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 1371038 : auto_purge_vect_location::~auto_purge_vect_location ()
96 : {
97 1371038 : vect_location = dump_user_location_t ();
98 1371038 : }
99 :
100 : /* Dump a cost entry according to args to F. */
101 :
102 : void
103 219989 : 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 219989 : if (stmt_info)
109 : {
110 202587 : print_gimple_expr (f, STMT_VINFO_STMT (stmt_info), 0, TDF_SLIM);
111 202587 : fprintf (f, " ");
112 : }
113 17402 : else if (node)
114 3493 : fprintf (f, "node %p ", (void *)node);
115 : else
116 13909 : fprintf (f, "<unknown> ");
117 219989 : fprintf (f, "%d times ", count);
118 219989 : const char *ks = "unknown";
119 219989 : switch (kind)
120 : {
121 49593 : case scalar_stmt:
122 49593 : ks = "scalar_stmt";
123 49593 : break;
124 37308 : case scalar_load:
125 37308 : ks = "scalar_load";
126 37308 : break;
127 28119 : case scalar_store:
128 28119 : ks = "scalar_store";
129 28119 : break;
130 34994 : case vector_stmt:
131 34994 : ks = "vector_stmt";
132 34994 : break;
133 22101 : case vector_load:
134 22101 : ks = "vector_load";
135 22101 : break;
136 0 : case vector_gather_load:
137 0 : ks = "vector_gather_load";
138 0 : break;
139 7876 : case unaligned_load:
140 7876 : ks = "unaligned_load";
141 7876 : break;
142 4666 : case unaligned_store:
143 4666 : ks = "unaligned_store";
144 4666 : break;
145 8891 : case vector_store:
146 8891 : ks = "vector_store";
147 8891 : break;
148 0 : case vector_scatter_store:
149 0 : ks = "vector_scatter_store";
150 0 : break;
151 2265 : case vec_to_scalar:
152 2265 : ks = "vec_to_scalar";
153 2265 : break;
154 9433 : case scalar_to_vec:
155 9433 : ks = "scalar_to_vec";
156 9433 : break;
157 12 : case cond_branch_not_taken:
158 12 : ks = "cond_branch_not_taken";
159 12 : break;
160 544 : case cond_branch_taken:
161 544 : ks = "cond_branch_taken";
162 544 : break;
163 6518 : case vec_perm:
164 6518 : ks = "vec_perm";
165 6518 : break;
166 5418 : case vec_promote_demote:
167 5418 : ks = "vec_promote_demote";
168 5418 : break;
169 1129 : case vec_construct:
170 1129 : ks = "vec_construct";
171 1129 : break;
172 1122 : case vec_deconstruct:
173 1122 : ks = "vec_deconstruct";
174 1122 : break;
175 : }
176 219989 : fprintf (f, "%s ", ks);
177 219989 : if (kind == unaligned_load || kind == unaligned_store)
178 12542 : fprintf (f, "(misalign %d) ", misalign);
179 219989 : fprintf (f, "costs %u ", cost);
180 219989 : const char *ws = "unknown";
181 219989 : switch (where)
182 : {
183 23306 : case vect_prologue:
184 23306 : ws = "prologue";
185 23306 : break;
186 189229 : case vect_body:
187 189229 : ws = "body";
188 189229 : break;
189 7454 : case vect_epilogue:
190 7454 : ws = "epilogue";
191 7454 : break;
192 : }
193 219989 : fprintf (f, "in %s\n", ws);
194 219989 : }
195 :
196 : /* For mapping simduid to vectorization factor. */
197 :
198 : class simduid_to_vf : public free_ptr_hash<simduid_to_vf>
199 : {
200 : public:
201 : unsigned int simduid;
202 : poly_uint64 vf;
203 :
204 : /* hash_table support. */
205 : static inline hashval_t hash (const simduid_to_vf *);
206 : static inline int equal (const simduid_to_vf *, const simduid_to_vf *);
207 : };
208 :
209 : inline hashval_t
210 7664 : simduid_to_vf::hash (const simduid_to_vf *p)
211 : {
212 7664 : return p->simduid;
213 : }
214 :
215 : inline int
216 14168 : simduid_to_vf::equal (const simduid_to_vf *p1, const simduid_to_vf *p2)
217 : {
218 14168 : return p1->simduid == p2->simduid;
219 : }
220 :
221 : /* This hash maps the OMP simd array to the corresponding simduid used
222 : to index into it. Like thus,
223 :
224 : _7 = GOMP_SIMD_LANE (simduid.0)
225 : ...
226 : ...
227 : D.1737[_7] = stuff;
228 :
229 :
230 : This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
231 : simduid.0. */
232 :
233 : struct simd_array_to_simduid : free_ptr_hash<simd_array_to_simduid>
234 : {
235 : tree decl;
236 : unsigned int simduid;
237 :
238 : /* hash_table support. */
239 : static inline hashval_t hash (const simd_array_to_simduid *);
240 : static inline int equal (const simd_array_to_simduid *,
241 : const simd_array_to_simduid *);
242 : };
243 :
244 : inline hashval_t
245 23644 : simd_array_to_simduid::hash (const simd_array_to_simduid *p)
246 : {
247 23644 : return DECL_UID (p->decl);
248 : }
249 :
250 : inline int
251 16566 : simd_array_to_simduid::equal (const simd_array_to_simduid *p1,
252 : const simd_array_to_simduid *p2)
253 : {
254 16566 : return p1->decl == p2->decl;
255 : }
256 :
257 : /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LAST_LANE,
258 : into their corresponding constants and remove
259 : IFN_GOMP_SIMD_ORDERED_{START,END}. */
260 :
261 : static void
262 7809 : adjust_simduid_builtins (hash_table<simduid_to_vf> *htab, function *fun)
263 : {
264 7809 : basic_block bb;
265 :
266 118569 : FOR_EACH_BB_FN (bb, fun)
267 : {
268 110760 : gimple_stmt_iterator i;
269 :
270 738824 : for (i = gsi_start_bb (bb); !gsi_end_p (i); )
271 : {
272 517304 : poly_uint64 vf = 1;
273 517304 : enum internal_fn ifn;
274 517304 : gimple *stmt = gsi_stmt (i);
275 517304 : tree t;
276 517304 : if (!is_gimple_call (stmt)
277 517304 : || !gimple_call_internal_p (stmt))
278 : {
279 509741 : gsi_next (&i);
280 510657 : continue;
281 : }
282 7563 : ifn = gimple_call_internal_fn (stmt);
283 7563 : switch (ifn)
284 : {
285 6647 : case IFN_GOMP_SIMD_LANE:
286 6647 : case IFN_GOMP_SIMD_VF:
287 6647 : case IFN_GOMP_SIMD_LAST_LANE:
288 6647 : break;
289 158 : case IFN_GOMP_SIMD_ORDERED_START:
290 158 : case IFN_GOMP_SIMD_ORDERED_END:
291 158 : if (integer_onep (gimple_call_arg (stmt, 0)))
292 : {
293 5 : enum built_in_function bcode
294 : = (ifn == IFN_GOMP_SIMD_ORDERED_START
295 10 : ? BUILT_IN_GOMP_ORDERED_START
296 : : BUILT_IN_GOMP_ORDERED_END);
297 10 : gimple *g
298 10 : = gimple_build_call (builtin_decl_explicit (bcode), 0);
299 10 : gimple_move_vops (g, stmt);
300 10 : gsi_replace (&i, g, true);
301 10 : continue;
302 10 : }
303 148 : gsi_remove (&i, true);
304 148 : unlink_stmt_vdef (stmt);
305 148 : continue;
306 758 : default:
307 758 : gsi_next (&i);
308 758 : continue;
309 906 : }
310 6647 : tree arg = gimple_call_arg (stmt, 0);
311 6647 : gcc_assert (arg != NULL_TREE);
312 6647 : gcc_assert (TREE_CODE (arg) == SSA_NAME);
313 6647 : simduid_to_vf *p = NULL, data;
314 6647 : data.simduid = DECL_UID (SSA_NAME_VAR (arg));
315 : /* Need to nullify loop safelen field since it's value is not
316 : valid after transformation. */
317 6647 : if (bb->loop_father && bb->loop_father->safelen > 0)
318 2169 : bb->loop_father->safelen = 0;
319 6647 : if (htab)
320 : {
321 4742 : p = htab->find (&data);
322 4742 : if (p)
323 4701 : vf = p->vf;
324 : }
325 6647 : switch (ifn)
326 : {
327 969 : case IFN_GOMP_SIMD_VF:
328 969 : t = build_int_cst (unsigned_type_node, vf);
329 969 : break;
330 3484 : case IFN_GOMP_SIMD_LANE:
331 3484 : t = build_int_cst (unsigned_type_node, 0);
332 3484 : break;
333 2194 : case IFN_GOMP_SIMD_LAST_LANE:
334 2194 : t = gimple_call_arg (stmt, 1);
335 2194 : break;
336 : default:
337 : gcc_unreachable ();
338 : }
339 6647 : tree lhs = gimple_call_lhs (stmt);
340 6647 : if (lhs)
341 6590 : replace_uses_by (lhs, t);
342 6647 : release_defs (stmt);
343 6647 : gsi_remove (&i, true);
344 : }
345 : }
346 7809 : }
347 :
348 : /* Helper structure for note_simd_array_uses. */
349 :
350 : struct note_simd_array_uses_struct
351 : {
352 : hash_table<simd_array_to_simduid> **htab;
353 : unsigned int simduid;
354 : };
355 :
356 : /* Callback for note_simd_array_uses, called through walk_gimple_op. */
357 :
358 : static tree
359 65131 : note_simd_array_uses_cb (tree *tp, int *walk_subtrees, void *data)
360 : {
361 65131 : struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
362 65131 : struct note_simd_array_uses_struct *ns
363 : = (struct note_simd_array_uses_struct *) wi->info;
364 :
365 65131 : if (TYPE_P (*tp))
366 0 : *walk_subtrees = 0;
367 65131 : else if (VAR_P (*tp)
368 12360 : && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp))
369 77491 : && DECL_CONTEXT (*tp) == current_function_decl)
370 : {
371 12360 : simd_array_to_simduid data;
372 12360 : if (!*ns->htab)
373 2208 : *ns->htab = new hash_table<simd_array_to_simduid> (15);
374 12360 : data.decl = *tp;
375 12360 : data.simduid = ns->simduid;
376 12360 : simd_array_to_simduid **slot = (*ns->htab)->find_slot (&data, INSERT);
377 12360 : if (*slot == NULL)
378 : {
379 5501 : simd_array_to_simduid *p = XNEW (simd_array_to_simduid);
380 5501 : *p = data;
381 5501 : *slot = p;
382 : }
383 6859 : else if ((*slot)->simduid != ns->simduid)
384 0 : (*slot)->simduid = -1U;
385 12360 : *walk_subtrees = 0;
386 : }
387 65131 : return NULL_TREE;
388 : }
389 :
390 : /* Find "omp simd array" temporaries and map them to corresponding
391 : simduid. */
392 :
393 : static void
394 7809 : note_simd_array_uses (hash_table<simd_array_to_simduid> **htab, function *fun)
395 : {
396 7809 : basic_block bb;
397 7809 : gimple_stmt_iterator gsi;
398 7809 : struct walk_stmt_info wi;
399 7809 : struct note_simd_array_uses_struct ns;
400 :
401 7809 : memset (&wi, 0, sizeof (wi));
402 7809 : wi.info = &ns;
403 7809 : ns.htab = htab;
404 :
405 102453 : FOR_EACH_BB_FN (bb, fun)
406 583018 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
407 : {
408 393730 : gimple *stmt = gsi_stmt (gsi);
409 393730 : if (!is_gimple_call (stmt) || !gimple_call_internal_p (stmt))
410 387146 : continue;
411 7509 : switch (gimple_call_internal_fn (stmt))
412 : {
413 6620 : case IFN_GOMP_SIMD_LANE:
414 6620 : case IFN_GOMP_SIMD_VF:
415 6620 : case IFN_GOMP_SIMD_LAST_LANE:
416 6620 : break;
417 889 : default:
418 889 : continue;
419 : }
420 6620 : tree lhs = gimple_call_lhs (stmt);
421 6620 : if (lhs == NULL_TREE)
422 36 : continue;
423 6584 : imm_use_iterator use_iter;
424 6584 : gimple *use_stmt;
425 6584 : ns.simduid = DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt, 0)));
426 31198 : FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, lhs)
427 18030 : if (!is_gimple_debug (use_stmt))
428 24522 : walk_gimple_op (use_stmt, note_simd_array_uses_cb, &wi);
429 : }
430 7809 : }
431 :
432 : /* Shrink arrays with "omp simd array" attribute to the corresponding
433 : vectorization factor. */
434 :
435 : static void
436 2208 : shrink_simd_arrays
437 : (hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab,
438 : hash_table<simduid_to_vf> *simduid_to_vf_htab)
439 : {
440 7709 : for (hash_table<simd_array_to_simduid>::iterator iter
441 2208 : = simd_array_to_simduid_htab->begin ();
442 13210 : iter != simd_array_to_simduid_htab->end (); ++iter)
443 5501 : if ((*iter)->simduid != -1U)
444 : {
445 5501 : tree decl = (*iter)->decl;
446 5501 : poly_uint64 vf = 1;
447 5501 : if (simduid_to_vf_htab)
448 : {
449 4574 : simduid_to_vf *p = NULL, data;
450 4574 : data.simduid = (*iter)->simduid;
451 4574 : p = simduid_to_vf_htab->find (&data);
452 4574 : if (p)
453 4540 : vf = p->vf;
454 : }
455 5501 : tree atype
456 5501 : = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl)), vf);
457 5501 : TREE_TYPE (decl) = atype;
458 5501 : relayout_decl (decl);
459 : }
460 :
461 2208 : delete simd_array_to_simduid_htab;
462 2208 : }
463 :
464 : /* Initialize the vec_info with kind KIND_IN and target cost data
465 : TARGET_COST_DATA_IN. */
466 :
467 2759047 : vec_info::vec_info (vec_info::vec_kind kind_in, vec_info_shared *shared_)
468 2759047 : : kind (kind_in),
469 2759047 : shared (shared_),
470 2759047 : stmt_vec_info_ro (false),
471 2759047 : bbs (NULL),
472 2759047 : nbbs (0),
473 2759047 : inv_pattern_def_seq (NULL)
474 : {
475 2759047 : stmt_vec_infos.create (50);
476 2759047 : }
477 :
478 2759047 : vec_info::~vec_info ()
479 : {
480 4969430 : for (slp_instance &instance : slp_instances)
481 1120361 : vect_free_slp_instance (instance);
482 :
483 2759047 : free_stmt_vec_infos ();
484 2759047 : }
485 :
486 2328073 : vec_info_shared::vec_info_shared ()
487 2328073 : : datarefs (vNULL),
488 2328073 : datarefs_copy (vNULL),
489 2328073 : ddrs (vNULL)
490 : {
491 2328073 : }
492 :
493 2328073 : vec_info_shared::~vec_info_shared ()
494 : {
495 2328073 : free_data_refs (datarefs);
496 2328073 : free_dependence_relations (ddrs);
497 2328073 : datarefs_copy.release ();
498 2328073 : }
499 :
500 : void
501 2099616 : vec_info_shared::save_datarefs ()
502 : {
503 2099616 : if (!flag_checking)
504 : return;
505 3150377 : datarefs_copy.reserve_exact (datarefs.length ());
506 12712913 : for (unsigned i = 0; i < datarefs.length (); ++i)
507 10613314 : datarefs_copy.quick_push (*datarefs[i]);
508 : }
509 :
510 : void
511 886419 : vec_info_shared::check_datarefs ()
512 : {
513 886419 : if (!flag_checking)
514 : return;
515 2656111 : gcc_assert (datarefs.length () == datarefs_copy.length ());
516 12235355 : for (unsigned i = 0; i < datarefs.length (); ++i)
517 11348936 : if (memcmp (&datarefs_copy[i], datarefs[i],
518 : offsetof (data_reference, alt_indices)) != 0)
519 0 : gcc_unreachable ();
520 : }
521 :
522 : /* Record that STMT belongs to the vectorizable region. Create and return
523 : an associated stmt_vec_info. */
524 :
525 : stmt_vec_info
526 61746381 : vec_info::add_stmt (gimple *stmt)
527 : {
528 61746381 : stmt_vec_info res = new_stmt_vec_info (stmt);
529 61746381 : set_vinfo_for_stmt (stmt, res);
530 61746381 : return res;
531 : }
532 :
533 : /* Record that STMT belongs to the vectorizable region. Create a new
534 : stmt_vec_info and mark VECINFO as being related and return the new
535 : stmt_vec_info. */
536 :
537 : stmt_vec_info
538 1107 : vec_info::add_pattern_stmt (gimple *stmt, stmt_vec_info stmt_info)
539 : {
540 1107 : stmt_vec_info res = new_stmt_vec_info (stmt);
541 1107 : res->pattern_stmt_p = true;
542 1107 : set_vinfo_for_stmt (stmt, res, false);
543 1107 : STMT_VINFO_RELATED_STMT (res) = stmt_info;
544 1107 : return res;
545 : }
546 :
547 : /* If STMT was previously associated with a stmt_vec_info and STMT now resides
548 : at a different address than before (e.g., because STMT is a phi node that has
549 : been resized), update the stored address to match the new one. It is not
550 : possible to use lookup_stmt () to perform this task, because that function
551 : returns NULL if the stored stmt pointer does not match the one being looked
552 : up. */
553 :
554 : stmt_vec_info
555 10865 : vec_info::resync_stmt_addr (gimple *stmt)
556 : {
557 10865 : unsigned int uid = gimple_uid (stmt);
558 10865 : if (uid > 0 && uid - 1 < stmt_vec_infos.length ())
559 : {
560 10865 : stmt_vec_info res = stmt_vec_infos[uid - 1];
561 10865 : if (res && res->stmt)
562 : {
563 10865 : res->stmt = stmt;
564 10865 : return res;
565 : }
566 : }
567 : return nullptr;
568 : }
569 :
570 : /* If STMT has an associated stmt_vec_info, return that vec_info, otherwise
571 : return null. It is safe to call this function on any statement, even if
572 : it might not be part of the vectorizable region. */
573 :
574 : stmt_vec_info
575 479889027 : vec_info::lookup_stmt (gimple *stmt)
576 : {
577 479889027 : unsigned int uid = gimple_uid (stmt);
578 479889027 : if (uid > 0 && uid - 1 < stmt_vec_infos.length ())
579 : {
580 302633279 : stmt_vec_info res = stmt_vec_infos[uid - 1];
581 302633279 : if (res && res->stmt == stmt)
582 302275197 : return res;
583 : }
584 : return NULL;
585 : }
586 :
587 : /* If NAME is an SSA_NAME and its definition has an associated stmt_vec_info,
588 : return that stmt_vec_info, otherwise return null. It is safe to call
589 : this on arbitrary operands. */
590 :
591 : stmt_vec_info
592 56559070 : vec_info::lookup_def (tree name)
593 : {
594 56559070 : if (TREE_CODE (name) == SSA_NAME
595 56559070 : && !SSA_NAME_IS_DEFAULT_DEF (name))
596 51306109 : return lookup_stmt (SSA_NAME_DEF_STMT (name));
597 : return NULL;
598 : }
599 :
600 : /* See whether there is a single non-debug statement that uses LHS and
601 : whether that statement has an associated stmt_vec_info. Return the
602 : stmt_vec_info if so, otherwise return null. */
603 :
604 : stmt_vec_info
605 1908 : vec_info::lookup_single_use (tree lhs)
606 : {
607 1908 : use_operand_p dummy;
608 1908 : gimple *use_stmt;
609 1908 : if (single_imm_use (lhs, &dummy, &use_stmt))
610 1759 : return lookup_stmt (use_stmt);
611 : return NULL;
612 : }
613 :
614 : /* Return vectorization information about DR. */
615 :
616 : dr_vec_info *
617 49911251 : vec_info::lookup_dr (data_reference *dr)
618 : {
619 49911251 : stmt_vec_info stmt_info = lookup_stmt (DR_STMT (dr));
620 : /* DR_STMT should never refer to a stmt in a pattern replacement. */
621 49911251 : gcc_checking_assert (!is_pattern_stmt_p (stmt_info));
622 49911251 : return STMT_VINFO_DR_INFO (stmt_info->dr_aux.stmt);
623 : }
624 :
625 : /* Record that NEW_STMT_INFO now implements the same data reference
626 : as OLD_STMT_INFO. */
627 :
628 : void
629 6410 : vec_info::move_dr (stmt_vec_info new_stmt_info, stmt_vec_info old_stmt_info)
630 : {
631 6410 : gcc_assert (!is_pattern_stmt_p (old_stmt_info));
632 6410 : STMT_VINFO_DR_INFO (old_stmt_info)->stmt = new_stmt_info;
633 6410 : new_stmt_info->dr_aux = old_stmt_info->dr_aux;
634 6410 : STMT_VINFO_DR_WRT_VEC_LOOP (new_stmt_info)
635 6410 : = STMT_VINFO_DR_WRT_VEC_LOOP (old_stmt_info);
636 6410 : STMT_VINFO_GATHER_SCATTER_P (new_stmt_info)
637 6410 : = STMT_VINFO_GATHER_SCATTER_P (old_stmt_info);
638 6410 : STMT_VINFO_STRIDED_P (new_stmt_info)
639 6410 : = STMT_VINFO_STRIDED_P (old_stmt_info);
640 6410 : STMT_VINFO_SIMD_LANE_ACCESS_P (new_stmt_info)
641 6410 : = STMT_VINFO_SIMD_LANE_ACCESS_P (old_stmt_info);
642 6410 : }
643 :
644 : /* Permanently remove the statement described by STMT_INFO from the
645 : function. */
646 :
647 : void
648 1495614 : vec_info::remove_stmt (stmt_vec_info stmt_info)
649 : {
650 1495614 : gcc_assert (!stmt_info->pattern_stmt_p);
651 1495614 : set_vinfo_for_stmt (stmt_info->stmt, NULL);
652 1495614 : unlink_stmt_vdef (stmt_info->stmt);
653 1495614 : gimple_stmt_iterator si = gsi_for_stmt (stmt_info->stmt);
654 1495614 : gsi_remove (&si, true);
655 1495614 : release_defs (stmt_info->stmt);
656 1495614 : free_stmt_vec_info (stmt_info);
657 1495614 : }
658 :
659 : /* Replace the statement at GSI by NEW_STMT, both the vectorization
660 : information and the function itself. STMT_INFO describes the statement
661 : at GSI. */
662 :
663 : void
664 5193 : vec_info::replace_stmt (gimple_stmt_iterator *gsi, stmt_vec_info stmt_info,
665 : gimple *new_stmt)
666 : {
667 5193 : gimple *old_stmt = stmt_info->stmt;
668 5193 : gcc_assert (!stmt_info->pattern_stmt_p && old_stmt == gsi_stmt (*gsi));
669 5193 : gimple_set_uid (new_stmt, gimple_uid (old_stmt));
670 5193 : stmt_info->stmt = new_stmt;
671 5193 : gsi_replace (gsi, new_stmt, true);
672 5193 : }
673 :
674 : /* Insert stmts in SEQ on the VEC_INFO region entry. If CONTEXT is
675 : not NULL it specifies whether to use the sub-region entry
676 : determined by it, currently used for loop vectorization to insert
677 : on the inner loop entry vs. the outer loop entry. */
678 :
679 : void
680 101860 : vec_info::insert_seq_on_entry (stmt_vec_info context, gimple_seq seq)
681 : {
682 101860 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (this))
683 : {
684 19163 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
685 19163 : basic_block new_bb;
686 19163 : edge pe;
687 :
688 19163 : if (context && nested_in_vect_loop_p (loop, context))
689 : loop = loop->inner;
690 :
691 19163 : pe = loop_preheader_edge (loop);
692 19163 : new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
693 19163 : gcc_assert (!new_bb);
694 : }
695 : else
696 : {
697 82697 : gimple_stmt_iterator gsi_region_begin
698 82697 : = gsi_after_labels (bbs[0]);
699 82697 : gsi_insert_seq_before (&gsi_region_begin, seq, GSI_SAME_STMT);
700 : }
701 101860 : }
702 :
703 : /* Like insert_seq_on_entry but just inserts the single stmt NEW_STMT. */
704 :
705 : void
706 3299 : vec_info::insert_on_entry (stmt_vec_info context, gimple *new_stmt)
707 : {
708 3299 : gimple_seq seq = NULL;
709 3299 : gimple_stmt_iterator gsi = gsi_start (seq);
710 3299 : gsi_insert_before_without_update (&gsi, new_stmt, GSI_SAME_STMT);
711 3299 : insert_seq_on_entry (context, seq);
712 3299 : }
713 :
714 : /* Create and initialize a new stmt_vec_info struct for STMT. */
715 :
716 : stmt_vec_info
717 61747488 : vec_info::new_stmt_vec_info (gimple *stmt)
718 : {
719 61747488 : stmt_vec_info res = XCNEW (class _stmt_vec_info);
720 61747488 : res->stmt = stmt;
721 :
722 61747488 : STMT_VINFO_RELEVANT (res) = vect_unused_in_scope;
723 61747488 : STMT_VINFO_VECTORIZABLE (res) = true;
724 61747488 : STMT_VINFO_REDUC_TYPE (res) = TREE_CODE_REDUCTION;
725 61747488 : STMT_VINFO_REDUC_CODE (res) = ERROR_MARK;
726 61747488 : STMT_VINFO_REDUC_IDX (res) = -1;
727 61747488 : STMT_VINFO_REDUC_DEF (res) = NULL;
728 61747488 : STMT_VINFO_SLP_VECT_ONLY (res) = false;
729 :
730 61747488 : if (is_a <loop_vec_info> (this)
731 8597686 : && gimple_code (stmt) == GIMPLE_PHI
732 63271136 : && is_loop_header_bb_p (gimple_bb (stmt)))
733 1512904 : STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
734 : else
735 60234584 : STMT_VINFO_DEF_TYPE (res) = vect_internal_def;
736 :
737 61747488 : STMT_SLP_TYPE (res) = not_vect;
738 :
739 : /* This is really "uninitialized" until vect_compute_data_ref_alignment. */
740 61747488 : res->dr_aux.misalignment = DR_MISALIGNMENT_UNINITIALIZED;
741 :
742 61747488 : return res;
743 : }
744 :
745 : /* Associate STMT with INFO. */
746 :
747 : void
748 63243102 : vec_info::set_vinfo_for_stmt (gimple *stmt, stmt_vec_info info, bool check_ro)
749 : {
750 63243102 : unsigned int uid = gimple_uid (stmt);
751 63243102 : if (uid == 0)
752 : {
753 61747488 : gcc_assert (!check_ro || !stmt_vec_info_ro);
754 61747488 : gcc_checking_assert (info);
755 61747488 : uid = stmt_vec_infos.length () + 1;
756 61747488 : gimple_set_uid (stmt, uid);
757 61747488 : stmt_vec_infos.safe_push (info);
758 : }
759 : else
760 : {
761 1495614 : gcc_checking_assert (info == NULL);
762 1495614 : stmt_vec_infos[uid - 1] = info;
763 : }
764 63243102 : }
765 :
766 : /* Free the contents of stmt_vec_infos. */
767 :
768 : void
769 2759047 : vec_info::free_stmt_vec_infos (void)
770 : {
771 70024629 : for (stmt_vec_info &info : stmt_vec_infos)
772 61747488 : if (info != NULL)
773 60251874 : free_stmt_vec_info (info);
774 2759047 : stmt_vec_infos.release ();
775 2759047 : }
776 :
777 : /* Free STMT_INFO. */
778 :
779 : void
780 61747488 : vec_info::free_stmt_vec_info (stmt_vec_info stmt_info)
781 : {
782 61747488 : if (stmt_info->pattern_stmt_p)
783 : {
784 2428948 : gimple_set_bb (stmt_info->stmt, NULL);
785 2428948 : tree lhs = gimple_get_lhs (stmt_info->stmt);
786 2428948 : if (lhs && TREE_CODE (lhs) == SSA_NAME)
787 2060688 : release_ssa_name (lhs);
788 : }
789 :
790 61747488 : free (stmt_info);
791 61747488 : }
792 :
793 : /* Returns true if S1 dominates S2. */
794 :
795 : bool
796 553096 : vect_stmt_dominates_stmt_p (gimple *s1, gimple *s2)
797 : {
798 553096 : basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
799 :
800 : /* If bb1 is NULL, it should be a GIMPLE_NOP def stmt of an (D)
801 : SSA_NAME. Assume it lives at the beginning of function and
802 : thus dominates everything. */
803 553096 : if (!bb1 || s1 == s2)
804 : return true;
805 :
806 : /* If bb2 is NULL, it doesn't dominate any stmt with a bb. */
807 551062 : if (!bb2)
808 : return false;
809 :
810 551062 : if (bb1 != bb2)
811 189469 : return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
812 :
813 : /* PHIs in the same basic block are assumed to be
814 : executed all in parallel, if only one stmt is a PHI,
815 : it dominates the other stmt in the same basic block. */
816 361593 : if (gimple_code (s1) == GIMPLE_PHI)
817 : return true;
818 :
819 325136 : if (gimple_code (s2) == GIMPLE_PHI)
820 : return false;
821 :
822 : /* Inserted vectorized stmts all have UID 0 while the original stmts
823 : in the IL have UID increasing within a BB. Walk from both sides
824 : until we find the other stmt or a stmt with UID != 0. */
825 307487 : gimple_stmt_iterator gsi1 = gsi_for_stmt (s1);
826 777819 : while (gimple_uid (gsi_stmt (gsi1)) == 0)
827 : {
828 587773 : gsi_next (&gsi1);
829 587773 : if (gsi_end_p (gsi1))
830 : return false;
831 586059 : if (gsi_stmt (gsi1) == s2)
832 : return true;
833 : }
834 190046 : if (gimple_uid (gsi_stmt (gsi1)) == -1u)
835 : return false;
836 :
837 190046 : gimple_stmt_iterator gsi2 = gsi_for_stmt (s2);
838 797566 : while (gimple_uid (gsi_stmt (gsi2)) == 0)
839 : {
840 619683 : gsi_prev (&gsi2);
841 619683 : if (gsi_end_p (gsi2))
842 : return false;
843 607528 : if (gsi_stmt (gsi2) == s1)
844 : return true;
845 : }
846 177883 : if (gimple_uid (gsi_stmt (gsi2)) == -1u)
847 : return false;
848 :
849 177883 : if (gimple_uid (gsi_stmt (gsi1)) <= gimple_uid (gsi_stmt (gsi2)))
850 : return true;
851 : return false;
852 : }
853 :
854 : /* A helper function to free scev and LOOP niter information, as well as
855 : clear loop constraint LOOP_C_FINITE. */
856 :
857 : void
858 44151 : vect_free_loop_info_assumptions (class loop *loop)
859 : {
860 44151 : scev_reset_htab ();
861 : /* We need to explicitly reset upper bound information since they are
862 : used even after free_numbers_of_iterations_estimates. */
863 44151 : loop->any_upper_bound = false;
864 44151 : loop->any_likely_upper_bound = false;
865 44151 : free_numbers_of_iterations_estimates (loop);
866 44151 : loop_constraint_clear (loop, LOOP_C_FINITE);
867 44151 : }
868 :
869 : /* If LOOP has been versioned during ifcvt, return the internal call
870 : guarding it. */
871 :
872 : gimple *
873 518441 : vect_loop_vectorized_call (class loop *loop, gcond **cond)
874 : {
875 518441 : basic_block bb = loop_preheader_edge (loop)->src;
876 957894 : gimple *g;
877 1397347 : do
878 : {
879 957894 : g = *gsi_last_bb (bb);
880 608080 : if ((g && gimple_code (g) == GIMPLE_COND)
881 2055052 : || !single_succ_p (bb))
882 : break;
883 578717 : if (!single_pred_p (bb))
884 : break;
885 439453 : bb = single_pred (bb);
886 : }
887 : while (1);
888 518441 : if (g && gimple_code (g) == GIMPLE_COND)
889 : {
890 372959 : if (cond)
891 0 : *cond = as_a <gcond *> (g);
892 372959 : gimple_stmt_iterator gsi = gsi_for_stmt (g);
893 372959 : gsi_prev (&gsi);
894 372959 : if (!gsi_end_p (gsi))
895 : {
896 341537 : g = gsi_stmt (gsi);
897 341537 : if (gimple_call_internal_p (g, IFN_LOOP_VECTORIZED)
898 341537 : && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->num
899 30169 : || tree_to_shwi (gimple_call_arg (g, 1)) == loop->num))
900 60621 : return g;
901 : }
902 : }
903 : return NULL;
904 : }
905 :
906 : /* If LOOP has been versioned during loop distribution, return the gurading
907 : internal call. */
908 :
909 : static gimple *
910 481702 : vect_loop_dist_alias_call (class loop *loop, function *fun)
911 : {
912 481702 : basic_block bb;
913 481702 : basic_block entry;
914 481702 : class loop *outer, *orig;
915 :
916 481702 : if (loop->orig_loop_num == 0)
917 : return NULL;
918 :
919 168 : orig = get_loop (fun, loop->orig_loop_num);
920 168 : if (orig == NULL)
921 : {
922 : /* The original loop is somehow destroyed. Clear the information. */
923 0 : loop->orig_loop_num = 0;
924 0 : return NULL;
925 : }
926 :
927 168 : if (loop != orig)
928 95 : bb = nearest_common_dominator (CDI_DOMINATORS, loop->header, orig->header);
929 : else
930 73 : bb = loop_preheader_edge (loop)->src;
931 :
932 168 : outer = bb->loop_father;
933 168 : entry = ENTRY_BLOCK_PTR_FOR_FN (fun);
934 :
935 : /* Look upward in dominance tree. */
936 751 : for (; bb != entry && flow_bb_inside_loop_p (outer, bb);
937 583 : bb = get_immediate_dominator (CDI_DOMINATORS, bb))
938 : {
939 699 : gimple_stmt_iterator gsi = gsi_last_bb (bb);
940 699 : if (!safe_is_a <gcond *> (*gsi))
941 583 : continue;
942 :
943 514 : gsi_prev (&gsi);
944 514 : if (gsi_end_p (gsi))
945 8 : continue;
946 :
947 506 : gimple *g = gsi_stmt (gsi);
948 : /* The guarding internal function call must have the same distribution
949 : alias id. */
950 506 : if (gimple_call_internal_p (g, IFN_LOOP_DIST_ALIAS)
951 506 : && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->orig_loop_num))
952 481702 : return g;
953 : }
954 : return NULL;
955 : }
956 :
957 : /* Set the uids of all the statements in basic blocks inside loop
958 : represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
959 : call guarding the loop which has been if converted. */
960 : static void
961 7807 : set_uid_loop_bbs (loop_vec_info loop_vinfo, gimple *loop_vectorized_call,
962 : function *fun)
963 : {
964 7807 : tree arg = gimple_call_arg (loop_vectorized_call, 1);
965 7807 : basic_block *bbs;
966 7807 : unsigned int i;
967 7807 : class loop *scalar_loop = get_loop (fun, tree_to_shwi (arg));
968 :
969 7807 : LOOP_VINFO_SCALAR_LOOP (loop_vinfo) = scalar_loop;
970 7807 : LOOP_VINFO_SCALAR_MAIN_EXIT (loop_vinfo)
971 7807 : = vec_init_loop_exit_info (scalar_loop);
972 7807 : gcc_checking_assert (vect_loop_vectorized_call (scalar_loop)
973 : == loop_vectorized_call);
974 : /* If we are going to vectorize outer loop, prevent vectorization
975 : of the inner loop in the scalar loop - either the scalar loop is
976 : thrown away, so it is a wasted work, or is used only for
977 : a few iterations. */
978 7807 : if (scalar_loop->inner)
979 : {
980 120 : gimple *g = vect_loop_vectorized_call (scalar_loop->inner);
981 120 : if (g)
982 : {
983 120 : arg = gimple_call_arg (g, 0);
984 120 : get_loop (fun, tree_to_shwi (arg))->dont_vectorize = true;
985 120 : fold_loop_internal_call (g, boolean_false_node);
986 : }
987 : }
988 7807 : bbs = get_loop_body (scalar_loop);
989 42491 : for (i = 0; i < scalar_loop->num_nodes; i++)
990 : {
991 34684 : basic_block bb = bbs[i];
992 34684 : gimple_stmt_iterator gsi;
993 66841 : for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
994 : {
995 32157 : gimple *phi = gsi_stmt (gsi);
996 32157 : gimple_set_uid (phi, 0);
997 : }
998 175231 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
999 : {
1000 105863 : gimple *stmt = gsi_stmt (gsi);
1001 105863 : gimple_set_uid (stmt, 0);
1002 : }
1003 : }
1004 7807 : free (bbs);
1005 7807 : }
1006 :
1007 : /* Generate vectorized code for LOOP and its epilogues. */
1008 :
1009 : static unsigned
1010 61803 : vect_transform_loops (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1011 : loop_p loop, gimple *loop_vectorized_call,
1012 : function *fun)
1013 : {
1014 61803 : loop_vec_info loop_vinfo = loop_vec_info_for_loop (loop);
1015 :
1016 61803 : if (loop_vectorized_call)
1017 7807 : set_uid_loop_bbs (loop_vinfo, loop_vectorized_call, fun);
1018 :
1019 61803 : unsigned HOST_WIDE_INT bytes;
1020 61803 : if (dump_enabled_p ())
1021 : {
1022 22030 : if (GET_MODE_SIZE (loop_vinfo->vector_mode).is_constant (&bytes))
1023 11015 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
1024 : "%sloop vectorized using %s%wu byte vectors and"
1025 : " unroll factor %u\n",
1026 11015 : LOOP_VINFO_EPILOGUE_P (loop_vinfo)
1027 : ? "epilogue " : "",
1028 11015 : LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)
1029 : ? "masked " : "", bytes,
1030 : (unsigned int) LOOP_VINFO_VECT_FACTOR
1031 11015 : (loop_vinfo).to_constant ());
1032 : else
1033 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
1034 : "%sloop vectorized using variable length vectors\n",
1035 : LOOP_VINFO_EPILOGUE_P (loop_vinfo)
1036 : ? "epilogue " : "");
1037 : }
1038 :
1039 61803 : loop_p new_loop = vect_transform_loop (loop_vinfo,
1040 : loop_vectorized_call);
1041 : /* Now that the loop has been vectorized, allow it to be unrolled
1042 : etc. */
1043 61803 : loop->force_vectorize = false;
1044 :
1045 61803 : if (loop->simduid)
1046 : {
1047 1895 : simduid_to_vf *simduid_to_vf_data = XNEW (simduid_to_vf);
1048 1895 : if (!simduid_to_vf_htab)
1049 1535 : simduid_to_vf_htab = new hash_table<simduid_to_vf> (15);
1050 1895 : simduid_to_vf_data->simduid = DECL_UID (loop->simduid);
1051 1895 : simduid_to_vf_data->vf = loop_vinfo->vectorization_factor;
1052 1895 : *simduid_to_vf_htab->find_slot (simduid_to_vf_data, INSERT)
1053 1895 : = simduid_to_vf_data;
1054 : }
1055 :
1056 : /* We should not have to update virtual SSA form here but some
1057 : transforms involve creating new virtual definitions which makes
1058 : updating difficult.
1059 : We delay the actual update to the end of the pass but avoid
1060 : confusing ourselves by forcing need_ssa_update_p () to false. */
1061 61803 : unsigned todo = 0;
1062 61803 : if (need_ssa_update_p (cfun))
1063 : {
1064 119 : gcc_assert (loop_vinfo->any_known_not_updated_vssa);
1065 119 : fun->gimple_df->ssa_renaming_needed = false;
1066 119 : todo |= TODO_update_ssa_only_virtuals;
1067 : }
1068 61803 : gcc_assert (!need_ssa_update_p (cfun));
1069 :
1070 : /* Epilogue of vectorized loop must be vectorized too. */
1071 61803 : if (new_loop)
1072 6847 : todo |= vect_transform_loops (simduid_to_vf_htab, new_loop, NULL, fun);
1073 :
1074 61803 : return todo;
1075 : }
1076 :
1077 : /* Try to vectorize LOOP. */
1078 :
1079 : static unsigned
1080 464638 : try_vectorize_loop_1 (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1081 : unsigned *num_vectorized_loops, loop_p loop,
1082 : gimple *loop_vectorized_call,
1083 : gimple *loop_dist_alias_call,
1084 : function *fun)
1085 : {
1086 464638 : unsigned ret = 0;
1087 464638 : vec_info_shared shared;
1088 464638 : auto_purge_vect_location sentinel;
1089 464638 : vect_location = find_loop_location (loop);
1090 :
1091 464638 : if (LOCATION_LOCUS (vect_location.get_location_t ()) != UNKNOWN_LOCATION
1092 464638 : && dump_enabled_p ())
1093 15028 : dump_printf (MSG_NOTE | MSG_PRIORITY_INTERNALS,
1094 : "\nAnalyzing loop at %s:%d\n",
1095 15028 : LOCATION_FILE (vect_location.get_location_t ()),
1096 30056 : LOCATION_LINE (vect_location.get_location_t ()));
1097 :
1098 : /* Try to analyze the loop, retaining an opt_problem if dump_enabled_p. */
1099 464638 : opt_loop_vec_info loop_vinfo = vect_analyze_loop (loop, loop_vectorized_call,
1100 : &shared);
1101 464638 : loop->aux = loop_vinfo;
1102 :
1103 464638 : if (!loop_vinfo)
1104 409678 : if (dump_enabled_p ())
1105 5782 : if (opt_problem *problem = loop_vinfo.get_problem ())
1106 : {
1107 5782 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1108 : "couldn't vectorize loop\n");
1109 5782 : problem->emit_and_clear ();
1110 : }
1111 :
1112 464638 : if (!loop_vinfo || !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo))
1113 : {
1114 : /* Free existing information if loop is analyzed with some
1115 : assumptions. */
1116 409678 : if (loop_constraint_set_p (loop, LOOP_C_FINITE))
1117 7990 : vect_free_loop_info_assumptions (loop);
1118 :
1119 : /* If we applied if-conversion then try to vectorize the
1120 : BB of innermost loops.
1121 : ??? Ideally BB vectorization would learn to vectorize
1122 : control flow by applying if-conversion on-the-fly, the
1123 : following retains the if-converted loop body even when
1124 : only non-if-converted parts took part in BB vectorization. */
1125 409678 : if (flag_tree_slp_vectorize != 0
1126 408660 : && loop_vectorized_call
1127 20981 : && ! loop->inner)
1128 : {
1129 20180 : basic_block bb = loop->header;
1130 20180 : bool require_loop_vectorize = false;
1131 40360 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
1132 735064 : !gsi_end_p (gsi); gsi_next (&gsi))
1133 : {
1134 715705 : gimple *stmt = gsi_stmt (gsi);
1135 715705 : gcall *call = dyn_cast <gcall *> (stmt);
1136 1243 : if (call && gimple_call_internal_p (call))
1137 : {
1138 1146 : internal_fn ifn = gimple_call_internal_fn (call);
1139 1146 : if (ifn == IFN_MASK_LOAD
1140 1146 : || ifn == IFN_MASK_STORE
1141 729 : || ifn == IFN_MASK_CALL
1142 : /* Don't keep the if-converted parts when the ifn with
1143 : specific type is not supported by the backend. */
1144 1856 : || (direct_internal_fn_p (ifn)
1145 385 : && !direct_internal_fn_supported_p
1146 385 : (call, OPTIMIZE_FOR_SPEED)))
1147 : {
1148 : require_loop_vectorize = true;
1149 : break;
1150 : }
1151 : }
1152 714884 : gimple_set_uid (stmt, -1);
1153 714884 : gimple_set_visited (stmt, false);
1154 : }
1155 20180 : if (!require_loop_vectorize)
1156 : {
1157 19359 : tree arg = gimple_call_arg (loop_vectorized_call, 1);
1158 19359 : class loop *scalar_loop = get_loop (fun, tree_to_shwi (arg));
1159 19359 : if (vect_slp_if_converted_bb (bb, scalar_loop))
1160 : {
1161 75 : fold_loop_internal_call (loop_vectorized_call,
1162 : boolean_true_node);
1163 75 : loop_vectorized_call = NULL;
1164 75 : ret |= TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1165 : }
1166 : }
1167 : }
1168 : /* If outer loop vectorization fails for LOOP_VECTORIZED guarded
1169 : loop, don't vectorize its inner loop; we'll attempt to
1170 : vectorize LOOP_VECTORIZED guarded inner loop of the scalar
1171 : loop version. */
1172 21999 : if (loop_vectorized_call && loop->inner)
1173 803 : loop->inner->dont_vectorize = true;
1174 409678 : return ret;
1175 : }
1176 :
1177 54960 : if (!dbg_cnt (vect_loop))
1178 : {
1179 : /* Free existing information if loop is analyzed with some
1180 : assumptions. */
1181 4 : if (loop_constraint_set_p (loop, LOOP_C_FINITE))
1182 0 : vect_free_loop_info_assumptions (loop);
1183 4 : return ret;
1184 : }
1185 :
1186 54956 : (*num_vectorized_loops)++;
1187 : /* Transform LOOP and its epilogues. */
1188 54956 : ret |= vect_transform_loops (simduid_to_vf_htab, loop,
1189 : loop_vectorized_call, fun);
1190 :
1191 54956 : if (loop_vectorized_call)
1192 : {
1193 7807 : fold_loop_internal_call (loop_vectorized_call, boolean_true_node);
1194 7807 : ret |= TODO_cleanup_cfg;
1195 : }
1196 54956 : if (loop_dist_alias_call)
1197 : {
1198 8 : tree value = gimple_call_arg (loop_dist_alias_call, 1);
1199 8 : fold_loop_internal_call (loop_dist_alias_call, value);
1200 8 : ret |= TODO_cleanup_cfg;
1201 : }
1202 :
1203 : return ret;
1204 464638 : }
1205 :
1206 : /* Try to vectorize LOOP. */
1207 :
1208 : static unsigned
1209 495807 : try_vectorize_loop (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1210 : unsigned *num_vectorized_loops, loop_p loop,
1211 : function *fun)
1212 : {
1213 495807 : if (!((flag_tree_loop_vectorize
1214 491235 : && optimize_loop_nest_for_speed_p (loop))
1215 32972 : || loop->force_vectorize))
1216 : return 0;
1217 :
1218 464638 : return try_vectorize_loop_1 (simduid_to_vf_htab, num_vectorized_loops, loop,
1219 : vect_loop_vectorized_call (loop),
1220 464638 : vect_loop_dist_alias_call (loop, fun), fun);
1221 : }
1222 :
1223 :
1224 : /* Loop autovectorization. */
1225 :
1226 : namespace {
1227 :
1228 : const pass_data pass_data_vectorize =
1229 : {
1230 : GIMPLE_PASS, /* type */
1231 : "vect", /* name */
1232 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1233 : TV_TREE_VECTORIZATION, /* tv_id */
1234 : ( PROP_cfg | PROP_ssa ), /* properties_required */
1235 : 0, /* properties_provided */
1236 : 0, /* properties_destroyed */
1237 : 0, /* todo_flags_start */
1238 : 0, /* todo_flags_finish */
1239 : };
1240 :
1241 : class pass_vectorize : public gimple_opt_pass
1242 : {
1243 : public:
1244 298828 : pass_vectorize (gcc::context *ctxt)
1245 597656 : : gimple_opt_pass (pass_data_vectorize, ctxt)
1246 : {}
1247 :
1248 : /* opt_pass methods: */
1249 240924 : bool gate (function *fun) final override
1250 : {
1251 240924 : return flag_tree_loop_vectorize || fun->has_force_vectorize_loops;
1252 : }
1253 :
1254 : unsigned int execute (function *) final override;
1255 :
1256 : }; // class pass_vectorize
1257 :
1258 : /* Function vectorize_loops.
1259 :
1260 : Entry point to loop vectorization phase. */
1261 :
1262 : unsigned
1263 207522 : pass_vectorize::execute (function *fun)
1264 : {
1265 207522 : unsigned int i;
1266 207522 : unsigned int num_vectorized_loops = 0;
1267 207522 : unsigned int vect_loops_num;
1268 207522 : hash_table<simduid_to_vf> *simduid_to_vf_htab = NULL;
1269 207522 : hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
1270 207522 : bool any_ifcvt_loops = false;
1271 207522 : unsigned ret = 0;
1272 :
1273 207522 : vect_loops_num = number_of_loops (fun);
1274 :
1275 : /* Bail out if there are no loops. */
1276 207522 : if (vect_loops_num <= 1)
1277 : return 0;
1278 :
1279 207522 : vect_slp_init ();
1280 :
1281 207522 : if (fun->has_simduid_loops)
1282 5601 : note_simd_array_uses (&simd_array_to_simduid_htab, fun);
1283 :
1284 : /* ----------- Analyze loops. ----------- */
1285 207522 : enable_ranger (fun);
1286 :
1287 : /* If some loop was duplicated, it gets bigger number
1288 : than all previously defined loops. This fact allows us to run
1289 : only over initial loops skipping newly generated ones. */
1290 1153140 : for (auto loop : loops_list (fun, 0))
1291 530574 : if (loop->dont_vectorize)
1292 : {
1293 35690 : any_ifcvt_loops = true;
1294 : /* If-conversion sometimes versions both the outer loop
1295 : (for the case when outer loop vectorization might be
1296 : desirable) as well as the inner loop in the scalar version
1297 : of the loop. So we have:
1298 : if (LOOP_VECTORIZED (1, 3))
1299 : {
1300 : loop1
1301 : loop2
1302 : }
1303 : else
1304 : loop3 (copy of loop1)
1305 : if (LOOP_VECTORIZED (4, 5))
1306 : loop4 (copy of loop2)
1307 : else
1308 : loop5 (copy of loop4)
1309 : If loops' iteration gives us loop3 first (which has
1310 : dont_vectorize set), make sure to process loop1 before loop4;
1311 : so that we can prevent vectorization of loop4 if loop1
1312 : is successfully vectorized. */
1313 35690 : if (loop->inner)
1314 : {
1315 2168 : gimple *loop_vectorized_call
1316 2168 : = vect_loop_vectorized_call (loop);
1317 2168 : if (loop_vectorized_call
1318 2168 : && vect_loop_vectorized_call (loop->inner))
1319 : {
1320 923 : tree arg = gimple_call_arg (loop_vectorized_call, 0);
1321 923 : class loop *vector_loop
1322 923 : = get_loop (fun, tree_to_shwi (arg));
1323 923 : if (vector_loop && vector_loop != loop)
1324 : {
1325 : /* Make sure we don't vectorize it twice. */
1326 923 : vector_loop->dont_vectorize = true;
1327 923 : ret |= try_vectorize_loop (simduid_to_vf_htab,
1328 : &num_vectorized_loops,
1329 : vector_loop, fun);
1330 : }
1331 : }
1332 : }
1333 : }
1334 : else
1335 494884 : ret |= try_vectorize_loop (simduid_to_vf_htab, &num_vectorized_loops,
1336 207522 : loop, fun);
1337 :
1338 207522 : vect_location = dump_user_location_t ();
1339 :
1340 207522 : statistics_counter_event (fun, "Vectorized loops", num_vectorized_loops);
1341 207522 : if (dump_enabled_p ()
1342 207522 : || (num_vectorized_loops > 0 && dump_enabled_p ()))
1343 11662 : dump_printf_loc (MSG_NOTE, vect_location,
1344 : "vectorized %u loops in function.\n",
1345 : num_vectorized_loops);
1346 :
1347 : /* ----------- Finalize. ----------- */
1348 207522 : disable_ranger (fun);
1349 :
1350 207522 : if (any_ifcvt_loops)
1351 322930 : for (i = 1; i < number_of_loops (fun); i++)
1352 : {
1353 139705 : class loop *loop = get_loop (fun, i);
1354 139705 : if (loop && loop->dont_vectorize)
1355 : {
1356 38263 : gimple *g = vect_loop_vectorized_call (loop);
1357 38263 : if (g)
1358 : {
1359 21199 : fold_loop_internal_call (g, boolean_false_node);
1360 21199 : loop->dont_vectorize = false;
1361 21199 : ret |= TODO_cleanup_cfg;
1362 21199 : g = NULL;
1363 : }
1364 : else
1365 17064 : g = vect_loop_dist_alias_call (loop, fun);
1366 :
1367 38263 : if (g)
1368 : {
1369 30 : fold_loop_internal_call (g, boolean_false_node);
1370 30 : loop->dont_vectorize = false;
1371 30 : ret |= TODO_cleanup_cfg;
1372 : }
1373 : }
1374 : }
1375 :
1376 : /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1377 207522 : if (fun->has_simduid_loops)
1378 : {
1379 5601 : adjust_simduid_builtins (simduid_to_vf_htab, fun);
1380 : /* Avoid stale SCEV cache entries for the SIMD_LANE defs. */
1381 5601 : scev_reset ();
1382 : }
1383 : /* Shrink any "omp array simd" temporary arrays to the
1384 : actual vectorization factors. */
1385 207522 : if (simd_array_to_simduid_htab)
1386 2204 : shrink_simd_arrays (simd_array_to_simduid_htab, simduid_to_vf_htab);
1387 207522 : delete simduid_to_vf_htab;
1388 207522 : fun->has_simduid_loops = false;
1389 :
1390 207522 : if (num_vectorized_loops > 0)
1391 : {
1392 : /* We are collecting some corner cases where we need to update
1393 : virtual SSA form via the TODO but delete the queued update-SSA
1394 : state. Force renaming if we think that might be necessary. */
1395 37281 : if (ret & TODO_update_ssa_only_virtuals)
1396 89 : mark_virtual_operands_for_renaming (cfun);
1397 : /* If we vectorized any loop only virtual SSA form needs to be updated.
1398 : ??? Also while we try hard to update loop-closed SSA form we fail
1399 : to properly do this in some corner-cases (see PR56286). */
1400 37281 : rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_only_virtuals);
1401 37281 : ret |= TODO_cleanup_cfg;
1402 : }
1403 :
1404 1852228 : for (i = 1; i < number_of_loops (fun); i++)
1405 : {
1406 718592 : loop_vec_info loop_vinfo;
1407 718592 : bool has_mask_store;
1408 :
1409 718592 : class loop *loop = get_loop (fun, i);
1410 718592 : if (!loop || !loop->aux)
1411 656785 : continue;
1412 61807 : loop_vinfo = (loop_vec_info) loop->aux;
1413 61807 : has_mask_store = LOOP_VINFO_HAS_MASK_STORE (loop_vinfo);
1414 61807 : delete loop_vinfo;
1415 61807 : if (has_mask_store
1416 61807 : && targetm.vectorize.empty_mask_is_expensive (IFN_MASK_STORE))
1417 493 : optimize_mask_stores (loop);
1418 :
1419 61807 : auto_bitmap exit_bbs;
1420 : /* Perform local CSE, this esp. helps because we emit code for
1421 : predicates that need to be shared for optimal predicate usage.
1422 : However reassoc will re-order them and prevent CSE from working
1423 : as it should. CSE only the loop body, not the entry. */
1424 61807 : auto_vec<edge> exits = get_loop_exit_edges (loop);
1425 248755 : for (edge exit : exits)
1426 63334 : bitmap_set_bit (exit_bbs, exit->dest->index);
1427 :
1428 61807 : edge entry = EDGE_PRED (loop_preheader_edge (loop)->src, 0);
1429 61807 : do_rpo_vn (fun, entry, exit_bbs);
1430 :
1431 61807 : loop->aux = NULL;
1432 61807 : }
1433 :
1434 207522 : vect_slp_fini ();
1435 :
1436 207522 : return ret;
1437 : }
1438 :
1439 : } // anon namespace
1440 :
1441 : gimple_opt_pass *
1442 298828 : make_pass_vectorize (gcc::context *ctxt)
1443 : {
1444 298828 : return new pass_vectorize (ctxt);
1445 : }
1446 :
1447 : /* Entry point to the simduid cleanup pass. */
1448 :
1449 : namespace {
1450 :
1451 : const pass_data pass_data_simduid_cleanup =
1452 : {
1453 : GIMPLE_PASS, /* type */
1454 : "simduid", /* name */
1455 : OPTGROUP_NONE, /* optinfo_flags */
1456 : TV_NONE, /* tv_id */
1457 : ( PROP_ssa | PROP_cfg ), /* properties_required */
1458 : 0, /* properties_provided */
1459 : 0, /* properties_destroyed */
1460 : 0, /* todo_flags_start */
1461 : 0, /* todo_flags_finish */
1462 : };
1463 :
1464 : class pass_simduid_cleanup : public gimple_opt_pass
1465 : {
1466 : public:
1467 597656 : pass_simduid_cleanup (gcc::context *ctxt)
1468 1195312 : : gimple_opt_pass (pass_data_simduid_cleanup, ctxt)
1469 : {}
1470 :
1471 : /* opt_pass methods: */
1472 298828 : opt_pass * clone () final override
1473 : {
1474 298828 : return new pass_simduid_cleanup (m_ctxt);
1475 : }
1476 2528978 : bool gate (function *fun) final override { return fun->has_simduid_loops; }
1477 : unsigned int execute (function *) final override;
1478 :
1479 : }; // class pass_simduid_cleanup
1480 :
1481 : unsigned int
1482 2208 : pass_simduid_cleanup::execute (function *fun)
1483 : {
1484 2208 : hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
1485 :
1486 2208 : note_simd_array_uses (&simd_array_to_simduid_htab, fun);
1487 :
1488 : /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1489 2208 : adjust_simduid_builtins (NULL, fun);
1490 :
1491 : /* Shrink any "omp array simd" temporary arrays to the
1492 : actual vectorization factors. */
1493 2208 : if (simd_array_to_simduid_htab)
1494 4 : shrink_simd_arrays (simd_array_to_simduid_htab, NULL);
1495 2208 : fun->has_simduid_loops = false;
1496 2208 : return 0;
1497 : }
1498 :
1499 : } // anon namespace
1500 :
1501 : gimple_opt_pass *
1502 298828 : make_pass_simduid_cleanup (gcc::context *ctxt)
1503 : {
1504 298828 : return new pass_simduid_cleanup (ctxt);
1505 : }
1506 :
1507 :
1508 : /* Entry point to basic block SLP phase. */
1509 :
1510 : namespace {
1511 :
1512 : const pass_data pass_data_slp_vectorize =
1513 : {
1514 : GIMPLE_PASS, /* type */
1515 : "slp", /* name */
1516 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1517 : TV_TREE_SLP_VECTORIZATION, /* tv_id */
1518 : ( PROP_ssa | PROP_cfg ), /* properties_required */
1519 : 0, /* properties_provided */
1520 : 0, /* properties_destroyed */
1521 : 0, /* todo_flags_start */
1522 : TODO_update_ssa, /* todo_flags_finish */
1523 : };
1524 :
1525 : class pass_slp_vectorize : public gimple_opt_pass
1526 : {
1527 : public:
1528 597656 : pass_slp_vectorize (gcc::context *ctxt)
1529 1195312 : : gimple_opt_pass (pass_data_slp_vectorize, ctxt)
1530 : {}
1531 :
1532 : /* opt_pass methods: */
1533 298828 : opt_pass * clone () final override { return new pass_slp_vectorize (m_ctxt); }
1534 1039824 : bool gate (function *) final override { return flag_tree_slp_vectorize != 0; }
1535 : unsigned int execute (function *) final override;
1536 :
1537 : }; // class pass_slp_vectorize
1538 :
1539 : unsigned int
1540 905907 : pass_slp_vectorize::execute (function *fun)
1541 : {
1542 905907 : auto_purge_vect_location sentinel;
1543 905907 : basic_block bb;
1544 :
1545 905907 : bool in_loop_pipeline = scev_initialized_p ();
1546 905907 : if (!in_loop_pipeline)
1547 : {
1548 700016 : loop_optimizer_init (LOOPS_NORMAL);
1549 700016 : scev_initialize ();
1550 : }
1551 :
1552 : /* Mark all stmts as not belonging to the current region and unvisited. */
1553 11393593 : FOR_EACH_BB_FN (bb, fun)
1554 : {
1555 15156288 : for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
1556 4668602 : gsi_next (&gsi))
1557 : {
1558 4668602 : gphi *stmt = gsi.phi ();
1559 4668602 : gimple_set_uid (stmt, -1);
1560 4668602 : gimple_set_visited (stmt, false);
1561 : }
1562 98624488 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1563 77649116 : gsi_next (&gsi))
1564 : {
1565 77649116 : gimple *stmt = gsi_stmt (gsi);
1566 77649116 : gimple_set_uid (stmt, -1);
1567 77649116 : gimple_set_visited (stmt, false);
1568 : }
1569 : }
1570 :
1571 905907 : vect_slp_init ();
1572 :
1573 905907 : vect_slp_function (fun);
1574 :
1575 905907 : vect_slp_fini ();
1576 :
1577 905907 : if (!in_loop_pipeline)
1578 : {
1579 700016 : scev_finalize ();
1580 700016 : loop_optimizer_finalize ();
1581 : }
1582 :
1583 1811814 : return 0;
1584 905907 : }
1585 :
1586 : } // anon namespace
1587 :
1588 : gimple_opt_pass *
1589 298828 : make_pass_slp_vectorize (gcc::context *ctxt)
1590 : {
1591 298828 : return new pass_slp_vectorize (ctxt);
1592 : }
1593 :
1594 :
1595 : /* Increase alignment of global arrays to improve vectorization potential.
1596 : TODO:
1597 : - Consider also structs that have an array field.
1598 : - Use ipa analysis to prune arrays that can't be vectorized?
1599 : This should involve global alignment analysis and in the future also
1600 : array padding. */
1601 :
1602 : static unsigned get_vec_alignment_for_type (tree);
1603 : static hash_map<tree, unsigned> *type_align_map;
1604 :
1605 : /* Return alignment of array's vector type corresponding to scalar type.
1606 : 0 if no vector type exists. */
1607 : static unsigned
1608 0 : get_vec_alignment_for_array_type (tree type)
1609 : {
1610 0 : gcc_assert (TREE_CODE (type) == ARRAY_TYPE);
1611 0 : poly_uint64 array_size, vector_size;
1612 :
1613 0 : tree scalar_type = strip_array_types (type);
1614 0 : tree vectype = get_related_vectype_for_scalar_type (VOIDmode, scalar_type);
1615 0 : if (!vectype
1616 0 : || !poly_int_tree_p (TYPE_SIZE (type), &array_size)
1617 0 : || !poly_int_tree_p (TYPE_SIZE (vectype), &vector_size)
1618 0 : || maybe_lt (array_size, vector_size))
1619 0 : return 0;
1620 :
1621 0 : return TYPE_ALIGN (vectype);
1622 : }
1623 :
1624 : /* Return alignment of field having maximum alignment of vector type
1625 : corresponding to it's scalar type. For now, we only consider fields whose
1626 : offset is a multiple of it's vector alignment.
1627 : 0 if no suitable field is found. */
1628 : static unsigned
1629 0 : get_vec_alignment_for_record_type (tree type)
1630 : {
1631 0 : gcc_assert (TREE_CODE (type) == RECORD_TYPE);
1632 :
1633 0 : unsigned max_align = 0, alignment;
1634 0 : HOST_WIDE_INT offset;
1635 0 : tree offset_tree;
1636 :
1637 0 : if (TYPE_PACKED (type))
1638 : return 0;
1639 :
1640 0 : unsigned *slot = type_align_map->get (type);
1641 0 : if (slot)
1642 0 : return *slot;
1643 :
1644 0 : for (tree field = first_field (type);
1645 0 : field != NULL_TREE;
1646 0 : field = DECL_CHAIN (field))
1647 : {
1648 : /* Skip if not FIELD_DECL or if alignment is set by user. */
1649 0 : if (TREE_CODE (field) != FIELD_DECL
1650 0 : || DECL_USER_ALIGN (field)
1651 0 : || DECL_ARTIFICIAL (field))
1652 0 : continue;
1653 :
1654 : /* We don't need to process the type further if offset is variable,
1655 : since the offsets of remaining members will also be variable. */
1656 0 : if (TREE_CODE (DECL_FIELD_OFFSET (field)) != INTEGER_CST
1657 0 : || TREE_CODE (DECL_FIELD_BIT_OFFSET (field)) != INTEGER_CST)
1658 : break;
1659 :
1660 : /* Similarly stop processing the type if offset_tree
1661 : does not fit in unsigned HOST_WIDE_INT. */
1662 0 : offset_tree = bit_position (field);
1663 0 : if (!tree_fits_uhwi_p (offset_tree))
1664 : break;
1665 :
1666 0 : offset = tree_to_uhwi (offset_tree);
1667 0 : alignment = get_vec_alignment_for_type (TREE_TYPE (field));
1668 :
1669 : /* Get maximum alignment of vectorized field/array among those members
1670 : whose offset is multiple of the vector alignment. */
1671 0 : if (alignment
1672 0 : && (offset % alignment == 0)
1673 0 : && (alignment > max_align))
1674 0 : max_align = alignment;
1675 : }
1676 :
1677 0 : type_align_map->put (type, max_align);
1678 0 : return max_align;
1679 : }
1680 :
1681 : /* Return alignment of vector type corresponding to decl's scalar type
1682 : or 0 if it doesn't exist or the vector alignment is lesser than
1683 : decl's alignment. */
1684 : static unsigned
1685 0 : get_vec_alignment_for_type (tree type)
1686 : {
1687 0 : if (type == NULL_TREE)
1688 : return 0;
1689 :
1690 0 : gcc_assert (TYPE_P (type));
1691 :
1692 0 : static unsigned alignment = 0;
1693 0 : switch (TREE_CODE (type))
1694 : {
1695 0 : case ARRAY_TYPE:
1696 0 : alignment = get_vec_alignment_for_array_type (type);
1697 0 : break;
1698 0 : case RECORD_TYPE:
1699 0 : alignment = get_vec_alignment_for_record_type (type);
1700 0 : break;
1701 0 : default:
1702 0 : alignment = 0;
1703 0 : break;
1704 : }
1705 :
1706 0 : return (alignment > TYPE_ALIGN (type)) ? alignment : 0;
1707 : }
1708 :
1709 : /* Entry point to increase_alignment pass. */
1710 : static unsigned int
1711 0 : increase_alignment (void)
1712 : {
1713 0 : varpool_node *vnode;
1714 :
1715 0 : vect_location = dump_user_location_t ();
1716 0 : type_align_map = new hash_map<tree, unsigned>;
1717 :
1718 : /* Increase the alignment of all global arrays for vectorization. */
1719 0 : FOR_EACH_DEFINED_VARIABLE (vnode)
1720 : {
1721 0 : tree decl = vnode->decl;
1722 0 : unsigned int alignment;
1723 :
1724 0 : if ((decl_in_symtab_p (decl)
1725 0 : && !symtab_node::get (decl)->can_increase_alignment_p ())
1726 0 : || DECL_USER_ALIGN (decl) || DECL_ARTIFICIAL (decl))
1727 0 : continue;
1728 :
1729 0 : alignment = get_vec_alignment_for_type (TREE_TYPE (decl));
1730 0 : if (alignment && vect_can_force_dr_alignment_p (decl, alignment))
1731 : {
1732 0 : vnode->increase_alignment (alignment);
1733 0 : if (dump_enabled_p ())
1734 0 : dump_printf (MSG_NOTE, "Increasing alignment of decl: %T\n", decl);
1735 : }
1736 : }
1737 :
1738 0 : delete type_align_map;
1739 0 : return 0;
1740 : }
1741 :
1742 :
1743 : namespace {
1744 :
1745 : const pass_data pass_data_ipa_increase_alignment =
1746 : {
1747 : SIMPLE_IPA_PASS, /* type */
1748 : "increase_alignment", /* name */
1749 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1750 : TV_IPA_OPT, /* tv_id */
1751 : 0, /* properties_required */
1752 : 0, /* properties_provided */
1753 : 0, /* properties_destroyed */
1754 : 0, /* todo_flags_start */
1755 : 0, /* todo_flags_finish */
1756 : };
1757 :
1758 : class pass_ipa_increase_alignment : public simple_ipa_opt_pass
1759 : {
1760 : public:
1761 298828 : pass_ipa_increase_alignment (gcc::context *ctxt)
1762 597656 : : simple_ipa_opt_pass (pass_data_ipa_increase_alignment, ctxt)
1763 : {}
1764 :
1765 : /* opt_pass methods: */
1766 238720 : bool gate (function *) final override
1767 : {
1768 238720 : return flag_section_anchors && flag_tree_loop_vectorize;
1769 : }
1770 :
1771 0 : unsigned int execute (function *) final override
1772 : {
1773 0 : return increase_alignment ();
1774 : }
1775 :
1776 : }; // class pass_ipa_increase_alignment
1777 :
1778 : } // anon namespace
1779 :
1780 : simple_ipa_opt_pass *
1781 298828 : make_pass_ipa_increase_alignment (gcc::context *ctxt)
1782 : {
1783 298828 : return new pass_ipa_increase_alignment (ctxt);
1784 : }
1785 :
1786 : /* If the condition represented by T is a comparison or the SSA name
1787 : result of a comparison, extract the comparison's operands. Represent
1788 : T as NE_EXPR <T, 0> otherwise. */
1789 :
1790 : void
1791 63801 : scalar_cond_masked_key::get_cond_ops_from_tree (tree t)
1792 : {
1793 63801 : if (TREE_CODE_CLASS (TREE_CODE (t)) == tcc_comparison)
1794 : {
1795 0 : this->code = TREE_CODE (t);
1796 0 : this->op0 = TREE_OPERAND (t, 0);
1797 0 : this->op1 = TREE_OPERAND (t, 1);
1798 0 : this->inverted_p = false;
1799 0 : return;
1800 : }
1801 :
1802 63801 : if (TREE_CODE (t) == SSA_NAME)
1803 26230 : if (gassign *stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (t)))
1804 : {
1805 26230 : tree_code code = gimple_assign_rhs_code (stmt);
1806 26230 : if (TREE_CODE_CLASS (code) == tcc_comparison)
1807 : {
1808 17859 : this->code = code;
1809 17859 : this->op0 = gimple_assign_rhs1 (stmt);
1810 17859 : this->op1 = gimple_assign_rhs2 (stmt);
1811 17859 : this->inverted_p = false;
1812 17859 : return;
1813 : }
1814 8371 : else if (code == BIT_NOT_EXPR)
1815 : {
1816 3541 : tree n_op = gimple_assign_rhs1 (stmt);
1817 3541 : if ((stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (n_op))))
1818 : {
1819 3541 : code = gimple_assign_rhs_code (stmt);
1820 3541 : if (TREE_CODE_CLASS (code) == tcc_comparison)
1821 : {
1822 3503 : this->code = code;
1823 3503 : this->op0 = gimple_assign_rhs1 (stmt);
1824 3503 : this->op1 = gimple_assign_rhs2 (stmt);
1825 3503 : this->inverted_p = true;
1826 3503 : return;
1827 : }
1828 : }
1829 : }
1830 : }
1831 :
1832 42439 : this->code = NE_EXPR;
1833 42439 : this->op0 = t;
1834 42439 : this->op1 = build_zero_cst (TREE_TYPE (t));
1835 42439 : this->inverted_p = false;
1836 : }
1837 :
1838 : /* See the comment above the declaration for details. */
1839 :
1840 : unsigned int
1841 0 : vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
1842 : stmt_vec_info stmt_info, slp_tree,
1843 : tree vectype, int misalign,
1844 : vect_cost_model_location where)
1845 : {
1846 0 : unsigned int cost
1847 0 : = builtin_vectorization_cost (kind, vectype, misalign) * count;
1848 0 : return record_stmt_cost (stmt_info, where, cost);
1849 : }
1850 :
1851 : unsigned int
1852 2620063 : vector_costs::add_slp_cost (slp_tree,
1853 : const array_slice<stmt_info_for_cost> &cost_vec)
1854 : {
1855 2620063 : unsigned int sum = 0;
1856 5819165 : for (auto item : cost_vec)
1857 3199102 : sum += ::add_stmt_cost (this, item.count, item.kind, item.stmt_info,
1858 : item.node, item.vectype, item.misalign, item.where);
1859 2620063 : return sum;
1860 : }
1861 :
1862 : /* See the comment above the declaration for details. */
1863 :
1864 : void
1865 1807703 : vector_costs::finish_cost (const vector_costs *)
1866 : {
1867 1807703 : gcc_assert (!m_finished);
1868 1807703 : m_finished = true;
1869 1807703 : }
1870 :
1871 : /* Record a base cost of COST units against WHERE. If STMT_INFO is
1872 : nonnull, use it to adjust the cost based on execution frequency
1873 : (where appropriate). */
1874 :
1875 : unsigned int
1876 0 : vector_costs::record_stmt_cost (stmt_vec_info stmt_info,
1877 : vect_cost_model_location where,
1878 : unsigned int cost)
1879 : {
1880 0 : cost = adjust_cost_for_freq (stmt_info, where, cost);
1881 0 : m_costs[where] += cost;
1882 0 : return cost;
1883 : }
1884 :
1885 : /* COST is the base cost we have calculated for an operation in location WHERE.
1886 : If STMT_INFO is nonnull, use it to adjust the cost based on execution
1887 : frequency (where appropriate). Return the adjusted cost. */
1888 :
1889 : unsigned int
1890 7470412 : vector_costs::adjust_cost_for_freq (stmt_vec_info stmt_info,
1891 : vect_cost_model_location where,
1892 : unsigned int cost)
1893 : {
1894 : /* Statements in an inner loop relative to the loop being
1895 : vectorized are weighted more heavily. The value here is
1896 : arbitrary and could potentially be improved with analysis. */
1897 7470412 : if (where == vect_body
1898 7470412 : && stmt_info
1899 7470412 : && stmt_in_inner_loop_p (m_vinfo, stmt_info))
1900 : {
1901 11885 : loop_vec_info loop_vinfo = as_a<loop_vec_info> (m_vinfo);
1902 11885 : cost *= LOOP_VINFO_INNER_LOOP_COST_FACTOR (loop_vinfo);
1903 : }
1904 7470412 : return cost;
1905 : }
1906 :
1907 : /* See the comment above the declaration for details. */
1908 :
1909 : bool
1910 30975 : vector_costs::better_main_loop_than_p (const vector_costs *other) const
1911 : {
1912 30975 : int diff = compare_inside_loop_cost (other);
1913 30975 : if (diff != 0)
1914 30809 : return diff < 0;
1915 :
1916 : /* If there's nothing to choose between the loop bodies, see whether
1917 : there's a difference in the prologue and epilogue costs. */
1918 166 : diff = compare_outside_loop_cost (other);
1919 166 : if (diff != 0)
1920 138 : return diff < 0;
1921 :
1922 : return false;
1923 : }
1924 :
1925 :
1926 : /* See the comment above the declaration for details. */
1927 :
1928 : bool
1929 1409 : vector_costs::better_epilogue_loop_than_p (const vector_costs *other,
1930 : loop_vec_info main_loop) const
1931 : {
1932 1409 : loop_vec_info this_loop_vinfo = as_a<loop_vec_info> (this->m_vinfo);
1933 1409 : loop_vec_info other_loop_vinfo = as_a<loop_vec_info> (other->m_vinfo);
1934 :
1935 1409 : poly_int64 this_vf = LOOP_VINFO_VECT_FACTOR (this_loop_vinfo);
1936 1409 : poly_int64 other_vf = LOOP_VINFO_VECT_FACTOR (other_loop_vinfo);
1937 :
1938 1409 : poly_uint64 main_poly_vf = LOOP_VINFO_VECT_FACTOR (main_loop);
1939 1409 : unsigned HOST_WIDE_INT main_vf;
1940 1409 : unsigned HOST_WIDE_INT other_factor, this_factor, other_cost, this_cost;
1941 : /* If we can determine how many iterations are left for the epilogue
1942 : loop, that is if both the main loop's vectorization factor and number
1943 : of iterations are constant, then we use them to calculate the cost of
1944 : the epilogue loop together with a 'likely value' for the epilogues
1945 : vectorization factor. Otherwise we use the main loop's vectorization
1946 : factor and the maximum poly value for the epilogue's. If the target
1947 : has not provided with a sensible upper bound poly vectorization
1948 : factors are likely to be favored over constant ones. */
1949 1409 : if (main_poly_vf.is_constant (&main_vf)
1950 1409 : && LOOP_VINFO_NITERS_KNOWN_P (main_loop))
1951 : {
1952 94 : unsigned HOST_WIDE_INT niters
1953 94 : = LOOP_VINFO_INT_NITERS (main_loop) % main_vf;
1954 94 : HOST_WIDE_INT other_likely_vf
1955 94 : = estimated_poly_value (other_vf, POLY_VALUE_LIKELY);
1956 94 : HOST_WIDE_INT this_likely_vf
1957 94 : = estimated_poly_value (this_vf, POLY_VALUE_LIKELY);
1958 :
1959 : /* If the epilogue is using partial vectors we account for the
1960 : partial iteration here too. */
1961 94 : other_factor = niters / other_likely_vf;
1962 94 : if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (other_loop_vinfo)
1963 0 : && niters % other_likely_vf != 0)
1964 0 : other_factor++;
1965 :
1966 94 : this_factor = niters / this_likely_vf;
1967 94 : if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (this_loop_vinfo)
1968 0 : && niters % this_likely_vf != 0)
1969 0 : this_factor++;
1970 : }
1971 : else
1972 : {
1973 1315 : unsigned HOST_WIDE_INT main_vf_max
1974 1315 : = estimated_poly_value (main_poly_vf, POLY_VALUE_MAX);
1975 1315 : unsigned HOST_WIDE_INT other_vf_max
1976 1315 : = estimated_poly_value (other_vf, POLY_VALUE_MAX);
1977 1315 : unsigned HOST_WIDE_INT this_vf_max
1978 1315 : = estimated_poly_value (this_vf, POLY_VALUE_MAX);
1979 :
1980 1315 : other_factor = CEIL (main_vf_max, other_vf_max);
1981 1315 : this_factor = CEIL (main_vf_max, this_vf_max);
1982 :
1983 : /* If the loop is not using partial vectors then it will iterate one
1984 : time less than one that does. It is safe to subtract one here,
1985 : because the main loop's vf is always at least 2x bigger than that
1986 : of an epilogue. */
1987 1315 : if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (other_loop_vinfo))
1988 1302 : other_factor -= 1;
1989 1315 : if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (this_loop_vinfo))
1990 1313 : this_factor -= 1;
1991 : }
1992 :
1993 : /* Compute the costs by multiplying the inside costs with the factor and
1994 : add the outside costs for a more complete picture. The factor is the
1995 : amount of times we are expecting to iterate this epilogue. */
1996 1409 : other_cost = other->body_cost () * other_factor;
1997 1409 : this_cost = this->body_cost () * this_factor;
1998 1409 : other_cost += other->outside_cost ();
1999 1409 : this_cost += this->outside_cost ();
2000 1409 : return this_cost < other_cost;
2001 : }
2002 :
2003 : /* A <=>-style subroutine of better_main_loop_than_p. Check whether we can
2004 : determine the return value of better_main_loop_than_p by comparing the
2005 : inside (loop body) costs of THIS and OTHER. Return:
2006 :
2007 : * -1 if better_main_loop_than_p should return true.
2008 : * 1 if better_main_loop_than_p should return false.
2009 : * 0 if we can't decide. */
2010 :
2011 : int
2012 30975 : vector_costs::compare_inside_loop_cost (const vector_costs *other) const
2013 : {
2014 30975 : loop_vec_info this_loop_vinfo = as_a<loop_vec_info> (this->m_vinfo);
2015 30975 : loop_vec_info other_loop_vinfo = as_a<loop_vec_info> (other->m_vinfo);
2016 :
2017 30975 : struct loop *loop = LOOP_VINFO_LOOP (this_loop_vinfo);
2018 30975 : gcc_assert (LOOP_VINFO_LOOP (other_loop_vinfo) == loop);
2019 :
2020 30975 : poly_int64 this_vf = LOOP_VINFO_VECT_FACTOR (this_loop_vinfo);
2021 30975 : poly_int64 other_vf = LOOP_VINFO_VECT_FACTOR (other_loop_vinfo);
2022 :
2023 : /* Limit the VFs to what is likely to be the maximum number of iterations,
2024 : to handle cases in which at least one loop_vinfo is fully-masked. */
2025 30975 : HOST_WIDE_INT estimated_max_niter = likely_max_stmt_executions_int (loop);
2026 30975 : if (estimated_max_niter != -1)
2027 : {
2028 29883 : if (estimated_poly_value (this_vf, POLY_VALUE_MIN)
2029 : >= estimated_max_niter)
2030 : this_vf = estimated_max_niter;
2031 29883 : if (estimated_poly_value (other_vf, POLY_VALUE_MIN)
2032 : >= estimated_max_niter)
2033 : other_vf = estimated_max_niter;
2034 : }
2035 :
2036 : /* Check whether the (fractional) cost per scalar iteration is lower or
2037 : higher: this_inside_cost / this_vf vs. other_inside_cost / other_vf. */
2038 30975 : poly_int64 rel_this = this_loop_vinfo->vector_costs->body_cost () * other_vf;
2039 30975 : poly_int64 rel_other
2040 30975 : = other_loop_vinfo->vector_costs->body_cost () * this_vf;
2041 :
2042 30975 : HOST_WIDE_INT est_rel_this_min
2043 30975 : = estimated_poly_value (rel_this, POLY_VALUE_MIN);
2044 30975 : HOST_WIDE_INT est_rel_this_max
2045 30975 : = estimated_poly_value (rel_this, POLY_VALUE_MAX);
2046 :
2047 30975 : HOST_WIDE_INT est_rel_other_min
2048 30975 : = estimated_poly_value (rel_other, POLY_VALUE_MIN);
2049 30975 : HOST_WIDE_INT est_rel_other_max
2050 30975 : = estimated_poly_value (rel_other, POLY_VALUE_MAX);
2051 :
2052 : /* Check first if we can make out an unambiguous total order from the minimum
2053 : and maximum estimates. */
2054 30975 : if (est_rel_this_min < est_rel_other_min
2055 : && est_rel_this_max < est_rel_other_max)
2056 : return -1;
2057 :
2058 29935 : if (est_rel_other_min < est_rel_this_min
2059 : && est_rel_other_max < est_rel_this_max)
2060 29769 : return 1;
2061 :
2062 : /* When other_loop_vinfo uses a variable vectorization factor,
2063 : we know that it has a lower cost for at least one runtime VF.
2064 : However, we don't know how likely that VF is.
2065 :
2066 : One option would be to compare the costs for the estimated VFs.
2067 : The problem is that that can put too much pressure on the cost
2068 : model. E.g. if the estimated VF is also the lowest possible VF,
2069 : and if other_loop_vinfo is 1 unit worse than this_loop_vinfo
2070 : for the estimated VF, we'd then choose this_loop_vinfo even
2071 : though (a) this_loop_vinfo might not actually be better than
2072 : other_loop_vinfo for that VF and (b) it would be significantly
2073 : worse at larger VFs.
2074 :
2075 : Here we go for a hacky compromise: pick this_loop_vinfo if it is
2076 : no more expensive than other_loop_vinfo even after doubling the
2077 : estimated other_loop_vinfo VF. For all but trivial loops, this
2078 : ensures that we only pick this_loop_vinfo if it is significantly
2079 : better than other_loop_vinfo at the estimated VF. */
2080 : if (est_rel_other_min != est_rel_this_min
2081 : || est_rel_other_max != est_rel_this_max)
2082 : {
2083 : HOST_WIDE_INT est_rel_this_likely
2084 : = estimated_poly_value (rel_this, POLY_VALUE_LIKELY);
2085 : HOST_WIDE_INT est_rel_other_likely
2086 : = estimated_poly_value (rel_other, POLY_VALUE_LIKELY);
2087 :
2088 : return est_rel_this_likely * 2 <= est_rel_other_likely ? -1 : 1;
2089 : }
2090 :
2091 : return 0;
2092 : }
2093 :
2094 : /* A <=>-style subroutine of better_main_loop_than_p, used when there is
2095 : nothing to choose between the inside (loop body) costs of THIS and OTHER.
2096 : Check whether we can determine the return value of better_main_loop_than_p
2097 : by comparing the outside (prologue and epilogue) costs of THIS and OTHER.
2098 : Return:
2099 :
2100 : * -1 if better_main_loop_than_p should return true.
2101 : * 1 if better_main_loop_than_p should return false.
2102 : * 0 if we can't decide. */
2103 :
2104 : int
2105 166 : vector_costs::compare_outside_loop_cost (const vector_costs *other) const
2106 : {
2107 166 : auto this_outside_cost = this->outside_cost ();
2108 166 : auto other_outside_cost = other->outside_cost ();
2109 166 : if (this_outside_cost != other_outside_cost)
2110 138 : return this_outside_cost < other_outside_cost ? -1 : 1;
2111 :
2112 : return 0;
2113 : }
|
