Line data Source code
1 : /* Vectorizer
2 : Copyright (C) 2003-2026 Free Software Foundation, Inc.
3 : Contributed by Dorit Naishlos <dorit@il.ibm.com>
4 :
5 : This file is part of GCC.
6 :
7 : GCC is free software; you can redistribute it and/or modify it under
8 : the terms of the GNU General Public License as published by the Free
9 : Software Foundation; either version 3, or (at your option) any later
10 : version.
11 :
12 : GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 : WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 : FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 : for more details.
16 :
17 : You should have received a copy of the GNU General Public License
18 : along with GCC; see the file COPYING3. If not see
19 : <http://www.gnu.org/licenses/>. */
20 :
21 : /* Loop and basic block vectorizer.
22 :
23 : This file contains drivers for the three vectorizers:
24 : (1) loop vectorizer (inter-iteration parallelism),
25 : (2) loop-aware SLP (intra-iteration parallelism) (invoked by the loop
26 : vectorizer)
27 : (3) BB vectorizer (out-of-loops), aka SLP
28 :
29 : The rest of the vectorizer's code is organized as follows:
30 : - tree-vect-loop.cc - loop specific parts such as reductions, etc. These are
31 : used by drivers (1) and (2).
32 : - tree-vect-loop-manip.cc - vectorizer's loop control-flow utilities, used by
33 : drivers (1) and (2).
34 : - tree-vect-slp.cc - BB vectorization specific analysis and transformation,
35 : used by drivers (2) and (3).
36 : - tree-vect-stmts.cc - statements analysis and transformation (used by all).
37 : - tree-vect-data-refs.cc - vectorizer specific data-refs analysis and
38 : manipulations (used by all).
39 : - tree-vect-patterns.cc - vectorizable code patterns detector (used by all)
40 :
41 : Here's a poor attempt at illustrating that:
42 :
43 : tree-vectorizer.cc:
44 : loop_vect() loop_aware_slp() slp_vect()
45 : | / \ /
46 : | / \ /
47 : tree-vect-loop.cc tree-vect-slp.cc
48 : | \ \ / / |
49 : | \ \/ / |
50 : | \ /\ / |
51 : | \ / \ / |
52 : tree-vect-stmts.cc tree-vect-data-refs.cc
53 : \ /
54 : tree-vect-patterns.cc
55 : */
56 :
57 : #include "config.h"
58 : #include "system.h"
59 : #include "coretypes.h"
60 : #include "backend.h"
61 : #include "tree.h"
62 : #include "gimple.h"
63 : #include "predict.h"
64 : #include "tree-pass.h"
65 : #include "ssa.h"
66 : #include "cgraph.h"
67 : #include "fold-const.h"
68 : #include "stor-layout.h"
69 : #include "gimple-iterator.h"
70 : #include "gimple-walk.h"
71 : #include "tree-ssa-loop-manip.h"
72 : #include "tree-ssa-loop-niter.h"
73 : #include "tree-cfg.h"
74 : #include "cfgloop.h"
75 : #include "tree-vectorizer.h"
76 : #include "tree-ssa-propagate.h"
77 : #include "dbgcnt.h"
78 : #include "tree-scalar-evolution.h"
79 : #include "stringpool.h"
80 : #include "attribs.h"
81 : #include "gimple-pretty-print.h"
82 : #include "opt-problem.h"
83 : #include "internal-fn.h"
84 : #include "tree-ssa-sccvn.h"
85 : #include "tree-into-ssa.h"
86 : #include "gimple-range.h"
87 :
88 : /* Loop or bb location, with hotness information. */
89 : dump_user_location_t vect_location;
90 :
91 : /* auto_purge_vect_location's dtor: reset the vect_location
92 : global, to avoid stale location_t values that could reference
93 : GC-ed blocks. */
94 :
95 1379087 : auto_purge_vect_location::~auto_purge_vect_location ()
96 : {
97 1379087 : vect_location = dump_user_location_t ();
98 1379087 : }
99 :
100 : /* Dump a cost entry according to args to F. */
101 :
102 : void
103 214045 : dump_stmt_cost (FILE *f, int count, enum vect_cost_for_stmt kind,
104 : stmt_vec_info stmt_info, slp_tree node, tree,
105 : int misalign, unsigned cost,
106 : enum vect_cost_model_location where)
107 : {
108 214045 : if (stmt_info)
109 : {
110 197094 : print_gimple_expr (f, STMT_VINFO_STMT (stmt_info), 0, TDF_SLIM);
111 197094 : fprintf (f, " ");
112 : }
113 16951 : else if (node)
114 3339 : fprintf (f, "node %p ", (void *)node);
115 : else
116 13612 : fprintf (f, "<unknown> ");
117 214045 : fprintf (f, "%d times ", count);
118 214045 : const char *ks = "unknown";
119 214045 : switch (kind)
120 : {
121 48646 : case scalar_stmt:
122 48646 : ks = "scalar_stmt";
123 48646 : break;
124 36048 : case scalar_load:
125 36048 : ks = "scalar_load";
126 36048 : break;
127 27140 : case scalar_store:
128 27140 : ks = "scalar_store";
129 27140 : break;
130 34384 : case vector_stmt:
131 34384 : ks = "vector_stmt";
132 34384 : break;
133 21736 : case vector_load:
134 21736 : ks = "vector_load";
135 21736 : break;
136 0 : case vector_gather_load:
137 0 : ks = "vector_gather_load";
138 0 : break;
139 7501 : case unaligned_load:
140 7501 : ks = "unaligned_load";
141 7501 : break;
142 4535 : case unaligned_store:
143 4535 : ks = "unaligned_store";
144 4535 : break;
145 8618 : case vector_store:
146 8618 : ks = "vector_store";
147 8618 : break;
148 0 : case vector_scatter_store:
149 0 : ks = "vector_scatter_store";
150 0 : break;
151 3221 : case vec_to_scalar:
152 3221 : ks = "vec_to_scalar";
153 3221 : break;
154 9113 : case scalar_to_vec:
155 9113 : ks = "scalar_to_vec";
156 9113 : break;
157 8 : case cond_branch_not_taken:
158 8 : ks = "cond_branch_not_taken";
159 8 : break;
160 363 : case cond_branch_taken:
161 363 : ks = "cond_branch_taken";
162 363 : break;
163 6320 : case vec_perm:
164 6320 : ks = "vec_perm";
165 6320 : break;
166 5224 : case vec_promote_demote:
167 5224 : ks = "vec_promote_demote";
168 5224 : break;
169 1188 : case vec_construct:
170 1188 : ks = "vec_construct";
171 1188 : break;
172 : }
173 214045 : fprintf (f, "%s ", ks);
174 214045 : if (kind == unaligned_load || kind == unaligned_store)
175 12036 : fprintf (f, "(misalign %d) ", misalign);
176 214045 : fprintf (f, "costs %u ", cost);
177 214045 : const char *ws = "unknown";
178 214045 : switch (where)
179 : {
180 128191 : case vect_prologue:
181 128191 : ws = "prologue";
182 128191 : break;
183 79125 : case vect_body:
184 79125 : ws = "body";
185 79125 : break;
186 6729 : case vect_epilogue:
187 6729 : ws = "epilogue";
188 6729 : break;
189 : }
190 214045 : fprintf (f, "in %s\n", ws);
191 214045 : }
192 :
193 : /* For mapping simduid to vectorization factor. */
194 :
195 : class simduid_to_vf : public free_ptr_hash<simduid_to_vf>
196 : {
197 : public:
198 : unsigned int simduid;
199 : poly_uint64 vf;
200 :
201 : /* hash_table support. */
202 : static inline hashval_t hash (const simduid_to_vf *);
203 : static inline int equal (const simduid_to_vf *, const simduid_to_vf *);
204 : };
205 :
206 : inline hashval_t
207 8414 : simduid_to_vf::hash (const simduid_to_vf *p)
208 : {
209 8414 : return p->simduid;
210 : }
211 :
212 : inline int
213 14768 : simduid_to_vf::equal (const simduid_to_vf *p1, const simduid_to_vf *p2)
214 : {
215 14768 : return p1->simduid == p2->simduid;
216 : }
217 :
218 : /* This hash maps the OMP simd array to the corresponding simduid used
219 : to index into it. Like thus,
220 :
221 : _7 = GOMP_SIMD_LANE (simduid.0)
222 : ...
223 : ...
224 : D.1737[_7] = stuff;
225 :
226 :
227 : This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
228 : simduid.0. */
229 :
230 : struct simd_array_to_simduid : free_ptr_hash<simd_array_to_simduid>
231 : {
232 : tree decl;
233 : unsigned int simduid;
234 :
235 : /* hash_table support. */
236 : static inline hashval_t hash (const simd_array_to_simduid *);
237 : static inline int equal (const simd_array_to_simduid *,
238 : const simd_array_to_simduid *);
239 : };
240 :
241 : inline hashval_t
242 23846 : simd_array_to_simduid::hash (const simd_array_to_simduid *p)
243 : {
244 23846 : return DECL_UID (p->decl);
245 : }
246 :
247 : inline int
248 16545 : simd_array_to_simduid::equal (const simd_array_to_simduid *p1,
249 : const simd_array_to_simduid *p2)
250 : {
251 16545 : return p1->decl == p2->decl;
252 : }
253 :
254 : /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LAST_LANE,
255 : into their corresponding constants and remove
256 : IFN_GOMP_SIMD_ORDERED_{START,END}. */
257 :
258 : static void
259 7809 : adjust_simduid_builtins (hash_table<simduid_to_vf> *htab, function *fun)
260 : {
261 7809 : basic_block bb;
262 :
263 118628 : FOR_EACH_BB_FN (bb, fun)
264 : {
265 110819 : gimple_stmt_iterator i;
266 :
267 739076 : for (i = gsi_start_bb (bb); !gsi_end_p (i); )
268 : {
269 517438 : poly_uint64 vf = 1;
270 517438 : enum internal_fn ifn;
271 517438 : gimple *stmt = gsi_stmt (i);
272 517438 : tree t;
273 517438 : if (!is_gimple_call (stmt)
274 517438 : || !gimple_call_internal_p (stmt))
275 : {
276 509875 : gsi_next (&i);
277 510791 : continue;
278 : }
279 7563 : ifn = gimple_call_internal_fn (stmt);
280 7563 : switch (ifn)
281 : {
282 6647 : case IFN_GOMP_SIMD_LANE:
283 6647 : case IFN_GOMP_SIMD_VF:
284 6647 : case IFN_GOMP_SIMD_LAST_LANE:
285 6647 : break;
286 158 : case IFN_GOMP_SIMD_ORDERED_START:
287 158 : case IFN_GOMP_SIMD_ORDERED_END:
288 158 : if (integer_onep (gimple_call_arg (stmt, 0)))
289 : {
290 5 : enum built_in_function bcode
291 : = (ifn == IFN_GOMP_SIMD_ORDERED_START
292 10 : ? BUILT_IN_GOMP_ORDERED_START
293 : : BUILT_IN_GOMP_ORDERED_END);
294 10 : gimple *g
295 10 : = gimple_build_call (builtin_decl_explicit (bcode), 0);
296 10 : gimple_move_vops (g, stmt);
297 10 : gsi_replace (&i, g, true);
298 10 : continue;
299 10 : }
300 148 : gsi_remove (&i, true);
301 148 : unlink_stmt_vdef (stmt);
302 148 : continue;
303 758 : default:
304 758 : gsi_next (&i);
305 758 : continue;
306 906 : }
307 6647 : tree arg = gimple_call_arg (stmt, 0);
308 6647 : gcc_assert (arg != NULL_TREE);
309 6647 : gcc_assert (TREE_CODE (arg) == SSA_NAME);
310 6647 : simduid_to_vf *p = NULL, data;
311 6647 : data.simduid = DECL_UID (SSA_NAME_VAR (arg));
312 : /* Need to nullify loop safelen field since it's value is not
313 : valid after transformation. */
314 6647 : if (bb->loop_father && bb->loop_father->safelen > 0)
315 2169 : bb->loop_father->safelen = 0;
316 6647 : if (htab)
317 : {
318 4742 : p = htab->find (&data);
319 4742 : if (p)
320 4701 : vf = p->vf;
321 : }
322 6647 : switch (ifn)
323 : {
324 969 : case IFN_GOMP_SIMD_VF:
325 969 : t = build_int_cst (unsigned_type_node, vf);
326 969 : break;
327 3484 : case IFN_GOMP_SIMD_LANE:
328 3484 : t = build_int_cst (unsigned_type_node, 0);
329 3484 : break;
330 2194 : case IFN_GOMP_SIMD_LAST_LANE:
331 2194 : t = gimple_call_arg (stmt, 1);
332 2194 : break;
333 : default:
334 : gcc_unreachable ();
335 : }
336 6647 : tree lhs = gimple_call_lhs (stmt);
337 6647 : if (lhs)
338 6590 : replace_uses_by (lhs, t);
339 6647 : release_defs (stmt);
340 6647 : gsi_remove (&i, true);
341 : }
342 : }
343 7809 : }
344 :
345 : /* Helper structure for note_simd_array_uses. */
346 :
347 : struct note_simd_array_uses_struct
348 : {
349 : hash_table<simd_array_to_simduid> **htab;
350 : unsigned int simduid;
351 : };
352 :
353 : /* Callback for note_simd_array_uses, called through walk_gimple_op. */
354 :
355 : static tree
356 65131 : note_simd_array_uses_cb (tree *tp, int *walk_subtrees, void *data)
357 : {
358 65131 : struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
359 65131 : struct note_simd_array_uses_struct *ns
360 : = (struct note_simd_array_uses_struct *) wi->info;
361 :
362 65131 : if (TYPE_P (*tp))
363 0 : *walk_subtrees = 0;
364 65131 : else if (VAR_P (*tp)
365 12360 : && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp))
366 77491 : && DECL_CONTEXT (*tp) == current_function_decl)
367 : {
368 12360 : simd_array_to_simduid data;
369 12360 : if (!*ns->htab)
370 2208 : *ns->htab = new hash_table<simd_array_to_simduid> (15);
371 12360 : data.decl = *tp;
372 12360 : data.simduid = ns->simduid;
373 12360 : simd_array_to_simduid **slot = (*ns->htab)->find_slot (&data, INSERT);
374 12360 : if (*slot == NULL)
375 : {
376 5501 : simd_array_to_simduid *p = XNEW (simd_array_to_simduid);
377 5501 : *p = data;
378 5501 : *slot = p;
379 : }
380 6859 : else if ((*slot)->simduid != ns->simduid)
381 0 : (*slot)->simduid = -1U;
382 12360 : *walk_subtrees = 0;
383 : }
384 65131 : return NULL_TREE;
385 : }
386 :
387 : /* Find "omp simd array" temporaries and map them to corresponding
388 : simduid. */
389 :
390 : static void
391 7809 : note_simd_array_uses (hash_table<simd_array_to_simduid> **htab, function *fun)
392 : {
393 7809 : basic_block bb;
394 7809 : gimple_stmt_iterator gsi;
395 7809 : struct walk_stmt_info wi;
396 7809 : struct note_simd_array_uses_struct ns;
397 :
398 7809 : memset (&wi, 0, sizeof (wi));
399 7809 : wi.info = &ns;
400 7809 : ns.htab = htab;
401 :
402 102471 : FOR_EACH_BB_FN (bb, fun)
403 582996 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
404 : {
405 393672 : gimple *stmt = gsi_stmt (gsi);
406 393672 : if (!is_gimple_call (stmt) || !gimple_call_internal_p (stmt))
407 387088 : continue;
408 7509 : switch (gimple_call_internal_fn (stmt))
409 : {
410 6620 : case IFN_GOMP_SIMD_LANE:
411 6620 : case IFN_GOMP_SIMD_VF:
412 6620 : case IFN_GOMP_SIMD_LAST_LANE:
413 6620 : break;
414 889 : default:
415 889 : continue;
416 : }
417 6620 : tree lhs = gimple_call_lhs (stmt);
418 6620 : if (lhs == NULL_TREE)
419 36 : continue;
420 6584 : imm_use_iterator use_iter;
421 6584 : gimple *use_stmt;
422 6584 : ns.simduid = DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt, 0)));
423 31195 : FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, lhs)
424 18027 : if (!is_gimple_debug (use_stmt))
425 24522 : walk_gimple_op (use_stmt, note_simd_array_uses_cb, &wi);
426 : }
427 7809 : }
428 :
429 : /* Shrink arrays with "omp simd array" attribute to the corresponding
430 : vectorization factor. */
431 :
432 : static void
433 2208 : shrink_simd_arrays
434 : (hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab,
435 : hash_table<simduid_to_vf> *simduid_to_vf_htab)
436 : {
437 7709 : for (hash_table<simd_array_to_simduid>::iterator iter
438 2208 : = simd_array_to_simduid_htab->begin ();
439 13210 : iter != simd_array_to_simduid_htab->end (); ++iter)
440 5501 : if ((*iter)->simduid != -1U)
441 : {
442 5501 : tree decl = (*iter)->decl;
443 5501 : poly_uint64 vf = 1;
444 5501 : if (simduid_to_vf_htab)
445 : {
446 4574 : simduid_to_vf *p = NULL, data;
447 4574 : data.simduid = (*iter)->simduid;
448 4574 : p = simduid_to_vf_htab->find (&data);
449 4574 : if (p)
450 4540 : vf = p->vf;
451 : }
452 5501 : tree atype
453 5501 : = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl)), vf);
454 5501 : TREE_TYPE (decl) = atype;
455 5501 : relayout_decl (decl);
456 : }
457 :
458 2208 : delete simd_array_to_simduid_htab;
459 2208 : }
460 :
461 : /* Initialize the vec_info with kind KIND_IN and target cost data
462 : TARGET_COST_DATA_IN. */
463 :
464 2697078 : vec_info::vec_info (vec_info::vec_kind kind_in, vec_info_shared *shared_)
465 2697078 : : kind (kind_in),
466 2697078 : shared (shared_),
467 2697078 : stmt_vec_info_ro (false),
468 2697078 : bbs (NULL),
469 2697078 : nbbs (0),
470 2697078 : inv_pattern_def_seq (NULL)
471 : {
472 2697078 : stmt_vec_infos.create (50);
473 2697078 : }
474 :
475 2697078 : vec_info::~vec_info ()
476 : {
477 4687672 : for (slp_instance &instance : slp_instances)
478 1027344 : vect_free_slp_instance (instance);
479 :
480 2697078 : free_stmt_vec_infos ();
481 2697078 : }
482 :
483 2348666 : vec_info_shared::vec_info_shared ()
484 2348666 : : datarefs (vNULL),
485 2348666 : datarefs_copy (vNULL),
486 2348666 : ddrs (vNULL)
487 : {
488 2348666 : }
489 :
490 2348666 : vec_info_shared::~vec_info_shared ()
491 : {
492 2348666 : free_data_refs (datarefs);
493 2348666 : free_dependence_relations (ddrs);
494 2348666 : datarefs_copy.release ();
495 2348666 : }
496 :
497 : void
498 2115011 : vec_info_shared::save_datarefs ()
499 : {
500 2115011 : if (!flag_checking)
501 : return;
502 3171744 : datarefs_copy.reserve_exact (datarefs.length ());
503 12783438 : for (unsigned i = 0; i < datarefs.length (); ++i)
504 10668444 : datarefs_copy.quick_push (*datarefs[i]);
505 : }
506 :
507 : void
508 805721 : vec_info_shared::check_datarefs ()
509 : {
510 805721 : if (!flag_checking)
511 : return;
512 2414101 : gcc_assert (datarefs.length () == datarefs_copy.length ());
513 11956606 : for (unsigned i = 0; i < datarefs.length (); ++i)
514 11150885 : if (memcmp (&datarefs_copy[i], datarefs[i],
515 : offsetof (data_reference, alt_indices)) != 0)
516 0 : gcc_unreachable ();
517 : }
518 :
519 : /* Record that STMT belongs to the vectorizable region. Create and return
520 : an associated stmt_vec_info. */
521 :
522 : stmt_vec_info
523 60966745 : vec_info::add_stmt (gimple *stmt)
524 : {
525 60966745 : stmt_vec_info res = new_stmt_vec_info (stmt);
526 60966745 : set_vinfo_for_stmt (stmt, res);
527 60966745 : return res;
528 : }
529 :
530 : /* Record that STMT belongs to the vectorizable region. Create a new
531 : stmt_vec_info and mark VECINFO as being related and return the new
532 : stmt_vec_info. */
533 :
534 : stmt_vec_info
535 1081 : vec_info::add_pattern_stmt (gimple *stmt, stmt_vec_info stmt_info)
536 : {
537 1081 : stmt_vec_info res = new_stmt_vec_info (stmt);
538 1081 : res->pattern_stmt_p = true;
539 1081 : set_vinfo_for_stmt (stmt, res, false);
540 1081 : STMT_VINFO_RELATED_STMT (res) = stmt_info;
541 1081 : return res;
542 : }
543 :
544 : /* If STMT was previously associated with a stmt_vec_info and STMT now resides
545 : at a different address than before (e.g., because STMT is a phi node that has
546 : been resized), update the stored address to match the new one. It is not
547 : possible to use lookup_stmt () to perform this task, because that function
548 : returns NULL if the stored stmt pointer does not match the one being looked
549 : up. */
550 :
551 : stmt_vec_info
552 10876 : vec_info::resync_stmt_addr (gimple *stmt)
553 : {
554 10876 : unsigned int uid = gimple_uid (stmt);
555 10876 : if (uid > 0 && uid - 1 < stmt_vec_infos.length ())
556 : {
557 10876 : stmt_vec_info res = stmt_vec_infos[uid - 1];
558 10876 : if (res && res->stmt)
559 : {
560 10876 : res->stmt = stmt;
561 10876 : return res;
562 : }
563 : }
564 : return nullptr;
565 : }
566 :
567 : /* If STMT has an associated stmt_vec_info, return that vec_info, otherwise
568 : return null. It is safe to call this function on any statement, even if
569 : it might not be part of the vectorizable region. */
570 :
571 : stmt_vec_info
572 466709285 : vec_info::lookup_stmt (gimple *stmt)
573 : {
574 466709285 : unsigned int uid = gimple_uid (stmt);
575 466709285 : if (uid > 0 && uid - 1 < stmt_vec_infos.length ())
576 : {
577 291132812 : stmt_vec_info res = stmt_vec_infos[uid - 1];
578 291132812 : if (res && res->stmt == stmt)
579 290805594 : return res;
580 : }
581 : return NULL;
582 : }
583 :
584 : /* If NAME is an SSA_NAME and its definition has an associated stmt_vec_info,
585 : return that stmt_vec_info, otherwise return null. It is safe to call
586 : this on arbitrary operands. */
587 :
588 : stmt_vec_info
589 49135217 : vec_info::lookup_def (tree name)
590 : {
591 49135217 : if (TREE_CODE (name) == SSA_NAME
592 49135217 : && !SSA_NAME_IS_DEFAULT_DEF (name))
593 46928194 : return lookup_stmt (SSA_NAME_DEF_STMT (name));
594 : return NULL;
595 : }
596 :
597 : /* See whether there is a single non-debug statement that uses LHS and
598 : whether that statement has an associated stmt_vec_info. Return the
599 : stmt_vec_info if so, otherwise return null. */
600 :
601 : stmt_vec_info
602 1991 : vec_info::lookup_single_use (tree lhs)
603 : {
604 1991 : use_operand_p dummy;
605 1991 : gimple *use_stmt;
606 1991 : if (single_imm_use (lhs, &dummy, &use_stmt))
607 1843 : return lookup_stmt (use_stmt);
608 : return NULL;
609 : }
610 :
611 : /* Return vectorization information about DR. */
612 :
613 : dr_vec_info *
614 47359842 : vec_info::lookup_dr (data_reference *dr)
615 : {
616 47359842 : stmt_vec_info stmt_info = lookup_stmt (DR_STMT (dr));
617 : /* DR_STMT should never refer to a stmt in a pattern replacement. */
618 47359842 : gcc_checking_assert (!is_pattern_stmt_p (stmt_info));
619 47359842 : return STMT_VINFO_DR_INFO (stmt_info->dr_aux.stmt);
620 : }
621 :
622 : /* Record that NEW_STMT_INFO now implements the same data reference
623 : as OLD_STMT_INFO. */
624 :
625 : void
626 6130 : vec_info::move_dr (stmt_vec_info new_stmt_info, stmt_vec_info old_stmt_info)
627 : {
628 6130 : gcc_assert (!is_pattern_stmt_p (old_stmt_info));
629 6130 : STMT_VINFO_DR_INFO (old_stmt_info)->stmt = new_stmt_info;
630 6130 : new_stmt_info->dr_aux = old_stmt_info->dr_aux;
631 6130 : STMT_VINFO_DR_WRT_VEC_LOOP (new_stmt_info)
632 6130 : = STMT_VINFO_DR_WRT_VEC_LOOP (old_stmt_info);
633 6130 : STMT_VINFO_GATHER_SCATTER_P (new_stmt_info)
634 6130 : = STMT_VINFO_GATHER_SCATTER_P (old_stmt_info);
635 6130 : STMT_VINFO_STRIDED_P (new_stmt_info)
636 6130 : = STMT_VINFO_STRIDED_P (old_stmt_info);
637 6130 : STMT_VINFO_SIMD_LANE_ACCESS_P (new_stmt_info)
638 6130 : = STMT_VINFO_SIMD_LANE_ACCESS_P (old_stmt_info);
639 6130 : }
640 :
641 : /* Permanently remove the statement described by STMT_INFO from the
642 : function. */
643 :
644 : void
645 1481702 : vec_info::remove_stmt (stmt_vec_info stmt_info)
646 : {
647 1481702 : gcc_assert (!stmt_info->pattern_stmt_p);
648 1481702 : set_vinfo_for_stmt (stmt_info->stmt, NULL);
649 1481702 : unlink_stmt_vdef (stmt_info->stmt);
650 1481702 : gimple_stmt_iterator si = gsi_for_stmt (stmt_info->stmt);
651 1481702 : gsi_remove (&si, true);
652 1481702 : release_defs (stmt_info->stmt);
653 1481702 : free_stmt_vec_info (stmt_info);
654 1481702 : }
655 :
656 : /* Replace the statement at GSI by NEW_STMT, both the vectorization
657 : information and the function itself. STMT_INFO describes the statement
658 : at GSI. */
659 :
660 : void
661 5193 : vec_info::replace_stmt (gimple_stmt_iterator *gsi, stmt_vec_info stmt_info,
662 : gimple *new_stmt)
663 : {
664 5193 : gimple *old_stmt = stmt_info->stmt;
665 5193 : gcc_assert (!stmt_info->pattern_stmt_p && old_stmt == gsi_stmt (*gsi));
666 5193 : gimple_set_uid (new_stmt, gimple_uid (old_stmt));
667 5193 : stmt_info->stmt = new_stmt;
668 5193 : gsi_replace (gsi, new_stmt, true);
669 5193 : }
670 :
671 : /* Insert stmts in SEQ on the VEC_INFO region entry. If CONTEXT is
672 : not NULL it specifies whether to use the sub-region entry
673 : determined by it, currently used for loop vectorization to insert
674 : on the inner loop entry vs. the outer loop entry. */
675 :
676 : void
677 101510 : vec_info::insert_seq_on_entry (stmt_vec_info context, gimple_seq seq)
678 : {
679 101510 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (this))
680 : {
681 19134 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
682 19134 : basic_block new_bb;
683 19134 : edge pe;
684 :
685 19134 : if (context && nested_in_vect_loop_p (loop, context))
686 : loop = loop->inner;
687 :
688 19134 : pe = loop_preheader_edge (loop);
689 19134 : new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
690 19134 : gcc_assert (!new_bb);
691 : }
692 : else
693 : {
694 82376 : gimple_stmt_iterator gsi_region_begin
695 82376 : = gsi_after_labels (bbs[0]);
696 82376 : gsi_insert_seq_before (&gsi_region_begin, seq, GSI_SAME_STMT);
697 : }
698 101510 : }
699 :
700 : /* Like insert_seq_on_entry but just inserts the single stmt NEW_STMT. */
701 :
702 : void
703 3323 : vec_info::insert_on_entry (stmt_vec_info context, gimple *new_stmt)
704 : {
705 3323 : gimple_seq seq = NULL;
706 3323 : gimple_stmt_iterator gsi = gsi_start (seq);
707 3323 : gsi_insert_before_without_update (&gsi, new_stmt, GSI_SAME_STMT);
708 3323 : insert_seq_on_entry (context, seq);
709 3323 : }
710 :
711 : /* Create and initialize a new stmt_vec_info struct for STMT. */
712 :
713 : stmt_vec_info
714 60967826 : vec_info::new_stmt_vec_info (gimple *stmt)
715 : {
716 60967826 : stmt_vec_info res = XCNEW (class _stmt_vec_info);
717 60967826 : res->stmt = stmt;
718 :
719 60967826 : STMT_VINFO_RELEVANT (res) = vect_unused_in_scope;
720 60967826 : STMT_VINFO_VECTORIZABLE (res) = true;
721 60967826 : STMT_VINFO_REDUC_TYPE (res) = TREE_CODE_REDUCTION;
722 60967826 : STMT_VINFO_REDUC_CODE (res) = ERROR_MARK;
723 60967826 : STMT_VINFO_REDUC_IDX (res) = -1;
724 60967826 : STMT_VINFO_REDUC_DEF (res) = NULL;
725 60967826 : STMT_VINFO_SLP_VECT_ONLY (res) = false;
726 60967826 : STMT_VINFO_SLP_VECT_ONLY_PATTERN (res) = false;
727 :
728 60967826 : if (is_a <loop_vec_info> (this)
729 7452589 : && gimple_code (stmt) == GIMPLE_PHI
730 62262505 : && is_loop_header_bb_p (gimple_bb (stmt)))
731 1284366 : STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
732 : else
733 59683460 : STMT_VINFO_DEF_TYPE (res) = vect_internal_def;
734 :
735 60967826 : STMT_SLP_TYPE (res) = not_vect;
736 :
737 : /* This is really "uninitialized" until vect_compute_data_ref_alignment. */
738 60967826 : res->dr_aux.misalignment = DR_MISALIGNMENT_UNINITIALIZED;
739 :
740 60967826 : return res;
741 : }
742 :
743 : /* Associate STMT with INFO. */
744 :
745 : void
746 62449528 : vec_info::set_vinfo_for_stmt (gimple *stmt, stmt_vec_info info, bool check_ro)
747 : {
748 62449528 : unsigned int uid = gimple_uid (stmt);
749 62449528 : if (uid == 0)
750 : {
751 60967826 : gcc_assert (!check_ro || !stmt_vec_info_ro);
752 60967826 : gcc_checking_assert (info);
753 60967826 : uid = stmt_vec_infos.length () + 1;
754 60967826 : gimple_set_uid (stmt, uid);
755 60967826 : stmt_vec_infos.safe_push (info);
756 : }
757 : else
758 : {
759 1481702 : gcc_checking_assert (info == NULL);
760 1481702 : stmt_vec_infos[uid - 1] = info;
761 : }
762 62449528 : }
763 :
764 : /* Free the contents of stmt_vec_infos. */
765 :
766 : void
767 2697078 : vec_info::free_stmt_vec_infos (void)
768 : {
769 69059060 : for (stmt_vec_info &info : stmt_vec_infos)
770 60967826 : if (info != NULL)
771 59486124 : free_stmt_vec_info (info);
772 2697078 : stmt_vec_infos.release ();
773 2697078 : }
774 :
775 : /* Free STMT_INFO. */
776 :
777 : void
778 60967826 : vec_info::free_stmt_vec_info (stmt_vec_info stmt_info)
779 : {
780 60967826 : if (stmt_info->pattern_stmt_p)
781 : {
782 2312539 : gimple_set_bb (stmt_info->stmt, NULL);
783 2312539 : tree lhs = gimple_get_lhs (stmt_info->stmt);
784 2312539 : if (lhs && TREE_CODE (lhs) == SSA_NAME)
785 1948688 : release_ssa_name (lhs);
786 : }
787 :
788 60967826 : free (stmt_info);
789 60967826 : }
790 :
791 : /* Returns true if S1 dominates S2. */
792 :
793 : bool
794 586716 : vect_stmt_dominates_stmt_p (gimple *s1, gimple *s2)
795 : {
796 586716 : basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
797 :
798 : /* If bb1 is NULL, it should be a GIMPLE_NOP def stmt of an (D)
799 : SSA_NAME. Assume it lives at the beginning of function and
800 : thus dominates everything. */
801 586716 : if (!bb1 || s1 == s2)
802 : return true;
803 :
804 : /* If bb2 is NULL, it doesn't dominate any stmt with a bb. */
805 584801 : if (!bb2)
806 : return false;
807 :
808 584801 : if (bb1 != bb2)
809 215466 : return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
810 :
811 : /* PHIs in the same basic block are assumed to be
812 : executed all in parallel, if only one stmt is a PHI,
813 : it dominates the other stmt in the same basic block. */
814 369335 : if (gimple_code (s1) == GIMPLE_PHI)
815 : return true;
816 :
817 332900 : if (gimple_code (s2) == GIMPLE_PHI)
818 : return false;
819 :
820 : /* Inserted vectorized stmts all have UID 0 while the original stmts
821 : in the IL have UID increasing within a BB. Walk from both sides
822 : until we find the other stmt or a stmt with UID != 0. */
823 315145 : gimple_stmt_iterator gsi1 = gsi_for_stmt (s1);
824 799530 : while (gimple_uid (gsi_stmt (gsi1)) == 0)
825 : {
826 601108 : gsi_next (&gsi1);
827 601108 : if (gsi_end_p (gsi1))
828 : return false;
829 599367 : if (gsi_stmt (gsi1) == s2)
830 : return true;
831 : }
832 198422 : if (gimple_uid (gsi_stmt (gsi1)) == -1u)
833 : return false;
834 :
835 198422 : gimple_stmt_iterator gsi2 = gsi_for_stmt (s2);
836 804478 : while (gimple_uid (gsi_stmt (gsi2)) == 0)
837 : {
838 617995 : gsi_prev (&gsi2);
839 617995 : if (gsi_end_p (gsi2))
840 : return false;
841 606076 : if (gsi_stmt (gsi2) == s1)
842 : return true;
843 : }
844 186483 : if (gimple_uid (gsi_stmt (gsi2)) == -1u)
845 : return false;
846 :
847 186483 : if (gimple_uid (gsi_stmt (gsi1)) <= gimple_uid (gsi_stmt (gsi2)))
848 : return true;
849 : return false;
850 : }
851 :
852 : /* A helper function to free scev and LOOP niter information, as well as
853 : clear loop constraint LOOP_C_FINITE. */
854 :
855 : void
856 58042 : vect_free_loop_info_assumptions (class loop *loop)
857 : {
858 58042 : scev_reset_htab ();
859 : /* We need to explicitly reset upper bound information since they are
860 : used even after free_numbers_of_iterations_estimates. */
861 58042 : loop->any_upper_bound = false;
862 58042 : loop->any_likely_upper_bound = false;
863 58042 : free_numbers_of_iterations_estimates (loop);
864 58042 : loop_constraint_clear (loop, LOOP_C_FINITE);
865 58042 : }
866 :
867 : /* If LOOP has been versioned during ifcvt, return the internal call
868 : guarding it. */
869 :
870 : gimple *
871 523177 : vect_loop_vectorized_call (class loop *loop, gcond **cond)
872 : {
873 523177 : basic_block bb = loop_preheader_edge (loop)->src;
874 967239 : gimple *g;
875 1411301 : do
876 : {
877 967239 : g = *gsi_last_bb (bb);
878 613020 : if ((g && gimple_code (g) == GIMPLE_COND)
879 2073338 : || !single_succ_p (bb))
880 : break;
881 582922 : if (!single_pred_p (bb))
882 : break;
883 444062 : bb = single_pred (bb);
884 : }
885 : while (1);
886 523177 : if (g && gimple_code (g) == GIMPLE_COND)
887 : {
888 377740 : if (cond)
889 0 : *cond = as_a <gcond *> (g);
890 377740 : gimple_stmt_iterator gsi = gsi_for_stmt (g);
891 377740 : gsi_prev (&gsi);
892 377740 : if (!gsi_end_p (gsi))
893 : {
894 346687 : g = gsi_stmt (gsi);
895 346687 : if (gimple_call_internal_p (g, IFN_LOOP_VECTORIZED)
896 346687 : && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->num
897 30140 : || tree_to_shwi (gimple_call_arg (g, 1)) == loop->num))
898 60625 : return g;
899 : }
900 : }
901 : return NULL;
902 : }
903 :
904 : /* If LOOP has been versioned during loop distribution, return the gurading
905 : internal call. */
906 :
907 : static gimple *
908 486396 : vect_loop_dist_alias_call (class loop *loop, function *fun)
909 : {
910 486396 : basic_block bb;
911 486396 : basic_block entry;
912 486396 : class loop *outer, *orig;
913 :
914 486396 : if (loop->orig_loop_num == 0)
915 : return NULL;
916 :
917 162 : orig = get_loop (fun, loop->orig_loop_num);
918 162 : if (orig == NULL)
919 : {
920 : /* The original loop is somehow destroyed. Clear the information. */
921 0 : loop->orig_loop_num = 0;
922 0 : return NULL;
923 : }
924 :
925 162 : if (loop != orig)
926 91 : bb = nearest_common_dominator (CDI_DOMINATORS, loop->header, orig->header);
927 : else
928 71 : bb = loop_preheader_edge (loop)->src;
929 :
930 162 : outer = bb->loop_father;
931 162 : entry = ENTRY_BLOCK_PTR_FOR_FN (fun);
932 :
933 : /* Look upward in dominance tree. */
934 743 : for (; bb != entry && flow_bb_inside_loop_p (outer, bb);
935 581 : bb = get_immediate_dominator (CDI_DOMINATORS, bb))
936 : {
937 691 : gimple_stmt_iterator gsi = gsi_last_bb (bb);
938 691 : if (!safe_is_a <gcond *> (*gsi))
939 581 : continue;
940 :
941 508 : gsi_prev (&gsi);
942 508 : if (gsi_end_p (gsi))
943 8 : continue;
944 :
945 500 : gimple *g = gsi_stmt (gsi);
946 : /* The guarding internal function call must have the same distribution
947 : alias id. */
948 500 : if (gimple_call_internal_p (g, IFN_LOOP_DIST_ALIAS)
949 500 : && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->orig_loop_num))
950 486396 : return g;
951 : }
952 : return NULL;
953 : }
954 :
955 : /* Set the uids of all the statements in basic blocks inside loop
956 : represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
957 : call guarding the loop which has been if converted. */
958 : static void
959 7768 : set_uid_loop_bbs (loop_vec_info loop_vinfo, gimple *loop_vectorized_call,
960 : function *fun)
961 : {
962 7768 : tree arg = gimple_call_arg (loop_vectorized_call, 1);
963 7768 : basic_block *bbs;
964 7768 : unsigned int i;
965 7768 : class loop *scalar_loop = get_loop (fun, tree_to_shwi (arg));
966 :
967 7768 : LOOP_VINFO_SCALAR_LOOP (loop_vinfo) = scalar_loop;
968 7768 : LOOP_VINFO_SCALAR_MAIN_EXIT (loop_vinfo)
969 7768 : = vec_init_loop_exit_info (scalar_loop);
970 7768 : gcc_checking_assert (vect_loop_vectorized_call (scalar_loop)
971 : == loop_vectorized_call);
972 : /* If we are going to vectorize outer loop, prevent vectorization
973 : of the inner loop in the scalar loop - either the scalar loop is
974 : thrown away, so it is a wasted work, or is used only for
975 : a few iterations. */
976 7768 : if (scalar_loop->inner)
977 : {
978 102 : gimple *g = vect_loop_vectorized_call (scalar_loop->inner);
979 102 : if (g)
980 : {
981 102 : arg = gimple_call_arg (g, 0);
982 102 : get_loop (fun, tree_to_shwi (arg))->dont_vectorize = true;
983 102 : fold_loop_internal_call (g, boolean_false_node);
984 : }
985 : }
986 7768 : bbs = get_loop_body (scalar_loop);
987 42237 : for (i = 0; i < scalar_loop->num_nodes; i++)
988 : {
989 34469 : basic_block bb = bbs[i];
990 34469 : gimple_stmt_iterator gsi;
991 66243 : for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
992 : {
993 31774 : gimple *phi = gsi_stmt (gsi);
994 31774 : gimple_set_uid (phi, 0);
995 : }
996 174952 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
997 : {
998 106014 : gimple *stmt = gsi_stmt (gsi);
999 106014 : gimple_set_uid (stmt, 0);
1000 : }
1001 : }
1002 7768 : free (bbs);
1003 7768 : }
1004 :
1005 : /* Generate vectorized code for LOOP and its epilogues. */
1006 :
1007 : static unsigned
1008 61418 : vect_transform_loops (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1009 : loop_p loop, gimple *loop_vectorized_call,
1010 : function *fun)
1011 : {
1012 61418 : loop_vec_info loop_vinfo = loop_vec_info_for_loop (loop);
1013 :
1014 61418 : if (loop_vectorized_call)
1015 7768 : set_uid_loop_bbs (loop_vinfo, loop_vectorized_call, fun);
1016 :
1017 61418 : unsigned HOST_WIDE_INT bytes;
1018 61418 : if (dump_enabled_p ())
1019 : {
1020 21894 : if (GET_MODE_SIZE (loop_vinfo->vector_mode).is_constant (&bytes))
1021 10947 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
1022 : "%sloop vectorized using %s%wu byte vectors and"
1023 : " unroll factor %u\n",
1024 10947 : LOOP_VINFO_EPILOGUE_P (loop_vinfo)
1025 : ? "epilogue " : "",
1026 10947 : LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)
1027 : ? "masked " : "", bytes,
1028 : (unsigned int) LOOP_VINFO_VECT_FACTOR
1029 10947 : (loop_vinfo).to_constant ());
1030 : else
1031 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
1032 : "%sloop vectorized using variable length vectors\n",
1033 : LOOP_VINFO_EPILOGUE_P (loop_vinfo)
1034 : ? "epilogue " : "");
1035 : }
1036 :
1037 61418 : loop_p new_loop = vect_transform_loop (loop_vinfo,
1038 : loop_vectorized_call);
1039 : /* Now that the loop has been vectorized, allow it to be unrolled
1040 : etc. */
1041 61418 : loop->force_vectorize = false;
1042 :
1043 61418 : if (loop->simduid)
1044 : {
1045 1895 : simduid_to_vf *simduid_to_vf_data = XNEW (simduid_to_vf);
1046 1895 : if (!simduid_to_vf_htab)
1047 1535 : simduid_to_vf_htab = new hash_table<simduid_to_vf> (15);
1048 1895 : simduid_to_vf_data->simduid = DECL_UID (loop->simduid);
1049 1895 : simduid_to_vf_data->vf = loop_vinfo->vectorization_factor;
1050 1895 : *simduid_to_vf_htab->find_slot (simduid_to_vf_data, INSERT)
1051 1895 : = simduid_to_vf_data;
1052 : }
1053 :
1054 : /* We should not have to update virtual SSA form here but some
1055 : transforms involve creating new virtual definitions which makes
1056 : updating difficult.
1057 : We delay the actual update to the end of the pass but avoid
1058 : confusing ourselves by forcing need_ssa_update_p () to false. */
1059 61418 : unsigned todo = 0;
1060 61418 : if (need_ssa_update_p (cfun))
1061 : {
1062 119 : gcc_assert (loop_vinfo->any_known_not_updated_vssa);
1063 119 : fun->gimple_df->ssa_renaming_needed = false;
1064 119 : todo |= TODO_update_ssa_only_virtuals;
1065 : }
1066 61418 : gcc_assert (!need_ssa_update_p (cfun));
1067 :
1068 : /* Epilogue of vectorized loop must be vectorized too. */
1069 61418 : if (new_loop)
1070 7056 : todo |= vect_transform_loops (simduid_to_vf_htab, new_loop, NULL, fun);
1071 :
1072 61418 : return todo;
1073 : }
1074 :
1075 : /* Try to vectorize LOOP. */
1076 :
1077 : static unsigned
1078 469419 : try_vectorize_loop_1 (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1079 : unsigned *num_vectorized_loops, loop_p loop,
1080 : gimple *loop_vectorized_call,
1081 : gimple *loop_dist_alias_call,
1082 : function *fun)
1083 : {
1084 469419 : unsigned ret = 0;
1085 469419 : vec_info_shared shared;
1086 469419 : auto_purge_vect_location sentinel;
1087 469419 : vect_location = find_loop_location (loop);
1088 :
1089 469419 : if (LOCATION_LOCUS (vect_location.get_location_t ()) != UNKNOWN_LOCATION
1090 469419 : && dump_enabled_p ())
1091 14954 : dump_printf (MSG_NOTE | MSG_PRIORITY_INTERNALS,
1092 : "\nAnalyzing loop at %s:%d\n",
1093 14954 : LOCATION_FILE (vect_location.get_location_t ()),
1094 29908 : LOCATION_LINE (vect_location.get_location_t ()));
1095 :
1096 : /* Try to analyze the loop, retaining an opt_problem if dump_enabled_p. */
1097 469419 : opt_loop_vec_info loop_vinfo = vect_analyze_loop (loop, loop_vectorized_call,
1098 : &shared);
1099 469419 : loop->aux = loop_vinfo;
1100 :
1101 469419 : if (!loop_vinfo)
1102 415053 : if (dump_enabled_p ())
1103 5764 : if (opt_problem *problem = loop_vinfo.get_problem ())
1104 : {
1105 5764 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1106 : "couldn't vectorize loop\n");
1107 5764 : problem->emit_and_clear ();
1108 : }
1109 :
1110 469419 : if (!loop_vinfo || !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo))
1111 : {
1112 : /* Free existing information if loop is analyzed with some
1113 : assumptions. */
1114 415053 : if (loop_constraint_set_p (loop, LOOP_C_FINITE))
1115 7872 : vect_free_loop_info_assumptions (loop);
1116 :
1117 : /* If we applied if-conversion then try to vectorize the
1118 : BB of innermost loops.
1119 : ??? Ideally BB vectorization would learn to vectorize
1120 : control flow by applying if-conversion on-the-fly, the
1121 : following retains the if-converted loop body even when
1122 : only non-if-converted parts took part in BB vectorization. */
1123 415053 : if (flag_tree_slp_vectorize != 0
1124 414027 : && loop_vectorized_call
1125 21021 : && ! loop->inner)
1126 : {
1127 20204 : basic_block bb = loop->header;
1128 20204 : bool require_loop_vectorize = false;
1129 40408 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
1130 697238 : !gsi_end_p (gsi); gsi_next (&gsi))
1131 : {
1132 677855 : gimple *stmt = gsi_stmt (gsi);
1133 677855 : gcall *call = dyn_cast <gcall *> (stmt);
1134 1199 : if (call && gimple_call_internal_p (call))
1135 : {
1136 1102 : internal_fn ifn = gimple_call_internal_fn (call);
1137 1102 : if (ifn == IFN_MASK_LOAD
1138 1102 : || ifn == IFN_MASK_STORE
1139 684 : || ifn == IFN_MASK_CALL
1140 : /* Don't keep the if-converted parts when the ifn with
1141 : specific type is not supported by the backend. */
1142 1767 : || (direct_internal_fn_p (ifn)
1143 384 : && !direct_internal_fn_supported_p
1144 384 : (call, OPTIMIZE_FOR_SPEED)))
1145 : {
1146 : require_loop_vectorize = true;
1147 : break;
1148 : }
1149 : }
1150 677034 : gimple_set_uid (stmt, -1);
1151 677034 : gimple_set_visited (stmt, false);
1152 : }
1153 20204 : if (!require_loop_vectorize)
1154 : {
1155 19383 : tree arg = gimple_call_arg (loop_vectorized_call, 1);
1156 19383 : class loop *scalar_loop = get_loop (fun, tree_to_shwi (arg));
1157 19383 : if (vect_slp_if_converted_bb (bb, scalar_loop))
1158 : {
1159 75 : fold_loop_internal_call (loop_vectorized_call,
1160 : boolean_true_node);
1161 75 : loop_vectorized_call = NULL;
1162 75 : ret |= TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1163 : }
1164 : }
1165 : }
1166 : /* If outer loop vectorization fails for LOOP_VECTORIZED guarded
1167 : loop, don't vectorize its inner loop; we'll attempt to
1168 : vectorize LOOP_VECTORIZED guarded inner loop of the scalar
1169 : loop version. */
1170 22047 : if (loop_vectorized_call && loop->inner)
1171 819 : loop->inner->dont_vectorize = true;
1172 415053 : return ret;
1173 : }
1174 :
1175 54366 : if (!dbg_cnt (vect_loop))
1176 : {
1177 : /* Free existing information if loop is analyzed with some
1178 : assumptions. */
1179 4 : if (loop_constraint_set_p (loop, LOOP_C_FINITE))
1180 0 : vect_free_loop_info_assumptions (loop);
1181 4 : return ret;
1182 : }
1183 :
1184 54362 : (*num_vectorized_loops)++;
1185 : /* Transform LOOP and its epilogues. */
1186 54362 : ret |= vect_transform_loops (simduid_to_vf_htab, loop,
1187 : loop_vectorized_call, fun);
1188 :
1189 54362 : if (loop_vectorized_call)
1190 : {
1191 7768 : fold_loop_internal_call (loop_vectorized_call, boolean_true_node);
1192 7768 : ret |= TODO_cleanup_cfg;
1193 : }
1194 54362 : if (loop_dist_alias_call)
1195 : {
1196 8 : tree value = gimple_call_arg (loop_dist_alias_call, 1);
1197 8 : fold_loop_internal_call (loop_dist_alias_call, value);
1198 8 : ret |= TODO_cleanup_cfg;
1199 : }
1200 :
1201 : return ret;
1202 469419 : }
1203 :
1204 : /* Try to vectorize LOOP. */
1205 :
1206 : static unsigned
1207 500741 : try_vectorize_loop (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1208 : unsigned *num_vectorized_loops, loop_p loop,
1209 : function *fun)
1210 : {
1211 500741 : if (!((flag_tree_loop_vectorize
1212 496159 : && optimize_loop_nest_for_speed_p (loop))
1213 33125 : || loop->force_vectorize))
1214 : return 0;
1215 :
1216 469419 : return try_vectorize_loop_1 (simduid_to_vf_htab, num_vectorized_loops, loop,
1217 : vect_loop_vectorized_call (loop),
1218 469419 : vect_loop_dist_alias_call (loop, fun), fun);
1219 : }
1220 :
1221 :
1222 : /* Loop autovectorization. */
1223 :
1224 : namespace {
1225 :
1226 : const pass_data pass_data_vectorize =
1227 : {
1228 : GIMPLE_PASS, /* type */
1229 : "vect", /* name */
1230 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1231 : TV_TREE_VECTORIZATION, /* tv_id */
1232 : ( PROP_cfg | PROP_ssa ), /* properties_required */
1233 : 0, /* properties_provided */
1234 : 0, /* properties_destroyed */
1235 : 0, /* todo_flags_start */
1236 : 0, /* todo_flags_finish */
1237 : };
1238 :
1239 : class pass_vectorize : public gimple_opt_pass
1240 : {
1241 : public:
1242 285722 : pass_vectorize (gcc::context *ctxt)
1243 571444 : : gimple_opt_pass (pass_data_vectorize, ctxt)
1244 : {}
1245 :
1246 : /* opt_pass methods: */
1247 241458 : bool gate (function *fun) final override
1248 : {
1249 241458 : return flag_tree_loop_vectorize || fun->has_force_vectorize_loops;
1250 : }
1251 :
1252 : unsigned int execute (function *) final override;
1253 :
1254 : }; // class pass_vectorize
1255 :
1256 : /* Function vectorize_loops.
1257 :
1258 : Entry point to loop vectorization phase. */
1259 :
1260 : unsigned
1261 208529 : pass_vectorize::execute (function *fun)
1262 : {
1263 208529 : unsigned int i;
1264 208529 : unsigned int num_vectorized_loops = 0;
1265 208529 : unsigned int vect_loops_num;
1266 208529 : hash_table<simduid_to_vf> *simduid_to_vf_htab = NULL;
1267 208529 : hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
1268 208529 : bool any_ifcvt_loops = false;
1269 208529 : unsigned ret = 0;
1270 :
1271 208529 : vect_loops_num = number_of_loops (fun);
1272 :
1273 : /* Bail out if there are no loops. */
1274 208529 : if (vect_loops_num <= 1)
1275 : return 0;
1276 :
1277 208529 : vect_slp_init ();
1278 :
1279 208529 : if (fun->has_simduid_loops)
1280 5601 : note_simd_array_uses (&simd_array_to_simduid_htab, fun);
1281 :
1282 : /* ----------- Analyze loops. ----------- */
1283 208529 : enable_ranger (fun);
1284 :
1285 : /* If some loop was duplicated, it gets bigger number
1286 : than all previously defined loops. This fact allows us to run
1287 : only over initial loops skipping newly generated ones. */
1288 1161078 : for (auto loop : loops_list (fun, 0))
1289 535491 : if (loop->dont_vectorize)
1290 : {
1291 35671 : any_ifcvt_loops = true;
1292 : /* If-conversion sometimes versions both the outer loop
1293 : (for the case when outer loop vectorization might be
1294 : desirable) as well as the inner loop in the scalar version
1295 : of the loop. So we have:
1296 : if (LOOP_VECTORIZED (1, 3))
1297 : {
1298 : loop1
1299 : loop2
1300 : }
1301 : else
1302 : loop3 (copy of loop1)
1303 : if (LOOP_VECTORIZED (4, 5))
1304 : loop4 (copy of loop2)
1305 : else
1306 : loop5 (copy of loop4)
1307 : If loops' iteration gives us loop3 first (which has
1308 : dont_vectorize set), make sure to process loop1 before loop4;
1309 : so that we can prevent vectorization of loop4 if loop1
1310 : is successfully vectorized. */
1311 35671 : if (loop->inner)
1312 : {
1313 2164 : gimple *loop_vectorized_call
1314 2164 : = vect_loop_vectorized_call (loop);
1315 2164 : if (loop_vectorized_call
1316 2164 : && vect_loop_vectorized_call (loop->inner))
1317 : {
1318 921 : tree arg = gimple_call_arg (loop_vectorized_call, 0);
1319 921 : class loop *vector_loop
1320 921 : = get_loop (fun, tree_to_shwi (arg));
1321 921 : if (vector_loop && vector_loop != loop)
1322 : {
1323 : /* Make sure we don't vectorize it twice. */
1324 921 : vector_loop->dont_vectorize = true;
1325 921 : ret |= try_vectorize_loop (simduid_to_vf_htab,
1326 : &num_vectorized_loops,
1327 : vector_loop, fun);
1328 : }
1329 : }
1330 : }
1331 : }
1332 : else
1333 499820 : ret |= try_vectorize_loop (simduid_to_vf_htab, &num_vectorized_loops,
1334 208529 : loop, fun);
1335 :
1336 208529 : vect_location = dump_user_location_t ();
1337 :
1338 208529 : statistics_counter_event (fun, "Vectorized loops", num_vectorized_loops);
1339 208529 : if (dump_enabled_p ()
1340 208529 : || (num_vectorized_loops > 0 && dump_enabled_p ()))
1341 11599 : dump_printf_loc (MSG_NOTE, vect_location,
1342 : "vectorized %u loops in function.\n",
1343 : num_vectorized_loops);
1344 :
1345 : /* ----------- Finalize. ----------- */
1346 208529 : disable_ranger (fun);
1347 :
1348 208529 : if (any_ifcvt_loops)
1349 322008 : for (i = 1; i < number_of_loops (fun); i++)
1350 : {
1351 139332 : class loop *loop = get_loop (fun, i);
1352 139332 : if (loop && loop->dont_vectorize)
1353 : {
1354 38224 : gimple *g = vect_loop_vectorized_call (loop);
1355 38224 : if (g)
1356 : {
1357 21247 : fold_loop_internal_call (g, boolean_false_node);
1358 21247 : loop->dont_vectorize = false;
1359 21247 : ret |= TODO_cleanup_cfg;
1360 21247 : g = NULL;
1361 : }
1362 : else
1363 16977 : g = vect_loop_dist_alias_call (loop, fun);
1364 :
1365 38224 : if (g)
1366 : {
1367 28 : fold_loop_internal_call (g, boolean_false_node);
1368 28 : loop->dont_vectorize = false;
1369 28 : ret |= TODO_cleanup_cfg;
1370 : }
1371 : }
1372 : }
1373 :
1374 : /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1375 208529 : if (fun->has_simduid_loops)
1376 : {
1377 5601 : adjust_simduid_builtins (simduid_to_vf_htab, fun);
1378 : /* Avoid stale SCEV cache entries for the SIMD_LANE defs. */
1379 5601 : scev_reset ();
1380 : }
1381 : /* Shrink any "omp array simd" temporary arrays to the
1382 : actual vectorization factors. */
1383 208529 : if (simd_array_to_simduid_htab)
1384 2204 : shrink_simd_arrays (simd_array_to_simduid_htab, simduid_to_vf_htab);
1385 208529 : delete simduid_to_vf_htab;
1386 208529 : fun->has_simduid_loops = false;
1387 :
1388 208529 : if (num_vectorized_loops > 0)
1389 : {
1390 : /* We are collecting some corner cases where we need to update
1391 : virtual SSA form via the TODO but delete the queued update-SSA
1392 : state. Force renaming if we think that might be necessary. */
1393 36820 : if (ret & TODO_update_ssa_only_virtuals)
1394 89 : mark_virtual_operands_for_renaming (cfun);
1395 : /* If we vectorized any loop only virtual SSA form needs to be updated.
1396 : ??? Also while we try hard to update loop-closed SSA form we fail
1397 : to properly do this in some corner-cases (see PR56286). */
1398 36820 : rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_only_virtuals);
1399 36820 : ret |= TODO_cleanup_cfg;
1400 : }
1401 :
1402 1873052 : for (i = 1; i < number_of_loops (fun); i++)
1403 : {
1404 727997 : loop_vec_info loop_vinfo;
1405 727997 : bool has_mask_store;
1406 :
1407 727997 : class loop *loop = get_loop (fun, i);
1408 727997 : if (!loop || !loop->aux)
1409 666575 : continue;
1410 61422 : loop_vinfo = (loop_vec_info) loop->aux;
1411 61422 : has_mask_store = LOOP_VINFO_HAS_MASK_STORE (loop_vinfo);
1412 61422 : delete loop_vinfo;
1413 61422 : if (has_mask_store
1414 61422 : && targetm.vectorize.empty_mask_is_expensive (IFN_MASK_STORE))
1415 499 : optimize_mask_stores (loop);
1416 :
1417 61422 : auto_bitmap exit_bbs;
1418 : /* Perform local CSE, this esp. helps because we emit code for
1419 : predicates that need to be shared for optimal predicate usage.
1420 : However reassoc will re-order them and prevent CSE from working
1421 : as it should. CSE only the loop body, not the entry. */
1422 61422 : auto_vec<edge> exits = get_loop_exit_edges (loop);
1423 247234 : for (edge exit : exits)
1424 62968 : bitmap_set_bit (exit_bbs, exit->dest->index);
1425 :
1426 61422 : edge entry = EDGE_PRED (loop_preheader_edge (loop)->src, 0);
1427 61422 : do_rpo_vn (fun, entry, exit_bbs);
1428 :
1429 61422 : loop->aux = NULL;
1430 61422 : }
1431 :
1432 208529 : vect_slp_fini ();
1433 :
1434 208529 : return ret;
1435 : }
1436 :
1437 : } // anon namespace
1438 :
1439 : gimple_opt_pass *
1440 285722 : make_pass_vectorize (gcc::context *ctxt)
1441 : {
1442 285722 : return new pass_vectorize (ctxt);
1443 : }
1444 :
1445 : /* Entry point to the simduid cleanup pass. */
1446 :
1447 : namespace {
1448 :
1449 : const pass_data pass_data_simduid_cleanup =
1450 : {
1451 : GIMPLE_PASS, /* type */
1452 : "simduid", /* name */
1453 : OPTGROUP_NONE, /* optinfo_flags */
1454 : TV_NONE, /* tv_id */
1455 : ( PROP_ssa | PROP_cfg ), /* properties_required */
1456 : 0, /* properties_provided */
1457 : 0, /* properties_destroyed */
1458 : 0, /* todo_flags_start */
1459 : 0, /* todo_flags_finish */
1460 : };
1461 :
1462 : class pass_simduid_cleanup : public gimple_opt_pass
1463 : {
1464 : public:
1465 571444 : pass_simduid_cleanup (gcc::context *ctxt)
1466 1142888 : : gimple_opt_pass (pass_data_simduid_cleanup, ctxt)
1467 : {}
1468 :
1469 : /* opt_pass methods: */
1470 285722 : opt_pass * clone () final override
1471 : {
1472 285722 : return new pass_simduid_cleanup (m_ctxt);
1473 : }
1474 2513634 : bool gate (function *fun) final override { return fun->has_simduid_loops; }
1475 : unsigned int execute (function *) final override;
1476 :
1477 : }; // class pass_simduid_cleanup
1478 :
1479 : unsigned int
1480 2208 : pass_simduid_cleanup::execute (function *fun)
1481 : {
1482 2208 : hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
1483 :
1484 2208 : note_simd_array_uses (&simd_array_to_simduid_htab, fun);
1485 :
1486 : /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1487 2208 : adjust_simduid_builtins (NULL, fun);
1488 :
1489 : /* Shrink any "omp array simd" temporary arrays to the
1490 : actual vectorization factors. */
1491 2208 : if (simd_array_to_simduid_htab)
1492 4 : shrink_simd_arrays (simd_array_to_simduid_htab, NULL);
1493 2208 : fun->has_simduid_loops = false;
1494 2208 : return 0;
1495 : }
1496 :
1497 : } // anon namespace
1498 :
1499 : gimple_opt_pass *
1500 285722 : make_pass_simduid_cleanup (gcc::context *ctxt)
1501 : {
1502 285722 : return new pass_simduid_cleanup (ctxt);
1503 : }
1504 :
1505 :
1506 : /* Entry point to basic block SLP phase. */
1507 :
1508 : namespace {
1509 :
1510 : const pass_data pass_data_slp_vectorize =
1511 : {
1512 : GIMPLE_PASS, /* type */
1513 : "slp", /* name */
1514 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1515 : TV_TREE_SLP_VECTORIZATION, /* tv_id */
1516 : ( PROP_ssa | PROP_cfg ), /* properties_required */
1517 : 0, /* properties_provided */
1518 : 0, /* properties_destroyed */
1519 : 0, /* todo_flags_start */
1520 : TODO_update_ssa, /* todo_flags_finish */
1521 : };
1522 :
1523 : class pass_slp_vectorize : public gimple_opt_pass
1524 : {
1525 : public:
1526 571444 : pass_slp_vectorize (gcc::context *ctxt)
1527 1142888 : : gimple_opt_pass (pass_data_slp_vectorize, ctxt)
1528 : {}
1529 :
1530 : /* opt_pass methods: */
1531 285722 : opt_pass * clone () final override { return new pass_slp_vectorize (m_ctxt); }
1532 1041489 : bool gate (function *) final override { return flag_tree_slp_vectorize != 0; }
1533 : unsigned int execute (function *) final override;
1534 :
1535 : }; // class pass_slp_vectorize
1536 :
1537 : unsigned int
1538 909169 : pass_slp_vectorize::execute (function *fun)
1539 : {
1540 909169 : auto_purge_vect_location sentinel;
1541 909169 : basic_block bb;
1542 :
1543 909169 : bool in_loop_pipeline = scev_initialized_p ();
1544 909169 : if (!in_loop_pipeline)
1545 : {
1546 702269 : loop_optimizer_init (LOOPS_NORMAL);
1547 702269 : scev_initialize ();
1548 : }
1549 :
1550 : /* Mark all stmts as not belonging to the current region and unvisited. */
1551 11588604 : FOR_EACH_BB_FN (bb, fun)
1552 : {
1553 15461299 : for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
1554 4781864 : gsi_next (&gsi))
1555 : {
1556 4781864 : gphi *stmt = gsi.phi ();
1557 4781864 : gimple_set_uid (stmt, -1);
1558 4781864 : gimple_set_visited (stmt, false);
1559 : }
1560 100041640 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1561 78682770 : gsi_next (&gsi))
1562 : {
1563 78682770 : gimple *stmt = gsi_stmt (gsi);
1564 78682770 : gimple_set_uid (stmt, -1);
1565 78682770 : gimple_set_visited (stmt, false);
1566 : }
1567 : }
1568 :
1569 909169 : vect_slp_init ();
1570 :
1571 909169 : vect_slp_function (fun);
1572 :
1573 909169 : vect_slp_fini ();
1574 :
1575 909169 : if (!in_loop_pipeline)
1576 : {
1577 702269 : scev_finalize ();
1578 702269 : loop_optimizer_finalize ();
1579 : }
1580 :
1581 1818338 : return 0;
1582 909169 : }
1583 :
1584 : } // anon namespace
1585 :
1586 : gimple_opt_pass *
1587 285722 : make_pass_slp_vectorize (gcc::context *ctxt)
1588 : {
1589 285722 : return new pass_slp_vectorize (ctxt);
1590 : }
1591 :
1592 :
1593 : /* Increase alignment of global arrays to improve vectorization potential.
1594 : TODO:
1595 : - Consider also structs that have an array field.
1596 : - Use ipa analysis to prune arrays that can't be vectorized?
1597 : This should involve global alignment analysis and in the future also
1598 : array padding. */
1599 :
1600 : static unsigned get_vec_alignment_for_type (tree);
1601 : static hash_map<tree, unsigned> *type_align_map;
1602 :
1603 : /* Return alignment of array's vector type corresponding to scalar type.
1604 : 0 if no vector type exists. */
1605 : static unsigned
1606 0 : get_vec_alignment_for_array_type (tree type)
1607 : {
1608 0 : gcc_assert (TREE_CODE (type) == ARRAY_TYPE);
1609 0 : poly_uint64 array_size, vector_size;
1610 :
1611 0 : tree scalar_type = strip_array_types (type);
1612 0 : tree vectype = get_related_vectype_for_scalar_type (VOIDmode, scalar_type);
1613 0 : if (!vectype
1614 0 : || !poly_int_tree_p (TYPE_SIZE (type), &array_size)
1615 0 : || !poly_int_tree_p (TYPE_SIZE (vectype), &vector_size)
1616 0 : || maybe_lt (array_size, vector_size))
1617 0 : return 0;
1618 :
1619 0 : return TYPE_ALIGN (vectype);
1620 : }
1621 :
1622 : /* Return alignment of field having maximum alignment of vector type
1623 : corresponding to it's scalar type. For now, we only consider fields whose
1624 : offset is a multiple of it's vector alignment.
1625 : 0 if no suitable field is found. */
1626 : static unsigned
1627 0 : get_vec_alignment_for_record_type (tree type)
1628 : {
1629 0 : gcc_assert (TREE_CODE (type) == RECORD_TYPE);
1630 :
1631 0 : unsigned max_align = 0, alignment;
1632 0 : HOST_WIDE_INT offset;
1633 0 : tree offset_tree;
1634 :
1635 0 : if (TYPE_PACKED (type))
1636 : return 0;
1637 :
1638 0 : unsigned *slot = type_align_map->get (type);
1639 0 : if (slot)
1640 0 : return *slot;
1641 :
1642 0 : for (tree field = first_field (type);
1643 0 : field != NULL_TREE;
1644 0 : field = DECL_CHAIN (field))
1645 : {
1646 : /* Skip if not FIELD_DECL or if alignment is set by user. */
1647 0 : if (TREE_CODE (field) != FIELD_DECL
1648 0 : || DECL_USER_ALIGN (field)
1649 0 : || DECL_ARTIFICIAL (field))
1650 0 : continue;
1651 :
1652 : /* We don't need to process the type further if offset is variable,
1653 : since the offsets of remaining members will also be variable. */
1654 0 : if (TREE_CODE (DECL_FIELD_OFFSET (field)) != INTEGER_CST
1655 0 : || TREE_CODE (DECL_FIELD_BIT_OFFSET (field)) != INTEGER_CST)
1656 : break;
1657 :
1658 : /* Similarly stop processing the type if offset_tree
1659 : does not fit in unsigned HOST_WIDE_INT. */
1660 0 : offset_tree = bit_position (field);
1661 0 : if (!tree_fits_uhwi_p (offset_tree))
1662 : break;
1663 :
1664 0 : offset = tree_to_uhwi (offset_tree);
1665 0 : alignment = get_vec_alignment_for_type (TREE_TYPE (field));
1666 :
1667 : /* Get maximum alignment of vectorized field/array among those members
1668 : whose offset is multiple of the vector alignment. */
1669 0 : if (alignment
1670 0 : && (offset % alignment == 0)
1671 0 : && (alignment > max_align))
1672 0 : max_align = alignment;
1673 : }
1674 :
1675 0 : type_align_map->put (type, max_align);
1676 0 : return max_align;
1677 : }
1678 :
1679 : /* Return alignment of vector type corresponding to decl's scalar type
1680 : or 0 if it doesn't exist or the vector alignment is lesser than
1681 : decl's alignment. */
1682 : static unsigned
1683 0 : get_vec_alignment_for_type (tree type)
1684 : {
1685 0 : if (type == NULL_TREE)
1686 : return 0;
1687 :
1688 0 : gcc_assert (TYPE_P (type));
1689 :
1690 0 : static unsigned alignment = 0;
1691 0 : switch (TREE_CODE (type))
1692 : {
1693 0 : case ARRAY_TYPE:
1694 0 : alignment = get_vec_alignment_for_array_type (type);
1695 0 : break;
1696 0 : case RECORD_TYPE:
1697 0 : alignment = get_vec_alignment_for_record_type (type);
1698 0 : break;
1699 0 : default:
1700 0 : alignment = 0;
1701 0 : break;
1702 : }
1703 :
1704 0 : return (alignment > TYPE_ALIGN (type)) ? alignment : 0;
1705 : }
1706 :
1707 : /* Entry point to increase_alignment pass. */
1708 : static unsigned int
1709 0 : increase_alignment (void)
1710 : {
1711 0 : varpool_node *vnode;
1712 :
1713 0 : vect_location = dump_user_location_t ();
1714 0 : type_align_map = new hash_map<tree, unsigned>;
1715 :
1716 : /* Increase the alignment of all global arrays for vectorization. */
1717 0 : FOR_EACH_DEFINED_VARIABLE (vnode)
1718 : {
1719 0 : tree decl = vnode->decl;
1720 0 : unsigned int alignment;
1721 :
1722 0 : if ((decl_in_symtab_p (decl)
1723 0 : && !symtab_node::get (decl)->can_increase_alignment_p ())
1724 0 : || DECL_USER_ALIGN (decl) || DECL_ARTIFICIAL (decl))
1725 0 : continue;
1726 :
1727 0 : alignment = get_vec_alignment_for_type (TREE_TYPE (decl));
1728 0 : if (alignment && vect_can_force_dr_alignment_p (decl, alignment))
1729 : {
1730 0 : vnode->increase_alignment (alignment);
1731 0 : if (dump_enabled_p ())
1732 0 : dump_printf (MSG_NOTE, "Increasing alignment of decl: %T\n", decl);
1733 : }
1734 : }
1735 :
1736 0 : delete type_align_map;
1737 0 : return 0;
1738 : }
1739 :
1740 :
1741 : namespace {
1742 :
1743 : const pass_data pass_data_ipa_increase_alignment =
1744 : {
1745 : SIMPLE_IPA_PASS, /* type */
1746 : "increase_alignment", /* name */
1747 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1748 : TV_IPA_OPT, /* tv_id */
1749 : 0, /* properties_required */
1750 : 0, /* properties_provided */
1751 : 0, /* properties_destroyed */
1752 : 0, /* todo_flags_start */
1753 : 0, /* todo_flags_finish */
1754 : };
1755 :
1756 : class pass_ipa_increase_alignment : public simple_ipa_opt_pass
1757 : {
1758 : public:
1759 285722 : pass_ipa_increase_alignment (gcc::context *ctxt)
1760 571444 : : simple_ipa_opt_pass (pass_data_ipa_increase_alignment, ctxt)
1761 : {}
1762 :
1763 : /* opt_pass methods: */
1764 229960 : bool gate (function *) final override
1765 : {
1766 229960 : return flag_section_anchors && flag_tree_loop_vectorize;
1767 : }
1768 :
1769 0 : unsigned int execute (function *) final override
1770 : {
1771 0 : return increase_alignment ();
1772 : }
1773 :
1774 : }; // class pass_ipa_increase_alignment
1775 :
1776 : } // anon namespace
1777 :
1778 : simple_ipa_opt_pass *
1779 285722 : make_pass_ipa_increase_alignment (gcc::context *ctxt)
1780 : {
1781 285722 : return new pass_ipa_increase_alignment (ctxt);
1782 : }
1783 :
1784 : /* If the condition represented by T is a comparison or the SSA name
1785 : result of a comparison, extract the comparison's operands. Represent
1786 : T as NE_EXPR <T, 0> otherwise. */
1787 :
1788 : void
1789 61908 : scalar_cond_masked_key::get_cond_ops_from_tree (tree t)
1790 : {
1791 61908 : if (TREE_CODE_CLASS (TREE_CODE (t)) == tcc_comparison)
1792 : {
1793 0 : this->code = TREE_CODE (t);
1794 0 : this->op0 = TREE_OPERAND (t, 0);
1795 0 : this->op1 = TREE_OPERAND (t, 1);
1796 0 : this->inverted_p = false;
1797 0 : return;
1798 : }
1799 :
1800 61908 : if (TREE_CODE (t) == SSA_NAME)
1801 24620 : if (gassign *stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (t)))
1802 : {
1803 24620 : tree_code code = gimple_assign_rhs_code (stmt);
1804 24620 : if (TREE_CODE_CLASS (code) == tcc_comparison)
1805 : {
1806 16736 : this->code = code;
1807 16736 : this->op0 = gimple_assign_rhs1 (stmt);
1808 16736 : this->op1 = gimple_assign_rhs2 (stmt);
1809 16736 : this->inverted_p = false;
1810 16736 : return;
1811 : }
1812 7884 : else if (code == BIT_NOT_EXPR)
1813 : {
1814 3502 : tree n_op = gimple_assign_rhs1 (stmt);
1815 3502 : if ((stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (n_op))))
1816 : {
1817 3502 : code = gimple_assign_rhs_code (stmt);
1818 3502 : if (TREE_CODE_CLASS (code) == tcc_comparison)
1819 : {
1820 3468 : this->code = code;
1821 3468 : this->op0 = gimple_assign_rhs1 (stmt);
1822 3468 : this->op1 = gimple_assign_rhs2 (stmt);
1823 3468 : this->inverted_p = true;
1824 3468 : return;
1825 : }
1826 : }
1827 : }
1828 : }
1829 :
1830 41704 : this->code = NE_EXPR;
1831 41704 : this->op0 = t;
1832 41704 : this->op1 = build_zero_cst (TREE_TYPE (t));
1833 41704 : this->inverted_p = false;
1834 : }
1835 :
1836 : /* See the comment above the declaration for details. */
1837 :
1838 : unsigned int
1839 0 : vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
1840 : stmt_vec_info stmt_info, slp_tree,
1841 : tree vectype, int misalign,
1842 : vect_cost_model_location where)
1843 : {
1844 0 : unsigned int cost
1845 0 : = builtin_vectorization_cost (kind, vectype, misalign) * count;
1846 0 : return record_stmt_cost (stmt_info, where, cost);
1847 : }
1848 :
1849 : /* See the comment above the declaration for details. */
1850 :
1851 : void
1852 1690129 : vector_costs::finish_cost (const vector_costs *)
1853 : {
1854 1690129 : gcc_assert (!m_finished);
1855 1690129 : m_finished = true;
1856 1690129 : }
1857 :
1858 : /* Record a base cost of COST units against WHERE. If STMT_INFO is
1859 : nonnull, use it to adjust the cost based on execution frequency
1860 : (where appropriate). */
1861 :
1862 : unsigned int
1863 0 : vector_costs::record_stmt_cost (stmt_vec_info stmt_info,
1864 : vect_cost_model_location where,
1865 : unsigned int cost)
1866 : {
1867 0 : cost = adjust_cost_for_freq (stmt_info, where, cost);
1868 0 : m_costs[where] += cost;
1869 0 : return cost;
1870 : }
1871 :
1872 : /* COST is the base cost we have calculated for an operation in location WHERE.
1873 : If STMT_INFO is nonnull, use it to adjust the cost based on execution
1874 : frequency (where appropriate). Return the adjusted cost. */
1875 :
1876 : unsigned int
1877 6656078 : vector_costs::adjust_cost_for_freq (stmt_vec_info stmt_info,
1878 : vect_cost_model_location where,
1879 : unsigned int cost)
1880 : {
1881 : /* Statements in an inner loop relative to the loop being
1882 : vectorized are weighted more heavily. The value here is
1883 : arbitrary and could potentially be improved with analysis. */
1884 6656078 : if (where == vect_body
1885 6656078 : && stmt_info
1886 6656078 : && stmt_in_inner_loop_p (m_vinfo, stmt_info))
1887 : {
1888 4578 : loop_vec_info loop_vinfo = as_a<loop_vec_info> (m_vinfo);
1889 4578 : cost *= LOOP_VINFO_INNER_LOOP_COST_FACTOR (loop_vinfo);
1890 : }
1891 6656078 : return cost;
1892 : }
1893 :
1894 : /* See the comment above the declaration for details. */
1895 :
1896 : bool
1897 5 : vector_costs::better_main_loop_than_p (const vector_costs *other) const
1898 : {
1899 5 : int diff = compare_inside_loop_cost (other);
1900 5 : if (diff != 0)
1901 5 : return diff < 0;
1902 :
1903 : /* If there's nothing to choose between the loop bodies, see whether
1904 : there's a difference in the prologue and epilogue costs. */
1905 0 : diff = compare_outside_loop_cost (other);
1906 0 : if (diff != 0)
1907 0 : return diff < 0;
1908 :
1909 : return false;
1910 : }
1911 :
1912 :
1913 : /* See the comment above the declaration for details. */
1914 :
1915 : bool
1916 0 : vector_costs::better_epilogue_loop_than_p (const vector_costs *other,
1917 : loop_vec_info main_loop) const
1918 : {
1919 0 : loop_vec_info this_loop_vinfo = as_a<loop_vec_info> (this->m_vinfo);
1920 0 : loop_vec_info other_loop_vinfo = as_a<loop_vec_info> (other->m_vinfo);
1921 :
1922 0 : poly_int64 this_vf = LOOP_VINFO_VECT_FACTOR (this_loop_vinfo);
1923 0 : poly_int64 other_vf = LOOP_VINFO_VECT_FACTOR (other_loop_vinfo);
1924 :
1925 0 : poly_uint64 main_poly_vf = LOOP_VINFO_VECT_FACTOR (main_loop);
1926 0 : unsigned HOST_WIDE_INT main_vf;
1927 0 : unsigned HOST_WIDE_INT other_factor, this_factor, other_cost, this_cost;
1928 : /* If we can determine how many iterations are left for the epilogue
1929 : loop, that is if both the main loop's vectorization factor and number
1930 : of iterations are constant, then we use them to calculate the cost of
1931 : the epilogue loop together with a 'likely value' for the epilogues
1932 : vectorization factor. Otherwise we use the main loop's vectorization
1933 : factor and the maximum poly value for the epilogue's. If the target
1934 : has not provided with a sensible upper bound poly vectorization
1935 : factors are likely to be favored over constant ones. */
1936 0 : if (main_poly_vf.is_constant (&main_vf)
1937 0 : && LOOP_VINFO_NITERS_KNOWN_P (main_loop))
1938 : {
1939 0 : unsigned HOST_WIDE_INT niters
1940 0 : = LOOP_VINFO_INT_NITERS (main_loop) % main_vf;
1941 0 : HOST_WIDE_INT other_likely_vf
1942 0 : = estimated_poly_value (other_vf, POLY_VALUE_LIKELY);
1943 0 : HOST_WIDE_INT this_likely_vf
1944 0 : = estimated_poly_value (this_vf, POLY_VALUE_LIKELY);
1945 :
1946 : /* If the epilogue is using partial vectors we account for the
1947 : partial iteration here too. */
1948 0 : other_factor = niters / other_likely_vf;
1949 0 : if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (other_loop_vinfo)
1950 0 : && niters % other_likely_vf != 0)
1951 0 : other_factor++;
1952 :
1953 0 : this_factor = niters / this_likely_vf;
1954 0 : if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (this_loop_vinfo)
1955 0 : && niters % this_likely_vf != 0)
1956 0 : this_factor++;
1957 : }
1958 : else
1959 : {
1960 0 : unsigned HOST_WIDE_INT main_vf_max
1961 0 : = estimated_poly_value (main_poly_vf, POLY_VALUE_MAX);
1962 0 : unsigned HOST_WIDE_INT other_vf_max
1963 0 : = estimated_poly_value (other_vf, POLY_VALUE_MAX);
1964 0 : unsigned HOST_WIDE_INT this_vf_max
1965 0 : = estimated_poly_value (this_vf, POLY_VALUE_MAX);
1966 :
1967 0 : other_factor = CEIL (main_vf_max, other_vf_max);
1968 0 : this_factor = CEIL (main_vf_max, this_vf_max);
1969 :
1970 : /* If the loop is not using partial vectors then it will iterate one
1971 : time less than one that does. It is safe to subtract one here,
1972 : because the main loop's vf is always at least 2x bigger than that
1973 : of an epilogue. */
1974 0 : if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (other_loop_vinfo))
1975 0 : other_factor -= 1;
1976 0 : if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (this_loop_vinfo))
1977 0 : this_factor -= 1;
1978 : }
1979 :
1980 : /* Compute the costs by multiplying the inside costs with the factor and
1981 : add the outside costs for a more complete picture. The factor is the
1982 : amount of times we are expecting to iterate this epilogue. */
1983 0 : other_cost = other->body_cost () * other_factor;
1984 0 : this_cost = this->body_cost () * this_factor;
1985 0 : other_cost += other->outside_cost ();
1986 0 : this_cost += this->outside_cost ();
1987 0 : return this_cost < other_cost;
1988 : }
1989 :
1990 : /* A <=>-style subroutine of better_main_loop_than_p. Check whether we can
1991 : determine the return value of better_main_loop_than_p by comparing the
1992 : inside (loop body) costs of THIS and OTHER. Return:
1993 :
1994 : * -1 if better_main_loop_than_p should return true.
1995 : * 1 if better_main_loop_than_p should return false.
1996 : * 0 if we can't decide. */
1997 :
1998 : int
1999 5 : vector_costs::compare_inside_loop_cost (const vector_costs *other) const
2000 : {
2001 5 : loop_vec_info this_loop_vinfo = as_a<loop_vec_info> (this->m_vinfo);
2002 5 : loop_vec_info other_loop_vinfo = as_a<loop_vec_info> (other->m_vinfo);
2003 :
2004 5 : struct loop *loop = LOOP_VINFO_LOOP (this_loop_vinfo);
2005 5 : gcc_assert (LOOP_VINFO_LOOP (other_loop_vinfo) == loop);
2006 :
2007 5 : poly_int64 this_vf = LOOP_VINFO_VECT_FACTOR (this_loop_vinfo);
2008 5 : poly_int64 other_vf = LOOP_VINFO_VECT_FACTOR (other_loop_vinfo);
2009 :
2010 : /* Limit the VFs to what is likely to be the maximum number of iterations,
2011 : to handle cases in which at least one loop_vinfo is fully-masked. */
2012 5 : HOST_WIDE_INT estimated_max_niter = likely_max_stmt_executions_int (loop);
2013 5 : if (estimated_max_niter != -1)
2014 : {
2015 5 : if (estimated_poly_value (this_vf, POLY_VALUE_MIN)
2016 : >= estimated_max_niter)
2017 : this_vf = estimated_max_niter;
2018 5 : if (estimated_poly_value (other_vf, POLY_VALUE_MIN)
2019 : >= estimated_max_niter)
2020 : other_vf = estimated_max_niter;
2021 : }
2022 :
2023 : /* Check whether the (fractional) cost per scalar iteration is lower or
2024 : higher: this_inside_cost / this_vf vs. other_inside_cost / other_vf. */
2025 5 : poly_int64 rel_this = this_loop_vinfo->vector_costs->body_cost () * other_vf;
2026 5 : poly_int64 rel_other
2027 5 : = other_loop_vinfo->vector_costs->body_cost () * this_vf;
2028 :
2029 5 : HOST_WIDE_INT est_rel_this_min
2030 5 : = estimated_poly_value (rel_this, POLY_VALUE_MIN);
2031 5 : HOST_WIDE_INT est_rel_this_max
2032 5 : = estimated_poly_value (rel_this, POLY_VALUE_MAX);
2033 :
2034 5 : HOST_WIDE_INT est_rel_other_min
2035 5 : = estimated_poly_value (rel_other, POLY_VALUE_MIN);
2036 5 : HOST_WIDE_INT est_rel_other_max
2037 5 : = estimated_poly_value (rel_other, POLY_VALUE_MAX);
2038 :
2039 : /* Check first if we can make out an unambigous total order from the minimum
2040 : and maximum estimates. */
2041 5 : if (est_rel_this_min < est_rel_other_min
2042 : && est_rel_this_max < est_rel_other_max)
2043 : return -1;
2044 :
2045 4 : if (est_rel_other_min < est_rel_this_min
2046 : && est_rel_other_max < est_rel_this_max)
2047 4 : return 1;
2048 :
2049 : /* When other_loop_vinfo uses a variable vectorization factor,
2050 : we know that it has a lower cost for at least one runtime VF.
2051 : However, we don't know how likely that VF is.
2052 :
2053 : One option would be to compare the costs for the estimated VFs.
2054 : The problem is that that can put too much pressure on the cost
2055 : model. E.g. if the estimated VF is also the lowest possible VF,
2056 : and if other_loop_vinfo is 1 unit worse than this_loop_vinfo
2057 : for the estimated VF, we'd then choose this_loop_vinfo even
2058 : though (a) this_loop_vinfo might not actually be better than
2059 : other_loop_vinfo for that VF and (b) it would be significantly
2060 : worse at larger VFs.
2061 :
2062 : Here we go for a hacky compromise: pick this_loop_vinfo if it is
2063 : no more expensive than other_loop_vinfo even after doubling the
2064 : estimated other_loop_vinfo VF. For all but trivial loops, this
2065 : ensures that we only pick this_loop_vinfo if it is significantly
2066 : better than other_loop_vinfo at the estimated VF. */
2067 : if (est_rel_other_min != est_rel_this_min
2068 : || est_rel_other_max != est_rel_this_max)
2069 : {
2070 : HOST_WIDE_INT est_rel_this_likely
2071 : = estimated_poly_value (rel_this, POLY_VALUE_LIKELY);
2072 : HOST_WIDE_INT est_rel_other_likely
2073 : = estimated_poly_value (rel_other, POLY_VALUE_LIKELY);
2074 :
2075 : return est_rel_this_likely * 2 <= est_rel_other_likely ? -1 : 1;
2076 : }
2077 :
2078 : return 0;
2079 : }
2080 :
2081 : /* A <=>-style subroutine of better_main_loop_than_p, used when there is
2082 : nothing to choose between the inside (loop body) costs of THIS and OTHER.
2083 : Check whether we can determine the return value of better_main_loop_than_p
2084 : by comparing the outside (prologue and epilogue) costs of THIS and OTHER.
2085 : Return:
2086 :
2087 : * -1 if better_main_loop_than_p should return true.
2088 : * 1 if better_main_loop_than_p should return false.
2089 : * 0 if we can't decide. */
2090 :
2091 : int
2092 0 : vector_costs::compare_outside_loop_cost (const vector_costs *other) const
2093 : {
2094 0 : auto this_outside_cost = this->outside_cost ();
2095 0 : auto other_outside_cost = other->outside_cost ();
2096 0 : if (this_outside_cost != other_outside_cost)
2097 0 : return this_outside_cost < other_outside_cost ? -1 : 1;
2098 :
2099 : return 0;
2100 : }
|
