stmt (load_bb)); |
9567 | gomp_atomic_store *store |
9568 | = as_a <gomp_atomic_store *> (p: last_nondebug_stmt (store_bb)); |
9569 | tree loaded_val = gimple_omp_atomic_load_lhs (load_stmt: load); |
9570 | tree addr = gimple_omp_atomic_load_rhs (load_stmt: load); |
9571 | tree stored_val = gimple_omp_atomic_store_val (store_stmt: store); |
9572 | tree type = TYPE_MAIN_VARIANT (TREE_TYPE (loaded_val)); |
9573 | HOST_WIDE_INT index; |
9574 | |
9575 | /* Make sure the type is one of the supported sizes. */ |
9576 | index = tree_to_uhwi (TYPE_SIZE_UNIT (type)); |
9577 | index = exact_log2 (x: index); |
9578 | if (index >= 0 && index <= 4) |
9579 | { |
9580 | unsigned int align = TYPE_ALIGN_UNIT (type); |
9581 | |
9582 | /* __sync builtins require strict data alignment. */ |
9583 | if (exact_log2 (x: align) >= index) |
9584 | { |
9585 | /* Atomic load. */ |
9586 | scalar_mode smode; |
9587 | if (loaded_val == stored_val |
9588 | && (is_int_mode (TYPE_MODE (type), int_mode: &smode) |
9589 | || is_float_mode (TYPE_MODE (type), float_mode: &smode)) |
9590 | && GET_MODE_BITSIZE (mode: smode) <= BITS_PER_WORD |
9591 | && expand_omp_atomic_load (load_bb, addr, loaded_val, index)) |
9592 | return; |
9593 | |
9594 | /* Atomic store. */ |
9595 | if ((is_int_mode (TYPE_MODE (type), int_mode: &smode) |
9596 | || is_float_mode (TYPE_MODE (type), float_mode: &smode)) |
9597 | && GET_MODE_BITSIZE (mode: smode) <= BITS_PER_WORD |
9598 | && store_bb == single_succ (bb: load_bb) |
9599 | && first_stmt (store_bb) == store |
9600 | && expand_omp_atomic_store (load_bb, addr, loaded_val, |
9601 | stored_val, index)) |
9602 | return; |
9603 | |
9604 | /* When possible, use specialized atomic update functions. */ |
9605 | if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type)) |
9606 | && store_bb == single_succ (bb: load_bb) |
9607 | && expand_omp_atomic_fetch_op (load_bb, addr, |
9608 | loaded_val, stored_val, index)) |
9609 | return; |
9610 | |
9611 | /* When possible, use ATOMIC_COMPARE_EXCHANGE ifn without a loop. */ |
9612 | if (store_bb == single_succ (bb: load_bb) |
9613 | && !gimple_in_ssa_p (cfun) |
9614 | && expand_omp_atomic_cas (load_bb, addr, loaded_val, stored_val, |
9615 | index)) |
9616 | return; |
9617 | |
9618 | /* If we don't have specialized __sync builtins, try and implement |
9619 | as a compare and swap loop. */ |
9620 | if (expand_omp_atomic_pipeline (load_bb, store_bb, addr, |
9621 | loaded_val, stored_val, index)) |
9622 | return; |
9623 | } |
9624 | } |
9625 | |
9626 | /* The ultimate fallback is wrapping the operation in a mutex. */ |
9627 | expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val); |
9628 | } |
9629 | |
9630 | /* Mark the loops inside the kernels region starting at REGION_ENTRY and ending |
9631 | at REGION_EXIT. */ |
9632 | |
9633 | static void |
9634 | mark_loops_in_oacc_kernels_region (basic_block region_entry, |
9635 | basic_block region_exit) |
9636 | { |
9637 | class loop *outer = region_entry->loop_father; |
9638 | gcc_assert (region_exit == NULL || outer == region_exit->loop_father); |
9639 | |
9640 | /* Don't parallelize the kernels region if it contains more than one outer |
9641 | loop. */ |
9642 | unsigned int nr_outer_loops = 0; |
9643 | class loop *single_outer = NULL; |
9644 | for (class loop *loop = outer->inner; loop != NULL; loop = loop->next) |
9645 | { |
9646 | gcc_assert (loop_outer (loop) == outer); |
9647 | |
9648 | if (!dominated_by_p (CDI_DOMINATORS, loop->header, region_entry)) |
9649 | continue; |
9650 | |
9651 | if (region_exit != NULL |
9652 | && dominated_by_p (CDI_DOMINATORS, loop->header, region_exit)) |
9653 | continue; |
9654 | |
9655 | nr_outer_loops++; |
9656 | single_outer = loop; |
9657 | } |
9658 | if (nr_outer_loops != 1) |
9659 | return; |
9660 | |
9661 | for (class loop *loop = single_outer->inner; |
9662 | loop != NULL; |
9663 | loop = loop->inner) |
9664 | if (loop->next) |
9665 | return; |
9666 | |
9667 | /* Mark the loops in the region. */ |
9668 | for (class loop *loop = single_outer; loop != NULL; loop = loop->inner) |
9669 | loop->in_oacc_kernels_region = true; |
9670 | } |
9671 | |
9672 | /* Build target argument identifier from the DEVICE identifier, value |
9673 | identifier ID and whether the element also has a SUBSEQUENT_PARAM. */ |
9674 | |
9675 | static tree |
9676 | get_target_argument_identifier_1 (int device, bool subseqent_param, int id) |
9677 | { |
9678 | tree t = build_int_cst (integer_type_node, device); |
9679 | if (subseqent_param) |
9680 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, |
9681 | build_int_cst (integer_type_node, |
9682 | GOMP_TARGET_ARG_SUBSEQUENT_PARAM)); |
9683 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, |
9684 | build_int_cst (integer_type_node, id)); |
9685 | return t; |
9686 | } |
9687 | |
9688 | /* Like above but return it in type that can be directly stored as an element |
9689 | of the argument array. */ |
9690 | |
9691 | static tree |
9692 | get_target_argument_identifier (int device, bool subseqent_param, int id) |
9693 | { |
9694 | tree t = get_target_argument_identifier_1 (device, subseqent_param, id); |
9695 | return fold_convert (ptr_type_node, t); |
9696 | } |
9697 | |
9698 | /* Return a target argument consisting of DEVICE identifier, value identifier |
9699 | ID, and the actual VALUE. */ |
9700 | |
9701 | static tree |
9702 | get_target_argument_value (gimple_stmt_iterator *gsi, int device, int id, |
9703 | tree value) |
9704 | { |
9705 | tree t = fold_build2 (LSHIFT_EXPR, integer_type_node, |
9706 | fold_convert (integer_type_node, value), |
9707 | build_int_cst (unsigned_type_node, |
9708 | GOMP_TARGET_ARG_VALUE_SHIFT)); |
9709 | t = fold_build2 (BIT_IOR_EXPR, integer_type_node, t, |
9710 | get_target_argument_identifier_1 (device, false, id)); |
9711 | t = fold_convert (ptr_type_node, t); |
9712 | return force_gimple_operand_gsi (gsi, t, true, NULL, true, GSI_SAME_STMT); |
9713 | } |
9714 | |
9715 | /* If VALUE is an integer constant greater than -2^15 and smaller than 2^15, |
9716 | push one argument to ARGS with both the DEVICE, ID and VALUE embedded in it, |
9717 | otherwise push an identifier (with DEVICE and ID) and the VALUE in two |
9718 | arguments. */ |
9719 | |
9720 | static void |
9721 | push_target_argument_according_to_value (gimple_stmt_iterator *gsi, int device, |
9722 | int id, tree value, vec <tree> *args) |
9723 | { |
9724 | if (tree_fits_shwi_p (value) |
9725 | && tree_to_shwi (value) > -(1 << 15) |
9726 | && tree_to_shwi (value) < (1 << 15)) |
9727 | args->quick_push (obj: get_target_argument_value (gsi, device, id, value)); |
9728 | else |
9729 | { |
9730 | args->quick_push (obj: get_target_argument_identifier (device, subseqent_param: true, id)); |
9731 | value = fold_convert (ptr_type_node, value); |
9732 | value = force_gimple_operand_gsi (gsi, value, true, NULL, true, |
9733 | GSI_SAME_STMT); |
9734 | args->quick_push (obj: value); |
9735 | } |
9736 | } |
9737 | |
9738 | /* Create an array of arguments that is then passed to GOMP_target. */ |
9739 | |
9740 | static tree |
9741 | get_target_arguments (gimple_stmt_iterator *gsi, gomp_target *tgt_stmt) |
9742 | { |
9743 | auto_vec <tree, 6> args; |
9744 | tree clauses = gimple_omp_target_clauses (gs: tgt_stmt); |
9745 | tree t, c = omp_find_clause (clauses, kind: OMP_CLAUSE_NUM_TEAMS); |
9746 | if (c) |
9747 | t = OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR (c); |
9748 | else |
9749 | t = integer_minus_one_node; |
9750 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, |
9751 | GOMP_TARGET_ARG_NUM_TEAMS, value: t, args: &args); |
9752 | |
9753 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_THREAD_LIMIT); |
9754 | if (c) |
9755 | t = OMP_CLAUSE_THREAD_LIMIT_EXPR (c); |
9756 | else |
9757 | t = integer_minus_one_node; |
9758 | push_target_argument_according_to_value (gsi, GOMP_TARGET_ARG_DEVICE_ALL, |
9759 | GOMP_TARGET_ARG_THREAD_LIMIT, value: t, |
9760 | args: &args); |
9761 | |
9762 | /* Produce more, perhaps device specific, arguments here. */ |
9763 | |
9764 | tree argarray = create_tmp_var (build_array_type_nelts (ptr_type_node, |
9765 | args.length () + 1), |
9766 | ".omp_target_args" ); |
9767 | for (unsigned i = 0; i < args.length (); i++) |
9768 | { |
9769 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, |
9770 | build_int_cst (integer_type_node, i), |
9771 | NULL_TREE, NULL_TREE); |
9772 | gsi_insert_before (gsi, gimple_build_assign (ref, args[i]), |
9773 | GSI_SAME_STMT); |
9774 | } |
9775 | tree ref = build4 (ARRAY_REF, ptr_type_node, argarray, |
9776 | build_int_cst (integer_type_node, args.length ()), |
9777 | NULL_TREE, NULL_TREE); |
9778 | gsi_insert_before (gsi, gimple_build_assign (ref, null_pointer_node), |
9779 | GSI_SAME_STMT); |
9780 | TREE_ADDRESSABLE (argarray) = 1; |
9781 | return build_fold_addr_expr (argarray); |
9782 | } |
9783 | |
9784 | /* Expand the GIMPLE_OMP_TARGET starting at REGION. */ |
9785 | |
9786 | static void |
9787 | expand_omp_target (struct omp_region *region) |
9788 | { |
9789 | basic_block entry_bb, exit_bb, new_bb; |
9790 | struct function *child_cfun; |
9791 | tree child_fn, child_fn2, block, t, c; |
9792 | gimple_stmt_iterator gsi; |
9793 | gomp_target *entry_stmt; |
9794 | gimple *stmt; |
9795 | edge e; |
9796 | bool offloaded; |
9797 | int target_kind; |
9798 | |
9799 | entry_stmt = as_a <gomp_target *> (p: last_nondebug_stmt (region->entry)); |
9800 | target_kind = gimple_omp_target_kind (g: entry_stmt); |
9801 | new_bb = region->entry; |
9802 | |
9803 | offloaded = is_gimple_omp_offloaded (stmt: entry_stmt); |
9804 | switch (target_kind) |
9805 | { |
9806 | case GF_OMP_TARGET_KIND_REGION: |
9807 | case GF_OMP_TARGET_KIND_UPDATE: |
9808 | case GF_OMP_TARGET_KIND_ENTER_DATA: |
9809 | case GF_OMP_TARGET_KIND_EXIT_DATA: |
9810 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
9811 | case GF_OMP_TARGET_KIND_OACC_KERNELS: |
9812 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
9813 | case GF_OMP_TARGET_KIND_OACC_UPDATE: |
9814 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
9815 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: |
9816 | case GF_OMP_TARGET_KIND_OACC_DECLARE: |
9817 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
9818 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: |
9819 | case GF_OMP_TARGET_KIND_DATA: |
9820 | case GF_OMP_TARGET_KIND_OACC_DATA: |
9821 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: |
9822 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: |
9823 | break; |
9824 | default: |
9825 | gcc_unreachable (); |
9826 | } |
9827 | |
9828 | tree clauses = gimple_omp_target_clauses (gs: entry_stmt); |
9829 | |
9830 | bool is_ancestor = false; |
9831 | child_fn = child_fn2 = NULL_TREE; |
9832 | child_cfun = NULL; |
9833 | if (offloaded) |
9834 | { |
9835 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_DEVICE); |
9836 | if (ENABLE_OFFLOADING && c) |
9837 | is_ancestor = OMP_CLAUSE_DEVICE_ANCESTOR (c); |
9838 | child_fn = gimple_omp_target_child_fn (omp_target_stmt: entry_stmt); |
9839 | child_cfun = DECL_STRUCT_FUNCTION (child_fn); |
9840 | } |
9841 | |
9842 | /* Supported by expand_omp_taskreg, but not here. */ |
9843 | if (child_cfun != NULL) |
9844 | gcc_checking_assert (!child_cfun->cfg); |
9845 | gcc_checking_assert (!gimple_in_ssa_p (cfun)); |
9846 | |
9847 | entry_bb = region->entry; |
9848 | exit_bb = region->exit; |
9849 | |
9850 | if (target_kind == GF_OMP_TARGET_KIND_OACC_KERNELS) |
9851 | mark_loops_in_oacc_kernels_region (region_entry: region->entry, region_exit: region->exit); |
9852 | |
9853 | /* Going on, all OpenACC compute constructs are mapped to |
9854 | 'BUILT_IN_GOACC_PARALLEL', and get their compute regions outlined. |
9855 | To distinguish between them, we attach attributes. */ |
9856 | switch (target_kind) |
9857 | { |
9858 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
9859 | DECL_ATTRIBUTES (child_fn) |
9860 | = tree_cons (get_identifier ("oacc parallel" ), |
9861 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); |
9862 | break; |
9863 | case GF_OMP_TARGET_KIND_OACC_KERNELS: |
9864 | DECL_ATTRIBUTES (child_fn) |
9865 | = tree_cons (get_identifier ("oacc kernels" ), |
9866 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); |
9867 | break; |
9868 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
9869 | DECL_ATTRIBUTES (child_fn) |
9870 | = tree_cons (get_identifier ("oacc serial" ), |
9871 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); |
9872 | break; |
9873 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
9874 | DECL_ATTRIBUTES (child_fn) |
9875 | = tree_cons (get_identifier ("oacc parallel_kernels_parallelized" ), |
9876 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); |
9877 | break; |
9878 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: |
9879 | DECL_ATTRIBUTES (child_fn) |
9880 | = tree_cons (get_identifier ("oacc parallel_kernels_gang_single" ), |
9881 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); |
9882 | break; |
9883 | default: |
9884 | /* Make sure we don't miss any. */ |
9885 | gcc_checking_assert (!(is_gimple_omp_oacc (entry_stmt) |
9886 | && is_gimple_omp_offloaded (entry_stmt))); |
9887 | break; |
9888 | } |
9889 | |
9890 | if (offloaded) |
9891 | { |
9892 | unsigned srcidx, dstidx, num; |
9893 | |
9894 | /* If the offloading region needs data sent from the parent |
9895 | function, then the very first statement (except possible |
9896 | tree profile counter updates) of the offloading body |
9897 | is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since |
9898 | &.OMP_DATA_O is passed as an argument to the child function, |
9899 | we need to replace it with the argument as seen by the child |
9900 | function. |
9901 | |
9902 | In most cases, this will end up being the identity assignment |
9903 | .OMP_DATA_I = .OMP_DATA_I. However, if the offloading body had |
9904 | a function call that has been inlined, the original PARM_DECL |
9905 | .OMP_DATA_I may have been converted into a different local |
9906 | variable. In which case, we need to keep the assignment. */ |
9907 | tree data_arg = gimple_omp_target_data_arg (omp_target_stmt: entry_stmt); |
9908 | if (data_arg) |
9909 | { |
9910 | basic_block entry_succ_bb = single_succ (bb: entry_bb); |
9911 | gimple_stmt_iterator gsi; |
9912 | tree arg; |
9913 | gimple *tgtcopy_stmt = NULL; |
9914 | tree sender = TREE_VEC_ELT (data_arg, 0); |
9915 | |
9916 | for (gsi = gsi_start_bb (bb: entry_succ_bb); ; gsi_next (i: &gsi)) |
9917 | { |
9918 | gcc_assert (!gsi_end_p (gsi)); |
9919 | stmt = gsi_stmt (i: gsi); |
9920 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
9921 | continue; |
9922 | |
9923 | if (gimple_num_ops (gs: stmt) == 2) |
9924 | { |
9925 | tree arg = gimple_assign_rhs1 (gs: stmt); |
9926 | |
9927 | /* We're ignoring the subcode because we're |
9928 | effectively doing a STRIP_NOPS. */ |
9929 | |
9930 | if (TREE_CODE (arg) == ADDR_EXPR |
9931 | && TREE_OPERAND (arg, 0) == sender) |
9932 | { |
9933 | tgtcopy_stmt = stmt; |
9934 | break; |
9935 | } |
9936 | } |
9937 | } |
9938 | |
9939 | gcc_assert (tgtcopy_stmt != NULL); |
9940 | arg = DECL_ARGUMENTS (child_fn); |
9941 | |
9942 | gcc_assert (gimple_assign_lhs (tgtcopy_stmt) == arg); |
9943 | gsi_remove (&gsi, true); |
9944 | } |
9945 | |
9946 | /* Declare local variables needed in CHILD_CFUN. */ |
9947 | block = DECL_INITIAL (child_fn); |
9948 | BLOCK_VARS (block) = vec2chain (v: child_cfun->local_decls); |
9949 | /* The gimplifier could record temporaries in the offloading block |
9950 | rather than in containing function's local_decls chain, |
9951 | which would mean cgraph missed finalizing them. Do it now. */ |
9952 | for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t)) |
9953 | if (VAR_P (t) && TREE_STATIC (t) && !DECL_EXTERNAL (t)) |
9954 | varpool_node::finalize_decl (decl: t); |
9955 | DECL_SAVED_TREE (child_fn) = NULL; |
9956 | /* We'll create a CFG for child_fn, so no gimple body is needed. */ |
9957 | gimple_set_body (child_fn, NULL); |
9958 | TREE_USED (block) = 1; |
9959 | |
9960 | /* Reset DECL_CONTEXT on function arguments. */ |
9961 | for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t)) |
9962 | DECL_CONTEXT (t) = child_fn; |
9963 | |
9964 | /* Split ENTRY_BB at GIMPLE_*, |
9965 | so that it can be moved to the child function. */ |
9966 | gsi = gsi_last_nondebug_bb (bb: entry_bb); |
9967 | stmt = gsi_stmt (i: gsi); |
9968 | gcc_assert (stmt |
9969 | && gimple_code (stmt) == gimple_code (entry_stmt)); |
9970 | e = split_block (entry_bb, stmt); |
9971 | gsi_remove (&gsi, true); |
9972 | entry_bb = e->dest; |
9973 | single_succ_edge (bb: entry_bb)->flags = EDGE_FALLTHRU; |
9974 | |
9975 | /* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */ |
9976 | if (exit_bb) |
9977 | { |
9978 | gsi = gsi_last_nondebug_bb (bb: exit_bb); |
9979 | gcc_assert (!gsi_end_p (gsi) |
9980 | && gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN); |
9981 | stmt = gimple_build_return (NULL); |
9982 | gsi_insert_after (&gsi, stmt, GSI_SAME_STMT); |
9983 | gsi_remove (&gsi, true); |
9984 | } |
9985 | |
9986 | /* Move the offloading region into CHILD_CFUN. */ |
9987 | |
9988 | block = gimple_block (g: entry_stmt); |
9989 | |
9990 | new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block); |
9991 | if (exit_bb) |
9992 | single_succ_edge (bb: new_bb)->flags = EDGE_FALLTHRU; |
9993 | /* When the OMP expansion process cannot guarantee an up-to-date |
9994 | loop tree arrange for the child function to fixup loops. */ |
9995 | if (loops_state_satisfies_p (flags: LOOPS_NEED_FIXUP)) |
9996 | child_cfun->x_current_loops->state |= LOOPS_NEED_FIXUP; |
9997 | |
9998 | /* Remove non-local VAR_DECLs from child_cfun->local_decls list. */ |
9999 | num = vec_safe_length (v: child_cfun->local_decls); |
10000 | for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++) |
10001 | { |
10002 | t = (*child_cfun->local_decls)[srcidx]; |
10003 | if (DECL_CONTEXT (t) == cfun->decl) |
10004 | continue; |
10005 | if (srcidx != dstidx) |
10006 | (*child_cfun->local_decls)[dstidx] = t; |
10007 | dstidx++; |
10008 | } |
10009 | if (dstidx != num) |
10010 | vec_safe_truncate (v: child_cfun->local_decls, size: dstidx); |
10011 | |
10012 | /* Inform the callgraph about the new function. */ |
10013 | child_cfun->curr_properties = cfun->curr_properties; |
10014 | child_cfun->has_simduid_loops |= cfun->has_simduid_loops; |
10015 | child_cfun->has_force_vectorize_loops |= cfun->has_force_vectorize_loops; |
10016 | cgraph_node *node = cgraph_node::get_create (child_fn); |
10017 | node->parallelized_function = 1; |
10018 | cgraph_node::add_new_function (fndecl: child_fn, lowered: true); |
10019 | |
10020 | /* Add the new function to the offload table. */ |
10021 | if (ENABLE_OFFLOADING) |
10022 | { |
10023 | if (in_lto_p) |
10024 | DECL_PRESERVE_P (child_fn) = 1; |
10025 | if (!is_ancestor) |
10026 | vec_safe_push (v&: offload_funcs, obj: child_fn); |
10027 | } |
10028 | |
10029 | bool need_asm = DECL_ASSEMBLER_NAME_SET_P (current_function_decl) |
10030 | && !DECL_ASSEMBLER_NAME_SET_P (child_fn); |
10031 | |
10032 | /* Fix the callgraph edges for child_cfun. Those for cfun will be |
10033 | fixed in a following pass. */ |
10034 | push_cfun (new_cfun: child_cfun); |
10035 | if (need_asm) |
10036 | assign_assembler_name_if_needed (child_fn); |
10037 | cgraph_edge::rebuild_edges (); |
10038 | |
10039 | /* Some EH regions might become dead, see PR34608. If |
10040 | pass_cleanup_cfg isn't the first pass to happen with the |
10041 | new child, these dead EH edges might cause problems. |
10042 | Clean them up now. */ |
10043 | if (flag_exceptions) |
10044 | { |
10045 | basic_block bb; |
10046 | bool changed = false; |
10047 | |
10048 | FOR_EACH_BB_FN (bb, cfun) |
10049 | changed |= gimple_purge_dead_eh_edges (bb); |
10050 | if (changed) |
10051 | cleanup_tree_cfg (); |
10052 | } |
10053 | if (flag_checking && !loops_state_satisfies_p (flags: LOOPS_NEED_FIXUP)) |
10054 | verify_loop_structure (); |
10055 | pop_cfun (); |
10056 | |
10057 | if (dump_file && !gimple_in_ssa_p (cfun)) |
10058 | { |
10059 | omp_any_child_fn_dumped = true; |
10060 | dump_function_header (dump_file, child_fn, dump_flags); |
10061 | dump_function_to_file (child_fn, dump_file, dump_flags); |
10062 | } |
10063 | |
10064 | adjust_context_and_scope (region, entry_block: gimple_block (g: entry_stmt), child_fndecl: child_fn); |
10065 | |
10066 | /* Handle the case that an inner ancestor:1 target is called by an outer |
10067 | target region. */ |
10068 | if (is_ancestor) |
10069 | { |
10070 | cgraph_node *fn2_node; |
10071 | child_fn2 = build_decl (DECL_SOURCE_LOCATION (child_fn), |
10072 | FUNCTION_DECL, |
10073 | clone_function_name (decl: child_fn, suffix: "nohost" ), |
10074 | TREE_TYPE (child_fn)); |
10075 | if (in_lto_p) |
10076 | DECL_PRESERVE_P (child_fn2) = 1; |
10077 | TREE_STATIC (child_fn2) = 1; |
10078 | DECL_ARTIFICIAL (child_fn2) = 1; |
10079 | DECL_IGNORED_P (child_fn2) = 0; |
10080 | TREE_PUBLIC (child_fn2) = 0; |
10081 | DECL_UNINLINABLE (child_fn2) = 1; |
10082 | DECL_EXTERNAL (child_fn2) = 0; |
10083 | DECL_CONTEXT (child_fn2) = DECL_CONTEXT (child_fn); |
10084 | DECL_INITIAL (child_fn2) = make_node (BLOCK); |
10085 | BLOCK_SUPERCONTEXT (DECL_INITIAL (child_fn2)) = child_fn2; |
10086 | DECL_ATTRIBUTES (child_fn) |
10087 | = remove_attribute ("omp target entrypoint" , |
10088 | DECL_ATTRIBUTES (child_fn)); |
10089 | DECL_ATTRIBUTES (child_fn2) |
10090 | = tree_cons (get_identifier ("omp target device_ancestor_nohost" ), |
10091 | NULL_TREE, copy_list (DECL_ATTRIBUTES (child_fn))); |
10092 | DECL_ATTRIBUTES (child_fn) |
10093 | = tree_cons (get_identifier ("omp target device_ancestor_host" ), |
10094 | NULL_TREE, DECL_ATTRIBUTES (child_fn)); |
10095 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (child_fn2) |
10096 | = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (current_function_decl); |
10097 | DECL_FUNCTION_SPECIFIC_TARGET (child_fn2) |
10098 | = DECL_FUNCTION_SPECIFIC_TARGET (current_function_decl); |
10099 | DECL_FUNCTION_VERSIONED (child_fn2) |
10100 | = DECL_FUNCTION_VERSIONED (current_function_decl); |
10101 | |
10102 | fn2_node = cgraph_node::get_create (child_fn2); |
10103 | fn2_node->offloadable = 1; |
10104 | fn2_node->force_output = 1; |
10105 | node->offloadable = 0; |
10106 | |
10107 | /* Enable pass_omp_device_lower pass. */ |
10108 | fn2_node = cgraph_node::get (DECL_CONTEXT (child_fn)); |
10109 | fn2_node->calls_declare_variant_alt = 1; |
10110 | |
10111 | t = build_decl (DECL_SOURCE_LOCATION (child_fn), |
10112 | RESULT_DECL, NULL_TREE, void_type_node); |
10113 | DECL_ARTIFICIAL (t) = 1; |
10114 | DECL_IGNORED_P (t) = 1; |
10115 | DECL_CONTEXT (t) = child_fn2; |
10116 | DECL_RESULT (child_fn2) = t; |
10117 | DECL_SAVED_TREE (child_fn2) = build1 (RETURN_EXPR, |
10118 | void_type_node, NULL); |
10119 | tree tmp = DECL_ARGUMENTS (child_fn); |
10120 | t = build_decl (DECL_SOURCE_LOCATION (child_fn), PARM_DECL, |
10121 | DECL_NAME (tmp), TREE_TYPE (tmp)); |
10122 | DECL_ARTIFICIAL (t) = 1; |
10123 | DECL_NAMELESS (t) = 1; |
10124 | DECL_ARG_TYPE (t) = ptr_type_node; |
10125 | DECL_CONTEXT (t) = current_function_decl; |
10126 | TREE_USED (t) = 1; |
10127 | TREE_READONLY (t) = 1; |
10128 | DECL_ARGUMENTS (child_fn2) = t; |
10129 | gcc_assert (TREE_CHAIN (tmp) == NULL_TREE); |
10130 | |
10131 | gimplify_function_tree (child_fn2); |
10132 | cgraph_node::add_new_function (fndecl: child_fn2, lowered: true); |
10133 | |
10134 | vec_safe_push (v&: offload_funcs, obj: child_fn2); |
10135 | if (dump_file && !gimple_in_ssa_p (cfun)) |
10136 | { |
10137 | dump_function_header (dump_file, child_fn2, dump_flags); |
10138 | dump_function_to_file (child_fn2, dump_file, dump_flags); |
10139 | } |
10140 | } |
10141 | } |
10142 | |
10143 | /* Emit a library call to launch the offloading region, or do data |
10144 | transfers. */ |
10145 | tree t1, t2, t3, t4, depend; |
10146 | enum built_in_function start_ix; |
10147 | unsigned int flags_i = 0; |
10148 | |
10149 | switch (gimple_omp_target_kind (g: entry_stmt)) |
10150 | { |
10151 | case GF_OMP_TARGET_KIND_REGION: |
10152 | start_ix = BUILT_IN_GOMP_TARGET; |
10153 | break; |
10154 | case GF_OMP_TARGET_KIND_DATA: |
10155 | start_ix = BUILT_IN_GOMP_TARGET_DATA; |
10156 | break; |
10157 | case GF_OMP_TARGET_KIND_UPDATE: |
10158 | start_ix = BUILT_IN_GOMP_TARGET_UPDATE; |
10159 | break; |
10160 | case GF_OMP_TARGET_KIND_ENTER_DATA: |
10161 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; |
10162 | break; |
10163 | case GF_OMP_TARGET_KIND_EXIT_DATA: |
10164 | start_ix = BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA; |
10165 | flags_i |= GOMP_TARGET_FLAG_EXIT_DATA; |
10166 | break; |
10167 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
10168 | case GF_OMP_TARGET_KIND_OACC_KERNELS: |
10169 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
10170 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
10171 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: |
10172 | start_ix = BUILT_IN_GOACC_PARALLEL; |
10173 | break; |
10174 | case GF_OMP_TARGET_KIND_OACC_DATA: |
10175 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: |
10176 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: |
10177 | start_ix = BUILT_IN_GOACC_DATA_START; |
10178 | break; |
10179 | case GF_OMP_TARGET_KIND_OACC_UPDATE: |
10180 | start_ix = BUILT_IN_GOACC_UPDATE; |
10181 | break; |
10182 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
10183 | start_ix = BUILT_IN_GOACC_ENTER_DATA; |
10184 | break; |
10185 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: |
10186 | start_ix = BUILT_IN_GOACC_EXIT_DATA; |
10187 | break; |
10188 | case GF_OMP_TARGET_KIND_OACC_DECLARE: |
10189 | start_ix = BUILT_IN_GOACC_DECLARE; |
10190 | break; |
10191 | default: |
10192 | gcc_unreachable (); |
10193 | } |
10194 | |
10195 | tree device = NULL_TREE; |
10196 | location_t device_loc = UNKNOWN_LOCATION; |
10197 | tree goacc_flags = NULL_TREE; |
10198 | bool need_device_adjustment = false; |
10199 | gimple_stmt_iterator adj_gsi; |
10200 | if (is_gimple_omp_oacc (stmt: entry_stmt)) |
10201 | { |
10202 | /* By default, no GOACC_FLAGs are set. */ |
10203 | goacc_flags = integer_zero_node; |
10204 | } |
10205 | else |
10206 | { |
10207 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_DEVICE); |
10208 | if (c) |
10209 | { |
10210 | device = OMP_CLAUSE_DEVICE_ID (c); |
10211 | /* Ensure 'device' is of the correct type. */ |
10212 | device = fold_convert_loc (device_loc, integer_type_node, device); |
10213 | if (TREE_CODE (device) == INTEGER_CST) |
10214 | { |
10215 | if (wi::to_wide (t: device) == GOMP_DEVICE_ICV) |
10216 | device = build_int_cst (integer_type_node, |
10217 | GOMP_DEVICE_HOST_FALLBACK); |
10218 | else if (wi::to_wide (t: device) == GOMP_DEVICE_HOST_FALLBACK) |
10219 | device = build_int_cst (integer_type_node, |
10220 | GOMP_DEVICE_HOST_FALLBACK - 1); |
10221 | } |
10222 | else |
10223 | need_device_adjustment = true; |
10224 | device_loc = OMP_CLAUSE_LOCATION (c); |
10225 | if (OMP_CLAUSE_DEVICE_ANCESTOR (c)) |
10226 | device = build_int_cst (integer_type_node, |
10227 | GOMP_DEVICE_HOST_FALLBACK); |
10228 | } |
10229 | else |
10230 | { |
10231 | /* By default, the value of DEVICE is GOMP_DEVICE_ICV (let runtime |
10232 | library choose). */ |
10233 | device = build_int_cst (integer_type_node, GOMP_DEVICE_ICV); |
10234 | device_loc = gimple_location (g: entry_stmt); |
10235 | } |
10236 | |
10237 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_NOWAIT); |
10238 | /* FIXME: in_reduction(...) nowait is unimplemented yet, pretend |
10239 | nowait doesn't appear. */ |
10240 | if (c && omp_find_clause (clauses, kind: OMP_CLAUSE_IN_REDUCTION)) |
10241 | c = NULL; |
10242 | if (c) |
10243 | flags_i |= GOMP_TARGET_FLAG_NOWAIT; |
10244 | } |
10245 | |
10246 | /* By default, there is no conditional. */ |
10247 | tree cond = NULL_TREE; |
10248 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_IF); |
10249 | if (c) |
10250 | cond = OMP_CLAUSE_IF_EXPR (c); |
10251 | /* If we found the clause 'if (cond)', build: |
10252 | OpenACC: goacc_flags = (cond ? goacc_flags |
10253 | : goacc_flags | GOACC_FLAG_HOST_FALLBACK) |
10254 | OpenMP: device = (cond ? device : GOMP_DEVICE_HOST_FALLBACK) */ |
10255 | if (cond) |
10256 | { |
10257 | tree *tp; |
10258 | if (is_gimple_omp_oacc (stmt: entry_stmt)) |
10259 | tp = &goacc_flags; |
10260 | else |
10261 | tp = &device; |
10262 | |
10263 | cond = gimple_boolify (cond); |
10264 | |
10265 | basic_block cond_bb, then_bb, else_bb; |
10266 | edge e; |
10267 | tree tmp_var = create_tmp_var (TREE_TYPE (*tp)); |
10268 | if (offloaded) |
10269 | e = split_block_after_labels (new_bb); |
10270 | else |
10271 | { |
10272 | gsi = gsi_last_nondebug_bb (bb: new_bb); |
10273 | gsi_prev (i: &gsi); |
10274 | e = split_block (new_bb, gsi_stmt (i: gsi)); |
10275 | } |
10276 | cond_bb = e->src; |
10277 | new_bb = e->dest; |
10278 | remove_edge (e); |
10279 | |
10280 | then_bb = create_empty_bb (cond_bb); |
10281 | else_bb = create_empty_bb (then_bb); |
10282 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
10283 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); |
10284 | |
10285 | stmt = gimple_build_cond_empty (cond); |
10286 | gsi = gsi_last_bb (bb: cond_bb); |
10287 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
10288 | |
10289 | gsi = gsi_start_bb (bb: then_bb); |
10290 | stmt = gimple_build_assign (tmp_var, *tp); |
10291 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
10292 | adj_gsi = gsi; |
10293 | |
10294 | gsi = gsi_start_bb (bb: else_bb); |
10295 | if (is_gimple_omp_oacc (stmt: entry_stmt)) |
10296 | stmt = gimple_build_assign (tmp_var, |
10297 | BIT_IOR_EXPR, |
10298 | *tp, |
10299 | build_int_cst (integer_type_node, |
10300 | GOACC_FLAG_HOST_FALLBACK)); |
10301 | else |
10302 | stmt = gimple_build_assign (tmp_var, |
10303 | build_int_cst (integer_type_node, |
10304 | GOMP_DEVICE_HOST_FALLBACK)); |
10305 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
10306 | |
10307 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); |
10308 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); |
10309 | add_bb_to_loop (then_bb, cond_bb->loop_father); |
10310 | add_bb_to_loop (else_bb, cond_bb->loop_father); |
10311 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); |
10312 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); |
10313 | |
10314 | *tp = tmp_var; |
10315 | |
10316 | gsi = gsi_last_nondebug_bb (bb: new_bb); |
10317 | } |
10318 | else |
10319 | { |
10320 | gsi = gsi_last_nondebug_bb (bb: new_bb); |
10321 | |
10322 | if (device != NULL_TREE) |
10323 | device = force_gimple_operand_gsi (&gsi, device, true, NULL_TREE, |
10324 | true, GSI_SAME_STMT); |
10325 | if (need_device_adjustment) |
10326 | { |
10327 | tree tmp_var = create_tmp_var (TREE_TYPE (device)); |
10328 | stmt = gimple_build_assign (tmp_var, device); |
10329 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
10330 | adj_gsi = gsi_for_stmt (stmt); |
10331 | device = tmp_var; |
10332 | } |
10333 | } |
10334 | |
10335 | if ((c = omp_find_clause (clauses, kind: OMP_CLAUSE_SELF)) != NULL_TREE) |
10336 | { |
10337 | gcc_assert ((is_gimple_omp_oacc (entry_stmt) && offloaded) |
10338 | || (gimple_omp_target_kind (entry_stmt) |
10339 | == GF_OMP_TARGET_KIND_OACC_DATA_KERNELS)); |
10340 | |
10341 | edge e; |
10342 | if (offloaded) |
10343 | e = split_block_after_labels (new_bb); |
10344 | else |
10345 | { |
10346 | gsi = gsi_last_nondebug_bb (bb: new_bb); |
10347 | gsi_prev (i: &gsi); |
10348 | e = split_block (new_bb, gsi_stmt (i: gsi)); |
10349 | } |
10350 | basic_block cond_bb = e->src; |
10351 | new_bb = e->dest; |
10352 | remove_edge (e); |
10353 | |
10354 | basic_block then_bb = create_empty_bb (cond_bb); |
10355 | basic_block else_bb = create_empty_bb (then_bb); |
10356 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
10357 | set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb); |
10358 | |
10359 | tree self_cond = gimple_boolify (OMP_CLAUSE_SELF_EXPR (c)); |
10360 | stmt = gimple_build_cond_empty (cond: self_cond); |
10361 | gsi = gsi_last_bb (bb: cond_bb); |
10362 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
10363 | |
10364 | tree tmp_var = create_tmp_var (TREE_TYPE (goacc_flags)); |
10365 | stmt = gimple_build_assign (tmp_var, BIT_IOR_EXPR, goacc_flags, |
10366 | build_int_cst (integer_type_node, |
10367 | GOACC_FLAG_LOCAL_DEVICE)); |
10368 | gsi = gsi_start_bb (bb: then_bb); |
10369 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
10370 | |
10371 | gsi = gsi_start_bb (bb: else_bb); |
10372 | stmt = gimple_build_assign (tmp_var, goacc_flags); |
10373 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
10374 | |
10375 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); |
10376 | make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE); |
10377 | add_bb_to_loop (then_bb, cond_bb->loop_father); |
10378 | add_bb_to_loop (else_bb, cond_bb->loop_father); |
10379 | make_edge (then_bb, new_bb, EDGE_FALLTHRU); |
10380 | make_edge (else_bb, new_bb, EDGE_FALLTHRU); |
10381 | |
10382 | goacc_flags = tmp_var; |
10383 | gsi = gsi_last_nondebug_bb (bb: new_bb); |
10384 | } |
10385 | |
10386 | if (need_device_adjustment) |
10387 | { |
10388 | tree uns = fold_convert (unsigned_type_node, device); |
10389 | uns = force_gimple_operand_gsi (&adj_gsi, uns, true, NULL_TREE, |
10390 | false, GSI_CONTINUE_LINKING); |
10391 | edge e = split_block (gsi_bb (i: adj_gsi), gsi_stmt (i: adj_gsi)); |
10392 | basic_block cond_bb = e->src; |
10393 | basic_block else_bb = e->dest; |
10394 | if (gsi_bb (i: adj_gsi) == new_bb) |
10395 | { |
10396 | new_bb = else_bb; |
10397 | gsi = gsi_last_nondebug_bb (bb: new_bb); |
10398 | } |
10399 | |
10400 | basic_block then_bb = create_empty_bb (cond_bb); |
10401 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
10402 | |
10403 | cond = build2 (GT_EXPR, boolean_type_node, uns, |
10404 | build_int_cst (unsigned_type_node, |
10405 | GOMP_DEVICE_HOST_FALLBACK - 1)); |
10406 | stmt = gimple_build_cond_empty (cond); |
10407 | adj_gsi = gsi_last_bb (bb: cond_bb); |
10408 | gsi_insert_after (&adj_gsi, stmt, GSI_CONTINUE_LINKING); |
10409 | |
10410 | adj_gsi = gsi_start_bb (bb: then_bb); |
10411 | tree add = build2 (PLUS_EXPR, integer_type_node, device, |
10412 | build_int_cst (integer_type_node, -1)); |
10413 | stmt = gimple_build_assign (device, add); |
10414 | gsi_insert_after (&adj_gsi, stmt, GSI_CONTINUE_LINKING); |
10415 | |
10416 | make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE); |
10417 | e->flags = EDGE_FALSE_VALUE; |
10418 | add_bb_to_loop (then_bb, cond_bb->loop_father); |
10419 | make_edge (then_bb, else_bb, EDGE_FALLTHRU); |
10420 | } |
10421 | |
10422 | t = gimple_omp_target_data_arg (omp_target_stmt: entry_stmt); |
10423 | if (t == NULL) |
10424 | { |
10425 | t1 = size_zero_node; |
10426 | t2 = build_zero_cst (ptr_type_node); |
10427 | t3 = t2; |
10428 | t4 = t2; |
10429 | } |
10430 | else |
10431 | { |
10432 | t1 = TYPE_MAX_VALUE (TYPE_DOMAIN (TREE_TYPE (TREE_VEC_ELT (t, 1)))); |
10433 | t1 = size_binop (PLUS_EXPR, t1, size_int (1)); |
10434 | t2 = build_fold_addr_expr (TREE_VEC_ELT (t, 0)); |
10435 | t3 = build_fold_addr_expr (TREE_VEC_ELT (t, 1)); |
10436 | t4 = build_fold_addr_expr (TREE_VEC_ELT (t, 2)); |
10437 | } |
10438 | |
10439 | gimple *g; |
10440 | bool tagging = false; |
10441 | /* The maximum number used by any start_ix, without varargs. */ |
10442 | auto_vec<tree, 11> args; |
10443 | if (is_gimple_omp_oacc (stmt: entry_stmt)) |
10444 | { |
10445 | tree goacc_flags_m = fold_build1 (GOACC_FLAGS_MARSHAL_OP, |
10446 | TREE_TYPE (goacc_flags), goacc_flags); |
10447 | goacc_flags_m = force_gimple_operand_gsi (&gsi, goacc_flags_m, true, |
10448 | NULL_TREE, true, |
10449 | GSI_SAME_STMT); |
10450 | args.quick_push (obj: goacc_flags_m); |
10451 | } |
10452 | else |
10453 | args.quick_push (obj: device); |
10454 | if (offloaded) |
10455 | args.quick_push (build_fold_addr_expr (child_fn2 ? child_fn2 : child_fn)); |
10456 | args.quick_push (obj: t1); |
10457 | args.quick_push (obj: t2); |
10458 | args.quick_push (obj: t3); |
10459 | args.quick_push (obj: t4); |
10460 | switch (start_ix) |
10461 | { |
10462 | case BUILT_IN_GOACC_DATA_START: |
10463 | case BUILT_IN_GOACC_DECLARE: |
10464 | case BUILT_IN_GOMP_TARGET_DATA: |
10465 | break; |
10466 | case BUILT_IN_GOMP_TARGET: |
10467 | case BUILT_IN_GOMP_TARGET_UPDATE: |
10468 | case BUILT_IN_GOMP_TARGET_ENTER_EXIT_DATA: |
10469 | args.quick_push (obj: build_int_cst (unsigned_type_node, flags_i)); |
10470 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_DEPEND); |
10471 | if (c) |
10472 | depend = OMP_CLAUSE_DECL (c); |
10473 | else |
10474 | depend = build_int_cst (ptr_type_node, 0); |
10475 | args.quick_push (obj: depend); |
10476 | if (start_ix == BUILT_IN_GOMP_TARGET) |
10477 | args.quick_push (obj: get_target_arguments (gsi: &gsi, tgt_stmt: entry_stmt)); |
10478 | break; |
10479 | case BUILT_IN_GOACC_PARALLEL: |
10480 | if (lookup_attribute (attr_name: "oacc serial" , DECL_ATTRIBUTES (child_fn)) != NULL) |
10481 | { |
10482 | tree dims = NULL_TREE; |
10483 | unsigned int ix; |
10484 | |
10485 | /* For serial constructs we set all dimensions to 1. */ |
10486 | for (ix = GOMP_DIM_MAX; ix--;) |
10487 | dims = tree_cons (NULL_TREE, integer_one_node, dims); |
10488 | oacc_replace_fn_attrib (fn: child_fn, dims); |
10489 | } |
10490 | else |
10491 | oacc_set_fn_attrib (fn: child_fn, clauses, args: &args); |
10492 | tagging = true; |
10493 | /* FALLTHRU */ |
10494 | case BUILT_IN_GOACC_ENTER_DATA: |
10495 | case BUILT_IN_GOACC_EXIT_DATA: |
10496 | case BUILT_IN_GOACC_UPDATE: |
10497 | { |
10498 | tree t_async = NULL_TREE; |
10499 | |
10500 | /* If present, use the value specified by the respective |
10501 | clause, making sure that is of the correct type. */ |
10502 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_ASYNC); |
10503 | if (c) |
10504 | t_async = fold_convert_loc (OMP_CLAUSE_LOCATION (c), |
10505 | integer_type_node, |
10506 | OMP_CLAUSE_ASYNC_EXPR (c)); |
10507 | else if (!tagging) |
10508 | /* Default values for t_async. */ |
10509 | t_async = fold_convert_loc (gimple_location (g: entry_stmt), |
10510 | integer_type_node, |
10511 | build_int_cst (integer_type_node, |
10512 | GOMP_ASYNC_SYNC)); |
10513 | if (tagging && t_async) |
10514 | { |
10515 | unsigned HOST_WIDE_INT i_async = GOMP_LAUNCH_OP_MAX; |
10516 | |
10517 | if (TREE_CODE (t_async) == INTEGER_CST) |
10518 | { |
10519 | /* See if we can pack the async arg in to the tag's |
10520 | operand. */ |
10521 | i_async = TREE_INT_CST_LOW (t_async); |
10522 | if (i_async < GOMP_LAUNCH_OP_MAX) |
10523 | t_async = NULL_TREE; |
10524 | else |
10525 | i_async = GOMP_LAUNCH_OP_MAX; |
10526 | } |
10527 | args.safe_push (obj: oacc_launch_pack (GOMP_LAUNCH_ASYNC, NULL_TREE, |
10528 | op: i_async)); |
10529 | } |
10530 | if (t_async) |
10531 | args.safe_push (obj: force_gimple_operand_gsi (&gsi, t_async, true, |
10532 | NULL_TREE, true, |
10533 | GSI_SAME_STMT)); |
10534 | |
10535 | /* Save the argument index, and ... */ |
10536 | unsigned t_wait_idx = args.length (); |
10537 | unsigned num_waits = 0; |
10538 | c = omp_find_clause (clauses, kind: OMP_CLAUSE_WAIT); |
10539 | if (!tagging || c) |
10540 | /* ... push a placeholder. */ |
10541 | args.safe_push (integer_zero_node); |
10542 | |
10543 | for (; c; c = OMP_CLAUSE_CHAIN (c)) |
10544 | if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_WAIT) |
10545 | { |
10546 | tree arg = fold_convert_loc (OMP_CLAUSE_LOCATION (c), |
10547 | integer_type_node, |
10548 | OMP_CLAUSE_WAIT_EXPR (c)); |
10549 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, true, |
10550 | GSI_SAME_STMT); |
10551 | args.safe_push (obj: arg); |
10552 | num_waits++; |
10553 | } |
10554 | |
10555 | if (!tagging || num_waits) |
10556 | { |
10557 | tree len; |
10558 | |
10559 | /* Now that we know the number, update the placeholder. */ |
10560 | if (tagging) |
10561 | len = oacc_launch_pack (GOMP_LAUNCH_WAIT, NULL_TREE, op: num_waits); |
10562 | else |
10563 | len = build_int_cst (integer_type_node, num_waits); |
10564 | len = fold_convert_loc (gimple_location (g: entry_stmt), |
10565 | unsigned_type_node, len); |
10566 | args[t_wait_idx] = len; |
10567 | } |
10568 | } |
10569 | break; |
10570 | default: |
10571 | gcc_unreachable (); |
10572 | } |
10573 | if (tagging) |
10574 | /* Push terminal marker - zero. */ |
10575 | args.safe_push (obj: oacc_launch_pack (code: 0, NULL_TREE, op: 0)); |
10576 | |
10577 | if (child_fn2) |
10578 | { |
10579 | g = gimple_build_call_internal (IFN_GOMP_TARGET_REV, 1, |
10580 | build_fold_addr_expr (child_fn)); |
10581 | gimple_set_location (g, location: gimple_location (g: entry_stmt)); |
10582 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
10583 | } |
10584 | |
10585 | g = gimple_build_call_vec (builtin_decl_explicit (fncode: start_ix), args); |
10586 | gimple_set_location (g, location: gimple_location (g: entry_stmt)); |
10587 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
10588 | if (!offloaded) |
10589 | { |
10590 | g = gsi_stmt (i: gsi); |
10591 | gcc_assert (g && gimple_code (g) == GIMPLE_OMP_TARGET); |
10592 | gsi_remove (&gsi, true); |
10593 | } |
10594 | } |
10595 | |
10596 | /* Expand the parallel region tree rooted at REGION. Expansion |
10597 | proceeds in depth-first order. Innermost regions are expanded |
10598 | first. This way, parallel regions that require a new function to |
10599 | be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any |
10600 | internal dependencies in their body. */ |
10601 | |
10602 | static void |
10603 | expand_omp (struct omp_region *region) |
10604 | { |
10605 | omp_any_child_fn_dumped = false; |
10606 | while (region) |
10607 | { |
10608 | location_t saved_location; |
10609 | gimple *inner_stmt = NULL; |
10610 | |
10611 | /* First, determine whether this is a combined parallel+workshare |
10612 | region. */ |
10613 | if (region->type == GIMPLE_OMP_PARALLEL) |
10614 | determine_parallel_type (region); |
10615 | |
10616 | if (region->type == GIMPLE_OMP_FOR |
10617 | && gimple_omp_for_combined_p (g: last_nondebug_stmt (region->entry))) |
10618 | inner_stmt = last_nondebug_stmt (region->inner->entry); |
10619 | |
10620 | if (region->inner) |
10621 | expand_omp (region: region->inner); |
10622 | |
10623 | saved_location = input_location; |
10624 | if (gimple_has_location (g: last_nondebug_stmt (region->entry))) |
10625 | input_location = gimple_location (g: last_nondebug_stmt (region->entry)); |
10626 | |
10627 | switch (region->type) |
10628 | { |
10629 | case GIMPLE_OMP_PARALLEL: |
10630 | case GIMPLE_OMP_TASK: |
10631 | expand_omp_taskreg (region); |
10632 | break; |
10633 | |
10634 | case GIMPLE_OMP_FOR: |
10635 | expand_omp_for (region, inner_stmt); |
10636 | break; |
10637 | |
10638 | case GIMPLE_OMP_SECTIONS: |
10639 | expand_omp_sections (region); |
10640 | break; |
10641 | |
10642 | case GIMPLE_OMP_SECTION: |
10643 | /* Individual omp sections are handled together with their |
10644 | parent GIMPLE_OMP_SECTIONS region. */ |
10645 | break; |
10646 | |
10647 | case GIMPLE_OMP_STRUCTURED_BLOCK: |
10648 | /* We should have gotten rid of these in gimple lowering. */ |
10649 | gcc_unreachable (); |
10650 | |
10651 | case GIMPLE_OMP_SINGLE: |
10652 | case GIMPLE_OMP_SCOPE: |
10653 | expand_omp_single (region); |
10654 | break; |
10655 | |
10656 | case GIMPLE_OMP_ORDERED: |
10657 | { |
10658 | gomp_ordered *ord_stmt |
10659 | = as_a <gomp_ordered *> (p: last_nondebug_stmt (region->entry)); |
10660 | if (gimple_omp_ordered_standalone_p (g: ord_stmt)) |
10661 | { |
10662 | /* We'll expand these when expanding corresponding |
10663 | worksharing region with ordered(n) clause. */ |
10664 | gcc_assert (region->outer |
10665 | && region->outer->type == GIMPLE_OMP_FOR); |
10666 | region->ord_stmt = ord_stmt; |
10667 | break; |
10668 | } |
10669 | } |
10670 | /* FALLTHRU */ |
10671 | case GIMPLE_OMP_MASTER: |
10672 | case GIMPLE_OMP_MASKED: |
10673 | case GIMPLE_OMP_TASKGROUP: |
10674 | case GIMPLE_OMP_CRITICAL: |
10675 | case GIMPLE_OMP_TEAMS: |
10676 | expand_omp_synch (region); |
10677 | break; |
10678 | |
10679 | case GIMPLE_OMP_ATOMIC_LOAD: |
10680 | expand_omp_atomic (region); |
10681 | break; |
10682 | |
10683 | case GIMPLE_OMP_TARGET: |
10684 | expand_omp_target (region); |
10685 | break; |
10686 | |
10687 | default: |
10688 | gcc_unreachable (); |
10689 | } |
10690 | |
10691 | input_location = saved_location; |
10692 | region = region->next; |
10693 | } |
10694 | if (omp_any_child_fn_dumped) |
10695 | { |
10696 | if (dump_file) |
10697 | dump_function_header (dump_file, current_function_decl, dump_flags); |
10698 | omp_any_child_fn_dumped = false; |
10699 | } |
10700 | } |
10701 | |
10702 | /* Helper for build_omp_regions. Scan the dominator tree starting at |
10703 | block BB. PARENT is the region that contains BB. If SINGLE_TREE is |
10704 | true, the function ends once a single tree is built (otherwise, whole |
10705 | forest of OMP constructs may be built). */ |
10706 | |
10707 | static void |
10708 | build_omp_regions_1 (basic_block bb, struct omp_region *parent, |
10709 | bool single_tree) |
10710 | { |
10711 | gimple_stmt_iterator gsi; |
10712 | gimple *stmt; |
10713 | basic_block son; |
10714 | |
10715 | gsi = gsi_last_nondebug_bb (bb); |
10716 | if (!gsi_end_p (i: gsi) && is_gimple_omp (stmt: gsi_stmt (i: gsi))) |
10717 | { |
10718 | struct omp_region *region; |
10719 | enum gimple_code code; |
10720 | |
10721 | stmt = gsi_stmt (i: gsi); |
10722 | code = gimple_code (g: stmt); |
10723 | if (code == GIMPLE_OMP_RETURN) |
10724 | { |
10725 | /* STMT is the return point out of region PARENT. Mark it |
10726 | as the exit point and make PARENT the immediately |
10727 | enclosing region. */ |
10728 | gcc_assert (parent); |
10729 | region = parent; |
10730 | region->exit = bb; |
10731 | parent = parent->outer; |
10732 | } |
10733 | else if (code == GIMPLE_OMP_ATOMIC_STORE) |
10734 | { |
10735 | /* GIMPLE_OMP_ATOMIC_STORE is analogous to |
10736 | GIMPLE_OMP_RETURN, but matches with |
10737 | GIMPLE_OMP_ATOMIC_LOAD. */ |
10738 | gcc_assert (parent); |
10739 | gcc_assert (parent->type == GIMPLE_OMP_ATOMIC_LOAD); |
10740 | region = parent; |
10741 | region->exit = bb; |
10742 | parent = parent->outer; |
10743 | } |
10744 | else if (code == GIMPLE_OMP_CONTINUE) |
10745 | { |
10746 | gcc_assert (parent); |
10747 | parent->cont = bb; |
10748 | } |
10749 | else if (code == GIMPLE_OMP_SECTIONS_SWITCH) |
10750 | { |
10751 | /* GIMPLE_OMP_SECTIONS_SWITCH is part of |
10752 | GIMPLE_OMP_SECTIONS, and we do nothing for it. */ |
10753 | } |
10754 | else |
10755 | { |
10756 | region = new_omp_region (bb, type: code, parent); |
10757 | /* Otherwise... */ |
10758 | if (code == GIMPLE_OMP_TARGET) |
10759 | { |
10760 | switch (gimple_omp_target_kind (g: stmt)) |
10761 | { |
10762 | case GF_OMP_TARGET_KIND_REGION: |
10763 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
10764 | case GF_OMP_TARGET_KIND_OACC_KERNELS: |
10765 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
10766 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
10767 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: |
10768 | break; |
10769 | case GF_OMP_TARGET_KIND_UPDATE: |
10770 | case GF_OMP_TARGET_KIND_ENTER_DATA: |
10771 | case GF_OMP_TARGET_KIND_EXIT_DATA: |
10772 | case GF_OMP_TARGET_KIND_DATA: |
10773 | case GF_OMP_TARGET_KIND_OACC_DATA: |
10774 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: |
10775 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: |
10776 | case GF_OMP_TARGET_KIND_OACC_UPDATE: |
10777 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
10778 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: |
10779 | case GF_OMP_TARGET_KIND_OACC_DECLARE: |
10780 | /* ..., other than for those stand-alone directives... |
10781 | To be precise, target data isn't stand-alone, but |
10782 | gimplifier put the end API call into try finally block |
10783 | for it, so omp expansion can treat it as such. */ |
10784 | region = NULL; |
10785 | break; |
10786 | default: |
10787 | gcc_unreachable (); |
10788 | } |
10789 | } |
10790 | else if (code == GIMPLE_OMP_ORDERED |
10791 | && gimple_omp_ordered_standalone_p (g: stmt)) |
10792 | /* #pragma omp ordered depend is also just a stand-alone |
10793 | directive. */ |
10794 | region = NULL; |
10795 | else if (code == GIMPLE_OMP_TASK |
10796 | && gimple_omp_task_taskwait_p (g: stmt)) |
10797 | /* #pragma omp taskwait depend(...) is a stand-alone directive. */ |
10798 | region = NULL; |
10799 | else if (code == GIMPLE_OMP_TASKGROUP) |
10800 | /* #pragma omp taskgroup isn't a stand-alone directive, but |
10801 | gimplifier put the end API call into try finall block |
10802 | for it, so omp expansion can treat it as such. */ |
10803 | region = NULL; |
10804 | /* ..., this directive becomes the parent for a new region. */ |
10805 | if (region) |
10806 | parent = region; |
10807 | } |
10808 | } |
10809 | |
10810 | if (single_tree && !parent) |
10811 | return; |
10812 | |
10813 | for (son = first_dom_son (CDI_DOMINATORS, bb); |
10814 | son; |
10815 | son = next_dom_son (CDI_DOMINATORS, son)) |
10816 | build_omp_regions_1 (bb: son, parent, single_tree); |
10817 | } |
10818 | |
10819 | /* Builds the tree of OMP regions rooted at ROOT, storing it to |
10820 | root_omp_region. */ |
10821 | |
10822 | static void |
10823 | build_omp_regions_root (basic_block root) |
10824 | { |
10825 | gcc_assert (root_omp_region == NULL); |
10826 | build_omp_regions_1 (bb: root, NULL, single_tree: true); |
10827 | gcc_assert (root_omp_region != NULL); |
10828 | } |
10829 | |
10830 | /* Expands omp construct (and its subconstructs) starting in HEAD. */ |
10831 | |
10832 | void |
10833 | omp_expand_local (basic_block head) |
10834 | { |
10835 | build_omp_regions_root (root: head); |
10836 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10837 | { |
10838 | fprintf (stream: dump_file, format: "\nOMP region tree\n\n" ); |
10839 | dump_omp_region (file: dump_file, region: root_omp_region, indent: 0); |
10840 | fprintf (stream: dump_file, format: "\n" ); |
10841 | } |
10842 | |
10843 | remove_exit_barriers (region: root_omp_region); |
10844 | expand_omp (region: root_omp_region); |
10845 | |
10846 | omp_free_regions (); |
10847 | } |
10848 | |
10849 | /* Scan the CFG and build a tree of OMP regions. Return the root of |
10850 | the OMP region tree. */ |
10851 | |
10852 | static void |
10853 | build_omp_regions (void) |
10854 | { |
10855 | gcc_assert (root_omp_region == NULL); |
10856 | calculate_dominance_info (CDI_DOMINATORS); |
10857 | build_omp_regions_1 (ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, single_tree: false); |
10858 | } |
10859 | |
10860 | /* Main entry point for expanding OMP-GIMPLE into runtime calls. */ |
10861 | |
10862 | static unsigned int |
10863 | execute_expand_omp (void) |
10864 | { |
10865 | build_omp_regions (); |
10866 | |
10867 | if (!root_omp_region) |
10868 | return 0; |
10869 | |
10870 | if (dump_file) |
10871 | { |
10872 | fprintf (stream: dump_file, format: "\nOMP region tree\n\n" ); |
10873 | dump_omp_region (file: dump_file, region: root_omp_region, indent: 0); |
10874 | fprintf (stream: dump_file, format: "\n" ); |
10875 | } |
10876 | |
10877 | remove_exit_barriers (region: root_omp_region); |
10878 | |
10879 | expand_omp (region: root_omp_region); |
10880 | |
10881 | omp_free_regions (); |
10882 | |
10883 | return (TODO_cleanup_cfg |
10884 | | (gimple_in_ssa_p (cfun) ? TODO_update_ssa_only_virtuals : 0)); |
10885 | } |
10886 | |
10887 | /* OMP expansion -- the default pass, run before creation of SSA form. */ |
10888 | |
10889 | namespace { |
10890 | |
10891 | const pass_data pass_data_expand_omp = |
10892 | { |
10893 | .type: GIMPLE_PASS, /* type */ |
10894 | .name: "ompexp" , /* name */ |
10895 | .optinfo_flags: OPTGROUP_OMP, /* optinfo_flags */ |
10896 | .tv_id: TV_NONE, /* tv_id */ |
10897 | PROP_gimple_any, /* properties_required */ |
10898 | PROP_gimple_eomp, /* properties_provided */ |
10899 | .properties_destroyed: 0, /* properties_destroyed */ |
10900 | .todo_flags_start: 0, /* todo_flags_start */ |
10901 | .todo_flags_finish: 0, /* todo_flags_finish */ |
10902 | }; |
10903 | |
10904 | class pass_expand_omp : public gimple_opt_pass |
10905 | { |
10906 | public: |
10907 | pass_expand_omp (gcc::context *ctxt) |
10908 | : gimple_opt_pass (pass_data_expand_omp, ctxt) |
10909 | {} |
10910 | |
10911 | /* opt_pass methods: */ |
10912 | unsigned int execute (function *) final override |
10913 | { |
10914 | bool gate = ((flag_openacc != 0 || flag_openmp != 0 |
10915 | || flag_openmp_simd != 0) |
10916 | && !seen_error ()); |
10917 | |
10918 | /* This pass always runs, to provide PROP_gimple_eomp. |
10919 | But often, there is nothing to do. */ |
10920 | if (!gate) |
10921 | return 0; |
10922 | |
10923 | return execute_expand_omp (); |
10924 | } |
10925 | |
10926 | }; // class pass_expand_omp |
10927 | |
10928 | } // anon namespace |
10929 | |
10930 | gimple_opt_pass * |
10931 | make_pass_expand_omp (gcc::context *ctxt) |
10932 | { |
10933 | return new pass_expand_omp (ctxt); |
10934 | } |
10935 | |
10936 | namespace { |
10937 | |
10938 | const pass_data pass_data_expand_omp_ssa = |
10939 | { |
10940 | .type: GIMPLE_PASS, /* type */ |
10941 | .name: "ompexpssa" , /* name */ |
10942 | .optinfo_flags: OPTGROUP_OMP, /* optinfo_flags */ |
10943 | .tv_id: TV_NONE, /* tv_id */ |
10944 | PROP_cfg | PROP_ssa, /* properties_required */ |
10945 | PROP_gimple_eomp, /* properties_provided */ |
10946 | .properties_destroyed: 0, /* properties_destroyed */ |
10947 | .todo_flags_start: 0, /* todo_flags_start */ |
10948 | TODO_cleanup_cfg | TODO_rebuild_alias, /* todo_flags_finish */ |
10949 | }; |
10950 | |
10951 | class pass_expand_omp_ssa : public gimple_opt_pass |
10952 | { |
10953 | public: |
10954 | pass_expand_omp_ssa (gcc::context *ctxt) |
10955 | : gimple_opt_pass (pass_data_expand_omp_ssa, ctxt) |
10956 | {} |
10957 | |
10958 | /* opt_pass methods: */ |
10959 | bool gate (function *fun) final override |
10960 | { |
10961 | return !(fun->curr_properties & PROP_gimple_eomp); |
10962 | } |
10963 | unsigned int execute (function *) final override |
10964 | { |
10965 | return execute_expand_omp (); |
10966 | } |
10967 | opt_pass * clone () final override |
10968 | { |
10969 | return new pass_expand_omp_ssa (m_ctxt); |
10970 | } |
10971 | |
10972 | }; // class pass_expand_omp_ssa |
10973 | |
10974 | } // anon namespace |
10975 | |
10976 | gimple_opt_pass * |
10977 | make_pass_expand_omp_ssa (gcc::context *ctxt) |
10978 | { |
10979 | return new pass_expand_omp_ssa (ctxt); |
10980 | } |
10981 | |
10982 | /* Called from tree-cfg.cc::make_edges to create cfg edges for all relevant |
10983 | GIMPLE_* codes. */ |
10984 | |
10985 | bool |
10986 | omp_make_gimple_edges (basic_block bb, struct omp_region **region, |
10987 | int *region_idx) |
10988 | { |
10989 | gimple *last = last_nondebug_stmt (bb); |
10990 | enum gimple_code code = gimple_code (g: last); |
10991 | struct omp_region *cur_region = *region; |
10992 | bool fallthru = false; |
10993 | |
10994 | switch (code) |
10995 | { |
10996 | case GIMPLE_OMP_PARALLEL: |
10997 | case GIMPLE_OMP_FOR: |
10998 | case GIMPLE_OMP_SINGLE: |
10999 | case GIMPLE_OMP_TEAMS: |
11000 | case GIMPLE_OMP_MASTER: |
11001 | case GIMPLE_OMP_MASKED: |
11002 | case GIMPLE_OMP_SCOPE: |
11003 | case GIMPLE_OMP_CRITICAL: |
11004 | case GIMPLE_OMP_SECTION: |
11005 | cur_region = new_omp_region (bb, type: code, parent: cur_region); |
11006 | fallthru = true; |
11007 | break; |
11008 | |
11009 | case GIMPLE_OMP_TASKGROUP: |
11010 | cur_region = new_omp_region (bb, type: code, parent: cur_region); |
11011 | fallthru = true; |
11012 | cur_region = cur_region->outer; |
11013 | break; |
11014 | |
11015 | case GIMPLE_OMP_TASK: |
11016 | cur_region = new_omp_region (bb, type: code, parent: cur_region); |
11017 | fallthru = true; |
11018 | if (gimple_omp_task_taskwait_p (g: last)) |
11019 | cur_region = cur_region->outer; |
11020 | break; |
11021 | |
11022 | case GIMPLE_OMP_ORDERED: |
11023 | cur_region = new_omp_region (bb, type: code, parent: cur_region); |
11024 | fallthru = true; |
11025 | if (gimple_omp_ordered_standalone_p (g: last)) |
11026 | cur_region = cur_region->outer; |
11027 | break; |
11028 | |
11029 | case GIMPLE_OMP_TARGET: |
11030 | cur_region = new_omp_region (bb, type: code, parent: cur_region); |
11031 | fallthru = true; |
11032 | switch (gimple_omp_target_kind (g: last)) |
11033 | { |
11034 | case GF_OMP_TARGET_KIND_REGION: |
11035 | case GF_OMP_TARGET_KIND_OACC_PARALLEL: |
11036 | case GF_OMP_TARGET_KIND_OACC_KERNELS: |
11037 | case GF_OMP_TARGET_KIND_OACC_SERIAL: |
11038 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED: |
11039 | case GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE: |
11040 | break; |
11041 | case GF_OMP_TARGET_KIND_UPDATE: |
11042 | case GF_OMP_TARGET_KIND_ENTER_DATA: |
11043 | case GF_OMP_TARGET_KIND_EXIT_DATA: |
11044 | case GF_OMP_TARGET_KIND_DATA: |
11045 | case GF_OMP_TARGET_KIND_OACC_DATA: |
11046 | case GF_OMP_TARGET_KIND_OACC_HOST_DATA: |
11047 | case GF_OMP_TARGET_KIND_OACC_DATA_KERNELS: |
11048 | case GF_OMP_TARGET_KIND_OACC_UPDATE: |
11049 | case GF_OMP_TARGET_KIND_OACC_ENTER_DATA: |
11050 | case GF_OMP_TARGET_KIND_OACC_EXIT_DATA: |
11051 | case GF_OMP_TARGET_KIND_OACC_DECLARE: |
11052 | cur_region = cur_region->outer; |
11053 | break; |
11054 | default: |
11055 | gcc_unreachable (); |
11056 | } |
11057 | break; |
11058 | |
11059 | case GIMPLE_OMP_SECTIONS: |
11060 | cur_region = new_omp_region (bb, type: code, parent: cur_region); |
11061 | fallthru = true; |
11062 | break; |
11063 | |
11064 | case GIMPLE_OMP_SECTIONS_SWITCH: |
11065 | fallthru = false; |
11066 | break; |
11067 | |
11068 | case GIMPLE_OMP_ATOMIC_LOAD: |
11069 | case GIMPLE_OMP_ATOMIC_STORE: |
11070 | fallthru = true; |
11071 | break; |
11072 | |
11073 | case GIMPLE_OMP_RETURN: |
11074 | /* In the case of a GIMPLE_OMP_SECTION, the edge will go |
11075 | somewhere other than the next block. This will be |
11076 | created later. */ |
11077 | cur_region->exit = bb; |
11078 | if (cur_region->type == GIMPLE_OMP_TASK) |
11079 | /* Add an edge corresponding to not scheduling the task |
11080 | immediately. */ |
11081 | make_edge (cur_region->entry, bb, EDGE_ABNORMAL); |
11082 | fallthru = cur_region->type != GIMPLE_OMP_SECTION; |
11083 | cur_region = cur_region->outer; |
11084 | break; |
11085 | |
11086 | case GIMPLE_OMP_CONTINUE: |
11087 | cur_region->cont = bb; |
11088 | switch (cur_region->type) |
11089 | { |
11090 | case GIMPLE_OMP_FOR: |
11091 | /* Mark all GIMPLE_OMP_FOR and GIMPLE_OMP_CONTINUE |
11092 | succs edges as abnormal to prevent splitting |
11093 | them. */ |
11094 | single_succ_edge (bb: cur_region->entry)->flags |= EDGE_ABNORMAL; |
11095 | /* Make the loopback edge. */ |
11096 | make_edge (bb, single_succ (bb: cur_region->entry), |
11097 | EDGE_ABNORMAL); |
11098 | |
11099 | /* Create an edge from GIMPLE_OMP_FOR to exit, which |
11100 | corresponds to the case that the body of the loop |
11101 | is not executed at all. */ |
11102 | make_edge (cur_region->entry, bb->next_bb, EDGE_ABNORMAL); |
11103 | make_edge (bb, bb->next_bb, EDGE_FALLTHRU | EDGE_ABNORMAL); |
11104 | fallthru = false; |
11105 | break; |
11106 | |
11107 | case GIMPLE_OMP_SECTIONS: |
11108 | /* Wire up the edges into and out of the nested sections. */ |
11109 | { |
11110 | basic_block switch_bb = single_succ (bb: cur_region->entry); |
11111 | |
11112 | struct omp_region *i; |
11113 | for (i = cur_region->inner; i ; i = i->next) |
11114 | { |
11115 | gcc_assert (i->type == GIMPLE_OMP_SECTION); |
11116 | make_edge (switch_bb, i->entry, 0); |
11117 | make_edge (i->exit, bb, EDGE_FALLTHRU); |
11118 | } |
11119 | |
11120 | /* Make the loopback edge to the block with |
11121 | GIMPLE_OMP_SECTIONS_SWITCH. */ |
11122 | make_edge (bb, switch_bb, 0); |
11123 | |
11124 | /* Make the edge from the switch to exit. */ |
11125 | make_edge (switch_bb, bb->next_bb, 0); |
11126 | fallthru = false; |
11127 | } |
11128 | break; |
11129 | |
11130 | case GIMPLE_OMP_TASK: |
11131 | fallthru = true; |
11132 | break; |
11133 | |
11134 | default: |
11135 | gcc_unreachable (); |
11136 | } |
11137 | break; |
11138 | |
11139 | default: |
11140 | gcc_unreachable (); |
11141 | } |
11142 | |
11143 | if (*region != cur_region) |
11144 | { |
11145 | *region = cur_region; |
11146 | if (cur_region) |
11147 | *region_idx = cur_region->entry->index; |
11148 | else |
11149 | *region_idx = 0; |
11150 | } |
11151 | |
11152 | return fallthru; |
11153 | } |
11154 | |