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 1370994 : auto_purge_vect_location::~auto_purge_vect_location ()
96 : {
97 1370994 : vect_location = dump_user_location_t ();
98 1370994 : }
99 :
100 : /* Dump a cost entry according to args to F. */
101 :
102 : void
103 218578 : 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 218578 : if (stmt_info)
109 : {
110 201261 : print_gimple_expr (f, STMT_VINFO_STMT (stmt_info), 0, TDF_SLIM);
111 201261 : fprintf (f, " ");
112 : }
113 17317 : else if (node)
114 3457 : fprintf (f, "node %p ", (void *)node);
115 : else
116 13860 : fprintf (f, "<unknown> ");
117 218578 : fprintf (f, "%d times ", count);
118 218578 : const char *ks = "unknown";
119 218578 : switch (kind)
120 : {
121 49214 : case scalar_stmt:
122 49214 : ks = "scalar_stmt";
123 49214 : break;
124 37052 : case scalar_load:
125 37052 : ks = "scalar_load";
126 37052 : break;
127 27930 : case scalar_store:
128 27930 : ks = "scalar_store";
129 27930 : break;
130 34777 : case vector_stmt:
131 34777 : ks = "vector_stmt";
132 34777 : break;
133 22008 : case vector_load:
134 22008 : ks = "vector_load";
135 22008 : break;
136 0 : case vector_gather_load:
137 0 : ks = "vector_gather_load";
138 0 : break;
139 7846 : case unaligned_load:
140 7846 : ks = "unaligned_load";
141 7846 : break;
142 4648 : case unaligned_store:
143 4648 : ks = "unaligned_store";
144 4648 : break;
145 8881 : case vector_store:
146 8881 : ks = "vector_store";
147 8881 : break;
148 0 : case vector_scatter_store:
149 0 : ks = "vector_scatter_store";
150 0 : break;
151 3337 : case vec_to_scalar:
152 3337 : ks = "vec_to_scalar";
153 3337 : break;
154 9385 : case scalar_to_vec:
155 9385 : ks = "scalar_to_vec";
156 9385 : 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 6432 : case vec_perm:
164 6432 : ks = "vec_perm";
165 6432 : break;
166 5408 : case vec_promote_demote:
167 5408 : ks = "vec_promote_demote";
168 5408 : break;
169 1104 : case vec_construct:
170 1104 : ks = "vec_construct";
171 1104 : break;
172 : }
173 218578 : fprintf (f, "%s ", ks);
174 218578 : if (kind == unaligned_load || kind == unaligned_store)
175 12494 : fprintf (f, "(misalign %d) ", misalign);
176 218578 : fprintf (f, "costs %u ", cost);
177 218578 : const char *ws = "unknown";
178 218578 : switch (where)
179 : {
180 23184 : case vect_prologue:
181 23184 : ws = "prologue";
182 23184 : break;
183 187992 : case vect_body:
184 187992 : ws = "body";
185 187992 : break;
186 7402 : case vect_epilogue:
187 7402 : ws = "epilogue";
188 7402 : break;
189 : }
190 218578 : fprintf (f, "in %s\n", ws);
191 218578 : }
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 7776 : simduid_to_vf::hash (const simduid_to_vf *p)
208 : {
209 7776 : return p->simduid;
210 : }
211 :
212 : inline int
213 14243 : simduid_to_vf::equal (const simduid_to_vf *p1, const simduid_to_vf *p2)
214 : {
215 14243 : 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 24070 : simd_array_to_simduid::hash (const simd_array_to_simduid *p)
243 : {
244 24070 : return DECL_UID (p->decl);
245 : }
246 :
247 : inline int
248 16817 : simd_array_to_simduid::equal (const simd_array_to_simduid *p1,
249 : const simd_array_to_simduid *p2)
250 : {
251 16817 : 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 118627 : FOR_EACH_BB_FN (bb, fun)
264 : {
265 110818 : gimple_stmt_iterator i;
266 :
267 739164 : for (i = gsi_start_bb (bb); !gsi_end_p (i); )
268 : {
269 517528 : poly_uint64 vf = 1;
270 517528 : enum internal_fn ifn;
271 517528 : gimple *stmt = gsi_stmt (i);
272 517528 : tree t;
273 517528 : if (!is_gimple_call (stmt)
274 517528 : || !gimple_call_internal_p (stmt))
275 : {
276 509965 : gsi_next (&i);
277 510881 : 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 102470 : FOR_EACH_BB_FN (bb, fun)
403 583066 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
404 : {
405 393744 : gimple *stmt = gsi_stmt (gsi);
406 393744 : if (!is_gimple_call (stmt) || !gimple_call_internal_p (stmt))
407 387160 : 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 31198 : FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, lhs)
424 18030 : 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 2763761 : vec_info::vec_info (vec_info::vec_kind kind_in, vec_info_shared *shared_)
465 2763761 : : kind (kind_in),
466 2763761 : shared (shared_),
467 2763761 : stmt_vec_info_ro (false),
468 2763761 : bbs (NULL),
469 2763761 : nbbs (0),
470 2763761 : inv_pattern_def_seq (NULL)
471 : {
472 2763761 : stmt_vec_infos.create (50);
473 2763761 : }
474 :
475 2763761 : vec_info::~vec_info ()
476 : {
477 4963107 : for (slp_instance &instance : slp_instances)
478 1113502 : vect_free_slp_instance (instance);
479 :
480 2763761 : free_stmt_vec_infos ();
481 2763761 : }
482 :
483 2336857 : vec_info_shared::vec_info_shared ()
484 2336857 : : datarefs (vNULL),
485 2336857 : datarefs_copy (vNULL),
486 2336857 : ddrs (vNULL)
487 : {
488 2336857 : }
489 :
490 2336857 : vec_info_shared::~vec_info_shared ()
491 : {
492 2336857 : free_data_refs (datarefs);
493 2336857 : free_dependence_relations (ddrs);
494 2336857 : datarefs_copy.release ();
495 2336857 : }
496 :
497 : void
498 2107125 : vec_info_shared::save_datarefs ()
499 : {
500 2107125 : if (!flag_checking)
501 : return;
502 3161549 : datarefs_copy.reserve_exact (datarefs.length ());
503 12725637 : for (unsigned i = 0; i < datarefs.length (); ++i)
504 10618529 : datarefs_copy.quick_push (*datarefs[i]);
505 : }
506 :
507 : void
508 882306 : vec_info_shared::check_datarefs ()
509 : {
510 882306 : if (!flag_checking)
511 : return;
512 2643806 : gcc_assert (datarefs.length () == datarefs_copy.length ());
513 12197366 : for (unsigned i = 0; i < datarefs.length (); ++i)
514 11315060 : 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 61479081 : vec_info::add_stmt (gimple *stmt)
524 : {
525 61479081 : stmt_vec_info res = new_stmt_vec_info (stmt);
526 61479081 : set_vinfo_for_stmt (stmt, res);
527 61479081 : 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 1171 : vec_info::add_pattern_stmt (gimple *stmt, stmt_vec_info stmt_info)
536 : {
537 1171 : stmt_vec_info res = new_stmt_vec_info (stmt);
538 1171 : res->pattern_stmt_p = true;
539 1171 : set_vinfo_for_stmt (stmt, res, false);
540 1171 : STMT_VINFO_RELATED_STMT (res) = stmt_info;
541 1171 : 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 10856 : vec_info::resync_stmt_addr (gimple *stmt)
553 : {
554 10856 : unsigned int uid = gimple_uid (stmt);
555 10856 : if (uid > 0 && uid - 1 < stmt_vec_infos.length ())
556 : {
557 10856 : stmt_vec_info res = stmt_vec_infos[uid - 1];
558 10856 : if (res && res->stmt)
559 : {
560 10856 : res->stmt = stmt;
561 10856 : 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 475123195 : vec_info::lookup_stmt (gimple *stmt)
573 : {
574 475123195 : unsigned int uid = gimple_uid (stmt);
575 475123195 : if (uid > 0 && uid - 1 < stmt_vec_infos.length ())
576 : {
577 300345882 : stmt_vec_info res = stmt_vec_infos[uid - 1];
578 300345882 : if (res && res->stmt == stmt)
579 299992768 : 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 56176798 : vec_info::lookup_def (tree name)
590 : {
591 56176798 : if (TREE_CODE (name) == SSA_NAME
592 56176798 : && !SSA_NAME_IS_DEFAULT_DEF (name))
593 50939781 : 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 1914 : vec_info::lookup_single_use (tree lhs)
603 : {
604 1914 : use_operand_p dummy;
605 1914 : gimple *use_stmt;
606 1914 : if (single_imm_use (lhs, &dummy, &use_stmt))
607 1765 : return lookup_stmt (use_stmt);
608 : return NULL;
609 : }
610 :
611 : /* Return vectorization information about DR. */
612 :
613 : dr_vec_info *
614 49865160 : vec_info::lookup_dr (data_reference *dr)
615 : {
616 49865160 : stmt_vec_info stmt_info = lookup_stmt (DR_STMT (dr));
617 : /* DR_STMT should never refer to a stmt in a pattern replacement. */
618 49865160 : gcc_checking_assert (!is_pattern_stmt_p (stmt_info));
619 49865160 : 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 6385 : vec_info::move_dr (stmt_vec_info new_stmt_info, stmt_vec_info old_stmt_info)
627 : {
628 6385 : gcc_assert (!is_pattern_stmt_p (old_stmt_info));
629 6385 : STMT_VINFO_DR_INFO (old_stmt_info)->stmt = new_stmt_info;
630 6385 : new_stmt_info->dr_aux = old_stmt_info->dr_aux;
631 6385 : STMT_VINFO_DR_WRT_VEC_LOOP (new_stmt_info)
632 6385 : = STMT_VINFO_DR_WRT_VEC_LOOP (old_stmt_info);
633 6385 : STMT_VINFO_GATHER_SCATTER_P (new_stmt_info)
634 6385 : = STMT_VINFO_GATHER_SCATTER_P (old_stmt_info);
635 6385 : STMT_VINFO_STRIDED_P (new_stmt_info)
636 6385 : = STMT_VINFO_STRIDED_P (old_stmt_info);
637 6385 : STMT_VINFO_SIMD_LANE_ACCESS_P (new_stmt_info)
638 6385 : = STMT_VINFO_SIMD_LANE_ACCESS_P (old_stmt_info);
639 6385 : }
640 :
641 : /* Permanently remove the statement described by STMT_INFO from the
642 : function. */
643 :
644 : void
645 1489586 : vec_info::remove_stmt (stmt_vec_info stmt_info)
646 : {
647 1489586 : gcc_assert (!stmt_info->pattern_stmt_p);
648 1489586 : set_vinfo_for_stmt (stmt_info->stmt, NULL);
649 1489586 : unlink_stmt_vdef (stmt_info->stmt);
650 1489586 : gimple_stmt_iterator si = gsi_for_stmt (stmt_info->stmt);
651 1489586 : gsi_remove (&si, true);
652 1489586 : release_defs (stmt_info->stmt);
653 1489586 : free_stmt_vec_info (stmt_info);
654 1489586 : }
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 100872 : vec_info::insert_seq_on_entry (stmt_vec_info context, gimple_seq seq)
678 : {
679 100872 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (this))
680 : {
681 19154 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
682 19154 : basic_block new_bb;
683 19154 : edge pe;
684 :
685 19154 : if (context && nested_in_vect_loop_p (loop, context))
686 : loop = loop->inner;
687 :
688 19154 : pe = loop_preheader_edge (loop);
689 19154 : new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
690 19154 : gcc_assert (!new_bb);
691 : }
692 : else
693 : {
694 81718 : gimple_stmt_iterator gsi_region_begin
695 81718 : = gsi_after_labels (bbs[0]);
696 81718 : gsi_insert_seq_before (&gsi_region_begin, seq, GSI_SAME_STMT);
697 : }
698 100872 : }
699 :
700 : /* Like insert_seq_on_entry but just inserts the single stmt NEW_STMT. */
701 :
702 : void
703 3305 : vec_info::insert_on_entry (stmt_vec_info context, gimple *new_stmt)
704 : {
705 3305 : gimple_seq seq = NULL;
706 3305 : gimple_stmt_iterator gsi = gsi_start (seq);
707 3305 : gsi_insert_before_without_update (&gsi, new_stmt, GSI_SAME_STMT);
708 3305 : insert_seq_on_entry (context, seq);
709 3305 : }
710 :
711 : /* Create and initialize a new stmt_vec_info struct for STMT. */
712 :
713 : stmt_vec_info
714 61480252 : vec_info::new_stmt_vec_info (gimple *stmt)
715 : {
716 61480252 : stmt_vec_info res = XCNEW (class _stmt_vec_info);
717 61480252 : res->stmt = stmt;
718 :
719 61480252 : STMT_VINFO_RELEVANT (res) = vect_unused_in_scope;
720 61480252 : STMT_VINFO_VECTORIZABLE (res) = true;
721 61480252 : STMT_VINFO_REDUC_TYPE (res) = TREE_CODE_REDUCTION;
722 61480252 : STMT_VINFO_REDUC_CODE (res) = ERROR_MARK;
723 61480252 : STMT_VINFO_REDUC_IDX (res) = -1;
724 61480252 : STMT_VINFO_REDUC_DEF (res) = NULL;
725 61480252 : STMT_VINFO_SLP_VECT_ONLY (res) = false;
726 :
727 61480252 : if (is_a <loop_vec_info> (this)
728 8552402 : && gimple_code (stmt) == GIMPLE_PHI
729 62996058 : && is_loop_header_bb_p (gimple_bb (stmt)))
730 1505126 : STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
731 : else
732 59975126 : STMT_VINFO_DEF_TYPE (res) = vect_internal_def;
733 :
734 61480252 : STMT_SLP_TYPE (res) = not_vect;
735 :
736 : /* This is really "uninitialized" until vect_compute_data_ref_alignment. */
737 61480252 : res->dr_aux.misalignment = DR_MISALIGNMENT_UNINITIALIZED;
738 :
739 61480252 : return res;
740 : }
741 :
742 : /* Associate STMT with INFO. */
743 :
744 : void
745 62969838 : vec_info::set_vinfo_for_stmt (gimple *stmt, stmt_vec_info info, bool check_ro)
746 : {
747 62969838 : unsigned int uid = gimple_uid (stmt);
748 62969838 : if (uid == 0)
749 : {
750 61480252 : gcc_assert (!check_ro || !stmt_vec_info_ro);
751 61480252 : gcc_checking_assert (info);
752 61480252 : uid = stmt_vec_infos.length () + 1;
753 61480252 : gimple_set_uid (stmt, uid);
754 61480252 : stmt_vec_infos.safe_push (info);
755 : }
756 : else
757 : {
758 1489586 : gcc_checking_assert (info == NULL);
759 1489586 : stmt_vec_infos[uid - 1] = info;
760 : }
761 62969838 : }
762 :
763 : /* Free the contents of stmt_vec_infos. */
764 :
765 : void
766 2763761 : vec_info::free_stmt_vec_infos (void)
767 : {
768 69771535 : for (stmt_vec_info &info : stmt_vec_infos)
769 61480252 : if (info != NULL)
770 59990666 : free_stmt_vec_info (info);
771 2763761 : stmt_vec_infos.release ();
772 2763761 : }
773 :
774 : /* Free STMT_INFO. */
775 :
776 : void
777 61480252 : vec_info::free_stmt_vec_info (stmt_vec_info stmt_info)
778 : {
779 61480252 : if (stmt_info->pattern_stmt_p)
780 : {
781 2372118 : gimple_set_bb (stmt_info->stmt, NULL);
782 2372118 : tree lhs = gimple_get_lhs (stmt_info->stmt);
783 2372118 : if (lhs && TREE_CODE (lhs) == SSA_NAME)
784 2008464 : release_ssa_name (lhs);
785 : }
786 :
787 61480252 : free (stmt_info);
788 61480252 : }
789 :
790 : /* Returns true if S1 dominates S2. */
791 :
792 : bool
793 549136 : vect_stmt_dominates_stmt_p (gimple *s1, gimple *s2)
794 : {
795 549136 : basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
796 :
797 : /* If bb1 is NULL, it should be a GIMPLE_NOP def stmt of an (D)
798 : SSA_NAME. Assume it lives at the beginning of function and
799 : thus dominates everything. */
800 549136 : if (!bb1 || s1 == s2)
801 : return true;
802 :
803 : /* If bb2 is NULL, it doesn't dominate any stmt with a bb. */
804 547113 : if (!bb2)
805 : return false;
806 :
807 547113 : if (bb1 != bb2)
808 188747 : return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
809 :
810 : /* PHIs in the same basic block are assumed to be
811 : executed all in parallel, if only one stmt is a PHI,
812 : it dominates the other stmt in the same basic block. */
813 358366 : if (gimple_code (s1) == GIMPLE_PHI)
814 : return true;
815 :
816 322646 : if (gimple_code (s2) == GIMPLE_PHI)
817 : return false;
818 :
819 : /* Inserted vectorized stmts all have UID 0 while the original stmts
820 : in the IL have UID increasing within a BB. Walk from both sides
821 : until we find the other stmt or a stmt with UID != 0. */
822 305014 : gimple_stmt_iterator gsi1 = gsi_for_stmt (s1);
823 775064 : while (gimple_uid (gsi_stmt (gsi1)) == 0)
824 : {
825 587222 : gsi_next (&gsi1);
826 587222 : if (gsi_end_p (gsi1))
827 : return false;
828 585506 : if (gsi_stmt (gsi1) == s2)
829 : return true;
830 : }
831 187842 : if (gimple_uid (gsi_stmt (gsi1)) == -1u)
832 : return false;
833 :
834 187842 : gimple_stmt_iterator gsi2 = gsi_for_stmt (s2);
835 786120 : while (gimple_uid (gsi_stmt (gsi2)) == 0)
836 : {
837 610301 : gsi_prev (&gsi2);
838 610301 : if (gsi_end_p (gsi2))
839 : return false;
840 598286 : if (gsi_stmt (gsi2) == s1)
841 : return true;
842 : }
843 175819 : if (gimple_uid (gsi_stmt (gsi2)) == -1u)
844 : return false;
845 :
846 175819 : if (gimple_uid (gsi_stmt (gsi1)) <= gimple_uid (gsi_stmt (gsi2)))
847 : return true;
848 : return false;
849 : }
850 :
851 : /* A helper function to free scev and LOOP niter information, as well as
852 : clear loop constraint LOOP_C_FINITE. */
853 :
854 : void
855 43998 : vect_free_loop_info_assumptions (class loop *loop)
856 : {
857 43998 : scev_reset_htab ();
858 : /* We need to explicitly reset upper bound information since they are
859 : used even after free_numbers_of_iterations_estimates. */
860 43998 : loop->any_upper_bound = false;
861 43998 : loop->any_likely_upper_bound = false;
862 43998 : free_numbers_of_iterations_estimates (loop);
863 43998 : loop_constraint_clear (loop, LOOP_C_FINITE);
864 43998 : }
865 :
866 : /* If LOOP has been versioned during ifcvt, return the internal call
867 : guarding it. */
868 :
869 : gimple *
870 517653 : vect_loop_vectorized_call (class loop *loop, gcond **cond)
871 : {
872 517653 : basic_block bb = loop_preheader_edge (loop)->src;
873 956644 : gimple *g;
874 1395635 : do
875 : {
876 956644 : g = *gsi_last_bb (bb);
877 607076 : if ((g && gimple_code (g) == GIMPLE_COND)
878 2051927 : || !single_succ_p (bb))
879 : break;
880 577630 : if (!single_pred_p (bb))
881 : break;
882 438991 : bb = single_pred (bb);
883 : }
884 : while (1);
885 517653 : if (g && gimple_code (g) == GIMPLE_COND)
886 : {
887 372699 : if (cond)
888 0 : *cond = as_a <gcond *> (g);
889 372699 : gimple_stmt_iterator gsi = gsi_for_stmt (g);
890 372699 : gsi_prev (&gsi);
891 372699 : if (!gsi_end_p (gsi))
892 : {
893 341536 : g = gsi_stmt (gsi);
894 341536 : if (gimple_call_internal_p (g, IFN_LOOP_VECTORIZED)
895 341536 : && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->num
896 30010 : || tree_to_shwi (gimple_call_arg (g, 1)) == loop->num))
897 60295 : return g;
898 : }
899 : }
900 : return NULL;
901 : }
902 :
903 : /* If LOOP has been versioned during loop distribution, return the gurading
904 : internal call. */
905 :
906 : static gimple *
907 481048 : vect_loop_dist_alias_call (class loop *loop, function *fun)
908 : {
909 481048 : basic_block bb;
910 481048 : basic_block entry;
911 481048 : class loop *outer, *orig;
912 :
913 481048 : if (loop->orig_loop_num == 0)
914 : return NULL;
915 :
916 162 : orig = get_loop (fun, loop->orig_loop_num);
917 162 : if (orig == NULL)
918 : {
919 : /* The original loop is somehow destroyed. Clear the information. */
920 0 : loop->orig_loop_num = 0;
921 0 : return NULL;
922 : }
923 :
924 162 : if (loop != orig)
925 91 : bb = nearest_common_dominator (CDI_DOMINATORS, loop->header, orig->header);
926 : else
927 71 : bb = loop_preheader_edge (loop)->src;
928 :
929 162 : outer = bb->loop_father;
930 162 : entry = ENTRY_BLOCK_PTR_FOR_FN (fun);
931 :
932 : /* Look upward in dominance tree. */
933 743 : for (; bb != entry && flow_bb_inside_loop_p (outer, bb);
934 581 : bb = get_immediate_dominator (CDI_DOMINATORS, bb))
935 : {
936 691 : gimple_stmt_iterator gsi = gsi_last_bb (bb);
937 691 : if (!safe_is_a <gcond *> (*gsi))
938 581 : continue;
939 :
940 508 : gsi_prev (&gsi);
941 508 : if (gsi_end_p (gsi))
942 8 : continue;
943 :
944 500 : gimple *g = gsi_stmt (gsi);
945 : /* The guarding internal function call must have the same distribution
946 : alias id. */
947 500 : if (gimple_call_internal_p (g, IFN_LOOP_DIST_ALIAS)
948 500 : && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->orig_loop_num))
949 481048 : return g;
950 : }
951 : return NULL;
952 : }
953 :
954 : /* Set the uids of all the statements in basic blocks inside loop
955 : represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
956 : call guarding the loop which has been if converted. */
957 : static void
958 7780 : set_uid_loop_bbs (loop_vec_info loop_vinfo, gimple *loop_vectorized_call,
959 : function *fun)
960 : {
961 7780 : tree arg = gimple_call_arg (loop_vectorized_call, 1);
962 7780 : basic_block *bbs;
963 7780 : unsigned int i;
964 7780 : class loop *scalar_loop = get_loop (fun, tree_to_shwi (arg));
965 :
966 7780 : LOOP_VINFO_SCALAR_LOOP (loop_vinfo) = scalar_loop;
967 7780 : LOOP_VINFO_SCALAR_MAIN_EXIT (loop_vinfo)
968 7780 : = vec_init_loop_exit_info (scalar_loop);
969 7780 : gcc_checking_assert (vect_loop_vectorized_call (scalar_loop)
970 : == loop_vectorized_call);
971 : /* If we are going to vectorize outer loop, prevent vectorization
972 : of the inner loop in the scalar loop - either the scalar loop is
973 : thrown away, so it is a wasted work, or is used only for
974 : a few iterations. */
975 7780 : if (scalar_loop->inner)
976 : {
977 120 : gimple *g = vect_loop_vectorized_call (scalar_loop->inner);
978 120 : if (g)
979 : {
980 120 : arg = gimple_call_arg (g, 0);
981 120 : get_loop (fun, tree_to_shwi (arg))->dont_vectorize = true;
982 120 : fold_loop_internal_call (g, boolean_false_node);
983 : }
984 : }
985 7780 : bbs = get_loop_body (scalar_loop);
986 42427 : for (i = 0; i < scalar_loop->num_nodes; i++)
987 : {
988 34647 : basic_block bb = bbs[i];
989 34647 : gimple_stmt_iterator gsi;
990 66597 : for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
991 : {
992 31950 : gimple *phi = gsi_stmt (gsi);
993 31950 : gimple_set_uid (phi, 0);
994 : }
995 175231 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
996 : {
997 105937 : gimple *stmt = gsi_stmt (gsi);
998 105937 : gimple_set_uid (stmt, 0);
999 : }
1000 : }
1001 7780 : free (bbs);
1002 7780 : }
1003 :
1004 : /* Generate vectorized code for LOOP and its epilogues. */
1005 :
1006 : static unsigned
1007 61412 : vect_transform_loops (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1008 : loop_p loop, gimple *loop_vectorized_call,
1009 : function *fun)
1010 : {
1011 61412 : loop_vec_info loop_vinfo = loop_vec_info_for_loop (loop);
1012 :
1013 61412 : if (loop_vectorized_call)
1014 7780 : set_uid_loop_bbs (loop_vinfo, loop_vectorized_call, fun);
1015 :
1016 61412 : unsigned HOST_WIDE_INT bytes;
1017 61412 : if (dump_enabled_p ())
1018 : {
1019 21930 : if (GET_MODE_SIZE (loop_vinfo->vector_mode).is_constant (&bytes))
1020 10965 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
1021 : "%sloop vectorized using %s%wu byte vectors and"
1022 : " unroll factor %u\n",
1023 10965 : LOOP_VINFO_EPILOGUE_P (loop_vinfo)
1024 : ? "epilogue " : "",
1025 10965 : LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)
1026 : ? "masked " : "", bytes,
1027 : (unsigned int) LOOP_VINFO_VECT_FACTOR
1028 10965 : (loop_vinfo).to_constant ());
1029 : else
1030 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
1031 : "%sloop vectorized using variable length vectors\n",
1032 : LOOP_VINFO_EPILOGUE_P (loop_vinfo)
1033 : ? "epilogue " : "");
1034 : }
1035 :
1036 61412 : loop_p new_loop = vect_transform_loop (loop_vinfo,
1037 : loop_vectorized_call);
1038 : /* Now that the loop has been vectorized, allow it to be unrolled
1039 : etc. */
1040 61412 : loop->force_vectorize = false;
1041 :
1042 61412 : if (loop->simduid)
1043 : {
1044 1895 : simduid_to_vf *simduid_to_vf_data = XNEW (simduid_to_vf);
1045 1895 : if (!simduid_to_vf_htab)
1046 1535 : simduid_to_vf_htab = new hash_table<simduid_to_vf> (15);
1047 1895 : simduid_to_vf_data->simduid = DECL_UID (loop->simduid);
1048 1895 : simduid_to_vf_data->vf = loop_vinfo->vectorization_factor;
1049 1895 : *simduid_to_vf_htab->find_slot (simduid_to_vf_data, INSERT)
1050 1895 : = simduid_to_vf_data;
1051 : }
1052 :
1053 : /* We should not have to update virtual SSA form here but some
1054 : transforms involve creating new virtual definitions which makes
1055 : updating difficult.
1056 : We delay the actual update to the end of the pass but avoid
1057 : confusing ourselves by forcing need_ssa_update_p () to false. */
1058 61412 : unsigned todo = 0;
1059 61412 : if (need_ssa_update_p (cfun))
1060 : {
1061 119 : gcc_assert (loop_vinfo->any_known_not_updated_vssa);
1062 119 : fun->gimple_df->ssa_renaming_needed = false;
1063 119 : todo |= TODO_update_ssa_only_virtuals;
1064 : }
1065 61412 : gcc_assert (!need_ssa_update_p (cfun));
1066 :
1067 : /* Epilogue of vectorized loop must be vectorized too. */
1068 61412 : if (new_loop)
1069 6824 : todo |= vect_transform_loops (simduid_to_vf_htab, new_loop, NULL, fun);
1070 :
1071 61412 : return todo;
1072 : }
1073 :
1074 : /* Try to vectorize LOOP. */
1075 :
1076 : static unsigned
1077 464011 : try_vectorize_loop_1 (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1078 : unsigned *num_vectorized_loops, loop_p loop,
1079 : gimple *loop_vectorized_call,
1080 : gimple *loop_dist_alias_call,
1081 : function *fun)
1082 : {
1083 464011 : unsigned ret = 0;
1084 464011 : vec_info_shared shared;
1085 464011 : auto_purge_vect_location sentinel;
1086 464011 : vect_location = find_loop_location (loop);
1087 :
1088 464011 : if (LOCATION_LOCUS (vect_location.get_location_t ()) != UNKNOWN_LOCATION
1089 464011 : && dump_enabled_p ())
1090 14989 : dump_printf (MSG_NOTE | MSG_PRIORITY_INTERNALS,
1091 : "\nAnalyzing loop at %s:%d\n",
1092 14989 : LOCATION_FILE (vect_location.get_location_t ()),
1093 29978 : LOCATION_LINE (vect_location.get_location_t ()));
1094 :
1095 : /* Try to analyze the loop, retaining an opt_problem if dump_enabled_p. */
1096 464011 : opt_loop_vec_info loop_vinfo = vect_analyze_loop (loop, loop_vectorized_call,
1097 : &shared);
1098 464011 : loop->aux = loop_vinfo;
1099 :
1100 464011 : if (!loop_vinfo)
1101 409419 : if (dump_enabled_p ())
1102 5789 : if (opt_problem *problem = loop_vinfo.get_problem ())
1103 : {
1104 5789 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1105 : "couldn't vectorize loop\n");
1106 5789 : problem->emit_and_clear ();
1107 : }
1108 :
1109 464011 : if (!loop_vinfo || !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo))
1110 : {
1111 : /* Free existing information if loop is analyzed with some
1112 : assumptions. */
1113 409419 : if (loop_constraint_set_p (loop, LOOP_C_FINITE))
1114 7876 : vect_free_loop_info_assumptions (loop);
1115 :
1116 : /* If we applied if-conversion then try to vectorize the
1117 : BB of innermost loops.
1118 : ??? Ideally BB vectorization would learn to vectorize
1119 : control flow by applying if-conversion on-the-fly, the
1120 : following retains the if-converted loop body even when
1121 : only non-if-converted parts took part in BB vectorization. */
1122 409419 : if (flag_tree_slp_vectorize != 0
1123 408393 : && loop_vectorized_call
1124 20841 : && ! loop->inner)
1125 : {
1126 20040 : basic_block bb = loop->header;
1127 20040 : bool require_loop_vectorize = false;
1128 40080 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
1129 688497 : !gsi_end_p (gsi); gsi_next (&gsi))
1130 : {
1131 669278 : gimple *stmt = gsi_stmt (gsi);
1132 669278 : gcall *call = dyn_cast <gcall *> (stmt);
1133 1243 : if (call && gimple_call_internal_p (call))
1134 : {
1135 1146 : internal_fn ifn = gimple_call_internal_fn (call);
1136 1146 : if (ifn == IFN_MASK_LOAD
1137 1146 : || ifn == IFN_MASK_STORE
1138 728 : || ifn == IFN_MASK_CALL
1139 : /* Don't keep the if-converted parts when the ifn with
1140 : specific type is not supported by the backend. */
1141 1855 : || (direct_internal_fn_p (ifn)
1142 384 : && !direct_internal_fn_supported_p
1143 384 : (call, OPTIMIZE_FOR_SPEED)))
1144 : {
1145 : require_loop_vectorize = true;
1146 : break;
1147 : }
1148 : }
1149 668457 : gimple_set_uid (stmt, -1);
1150 668457 : gimple_set_visited (stmt, false);
1151 : }
1152 20040 : if (!require_loop_vectorize)
1153 : {
1154 19219 : tree arg = gimple_call_arg (loop_vectorized_call, 1);
1155 19219 : class loop *scalar_loop = get_loop (fun, tree_to_shwi (arg));
1156 19219 : if (vect_slp_if_converted_bb (bb, scalar_loop))
1157 : {
1158 75 : fold_loop_internal_call (loop_vectorized_call,
1159 : boolean_true_node);
1160 75 : loop_vectorized_call = NULL;
1161 75 : ret |= TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1162 : }
1163 : }
1164 : }
1165 : /* If outer loop vectorization fails for LOOP_VECTORIZED guarded
1166 : loop, don't vectorize its inner loop; we'll attempt to
1167 : vectorize LOOP_VECTORIZED guarded inner loop of the scalar
1168 : loop version. */
1169 21867 : if (loop_vectorized_call && loop->inner)
1170 803 : loop->inner->dont_vectorize = true;
1171 409419 : return ret;
1172 : }
1173 :
1174 54592 : if (!dbg_cnt (vect_loop))
1175 : {
1176 : /* Free existing information if loop is analyzed with some
1177 : assumptions. */
1178 4 : if (loop_constraint_set_p (loop, LOOP_C_FINITE))
1179 0 : vect_free_loop_info_assumptions (loop);
1180 4 : return ret;
1181 : }
1182 :
1183 54588 : (*num_vectorized_loops)++;
1184 : /* Transform LOOP and its epilogues. */
1185 54588 : ret |= vect_transform_loops (simduid_to_vf_htab, loop,
1186 : loop_vectorized_call, fun);
1187 :
1188 54588 : if (loop_vectorized_call)
1189 : {
1190 7780 : fold_loop_internal_call (loop_vectorized_call, boolean_true_node);
1191 7780 : ret |= TODO_cleanup_cfg;
1192 : }
1193 54588 : if (loop_dist_alias_call)
1194 : {
1195 8 : tree value = gimple_call_arg (loop_dist_alias_call, 1);
1196 8 : fold_loop_internal_call (loop_dist_alias_call, value);
1197 8 : ret |= TODO_cleanup_cfg;
1198 : }
1199 :
1200 : return ret;
1201 464011 : }
1202 :
1203 : /* Try to vectorize LOOP. */
1204 :
1205 : static unsigned
1206 495356 : try_vectorize_loop (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
1207 : unsigned *num_vectorized_loops, loop_p loop,
1208 : function *fun)
1209 : {
1210 495356 : if (!((flag_tree_loop_vectorize
1211 490774 : && optimize_loop_nest_for_speed_p (loop))
1212 33148 : || loop->force_vectorize))
1213 : return 0;
1214 :
1215 464011 : return try_vectorize_loop_1 (simduid_to_vf_htab, num_vectorized_loops, loop,
1216 : vect_loop_vectorized_call (loop),
1217 464011 : vect_loop_dist_alias_call (loop, fun), fun);
1218 : }
1219 :
1220 :
1221 : /* Loop autovectorization. */
1222 :
1223 : namespace {
1224 :
1225 : const pass_data pass_data_vectorize =
1226 : {
1227 : GIMPLE_PASS, /* type */
1228 : "vect", /* name */
1229 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1230 : TV_TREE_VECTORIZATION, /* tv_id */
1231 : ( PROP_cfg | PROP_ssa ), /* properties_required */
1232 : 0, /* properties_provided */
1233 : 0, /* properties_destroyed */
1234 : 0, /* todo_flags_start */
1235 : 0, /* todo_flags_finish */
1236 : };
1237 :
1238 : class pass_vectorize : public gimple_opt_pass
1239 : {
1240 : public:
1241 288047 : pass_vectorize (gcc::context *ctxt)
1242 576094 : : gimple_opt_pass (pass_data_vectorize, ctxt)
1243 : {}
1244 :
1245 : /* opt_pass methods: */
1246 240015 : bool gate (function *fun) final override
1247 : {
1248 240015 : return flag_tree_loop_vectorize || fun->has_force_vectorize_loops;
1249 : }
1250 :
1251 : unsigned int execute (function *) final override;
1252 :
1253 : }; // class pass_vectorize
1254 :
1255 : /* Function vectorize_loops.
1256 :
1257 : Entry point to loop vectorization phase. */
1258 :
1259 : unsigned
1260 206946 : pass_vectorize::execute (function *fun)
1261 : {
1262 206946 : unsigned int i;
1263 206946 : unsigned int num_vectorized_loops = 0;
1264 206946 : unsigned int vect_loops_num;
1265 206946 : hash_table<simduid_to_vf> *simduid_to_vf_htab = NULL;
1266 206946 : hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
1267 206946 : bool any_ifcvt_loops = false;
1268 206946 : unsigned ret = 0;
1269 :
1270 206946 : vect_loops_num = number_of_loops (fun);
1271 :
1272 : /* Bail out if there are no loops. */
1273 206946 : if (vect_loops_num <= 1)
1274 : return 0;
1275 :
1276 206946 : vect_slp_init ();
1277 :
1278 206946 : if (fun->has_simduid_loops)
1279 5601 : note_simd_array_uses (&simd_array_to_simduid_htab, fun);
1280 :
1281 : /* ----------- Analyze loops. ----------- */
1282 206946 : enable_ranger (fun);
1283 :
1284 : /* If some loop was duplicated, it gets bigger number
1285 : than all previously defined loops. This fact allows us to run
1286 : only over initial loops skipping newly generated ones. */
1287 1150800 : for (auto loop : loops_list (fun, 0))
1288 529962 : if (loop->dont_vectorize)
1289 : {
1290 35529 : any_ifcvt_loops = true;
1291 : /* If-conversion sometimes versions both the outer loop
1292 : (for the case when outer loop vectorization might be
1293 : desirable) as well as the inner loop in the scalar version
1294 : of the loop. So we have:
1295 : if (LOOP_VECTORIZED (1, 3))
1296 : {
1297 : loop1
1298 : loop2
1299 : }
1300 : else
1301 : loop3 (copy of loop1)
1302 : if (LOOP_VECTORIZED (4, 5))
1303 : loop4 (copy of loop2)
1304 : else
1305 : loop5 (copy of loop4)
1306 : If loops' iteration gives us loop3 first (which has
1307 : dont_vectorize set), make sure to process loop1 before loop4;
1308 : so that we can prevent vectorization of loop4 if loop1
1309 : is successfully vectorized. */
1310 35529 : if (loop->inner)
1311 : {
1312 2168 : gimple *loop_vectorized_call
1313 2168 : = vect_loop_vectorized_call (loop);
1314 2168 : if (loop_vectorized_call
1315 2168 : && vect_loop_vectorized_call (loop->inner))
1316 : {
1317 923 : tree arg = gimple_call_arg (loop_vectorized_call, 0);
1318 923 : class loop *vector_loop
1319 923 : = get_loop (fun, tree_to_shwi (arg));
1320 923 : if (vector_loop && vector_loop != loop)
1321 : {
1322 : /* Make sure we don't vectorize it twice. */
1323 923 : vector_loop->dont_vectorize = true;
1324 923 : ret |= try_vectorize_loop (simduid_to_vf_htab,
1325 : &num_vectorized_loops,
1326 : vector_loop, fun);
1327 : }
1328 : }
1329 : }
1330 : }
1331 : else
1332 494433 : ret |= try_vectorize_loop (simduid_to_vf_htab, &num_vectorized_loops,
1333 206946 : loop, fun);
1334 :
1335 206946 : vect_location = dump_user_location_t ();
1336 :
1337 206946 : statistics_counter_event (fun, "Vectorized loops", num_vectorized_loops);
1338 206946 : if (dump_enabled_p ()
1339 206946 : || (num_vectorized_loops > 0 && dump_enabled_p ()))
1340 11624 : dump_printf_loc (MSG_NOTE, vect_location,
1341 : "vectorized %u loops in function.\n",
1342 : num_vectorized_loops);
1343 :
1344 : /* ----------- Finalize. ----------- */
1345 206946 : disable_ranger (fun);
1346 :
1347 206946 : if (any_ifcvt_loops)
1348 321090 : for (i = 1; i < number_of_loops (fun); i++)
1349 : {
1350 138919 : class loop *loop = get_loop (fun, i);
1351 138919 : if (loop && loop->dont_vectorize)
1352 : {
1353 38104 : gimple *g = vect_loop_vectorized_call (loop);
1354 38104 : if (g)
1355 : {
1356 21067 : fold_loop_internal_call (g, boolean_false_node);
1357 21067 : loop->dont_vectorize = false;
1358 21067 : ret |= TODO_cleanup_cfg;
1359 21067 : g = NULL;
1360 : }
1361 : else
1362 17037 : g = vect_loop_dist_alias_call (loop, fun);
1363 :
1364 38104 : if (g)
1365 : {
1366 28 : fold_loop_internal_call (g, boolean_false_node);
1367 28 : loop->dont_vectorize = false;
1368 28 : ret |= TODO_cleanup_cfg;
1369 : }
1370 : }
1371 : }
1372 :
1373 : /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1374 206946 : if (fun->has_simduid_loops)
1375 : {
1376 5601 : adjust_simduid_builtins (simduid_to_vf_htab, fun);
1377 : /* Avoid stale SCEV cache entries for the SIMD_LANE defs. */
1378 5601 : scev_reset ();
1379 : }
1380 : /* Shrink any "omp array simd" temporary arrays to the
1381 : actual vectorization factors. */
1382 206946 : if (simd_array_to_simduid_htab)
1383 2204 : shrink_simd_arrays (simd_array_to_simduid_htab, simduid_to_vf_htab);
1384 206946 : delete simduid_to_vf_htab;
1385 206946 : fun->has_simduid_loops = false;
1386 :
1387 206946 : if (num_vectorized_loops > 0)
1388 : {
1389 : /* We are collecting some corner cases where we need to update
1390 : virtual SSA form via the TODO but delete the queued update-SSA
1391 : state. Force renaming if we think that might be necessary. */
1392 37010 : if (ret & TODO_update_ssa_only_virtuals)
1393 89 : mark_virtual_operands_for_renaming (cfun);
1394 : /* If we vectorized any loop only virtual SSA form needs to be updated.
1395 : ??? Also while we try hard to update loop-closed SSA form we fail
1396 : to properly do this in some corner-cases (see PR56286). */
1397 37010 : rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_only_virtuals);
1398 37010 : ret |= TODO_cleanup_cfg;
1399 : }
1400 :
1401 1849488 : for (i = 1; i < number_of_loops (fun); i++)
1402 : {
1403 717798 : loop_vec_info loop_vinfo;
1404 717798 : bool has_mask_store;
1405 :
1406 717798 : class loop *loop = get_loop (fun, i);
1407 717798 : if (!loop || !loop->aux)
1408 656382 : continue;
1409 61416 : loop_vinfo = (loop_vec_info) loop->aux;
1410 61416 : has_mask_store = LOOP_VINFO_HAS_MASK_STORE (loop_vinfo);
1411 61416 : delete loop_vinfo;
1412 61416 : if (has_mask_store
1413 61416 : && targetm.vectorize.empty_mask_is_expensive (IFN_MASK_STORE))
1414 493 : optimize_mask_stores (loop);
1415 :
1416 61416 : auto_bitmap exit_bbs;
1417 : /* Perform local CSE, this esp. helps because we emit code for
1418 : predicates that need to be shared for optimal predicate usage.
1419 : However reassoc will re-order them and prevent CSE from working
1420 : as it should. CSE only the loop body, not the entry. */
1421 61416 : auto_vec<edge> exits = get_loop_exit_edges (loop);
1422 247185 : for (edge exit : exits)
1423 62937 : bitmap_set_bit (exit_bbs, exit->dest->index);
1424 :
1425 61416 : edge entry = EDGE_PRED (loop_preheader_edge (loop)->src, 0);
1426 61416 : do_rpo_vn (fun, entry, exit_bbs);
1427 :
1428 61416 : loop->aux = NULL;
1429 61416 : }
1430 :
1431 206946 : vect_slp_fini ();
1432 :
1433 206946 : return ret;
1434 : }
1435 :
1436 : } // anon namespace
1437 :
1438 : gimple_opt_pass *
1439 288047 : make_pass_vectorize (gcc::context *ctxt)
1440 : {
1441 288047 : return new pass_vectorize (ctxt);
1442 : }
1443 :
1444 : /* Entry point to the simduid cleanup pass. */
1445 :
1446 : namespace {
1447 :
1448 : const pass_data pass_data_simduid_cleanup =
1449 : {
1450 : GIMPLE_PASS, /* type */
1451 : "simduid", /* name */
1452 : OPTGROUP_NONE, /* optinfo_flags */
1453 : TV_NONE, /* tv_id */
1454 : ( PROP_ssa | PROP_cfg ), /* properties_required */
1455 : 0, /* properties_provided */
1456 : 0, /* properties_destroyed */
1457 : 0, /* todo_flags_start */
1458 : 0, /* todo_flags_finish */
1459 : };
1460 :
1461 : class pass_simduid_cleanup : public gimple_opt_pass
1462 : {
1463 : public:
1464 576094 : pass_simduid_cleanup (gcc::context *ctxt)
1465 1152188 : : gimple_opt_pass (pass_data_simduid_cleanup, ctxt)
1466 : {}
1467 :
1468 : /* opt_pass methods: */
1469 288047 : opt_pass * clone () final override
1470 : {
1471 288047 : return new pass_simduid_cleanup (m_ctxt);
1472 : }
1473 2514741 : bool gate (function *fun) final override { return fun->has_simduid_loops; }
1474 : unsigned int execute (function *) final override;
1475 :
1476 : }; // class pass_simduid_cleanup
1477 :
1478 : unsigned int
1479 2208 : pass_simduid_cleanup::execute (function *fun)
1480 : {
1481 2208 : hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
1482 :
1483 2208 : note_simd_array_uses (&simd_array_to_simduid_htab, fun);
1484 :
1485 : /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1486 2208 : adjust_simduid_builtins (NULL, fun);
1487 :
1488 : /* Shrink any "omp array simd" temporary arrays to the
1489 : actual vectorization factors. */
1490 2208 : if (simd_array_to_simduid_htab)
1491 4 : shrink_simd_arrays (simd_array_to_simduid_htab, NULL);
1492 2208 : fun->has_simduid_loops = false;
1493 2208 : return 0;
1494 : }
1495 :
1496 : } // anon namespace
1497 :
1498 : gimple_opt_pass *
1499 288047 : make_pass_simduid_cleanup (gcc::context *ctxt)
1500 : {
1501 288047 : return new pass_simduid_cleanup (ctxt);
1502 : }
1503 :
1504 :
1505 : /* Entry point to basic block SLP phase. */
1506 :
1507 : namespace {
1508 :
1509 : const pass_data pass_data_slp_vectorize =
1510 : {
1511 : GIMPLE_PASS, /* type */
1512 : "slp", /* name */
1513 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1514 : TV_TREE_SLP_VECTORIZATION, /* tv_id */
1515 : ( PROP_ssa | PROP_cfg ), /* properties_required */
1516 : 0, /* properties_provided */
1517 : 0, /* properties_destroyed */
1518 : 0, /* todo_flags_start */
1519 : TODO_update_ssa, /* todo_flags_finish */
1520 : };
1521 :
1522 : class pass_slp_vectorize : public gimple_opt_pass
1523 : {
1524 : public:
1525 576094 : pass_slp_vectorize (gcc::context *ctxt)
1526 1152188 : : gimple_opt_pass (pass_data_slp_vectorize, ctxt)
1527 : {}
1528 :
1529 : /* opt_pass methods: */
1530 288047 : opt_pass * clone () final override { return new pass_slp_vectorize (m_ctxt); }
1531 1039544 : bool gate (function *) final override { return flag_tree_slp_vectorize != 0; }
1532 : unsigned int execute (function *) final override;
1533 :
1534 : }; // class pass_slp_vectorize
1535 :
1536 : unsigned int
1537 906490 : pass_slp_vectorize::execute (function *fun)
1538 : {
1539 906490 : auto_purge_vect_location sentinel;
1540 906490 : basic_block bb;
1541 :
1542 906490 : bool in_loop_pipeline = scev_initialized_p ();
1543 906490 : if (!in_loop_pipeline)
1544 : {
1545 701175 : loop_optimizer_init (LOOPS_NORMAL);
1546 701175 : scev_initialize ();
1547 : }
1548 :
1549 : /* Mark all stmts as not belonging to the current region and unvisited. */
1550 11439017 : FOR_EACH_BB_FN (bb, fun)
1551 : {
1552 15210910 : for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
1553 4678383 : gsi_next (&gsi))
1554 : {
1555 4678383 : gphi *stmt = gsi.phi ();
1556 4678383 : gimple_set_uid (stmt, -1);
1557 4678383 : gimple_set_visited (stmt, false);
1558 : }
1559 99112504 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1560 78047450 : gsi_next (&gsi))
1561 : {
1562 78047450 : gimple *stmt = gsi_stmt (gsi);
1563 78047450 : gimple_set_uid (stmt, -1);
1564 78047450 : gimple_set_visited (stmt, false);
1565 : }
1566 : }
1567 :
1568 906490 : vect_slp_init ();
1569 :
1570 906490 : vect_slp_function (fun);
1571 :
1572 906490 : vect_slp_fini ();
1573 :
1574 906490 : if (!in_loop_pipeline)
1575 : {
1576 701175 : scev_finalize ();
1577 701175 : loop_optimizer_finalize ();
1578 : }
1579 :
1580 1812980 : return 0;
1581 906490 : }
1582 :
1583 : } // anon namespace
1584 :
1585 : gimple_opt_pass *
1586 288047 : make_pass_slp_vectorize (gcc::context *ctxt)
1587 : {
1588 288047 : return new pass_slp_vectorize (ctxt);
1589 : }
1590 :
1591 :
1592 : /* Increase alignment of global arrays to improve vectorization potential.
1593 : TODO:
1594 : - Consider also structs that have an array field.
1595 : - Use ipa analysis to prune arrays that can't be vectorized?
1596 : This should involve global alignment analysis and in the future also
1597 : array padding. */
1598 :
1599 : static unsigned get_vec_alignment_for_type (tree);
1600 : static hash_map<tree, unsigned> *type_align_map;
1601 :
1602 : /* Return alignment of array's vector type corresponding to scalar type.
1603 : 0 if no vector type exists. */
1604 : static unsigned
1605 0 : get_vec_alignment_for_array_type (tree type)
1606 : {
1607 0 : gcc_assert (TREE_CODE (type) == ARRAY_TYPE);
1608 0 : poly_uint64 array_size, vector_size;
1609 :
1610 0 : tree scalar_type = strip_array_types (type);
1611 0 : tree vectype = get_related_vectype_for_scalar_type (VOIDmode, scalar_type);
1612 0 : if (!vectype
1613 0 : || !poly_int_tree_p (TYPE_SIZE (type), &array_size)
1614 0 : || !poly_int_tree_p (TYPE_SIZE (vectype), &vector_size)
1615 0 : || maybe_lt (array_size, vector_size))
1616 0 : return 0;
1617 :
1618 0 : return TYPE_ALIGN (vectype);
1619 : }
1620 :
1621 : /* Return alignment of field having maximum alignment of vector type
1622 : corresponding to it's scalar type. For now, we only consider fields whose
1623 : offset is a multiple of it's vector alignment.
1624 : 0 if no suitable field is found. */
1625 : static unsigned
1626 0 : get_vec_alignment_for_record_type (tree type)
1627 : {
1628 0 : gcc_assert (TREE_CODE (type) == RECORD_TYPE);
1629 :
1630 0 : unsigned max_align = 0, alignment;
1631 0 : HOST_WIDE_INT offset;
1632 0 : tree offset_tree;
1633 :
1634 0 : if (TYPE_PACKED (type))
1635 : return 0;
1636 :
1637 0 : unsigned *slot = type_align_map->get (type);
1638 0 : if (slot)
1639 0 : return *slot;
1640 :
1641 0 : for (tree field = first_field (type);
1642 0 : field != NULL_TREE;
1643 0 : field = DECL_CHAIN (field))
1644 : {
1645 : /* Skip if not FIELD_DECL or if alignment is set by user. */
1646 0 : if (TREE_CODE (field) != FIELD_DECL
1647 0 : || DECL_USER_ALIGN (field)
1648 0 : || DECL_ARTIFICIAL (field))
1649 0 : continue;
1650 :
1651 : /* We don't need to process the type further if offset is variable,
1652 : since the offsets of remaining members will also be variable. */
1653 0 : if (TREE_CODE (DECL_FIELD_OFFSET (field)) != INTEGER_CST
1654 0 : || TREE_CODE (DECL_FIELD_BIT_OFFSET (field)) != INTEGER_CST)
1655 : break;
1656 :
1657 : /* Similarly stop processing the type if offset_tree
1658 : does not fit in unsigned HOST_WIDE_INT. */
1659 0 : offset_tree = bit_position (field);
1660 0 : if (!tree_fits_uhwi_p (offset_tree))
1661 : break;
1662 :
1663 0 : offset = tree_to_uhwi (offset_tree);
1664 0 : alignment = get_vec_alignment_for_type (TREE_TYPE (field));
1665 :
1666 : /* Get maximum alignment of vectorized field/array among those members
1667 : whose offset is multiple of the vector alignment. */
1668 0 : if (alignment
1669 0 : && (offset % alignment == 0)
1670 0 : && (alignment > max_align))
1671 0 : max_align = alignment;
1672 : }
1673 :
1674 0 : type_align_map->put (type, max_align);
1675 0 : return max_align;
1676 : }
1677 :
1678 : /* Return alignment of vector type corresponding to decl's scalar type
1679 : or 0 if it doesn't exist or the vector alignment is lesser than
1680 : decl's alignment. */
1681 : static unsigned
1682 0 : get_vec_alignment_for_type (tree type)
1683 : {
1684 0 : if (type == NULL_TREE)
1685 : return 0;
1686 :
1687 0 : gcc_assert (TYPE_P (type));
1688 :
1689 0 : static unsigned alignment = 0;
1690 0 : switch (TREE_CODE (type))
1691 : {
1692 0 : case ARRAY_TYPE:
1693 0 : alignment = get_vec_alignment_for_array_type (type);
1694 0 : break;
1695 0 : case RECORD_TYPE:
1696 0 : alignment = get_vec_alignment_for_record_type (type);
1697 0 : break;
1698 0 : default:
1699 0 : alignment = 0;
1700 0 : break;
1701 : }
1702 :
1703 0 : return (alignment > TYPE_ALIGN (type)) ? alignment : 0;
1704 : }
1705 :
1706 : /* Entry point to increase_alignment pass. */
1707 : static unsigned int
1708 0 : increase_alignment (void)
1709 : {
1710 0 : varpool_node *vnode;
1711 :
1712 0 : vect_location = dump_user_location_t ();
1713 0 : type_align_map = new hash_map<tree, unsigned>;
1714 :
1715 : /* Increase the alignment of all global arrays for vectorization. */
1716 0 : FOR_EACH_DEFINED_VARIABLE (vnode)
1717 : {
1718 0 : tree decl = vnode->decl;
1719 0 : unsigned int alignment;
1720 :
1721 0 : if ((decl_in_symtab_p (decl)
1722 0 : && !symtab_node::get (decl)->can_increase_alignment_p ())
1723 0 : || DECL_USER_ALIGN (decl) || DECL_ARTIFICIAL (decl))
1724 0 : continue;
1725 :
1726 0 : alignment = get_vec_alignment_for_type (TREE_TYPE (decl));
1727 0 : if (alignment && vect_can_force_dr_alignment_p (decl, alignment))
1728 : {
1729 0 : vnode->increase_alignment (alignment);
1730 0 : if (dump_enabled_p ())
1731 0 : dump_printf (MSG_NOTE, "Increasing alignment of decl: %T\n", decl);
1732 : }
1733 : }
1734 :
1735 0 : delete type_align_map;
1736 0 : return 0;
1737 : }
1738 :
1739 :
1740 : namespace {
1741 :
1742 : const pass_data pass_data_ipa_increase_alignment =
1743 : {
1744 : SIMPLE_IPA_PASS, /* type */
1745 : "increase_alignment", /* name */
1746 : OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1747 : TV_IPA_OPT, /* tv_id */
1748 : 0, /* properties_required */
1749 : 0, /* properties_provided */
1750 : 0, /* properties_destroyed */
1751 : 0, /* todo_flags_start */
1752 : 0, /* todo_flags_finish */
1753 : };
1754 :
1755 : class pass_ipa_increase_alignment : public simple_ipa_opt_pass
1756 : {
1757 : public:
1758 288047 : pass_ipa_increase_alignment (gcc::context *ctxt)
1759 576094 : : simple_ipa_opt_pass (pass_data_ipa_increase_alignment, ctxt)
1760 : {}
1761 :
1762 : /* opt_pass methods: */
1763 231936 : bool gate (function *) final override
1764 : {
1765 231936 : return flag_section_anchors && flag_tree_loop_vectorize;
1766 : }
1767 :
1768 0 : unsigned int execute (function *) final override
1769 : {
1770 0 : return increase_alignment ();
1771 : }
1772 :
1773 : }; // class pass_ipa_increase_alignment
1774 :
1775 : } // anon namespace
1776 :
1777 : simple_ipa_opt_pass *
1778 288047 : make_pass_ipa_increase_alignment (gcc::context *ctxt)
1779 : {
1780 288047 : return new pass_ipa_increase_alignment (ctxt);
1781 : }
1782 :
1783 : /* If the condition represented by T is a comparison or the SSA name
1784 : result of a comparison, extract the comparison's operands. Represent
1785 : T as NE_EXPR <T, 0> otherwise. */
1786 :
1787 : void
1788 63264 : scalar_cond_masked_key::get_cond_ops_from_tree (tree t)
1789 : {
1790 63264 : if (TREE_CODE_CLASS (TREE_CODE (t)) == tcc_comparison)
1791 : {
1792 0 : this->code = TREE_CODE (t);
1793 0 : this->op0 = TREE_OPERAND (t, 0);
1794 0 : this->op1 = TREE_OPERAND (t, 1);
1795 0 : this->inverted_p = false;
1796 0 : return;
1797 : }
1798 :
1799 63264 : if (TREE_CODE (t) == SSA_NAME)
1800 26063 : if (gassign *stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (t)))
1801 : {
1802 26063 : tree_code code = gimple_assign_rhs_code (stmt);
1803 26063 : if (TREE_CODE_CLASS (code) == tcc_comparison)
1804 : {
1805 17666 : this->code = code;
1806 17666 : this->op0 = gimple_assign_rhs1 (stmt);
1807 17666 : this->op1 = gimple_assign_rhs2 (stmt);
1808 17666 : this->inverted_p = false;
1809 17666 : return;
1810 : }
1811 8397 : else if (code == BIT_NOT_EXPR)
1812 : {
1813 3565 : tree n_op = gimple_assign_rhs1 (stmt);
1814 3565 : if ((stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (n_op))))
1815 : {
1816 3565 : code = gimple_assign_rhs_code (stmt);
1817 3565 : if (TREE_CODE_CLASS (code) == tcc_comparison)
1818 : {
1819 3531 : this->code = code;
1820 3531 : this->op0 = gimple_assign_rhs1 (stmt);
1821 3531 : this->op1 = gimple_assign_rhs2 (stmt);
1822 3531 : this->inverted_p = true;
1823 3531 : return;
1824 : }
1825 : }
1826 : }
1827 : }
1828 :
1829 42067 : this->code = NE_EXPR;
1830 42067 : this->op0 = t;
1831 42067 : this->op1 = build_zero_cst (TREE_TYPE (t));
1832 42067 : this->inverted_p = false;
1833 : }
1834 :
1835 : /* See the comment above the declaration for details. */
1836 :
1837 : unsigned int
1838 0 : vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
1839 : stmt_vec_info stmt_info, slp_tree,
1840 : tree vectype, int misalign,
1841 : vect_cost_model_location where)
1842 : {
1843 0 : unsigned int cost
1844 0 : = builtin_vectorization_cost (kind, vectype, misalign) * count;
1845 0 : return record_stmt_cost (stmt_info, where, cost);
1846 : }
1847 :
1848 : /* See the comment above the declaration for details. */
1849 :
1850 : void
1851 1801339 : vector_costs::finish_cost (const vector_costs *)
1852 : {
1853 1801339 : gcc_assert (!m_finished);
1854 1801339 : m_finished = true;
1855 1801339 : }
1856 :
1857 : /* Record a base cost of COST units against WHERE. If STMT_INFO is
1858 : nonnull, use it to adjust the cost based on execution frequency
1859 : (where appropriate). */
1860 :
1861 : unsigned int
1862 0 : vector_costs::record_stmt_cost (stmt_vec_info stmt_info,
1863 : vect_cost_model_location where,
1864 : unsigned int cost)
1865 : {
1866 0 : cost = adjust_cost_for_freq (stmt_info, where, cost);
1867 0 : m_costs[where] += cost;
1868 0 : return cost;
1869 : }
1870 :
1871 : /* COST is the base cost we have calculated for an operation in location WHERE.
1872 : If STMT_INFO is nonnull, use it to adjust the cost based on execution
1873 : frequency (where appropriate). Return the adjusted cost. */
1874 :
1875 : unsigned int
1876 7447795 : vector_costs::adjust_cost_for_freq (stmt_vec_info stmt_info,
1877 : vect_cost_model_location where,
1878 : unsigned int cost)
1879 : {
1880 : /* Statements in an inner loop relative to the loop being
1881 : vectorized are weighted more heavily. The value here is
1882 : arbitrary and could potentially be improved with analysis. */
1883 7447795 : if (where == vect_body
1884 7447795 : && stmt_info
1885 7447795 : && stmt_in_inner_loop_p (m_vinfo, stmt_info))
1886 : {
1887 11870 : loop_vec_info loop_vinfo = as_a<loop_vec_info> (m_vinfo);
1888 11870 : cost *= LOOP_VINFO_INNER_LOOP_COST_FACTOR (loop_vinfo);
1889 : }
1890 7447795 : return cost;
1891 : }
1892 :
1893 : /* See the comment above the declaration for details. */
1894 :
1895 : bool
1896 30777 : vector_costs::better_main_loop_than_p (const vector_costs *other) const
1897 : {
1898 30777 : int diff = compare_inside_loop_cost (other);
1899 30777 : if (diff != 0)
1900 30690 : return diff < 0;
1901 :
1902 : /* If there's nothing to choose between the loop bodies, see whether
1903 : there's a difference in the prologue and epilogue costs. */
1904 87 : diff = compare_outside_loop_cost (other);
1905 87 : if (diff != 0)
1906 61 : return diff < 0;
1907 :
1908 : return false;
1909 : }
1910 :
1911 :
1912 : /* See the comment above the declaration for details. */
1913 :
1914 : bool
1915 1439 : vector_costs::better_epilogue_loop_than_p (const vector_costs *other,
1916 : loop_vec_info main_loop) const
1917 : {
1918 1439 : loop_vec_info this_loop_vinfo = as_a<loop_vec_info> (this->m_vinfo);
1919 1439 : loop_vec_info other_loop_vinfo = as_a<loop_vec_info> (other->m_vinfo);
1920 :
1921 1439 : poly_int64 this_vf = LOOP_VINFO_VECT_FACTOR (this_loop_vinfo);
1922 1439 : poly_int64 other_vf = LOOP_VINFO_VECT_FACTOR (other_loop_vinfo);
1923 :
1924 1439 : poly_uint64 main_poly_vf = LOOP_VINFO_VECT_FACTOR (main_loop);
1925 1439 : unsigned HOST_WIDE_INT main_vf;
1926 1439 : unsigned HOST_WIDE_INT other_factor, this_factor, other_cost, this_cost;
1927 : /* If we can determine how many iterations are left for the epilogue
1928 : loop, that is if both the main loop's vectorization factor and number
1929 : of iterations are constant, then we use them to calculate the cost of
1930 : the epilogue loop together with a 'likely value' for the epilogues
1931 : vectorization factor. Otherwise we use the main loop's vectorization
1932 : factor and the maximum poly value for the epilogue's. If the target
1933 : has not provided with a sensible upper bound poly vectorization
1934 : factors are likely to be favored over constant ones. */
1935 1439 : if (main_poly_vf.is_constant (&main_vf)
1936 1439 : && LOOP_VINFO_NITERS_KNOWN_P (main_loop))
1937 : {
1938 94 : unsigned HOST_WIDE_INT niters
1939 94 : = LOOP_VINFO_INT_NITERS (main_loop) % main_vf;
1940 94 : HOST_WIDE_INT other_likely_vf
1941 94 : = estimated_poly_value (other_vf, POLY_VALUE_LIKELY);
1942 94 : HOST_WIDE_INT this_likely_vf
1943 94 : = estimated_poly_value (this_vf, POLY_VALUE_LIKELY);
1944 :
1945 : /* If the epilogue is using partial vectors we account for the
1946 : partial iteration here too. */
1947 94 : other_factor = niters / other_likely_vf;
1948 94 : if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (other_loop_vinfo)
1949 0 : && niters % other_likely_vf != 0)
1950 0 : other_factor++;
1951 :
1952 94 : this_factor = niters / this_likely_vf;
1953 94 : if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (this_loop_vinfo)
1954 0 : && niters % this_likely_vf != 0)
1955 0 : this_factor++;
1956 : }
1957 : else
1958 : {
1959 1345 : unsigned HOST_WIDE_INT main_vf_max
1960 1345 : = estimated_poly_value (main_poly_vf, POLY_VALUE_MAX);
1961 1345 : unsigned HOST_WIDE_INT other_vf_max
1962 1345 : = estimated_poly_value (other_vf, POLY_VALUE_MAX);
1963 1345 : unsigned HOST_WIDE_INT this_vf_max
1964 1345 : = estimated_poly_value (this_vf, POLY_VALUE_MAX);
1965 :
1966 1345 : other_factor = CEIL (main_vf_max, other_vf_max);
1967 1345 : this_factor = CEIL (main_vf_max, this_vf_max);
1968 :
1969 : /* If the loop is not using partial vectors then it will iterate one
1970 : time less than one that does. It is safe to subtract one here,
1971 : because the main loop's vf is always at least 2x bigger than that
1972 : of an epilogue. */
1973 1345 : if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (other_loop_vinfo))
1974 1332 : other_factor -= 1;
1975 1345 : if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (this_loop_vinfo))
1976 1343 : this_factor -= 1;
1977 : }
1978 :
1979 : /* Compute the costs by multiplying the inside costs with the factor and
1980 : add the outside costs for a more complete picture. The factor is the
1981 : amount of times we are expecting to iterate this epilogue. */
1982 1439 : other_cost = other->body_cost () * other_factor;
1983 1439 : this_cost = this->body_cost () * this_factor;
1984 1439 : other_cost += other->outside_cost ();
1985 1439 : this_cost += this->outside_cost ();
1986 1439 : return this_cost < other_cost;
1987 : }
1988 :
1989 : /* A <=>-style subroutine of better_main_loop_than_p. Check whether we can
1990 : determine the return value of better_main_loop_than_p by comparing the
1991 : inside (loop body) costs of THIS and OTHER. Return:
1992 :
1993 : * -1 if better_main_loop_than_p should return true.
1994 : * 1 if better_main_loop_than_p should return false.
1995 : * 0 if we can't decide. */
1996 :
1997 : int
1998 30777 : vector_costs::compare_inside_loop_cost (const vector_costs *other) const
1999 : {
2000 30777 : loop_vec_info this_loop_vinfo = as_a<loop_vec_info> (this->m_vinfo);
2001 30777 : loop_vec_info other_loop_vinfo = as_a<loop_vec_info> (other->m_vinfo);
2002 :
2003 30777 : struct loop *loop = LOOP_VINFO_LOOP (this_loop_vinfo);
2004 30777 : gcc_assert (LOOP_VINFO_LOOP (other_loop_vinfo) == loop);
2005 :
2006 30777 : poly_int64 this_vf = LOOP_VINFO_VECT_FACTOR (this_loop_vinfo);
2007 30777 : poly_int64 other_vf = LOOP_VINFO_VECT_FACTOR (other_loop_vinfo);
2008 :
2009 : /* Limit the VFs to what is likely to be the maximum number of iterations,
2010 : to handle cases in which at least one loop_vinfo is fully-masked. */
2011 30777 : HOST_WIDE_INT estimated_max_niter = likely_max_stmt_executions_int (loop);
2012 30777 : if (estimated_max_niter != -1)
2013 : {
2014 29703 : if (estimated_poly_value (this_vf, POLY_VALUE_MIN)
2015 : >= estimated_max_niter)
2016 : this_vf = estimated_max_niter;
2017 29703 : if (estimated_poly_value (other_vf, POLY_VALUE_MIN)
2018 : >= estimated_max_niter)
2019 : other_vf = estimated_max_niter;
2020 : }
2021 :
2022 : /* Check whether the (fractional) cost per scalar iteration is lower or
2023 : higher: this_inside_cost / this_vf vs. other_inside_cost / other_vf. */
2024 30777 : poly_int64 rel_this = this_loop_vinfo->vector_costs->body_cost () * other_vf;
2025 30777 : poly_int64 rel_other
2026 30777 : = other_loop_vinfo->vector_costs->body_cost () * this_vf;
2027 :
2028 30777 : HOST_WIDE_INT est_rel_this_min
2029 30777 : = estimated_poly_value (rel_this, POLY_VALUE_MIN);
2030 30777 : HOST_WIDE_INT est_rel_this_max
2031 30777 : = estimated_poly_value (rel_this, POLY_VALUE_MAX);
2032 :
2033 30777 : HOST_WIDE_INT est_rel_other_min
2034 30777 : = estimated_poly_value (rel_other, POLY_VALUE_MIN);
2035 30777 : HOST_WIDE_INT est_rel_other_max
2036 30777 : = estimated_poly_value (rel_other, POLY_VALUE_MAX);
2037 :
2038 : /* Check first if we can make out an unambigous total order from the minimum
2039 : and maximum estimates. */
2040 30777 : if (est_rel_this_min < est_rel_other_min
2041 : && est_rel_this_max < est_rel_other_max)
2042 : return -1;
2043 :
2044 29903 : if (est_rel_other_min < est_rel_this_min
2045 : && est_rel_other_max < est_rel_this_max)
2046 29816 : return 1;
2047 :
2048 : /* When other_loop_vinfo uses a variable vectorization factor,
2049 : we know that it has a lower cost for at least one runtime VF.
2050 : However, we don't know how likely that VF is.
2051 :
2052 : One option would be to compare the costs for the estimated VFs.
2053 : The problem is that that can put too much pressure on the cost
2054 : model. E.g. if the estimated VF is also the lowest possible VF,
2055 : and if other_loop_vinfo is 1 unit worse than this_loop_vinfo
2056 : for the estimated VF, we'd then choose this_loop_vinfo even
2057 : though (a) this_loop_vinfo might not actually be better than
2058 : other_loop_vinfo for that VF and (b) it would be significantly
2059 : worse at larger VFs.
2060 :
2061 : Here we go for a hacky compromise: pick this_loop_vinfo if it is
2062 : no more expensive than other_loop_vinfo even after doubling the
2063 : estimated other_loop_vinfo VF. For all but trivial loops, this
2064 : ensures that we only pick this_loop_vinfo if it is significantly
2065 : better than other_loop_vinfo at the estimated VF. */
2066 : if (est_rel_other_min != est_rel_this_min
2067 : || est_rel_other_max != est_rel_this_max)
2068 : {
2069 : HOST_WIDE_INT est_rel_this_likely
2070 : = estimated_poly_value (rel_this, POLY_VALUE_LIKELY);
2071 : HOST_WIDE_INT est_rel_other_likely
2072 : = estimated_poly_value (rel_other, POLY_VALUE_LIKELY);
2073 :
2074 : return est_rel_this_likely * 2 <= est_rel_other_likely ? -1 : 1;
2075 : }
2076 :
2077 : return 0;
2078 : }
2079 :
2080 : /* A <=>-style subroutine of better_main_loop_than_p, used when there is
2081 : nothing to choose between the inside (loop body) costs of THIS and OTHER.
2082 : Check whether we can determine the return value of better_main_loop_than_p
2083 : by comparing the outside (prologue and epilogue) costs of THIS and OTHER.
2084 : Return:
2085 :
2086 : * -1 if better_main_loop_than_p should return true.
2087 : * 1 if better_main_loop_than_p should return false.
2088 : * 0 if we can't decide. */
2089 :
2090 : int
2091 87 : vector_costs::compare_outside_loop_cost (const vector_costs *other) const
2092 : {
2093 87 : auto this_outside_cost = this->outside_cost ();
2094 87 : auto other_outside_cost = other->outside_cost ();
2095 87 : if (this_outside_cost != other_outside_cost)
2096 61 : return this_outside_cost < other_outside_cost ? -1 : 1;
2097 :
2098 : return 0;
2099 : }
|
