Line data Source code
1 : /* Statement Analysis and Transformation for Vectorization
2 : Copyright (C) 2003-2026 Free Software Foundation, Inc.
3 : Contributed by Dorit Naishlos <dorit@il.ibm.com>
4 : and Ira Rosen <irar@il.ibm.com>
5 :
6 : This file is part of GCC.
7 :
8 : GCC is free software; you can redistribute it and/or modify it under
9 : the terms of the GNU General Public License as published by the Free
10 : Software Foundation; either version 3, or (at your option) any later
11 : version.
12 :
13 : GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 : WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 : FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 : for more details.
17 :
18 : You should have received a copy of the GNU General Public License
19 : along with GCC; see the file COPYING3. If not see
20 : <http://www.gnu.org/licenses/>. */
21 :
22 : #include "config.h"
23 : #include "system.h"
24 : #include "coretypes.h"
25 : #include "backend.h"
26 : #include "target.h"
27 : #include "rtl.h"
28 : #include "tree.h"
29 : #include "gimple.h"
30 : #include "ssa.h"
31 : #include "optabs-tree.h"
32 : #include "insn-config.h"
33 : #include "recog.h" /* FIXME: for insn_data */
34 : #include "cgraph.h"
35 : #include "dumpfile.h"
36 : #include "alias.h"
37 : #include "fold-const.h"
38 : #include "stor-layout.h"
39 : #include "tree-eh.h"
40 : #include "gimplify.h"
41 : #include "gimple-iterator.h"
42 : #include "gimplify-me.h"
43 : #include "tree-cfg.h"
44 : #include "tree-ssa-loop-manip.h"
45 : #include "cfgloop.h"
46 : #include "explow.h"
47 : #include "tree-ssa-loop.h"
48 : #include "tree-scalar-evolution.h"
49 : #include "tree-vectorizer.h"
50 : #include "builtins.h"
51 : #include "internal-fn.h"
52 : #include "tree-vector-builder.h"
53 : #include "vec-perm-indices.h"
54 : #include "gimple-range.h"
55 : #include "tree-ssa-loop-niter.h"
56 : #include "gimple-fold.h"
57 : #include "regs.h"
58 : #include "attribs.h"
59 : #include "optabs-libfuncs.h"
60 : #include "tree-dfa.h"
61 :
62 : /* For lang_hooks.types.type_for_mode. */
63 : #include "langhooks.h"
64 :
65 : static tree vector_vector_composition_type (tree, poly_uint64, tree *,
66 : bool = false);
67 :
68 : /* Return TRUE iff the given statement is in an inner loop relative to
69 : the loop being vectorized. */
70 : bool
71 3823896 : stmt_in_inner_loop_p (vec_info *vinfo, class _stmt_vec_info *stmt_info)
72 : {
73 3823896 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
74 3823896 : basic_block bb = gimple_bb (stmt);
75 3823896 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
76 785694 : class loop* loop;
77 :
78 785694 : if (!loop_vinfo)
79 : return false;
80 :
81 785694 : loop = LOOP_VINFO_LOOP (loop_vinfo);
82 :
83 785694 : return (bb->loop_father == loop->inner);
84 : }
85 :
86 : /* Record the cost of a statement, either by directly informing the
87 : target model or by saving it in a vector for later processing.
88 : Return a preliminary estimate of the statement's cost. */
89 :
90 : unsigned
91 9257276 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
92 : enum vect_cost_for_stmt kind,
93 : stmt_vec_info stmt_info, slp_tree node,
94 : tree vectype, int misalign,
95 : enum vect_cost_model_location where)
96 : {
97 9257276 : if ((kind == vector_load || kind == unaligned_load)
98 1354748 : && (stmt_info && STMT_VINFO_GATHER_SCATTER_P (stmt_info)))
99 : kind = vector_gather_load;
100 9257276 : if ((kind == vector_store || kind == unaligned_store)
101 904203 : && (stmt_info && STMT_VINFO_GATHER_SCATTER_P (stmt_info)))
102 9257276 : kind = vector_scatter_store;
103 :
104 9257276 : stmt_info_for_cost si
105 9257276 : = { count, kind, where, stmt_info, node, vectype, misalign };
106 9257276 : body_cost_vec->safe_push (si);
107 :
108 9257276 : return (unsigned)
109 9257276 : (builtin_vectorization_cost (kind, vectype, misalign) * count);
110 : }
111 :
112 : unsigned
113 4864480 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
114 : enum vect_cost_for_stmt kind, stmt_vec_info stmt_info,
115 : tree vectype, int misalign,
116 : enum vect_cost_model_location where)
117 : {
118 4864480 : return record_stmt_cost (body_cost_vec, count, kind, stmt_info, NULL,
119 4864480 : vectype, misalign, where);
120 : }
121 :
122 : unsigned
123 1514167 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
124 : enum vect_cost_for_stmt kind, slp_tree node,
125 : tree vectype, int misalign,
126 : enum vect_cost_model_location where)
127 : {
128 1514167 : return record_stmt_cost (body_cost_vec, count, kind,
129 : SLP_TREE_REPRESENTATIVE (node), node,
130 1514167 : vectype, misalign, where);
131 : }
132 :
133 : unsigned
134 177382 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
135 : enum vect_cost_for_stmt kind,
136 : enum vect_cost_model_location where)
137 : {
138 177382 : gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken
139 : || kind == scalar_stmt);
140 177382 : return record_stmt_cost (body_cost_vec, count, kind, NULL, NULL,
141 177382 : NULL_TREE, 0, where);
142 : }
143 :
144 : /* Return a variable of type ELEM_TYPE[NELEMS]. */
145 :
146 : static tree
147 0 : create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems)
148 : {
149 0 : return create_tmp_var (build_array_type_nelts (elem_type, nelems),
150 0 : "vect_array");
151 : }
152 :
153 : /* ARRAY is an array of vectors created by create_vector_array.
154 : Return an SSA_NAME for the vector in index N. The reference
155 : is part of the vectorization of STMT_INFO and the vector is associated
156 : with scalar destination SCALAR_DEST.
157 : If we need to ensure that inactive elements are set to zero,
158 : NEED_ZEROING is true, MASK contains the loop mask to be used. */
159 :
160 : static tree
161 0 : read_vector_array (vec_info *vinfo,
162 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
163 : tree scalar_dest, tree array, unsigned HOST_WIDE_INT n,
164 : bool need_zeroing, tree mask)
165 : {
166 0 : tree vect_type, vect, vect_name, tmp, tmp_name, array_ref;
167 0 : gimple *new_stmt;
168 :
169 0 : gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE);
170 0 : vect_type = TREE_TYPE (TREE_TYPE (array));
171 0 : tmp = vect_create_destination_var (scalar_dest, vect_type);
172 0 : vect = vect_create_destination_var (scalar_dest, vect_type);
173 0 : array_ref = build4 (ARRAY_REF, vect_type, array,
174 0 : build_int_cst (size_type_node, n),
175 : NULL_TREE, NULL_TREE);
176 :
177 0 : new_stmt = gimple_build_assign (tmp, array_ref);
178 0 : tmp_name = make_ssa_name (vect, new_stmt);
179 0 : gimple_assign_set_lhs (new_stmt, tmp_name);
180 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
181 :
182 0 : if (need_zeroing)
183 : {
184 0 : tree vec_els = vect_get_mask_load_else (MASK_LOAD_ELSE_ZERO,
185 : vect_type);
186 0 : vect_name = make_ssa_name (vect, new_stmt);
187 0 : new_stmt
188 0 : = gimple_build_assign (vect_name, VEC_COND_EXPR,
189 : mask, tmp_name, vec_els);
190 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
191 : }
192 : else
193 : vect_name = tmp_name;
194 :
195 0 : return vect_name;
196 : }
197 :
198 : /* ARRAY is an array of vectors created by create_vector_array.
199 : Emit code to store SSA_NAME VECT in index N of the array.
200 : The store is part of the vectorization of STMT_INFO. */
201 :
202 : static void
203 0 : write_vector_array (vec_info *vinfo,
204 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
205 : tree vect, tree array, unsigned HOST_WIDE_INT n)
206 : {
207 0 : tree array_ref;
208 0 : gimple *new_stmt;
209 :
210 0 : array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array,
211 0 : build_int_cst (size_type_node, n),
212 : NULL_TREE, NULL_TREE);
213 :
214 0 : new_stmt = gimple_build_assign (array_ref, vect);
215 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
216 0 : }
217 :
218 : /* PTR is a pointer to an array of type TYPE. Return a representation
219 : of *PTR. The memory reference replaces those in FIRST_DR
220 : (and its group). */
221 :
222 : static tree
223 0 : create_array_ref (tree type, tree ptr, tree alias_ptr_type)
224 : {
225 0 : tree mem_ref;
226 :
227 0 : mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0));
228 : /* Arrays have the same alignment as their type. */
229 0 : set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0);
230 0 : return mem_ref;
231 : }
232 :
233 : /* Add a clobber of variable VAR to the vectorization of STMT_INFO.
234 : Emit the clobber before *GSI. */
235 :
236 : static void
237 15 : vect_clobber_variable (vec_info *vinfo, stmt_vec_info stmt_info,
238 : gimple_stmt_iterator *gsi, tree var)
239 : {
240 15 : tree clobber = build_clobber (TREE_TYPE (var));
241 15 : gimple *new_stmt = gimple_build_assign (var, clobber);
242 15 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
243 15 : }
244 :
245 : /* Utility functions used by vect_mark_stmts_to_be_vectorized. */
246 :
247 : /* Function vect_mark_relevant.
248 :
249 : Mark STMT_INFO as "relevant for vectorization" and add it to WORKLIST. */
250 :
251 : static void
252 2762676 : vect_mark_relevant (vec<stmt_vec_info> *worklist, stmt_vec_info stmt_info,
253 : enum vect_relevant relevant, bool live_p)
254 : {
255 2762676 : enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info);
256 2762676 : bool save_live_p = STMT_VINFO_LIVE_P (stmt_info);
257 :
258 2762676 : if (dump_enabled_p ())
259 159761 : dump_printf_loc (MSG_NOTE, vect_location,
260 : "mark relevant %d, live %d: %G", relevant, live_p,
261 : stmt_info->stmt);
262 :
263 : /* If this stmt is an original stmt in a pattern, we might need to mark its
264 : related pattern stmt instead of the original stmt. However, such stmts
265 : may have their own uses that are not in any pattern, in such cases the
266 : stmt itself should be marked. */
267 2762676 : if (STMT_VINFO_IN_PATTERN_P (stmt_info))
268 : {
269 : /* This is the last stmt in a sequence that was detected as a
270 : pattern that can potentially be vectorized. Don't mark the stmt
271 : as relevant/live because it's not going to be vectorized.
272 : Instead mark the pattern-stmt that replaces it. */
273 :
274 241306 : if (dump_enabled_p ())
275 2711 : dump_printf_loc (MSG_NOTE, vect_location,
276 : "last stmt in pattern. don't mark"
277 : " relevant/live.\n");
278 :
279 241306 : stmt_vec_info old_stmt_info = stmt_info;
280 241306 : stmt_info = STMT_VINFO_RELATED_STMT (stmt_info);
281 241306 : gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == old_stmt_info);
282 241306 : save_relevant = STMT_VINFO_RELEVANT (stmt_info);
283 241306 : save_live_p = STMT_VINFO_LIVE_P (stmt_info);
284 :
285 241306 : if (live_p && relevant == vect_unused_in_scope)
286 : {
287 122 : if (dump_enabled_p ())
288 10 : dump_printf_loc (MSG_NOTE, vect_location,
289 : "vec_stmt_relevant_p: forcing live pattern stmt "
290 : "relevant.\n");
291 : relevant = vect_used_only_live;
292 : }
293 :
294 241306 : if (dump_enabled_p ())
295 2711 : dump_printf_loc (MSG_NOTE, vect_location,
296 : "mark relevant %d, live %d: %G", relevant, live_p,
297 : stmt_info->stmt);
298 : }
299 :
300 2762676 : STMT_VINFO_LIVE_P (stmt_info) |= live_p;
301 2762676 : if (relevant > STMT_VINFO_RELEVANT (stmt_info))
302 2498857 : STMT_VINFO_RELEVANT (stmt_info) = relevant;
303 :
304 2762676 : if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant
305 263819 : && STMT_VINFO_LIVE_P (stmt_info) == save_live_p)
306 : {
307 263103 : if (dump_enabled_p ())
308 19145 : dump_printf_loc (MSG_NOTE, vect_location,
309 : "already marked relevant/live.\n");
310 263103 : return;
311 : }
312 :
313 2499573 : worklist->safe_push (stmt_info);
314 : }
315 :
316 :
317 : /* Function is_simple_and_all_uses_invariant
318 :
319 : Return true if STMT_INFO is simple and all uses of it are invariant. */
320 :
321 : bool
322 226309 : is_simple_and_all_uses_invariant (stmt_vec_info stmt_info,
323 : loop_vec_info loop_vinfo)
324 : {
325 226309 : tree op;
326 226309 : ssa_op_iter iter;
327 :
328 399871 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
329 174400 : if (!stmt)
330 : return false;
331 :
332 181944 : FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
333 : {
334 181106 : enum vect_def_type dt = vect_uninitialized_def;
335 :
336 181106 : if (!vect_is_simple_use (op, loop_vinfo, &dt))
337 : {
338 5626 : if (dump_enabled_p ())
339 16 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
340 : "use not simple.\n");
341 173562 : return false;
342 : }
343 :
344 175480 : if (dt != vect_external_def && dt != vect_constant_def)
345 : return false;
346 : }
347 : return true;
348 : }
349 :
350 : /* Function vect_stmt_relevant_p.
351 :
352 : Return true if STMT_INFO, in the loop that is represented by LOOP_VINFO,
353 : is "relevant for vectorization".
354 :
355 : A stmt is considered "relevant for vectorization" if:
356 : - it has uses outside the loop.
357 : - it has vdefs (it alters memory).
358 : - control stmts in the loop (except for the exit condition).
359 :
360 : CHECKME: what other side effects would the vectorizer allow? */
361 :
362 : static bool
363 4343582 : vect_stmt_relevant_p (stmt_vec_info stmt_info, loop_vec_info loop_vinfo,
364 : enum vect_relevant *relevant, bool *live_p)
365 : {
366 4343582 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
367 4343582 : ssa_op_iter op_iter;
368 4343582 : imm_use_iterator imm_iter;
369 4343582 : use_operand_p use_p;
370 4343582 : def_operand_p def_p;
371 :
372 4343582 : *relevant = vect_unused_in_scope;
373 4343582 : *live_p = false;
374 :
375 : /* cond stmt other than loop exit cond. */
376 4343582 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
377 4343582 : if (is_ctrl_stmt (stmt)
378 542234 : && LOOP_VINFO_LOOP_IV_COND (loop_vinfo) != stmt
379 4574671 : && (!loop->inner || gimple_bb (stmt)->loop_father == loop))
380 229358 : *relevant = vect_used_in_scope;
381 :
382 : /* changing memory. */
383 4343582 : if (gimple_code (stmt_info->stmt) != GIMPLE_PHI)
384 3604578 : if (gimple_vdef (stmt_info->stmt)
385 3062344 : && !gimple_clobber_p (stmt_info->stmt))
386 : {
387 293792 : if (dump_enabled_p ())
388 27202 : dump_printf_loc (MSG_NOTE, vect_location,
389 : "vec_stmt_relevant_p: stmt has vdefs.\n");
390 293792 : *relevant = vect_used_in_scope;
391 293792 : if (! STMT_VINFO_DATA_REF (stmt_info)
392 293792 : && zero_ssa_operands (stmt_info->stmt, SSA_OP_DEF))
393 20 : LOOP_VINFO_ALTERNATE_DEFS (loop_vinfo).safe_push (stmt_info);
394 : }
395 :
396 : /* uses outside the loop. */
397 12188340 : FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt_info->stmt, op_iter, SSA_OP_DEF)
398 : {
399 12973872 : FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p))
400 : {
401 5971520 : basic_block bb = gimple_bb (USE_STMT (use_p));
402 5971520 : if (!flow_bb_inside_loop_p (loop, bb))
403 : {
404 242675 : if (is_gimple_debug (USE_STMT (use_p)))
405 1081 : continue;
406 :
407 241594 : if (dump_enabled_p ())
408 5778 : dump_printf_loc (MSG_NOTE, vect_location,
409 : "vec_stmt_relevant_p: used out of loop.\n");
410 :
411 : /* We expect all such uses to be in the loop exit phis
412 : (because of loop closed form) */
413 241594 : gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI);
414 :
415 241594 : *live_p = true;
416 : }
417 3501176 : }
418 : }
419 :
420 226311 : if (*live_p && *relevant == vect_unused_in_scope
421 4569891 : && !is_simple_and_all_uses_invariant (stmt_info, loop_vinfo))
422 : {
423 225471 : if (dump_enabled_p ())
424 5638 : dump_printf_loc (MSG_NOTE, vect_location,
425 : "vec_stmt_relevant_p: stmt live but not relevant.\n");
426 225471 : *relevant = vect_used_only_live;
427 : }
428 :
429 4343582 : return (*live_p || *relevant);
430 : }
431 :
432 :
433 : /* Function exist_non_indexing_operands_for_use_p
434 :
435 : USE is one of the uses attached to STMT_INFO. Check if USE is
436 : used in STMT_INFO for anything other than indexing an array. */
437 :
438 : static bool
439 3686978 : exist_non_indexing_operands_for_use_p (tree use, stmt_vec_info stmt_info)
440 : {
441 3686978 : tree operand;
442 :
443 : /* USE corresponds to some operand in STMT. If there is no data
444 : reference in STMT, then any operand that corresponds to USE
445 : is not indexing an array. */
446 3686978 : if (!STMT_VINFO_DATA_REF (stmt_info))
447 : return true;
448 :
449 : /* STMT has a data_ref. FORNOW this means that its of one of
450 : the following forms:
451 : -1- ARRAY_REF = var
452 : -2- var = ARRAY_REF
453 : (This should have been verified in analyze_data_refs).
454 :
455 : 'var' in the second case corresponds to a def, not a use,
456 : so USE cannot correspond to any operands that are not used
457 : for array indexing.
458 :
459 : Therefore, all we need to check is if STMT falls into the
460 : first case, and whether var corresponds to USE. */
461 :
462 1214074 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
463 1200257 : if (!assign || !gimple_assign_copy_p (assign))
464 : {
465 679039 : gcall *call = dyn_cast <gcall *> (stmt_info->stmt);
466 13817 : if (call && gimple_call_internal_p (call))
467 : {
468 13817 : internal_fn ifn = gimple_call_internal_fn (call);
469 13817 : int mask_index = internal_fn_mask_index (ifn);
470 13817 : if (mask_index >= 0
471 13817 : && use == gimple_call_arg (call, mask_index))
472 : return true;
473 9057 : int els_index = internal_fn_else_index (ifn);
474 9057 : if (els_index >= 0
475 9057 : && use == gimple_call_arg (call, els_index))
476 : return true;
477 7755 : int stored_value_index = internal_fn_stored_value_index (ifn);
478 7755 : if (stored_value_index >= 0
479 7755 : && use == gimple_call_arg (call, stored_value_index))
480 : return true;
481 6258 : if (internal_gather_scatter_fn_p (ifn)
482 6258 : && use == gimple_call_arg (call, 1))
483 : return true;
484 : }
485 671480 : return false;
486 : }
487 :
488 535035 : if (TREE_CODE (gimple_assign_lhs (assign)) == SSA_NAME)
489 : return false;
490 535035 : operand = gimple_assign_rhs1 (assign);
491 535035 : if (TREE_CODE (operand) != SSA_NAME)
492 : return false;
493 :
494 464573 : if (operand == use)
495 : return true;
496 :
497 : return false;
498 : }
499 :
500 :
501 : /*
502 : Function process_use.
503 :
504 : Inputs:
505 : - a USE in STMT_VINFO in a loop represented by LOOP_VINFO
506 : - RELEVANT - enum value to be set in the STMT_VINFO of the stmt
507 : that defined USE. This is done by calling mark_relevant and passing it
508 : the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant).
509 : - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't
510 : be performed.
511 :
512 : Outputs:
513 : Generally, LIVE_P and RELEVANT are used to define the liveness and
514 : relevance info of the DEF_STMT of this USE:
515 : STMT_VINFO_LIVE_P (DEF_stmt_vinfo) <-- live_p
516 : STMT_VINFO_RELEVANT (DEF_stmt_vinfo) <-- relevant
517 : Exceptions:
518 : - case 1: If USE is used only for address computations (e.g. array indexing),
519 : which does not need to be directly vectorized, then the liveness/relevance
520 : of the respective DEF_STMT is left unchanged.
521 : - case 2: If STMT_VINFO is a reduction phi and DEF_STMT is a reduction stmt,
522 : we skip DEF_STMT cause it had already been processed.
523 : - case 3: If DEF_STMT and STMT_VINFO are in different nests, then
524 : "relevant" will be modified accordingly.
525 :
526 : Return true if everything is as expected. Return false otherwise. */
527 :
528 : static opt_result
529 3748469 : process_use (stmt_vec_info stmt_vinfo, tree use, loop_vec_info loop_vinfo,
530 : enum vect_relevant relevant, vec<stmt_vec_info> *worklist,
531 : bool force)
532 : {
533 3748469 : stmt_vec_info dstmt_vinfo;
534 3748469 : enum vect_def_type dt;
535 :
536 : /* case 1: we are only interested in uses that need to be vectorized. Uses
537 : that are used for address computation are not considered relevant. */
538 3748469 : if (!force && !exist_non_indexing_operands_for_use_p (use, stmt_vinfo))
539 987431 : return opt_result::success ();
540 :
541 2761038 : if (!vect_is_simple_use (use, loop_vinfo, &dt, &dstmt_vinfo))
542 35884 : return opt_result::failure_at (stmt_vinfo->stmt,
543 : "not vectorized:"
544 : " unsupported use in stmt.\n");
545 :
546 2725154 : if (!dstmt_vinfo)
547 538133 : return opt_result::success ();
548 :
549 2187021 : basic_block def_bb = gimple_bb (dstmt_vinfo->stmt);
550 2187021 : basic_block bb = gimple_bb (stmt_vinfo->stmt);
551 :
552 : /* case 2: A reduction phi (STMT) defined by a reduction stmt (DSTMT_VINFO).
553 : We have to force the stmt live since the epilogue loop needs it to
554 : continue computing the reduction. */
555 2187021 : if (gimple_code (stmt_vinfo->stmt) == GIMPLE_PHI
556 228686 : && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def
557 59373 : && gimple_code (dstmt_vinfo->stmt) != GIMPLE_PHI
558 59373 : && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def
559 2246394 : && bb->loop_father == def_bb->loop_father)
560 : {
561 59373 : if (dump_enabled_p ())
562 3744 : dump_printf_loc (MSG_NOTE, vect_location,
563 : "reduc-stmt defining reduc-phi in the same nest.\n");
564 59373 : vect_mark_relevant (worklist, dstmt_vinfo, relevant, true);
565 59373 : return opt_result::success ();
566 : }
567 :
568 : /* case 3a: outer-loop stmt defining an inner-loop stmt:
569 : outer-loop-header-bb:
570 : d = dstmt_vinfo
571 : inner-loop:
572 : stmt # use (d)
573 : outer-loop-tail-bb:
574 : ... */
575 2127648 : if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father))
576 : {
577 1899 : if (dump_enabled_p ())
578 320 : dump_printf_loc (MSG_NOTE, vect_location,
579 : "outer-loop def-stmt defining inner-loop stmt.\n");
580 :
581 1899 : switch (relevant)
582 : {
583 0 : case vect_unused_in_scope:
584 0 : relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ?
585 : vect_used_in_scope : vect_unused_in_scope;
586 : break;
587 :
588 637 : case vect_used_in_outer_by_reduction:
589 637 : gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def);
590 : relevant = vect_used_by_reduction;
591 : break;
592 :
593 1032 : case vect_used_in_outer:
594 1032 : gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def);
595 : relevant = vect_used_in_scope;
596 : break;
597 :
598 : case vect_used_in_scope:
599 : break;
600 :
601 0 : default:
602 0 : gcc_unreachable ();
603 : }
604 : }
605 :
606 : /* case 3b: inner-loop stmt defining an outer-loop stmt:
607 : outer-loop-header-bb:
608 : ...
609 : inner-loop:
610 : d = dstmt_vinfo
611 : outer-loop-tail-bb (or outer-loop-exit-bb in double reduction):
612 : stmt # use (d) */
613 2125749 : else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father))
614 : {
615 1833 : if (dump_enabled_p ())
616 625 : dump_printf_loc (MSG_NOTE, vect_location,
617 : "inner-loop def-stmt defining outer-loop stmt.\n");
618 :
619 1833 : switch (relevant)
620 : {
621 0 : case vect_unused_in_scope:
622 0 : relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def
623 0 : || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ?
624 : vect_used_in_outer_by_reduction : vect_unused_in_scope;
625 : break;
626 :
627 : case vect_used_by_reduction:
628 : case vect_used_only_live:
629 : relevant = vect_used_in_outer_by_reduction;
630 : break;
631 :
632 : case vect_used_in_scope:
633 1964031 : relevant = vect_used_in_outer;
634 : break;
635 :
636 0 : default:
637 0 : gcc_unreachable ();
638 : }
639 : }
640 : /* We are also not interested in uses on loop PHI backedges that are
641 : inductions. Otherwise we'll needlessly vectorize the IV increment
642 : and cause hybrid SLP for SLP inductions. */
643 2123916 : else if (gimple_code (stmt_vinfo->stmt) == GIMPLE_PHI
644 166367 : && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def
645 2287533 : && (PHI_ARG_DEF_FROM_EDGE (stmt_vinfo->stmt,
646 : loop_latch_edge (bb->loop_father))
647 : == use))
648 : {
649 163617 : if (dump_enabled_p ())
650 4789 : dump_printf_loc (MSG_NOTE, vect_location,
651 : "induction value on backedge.\n");
652 163617 : return opt_result::success ();
653 : }
654 :
655 1964031 : vect_mark_relevant (worklist, dstmt_vinfo, relevant, false);
656 1964031 : return opt_result::success ();
657 : }
658 :
659 :
660 : /* Function vect_mark_stmts_to_be_vectorized.
661 :
662 : Not all stmts in the loop need to be vectorized. For example:
663 :
664 : for i...
665 : for j...
666 : 1. T0 = i + j
667 : 2. T1 = a[T0]
668 :
669 : 3. j = j + 1
670 :
671 : Stmt 1 and 3 do not need to be vectorized, because loop control and
672 : addressing of vectorized data-refs are handled differently.
673 :
674 : This pass detects such stmts. */
675 :
676 : opt_result
677 369726 : vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo, bool *fatal)
678 : {
679 369726 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
680 369726 : basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
681 369726 : unsigned int nbbs = loop->num_nodes;
682 369726 : gimple_stmt_iterator si;
683 369726 : unsigned int i;
684 369726 : basic_block bb;
685 369726 : bool live_p;
686 369726 : enum vect_relevant relevant;
687 :
688 369726 : DUMP_VECT_SCOPE ("vect_mark_stmts_to_be_vectorized");
689 :
690 369726 : auto_vec<stmt_vec_info, 64> worklist;
691 :
692 : /* 1. Init worklist. */
693 1274841 : for (i = 0; i < nbbs; i++)
694 : {
695 915446 : bb = bbs[i];
696 1828248 : for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
697 : {
698 1845978 : if (virtual_operand_p (gimple_phi_result (gsi_stmt (si))))
699 183985 : continue;
700 739004 : stmt_vec_info phi_info = loop_vinfo->lookup_stmt (gsi_stmt (si));
701 739004 : if (dump_enabled_p ())
702 40259 : dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? %G",
703 : phi_info->stmt);
704 :
705 739004 : if (vect_stmt_relevant_p (phi_info, loop_vinfo, &relevant, &live_p))
706 : {
707 45703 : if (STMT_VINFO_DEF_TYPE (phi_info) == vect_unknown_def_type)
708 10187 : return opt_result::failure_at
709 10187 : (*si, "not vectorized: unhandled relevant PHI: %G", *si);
710 35516 : vect_mark_relevant (&worklist, phi_info, relevant, live_p);
711 : }
712 : }
713 7113976 : for (si = gsi_after_labels (bb); !gsi_end_p (si); gsi_next (&si))
714 : {
715 6208861 : gimple *stmt = gsi_stmt (si);
716 6208861 : if (is_gimple_debug (stmt))
717 2604139 : continue;
718 3604722 : stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (stmt);
719 3604722 : if (dump_enabled_p ())
720 215671 : dump_printf_loc (MSG_NOTE, vect_location,
721 : "init: stmt relevant? %G", stmt);
722 :
723 3604722 : if (gimple_get_lhs (stmt) == NULL_TREE
724 547759 : && !is_a <gcond *> (stmt)
725 3610247 : && !is_a <gcall *> (stmt))
726 144 : return opt_result::failure_at
727 144 : (stmt, "not vectorized: irregular stmt: %G", stmt);
728 :
729 3604578 : if (vect_stmt_relevant_p (stmt_info, loop_vinfo, &relevant, &live_p))
730 703756 : vect_mark_relevant (&worklist, stmt_info, relevant, live_p);
731 : }
732 : }
733 :
734 : /* 2. Process_worklist */
735 2730156 : while (worklist.length () > 0)
736 : {
737 2406647 : use_operand_p use_p;
738 2406647 : ssa_op_iter iter;
739 :
740 2406647 : stmt_vec_info stmt_vinfo = worklist.pop ();
741 2406647 : if (dump_enabled_p ())
742 140006 : dump_printf_loc (MSG_NOTE, vect_location,
743 : "worklist: examine stmt: %G", stmt_vinfo->stmt);
744 :
745 : /* Examine the USEs of STMT. For each USE, mark the stmt that defines it
746 : (DEF_STMT) as relevant/irrelevant according to the relevance property
747 : of STMT. */
748 2406647 : relevant = STMT_VINFO_RELEVANT (stmt_vinfo);
749 :
750 : /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is
751 : propagated as is to the DEF_STMTs of its USEs.
752 :
753 : One exception is when STMT has been identified as defining a reduction
754 : variable; in this case we set the relevance to vect_used_by_reduction.
755 : This is because we distinguish between two kinds of relevant stmts -
756 : those that are used by a reduction computation, and those that are
757 : (also) used by a regular computation. This allows us later on to
758 : identify stmts that are used solely by a reduction, and therefore the
759 : order of the results that they produce does not have to be kept. */
760 :
761 2406647 : switch (STMT_VINFO_DEF_TYPE (stmt_vinfo))
762 : {
763 120835 : case vect_reduction_def:
764 120835 : gcc_assert (relevant != vect_unused_in_scope);
765 120835 : if (relevant != vect_unused_in_scope
766 120835 : && relevant != vect_used_in_scope
767 120835 : && relevant != vect_used_by_reduction
768 120835 : && relevant != vect_used_only_live)
769 0 : return opt_result::failure_at
770 0 : (stmt_vinfo->stmt, "unsupported use of reduction.\n");
771 : break;
772 :
773 1996 : case vect_nested_cycle:
774 1996 : if (relevant != vect_unused_in_scope
775 1996 : && relevant != vect_used_in_outer_by_reduction
776 1482 : && relevant != vect_used_in_outer)
777 2 : return opt_result::failure_at
778 2 : (stmt_vinfo->stmt, "unsupported use of nested cycle.\n");
779 : break;
780 :
781 1035 : case vect_double_reduction_def:
782 1035 : if (relevant != vect_unused_in_scope
783 1035 : && relevant != vect_used_by_reduction
784 351 : && relevant != vect_used_only_live)
785 0 : return opt_result::failure_at
786 0 : (stmt_vinfo->stmt, "unsupported use of double reduction.\n");
787 : break;
788 :
789 : default:
790 : break;
791 : }
792 :
793 2406645 : if (is_pattern_stmt_p (stmt_vinfo))
794 : {
795 : /* Pattern statements are not inserted into the code, so
796 : FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we
797 : have to scan the RHS or function arguments instead. */
798 589684 : if (gassign *assign = dyn_cast <gassign *> (stmt_vinfo->stmt))
799 : {
800 374658 : enum tree_code rhs_code = gimple_assign_rhs_code (assign);
801 374658 : tree op = gimple_assign_rhs1 (assign);
802 :
803 374658 : i = 1;
804 374658 : if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op))
805 : {
806 0 : opt_result res
807 0 : = process_use (stmt_vinfo, TREE_OPERAND (op, 0),
808 : loop_vinfo, relevant, &worklist, false);
809 0 : if (!res)
810 0 : return res;
811 0 : res = process_use (stmt_vinfo, TREE_OPERAND (op, 1),
812 : loop_vinfo, relevant, &worklist, false);
813 0 : if (!res)
814 0 : return res;
815 : i = 2;
816 : }
817 1079913 : for (; i < gimple_num_ops (assign); i++)
818 : {
819 709501 : op = gimple_op (assign, i);
820 709501 : if (TREE_CODE (op) == SSA_NAME)
821 : {
822 541094 : opt_result res
823 541094 : = process_use (stmt_vinfo, op, loop_vinfo, relevant,
824 : &worklist, false);
825 541094 : if (!res)
826 4246 : return res;
827 : }
828 : }
829 : }
830 215026 : else if (gcond *cond = dyn_cast <gcond *> (stmt_vinfo->stmt))
831 : {
832 210672 : tree_code rhs_code = gimple_cond_code (cond);
833 210672 : gcc_assert (TREE_CODE_CLASS (rhs_code) == tcc_comparison);
834 210672 : opt_result res
835 210672 : = process_use (stmt_vinfo, gimple_cond_lhs (cond),
836 : loop_vinfo, relevant, &worklist, false);
837 210672 : if (!res)
838 35886 : return res;
839 210672 : res = process_use (stmt_vinfo, gimple_cond_rhs (cond),
840 : loop_vinfo, relevant, &worklist, false);
841 210672 : if (!res)
842 0 : return res;
843 : }
844 4354 : else if (gcall *call = dyn_cast <gcall *> (stmt_vinfo->stmt))
845 : {
846 21018 : for (i = 0; i < gimple_call_num_args (call); i++)
847 : {
848 16664 : tree arg = gimple_call_arg (call, i);
849 16664 : opt_result res
850 16664 : = process_use (stmt_vinfo, arg, loop_vinfo, relevant,
851 : &worklist, false);
852 16664 : if (!res)
853 0 : return res;
854 : }
855 : }
856 : else
857 0 : gcc_unreachable ();
858 : }
859 : else
860 6323373 : FOR_EACH_PHI_OR_STMT_USE (use_p, stmt_vinfo->stmt, iter, SSA_OP_USE)
861 : {
862 2707876 : tree op = USE_FROM_PTR (use_p);
863 2707876 : opt_result res
864 2707876 : = process_use (stmt_vinfo, op, loop_vinfo, relevant,
865 : &worklist, false);
866 2707876 : if (!res)
867 18425 : return res;
868 : }
869 :
870 2383974 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo))
871 : {
872 61491 : gather_scatter_info gs_info;
873 61491 : if (!vect_check_gather_scatter (stmt_vinfo,
874 : STMT_VINFO_VECTYPE (stmt_vinfo),
875 : loop_vinfo, &gs_info))
876 0 : gcc_unreachable ();
877 61491 : opt_result res
878 61491 : = process_use (stmt_vinfo, gs_info.offset, loop_vinfo, relevant,
879 : &worklist, true);
880 61491 : if (!res)
881 : {
882 13213 : if (fatal)
883 13213 : *fatal = false;
884 13213 : return res;
885 : }
886 : }
887 : } /* while worklist */
888 :
889 323509 : return opt_result::success ();
890 369726 : }
891 :
892 : /* Function vect_model_simple_cost.
893 :
894 : Models cost for simple operations, i.e. those that only emit N operations
895 : of the same KIND. */
896 :
897 : static void
898 622183 : vect_model_simple_cost (vec_info *vinfo, int n, slp_tree node,
899 : stmt_vector_for_cost *cost_vec,
900 : vect_cost_for_stmt kind = vector_stmt)
901 : {
902 622183 : int inside_cost = 0, prologue_cost = 0;
903 :
904 622183 : gcc_assert (cost_vec != NULL);
905 :
906 622183 : n *= vect_get_num_copies (vinfo, node);
907 :
908 : /* Pass the inside-of-loop statements to the target-specific cost model. */
909 622183 : inside_cost += record_stmt_cost (cost_vec, n, kind, node, 0, vect_body);
910 :
911 622183 : if (dump_enabled_p ())
912 32520 : dump_printf_loc (MSG_NOTE, vect_location,
913 : "vect_model_simple_cost: inside_cost = %d, "
914 : "prologue_cost = %d .\n", inside_cost, prologue_cost);
915 622183 : }
916 :
917 :
918 : /* Model cost for type demotion and promotion operations. PWR is
919 : normally zero for single-step promotions and demotions. It will be
920 : one if two-step promotion/demotion is required, and so on. NCOPIES
921 : is the number of vector results (and thus number of instructions)
922 : for the narrowest end of the operation chain. Each additional
923 : step doubles the number of instructions required. If WIDEN_ARITH
924 : is true the stmt is doing widening arithmetic. */
925 :
926 : static void
927 53507 : vect_model_promotion_demotion_cost (slp_tree slp_node,
928 : unsigned int ncopies, int pwr,
929 : stmt_vector_for_cost *cost_vec,
930 : bool widen_arith)
931 : {
932 53507 : int i;
933 53507 : int inside_cost = 0, prologue_cost = 0;
934 :
935 125836 : for (i = 0; i < pwr + 1; i++)
936 : {
937 142870 : inside_cost += record_stmt_cost (cost_vec, ncopies,
938 : widen_arith
939 : ? vector_stmt : vec_promote_demote,
940 : slp_node, 0, vect_body);
941 72329 : ncopies *= 2;
942 : }
943 :
944 53507 : if (dump_enabled_p ())
945 6155 : dump_printf_loc (MSG_NOTE, vect_location,
946 : "vect_model_promotion_demotion_cost: inside_cost = %d, "
947 : "prologue_cost = %d .\n", inside_cost, prologue_cost);
948 53507 : }
949 :
950 : /* Returns true if the current function returns DECL. */
951 :
952 : static bool
953 534724 : cfun_returns (tree decl)
954 : {
955 534724 : edge_iterator ei;
956 534724 : edge e;
957 1052417 : FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
958 : {
959 1058636 : greturn *ret = safe_dyn_cast <greturn *> (*gsi_last_bb (e->src));
960 529318 : if (!ret)
961 0 : continue;
962 529318 : if (gimple_return_retval (ret) == decl)
963 : return true;
964 : /* We often end up with an aggregate copy to the result decl,
965 : handle that case as well. First skip intermediate clobbers
966 : though. */
967 : gimple *def = ret;
968 1562465 : do
969 : {
970 3124930 : def = SSA_NAME_DEF_STMT (gimple_vuse (def));
971 : }
972 1562465 : while (gimple_clobber_p (def));
973 518377 : if (is_a <gassign *> (def)
974 59845 : && gimple_assign_lhs (def) == gimple_return_retval (ret)
975 525307 : && gimple_assign_rhs1 (def) == decl)
976 : return true;
977 : }
978 : return false;
979 : }
980 :
981 : /* Calculate cost of DR's memory access. */
982 : void
983 894178 : vect_get_store_cost (vec_info *, stmt_vec_info stmt_info, slp_tree slp_node,
984 : int ncopies, dr_alignment_support alignment_support_scheme,
985 : int misalignment,
986 : unsigned int *inside_cost,
987 : stmt_vector_for_cost *body_cost_vec)
988 : {
989 894178 : tree vectype
990 894178 : = slp_node ? SLP_TREE_VECTYPE (slp_node) : STMT_VINFO_VECTYPE (stmt_info);
991 894178 : switch (alignment_support_scheme)
992 : {
993 472818 : case dr_aligned:
994 472818 : {
995 472818 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies,
996 : vector_store, stmt_info, slp_node,
997 : vectype, 0, vect_body);
998 :
999 472818 : if (dump_enabled_p ())
1000 13674 : dump_printf_loc (MSG_NOTE, vect_location,
1001 : "vect_model_store_cost: aligned.\n");
1002 : break;
1003 : }
1004 :
1005 421360 : case dr_unaligned_supported:
1006 421360 : {
1007 : /* Here, we assign an additional cost for the unaligned store. */
1008 421360 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies,
1009 : unaligned_store, stmt_info, slp_node,
1010 : vectype, misalignment, vect_body);
1011 421360 : if (dump_enabled_p ())
1012 12573 : dump_printf_loc (MSG_NOTE, vect_location,
1013 : "vect_model_store_cost: unaligned supported by "
1014 : "hardware.\n");
1015 : break;
1016 : }
1017 :
1018 0 : case dr_unaligned_unsupported:
1019 0 : {
1020 0 : *inside_cost = VECT_MAX_COST;
1021 :
1022 0 : if (dump_enabled_p ())
1023 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1024 : "vect_model_store_cost: unsupported access.\n");
1025 : break;
1026 : }
1027 :
1028 0 : default:
1029 0 : gcc_unreachable ();
1030 : }
1031 894178 : }
1032 :
1033 : /* Calculate cost of DR's memory access. */
1034 : void
1035 741597 : vect_get_load_cost (vec_info *, stmt_vec_info stmt_info, slp_tree slp_node,
1036 : int ncopies, dr_alignment_support alignment_support_scheme,
1037 : int misalignment,
1038 : bool add_realign_cost, unsigned int *inside_cost,
1039 : unsigned int *prologue_cost,
1040 : stmt_vector_for_cost *prologue_cost_vec,
1041 : stmt_vector_for_cost *body_cost_vec,
1042 : bool record_prologue_costs)
1043 : {
1044 741597 : tree vectype
1045 741597 : = slp_node ? SLP_TREE_VECTYPE (slp_node) : STMT_VINFO_VECTYPE (stmt_info);
1046 741597 : switch (alignment_support_scheme)
1047 : {
1048 428676 : case dr_aligned:
1049 428676 : {
1050 428676 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load,
1051 : stmt_info, slp_node, vectype,
1052 : 0, vect_body);
1053 :
1054 428676 : if (dump_enabled_p ())
1055 17697 : dump_printf_loc (MSG_NOTE, vect_location,
1056 : "vect_model_load_cost: aligned.\n");
1057 :
1058 : break;
1059 : }
1060 257693 : case dr_unaligned_supported:
1061 257693 : {
1062 : /* Here, we assign an additional cost for the unaligned load. */
1063 257693 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies,
1064 : unaligned_load, stmt_info, slp_node,
1065 : vectype, misalignment, vect_body);
1066 :
1067 257693 : if (dump_enabled_p ())
1068 20881 : dump_printf_loc (MSG_NOTE, vect_location,
1069 : "vect_model_load_cost: unaligned supported by "
1070 : "hardware.\n");
1071 :
1072 : break;
1073 : }
1074 0 : case dr_explicit_realign:
1075 0 : {
1076 0 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2,
1077 : vector_load, stmt_info, slp_node,
1078 : vectype, 0, vect_body);
1079 0 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies,
1080 : vec_perm, stmt_info, slp_node,
1081 : vectype, 0, vect_body);
1082 :
1083 : /* FIXME: If the misalignment remains fixed across the iterations of
1084 : the containing loop, the following cost should be added to the
1085 : prologue costs. */
1086 0 : if (targetm.vectorize.builtin_mask_for_load)
1087 0 : *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt,
1088 : stmt_info, slp_node, vectype,
1089 : 0, vect_body);
1090 :
1091 0 : if (dump_enabled_p ())
1092 0 : dump_printf_loc (MSG_NOTE, vect_location,
1093 : "vect_model_load_cost: explicit realign\n");
1094 :
1095 : break;
1096 : }
1097 0 : case dr_explicit_realign_optimized:
1098 0 : {
1099 0 : if (dump_enabled_p ())
1100 0 : dump_printf_loc (MSG_NOTE, vect_location,
1101 : "vect_model_load_cost: unaligned software "
1102 : "pipelined.\n");
1103 :
1104 : /* Unaligned software pipeline has a load of an address, an initial
1105 : load, and possibly a mask operation to "prime" the loop. However,
1106 : if this is an access in a group of loads, which provide grouped
1107 : access, then the above cost should only be considered for one
1108 : access in the group. Inside the loop, there is a load op
1109 : and a realignment op. */
1110 :
1111 0 : if (add_realign_cost && record_prologue_costs)
1112 : {
1113 0 : *prologue_cost += record_stmt_cost (prologue_cost_vec, 2,
1114 : vector_stmt, stmt_info,
1115 : slp_node, vectype,
1116 : 0, vect_prologue);
1117 0 : if (targetm.vectorize.builtin_mask_for_load)
1118 0 : *prologue_cost += record_stmt_cost (prologue_cost_vec, 1,
1119 : vector_stmt, stmt_info,
1120 : slp_node, vectype,
1121 : 0, vect_prologue);
1122 : }
1123 :
1124 0 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load,
1125 : stmt_info, slp_node, vectype,
1126 : 0, vect_body);
1127 0 : *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm,
1128 : stmt_info, slp_node, vectype,
1129 : 0, vect_body);
1130 :
1131 0 : if (dump_enabled_p ())
1132 0 : dump_printf_loc (MSG_NOTE, vect_location,
1133 : "vect_model_load_cost: explicit realign optimized"
1134 : "\n");
1135 :
1136 : break;
1137 : }
1138 :
1139 55228 : case dr_unaligned_unsupported:
1140 55228 : {
1141 55228 : *inside_cost = VECT_MAX_COST;
1142 :
1143 55228 : if (dump_enabled_p ())
1144 94 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1145 : "vect_model_load_cost: unsupported access.\n");
1146 : break;
1147 : }
1148 :
1149 0 : default:
1150 0 : gcc_unreachable ();
1151 : }
1152 741597 : }
1153 :
1154 : /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in
1155 : the loop preheader for the vectorized stmt STMT_VINFO. */
1156 :
1157 : static void
1158 6358 : vect_init_vector_1 (vec_info *vinfo, stmt_vec_info stmt_vinfo, gimple *new_stmt,
1159 : gimple_stmt_iterator *gsi)
1160 : {
1161 6358 : if (gsi)
1162 3035 : vect_finish_stmt_generation (vinfo, stmt_vinfo, new_stmt, gsi);
1163 : else
1164 3323 : vinfo->insert_on_entry (stmt_vinfo, new_stmt);
1165 :
1166 6358 : if (dump_enabled_p ())
1167 1807 : dump_printf_loc (MSG_NOTE, vect_location,
1168 : "created new init_stmt: %G", new_stmt);
1169 6358 : }
1170 :
1171 : /* Function vect_init_vector.
1172 :
1173 : Insert a new stmt (INIT_STMT) that initializes a new variable of type
1174 : TYPE with the value VAL. If TYPE is a vector type and VAL does not have
1175 : vector type a vector with all elements equal to VAL is created first.
1176 : Place the initialization at GSI if it is not NULL. Otherwise, place the
1177 : initialization at the loop preheader.
1178 : Return the DEF of INIT_STMT.
1179 : It will be used in the vectorization of STMT_INFO. */
1180 :
1181 : tree
1182 4642 : vect_init_vector (vec_info *vinfo, stmt_vec_info stmt_info, tree val, tree type,
1183 : gimple_stmt_iterator *gsi)
1184 : {
1185 4642 : gimple *init_stmt;
1186 4642 : tree new_temp;
1187 :
1188 : /* We abuse this function to push sth to a SSA name with initial 'val'. */
1189 4642 : if (! useless_type_conversion_p (type, TREE_TYPE (val)))
1190 : {
1191 1361 : gcc_assert (VECTOR_TYPE_P (type));
1192 1361 : if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val)))
1193 : {
1194 : /* Scalar boolean value should be transformed into
1195 : all zeros or all ones value before building a vector. */
1196 8 : if (VECTOR_BOOLEAN_TYPE_P (type))
1197 : {
1198 0 : tree true_val = build_all_ones_cst (TREE_TYPE (type));
1199 0 : tree false_val = build_zero_cst (TREE_TYPE (type));
1200 :
1201 0 : if (CONSTANT_CLASS_P (val))
1202 0 : val = integer_zerop (val) ? false_val : true_val;
1203 : else
1204 : {
1205 0 : new_temp = make_ssa_name (TREE_TYPE (type));
1206 0 : init_stmt = gimple_build_assign (new_temp, COND_EXPR,
1207 : val, true_val, false_val);
1208 0 : vect_init_vector_1 (vinfo, stmt_info, init_stmt, gsi);
1209 0 : val = new_temp;
1210 : }
1211 : }
1212 : else
1213 : {
1214 8 : gimple_seq stmts = NULL;
1215 8 : if (! INTEGRAL_TYPE_P (TREE_TYPE (val)))
1216 8 : val = gimple_build (&stmts, VIEW_CONVERT_EXPR,
1217 8 : TREE_TYPE (type), val);
1218 : else
1219 : /* ??? Condition vectorization expects us to do
1220 : promotion of invariant/external defs. */
1221 0 : val = gimple_convert (&stmts, TREE_TYPE (type), val);
1222 16 : for (gimple_stmt_iterator gsi2 = gsi_start (stmts);
1223 16 : !gsi_end_p (gsi2); )
1224 : {
1225 8 : init_stmt = gsi_stmt (gsi2);
1226 8 : gsi_remove (&gsi2, false);
1227 8 : vect_init_vector_1 (vinfo, stmt_info, init_stmt, gsi);
1228 : }
1229 : }
1230 : }
1231 1361 : val = build_vector_from_val (type, val);
1232 : }
1233 :
1234 4642 : new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_");
1235 4642 : init_stmt = gimple_build_assign (new_temp, val);
1236 4642 : vect_init_vector_1 (vinfo, stmt_info, init_stmt, gsi);
1237 4642 : return new_temp;
1238 : }
1239 :
1240 :
1241 : /* Get vectorized definitions for OP0 and OP1. */
1242 :
1243 : void
1244 186520 : vect_get_vec_defs (vec_info *, slp_tree slp_node,
1245 : tree op0, vec<tree> *vec_oprnds0,
1246 : tree op1, vec<tree> *vec_oprnds1,
1247 : tree op2, vec<tree> *vec_oprnds2,
1248 : tree op3, vec<tree> *vec_oprnds3)
1249 : {
1250 186520 : if (op0)
1251 184873 : vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[0], vec_oprnds0);
1252 186520 : if (op1)
1253 138778 : vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[1], vec_oprnds1);
1254 186520 : if (op2)
1255 9186 : vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[2], vec_oprnds2);
1256 186520 : if (op3)
1257 0 : vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[3], vec_oprnds3);
1258 186520 : }
1259 :
1260 : /* Helper function called by vect_finish_replace_stmt and
1261 : vect_finish_stmt_generation. Set the location of the new
1262 : statement and create and return a stmt_vec_info for it. */
1263 :
1264 : static void
1265 1428291 : vect_finish_stmt_generation_1 (vec_info *,
1266 : stmt_vec_info stmt_info, gimple *vec_stmt)
1267 : {
1268 1428291 : if (dump_enabled_p ())
1269 148093 : dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: %G", vec_stmt);
1270 :
1271 1428291 : if (stmt_info)
1272 : {
1273 1396890 : gimple_set_location (vec_stmt, gimple_location (stmt_info->stmt));
1274 :
1275 : /* While EH edges will generally prevent vectorization, stmt might
1276 : e.g. be in a must-not-throw region. Ensure newly created stmts
1277 : that could throw are part of the same region. */
1278 1396890 : int lp_nr = lookup_stmt_eh_lp (stmt_info->stmt);
1279 1396890 : if (lp_nr != 0 && stmt_could_throw_p (cfun, vec_stmt))
1280 48 : add_stmt_to_eh_lp (vec_stmt, lp_nr);
1281 : }
1282 : else
1283 31401 : gcc_assert (!stmt_could_throw_p (cfun, vec_stmt));
1284 1428291 : }
1285 :
1286 : /* Replace the scalar statement STMT_INFO with a new vector statement VEC_STMT,
1287 : which sets the same scalar result as STMT_INFO did. Create and return a
1288 : stmt_vec_info for VEC_STMT. */
1289 :
1290 : void
1291 843 : vect_finish_replace_stmt (vec_info *vinfo,
1292 : stmt_vec_info stmt_info, gimple *vec_stmt)
1293 : {
1294 843 : gimple *scalar_stmt = vect_orig_stmt (stmt_info)->stmt;
1295 843 : gcc_assert (gimple_get_lhs (scalar_stmt) == gimple_get_lhs (vec_stmt));
1296 :
1297 843 : gimple_stmt_iterator gsi = gsi_for_stmt (scalar_stmt);
1298 843 : gsi_replace (&gsi, vec_stmt, true);
1299 :
1300 843 : vect_finish_stmt_generation_1 (vinfo, stmt_info, vec_stmt);
1301 843 : }
1302 :
1303 : /* Add VEC_STMT to the vectorized implementation of STMT_INFO and insert it
1304 : before *GSI. Create and return a stmt_vec_info for VEC_STMT. */
1305 :
1306 : void
1307 1427448 : vect_finish_stmt_generation (vec_info *vinfo,
1308 : stmt_vec_info stmt_info, gimple *vec_stmt,
1309 : gimple_stmt_iterator *gsi)
1310 : {
1311 1427448 : gcc_assert (!stmt_info || gimple_code (stmt_info->stmt) != GIMPLE_LABEL);
1312 :
1313 1427448 : if (!gsi_end_p (*gsi)
1314 2853788 : && gimple_has_mem_ops (vec_stmt))
1315 : {
1316 1426340 : gimple *at_stmt = gsi_stmt (*gsi);
1317 1426340 : tree vuse = gimple_vuse (at_stmt);
1318 1420100 : if (vuse && TREE_CODE (vuse) == SSA_NAME)
1319 : {
1320 1279180 : tree vdef = gimple_vdef (at_stmt);
1321 1279180 : gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt));
1322 1279180 : gimple_set_modified (vec_stmt, true);
1323 : /* If we have an SSA vuse and insert a store, update virtual
1324 : SSA form to avoid triggering the renamer. Do so only
1325 : if we can easily see all uses - which is what almost always
1326 : happens with the way vectorized stmts are inserted. */
1327 751601 : if ((vdef && TREE_CODE (vdef) == SSA_NAME)
1328 2030745 : && ((is_gimple_assign (vec_stmt)
1329 750687 : && !is_gimple_reg (gimple_assign_lhs (vec_stmt)))
1330 66317 : || (is_gimple_call (vec_stmt)
1331 878 : && (!(gimple_call_flags (vec_stmt)
1332 878 : & (ECF_CONST|ECF_PURE|ECF_NOVOPS))
1333 1 : || (gimple_call_lhs (vec_stmt)
1334 1 : && !is_gimple_reg (gimple_call_lhs (vec_stmt)))))))
1335 : {
1336 686125 : tree new_vdef = copy_ssa_name (vuse, vec_stmt);
1337 686125 : gimple_set_vdef (vec_stmt, new_vdef);
1338 686125 : SET_USE (gimple_vuse_op (at_stmt), new_vdef);
1339 : }
1340 : }
1341 : }
1342 1427448 : gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT);
1343 1427448 : vect_finish_stmt_generation_1 (vinfo, stmt_info, vec_stmt);
1344 1427448 : }
1345 :
1346 : /* We want to vectorize a call to combined function CFN with function
1347 : decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN
1348 : as the types of all inputs. Check whether this is possible using
1349 : an internal function, returning its code if so or IFN_LAST if not. */
1350 :
1351 : static internal_fn
1352 11866 : vectorizable_internal_function (combined_fn cfn, tree fndecl,
1353 : tree vectype_out, tree vectype_in)
1354 : {
1355 11866 : internal_fn ifn;
1356 11866 : if (internal_fn_p (cfn))
1357 9929 : ifn = as_internal_fn (cfn);
1358 : else
1359 1937 : ifn = associated_internal_fn (fndecl);
1360 11866 : if (ifn != IFN_LAST && direct_internal_fn_p (ifn))
1361 : {
1362 8422 : const direct_internal_fn_info &info = direct_internal_fn (ifn);
1363 8422 : if (info.vectorizable)
1364 : {
1365 8422 : bool same_size_p = TYPE_SIZE (vectype_in) == TYPE_SIZE (vectype_out);
1366 8422 : tree type0 = (info.type0 < 0 ? vectype_out : vectype_in);
1367 8422 : tree type1 = (info.type1 < 0 ? vectype_out : vectype_in);
1368 :
1369 : /* The type size of both the vectype_in and vectype_out should be
1370 : exactly the same when vectype_out isn't participating the optab.
1371 : While there is no restriction for type size when vectype_out
1372 : is part of the optab query. */
1373 8422 : if (type0 != vectype_out && type1 != vectype_out && !same_size_p)
1374 : return IFN_LAST;
1375 :
1376 8402 : if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1),
1377 : OPTIMIZE_FOR_SPEED))
1378 : return ifn;
1379 : }
1380 : }
1381 : return IFN_LAST;
1382 : }
1383 :
1384 :
1385 : static tree permute_vec_elements (vec_info *, tree, tree, tree, stmt_vec_info,
1386 : gimple_stmt_iterator *);
1387 :
1388 : /* Check whether a load or store statement in the loop described by
1389 : LOOP_VINFO is possible in a loop using partial vectors. This is
1390 : testing whether the vectorizer pass has the appropriate support,
1391 : as well as whether the target does.
1392 :
1393 : VLS_TYPE says whether the statement is a load or store and VECTYPE
1394 : is the type of the vector being loaded or stored. SLP_NODE is the SLP
1395 : node that contains the statement, or null if none. MEMORY_ACCESS_TYPE
1396 : says how the load or store is going to be implemented and GROUP_SIZE
1397 : is the number of load or store statements in the containing group.
1398 : If the access is a gather load or scatter store, GS_INFO describes
1399 : its arguments. If the load or store is conditional, SCALAR_MASK is the
1400 : condition under which it occurs.
1401 :
1402 : Clear LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P if a loop using partial
1403 : vectors is not supported, otherwise record the required rgroup control
1404 : types.
1405 :
1406 : If partial vectors can be used and ELSVALS is nonzero the supported
1407 : else values will be added to the vector ELSVALS points to. */
1408 :
1409 : static void
1410 224395 : check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype,
1411 : slp_tree slp_node,
1412 : vec_load_store_type vls_type,
1413 : int group_size,
1414 : vect_load_store_data *ls,
1415 : slp_tree mask_node,
1416 : vec<int> *elsvals = nullptr)
1417 : {
1418 224395 : vect_memory_access_type memory_access_type = ls->memory_access_type;
1419 :
1420 : /* Invariant loads need no special support. */
1421 224395 : if (memory_access_type == VMAT_INVARIANT)
1422 27239 : return;
1423 :
1424 : /* Figure whether the mask is uniform. scalar_mask is used to
1425 : populate the scalar_cond_masked_set. */
1426 223367 : tree scalar_mask = NULL_TREE;
1427 223367 : if (mask_node)
1428 3538 : for (unsigned i = 0; i < SLP_TREE_LANES (mask_node); ++i)
1429 : {
1430 1820 : tree def = vect_get_slp_scalar_def (mask_node, i);
1431 1820 : if (!def
1432 1820 : || (scalar_mask && def != scalar_mask))
1433 : {
1434 : scalar_mask = NULL;
1435 : break;
1436 : }
1437 : else
1438 1789 : scalar_mask = def;
1439 : }
1440 :
1441 223367 : unsigned int nvectors = vect_get_num_copies (loop_vinfo, slp_node);
1442 223367 : vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
1443 223367 : vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
1444 223367 : machine_mode vecmode = TYPE_MODE (vectype);
1445 223367 : bool is_load = (vls_type == VLS_LOAD);
1446 223367 : if (memory_access_type == VMAT_LOAD_STORE_LANES)
1447 : {
1448 0 : nvectors /= group_size;
1449 0 : internal_fn ifn
1450 0 : = (is_load ? vect_load_lanes_supported (vectype, group_size, true,
1451 : elsvals)
1452 0 : : vect_store_lanes_supported (vectype, group_size, true));
1453 0 : if (ifn == IFN_MASK_LEN_LOAD_LANES || ifn == IFN_MASK_LEN_STORE_LANES)
1454 0 : vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1);
1455 0 : else if (ifn == IFN_MASK_LOAD_LANES || ifn == IFN_MASK_STORE_LANES)
1456 0 : vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
1457 : scalar_mask);
1458 : else
1459 : {
1460 0 : if (dump_enabled_p ())
1461 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1462 : "can't operate on partial vectors because"
1463 : " the target doesn't have an appropriate"
1464 : " load/store-lanes instruction.\n");
1465 0 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
1466 : }
1467 0 : return;
1468 : }
1469 :
1470 223367 : if (mat_gather_scatter_p (memory_access_type))
1471 : {
1472 1441 : internal_fn ifn = (is_load
1473 1441 : ? IFN_MASK_GATHER_LOAD
1474 : : IFN_MASK_SCATTER_STORE);
1475 334 : internal_fn len_ifn = (is_load
1476 : ? IFN_MASK_LEN_GATHER_LOAD
1477 : : IFN_MASK_LEN_SCATTER_STORE);
1478 1441 : stmt_vec_info repr = SLP_TREE_REPRESENTATIVE (slp_node);
1479 1441 : tree off_vectype = (STMT_VINFO_GATHER_SCATTER_P (repr)
1480 1441 : ? SLP_TREE_VECTYPE (SLP_TREE_CHILDREN (slp_node)[0])
1481 1441 : : ls->strided_offset_vectype);
1482 1441 : tree memory_type = TREE_TYPE (DR_REF (STMT_VINFO_DR_INFO (repr)->dr));
1483 1441 : int scale = SLP_TREE_GS_SCALE (slp_node);
1484 :
1485 : /* The following "supported" checks just verify what we established in
1486 : get_load_store_type and don't try different offset types.
1487 : Therefore, off_vectype must be a supported offset type. In case
1488 : we chose a different one use this instead. */
1489 1441 : if (ls->supported_offset_vectype)
1490 0 : off_vectype = ls->supported_offset_vectype;
1491 : /* Same for scale. */
1492 1441 : if (ls->supported_scale)
1493 0 : scale = ls->supported_scale;
1494 :
1495 1441 : if (internal_gather_scatter_fn_supported_p (len_ifn, vectype,
1496 : memory_type,
1497 : off_vectype, scale,
1498 : elsvals))
1499 0 : vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1);
1500 1441 : else if (internal_gather_scatter_fn_supported_p (ifn, vectype,
1501 : memory_type,
1502 : off_vectype, scale,
1503 : elsvals)
1504 1441 : || memory_access_type == VMAT_GATHER_SCATTER_LEGACY)
1505 339 : vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype,
1506 : scalar_mask);
1507 : else
1508 : {
1509 1102 : if (dump_enabled_p ())
1510 22 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1511 : "can't operate on partial vectors because"
1512 : " the target doesn't have an appropriate"
1513 : " gather load or scatter store instruction.\n");
1514 1102 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
1515 : }
1516 1441 : return;
1517 : }
1518 :
1519 221926 : if (memory_access_type != VMAT_CONTIGUOUS)
1520 : {
1521 : /* Element X of the data must come from iteration i * VF + X of the
1522 : scalar loop. We need more work to support other mappings. */
1523 24770 : if (dump_enabled_p ())
1524 720 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1525 : "can't operate on partial vectors because an"
1526 : " access isn't contiguous.\n");
1527 24770 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
1528 24770 : return;
1529 : }
1530 :
1531 197156 : if (!VECTOR_MODE_P (vecmode))
1532 : {
1533 0 : if (dump_enabled_p ())
1534 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1535 : "can't operate on partial vectors when emulating"
1536 : " vector operations.\n");
1537 0 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
1538 0 : return;
1539 : }
1540 :
1541 : /* We might load more scalars than we need for permuting SLP loads.
1542 : We checked in get_load_store_type that the extra elements
1543 : don't leak into a new vector. */
1544 262763 : auto group_memory_nvectors = [](poly_uint64 size, poly_uint64 nunits)
1545 : {
1546 65607 : unsigned int nvectors;
1547 131214 : if (can_div_away_from_zero_p (size, nunits, &nvectors))
1548 65607 : return nvectors;
1549 : gcc_unreachable ();
1550 : };
1551 :
1552 197156 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
1553 197156 : poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
1554 197156 : machine_mode mask_mode;
1555 197156 : machine_mode vmode;
1556 197156 : bool using_partial_vectors_p = false;
1557 197156 : if (get_len_load_store_mode
1558 197156 : (vecmode, is_load, nullptr, elsvals).exists (&vmode))
1559 : {
1560 0 : nvectors = group_memory_nvectors (group_size * vf, nunits);
1561 0 : unsigned factor = (vecmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vecmode);
1562 0 : vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, factor);
1563 0 : using_partial_vectors_p = true;
1564 : }
1565 262763 : else if (targetm.vectorize.get_mask_mode (vecmode).exists (&mask_mode)
1566 197156 : && can_vec_mask_load_store_p (vecmode, mask_mode, is_load, NULL,
1567 : elsvals))
1568 : {
1569 65607 : nvectors = group_memory_nvectors (group_size * vf, nunits);
1570 65607 : vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype, scalar_mask);
1571 65607 : using_partial_vectors_p = true;
1572 : }
1573 :
1574 65607 : if (!using_partial_vectors_p)
1575 : {
1576 131549 : if (dump_enabled_p ())
1577 11251 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1578 : "can't operate on partial vectors because the"
1579 : " target doesn't have the appropriate partial"
1580 : " vectorization load or store.\n");
1581 131549 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
1582 : }
1583 : }
1584 :
1585 : /* Return the mask input to a masked load or store. VEC_MASK is the vectorized
1586 : form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask
1587 : that needs to be applied to all loads and stores in a vectorized loop.
1588 : Return VEC_MASK if LOOP_MASK is null or if VEC_MASK is already masked,
1589 : otherwise return VEC_MASK & LOOP_MASK.
1590 :
1591 : MASK_TYPE is the type of both masks. If new statements are needed,
1592 : insert them before GSI. */
1593 :
1594 : tree
1595 1706 : prepare_vec_mask (loop_vec_info loop_vinfo, tree mask_type, tree loop_mask,
1596 : tree vec_mask, gimple_stmt_iterator *gsi)
1597 : {
1598 1706 : gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask)));
1599 1706 : if (!loop_mask)
1600 : return vec_mask;
1601 :
1602 139 : gcc_assert (TREE_TYPE (loop_mask) == mask_type);
1603 :
1604 139 : if (loop_vinfo->vec_cond_masked_set.contains ({ vec_mask, loop_mask }))
1605 : return vec_mask;
1606 :
1607 139 : tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and");
1608 139 : gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR,
1609 : vec_mask, loop_mask);
1610 :
1611 139 : gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT);
1612 139 : return and_res;
1613 : }
1614 :
1615 : /* Determine whether we can use a gather load or scatter store to vectorize
1616 : strided load or store STMT_INFO by truncating the current offset to a
1617 : smaller width. We need to be able to construct an offset vector:
1618 :
1619 : { 0, X, X*2, X*3, ... }
1620 :
1621 : without loss of precision, where X is STMT_INFO's DR_STEP.
1622 :
1623 : Return true if this is possible, describing the gather load or scatter
1624 : store in GS_INFO. MASKED_P is true if the load or store is conditional.
1625 :
1626 : If we can use gather/scatter and ELSVALS is nonzero the supported
1627 : else values will be stored in the vector ELSVALS points to. */
1628 :
1629 : static bool
1630 62917 : vect_truncate_gather_scatter_offset (stmt_vec_info stmt_info, tree vectype,
1631 : loop_vec_info loop_vinfo, bool masked_p,
1632 : gather_scatter_info *gs_info,
1633 : vec<int> *elsvals)
1634 : {
1635 62917 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info);
1636 62917 : data_reference *dr = dr_info->dr;
1637 62917 : tree step = DR_STEP (dr);
1638 62917 : if (TREE_CODE (step) != INTEGER_CST)
1639 : {
1640 : /* ??? Perhaps we could use range information here? */
1641 28184 : if (dump_enabled_p ())
1642 229 : dump_printf_loc (MSG_NOTE, vect_location,
1643 : "cannot truncate variable step.\n");
1644 28184 : return false;
1645 : }
1646 :
1647 : /* Get the number of bits in an element. */
1648 34733 : scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype));
1649 34733 : unsigned int element_bits = GET_MODE_BITSIZE (element_mode);
1650 :
1651 : /* Set COUNT to the upper limit on the number of elements - 1.
1652 : Start with the maximum vectorization factor. */
1653 34733 : unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1;
1654 :
1655 : /* Try lowering COUNT to the number of scalar latch iterations. */
1656 34733 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1657 34733 : widest_int max_iters;
1658 34733 : if (max_loop_iterations (loop, &max_iters)
1659 68753 : && max_iters < count)
1660 2115 : count = max_iters.to_shwi ();
1661 :
1662 : /* Try scales of 1 and the element size. */
1663 34733 : unsigned int scales[] = { 1, vect_get_scalar_dr_size (dr_info) };
1664 34733 : wi::overflow_type overflow = wi::OVF_NONE;
1665 104199 : for (int i = 0; i < 2; ++i)
1666 : {
1667 69466 : unsigned int scale = scales[i];
1668 69466 : widest_int factor;
1669 69466 : if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor))
1670 0 : continue;
1671 :
1672 : /* Determine the minimum precision of (COUNT - 1) * STEP / SCALE. */
1673 69466 : widest_int range = wi::mul (count, factor, SIGNED, &overflow);
1674 69466 : if (overflow)
1675 0 : continue;
1676 69466 : signop sign = range >= 0 ? UNSIGNED : SIGNED;
1677 69466 : unsigned int min_offset_bits = wi::min_precision (range, sign);
1678 :
1679 : /* Find the narrowest viable offset type. */
1680 69466 : unsigned int offset_bits = 1U << ceil_log2 (min_offset_bits);
1681 69466 : tree offset_type = build_nonstandard_integer_type (offset_bits,
1682 : sign == UNSIGNED);
1683 :
1684 : /* See whether the target supports the operation with an offset
1685 : no narrower than OFFSET_TYPE. */
1686 69466 : tree memory_type = TREE_TYPE (DR_REF (dr));
1687 69466 : tree tmp_offset_vectype;
1688 69466 : int tmp_scale;
1689 69466 : if (!vect_gather_scatter_fn_p (loop_vinfo, DR_IS_READ (dr), masked_p,
1690 : vectype, memory_type, offset_type,
1691 : scale, &tmp_scale,
1692 : &gs_info->ifn, &gs_info->offset_vectype,
1693 : &tmp_offset_vectype, elsvals)
1694 69466 : || gs_info->ifn == IFN_LAST)
1695 69466 : continue;
1696 :
1697 0 : gs_info->decl = NULL_TREE;
1698 : /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET,
1699 : but we don't need to store that here. */
1700 0 : gs_info->base = NULL_TREE;
1701 0 : gs_info->alias_ptr = build_int_cst
1702 0 : (reference_alias_ptr_type (DR_REF (dr)),
1703 0 : get_object_alignment (DR_REF (dr)));
1704 0 : gs_info->element_type = TREE_TYPE (vectype);
1705 0 : gs_info->offset = fold_convert (offset_type, step);
1706 0 : gs_info->scale = scale;
1707 0 : gs_info->memory_type = memory_type;
1708 0 : return true;
1709 138932 : }
1710 :
1711 34733 : if (overflow && dump_enabled_p ())
1712 0 : dump_printf_loc (MSG_NOTE, vect_location,
1713 : "truncating gather/scatter offset to %d bits"
1714 : " might change its value.\n", element_bits);
1715 :
1716 : return false;
1717 34733 : }
1718 :
1719 : /* Return true if we can use gather/scatter or strided internal functions
1720 : to vectorize STMT_INFO, which is a grouped or strided load or store
1721 : with multiple lanes and will be implemented by a type-punned access
1722 : of a vector with element size that matches the number of lanes.
1723 :
1724 : MASKED_P is true if load or store is conditional.
1725 : When returning true, fill in GS_INFO with the information required to
1726 : perform the operation. Also, store the punning type in PUNNED_VECTYPE.
1727 :
1728 : If successful and ELSVALS is nonzero the supported
1729 : else values will be stored in the vector ELSVALS points to. */
1730 :
1731 : static bool
1732 3663 : vect_use_grouped_gather (dr_vec_info *dr_info, tree vectype,
1733 : loop_vec_info loop_vinfo, bool masked_p,
1734 : unsigned int nelts,
1735 : gather_scatter_info *info, vec<int> *elsvals,
1736 : tree *pun_vectype)
1737 : {
1738 3663 : data_reference *dr = dr_info->dr;
1739 :
1740 : /* TODO: We can support nelts > BITS_PER_UNIT or non-power-of-two by
1741 : multiple gathers/scatter. */
1742 7024 : if (nelts > BITS_PER_UNIT || !pow2p_hwi (nelts))
1743 : return false;
1744 :
1745 : /* Pun the vectype with one of the same size but an element spanning
1746 : NELTS elements of VECTYPE.
1747 : The punned type of a V16QI with NELTS = 4 would be V4SI.
1748 : */
1749 3097 : tree tmp;
1750 3097 : unsigned int pieces;
1751 3097 : if (!can_div_trunc_p (TYPE_VECTOR_SUBPARTS (vectype), nelts, &pieces)
1752 3097 : || !pieces)
1753 193 : return false;
1754 :
1755 2904 : *pun_vectype = vector_vector_composition_type (vectype, pieces, &tmp, true);
1756 :
1757 2904 : if (!*pun_vectype || !VECTOR_TYPE_P (*pun_vectype))
1758 : return false;
1759 :
1760 2540 : internal_fn ifn;
1761 2540 : tree offset_vectype = *pun_vectype;
1762 :
1763 1641 : internal_fn strided_ifn = DR_IS_READ (dr)
1764 2540 : ? IFN_MASK_LEN_STRIDED_LOAD : IFN_MASK_LEN_STRIDED_STORE;
1765 :
1766 : /* Check if we have a gather/scatter with the new type. We're just trying
1767 : with the type itself as offset for now. If not, check if we have a
1768 : strided load/store. These have fewer constraints (for example no offset
1769 : type must exist) so it is possible that even though a gather/scatter is
1770 : not available we still have a strided load/store. */
1771 2540 : bool ok = false;
1772 2540 : tree tmp_vectype;
1773 2540 : int tmp_scale;
1774 2540 : if (vect_gather_scatter_fn_p
1775 2540 : (loop_vinfo, DR_IS_READ (dr), masked_p, *pun_vectype,
1776 2540 : TREE_TYPE (*pun_vectype), *pun_vectype, 1, &tmp_scale, &ifn,
1777 : &offset_vectype, &tmp_vectype, elsvals))
1778 : ok = true;
1779 2540 : else if (internal_strided_fn_supported_p (strided_ifn, *pun_vectype,
1780 : elsvals))
1781 : {
1782 : /* Use gather/scatter IFNs, vect_get_strided_load_store_ops
1783 : will switch back to the strided variants. */
1784 0 : ifn = DR_IS_READ (dr) ? IFN_MASK_LEN_GATHER_LOAD :
1785 : IFN_MASK_LEN_SCATTER_STORE;
1786 0 : ok = true;
1787 : }
1788 :
1789 0 : if (ok)
1790 : {
1791 0 : info->ifn = ifn;
1792 0 : info->decl = NULL_TREE;
1793 0 : info->base = dr->ref;
1794 0 : info->alias_ptr = build_int_cst
1795 0 : (reference_alias_ptr_type (DR_REF (dr)),
1796 0 : get_object_alignment (DR_REF (dr)));
1797 0 : info->element_type = TREE_TYPE (*pun_vectype);
1798 0 : info->offset_vectype = offset_vectype;
1799 : /* No need to set the offset, vect_get_strided_load_store_ops
1800 : will do that. */
1801 0 : info->scale = 1;
1802 0 : info->memory_type = TREE_TYPE (DR_REF (dr));
1803 0 : return true;
1804 : }
1805 :
1806 : return false;
1807 : }
1808 :
1809 :
1810 : /* Return true if we can use gather/scatter internal functions to
1811 : vectorize STMT_INFO, which is a grouped or strided load or store.
1812 : MASKED_P is true if load or store is conditional. When returning
1813 : true, fill in GS_INFO with the information required to perform the
1814 : operation.
1815 :
1816 : If we can use gather/scatter and ELSVALS is nonzero the supported
1817 : else values will be stored in the vector ELSVALS points to. */
1818 :
1819 : static bool
1820 62917 : vect_use_strided_gather_scatters_p (stmt_vec_info stmt_info, tree vectype,
1821 : loop_vec_info loop_vinfo, bool masked_p,
1822 : gather_scatter_info *gs_info,
1823 : vec<int> *elsvals,
1824 : unsigned int group_size,
1825 : bool single_element_p)
1826 : {
1827 62917 : if (!vect_check_gather_scatter (stmt_info, vectype,
1828 : loop_vinfo, gs_info, elsvals)
1829 62917 : || gs_info->ifn == IFN_LAST)
1830 : {
1831 62917 : if (!vect_truncate_gather_scatter_offset (stmt_info, vectype, loop_vinfo,
1832 : masked_p, gs_info, elsvals))
1833 : return false;
1834 : }
1835 :
1836 0 : if (!single_element_p
1837 0 : && !targetm.vectorize.prefer_gather_scatter (TYPE_MODE (vectype),
1838 : gs_info->scale,
1839 : group_size))
1840 : return false;
1841 :
1842 0 : if (dump_enabled_p ())
1843 0 : dump_printf_loc (MSG_NOTE, vect_location,
1844 : "using gather/scatter for strided/grouped access,"
1845 : " scale = %d\n", gs_info->scale);
1846 :
1847 : return true;
1848 : }
1849 :
1850 : /* STMT_INFO is a non-strided load or store, meaning that it accesses
1851 : elements with a known constant step. Return -1 if that step
1852 : is negative, 0 if it is zero, and 1 if it is greater than zero. */
1853 :
1854 : int
1855 1351389 : compare_step_with_zero (vec_info *vinfo, stmt_vec_info stmt_info)
1856 : {
1857 1351389 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info);
1858 1351389 : return tree_int_cst_compare (vect_dr_behavior (vinfo, dr_info)->step,
1859 1351389 : size_zero_node);
1860 : }
1861 :
1862 : /* If the target supports a permute mask that reverses the elements in
1863 : a vector of type VECTYPE, return that mask, otherwise return null. */
1864 :
1865 : tree
1866 8799 : perm_mask_for_reverse (tree vectype)
1867 : {
1868 8799 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
1869 :
1870 : /* The encoding has a single stepped pattern. */
1871 8799 : vec_perm_builder sel (nunits, 1, 3);
1872 35196 : for (int i = 0; i < 3; ++i)
1873 26397 : sel.quick_push (nunits - 1 - i);
1874 :
1875 8799 : vec_perm_indices indices (sel, 1, nunits);
1876 8799 : if (!can_vec_perm_const_p (TYPE_MODE (vectype), TYPE_MODE (vectype),
1877 : indices))
1878 : return NULL_TREE;
1879 7653 : return vect_gen_perm_mask_checked (vectype, indices);
1880 8799 : }
1881 :
1882 : /* A subroutine of get_load_store_type, with a subset of the same
1883 : arguments. Handle the case where STMT_INFO is a load or store that
1884 : accesses consecutive elements with a negative step. Sets *POFFSET
1885 : to the offset to be applied to the DR for the first access. */
1886 :
1887 : static vect_memory_access_type
1888 11435 : get_negative_load_store_type (vec_info *vinfo,
1889 : stmt_vec_info stmt_info, tree vectype,
1890 : vec_load_store_type vls_type,
1891 : unsigned int ncopies, poly_int64 *poffset)
1892 : {
1893 11435 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info);
1894 11435 : dr_alignment_support alignment_support_scheme;
1895 :
1896 11435 : if (ncopies > 1)
1897 : {
1898 0 : if (dump_enabled_p ())
1899 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1900 : "multiple types with negative step.\n");
1901 0 : return VMAT_ELEMENTWISE;
1902 : }
1903 :
1904 : /* For backward running DRs the first access in vectype actually is
1905 : N-1 elements before the address of the DR. */
1906 11435 : *poffset = ((-TYPE_VECTOR_SUBPARTS (vectype) + 1)
1907 11435 : * TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype))));
1908 :
1909 11435 : int misalignment = dr_misalignment (dr_info, vectype, *poffset);
1910 11435 : alignment_support_scheme
1911 11435 : = vect_supportable_dr_alignment (vinfo, dr_info, vectype, misalignment);
1912 11435 : if (alignment_support_scheme != dr_aligned
1913 11435 : && alignment_support_scheme != dr_unaligned_supported)
1914 : {
1915 4478 : if (dump_enabled_p ())
1916 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1917 : "negative step but alignment required.\n");
1918 4478 : *poffset = 0;
1919 4478 : return VMAT_ELEMENTWISE;
1920 : }
1921 :
1922 6957 : if (vls_type == VLS_STORE_INVARIANT)
1923 : {
1924 725 : if (dump_enabled_p ())
1925 21 : dump_printf_loc (MSG_NOTE, vect_location,
1926 : "negative step with invariant source;"
1927 : " no permute needed.\n");
1928 725 : return VMAT_CONTIGUOUS_DOWN;
1929 : }
1930 :
1931 6232 : if (!perm_mask_for_reverse (vectype))
1932 : {
1933 1146 : if (dump_enabled_p ())
1934 52 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1935 : "negative step and reversing not supported.\n");
1936 1146 : *poffset = 0;
1937 1146 : return VMAT_ELEMENTWISE;
1938 : }
1939 :
1940 : return VMAT_CONTIGUOUS_REVERSE;
1941 : }
1942 :
1943 : /* STMT_INFO is either a masked or unconditional store. Return the value
1944 : being stored. */
1945 :
1946 : tree
1947 0 : vect_get_store_rhs (stmt_vec_info stmt_info)
1948 : {
1949 0 : if (gassign *assign = dyn_cast <gassign *> (stmt_info->stmt))
1950 : {
1951 0 : gcc_assert (gimple_assign_single_p (assign));
1952 0 : return gimple_assign_rhs1 (assign);
1953 : }
1954 0 : if (gcall *call = dyn_cast <gcall *> (stmt_info->stmt))
1955 : {
1956 0 : internal_fn ifn = gimple_call_internal_fn (call);
1957 0 : int index = internal_fn_stored_value_index (ifn);
1958 0 : gcc_assert (index >= 0);
1959 0 : return gimple_call_arg (call, index);
1960 : }
1961 0 : gcc_unreachable ();
1962 : }
1963 :
1964 : /* Function VECTOR_VECTOR_COMPOSITION_TYPE
1965 :
1966 : This function returns a vector type which can be composed with NELTS pieces,
1967 : whose type is recorded in PTYPE. VTYPE should be a vector type, and has the
1968 : same vector size as the return vector. It checks target whether supports
1969 : pieces-size vector mode for construction firstly, if target fails to, check
1970 : pieces-size scalar mode for construction further. It returns NULL_TREE if
1971 : fails to find the available composition. If the caller only wants scalar
1972 : pieces where PTYPE e.g. is a possible gather/scatter element type
1973 : SCALAR_PTYPE_ONLY must be true.
1974 :
1975 : For example, for (vtype=V16QI, nelts=4), we can probably get:
1976 : - V16QI with PTYPE V4QI.
1977 : - V4SI with PTYPE SI.
1978 : - NULL_TREE. */
1979 :
1980 : static tree
1981 13078 : vector_vector_composition_type (tree vtype, poly_uint64 nelts, tree *ptype,
1982 : bool scalar_ptype_only)
1983 : {
1984 13078 : gcc_assert (VECTOR_TYPE_P (vtype));
1985 13078 : gcc_assert (known_gt (nelts, 0U));
1986 :
1987 13078 : machine_mode vmode = TYPE_MODE (vtype);
1988 13078 : if (!VECTOR_MODE_P (vmode))
1989 : return NULL_TREE;
1990 :
1991 : /* When we are asked to compose the vector from its components let
1992 : that happen directly. */
1993 13078 : if (known_eq (TYPE_VECTOR_SUBPARTS (vtype), nelts))
1994 : {
1995 5804 : *ptype = TREE_TYPE (vtype);
1996 5804 : return vtype;
1997 : }
1998 :
1999 14548 : poly_uint64 vbsize = GET_MODE_BITSIZE (vmode);
2000 7274 : unsigned int pbsize;
2001 7274 : if (constant_multiple_p (vbsize, nelts, &pbsize))
2002 : {
2003 : /* First check if vec_init optab supports construction from
2004 : vector pieces directly. */
2005 7274 : scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vtype));
2006 14548 : poly_uint64 inelts = pbsize / GET_MODE_BITSIZE (elmode);
2007 7274 : machine_mode rmode;
2008 7274 : if (!scalar_ptype_only
2009 4370 : && related_vector_mode (vmode, elmode, inelts).exists (&rmode)
2010 11266 : && (convert_optab_handler (vec_init_optab, vmode, rmode)
2011 : != CODE_FOR_nothing))
2012 : {
2013 3375 : *ptype = build_vector_type (TREE_TYPE (vtype), inelts);
2014 3375 : return vtype;
2015 : }
2016 :
2017 : /* Otherwise check if exists an integer type of the same piece size and
2018 : if vec_init optab supports construction from it directly. */
2019 3899 : if (int_mode_for_size (pbsize, 0).exists (&elmode)
2020 3899 : && related_vector_mode (vmode, elmode, nelts).exists (&rmode))
2021 : {
2022 3499 : if (scalar_ptype_only
2023 3499 : || convert_optab_handler (vec_init_optab, rmode, elmode)
2024 : != CODE_FOR_nothing)
2025 : {
2026 3499 : *ptype = build_nonstandard_integer_type (pbsize, 1);
2027 3499 : return build_vector_type (*ptype, nelts);
2028 : }
2029 : }
2030 : }
2031 :
2032 : return NULL_TREE;
2033 : }
2034 :
2035 : /* Check if the load permutation of NODE only refers to a consecutive
2036 : subset of the group indices where GROUP_SIZE is the size of the
2037 : dataref's group. We also assert that the length of the permutation
2038 : divides the group size and is a power of two.
2039 : Such load permutations can be elided in strided access schemes as
2040 : we can "jump over" the gap they leave. */
2041 :
2042 : bool
2043 43043 : has_consecutive_load_permutation (slp_tree node, unsigned group_size)
2044 : {
2045 43043 : load_permutation_t perm = SLP_TREE_LOAD_PERMUTATION (node);
2046 43043 : if (!perm.exists ()
2047 1839 : || perm.length () <= 1
2048 411 : || !pow2p_hwi (perm.length ())
2049 43440 : || group_size % perm.length ())
2050 : return false;
2051 :
2052 354 : return vect_load_perm_consecutive_p (node);
2053 : }
2054 :
2055 :
2056 : /* Analyze load or store SLP_NODE of type VLS_TYPE. Return true
2057 : if there is a memory access type that the vectorized form can use,
2058 : storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers
2059 : or scatters, fill in GS_INFO accordingly. In addition
2060 : *ALIGNMENT_SUPPORT_SCHEME is filled out and false is returned if
2061 : the target does not support the alignment scheme. *MISALIGNMENT
2062 : is set according to the alignment of the access (including
2063 : DR_MISALIGNMENT_UNKNOWN when it is unknown).
2064 :
2065 : MASKED_P is true if the statement is conditional on a vectorized mask.
2066 : VECTYPE is the vector type that the vectorized statements will use.
2067 :
2068 : If ELSVALS is nonzero the supported else values will be stored in the
2069 : vector ELSVALS points to. */
2070 :
2071 : static bool
2072 1246618 : get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
2073 : tree vectype, slp_tree slp_node,
2074 : bool masked_p, vec_load_store_type vls_type,
2075 : vect_load_store_data *ls)
2076 : {
2077 1246618 : vect_memory_access_type *memory_access_type = &ls->memory_access_type;
2078 1246618 : poly_int64 *poffset = &ls->poffset;
2079 1246618 : dr_alignment_support *alignment_support_scheme
2080 : = &ls->alignment_support_scheme;
2081 1246618 : int *misalignment = &ls->misalignment;
2082 1246618 : internal_fn *lanes_ifn = &ls->lanes_ifn;
2083 1246618 : vec<int> *elsvals = &ls->elsvals;
2084 1246618 : tree *ls_type = &ls->ls_type;
2085 1246618 : bool *slp_perm = &ls->slp_perm;
2086 1246618 : unsigned *n_perms = &ls->n_perms;
2087 1246618 : unsigned *n_loads = &ls->n_loads;
2088 1246618 : tree *supported_offset_vectype = &ls->supported_offset_vectype;
2089 1246618 : int *supported_scale = &ls->supported_scale;
2090 1246618 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
2091 1246618 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
2092 1246618 : class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
2093 1246618 : stmt_vec_info first_stmt_info;
2094 1246618 : unsigned int group_size;
2095 1246618 : unsigned HOST_WIDE_INT gap;
2096 1246618 : bool single_element_p;
2097 1246618 : poly_int64 neg_ldst_offset = 0;
2098 :
2099 1246618 : *misalignment = DR_MISALIGNMENT_UNKNOWN;
2100 1246618 : *poffset = 0;
2101 1246618 : *ls_type = NULL_TREE;
2102 1246618 : *slp_perm = false;
2103 1246618 : *n_perms = -1U;
2104 1246618 : *n_loads = -1U;
2105 1246618 : ls->subchain_p = false;
2106 :
2107 1246618 : bool perm_ok = true;
2108 1246618 : poly_int64 vf = loop_vinfo ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) : 1;
2109 :
2110 1246618 : if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())
2111 66547 : perm_ok = vect_transform_slp_perm_load (vinfo, slp_node, vNULL, NULL,
2112 66547 : vf, true, n_perms, n_loads);
2113 :
2114 1246618 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2115 : {
2116 858534 : first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
2117 858534 : group_size = DR_GROUP_SIZE (first_stmt_info);
2118 858534 : gap = DR_GROUP_GAP (first_stmt_info);
2119 858534 : single_element_p = (stmt_info == first_stmt_info
2120 858534 : && !DR_GROUP_NEXT_ELEMENT (stmt_info));
2121 : }
2122 : else
2123 : {
2124 : first_stmt_info = stmt_info;
2125 : group_size = 1;
2126 : gap = 0;
2127 : single_element_p = true;
2128 : }
2129 1246618 : dr_vec_info *first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info);
2130 :
2131 : /* True if the vectorized statements would access beyond the last
2132 : statement in the group. */
2133 1246618 : bool overrun_p = false;
2134 :
2135 : /* True if we can cope with such overrun by peeling for gaps, so that
2136 : there is at least one final scalar iteration after the vector loop. */
2137 2493236 : bool can_overrun_p = (!masked_p
2138 1246618 : && vls_type == VLS_LOAD
2139 471984 : && loop_vinfo
2140 1588700 : && !loop->inner);
2141 :
2142 : /* There can only be a gap at the end of the group if the stride is
2143 : known at compile time. */
2144 1246618 : gcc_assert (!STMT_VINFO_STRIDED_P (first_stmt_info) || gap == 0);
2145 :
2146 : /* For SLP vectorization we directly vectorize a subchain
2147 : without permutation. */
2148 1246618 : if (! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())
2149 1180071 : first_dr_info = STMT_VINFO_DR_INFO (SLP_TREE_SCALAR_STMTS (slp_node)[0]);
2150 :
2151 1246618 : if (STMT_VINFO_STRIDED_P (first_stmt_info))
2152 : {
2153 : /* Try to use consecutive accesses of as many elements as possible,
2154 : separated by the stride, until we have a complete vector.
2155 : Fall back to scalar accesses if that isn't possible. */
2156 43043 : *memory_access_type = VMAT_STRIDED_SLP;
2157 :
2158 : /* If the load permutation is consecutive we can reduce the group to
2159 : the elements the permutation accesses. Then we release the
2160 : permutation. */
2161 43043 : if (has_consecutive_load_permutation (slp_node, group_size))
2162 : {
2163 20 : ls->subchain_p = true;
2164 20 : group_size = SLP_TREE_LANES (slp_node);
2165 20 : SLP_TREE_LOAD_PERMUTATION (slp_node).release ();
2166 : }
2167 : }
2168 1203575 : else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
2169 : {
2170 10330 : slp_tree offset_node = SLP_TREE_CHILDREN (slp_node)[0];
2171 10330 : tree offset_vectype = SLP_TREE_VECTYPE (offset_node);
2172 10330 : int scale = SLP_TREE_GS_SCALE (slp_node);
2173 10330 : tree memory_type = TREE_TYPE (DR_REF (first_dr_info->dr));
2174 10330 : tree tem;
2175 10330 : if (vect_gather_scatter_fn_p (loop_vinfo, vls_type == VLS_LOAD,
2176 : masked_p, vectype, memory_type,
2177 : offset_vectype, scale, supported_scale,
2178 : &ls->gs.ifn, &tem,
2179 : supported_offset_vectype, elsvals))
2180 : {
2181 0 : if (dump_enabled_p ())
2182 : {
2183 0 : dump_printf_loc (MSG_NOTE, vect_location,
2184 : "gather/scatter with required "
2185 : "offset type "
2186 : "%T and offset scale %d.\n",
2187 : offset_vectype, scale);
2188 0 : if (*supported_offset_vectype)
2189 0 : dump_printf_loc (MSG_NOTE, vect_location,
2190 : " target supports offset type %T.\n",
2191 : *supported_offset_vectype);
2192 0 : if (*supported_scale)
2193 0 : dump_printf_loc (MSG_NOTE, vect_location,
2194 : " target supports offset scale %d.\n",
2195 : *supported_scale);
2196 : }
2197 0 : *memory_access_type = VMAT_GATHER_SCATTER_IFN;
2198 : }
2199 10330 : else if (vls_type == VLS_LOAD
2200 10330 : ? (targetm.vectorize.builtin_gather
2201 8912 : && (ls->gs.decl
2202 8912 : = targetm.vectorize.builtin_gather (vectype,
2203 8912 : TREE_TYPE
2204 : (offset_vectype),
2205 : scale)))
2206 1418 : : (targetm.vectorize.builtin_scatter
2207 1418 : && (ls->gs.decl
2208 1418 : = targetm.vectorize.builtin_scatter (vectype,
2209 1418 : TREE_TYPE
2210 : (offset_vectype),
2211 : scale))))
2212 345 : *memory_access_type = VMAT_GATHER_SCATTER_LEGACY;
2213 : else
2214 : {
2215 : /* GATHER_SCATTER_EMULATED_P. */
2216 9985 : if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant ()
2217 9985 : || !TYPE_VECTOR_SUBPARTS (offset_vectype).is_constant ()
2218 9985 : || VECTOR_BOOLEAN_TYPE_P (offset_vectype)
2219 9985 : || !constant_multiple_p (TYPE_VECTOR_SUBPARTS (offset_vectype),
2220 9985 : TYPE_VECTOR_SUBPARTS (vectype)))
2221 : {
2222 2600 : if (dump_enabled_p ())
2223 450 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2224 : "unsupported vector types for emulated "
2225 : "gather.\n");
2226 2600 : return false;
2227 : }
2228 7385 : *memory_access_type = VMAT_GATHER_SCATTER_EMULATED;
2229 : }
2230 : }
2231 : else
2232 : {
2233 1193245 : int cmp = compare_step_with_zero (vinfo, stmt_info);
2234 1193245 : if (cmp < 0)
2235 : {
2236 11574 : if (single_element_p)
2237 : /* ??? The VMAT_CONTIGUOUS_REVERSE code generation is
2238 : only correct for single element "interleaving" SLP. */
2239 11435 : *memory_access_type = get_negative_load_store_type
2240 11435 : (vinfo, stmt_info, vectype, vls_type, 1,
2241 : &neg_ldst_offset);
2242 : else
2243 : /* We can fall back to VMAT_STRIDED_SLP since that does
2244 : not care whether the stride between the group instances
2245 : is positive or negative. */
2246 139 : *memory_access_type = VMAT_STRIDED_SLP;
2247 : }
2248 1181671 : else if (cmp == 0 && loop_vinfo)
2249 : {
2250 3046 : gcc_assert (vls_type == VLS_LOAD);
2251 3046 : *memory_access_type = VMAT_INVARIANT;
2252 : }
2253 : /* Try using LOAD/STORE_LANES. */
2254 1178625 : else if (slp_node->ldst_lanes
2255 1178625 : && (*lanes_ifn
2256 0 : = (vls_type == VLS_LOAD
2257 0 : ? vect_load_lanes_supported (vectype, group_size,
2258 : masked_p, elsvals)
2259 0 : : vect_store_lanes_supported (vectype, group_size,
2260 : masked_p))) != IFN_LAST)
2261 0 : *memory_access_type = VMAT_LOAD_STORE_LANES;
2262 1178625 : else if (!loop_vinfo && slp_node->avoid_stlf_fail)
2263 : {
2264 74 : *memory_access_type = VMAT_ELEMENTWISE;
2265 74 : if (dump_enabled_p ())
2266 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2267 : "using element-wise load to avoid disrupting "
2268 : "cross iteration store-to-load forwarding\n");
2269 : }
2270 : else
2271 1178551 : *memory_access_type = VMAT_CONTIGUOUS;
2272 :
2273 : /* If this is single-element interleaving with an element
2274 : distance that leaves unused vector loads around fall back
2275 : to elementwise access if possible - we otherwise least
2276 : create very sub-optimal code in that case (and
2277 : blow up memory, see PR65518). */
2278 1193245 : if (loop_vinfo
2279 1193245 : && single_element_p
2280 369777 : && (*memory_access_type == VMAT_CONTIGUOUS
2281 14481 : || *memory_access_type == VMAT_CONTIGUOUS_REVERSE)
2282 1563022 : && maybe_gt (group_size, TYPE_VECTOR_SUBPARTS (vectype)))
2283 : {
2284 17760 : *memory_access_type = VMAT_ELEMENTWISE;
2285 17760 : if (dump_enabled_p ())
2286 196 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2287 : "single-element interleaving not supported "
2288 : "for not adjacent vector loads, using "
2289 : "elementwise access\n");
2290 : }
2291 :
2292 : /* Also fall back to elementwise access in case we did not lower a
2293 : permutation and cannot code generate it. */
2294 1193245 : if (loop_vinfo
2295 415449 : && *memory_access_type != VMAT_ELEMENTWISE
2296 392065 : && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()
2297 1215723 : && !perm_ok)
2298 : {
2299 1998 : *memory_access_type = VMAT_ELEMENTWISE;
2300 1998 : if (dump_enabled_p ())
2301 246 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2302 : "permutation not supported, using elementwise "
2303 : "access\n");
2304 : }
2305 :
2306 415449 : overrun_p = (loop_vinfo && gap != 0
2307 1235748 : && *memory_access_type != VMAT_ELEMENTWISE);
2308 1193245 : if (overrun_p && vls_type != VLS_LOAD)
2309 : {
2310 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2311 : "Grouped store with gaps requires"
2312 : " non-consecutive accesses\n");
2313 9 : return false;
2314 : }
2315 :
2316 1193245 : unsigned HOST_WIDE_INT dr_size = vect_get_scalar_dr_size (first_dr_info);
2317 1193245 : poly_int64 off = 0;
2318 1193245 : if (*memory_access_type == VMAT_CONTIGUOUS_REVERSE)
2319 4927 : off = (TYPE_VECTOR_SUBPARTS (vectype) - 1) * -dr_size;
2320 :
2321 : /* An overrun is fine if the trailing elements are smaller
2322 : than the alignment boundary B. Every vector access will
2323 : be a multiple of B and so we are guaranteed to access a
2324 : non-gap element in the same B-sized block. */
2325 1193245 : if (overrun_p
2326 1193245 : && gap < (vect_known_alignment_in_bytes (first_dr_info,
2327 22649 : vectype, off) / dr_size))
2328 : overrun_p = false;
2329 :
2330 : /* When we have a contiguous access across loop iterations
2331 : but the access in the loop doesn't cover the full vector
2332 : we can end up with no gap recorded but still excess
2333 : elements accessed, see PR103116. Make sure we peel for
2334 : gaps if necessary and sufficient and give up if not.
2335 :
2336 : If there is a combination of the access not covering the full
2337 : vector and a gap recorded then we may need to peel twice. */
2338 1193245 : bool large_vector_overrun_p = false;
2339 1193245 : if (loop_vinfo
2340 415449 : && (*memory_access_type == VMAT_CONTIGUOUS
2341 34216 : || *memory_access_type == VMAT_CONTIGUOUS_REVERSE)
2342 386160 : && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()
2343 1213423 : && !multiple_p (group_size * LOOP_VINFO_VECT_FACTOR (loop_vinfo),
2344 : nunits))
2345 : large_vector_overrun_p = overrun_p = true;
2346 :
2347 : /* If the gap splits the vector in half and the target
2348 : can do half-vector operations avoid the epilogue peeling
2349 : by simply loading half of the vector only. Usually
2350 : the construction with an upper zero half will be elided. */
2351 1193245 : dr_alignment_support alss;
2352 1193245 : int misalign = dr_misalignment (first_dr_info, vectype, off);
2353 1193245 : tree half_vtype;
2354 1193245 : poly_uint64 remain;
2355 1193245 : unsigned HOST_WIDE_INT tem, num;
2356 1193245 : if (overrun_p
2357 1193245 : && !masked_p
2358 17178 : && *memory_access_type != VMAT_LOAD_STORE_LANES
2359 17178 : && (((alss = vect_supportable_dr_alignment (vinfo, first_dr_info,
2360 : vectype, misalign)))
2361 : == dr_aligned
2362 14737 : || alss == dr_unaligned_supported)
2363 9636 : && can_div_trunc_p (group_size
2364 9636 : * LOOP_VINFO_VECT_FACTOR (loop_vinfo) - gap,
2365 : nunits, &tem, &remain)
2366 1202881 : && (known_eq (remain, 0u)
2367 7252 : || (known_ne (remain, 0u)
2368 5621 : && constant_multiple_p (nunits, remain, &num)
2369 1190861 : && (vector_vector_composition_type (vectype, num, &half_vtype)
2370 : != NULL_TREE))))
2371 8005 : overrun_p = false;
2372 :
2373 1193245 : if (overrun_p && !can_overrun_p)
2374 : {
2375 6 : if (dump_enabled_p ())
2376 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2377 : "Peeling for outer loop is not supported\n");
2378 6 : return false;
2379 : }
2380 :
2381 : /* Peeling for gaps assumes that a single scalar iteration
2382 : is enough to make sure the last vector iteration doesn't
2383 : access excess elements. */
2384 1193239 : if (overrun_p
2385 1193239 : && (!can_div_trunc_p (group_size
2386 9167 : * LOOP_VINFO_VECT_FACTOR (loop_vinfo) - gap,
2387 : nunits, &tem, &remain)
2388 9167 : || maybe_lt (remain + group_size, nunits)))
2389 : {
2390 : /* But peeling a single scalar iteration is enough if
2391 : we can use the next power-of-two sized partial
2392 : access and that is sufficiently small to be covered
2393 : by the single scalar iteration. */
2394 16 : unsigned HOST_WIDE_INT cnunits, cvf, cremain, cpart_size;
2395 16 : if (masked_p
2396 16 : || !nunits.is_constant (&cnunits)
2397 16 : || !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&cvf)
2398 16 : || (((cremain = (group_size * cvf - gap) % cnunits), true)
2399 16 : && ((cpart_size = (1 << ceil_log2 (cremain))), true)
2400 16 : && (cremain + group_size < cpart_size
2401 13 : || (vector_vector_composition_type (vectype,
2402 13 : cnunits / cpart_size,
2403 : &half_vtype)
2404 : == NULL_TREE))))
2405 : {
2406 : /* If all fails we can still resort to niter masking unless
2407 : the vectors used are too big, so enforce the use of
2408 : partial vectors. */
2409 3 : if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
2410 3 : && !large_vector_overrun_p)
2411 : {
2412 0 : if (dump_enabled_p ())
2413 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2414 : "peeling for gaps insufficient for "
2415 : "access unless using partial "
2416 : "vectors\n");
2417 0 : LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P (loop_vinfo) = true;
2418 : }
2419 : else
2420 : {
2421 3 : if (dump_enabled_p ())
2422 3 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2423 : "peeling for gaps insufficient for "
2424 : "access\n");
2425 3 : return false;
2426 : }
2427 : }
2428 13 : else if (large_vector_overrun_p)
2429 : {
2430 13 : if (dump_enabled_p ())
2431 12 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2432 : "can't operate on partial vectors because "
2433 : "only unmasked loads handle access "
2434 : "shortening required because of gaps at "
2435 : "the end of the access\n");
2436 13 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
2437 : }
2438 : }
2439 : }
2440 :
2441 : /* As a last resort, trying using a gather load or scatter store.
2442 :
2443 : ??? Although the code can handle all group sizes correctly,
2444 : it probably isn't a win to use separate strided accesses based
2445 : on nearby locations. Or, even if it's a win over scalar code,
2446 : it might not be a win over vectorizing at a lower VF, if that
2447 : allows us to use contiguous accesses. */
2448 1244009 : vect_memory_access_type grouped_gather_fallback = VMAT_UNINITIALIZED;
2449 1244009 : if (loop_vinfo
2450 466213 : && (*memory_access_type == VMAT_ELEMENTWISE
2451 466213 : || *memory_access_type == VMAT_STRIDED_SLP))
2452 : {
2453 68561 : gather_scatter_info gs_info;
2454 68561 : if (SLP_TREE_LANES (slp_node) == 1
2455 64707 : && (!SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()
2456 21262 : || single_element_p)
2457 131478 : && vect_use_strided_gather_scatters_p (stmt_info, vectype, loop_vinfo,
2458 : masked_p, &gs_info, elsvals,
2459 : group_size, single_element_p))
2460 : {
2461 : /* vect_use_strided_gather_scatters_p does not save the actually
2462 : supported scale and offset type so do that here.
2463 : We need it later in check_load_store_for_partial_vectors
2464 : where we only check if the given internal function is supported
2465 : (to choose whether to use the IFN, LEGACY, or EMULATED flavor
2466 : of gather/scatter) and don't re-do the full analysis. */
2467 0 : tree tmp;
2468 0 : gcc_assert (vect_gather_scatter_fn_p
2469 : (loop_vinfo, vls_type == VLS_LOAD, masked_p, vectype,
2470 : gs_info.memory_type, TREE_TYPE (gs_info.offset),
2471 : gs_info.scale, supported_scale, &gs_info.ifn,
2472 : &tmp, supported_offset_vectype, elsvals));
2473 :
2474 0 : SLP_TREE_GS_SCALE (slp_node) = gs_info.scale;
2475 0 : SLP_TREE_GS_BASE (slp_node) = error_mark_node;
2476 0 : ls->gs.ifn = gs_info.ifn;
2477 0 : ls->strided_offset_vectype = gs_info.offset_vectype;
2478 0 : *memory_access_type = VMAT_GATHER_SCATTER_IFN;
2479 : }
2480 68561 : else if (SLP_TREE_LANES (slp_node) > 1
2481 : && !masked_p
2482 3854 : && !single_element_p
2483 72224 : && vect_use_grouped_gather (STMT_VINFO_DR_INFO (stmt_info),
2484 : vectype, loop_vinfo,
2485 : masked_p, group_size,
2486 : &gs_info, elsvals, ls_type))
2487 : {
2488 0 : SLP_TREE_GS_SCALE (slp_node) = gs_info.scale;
2489 0 : SLP_TREE_GS_BASE (slp_node) = error_mark_node;
2490 0 : grouped_gather_fallback = *memory_access_type;
2491 0 : *memory_access_type = VMAT_GATHER_SCATTER_IFN;
2492 0 : ls->gs.ifn = gs_info.ifn;
2493 0 : vectype = *ls_type;
2494 0 : ls->strided_offset_vectype = gs_info.offset_vectype;
2495 : }
2496 : }
2497 :
2498 1244009 : if (*memory_access_type == VMAT_CONTIGUOUS_DOWN
2499 1244009 : || *memory_access_type == VMAT_CONTIGUOUS_REVERSE)
2500 5648 : *poffset = neg_ldst_offset;
2501 :
2502 1244009 : if (*memory_access_type == VMAT_ELEMENTWISE
2503 1218553 : || *memory_access_type == VMAT_GATHER_SCATTER_LEGACY
2504 1218208 : || *memory_access_type == VMAT_STRIDED_SLP
2505 1175029 : || *memory_access_type == VMAT_INVARIANT)
2506 : {
2507 72026 : *alignment_support_scheme = dr_unaligned_supported;
2508 72026 : *misalignment = DR_MISALIGNMENT_UNKNOWN;
2509 : }
2510 : else
2511 : {
2512 1171983 : if (mat_gather_scatter_p (*memory_access_type)
2513 : && !first_dr_info)
2514 : *misalignment = DR_MISALIGNMENT_UNKNOWN;
2515 : else
2516 1171983 : *misalignment = dr_misalignment (first_dr_info, vectype, *poffset);
2517 1171983 : *alignment_support_scheme
2518 1171983 : = vect_supportable_dr_alignment
2519 1171983 : (vinfo, first_dr_info, vectype, *misalignment,
2520 1171983 : mat_gather_scatter_p (*memory_access_type));
2521 1171983 : if (grouped_gather_fallback != VMAT_UNINITIALIZED
2522 0 : && *alignment_support_scheme != dr_aligned
2523 0 : && *alignment_support_scheme != dr_unaligned_supported)
2524 : {
2525 : /* No supportable alignment for a grouped gather, fall back to the
2526 : original memory access type. Even though VMAT_STRIDED_SLP might
2527 : also try aligned vector loads it can still choose vector
2528 : construction from scalars. */
2529 0 : *memory_access_type = grouped_gather_fallback;
2530 0 : *alignment_support_scheme = dr_unaligned_supported;
2531 0 : *misalignment = DR_MISALIGNMENT_UNKNOWN;
2532 : }
2533 : }
2534 :
2535 1244009 : if (overrun_p)
2536 : {
2537 9164 : gcc_assert (can_overrun_p);
2538 9164 : if (dump_enabled_p ())
2539 503 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2540 : "Data access with gaps requires scalar "
2541 : "epilogue loop\n");
2542 9164 : LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true;
2543 : }
2544 :
2545 1244009 : if ((*memory_access_type == VMAT_ELEMENTWISE
2546 1244009 : || *memory_access_type == VMAT_STRIDED_SLP)
2547 : && !nunits.is_constant ())
2548 : {
2549 : if (dump_enabled_p ())
2550 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2551 : "Not using elementwise accesses due to variable "
2552 : "vectorization factor.\n");
2553 : return false;
2554 : }
2555 :
2556 : /* Checks if all scalar iterations are known to be inbounds. */
2557 1244009 : bool inbounds = DR_SCALAR_KNOWN_BOUNDS (STMT_VINFO_DR_INFO (stmt_info));
2558 :
2559 : /* Check if we support the operation if early breaks are needed. Here we
2560 : must ensure that we don't access any more than the scalar code would
2561 : have. A masked operation would ensure this, so for these load types
2562 : force masking. */
2563 1244009 : if (loop_vinfo
2564 466213 : && dr_safe_speculative_read_required (stmt_info)
2565 1416594 : && LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
2566 : {
2567 172585 : if (mat_gather_scatter_p (*memory_access_type)
2568 172585 : || *memory_access_type == VMAT_STRIDED_SLP)
2569 : {
2570 9006 : if (dump_enabled_p ())
2571 8 : dump_printf_loc (MSG_NOTE, vect_location,
2572 : "early break not supported: cannot peel for "
2573 : "alignment. With non-contiguous memory vectorization"
2574 : " could read out of bounds at %G ",
2575 : STMT_VINFO_STMT (stmt_info));
2576 9006 : if (inbounds)
2577 0 : LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P (loop_vinfo) = true;
2578 : else
2579 : return false;
2580 : }
2581 : /* Block-level alignment: Even though individual accesses of
2582 : VMAT_ELEMENTWISE type do not cause alignment problems, loading the
2583 : whole vector's worth of values in a speculative early-break context
2584 : might cross a page boundary. Set the alignment scheme to `dr_aligned'
2585 : here in order to force checking of whether such accesses meet
2586 : alignment criteria. */
2587 163579 : else if (*memory_access_type == VMAT_ELEMENTWISE && !inbounds)
2588 14981 : *alignment_support_scheme = dr_aligned;
2589 : }
2590 :
2591 : /* If this DR needs alignment for correctness, we must ensure the target
2592 : alignment is a constant power-of-two multiple of the amount read per
2593 : vector iteration or force masking. */
2594 1235003 : if (dr_safe_speculative_read_required (stmt_info)
2595 1235003 : && (*alignment_support_scheme == dr_aligned
2596 96038 : && !mat_gather_scatter_p (*memory_access_type)))
2597 : {
2598 : /* We can only peel for loops, of course. */
2599 96038 : gcc_checking_assert (loop_vinfo);
2600 :
2601 96038 : poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2602 96038 : poly_uint64 read_amount
2603 96038 : = vf * TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
2604 96038 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2605 96038 : read_amount *= group_size;
2606 :
2607 96038 : auto target_alignment
2608 96038 : = DR_TARGET_ALIGNMENT (STMT_VINFO_DR_INFO (stmt_info));
2609 96038 : if (!multiple_p (target_alignment, read_amount))
2610 : {
2611 12716 : if (dump_enabled_p ())
2612 : {
2613 28 : dump_printf_loc (MSG_NOTE, vect_location,
2614 : "desired alignment not met, target was ");
2615 28 : dump_dec (MSG_NOTE, target_alignment);
2616 28 : dump_printf (MSG_NOTE, " previously, but read amount is ");
2617 28 : dump_dec (MSG_NOTE, read_amount);
2618 28 : dump_printf (MSG_NOTE, " at %G.\n", STMT_VINFO_STMT (stmt_info));
2619 : }
2620 14901 : return false;
2621 : }
2622 :
2623 : /* When using a group access the first element may be aligned but the
2624 : subsequent loads may not be. For LOAD_LANES since the loads are based
2625 : on the first DR then all loads in the group are aligned. For
2626 : non-LOAD_LANES this is not the case. In particular a load + blend when
2627 : there are gaps can have the non first loads issued unaligned, even
2628 : partially overlapping the memory of the first load in order to simplify
2629 : the blend. This is what the x86_64 backend does for instance. As
2630 : such only the first load in the group is aligned, the rest are not.
2631 : Because of this the permutes may break the alignment requirements that
2632 : have been set, and as such we should for now, reject them. */
2633 83322 : if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())
2634 : {
2635 2185 : if (dump_enabled_p ())
2636 75 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2637 : "loads with load permutations not supported for "
2638 : "speculative early break loads for %G",
2639 : STMT_VINFO_STMT (stmt_info));
2640 2185 : return false;
2641 : }
2642 :
2643 : /* Reject vectorization if we know the read mount per vector iteration
2644 : exceeds the min page size. */
2645 81137 : if (known_gt (read_amount, (unsigned) param_min_pagesize))
2646 : {
2647 0 : if (dump_enabled_p ())
2648 : {
2649 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2650 : "alignment required for correctness (");
2651 0 : dump_dec (MSG_MISSED_OPTIMIZATION, read_amount);
2652 0 : dump_printf (MSG_NOTE, ") may exceed page size.\n");
2653 : }
2654 0 : return false;
2655 : }
2656 :
2657 81137 : if (!vf.is_constant ())
2658 : {
2659 : /* For VLA modes, we need a runtime check to ensure any speculative
2660 : read amount does not exceed the page size. Here we record the max
2661 : possible read amount for the check. */
2662 : if (maybe_gt (read_amount,
2663 : LOOP_VINFO_MAX_SPEC_READ_AMOUNT (loop_vinfo)))
2664 : LOOP_VINFO_MAX_SPEC_READ_AMOUNT (loop_vinfo) = read_amount;
2665 :
2666 : /* For VLA modes, we must use partial vectors. */
2667 : LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P (loop_vinfo) = true;
2668 : }
2669 : }
2670 :
2671 1220102 : if (*alignment_support_scheme == dr_unaligned_unsupported)
2672 : {
2673 64217 : if (dump_enabled_p ())
2674 248 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2675 : "unsupported unaligned access\n");
2676 64217 : return false;
2677 : }
2678 :
2679 : /* FIXME: At the moment the cost model seems to underestimate the
2680 : cost of using elementwise accesses. This check preserves the
2681 : traditional behavior until that can be fixed. */
2682 1155885 : if (*memory_access_type == VMAT_ELEMENTWISE
2683 14719 : && !STMT_VINFO_STRIDED_P (first_stmt_info)
2684 1170604 : && !(STMT_VINFO_GROUPED_ACCESS (stmt_info)
2685 9532 : && single_element_p
2686 8906 : && !pow2p_hwi (group_size)))
2687 : {
2688 9090 : if (dump_enabled_p ())
2689 362 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2690 : "not falling back to elementwise accesses\n");
2691 9090 : return false;
2692 : }
2693 :
2694 : /* For BB vectorization build up the vector from existing scalar defs. */
2695 1146795 : if (!loop_vinfo && *memory_access_type == VMAT_ELEMENTWISE)
2696 : return false;
2697 :
2698 : /* Some loads need to explicitly permute the loaded data if there
2699 : is a load permutation. Among those are:
2700 : - VMAT_ELEMENTWISE.
2701 : - VMAT_STRIDED_SLP.
2702 : - VMAT_GATHER_SCATTER:
2703 : - Strided gather (fallback for VMAT_STRIDED_SLP if #lanes == 1).
2704 : - Grouped strided gather (ditto but for #lanes > 1).
2705 :
2706 : For VMAT_ELEMENTWISE we can fold the load permutation into the
2707 : individual indices we access directly, eliding the permutation.
2708 : Strided gather only allows load permutations for the
2709 : single-element case. */
2710 :
2711 1146795 : if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()
2712 1146795 : && !(*memory_access_type == VMAT_ELEMENTWISE
2713 39806 : || (mat_gather_scatter_p (*memory_access_type)
2714 0 : && SLP_TREE_LANES (slp_node) == 1
2715 0 : && single_element_p)))
2716 : {
2717 39806 : if (!loop_vinfo)
2718 : {
2719 : /* In BB vectorization we may not actually use a loaded vector
2720 : accessing elements in excess of DR_GROUP_SIZE. */
2721 24148 : stmt_vec_info group_info = SLP_TREE_SCALAR_STMTS (slp_node)[0];
2722 24148 : group_info = DR_GROUP_FIRST_ELEMENT (group_info);
2723 24148 : unsigned HOST_WIDE_INT nunits;
2724 24148 : unsigned j, k, maxk = 0;
2725 85654 : FOR_EACH_VEC_ELT (SLP_TREE_LOAD_PERMUTATION (slp_node), j, k)
2726 61506 : if (k > maxk)
2727 : maxk = k;
2728 24148 : tree vectype = SLP_TREE_VECTYPE (slp_node);
2729 44044 : if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)
2730 24148 : || maxk >= (DR_GROUP_SIZE (group_info) & ~(nunits - 1)))
2731 : {
2732 4252 : if (dump_enabled_p ())
2733 31 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2734 : "BB vectorization with gaps at the end of "
2735 : "a load is not supported\n");
2736 4252 : return false;
2737 : }
2738 : }
2739 :
2740 35554 : if (!perm_ok)
2741 : {
2742 1985 : if (dump_enabled_p ())
2743 8 : dump_printf_loc (MSG_MISSED_OPTIMIZATION,
2744 : vect_location,
2745 : "unsupported load permutation\n");
2746 1985 : return false;
2747 : }
2748 :
2749 33569 : *slp_perm = true;
2750 : }
2751 :
2752 : return true;
2753 : }
2754 :
2755 : /* Return true if boolean argument at MASK_INDEX is suitable for vectorizing
2756 : conditional operation STMT_INFO. When returning true, store the mask
2757 : in *MASK_NODE, the type of its definition in *MASK_DT_OUT and the type of
2758 : the vectorized mask in *MASK_VECTYPE_OUT. */
2759 :
2760 : static bool
2761 8097 : vect_check_scalar_mask (vec_info *vinfo,
2762 : slp_tree slp_node, unsigned mask_index,
2763 : slp_tree *mask_node,
2764 : vect_def_type *mask_dt_out, tree *mask_vectype_out)
2765 : {
2766 8097 : enum vect_def_type mask_dt;
2767 8097 : tree mask_vectype;
2768 8097 : slp_tree mask_node_1;
2769 8097 : tree mask_;
2770 8097 : if (!vect_is_simple_use (vinfo, slp_node, mask_index,
2771 : &mask_, &mask_node_1, &mask_dt, &mask_vectype))
2772 : {
2773 0 : if (dump_enabled_p ())
2774 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2775 : "mask use not simple.\n");
2776 0 : return false;
2777 : }
2778 :
2779 8097 : if ((mask_dt == vect_constant_def || mask_dt == vect_external_def)
2780 8097 : && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask_)))
2781 : {
2782 0 : if (dump_enabled_p ())
2783 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2784 : "mask argument is not a boolean.\n");
2785 0 : return false;
2786 : }
2787 :
2788 8097 : tree vectype = SLP_TREE_VECTYPE (slp_node);
2789 8097 : if (!mask_vectype)
2790 17 : mask_vectype = get_mask_type_for_scalar_type (vinfo, TREE_TYPE (vectype),
2791 : mask_node_1);
2792 :
2793 8097 : if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype))
2794 : {
2795 0 : if (dump_enabled_p ())
2796 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2797 : "could not find an appropriate vector mask type.\n");
2798 0 : return false;
2799 : }
2800 :
2801 8097 : if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype),
2802 16194 : TYPE_VECTOR_SUBPARTS (vectype)))
2803 : {
2804 0 : if (dump_enabled_p ())
2805 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2806 : "vector mask type %T"
2807 : " does not match vector data type %T.\n",
2808 : mask_vectype, vectype);
2809 :
2810 0 : return false;
2811 : }
2812 :
2813 8097 : *mask_dt_out = mask_dt;
2814 8097 : *mask_vectype_out = mask_vectype;
2815 8097 : *mask_node = mask_node_1;
2816 8097 : return true;
2817 : }
2818 :
2819 :
2820 : /* Return true if stored value is suitable for vectorizing store
2821 : statement STMT_INFO. When returning true, store the scalar stored
2822 : in *RHS and *RHS_NODE, the type of the definition in *RHS_DT_OUT,
2823 : the type of the vectorized store value in
2824 : *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */
2825 :
2826 : static bool
2827 1316572 : vect_check_store_rhs (vec_info *vinfo, stmt_vec_info stmt_info,
2828 : slp_tree slp_node, slp_tree *rhs_node,
2829 : vect_def_type *rhs_dt_out, tree *rhs_vectype_out,
2830 : vec_load_store_type *vls_type_out)
2831 : {
2832 1316572 : int op_no = 0;
2833 1316572 : if (gcall *call = dyn_cast <gcall *> (stmt_info->stmt))
2834 : {
2835 1470 : if (gimple_call_internal_p (call)
2836 1470 : && internal_store_fn_p (gimple_call_internal_fn (call)))
2837 1470 : op_no = internal_fn_stored_value_index (gimple_call_internal_fn (call));
2838 : }
2839 1316572 : op_no = vect_slp_child_index_for_operand
2840 1316572 : (stmt_info->stmt, op_no, STMT_VINFO_GATHER_SCATTER_P (stmt_info));
2841 :
2842 1316572 : enum vect_def_type rhs_dt;
2843 1316572 : tree rhs_vectype;
2844 1316572 : tree rhs;
2845 1316572 : if (!vect_is_simple_use (vinfo, slp_node, op_no,
2846 : &rhs, rhs_node, &rhs_dt, &rhs_vectype))
2847 : {
2848 0 : if (dump_enabled_p ())
2849 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2850 : "use not simple.\n");
2851 0 : return false;
2852 : }
2853 :
2854 : /* In the case this is a store from a constant make sure
2855 : native_encode_expr can handle it. */
2856 1316572 : if (rhs_dt == vect_constant_def
2857 1316572 : && CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0)
2858 : {
2859 0 : if (dump_enabled_p ())
2860 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2861 : "cannot encode constant as a byte sequence.\n");
2862 0 : return false;
2863 : }
2864 :
2865 1316572 : tree vectype = SLP_TREE_VECTYPE (slp_node);
2866 1316572 : if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype))
2867 : {
2868 24 : if (dump_enabled_p ())
2869 24 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2870 : "incompatible vector types.\n");
2871 24 : return false;
2872 : }
2873 :
2874 1316548 : *rhs_dt_out = rhs_dt;
2875 1316548 : *rhs_vectype_out = rhs_vectype;
2876 1316548 : if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def)
2877 990936 : *vls_type_out = VLS_STORE_INVARIANT;
2878 : else
2879 325612 : *vls_type_out = VLS_STORE;
2880 : return true;
2881 : }
2882 :
2883 : /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT_INFO.
2884 : Note that we support masks with floating-point type, in which case the
2885 : floats are interpreted as a bitmask. */
2886 :
2887 : static tree
2888 165 : vect_build_all_ones_mask (vec_info *vinfo,
2889 : stmt_vec_info stmt_info, tree masktype)
2890 : {
2891 165 : if (TREE_CODE (masktype) == INTEGER_TYPE)
2892 98 : return build_int_cst (masktype, -1);
2893 67 : else if (VECTOR_BOOLEAN_TYPE_P (masktype)
2894 134 : || TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE)
2895 : {
2896 14 : tree mask = build_int_cst (TREE_TYPE (masktype), -1);
2897 14 : mask = build_vector_from_val (masktype, mask);
2898 14 : return vect_init_vector (vinfo, stmt_info, mask, masktype, NULL);
2899 : }
2900 53 : else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype)))
2901 : {
2902 : REAL_VALUE_TYPE r;
2903 : long tmp[6];
2904 371 : for (int j = 0; j < 6; ++j)
2905 318 : tmp[j] = -1;
2906 53 : real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype)));
2907 53 : tree mask = build_real (TREE_TYPE (masktype), r);
2908 53 : mask = build_vector_from_val (masktype, mask);
2909 53 : return vect_init_vector (vinfo, stmt_info, mask, masktype, NULL);
2910 : }
2911 0 : gcc_unreachable ();
2912 : }
2913 :
2914 : /* Build an all-zero merge value of type VECTYPE while vectorizing
2915 : STMT_INFO as a gather load. */
2916 :
2917 : static tree
2918 158 : vect_build_zero_merge_argument (vec_info *vinfo,
2919 : stmt_vec_info stmt_info, tree vectype)
2920 : {
2921 158 : tree merge;
2922 158 : if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE)
2923 49 : merge = build_int_cst (TREE_TYPE (vectype), 0);
2924 109 : else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype)))
2925 : {
2926 : REAL_VALUE_TYPE r;
2927 : long tmp[6];
2928 763 : for (int j = 0; j < 6; ++j)
2929 654 : tmp[j] = 0;
2930 109 : real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype)));
2931 109 : merge = build_real (TREE_TYPE (vectype), r);
2932 : }
2933 : else
2934 0 : gcc_unreachable ();
2935 158 : merge = build_vector_from_val (vectype, merge);
2936 158 : return vect_init_vector (vinfo, stmt_info, merge, vectype, NULL);
2937 : }
2938 :
2939 : /* Return the corresponding else value for an else value constant
2940 : ELSVAL with type TYPE. */
2941 :
2942 : tree
2943 1941 : vect_get_mask_load_else (int elsval, tree type)
2944 : {
2945 1941 : tree els;
2946 1941 : if (elsval == MASK_LOAD_ELSE_UNDEFINED)
2947 : {
2948 0 : tree tmp = create_tmp_var (type);
2949 : /* No need to warn about anything. */
2950 0 : TREE_NO_WARNING (tmp) = 1;
2951 0 : els = get_or_create_ssa_default_def (cfun, tmp);
2952 : }
2953 1941 : else if (elsval == MASK_LOAD_ELSE_M1)
2954 0 : els = build_minus_one_cst (type);
2955 1941 : else if (elsval == MASK_LOAD_ELSE_ZERO)
2956 1941 : els = build_zero_cst (type);
2957 : else
2958 0 : gcc_unreachable ();
2959 :
2960 1941 : return els;
2961 : }
2962 :
2963 : /* Build a gather load call while vectorizing STMT_INFO. Insert new
2964 : instructions before GSI and add them to VEC_STMT. GS_INFO describes
2965 : the gather load operation. If the load is conditional, MASK is the
2966 : vectorized condition, otherwise MASK is null. PTR is the base
2967 : pointer and OFFSET is the vectorized offset. */
2968 :
2969 : static gimple *
2970 346 : vect_build_one_gather_load_call (vec_info *vinfo, stmt_vec_info stmt_info,
2971 : slp_tree slp_node, tree vectype,
2972 : gimple_stmt_iterator *gsi, tree decl,
2973 : tree ptr, tree offset, tree mask)
2974 : {
2975 346 : tree arglist = TYPE_ARG_TYPES (TREE_TYPE (decl));
2976 346 : tree rettype = TREE_TYPE (TREE_TYPE (decl));
2977 346 : tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
2978 346 : /* ptrtype */ arglist = TREE_CHAIN (arglist);
2979 346 : tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
2980 346 : tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
2981 346 : tree scaletype = TREE_VALUE (arglist);
2982 346 : tree var;
2983 346 : gcc_checking_assert (types_compatible_p (srctype, rettype)
2984 : && (!mask
2985 : || TREE_CODE (masktype) == INTEGER_TYPE
2986 : || types_compatible_p (srctype, masktype)));
2987 :
2988 346 : tree op = offset;
2989 346 : if (!useless_type_conversion_p (idxtype, TREE_TYPE (op)))
2990 : {
2991 100 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)),
2992 : TYPE_VECTOR_SUBPARTS (idxtype)));
2993 100 : var = vect_get_new_ssa_name (idxtype, vect_simple_var);
2994 100 : op = build1 (VIEW_CONVERT_EXPR, idxtype, op);
2995 100 : gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op);
2996 100 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
2997 100 : op = var;
2998 : }
2999 :
3000 346 : tree src_op = NULL_TREE;
3001 346 : tree mask_op = NULL_TREE;
3002 346 : if (mask)
3003 : {
3004 188 : if (!useless_type_conversion_p (masktype, TREE_TYPE (mask)))
3005 : {
3006 188 : tree utype, optype = TREE_TYPE (mask);
3007 188 : if (VECTOR_TYPE_P (masktype)
3008 188 : || TYPE_MODE (masktype) == TYPE_MODE (optype))
3009 : utype = masktype;
3010 : else
3011 6 : utype = lang_hooks.types.type_for_mode (TYPE_MODE (optype), 1);
3012 188 : var = vect_get_new_ssa_name (utype, vect_scalar_var);
3013 188 : tree mask_arg = build1 (VIEW_CONVERT_EXPR, utype, mask);
3014 188 : gassign *new_stmt
3015 188 : = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_arg);
3016 188 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3017 188 : mask_arg = var;
3018 188 : if (!useless_type_conversion_p (masktype, utype))
3019 : {
3020 6 : gcc_assert (TYPE_PRECISION (utype)
3021 : <= TYPE_PRECISION (masktype));
3022 6 : var = vect_get_new_ssa_name (masktype, vect_scalar_var);
3023 6 : new_stmt = gimple_build_assign (var, NOP_EXPR, mask_arg);
3024 6 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3025 6 : mask_arg = var;
3026 : }
3027 188 : src_op = build_zero_cst (srctype);
3028 188 : mask_op = mask_arg;
3029 : }
3030 : else
3031 : {
3032 : src_op = mask;
3033 : mask_op = mask;
3034 : }
3035 : }
3036 : else
3037 : {
3038 158 : src_op = vect_build_zero_merge_argument (vinfo, stmt_info, rettype);
3039 158 : mask_op = vect_build_all_ones_mask (vinfo, stmt_info, masktype);
3040 : }
3041 :
3042 346 : tree scale = build_int_cst (scaletype, SLP_TREE_GS_SCALE (slp_node));
3043 346 : gimple *new_stmt = gimple_build_call (decl, 5, src_op, ptr, op,
3044 : mask_op, scale);
3045 :
3046 346 : if (!useless_type_conversion_p (vectype, rettype))
3047 : {
3048 49 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype),
3049 : TYPE_VECTOR_SUBPARTS (rettype)));
3050 49 : op = vect_get_new_ssa_name (rettype, vect_simple_var);
3051 49 : gimple_call_set_lhs (new_stmt, op);
3052 49 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3053 49 : op = build1 (VIEW_CONVERT_EXPR, vectype, op);
3054 49 : new_stmt = gimple_build_assign (NULL_TREE, VIEW_CONVERT_EXPR, op);
3055 : }
3056 :
3057 346 : return new_stmt;
3058 : }
3059 :
3060 : /* Build a scatter store call while vectorizing STMT_INFO. Insert new
3061 : instructions before GSI. GS_INFO describes the scatter store operation.
3062 : PTR is the base pointer, OFFSET the vectorized offsets and OPRND the
3063 : vectorized data to store.
3064 : If the store is conditional, MASK is the vectorized condition, otherwise
3065 : MASK is null. */
3066 :
3067 : static gimple *
3068 161 : vect_build_one_scatter_store_call (vec_info *vinfo, stmt_vec_info stmt_info,
3069 : slp_tree slp_node,
3070 : gimple_stmt_iterator *gsi,
3071 : tree decl,
3072 : tree ptr, tree offset, tree oprnd, tree mask)
3073 : {
3074 161 : tree rettype = TREE_TYPE (TREE_TYPE (decl));
3075 161 : tree arglist = TYPE_ARG_TYPES (TREE_TYPE (decl));
3076 161 : /* tree ptrtype = TREE_VALUE (arglist); */ arglist = TREE_CHAIN (arglist);
3077 161 : tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
3078 161 : tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
3079 161 : tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
3080 161 : tree scaletype = TREE_VALUE (arglist);
3081 161 : gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE
3082 : && TREE_CODE (rettype) == VOID_TYPE);
3083 :
3084 161 : tree mask_arg = NULL_TREE;
3085 161 : if (mask)
3086 : {
3087 110 : mask_arg = mask;
3088 110 : tree optype = TREE_TYPE (mask_arg);
3089 110 : tree utype;
3090 110 : if (TYPE_MODE (masktype) == TYPE_MODE (optype))
3091 : utype = masktype;
3092 : else
3093 8 : utype = lang_hooks.types.type_for_mode (TYPE_MODE (optype), 1);
3094 110 : tree var = vect_get_new_ssa_name (utype, vect_scalar_var);
3095 110 : mask_arg = build1 (VIEW_CONVERT_EXPR, utype, mask_arg);
3096 110 : gassign *new_stmt
3097 110 : = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_arg);
3098 110 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3099 110 : mask_arg = var;
3100 110 : if (!useless_type_conversion_p (masktype, utype))
3101 : {
3102 8 : gcc_assert (TYPE_PRECISION (utype) <= TYPE_PRECISION (masktype));
3103 8 : tree var = vect_get_new_ssa_name (masktype, vect_scalar_var);
3104 8 : new_stmt = gimple_build_assign (var, NOP_EXPR, mask_arg);
3105 8 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3106 8 : mask_arg = var;
3107 : }
3108 : }
3109 : else
3110 : {
3111 51 : mask_arg = build_int_cst (masktype, -1);
3112 51 : mask_arg = vect_init_vector (vinfo, stmt_info, mask_arg, masktype, NULL);
3113 : }
3114 :
3115 161 : tree src = oprnd;
3116 161 : if (!useless_type_conversion_p (srctype, TREE_TYPE (src)))
3117 : {
3118 0 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)),
3119 : TYPE_VECTOR_SUBPARTS (srctype)));
3120 0 : tree var = vect_get_new_ssa_name (srctype, vect_simple_var);
3121 0 : src = build1 (VIEW_CONVERT_EXPR, srctype, src);
3122 0 : gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src);
3123 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3124 0 : src = var;
3125 : }
3126 :
3127 161 : tree op = offset;
3128 161 : if (!useless_type_conversion_p (idxtype, TREE_TYPE (op)))
3129 : {
3130 16 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)),
3131 : TYPE_VECTOR_SUBPARTS (idxtype)));
3132 16 : tree var = vect_get_new_ssa_name (idxtype, vect_simple_var);
3133 16 : op = build1 (VIEW_CONVERT_EXPR, idxtype, op);
3134 16 : gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op);
3135 16 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3136 16 : op = var;
3137 : }
3138 :
3139 161 : tree scale = build_int_cst (scaletype, SLP_TREE_GS_SCALE (slp_node));
3140 161 : gcall *new_stmt
3141 161 : = gimple_build_call (decl, 5, ptr, mask_arg, op, src, scale);
3142 161 : return new_stmt;
3143 : }
3144 :
3145 : /* Prepare the base and offset in GS_INFO for vectorization.
3146 : Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET
3147 : to the vectorized offset argument for the first copy of STMT_INFO.
3148 : STMT_INFO is the statement described by GS_INFO and LOOP is the
3149 : containing loop. */
3150 :
3151 : static void
3152 1219 : vect_get_gather_scatter_ops (class loop *loop, slp_tree slp_node,
3153 : tree *dataref_ptr, vec<tree> *vec_offset)
3154 : {
3155 1219 : gimple_seq stmts = NULL;
3156 1219 : *dataref_ptr = force_gimple_operand (SLP_TREE_GS_BASE (slp_node),
3157 : &stmts, true, NULL_TREE);
3158 1219 : if (stmts != NULL)
3159 : {
3160 986 : basic_block new_bb;
3161 986 : edge pe = loop_preheader_edge (loop);
3162 986 : new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts);
3163 986 : gcc_assert (!new_bb);
3164 : }
3165 1219 : vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[0], vec_offset);
3166 1219 : }
3167 :
3168 : /* Prepare to implement a grouped or strided load or store using
3169 : the gather load or scatter store operation described by GS_INFO.
3170 : STMT_INFO is the load or store statement.
3171 :
3172 : Set *DATAREF_BUMP to the amount that should be added to the base
3173 : address after each copy of the vectorized statement. Set *VEC_OFFSET
3174 : to an invariant offset vector in which element I has the value
3175 : I * DR_STEP / SCALE. */
3176 :
3177 : static void
3178 0 : vect_get_strided_load_store_ops (stmt_vec_info stmt_info, slp_tree node,
3179 : tree vectype, tree offset_vectype,
3180 : loop_vec_info loop_vinfo,
3181 : gimple_stmt_iterator *gsi,
3182 : tree *dataref_bump, tree *vec_offset,
3183 : vec_loop_lens *loop_lens)
3184 : {
3185 0 : struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
3186 :
3187 0 : if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
3188 : {
3189 : /* _31 = .SELECT_VL (ivtmp_29, POLY_INT_CST [4, 4]);
3190 : ivtmp_8 = _31 * 16 (step in bytes);
3191 : .MASK_LEN_SCATTER_STORE (vectp_a.9_7, ... );
3192 : vectp_a.9_26 = vectp_a.9_7 + ivtmp_8; */
3193 0 : tree loop_len
3194 0 : = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, vectype, 0, 0, true);
3195 0 : tree tmp
3196 0 : = fold_build2 (MULT_EXPR, sizetype,
3197 : fold_convert (sizetype, unshare_expr (DR_STEP (dr))),
3198 : loop_len);
3199 0 : *dataref_bump = force_gimple_operand_gsi (gsi, tmp, true, NULL_TREE, true,
3200 : GSI_SAME_STMT);
3201 : }
3202 : else
3203 : {
3204 0 : tree bump
3205 0 : = size_binop (MULT_EXPR,
3206 : fold_convert (sizetype, unshare_expr (DR_STEP (dr))),
3207 : size_int (TYPE_VECTOR_SUBPARTS (vectype)));
3208 0 : *dataref_bump = cse_and_gimplify_to_preheader (loop_vinfo, bump);
3209 : }
3210 :
3211 0 : internal_fn ifn
3212 0 : = DR_IS_READ (dr) ? IFN_MASK_LEN_STRIDED_LOAD : IFN_MASK_LEN_STRIDED_STORE;
3213 0 : if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED))
3214 : {
3215 0 : *vec_offset = cse_and_gimplify_to_preheader (loop_vinfo,
3216 : unshare_expr (DR_STEP (dr)));
3217 0 : return;
3218 : }
3219 :
3220 : /* The offset given in GS_INFO can have pointer type, so use the element
3221 : type of the vector instead. */
3222 0 : tree offset_type = TREE_TYPE (offset_vectype);
3223 :
3224 : /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */
3225 0 : tree step = size_binop (EXACT_DIV_EXPR, unshare_expr (DR_STEP (dr)),
3226 : ssize_int (SLP_TREE_GS_SCALE (node)));
3227 0 : step = fold_convert (offset_type, step);
3228 :
3229 : /* Create {0, X, X*2, X*3, ...}. */
3230 0 : tree offset = fold_build2 (VEC_SERIES_EXPR, offset_vectype,
3231 : build_zero_cst (offset_type), step);
3232 0 : *vec_offset = cse_and_gimplify_to_preheader (loop_vinfo, offset);
3233 : }
3234 :
3235 : /* Prepare the pointer IVs which needs to be updated by a variable amount.
3236 : Such variable amount is the outcome of .SELECT_VL. In this case, we can
3237 : allow each iteration process the flexible number of elements as long as
3238 : the number <= vf elments.
3239 :
3240 : Return data reference according to SELECT_VL.
3241 : If new statements are needed, insert them before GSI. */
3242 :
3243 : static tree
3244 0 : vect_get_loop_variant_data_ptr_increment (
3245 : vec_info *vinfo, tree aggr_type, gimple_stmt_iterator *gsi,
3246 : vec_loop_lens *loop_lens, dr_vec_info *dr_info,
3247 : vect_memory_access_type memory_access_type)
3248 : {
3249 0 : loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
3250 0 : tree step = vect_dr_behavior (vinfo, dr_info)->step;
3251 :
3252 : /* gather/scatter never reach here. */
3253 0 : gcc_assert (!mat_gather_scatter_p (memory_access_type));
3254 :
3255 : /* When we support SELECT_VL pattern, we dynamic adjust
3256 : the memory address by .SELECT_VL result.
3257 :
3258 : The result of .SELECT_VL is the number of elements to
3259 : be processed of each iteration. So the memory address
3260 : adjustment operation should be:
3261 :
3262 : addr = addr + .SELECT_VL (ARG..) * step;
3263 : */
3264 0 : tree loop_len
3265 0 : = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0, 0, true);
3266 0 : tree len_type = TREE_TYPE (loop_len);
3267 : /* Since the outcome of .SELECT_VL is element size, we should adjust
3268 : it into bytesize so that it can be used in address pointer variable
3269 : amount IVs adjustment. */
3270 0 : tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
3271 : wide_int_to_tree (len_type, wi::to_widest (step)));
3272 0 : tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
3273 0 : gassign *assign = gimple_build_assign (bump, tmp);
3274 0 : gsi_insert_before (gsi, assign, GSI_SAME_STMT);
3275 0 : return bump;
3276 : }
3277 :
3278 : /* Return the amount that should be added to a vector pointer to move
3279 : to the next or previous copy of AGGR_TYPE. DR_INFO is the data reference
3280 : being vectorized and MEMORY_ACCESS_TYPE describes the type of
3281 : vectorization. */
3282 :
3283 : static tree
3284 700634 : vect_get_data_ptr_increment (vec_info *vinfo, gimple_stmt_iterator *gsi,
3285 : dr_vec_info *dr_info, tree aggr_type,
3286 : vect_memory_access_type memory_access_type,
3287 : vec_loop_lens *loop_lens)
3288 : {
3289 700634 : if (memory_access_type == VMAT_INVARIANT)
3290 0 : return size_zero_node;
3291 :
3292 700634 : loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
3293 134912 : if (loop_vinfo && LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
3294 0 : return vect_get_loop_variant_data_ptr_increment (vinfo, aggr_type, gsi,
3295 : loop_lens, dr_info,
3296 0 : memory_access_type);
3297 :
3298 700634 : tree iv_step = TYPE_SIZE_UNIT (aggr_type);
3299 700634 : tree step = vect_dr_behavior (vinfo, dr_info)->step;
3300 700634 : if (tree_int_cst_sgn (step) == -1)
3301 2828 : iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step);
3302 : return iv_step;
3303 : }
3304 :
3305 : /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
3306 :
3307 : static bool
3308 206 : vectorizable_bswap (vec_info *vinfo,
3309 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
3310 : slp_tree slp_node,
3311 : slp_tree *slp_op,
3312 : tree vectype_in, stmt_vector_for_cost *cost_vec)
3313 : {
3314 206 : tree op, vectype;
3315 206 : gcall *stmt = as_a <gcall *> (stmt_info->stmt);
3316 :
3317 206 : op = gimple_call_arg (stmt, 0);
3318 206 : vectype = SLP_TREE_VECTYPE (slp_node);
3319 206 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3320 :
3321 206 : if (TYPE_SIZE (vectype_in) != TYPE_SIZE (vectype))
3322 : {
3323 0 : if (dump_enabled_p ())
3324 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3325 : "mismatched vector sizes %T and %T\n",
3326 : vectype_in, vectype);
3327 0 : return false;
3328 : }
3329 :
3330 206 : tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in);
3331 206 : if (! char_vectype)
3332 : return false;
3333 :
3334 206 : poly_uint64 num_bytes = TYPE_VECTOR_SUBPARTS (char_vectype);
3335 206 : unsigned word_bytes;
3336 206 : if (!constant_multiple_p (num_bytes, nunits, &word_bytes))
3337 : return false;
3338 :
3339 : /* The encoding uses one stepped pattern for each byte in the word. */
3340 206 : vec_perm_builder elts (num_bytes, word_bytes, 3);
3341 824 : for (unsigned i = 0; i < 3; ++i)
3342 3318 : for (unsigned j = 0; j < word_bytes; ++j)
3343 2700 : elts.quick_push ((i + 1) * word_bytes - j - 1);
3344 :
3345 206 : vec_perm_indices indices (elts, 1, num_bytes);
3346 206 : machine_mode vmode = TYPE_MODE (char_vectype);
3347 206 : if (!can_vec_perm_const_p (vmode, vmode, indices))
3348 : return false;
3349 :
3350 152 : if (cost_vec)
3351 : {
3352 140 : if (!vect_maybe_update_slp_op_vectype (slp_op[0], vectype_in))
3353 : {
3354 0 : if (dump_enabled_p ())
3355 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3356 : "incompatible vector types for invariants\n");
3357 0 : return false;
3358 : }
3359 :
3360 140 : SLP_TREE_TYPE (slp_node) = call_vec_info_type;
3361 140 : DUMP_VECT_SCOPE ("vectorizable_bswap");
3362 140 : record_stmt_cost (cost_vec,
3363 : 1, vector_stmt, slp_node, 0, vect_prologue);
3364 140 : record_stmt_cost (cost_vec,
3365 140 : vect_get_num_copies (vinfo, slp_node),
3366 : vec_perm, slp_node, 0, vect_body);
3367 140 : return true;
3368 : }
3369 :
3370 12 : tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices);
3371 :
3372 : /* Transform. */
3373 12 : vec<tree> vec_oprnds = vNULL;
3374 12 : vect_get_vec_defs (vinfo, slp_node, op, &vec_oprnds);
3375 : /* Arguments are ready. create the new vector stmt. */
3376 12 : unsigned i;
3377 12 : tree vop;
3378 24 : FOR_EACH_VEC_ELT (vec_oprnds, i, vop)
3379 : {
3380 12 : gimple *new_stmt;
3381 12 : tree tem = make_ssa_name (char_vectype);
3382 12 : new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR,
3383 : char_vectype, vop));
3384 12 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3385 12 : tree tem2 = make_ssa_name (char_vectype);
3386 12 : new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR,
3387 : tem, tem, bswap_vconst);
3388 12 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3389 12 : tem = make_ssa_name (vectype);
3390 12 : new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR,
3391 : vectype, tem2));
3392 12 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3393 12 : slp_node->push_vec_def (new_stmt);
3394 : }
3395 :
3396 12 : vec_oprnds.release ();
3397 12 : return true;
3398 206 : }
3399 :
3400 : /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have
3401 : integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT
3402 : in a single step. On success, store the binary pack code in
3403 : *CONVERT_CODE. */
3404 :
3405 : static bool
3406 156 : simple_integer_narrowing (tree vectype_out, tree vectype_in,
3407 : code_helper *convert_code)
3408 : {
3409 312 : if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out))
3410 312 : || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in)))
3411 : return false;
3412 :
3413 66 : code_helper code;
3414 66 : int multi_step_cvt = 0;
3415 66 : auto_vec <tree, 8> interm_types;
3416 97 : if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in,
3417 : &code, &multi_step_cvt, &interm_types)
3418 66 : || multi_step_cvt)
3419 31 : return false;
3420 :
3421 35 : *convert_code = code;
3422 35 : return true;
3423 66 : }
3424 :
3425 : /* Function vectorizable_call.
3426 :
3427 : Check if STMT_INFO performs a function call that can be vectorized.
3428 : If COST_VEC is passed, calculate costs but don't change anything,
3429 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
3430 : it, and insert it at GSI.
3431 : Return true if STMT_INFO is vectorizable in this way. */
3432 :
3433 : static bool
3434 2334134 : vectorizable_call (vec_info *vinfo,
3435 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
3436 : slp_tree slp_node,
3437 : stmt_vector_for_cost *cost_vec)
3438 : {
3439 2334134 : gcall *stmt;
3440 2334134 : tree vec_dest;
3441 2334134 : tree scalar_dest;
3442 2334134 : tree op;
3443 2334134 : tree vec_oprnd0 = NULL_TREE;
3444 2334134 : tree vectype_out, vectype_in;
3445 2334134 : poly_uint64 nunits_in;
3446 2334134 : poly_uint64 nunits_out;
3447 2334134 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
3448 2334134 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
3449 2334134 : tree fndecl, new_temp, rhs_type;
3450 2334134 : enum vect_def_type dt[5]
3451 : = { vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type,
3452 : vect_unknown_def_type, vect_unknown_def_type };
3453 2334134 : tree vectypes[ARRAY_SIZE (dt)] = {};
3454 2334134 : slp_tree slp_op[ARRAY_SIZE (dt)] = {};
3455 2334134 : auto_vec<tree, 8> vargs;
3456 2334134 : enum { NARROW, NONE, WIDEN } modifier;
3457 2334134 : size_t i, nargs;
3458 2334134 : tree clz_ctz_arg1 = NULL_TREE;
3459 :
3460 2334134 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3461 : return false;
3462 :
3463 2334134 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
3464 194522 : && cost_vec)
3465 : return false;
3466 :
3467 : /* Is STMT_INFO a vectorizable call? */
3468 2343683 : stmt = dyn_cast <gcall *> (stmt_info->stmt);
3469 19326 : if (!stmt)
3470 : return false;
3471 :
3472 19326 : if (gimple_call_internal_p (stmt)
3473 19326 : && (internal_load_fn_p (gimple_call_internal_fn (stmt))
3474 12184 : || internal_store_fn_p (gimple_call_internal_fn (stmt))))
3475 : /* Handled by vectorizable_load and vectorizable_store. */
3476 2692 : return false;
3477 :
3478 16634 : if (gimple_call_lhs (stmt) == NULL_TREE
3479 16634 : || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
3480 : return false;
3481 :
3482 16628 : gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt));
3483 :
3484 16628 : vectype_out = SLP_TREE_VECTYPE (slp_node);
3485 :
3486 : /* Process function arguments. */
3487 16628 : rhs_type = NULL_TREE;
3488 16628 : vectype_in = NULL_TREE;
3489 16628 : nargs = gimple_call_num_args (stmt);
3490 :
3491 : /* Bail out if the function has more than four arguments, we do not have
3492 : interesting builtin functions to vectorize with more than two arguments
3493 : except for fma (cond_fma has more). No arguments is also not good. */
3494 16628 : if (nargs == 0 || nargs > 5)
3495 : return false;
3496 :
3497 : /* Ignore the arguments of IFN_GOMP_SIMD_LANE, they are magic. */
3498 16548 : combined_fn cfn = gimple_call_combined_fn (stmt);
3499 16548 : if (cfn == CFN_GOMP_SIMD_LANE)
3500 : {
3501 3207 : nargs = 0;
3502 3207 : rhs_type = unsigned_type_node;
3503 : }
3504 : /* Similarly pretend IFN_CLZ and IFN_CTZ only has one argument, the second
3505 : argument just says whether it is well-defined at zero or not and what
3506 : value should be returned for it. */
3507 16548 : if ((cfn == CFN_CLZ || cfn == CFN_CTZ) && nargs == 2)
3508 : {
3509 118 : nargs = 1;
3510 118 : clz_ctz_arg1 = gimple_call_arg (stmt, 1);
3511 : }
3512 :
3513 16548 : int mask_opno = -1;
3514 16548 : if (internal_fn_p (cfn))
3515 : {
3516 : /* We can only handle direct internal masked calls here,
3517 : vectorizable_simd_clone_call is for the rest. */
3518 14081 : if (cfn == CFN_MASK_CALL)
3519 : return false;
3520 13927 : mask_opno = internal_fn_mask_index (as_internal_fn (cfn));
3521 : }
3522 :
3523 45912 : for (i = 0; i < nargs; i++)
3524 : {
3525 30762 : if ((int) i == mask_opno)
3526 : {
3527 4220 : if (!vect_check_scalar_mask (vinfo, slp_node, mask_opno,
3528 : &slp_op[i], &dt[i], &vectypes[i]))
3529 : return false;
3530 4220 : continue;
3531 : }
3532 :
3533 26542 : if (!vect_is_simple_use (vinfo, slp_node,
3534 : i, &op, &slp_op[i], &dt[i], &vectypes[i]))
3535 : {
3536 0 : if (dump_enabled_p ())
3537 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3538 : "use not simple.\n");
3539 0 : return false;
3540 : }
3541 :
3542 : /* We can only handle calls with arguments of the same type. */
3543 26542 : if (rhs_type
3544 26542 : && !types_compatible_p (rhs_type, TREE_TYPE (op)))
3545 : {
3546 1244 : if (dump_enabled_p ())
3547 200 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3548 : "argument types differ.\n");
3549 1244 : return false;
3550 : }
3551 25298 : if (!rhs_type)
3552 13187 : rhs_type = TREE_TYPE (op);
3553 :
3554 25298 : if (!vectype_in)
3555 13642 : vectype_in = vectypes[i];
3556 11656 : else if (vectypes[i]
3557 11656 : && !types_compatible_p (vectypes[i], vectype_in))
3558 : {
3559 0 : if (dump_enabled_p ())
3560 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3561 : "argument vector types differ.\n");
3562 0 : return false;
3563 : }
3564 : }
3565 : /* If all arguments are external or constant defs, infer the vector type
3566 : from the scalar type. */
3567 15150 : if (!vectype_in)
3568 5554 : vectype_in = get_vectype_for_scalar_type (vinfo, rhs_type, slp_node);
3569 15150 : if (!cost_vec)
3570 4210 : gcc_assert (vectype_in);
3571 10940 : if (!vectype_in)
3572 : {
3573 1065 : if (dump_enabled_p ())
3574 4 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3575 : "no vectype for scalar type %T\n", rhs_type);
3576 :
3577 1065 : return false;
3578 : }
3579 :
3580 28170 : if (VECTOR_BOOLEAN_TYPE_P (vectype_out)
3581 14085 : != VECTOR_BOOLEAN_TYPE_P (vectype_in))
3582 : {
3583 12 : if (dump_enabled_p ())
3584 12 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3585 : "mixed mask and nonmask vector types\n");
3586 12 : return false;
3587 : }
3588 :
3589 14073 : if (vect_emulated_vector_p (vectype_in)
3590 14073 : || vect_emulated_vector_p (vectype_out))
3591 : {
3592 0 : if (dump_enabled_p ())
3593 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3594 : "use emulated vector type for call\n");
3595 0 : return false;
3596 : }
3597 :
3598 : /* FORNOW */
3599 14073 : nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
3600 14073 : nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3601 14073 : if (known_eq (nunits_in * 2, nunits_out))
3602 : modifier = NARROW;
3603 13530 : else if (known_eq (nunits_out, nunits_in))
3604 : modifier = NONE;
3605 45 : else if (known_eq (nunits_out * 2, nunits_in))
3606 : modifier = WIDEN;
3607 : else
3608 : return false;
3609 :
3610 : /* We only handle functions that do not read or clobber memory. */
3611 28146 : if (gimple_vuse (stmt))
3612 : {
3613 1241 : if (dump_enabled_p ())
3614 14 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3615 : "function reads from or writes to memory.\n");
3616 1241 : return false;
3617 : }
3618 :
3619 : /* For now, we only vectorize functions if a target specific builtin
3620 : is available. TODO -- in some cases, it might be profitable to
3621 : insert the calls for pieces of the vector, in order to be able
3622 : to vectorize other operations in the loop. */
3623 12832 : fndecl = NULL_TREE;
3624 12832 : internal_fn ifn = IFN_LAST;
3625 12832 : tree callee = gimple_call_fndecl (stmt);
3626 :
3627 : /* First try using an internal function. */
3628 12832 : code_helper convert_code = MAX_TREE_CODES;
3629 12832 : if (cfn != CFN_LAST
3630 12832 : && (modifier == NONE
3631 168 : || (modifier == NARROW
3632 156 : && simple_integer_narrowing (vectype_out, vectype_in,
3633 : &convert_code))))
3634 11866 : ifn = vectorizable_internal_function (cfn, callee, vectype_out,
3635 : vectype_in);
3636 :
3637 : /* Check if the operation traps. */
3638 12832 : bool could_trap = gimple_could_trap_p (STMT_VINFO_STMT (stmt_info));
3639 12832 : if (could_trap && cost_vec && loop_vinfo)
3640 : {
3641 : /* If the operation can trap it must be conditional, otherwise fail. */
3642 414 : internal_fn cond_fn = (internal_fn_mask_index (ifn) != -1
3643 414 : ? ifn : get_conditional_internal_fn (ifn));
3644 414 : internal_fn cond_len_fn = get_len_internal_fn (cond_fn);
3645 414 : if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
3646 : {
3647 : /* We assume that BB SLP fills all lanes, so no inactive lanes can
3648 : cause issues. */
3649 52 : if ((cond_fn == IFN_LAST
3650 35 : || !direct_internal_fn_supported_p (cond_fn, vectype_out,
3651 : OPTIMIZE_FOR_SPEED))
3652 87 : && (cond_len_fn == IFN_LAST
3653 35 : || !direct_internal_fn_supported_p (cond_len_fn, vectype_out,
3654 : OPTIMIZE_FOR_SPEED)))
3655 : {
3656 52 : if (dump_enabled_p ())
3657 10 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3658 : "can't use a fully-masked loop because no"
3659 : " conditional operation is available.\n");
3660 52 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
3661 : }
3662 : }
3663 : }
3664 :
3665 : /* If that fails, try asking for a target-specific built-in function. */
3666 12832 : if (ifn == IFN_LAST)
3667 : {
3668 6762 : if (cfn != CFN_LAST)
3669 5929 : fndecl = targetm.vectorize.builtin_vectorized_function
3670 5929 : (cfn, vectype_out, vectype_in);
3671 833 : else if (callee && fndecl_built_in_p (callee, BUILT_IN_MD))
3672 24 : fndecl = targetm.vectorize.builtin_md_vectorized_function
3673 24 : (callee, vectype_out, vectype_in);
3674 : }
3675 :
3676 12832 : if (ifn == IFN_LAST && !fndecl)
3677 : {
3678 6468 : if (cfn == CFN_GOMP_SIMD_LANE
3679 3207 : && SLP_TREE_LANES (slp_node) == 1
3680 3207 : && loop_vinfo
3681 3207 : && LOOP_VINFO_LOOP (loop_vinfo)->simduid
3682 3207 : && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME
3683 12882 : && LOOP_VINFO_LOOP (loop_vinfo)->simduid
3684 3207 : == SSA_NAME_VAR (gimple_call_arg (stmt, 0)))
3685 : {
3686 : /* We can handle IFN_GOMP_SIMD_LANE by returning a
3687 : { 0, 1, 2, ... vf - 1 } vector. */
3688 3207 : gcc_assert (nargs == 0);
3689 : }
3690 3261 : else if (modifier == NONE
3691 3261 : && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16)
3692 2931 : || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32)
3693 2783 : || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64)
3694 2751 : || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP128)))
3695 206 : return vectorizable_bswap (vinfo, stmt_info, gsi, slp_node,
3696 206 : slp_op, vectype_in, cost_vec);
3697 : else
3698 : {
3699 3055 : if (dump_enabled_p ())
3700 266 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3701 : "function is not vectorizable.\n");
3702 3055 : return false;
3703 : }
3704 : }
3705 :
3706 9571 : int reduc_idx = SLP_TREE_REDUC_IDX (slp_node);
3707 9571 : internal_fn cond_fn = (internal_fn_mask_index (ifn) != -1
3708 9571 : ? ifn : get_conditional_internal_fn (ifn));
3709 9571 : internal_fn cond_len_fn = get_len_internal_fn (cond_fn);
3710 9571 : vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL);
3711 7691 : vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL);
3712 9571 : unsigned int nvectors = vect_get_num_copies (vinfo, slp_node);
3713 9571 : if (cost_vec) /* transformation not required. */
3714 : {
3715 16386 : for (i = 0; i < nargs; ++i)
3716 11013 : if (!vect_maybe_update_slp_op_vectype (slp_op[i],
3717 11013 : vectypes[i]
3718 : ? vectypes[i] : vectype_in))
3719 : {
3720 0 : if (dump_enabled_p ())
3721 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3722 : "incompatible vector types for invariants\n");
3723 0 : return false;
3724 : }
3725 5373 : SLP_TREE_TYPE (slp_node) = call_vec_info_type;
3726 5373 : DUMP_VECT_SCOPE ("vectorizable_call");
3727 5373 : vect_model_simple_cost (vinfo, 1, slp_node, cost_vec);
3728 :
3729 5373 : if (loop_vinfo
3730 4428 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
3731 3070 : && (reduc_idx >= 0 || could_trap || mask_opno >= 0))
3732 : {
3733 1930 : if (reduc_idx >= 0
3734 1438 : && (cond_fn == IFN_LAST
3735 1438 : || !direct_internal_fn_supported_p (cond_fn, vectype_out,
3736 : OPTIMIZE_FOR_SPEED))
3737 1940 : && (cond_len_fn == IFN_LAST
3738 10 : || !direct_internal_fn_supported_p (cond_len_fn, vectype_out,
3739 : OPTIMIZE_FOR_SPEED)))
3740 : {
3741 10 : if (dump_enabled_p ())
3742 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3743 : "can't use a fully-masked loop because no"
3744 : " conditional operation is available.\n");
3745 10 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
3746 : }
3747 : else
3748 : {
3749 1920 : tree scalar_mask = NULL_TREE;
3750 1920 : if (mask_opno >= 0)
3751 1920 : scalar_mask = gimple_call_arg (stmt_info->stmt, mask_opno);
3752 1920 : if (cond_len_fn != IFN_LAST
3753 1920 : && direct_internal_fn_supported_p (cond_len_fn, vectype_out,
3754 : OPTIMIZE_FOR_SPEED))
3755 0 : vect_record_loop_len (loop_vinfo, lens, nvectors, vectype_out,
3756 : 1);
3757 : else
3758 1920 : vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out,
3759 : scalar_mask);
3760 : }
3761 : }
3762 5373 : return true;
3763 : }
3764 :
3765 : /* Transform. */
3766 :
3767 4198 : if (dump_enabled_p ())
3768 415 : dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n");
3769 :
3770 : /* Handle def. */
3771 4198 : scalar_dest = gimple_call_lhs (stmt);
3772 4198 : vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
3773 :
3774 4198 : bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
3775 3263 : bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo);
3776 4198 : unsigned int vect_nargs = nargs;
3777 4198 : if (len_loop_p && (reduc_idx >= 0 || could_trap || mask_opno >= 0))
3778 : {
3779 0 : ifn = cond_len_fn;
3780 : /* COND_* -> COND_LEN_* takes 2 extra arguments:LEN,BIAS. */
3781 0 : vect_nargs += 2;
3782 : /* But unless there's a mask argument already we need that
3783 : as well, and an else value. */
3784 0 : if (mask_opno == -1)
3785 0 : vect_nargs += 2;
3786 : }
3787 4198 : else if (masked_loop_p && mask_opno == -1 && (reduc_idx >= 0 || could_trap))
3788 : {
3789 0 : ifn = cond_fn;
3790 0 : vect_nargs += 2;
3791 : }
3792 4198 : int len_opno = internal_fn_len_index (ifn);
3793 4198 : if (clz_ctz_arg1)
3794 59 : ++vect_nargs;
3795 :
3796 4198 : if (modifier == NONE || ifn != IFN_LAST)
3797 : {
3798 4166 : tree prev_res = NULL_TREE;
3799 4166 : vargs.safe_grow (vect_nargs, true);
3800 4166 : auto_vec<vec<tree> > vec_defs (nargs);
3801 :
3802 : /* Build argument list for the vectorized call. */
3803 4166 : if (cfn == CFN_GOMP_SIMD_LANE)
3804 : {
3805 3308 : for (i = 0; i < nvectors; ++i)
3806 : {
3807 : /* ??? For multi-lane SLP we'd need to build
3808 : { 0, 0, .., 1, 1, ... }. */
3809 1708 : tree cst = build_index_vector (vectype_out,
3810 : i * nunits_out, 1);
3811 1708 : tree new_var
3812 1708 : = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_");
3813 1708 : gimple *init_stmt = gimple_build_assign (new_var, cst);
3814 1708 : vect_init_vector_1 (vinfo, stmt_info, init_stmt, NULL);
3815 1708 : new_temp = make_ssa_name (vec_dest);
3816 1708 : gimple *new_stmt = gimple_build_assign (new_temp, new_var);
3817 1708 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3818 1708 : slp_node->push_vec_def (new_stmt);
3819 : }
3820 : }
3821 : else
3822 : {
3823 2566 : vec<tree> vec_oprnds0;
3824 2566 : vect_get_slp_defs (vinfo, slp_node, &vec_defs);
3825 2566 : vec_oprnds0 = vec_defs[0];
3826 :
3827 : /* Arguments are ready. Create the new vector stmt. */
3828 5279 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0)
3829 : {
3830 2713 : int varg = 0;
3831 : /* Add the mask if necessary. */
3832 38 : if ((masked_loop_p || len_loop_p) && mask_opno == -1
3833 2715 : && internal_fn_mask_index (ifn) != -1)
3834 : {
3835 0 : gcc_assert (internal_fn_mask_index (ifn) == varg);
3836 0 : if (masked_loop_p)
3837 : {
3838 0 : unsigned int vec_num = vec_oprnds0.length ();
3839 0 : vargs[varg++] = vect_get_loop_mask (loop_vinfo, gsi,
3840 : masks, vec_num,
3841 : vectype_out, i);
3842 : }
3843 : else
3844 : {
3845 0 : tree mask_vectype = truth_type_for (vectype_out);
3846 0 : vargs[varg++] = vect_build_all_ones_mask (loop_vinfo,
3847 : stmt_info,
3848 : mask_vectype);
3849 : }
3850 : }
3851 : size_t k;
3852 9932 : for (k = 0; k < nargs; k++)
3853 : {
3854 7219 : vec<tree> vec_oprndsk = vec_defs[k];
3855 7219 : vargs[varg++] = vec_oprndsk[i];
3856 : }
3857 : /* Add the else value if necessary. */
3858 38 : if ((masked_loop_p || len_loop_p) && mask_opno == -1
3859 2715 : && internal_fn_else_index (ifn) != -1)
3860 : {
3861 0 : gcc_assert (internal_fn_else_index (ifn) == varg);
3862 0 : if (reduc_idx >= 0)
3863 0 : vargs[varg++] = vargs[reduc_idx + 1];
3864 : else
3865 : {
3866 0 : auto else_value = targetm.preferred_else_value
3867 0 : (ifn, vectype_out, varg - 1, &vargs[1]);
3868 0 : vargs[varg++] = else_value;
3869 : }
3870 : }
3871 2713 : if (clz_ctz_arg1)
3872 59 : vargs[varg++] = clz_ctz_arg1;
3873 :
3874 2713 : gimple *new_stmt;
3875 2713 : if (modifier == NARROW)
3876 : {
3877 : /* We don't define any narrowing conditional functions
3878 : at present. */
3879 0 : gcc_assert (mask_opno < 0);
3880 0 : tree half_res = make_ssa_name (vectype_in);
3881 0 : gcall *call = gimple_build_call_internal_vec (ifn, vargs);
3882 0 : gimple_call_set_lhs (call, half_res);
3883 0 : gimple_call_set_nothrow (call, true);
3884 0 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
3885 0 : if ((i & 1) == 0)
3886 : {
3887 0 : prev_res = half_res;
3888 0 : continue;
3889 : }
3890 0 : new_temp = make_ssa_name (vec_dest);
3891 0 : new_stmt = vect_gimple_build (new_temp, convert_code,
3892 : prev_res, half_res);
3893 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
3894 : }
3895 : else
3896 : {
3897 2713 : if (len_opno >= 0 && len_loop_p)
3898 : {
3899 0 : unsigned int vec_num = vec_oprnds0.length ();
3900 0 : tree len = vect_get_loop_len (loop_vinfo, gsi, lens,
3901 : vec_num, vectype_out, i, 1, true);
3902 0 : signed char biasval
3903 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
3904 0 : tree bias = build_int_cst (intQI_type_node, biasval);
3905 0 : vargs[len_opno] = len;
3906 0 : vargs[len_opno + 1] = bias;
3907 : }
3908 2713 : else if (mask_opno >= 0 && masked_loop_p)
3909 : {
3910 36 : unsigned int vec_num = vec_oprnds0.length ();
3911 36 : tree mask = vect_get_loop_mask (loop_vinfo, gsi, masks,
3912 : vec_num, vectype_out, i);
3913 36 : vargs[mask_opno]
3914 72 : = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), mask,
3915 36 : vargs[mask_opno], gsi);
3916 : }
3917 :
3918 2713 : gcall *call;
3919 2713 : if (ifn != IFN_LAST)
3920 2632 : call = gimple_build_call_internal_vec (ifn, vargs);
3921 : else
3922 81 : call = gimple_build_call_vec (fndecl, vargs);
3923 2713 : new_temp = make_ssa_name (vec_dest, call);
3924 2713 : gimple_call_set_lhs (call, new_temp);
3925 2713 : gimple_call_set_nothrow (call, true);
3926 2713 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
3927 2713 : new_stmt = call;
3928 : }
3929 2713 : slp_node->push_vec_def (new_stmt);
3930 : }
3931 : }
3932 :
3933 11035 : for (i = 0; i < nargs; i++)
3934 : {
3935 6869 : vec<tree> vec_oprndsi = vec_defs[i];
3936 6869 : vec_oprndsi.release ();
3937 : }
3938 4166 : }
3939 32 : else if (modifier == NARROW)
3940 : {
3941 32 : auto_vec<vec<tree> > vec_defs (nargs);
3942 : /* We don't define any narrowing conditional functions at present. */
3943 32 : gcc_assert (mask_opno < 0);
3944 :
3945 : /* Build argument list for the vectorized call. */
3946 32 : vargs.create (nargs * 2);
3947 :
3948 32 : vect_get_slp_defs (vinfo, slp_node, &vec_defs);
3949 32 : vec<tree> vec_oprnds0 = vec_defs[0];
3950 :
3951 : /* Arguments are ready. Create the new vector stmt. */
3952 64 : for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2)
3953 : {
3954 32 : size_t k;
3955 32 : vargs.truncate (0);
3956 64 : for (k = 0; k < nargs; k++)
3957 : {
3958 32 : vec<tree> vec_oprndsk = vec_defs[k];
3959 32 : vargs.quick_push (vec_oprndsk[i]);
3960 32 : vargs.quick_push (vec_oprndsk[i + 1]);
3961 : }
3962 32 : gcall *call;
3963 32 : if (ifn != IFN_LAST)
3964 : call = gimple_build_call_internal_vec (ifn, vargs);
3965 : else
3966 32 : call = gimple_build_call_vec (fndecl, vargs);
3967 32 : new_temp = make_ssa_name (vec_dest, call);
3968 32 : gimple_call_set_lhs (call, new_temp);
3969 32 : gimple_call_set_nothrow (call, true);
3970 32 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
3971 32 : slp_node->push_vec_def (call);
3972 : }
3973 :
3974 64 : for (i = 0; i < nargs; i++)
3975 : {
3976 32 : vec<tree> vec_oprndsi = vec_defs[i];
3977 32 : vec_oprndsi.release ();
3978 : }
3979 32 : }
3980 : else
3981 : /* No current target implements this case. */
3982 : return false;
3983 :
3984 4198 : vargs.release ();
3985 :
3986 4198 : return true;
3987 2334134 : }
3988 :
3989 :
3990 : struct simd_call_arg_info
3991 : {
3992 : tree vectype;
3993 : tree op;
3994 : HOST_WIDE_INT linear_step;
3995 : enum vect_def_type dt;
3996 : unsigned int align;
3997 : bool simd_lane_linear;
3998 : };
3999 :
4000 : /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME,
4001 : is linear within simd lane (but not within whole loop), note it in
4002 : *ARGINFO. */
4003 :
4004 : static void
4005 15 : vect_simd_lane_linear (tree op, class loop *loop,
4006 : struct simd_call_arg_info *arginfo)
4007 : {
4008 15 : gimple *def_stmt = SSA_NAME_DEF_STMT (op);
4009 :
4010 15 : if (!is_gimple_assign (def_stmt)
4011 15 : || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR
4012 27 : || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt)))
4013 3 : return;
4014 :
4015 12 : tree base = gimple_assign_rhs1 (def_stmt);
4016 12 : HOST_WIDE_INT linear_step = 0;
4017 12 : tree v = gimple_assign_rhs2 (def_stmt);
4018 48 : while (TREE_CODE (v) == SSA_NAME)
4019 : {
4020 36 : tree t;
4021 36 : def_stmt = SSA_NAME_DEF_STMT (v);
4022 36 : if (is_gimple_assign (def_stmt))
4023 24 : switch (gimple_assign_rhs_code (def_stmt))
4024 : {
4025 0 : case PLUS_EXPR:
4026 0 : t = gimple_assign_rhs2 (def_stmt);
4027 0 : if (linear_step || TREE_CODE (t) != INTEGER_CST)
4028 : return;
4029 0 : base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t);
4030 0 : v = gimple_assign_rhs1 (def_stmt);
4031 0 : continue;
4032 12 : case MULT_EXPR:
4033 12 : t = gimple_assign_rhs2 (def_stmt);
4034 12 : if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t))
4035 0 : return;
4036 12 : linear_step = tree_to_shwi (t);
4037 12 : v = gimple_assign_rhs1 (def_stmt);
4038 12 : continue;
4039 12 : CASE_CONVERT:
4040 12 : t = gimple_assign_rhs1 (def_stmt);
4041 12 : if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE
4042 12 : || (TYPE_PRECISION (TREE_TYPE (v))
4043 12 : < TYPE_PRECISION (TREE_TYPE (t))))
4044 : return;
4045 12 : if (!linear_step)
4046 0 : linear_step = 1;
4047 12 : v = t;
4048 12 : continue;
4049 : default:
4050 : return;
4051 : }
4052 12 : else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE)
4053 12 : && loop->simduid
4054 12 : && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME
4055 24 : && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0))
4056 : == loop->simduid))
4057 : {
4058 12 : if (!linear_step)
4059 0 : linear_step = 1;
4060 12 : arginfo->linear_step = linear_step;
4061 12 : arginfo->op = base;
4062 12 : arginfo->simd_lane_linear = true;
4063 12 : return;
4064 : }
4065 : }
4066 : }
4067 :
4068 : /* Function vectorizable_simd_clone_call.
4069 :
4070 : Check if STMT_INFO performs a function call that can be vectorized
4071 : by calling a simd clone of the function.
4072 : If COST_VEC is passed, calculate costs but don't change anything,
4073 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
4074 : it, and insert it at GSI.
4075 : Return true if STMT_INFO is vectorizable in this way. */
4076 :
4077 : static bool
4078 2324768 : vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
4079 : gimple_stmt_iterator *gsi,
4080 : slp_tree slp_node,
4081 : stmt_vector_for_cost *cost_vec)
4082 : {
4083 2324768 : tree vec_dest;
4084 2324768 : tree scalar_dest;
4085 2324768 : tree vec_oprnd0 = NULL_TREE;
4086 2324768 : tree vectype;
4087 2324768 : poly_uint64 nunits;
4088 2324768 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
4089 2324768 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
4090 2324768 : class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
4091 2324768 : tree fndecl, new_temp;
4092 2324768 : int j;
4093 2324768 : auto_vec<simd_call_arg_info> arginfo;
4094 2324768 : vec<tree> vargs = vNULL;
4095 2324768 : size_t i, nargs;
4096 2324768 : tree rtype, ratype;
4097 2324768 : vec<constructor_elt, va_gc> *ret_ctor_elts = NULL;
4098 2324768 : int masked_call_offset = 0;
4099 :
4100 : /* Is STMT a vectorizable call? */
4101 2324768 : gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt);
4102 10894 : if (!stmt)
4103 : return false;
4104 :
4105 10894 : fndecl = gimple_call_fndecl (stmt);
4106 10894 : if (fndecl == NULL_TREE
4107 10894 : && gimple_call_internal_p (stmt, IFN_MASK_CALL))
4108 : {
4109 220 : fndecl = gimple_call_arg (stmt, 0);
4110 220 : gcc_checking_assert (TREE_CODE (fndecl) == ADDR_EXPR);
4111 220 : fndecl = TREE_OPERAND (fndecl, 0);
4112 220 : gcc_checking_assert (TREE_CODE (fndecl) == FUNCTION_DECL);
4113 : masked_call_offset = 1;
4114 : }
4115 10674 : if (fndecl == NULL_TREE)
4116 : return false;
4117 :
4118 4655 : struct cgraph_node *node = cgraph_node::get (fndecl);
4119 4655 : if (node == NULL || node->simd_clones == NULL)
4120 : return false;
4121 :
4122 1454 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
4123 : return false;
4124 :
4125 1454 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
4126 0 : && cost_vec)
4127 : return false;
4128 :
4129 1454 : if (gimple_call_lhs (stmt)
4130 1454 : && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
4131 : return false;
4132 :
4133 1454 : gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt));
4134 :
4135 1454 : vectype = SLP_TREE_VECTYPE (slp_node);
4136 :
4137 2324832 : if (loop_vinfo && nested_in_vect_loop_p (loop, stmt_info))
4138 : return false;
4139 :
4140 : /* Process function arguments. */
4141 1454 : nargs = gimple_call_num_args (stmt) - masked_call_offset;
4142 :
4143 : /* Bail out if the function has zero arguments. */
4144 1454 : if (nargs == 0)
4145 : return false;
4146 :
4147 1390 : vect_simd_clone_data _data;
4148 1390 : vect_simd_clone_data &data = slp_node->get_data (_data);
4149 1390 : vec<tree>& simd_clone_info = data.simd_clone_info;
4150 1390 : arginfo.reserve (nargs, true);
4151 1390 : auto_vec<slp_tree> slp_op;
4152 1390 : slp_op.safe_grow_cleared (nargs);
4153 :
4154 3996 : for (i = 0; i < nargs; i++)
4155 : {
4156 2606 : simd_call_arg_info thisarginfo;
4157 2606 : affine_iv iv;
4158 2606 : tree op;
4159 :
4160 2606 : thisarginfo.linear_step = 0;
4161 2606 : thisarginfo.align = 0;
4162 2606 : thisarginfo.op = NULL_TREE;
4163 2606 : thisarginfo.simd_lane_linear = false;
4164 :
4165 5212 : int op_no = vect_slp_child_index_for_operand (stmt,
4166 2606 : i + masked_call_offset,
4167 : false);
4168 5212 : if (!vect_is_simple_use (vinfo, slp_node,
4169 2606 : op_no, &op, &slp_op[i],
4170 : &thisarginfo.dt, &thisarginfo.vectype)
4171 2606 : || thisarginfo.dt == vect_uninitialized_def)
4172 : {
4173 0 : if (dump_enabled_p ())
4174 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
4175 : "use not simple.\n");
4176 0 : return false;
4177 : }
4178 :
4179 2606 : if (thisarginfo.dt == vect_constant_def
4180 2606 : || thisarginfo.dt == vect_external_def)
4181 : {
4182 : /* With SLP we determine the vector type of constants/externals
4183 : at analysis time, handling conflicts via
4184 : vect_maybe_update_slp_op_vectype. At transform time
4185 : we have a vector type recorded for SLP. */
4186 699 : gcc_assert (cost_vec
4187 : || thisarginfo.vectype != NULL_TREE);
4188 : if (cost_vec)
4189 568 : thisarginfo.vectype = get_vectype_for_scalar_type (vinfo,
4190 568 : TREE_TYPE (op),
4191 : slp_node);
4192 : }
4193 : else
4194 1907 : gcc_assert (thisarginfo.vectype != NULL_TREE);
4195 :
4196 : /* For linear arguments, the analyze phase should have saved
4197 : the base and step. */
4198 2475 : if (!cost_vec
4199 1584 : && i * 3 + 4 <= simd_clone_info.length ()
4200 2685 : && simd_clone_info[i * 3 + 2])
4201 : {
4202 118 : thisarginfo.linear_step = tree_to_shwi (simd_clone_info[i * 3 + 2]);
4203 118 : thisarginfo.op = simd_clone_info[i * 3 + 1];
4204 118 : thisarginfo.simd_lane_linear
4205 118 : = (simd_clone_info[i * 3 + 3] == boolean_true_node);
4206 : /* If loop has been peeled for alignment, we need to adjust it. */
4207 118 : tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo);
4208 118 : tree n2 = LOOP_VINFO_NITERS (loop_vinfo);
4209 118 : if (n1 != n2 && !thisarginfo.simd_lane_linear)
4210 : {
4211 0 : tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2);
4212 0 : tree step = simd_clone_info[i * 3 + 2];
4213 0 : tree opt = TREE_TYPE (thisarginfo.op);
4214 0 : bias = fold_convert (TREE_TYPE (step), bias);
4215 0 : bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step);
4216 0 : thisarginfo.op
4217 0 : = fold_build2 (POINTER_TYPE_P (opt)
4218 : ? POINTER_PLUS_EXPR : PLUS_EXPR, opt,
4219 : thisarginfo.op, bias);
4220 : }
4221 : }
4222 2488 : else if (cost_vec
4223 1814 : && thisarginfo.dt != vect_constant_def
4224 1692 : && thisarginfo.dt != vect_external_def
4225 1246 : && loop_vinfo
4226 1241 : && SLP_TREE_LANES (slp_node) == 1
4227 1217 : && TREE_CODE (op) == SSA_NAME
4228 2434 : && simple_iv (loop, loop_containing_stmt (stmt), op,
4229 : &iv, false)
4230 2694 : && tree_fits_shwi_p (iv.step))
4231 : {
4232 206 : thisarginfo.linear_step = tree_to_shwi (iv.step);
4233 206 : thisarginfo.op = iv.base;
4234 : }
4235 2282 : else if ((thisarginfo.dt == vect_constant_def
4236 2282 : || thisarginfo.dt == vect_external_def)
4237 699 : && SLP_TREE_LANES (slp_node) == 1
4238 2583 : && POINTER_TYPE_P (TREE_TYPE (op)))
4239 86 : thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT;
4240 : /* Addresses of array elements indexed by GOMP_SIMD_LANE are
4241 : linear too. */
4242 2606 : if (SLP_TREE_LANES (slp_node) == 1
4243 2162 : && POINTER_TYPE_P (TREE_TYPE (op))
4244 196 : && !thisarginfo.linear_step
4245 112 : && cost_vec
4246 58 : && thisarginfo.dt != vect_constant_def
4247 58 : && thisarginfo.dt != vect_external_def
4248 15 : && loop_vinfo
4249 2621 : && TREE_CODE (op) == SSA_NAME)
4250 15 : vect_simd_lane_linear (op, loop, &thisarginfo);
4251 :
4252 2606 : if (!vectype)
4253 12 : vectype = thisarginfo.vectype;
4254 2606 : arginfo.quick_push (thisarginfo);
4255 : }
4256 :
4257 1390 : poly_uint64 vf = loop_vinfo ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) : 1;
4258 1390 : unsigned group_size = SLP_TREE_LANES (slp_node);
4259 1390 : unsigned int badness = 0;
4260 1390 : unsigned int badness_inbranch = 0;
4261 1390 : struct cgraph_node *bestn = NULL;
4262 1390 : struct cgraph_node *bestn_inbranch = NULL;
4263 1390 : if (!cost_vec)
4264 357 : bestn = ((loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
4265 357 : ? data.clone_inbranch : data.clone);
4266 : else
4267 5979 : for (struct cgraph_node *n = node->simd_clones; n != NULL;
4268 4946 : n = n->simdclone->next_clone)
4269 : {
4270 4946 : unsigned int this_badness = 0;
4271 4946 : unsigned int num_calls;
4272 : /* The number of arguments in the call and the number of parameters in
4273 : the simdclone should match. However, when the simdclone is
4274 : 'inbranch', it could have one more paramater than nargs when using
4275 : an inbranch simdclone to call a non-inbranch call, either in a
4276 : non-masked loop using a all true constant mask, or inside a masked
4277 : loop using it's mask. */
4278 4946 : size_t simd_nargs = n->simdclone->nargs;
4279 4946 : if (!masked_call_offset && n->simdclone->inbranch)
4280 2333 : simd_nargs--;
4281 4946 : if (!constant_multiple_p (vf * group_size, n->simdclone->simdlen,
4282 : &num_calls)
4283 1952 : || (!n->simdclone->inbranch && (masked_call_offset > 0))
4284 1768 : || (nargs != simd_nargs))
4285 3178 : continue;
4286 1768 : if (num_calls != 1)
4287 1136 : this_badness += floor_log2 (num_calls) * 4096;
4288 1768 : if (n->simdclone->inbranch)
4289 764 : this_badness += 8192;
4290 :
4291 : /* If SLP_TREE_VECTYPE has not been set yet pass the general vector
4292 : mode, which for targets that use it will determine what ISA we can
4293 : vectorize this code with. */
4294 1768 : machine_mode vector_mode = vinfo->vector_mode;
4295 1768 : if (vectype)
4296 1768 : vector_mode = TYPE_MODE (vectype);
4297 1768 : int target_badness = targetm.simd_clone.usable (n, vector_mode);
4298 1768 : if (target_badness < 0)
4299 368 : continue;
4300 1400 : this_badness += target_badness * 512;
4301 4146 : for (i = 0; i < nargs; i++)
4302 : {
4303 2994 : switch (n->simdclone->args[i].arg_type)
4304 : {
4305 2064 : case SIMD_CLONE_ARG_TYPE_VECTOR:
4306 2064 : if (VECTOR_BOOLEAN_TYPE_P (n->simdclone->args[i].vector_type))
4307 : /* Vector mask arguments are not supported. */
4308 : i = -1;
4309 2056 : else if (!useless_type_conversion_p
4310 2056 : (n->simdclone->args[i].orig_type,
4311 2056 : TREE_TYPE (gimple_call_arg (stmt,
4312 : i + masked_call_offset))))
4313 : i = -1;
4314 2056 : else if (arginfo[i].dt == vect_constant_def
4315 1951 : || arginfo[i].dt == vect_external_def
4316 3943 : || arginfo[i].linear_step)
4317 392 : this_badness += 64;
4318 : break;
4319 310 : case SIMD_CLONE_ARG_TYPE_UNIFORM:
4320 310 : if ((arginfo[i].dt != vect_constant_def
4321 145 : && arginfo[i].dt != vect_external_def)
4322 410 : || SLP_TREE_LANES (slp_node) != 1)
4323 : i = -1;
4324 : break;
4325 324 : case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP:
4326 324 : case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP:
4327 324 : if (arginfo[i].dt == vect_constant_def
4328 324 : || arginfo[i].dt == vect_external_def
4329 324 : || (arginfo[i].linear_step
4330 324 : != n->simdclone->args[i].linear_step))
4331 : i = -1;
4332 : break;
4333 : case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP:
4334 : case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP:
4335 : case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP:
4336 : case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP:
4337 : case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP:
4338 : case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP:
4339 : /* FORNOW */
4340 : i = -1;
4341 : break;
4342 296 : case SIMD_CLONE_ARG_TYPE_MASK:
4343 296 : if (!SCALAR_INT_MODE_P (n->simdclone->mask_mode)
4344 264 : && n->simdclone->mask_mode != VOIDmode)
4345 : i = -1;
4346 : /* While we can create a traditional data vector from
4347 : an incoming integer mode mask we have no good way to
4348 : force generate an integer mode mask from a traditional
4349 : boolean vector input. */
4350 296 : else if (SCALAR_INT_MODE_P (n->simdclone->mask_mode)
4351 296 : && !SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype)))
4352 : i = -1;
4353 290 : else if (n->simdclone->mask_mode == VOIDmode
4354 : /* FORNOW we only have partial support for vector-type
4355 : masks that can't hold all of simdlen. */
4356 554 : && (maybe_ne (TYPE_VECTOR_SUBPARTS (n->simdclone->args[i].vector_type),
4357 264 : TYPE_VECTOR_SUBPARTS (arginfo[i].vectype))
4358 : /* Verify we can compute the mask argument. */
4359 111 : || !expand_vec_cond_expr_p (n->simdclone->args[i].vector_type,
4360 111 : arginfo[i].vectype)))
4361 : i = -1;
4362 125 : else if (SCALAR_INT_MODE_P (n->simdclone->mask_mode)
4363 : /* FORNOW we only have partial support for
4364 : integer-type masks that represent the same number
4365 : of lanes as the vectorized mask inputs. */
4366 151 : && maybe_ne (exact_div (n->simdclone->simdlen,
4367 : n->simdclone->args[i].linear_step),
4368 26 : TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)))
4369 : i = -1;
4370 107 : else if (!SCALAR_INT_MODE_P (n->simdclone->mask_mode)
4371 107 : && SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype)))
4372 8 : this_badness += 2048;
4373 : break;
4374 : }
4375 183 : if (i == (size_t) -1)
4376 : break;
4377 2746 : if (n->simdclone->args[i].alignment > arginfo[i].align)
4378 : {
4379 : i = -1;
4380 : break;
4381 : }
4382 2746 : if (arginfo[i].align)
4383 110 : this_badness += (exact_log2 (arginfo[i].align)
4384 160 : - exact_log2 (n->simdclone->args[i].alignment));
4385 : }
4386 1400 : if (i == (size_t) -1)
4387 248 : continue;
4388 1152 : if (masked_call_offset == 0
4389 1045 : && n->simdclone->inbranch
4390 340 : && n->simdclone->nargs > nargs)
4391 : {
4392 340 : gcc_assert (n->simdclone->args[n->simdclone->nargs - 1].arg_type ==
4393 : SIMD_CLONE_ARG_TYPE_MASK);
4394 : /* Penalize using a masked SIMD clone in a non-masked loop, that is
4395 : not in a branch, as we'd have to construct an all-true mask. */
4396 340 : this_badness += 64;
4397 : }
4398 1152 : if (bestn == NULL || this_badness < badness)
4399 : {
4400 797 : bestn = n;
4401 797 : badness = this_badness;
4402 : }
4403 1152 : if (n->simdclone->inbranch
4404 447 : && (bestn_inbranch == NULL || this_badness < badness_inbranch))
4405 : {
4406 4946 : bestn_inbranch = n;
4407 4946 : badness_inbranch = this_badness;
4408 : }
4409 : }
4410 :
4411 1390 : if (bestn == NULL)
4412 : return false;
4413 :
4414 809 : fndecl = bestn->decl;
4415 809 : nunits = bestn->simdclone->simdlen;
4416 809 : int ncopies = vector_unroll_factor (vf * group_size, nunits);
4417 :
4418 : /* If the function isn't const, only allow it in simd loops where user
4419 : has asserted that at least nunits consecutive iterations can be
4420 : performed using SIMD instructions. */
4421 804 : if ((loop == NULL || maybe_lt ((unsigned) loop->safelen, nunits))
4422 966 : && gimple_vuse (stmt))
4423 : return false;
4424 :
4425 : /* ncopies is the number of SIMD clone calls we create, since simdlen
4426 : is not necessarily matching nunits of the vector types used, track
4427 : that in ncopies_in. */
4428 809 : int ncopies_in = vect_get_num_vectors (vf * group_size, vectype);
4429 :
4430 : /* Sanity check: make sure that at least one copy of the vectorized stmt
4431 : needs to be generated. */
4432 809 : gcc_assert (ncopies >= 1);
4433 :
4434 809 : if (cost_vec) /* transformation not required. */
4435 : {
4436 1483 : for (unsigned i = 0; i < nargs; ++i)
4437 1031 : if (!vect_maybe_update_slp_op_vectype (slp_op[i], arginfo[i].vectype))
4438 : {
4439 0 : if (dump_enabled_p ())
4440 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
4441 : "incompatible vector types for invariants\n");
4442 0 : return false;
4443 : }
4444 :
4445 452 : if (!bestn_inbranch && loop_vinfo)
4446 : {
4447 237 : if (dump_enabled_p ()
4448 237 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
4449 171 : dump_printf_loc (MSG_NOTE, vect_location,
4450 : "can't use a fully-masked loop because no"
4451 : " masked simd clone was available.\n");
4452 237 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
4453 : }
4454 :
4455 : /* When the original call is pure or const but the SIMD ABI dictates
4456 : an aggregate return we will have to use a virtual definition and
4457 : in a loop eventually even need to add a virtual PHI. That's
4458 : not straight-forward so allow to fix this up via renaming. */
4459 452 : if (gimple_call_lhs (stmt)
4460 446 : && !gimple_vdef (stmt)
4461 802 : && TREE_CODE (TREE_TYPE (TREE_TYPE (bestn->decl))) == ARRAY_TYPE)
4462 27 : vinfo->any_known_not_updated_vssa = true;
4463 : /* ??? For SLP code-gen we end up inserting after the last
4464 : vector argument def rather than at the original call position
4465 : so automagic virtual operand updating doesn't work. */
4466 904 : if (gimple_vuse (stmt))
4467 139 : vinfo->any_known_not_updated_vssa = true;
4468 :
4469 452 : data.clone = bestn;
4470 452 : data.clone_inbranch = bestn_inbranch;
4471 :
4472 452 : simd_clone_info.safe_push (NULL_TREE);
4473 1628 : for (i = 0;
4474 2445 : i < (bestn_inbranch ? bestn_inbranch : bestn)->simdclone->nargs; i++)
4475 : {
4476 1176 : if (loop_vinfo
4477 1170 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
4478 473 : && (bestn_inbranch->simdclone->args[i].arg_type
4479 : == SIMD_CLONE_ARG_TYPE_MASK))
4480 : {
4481 170 : if (masked_call_offset)
4482 : /* When there is an explicit mask we require the
4483 : number of elements to match up. */
4484 49 : vect_record_loop_mask (loop_vinfo,
4485 : &LOOP_VINFO_MASKS (loop_vinfo),
4486 : ncopies_in, vectype, NULL_TREE);
4487 : else
4488 : {
4489 : /* When there is no explicit mask on the call we have
4490 : more relaxed requirements. */
4491 121 : tree masktype;
4492 121 : poly_uint64 callee_nelements;
4493 121 : if (SCALAR_INT_MODE_P (bestn_inbranch->simdclone->mask_mode))
4494 : {
4495 12 : callee_nelements
4496 12 : = exact_div (bestn_inbranch->simdclone->simdlen,
4497 : bestn_inbranch->simdclone->args[i].linear_step);
4498 12 : masktype = get_related_vectype_for_scalar_type
4499 12 : (vinfo->vector_mode, TREE_TYPE (vectype),
4500 : callee_nelements);
4501 : }
4502 : else
4503 : {
4504 109 : masktype = bestn_inbranch->simdclone->args[i].vector_type;
4505 : /* The aarch64 port will add custom attributes to types
4506 : for SVE simdclones which make the types different. We
4507 : should use canonincal types for masks within the
4508 : vectorizer, hence we construct the related vectype
4509 : here. */
4510 109 : masktype
4511 : = build_truth_vector_type_for_mode
4512 109 : (TYPE_VECTOR_SUBPARTS (masktype),
4513 109 : TYPE_MODE (masktype));
4514 109 : callee_nelements = TYPE_VECTOR_SUBPARTS (masktype);
4515 : }
4516 121 : auto o = vector_unroll_factor (nunits, callee_nelements);
4517 121 : vect_record_loop_mask (loop_vinfo,
4518 : &LOOP_VINFO_MASKS (loop_vinfo),
4519 : ncopies * o, masktype, NULL_TREE);
4520 : }
4521 : }
4522 1006 : else if ((bestn->simdclone->args[i].arg_type
4523 : == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP)
4524 899 : || (bestn->simdclone->args[i].arg_type
4525 : == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)
4526 888 : || (bestn_inbranch
4527 359 : && ((bestn_inbranch->simdclone->args[i].arg_type
4528 : == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP)
4529 359 : || (bestn_inbranch->simdclone->args[i].arg_type
4530 : == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP))))
4531 : {
4532 118 : simd_clone_info.safe_grow_cleared (i * 3 + 1, true);
4533 118 : simd_clone_info.safe_push (arginfo[i].op);
4534 202 : tree lst = (POINTER_TYPE_P (TREE_TYPE (arginfo[i].op))
4535 202 : ? size_type_node : TREE_TYPE (arginfo[i].op));
4536 118 : tree ls = build_int_cst (lst, arginfo[i].linear_step);
4537 118 : simd_clone_info.safe_push (ls);
4538 118 : tree sll = (arginfo[i].simd_lane_linear
4539 118 : ? boolean_true_node : boolean_false_node);
4540 118 : simd_clone_info.safe_push (sll);
4541 : }
4542 : }
4543 :
4544 452 : SLP_TREE_TYPE (slp_node) = call_simd_clone_vec_info_type;
4545 452 : slp_node->data = new vect_simd_clone_data (std::move (_data));
4546 452 : DUMP_VECT_SCOPE ("vectorizable_simd_clone_call");
4547 : /* vect_model_simple_cost (vinfo, 1, slp_node, cost_vec); */
4548 452 : return true;
4549 : }
4550 :
4551 : /* Transform. */
4552 :
4553 357 : if (dump_enabled_p ())
4554 246 : dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n");
4555 :
4556 : /* Handle def. */
4557 357 : scalar_dest = gimple_call_lhs (stmt);
4558 357 : vec_dest = NULL_TREE;
4559 357 : rtype = NULL_TREE;
4560 357 : ratype = NULL_TREE;
4561 357 : if (scalar_dest)
4562 : {
4563 351 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
4564 351 : rtype = TREE_TYPE (TREE_TYPE (fndecl));
4565 351 : if (TREE_CODE (rtype) == ARRAY_TYPE)
4566 : {
4567 9 : ratype = rtype;
4568 9 : rtype = TREE_TYPE (ratype);
4569 : }
4570 : }
4571 :
4572 714 : auto_vec<vec<tree> > vec_oprnds;
4573 357 : auto_vec<unsigned> vec_oprnds_i;
4574 357 : vec_oprnds_i.safe_grow_cleared (nargs, true);
4575 357 : vec_oprnds.reserve_exact (nargs);
4576 357 : vect_get_slp_defs (vinfo, slp_node, &vec_oprnds);
4577 823 : for (j = 0; j < ncopies; ++j)
4578 : {
4579 466 : poly_uint64 callee_nelements;
4580 466 : poly_uint64 caller_nelements;
4581 : /* Build argument list for the vectorized call. */
4582 466 : if (j == 0)
4583 357 : vargs.create (nargs);
4584 : else
4585 109 : vargs.truncate (0);
4586 :
4587 1570 : for (i = 0; i < nargs; i++)
4588 : {
4589 1104 : unsigned int k, l, m, o;
4590 1104 : tree atype;
4591 1104 : tree op = gimple_call_arg (stmt, i + masked_call_offset);
4592 1104 : switch (bestn->simdclone->args[i].arg_type)
4593 : {
4594 815 : case SIMD_CLONE_ARG_TYPE_VECTOR:
4595 815 : atype = bestn->simdclone->args[i].vector_type;
4596 815 : caller_nelements = TYPE_VECTOR_SUBPARTS (arginfo[i].vectype);
4597 815 : callee_nelements = TYPE_VECTOR_SUBPARTS (atype);
4598 815 : o = vector_unroll_factor (nunits, callee_nelements);
4599 1860 : for (m = j * o; m < (j + 1) * o; m++)
4600 : {
4601 1045 : if (known_lt (callee_nelements, caller_nelements))
4602 : {
4603 516 : poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype));
4604 258 : if (!constant_multiple_p (caller_nelements,
4605 : callee_nelements, &k))
4606 0 : gcc_unreachable ();
4607 :
4608 258 : gcc_assert ((k & (k - 1)) == 0);
4609 258 : if (m == 0)
4610 : {
4611 57 : vec_oprnds_i[i] = 0;
4612 57 : vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
4613 : }
4614 : else
4615 : {
4616 201 : vec_oprnd0 = arginfo[i].op;
4617 201 : if ((m & (k - 1)) == 0)
4618 72 : vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
4619 : }
4620 258 : arginfo[i].op = vec_oprnd0;
4621 258 : vec_oprnd0
4622 258 : = build3 (BIT_FIELD_REF, atype, vec_oprnd0,
4623 258 : bitsize_int (prec),
4624 258 : bitsize_int ((m & (k - 1)) * prec));
4625 258 : gassign *new_stmt
4626 258 : = gimple_build_assign (make_ssa_name (atype),
4627 : vec_oprnd0);
4628 258 : vect_finish_stmt_generation (vinfo, stmt_info,
4629 : new_stmt, gsi);
4630 258 : vargs.safe_push (gimple_assign_lhs (new_stmt));
4631 : }
4632 : else
4633 : {
4634 787 : if (!constant_multiple_p (callee_nelements,
4635 : caller_nelements, &k))
4636 0 : gcc_unreachable ();
4637 787 : gcc_assert ((k & (k - 1)) == 0);
4638 787 : vec<constructor_elt, va_gc> *ctor_elts;
4639 787 : if (k != 1)
4640 14 : vec_alloc (ctor_elts, k);
4641 : else
4642 773 : ctor_elts = NULL;
4643 815 : for (l = 0; l < k; l++)
4644 : {
4645 801 : if (m == 0 && l == 0)
4646 : {
4647 449 : vec_oprnds_i[i] = 0;
4648 449 : vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
4649 : }
4650 : else
4651 352 : vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
4652 801 : arginfo[i].op = vec_oprnd0;
4653 801 : if (k == 1)
4654 : break;
4655 28 : CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE,
4656 : vec_oprnd0);
4657 : }
4658 787 : if (k == 1)
4659 773 : if (!useless_type_conversion_p (TREE_TYPE (vec_oprnd0),
4660 : atype))
4661 : {
4662 0 : vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, atype,
4663 : vec_oprnd0);
4664 0 : gassign *new_stmt
4665 0 : = gimple_build_assign (make_ssa_name (atype),
4666 : vec_oprnd0);
4667 0 : vect_finish_stmt_generation (vinfo, stmt_info,
4668 : new_stmt, gsi);
4669 0 : vargs.safe_push (gimple_get_lhs (new_stmt));
4670 : }
4671 : else
4672 773 : vargs.safe_push (vec_oprnd0);
4673 : else
4674 : {
4675 14 : vec_oprnd0 = build_constructor (atype, ctor_elts);
4676 14 : gassign *new_stmt
4677 14 : = gimple_build_assign (make_ssa_name (atype),
4678 : vec_oprnd0);
4679 14 : vect_finish_stmt_generation (vinfo, stmt_info,
4680 : new_stmt, gsi);
4681 14 : vargs.safe_push (gimple_assign_lhs (new_stmt));
4682 : }
4683 : }
4684 : }
4685 : break;
4686 66 : case SIMD_CLONE_ARG_TYPE_MASK:
4687 66 : if (bestn->simdclone->mask_mode == VOIDmode)
4688 : {
4689 60 : atype = bestn->simdclone->args[i].vector_type;
4690 60 : tree elt_type = TREE_TYPE (atype);
4691 60 : tree one = fold_convert (elt_type, integer_one_node);
4692 60 : tree zero = fold_convert (elt_type, integer_zero_node);
4693 60 : callee_nelements = TYPE_VECTOR_SUBPARTS (atype);
4694 60 : caller_nelements = TYPE_VECTOR_SUBPARTS (arginfo[i].vectype);
4695 60 : o = vector_unroll_factor (nunits, callee_nelements);
4696 120 : for (m = j * o; m < (j + 1) * o; m++)
4697 : {
4698 60 : if (maybe_lt (callee_nelements, caller_nelements))
4699 : {
4700 : /* The mask type has fewer elements than simdlen. */
4701 :
4702 : /* FORNOW */
4703 0 : gcc_unreachable ();
4704 : }
4705 60 : else if (known_eq (callee_nelements, caller_nelements))
4706 : {
4707 : /* The SIMD clone function has the same number of
4708 : elements as the current function. */
4709 60 : if (m == 0)
4710 60 : vec_oprnds_i[i] = 0;
4711 60 : vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
4712 60 : if (loop_vinfo
4713 60 : && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
4714 : {
4715 0 : vec_loop_masks *loop_masks
4716 : = &LOOP_VINFO_MASKS (loop_vinfo);
4717 0 : tree loop_mask
4718 0 : = vect_get_loop_mask (loop_vinfo, gsi,
4719 : loop_masks, ncopies_in,
4720 0 : vectype, j);
4721 0 : vec_oprnd0
4722 0 : = prepare_vec_mask (loop_vinfo,
4723 0 : TREE_TYPE (loop_mask),
4724 : loop_mask, vec_oprnd0,
4725 : gsi);
4726 0 : loop_vinfo->vec_cond_masked_set.add ({ vec_oprnd0,
4727 : loop_mask });
4728 :
4729 : }
4730 60 : vec_oprnd0
4731 60 : = build3 (VEC_COND_EXPR, atype, vec_oprnd0,
4732 : build_vector_from_val (atype, one),
4733 : build_vector_from_val (atype, zero));
4734 60 : gassign *new_stmt
4735 60 : = gimple_build_assign (make_ssa_name (atype),
4736 : vec_oprnd0);
4737 60 : vect_finish_stmt_generation (vinfo, stmt_info,
4738 : new_stmt, gsi);
4739 60 : vargs.safe_push (gimple_assign_lhs (new_stmt));
4740 : }
4741 : else
4742 : {
4743 : /* The mask type has more elements than simdlen. */
4744 :
4745 : /* FORNOW */
4746 0 : gcc_unreachable ();
4747 : }
4748 : }
4749 : }
4750 6 : else if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode))
4751 : {
4752 6 : atype = bestn->simdclone->args[i].vector_type;
4753 6 : poly_uint64 atype_subparts
4754 6 : = exact_div (bestn->simdclone->simdlen,
4755 : bestn->simdclone->args[i].linear_step);
4756 6 : o = bestn->simdclone->args[i].linear_step;
4757 12 : for (m = j * o; m < (j + 1) * o; m++)
4758 : {
4759 6 : if (m == 0)
4760 6 : vec_oprnds_i[i] = 0;
4761 6 : if (maybe_lt (atype_subparts,
4762 6 : TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)))
4763 : {
4764 : /* The mask argument has fewer elements than the
4765 : input vector. */
4766 : /* FORNOW */
4767 0 : gcc_unreachable ();
4768 : }
4769 6 : else if (known_eq (atype_subparts,
4770 : TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)))
4771 : {
4772 6 : vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
4773 6 : if (loop_vinfo
4774 6 : && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
4775 : {
4776 1 : vec_loop_masks *loop_masks
4777 : = &LOOP_VINFO_MASKS (loop_vinfo);
4778 1 : tree loop_mask
4779 1 : = vect_get_loop_mask (loop_vinfo, gsi,
4780 : loop_masks, ncopies_in,
4781 : vectype, j);
4782 1 : vec_oprnd0
4783 1 : = prepare_vec_mask (loop_vinfo,
4784 1 : TREE_TYPE (loop_mask),
4785 : loop_mask, vec_oprnd0,
4786 : gsi);
4787 : }
4788 : /* The vector mask argument matches the input
4789 : in the number of lanes, but not necessarily
4790 : in the mode. */
4791 6 : tree st = lang_hooks.types.type_for_mode
4792 6 : (TYPE_MODE (TREE_TYPE (vec_oprnd0)), 1);
4793 6 : vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, st,
4794 : vec_oprnd0);
4795 6 : gassign *new_stmt
4796 6 : = gimple_build_assign (make_ssa_name (st),
4797 : vec_oprnd0);
4798 6 : vect_finish_stmt_generation (vinfo, stmt_info,
4799 : new_stmt, gsi);
4800 6 : if (!types_compatible_p (atype, st))
4801 : {
4802 6 : new_stmt
4803 6 : = gimple_build_assign (make_ssa_name (atype),
4804 : NOP_EXPR,
4805 : gimple_assign_lhs
4806 : (new_stmt));
4807 6 : vect_finish_stmt_generation (vinfo, stmt_info,
4808 : new_stmt, gsi);
4809 : }
4810 6 : vargs.safe_push (gimple_assign_lhs (new_stmt));
4811 : }
4812 : else
4813 : {
4814 : /* The mask argument has more elements than the
4815 : input vector. */
4816 : /* FORNOW */
4817 0 : gcc_unreachable ();
4818 : }
4819 : }
4820 : }
4821 : else
4822 0 : gcc_unreachable ();
4823 : break;
4824 102 : case SIMD_CLONE_ARG_TYPE_UNIFORM:
4825 102 : vargs.safe_push (op);
4826 102 : break;
4827 121 : case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP:
4828 121 : case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP:
4829 121 : if (j == 0)
4830 : {
4831 118 : gimple_seq stmts;
4832 118 : arginfo[i].op
4833 118 : = force_gimple_operand (unshare_expr (arginfo[i].op),
4834 : &stmts, true, NULL_TREE);
4835 118 : if (stmts != NULL)
4836 : {
4837 0 : basic_block new_bb;
4838 0 : edge pe = loop_preheader_edge (loop);
4839 0 : new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts);
4840 0 : gcc_assert (!new_bb);
4841 : }
4842 118 : if (arginfo[i].simd_lane_linear)
4843 : {
4844 6 : vargs.safe_push (arginfo[i].op);
4845 6 : break;
4846 : }
4847 112 : tree phi_res = copy_ssa_name (op);
4848 112 : gphi *new_phi = create_phi_node (phi_res, loop->header);
4849 112 : add_phi_arg (new_phi, arginfo[i].op,
4850 : loop_preheader_edge (loop), UNKNOWN_LOCATION);
4851 112 : enum tree_code code
4852 196 : = POINTER_TYPE_P (TREE_TYPE (op))
4853 112 : ? POINTER_PLUS_EXPR : PLUS_EXPR;
4854 196 : tree type = POINTER_TYPE_P (TREE_TYPE (op))
4855 196 : ? sizetype : TREE_TYPE (op);
4856 112 : poly_widest_int cst
4857 112 : = wi::mul (bestn->simdclone->args[i].linear_step,
4858 112 : ncopies * nunits);
4859 112 : tree tcst = wide_int_to_tree (type, cst);
4860 112 : tree phi_arg = copy_ssa_name (op);
4861 112 : gassign *new_stmt
4862 112 : = gimple_build_assign (phi_arg, code, phi_res, tcst);
4863 112 : gimple_stmt_iterator si = gsi_after_labels (loop->header);
4864 112 : gsi_insert_after (&si, new_stmt, GSI_NEW_STMT);
4865 112 : add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop),
4866 : UNKNOWN_LOCATION);
4867 112 : arginfo[i].op = phi_res;
4868 112 : vargs.safe_push (phi_res);
4869 112 : }
4870 : else
4871 : {
4872 3 : enum tree_code code
4873 6 : = POINTER_TYPE_P (TREE_TYPE (op))
4874 3 : ? POINTER_PLUS_EXPR : PLUS_EXPR;
4875 6 : tree type = POINTER_TYPE_P (TREE_TYPE (op))
4876 6 : ? sizetype : TREE_TYPE (op);
4877 3 : poly_widest_int cst
4878 3 : = wi::mul (bestn->simdclone->args[i].linear_step,
4879 3 : j * nunits);
4880 3 : tree tcst = wide_int_to_tree (type, cst);
4881 3 : new_temp = make_ssa_name (TREE_TYPE (op));
4882 3 : gassign *new_stmt
4883 6 : = gimple_build_assign (new_temp, code,
4884 3 : arginfo[i].op, tcst);
4885 3 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
4886 3 : vargs.safe_push (new_temp);
4887 3 : }
4888 : break;
4889 0 : case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP:
4890 0 : case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP:
4891 0 : case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP:
4892 0 : case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP:
4893 0 : case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP:
4894 0 : case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP:
4895 0 : default:
4896 0 : gcc_unreachable ();
4897 : }
4898 : }
4899 :
4900 466 : if (masked_call_offset == 0
4901 400 : && bestn->simdclone->inbranch
4902 8 : && bestn->simdclone->nargs > nargs)
4903 : {
4904 8 : unsigned long m, o;
4905 8 : size_t mask_i = bestn->simdclone->nargs - 1;
4906 8 : tree mask;
4907 8 : gcc_assert (bestn->simdclone->args[mask_i].arg_type ==
4908 : SIMD_CLONE_ARG_TYPE_MASK);
4909 :
4910 8 : tree mask_argtype = bestn->simdclone->args[mask_i].vector_type;
4911 8 : tree mask_vectype;
4912 8 : if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode))
4913 : {
4914 2 : callee_nelements = exact_div (bestn->simdclone->simdlen,
4915 : bestn->simdclone->args[i].linear_step);
4916 2 : mask_vectype = get_related_vectype_for_scalar_type
4917 2 : (vinfo->vector_mode, TREE_TYPE (vectype), callee_nelements);
4918 : }
4919 : else
4920 : {
4921 6 : mask_vectype = mask_argtype;
4922 6 : callee_nelements = TYPE_VECTOR_SUBPARTS (mask_vectype);
4923 : }
4924 8 : o = vector_unroll_factor (nunits, callee_nelements);
4925 16 : for (m = j * o; m < (j + 1) * o; m++)
4926 : {
4927 8 : if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
4928 : {
4929 1 : vec_loop_masks *loop_masks = &LOOP_VINFO_MASKS (loop_vinfo);
4930 1 : mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
4931 : ncopies * o, mask_vectype, m);
4932 : }
4933 : else
4934 7 : mask = vect_build_all_ones_mask (vinfo, stmt_info,
4935 : mask_argtype);
4936 :
4937 8 : gassign *new_stmt;
4938 8 : if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode))
4939 : {
4940 : /* This means we are dealing with integer mask modes.
4941 : First convert to an integer type with the same size as
4942 : the current vector type. */
4943 2 : unsigned HOST_WIDE_INT intermediate_size
4944 2 : = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (mask)));
4945 2 : tree mid_int_type =
4946 2 : build_nonstandard_integer_type (intermediate_size, 1);
4947 2 : mask = build1 (VIEW_CONVERT_EXPR, mid_int_type, mask);
4948 2 : new_stmt
4949 2 : = gimple_build_assign (make_ssa_name (mid_int_type),
4950 : mask);
4951 2 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4952 : /* Then zero-extend to the mask mode. */
4953 2 : mask = fold_build1 (NOP_EXPR, mask_argtype,
4954 : gimple_get_lhs (new_stmt));
4955 : }
4956 6 : else if (bestn->simdclone->mask_mode == VOIDmode)
4957 6 : mask = build3 (VEC_COND_EXPR, mask_argtype, mask,
4958 : build_one_cst (mask_argtype),
4959 : build_zero_cst (mask_argtype));
4960 : else
4961 0 : gcc_unreachable ();
4962 :
4963 8 : new_stmt = gimple_build_assign (make_ssa_name (mask_argtype),
4964 : mask);
4965 8 : vect_finish_stmt_generation (vinfo, stmt_info,
4966 : new_stmt, gsi);
4967 8 : mask = gimple_assign_lhs (new_stmt);
4968 8 : vargs.safe_push (mask);
4969 : }
4970 : }
4971 :
4972 466 : gcall *new_call = gimple_build_call_vec (fndecl, vargs);
4973 466 : if (vec_dest)
4974 : {
4975 460 : gcc_assert (ratype
4976 : || known_eq (TYPE_VECTOR_SUBPARTS (rtype), nunits));
4977 460 : if (ratype)
4978 15 : new_temp = create_tmp_var (ratype);
4979 445 : else if (useless_type_conversion_p (vectype, rtype))
4980 423 : new_temp = make_ssa_name (vec_dest, new_call);
4981 : else
4982 22 : new_temp = make_ssa_name (rtype, new_call);
4983 460 : gimple_call_set_lhs (new_call, new_temp);
4984 : }
4985 466 : vect_finish_stmt_generation (vinfo, stmt_info, new_call, gsi);
4986 466 : gimple *new_stmt = new_call;
4987 :
4988 466 : if (vec_dest)
4989 : {
4990 460 : if (!multiple_p (TYPE_VECTOR_SUBPARTS (vectype), nunits))
4991 : {
4992 21 : unsigned int k, l;
4993 42 : poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype));
4994 42 : poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype));
4995 21 : k = vector_unroll_factor (nunits,
4996 : TYPE_VECTOR_SUBPARTS (vectype));
4997 21 : gcc_assert ((k & (k - 1)) == 0);
4998 75 : for (l = 0; l < k; l++)
4999 : {
5000 54 : tree t;
5001 54 : if (ratype)
5002 : {
5003 42 : t = build_fold_addr_expr (new_temp);
5004 42 : t = build2 (MEM_REF, vectype, t,
5005 42 : build_int_cst (TREE_TYPE (t), l * bytes));
5006 : }
5007 : else
5008 12 : t = build3 (BIT_FIELD_REF, vectype, new_temp,
5009 12 : bitsize_int (prec), bitsize_int (l * prec));
5010 54 : new_stmt = gimple_build_assign (make_ssa_name (vectype), t);
5011 54 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5012 :
5013 54 : SLP_TREE_VEC_DEFS (slp_node)
5014 54 : .quick_push (gimple_assign_lhs (new_stmt));
5015 : }
5016 :
5017 21 : if (ratype)
5018 15 : vect_clobber_variable (vinfo, stmt_info, gsi, new_temp);
5019 21 : continue;
5020 21 : }
5021 439 : else if (!multiple_p (nunits, TYPE_VECTOR_SUBPARTS (vectype)))
5022 : {
5023 16 : unsigned int k;
5024 16 : if (!constant_multiple_p (TYPE_VECTOR_SUBPARTS (vectype),
5025 16 : TYPE_VECTOR_SUBPARTS (rtype), &k))
5026 0 : gcc_unreachable ();
5027 16 : gcc_assert ((k & (k - 1)) == 0);
5028 16 : if ((j & (k - 1)) == 0)
5029 8 : vec_alloc (ret_ctor_elts, k);
5030 16 : if (ratype)
5031 : {
5032 0 : unsigned int m, o;
5033 0 : o = vector_unroll_factor (nunits,
5034 : TYPE_VECTOR_SUBPARTS (rtype));
5035 0 : for (m = 0; m < o; m++)
5036 : {
5037 0 : tree tem = build4 (ARRAY_REF, rtype, new_temp,
5038 0 : size_int (m), NULL_TREE, NULL_TREE);
5039 0 : new_stmt = gimple_build_assign (make_ssa_name (rtype),
5040 : tem);
5041 0 : vect_finish_stmt_generation (vinfo, stmt_info,
5042 : new_stmt, gsi);
5043 0 : CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE,
5044 : gimple_assign_lhs (new_stmt));
5045 : }
5046 0 : vect_clobber_variable (vinfo, stmt_info, gsi, new_temp);
5047 : }
5048 : else
5049 16 : CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp);
5050 16 : if ((j & (k - 1)) != k - 1)
5051 8 : continue;
5052 8 : vec_oprnd0 = build_constructor (vectype, ret_ctor_elts);
5053 8 : new_stmt
5054 8 : = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0);
5055 8 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5056 :
5057 8 : SLP_TREE_VEC_DEFS (slp_node)
5058 8 : .quick_push (gimple_assign_lhs (new_stmt));
5059 8 : continue;
5060 8 : }
5061 423 : else if (ratype)
5062 : {
5063 0 : tree t = build_fold_addr_expr (new_temp);
5064 0 : t = build2 (MEM_REF, vectype, t,
5065 0 : build_int_cst (TREE_TYPE (t), 0));
5066 0 : new_stmt = gimple_build_assign (make_ssa_name (vec_dest), t);
5067 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5068 0 : vect_clobber_variable (vinfo, stmt_info, gsi, new_temp);
5069 : }
5070 423 : else if (!useless_type_conversion_p (vectype, rtype))
5071 : {
5072 0 : vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, vectype, new_temp);
5073 0 : new_stmt
5074 0 : = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0);
5075 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5076 : }
5077 : }
5078 :
5079 429 : if (gimple_get_lhs (new_stmt))
5080 423 : SLP_TREE_VEC_DEFS (slp_node).quick_push (gimple_get_lhs (new_stmt));
5081 : }
5082 :
5083 1149 : for (i = 0; i < nargs; ++i)
5084 : {
5085 792 : vec<tree> oprndsi = vec_oprnds[i];
5086 792 : oprndsi.release ();
5087 : }
5088 357 : vargs.release ();
5089 :
5090 : /* Mark the clone as no longer being a candidate for GC. */
5091 357 : bestn->gc_candidate = false;
5092 :
5093 357 : return true;
5094 1390 : }
5095 :
5096 :
5097 : /* Function vect_gen_widened_results_half
5098 :
5099 : Create a vector stmt whose code, type, number of arguments, and result
5100 : variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are
5101 : VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at GSI.
5102 : In the case that CODE is a CALL_EXPR, this means that a call to DECL
5103 : needs to be created (DECL is a function-decl of a target-builtin).
5104 : STMT_INFO is the original scalar stmt that we are vectorizing. */
5105 :
5106 : static gimple *
5107 32664 : vect_gen_widened_results_half (vec_info *vinfo, code_helper ch,
5108 : tree vec_oprnd0, tree vec_oprnd1, int op_type,
5109 : tree vec_dest, gimple_stmt_iterator *gsi,
5110 : stmt_vec_info stmt_info)
5111 : {
5112 32664 : gimple *new_stmt;
5113 32664 : tree new_temp;
5114 :
5115 : /* Generate half of the widened result: */
5116 32664 : if (op_type != binary_op)
5117 31544 : vec_oprnd1 = NULL;
5118 32664 : new_stmt = vect_gimple_build (vec_dest, ch, vec_oprnd0, vec_oprnd1);
5119 32664 : new_temp = make_ssa_name (vec_dest, new_stmt);
5120 32664 : gimple_set_lhs (new_stmt, new_temp);
5121 32664 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5122 :
5123 32664 : return new_stmt;
5124 : }
5125 :
5126 :
5127 : /* Create vectorized demotion statements for vector operands from VEC_OPRNDS.
5128 : For multi-step conversions store the resulting vectors and call the function
5129 : recursively. When NARROW_SRC_P is true, there's still a conversion after
5130 : narrowing, don't store the vectors in the SLP_NODE or in vector info of
5131 : the scalar statement(or in STMT_VINFO_RELATED_STMT chain). */
5132 :
5133 : static void
5134 11978 : vect_create_vectorized_demotion_stmts (vec_info *vinfo, vec<tree> *vec_oprnds,
5135 : int multi_step_cvt,
5136 : stmt_vec_info stmt_info,
5137 : vec<tree> &vec_dsts,
5138 : gimple_stmt_iterator *gsi,
5139 : slp_tree slp_node, code_helper code,
5140 : bool narrow_src_p)
5141 : {
5142 11978 : unsigned int i;
5143 11978 : tree vop0, vop1, new_tmp, vec_dest;
5144 :
5145 11978 : vec_dest = vec_dsts.pop ();
5146 :
5147 28350 : for (i = 0; i < vec_oprnds->length (); i += 2)
5148 : {
5149 : /* Create demotion operation. */
5150 16372 : vop0 = (*vec_oprnds)[i];
5151 16372 : vop1 = (*vec_oprnds)[i + 1];
5152 16372 : gimple *new_stmt = vect_gimple_build (vec_dest, code, vop0, vop1);
5153 16372 : new_tmp = make_ssa_name (vec_dest, new_stmt);
5154 16372 : gimple_set_lhs (new_stmt, new_tmp);
5155 16372 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5156 16372 : if (multi_step_cvt || narrow_src_p)
5157 : /* Store the resulting vector for next recursive call,
5158 : or return the resulting vector_tmp for NARROW FLOAT_EXPR. */
5159 6699 : (*vec_oprnds)[i/2] = new_tmp;
5160 : else
5161 : {
5162 : /* This is the last step of the conversion sequence. Store the
5163 : vectors in SLP_NODE. */
5164 9673 : slp_node->push_vec_def (new_stmt);
5165 : }
5166 : }
5167 :
5168 : /* For multi-step demotion operations we first generate demotion operations
5169 : from the source type to the intermediate types, and then combine the
5170 : results (stored in VEC_OPRNDS) in demotion operation to the destination
5171 : type. */
5172 11978 : if (multi_step_cvt)
5173 : {
5174 : /* At each level of recursion we have half of the operands we had at the
5175 : previous level. */
5176 2971 : vec_oprnds->truncate ((i+1)/2);
5177 2971 : vect_create_vectorized_demotion_stmts (vinfo, vec_oprnds,
5178 : multi_step_cvt - 1,
5179 : stmt_info, vec_dsts, gsi,
5180 2971 : slp_node, VEC_PACK_TRUNC_EXPR,
5181 : narrow_src_p);
5182 : }
5183 :
5184 11978 : vec_dsts.quick_push (vec_dest);
5185 11978 : }
5186 :
5187 :
5188 : /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0
5189 : and VEC_OPRNDS1, for a binary operation associated with scalar statement
5190 : STMT_INFO. For multi-step conversions store the resulting vectors and
5191 : call the function recursively. */
5192 :
5193 : static void
5194 11861 : vect_create_vectorized_promotion_stmts (vec_info *vinfo,
5195 : vec<tree> *vec_oprnds0,
5196 : vec<tree> *vec_oprnds1,
5197 : stmt_vec_info stmt_info, tree vec_dest,
5198 : gimple_stmt_iterator *gsi,
5199 : code_helper ch1,
5200 : code_helper ch2, int op_type)
5201 : {
5202 11861 : int i;
5203 11861 : tree vop0, vop1, new_tmp1, new_tmp2;
5204 11861 : gimple *new_stmt1, *new_stmt2;
5205 11861 : vec<tree> vec_tmp = vNULL;
5206 :
5207 11861 : vec_tmp.create (vec_oprnds0->length () * 2);
5208 40054 : FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0)
5209 : {
5210 16332 : if (op_type == binary_op)
5211 560 : vop1 = (*vec_oprnds1)[i];
5212 : else
5213 : vop1 = NULL_TREE;
5214 :
5215 : /* Generate the two halves of promotion operation. */
5216 16332 : new_stmt1 = vect_gen_widened_results_half (vinfo, ch1, vop0, vop1,
5217 : op_type, vec_dest, gsi,
5218 : stmt_info);
5219 16332 : new_stmt2 = vect_gen_widened_results_half (vinfo, ch2, vop0, vop1,
5220 : op_type, vec_dest, gsi,
5221 : stmt_info);
5222 16332 : if (is_gimple_call (new_stmt1))
5223 : {
5224 0 : new_tmp1 = gimple_call_lhs (new_stmt1);
5225 0 : new_tmp2 = gimple_call_lhs (new_stmt2);
5226 : }
5227 : else
5228 : {
5229 16332 : new_tmp1 = gimple_assign_lhs (new_stmt1);
5230 16332 : new_tmp2 = gimple_assign_lhs (new_stmt2);
5231 : }
5232 :
5233 : /* Store the results for the next step. */
5234 16332 : vec_tmp.quick_push (new_tmp1);
5235 16332 : vec_tmp.quick_push (new_tmp2);
5236 : }
5237 :
5238 11861 : vec_oprnds0->release ();
5239 11861 : *vec_oprnds0 = vec_tmp;
5240 11861 : }
5241 :
5242 : /* Create vectorized promotion stmts for widening stmts using only half the
5243 : potential vector size for input. */
5244 : static void
5245 14 : vect_create_half_widening_stmts (vec_info *vinfo,
5246 : vec<tree> *vec_oprnds0,
5247 : vec<tree> *vec_oprnds1,
5248 : stmt_vec_info stmt_info, tree vec_dest,
5249 : gimple_stmt_iterator *gsi,
5250 : code_helper code1,
5251 : int op_type)
5252 : {
5253 14 : int i;
5254 14 : tree vop0, vop1;
5255 14 : gimple *new_stmt1;
5256 14 : gimple *new_stmt2;
5257 14 : gimple *new_stmt3;
5258 14 : vec<tree> vec_tmp = vNULL;
5259 :
5260 14 : vec_tmp.create (vec_oprnds0->length ());
5261 28 : FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0)
5262 : {
5263 14 : tree new_tmp1, new_tmp2, new_tmp3, out_type;
5264 :
5265 14 : gcc_assert (op_type == binary_op);
5266 14 : vop1 = (*vec_oprnds1)[i];
5267 :
5268 : /* Widen the first vector input. */
5269 14 : out_type = TREE_TYPE (vec_dest);
5270 14 : new_tmp1 = make_ssa_name (out_type);
5271 14 : new_stmt1 = gimple_build_assign (new_tmp1, NOP_EXPR, vop0);
5272 14 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt1, gsi);
5273 14 : if (VECTOR_TYPE_P (TREE_TYPE (vop1)))
5274 : {
5275 : /* Widen the second vector input. */
5276 14 : new_tmp2 = make_ssa_name (out_type);
5277 14 : new_stmt2 = gimple_build_assign (new_tmp2, NOP_EXPR, vop1);
5278 14 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt2, gsi);
5279 : /* Perform the operation. With both vector inputs widened. */
5280 14 : new_stmt3 = vect_gimple_build (vec_dest, code1, new_tmp1, new_tmp2);
5281 : }
5282 : else
5283 : {
5284 : /* Perform the operation. With the single vector input widened. */
5285 0 : new_stmt3 = vect_gimple_build (vec_dest, code1, new_tmp1, vop1);
5286 : }
5287 :
5288 14 : new_tmp3 = make_ssa_name (vec_dest, new_stmt3);
5289 14 : gimple_assign_set_lhs (new_stmt3, new_tmp3);
5290 14 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt3, gsi);
5291 :
5292 : /* Store the results for the next step. */
5293 14 : vec_tmp.quick_push (new_tmp3);
5294 : }
5295 :
5296 14 : vec_oprnds0->release ();
5297 14 : *vec_oprnds0 = vec_tmp;
5298 14 : }
5299 :
5300 :
5301 : /* Check if STMT_INFO performs a conversion operation that can be vectorized.
5302 : If COST_VEC is passed, calculate costs but don't change anything,
5303 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
5304 : it, and insert it at GSI.
5305 : Return true if STMT_INFO is vectorizable in this way. */
5306 :
5307 : static bool
5308 2346893 : vectorizable_conversion (vec_info *vinfo,
5309 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
5310 : slp_tree slp_node,
5311 : stmt_vector_for_cost *cost_vec)
5312 : {
5313 2346893 : tree vec_dest, cvt_op = NULL_TREE;
5314 2346893 : tree scalar_dest;
5315 2346893 : tree op0, op1 = NULL_TREE;
5316 2346893 : tree_code tc1;
5317 2346893 : code_helper code, code1, code2;
5318 2346893 : code_helper codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK;
5319 2346893 : tree new_temp;
5320 2346893 : enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
5321 2346893 : poly_uint64 nunits_in;
5322 2346893 : poly_uint64 nunits_out;
5323 2346893 : tree vectype_out, vectype_in;
5324 2346893 : int i;
5325 2346893 : tree lhs_type, rhs_type;
5326 : /* For conversions between floating point and integer, there're 2 NARROW
5327 : cases. NARROW_SRC is for FLOAT_EXPR, means
5328 : integer --DEMOTION--> integer --FLOAT_EXPR--> floating point.
5329 : This is safe when the range of the source integer can fit into the lower
5330 : precision. NARROW_DST is for FIX_TRUNC_EXPR, means
5331 : floating point --FIX_TRUNC_EXPR--> integer --DEMOTION--> INTEGER.
5332 : For other conversions, when there's narrowing, NARROW_DST is used as
5333 : default. */
5334 2346893 : enum { NARROW_SRC, NARROW_DST, NONE, WIDEN } modifier;
5335 2346893 : vec<tree> vec_oprnds0 = vNULL;
5336 2346893 : vec<tree> vec_oprnds1 = vNULL;
5337 2346893 : tree vop0;
5338 2346893 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
5339 2346893 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
5340 2346893 : int multi_step_cvt = 0;
5341 2346893 : vec<tree> interm_types = vNULL;
5342 2346893 : tree intermediate_type, cvt_type = NULL_TREE;
5343 2346893 : int op_type;
5344 2346893 : unsigned short fltsz;
5345 :
5346 : /* Is STMT a vectorizable conversion? */
5347 :
5348 2346893 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
5349 : return false;
5350 :
5351 2346893 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
5352 194522 : && cost_vec)
5353 : return false;
5354 :
5355 2152371 : gimple* stmt = stmt_info->stmt;
5356 2152371 : if (!(is_gimple_assign (stmt) || is_gimple_call (stmt)))
5357 : return false;
5358 :
5359 2094595 : if (gimple_get_lhs (stmt) == NULL_TREE
5360 2094595 : || TREE_CODE (gimple_get_lhs (stmt)) != SSA_NAME)
5361 772893 : return false;
5362 :
5363 1321702 : if (TREE_CODE (gimple_get_lhs (stmt)) != SSA_NAME)
5364 : return false;
5365 :
5366 1321702 : if (is_gimple_assign (stmt))
5367 : {
5368 1313470 : code = gimple_assign_rhs_code (stmt);
5369 1313470 : op_type = TREE_CODE_LENGTH ((tree_code) code);
5370 : }
5371 8232 : else if (gimple_call_internal_p (stmt))
5372 : {
5373 4359 : code = gimple_call_internal_fn (stmt);
5374 4359 : op_type = gimple_call_num_args (stmt);
5375 : }
5376 : else
5377 : return false;
5378 :
5379 1317829 : bool widen_arith = (code == WIDEN_MULT_EXPR
5380 1315426 : || code == WIDEN_LSHIFT_EXPR
5381 2633255 : || widening_fn_p (code));
5382 :
5383 1315426 : if (!widen_arith
5384 1315426 : && !CONVERT_EXPR_CODE_P (code)
5385 1177797 : && code != FIX_TRUNC_EXPR
5386 1176315 : && code != FLOAT_EXPR)
5387 : return false;
5388 :
5389 : /* Check types of lhs and rhs. */
5390 154935 : scalar_dest = gimple_get_lhs (stmt);
5391 154935 : lhs_type = TREE_TYPE (scalar_dest);
5392 154935 : vectype_out = SLP_TREE_VECTYPE (slp_node);
5393 :
5394 : /* Check the operands of the operation. */
5395 154935 : slp_tree slp_op0, slp_op1 = NULL;
5396 154935 : if (!vect_is_simple_use (vinfo, slp_node,
5397 : 0, &op0, &slp_op0, &dt[0], &vectype_in))
5398 : {
5399 0 : if (dump_enabled_p ())
5400 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5401 : "use not simple.\n");
5402 0 : return false;
5403 : }
5404 :
5405 154935 : rhs_type = TREE_TYPE (op0);
5406 153453 : if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
5407 298321 : && !((INTEGRAL_TYPE_P (lhs_type)
5408 131799 : && INTEGRAL_TYPE_P (rhs_type))
5409 : || (SCALAR_FLOAT_TYPE_P (lhs_type)
5410 7162 : && SCALAR_FLOAT_TYPE_P (rhs_type))))
5411 : return false;
5412 :
5413 150510 : if (!VECTOR_BOOLEAN_TYPE_P (vectype_out)
5414 134739 : && INTEGRAL_TYPE_P (lhs_type)
5415 264666 : && !type_has_mode_precision_p (lhs_type))
5416 : {
5417 445 : if (dump_enabled_p ())
5418 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5419 : "type conversion to bit-precision unsupported\n");
5420 445 : return false;
5421 : }
5422 :
5423 150065 : if (op_type == binary_op)
5424 : {
5425 2403 : gcc_assert (code == WIDEN_MULT_EXPR
5426 : || code == WIDEN_LSHIFT_EXPR
5427 : || widening_fn_p (code));
5428 :
5429 2403 : op1 = is_gimple_assign (stmt) ? gimple_assign_rhs2 (stmt) :
5430 0 : gimple_call_arg (stmt, 0);
5431 2403 : tree vectype1_in;
5432 2403 : if (!vect_is_simple_use (vinfo, slp_node, 1,
5433 : &op1, &slp_op1, &dt[1], &vectype1_in))
5434 : {
5435 0 : if (dump_enabled_p ())
5436 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5437 : "use not simple.\n");
5438 0 : return false;
5439 : }
5440 : /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of
5441 : OP1. */
5442 2403 : if (!vectype_in)
5443 101 : vectype_in = vectype1_in;
5444 : }
5445 :
5446 : /* If op0 is an external or constant def, infer the vector type
5447 : from the scalar type. */
5448 150065 : if (!vectype_in)
5449 20024 : vectype_in = get_vectype_for_scalar_type (vinfo, rhs_type, slp_node);
5450 150065 : if (!cost_vec)
5451 22934 : gcc_assert (vectype_in);
5452 150065 : if (!vectype_in)
5453 : {
5454 252 : if (dump_enabled_p ())
5455 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5456 : "no vectype for scalar type %T\n", rhs_type);
5457 :
5458 252 : return false;
5459 : }
5460 :
5461 299626 : if (VECTOR_BOOLEAN_TYPE_P (vectype_out)
5462 149813 : != VECTOR_BOOLEAN_TYPE_P (vectype_in))
5463 : {
5464 229 : if (dump_enabled_p ())
5465 36 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5466 : "can't convert between boolean and non "
5467 : "boolean vectors %T\n", rhs_type);
5468 :
5469 229 : return false;
5470 : }
5471 :
5472 149584 : nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
5473 149584 : nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
5474 149584 : if (known_eq (nunits_out, nunits_in))
5475 70926 : if (widen_arith)
5476 : modifier = WIDEN;
5477 : else
5478 149584 : modifier = NONE;
5479 78658 : else if (multiple_p (nunits_out, nunits_in))
5480 : modifier = NARROW_DST;
5481 : else
5482 : {
5483 43987 : gcc_checking_assert (multiple_p (nunits_in, nunits_out));
5484 : modifier = WIDEN;
5485 : }
5486 :
5487 149584 : bool found_mode = false;
5488 149584 : scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type);
5489 149584 : scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type);
5490 149584 : opt_scalar_mode rhs_mode_iter;
5491 149584 : auto_vec<std::pair<tree, tree_code>, 2> converts;
5492 149584 : bool evenodd_ok = false;
5493 :
5494 : /* Supportable by target? */
5495 149584 : switch (modifier)
5496 : {
5497 70683 : case NONE:
5498 70683 : if (code != FIX_TRUNC_EXPR
5499 69808 : && code != FLOAT_EXPR
5500 134307 : && !CONVERT_EXPR_CODE_P (code))
5501 : return false;
5502 70683 : gcc_assert (code.is_tree_code ());
5503 70683 : if (supportable_indirect_convert_operation (code,
5504 : vectype_out, vectype_in,
5505 : converts, op0, slp_op0))
5506 : {
5507 16045 : gcc_assert (converts.length () <= 2);
5508 16045 : if (converts.length () == 1)
5509 15971 : code1 = converts[0].second;
5510 : else
5511 : {
5512 74 : cvt_type = NULL_TREE;
5513 74 : multi_step_cvt = converts.length () - 1;
5514 74 : codecvt1 = converts[0].second;
5515 74 : code1 = converts[1].second;
5516 74 : interm_types.safe_push (converts[0].first);
5517 : }
5518 : break;
5519 : }
5520 :
5521 : /* FALLTHRU */
5522 54638 : unsupported:
5523 61076 : if (dump_enabled_p ())
5524 5888 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5525 : "conversion not supported by target.\n");
5526 : return false;
5527 :
5528 44230 : case WIDEN:
5529 44230 : if (known_eq (nunits_in, nunits_out))
5530 : {
5531 486 : if (!(code.is_tree_code ()
5532 243 : && supportable_half_widening_operation ((tree_code) code,
5533 : vectype_out, vectype_in,
5534 : &tc1)))
5535 74 : goto unsupported;
5536 169 : code1 = tc1;
5537 169 : gcc_assert (!(multi_step_cvt && op_type == binary_op));
5538 : break;
5539 : }
5540 : /* Elements in a vector can only be reordered if used in a reduction
5541 : operation only. */
5542 43987 : if (code == WIDEN_MULT_EXPR
5543 2160 : && loop_vinfo
5544 2111 : && !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info)
5545 : /* For a SLP reduction we cannot swizzle lanes, detecting a
5546 : reduction chain isn't possible here. */
5547 46076 : && SLP_TREE_LANES (slp_node) == 1)
5548 : {
5549 : /* ??? There is no way to look for SLP uses, so work on
5550 : the stmt and what the stmt-based cycle detection gives us. */
5551 1991 : tree lhs = gimple_get_lhs (vect_orig_stmt (stmt_info)->stmt);
5552 1991 : stmt_vec_info use_stmt_info
5553 1991 : = lhs ? loop_vinfo->lookup_single_use (lhs) : NULL;
5554 1991 : if (use_stmt_info
5555 1843 : && STMT_VINFO_REDUC_DEF (use_stmt_info))
5556 43987 : evenodd_ok = true;
5557 : }
5558 43987 : if (supportable_widening_operation (code, vectype_out, vectype_in,
5559 : evenodd_ok, &code1,
5560 : &code2, &multi_step_cvt,
5561 : &interm_types))
5562 : {
5563 : /* Binary widening operation can only be supported directly by the
5564 : architecture. */
5565 42285 : gcc_assert (!(multi_step_cvt && op_type == binary_op));
5566 : break;
5567 : }
5568 :
5569 1702 : if (code != FLOAT_EXPR
5570 2026 : || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode))
5571 1540 : goto unsupported;
5572 :
5573 162 : fltsz = GET_MODE_SIZE (lhs_mode);
5574 237 : FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode)
5575 : {
5576 237 : rhs_mode = rhs_mode_iter.require ();
5577 474 : if (GET_MODE_SIZE (rhs_mode) > fltsz)
5578 : break;
5579 :
5580 237 : cvt_type
5581 237 : = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0);
5582 237 : cvt_type = get_same_sized_vectype (cvt_type, vectype_in);
5583 237 : if (cvt_type == NULL_TREE)
5584 0 : goto unsupported;
5585 :
5586 474 : if (GET_MODE_SIZE (rhs_mode) == fltsz)
5587 : {
5588 57 : tc1 = ERROR_MARK;
5589 57 : gcc_assert (code.is_tree_code ());
5590 57 : if (!supportable_convert_operation ((tree_code) code, vectype_out,
5591 : cvt_type, &tc1))
5592 22 : goto unsupported;
5593 35 : codecvt1 = tc1;
5594 : }
5595 180 : else if (!supportable_widening_operation (code, vectype_out,
5596 : cvt_type, evenodd_ok,
5597 : &codecvt1,
5598 : &codecvt2, &multi_step_cvt,
5599 : &interm_types))
5600 75 : continue;
5601 : else
5602 105 : gcc_assert (multi_step_cvt == 0);
5603 :
5604 140 : if (supportable_widening_operation (NOP_EXPR, cvt_type,
5605 : vectype_in, evenodd_ok, &code1,
5606 : &code2, &multi_step_cvt,
5607 : &interm_types))
5608 : {
5609 : found_mode = true;
5610 : break;
5611 : }
5612 : }
5613 :
5614 140 : if (!found_mode)
5615 0 : goto unsupported;
5616 :
5617 280 : if (GET_MODE_SIZE (rhs_mode) == fltsz)
5618 35 : codecvt2 = ERROR_MARK;
5619 : else
5620 : {
5621 105 : multi_step_cvt++;
5622 105 : interm_types.safe_push (cvt_type);
5623 105 : cvt_type = NULL_TREE;
5624 : }
5625 : break;
5626 :
5627 34671 : case NARROW_DST:
5628 34671 : gcc_assert (op_type == unary_op);
5629 34671 : if (supportable_narrowing_operation (code, vectype_out, vectype_in,
5630 : &code1, &multi_step_cvt,
5631 : &interm_types))
5632 : break;
5633 :
5634 14664 : if (GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode))
5635 966 : goto unsupported;
5636 :
5637 3922 : if (code == FIX_TRUNC_EXPR)
5638 : {
5639 82 : cvt_type
5640 82 : = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0);
5641 82 : cvt_type = get_same_sized_vectype (cvt_type, vectype_in);
5642 82 : if (cvt_type == NULL_TREE)
5643 0 : goto unsupported;
5644 82 : if (supportable_convert_operation ((tree_code) code, cvt_type, vectype_in,
5645 : &tc1))
5646 80 : codecvt1 = tc1;
5647 : else
5648 2 : goto unsupported;
5649 80 : if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type,
5650 : &code1, &multi_step_cvt,
5651 : &interm_types))
5652 : break;
5653 : }
5654 : /* If op0 can be represented with low precision integer,
5655 : truncate it to cvt_type and the do FLOAT_EXPR. */
5656 3840 : else if (code == FLOAT_EXPR)
5657 : {
5658 96 : if (cost_vec)
5659 : {
5660 91 : wide_int op_min_value, op_max_value;
5661 91 : tree def;
5662 :
5663 : /* ??? Merge ranges in case of more than one lane. */
5664 91 : if (SLP_TREE_LANES (slp_op0) != 1
5665 89 : || !(def = vect_get_slp_scalar_def (slp_op0, 0))
5666 180 : || !vect_get_range_info (def, &op_min_value, &op_max_value))
5667 86 : goto unsupported;
5668 :
5669 5 : if ((wi::min_precision (op_max_value, SIGNED)
5670 5 : > GET_MODE_BITSIZE (lhs_mode))
5671 5 : || (wi::min_precision (op_min_value, SIGNED)
5672 5 : > GET_MODE_BITSIZE (lhs_mode)))
5673 0 : goto unsupported;
5674 91 : }
5675 :
5676 10 : cvt_type
5677 10 : = build_nonstandard_integer_type (GET_MODE_BITSIZE (lhs_mode), 0);
5678 10 : cvt_type = get_same_sized_vectype (cvt_type, vectype_out);
5679 10 : if (cvt_type == NULL_TREE)
5680 0 : goto unsupported;
5681 10 : if (!supportable_narrowing_operation (NOP_EXPR, cvt_type, vectype_in,
5682 : &code1, &multi_step_cvt,
5683 : &interm_types))
5684 0 : goto unsupported;
5685 10 : if (supportable_convert_operation ((tree_code) code, vectype_out,
5686 : cvt_type, &tc1))
5687 : {
5688 10 : codecvt1 = tc1;
5689 10 : modifier = NARROW_SRC;
5690 10 : break;
5691 : }
5692 : }
5693 :
5694 3748 : goto unsupported;
5695 :
5696 : default:
5697 : gcc_unreachable ();
5698 : }
5699 :
5700 88508 : if (modifier == WIDEN
5701 88508 : && loop_vinfo
5702 41447 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
5703 105672 : && (code1 == VEC_WIDEN_MULT_EVEN_EXPR
5704 17148 : || widening_evenodd_fn_p (code1)))
5705 : {
5706 16 : if (dump_enabled_p ())
5707 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5708 : "can't use a fully-masked loop because"
5709 : " widening operation on even/odd elements"
5710 : " mixes up lanes.\n");
5711 16 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
5712 : }
5713 :
5714 88508 : if (cost_vec) /* transformation not required. */
5715 : {
5716 65574 : if (!vect_maybe_update_slp_op_vectype (slp_op0, vectype_in)
5717 65574 : || !vect_maybe_update_slp_op_vectype (slp_op1, vectype_in))
5718 : {
5719 0 : if (dump_enabled_p ())
5720 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
5721 : "incompatible vector types for invariants\n");
5722 0 : return false;
5723 : }
5724 65574 : DUMP_VECT_SCOPE ("vectorizable_conversion");
5725 65574 : unsigned int nvectors = vect_get_num_copies (vinfo, slp_node);
5726 65574 : if (modifier == NONE)
5727 : {
5728 12067 : SLP_TREE_TYPE (slp_node) = type_conversion_vec_info_type;
5729 12067 : vect_model_simple_cost (vinfo, (1 + multi_step_cvt),
5730 : slp_node, cost_vec);
5731 : }
5732 53507 : else if (modifier == NARROW_SRC || modifier == NARROW_DST)
5733 : {
5734 20862 : SLP_TREE_TYPE (slp_node) = type_demotion_vec_info_type;
5735 : /* The final packing step produces one vector result per copy. */
5736 20862 : vect_model_promotion_demotion_cost (slp_node, nvectors,
5737 : multi_step_cvt, cost_vec,
5738 : widen_arith);
5739 : }
5740 : else
5741 : {
5742 32645 : SLP_TREE_TYPE (slp_node) = type_promotion_vec_info_type;
5743 : /* The initial unpacking step produces two vector results
5744 : per copy. MULTI_STEP_CVT is 0 for a single conversion,
5745 : so >> MULTI_STEP_CVT divides by 2^(number of steps - 1). */
5746 32645 : vect_model_promotion_demotion_cost (slp_node,
5747 : nvectors >> multi_step_cvt,
5748 : multi_step_cvt, cost_vec,
5749 : widen_arith);
5750 : }
5751 65574 : interm_types.release ();
5752 65574 : return true;
5753 65574 : }
5754 :
5755 : /* Transform. */
5756 22934 : if (dump_enabled_p ())
5757 4266 : dump_printf_loc (MSG_NOTE, vect_location, "transform conversion.\n");
5758 :
5759 22934 : if (op_type == binary_op)
5760 : {
5761 513 : if (CONSTANT_CLASS_P (op0))
5762 0 : op0 = fold_convert (TREE_TYPE (op1), op0);
5763 513 : else if (CONSTANT_CLASS_P (op1))
5764 237 : op1 = fold_convert (TREE_TYPE (op0), op1);
5765 : }
5766 :
5767 : /* In case of multi-step conversion, we first generate conversion operations
5768 : to the intermediate types, and then from that types to the final one.
5769 : We create vector destinations for the intermediate type (TYPES) received
5770 : from supportable_*_operation, and store them in the correct order
5771 : for future use in vect_create_vectorized_*_stmts (). */
5772 22934 : auto_vec<tree> vec_dsts (multi_step_cvt + 1);
5773 22934 : bool widen_or_narrow_float_p
5774 22934 : = cvt_type && (modifier == WIDEN || modifier == NARROW_SRC);
5775 22934 : vec_dest = vect_create_destination_var (scalar_dest,
5776 : widen_or_narrow_float_p
5777 : ? cvt_type : vectype_out);
5778 22934 : vec_dsts.quick_push (vec_dest);
5779 :
5780 22934 : if (multi_step_cvt)
5781 : {
5782 9344 : for (i = interm_types.length () - 1;
5783 9344 : interm_types.iterate (i, &intermediate_type); i--)
5784 : {
5785 4918 : vec_dest = vect_create_destination_var (scalar_dest,
5786 : intermediate_type);
5787 4918 : vec_dsts.quick_push (vec_dest);
5788 : }
5789 : }
5790 :
5791 22934 : if (cvt_type)
5792 73 : vec_dest = vect_create_destination_var (scalar_dest,
5793 : widen_or_narrow_float_p
5794 : ? vectype_out : cvt_type);
5795 :
5796 22934 : switch (modifier)
5797 : {
5798 3978 : case NONE:
5799 3978 : vect_get_vec_defs (vinfo, slp_node, op0, &vec_oprnds0);
5800 : /* vec_dest is intermediate type operand when multi_step_cvt. */
5801 3978 : if (multi_step_cvt)
5802 : {
5803 21 : cvt_op = vec_dest;
5804 21 : vec_dest = vec_dsts[0];
5805 : }
5806 :
5807 8332 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0)
5808 : {
5809 : /* Arguments are ready, create the new vector stmt. */
5810 4354 : gimple* new_stmt;
5811 4354 : if (multi_step_cvt)
5812 : {
5813 21 : gcc_assert (multi_step_cvt == 1);
5814 21 : new_stmt = vect_gimple_build (cvt_op, codecvt1, vop0);
5815 21 : new_temp = make_ssa_name (cvt_op, new_stmt);
5816 21 : gimple_assign_set_lhs (new_stmt, new_temp);
5817 21 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5818 21 : vop0 = new_temp;
5819 : }
5820 4354 : new_stmt = vect_gimple_build (vec_dest, code1, vop0);
5821 4354 : new_temp = make_ssa_name (vec_dest, new_stmt);
5822 4354 : gimple_set_lhs (new_stmt, new_temp);
5823 4354 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5824 :
5825 4354 : slp_node->push_vec_def (new_stmt);
5826 : }
5827 : break;
5828 :
5829 9949 : case WIDEN:
5830 : /* In case the vectorization factor (VF) is bigger than the number
5831 : of elements that we can fit in a vectype (nunits), we have to
5832 : generate more than one vector stmt - i.e - we need to "unroll"
5833 : the vector stmt by a factor VF/nunits. */
5834 9949 : vect_get_vec_defs (vinfo, slp_node, op0, &vec_oprnds0,
5835 9949 : code == WIDEN_LSHIFT_EXPR ? NULL_TREE : op1,
5836 : &vec_oprnds1);
5837 9949 : if (code == WIDEN_LSHIFT_EXPR)
5838 : {
5839 0 : int oprnds_size = vec_oprnds0.length ();
5840 0 : vec_oprnds1.create (oprnds_size);
5841 0 : for (i = 0; i < oprnds_size; ++i)
5842 0 : vec_oprnds1.quick_push (op1);
5843 : }
5844 : /* Arguments are ready. Create the new vector stmts. */
5845 21824 : for (i = multi_step_cvt; i >= 0; i--)
5846 : {
5847 11875 : tree this_dest = vec_dsts[i];
5848 11875 : code_helper c1 = code1, c2 = code2;
5849 11875 : if (i == 0 && codecvt2 != ERROR_MARK)
5850 : {
5851 48 : c1 = codecvt1;
5852 48 : c2 = codecvt2;
5853 : }
5854 11875 : if (known_eq (nunits_out, nunits_in))
5855 14 : vect_create_half_widening_stmts (vinfo, &vec_oprnds0, &vec_oprnds1,
5856 : stmt_info, this_dest, gsi, c1,
5857 : op_type);
5858 : else
5859 11861 : vect_create_vectorized_promotion_stmts (vinfo, &vec_oprnds0,
5860 : &vec_oprnds1, stmt_info,
5861 : this_dest, gsi,
5862 : c1, c2, op_type);
5863 : }
5864 :
5865 38155 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0)
5866 : {
5867 28206 : gimple *new_stmt;
5868 28206 : if (cvt_type)
5869 : {
5870 120 : new_temp = make_ssa_name (vec_dest);
5871 120 : new_stmt = vect_gimple_build (new_temp, codecvt1, vop0);
5872 120 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5873 : }
5874 : else
5875 28086 : new_stmt = SSA_NAME_DEF_STMT (vop0);
5876 :
5877 28206 : slp_node->push_vec_def (new_stmt);
5878 : }
5879 : break;
5880 :
5881 9007 : case NARROW_SRC:
5882 9007 : case NARROW_DST:
5883 : /* In case the vectorization factor (VF) is bigger than the number
5884 : of elements that we can fit in a vectype (nunits), we have to
5885 : generate more than one vector stmt - i.e - we need to "unroll"
5886 : the vector stmt by a factor VF/nunits. */
5887 9007 : vect_get_vec_defs (vinfo, slp_node, op0, &vec_oprnds0);
5888 : /* Arguments are ready. Create the new vector stmts. */
5889 9007 : if (cvt_type && modifier == NARROW_DST)
5890 153 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0)
5891 : {
5892 124 : new_temp = make_ssa_name (vec_dest);
5893 124 : gimple *new_stmt = vect_gimple_build (new_temp, codecvt1, vop0);
5894 124 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5895 124 : vec_oprnds0[i] = new_temp;
5896 : }
5897 :
5898 9007 : vect_create_vectorized_demotion_stmts (vinfo, &vec_oprnds0,
5899 : multi_step_cvt,
5900 : stmt_info, vec_dsts, gsi,
5901 : slp_node, code1,
5902 : modifier == NARROW_SRC);
5903 : /* After demoting op0 to cvt_type, convert it to dest. */
5904 9007 : if (cvt_type && code == FLOAT_EXPR)
5905 : {
5906 10 : for (unsigned int i = 0; i != vec_oprnds0.length() / 2; i++)
5907 : {
5908 : /* Arguments are ready, create the new vector stmt. */
5909 5 : gcc_assert (TREE_CODE_LENGTH ((tree_code) codecvt1) == unary_op);
5910 5 : gimple *new_stmt
5911 5 : = vect_gimple_build (vec_dest, codecvt1, vec_oprnds0[i]);
5912 5 : new_temp = make_ssa_name (vec_dest, new_stmt);
5913 5 : gimple_set_lhs (new_stmt, new_temp);
5914 5 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
5915 :
5916 : /* This is the last step of the conversion sequence. Store the
5917 : vectors in SLP_NODE or in vector info of the scalar statement
5918 : (or in STMT_VINFO_RELATED_STMT chain). */
5919 5 : slp_node->push_vec_def (new_stmt);
5920 : }
5921 : }
5922 : break;
5923 : }
5924 :
5925 22934 : vec_oprnds0.release ();
5926 22934 : vec_oprnds1.release ();
5927 22934 : interm_types.release ();
5928 :
5929 22934 : return true;
5930 149584 : }
5931 :
5932 : /* Return true if we can assume from the scalar form of STMT_INFO that
5933 : neither the scalar nor the vector forms will generate code. STMT_INFO
5934 : is known not to involve a data reference. */
5935 :
5936 : bool
5937 1013205 : vect_nop_conversion_p (stmt_vec_info stmt_info)
5938 : {
5939 1013205 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
5940 749129 : if (!stmt)
5941 : return false;
5942 :
5943 749129 : tree lhs = gimple_assign_lhs (stmt);
5944 749129 : tree_code code = gimple_assign_rhs_code (stmt);
5945 749129 : tree rhs = gimple_assign_rhs1 (stmt);
5946 :
5947 749129 : if (code == SSA_NAME || code == VIEW_CONVERT_EXPR)
5948 : return true;
5949 :
5950 746836 : if (CONVERT_EXPR_CODE_P (code))
5951 194066 : return tree_nop_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs));
5952 :
5953 : return false;
5954 : }
5955 :
5956 : /* Function vectorizable_assignment.
5957 :
5958 : Check if STMT_INFO performs an assignment (copy) that can be vectorized.
5959 : If COST_VEC is passed, calculate costs but don't change anything,
5960 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
5961 : it, and insert it at GSI.
5962 : Return true if STMT_INFO is vectorizable in this way. */
5963 :
5964 : static bool
5965 1855434 : vectorizable_assignment (vec_info *vinfo,
5966 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
5967 : slp_tree slp_node,
5968 : stmt_vector_for_cost *cost_vec)
5969 : {
5970 1855434 : tree vec_dest;
5971 1855434 : tree scalar_dest;
5972 1855434 : tree op;
5973 1855434 : tree new_temp;
5974 1855434 : enum vect_def_type dt[1] = {vect_unknown_def_type};
5975 1855434 : int i;
5976 1855434 : vec<tree> vec_oprnds = vNULL;
5977 1855434 : tree vop;
5978 1855434 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
5979 1855434 : enum tree_code code;
5980 1855434 : tree vectype_in;
5981 :
5982 1855434 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
5983 : return false;
5984 :
5985 1855434 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
5986 194522 : && cost_vec)
5987 : return false;
5988 :
5989 : /* Is vectorizable assignment? */
5990 3377085 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
5991 1593985 : if (!stmt)
5992 : return false;
5993 :
5994 1593985 : scalar_dest = gimple_assign_lhs (stmt);
5995 1593985 : if (TREE_CODE (scalar_dest) != SSA_NAME)
5996 : return false;
5997 :
5998 822011 : if (STMT_VINFO_DATA_REF (stmt_info))
5999 : return false;
6000 :
6001 349632 : code = gimple_assign_rhs_code (stmt);
6002 349632 : if (!(gimple_assign_single_p (stmt)
6003 348280 : || code == PAREN_EXPR
6004 347226 : || CONVERT_EXPR_CODE_P (code)))
6005 : return false;
6006 :
6007 82525 : tree vectype = SLP_TREE_VECTYPE (slp_node);
6008 82525 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
6009 :
6010 82525 : slp_tree slp_op;
6011 82525 : if (!vect_is_simple_use (vinfo, slp_node, 0, &op, &slp_op,
6012 : &dt[0], &vectype_in))
6013 : {
6014 0 : if (dump_enabled_p ())
6015 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6016 : "use not simple.\n");
6017 0 : return false;
6018 : }
6019 82525 : if (!vectype_in)
6020 17786 : vectype_in = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op), slp_node);
6021 :
6022 : /* We can handle VIEW_CONVERT conversions that do not change the number
6023 : of elements or the vector size or other conversions when the component
6024 : types are nop-convertible. */
6025 82525 : if (!vectype_in
6026 82253 : || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits)
6027 75511 : || (code == VIEW_CONVERT_EXPR
6028 2448 : && maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)),
6029 2448 : GET_MODE_SIZE (TYPE_MODE (vectype_in))))
6030 158036 : || (CONVERT_EXPR_CODE_P (code)
6031 73137 : && !tree_nop_conversion_p (TREE_TYPE (vectype),
6032 73137 : TREE_TYPE (vectype_in))))
6033 9923 : return false;
6034 :
6035 217728 : if (VECTOR_BOOLEAN_TYPE_P (vectype) != VECTOR_BOOLEAN_TYPE_P (vectype_in))
6036 : {
6037 2 : if (dump_enabled_p ())
6038 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6039 : "can't convert between boolean and non "
6040 0 : "boolean vectors %T\n", TREE_TYPE (op));
6041 :
6042 2 : return false;
6043 : }
6044 :
6045 : /* We do not handle bit-precision changes. */
6046 72600 : if ((CONVERT_EXPR_CODE_P (code)
6047 2374 : || code == VIEW_CONVERT_EXPR)
6048 71450 : && ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
6049 70201 : && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)))
6050 71144 : || (INTEGRAL_TYPE_P (TREE_TYPE (op))
6051 66553 : && !type_has_mode_precision_p (TREE_TYPE (op))))
6052 : /* But a conversion that does not change the bit-pattern is ok. */
6053 73310 : && !(INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
6054 710 : && INTEGRAL_TYPE_P (TREE_TYPE (op))
6055 710 : && (((TYPE_PRECISION (TREE_TYPE (scalar_dest))
6056 710 : > TYPE_PRECISION (TREE_TYPE (op)))
6057 404 : && TYPE_UNSIGNED (TREE_TYPE (op)))
6058 322 : || (TYPE_PRECISION (TREE_TYPE (scalar_dest))
6059 322 : == TYPE_PRECISION (TREE_TYPE (op))))))
6060 : {
6061 266 : if (dump_enabled_p ())
6062 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6063 : "type conversion to/from bit-precision "
6064 : "unsupported.\n");
6065 266 : return false;
6066 : }
6067 :
6068 72334 : if (cost_vec) /* transformation not required. */
6069 : {
6070 57314 : if (!vect_maybe_update_slp_op_vectype (slp_op, vectype_in))
6071 : {
6072 0 : if (dump_enabled_p ())
6073 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6074 : "incompatible vector types for invariants\n");
6075 0 : return false;
6076 : }
6077 57314 : SLP_TREE_TYPE (slp_node) = assignment_vec_info_type;
6078 57314 : DUMP_VECT_SCOPE ("vectorizable_assignment");
6079 57314 : if (!vect_nop_conversion_p (stmt_info))
6080 827 : vect_model_simple_cost (vinfo, 1, slp_node, cost_vec);
6081 57314 : return true;
6082 : }
6083 :
6084 : /* Transform. */
6085 15020 : if (dump_enabled_p ())
6086 3591 : dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n");
6087 :
6088 : /* Handle def. */
6089 15020 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
6090 :
6091 : /* Handle use. */
6092 15020 : vect_get_vec_defs (vinfo, slp_node, op, &vec_oprnds);
6093 :
6094 : /* Arguments are ready. create the new vector stmt. */
6095 34150 : FOR_EACH_VEC_ELT (vec_oprnds, i, vop)
6096 : {
6097 19130 : if (CONVERT_EXPR_CODE_P (code)
6098 683 : || code == VIEW_CONVERT_EXPR)
6099 18581 : vop = build1 (VIEW_CONVERT_EXPR, vectype, vop);
6100 19130 : gassign *new_stmt = gimple_build_assign (vec_dest, vop);
6101 19130 : new_temp = make_ssa_name (vec_dest, new_stmt);
6102 19130 : gimple_assign_set_lhs (new_stmt, new_temp);
6103 19130 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
6104 19130 : slp_node->push_vec_def (new_stmt);
6105 : }
6106 :
6107 15020 : vec_oprnds.release ();
6108 15020 : return true;
6109 : }
6110 :
6111 :
6112 : /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE
6113 : either as shift by a scalar or by a vector. */
6114 :
6115 : bool
6116 279634 : vect_supportable_shift (vec_info *vinfo, enum tree_code code, tree scalar_type)
6117 : {
6118 279634 : optab optab;
6119 279634 : tree vectype;
6120 :
6121 279634 : vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
6122 279634 : if (!vectype)
6123 : return false;
6124 :
6125 279634 : optab = optab_for_tree_code (code, vectype, optab_scalar);
6126 279634 : if (optab && can_implement_p (optab, TYPE_MODE (vectype)))
6127 : return true;
6128 :
6129 246191 : optab = optab_for_tree_code (code, vectype, optab_vector);
6130 246191 : if (optab && can_implement_p (optab, TYPE_MODE (vectype)))
6131 : return true;
6132 :
6133 : return false;
6134 : }
6135 :
6136 :
6137 : /* Function vectorizable_shift.
6138 :
6139 : Check if STMT_INFO performs a shift operation that can be vectorized.
6140 : If COST_VEC is passed, calculate costs but don't change anything,
6141 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
6142 : it, and insert it at GSI.
6143 : Return true if STMT_INFO is vectorizable in this way. */
6144 :
6145 : static bool
6146 651074 : vectorizable_shift (vec_info *vinfo,
6147 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
6148 : slp_tree slp_node,
6149 : stmt_vector_for_cost *cost_vec)
6150 : {
6151 651074 : tree vec_dest;
6152 651074 : tree scalar_dest;
6153 651074 : tree op0, op1 = NULL;
6154 651074 : tree vec_oprnd1 = NULL_TREE;
6155 651074 : tree vectype;
6156 651074 : enum tree_code code;
6157 651074 : machine_mode vec_mode;
6158 651074 : tree new_temp;
6159 651074 : optab optab;
6160 651074 : int icode;
6161 651074 : machine_mode optab_op2_mode;
6162 651074 : enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
6163 651074 : poly_uint64 nunits_in;
6164 651074 : poly_uint64 nunits_out;
6165 651074 : tree vectype_out;
6166 651074 : tree op1_vectype;
6167 651074 : int i;
6168 651074 : vec<tree> vec_oprnds0 = vNULL;
6169 651074 : vec<tree> vec_oprnds1 = vNULL;
6170 651074 : tree vop0, vop1;
6171 651074 : unsigned int k;
6172 651074 : bool scalar_shift_arg = true;
6173 651074 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
6174 651074 : bool incompatible_op1_vectype_p = false;
6175 :
6176 651074 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
6177 : return false;
6178 :
6179 651074 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
6180 194522 : && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle
6181 193095 : && cost_vec)
6182 : return false;
6183 :
6184 : /* Is STMT a vectorizable binary/unary operation? */
6185 989190 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
6186 391398 : if (!stmt)
6187 : return false;
6188 :
6189 391398 : if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
6190 : return false;
6191 :
6192 390916 : code = gimple_assign_rhs_code (stmt);
6193 :
6194 390916 : if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR
6195 : || code == RROTATE_EXPR))
6196 : return false;
6197 :
6198 58608 : scalar_dest = gimple_assign_lhs (stmt);
6199 58608 : vectype_out = SLP_TREE_VECTYPE (slp_node);
6200 58608 : if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)))
6201 : {
6202 0 : if (dump_enabled_p ())
6203 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6204 : "bit-precision shifts not supported.\n");
6205 0 : return false;
6206 : }
6207 :
6208 58608 : slp_tree slp_op0;
6209 58608 : if (!vect_is_simple_use (vinfo, slp_node,
6210 : 0, &op0, &slp_op0, &dt[0], &vectype))
6211 : {
6212 0 : if (dump_enabled_p ())
6213 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6214 : "use not simple.\n");
6215 0 : return false;
6216 : }
6217 : /* If op0 is an external or constant def, infer the vector type
6218 : from the scalar type. */
6219 58608 : if (!vectype)
6220 12183 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op0), slp_node);
6221 58608 : if (!cost_vec)
6222 7821 : gcc_assert (vectype);
6223 58608 : if (!vectype)
6224 : {
6225 0 : if (dump_enabled_p ())
6226 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6227 : "no vectype for scalar type\n");
6228 0 : return false;
6229 : }
6230 :
6231 58608 : nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
6232 58608 : nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
6233 58608 : if (maybe_ne (nunits_out, nunits_in))
6234 : return false;
6235 :
6236 58608 : stmt_vec_info op1_def_stmt_info;
6237 58608 : slp_tree slp_op1;
6238 58608 : if (!vect_is_simple_use (vinfo, slp_node, 1, &op1, &slp_op1,
6239 : &dt[1], &op1_vectype, &op1_def_stmt_info))
6240 : {
6241 0 : if (dump_enabled_p ())
6242 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6243 : "use not simple.\n");
6244 0 : return false;
6245 : }
6246 :
6247 : /* Determine whether the shift amount is a vector, or scalar. If the
6248 : shift/rotate amount is a vector, use the vector/vector shift optabs. */
6249 :
6250 58608 : if ((dt[1] == vect_internal_def
6251 58608 : || dt[1] == vect_induction_def
6252 45699 : || dt[1] == vect_nested_cycle)
6253 12927 : && SLP_TREE_LANES (slp_node) == 1)
6254 : scalar_shift_arg = false;
6255 45736 : else if (dt[1] == vect_constant_def
6256 : || dt[1] == vect_external_def
6257 45736 : || dt[1] == vect_internal_def)
6258 : {
6259 : /* In SLP, need to check whether the shift count is the same,
6260 : in loops if it is a constant or invariant, it is always
6261 : a scalar shift. */
6262 45730 : vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node);
6263 45730 : stmt_vec_info slpstmt_info;
6264 :
6265 123714 : FOR_EACH_VEC_ELT (stmts, k, slpstmt_info)
6266 77984 : if (slpstmt_info)
6267 : {
6268 77984 : gassign *slpstmt = as_a <gassign *> (slpstmt_info->stmt);
6269 155968 : if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0))
6270 77984 : scalar_shift_arg = false;
6271 : }
6272 :
6273 : /* For internal SLP defs we have to make sure we see scalar stmts
6274 : for all vector elements.
6275 : ??? For different vectors we could resort to a different
6276 : scalar shift operand but code-generation below simply always
6277 : takes the first. */
6278 45730 : if (dt[1] == vect_internal_def
6279 45779 : && maybe_ne (nunits_out * vect_get_num_copies (vinfo, slp_node),
6280 49 : stmts.length ()))
6281 : scalar_shift_arg = false;
6282 :
6283 : /* If the shift amount is computed by a pattern stmt we cannot
6284 : use the scalar amount directly thus give up and use a vector
6285 : shift. */
6286 45730 : if (op1_def_stmt_info && is_pattern_stmt_p (op1_def_stmt_info))
6287 : scalar_shift_arg = false;
6288 : }
6289 : else
6290 : {
6291 6 : if (dump_enabled_p ())
6292 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6293 : "operand mode requires invariant argument.\n");
6294 6 : return false;
6295 : }
6296 :
6297 : /* Vector shifted by vector. */
6298 58640 : bool was_scalar_shift_arg = scalar_shift_arg;
6299 45721 : if (!scalar_shift_arg)
6300 : {
6301 12919 : optab = optab_for_tree_code (code, vectype, optab_vector);
6302 12919 : if (dump_enabled_p ())
6303 1196 : dump_printf_loc (MSG_NOTE, vect_location,
6304 : "vector/vector shift/rotate found.\n");
6305 :
6306 12919 : if (!op1_vectype)
6307 15 : op1_vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op1),
6308 : slp_op1);
6309 12919 : incompatible_op1_vectype_p
6310 25838 : = (op1_vectype == NULL_TREE
6311 12919 : || maybe_ne (TYPE_VECTOR_SUBPARTS (op1_vectype),
6312 12919 : TYPE_VECTOR_SUBPARTS (vectype))
6313 25836 : || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype));
6314 12912 : if (incompatible_op1_vectype_p
6315 7 : && (SLP_TREE_DEF_TYPE (slp_op1) != vect_constant_def
6316 1 : || slp_op1->refcnt != 1))
6317 : {
6318 6 : if (dump_enabled_p ())
6319 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6320 : "unusable type for last operand in"
6321 : " vector/vector shift/rotate.\n");
6322 6 : return false;
6323 : }
6324 : }
6325 : /* See if the machine has a vector shifted by scalar insn and if not
6326 : then see if it has a vector shifted by vector insn. */
6327 : else
6328 : {
6329 45683 : optab = optab_for_tree_code (code, vectype, optab_scalar);
6330 45683 : if (optab
6331 45683 : && can_implement_p (optab, TYPE_MODE (vectype)))
6332 : {
6333 45683 : if (dump_enabled_p ())
6334 4914 : dump_printf_loc (MSG_NOTE, vect_location,
6335 : "vector/scalar shift/rotate found.\n");
6336 : }
6337 : else
6338 : {
6339 0 : optab = optab_for_tree_code (code, vectype, optab_vector);
6340 0 : if (optab
6341 0 : && can_implement_p (optab, TYPE_MODE (vectype)))
6342 : {
6343 0 : scalar_shift_arg = false;
6344 :
6345 0 : if (dump_enabled_p ())
6346 0 : dump_printf_loc (MSG_NOTE, vect_location,
6347 : "vector/vector shift/rotate found.\n");
6348 :
6349 0 : if (!op1_vectype)
6350 0 : op1_vectype = get_vectype_for_scalar_type (vinfo,
6351 0 : TREE_TYPE (op1),
6352 : slp_op1);
6353 :
6354 : /* Unlike the other binary operators, shifts/rotates have
6355 : the rhs being int, instead of the same type as the lhs,
6356 : so make sure the scalar is the right type if we are
6357 : dealing with vectors of long long/long/short/char. */
6358 0 : incompatible_op1_vectype_p
6359 0 : = (!op1_vectype
6360 0 : || !tree_nop_conversion_p (TREE_TYPE (vectype),
6361 0 : TREE_TYPE (op1)));
6362 0 : if (incompatible_op1_vectype_p
6363 0 : && dt[1] == vect_internal_def)
6364 : {
6365 0 : if (dump_enabled_p ())
6366 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6367 : "unusable type for last operand in"
6368 : " vector/vector shift/rotate.\n");
6369 0 : return false;
6370 : }
6371 : }
6372 : }
6373 : }
6374 :
6375 : /* Supportable by target? */
6376 58596 : if (!optab)
6377 : {
6378 0 : if (dump_enabled_p ())
6379 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6380 : "no shift optab for %s and %T.\n",
6381 : get_tree_code_name (code), vectype);
6382 0 : return false;
6383 : }
6384 58596 : vec_mode = TYPE_MODE (vectype);
6385 58596 : icode = (int) optab_handler (optab, vec_mode);
6386 58596 : if (icode == CODE_FOR_nothing)
6387 : {
6388 5314 : if (dump_enabled_p ())
6389 886 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6390 : "shift op not supported by target.\n");
6391 5314 : return false;
6392 : }
6393 : /* vector lowering cannot optimize vector shifts using word arithmetic. */
6394 53282 : if (vect_emulated_vector_p (vectype))
6395 : return false;
6396 :
6397 53282 : if (cost_vec) /* transformation not required. */
6398 : {
6399 45461 : if (!vect_maybe_update_slp_op_vectype (slp_op0, vectype)
6400 45461 : || ((!scalar_shift_arg || dt[1] == vect_internal_def)
6401 5462 : && (!incompatible_op1_vectype_p
6402 1 : || dt[1] == vect_constant_def)
6403 5462 : && !vect_maybe_update_slp_op_vectype
6404 5462 : (slp_op1,
6405 : incompatible_op1_vectype_p ? vectype : op1_vectype)))
6406 : {
6407 0 : if (dump_enabled_p ())
6408 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6409 : "incompatible vector types for invariants\n");
6410 0 : return false;
6411 : }
6412 : /* Now adjust the constant shift amount in place. */
6413 45461 : if (incompatible_op1_vectype_p
6414 1 : && dt[1] == vect_constant_def)
6415 4 : for (unsigned i = 0;
6416 5 : i < SLP_TREE_SCALAR_OPS (slp_op1).length (); ++i)
6417 : {
6418 4 : SLP_TREE_SCALAR_OPS (slp_op1)[i]
6419 4 : = fold_convert (TREE_TYPE (vectype),
6420 : SLP_TREE_SCALAR_OPS (slp_op1)[i]);
6421 4 : gcc_assert ((TREE_CODE (SLP_TREE_SCALAR_OPS (slp_op1)[i])
6422 : == INTEGER_CST));
6423 : }
6424 45461 : SLP_TREE_TYPE (slp_node) = shift_vec_info_type;
6425 45461 : DUMP_VECT_SCOPE ("vectorizable_shift");
6426 45461 : vect_model_simple_cost (vinfo, 1, slp_node, cost_vec);
6427 45461 : return true;
6428 : }
6429 :
6430 : /* Transform. */
6431 :
6432 7821 : if (dump_enabled_p ())
6433 2014 : dump_printf_loc (MSG_NOTE, vect_location,
6434 : "transform binary/unary operation.\n");
6435 :
6436 : /* Handle def. */
6437 7821 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
6438 :
6439 7821 : unsigned nvectors = vect_get_num_copies (vinfo, slp_node);
6440 7821 : if (scalar_shift_arg && dt[1] != vect_internal_def)
6441 : {
6442 : /* Vector shl and shr insn patterns can be defined with scalar
6443 : operand 2 (shift operand). In this case, use constant or loop
6444 : invariant op1 directly, without extending it to vector mode
6445 : first. */
6446 5664 : optab_op2_mode = insn_data[icode].operand[2].mode;
6447 5664 : if (!VECTOR_MODE_P (optab_op2_mode))
6448 : {
6449 5664 : if (dump_enabled_p ())
6450 1899 : dump_printf_loc (MSG_NOTE, vect_location,
6451 : "operand 1 using scalar mode.\n");
6452 5664 : vec_oprnd1 = op1;
6453 5664 : vec_oprnds1.create (nvectors);
6454 5664 : vec_oprnds1.quick_push (vec_oprnd1);
6455 : /* Store vec_oprnd1 for every vector stmt to be created.
6456 : We check during the analysis that all the shift arguments
6457 : are the same.
6458 : TODO: Allow different constants for different vector
6459 : stmts generated for an SLP instance. */
6460 13355 : for (k = 0; k < nvectors - 1; k++)
6461 2027 : vec_oprnds1.quick_push (vec_oprnd1);
6462 : }
6463 : }
6464 2157 : else if (!scalar_shift_arg && incompatible_op1_vectype_p)
6465 : {
6466 0 : if (was_scalar_shift_arg)
6467 : {
6468 : /* If the argument was the same in all lanes create the
6469 : correctly typed vector shift amount directly. Note
6470 : we made SLP scheduling think we use the original scalars,
6471 : so place the compensation code next to the shift which
6472 : is conservative. See PR119640 where it otherwise breaks. */
6473 0 : op1 = fold_convert (TREE_TYPE (vectype), op1);
6474 0 : op1 = vect_init_vector (vinfo, stmt_info, op1, TREE_TYPE (vectype),
6475 : gsi);
6476 0 : vec_oprnd1 = vect_init_vector (vinfo, stmt_info, op1, vectype,
6477 : gsi);
6478 0 : vec_oprnds1.create (nvectors);
6479 0 : for (k = 0; k < nvectors; k++)
6480 0 : vec_oprnds1.quick_push (vec_oprnd1);
6481 : }
6482 0 : else if (dt[1] == vect_constant_def)
6483 : /* The constant shift amount has been adjusted in place. */
6484 : ;
6485 : else
6486 0 : gcc_assert (TYPE_MODE (op1_vectype) == TYPE_MODE (vectype));
6487 : }
6488 :
6489 : /* vec_oprnd1 is available if operand 1 should be of a scalar-type
6490 : (a special case for certain kind of vector shifts); otherwise,
6491 : operand 1 should be of a vector type (the usual case). */
6492 2157 : vect_get_vec_defs (vinfo, slp_node,
6493 : op0, &vec_oprnds0,
6494 7821 : vec_oprnd1 ? NULL_TREE : op1, &vec_oprnds1);
6495 :
6496 : /* Arguments are ready. Create the new vector stmt. */
6497 21171 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0)
6498 : {
6499 : /* For internal defs where we need to use a scalar shift arg
6500 : extract the first lane. */
6501 13350 : if (scalar_shift_arg && dt[1] == vect_internal_def)
6502 : {
6503 10 : vop1 = vec_oprnds1[0];
6504 10 : new_temp = make_ssa_name (TREE_TYPE (TREE_TYPE (vop1)));
6505 10 : gassign *new_stmt
6506 10 : = gimple_build_assign (new_temp,
6507 10 : build3 (BIT_FIELD_REF, TREE_TYPE (new_temp),
6508 : vop1,
6509 10 : TYPE_SIZE (TREE_TYPE (new_temp)),
6510 : bitsize_zero_node));
6511 10 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
6512 10 : vop1 = new_temp;
6513 10 : }
6514 : else
6515 13340 : vop1 = vec_oprnds1[i];
6516 13350 : gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1);
6517 13350 : new_temp = make_ssa_name (vec_dest, new_stmt);
6518 13350 : gimple_assign_set_lhs (new_stmt, new_temp);
6519 13350 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
6520 13350 : slp_node->push_vec_def (new_stmt);
6521 : }
6522 :
6523 7821 : vec_oprnds0.release ();
6524 7821 : vec_oprnds1.release ();
6525 :
6526 7821 : return true;
6527 : }
6528 :
6529 : /* Function vectorizable_operation.
6530 :
6531 : Check if STMT_INFO performs a binary, unary or ternary operation that can
6532 : be vectorized.
6533 : If COST_VEC is passed, calculate costs but don't change anything,
6534 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
6535 : it, and insert it at GSI.
6536 : Return true if STMT_INFO is vectorizable in this way. */
6537 :
6538 : static bool
6539 2373850 : vectorizable_operation (vec_info *vinfo,
6540 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
6541 : slp_tree slp_node,
6542 : stmt_vector_for_cost *cost_vec)
6543 : {
6544 2373850 : tree vec_dest;
6545 2373850 : tree scalar_dest;
6546 2373850 : tree op0, op1 = NULL_TREE, op2 = NULL_TREE;
6547 2373850 : tree vectype;
6548 2373850 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
6549 2373850 : enum tree_code code, orig_code;
6550 2373850 : machine_mode vec_mode;
6551 2373850 : tree new_temp;
6552 2373850 : int op_type;
6553 2373850 : optab optab;
6554 2373850 : bool target_support_p;
6555 2373850 : enum vect_def_type dt[3]
6556 : = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type};
6557 2373850 : poly_uint64 nunits_in;
6558 2373850 : poly_uint64 nunits_out;
6559 2373850 : tree vectype_out;
6560 2373850 : int i;
6561 2373850 : vec<tree> vec_oprnds0 = vNULL;
6562 2373850 : vec<tree> vec_oprnds1 = vNULL;
6563 2373850 : vec<tree> vec_oprnds2 = vNULL;
6564 2373850 : tree vop0, vop1, vop2;
6565 2373850 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
6566 :
6567 2373850 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
6568 : return false;
6569 :
6570 2373850 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
6571 194522 : && cost_vec)
6572 : return false;
6573 :
6574 : /* Is STMT a vectorizable binary/unary operation? */
6575 3952815 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
6576 2112401 : if (!stmt)
6577 : return false;
6578 :
6579 : /* Loads and stores are handled in vectorizable_{load,store}. */
6580 2112401 : if (STMT_VINFO_DATA_REF (stmt_info))
6581 : return false;
6582 :
6583 868048 : orig_code = code = gimple_assign_rhs_code (stmt);
6584 :
6585 : /* Shifts are handled in vectorizable_shift. */
6586 868048 : if (code == LSHIFT_EXPR
6587 : || code == RSHIFT_EXPR
6588 : || code == LROTATE_EXPR
6589 868048 : || code == RROTATE_EXPR)
6590 : return false;
6591 :
6592 : /* Comparisons are handled in vectorizable_comparison. */
6593 817261 : if (TREE_CODE_CLASS (code) == tcc_comparison)
6594 : return false;
6595 :
6596 : /* Conditions are handled in vectorizable_condition. */
6597 650233 : if (code == COND_EXPR)
6598 : return false;
6599 :
6600 : /* For pointer addition and subtraction, we should use the normal
6601 : plus and minus for the vector operation. */
6602 630860 : if (code == POINTER_PLUS_EXPR)
6603 : code = PLUS_EXPR;
6604 611988 : if (code == POINTER_DIFF_EXPR)
6605 1026 : code = MINUS_EXPR;
6606 :
6607 : /* Support only unary or binary operations. */
6608 630860 : op_type = TREE_CODE_LENGTH (code);
6609 630860 : if (op_type != unary_op && op_type != binary_op && op_type != ternary_op)
6610 : {
6611 0 : if (dump_enabled_p ())
6612 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6613 : "num. args = %d (not unary/binary/ternary op).\n",
6614 : op_type);
6615 0 : return false;
6616 : }
6617 :
6618 630860 : scalar_dest = gimple_assign_lhs (stmt);
6619 630860 : vectype_out = SLP_TREE_VECTYPE (slp_node);
6620 :
6621 : /* Most operations cannot handle bit-precision types without extra
6622 : truncations. */
6623 630860 : bool mask_op_p = VECTOR_BOOLEAN_TYPE_P (vectype_out);
6624 622257 : if (!mask_op_p
6625 622257 : && !type_has_mode_precision_p (TREE_TYPE (scalar_dest))
6626 : /* Exception are bitwise binary operations. */
6627 : && code != BIT_IOR_EXPR
6628 1560 : && code != BIT_XOR_EXPR
6629 1235 : && code != BIT_AND_EXPR)
6630 : {
6631 997 : if (dump_enabled_p ())
6632 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6633 : "bit-precision arithmetic not supported.\n");
6634 997 : return false;
6635 : }
6636 :
6637 629863 : slp_tree slp_op0;
6638 629863 : if (!vect_is_simple_use (vinfo, slp_node,
6639 : 0, &op0, &slp_op0, &dt[0], &vectype))
6640 : {
6641 0 : if (dump_enabled_p ())
6642 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6643 : "use not simple.\n");
6644 0 : return false;
6645 : }
6646 629863 : bool is_invariant = (dt[0] == vect_external_def
6647 629863 : || dt[0] == vect_constant_def);
6648 : /* If op0 is an external or constant def, infer the vector type
6649 : from the scalar type. */
6650 629863 : if (!vectype)
6651 : {
6652 : /* For boolean type we cannot determine vectype by
6653 : invariant value (don't know whether it is a vector
6654 : of booleans or vector of integers). We use output
6655 : vectype because operations on boolean don't change
6656 : type. */
6657 70705 : if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0)))
6658 : {
6659 1131 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest)))
6660 : {
6661 239 : if (dump_enabled_p ())
6662 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6663 : "not supported operation on bool value.\n");
6664 239 : return false;
6665 : }
6666 892 : vectype = vectype_out;
6667 : }
6668 : else
6669 69574 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op0),
6670 : slp_node);
6671 : }
6672 629624 : if (!cost_vec)
6673 115465 : gcc_assert (vectype);
6674 629624 : if (!vectype)
6675 : {
6676 284 : if (dump_enabled_p ())
6677 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6678 : "no vectype for scalar type %T\n",
6679 0 : TREE_TYPE (op0));
6680 :
6681 284 : return false;
6682 : }
6683 :
6684 629340 : nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
6685 629340 : nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
6686 629340 : if (maybe_ne (nunits_out, nunits_in)
6687 629340 : || !tree_nop_conversion_p (TREE_TYPE (vectype_out), TREE_TYPE (vectype)))
6688 10759 : return false;
6689 :
6690 618581 : tree vectype2 = NULL_TREE, vectype3 = NULL_TREE;
6691 618581 : slp_tree slp_op1 = NULL, slp_op2 = NULL;
6692 618581 : if (op_type == binary_op || op_type == ternary_op)
6693 : {
6694 552275 : if (!vect_is_simple_use (vinfo, slp_node,
6695 : 1, &op1, &slp_op1, &dt[1], &vectype2))
6696 : {
6697 0 : if (dump_enabled_p ())
6698 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6699 : "use not simple.\n");
6700 0 : return false;
6701 : }
6702 552275 : is_invariant &= (dt[1] == vect_external_def
6703 552275 : || dt[1] == vect_constant_def);
6704 552275 : if (vectype2
6705 919076 : && (maybe_ne (nunits_out, TYPE_VECTOR_SUBPARTS (vectype2))
6706 366801 : || !tree_nop_conversion_p (TREE_TYPE (vectype_out),
6707 366801 : TREE_TYPE (vectype2))))
6708 4 : return false;
6709 : }
6710 618577 : if (op_type == ternary_op)
6711 : {
6712 0 : if (!vect_is_simple_use (vinfo, slp_node,
6713 : 2, &op2, &slp_op2, &dt[2], &vectype3))
6714 : {
6715 0 : if (dump_enabled_p ())
6716 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6717 : "use not simple.\n");
6718 0 : return false;
6719 : }
6720 0 : is_invariant &= (dt[2] == vect_external_def
6721 0 : || dt[2] == vect_constant_def);
6722 0 : if (vectype3
6723 0 : && (maybe_ne (nunits_out, TYPE_VECTOR_SUBPARTS (vectype3))
6724 0 : || !tree_nop_conversion_p (TREE_TYPE (vectype_out),
6725 0 : TREE_TYPE (vectype3))))
6726 0 : return false;
6727 : }
6728 :
6729 : /* Multiple types in SLP are handled by creating the appropriate number of
6730 : vectorized stmts for each SLP node. */
6731 618577 : auto vec_num = vect_get_num_copies (vinfo, slp_node);
6732 :
6733 : /* Reject attempts to combine mask types with nonmask types, e.g. if
6734 : we have an AND between a (nonmask) boolean loaded from memory and
6735 : a (mask) boolean result of a comparison.
6736 :
6737 : TODO: We could easily fix these cases up using pattern statements. */
6738 618577 : if (VECTOR_BOOLEAN_TYPE_P (vectype) != mask_op_p
6739 979596 : || (vectype2 && VECTOR_BOOLEAN_TYPE_P (vectype2) != mask_op_p)
6740 1237154 : || (vectype3 && VECTOR_BOOLEAN_TYPE_P (vectype3) != mask_op_p))
6741 : {
6742 0 : if (dump_enabled_p ())
6743 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6744 : "mixed mask and nonmask vector types\n");
6745 0 : return false;
6746 : }
6747 :
6748 : /* Supportable by target? */
6749 :
6750 618577 : vec_mode = TYPE_MODE (vectype);
6751 618577 : optab = optab_for_tree_code (code, vectype, optab_default);
6752 618577 : if (!optab)
6753 : {
6754 56745 : if (dump_enabled_p ())
6755 5801 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6756 : "no optab for %s and %T.\n",
6757 : get_tree_code_name (code), vectype);
6758 56745 : return false;
6759 : }
6760 561832 : target_support_p = can_implement_p (optab, vec_mode);
6761 :
6762 561832 : bool using_emulated_vectors_p = vect_emulated_vector_p (vectype);
6763 561832 : if (!target_support_p || using_emulated_vectors_p)
6764 : {
6765 28765 : if (dump_enabled_p ())
6766 1112 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6767 : "op not supported by target.\n");
6768 : /* When vec_mode is not a vector mode and we verified ops we
6769 : do not have to lower like AND are natively supported let
6770 : those through even when the mode isn't word_mode. For
6771 : ops we have to lower the lowering code assumes we are
6772 : dealing with word_mode. */
6773 57530 : if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype))
6774 28671 : || !GET_MODE_SIZE (vec_mode).is_constant ()
6775 28671 : || (((code == PLUS_EXPR || code == MINUS_EXPR || code == NEGATE_EXPR)
6776 23525 : || !target_support_p)
6777 61172 : && maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD))
6778 : /* Check only during analysis. */
6779 39274 : || (cost_vec && !vect_can_vectorize_without_simd_p (code)))
6780 : {
6781 28217 : if (dump_enabled_p ())
6782 1112 : dump_printf (MSG_NOTE, "using word mode not possible.\n");
6783 28217 : return false;
6784 : }
6785 548 : if (dump_enabled_p ())
6786 0 : dump_printf_loc (MSG_NOTE, vect_location,
6787 : "proceeding using word mode.\n");
6788 : using_emulated_vectors_p = true;
6789 : }
6790 :
6791 533615 : int reduc_idx = SLP_TREE_REDUC_IDX (slp_node);
6792 533615 : vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL);
6793 354725 : vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL);
6794 533615 : internal_fn cond_fn = get_conditional_internal_fn (code);
6795 533615 : internal_fn cond_len_fn = get_conditional_len_internal_fn (code);
6796 :
6797 : /* If operating on inactive elements could generate spurious traps,
6798 : we need to restrict the operation to active lanes. Note that this
6799 : specifically doesn't apply to unhoisted invariants, since they
6800 : operate on the same value for every lane.
6801 :
6802 : Similarly, if this operation is part of a reduction, a fully-masked
6803 : loop should only change the active lanes of the reduction chain,
6804 : keeping the inactive lanes as-is. */
6805 506132 : bool mask_out_inactive = ((!is_invariant && gimple_could_trap_p (stmt))
6806 988469 : || reduc_idx >= 0);
6807 :
6808 533615 : if (cost_vec) /* transformation not required. */
6809 : {
6810 418150 : if (loop_vinfo
6811 250870 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
6812 68346 : && mask_out_inactive)
6813 : {
6814 15258 : if (cond_len_fn != IFN_LAST
6815 15258 : && direct_internal_fn_supported_p (cond_len_fn, vectype,
6816 : OPTIMIZE_FOR_SPEED))
6817 0 : vect_record_loop_len (loop_vinfo, lens, vec_num, vectype,
6818 : 1);
6819 15258 : else if (cond_fn != IFN_LAST
6820 15258 : && direct_internal_fn_supported_p (cond_fn, vectype,
6821 : OPTIMIZE_FOR_SPEED))
6822 6873 : vect_record_loop_mask (loop_vinfo, masks, vec_num,
6823 : vectype, NULL);
6824 : else
6825 : {
6826 8385 : if (dump_enabled_p ())
6827 591 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6828 : "can't use a fully-masked loop because no"
6829 : " conditional operation is available.\n");
6830 8385 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
6831 : }
6832 : }
6833 :
6834 : /* Put types on constant and invariant SLP children. */
6835 418150 : if (!vect_maybe_update_slp_op_vectype (slp_op0, vectype)
6836 418072 : || !vect_maybe_update_slp_op_vectype (slp_op1, vectype)
6837 836121 : || !vect_maybe_update_slp_op_vectype (slp_op2, vectype))
6838 : {
6839 179 : if (dump_enabled_p ())
6840 3 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
6841 : "incompatible vector types for invariants\n");
6842 179 : return false;
6843 : }
6844 :
6845 417971 : SLP_TREE_TYPE (slp_node) = op_vec_info_type;
6846 417971 : DUMP_VECT_SCOPE ("vectorizable_operation");
6847 417971 : vect_model_simple_cost (vinfo, 1, slp_node, cost_vec);
6848 417971 : if (using_emulated_vectors_p)
6849 : {
6850 : /* The above vect_model_simple_cost call handles constants
6851 : in the prologue and (mis-)costs one of the stmts as
6852 : vector stmt. See below for the actual lowering that will
6853 : be applied. */
6854 546 : unsigned n = vect_get_num_copies (vinfo, slp_node);
6855 546 : switch (code)
6856 : {
6857 195 : case PLUS_EXPR:
6858 195 : n *= 5;
6859 195 : break;
6860 324 : case MINUS_EXPR:
6861 324 : n *= 6;
6862 324 : break;
6863 0 : case NEGATE_EXPR:
6864 0 : n *= 4;
6865 0 : break;
6866 : default:
6867 : /* Bit operations do not have extra cost and are accounted
6868 : as vector stmt by vect_model_simple_cost. */
6869 : n = 0;
6870 : break;
6871 : }
6872 519 : if (n != 0)
6873 : {
6874 : /* We also need to materialize two large constants. */
6875 519 : record_stmt_cost (cost_vec, 2, scalar_stmt, stmt_info,
6876 : 0, vect_prologue);
6877 519 : record_stmt_cost (cost_vec, n, scalar_stmt, stmt_info,
6878 : 0, vect_body);
6879 : }
6880 : }
6881 417971 : return true;
6882 : }
6883 :
6884 : /* Transform. */
6885 :
6886 115465 : if (dump_enabled_p ())
6887 16854 : dump_printf_loc (MSG_NOTE, vect_location,
6888 : "transform binary/unary operation.\n");
6889 :
6890 115465 : bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
6891 103855 : bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo);
6892 :
6893 : /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as
6894 : vectors with unsigned elements, but the result is signed. So, we
6895 : need to compute the MINUS_EXPR into vectype temporary and
6896 : VIEW_CONVERT_EXPR it into the final vectype_out result. */
6897 115465 : tree vec_cvt_dest = NULL_TREE;
6898 115465 : if (orig_code == POINTER_DIFF_EXPR)
6899 : {
6900 126 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
6901 126 : vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out);
6902 : }
6903 : /* For reduction operations with undefined overflow behavior make sure to
6904 : pun them to unsigned since we change the order of evaluation.
6905 : ??? Avoid for in-order reductions? */
6906 115339 : else if (arith_code_with_undefined_signed_overflow (orig_code)
6907 98855 : && ANY_INTEGRAL_TYPE_P (vectype)
6908 48697 : && TYPE_OVERFLOW_UNDEFINED (vectype)
6909 141649 : && SLP_TREE_REDUC_IDX (slp_node) != -1)
6910 : {
6911 2492 : gcc_assert (orig_code == PLUS_EXPR || orig_code == MINUS_EXPR
6912 : || orig_code == MULT_EXPR || orig_code == POINTER_PLUS_EXPR);
6913 2492 : vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out);
6914 2492 : vectype = unsigned_type_for (vectype);
6915 2492 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
6916 : }
6917 : /* Handle def. */
6918 : else
6919 112847 : vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
6920 :
6921 115465 : vect_get_vec_defs (vinfo, slp_node,
6922 : op0, &vec_oprnds0, op1, &vec_oprnds1, op2, &vec_oprnds2);
6923 : /* Arguments are ready. Create the new vector stmt. */
6924 255069 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0)
6925 : {
6926 139604 : gimple *new_stmt = NULL;
6927 279208 : vop1 = ((op_type == binary_op || op_type == ternary_op)
6928 139604 : ? vec_oprnds1[i] : NULL_TREE);
6929 139604 : vop2 = ((op_type == ternary_op) ? vec_oprnds2[i] : NULL_TREE);
6930 :
6931 139604 : if (vec_cvt_dest
6932 139604 : && !useless_type_conversion_p (vectype, TREE_TYPE (vop0)))
6933 : {
6934 2920 : new_temp = build1 (VIEW_CONVERT_EXPR, vectype, vop0);
6935 2920 : new_stmt = gimple_build_assign (vec_dest, VIEW_CONVERT_EXPR,
6936 : new_temp);
6937 2920 : new_temp = make_ssa_name (vec_dest, new_stmt);
6938 2920 : gimple_assign_set_lhs (new_stmt, new_temp);
6939 2920 : vect_finish_stmt_generation (vinfo, stmt_info,
6940 : new_stmt, gsi);
6941 2920 : vop0 = new_temp;
6942 : }
6943 139604 : if (vop1
6944 137046 : && vec_cvt_dest
6945 142665 : && !useless_type_conversion_p (vectype, TREE_TYPE (vop1)))
6946 : {
6947 2920 : new_temp = build1 (VIEW_CONVERT_EXPR, vectype, vop1);
6948 2920 : new_stmt = gimple_build_assign (vec_dest, VIEW_CONVERT_EXPR,
6949 : new_temp);
6950 2920 : new_temp = make_ssa_name (vec_dest, new_stmt);
6951 2920 : gimple_assign_set_lhs (new_stmt, new_temp);
6952 2920 : vect_finish_stmt_generation (vinfo, stmt_info,
6953 : new_stmt, gsi);
6954 2920 : vop1 = new_temp;
6955 : }
6956 139604 : if (vop2
6957 0 : && vec_cvt_dest
6958 139604 : && !useless_type_conversion_p (vectype, TREE_TYPE (vop2)))
6959 : {
6960 0 : new_temp = build1 (VIEW_CONVERT_EXPR, vectype, vop2);
6961 0 : new_stmt = gimple_build_assign (vec_dest, VIEW_CONVERT_EXPR,
6962 : new_temp);
6963 0 : new_temp = make_ssa_name (vec_dest, new_stmt);
6964 0 : gimple_assign_set_lhs (new_stmt, new_temp);
6965 0 : vect_finish_stmt_generation (vinfo, stmt_info,
6966 : new_stmt, gsi);
6967 0 : vop2 = new_temp;
6968 : }
6969 :
6970 139604 : if (using_emulated_vectors_p)
6971 : {
6972 : /* Lower the operation. This follows vector lowering. */
6973 2 : tree word_type = build_nonstandard_integer_type
6974 2 : (GET_MODE_BITSIZE (vec_mode).to_constant (), 1);
6975 2 : tree wvop0 = make_ssa_name (word_type);
6976 2 : new_stmt = gimple_build_assign (wvop0, VIEW_CONVERT_EXPR,
6977 : build1 (VIEW_CONVERT_EXPR,
6978 : word_type, vop0));
6979 2 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
6980 2 : tree wvop1 = NULL_TREE;
6981 2 : if (vop1)
6982 : {
6983 2 : wvop1 = make_ssa_name (word_type);
6984 2 : new_stmt = gimple_build_assign (wvop1, VIEW_CONVERT_EXPR,
6985 : build1 (VIEW_CONVERT_EXPR,
6986 : word_type, vop1));
6987 2 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
6988 : }
6989 :
6990 2 : tree result_low;
6991 2 : if (code == PLUS_EXPR || code == MINUS_EXPR || code == NEGATE_EXPR)
6992 : {
6993 1 : unsigned int width = vector_element_bits (vectype);
6994 1 : tree inner_type = TREE_TYPE (vectype);
6995 1 : HOST_WIDE_INT max = GET_MODE_MASK (TYPE_MODE (inner_type));
6996 1 : tree low_bits
6997 1 : = build_replicated_int_cst (word_type, width, max >> 1);
6998 1 : tree high_bits
6999 2 : = build_replicated_int_cst (word_type,
7000 1 : width, max & ~(max >> 1));
7001 1 : tree signs;
7002 1 : if (code == PLUS_EXPR || code == MINUS_EXPR)
7003 : {
7004 1 : signs = make_ssa_name (word_type);
7005 1 : new_stmt = gimple_build_assign (signs,
7006 : BIT_XOR_EXPR, wvop0, wvop1);
7007 1 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7008 1 : tree b_low = make_ssa_name (word_type);
7009 1 : new_stmt = gimple_build_assign (b_low, BIT_AND_EXPR,
7010 : wvop1, low_bits);
7011 1 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7012 1 : tree a_low = make_ssa_name (word_type);
7013 1 : if (code == PLUS_EXPR)
7014 1 : new_stmt = gimple_build_assign (a_low, BIT_AND_EXPR,
7015 : wvop0, low_bits);
7016 : else
7017 0 : new_stmt = gimple_build_assign (a_low, BIT_IOR_EXPR,
7018 : wvop0, high_bits);
7019 1 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7020 1 : if (code == MINUS_EXPR)
7021 : {
7022 0 : new_stmt = gimple_build_assign (NULL_TREE,
7023 : BIT_NOT_EXPR, signs);
7024 0 : signs = make_ssa_name (word_type);
7025 0 : gimple_assign_set_lhs (new_stmt, signs);
7026 0 : vect_finish_stmt_generation (vinfo, stmt_info,
7027 : new_stmt, gsi);
7028 : }
7029 1 : new_stmt = gimple_build_assign (NULL_TREE, BIT_AND_EXPR,
7030 : signs, high_bits);
7031 1 : signs = make_ssa_name (word_type);
7032 1 : gimple_assign_set_lhs (new_stmt, signs);
7033 1 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7034 1 : result_low = make_ssa_name (word_type);
7035 1 : new_stmt = gimple_build_assign (result_low, code,
7036 : a_low, b_low);
7037 1 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7038 : }
7039 : else /* if (code == NEGATE_EXPR) */
7040 : {
7041 0 : tree a_low = make_ssa_name (word_type);
7042 0 : new_stmt = gimple_build_assign (a_low, BIT_AND_EXPR,
7043 : wvop0, low_bits);
7044 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7045 0 : signs = make_ssa_name (word_type);
7046 0 : new_stmt = gimple_build_assign (signs, BIT_NOT_EXPR, wvop0);
7047 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7048 0 : new_stmt = gimple_build_assign (NULL_TREE, BIT_AND_EXPR,
7049 : signs, high_bits);
7050 0 : signs = make_ssa_name (word_type);
7051 0 : gimple_assign_set_lhs (new_stmt, signs);
7052 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7053 0 : result_low = make_ssa_name (word_type);
7054 0 : new_stmt = gimple_build_assign (result_low,
7055 : MINUS_EXPR, high_bits, a_low);
7056 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7057 : }
7058 1 : new_stmt = gimple_build_assign (NULL_TREE, BIT_XOR_EXPR,
7059 : result_low, signs);
7060 1 : result_low = make_ssa_name (word_type);
7061 1 : gimple_assign_set_lhs (new_stmt, result_low);
7062 1 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7063 : }
7064 : else
7065 : {
7066 1 : new_stmt = gimple_build_assign (NULL_TREE, code, wvop0, wvop1);
7067 1 : result_low = make_ssa_name (word_type);
7068 1 : gimple_assign_set_lhs (new_stmt, result_low);
7069 1 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7070 :
7071 : }
7072 2 : new_stmt = gimple_build_assign (NULL_TREE, VIEW_CONVERT_EXPR,
7073 : build1 (VIEW_CONVERT_EXPR,
7074 : vectype, result_low));
7075 2 : new_temp = make_ssa_name (vectype);
7076 2 : gimple_assign_set_lhs (new_stmt, new_temp);
7077 2 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7078 : }
7079 139602 : else if ((masked_loop_p || len_loop_p) && mask_out_inactive)
7080 : {
7081 16 : tree mask;
7082 16 : if (masked_loop_p)
7083 16 : mask = vect_get_loop_mask (loop_vinfo, gsi, masks,
7084 : vec_num, vectype, i);
7085 : else
7086 : /* Dummy mask. */
7087 0 : mask = build_minus_one_cst (truth_type_for (vectype));
7088 16 : auto_vec<tree> vops (6);
7089 16 : vops.quick_push (mask);
7090 16 : vops.quick_push (vop0);
7091 16 : if (vop1)
7092 16 : vops.quick_push (vop1);
7093 16 : if (vop2)
7094 0 : vops.quick_push (vop2);
7095 16 : if (reduc_idx >= 0)
7096 : {
7097 : /* Perform the operation on active elements only and take
7098 : inactive elements from the reduction chain input. */
7099 8 : gcc_assert (!vop2);
7100 8 : vops.quick_push (reduc_idx == 1 ? vop1 : vop0);
7101 : }
7102 : else
7103 : {
7104 8 : auto else_value = targetm.preferred_else_value
7105 8 : (cond_fn, vectype, vops.length () - 1, &vops[1]);
7106 8 : vops.quick_push (else_value);
7107 : }
7108 16 : if (len_loop_p)
7109 : {
7110 0 : tree len = vect_get_loop_len (loop_vinfo, gsi, lens,
7111 0 : vec_num, vectype, i, 1, true);
7112 0 : signed char biasval
7113 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
7114 0 : tree bias = build_int_cst (intQI_type_node, biasval);
7115 0 : vops.quick_push (len);
7116 0 : vops.quick_push (bias);
7117 : }
7118 16 : gcall *call
7119 16 : = gimple_build_call_internal_vec (masked_loop_p ? cond_fn
7120 : : cond_len_fn,
7121 : vops);
7122 16 : new_temp = make_ssa_name (vec_dest, call);
7123 16 : gimple_call_set_lhs (call, new_temp);
7124 16 : gimple_call_set_nothrow (call, true);
7125 16 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
7126 16 : new_stmt = call;
7127 16 : }
7128 : else
7129 : {
7130 139586 : tree mask = NULL_TREE;
7131 : /* When combining two masks check if either of them is elsewhere
7132 : combined with a loop mask, if that's the case we can mark that the
7133 : new combined mask doesn't need to be combined with a loop mask. */
7134 139586 : if (masked_loop_p
7135 139586 : && code == BIT_AND_EXPR
7136 139586 : && VECTOR_BOOLEAN_TYPE_P (vectype))
7137 : {
7138 8 : if (loop_vinfo->scalar_cond_masked_set.contains ({ op0, vec_num }))
7139 : {
7140 0 : mask = vect_get_loop_mask (loop_vinfo, gsi, masks,
7141 : vec_num, vectype, i);
7142 :
7143 0 : vop0 = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), mask,
7144 : vop0, gsi);
7145 : }
7146 :
7147 8 : if (loop_vinfo->scalar_cond_masked_set.contains ({ op1, vec_num }))
7148 : {
7149 0 : mask = vect_get_loop_mask (loop_vinfo, gsi, masks,
7150 : vec_num, vectype, i);
7151 :
7152 0 : vop1 = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), mask,
7153 : vop1, gsi);
7154 : }
7155 : }
7156 :
7157 139586 : new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2);
7158 139586 : new_temp = make_ssa_name (vec_dest, new_stmt);
7159 139586 : gimple_assign_set_lhs (new_stmt, new_temp);
7160 139586 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
7161 139586 : if (using_emulated_vectors_p)
7162 : suppress_warning (new_stmt, OPT_Wvector_operation_performance);
7163 :
7164 : /* Enter the combined value into the vector cond hash so we don't
7165 : AND it with a loop mask again. */
7166 139586 : if (mask)
7167 0 : loop_vinfo->vec_cond_masked_set.add ({ new_temp, mask });
7168 : }
7169 :
7170 139604 : if (vec_cvt_dest)
7171 : {
7172 3061 : new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp);
7173 3061 : new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR,
7174 : new_temp);
7175 3061 : new_temp = make_ssa_name (vec_cvt_dest, new_stmt);
7176 3061 : gimple_assign_set_lhs (new_stmt, new_temp);
7177 3061 : vect_finish_stmt_generation (vinfo, stmt_info,
7178 : new_stmt, gsi);
7179 : }
7180 :
7181 139604 : slp_node->push_vec_def (new_stmt);
7182 : }
7183 :
7184 115465 : vec_oprnds0.release ();
7185 115465 : vec_oprnds1.release ();
7186 115465 : vec_oprnds2.release ();
7187 :
7188 115465 : return true;
7189 : }
7190 :
7191 : /* A helper function to ensure data reference DR_INFO's base alignment. */
7192 :
7193 : static void
7194 1849435 : ensure_base_align (dr_vec_info *dr_info)
7195 : {
7196 : /* Alignment is only analyzed for the first element of a DR group,
7197 : use that to look at base alignment we need to enforce. */
7198 1849435 : if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt))
7199 1412157 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt));
7200 :
7201 1849435 : gcc_assert (dr_info->misalignment != DR_MISALIGNMENT_UNINITIALIZED);
7202 :
7203 1849435 : if (dr_info->base_misaligned)
7204 : {
7205 168562 : tree base_decl = dr_info->base_decl;
7206 :
7207 : // We should only be able to increase the alignment of a base object if
7208 : // we know what its new alignment should be at compile time.
7209 168562 : unsigned HOST_WIDE_INT align_base_to =
7210 168562 : DR_TARGET_ALIGNMENT (dr_info).to_constant () * BITS_PER_UNIT;
7211 :
7212 168562 : if (decl_in_symtab_p (base_decl))
7213 4612 : symtab_node::get (base_decl)->increase_alignment (align_base_to);
7214 163950 : else if (DECL_ALIGN (base_decl) < align_base_to)
7215 : {
7216 131308 : SET_DECL_ALIGN (base_decl, align_base_to);
7217 131308 : DECL_USER_ALIGN (base_decl) = 1;
7218 : }
7219 168562 : dr_info->base_misaligned = false;
7220 : }
7221 1849435 : }
7222 :
7223 :
7224 : /* Function get_group_alias_ptr_type.
7225 :
7226 : Return the alias type for the group starting at FIRST_STMT_INFO. */
7227 :
7228 : static tree
7229 1581257 : get_group_alias_ptr_type (stmt_vec_info first_stmt_info)
7230 : {
7231 1581257 : struct data_reference *first_dr, *next_dr;
7232 :
7233 1581257 : first_dr = STMT_VINFO_DATA_REF (first_stmt_info);
7234 1581257 : stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (first_stmt_info);
7235 3854849 : while (next_stmt_info)
7236 : {
7237 2405384 : next_dr = STMT_VINFO_DATA_REF (next_stmt_info);
7238 4810768 : if (get_alias_set (DR_REF (first_dr))
7239 2405384 : != get_alias_set (DR_REF (next_dr)))
7240 : {
7241 131792 : if (dump_enabled_p ())
7242 30 : dump_printf_loc (MSG_NOTE, vect_location,
7243 : "conflicting alias set types.\n");
7244 131792 : return ptr_type_node;
7245 : }
7246 2273592 : next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
7247 : }
7248 1449465 : return reference_alias_ptr_type (DR_REF (first_dr));
7249 : }
7250 :
7251 :
7252 : /* Function scan_operand_equal_p.
7253 :
7254 : Helper function for check_scan_store. Compare two references
7255 : with .GOMP_SIMD_LANE bases. */
7256 :
7257 : static bool
7258 1284 : scan_operand_equal_p (tree ref1, tree ref2)
7259 : {
7260 1284 : tree ref[2] = { ref1, ref2 };
7261 1284 : poly_int64 bitsize[2], bitpos[2];
7262 : tree offset[2], base[2];
7263 3852 : for (int i = 0; i < 2; ++i)
7264 : {
7265 2568 : machine_mode mode;
7266 2568 : int unsignedp, reversep, volatilep = 0;
7267 2568 : base[i] = get_inner_reference (ref[i], &bitsize[i], &bitpos[i],
7268 : &offset[i], &mode, &unsignedp,
7269 : &reversep, &volatilep);
7270 2568 : if (reversep || volatilep || maybe_ne (bitpos[i], 0))
7271 0 : return false;
7272 2568 : if (TREE_CODE (base[i]) == MEM_REF
7273 42 : && offset[i] == NULL_TREE
7274 2610 : && TREE_CODE (TREE_OPERAND (base[i], 0)) == SSA_NAME)
7275 : {
7276 42 : gimple *def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (base[i], 0));
7277 42 : if (is_gimple_assign (def_stmt)
7278 42 : && gimple_assign_rhs_code (def_stmt) == POINTER_PLUS_EXPR
7279 42 : && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == ADDR_EXPR
7280 84 : && TREE_CODE (gimple_assign_rhs2 (def_stmt)) == SSA_NAME)
7281 : {
7282 42 : if (maybe_ne (mem_ref_offset (base[i]), 0))
7283 : return false;
7284 42 : base[i] = TREE_OPERAND (gimple_assign_rhs1 (def_stmt), 0);
7285 42 : offset[i] = gimple_assign_rhs2 (def_stmt);
7286 : }
7287 : }
7288 : }
7289 :
7290 1284 : if (!operand_equal_p (base[0], base[1], 0))
7291 : return false;
7292 934 : if (maybe_ne (bitsize[0], bitsize[1]))
7293 : return false;
7294 934 : if (offset[0] != offset[1])
7295 : {
7296 916 : if (!offset[0] || !offset[1])
7297 : return false;
7298 916 : if (!operand_equal_p (offset[0], offset[1], 0))
7299 : {
7300 : tree step[2];
7301 0 : for (int i = 0; i < 2; ++i)
7302 : {
7303 0 : step[i] = integer_one_node;
7304 0 : if (TREE_CODE (offset[i]) == SSA_NAME)
7305 : {
7306 0 : gimple *def_stmt = SSA_NAME_DEF_STMT (offset[i]);
7307 0 : if (is_gimple_assign (def_stmt)
7308 0 : && gimple_assign_rhs_code (def_stmt) == MULT_EXPR
7309 0 : && (TREE_CODE (gimple_assign_rhs2 (def_stmt))
7310 : == INTEGER_CST))
7311 : {
7312 0 : step[i] = gimple_assign_rhs2 (def_stmt);
7313 0 : offset[i] = gimple_assign_rhs1 (def_stmt);
7314 : }
7315 : }
7316 0 : else if (TREE_CODE (offset[i]) == MULT_EXPR)
7317 : {
7318 0 : step[i] = TREE_OPERAND (offset[i], 1);
7319 0 : offset[i] = TREE_OPERAND (offset[i], 0);
7320 : }
7321 0 : tree rhs1 = NULL_TREE;
7322 0 : if (TREE_CODE (offset[i]) == SSA_NAME)
7323 : {
7324 0 : gimple *def_stmt = SSA_NAME_DEF_STMT (offset[i]);
7325 0 : if (gimple_assign_cast_p (def_stmt))
7326 0 : rhs1 = gimple_assign_rhs1 (def_stmt);
7327 : }
7328 0 : else if (CONVERT_EXPR_P (offset[i]))
7329 0 : rhs1 = TREE_OPERAND (offset[i], 0);
7330 0 : if (rhs1
7331 0 : && INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
7332 0 : && INTEGRAL_TYPE_P (TREE_TYPE (offset[i]))
7333 0 : && (TYPE_PRECISION (TREE_TYPE (offset[i]))
7334 0 : >= TYPE_PRECISION (TREE_TYPE (rhs1))))
7335 0 : offset[i] = rhs1;
7336 : }
7337 0 : if (!operand_equal_p (offset[0], offset[1], 0)
7338 0 : || !operand_equal_p (step[0], step[1], 0))
7339 0 : return false;
7340 : }
7341 : }
7342 : return true;
7343 : }
7344 :
7345 :
7346 : enum scan_store_kind {
7347 : /* Normal permutation. */
7348 : scan_store_kind_perm,
7349 :
7350 : /* Whole vector left shift permutation with zero init. */
7351 : scan_store_kind_lshift_zero,
7352 :
7353 : /* Whole vector left shift permutation and VEC_COND_EXPR. */
7354 : scan_store_kind_lshift_cond
7355 : };
7356 :
7357 : /* Function check_scan_store.
7358 :
7359 : Verify if we can perform the needed permutations or whole vector shifts.
7360 : Return -1 on failure, otherwise exact log2 of vectype's nunits.
7361 : USE_WHOLE_VECTOR is a vector of enum scan_store_kind which operation
7362 : to do at each step. */
7363 :
7364 : static int
7365 1024 : scan_store_can_perm_p (tree vectype, tree init,
7366 : vec<enum scan_store_kind> *use_whole_vector = NULL)
7367 : {
7368 1024 : enum machine_mode vec_mode = TYPE_MODE (vectype);
7369 1024 : unsigned HOST_WIDE_INT nunits;
7370 1024 : if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits))
7371 : return -1;
7372 1024 : int units_log2 = exact_log2 (nunits);
7373 1024 : if (units_log2 <= 0)
7374 : return -1;
7375 :
7376 : int i;
7377 : enum scan_store_kind whole_vector_shift_kind = scan_store_kind_perm;
7378 4784 : for (i = 0; i <= units_log2; ++i)
7379 : {
7380 3760 : unsigned HOST_WIDE_INT j, k;
7381 3760 : enum scan_store_kind kind = scan_store_kind_perm;
7382 3760 : vec_perm_builder sel (nunits, nunits, 1);
7383 3760 : sel.quick_grow (nunits);
7384 3760 : if (i == units_log2)
7385 : {
7386 9728 : for (j = 0; j < nunits; ++j)
7387 8704 : sel[j] = nunits - 1;
7388 : }
7389 : else
7390 : {
7391 10416 : for (j = 0; j < (HOST_WIDE_INT_1U << i); ++j)
7392 7680 : sel[j] = j;
7393 26416 : for (k = 0; j < nunits; ++j, ++k)
7394 23680 : sel[j] = nunits + k;
7395 : }
7396 6496 : vec_perm_indices indices (sel, i == units_log2 ? 1 : 2, nunits);
7397 3760 : if (!can_vec_perm_const_p (vec_mode, vec_mode, indices))
7398 : {
7399 0 : if (i == units_log2)
7400 : return -1;
7401 :
7402 0 : if (whole_vector_shift_kind == scan_store_kind_perm)
7403 : {
7404 0 : if (!can_implement_p (vec_shl_optab, vec_mode))
7405 : return -1;
7406 0 : whole_vector_shift_kind = scan_store_kind_lshift_zero;
7407 : /* Whole vector shifts shift in zeros, so if init is all zero
7408 : constant, there is no need to do anything further. */
7409 0 : if ((TREE_CODE (init) != INTEGER_CST
7410 0 : && TREE_CODE (init) != REAL_CST)
7411 0 : || !initializer_zerop (init))
7412 : {
7413 0 : tree masktype = truth_type_for (vectype);
7414 0 : if (!expand_vec_cond_expr_p (vectype, masktype))
7415 : return -1;
7416 : whole_vector_shift_kind = scan_store_kind_lshift_cond;
7417 : }
7418 : }
7419 0 : kind = whole_vector_shift_kind;
7420 : }
7421 3760 : if (use_whole_vector)
7422 : {
7423 1880 : if (kind != scan_store_kind_perm && use_whole_vector->is_empty ())
7424 0 : use_whole_vector->safe_grow_cleared (i, true);
7425 5640 : if (kind != scan_store_kind_perm || !use_whole_vector->is_empty ())
7426 0 : use_whole_vector->safe_push (kind);
7427 : }
7428 3760 : }
7429 :
7430 : return units_log2;
7431 : }
7432 :
7433 :
7434 : /* Function check_scan_store.
7435 :
7436 : Check magic stores for #pragma omp scan {in,ex}clusive reductions. */
7437 :
7438 : static bool
7439 1076 : check_scan_store (vec_info *vinfo, stmt_vec_info stmt_info, tree vectype,
7440 : enum vect_def_type rhs_dt, slp_tree slp_node,
7441 : slp_tree mask_node,
7442 : vect_memory_access_type memory_access_type)
7443 : {
7444 1076 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
7445 1076 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info);
7446 1076 : tree ref_type;
7447 :
7448 1076 : gcc_assert (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) > 1);
7449 1076 : if (SLP_TREE_LANES (slp_node) > 1
7450 1076 : || mask_node
7451 1076 : || memory_access_type != VMAT_CONTIGUOUS
7452 1076 : || TREE_CODE (DR_BASE_ADDRESS (dr_info->dr)) != ADDR_EXPR
7453 1076 : || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0))
7454 1076 : || loop_vinfo == NULL
7455 1076 : || LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
7456 1076 : || LOOP_VINFO_EPILOGUE_P (loop_vinfo)
7457 1076 : || STMT_VINFO_GROUPED_ACCESS (stmt_info)
7458 1076 : || !integer_zerop (get_dr_vinfo_offset (vinfo, dr_info))
7459 1076 : || !integer_zerop (DR_INIT (dr_info->dr))
7460 1076 : || !(ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr)))
7461 2152 : || !alias_sets_conflict_p (get_alias_set (vectype),
7462 1076 : get_alias_set (TREE_TYPE (ref_type))))
7463 : {
7464 0 : if (dump_enabled_p ())
7465 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
7466 : "unsupported OpenMP scan store.\n");
7467 0 : return false;
7468 : }
7469 :
7470 : /* We need to pattern match code built by OpenMP lowering and simplified
7471 : by following optimizations into something we can handle.
7472 : #pragma omp simd reduction(inscan,+:r)
7473 : for (...)
7474 : {
7475 : r += something ();
7476 : #pragma omp scan inclusive (r)
7477 : use (r);
7478 : }
7479 : shall have body with:
7480 : // Initialization for input phase, store the reduction initializer:
7481 : _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0);
7482 : _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1);
7483 : D.2042[_21] = 0;
7484 : // Actual input phase:
7485 : ...
7486 : r.0_5 = D.2042[_20];
7487 : _6 = _4 + r.0_5;
7488 : D.2042[_20] = _6;
7489 : // Initialization for scan phase:
7490 : _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 2);
7491 : _26 = D.2043[_25];
7492 : _27 = D.2042[_25];
7493 : _28 = _26 + _27;
7494 : D.2043[_25] = _28;
7495 : D.2042[_25] = _28;
7496 : // Actual scan phase:
7497 : ...
7498 : r.1_8 = D.2042[_20];
7499 : ...
7500 : The "omp simd array" variable D.2042 holds the privatized copy used
7501 : inside of the loop and D.2043 is another one that holds copies of
7502 : the current original list item. The separate GOMP_SIMD_LANE ifn
7503 : kinds are there in order to allow optimizing the initializer store
7504 : and combiner sequence, e.g. if it is originally some C++ish user
7505 : defined reduction, but allow the vectorizer to pattern recognize it
7506 : and turn into the appropriate vectorized scan.
7507 :
7508 : For exclusive scan, this is slightly different:
7509 : #pragma omp simd reduction(inscan,+:r)
7510 : for (...)
7511 : {
7512 : use (r);
7513 : #pragma omp scan exclusive (r)
7514 : r += something ();
7515 : }
7516 : shall have body with:
7517 : // Initialization for input phase, store the reduction initializer:
7518 : _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0);
7519 : _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1);
7520 : D.2042[_21] = 0;
7521 : // Actual input phase:
7522 : ...
7523 : r.0_5 = D.2042[_20];
7524 : _6 = _4 + r.0_5;
7525 : D.2042[_20] = _6;
7526 : // Initialization for scan phase:
7527 : _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 3);
7528 : _26 = D.2043[_25];
7529 : D.2044[_25] = _26;
7530 : _27 = D.2042[_25];
7531 : _28 = _26 + _27;
7532 : D.2043[_25] = _28;
7533 : // Actual scan phase:
7534 : ...
7535 : r.1_8 = D.2044[_20];
7536 : ... */
7537 :
7538 1076 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 2)
7539 : {
7540 : /* Match the D.2042[_21] = 0; store above. Just require that
7541 : it is a constant or external definition store. */
7542 564 : if (rhs_dt != vect_constant_def && rhs_dt != vect_external_def)
7543 : {
7544 0 : fail_init:
7545 0 : if (dump_enabled_p ())
7546 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
7547 : "unsupported OpenMP scan initializer store.\n");
7548 0 : return false;
7549 : }
7550 :
7551 564 : if (! loop_vinfo->scan_map)
7552 322 : loop_vinfo->scan_map = new hash_map<tree, tree>;
7553 564 : tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0);
7554 564 : tree &cached = loop_vinfo->scan_map->get_or_insert (var);
7555 564 : if (cached)
7556 0 : goto fail_init;
7557 564 : cached = gimple_assign_rhs1 (STMT_VINFO_STMT (stmt_info));
7558 :
7559 : /* These stores can be vectorized normally. */
7560 564 : return true;
7561 : }
7562 :
7563 512 : if (rhs_dt != vect_internal_def)
7564 : {
7565 0 : fail:
7566 0 : if (dump_enabled_p ())
7567 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
7568 : "unsupported OpenMP scan combiner pattern.\n");
7569 0 : return false;
7570 : }
7571 :
7572 512 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
7573 512 : tree rhs = gimple_assign_rhs1 (stmt);
7574 512 : if (TREE_CODE (rhs) != SSA_NAME)
7575 0 : goto fail;
7576 :
7577 512 : gimple *other_store_stmt = NULL;
7578 512 : tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0);
7579 512 : bool inscan_var_store
7580 512 : = lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var)) != NULL;
7581 :
7582 512 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4)
7583 : {
7584 252 : if (!inscan_var_store)
7585 : {
7586 126 : use_operand_p use_p;
7587 126 : imm_use_iterator iter;
7588 378 : FOR_EACH_IMM_USE_FAST (use_p, iter, rhs)
7589 : {
7590 252 : gimple *use_stmt = USE_STMT (use_p);
7591 252 : if (use_stmt == stmt || is_gimple_debug (use_stmt))
7592 126 : continue;
7593 126 : if (gimple_bb (use_stmt) != gimple_bb (stmt)
7594 126 : || !is_gimple_assign (use_stmt)
7595 126 : || gimple_assign_rhs_class (use_stmt) != GIMPLE_BINARY_RHS
7596 126 : || other_store_stmt
7597 252 : || TREE_CODE (gimple_assign_lhs (use_stmt)) != SSA_NAME)
7598 0 : goto fail;
7599 126 : other_store_stmt = use_stmt;
7600 0 : }
7601 126 : if (other_store_stmt == NULL)
7602 0 : goto fail;
7603 126 : rhs = gimple_assign_lhs (other_store_stmt);
7604 126 : if (!single_imm_use (rhs, &use_p, &other_store_stmt))
7605 0 : goto fail;
7606 : }
7607 : }
7608 260 : else if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 3)
7609 : {
7610 260 : use_operand_p use_p;
7611 260 : imm_use_iterator iter;
7612 1040 : FOR_EACH_IMM_USE_FAST (use_p, iter, rhs)
7613 : {
7614 520 : gimple *use_stmt = USE_STMT (use_p);
7615 520 : if (use_stmt == stmt || is_gimple_debug (use_stmt))
7616 260 : continue;
7617 260 : if (other_store_stmt)
7618 0 : goto fail;
7619 260 : other_store_stmt = use_stmt;
7620 260 : }
7621 : }
7622 : else
7623 0 : goto fail;
7624 :
7625 512 : gimple *def_stmt = SSA_NAME_DEF_STMT (rhs);
7626 512 : if (gimple_bb (def_stmt) != gimple_bb (stmt)
7627 512 : || !is_gimple_assign (def_stmt)
7628 1024 : || gimple_assign_rhs_class (def_stmt) != GIMPLE_BINARY_RHS)
7629 0 : goto fail;
7630 :
7631 512 : enum tree_code code = gimple_assign_rhs_code (def_stmt);
7632 : /* For pointer addition, we should use the normal plus for the vector
7633 : operation. */
7634 512 : switch (code)
7635 : {
7636 0 : case POINTER_PLUS_EXPR:
7637 0 : code = PLUS_EXPR;
7638 0 : break;
7639 0 : case MULT_HIGHPART_EXPR:
7640 0 : goto fail;
7641 : default:
7642 : break;
7643 : }
7644 512 : if (TREE_CODE_LENGTH (code) != binary_op || !commutative_tree_code (code))
7645 0 : goto fail;
7646 :
7647 512 : tree rhs1 = gimple_assign_rhs1 (def_stmt);
7648 512 : tree rhs2 = gimple_assign_rhs2 (def_stmt);
7649 512 : if (TREE_CODE (rhs1) != SSA_NAME || TREE_CODE (rhs2) != SSA_NAME)
7650 0 : goto fail;
7651 :
7652 512 : gimple *load1_stmt = SSA_NAME_DEF_STMT (rhs1);
7653 512 : gimple *load2_stmt = SSA_NAME_DEF_STMT (rhs2);
7654 512 : if (gimple_bb (load1_stmt) != gimple_bb (stmt)
7655 512 : || !gimple_assign_load_p (load1_stmt)
7656 512 : || gimple_bb (load2_stmt) != gimple_bb (stmt)
7657 1024 : || !gimple_assign_load_p (load2_stmt))
7658 0 : goto fail;
7659 :
7660 512 : stmt_vec_info load1_stmt_info = loop_vinfo->lookup_stmt (load1_stmt);
7661 512 : stmt_vec_info load2_stmt_info = loop_vinfo->lookup_stmt (load2_stmt);
7662 512 : if (load1_stmt_info == NULL
7663 512 : || load2_stmt_info == NULL
7664 512 : || (STMT_VINFO_SIMD_LANE_ACCESS_P (load1_stmt_info)
7665 512 : != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info))
7666 512 : || (STMT_VINFO_SIMD_LANE_ACCESS_P (load2_stmt_info)
7667 512 : != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info)))
7668 0 : goto fail;
7669 :
7670 512 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && inscan_var_store)
7671 : {
7672 126 : dr_vec_info *load1_dr_info = STMT_VINFO_DR_INFO (load1_stmt_info);
7673 126 : if (TREE_CODE (DR_BASE_ADDRESS (load1_dr_info->dr)) != ADDR_EXPR
7674 126 : || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0)))
7675 0 : goto fail;
7676 126 : tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0);
7677 126 : tree lrhs;
7678 126 : if (lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var1)))
7679 : lrhs = rhs1;
7680 : else
7681 16 : lrhs = rhs2;
7682 126 : use_operand_p use_p;
7683 126 : imm_use_iterator iter;
7684 504 : FOR_EACH_IMM_USE_FAST (use_p, iter, lrhs)
7685 : {
7686 252 : gimple *use_stmt = USE_STMT (use_p);
7687 252 : if (use_stmt == def_stmt || is_gimple_debug (use_stmt))
7688 126 : continue;
7689 126 : if (other_store_stmt)
7690 0 : goto fail;
7691 126 : other_store_stmt = use_stmt;
7692 126 : }
7693 : }
7694 :
7695 512 : if (other_store_stmt == NULL)
7696 0 : goto fail;
7697 512 : if (gimple_bb (other_store_stmt) != gimple_bb (stmt)
7698 512 : || !gimple_store_p (other_store_stmt))
7699 0 : goto fail;
7700 :
7701 512 : stmt_vec_info other_store_stmt_info
7702 512 : = loop_vinfo->lookup_stmt (other_store_stmt);
7703 512 : if (other_store_stmt_info == NULL
7704 512 : || (STMT_VINFO_SIMD_LANE_ACCESS_P (other_store_stmt_info)
7705 512 : != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info)))
7706 0 : goto fail;
7707 :
7708 512 : gimple *stmt1 = stmt;
7709 512 : gimple *stmt2 = other_store_stmt;
7710 512 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store)
7711 : std::swap (stmt1, stmt2);
7712 512 : if (scan_operand_equal_p (gimple_assign_lhs (stmt1),
7713 : gimple_assign_rhs1 (load2_stmt)))
7714 : {
7715 162 : std::swap (rhs1, rhs2);
7716 162 : std::swap (load1_stmt, load2_stmt);
7717 162 : std::swap (load1_stmt_info, load2_stmt_info);
7718 : }
7719 512 : if (!scan_operand_equal_p (gimple_assign_lhs (stmt1),
7720 : gimple_assign_rhs1 (load1_stmt)))
7721 0 : goto fail;
7722 :
7723 512 : tree var3 = NULL_TREE;
7724 512 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 3
7725 512 : && !scan_operand_equal_p (gimple_assign_lhs (stmt2),
7726 : gimple_assign_rhs1 (load2_stmt)))
7727 0 : goto fail;
7728 512 : else if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4)
7729 : {
7730 252 : dr_vec_info *load2_dr_info = STMT_VINFO_DR_INFO (load2_stmt_info);
7731 252 : if (TREE_CODE (DR_BASE_ADDRESS (load2_dr_info->dr)) != ADDR_EXPR
7732 252 : || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0)))
7733 0 : goto fail;
7734 252 : var3 = TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0);
7735 252 : if (!lookup_attribute ("omp simd array", DECL_ATTRIBUTES (var3))
7736 252 : || lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var3))
7737 504 : || lookup_attribute ("omp simd inscan exclusive",
7738 252 : DECL_ATTRIBUTES (var3)))
7739 0 : goto fail;
7740 : }
7741 :
7742 512 : dr_vec_info *other_dr_info = STMT_VINFO_DR_INFO (other_store_stmt_info);
7743 512 : if (TREE_CODE (DR_BASE_ADDRESS (other_dr_info->dr)) != ADDR_EXPR
7744 512 : || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (other_dr_info->dr), 0)))
7745 0 : goto fail;
7746 :
7747 512 : tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0);
7748 512 : tree var2 = TREE_OPERAND (DR_BASE_ADDRESS (other_dr_info->dr), 0);
7749 512 : if (!lookup_attribute ("omp simd array", DECL_ATTRIBUTES (var1))
7750 512 : || !lookup_attribute ("omp simd array", DECL_ATTRIBUTES (var2))
7751 1024 : || (!lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var1)))
7752 512 : == (!lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var2))))
7753 0 : goto fail;
7754 :
7755 512 : if (lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var1)))
7756 256 : std::swap (var1, var2);
7757 :
7758 512 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4)
7759 : {
7760 252 : if (!lookup_attribute ("omp simd inscan exclusive",
7761 252 : DECL_ATTRIBUTES (var1)))
7762 0 : goto fail;
7763 252 : var1 = var3;
7764 : }
7765 :
7766 512 : if (loop_vinfo->scan_map == NULL)
7767 0 : goto fail;
7768 512 : tree *init = loop_vinfo->scan_map->get (var1);
7769 512 : if (init == NULL)
7770 0 : goto fail;
7771 :
7772 : /* The IL is as expected, now check if we can actually vectorize it.
7773 : Inclusive scan:
7774 : _26 = D.2043[_25];
7775 : _27 = D.2042[_25];
7776 : _28 = _26 + _27;
7777 : D.2043[_25] = _28;
7778 : D.2042[_25] = _28;
7779 : should be vectorized as (where _40 is the vectorized rhs
7780 : from the D.2042[_21] = 0; store):
7781 : _30 = MEM <vector(8) int> [(int *)&D.2043];
7782 : _31 = MEM <vector(8) int> [(int *)&D.2042];
7783 : _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>;
7784 : _33 = _31 + _32;
7785 : // _33 = { _31[0], _31[0]+_31[1], _31[1]+_31[2], ..., _31[6]+_31[7] };
7786 : _34 = VEC_PERM_EXPR <_40, _33, { 0, 1, 8, 9, 10, 11, 12, 13 }>;
7787 : _35 = _33 + _34;
7788 : // _35 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7789 : // _31[1]+.._31[4], ... _31[4]+.._31[7] };
7790 : _36 = VEC_PERM_EXPR <_40, _35, { 0, 1, 2, 3, 8, 9, 10, 11 }>;
7791 : _37 = _35 + _36;
7792 : // _37 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7793 : // _31[0]+.._31[4], ... _31[0]+.._31[7] };
7794 : _38 = _30 + _37;
7795 : _39 = VEC_PERM_EXPR <_38, _38, { 7, 7, 7, 7, 7, 7, 7, 7 }>;
7796 : MEM <vector(8) int> [(int *)&D.2043] = _39;
7797 : MEM <vector(8) int> [(int *)&D.2042] = _38;
7798 : Exclusive scan:
7799 : _26 = D.2043[_25];
7800 : D.2044[_25] = _26;
7801 : _27 = D.2042[_25];
7802 : _28 = _26 + _27;
7803 : D.2043[_25] = _28;
7804 : should be vectorized as (where _40 is the vectorized rhs
7805 : from the D.2042[_21] = 0; store):
7806 : _30 = MEM <vector(8) int> [(int *)&D.2043];
7807 : _31 = MEM <vector(8) int> [(int *)&D.2042];
7808 : _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>;
7809 : _33 = VEC_PERM_EXPR <_40, _32, { 0, 8, 9, 10, 11, 12, 13, 14 }>;
7810 : _34 = _32 + _33;
7811 : // _34 = { 0, _31[0], _31[0]+_31[1], _31[1]+_31[2], _31[2]+_31[3],
7812 : // _31[3]+_31[4], ... _31[5]+.._31[6] };
7813 : _35 = VEC_PERM_EXPR <_40, _34, { 0, 1, 8, 9, 10, 11, 12, 13 }>;
7814 : _36 = _34 + _35;
7815 : // _36 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7816 : // _31[1]+.._31[4], ... _31[3]+.._31[6] };
7817 : _37 = VEC_PERM_EXPR <_40, _36, { 0, 1, 2, 3, 8, 9, 10, 11 }>;
7818 : _38 = _36 + _37;
7819 : // _38 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7820 : // _31[0]+.._31[4], ... _31[0]+.._31[6] };
7821 : _39 = _30 + _38;
7822 : _50 = _31 + _39;
7823 : _51 = VEC_PERM_EXPR <_50, _50, { 7, 7, 7, 7, 7, 7, 7, 7 }>;
7824 : MEM <vector(8) int> [(int *)&D.2044] = _39;
7825 : MEM <vector(8) int> [(int *)&D.2042] = _51; */
7826 512 : enum machine_mode vec_mode = TYPE_MODE (vectype);
7827 512 : optab optab = optab_for_tree_code (code, vectype, optab_default);
7828 512 : if (!optab || !can_implement_p (optab, vec_mode))
7829 0 : goto fail;
7830 :
7831 512 : int units_log2 = scan_store_can_perm_p (vectype, *init);
7832 512 : if (units_log2 == -1)
7833 0 : goto fail;
7834 :
7835 : return true;
7836 : }
7837 :
7838 :
7839 : /* Function vectorizable_scan_store.
7840 :
7841 : Helper of vectorizable_score, arguments like on vectorizable_store.
7842 : Handle only the transformation, checking is done in check_scan_store. */
7843 :
7844 : static bool
7845 512 : vectorizable_scan_store (vec_info *vinfo, stmt_vec_info stmt_info,
7846 : slp_tree slp_node, gimple_stmt_iterator *gsi)
7847 : {
7848 512 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
7849 512 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info);
7850 512 : tree ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr));
7851 512 : tree vectype = SLP_TREE_VECTYPE (slp_node);
7852 :
7853 512 : if (dump_enabled_p ())
7854 492 : dump_printf_loc (MSG_NOTE, vect_location,
7855 : "transform scan store.\n");
7856 :
7857 512 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
7858 512 : tree rhs = gimple_assign_rhs1 (stmt);
7859 512 : gcc_assert (TREE_CODE (rhs) == SSA_NAME);
7860 :
7861 512 : tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0);
7862 512 : bool inscan_var_store
7863 512 : = lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var)) != NULL;
7864 :
7865 512 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store)
7866 : {
7867 126 : use_operand_p use_p;
7868 126 : imm_use_iterator iter;
7869 252 : FOR_EACH_IMM_USE_FAST (use_p, iter, rhs)
7870 : {
7871 126 : gimple *use_stmt = USE_STMT (use_p);
7872 126 : if (use_stmt == stmt || is_gimple_debug (use_stmt))
7873 0 : continue;
7874 126 : rhs = gimple_assign_lhs (use_stmt);
7875 126 : break;
7876 126 : }
7877 : }
7878 :
7879 512 : gimple *def_stmt = SSA_NAME_DEF_STMT (rhs);
7880 512 : enum tree_code code = gimple_assign_rhs_code (def_stmt);
7881 512 : if (code == POINTER_PLUS_EXPR)
7882 0 : code = PLUS_EXPR;
7883 512 : gcc_assert (TREE_CODE_LENGTH (code) == binary_op
7884 : && commutative_tree_code (code));
7885 512 : tree rhs1 = gimple_assign_rhs1 (def_stmt);
7886 512 : tree rhs2 = gimple_assign_rhs2 (def_stmt);
7887 512 : gcc_assert (TREE_CODE (rhs1) == SSA_NAME && TREE_CODE (rhs2) == SSA_NAME);
7888 512 : gimple *load1_stmt = SSA_NAME_DEF_STMT (rhs1);
7889 512 : gimple *load2_stmt = SSA_NAME_DEF_STMT (rhs2);
7890 512 : stmt_vec_info load1_stmt_info = loop_vinfo->lookup_stmt (load1_stmt);
7891 512 : stmt_vec_info load2_stmt_info = loop_vinfo->lookup_stmt (load2_stmt);
7892 512 : dr_vec_info *load1_dr_info = STMT_VINFO_DR_INFO (load1_stmt_info);
7893 512 : dr_vec_info *load2_dr_info = STMT_VINFO_DR_INFO (load2_stmt_info);
7894 512 : tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0);
7895 512 : tree var2 = TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0);
7896 :
7897 512 : if (lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var1)))
7898 : {
7899 436 : std::swap (rhs1, rhs2);
7900 436 : std::swap (var1, var2);
7901 436 : std::swap (load1_dr_info, load2_dr_info);
7902 : }
7903 :
7904 512 : tree *init = loop_vinfo->scan_map->get (var1);
7905 512 : gcc_assert (init);
7906 :
7907 512 : unsigned HOST_WIDE_INT nunits;
7908 512 : if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits))
7909 : gcc_unreachable ();
7910 512 : auto_vec<enum scan_store_kind, 16> use_whole_vector;
7911 512 : int units_log2 = scan_store_can_perm_p (vectype, *init, &use_whole_vector);
7912 512 : gcc_assert (units_log2 > 0);
7913 512 : auto_vec<tree, 16> perms;
7914 512 : perms.quick_grow (units_log2 + 1);
7915 512 : tree zero_vec = NULL_TREE, masktype = NULL_TREE;
7916 2392 : for (int i = 0; i <= units_log2; ++i)
7917 : {
7918 1880 : unsigned HOST_WIDE_INT j, k;
7919 1880 : vec_perm_builder sel (nunits, nunits, 1);
7920 1880 : sel.quick_grow (nunits);
7921 1880 : if (i == units_log2)
7922 4864 : for (j = 0; j < nunits; ++j)
7923 4352 : sel[j] = nunits - 1;
7924 : else
7925 : {
7926 5208 : for (j = 0; j < (HOST_WIDE_INT_1U << i); ++j)
7927 3840 : sel[j] = j;
7928 13208 : for (k = 0; j < nunits; ++j, ++k)
7929 11840 : sel[j] = nunits + k;
7930 : }
7931 3248 : vec_perm_indices indices (sel, i == units_log2 ? 1 : 2, nunits);
7932 1880 : if (!use_whole_vector.is_empty ()
7933 0 : && use_whole_vector[i] != scan_store_kind_perm)
7934 : {
7935 0 : if (zero_vec == NULL_TREE)
7936 0 : zero_vec = build_zero_cst (vectype);
7937 0 : if (masktype == NULL_TREE
7938 0 : && use_whole_vector[i] == scan_store_kind_lshift_cond)
7939 0 : masktype = truth_type_for (vectype);
7940 0 : perms[i] = vect_gen_perm_mask_any (vectype, indices);
7941 : }
7942 : else
7943 1880 : perms[i] = vect_gen_perm_mask_checked (vectype, indices);
7944 1880 : }
7945 :
7946 512 : vec_loop_lens *loop_lens
7947 512 : = (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)
7948 : ? &LOOP_VINFO_LENS (loop_vinfo)
7949 0 : : NULL);
7950 :
7951 512 : tree vec_oprnd1 = NULL_TREE;
7952 512 : tree vec_oprnd2 = NULL_TREE;
7953 512 : tree vec_oprnd3 = NULL_TREE;
7954 512 : tree dataref_ptr = DR_BASE_ADDRESS (dr_info->dr);
7955 512 : tree dataref_offset = build_int_cst (ref_type, 0);
7956 512 : tree bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info,
7957 : vectype, VMAT_CONTIGUOUS,
7958 : loop_lens);
7959 512 : tree ldataref_ptr = NULL_TREE;
7960 512 : tree orig = NULL_TREE;
7961 512 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store)
7962 126 : ldataref_ptr = DR_BASE_ADDRESS (load1_dr_info->dr);
7963 : /* The initialization is invariant. */
7964 512 : vec_oprnd1 = vect_init_vector (vinfo, stmt_info, *init, vectype, NULL);
7965 512 : auto_vec<tree> vec_oprnds2;
7966 512 : auto_vec<tree> vec_oprnds3;
7967 512 : if (ldataref_ptr == NULL)
7968 : {
7969 : /* We want to lookup the vector operands of the reduction, not those
7970 : of the store - for SLP we have to use the proper SLP node for the
7971 : lookup, which should be the single child of the scan store. */
7972 386 : vect_get_vec_defs (vinfo, SLP_TREE_CHILDREN (slp_node)[0],
7973 : rhs1, &vec_oprnds2, rhs2, &vec_oprnds3);
7974 : /* ??? For SLP we do not key the def on 'rhs1' or 'rhs2' but get
7975 : them in SLP child order. So we have to swap here with logic
7976 : similar to above. */
7977 386 : stmt_vec_info load
7978 386 : = SLP_TREE_SCALAR_STMTS (SLP_TREE_CHILDREN
7979 386 : (SLP_TREE_CHILDREN (slp_node)[0])[0])[0];
7980 386 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (load);
7981 386 : tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0);
7982 386 : if (lookup_attribute ("omp simd inscan", DECL_ATTRIBUTES (var)))
7983 820 : for (unsigned i = 0; i < vec_oprnds2.length (); ++i)
7984 494 : std::swap (vec_oprnds2[i], vec_oprnds3[i]);;
7985 : }
7986 : else
7987 126 : vect_get_vec_defs (vinfo, slp_node,
7988 : rhs2, &vec_oprnds3);
7989 1248 : for (unsigned j = 0; j < vec_oprnds3.length (); j++)
7990 : {
7991 736 : if (ldataref_ptr == NULL)
7992 554 : vec_oprnd2 = vec_oprnds2[j];
7993 736 : vec_oprnd3 = vec_oprnds3[j];
7994 736 : if (j == 0)
7995 : orig = vec_oprnd3;
7996 224 : else if (!inscan_var_store)
7997 112 : dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump);
7998 :
7999 736 : if (ldataref_ptr)
8000 : {
8001 182 : vec_oprnd2 = make_ssa_name (vectype);
8002 182 : tree data_ref = fold_build2 (MEM_REF, vectype,
8003 : unshare_expr (ldataref_ptr),
8004 : dataref_offset);
8005 182 : vect_copy_ref_info (data_ref, DR_REF (load1_dr_info->dr));
8006 182 : gimple *g = gimple_build_assign (vec_oprnd2, data_ref);
8007 182 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8008 : }
8009 :
8010 736 : tree v = vec_oprnd2;
8011 3068 : for (int i = 0; i < units_log2; ++i)
8012 : {
8013 2332 : tree new_temp = make_ssa_name (vectype);
8014 2332 : gimple *g = gimple_build_assign (new_temp, VEC_PERM_EXPR,
8015 : (zero_vec
8016 0 : && (use_whole_vector[i]
8017 0 : != scan_store_kind_perm))
8018 : ? zero_vec : vec_oprnd1, v,
8019 2332 : perms[i]);
8020 2332 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8021 :
8022 2332 : if (zero_vec && use_whole_vector[i] == scan_store_kind_lshift_cond)
8023 : {
8024 : /* Whole vector shift shifted in zero bits, but if *init
8025 : is not initializer_zerop, we need to replace those elements
8026 : with elements from vec_oprnd1. */
8027 0 : tree_vector_builder vb (masktype, nunits, 1);
8028 0 : for (unsigned HOST_WIDE_INT k = 0; k < nunits; ++k)
8029 0 : vb.quick_push (k < (HOST_WIDE_INT_1U << i)
8030 : ? boolean_false_node : boolean_true_node);
8031 :
8032 0 : tree new_temp2 = make_ssa_name (vectype);
8033 0 : g = gimple_build_assign (new_temp2, VEC_COND_EXPR, vb.build (),
8034 : new_temp, vec_oprnd1);
8035 0 : vect_finish_stmt_generation (vinfo, stmt_info,
8036 : g, gsi);
8037 0 : new_temp = new_temp2;
8038 0 : }
8039 :
8040 : /* For exclusive scan, perform the perms[i] permutation once
8041 : more. */
8042 2332 : if (i == 0
8043 1100 : && STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4
8044 728 : && v == vec_oprnd2)
8045 : {
8046 364 : v = new_temp;
8047 364 : --i;
8048 364 : continue;
8049 : }
8050 :
8051 1968 : tree new_temp2 = make_ssa_name (vectype);
8052 1968 : g = gimple_build_assign (new_temp2, code, v, new_temp);
8053 1968 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8054 :
8055 1968 : v = new_temp2;
8056 : }
8057 :
8058 736 : tree new_temp = make_ssa_name (vectype);
8059 736 : gimple *g = gimple_build_assign (new_temp, code, orig, v);
8060 736 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8061 :
8062 736 : tree last_perm_arg = new_temp;
8063 : /* For exclusive scan, new_temp computed above is the exclusive scan
8064 : prefix sum. Turn it into inclusive prefix sum for the broadcast
8065 : of the last element into orig. */
8066 736 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4)
8067 : {
8068 364 : last_perm_arg = make_ssa_name (vectype);
8069 364 : g = gimple_build_assign (last_perm_arg, code, new_temp, vec_oprnd2);
8070 364 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8071 : }
8072 :
8073 736 : orig = make_ssa_name (vectype);
8074 2208 : g = gimple_build_assign (orig, VEC_PERM_EXPR, last_perm_arg,
8075 736 : last_perm_arg, perms[units_log2]);
8076 736 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8077 :
8078 736 : if (!inscan_var_store)
8079 : {
8080 368 : tree data_ref = fold_build2 (MEM_REF, vectype,
8081 : unshare_expr (dataref_ptr),
8082 : dataref_offset);
8083 368 : vect_copy_ref_info (data_ref, DR_REF (dr_info->dr));
8084 368 : g = gimple_build_assign (data_ref, new_temp);
8085 368 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8086 : }
8087 : }
8088 :
8089 512 : if (inscan_var_store)
8090 624 : for (unsigned j = 0; j < vec_oprnds3.length (); j++)
8091 : {
8092 368 : if (j != 0)
8093 112 : dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump);
8094 :
8095 368 : tree data_ref = fold_build2 (MEM_REF, vectype,
8096 : unshare_expr (dataref_ptr),
8097 : dataref_offset);
8098 368 : vect_copy_ref_info (data_ref, DR_REF (dr_info->dr));
8099 368 : gimple *g = gimple_build_assign (data_ref, orig);
8100 368 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
8101 : }
8102 512 : return true;
8103 512 : }
8104 :
8105 :
8106 : /* Function vectorizable_store.
8107 :
8108 : Check if STMT_INFO defines a non scalar data-ref (array/pointer/structure)
8109 : that can be vectorized.
8110 : If COST_VEC is passed, calculate costs but don't change anything,
8111 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
8112 : it, and insert it at GSI.
8113 : Return true if STMT_INFO is vectorizable in this way. */
8114 :
8115 : static bool
8116 1959155 : vectorizable_store (vec_info *vinfo,
8117 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
8118 : slp_tree slp_node,
8119 : stmt_vector_for_cost *cost_vec)
8120 : {
8121 1959155 : tree data_ref;
8122 1959155 : tree vec_oprnd = NULL_TREE;
8123 1959155 : tree elem_type;
8124 1959155 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
8125 1959155 : class loop *loop = NULL;
8126 1959155 : machine_mode vec_mode;
8127 1959155 : tree dummy;
8128 1959155 : enum vect_def_type rhs_dt = vect_unknown_def_type;
8129 1959155 : enum vect_def_type mask_dt = vect_unknown_def_type;
8130 1959155 : tree dataref_ptr = NULL_TREE;
8131 1959155 : tree dataref_offset = NULL_TREE;
8132 1959155 : gimple *ptr_incr = NULL;
8133 1959155 : int j;
8134 1959155 : stmt_vec_info first_stmt_info;
8135 1959155 : bool grouped_store;
8136 1959155 : unsigned int group_size, i;
8137 1959155 : unsigned int vec_num;
8138 1959155 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
8139 1959155 : tree aggr_type;
8140 1959155 : poly_uint64 vf;
8141 1959155 : vec_load_store_type vls_type;
8142 1959155 : tree ref_type;
8143 :
8144 1959155 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
8145 : return false;
8146 :
8147 1959155 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
8148 194522 : && cost_vec)
8149 : return false;
8150 :
8151 : /* Is vectorizable store? */
8152 :
8153 1764633 : tree mask_vectype = NULL_TREE;
8154 1764633 : slp_tree mask_node = NULL;
8155 1764633 : if (gassign *assign = dyn_cast <gassign *> (stmt_info->stmt))
8156 : {
8157 1698205 : tree scalar_dest = gimple_assign_lhs (assign);
8158 1698205 : if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR
8159 1698205 : && is_pattern_stmt_p (stmt_info))
8160 1393 : scalar_dest = TREE_OPERAND (scalar_dest, 0);
8161 1698205 : if (TREE_CODE (scalar_dest) != ARRAY_REF
8162 1698205 : && TREE_CODE (scalar_dest) != BIT_FIELD_REF
8163 : && TREE_CODE (scalar_dest) != INDIRECT_REF
8164 : && TREE_CODE (scalar_dest) != COMPONENT_REF
8165 : && TREE_CODE (scalar_dest) != IMAGPART_EXPR
8166 : && TREE_CODE (scalar_dest) != REALPART_EXPR
8167 : && TREE_CODE (scalar_dest) != MEM_REF)
8168 : return false;
8169 : }
8170 : else
8171 : {
8172 651259 : gcall *call = dyn_cast <gcall *> (stmt_info->stmt);
8173 8652 : if (!call || !gimple_call_internal_p (call))
8174 : return false;
8175 :
8176 4779 : internal_fn ifn = gimple_call_internal_fn (call);
8177 4779 : if (!internal_store_fn_p (ifn))
8178 : return false;
8179 :
8180 1470 : int mask_index = internal_fn_mask_index (ifn);
8181 1470 : if (mask_index >= 0)
8182 1470 : mask_index = vect_slp_child_index_for_operand
8183 1470 : (call, mask_index, STMT_VINFO_GATHER_SCATTER_P (stmt_info));
8184 1470 : if (mask_index >= 0
8185 1470 : && !vect_check_scalar_mask (vinfo, slp_node, mask_index,
8186 : &mask_node, &mask_dt,
8187 : &mask_vectype))
8188 : return false;
8189 : }
8190 :
8191 1316580 : tree vectype = SLP_TREE_VECTYPE (slp_node), rhs_vectype = NULL_TREE;
8192 1316580 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
8193 :
8194 1316580 : if (loop_vinfo)
8195 : {
8196 188080 : loop = LOOP_VINFO_LOOP (loop_vinfo);
8197 188080 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
8198 : }
8199 : else
8200 : vf = 1;
8201 1316580 : vec_num = vect_get_num_copies (vinfo, slp_node);
8202 :
8203 : /* FORNOW. This restriction should be relaxed. */
8204 1316580 : if (loop
8205 1316805 : && nested_in_vect_loop_p (loop, stmt_info)
8206 1316813 : && vec_num > 1)
8207 : {
8208 8 : if (dump_enabled_p ())
8209 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8210 : "multiple types in nested loop.\n");
8211 8 : return false;
8212 : }
8213 :
8214 1316572 : slp_tree op_node;
8215 1316572 : if (!vect_check_store_rhs (vinfo, stmt_info, slp_node,
8216 : &op_node, &rhs_dt, &rhs_vectype, &vls_type))
8217 : return false;
8218 :
8219 1316548 : elem_type = TREE_TYPE (vectype);
8220 1316548 : vec_mode = TYPE_MODE (vectype);
8221 :
8222 1316548 : if (!STMT_VINFO_DATA_REF (stmt_info))
8223 : return false;
8224 :
8225 1316548 : vect_load_store_data _ls_data{};
8226 1316548 : vect_load_store_data &ls = slp_node->get_data (_ls_data);
8227 1316548 : if (cost_vec
8228 1316548 : && !get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask_node,
8229 : vls_type, &_ls_data))
8230 : return false;
8231 : /* Temporary aliases to analysis data, should not be modified through
8232 : these. */
8233 1316008 : const vect_memory_access_type memory_access_type = ls.memory_access_type;
8234 1316008 : const dr_alignment_support alignment_support_scheme
8235 : = ls.alignment_support_scheme;
8236 1316008 : const int misalignment = ls.misalignment;
8237 1316008 : const poly_int64 poffset = ls.poffset;
8238 :
8239 1316008 : if (slp_node->ldst_lanes
8240 0 : && memory_access_type != VMAT_LOAD_STORE_LANES)
8241 : {
8242 0 : if (dump_enabled_p ())
8243 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8244 : "discovered store-lane but cannot use it.\n");
8245 0 : return false;
8246 : }
8247 :
8248 1316008 : if (mask_node)
8249 : {
8250 1380 : if (memory_access_type == VMAT_CONTIGUOUS)
8251 : {
8252 459 : if (!VECTOR_MODE_P (vec_mode)
8253 2276 : || !can_vec_mask_load_store_p (vec_mode,
8254 1138 : TYPE_MODE (mask_vectype), false))
8255 18 : return false;
8256 : }
8257 242 : else if (memory_access_type != VMAT_LOAD_STORE_LANES
8258 242 : && (!mat_gather_scatter_p (memory_access_type)
8259 218 : || (memory_access_type == VMAT_GATHER_SCATTER_LEGACY
8260 154 : && !VECTOR_BOOLEAN_TYPE_P (mask_vectype))))
8261 : {
8262 24 : if (dump_enabled_p ())
8263 24 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8264 : "unsupported access type for masked store.\n");
8265 24 : return false;
8266 : }
8267 218 : else if (memory_access_type == VMAT_GATHER_SCATTER_EMULATED)
8268 : {
8269 64 : if (dump_enabled_p ())
8270 24 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8271 : "unsupported masked emulated scatter.\n");
8272 64 : return false;
8273 : }
8274 : }
8275 : else
8276 : {
8277 : /* FORNOW. In some cases can vectorize even if data-type not supported
8278 : (e.g. - array initialization with 0). */
8279 1314628 : if (!can_implement_p (mov_optab, vec_mode))
8280 : return false;
8281 : }
8282 :
8283 1315902 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL;
8284 1315902 : grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info)
8285 2462069 : && !mat_gather_scatter_p (memory_access_type));
8286 1146167 : if (grouped_store)
8287 : {
8288 1146167 : first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
8289 1146167 : first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info);
8290 1146167 : group_size = DR_GROUP_SIZE (first_stmt_info);
8291 : }
8292 : else
8293 : {
8294 1315902 : first_stmt_info = stmt_info;
8295 1315902 : first_dr_info = dr_info;
8296 : group_size = 1;
8297 : }
8298 :
8299 1315902 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) > 1 && cost_vec)
8300 : {
8301 1076 : if (!check_scan_store (vinfo, stmt_info, vectype, rhs_dt, slp_node,
8302 : mask_node, memory_access_type))
8303 : return false;
8304 : }
8305 :
8306 2631036 : bool costing_p = cost_vec;
8307 1315134 : if (costing_p) /* transformation not required. */
8308 : {
8309 772215 : if (loop_vinfo
8310 124324 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
8311 57355 : check_load_store_for_partial_vectors (loop_vinfo, vectype, slp_node,
8312 : vls_type, group_size, &ls,
8313 : mask_node);
8314 :
8315 772215 : if (!vect_maybe_update_slp_op_vectype (op_node, vectype)
8316 772215 : || (mask_node
8317 723 : && !vect_maybe_update_slp_op_vectype (mask_node,
8318 : mask_vectype)))
8319 : {
8320 0 : if (dump_enabled_p ())
8321 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8322 : "incompatible vector types for invariants\n");
8323 0 : return false;
8324 : }
8325 :
8326 772215 : if (dump_enabled_p ()
8327 : && memory_access_type != VMAT_ELEMENTWISE
8328 14643 : && memory_access_type != VMAT_STRIDED_SLP
8329 13970 : && memory_access_type != VMAT_INVARIANT
8330 786185 : && alignment_support_scheme != dr_aligned)
8331 4834 : dump_printf_loc (MSG_NOTE, vect_location,
8332 : "Vectorizing an unaligned access.\n");
8333 :
8334 772215 : SLP_TREE_TYPE (slp_node) = store_vec_info_type;
8335 772215 : slp_node->data = new vect_load_store_data (std::move (ls));
8336 : }
8337 :
8338 : /* Transform. */
8339 :
8340 1315902 : ensure_base_align (dr_info);
8341 :
8342 1315902 : if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) >= 3)
8343 : {
8344 1024 : gcc_assert (memory_access_type == VMAT_CONTIGUOUS);
8345 1024 : gcc_assert (SLP_TREE_LANES (slp_node) == 1);
8346 1024 : if (costing_p)
8347 : {
8348 512 : unsigned int inside_cost = 0, prologue_cost = 0;
8349 512 : if (vls_type == VLS_STORE_INVARIANT)
8350 0 : prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec,
8351 : slp_node, 0, vect_prologue);
8352 512 : vect_get_store_cost (vinfo, stmt_info, slp_node, 1,
8353 : alignment_support_scheme, misalignment,
8354 : &inside_cost, cost_vec);
8355 :
8356 512 : if (dump_enabled_p ())
8357 492 : dump_printf_loc (MSG_NOTE, vect_location,
8358 : "vect_model_store_cost: inside_cost = %d, "
8359 : "prologue_cost = %d .\n",
8360 : inside_cost, prologue_cost);
8361 :
8362 512 : return true;
8363 : }
8364 512 : return vectorizable_scan_store (vinfo, stmt_info, slp_node, gsi);
8365 : }
8366 :
8367 : /* FORNOW */
8368 1314878 : gcc_assert (!grouped_store
8369 : || !loop
8370 : || !nested_in_vect_loop_p (loop, stmt_info));
8371 :
8372 1314878 : grouped_store = false;
8373 1314878 : first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0];
8374 1314878 : gcc_assert (!STMT_VINFO_GROUPED_ACCESS (first_stmt_info)
8375 : || (DR_GROUP_FIRST_ELEMENT (first_stmt_info) == first_stmt_info));
8376 1314878 : first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info);
8377 :
8378 1314878 : ref_type = get_group_alias_ptr_type (first_stmt_info);
8379 :
8380 1314878 : if (!costing_p && dump_enabled_p ())
8381 12238 : dump_printf_loc (MSG_NOTE, vect_location, "transform store.\n");
8382 :
8383 1314878 : if (memory_access_type == VMAT_ELEMENTWISE
8384 1314878 : || memory_access_type == VMAT_STRIDED_SLP)
8385 : {
8386 28124 : unsigned inside_cost = 0, prologue_cost = 0;
8387 28124 : gimple_stmt_iterator incr_gsi;
8388 28124 : bool insert_after;
8389 28124 : tree offvar = NULL_TREE;
8390 28124 : tree ivstep;
8391 28124 : tree running_off;
8392 28124 : tree stride_base, stride_step, alias_off;
8393 28124 : tree vec_oprnd = NULL_TREE;
8394 28124 : tree dr_offset;
8395 : /* Checked by get_load_store_type. */
8396 28124 : unsigned int const_nunits = nunits.to_constant ();
8397 :
8398 28124 : gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
8399 28124 : gcc_assert (!nested_in_vect_loop_p (loop, stmt_info));
8400 :
8401 28124 : dr_offset = get_dr_vinfo_offset (vinfo, first_dr_info);
8402 28124 : stride_base
8403 28124 : = fold_build_pointer_plus
8404 : (DR_BASE_ADDRESS (first_dr_info->dr),
8405 : size_binop (PLUS_EXPR,
8406 : convert_to_ptrofftype (dr_offset),
8407 : convert_to_ptrofftype (DR_INIT (first_dr_info->dr))));
8408 28124 : stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr));
8409 :
8410 : /* For a store with loop-invariant (but other than power-of-2)
8411 : stride (i.e. not a grouped access) like so:
8412 :
8413 : for (i = 0; i < n; i += stride)
8414 : array[i] = ...;
8415 :
8416 : we generate a new induction variable and new stores from
8417 : the components of the (vectorized) rhs:
8418 :
8419 : for (j = 0; ; j += VF*stride)
8420 : vectemp = ...;
8421 : tmp1 = vectemp[0];
8422 : array[j] = tmp1;
8423 : tmp2 = vectemp[1];
8424 : array[j + stride] = tmp2;
8425 : ...
8426 : */
8427 :
8428 : /* ??? Modify local copies of alignment_support_scheme and
8429 : misalignment, but this part of analysis should be done
8430 : earlier and remembered, likewise the chosen load mode. */
8431 28124 : const dr_alignment_support tem = alignment_support_scheme;
8432 28124 : dr_alignment_support alignment_support_scheme = tem;
8433 28124 : const int tem2 = misalignment;
8434 28124 : int misalignment = tem2;
8435 :
8436 28124 : unsigned nstores = const_nunits;
8437 28124 : unsigned lnel = 1;
8438 28124 : tree ltype = elem_type;
8439 28124 : tree lvectype = vectype;
8440 28124 : HOST_WIDE_INT n = gcd (group_size, const_nunits);
8441 28124 : if (n == const_nunits)
8442 : {
8443 2426 : int mis_align = dr_misalignment (first_dr_info, vectype);
8444 : /* With VF > 1 we advance the DR by step, if that is constant
8445 : and only aligned when performed VF times, DR alignment
8446 : analysis can analyze this as aligned since it assumes
8447 : contiguous accesses. But that is not how we code generate
8448 : here, so adjust for this. */
8449 2426 : if (maybe_gt (vf, 1u)
8450 3736 : && !multiple_p (DR_STEP_ALIGNMENT (first_dr_info->dr),
8451 3510 : DR_TARGET_ALIGNMENT (first_dr_info)))
8452 226 : mis_align = -1;
8453 2426 : dr_alignment_support dr_align
8454 2426 : = vect_supportable_dr_alignment (vinfo, dr_info, vectype,
8455 : mis_align);
8456 2426 : if (dr_align == dr_aligned
8457 2426 : || dr_align == dr_unaligned_supported)
8458 : {
8459 28124 : nstores = 1;
8460 28124 : lnel = const_nunits;
8461 28124 : ltype = vectype;
8462 28124 : lvectype = vectype;
8463 28124 : alignment_support_scheme = dr_align;
8464 28124 : misalignment = mis_align;
8465 : }
8466 : }
8467 25698 : else if (n > 1)
8468 : {
8469 1932 : nstores = const_nunits / n;
8470 1932 : lnel = n;
8471 1932 : ltype = build_vector_type (elem_type, n);
8472 1932 : lvectype = vectype;
8473 1932 : int mis_align = dr_misalignment (first_dr_info, ltype);
8474 1932 : if (maybe_gt (vf, 1u)
8475 3864 : && !multiple_p (DR_STEP_ALIGNMENT (first_dr_info->dr),
8476 3048 : DR_TARGET_ALIGNMENT (first_dr_info)))
8477 816 : mis_align = -1;
8478 1932 : dr_alignment_support dr_align
8479 1932 : = vect_supportable_dr_alignment (vinfo, dr_info, ltype,
8480 : mis_align);
8481 1932 : alignment_support_scheme = dr_align;
8482 1932 : misalignment = mis_align;
8483 :
8484 : /* First check if vec_extract optab doesn't support extraction
8485 : of vector elts directly. */
8486 1932 : scalar_mode elmode = SCALAR_TYPE_MODE (elem_type);
8487 1932 : machine_mode vmode;
8488 3864 : if (!VECTOR_MODE_P (TYPE_MODE (vectype))
8489 2160 : || !related_vector_mode (TYPE_MODE (vectype), elmode,
8490 1932 : n).exists (&vmode)
8491 1756 : || (convert_optab_handler (vec_extract_optab,
8492 1756 : TYPE_MODE (vectype), vmode)
8493 : == CODE_FOR_nothing)
8494 1932 : || !(dr_align == dr_aligned
8495 228 : || dr_align == dr_unaligned_supported))
8496 : {
8497 : /* Try to avoid emitting an extract of vector elements
8498 : by performing the extracts using an integer type of the
8499 : same size, extracting from a vector of those and then
8500 : re-interpreting it as the original vector type if
8501 : supported. */
8502 1704 : unsigned lsize = n * GET_MODE_BITSIZE (elmode);
8503 1704 : unsigned int lnunits = const_nunits / n;
8504 : /* If we can't construct such a vector fall back to
8505 : element extracts from the original vector type and
8506 : element size stores. */
8507 1704 : if (int_mode_for_size (lsize, 0).exists (&elmode)
8508 1704 : && VECTOR_MODE_P (TYPE_MODE (vectype))
8509 1704 : && related_vector_mode (TYPE_MODE (vectype), elmode,
8510 1704 : lnunits).exists (&vmode)
8511 1670 : && (convert_optab_handler (vec_extract_optab,
8512 : vmode, elmode)
8513 : != CODE_FOR_nothing))
8514 : {
8515 1670 : nstores = lnunits;
8516 1670 : lnel = n;
8517 1670 : ltype = build_nonstandard_integer_type (lsize, 1);
8518 1670 : lvectype = build_vector_type (ltype, nstores);
8519 : }
8520 : /* Else fall back to vector extraction anyway.
8521 : Fewer stores are more important than avoiding spilling
8522 : of the vector we extract from. Compared to the
8523 : construction case in vectorizable_load no store-forwarding
8524 : issue exists here for reasonable archs. But only
8525 : if the store is supported. */
8526 34 : else if (!(dr_align == dr_aligned
8527 34 : || dr_align == dr_unaligned_supported))
8528 : {
8529 : nstores = const_nunits;
8530 : lnel = 1;
8531 : ltype = elem_type;
8532 : lvectype = vectype;
8533 : }
8534 : }
8535 : }
8536 28124 : unsigned align;
8537 28124 : if (alignment_support_scheme == dr_aligned)
8538 930 : align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info));
8539 : else
8540 27194 : align = dr_alignment (vect_dr_behavior (vinfo, first_dr_info));
8541 : /* Alignment is at most the access size if we do multiple stores. */
8542 28124 : if (nstores > 1)
8543 25698 : align = MIN (tree_to_uhwi (TYPE_SIZE_UNIT (ltype)), align);
8544 28124 : ltype = build_aligned_type (ltype, align * BITS_PER_UNIT);
8545 28124 : int ncopies = vec_num;
8546 :
8547 28124 : if (!costing_p)
8548 : {
8549 3517 : ivstep = stride_step;
8550 3517 : ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep,
8551 : build_int_cst (TREE_TYPE (ivstep), vf));
8552 :
8553 3517 : standard_iv_increment_position (loop, &incr_gsi, &insert_after);
8554 :
8555 3517 : stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
8556 3517 : ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
8557 3517 : create_iv (stride_base, PLUS_EXPR, ivstep, NULL, loop, &incr_gsi,
8558 : insert_after, &offvar, NULL);
8559 :
8560 3517 : stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step);
8561 : }
8562 :
8563 28124 : alias_off = build_int_cst (ref_type, 0);
8564 28124 : auto_vec<tree> vec_oprnds;
8565 : /* For costing some adjacent vector stores, we'd like to cost with
8566 : the total number of them once instead of cost each one by one. */
8567 28124 : unsigned int n_adjacent_stores = 0;
8568 28124 : running_off = offvar;
8569 28124 : if (!costing_p)
8570 3517 : vect_get_slp_defs (op_node, &vec_oprnds);
8571 28124 : unsigned int group_el = 0;
8572 28124 : unsigned HOST_WIDE_INT elsz
8573 28124 : = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
8574 66679 : for (j = 0; j < ncopies; j++)
8575 : {
8576 38555 : if (!costing_p)
8577 : {
8578 5416 : vec_oprnd = vec_oprnds[j];
8579 : /* Pun the vector to extract from if necessary. */
8580 5416 : if (lvectype != vectype)
8581 : {
8582 1302 : tree tem = make_ssa_name (lvectype);
8583 1302 : tree cvt = build1 (VIEW_CONVERT_EXPR, lvectype, vec_oprnd);
8584 1302 : gimple *pun = gimple_build_assign (tem, cvt);
8585 1302 : vect_finish_stmt_generation (vinfo, stmt_info, pun, gsi);
8586 1302 : vec_oprnd = tem;
8587 : }
8588 : }
8589 173830 : for (i = 0; i < nstores; i++)
8590 : {
8591 135275 : if (costing_p)
8592 : {
8593 118409 : n_adjacent_stores++;
8594 118409 : continue;
8595 : }
8596 16866 : tree newref, newoff;
8597 16866 : gimple *incr, *assign;
8598 16866 : tree size = TYPE_SIZE (ltype);
8599 : /* Extract the i'th component. */
8600 16866 : tree pos = fold_build2 (MULT_EXPR, bitsizetype,
8601 : bitsize_int (i), size);
8602 16866 : tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd,
8603 : size, pos);
8604 :
8605 16866 : elem = force_gimple_operand_gsi (gsi, elem, true, NULL_TREE, true,
8606 : GSI_SAME_STMT);
8607 :
8608 16866 : tree this_off = build_int_cst (TREE_TYPE (alias_off),
8609 16866 : group_el * elsz);
8610 16866 : newref = build2 (MEM_REF, ltype, running_off, this_off);
8611 16866 : vect_copy_ref_info (newref, DR_REF (first_dr_info->dr));
8612 :
8613 : /* And store it to *running_off. */
8614 16866 : assign = gimple_build_assign (newref, elem);
8615 16866 : vect_finish_stmt_generation (vinfo, stmt_info, assign, gsi);
8616 :
8617 16866 : group_el += lnel;
8618 16866 : if (group_el == group_size)
8619 : {
8620 15143 : newoff = copy_ssa_name (running_off, NULL);
8621 15143 : incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR,
8622 : running_off, stride_step);
8623 15143 : vect_finish_stmt_generation (vinfo, stmt_info, incr, gsi);
8624 :
8625 15143 : running_off = newoff;
8626 15143 : group_el = 0;
8627 : }
8628 : }
8629 : }
8630 :
8631 28124 : if (costing_p)
8632 : {
8633 24607 : if (n_adjacent_stores > 0)
8634 : {
8635 : /* Take a single lane vector type store as scalar
8636 : store to avoid ICE like 110776. */
8637 24607 : if (VECTOR_TYPE_P (ltype)
8638 24607 : && maybe_ne (TYPE_VECTOR_SUBPARTS (ltype), 1U))
8639 1197 : vect_get_store_cost (vinfo, stmt_info, slp_node,
8640 : n_adjacent_stores, alignment_support_scheme,
8641 : misalignment, &inside_cost, cost_vec);
8642 : else
8643 23410 : inside_cost
8644 23410 : += record_stmt_cost (cost_vec, n_adjacent_stores,
8645 : scalar_store, slp_node, 0, vect_body);
8646 : /* Only need vector extracting when there are more
8647 : than one stores. */
8648 24607 : if (nstores > 1)
8649 23021 : inside_cost
8650 23021 : += record_stmt_cost (cost_vec, n_adjacent_stores,
8651 : vec_to_scalar, slp_node, 0, vect_body);
8652 : }
8653 24607 : if (dump_enabled_p ())
8654 673 : dump_printf_loc (MSG_NOTE, vect_location,
8655 : "vect_model_store_cost: inside_cost = %d, "
8656 : "prologue_cost = %d .\n",
8657 : inside_cost, prologue_cost);
8658 : }
8659 :
8660 28124 : return true;
8661 28124 : }
8662 :
8663 1286754 : gcc_assert (alignment_support_scheme);
8664 1286754 : vec_loop_masks *loop_masks
8665 158254 : = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
8666 1286754 : ? &LOOP_VINFO_MASKS (loop_vinfo)
8667 11 : : NULL);
8668 11 : vec_loop_lens *loop_lens
8669 158254 : = (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)
8670 : ? &LOOP_VINFO_LENS (loop_vinfo)
8671 0 : : NULL);
8672 :
8673 : /* The vect_transform_stmt and vect_analyze_stmt will go here but there
8674 : are some difference here. We cannot enable both the lens and masks
8675 : during transform but it is allowed during analysis.
8676 : Shouldn't go with length-based approach if fully masked. */
8677 1286754 : if (cost_vec == NULL)
8678 : /* The cost_vec is NULL during transfrom. */
8679 539658 : gcc_assert ((!loop_lens || !loop_masks));
8680 :
8681 : /* Targets with store-lane instructions must not require explicit
8682 : realignment. vect_supportable_dr_alignment always returns either
8683 : dr_aligned or dr_unaligned_supported for masked operations. */
8684 1286754 : gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES
8685 : && !mask_node
8686 : && !loop_masks)
8687 : || alignment_support_scheme == dr_aligned
8688 : || alignment_support_scheme == dr_unaligned_supported);
8689 :
8690 1286754 : tree offset = NULL_TREE;
8691 1286754 : if (!known_eq (poffset, 0))
8692 4066 : offset = size_int (poffset);
8693 :
8694 1286754 : tree bump;
8695 1286754 : tree vec_offset = NULL_TREE;
8696 1286754 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
8697 : {
8698 1366 : aggr_type = NULL_TREE;
8699 1366 : bump = NULL_TREE;
8700 : }
8701 1285388 : else if (mat_gather_scatter_p (memory_access_type))
8702 : {
8703 0 : aggr_type = elem_type;
8704 0 : if (!costing_p)
8705 : {
8706 0 : tree vtype = ls.ls_type ? ls.ls_type : vectype;
8707 0 : vect_get_strided_load_store_ops (stmt_info, slp_node, vtype,
8708 : ls.strided_offset_vectype,
8709 : loop_vinfo, gsi,
8710 : &bump, &vec_offset, loop_lens);
8711 : }
8712 : }
8713 : else
8714 : {
8715 1285388 : if (memory_access_type == VMAT_LOAD_STORE_LANES)
8716 0 : aggr_type = build_array_type_nelts (elem_type, group_size * nunits);
8717 : else
8718 : aggr_type = vectype;
8719 1285388 : if (!costing_p)
8720 539188 : bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, aggr_type,
8721 : memory_access_type, loop_lens);
8722 : }
8723 :
8724 1286754 : if (loop_vinfo && mask_node && !costing_p)
8725 550 : LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true;
8726 :
8727 : /* In case the vectorization factor (VF) is bigger than the number
8728 : of elements that we can fit in a vectype (nunits), we have to generate
8729 : more than one vector stmt - i.e - we need to "unroll" the
8730 : vector stmt by a factor VF/nunits. */
8731 :
8732 1286754 : auto_vec<tree> dr_chain (group_size);
8733 1286754 : auto_vec<tree> vec_masks;
8734 1286754 : tree vec_mask = NULL;
8735 1286754 : auto_delete_vec<auto_vec<tree>> gvec_oprnds (group_size);
8736 5842279 : for (i = 0; i < group_size; i++)
8737 3268771 : gvec_oprnds.quick_push (new auto_vec<tree> ());
8738 :
8739 1286754 : if (memory_access_type == VMAT_LOAD_STORE_LANES)
8740 : {
8741 0 : const internal_fn lanes_ifn = ls.lanes_ifn;
8742 :
8743 0 : if (costing_p)
8744 : /* Update all incoming store operand nodes, the general handling
8745 : above only handles the mask and the first store operand node. */
8746 0 : for (slp_tree child : SLP_TREE_CHILDREN (slp_node))
8747 0 : if (child != mask_node
8748 0 : && !vect_maybe_update_slp_op_vectype (child, vectype))
8749 : {
8750 0 : if (dump_enabled_p ())
8751 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8752 : "incompatible vector types for invariants\n");
8753 0 : return false;
8754 : }
8755 0 : unsigned inside_cost = 0, prologue_cost = 0;
8756 : /* For costing some adjacent vector stores, we'd like to cost with
8757 : the total number of them once instead of cost each one by one. */
8758 0 : unsigned int n_adjacent_stores = 0;
8759 0 : int ncopies = vec_num / group_size;
8760 0 : for (j = 0; j < ncopies; j++)
8761 : {
8762 0 : if (j == 0)
8763 : {
8764 0 : if (!costing_p)
8765 : {
8766 0 : if (mask_node)
8767 : {
8768 0 : vect_get_slp_defs (mask_node, &vec_masks);
8769 0 : vec_mask = vec_masks[0];
8770 : }
8771 0 : dataref_ptr
8772 0 : = vect_create_data_ref_ptr (vinfo, first_stmt_info,
8773 : aggr_type, NULL, offset, &dummy,
8774 : gsi, &ptr_incr, false, bump);
8775 : }
8776 : }
8777 0 : else if (!costing_p)
8778 : {
8779 0 : gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
8780 0 : if (mask_node)
8781 0 : vec_mask = vec_masks[j];
8782 0 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
8783 : stmt_info, bump);
8784 : }
8785 :
8786 0 : if (costing_p)
8787 : {
8788 0 : n_adjacent_stores += group_size;
8789 0 : continue;
8790 : }
8791 :
8792 : /* Get an array into which we can store the individual vectors. */
8793 0 : tree vec_array = create_vector_array (vectype, group_size);
8794 :
8795 : /* Invalidate the current contents of VEC_ARRAY. This should
8796 : become an RTL clobber too, which prevents the vector registers
8797 : from being upward-exposed. */
8798 0 : vect_clobber_variable (vinfo, stmt_info, gsi, vec_array);
8799 :
8800 : /* Store the individual vectors into the array. */
8801 0 : for (i = 0; i < group_size; i++)
8802 : {
8803 0 : slp_tree child;
8804 0 : if (i == 0 || !mask_node)
8805 0 : child = SLP_TREE_CHILDREN (slp_node)[i];
8806 : else
8807 0 : child = SLP_TREE_CHILDREN (slp_node)[i + 1];
8808 0 : vec_oprnd = SLP_TREE_VEC_DEFS (child)[j];
8809 0 : write_vector_array (vinfo, stmt_info, gsi, vec_oprnd, vec_array,
8810 : i);
8811 : }
8812 :
8813 0 : tree final_mask = NULL;
8814 0 : tree final_len = NULL;
8815 0 : tree bias = NULL;
8816 0 : if (loop_masks)
8817 0 : final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
8818 : ncopies, vectype, j);
8819 0 : if (vec_mask)
8820 0 : final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
8821 : vec_mask, gsi);
8822 :
8823 0 : if (lanes_ifn == IFN_MASK_LEN_STORE_LANES)
8824 : {
8825 0 : if (loop_lens)
8826 0 : final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
8827 : ncopies, vectype, j, 1, true);
8828 : else
8829 0 : final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
8830 0 : signed char biasval
8831 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
8832 0 : bias = build_int_cst (intQI_type_node, biasval);
8833 0 : if (!final_mask)
8834 : {
8835 0 : mask_vectype = truth_type_for (vectype);
8836 0 : final_mask = build_minus_one_cst (mask_vectype);
8837 : }
8838 : }
8839 :
8840 0 : gcall *call;
8841 0 : if (final_len && final_mask)
8842 : {
8843 : /* Emit:
8844 : MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
8845 : LEN, BIAS, VEC_ARRAY). */
8846 0 : unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
8847 0 : tree alias_ptr = build_int_cst (ref_type, align);
8848 0 : call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6,
8849 : dataref_ptr, alias_ptr,
8850 : final_mask, final_len, bias,
8851 : vec_array);
8852 : }
8853 0 : else if (final_mask)
8854 : {
8855 : /* Emit:
8856 : MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
8857 : VEC_ARRAY). */
8858 0 : unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
8859 0 : tree alias_ptr = build_int_cst (ref_type, align);
8860 0 : call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4,
8861 : dataref_ptr, alias_ptr,
8862 : final_mask, vec_array);
8863 : }
8864 : else
8865 : {
8866 : /* Emit:
8867 : MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */
8868 0 : data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type);
8869 0 : call = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array);
8870 0 : gimple_call_set_lhs (call, data_ref);
8871 : }
8872 0 : gimple_call_set_nothrow (call, true);
8873 0 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
8874 :
8875 : /* Record that VEC_ARRAY is now dead. */
8876 0 : vect_clobber_variable (vinfo, stmt_info, gsi, vec_array);
8877 : }
8878 :
8879 0 : if (costing_p)
8880 : {
8881 0 : if (n_adjacent_stores > 0)
8882 0 : vect_get_store_cost (vinfo, stmt_info, slp_node, n_adjacent_stores,
8883 : alignment_support_scheme, misalignment,
8884 : &inside_cost, cost_vec);
8885 0 : if (dump_enabled_p ())
8886 0 : dump_printf_loc (MSG_NOTE, vect_location,
8887 : "vect_model_store_cost: inside_cost = %d, "
8888 : "prologue_cost = %d .\n",
8889 : inside_cost, prologue_cost);
8890 : }
8891 :
8892 0 : return true;
8893 : }
8894 :
8895 1286754 : if (mat_gather_scatter_p (memory_access_type))
8896 : {
8897 1366 : gcc_assert (!grouped_store || ls.ls_type);
8898 1366 : if (ls.ls_type)
8899 0 : vectype = ls.ls_type;
8900 1366 : auto_vec<tree> vec_offsets;
8901 1366 : unsigned int inside_cost = 0, prologue_cost = 0;
8902 1366 : int num_stmts = vec_num;
8903 3124 : for (j = 0; j < num_stmts; j++)
8904 : {
8905 1758 : gimple *new_stmt;
8906 1758 : if (j == 0)
8907 : {
8908 1366 : if (costing_p && vls_type == VLS_STORE_INVARIANT)
8909 210 : prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec,
8910 : slp_node, 0, vect_prologue);
8911 : else if (!costing_p)
8912 : {
8913 : /* Since the store is not grouped, DR_GROUP_SIZE is 1, and
8914 : DR_CHAIN is of size 1. */
8915 470 : gcc_assert (group_size == 1);
8916 470 : vect_get_slp_defs (op_node, gvec_oprnds[0]);
8917 470 : if (mask_node)
8918 70 : vect_get_slp_defs (mask_node, &vec_masks);
8919 :
8920 470 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
8921 470 : vect_get_gather_scatter_ops (loop, slp_node,
8922 : &dataref_ptr, &vec_offsets);
8923 : else
8924 0 : dataref_ptr
8925 0 : = vect_create_data_ref_ptr (vinfo, first_stmt_info,
8926 : aggr_type, NULL, offset,
8927 : &dummy, gsi, &ptr_incr, false,
8928 : bump);
8929 : }
8930 : }
8931 392 : else if (!costing_p)
8932 : {
8933 34 : gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
8934 34 : if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info))
8935 0 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr,
8936 : gsi, stmt_info, bump);
8937 : }
8938 :
8939 2472 : new_stmt = NULL;
8940 714 : if (!costing_p)
8941 : {
8942 504 : vec_oprnd = (*gvec_oprnds[0])[j];
8943 504 : if (mask_node)
8944 90 : vec_mask = vec_masks[j];
8945 : /* We should have caught mismatched types earlier. */
8946 504 : gcc_assert (ls.ls_type
8947 : || useless_type_conversion_p
8948 : (vectype, TREE_TYPE (vec_oprnd)));
8949 : }
8950 504 : tree final_mask = NULL_TREE;
8951 2262 : tree final_len = NULL_TREE;
8952 2262 : tree bias = NULL_TREE;
8953 504 : if (!costing_p)
8954 : {
8955 504 : if (loop_masks)
8956 0 : final_mask = vect_get_loop_mask (loop_vinfo, gsi,
8957 : loop_masks, num_stmts,
8958 : vectype, j);
8959 504 : if (vec_mask)
8960 90 : final_mask = prepare_vec_mask (loop_vinfo, mask_vectype,
8961 : final_mask, vec_mask, gsi);
8962 : }
8963 :
8964 1758 : unsigned align = get_object_alignment (DR_REF (first_dr_info->dr));
8965 1758 : tree alias_align_ptr = build_int_cst (ref_type, align);
8966 1758 : if (memory_access_type == VMAT_GATHER_SCATTER_IFN)
8967 : {
8968 0 : if (costing_p)
8969 : {
8970 0 : if (ls.supported_offset_vectype)
8971 0 : inside_cost
8972 0 : += record_stmt_cost (cost_vec, 1, vector_stmt,
8973 : slp_node, 0, vect_body);
8974 0 : if (ls.supported_scale)
8975 0 : inside_cost
8976 0 : += record_stmt_cost (cost_vec, 1, vector_stmt,
8977 : slp_node, 0, vect_body);
8978 :
8979 0 : unsigned int cnunits = vect_nunits_for_cost (vectype);
8980 0 : inside_cost
8981 0 : += record_stmt_cost (cost_vec, cnunits, scalar_store,
8982 : slp_node, 0, vect_body);
8983 1758 : continue;
8984 0 : }
8985 :
8986 0 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
8987 0 : vec_offset = vec_offsets[j];
8988 :
8989 0 : tree scale = size_int (SLP_TREE_GS_SCALE (slp_node));
8990 0 : bool strided = !VECTOR_TYPE_P (TREE_TYPE (vec_offset));
8991 :
8992 : /* Perform the offset conversion and scaling if necessary. */
8993 0 : if (!strided
8994 0 : && (ls.supported_offset_vectype || ls.supported_scale))
8995 : {
8996 0 : gimple_seq stmts = NULL;
8997 0 : if (ls.supported_offset_vectype)
8998 0 : vec_offset = gimple_convert
8999 0 : (&stmts, ls.supported_offset_vectype, vec_offset);
9000 0 : if (ls.supported_scale)
9001 : {
9002 : /* Only scale the vec_offset if we haven't already. */
9003 0 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
9004 0 : || j == 0)
9005 : {
9006 0 : tree mult_cst = build_int_cst
9007 0 : (TREE_TYPE (TREE_TYPE (vec_offset)),
9008 0 : SLP_TREE_GS_SCALE (slp_node) / ls.supported_scale);
9009 0 : tree mult = build_vector_from_val
9010 0 : (TREE_TYPE (vec_offset), mult_cst);
9011 0 : vec_offset = gimple_build
9012 0 : (&stmts, MULT_EXPR, TREE_TYPE (vec_offset),
9013 : vec_offset, mult);
9014 : }
9015 0 : scale = size_int (ls.supported_scale);
9016 : }
9017 0 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
9018 : }
9019 :
9020 0 : if (ls.gs.ifn == IFN_MASK_LEN_SCATTER_STORE)
9021 : {
9022 0 : if (loop_lens)
9023 0 : final_len = vect_get_loop_len (loop_vinfo, gsi,
9024 : loop_lens, num_stmts,
9025 : vectype, j, 1, true);
9026 : else
9027 0 : final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
9028 :
9029 0 : signed char biasval
9030 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
9031 0 : bias = build_int_cst (intQI_type_node, biasval);
9032 0 : if (!final_mask)
9033 : {
9034 0 : mask_vectype = truth_type_for (vectype);
9035 0 : final_mask = build_minus_one_cst (mask_vectype);
9036 : }
9037 : }
9038 :
9039 0 : if (ls.ls_type)
9040 : {
9041 0 : gimple *conv_stmt
9042 0 : = gimple_build_assign (make_ssa_name (vectype),
9043 : VIEW_CONVERT_EXPR,
9044 : build1 (VIEW_CONVERT_EXPR, vectype,
9045 : vec_oprnd));
9046 0 : vect_finish_stmt_generation (vinfo, stmt_info, conv_stmt,
9047 : gsi);
9048 0 : vec_oprnd = gimple_get_lhs (conv_stmt);
9049 : }
9050 :
9051 0 : gcall *call;
9052 0 : if (final_len && final_mask)
9053 : {
9054 0 : if (VECTOR_TYPE_P (TREE_TYPE (vec_offset)))
9055 0 : call = gimple_build_call_internal (
9056 : IFN_MASK_LEN_SCATTER_STORE, 8, dataref_ptr,
9057 : alias_align_ptr,
9058 : vec_offset, scale, vec_oprnd, final_mask, final_len,
9059 : bias);
9060 : else
9061 : /* Non-vector offset indicates that prefer to take
9062 : MASK_LEN_STRIDED_STORE instead of the
9063 : IFN_MASK_SCATTER_STORE with direct stride arg.
9064 : Similar to the gather case we have checked the
9065 : alignment for a scatter already and assume
9066 : that the strided store has the same requirements. */
9067 0 : call = gimple_build_call_internal (
9068 : IFN_MASK_LEN_STRIDED_STORE, 6, dataref_ptr,
9069 : vec_offset, vec_oprnd, final_mask, final_len, bias);
9070 : }
9071 0 : else if (final_mask)
9072 0 : call = gimple_build_call_internal
9073 0 : (IFN_MASK_SCATTER_STORE, 6, dataref_ptr,
9074 : alias_align_ptr,
9075 : vec_offset, scale, vec_oprnd, final_mask);
9076 : else
9077 0 : call = gimple_build_call_internal (IFN_SCATTER_STORE, 5,
9078 : dataref_ptr,
9079 : alias_align_ptr,
9080 : vec_offset,
9081 : scale, vec_oprnd);
9082 0 : gimple_call_set_nothrow (call, true);
9083 0 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
9084 0 : new_stmt = call;
9085 : }
9086 1758 : else if (memory_access_type == VMAT_GATHER_SCATTER_LEGACY)
9087 : {
9088 : /* The builtin decls path for scatter is legacy, x86 only. */
9089 310 : gcc_assert (nunits.is_constant ()
9090 : && (!final_mask
9091 : || SCALAR_INT_MODE_P
9092 : (TYPE_MODE (TREE_TYPE (final_mask)))));
9093 310 : if (costing_p)
9094 : {
9095 179 : unsigned int cnunits = vect_nunits_for_cost (vectype);
9096 179 : inside_cost
9097 179 : += record_stmt_cost (cost_vec, cnunits, scalar_store,
9098 : slp_node, 0, vect_body);
9099 179 : continue;
9100 179 : }
9101 :
9102 131 : tree offset_vectype = TREE_TYPE (vec_offsets[0]);
9103 131 : poly_uint64 offset_nunits
9104 131 : = TYPE_VECTOR_SUBPARTS (offset_vectype);
9105 131 : if (known_eq (nunits, offset_nunits))
9106 : {
9107 55 : new_stmt = vect_build_one_scatter_store_call
9108 110 : (vinfo, stmt_info, slp_node, gsi,
9109 55 : ls.gs.decl, dataref_ptr, vec_offsets[j],
9110 : vec_oprnd, final_mask);
9111 55 : vect_finish_stmt_generation (vinfo, stmt_info,
9112 : new_stmt, gsi);
9113 : }
9114 76 : else if (known_eq (nunits, offset_nunits * 2))
9115 : {
9116 : /* We have a offset vector with half the number of
9117 : lanes but the builtins will store full vectype
9118 : data from the lower lanes. */
9119 30 : new_stmt = vect_build_one_scatter_store_call
9120 60 : (vinfo, stmt_info, slp_node, gsi, ls.gs.decl,
9121 30 : dataref_ptr, vec_offsets[2 * j],
9122 : vec_oprnd, final_mask);
9123 30 : vect_finish_stmt_generation (vinfo, stmt_info,
9124 : new_stmt, gsi);
9125 30 : int count = nunits.to_constant ();
9126 30 : vec_perm_builder sel (count, count, 1);
9127 30 : sel.quick_grow (count);
9128 382 : for (int i = 0; i < count; ++i)
9129 352 : sel[i] = i | (count / 2);
9130 30 : vec_perm_indices indices (sel, 2, count);
9131 30 : tree perm_mask
9132 30 : = vect_gen_perm_mask_checked (vectype, indices);
9133 30 : new_stmt = gimple_build_assign (NULL_TREE, VEC_PERM_EXPR,
9134 : vec_oprnd, vec_oprnd,
9135 : perm_mask);
9136 30 : vec_oprnd = make_ssa_name (vectype);
9137 30 : gimple_set_lhs (new_stmt, vec_oprnd);
9138 30 : vect_finish_stmt_generation (vinfo, stmt_info,
9139 : new_stmt, gsi);
9140 30 : if (final_mask)
9141 : {
9142 20 : new_stmt = gimple_build_assign (NULL_TREE,
9143 : VEC_UNPACK_HI_EXPR,
9144 : final_mask);
9145 20 : final_mask = make_ssa_name
9146 20 : (truth_type_for (offset_vectype));
9147 20 : gimple_set_lhs (new_stmt, final_mask);
9148 20 : vect_finish_stmt_generation (vinfo, stmt_info,
9149 : new_stmt, gsi);
9150 : }
9151 :
9152 30 : new_stmt = vect_build_one_scatter_store_call
9153 60 : (vinfo, stmt_info, slp_node, gsi, ls.gs.decl,
9154 30 : dataref_ptr, vec_offsets[2 * j + 1],
9155 : vec_oprnd, final_mask);
9156 30 : vect_finish_stmt_generation (vinfo, stmt_info,
9157 : new_stmt, gsi);
9158 30 : }
9159 46 : else if (known_eq (nunits * 2, offset_nunits))
9160 : {
9161 : /* We have a offset vector with double the number of
9162 : lanes. Select the low/high part accordingly. */
9163 46 : vec_offset = vec_offsets[j / 2];
9164 46 : if (j & 1)
9165 : {
9166 23 : int count = offset_nunits.to_constant ();
9167 23 : vec_perm_builder sel (count, count, 1);
9168 23 : sel.quick_grow (count);
9169 263 : for (int i = 0; i < count; ++i)
9170 240 : sel[i] = i | (count / 2);
9171 23 : vec_perm_indices indices (sel, 2, count);
9172 23 : tree perm_mask = vect_gen_perm_mask_checked
9173 23 : (TREE_TYPE (vec_offset), indices);
9174 23 : new_stmt = gimple_build_assign (NULL_TREE,
9175 : VEC_PERM_EXPR,
9176 : vec_offset,
9177 : vec_offset,
9178 : perm_mask);
9179 23 : vec_offset = make_ssa_name (TREE_TYPE (vec_offset));
9180 23 : gimple_set_lhs (new_stmt, vec_offset);
9181 23 : vect_finish_stmt_generation (vinfo, stmt_info,
9182 : new_stmt, gsi);
9183 23 : }
9184 :
9185 46 : new_stmt = vect_build_one_scatter_store_call
9186 46 : (vinfo, stmt_info, slp_node, gsi,
9187 : ls.gs.decl, dataref_ptr, vec_offset,
9188 : vec_oprnd, final_mask);
9189 46 : vect_finish_stmt_generation (vinfo, stmt_info,
9190 : new_stmt, gsi);
9191 : }
9192 : else
9193 0 : gcc_unreachable ();
9194 : }
9195 : else
9196 : {
9197 : /* Emulated scatter. */
9198 1448 : gcc_assert (!final_mask);
9199 1448 : if (costing_p)
9200 : {
9201 1075 : unsigned int cnunits = vect_nunits_for_cost (vectype);
9202 : /* For emulated scatter N offset vector element extracts
9203 : (we assume the scalar scaling and ptr + offset add is
9204 : consumed by the load). */
9205 1075 : inside_cost
9206 1075 : += record_stmt_cost (cost_vec, cnunits, vec_to_scalar,
9207 : slp_node, 0, vect_body);
9208 : /* N scalar stores plus extracting the elements. */
9209 1075 : inside_cost
9210 1075 : += record_stmt_cost (cost_vec, cnunits, vec_to_scalar,
9211 : slp_node, 0, vect_body);
9212 1075 : inside_cost
9213 1075 : += record_stmt_cost (cost_vec, cnunits, scalar_store,
9214 : slp_node, 0, vect_body);
9215 1075 : continue;
9216 1075 : }
9217 :
9218 373 : tree offset_vectype = TREE_TYPE (vec_offsets[0]);
9219 373 : unsigned HOST_WIDE_INT const_nunits = nunits.to_constant ();
9220 373 : unsigned HOST_WIDE_INT const_offset_nunits
9221 373 : = TYPE_VECTOR_SUBPARTS (offset_vectype).to_constant ();
9222 373 : vec<constructor_elt, va_gc> *ctor_elts;
9223 373 : vec_alloc (ctor_elts, const_nunits);
9224 373 : gimple_seq stmts = NULL;
9225 373 : tree elt_type = TREE_TYPE (vectype);
9226 373 : unsigned HOST_WIDE_INT elt_size
9227 373 : = tree_to_uhwi (TYPE_SIZE (elt_type));
9228 : /* We support offset vectors with more elements
9229 : than the data vector for now. */
9230 373 : unsigned HOST_WIDE_INT factor
9231 : = const_offset_nunits / const_nunits;
9232 373 : vec_offset = vec_offsets[j / factor];
9233 373 : unsigned elt_offset
9234 373 : = (j % factor) * const_nunits;
9235 373 : tree idx_type = TREE_TYPE (TREE_TYPE (vec_offset));
9236 373 : tree scale = size_int (SLP_TREE_GS_SCALE (slp_node));
9237 373 : tree ltype = build_aligned_type (TREE_TYPE (vectype), align);
9238 1519 : for (unsigned k = 0; k < const_nunits; ++k)
9239 : {
9240 : /* Compute the offsetted pointer. */
9241 1146 : tree boff = size_binop (MULT_EXPR, TYPE_SIZE (idx_type),
9242 : bitsize_int (k + elt_offset));
9243 1146 : tree idx
9244 2292 : = gimple_build (&stmts, BIT_FIELD_REF, idx_type,
9245 1146 : vec_offset, TYPE_SIZE (idx_type), boff);
9246 1146 : idx = gimple_convert (&stmts, sizetype, idx);
9247 1146 : idx = gimple_build (&stmts, MULT_EXPR, sizetype,
9248 : idx, scale);
9249 1146 : tree ptr
9250 1146 : = gimple_build (&stmts, PLUS_EXPR,
9251 1146 : TREE_TYPE (dataref_ptr),
9252 : dataref_ptr, idx);
9253 1146 : ptr = gimple_convert (&stmts, ptr_type_node, ptr);
9254 : /* Extract the element to be stored. */
9255 1146 : tree elt
9256 2292 : = gimple_build (&stmts, BIT_FIELD_REF,
9257 1146 : TREE_TYPE (vectype),
9258 1146 : vec_oprnd, TYPE_SIZE (elt_type),
9259 1146 : bitsize_int (k * elt_size));
9260 1146 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
9261 1146 : stmts = NULL;
9262 1146 : tree ref
9263 1146 : = build2 (MEM_REF, ltype, ptr,
9264 : build_int_cst (ref_type, 0));
9265 1146 : new_stmt = gimple_build_assign (ref, elt);
9266 1146 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
9267 : }
9268 :
9269 373 : slp_node->push_vec_def (new_stmt);
9270 : }
9271 : }
9272 :
9273 1366 : if (costing_p && dump_enabled_p ())
9274 68 : dump_printf_loc (MSG_NOTE, vect_location,
9275 : "vect_model_store_cost: inside_cost = %d, "
9276 : "prologue_cost = %d .\n",
9277 : inside_cost, prologue_cost);
9278 :
9279 1366 : return true;
9280 1366 : }
9281 :
9282 1285388 : gcc_assert (memory_access_type == VMAT_CONTIGUOUS
9283 : || memory_access_type == VMAT_CONTIGUOUS_DOWN
9284 : || memory_access_type == VMAT_CONTIGUOUS_REVERSE);
9285 :
9286 1285388 : unsigned inside_cost = 0, prologue_cost = 0;
9287 : /* For costing some adjacent vector stores, we'd like to cost with
9288 : the total number of them once instead of cost each one by one. */
9289 1285388 : unsigned int n_adjacent_stores = 0;
9290 1285388 : auto_vec<tree> result_chain (group_size);
9291 1285388 : auto_vec<tree, 1> vec_oprnds;
9292 1285388 : gimple *new_stmt;
9293 1285388 : if (!costing_p)
9294 : {
9295 : /* Get vectorized arguments for SLP_NODE. */
9296 539188 : vect_get_slp_defs (op_node, &vec_oprnds);
9297 539188 : vec_oprnd = vec_oprnds[0];
9298 539188 : if (mask_node)
9299 : {
9300 481 : vect_get_slp_defs (mask_node, &vec_masks);
9301 481 : vec_mask = vec_masks[0];
9302 : }
9303 : }
9304 :
9305 : /* We should have caught mismatched types earlier. */
9306 539188 : gcc_assert (costing_p
9307 : || useless_type_conversion_p (vectype, TREE_TYPE (vec_oprnd)));
9308 1285388 : bool simd_lane_access_p
9309 1285388 : = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0;
9310 1285388 : if (!costing_p
9311 1285388 : && simd_lane_access_p
9312 4374 : && !loop_masks
9313 4374 : && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR
9314 4374 : && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0))
9315 4374 : && integer_zerop (get_dr_vinfo_offset (vinfo, first_dr_info))
9316 4374 : && integer_zerop (DR_INIT (first_dr_info->dr))
9317 1289762 : && alias_sets_conflict_p (get_alias_set (aggr_type),
9318 4374 : get_alias_set (TREE_TYPE (ref_type))))
9319 : {
9320 4366 : dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr));
9321 4366 : dataref_offset = build_int_cst (ref_type, 0);
9322 : }
9323 1281022 : else if (!costing_p)
9324 1069636 : dataref_ptr = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
9325 : simd_lane_access_p ? loop : NULL,
9326 : offset, &dummy, gsi, &ptr_incr,
9327 : simd_lane_access_p, bump);
9328 :
9329 1285388 : new_stmt = NULL;
9330 1285388 : gcc_assert (!grouped_store);
9331 2856338 : for (i = 0; i < vec_num; i++)
9332 : {
9333 1570950 : if (!costing_p)
9334 667214 : vec_oprnd = vec_oprnds[i];
9335 :
9336 1570950 : if (memory_access_type == VMAT_CONTIGUOUS_REVERSE)
9337 : {
9338 3108 : if (costing_p)
9339 1972 : inside_cost += record_stmt_cost (cost_vec, 1, vec_perm,
9340 : slp_node, 0, vect_body);
9341 : else
9342 : {
9343 1136 : tree perm_mask = perm_mask_for_reverse (vectype);
9344 1136 : tree new_temp = make_ssa_name (vectype);
9345 :
9346 : /* Generate the permute statement. */
9347 1136 : gimple *perm_stmt
9348 1136 : = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd,
9349 : vec_oprnd, perm_mask);
9350 1136 : vect_finish_stmt_generation (vinfo, stmt_info, perm_stmt, gsi);
9351 :
9352 1136 : perm_stmt = SSA_NAME_DEF_STMT (new_temp);
9353 1570950 : vec_oprnd = new_temp;
9354 : }
9355 : }
9356 :
9357 1570950 : if (costing_p)
9358 : {
9359 903736 : n_adjacent_stores++;
9360 903736 : continue;
9361 : }
9362 :
9363 667214 : tree final_mask = NULL_TREE;
9364 667214 : tree final_len = NULL_TREE;
9365 667214 : tree bias = NULL_TREE;
9366 667214 : if (loop_masks)
9367 77 : final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
9368 : vec_num, vectype, i);
9369 667214 : if (vec_mask)
9370 702 : vec_mask = vec_masks[i];
9371 702 : if (vec_mask)
9372 702 : final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
9373 : vec_mask, gsi);
9374 :
9375 667214 : if (i > 0)
9376 : /* Bump the vector pointer. */
9377 128026 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
9378 : stmt_info, bump);
9379 :
9380 667214 : unsigned misalign;
9381 667214 : unsigned HOST_WIDE_INT align;
9382 667214 : align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info));
9383 667214 : if (alignment_support_scheme == dr_aligned)
9384 : misalign = 0;
9385 308175 : else if (misalignment == DR_MISALIGNMENT_UNKNOWN)
9386 : {
9387 160198 : align = dr_alignment (vect_dr_behavior (vinfo, first_dr_info));
9388 160198 : misalign = 0;
9389 : }
9390 : else
9391 147977 : misalign = misalignment;
9392 667214 : if (dataref_offset == NULL_TREE
9393 661834 : && TREE_CODE (dataref_ptr) == SSA_NAME)
9394 182550 : set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, misalign);
9395 667214 : align = least_bit_hwi (misalign | align);
9396 :
9397 : /* Compute IFN when LOOP_LENS or final_mask valid. */
9398 667214 : machine_mode vmode = TYPE_MODE (vectype);
9399 667214 : machine_mode new_vmode = vmode;
9400 667214 : internal_fn partial_ifn = IFN_LAST;
9401 667214 : if (loop_lens)
9402 : {
9403 0 : opt_machine_mode new_ovmode
9404 0 : = get_len_load_store_mode (vmode, false, &partial_ifn);
9405 0 : new_vmode = new_ovmode.require ();
9406 0 : unsigned factor
9407 0 : = (new_ovmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vmode);
9408 0 : final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
9409 : vec_num, vectype, i, factor, true);
9410 : }
9411 667214 : else if (final_mask)
9412 : {
9413 714 : if (!can_vec_mask_load_store_p (vmode,
9414 714 : TYPE_MODE (TREE_TYPE (final_mask)),
9415 : false, &partial_ifn))
9416 0 : gcc_unreachable ();
9417 : }
9418 :
9419 667214 : if (partial_ifn == IFN_MASK_LEN_STORE)
9420 : {
9421 0 : if (!final_len)
9422 : {
9423 : /* Pass VF value to 'len' argument of
9424 : MASK_LEN_STORE if LOOP_LENS is invalid. */
9425 0 : final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
9426 : }
9427 0 : if (!final_mask)
9428 : {
9429 : /* Pass all ones value to 'mask' argument of
9430 : MASK_LEN_STORE if final_mask is invalid. */
9431 0 : mask_vectype = truth_type_for (vectype);
9432 0 : final_mask = build_minus_one_cst (mask_vectype);
9433 : }
9434 : }
9435 667214 : if (final_len)
9436 : {
9437 0 : signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
9438 0 : bias = build_int_cst (intQI_type_node, biasval);
9439 : }
9440 :
9441 : /* Arguments are ready. Create the new vector stmt. */
9442 667214 : if (final_len)
9443 : {
9444 0 : gcall *call;
9445 0 : tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
9446 : /* Need conversion if it's wrapped with VnQI. */
9447 0 : if (vmode != new_vmode)
9448 : {
9449 0 : tree new_vtype
9450 0 : = build_vector_type_for_mode (unsigned_intQI_type_node,
9451 : new_vmode);
9452 0 : tree var = vect_get_new_ssa_name (new_vtype, vect_simple_var);
9453 0 : vec_oprnd = build1 (VIEW_CONVERT_EXPR, new_vtype, vec_oprnd);
9454 0 : gassign *new_stmt
9455 0 : = gimple_build_assign (var, VIEW_CONVERT_EXPR, vec_oprnd);
9456 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
9457 0 : vec_oprnd = var;
9458 : }
9459 :
9460 0 : if (partial_ifn == IFN_MASK_LEN_STORE)
9461 0 : call = gimple_build_call_internal (IFN_MASK_LEN_STORE, 6,
9462 : dataref_ptr, ptr, final_mask,
9463 : final_len, bias, vec_oprnd);
9464 : else
9465 0 : call = gimple_build_call_internal (IFN_LEN_STORE, 5,
9466 : dataref_ptr, ptr, final_len,
9467 : bias, vec_oprnd);
9468 0 : gimple_call_set_nothrow (call, true);
9469 0 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
9470 0 : new_stmt = call;
9471 : }
9472 667214 : else if (final_mask)
9473 : {
9474 714 : tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
9475 714 : gcall *call
9476 714 : = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr,
9477 : ptr, final_mask, vec_oprnd);
9478 714 : gimple_call_set_nothrow (call, true);
9479 714 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
9480 714 : new_stmt = call;
9481 : }
9482 : else
9483 : {
9484 666500 : data_ref = fold_build2 (MEM_REF, vectype, dataref_ptr,
9485 : dataref_offset ? dataref_offset
9486 : : build_int_cst (ref_type, 0));
9487 666500 : if (alignment_support_scheme == dr_aligned
9488 666500 : && align >= TYPE_ALIGN_UNIT (vectype))
9489 : ;
9490 : else
9491 307631 : TREE_TYPE (data_ref)
9492 615262 : = build_aligned_type (TREE_TYPE (data_ref),
9493 : align * BITS_PER_UNIT);
9494 666500 : vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr));
9495 666500 : new_stmt = gimple_build_assign (data_ref, vec_oprnd);
9496 666500 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
9497 : }
9498 : }
9499 :
9500 1285388 : if (costing_p)
9501 : {
9502 746200 : if (n_adjacent_stores > 0)
9503 746200 : vect_get_store_cost (vinfo, stmt_info, slp_node, n_adjacent_stores,
9504 : alignment_support_scheme, misalignment,
9505 : &inside_cost, cost_vec);
9506 :
9507 : /* When vectorizing a store into the function result assign
9508 : a penalty if the function returns in a multi-register location.
9509 : In this case we assume we'll end up with having to spill the
9510 : vector result and do piecewise loads as a conservative estimate. */
9511 746200 : tree base = get_base_address (STMT_VINFO_DATA_REF (stmt_info)->ref);
9512 746200 : if (base
9513 746200 : && (TREE_CODE (base) == RESULT_DECL
9514 695958 : || (DECL_P (base) && cfun_returns (base)))
9515 808067 : && !aggregate_value_p (base, cfun->decl))
9516 : {
9517 11037 : rtx reg = hard_function_value (TREE_TYPE (base), cfun->decl, 0, 1);
9518 : /* ??? Handle PARALLEL in some way. */
9519 11037 : if (REG_P (reg))
9520 : {
9521 10835 : int nregs = hard_regno_nregs (REGNO (reg), GET_MODE (reg));
9522 : /* Assume that a single reg-reg move is possible and cheap,
9523 : do not account for vector to gp register move cost. */
9524 10835 : if (nregs > 1)
9525 : {
9526 : /* Spill. */
9527 10025 : prologue_cost
9528 10025 : += record_stmt_cost (cost_vec, 1, vector_store,
9529 : slp_node, 0, vect_epilogue);
9530 : /* Loads. */
9531 10025 : prologue_cost
9532 10025 : += record_stmt_cost (cost_vec, nregs, scalar_load,
9533 : slp_node, 0, vect_epilogue);
9534 : }
9535 : }
9536 : }
9537 746200 : if (dump_enabled_p ())
9538 13410 : dump_printf_loc (MSG_NOTE, vect_location,
9539 : "vect_model_store_cost: inside_cost = %d, "
9540 : "prologue_cost = %d .\n",
9541 : inside_cost, prologue_cost);
9542 : }
9543 :
9544 1285388 : return true;
9545 2601936 : }
9546 :
9547 : /* Given a vector type VECTYPE, turns permutation SEL into the equivalent
9548 : VECTOR_CST mask. No checks are made that the target platform supports the
9549 : mask, so callers may wish to test can_vec_perm_const_p separately, or use
9550 : vect_gen_perm_mask_checked. */
9551 :
9552 : tree
9553 61452 : vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel)
9554 : {
9555 61452 : tree mask_type;
9556 :
9557 61452 : poly_uint64 nunits = sel.length ();
9558 61452 : gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype)));
9559 :
9560 61452 : mask_type = build_vector_type (ssizetype, nunits);
9561 61452 : return vec_perm_indices_to_tree (mask_type, sel);
9562 : }
9563 :
9564 : /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p,
9565 : i.e. that the target supports the pattern _for arbitrary input vectors_. */
9566 :
9567 : tree
9568 58689 : vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel)
9569 : {
9570 58689 : machine_mode vmode = TYPE_MODE (vectype);
9571 58689 : gcc_assert (can_vec_perm_const_p (vmode, vmode, sel));
9572 58689 : return vect_gen_perm_mask_any (vectype, sel);
9573 : }
9574 :
9575 : /* Given a vector variable X and Y, that was generated for the scalar
9576 : STMT_INFO, generate instructions to permute the vector elements of X and Y
9577 : using permutation mask MASK_VEC, insert them at *GSI and return the
9578 : permuted vector variable. */
9579 :
9580 : static tree
9581 1431 : permute_vec_elements (vec_info *vinfo,
9582 : tree x, tree y, tree mask_vec, stmt_vec_info stmt_info,
9583 : gimple_stmt_iterator *gsi)
9584 : {
9585 1431 : tree vectype = TREE_TYPE (x);
9586 1431 : tree perm_dest, data_ref;
9587 1431 : gimple *perm_stmt;
9588 :
9589 1431 : tree scalar_dest = gimple_get_lhs (stmt_info->stmt);
9590 1431 : if (scalar_dest && TREE_CODE (scalar_dest) == SSA_NAME)
9591 1431 : perm_dest = vect_create_destination_var (scalar_dest, vectype);
9592 : else
9593 0 : perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL);
9594 1431 : data_ref = make_ssa_name (perm_dest);
9595 :
9596 : /* Generate the permute statement. */
9597 1431 : perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec);
9598 1431 : vect_finish_stmt_generation (vinfo, stmt_info, perm_stmt, gsi);
9599 :
9600 1431 : return data_ref;
9601 : }
9602 :
9603 : /* Hoist the definitions of all SSA uses on STMT_INFO out of the loop LOOP,
9604 : inserting them on the loops preheader edge. Returns true if we
9605 : were successful in doing so (and thus STMT_INFO can be moved then),
9606 : otherwise returns false. HOIST_P indicates if we want to hoist the
9607 : definitions of all SSA uses, it would be false when we are costing. */
9608 :
9609 : static bool
9610 3765 : hoist_defs_of_uses (gimple *stmt, class loop *loop, bool hoist_p)
9611 : {
9612 3765 : ssa_op_iter i;
9613 3765 : use_operand_p use_p;
9614 3765 : auto_vec<use_operand_p, 8> to_hoist;
9615 :
9616 7188 : FOR_EACH_SSA_USE_OPERAND (use_p, stmt, i, SSA_OP_USE)
9617 : {
9618 3451 : gimple *def_stmt = SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p));
9619 3451 : if (!gimple_nop_p (def_stmt)
9620 3451 : && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)))
9621 : {
9622 : /* Make sure we don't need to recurse. While we could do
9623 : so in simple cases when there are more complex use webs
9624 : we don't have an easy way to preserve stmt order to fulfil
9625 : dependencies within them. */
9626 111 : tree op2;
9627 111 : ssa_op_iter i2;
9628 111 : if (gimple_code (def_stmt) == GIMPLE_PHI
9629 111 : || (single_ssa_def_operand (def_stmt, SSA_OP_DEF)
9630 : == NULL_DEF_OPERAND_P))
9631 28 : return false;
9632 226 : FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE)
9633 : {
9634 143 : gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2);
9635 143 : if (!gimple_nop_p (def_stmt2)
9636 143 : && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2)))
9637 : return false;
9638 : }
9639 83 : to_hoist.safe_push (use_p);
9640 : }
9641 : }
9642 :
9643 7474 : if (to_hoist.is_empty ())
9644 : return true;
9645 :
9646 59 : if (!hoist_p)
9647 : return true;
9648 :
9649 : /* Instead of moving defs we copy them so we can zero their UID to not
9650 : confuse dominance queries in the preheader. */
9651 9 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
9652 36 : for (use_operand_p use_p : to_hoist)
9653 : {
9654 9 : gimple *def_stmt = SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p));
9655 9 : gimple *copy = gimple_copy (def_stmt);
9656 9 : gimple_set_uid (copy, 0);
9657 9 : def_operand_p def_p = single_ssa_def_operand (def_stmt, SSA_OP_DEF);
9658 9 : tree new_def = duplicate_ssa_name (DEF_FROM_PTR (def_p), copy);
9659 9 : update_stmt (copy);
9660 9 : def_p = single_ssa_def_operand (copy, SSA_OP_DEF);
9661 9 : SET_DEF (def_p, new_def);
9662 9 : SET_USE (use_p, new_def);
9663 9 : gsi_insert_before (&gsi, copy, GSI_SAME_STMT);
9664 : }
9665 :
9666 : return true;
9667 3765 : }
9668 :
9669 : /* vectorizable_load.
9670 :
9671 : Check if STMT_INFO reads a non scalar data-ref (array/pointer/structure)
9672 : that can be vectorized.
9673 : If COST_VEC is passed, calculate costs but don't change anything,
9674 : otherwise, vectorize STMT_INFO: create a vectorized stmt to replace
9675 : it, and insert it at GSI.
9676 : Return true if STMT_INFO is vectorizable in this way. */
9677 :
9678 : static bool
9679 1949001 : vectorizable_load (vec_info *vinfo,
9680 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
9681 : slp_tree slp_node,
9682 : stmt_vector_for_cost *cost_vec)
9683 : {
9684 1949001 : tree scalar_dest;
9685 1949001 : tree vec_dest = NULL;
9686 1949001 : tree data_ref = NULL;
9687 1949001 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
9688 1949001 : class loop *loop = NULL;
9689 1949001 : class loop *containing_loop = gimple_bb (stmt_info->stmt)->loop_father;
9690 1949001 : bool nested_in_vect_loop = false;
9691 1949001 : tree elem_type;
9692 : /* Avoid false positive uninitialized warning, see PR110652. */
9693 1949001 : tree new_temp = NULL_TREE;
9694 1949001 : machine_mode mode;
9695 1949001 : tree dummy;
9696 1949001 : tree dataref_ptr = NULL_TREE;
9697 1949001 : tree dataref_offset = NULL_TREE;
9698 1949001 : gimple *ptr_incr = NULL;
9699 1949001 : int i, j;
9700 1949001 : unsigned int group_size;
9701 1949001 : poly_uint64 group_gap_adj;
9702 1949001 : tree msq = NULL_TREE, lsq;
9703 1949001 : tree realignment_token = NULL_TREE;
9704 1949001 : gphi *phi = NULL;
9705 1949001 : bool grouped_load = false;
9706 1949001 : stmt_vec_info first_stmt_info;
9707 1949001 : stmt_vec_info first_stmt_info_for_drptr = NULL;
9708 1949001 : bool compute_in_loop = false;
9709 1949001 : class loop *at_loop;
9710 1949001 : int vec_num;
9711 1949001 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
9712 1949001 : poly_uint64 vf;
9713 1949001 : tree aggr_type;
9714 1949001 : tree ref_type;
9715 1949001 : enum vect_def_type mask_dt = vect_unknown_def_type;
9716 1949001 : enum vect_def_type els_dt = vect_unknown_def_type;
9717 :
9718 1949001 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
9719 : return false;
9720 :
9721 1949001 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
9722 194522 : && cost_vec)
9723 : return false;
9724 :
9725 1754479 : if (!STMT_VINFO_DATA_REF (stmt_info))
9726 : return false;
9727 :
9728 1412946 : tree mask_vectype = NULL_TREE;
9729 1412946 : tree els = NULL_TREE; tree els_vectype = NULL_TREE;
9730 :
9731 1412946 : int mask_index = -1;
9732 1412946 : int els_index = -1;
9733 1412946 : slp_tree mask_node = NULL;
9734 1412946 : slp_tree els_op = NULL;
9735 1412946 : if (gassign *assign = dyn_cast <gassign *> (stmt_info->stmt))
9736 : {
9737 1409620 : scalar_dest = gimple_assign_lhs (assign);
9738 1409620 : if (TREE_CODE (scalar_dest) != SSA_NAME)
9739 : return false;
9740 :
9741 637646 : tree_code code = gimple_assign_rhs_code (assign);
9742 637646 : if (code != ARRAY_REF
9743 637646 : && code != BIT_FIELD_REF
9744 637646 : && code != INDIRECT_REF
9745 443226 : && code != COMPONENT_REF
9746 443226 : && code != IMAGPART_EXPR
9747 310571 : && code != REALPART_EXPR
9748 310571 : && code != MEM_REF
9749 237 : && TREE_CODE_CLASS (code) != tcc_declaration)
9750 : return false;
9751 : }
9752 : else
9753 : {
9754 1312669 : gcall *call = dyn_cast <gcall *> (stmt_info->stmt);
9755 3326 : if (!call || !gimple_call_internal_p (call))
9756 : return false;
9757 :
9758 3326 : internal_fn ifn = gimple_call_internal_fn (call);
9759 3326 : if (!internal_load_fn_p (ifn))
9760 : return false;
9761 :
9762 2407 : scalar_dest = gimple_call_lhs (call);
9763 2407 : if (!scalar_dest)
9764 : return false;
9765 :
9766 2407 : mask_index = internal_fn_mask_index (ifn);
9767 2407 : if (mask_index >= 0)
9768 2407 : mask_index = vect_slp_child_index_for_operand
9769 2407 : (call, mask_index, STMT_VINFO_GATHER_SCATTER_P (stmt_info));
9770 2407 : if (mask_index >= 0
9771 2407 : && !vect_check_scalar_mask (vinfo, slp_node, mask_index,
9772 : &mask_node, &mask_dt, &mask_vectype))
9773 : return false;
9774 :
9775 2407 : els_index = internal_fn_else_index (ifn);
9776 2407 : if (els_index >= 0)
9777 2407 : els_index = vect_slp_child_index_for_operand
9778 2407 : (call, els_index, STMT_VINFO_GATHER_SCATTER_P (stmt_info));
9779 2407 : if (els_index >= 0
9780 2407 : && !vect_is_simple_use (vinfo, slp_node, els_index,
9781 : &els, &els_op, &els_dt, &els_vectype))
9782 : return false;
9783 : }
9784 :
9785 639986 : tree vectype = SLP_TREE_VECTYPE (slp_node);
9786 639986 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
9787 :
9788 639986 : if (loop_vinfo)
9789 : {
9790 424968 : loop = LOOP_VINFO_LOOP (loop_vinfo);
9791 424968 : nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt_info);
9792 424968 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
9793 : }
9794 : else
9795 : vf = 1;
9796 :
9797 639986 : vec_num = vect_get_num_copies (vinfo, slp_node);
9798 :
9799 : /* FORNOW. This restriction should be relaxed. */
9800 639986 : if (nested_in_vect_loop && vec_num > 1)
9801 : {
9802 316 : if (dump_enabled_p ())
9803 66 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9804 : "multiple types in nested loop.\n");
9805 316 : return false;
9806 : }
9807 :
9808 639670 : elem_type = TREE_TYPE (vectype);
9809 639670 : mode = TYPE_MODE (vectype);
9810 :
9811 : /* FORNOW. In some cases can vectorize even if data-type not supported
9812 : (e.g. - data copies). */
9813 639670 : if (!can_implement_p (mov_optab, mode))
9814 : {
9815 0 : if (dump_enabled_p ())
9816 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9817 : "Aligned load, but unsupported type.\n");
9818 0 : return false;
9819 : }
9820 :
9821 : /* Check if the load is a part of an interleaving chain. */
9822 639670 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
9823 : {
9824 301850 : grouped_load = true;
9825 : /* FORNOW */
9826 301850 : gcc_assert (!nested_in_vect_loop);
9827 301850 : gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info));
9828 :
9829 301850 : first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
9830 301850 : group_size = DR_GROUP_SIZE (first_stmt_info);
9831 :
9832 : /* Invalidate assumptions made by dependence analysis when vectorization
9833 : on the unrolled body effectively re-orders stmts. */
9834 301850 : if (STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0
9835 301850 : && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
9836 : STMT_VINFO_MIN_NEG_DIST (stmt_info)))
9837 : {
9838 12 : if (dump_enabled_p ())
9839 12 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9840 : "cannot perform implicit CSE when performing "
9841 : "group loads with negative dependence distance\n");
9842 12 : return false;
9843 : }
9844 : }
9845 : else
9846 : group_size = 1;
9847 :
9848 639658 : vect_load_store_data _ls_data{};
9849 639658 : vect_load_store_data &ls = slp_node->get_data (_ls_data);
9850 639658 : if (cost_vec
9851 639658 : && !get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask_node,
9852 : VLS_LOAD, &ls))
9853 : return false;
9854 : /* Temporary aliases to analysis data, should not be modified through
9855 : these. */
9856 534138 : const vect_memory_access_type memory_access_type = ls.memory_access_type;
9857 534138 : const dr_alignment_support alignment_support_scheme
9858 : = ls.alignment_support_scheme;
9859 534138 : const int misalignment = ls.misalignment;
9860 534138 : const poly_int64 poffset = ls.poffset;
9861 534138 : const vec<int> &elsvals = ls.elsvals;
9862 :
9863 534138 : int maskload_elsval = 0;
9864 534138 : bool need_zeroing = false;
9865 :
9866 : /* We might need to explicitly zero inactive elements if there are
9867 : padding bits in the type that might leak otherwise.
9868 : Refer to PR115336. */
9869 534138 : tree scalar_type = TREE_TYPE (scalar_dest);
9870 534138 : bool type_mode_padding_p
9871 1068276 : = TYPE_PRECISION (scalar_type) < GET_MODE_PRECISION (GET_MODE_INNER (mode));
9872 :
9873 534138 : if (slp_node->ldst_lanes
9874 0 : && memory_access_type != VMAT_LOAD_STORE_LANES)
9875 : {
9876 0 : if (dump_enabled_p ())
9877 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9878 : "discovered load-lane but cannot use it.\n");
9879 0 : return false;
9880 : }
9881 :
9882 534138 : if (mask_node)
9883 : {
9884 2289 : if (memory_access_type == VMAT_CONTIGUOUS)
9885 : {
9886 1507 : machine_mode vec_mode = TYPE_MODE (vectype);
9887 395 : if (!VECTOR_MODE_P (vec_mode)
9888 3014 : || !can_vec_mask_load_store_p (vec_mode,
9889 1507 : TYPE_MODE (mask_vectype),
9890 : true, NULL, &ls.elsvals))
9891 67 : return false;
9892 : }
9893 782 : else if (memory_access_type == VMAT_ELEMENTWISE
9894 782 : || memory_access_type == VMAT_STRIDED_SLP)
9895 : {
9896 0 : if (dump_enabled_p ())
9897 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9898 : "unsupported masked strided access.\n");
9899 0 : return false;
9900 : }
9901 782 : else if (memory_access_type != VMAT_LOAD_STORE_LANES
9902 782 : && !mat_gather_scatter_p (memory_access_type))
9903 : {
9904 62 : if (dump_enabled_p ())
9905 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9906 : "unsupported access type for masked load.\n");
9907 62 : return false;
9908 : }
9909 720 : else if (memory_access_type == VMAT_GATHER_SCATTER_EMULATED)
9910 : {
9911 476 : if (dump_enabled_p ())
9912 26 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9913 : "unsupported masked emulated gather.\n");
9914 476 : return false;
9915 : }
9916 : }
9917 :
9918 533533 : bool costing_p = cost_vec;
9919 :
9920 533533 : if (costing_p) /* transformation not required. */
9921 : {
9922 367632 : if (mask_node
9923 367632 : && !vect_maybe_update_slp_op_vectype (mask_node,
9924 : mask_vectype))
9925 : {
9926 0 : if (dump_enabled_p ())
9927 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9928 : "incompatible vector types for invariants\n");
9929 0 : return false;
9930 : }
9931 :
9932 367632 : if (loop_vinfo
9933 244029 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
9934 167040 : check_load_store_for_partial_vectors (loop_vinfo, vectype, slp_node,
9935 : VLS_LOAD, group_size, &ls,
9936 : mask_node, &ls.elsvals);
9937 :
9938 367632 : if (dump_enabled_p ()
9939 24393 : && memory_access_type != VMAT_ELEMENTWISE
9940 24284 : && !mat_gather_scatter_p (memory_access_type)
9941 24062 : && memory_access_type != VMAT_STRIDED_SLP
9942 24062 : && memory_access_type != VMAT_INVARIANT
9943 390801 : && alignment_support_scheme != dr_aligned)
9944 9405 : dump_printf_loc (MSG_NOTE, vect_location,
9945 : "Vectorizing an unaligned access.\n");
9946 :
9947 367632 : if (memory_access_type == VMAT_LOAD_STORE_LANES)
9948 0 : vinfo->any_known_not_updated_vssa = true;
9949 :
9950 367632 : SLP_TREE_TYPE (slp_node) = load_vec_info_type;
9951 367632 : slp_node->data = new vect_load_store_data (std::move (ls));
9952 : }
9953 :
9954 : /* If the type needs padding we must zero inactive elements.
9955 : Check if we can do that with a VEC_COND_EXPR and store the
9956 : elsval we choose in MASKLOAD_ELSVAL. */
9957 533533 : if (elsvals.length ()
9958 23231 : && type_mode_padding_p
9959 3 : && !elsvals.contains (MASK_LOAD_ELSE_ZERO)
9960 23231 : && !expand_vec_cond_expr_p (vectype, truth_type_for (vectype)))
9961 : {
9962 0 : if (dump_enabled_p ())
9963 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9964 : "cannot zero inactive elements.\n");
9965 0 : return false;
9966 : }
9967 :
9968 : /* For now just use the first available else value.
9969 : get_supported_else_vals tries MASK_LOAD_ELSE_ZERO first so we will
9970 : select it here if it is supported. */
9971 533533 : if (elsvals.length ())
9972 23231 : maskload_elsval = *elsvals.begin ();
9973 :
9974 533533 : if (dump_enabled_p () && !costing_p)
9975 16586 : dump_printf_loc (MSG_NOTE, vect_location, "transform load.\n");
9976 :
9977 : /* Transform. */
9978 :
9979 533533 : dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL;
9980 533533 : ensure_base_align (dr_info);
9981 :
9982 533533 : if (memory_access_type == VMAT_INVARIANT)
9983 : {
9984 3854 : gcc_assert (!grouped_load && !mask_node && !bb_vinfo);
9985 : /* If we have versioned for aliasing or the loop doesn't
9986 : have any data dependencies that would preclude this,
9987 : then we are sure this is a loop invariant load and
9988 : thus we can insert it on the preheader edge.
9989 : TODO: hoist_defs_of_uses should ideally be computed
9990 : once at analysis time, remembered and used in the
9991 : transform time. */
9992 7708 : bool hoist_p = (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo)
9993 3854 : && !nested_in_vect_loop);
9994 :
9995 3854 : bool uniform_p = true;
9996 16096 : for (stmt_vec_info sinfo : SLP_TREE_SCALAR_STMTS (slp_node))
9997 : {
9998 : /* It is unsafe to hoist a conditional load over the conditions that
9999 : make it valid. When early break this means that any invariant load
10000 : can't be hoisted unless it's in the loop header or if we know
10001 : something else has verified the load is valid to do. Alignment
10002 : peeling would do this since getting through the prologue means the
10003 : load was done at least once and so the vector main body is free to
10004 : hoist it. However today GCC will hoist the load above the PFA
10005 : loop. As such that makes it still invalid and so we can't allow it
10006 : today. */
10007 4534 : if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo)
10008 1050 : && !DR_SCALAR_KNOWN_BOUNDS (STMT_VINFO_DR_INFO (sinfo))
10009 5552 : && gimple_bb (STMT_VINFO_STMT (vect_orig_stmt (sinfo)))
10010 1018 : != loop->header)
10011 : {
10012 918 : if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)
10013 918 : && dump_enabled_p ())
10014 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10015 : "not hoisting invariant load due to early break"
10016 : "constraints\n");
10017 912 : else if (dump_enabled_p ())
10018 16 : dump_printf_loc (MSG_NOTE, vect_location,
10019 : "not hoisting invariant load due to early break"
10020 : "constraints\n");
10021 : hoist_p = false;
10022 : }
10023 :
10024 3616 : hoist_p = hoist_p && hoist_defs_of_uses (sinfo->stmt, loop, false);
10025 4534 : if (sinfo != SLP_TREE_SCALAR_STMTS (slp_node)[0])
10026 211 : uniform_p = false;
10027 : }
10028 3854 : if (costing_p)
10029 : {
10030 2984 : if (!uniform_p && (!hoist_p || !vf.is_constant ()))
10031 : {
10032 0 : if (dump_enabled_p ())
10033 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10034 : "not vectorizing non-uniform invariant "
10035 : "load\n");
10036 0 : return false;
10037 : }
10038 1379 : enum vect_cost_model_location cost_loc
10039 2984 : = hoist_p ? vect_prologue : vect_body;
10040 2984 : unsigned int cost = record_stmt_cost (cost_vec, 1, scalar_load,
10041 : slp_node, 0, cost_loc);
10042 2984 : cost += record_stmt_cost (cost_vec, 1, scalar_to_vec,
10043 : slp_node, 0, cost_loc);
10044 2984 : unsigned int prologue_cost = hoist_p ? cost : 0;
10045 1379 : unsigned int inside_cost = hoist_p ? 0 : cost;
10046 2984 : if (dump_enabled_p ())
10047 508 : dump_printf_loc (MSG_NOTE, vect_location,
10048 : "vect_model_load_cost: inside_cost = %d, "
10049 : "prologue_cost = %d .\n",
10050 : inside_cost, prologue_cost);
10051 2984 : return true;
10052 : }
10053 870 : if (hoist_p)
10054 : {
10055 : /* ??? For non-uniform lanes there could be still duplicates.
10056 : We're leaving those to post-vectorizer CSE for the moment. */
10057 675 : auto_vec<tree> scalar_defs (SLP_TREE_LANES (slp_node));
10058 2182 : for (stmt_vec_info sinfo : SLP_TREE_SCALAR_STMTS (slp_node))
10059 : {
10060 777 : gassign *stmt = as_a <gassign *> (sinfo->stmt);
10061 777 : if (dump_enabled_p ())
10062 368 : dump_printf_loc (MSG_NOTE, vect_location,
10063 : "hoisting out of the vectorized loop: %G",
10064 : (gimple *) stmt);
10065 777 : scalar_dest = copy_ssa_name (gimple_assign_lhs (stmt));
10066 777 : tree rhs = unshare_expr (gimple_assign_rhs1 (stmt));
10067 777 : edge pe = loop_preheader_edge (loop);
10068 777 : gphi *vphi = get_virtual_phi (loop->header);
10069 777 : tree vuse;
10070 777 : if (vphi)
10071 771 : vuse = PHI_ARG_DEF_FROM_EDGE (vphi, pe);
10072 : else
10073 6 : vuse = gimple_vuse (gsi_stmt (*gsi));
10074 777 : gimple *new_stmt = gimple_build_assign (scalar_dest, rhs);
10075 777 : gimple_set_vuse (new_stmt, vuse);
10076 777 : gsi_insert_on_edge_immediate (pe, new_stmt);
10077 777 : hoist_defs_of_uses (new_stmt, loop, true);
10078 777 : if (!useless_type_conversion_p (TREE_TYPE (vectype),
10079 777 : TREE_TYPE (scalar_dest)))
10080 : {
10081 12 : tree tem = make_ssa_name (TREE_TYPE (vectype));
10082 12 : new_stmt = gimple_build_assign (tem,
10083 : NOP_EXPR, scalar_dest);
10084 12 : gsi_insert_on_edge_immediate (pe, new_stmt);
10085 12 : scalar_dest = tem;
10086 : }
10087 777 : scalar_defs.quick_push (scalar_dest);
10088 777 : if (uniform_p)
10089 : break;
10090 : }
10091 675 : if (!uniform_p)
10092 : {
10093 55 : unsigned const_nunits
10094 55 : = TYPE_VECTOR_SUBPARTS (vectype).to_constant ();
10095 124 : for (j = 0; j < (int) vec_num; ++j)
10096 : {
10097 69 : vec<constructor_elt, va_gc> *v = NULL;
10098 69 : vec_safe_reserve (v, const_nunits, true);
10099 405 : for (unsigned i = 0; i < const_nunits; ++i)
10100 : {
10101 336 : unsigned def_idx
10102 336 : = (j * const_nunits + i) % SLP_TREE_LANES (slp_node);
10103 336 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
10104 : scalar_defs[def_idx]);
10105 : }
10106 69 : scalar_dest = build_constructor (vectype, v);
10107 69 : new_temp = vect_init_vector (vinfo, stmt_info, scalar_dest,
10108 : vectype, NULL);
10109 69 : slp_node->push_vec_def (new_temp);
10110 : }
10111 55 : return true;
10112 : }
10113 620 : new_temp = vect_init_vector (vinfo, stmt_info, scalar_dest,
10114 : vectype, NULL);
10115 675 : }
10116 : else
10117 : {
10118 195 : gcc_assert (uniform_p);
10119 195 : gimple_stmt_iterator gsi2 = *gsi;
10120 195 : gsi_next (&gsi2);
10121 195 : new_temp = vect_init_vector (vinfo, stmt_info, scalar_dest,
10122 : vectype, &gsi2);
10123 : }
10124 1704 : for (j = 0; j < (int) vec_num; ++j)
10125 889 : slp_node->push_vec_def (new_temp);
10126 : return true;
10127 : }
10128 :
10129 529679 : if (memory_access_type == VMAT_ELEMENTWISE
10130 529679 : || memory_access_type == VMAT_STRIDED_SLP)
10131 : {
10132 22822 : gimple_stmt_iterator incr_gsi;
10133 22822 : bool insert_after;
10134 22822 : tree offvar = NULL_TREE;
10135 22822 : tree ivstep;
10136 22822 : tree running_off;
10137 22822 : vec<constructor_elt, va_gc> *v = NULL;
10138 22822 : tree stride_base, stride_step, alias_off;
10139 : /* Checked by get_load_store_type. */
10140 22822 : unsigned int const_nunits = nunits.to_constant ();
10141 22822 : unsigned HOST_WIDE_INT cst_offset = 0;
10142 22822 : tree dr_offset;
10143 22822 : unsigned int inside_cost = 0;
10144 :
10145 22822 : gcc_assert (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo));
10146 22822 : gcc_assert (!nested_in_vect_loop);
10147 :
10148 22822 : if (grouped_load)
10149 : {
10150 : /* If we elided a consecutive load permutation, don't
10151 : use the original first statement (which could be elided)
10152 : but the one the load permutation starts with.
10153 : This ensures the stride_base below is correct. */
10154 10165 : if (!ls.subchain_p)
10155 10133 : first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
10156 : else
10157 32 : first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0];
10158 10165 : first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info);
10159 10165 : ref_type = get_group_alias_ptr_type (first_stmt_info);
10160 : }
10161 : else
10162 : {
10163 12657 : first_stmt_info = stmt_info;
10164 12657 : first_dr_info = dr_info;
10165 12657 : ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr));
10166 : }
10167 :
10168 22822 : if (grouped_load)
10169 : {
10170 10165 : if (memory_access_type == VMAT_STRIDED_SLP)
10171 : {
10172 : /* If we elided a consecutive load permutation, adjust
10173 : the group size here. */
10174 3713 : if (!ls.subchain_p)
10175 3681 : group_size = DR_GROUP_SIZE (first_stmt_info);
10176 : else
10177 32 : group_size = SLP_TREE_LANES (slp_node);
10178 : }
10179 : else /* VMAT_ELEMENTWISE */
10180 6452 : group_size = SLP_TREE_LANES (slp_node);
10181 : }
10182 : else
10183 : group_size = 1;
10184 :
10185 22822 : if (!costing_p)
10186 : {
10187 3348 : dr_offset = get_dr_vinfo_offset (vinfo, first_dr_info);
10188 3348 : stride_base = fold_build_pointer_plus (
10189 : DR_BASE_ADDRESS (first_dr_info->dr),
10190 : size_binop (PLUS_EXPR, convert_to_ptrofftype (dr_offset),
10191 : convert_to_ptrofftype (DR_INIT (first_dr_info->dr))));
10192 3348 : stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr));
10193 :
10194 : /* For a load with loop-invariant (but other than power-of-2)
10195 : stride (i.e. not a grouped access) like so:
10196 :
10197 : for (i = 0; i < n; i += stride)
10198 : ... = array[i];
10199 :
10200 : we generate a new induction variable and new accesses to
10201 : form a new vector (or vectors, depending on ncopies):
10202 :
10203 : for (j = 0; ; j += VF*stride)
10204 : tmp1 = array[j];
10205 : tmp2 = array[j + stride];
10206 : ...
10207 : vectemp = {tmp1, tmp2, ...}
10208 : */
10209 :
10210 3348 : ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step,
10211 : build_int_cst (TREE_TYPE (stride_step), vf));
10212 :
10213 3348 : standard_iv_increment_position (loop, &incr_gsi, &insert_after);
10214 :
10215 3348 : stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
10216 3348 : ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
10217 3348 : create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
10218 : loop, &incr_gsi, insert_after,
10219 : &offvar, NULL);
10220 :
10221 3348 : stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step);
10222 : }
10223 :
10224 22822 : running_off = offvar;
10225 22822 : alias_off = build_int_cst (ref_type, 0);
10226 22822 : int nloads = const_nunits;
10227 22822 : int lnel = 1;
10228 22822 : tree ltype = TREE_TYPE (vectype);
10229 22822 : tree lvectype = vectype;
10230 22822 : auto_vec<tree> dr_chain;
10231 : /* ??? Modify local copies of alignment_support_scheme and
10232 : misalignment, but this part of analysis should be done
10233 : earlier and remembered, likewise the chosen load mode. */
10234 22822 : const dr_alignment_support tem = alignment_support_scheme;
10235 22822 : dr_alignment_support alignment_support_scheme = tem;
10236 22822 : const int tem2 = misalignment;
10237 22822 : int misalignment = tem2;
10238 22822 : if (memory_access_type == VMAT_STRIDED_SLP)
10239 : {
10240 16370 : HOST_WIDE_INT n = gcd (group_size, const_nunits);
10241 : /* Use the target vector type if the group size is a multiple
10242 : of it. */
10243 16370 : if (n == const_nunits)
10244 : {
10245 1950 : int mis_align = dr_misalignment (first_dr_info, vectype);
10246 : /* With VF > 1 we advance the DR by step, if that is constant
10247 : and only aligned when performed VF times, DR alignment
10248 : analysis can analyze this as aligned since it assumes
10249 : contiguous accesses. But that is not how we code generate
10250 : here, so adjust for this. */
10251 1950 : if (maybe_gt (vf, 1u)
10252 3184 : && !multiple_p (DR_STEP_ALIGNMENT (first_dr_info->dr),
10253 2980 : DR_TARGET_ALIGNMENT (first_dr_info)))
10254 204 : mis_align = -1;
10255 1950 : dr_alignment_support dr_align
10256 1950 : = vect_supportable_dr_alignment (vinfo, dr_info, vectype,
10257 : mis_align);
10258 1950 : if (dr_align == dr_aligned
10259 1950 : || dr_align == dr_unaligned_supported)
10260 : {
10261 16370 : nloads = 1;
10262 16370 : lnel = const_nunits;
10263 16370 : ltype = vectype;
10264 16370 : alignment_support_scheme = dr_align;
10265 16370 : misalignment = mis_align;
10266 : }
10267 : }
10268 : /* Else use the biggest vector we can load the group without
10269 : accessing excess elements. */
10270 14420 : else if (n > 1)
10271 : {
10272 1770 : tree ptype;
10273 1770 : tree vtype
10274 1770 : = vector_vector_composition_type (vectype, const_nunits / n,
10275 : &ptype);
10276 1770 : if (vtype != NULL_TREE)
10277 : {
10278 1734 : dr_alignment_support dr_align;
10279 1734 : int mis_align = 0;
10280 1734 : if (VECTOR_TYPE_P (ptype))
10281 : {
10282 888 : mis_align = dr_misalignment (first_dr_info, ptype);
10283 888 : if (maybe_gt (vf, 1u)
10284 1748 : && !multiple_p (DR_STEP_ALIGNMENT (first_dr_info->dr),
10285 894 : DR_TARGET_ALIGNMENT (first_dr_info)))
10286 854 : mis_align = -1;
10287 888 : dr_align
10288 888 : = vect_supportable_dr_alignment (vinfo, dr_info, ptype,
10289 : mis_align);
10290 : }
10291 : else
10292 : dr_align = dr_unaligned_supported;
10293 1734 : if (dr_align == dr_aligned
10294 1734 : || dr_align == dr_unaligned_supported)
10295 : {
10296 1734 : nloads = const_nunits / n;
10297 1734 : lnel = n;
10298 1734 : lvectype = vtype;
10299 1734 : ltype = ptype;
10300 1734 : alignment_support_scheme = dr_align;
10301 1734 : misalignment = mis_align;
10302 : }
10303 : }
10304 : }
10305 16370 : unsigned align;
10306 16370 : if (alignment_support_scheme == dr_aligned)
10307 12 : align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info));
10308 : else
10309 16358 : align = dr_alignment (vect_dr_behavior (vinfo, first_dr_info));
10310 : /* Alignment is at most the access size if we do multiple loads. */
10311 16370 : if (nloads > 1)
10312 14420 : align = MIN (tree_to_uhwi (TYPE_SIZE_UNIT (ltype)), align);
10313 16370 : ltype = build_aligned_type (ltype, align * BITS_PER_UNIT);
10314 : }
10315 :
10316 : /* For SLP permutation support we need to load the whole group,
10317 : not only the number of vector stmts the permutation result
10318 : fits in. */
10319 22822 : int ncopies;
10320 22822 : if (ls.slp_perm)
10321 : {
10322 2384 : gcc_assert (memory_access_type != VMAT_ELEMENTWISE);
10323 : /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for
10324 : variable VF. */
10325 2384 : unsigned int const_vf = vf.to_constant ();
10326 2384 : ncopies = CEIL (group_size * const_vf, const_nunits);
10327 2384 : dr_chain.create (ncopies);
10328 : }
10329 : else
10330 : ncopies = vec_num;
10331 :
10332 22822 : unsigned int group_el = 0;
10333 22822 : unsigned HOST_WIDE_INT
10334 22822 : elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
10335 22822 : unsigned int n_groups = 0;
10336 : /* For costing some adjacent vector loads, we'd like to cost with
10337 : the total number of them once instead of cost each one by one. */
10338 22822 : unsigned int n_adjacent_loads = 0;
10339 54292 : for (j = 0; j < ncopies; j++)
10340 : {
10341 31470 : if (nloads > 1 && !costing_p)
10342 2836 : vec_alloc (v, nloads);
10343 : gimple *new_stmt = NULL;
10344 134752 : for (i = 0; i < nloads; i++)
10345 : {
10346 103282 : if (costing_p)
10347 : {
10348 : /* For VMAT_ELEMENTWISE, just cost it as scalar_load to
10349 : avoid ICE, see PR110776. */
10350 93390 : if (VECTOR_TYPE_P (ltype)
10351 4634 : && memory_access_type != VMAT_ELEMENTWISE)
10352 4634 : n_adjacent_loads++;
10353 : else
10354 88756 : inside_cost += record_stmt_cost (cost_vec, 1, scalar_load,
10355 : slp_node, 0, vect_body);
10356 93390 : continue;
10357 : }
10358 9892 : unsigned int load_el = group_el;
10359 : /* For elementwise accesses apply a load permutation directly. */
10360 9892 : if (memory_access_type == VMAT_ELEMENTWISE
10361 9892 : && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())
10362 1978 : load_el = SLP_TREE_LOAD_PERMUTATION (slp_node)[group_el];
10363 9892 : tree this_off = build_int_cst (TREE_TYPE (alias_off),
10364 9892 : load_el * elsz + cst_offset);
10365 9892 : tree data_ref = build2 (MEM_REF, ltype, running_off, this_off);
10366 9892 : vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr));
10367 9892 : new_temp = make_ssa_name (ltype);
10368 9892 : new_stmt = gimple_build_assign (new_temp, data_ref);
10369 9892 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
10370 9892 : if (nloads > 1)
10371 8238 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, new_temp);
10372 :
10373 9892 : group_el += lnel;
10374 9892 : if (group_el == group_size)
10375 : {
10376 9497 : n_groups++;
10377 : /* When doing SLP make sure to not load elements from
10378 : the next vector iteration, those will not be accessed
10379 : so just use the last element again. See PR107451. */
10380 9497 : if (known_lt (n_groups, vf))
10381 : {
10382 6127 : tree newoff = copy_ssa_name (running_off);
10383 6127 : gimple *incr
10384 6127 : = gimple_build_assign (newoff, POINTER_PLUS_EXPR,
10385 : running_off, stride_step);
10386 6127 : vect_finish_stmt_generation (vinfo, stmt_info, incr, gsi);
10387 6127 : running_off = newoff;
10388 : }
10389 : group_el = 0;
10390 : }
10391 : }
10392 :
10393 31470 : if (nloads > 1)
10394 : {
10395 23188 : if (costing_p)
10396 20352 : inside_cost += record_stmt_cost (cost_vec, 1, vec_construct,
10397 : slp_node, 0, vect_body);
10398 : else
10399 : {
10400 2836 : tree vec_inv = build_constructor (lvectype, v);
10401 2836 : new_temp = vect_init_vector (vinfo, stmt_info, vec_inv,
10402 : lvectype, gsi);
10403 2836 : new_stmt = SSA_NAME_DEF_STMT (new_temp);
10404 2836 : if (lvectype != vectype)
10405 : {
10406 239 : new_stmt
10407 239 : = gimple_build_assign (make_ssa_name (vectype),
10408 : VIEW_CONVERT_EXPR,
10409 : build1 (VIEW_CONVERT_EXPR,
10410 : vectype, new_temp));
10411 239 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt,
10412 : gsi);
10413 : }
10414 : }
10415 : }
10416 8282 : else if (!costing_p && ltype != vectype)
10417 : {
10418 1629 : new_stmt = gimple_build_assign (make_ssa_name (vectype),
10419 : VIEW_CONVERT_EXPR,
10420 : build1 (VIEW_CONVERT_EXPR,
10421 : vectype, new_temp));
10422 1629 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt,
10423 : gsi);
10424 : }
10425 :
10426 31470 : if (!costing_p)
10427 : {
10428 4490 : if (ls.slp_perm)
10429 1330 : dr_chain.quick_push (gimple_assign_lhs (new_stmt));
10430 : else
10431 3160 : slp_node->push_vec_def (new_stmt);
10432 : }
10433 : }
10434 22822 : if (ls.slp_perm)
10435 : {
10436 2384 : if (costing_p)
10437 : {
10438 1763 : gcc_assert (ls.n_perms != -1U);
10439 1763 : inside_cost += record_stmt_cost (cost_vec, ls.n_perms, vec_perm,
10440 : slp_node, 0, vect_body);
10441 : }
10442 : else
10443 : {
10444 621 : unsigned n_perms2;
10445 621 : vect_transform_slp_perm_load (vinfo, slp_node, dr_chain, gsi, vf,
10446 : false, &n_perms2);
10447 621 : gcc_assert (ls.n_perms == n_perms2);
10448 : }
10449 : }
10450 :
10451 22822 : if (costing_p)
10452 : {
10453 19474 : if (n_adjacent_loads > 0)
10454 1766 : vect_get_load_cost (vinfo, stmt_info, slp_node, n_adjacent_loads,
10455 : alignment_support_scheme, misalignment, false,
10456 : &inside_cost, nullptr, cost_vec, cost_vec,
10457 : true);
10458 19474 : if (dump_enabled_p ())
10459 494 : dump_printf_loc (MSG_NOTE, vect_location,
10460 : "vect_model_load_cost: inside_cost = %u, "
10461 : "prologue_cost = 0 .\n",
10462 : inside_cost);
10463 : }
10464 :
10465 22822 : return true;
10466 22822 : }
10467 :
10468 506857 : if (mat_gather_scatter_p (memory_access_type)
10469 506857 : && !ls.ls_type)
10470 : grouped_load = false;
10471 :
10472 504130 : if (grouped_load
10473 506857 : || SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())
10474 : {
10475 256214 : if (grouped_load)
10476 : {
10477 255825 : first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
10478 255825 : group_size = DR_GROUP_SIZE (first_stmt_info);
10479 : }
10480 : else
10481 : {
10482 : first_stmt_info = stmt_info;
10483 : group_size = 1;
10484 : }
10485 : /* For SLP vectorization we directly vectorize a subchain
10486 : without permutation. */
10487 256214 : if (! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())
10488 207522 : first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0];
10489 : /* For BB vectorization always use the first stmt to base
10490 : the data ref pointer on. */
10491 256214 : if (bb_vinfo)
10492 208716 : first_stmt_info_for_drptr
10493 208716 : = vect_find_first_scalar_stmt_in_slp (slp_node);
10494 :
10495 256214 : first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info);
10496 256214 : group_gap_adj = 0;
10497 :
10498 : /* VEC_NUM is the number of vect stmts to be created for this group. */
10499 256214 : grouped_load = false;
10500 : /* If an SLP permutation is from N elements to N elements,
10501 : and if one vector holds a whole number of N, we can load
10502 : the inputs to the permutation in the same way as an
10503 : unpermuted sequence. In other cases we need to load the
10504 : whole group, not only the number of vector stmts the
10505 : permutation result fits in. */
10506 256214 : unsigned scalar_lanes = SLP_TREE_LANES (slp_node);
10507 256214 : if (nested_in_vect_loop)
10508 : /* We do not support grouped accesses in a nested loop,
10509 : instead the access is contiguous but it might be
10510 : permuted. No gap adjustment is needed though. */
10511 : ;
10512 256212 : else if (ls.slp_perm
10513 256212 : && (group_size != scalar_lanes
10514 11088 : || !multiple_p (nunits, group_size)))
10515 : {
10516 : /* We don't yet generate such SLP_TREE_LOAD_PERMUTATIONs for
10517 : variable VF; see vect_transform_slp_perm_load. */
10518 38707 : unsigned int const_vf = vf.to_constant ();
10519 38707 : unsigned int const_nunits = nunits.to_constant ();
10520 38707 : vec_num = CEIL (group_size * const_vf, const_nunits);
10521 38707 : group_gap_adj = vf * group_size - nunits * vec_num;
10522 : }
10523 : else
10524 : {
10525 217505 : group_gap_adj = group_size - scalar_lanes;
10526 : }
10527 :
10528 256214 : ref_type = get_group_alias_ptr_type (first_stmt_info);
10529 : }
10530 : else
10531 : {
10532 250643 : first_stmt_info = stmt_info;
10533 250643 : first_dr_info = dr_info;
10534 250643 : group_size = 1;
10535 250643 : group_gap_adj = 0;
10536 250643 : ref_type = reference_alias_ptr_type (DR_REF (first_dr_info->dr));
10537 : }
10538 :
10539 506857 : vec_loop_masks *loop_masks
10540 298141 : = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
10541 506857 : ? &LOOP_VINFO_MASKS (loop_vinfo)
10542 31 : : NULL);
10543 31 : vec_loop_lens *loop_lens
10544 298141 : = (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)
10545 : ? &LOOP_VINFO_LENS (loop_vinfo)
10546 0 : : NULL);
10547 :
10548 : /* The vect_transform_stmt and vect_analyze_stmt will go here but there
10549 : are some difference here. We cannot enable both the lens and masks
10550 : during transform but it is allowed during analysis.
10551 : Shouldn't go with length-based approach if fully masked. */
10552 506857 : if (cost_vec == NULL)
10553 : /* The cost_vec is NULL during transfrom. */
10554 161683 : gcc_assert ((!loop_lens || !loop_masks));
10555 :
10556 : /* Targets with store-lane instructions must not require explicit
10557 : realignment. vect_supportable_dr_alignment always returns either
10558 : dr_aligned or dr_unaligned_supported for (non-length) masked
10559 : operations. */
10560 506857 : gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES
10561 : && !mask_node
10562 : && !loop_masks)
10563 : || mat_gather_scatter_p (memory_access_type)
10564 : || alignment_support_scheme == dr_aligned
10565 : || alignment_support_scheme == dr_unaligned_supported);
10566 :
10567 : /* In case the vectorization factor (VF) is bigger than the number
10568 : of elements that we can fit in a vectype (nunits), we have to generate
10569 : more than one vector stmt - i.e - we need to "unroll" the
10570 : vector stmt by a factor VF/nunits. In doing so, we record a pointer
10571 : from one copy of the vector stmt to the next, in the field
10572 : STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
10573 : stages to find the correct vector defs to be used when vectorizing
10574 : stmts that use the defs of the current stmt. The example below
10575 : illustrates the vectorization process when VF=16 and nunits=4 (i.e., we
10576 : need to create 4 vectorized stmts):
10577 :
10578 : before vectorization:
10579 : RELATED_STMT VEC_STMT
10580 : S1: x = memref - -
10581 : S2: z = x + 1 - -
10582 :
10583 : step 1: vectorize stmt S1:
10584 : We first create the vector stmt VS1_0, and, as usual, record a
10585 : pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
10586 : Next, we create the vector stmt VS1_1, and record a pointer to
10587 : it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
10588 : Similarly, for VS1_2 and VS1_3. This is the resulting chain of
10589 : stmts and pointers:
10590 : RELATED_STMT VEC_STMT
10591 : VS1_0: vx0 = memref0 VS1_1 -
10592 : VS1_1: vx1 = memref1 VS1_2 -
10593 : VS1_2: vx2 = memref2 VS1_3 -
10594 : VS1_3: vx3 = memref3 - -
10595 : S1: x = load - VS1_0
10596 : S2: z = x + 1 - -
10597 : */
10598 :
10599 : /* If the data reference is aligned (dr_aligned) or potentially unaligned
10600 : on a target that supports unaligned accesses (dr_unaligned_supported)
10601 : we generate the following code:
10602 : p = initial_addr;
10603 : indx = 0;
10604 : loop {
10605 : p = p + indx * vectype_size;
10606 : vec_dest = *(p);
10607 : indx = indx + 1;
10608 : }
10609 :
10610 : Otherwise, the data reference is potentially unaligned on a target that
10611 : does not support unaligned accesses (dr_explicit_realign_optimized) -
10612 : then generate the following code, in which the data in each iteration is
10613 : obtained by two vector loads, one from the previous iteration, and one
10614 : from the current iteration:
10615 : p1 = initial_addr;
10616 : msq_init = *(floor(p1))
10617 : p2 = initial_addr + VS - 1;
10618 : realignment_token = call target_builtin;
10619 : indx = 0;
10620 : loop {
10621 : p2 = p2 + indx * vectype_size
10622 : lsq = *(floor(p2))
10623 : vec_dest = realign_load (msq, lsq, realignment_token)
10624 : indx = indx + 1;
10625 : msq = lsq;
10626 : } */
10627 :
10628 : /* If the misalignment remains the same throughout the execution of the
10629 : loop, we can create the init_addr and permutation mask at the loop
10630 : preheader. Otherwise, it needs to be created inside the loop.
10631 : This can only occur when vectorizing memory accesses in the inner-loop
10632 : nested within an outer-loop that is being vectorized. */
10633 :
10634 506857 : if (nested_in_vect_loop
10635 506857 : && !multiple_p (DR_STEP_ALIGNMENT (dr_info->dr),
10636 1200 : GET_MODE_SIZE (TYPE_MODE (vectype))))
10637 : {
10638 191 : gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized);
10639 : compute_in_loop = true;
10640 : }
10641 :
10642 506857 : bool diff_first_stmt_info
10643 506857 : = first_stmt_info_for_drptr && first_stmt_info != first_stmt_info_for_drptr;
10644 :
10645 506857 : tree offset = NULL_TREE;
10646 506857 : if ((alignment_support_scheme == dr_explicit_realign_optimized
10647 506857 : || alignment_support_scheme == dr_explicit_realign)
10648 0 : && !compute_in_loop)
10649 : {
10650 : /* If we have different first_stmt_info, we can't set up realignment
10651 : here, since we can't guarantee first_stmt_info DR has been
10652 : initialized yet, use first_stmt_info_for_drptr DR by bumping the
10653 : distance from first_stmt_info DR instead as below. */
10654 0 : if (!costing_p)
10655 : {
10656 0 : if (!diff_first_stmt_info)
10657 0 : msq = vect_setup_realignment (vinfo, first_stmt_info, vectype, gsi,
10658 : &realignment_token,
10659 : alignment_support_scheme, NULL_TREE,
10660 : &at_loop);
10661 0 : if (alignment_support_scheme == dr_explicit_realign_optimized)
10662 : {
10663 0 : phi = as_a<gphi *> (SSA_NAME_DEF_STMT (msq));
10664 0 : offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype),
10665 : size_one_node);
10666 0 : gcc_assert (!first_stmt_info_for_drptr);
10667 : }
10668 : }
10669 : }
10670 : else
10671 506857 : at_loop = loop;
10672 :
10673 506857 : if (!known_eq (poffset, 0))
10674 4384 : offset = (offset
10675 4384 : ? size_binop (PLUS_EXPR, offset, size_int (poffset))
10676 4384 : : size_int (poffset));
10677 :
10678 506857 : tree bump;
10679 506857 : tree vec_offset = NULL_TREE;
10680 :
10681 506857 : auto_vec<tree> vec_offsets;
10682 506857 : auto_vec<tree> vec_masks;
10683 506857 : if (mask_node && !costing_p)
10684 634 : vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[mask_index],
10685 : &vec_masks);
10686 :
10687 506857 : tree vec_mask = NULL_TREE;
10688 506857 : tree vec_els = NULL_TREE;
10689 506857 : if (memory_access_type == VMAT_LOAD_STORE_LANES)
10690 : {
10691 0 : const internal_fn lanes_ifn = ls.lanes_ifn;
10692 :
10693 0 : gcc_assert (alignment_support_scheme == dr_aligned
10694 : || alignment_support_scheme == dr_unaligned_supported);
10695 :
10696 0 : aggr_type = build_array_type_nelts (elem_type, group_size * nunits);
10697 0 : if (!costing_p)
10698 0 : bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, aggr_type,
10699 : memory_access_type, loop_lens);
10700 :
10701 0 : unsigned int inside_cost = 0, prologue_cost = 0;
10702 : /* For costing some adjacent vector loads, we'd like to cost with
10703 : the total number of them once instead of cost each one by one. */
10704 0 : unsigned int n_adjacent_loads = 0;
10705 0 : int ncopies = vec_num / group_size;
10706 0 : for (j = 0; j < ncopies; j++)
10707 : {
10708 0 : if (costing_p)
10709 : {
10710 : /* An IFN_LOAD_LANES will load all its vector results,
10711 : regardless of which ones we actually need. Account
10712 : for the cost of unused results. */
10713 0 : if (first_stmt_info == stmt_info)
10714 : {
10715 0 : unsigned int gaps = DR_GROUP_SIZE (first_stmt_info);
10716 0 : stmt_vec_info next_stmt_info = first_stmt_info;
10717 0 : do
10718 : {
10719 0 : gaps -= 1;
10720 0 : next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
10721 : }
10722 0 : while (next_stmt_info);
10723 0 : if (gaps)
10724 : {
10725 0 : if (dump_enabled_p ())
10726 0 : dump_printf_loc (MSG_NOTE, vect_location,
10727 : "vect_model_load_cost: %d "
10728 : "unused vectors.\n",
10729 : gaps);
10730 0 : vect_get_load_cost (vinfo, stmt_info, slp_node, gaps,
10731 : alignment_support_scheme,
10732 : misalignment, false, &inside_cost,
10733 : &prologue_cost, cost_vec, cost_vec,
10734 : true);
10735 : }
10736 : }
10737 0 : n_adjacent_loads++;
10738 0 : continue;
10739 0 : }
10740 :
10741 : /* 1. Create the vector or array pointer update chain. */
10742 0 : if (j == 0)
10743 0 : dataref_ptr
10744 0 : = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
10745 : at_loop, offset, &dummy, gsi,
10746 : &ptr_incr, false, bump);
10747 : else
10748 : {
10749 0 : gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
10750 0 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
10751 : stmt_info, bump);
10752 : }
10753 0 : if (mask_node)
10754 0 : vec_mask = vec_masks[j];
10755 :
10756 0 : tree vec_array = create_vector_array (vectype, group_size);
10757 :
10758 0 : tree final_mask = NULL_TREE;
10759 0 : tree final_len = NULL_TREE;
10760 0 : tree bias = NULL_TREE;
10761 0 : if (loop_masks)
10762 0 : final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
10763 : ncopies, vectype, j);
10764 0 : if (vec_mask)
10765 0 : final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
10766 : vec_mask, gsi);
10767 :
10768 0 : if (lanes_ifn == IFN_MASK_LEN_LOAD_LANES)
10769 : {
10770 0 : if (loop_lens)
10771 0 : final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
10772 : ncopies, vectype, j, 1, true);
10773 : else
10774 0 : final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
10775 0 : signed char biasval
10776 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
10777 0 : bias = build_int_cst (intQI_type_node, biasval);
10778 0 : if (!final_mask)
10779 : {
10780 0 : mask_vectype = truth_type_for (vectype);
10781 0 : final_mask = build_minus_one_cst (mask_vectype);
10782 : }
10783 : }
10784 :
10785 0 : if (final_mask)
10786 : {
10787 0 : vec_els = vect_get_mask_load_else (maskload_elsval, vectype);
10788 0 : if (type_mode_padding_p
10789 0 : && maskload_elsval != MASK_LOAD_ELSE_ZERO)
10790 0 : need_zeroing = true;
10791 : }
10792 :
10793 0 : gcall *call;
10794 0 : if (final_len && final_mask)
10795 : {
10796 : /* Emit:
10797 : VEC_ARRAY = MASK_LEN_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
10798 : VEC_MASK, LEN, BIAS). */
10799 0 : unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
10800 0 : tree alias_ptr = build_int_cst (ref_type, align);
10801 0 : call = gimple_build_call_internal (IFN_MASK_LEN_LOAD_LANES, 6,
10802 : dataref_ptr, alias_ptr,
10803 : final_mask, vec_els,
10804 : final_len, bias);
10805 : }
10806 0 : else if (final_mask)
10807 : {
10808 : /* Emit:
10809 : VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
10810 : VEC_MASK). */
10811 0 : unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype));
10812 0 : tree alias_ptr = build_int_cst (ref_type, align);
10813 0 : call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 4,
10814 : dataref_ptr, alias_ptr,
10815 : final_mask, vec_els);
10816 : }
10817 : else
10818 : {
10819 : /* Emit:
10820 : VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */
10821 0 : data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type);
10822 0 : call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref);
10823 : }
10824 0 : gimple_call_set_lhs (call, vec_array);
10825 0 : gimple_call_set_nothrow (call, true);
10826 0 : vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
10827 :
10828 : /* Extract each vector into an SSA_NAME. */
10829 0 : for (unsigned i = 0; i < group_size; i++)
10830 : {
10831 0 : new_temp = read_vector_array (vinfo, stmt_info, gsi, scalar_dest,
10832 : vec_array, i, need_zeroing,
10833 : final_mask);
10834 0 : slp_node->push_vec_def (new_temp);
10835 : }
10836 :
10837 : /* Record that VEC_ARRAY is now dead. */
10838 0 : vect_clobber_variable (vinfo, stmt_info, gsi, vec_array);
10839 : }
10840 :
10841 0 : if (costing_p)
10842 : {
10843 0 : if (n_adjacent_loads > 0)
10844 0 : vect_get_load_cost (vinfo, stmt_info, slp_node, n_adjacent_loads,
10845 : alignment_support_scheme, misalignment, false,
10846 : &inside_cost, &prologue_cost, cost_vec,
10847 : cost_vec, true);
10848 0 : if (dump_enabled_p ())
10849 0 : dump_printf_loc (MSG_NOTE, vect_location,
10850 : "vect_model_load_cost: inside_cost = %u, "
10851 : "prologue_cost = %u .\n",
10852 : inside_cost, prologue_cost);
10853 : }
10854 :
10855 0 : return true;
10856 : }
10857 :
10858 506857 : if (mat_gather_scatter_p (memory_access_type))
10859 : {
10860 2727 : gcc_assert ((!grouped_load && !ls.slp_perm) || ls.ls_type);
10861 :
10862 2727 : auto_vec<tree> dr_chain (vec_num);
10863 :
10864 : /* If we pun the original vectype the loads as well as costing, length,
10865 : etc. is performed with the new type. After loading we VIEW_CONVERT
10866 : the data to the original vectype. */
10867 2727 : tree original_vectype = vectype;
10868 2727 : if (ls.ls_type)
10869 0 : vectype = ls.ls_type;
10870 :
10871 : /* 1. Create the vector or array pointer update chain. */
10872 2727 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
10873 : {
10874 2727 : aggr_type = NULL_TREE;
10875 2727 : bump = NULL_TREE;
10876 2727 : if (!costing_p)
10877 749 : vect_get_gather_scatter_ops (loop, slp_node, &dataref_ptr,
10878 : &vec_offsets);
10879 : }
10880 : else
10881 : {
10882 0 : aggr_type = elem_type;
10883 0 : if (!costing_p)
10884 : {
10885 0 : vect_get_strided_load_store_ops (stmt_info, slp_node, vectype,
10886 : ls.strided_offset_vectype,
10887 : loop_vinfo, gsi,
10888 : &bump, &vec_offset, loop_lens);
10889 0 : dataref_ptr
10890 0 : = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
10891 : at_loop, offset, &dummy, gsi,
10892 : &ptr_incr, false, bump);
10893 : }
10894 : }
10895 :
10896 : unsigned int inside_cost = 0, prologue_cost = 0;
10897 :
10898 6147 : gimple *new_stmt = NULL;
10899 6147 : for (i = 0; i < vec_num; i++)
10900 : {
10901 3420 : tree final_mask = NULL_TREE;
10902 3420 : tree final_len = NULL_TREE;
10903 3420 : tree bias = NULL_TREE;
10904 3420 : if (!costing_p)
10905 : {
10906 963 : if (mask_node)
10907 153 : vec_mask = vec_masks[i];
10908 963 : if (loop_masks)
10909 0 : final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
10910 : vec_num, vectype, i);
10911 963 : if (vec_mask)
10912 153 : final_mask = prepare_vec_mask (loop_vinfo, mask_vectype,
10913 : final_mask, vec_mask, gsi);
10914 :
10915 963 : if (i > 0 && !STMT_VINFO_GATHER_SCATTER_P (stmt_info))
10916 0 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr,
10917 : gsi, stmt_info, bump);
10918 : }
10919 :
10920 : /* 2. Create the vector-load in the loop. */
10921 3420 : unsigned align = get_object_alignment (DR_REF (first_dr_info->dr));
10922 3420 : tree alias_align_ptr = build_int_cst (ref_type, align);
10923 3420 : if (memory_access_type == VMAT_GATHER_SCATTER_IFN)
10924 : {
10925 0 : if (costing_p)
10926 : {
10927 0 : if (ls.supported_offset_vectype)
10928 0 : inside_cost
10929 0 : += record_stmt_cost (cost_vec, 1, vector_stmt,
10930 : slp_node, 0, vect_body);
10931 0 : if (ls.supported_scale)
10932 0 : inside_cost
10933 0 : += record_stmt_cost (cost_vec, 1, vector_stmt,
10934 : slp_node, 0, vect_body);
10935 :
10936 0 : unsigned int cnunits = vect_nunits_for_cost (vectype);
10937 0 : inside_cost
10938 0 : = record_stmt_cost (cost_vec, cnunits, scalar_load,
10939 : slp_node, 0, vect_body);
10940 3420 : continue;
10941 0 : }
10942 0 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
10943 0 : vec_offset = vec_offsets[i];
10944 0 : tree zero = build_zero_cst (vectype);
10945 0 : tree scale = size_int (SLP_TREE_GS_SCALE (slp_node));
10946 0 : bool strided = !VECTOR_TYPE_P (TREE_TYPE (vec_offset));
10947 :
10948 : /* Perform the offset conversion and scaling if necessary. */
10949 0 : if (!strided
10950 0 : && (ls.supported_offset_vectype || ls.supported_scale))
10951 : {
10952 0 : gimple_seq stmts = NULL;
10953 0 : if (ls.supported_offset_vectype)
10954 0 : vec_offset = gimple_convert
10955 0 : (&stmts, ls.supported_offset_vectype, vec_offset);
10956 0 : if (ls.supported_scale)
10957 : {
10958 : /* Only scale the vec_offset if we haven't already. */
10959 0 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
10960 0 : || i == 0)
10961 : {
10962 0 : tree mult_cst = build_int_cst
10963 0 : (TREE_TYPE (TREE_TYPE (vec_offset)),
10964 0 : SLP_TREE_GS_SCALE (slp_node) / ls.supported_scale);
10965 0 : tree mult = build_vector_from_val
10966 0 : (TREE_TYPE (vec_offset), mult_cst);
10967 0 : vec_offset = gimple_build
10968 0 : (&stmts, MULT_EXPR, TREE_TYPE (vec_offset),
10969 : vec_offset, mult);
10970 : }
10971 0 : scale = size_int (ls.supported_scale);
10972 : }
10973 0 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
10974 : }
10975 :
10976 0 : if (ls.gs.ifn == IFN_MASK_LEN_GATHER_LOAD)
10977 : {
10978 0 : if (loop_lens)
10979 0 : final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
10980 : vec_num, vectype, i, 1, true);
10981 : else
10982 0 : final_len = build_int_cst (sizetype,
10983 0 : TYPE_VECTOR_SUBPARTS (vectype));
10984 0 : signed char biasval
10985 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
10986 0 : bias = build_int_cst (intQI_type_node, biasval);
10987 0 : if (!final_mask)
10988 : {
10989 0 : mask_vectype = truth_type_for (vectype);
10990 0 : final_mask = build_minus_one_cst (mask_vectype);
10991 : }
10992 : }
10993 :
10994 0 : if (final_mask)
10995 : {
10996 0 : vec_els = vect_get_mask_load_else (maskload_elsval, vectype);
10997 0 : if (type_mode_padding_p
10998 0 : && maskload_elsval != MASK_LOAD_ELSE_ZERO)
10999 0 : need_zeroing = true;
11000 : }
11001 :
11002 0 : gcall *call;
11003 0 : if (final_len && final_mask)
11004 : {
11005 0 : if (VECTOR_TYPE_P (TREE_TYPE (vec_offset)))
11006 0 : call = gimple_build_call_internal (IFN_MASK_LEN_GATHER_LOAD,
11007 : 9, dataref_ptr,
11008 : alias_align_ptr,
11009 : vec_offset, scale, zero,
11010 : final_mask, vec_els,
11011 : final_len, bias);
11012 : else
11013 : /* Non-vector offset indicates that prefer to take
11014 : MASK_LEN_STRIDED_LOAD instead of the
11015 : MASK_LEN_GATHER_LOAD with direct stride arg. */
11016 0 : call = gimple_build_call_internal
11017 0 : (IFN_MASK_LEN_STRIDED_LOAD, 7, dataref_ptr,
11018 : vec_offset, zero, final_mask, vec_els, final_len,
11019 : bias);
11020 : }
11021 0 : else if (final_mask)
11022 0 : call = gimple_build_call_internal (IFN_MASK_GATHER_LOAD,
11023 : 7, dataref_ptr,
11024 : alias_align_ptr,
11025 : vec_offset, scale,
11026 : zero, final_mask, vec_els);
11027 : else
11028 0 : call = gimple_build_call_internal (IFN_GATHER_LOAD, 5,
11029 : dataref_ptr,
11030 : alias_align_ptr,
11031 : vec_offset, scale, zero);
11032 0 : gimple_call_set_nothrow (call, true);
11033 0 : new_stmt = call;
11034 0 : data_ref = NULL_TREE;
11035 : }
11036 3420 : else if (memory_access_type == VMAT_GATHER_SCATTER_LEGACY)
11037 : {
11038 : /* The builtin decls path for gather is legacy, x86 only. */
11039 570 : gcc_assert (!final_len && nunits.is_constant ());
11040 570 : if (costing_p)
11041 : {
11042 287 : unsigned int cnunits = vect_nunits_for_cost (vectype);
11043 287 : inside_cost
11044 287 : = record_stmt_cost (cost_vec, cnunits, scalar_load,
11045 : slp_node, 0, vect_body);
11046 287 : continue;
11047 287 : }
11048 283 : tree offset_vectype = TREE_TYPE (vec_offsets[0]);
11049 283 : poly_uint64 offset_nunits = TYPE_VECTOR_SUBPARTS (offset_vectype);
11050 283 : if (known_eq (nunits, offset_nunits))
11051 : {
11052 134 : new_stmt = vect_build_one_gather_load_call
11053 134 : (vinfo, stmt_info, slp_node, vectype, gsi,
11054 134 : ls.gs.decl, dataref_ptr, vec_offsets[i],
11055 : final_mask);
11056 134 : data_ref = NULL_TREE;
11057 : }
11058 149 : else if (known_eq (nunits, offset_nunits * 2))
11059 : {
11060 : /* We have a offset vector with half the number of
11061 : lanes but the builtins will produce full vectype
11062 : data with just the lower lanes filled. */
11063 63 : new_stmt = vect_build_one_gather_load_call
11064 126 : (vinfo, stmt_info, slp_node, vectype, gsi,
11065 63 : ls.gs.decl, dataref_ptr, vec_offsets[2 * i],
11066 : final_mask);
11067 63 : tree low = make_ssa_name (vectype);
11068 63 : gimple_set_lhs (new_stmt, low);
11069 63 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11070 :
11071 : /* now put upper half of final_mask in final_mask low. */
11072 63 : if (final_mask
11073 63 : && !SCALAR_INT_MODE_P (TYPE_MODE (TREE_TYPE (final_mask))))
11074 : {
11075 11 : int count = nunits.to_constant ();
11076 11 : vec_perm_builder sel (count, count, 1);
11077 11 : sel.quick_grow (count);
11078 87 : for (int i = 0; i < count; ++i)
11079 76 : sel[i] = i | (count / 2);
11080 11 : vec_perm_indices indices (sel, 2, count);
11081 11 : tree perm_mask = vect_gen_perm_mask_checked
11082 11 : (TREE_TYPE (final_mask), indices);
11083 11 : new_stmt = gimple_build_assign (NULL_TREE, VEC_PERM_EXPR,
11084 : final_mask, final_mask,
11085 : perm_mask);
11086 11 : final_mask = make_ssa_name (TREE_TYPE (final_mask));
11087 11 : gimple_set_lhs (new_stmt, final_mask);
11088 11 : vect_finish_stmt_generation (vinfo, stmt_info,
11089 : new_stmt, gsi);
11090 11 : }
11091 52 : else if (final_mask)
11092 : {
11093 24 : new_stmt = gimple_build_assign (NULL_TREE,
11094 : VEC_UNPACK_HI_EXPR,
11095 : final_mask);
11096 24 : final_mask = make_ssa_name
11097 24 : (truth_type_for (offset_vectype));
11098 24 : gimple_set_lhs (new_stmt, final_mask);
11099 24 : vect_finish_stmt_generation (vinfo, stmt_info,
11100 : new_stmt, gsi);
11101 : }
11102 :
11103 63 : new_stmt = vect_build_one_gather_load_call
11104 126 : (vinfo, stmt_info, slp_node, vectype, gsi,
11105 : ls.gs.decl, dataref_ptr,
11106 63 : vec_offsets[2 * i + 1], final_mask);
11107 63 : tree high = make_ssa_name (vectype);
11108 63 : gimple_set_lhs (new_stmt, high);
11109 63 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11110 :
11111 : /* compose low + high. */
11112 63 : int count = nunits.to_constant ();
11113 63 : vec_perm_builder sel (count, count, 1);
11114 63 : sel.quick_grow (count);
11115 647 : for (int i = 0; i < count; ++i)
11116 584 : sel[i] = i < count / 2 ? i : i + count / 2;
11117 63 : vec_perm_indices indices (sel, 2, count);
11118 63 : tree perm_mask
11119 63 : = vect_gen_perm_mask_checked (vectype, indices);
11120 63 : new_stmt = gimple_build_assign (NULL_TREE, VEC_PERM_EXPR,
11121 : low, high, perm_mask);
11122 63 : data_ref = NULL_TREE;
11123 63 : }
11124 86 : else if (known_eq (nunits * 2, offset_nunits))
11125 : {
11126 : /* We have a offset vector with double the number of
11127 : lanes. Select the low/high part accordingly. */
11128 86 : vec_offset = vec_offsets[i / 2];
11129 86 : if (i & 1)
11130 : {
11131 43 : int count = offset_nunits.to_constant ();
11132 43 : vec_perm_builder sel (count, count, 1);
11133 43 : sel.quick_grow (count);
11134 463 : for (int i = 0; i < count; ++i)
11135 420 : sel[i] = i | (count / 2);
11136 43 : vec_perm_indices indices (sel, 2, count);
11137 43 : tree perm_mask = vect_gen_perm_mask_checked
11138 43 : (TREE_TYPE (vec_offset), indices);
11139 43 : new_stmt = gimple_build_assign (NULL_TREE, VEC_PERM_EXPR,
11140 : vec_offset, vec_offset,
11141 : perm_mask);
11142 43 : vec_offset = make_ssa_name (TREE_TYPE (vec_offset));
11143 43 : gimple_set_lhs (new_stmt, vec_offset);
11144 43 : vect_finish_stmt_generation (vinfo, stmt_info,
11145 : new_stmt, gsi);
11146 43 : }
11147 86 : new_stmt = vect_build_one_gather_load_call
11148 86 : (vinfo, stmt_info, slp_node, vectype, gsi,
11149 : ls.gs.decl,
11150 : dataref_ptr, vec_offset, final_mask);
11151 86 : data_ref = NULL_TREE;
11152 : }
11153 : else
11154 0 : gcc_unreachable ();
11155 : }
11156 : else
11157 : {
11158 : /* Emulated gather-scatter. */
11159 2850 : gcc_assert (!final_mask);
11160 2850 : unsigned HOST_WIDE_INT const_nunits = nunits.to_constant ();
11161 2850 : if (costing_p)
11162 : {
11163 : /* For emulated gathers N offset vector element
11164 : offset add is consumed by the load). */
11165 2170 : inside_cost = record_stmt_cost (cost_vec, const_nunits,
11166 : vec_to_scalar,
11167 : slp_node, 0, vect_body);
11168 : /* N scalar loads plus gathering them into a
11169 : vector. */
11170 2170 : inside_cost
11171 2170 : = record_stmt_cost (cost_vec, const_nunits, scalar_load,
11172 : slp_node, 0, vect_body);
11173 2170 : inside_cost
11174 2170 : = record_stmt_cost (cost_vec, 1, vec_construct,
11175 : slp_node, 0, vect_body);
11176 2170 : continue;
11177 : }
11178 680 : tree offset_vectype = TREE_TYPE (vec_offsets[0]);
11179 680 : unsigned HOST_WIDE_INT const_offset_nunits
11180 680 : = TYPE_VECTOR_SUBPARTS (offset_vectype).to_constant ();
11181 680 : vec<constructor_elt, va_gc> *ctor_elts;
11182 680 : vec_alloc (ctor_elts, const_nunits);
11183 680 : gimple_seq stmts = NULL;
11184 : /* We support offset vectors with more elements
11185 : than the data vector for now. */
11186 680 : unsigned HOST_WIDE_INT factor
11187 : = const_offset_nunits / const_nunits;
11188 680 : vec_offset = vec_offsets[i / factor];
11189 680 : unsigned elt_offset = (i % factor) * const_nunits;
11190 680 : tree idx_type = TREE_TYPE (TREE_TYPE (vec_offset));
11191 680 : tree scale = size_int (SLP_TREE_GS_SCALE (slp_node));
11192 680 : tree ltype = build_aligned_type (TREE_TYPE (vectype), align);
11193 2828 : for (unsigned k = 0; k < const_nunits; ++k)
11194 : {
11195 2148 : tree boff = size_binop (MULT_EXPR, TYPE_SIZE (idx_type),
11196 : bitsize_int (k + elt_offset));
11197 6444 : tree idx = gimple_build (&stmts, BIT_FIELD_REF, idx_type,
11198 2148 : vec_offset, TYPE_SIZE (idx_type),
11199 : boff);
11200 2148 : idx = gimple_convert (&stmts, sizetype, idx);
11201 2148 : idx = gimple_build (&stmts, MULT_EXPR, sizetype, idx, scale);
11202 2148 : tree ptr = gimple_build (&stmts, PLUS_EXPR,
11203 2148 : TREE_TYPE (dataref_ptr),
11204 : dataref_ptr, idx);
11205 2148 : ptr = gimple_convert (&stmts, ptr_type_node, ptr);
11206 2148 : tree elt = make_ssa_name (TREE_TYPE (vectype));
11207 2148 : tree ref = build2 (MEM_REF, ltype, ptr,
11208 : build_int_cst (ref_type, 0));
11209 2148 : new_stmt = gimple_build_assign (elt, ref);
11210 4296 : gimple_set_vuse (new_stmt, gimple_vuse (gsi_stmt (*gsi)));
11211 2148 : gimple_seq_add_stmt (&stmts, new_stmt);
11212 2148 : CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, elt);
11213 : }
11214 680 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
11215 680 : new_stmt = gimple_build_assign (NULL_TREE,
11216 : build_constructor (vectype,
11217 : ctor_elts));
11218 680 : data_ref = NULL_TREE;
11219 : }
11220 :
11221 963 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
11222 : /* DATA_REF is null if we've already built the statement. */
11223 963 : if (data_ref)
11224 : {
11225 : vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr));
11226 : new_stmt = gimple_build_assign (vec_dest, data_ref);
11227 : }
11228 1926 : new_temp = (need_zeroing
11229 963 : ? make_ssa_name (vectype)
11230 963 : : make_ssa_name (vec_dest, new_stmt));
11231 963 : gimple_set_lhs (new_stmt, new_temp);
11232 963 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11233 :
11234 : /* If we need to explicitly zero inactive elements emit a
11235 : VEC_COND_EXPR that does so. */
11236 963 : if (need_zeroing)
11237 : {
11238 0 : vec_els = vect_get_mask_load_else (MASK_LOAD_ELSE_ZERO,
11239 : vectype);
11240 :
11241 0 : tree new_temp2 = make_ssa_name (vec_dest, new_stmt);
11242 0 : new_stmt = gimple_build_assign (new_temp2, VEC_COND_EXPR,
11243 : final_mask, new_temp, vec_els);
11244 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11245 0 : new_temp = new_temp2;
11246 : }
11247 :
11248 963 : if (ls.ls_type)
11249 : {
11250 0 : new_stmt = gimple_build_assign (make_ssa_name
11251 : (original_vectype),
11252 : VIEW_CONVERT_EXPR,
11253 : build1 (VIEW_CONVERT_EXPR,
11254 : original_vectype,
11255 : new_temp));
11256 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11257 : }
11258 :
11259 : /* Store vector loads in the corresponding SLP_NODE. */
11260 963 : if (!costing_p)
11261 : {
11262 963 : if (ls.slp_perm)
11263 0 : dr_chain.quick_push (gimple_assign_lhs (new_stmt));
11264 : else
11265 963 : slp_node->push_vec_def (new_stmt);
11266 : }
11267 : }
11268 :
11269 2727 : if (ls.slp_perm)
11270 : {
11271 0 : if (costing_p)
11272 : {
11273 0 : gcc_assert (ls.n_perms != -1U);
11274 0 : inside_cost += record_stmt_cost (cost_vec, ls.n_perms, vec_perm,
11275 : slp_node, 0, vect_body);
11276 : }
11277 : else
11278 : {
11279 0 : unsigned n_perms2;
11280 0 : vect_transform_slp_perm_load (vinfo, slp_node, dr_chain, gsi, vf,
11281 : false, &n_perms2);
11282 0 : gcc_assert (ls.n_perms == n_perms2);
11283 : }
11284 : }
11285 :
11286 2727 : if (costing_p && dump_enabled_p ())
11287 222 : dump_printf_loc (MSG_NOTE, vect_location,
11288 : "vect_model_load_cost: inside_cost = %u, "
11289 : "prologue_cost = %u .\n",
11290 : inside_cost, prologue_cost);
11291 2727 : return true;
11292 2727 : }
11293 :
11294 504130 : aggr_type = vectype;
11295 504130 : if (!costing_p)
11296 160934 : bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, aggr_type,
11297 : memory_access_type, loop_lens);
11298 :
11299 504130 : poly_uint64 group_elt = 0;
11300 504130 : unsigned int inside_cost = 0, prologue_cost = 0;
11301 : /* For costing some adjacent vector loads, we'd like to cost with
11302 : the total number of them once instead of cost each one by one. */
11303 504130 : unsigned int n_adjacent_loads = 0;
11304 :
11305 : /* 1. Create the vector or array pointer update chain. */
11306 504130 : if (!costing_p)
11307 : {
11308 160934 : bool simd_lane_access_p
11309 160934 : = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0;
11310 160934 : if (simd_lane_access_p
11311 1629 : && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR
11312 1629 : && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0))
11313 1629 : && integer_zerop (get_dr_vinfo_offset (vinfo, first_dr_info))
11314 1629 : && integer_zerop (DR_INIT (first_dr_info->dr))
11315 1629 : && alias_sets_conflict_p (get_alias_set (aggr_type),
11316 1629 : get_alias_set (TREE_TYPE (ref_type)))
11317 160934 : && (alignment_support_scheme == dr_aligned
11318 1629 : || alignment_support_scheme == dr_unaligned_supported))
11319 : {
11320 1629 : dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr));
11321 1629 : dataref_offset = build_int_cst (ref_type, 0);
11322 : }
11323 159305 : else if (diff_first_stmt_info)
11324 : {
11325 3478 : dataref_ptr
11326 3478 : = vect_create_data_ref_ptr (vinfo, first_stmt_info_for_drptr,
11327 : aggr_type, at_loop, offset, &dummy,
11328 : gsi, &ptr_incr, simd_lane_access_p,
11329 : bump);
11330 : /* Adjust the pointer by the difference to first_stmt. */
11331 3478 : data_reference_p ptrdr
11332 : = STMT_VINFO_DATA_REF (first_stmt_info_for_drptr);
11333 3478 : tree diff = fold_convert (sizetype,
11334 : size_binop (MINUS_EXPR,
11335 : DR_INIT (first_dr_info->dr),
11336 : DR_INIT (ptrdr)));
11337 3478 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
11338 : stmt_info, diff);
11339 3478 : if (alignment_support_scheme == dr_explicit_realign)
11340 : {
11341 0 : msq = vect_setup_realignment (vinfo, first_stmt_info_for_drptr,
11342 : vectype, gsi,
11343 : &realignment_token,
11344 : alignment_support_scheme,
11345 : dataref_ptr, &at_loop);
11346 0 : gcc_assert (!compute_in_loop);
11347 : }
11348 : }
11349 : else
11350 155827 : dataref_ptr
11351 155827 : = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
11352 : at_loop,
11353 : offset, &dummy, gsi, &ptr_incr,
11354 : simd_lane_access_p, bump);
11355 : }
11356 : else if (!costing_p)
11357 : {
11358 : gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
11359 : if (dataref_offset)
11360 : dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump);
11361 : else
11362 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
11363 : stmt_info, bump);
11364 : }
11365 :
11366 504130 : auto_vec<tree> dr_chain;
11367 504130 : if (grouped_load || ls.slp_perm)
11368 48692 : dr_chain.create (vec_num);
11369 :
11370 : gimple *new_stmt = NULL;
11371 1334175 : for (i = 0; i < vec_num; i++)
11372 : {
11373 830045 : tree final_mask = NULL_TREE;
11374 830045 : tree final_len = NULL_TREE;
11375 830045 : tree bias = NULL_TREE;
11376 :
11377 830045 : if (!costing_p)
11378 : {
11379 251108 : if (mask_node)
11380 707 : vec_mask = vec_masks[i];
11381 251108 : if (loop_masks)
11382 48 : final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
11383 : vec_num, vectype, i);
11384 251108 : if (vec_mask)
11385 707 : final_mask = prepare_vec_mask (loop_vinfo, mask_vectype,
11386 : final_mask, vec_mask, gsi);
11387 :
11388 251108 : if (i > 0)
11389 90174 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr,
11390 : gsi, stmt_info, bump);
11391 : }
11392 :
11393 : /* 2. Create the vector-load in the loop. */
11394 830045 : switch (alignment_support_scheme)
11395 : {
11396 830045 : case dr_aligned:
11397 830045 : case dr_unaligned_supported:
11398 830045 : {
11399 830045 : if (costing_p)
11400 : break;
11401 :
11402 251108 : unsigned int misalign;
11403 251108 : unsigned HOST_WIDE_INT align;
11404 251108 : align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info));
11405 251108 : if (alignment_support_scheme == dr_aligned)
11406 : misalign = 0;
11407 162843 : else if (misalignment == DR_MISALIGNMENT_UNKNOWN)
11408 : {
11409 124014 : align = dr_alignment (vect_dr_behavior (vinfo, first_dr_info));
11410 124014 : misalign = 0;
11411 : }
11412 : else
11413 38829 : misalign = misalignment;
11414 251108 : if (dataref_offset == NULL_TREE
11415 248981 : && TREE_CODE (dataref_ptr) == SSA_NAME)
11416 169910 : set_ptr_info_alignment (get_ptr_info (dataref_ptr), align,
11417 : misalign);
11418 251108 : align = least_bit_hwi (misalign | align);
11419 :
11420 : /* Compute IFN when LOOP_LENS or final_mask valid. */
11421 251108 : machine_mode vmode = TYPE_MODE (vectype);
11422 251108 : machine_mode new_vmode = vmode;
11423 251108 : internal_fn partial_ifn = IFN_LAST;
11424 251108 : if (loop_lens)
11425 : {
11426 0 : opt_machine_mode new_ovmode
11427 0 : = get_len_load_store_mode (vmode, true, &partial_ifn);
11428 0 : new_vmode = new_ovmode.require ();
11429 0 : unsigned factor
11430 0 : = (new_ovmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vmode);
11431 0 : final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
11432 : vec_num, vectype, i, factor, true);
11433 : }
11434 251108 : else if (final_mask)
11435 : {
11436 735 : if (!can_vec_mask_load_store_p (vmode,
11437 735 : TYPE_MODE
11438 : (TREE_TYPE (final_mask)),
11439 : true, &partial_ifn))
11440 0 : gcc_unreachable ();
11441 : }
11442 :
11443 251108 : if (partial_ifn == IFN_MASK_LEN_LOAD)
11444 : {
11445 0 : if (!final_len)
11446 : {
11447 : /* Pass VF value to 'len' argument of
11448 : MASK_LEN_LOAD if LOOP_LENS is invalid. */
11449 0 : final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
11450 : }
11451 0 : if (!final_mask)
11452 : {
11453 : /* Pass all ones value to 'mask' argument of
11454 : MASK_LEN_LOAD if final_mask is invalid. */
11455 0 : mask_vectype = truth_type_for (vectype);
11456 0 : final_mask = build_minus_one_cst (mask_vectype);
11457 : }
11458 : }
11459 251108 : if (final_len)
11460 : {
11461 0 : signed char biasval
11462 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
11463 0 : bias = build_int_cst (intQI_type_node, biasval);
11464 : }
11465 :
11466 251108 : tree vec_els;
11467 :
11468 251108 : if (final_len)
11469 : {
11470 0 : tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
11471 0 : gcall *call;
11472 :
11473 : /* Need conversion if the vectype is punned by VnQI. */
11474 0 : els_vectype = vectype;
11475 0 : if (vmode != new_vmode)
11476 0 : els_vectype
11477 0 : = build_vector_type_for_mode (unsigned_intQI_type_node,
11478 : new_vmode);
11479 0 : vec_els = vect_get_mask_load_else (maskload_elsval,
11480 : els_vectype);
11481 :
11482 0 : if (partial_ifn == IFN_MASK_LEN_LOAD)
11483 : {
11484 0 : if (type_mode_padding_p
11485 0 : && maskload_elsval != MASK_LOAD_ELSE_ZERO)
11486 0 : need_zeroing = true;
11487 0 : call = gimple_build_call_internal (IFN_MASK_LEN_LOAD,
11488 : 6, dataref_ptr, ptr,
11489 : final_mask, vec_els,
11490 : final_len, bias);
11491 : }
11492 : else
11493 0 : call = gimple_build_call_internal (IFN_LEN_LOAD, 5,
11494 : dataref_ptr, ptr,
11495 : vec_els, final_len,
11496 : bias);
11497 0 : gimple_call_set_nothrow (call, true);
11498 0 : new_stmt = call;
11499 0 : data_ref = NULL_TREE;
11500 :
11501 : /* Need conversion if it's wrapped with VnQI. */
11502 0 : if (vmode != new_vmode)
11503 : {
11504 0 : tree new_vtype
11505 0 : = build_vector_type_for_mode (unsigned_intQI_type_node,
11506 : new_vmode);
11507 0 : tree var = vect_get_new_ssa_name (new_vtype,
11508 : vect_simple_var);
11509 0 : gimple_set_lhs (call, var);
11510 0 : vect_finish_stmt_generation (vinfo, stmt_info, call,
11511 : gsi);
11512 0 : tree op = build1 (VIEW_CONVERT_EXPR, vectype, var);
11513 0 : new_stmt = gimple_build_assign (vec_dest,
11514 : VIEW_CONVERT_EXPR, op);
11515 : }
11516 : }
11517 251108 : else if (final_mask)
11518 : {
11519 735 : tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
11520 735 : vec_els = vect_get_mask_load_else (maskload_elsval, vectype);
11521 735 : if (type_mode_padding_p
11522 735 : && maskload_elsval != MASK_LOAD_ELSE_ZERO)
11523 0 : need_zeroing = true;
11524 735 : gcall *call = gimple_build_call_internal (IFN_MASK_LOAD, 4,
11525 : dataref_ptr, ptr,
11526 : final_mask,
11527 : vec_els);
11528 735 : gimple_call_set_nothrow (call, true);
11529 735 : new_stmt = call;
11530 735 : data_ref = NULL_TREE;
11531 : }
11532 : else
11533 : {
11534 250373 : tree ltype = vectype;
11535 250373 : tree new_vtype = NULL_TREE;
11536 250373 : unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (first_stmt_info);
11537 250373 : unsigned HOST_WIDE_INT dr_size
11538 250373 : = vect_get_scalar_dr_size (first_dr_info);
11539 250373 : poly_int64 off = 0;
11540 250373 : if (memory_access_type == VMAT_CONTIGUOUS_REVERSE)
11541 1431 : off = (TYPE_VECTOR_SUBPARTS (vectype) - 1) * -dr_size;
11542 250373 : unsigned int vect_align
11543 250373 : = vect_known_alignment_in_bytes (first_dr_info, vectype,
11544 250373 : off);
11545 : /* Try to use a single smaller load when we are about
11546 : to load excess elements compared to the unrolled
11547 : scalar loop. */
11548 250373 : if (known_gt ((i + 1) * nunits,
11549 : (group_size * vf - gap)))
11550 : {
11551 6850 : poly_uint64 remain = ((group_size * vf - gap) - i * nunits);
11552 6850 : if (known_ge ((i + 1) * nunits - (group_size * vf - gap),
11553 : nunits))
11554 : /* DR will be unused. */
11555 : ltype = NULL_TREE;
11556 2179 : else if (known_ge (vect_align,
11557 : tree_to_poly_uint64
11558 : (TYPE_SIZE_UNIT (vectype))))
11559 : /* Aligned access to excess elements is OK if
11560 : at least one element is accessed in the
11561 : scalar loop. */
11562 : ;
11563 1840 : else if (known_gt (vect_align,
11564 : ((nunits - remain) * dr_size)))
11565 : /* Aligned access to the gap area when there's
11566 : at least one element in it is OK. */
11567 : ;
11568 : else
11569 : {
11570 : /* remain should now be > 0 and < nunits. */
11571 1837 : unsigned num;
11572 1837 : if (known_ne (remain, 0u)
11573 1837 : && constant_multiple_p (nunits, remain, &num))
11574 : {
11575 1379 : tree ptype;
11576 1379 : new_vtype
11577 1379 : = vector_vector_composition_type (vectype, num,
11578 : &ptype);
11579 1379 : if (new_vtype)
11580 1379 : ltype = ptype;
11581 : }
11582 : /* Else use multiple loads or a masked load? */
11583 : /* For loop vectorization we now should have
11584 : an alternate type or LOOP_VINFO_PEELING_FOR_GAPS
11585 : set. */
11586 1837 : if (loop_vinfo)
11587 1598 : gcc_assert (new_vtype
11588 : || LOOP_VINFO_PEELING_FOR_GAPS
11589 : (loop_vinfo));
11590 : /* But still reduce the access size to the next
11591 : required power-of-two so peeling a single
11592 : scalar iteration is sufficient. */
11593 1837 : unsigned HOST_WIDE_INT cremain;
11594 1837 : if (remain.is_constant (&cremain))
11595 : {
11596 1837 : unsigned HOST_WIDE_INT cpart_size
11597 1837 : = 1 << ceil_log2 (cremain);
11598 1837 : if (known_gt (nunits, cpart_size)
11599 1837 : && constant_multiple_p (nunits, cpart_size,
11600 : &num))
11601 : {
11602 1391 : tree ptype;
11603 1391 : new_vtype
11604 2782 : = vector_vector_composition_type (vectype,
11605 1391 : num,
11606 : &ptype);
11607 1391 : if (new_vtype)
11608 1391 : ltype = ptype;
11609 : }
11610 : }
11611 : }
11612 : }
11613 250373 : tree offset = (dataref_offset ? dataref_offset
11614 248246 : : build_int_cst (ref_type, 0));
11615 250373 : if (!ltype)
11616 : ;
11617 245702 : else if (ltype != vectype
11618 245702 : && memory_access_type == VMAT_CONTIGUOUS_REVERSE)
11619 : {
11620 21 : poly_uint64 gap_offset
11621 21 : = (tree_to_poly_uint64 (TYPE_SIZE_UNIT (vectype))
11622 21 : - tree_to_poly_uint64 (TYPE_SIZE_UNIT (ltype)));
11623 21 : tree gapcst = build_int_cstu (ref_type, gap_offset);
11624 21 : offset = size_binop (PLUS_EXPR, offset, gapcst);
11625 : }
11626 250373 : if (ltype)
11627 : {
11628 245702 : data_ref = fold_build2 (MEM_REF, ltype,
11629 : dataref_ptr, offset);
11630 245702 : if (alignment_support_scheme == dr_aligned
11631 245702 : && align >= TYPE_ALIGN_UNIT (ltype))
11632 : ;
11633 : else
11634 161127 : TREE_TYPE (data_ref)
11635 322254 : = build_aligned_type (TREE_TYPE (data_ref),
11636 : align * BITS_PER_UNIT);
11637 : }
11638 250373 : if (!ltype)
11639 4671 : data_ref = build_constructor (vectype, NULL);
11640 245702 : else if (ltype != vectype)
11641 : {
11642 1391 : vect_copy_ref_info (data_ref,
11643 1391 : DR_REF (first_dr_info->dr));
11644 1391 : tree tem = make_ssa_name (ltype);
11645 1391 : new_stmt = gimple_build_assign (tem, data_ref);
11646 1391 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt,
11647 : gsi);
11648 1391 : data_ref = NULL;
11649 1391 : vec<constructor_elt, va_gc> *v;
11650 : /* We've computed 'num' above to statically two
11651 : or via constant_multiple_p. */
11652 1391 : unsigned num
11653 1391 : = (exact_div (tree_to_poly_uint64
11654 1391 : (TYPE_SIZE_UNIT (vectype)),
11655 : tree_to_poly_uint64
11656 1391 : (TYPE_SIZE_UNIT (ltype)))
11657 1391 : .to_constant ());
11658 1391 : vec_alloc (v, num);
11659 1391 : if (memory_access_type == VMAT_CONTIGUOUS_REVERSE)
11660 : {
11661 54 : while (--num)
11662 54 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
11663 : build_zero_cst (ltype));
11664 21 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem);
11665 : }
11666 : else
11667 : {
11668 1370 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem);
11669 1370 : while (--num)
11670 3094 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
11671 : build_zero_cst (ltype));
11672 : }
11673 1391 : gcc_assert (new_vtype != NULL_TREE);
11674 1391 : if (new_vtype == vectype)
11675 1361 : new_stmt
11676 1361 : = gimple_build_assign (vec_dest,
11677 : build_constructor (vectype, v));
11678 : else
11679 : {
11680 30 : tree new_vname = make_ssa_name (new_vtype);
11681 30 : new_stmt
11682 30 : = gimple_build_assign (new_vname,
11683 : build_constructor (new_vtype,
11684 : v));
11685 30 : vect_finish_stmt_generation (vinfo, stmt_info,
11686 : new_stmt, gsi);
11687 30 : new_stmt
11688 30 : = gimple_build_assign (vec_dest,
11689 : build1 (VIEW_CONVERT_EXPR,
11690 : vectype, new_vname));
11691 : }
11692 : }
11693 : }
11694 : break;
11695 : }
11696 0 : case dr_explicit_realign:
11697 0 : {
11698 0 : if (costing_p)
11699 : break;
11700 0 : tree ptr, bump;
11701 :
11702 0 : tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype));
11703 :
11704 0 : if (compute_in_loop)
11705 0 : msq = vect_setup_realignment (vinfo, first_stmt_info, vectype,
11706 : gsi, &realignment_token,
11707 : dr_explicit_realign,
11708 : dataref_ptr, NULL);
11709 :
11710 0 : if (TREE_CODE (dataref_ptr) == SSA_NAME)
11711 0 : ptr = copy_ssa_name (dataref_ptr);
11712 : else
11713 0 : ptr = make_ssa_name (TREE_TYPE (dataref_ptr));
11714 : // For explicit realign the target alignment should be
11715 : // known at compile time.
11716 0 : unsigned HOST_WIDE_INT align
11717 0 : = DR_TARGET_ALIGNMENT (first_dr_info).to_constant ();
11718 0 : new_stmt = gimple_build_assign (ptr, BIT_AND_EXPR, dataref_ptr,
11719 : build_int_cst
11720 0 : (TREE_TYPE (dataref_ptr),
11721 0 : -(HOST_WIDE_INT) align));
11722 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11723 0 : data_ref = build2 (MEM_REF, vectype,
11724 : ptr, build_int_cst (ref_type, 0));
11725 0 : vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr));
11726 0 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
11727 0 : new_stmt = gimple_build_assign (vec_dest, data_ref);
11728 0 : new_temp = make_ssa_name (vec_dest, new_stmt);
11729 0 : gimple_assign_set_lhs (new_stmt, new_temp);
11730 0 : gimple_move_vops (new_stmt, stmt_info->stmt);
11731 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11732 0 : msq = new_temp;
11733 :
11734 0 : bump = size_binop (MULT_EXPR, vs, TYPE_SIZE_UNIT (elem_type));
11735 0 : bump = size_binop (MINUS_EXPR, bump, size_one_node);
11736 0 : ptr = bump_vector_ptr (vinfo, dataref_ptr, NULL, gsi, stmt_info,
11737 : bump);
11738 0 : new_stmt = gimple_build_assign (NULL_TREE, BIT_AND_EXPR, ptr,
11739 0 : build_int_cst (TREE_TYPE (ptr),
11740 0 : -(HOST_WIDE_INT) align));
11741 0 : if (TREE_CODE (ptr) == SSA_NAME)
11742 0 : ptr = copy_ssa_name (ptr, new_stmt);
11743 : else
11744 0 : ptr = make_ssa_name (TREE_TYPE (ptr), new_stmt);
11745 0 : gimple_assign_set_lhs (new_stmt, ptr);
11746 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11747 0 : data_ref = build2 (MEM_REF, vectype,
11748 : ptr, build_int_cst (ref_type, 0));
11749 0 : break;
11750 : }
11751 0 : case dr_explicit_realign_optimized:
11752 0 : {
11753 0 : if (costing_p)
11754 : break;
11755 0 : if (TREE_CODE (dataref_ptr) == SSA_NAME)
11756 0 : new_temp = copy_ssa_name (dataref_ptr);
11757 : else
11758 0 : new_temp = make_ssa_name (TREE_TYPE (dataref_ptr));
11759 : // We should only be doing this if we know the target
11760 : // alignment at compile time.
11761 0 : unsigned HOST_WIDE_INT align
11762 0 : = DR_TARGET_ALIGNMENT (first_dr_info).to_constant ();
11763 0 : new_stmt = gimple_build_assign (new_temp, BIT_AND_EXPR, dataref_ptr,
11764 0 : build_int_cst (TREE_TYPE (dataref_ptr),
11765 0 : -(HOST_WIDE_INT) align));
11766 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11767 0 : data_ref = build2 (MEM_REF, vectype, new_temp,
11768 : build_int_cst (ref_type, 0));
11769 0 : break;
11770 : }
11771 0 : default:
11772 0 : gcc_unreachable ();
11773 : }
11774 :
11775 : /* One common place to cost the above vect load for different
11776 : alignment support schemes. */
11777 830045 : if (costing_p)
11778 : {
11779 : /* For the prologue cost for realign,
11780 : we only need to count it once for the whole group. */
11781 578937 : bool first_stmt_info_p = first_stmt_info == stmt_info;
11782 578937 : bool add_realign_cost = first_stmt_info_p && i == 0;
11783 578937 : if (memory_access_type == VMAT_CONTIGUOUS
11784 578937 : || memory_access_type == VMAT_CONTIGUOUS_REVERSE)
11785 : {
11786 : /* Leave realign cases alone to keep them simple. */
11787 578937 : if (alignment_support_scheme == dr_explicit_realign_optimized
11788 : || alignment_support_scheme == dr_explicit_realign)
11789 0 : vect_get_load_cost (vinfo, stmt_info, slp_node, 1,
11790 : alignment_support_scheme, misalignment,
11791 : add_realign_cost, &inside_cost,
11792 : &prologue_cost, cost_vec, cost_vec,
11793 : true);
11794 : else
11795 578937 : n_adjacent_loads++;
11796 : }
11797 : }
11798 : else
11799 : {
11800 251108 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
11801 : /* DATA_REF is null if we've already built the statement. */
11802 251108 : if (data_ref)
11803 : {
11804 248982 : vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr));
11805 248982 : new_stmt = gimple_build_assign (vec_dest, data_ref);
11806 : }
11807 :
11808 502216 : new_temp = (need_zeroing
11809 251108 : ? make_ssa_name (vectype)
11810 251108 : : make_ssa_name (vec_dest, new_stmt));
11811 251108 : gimple_set_lhs (new_stmt, new_temp);
11812 251108 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11813 :
11814 : /* If we need to explicitly zero inactive elements emit a
11815 : VEC_COND_EXPR that does so. */
11816 251108 : if (need_zeroing)
11817 : {
11818 0 : vec_els = vect_get_mask_load_else (MASK_LOAD_ELSE_ZERO,
11819 : vectype);
11820 :
11821 0 : tree new_temp2 = make_ssa_name (vec_dest, new_stmt);
11822 0 : new_stmt = gimple_build_assign (new_temp2, VEC_COND_EXPR,
11823 : final_mask, new_temp, vec_els);
11824 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt,
11825 : gsi);
11826 0 : new_temp = new_temp2;
11827 : }
11828 : }
11829 :
11830 : /* 3. Handle explicit realignment if necessary/supported.
11831 : Create in loop:
11832 : vec_dest = realign_load (msq, lsq, realignment_token) */
11833 830045 : if (!costing_p
11834 251108 : && (alignment_support_scheme == dr_explicit_realign_optimized
11835 : || alignment_support_scheme == dr_explicit_realign))
11836 : {
11837 0 : lsq = gimple_assign_lhs (new_stmt);
11838 0 : if (!realignment_token)
11839 0 : realignment_token = dataref_ptr;
11840 0 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
11841 0 : new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, msq,
11842 : lsq, realignment_token);
11843 0 : new_temp = make_ssa_name (vec_dest, new_stmt);
11844 0 : gimple_assign_set_lhs (new_stmt, new_temp);
11845 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
11846 :
11847 0 : if (alignment_support_scheme == dr_explicit_realign_optimized)
11848 : {
11849 0 : gcc_assert (phi);
11850 0 : if (i == vec_num - 1)
11851 0 : add_phi_arg (phi, lsq, loop_latch_edge (containing_loop),
11852 : UNKNOWN_LOCATION);
11853 : msq = lsq;
11854 : }
11855 : }
11856 :
11857 830045 : if (memory_access_type == VMAT_CONTIGUOUS_REVERSE)
11858 : {
11859 5580 : if (costing_p)
11860 4149 : inside_cost = record_stmt_cost (cost_vec, 1, vec_perm,
11861 : slp_node, 0, vect_body);
11862 : else
11863 : {
11864 1431 : tree perm_mask = perm_mask_for_reverse (vectype);
11865 1431 : new_temp = permute_vec_elements (vinfo, new_temp, new_temp,
11866 : perm_mask, stmt_info, gsi);
11867 1431 : new_stmt = SSA_NAME_DEF_STMT (new_temp);
11868 : }
11869 : }
11870 :
11871 : /* Collect vector loads and later create their permutation in
11872 : vect_transform_slp_perm_load. */
11873 830045 : if (!costing_p && (grouped_load || ls.slp_perm))
11874 72213 : dr_chain.quick_push (new_temp);
11875 :
11876 : /* Store vector loads in the corresponding SLP_NODE. */
11877 251108 : if (!costing_p && !ls.slp_perm)
11878 178895 : slp_node->push_vec_def (new_stmt);
11879 :
11880 : /* With SLP permutation we load the gaps as well, without
11881 : we need to skip the gaps after we manage to fully load
11882 : all elements. group_gap_adj is DR_GROUP_SIZE here. */
11883 830045 : group_elt += nunits;
11884 830045 : if (!costing_p
11885 251108 : && maybe_ne (group_gap_adj, 0U)
11886 44821 : && !ls.slp_perm
11887 850064 : && known_eq (group_elt, group_size - group_gap_adj))
11888 : {
11889 15831 : poly_wide_int bump_val
11890 15831 : = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) * group_gap_adj);
11891 15831 : if (tree_int_cst_sgn (vect_dr_behavior (vinfo, dr_info)->step) == -1)
11892 0 : bump_val = -bump_val;
11893 15831 : tree bump = wide_int_to_tree (sizetype, bump_val);
11894 15831 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
11895 : stmt_info, bump);
11896 15831 : group_elt = 0;
11897 15831 : }
11898 : }
11899 : /* Bump the vector pointer to account for a gap or for excess
11900 : elements loaded for a permuted SLP load. */
11901 504130 : if (!costing_p
11902 160934 : && maybe_ne (group_gap_adj, 0U)
11903 520395 : && ls.slp_perm)
11904 : {
11905 434 : poly_wide_int bump_val
11906 434 : = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) * group_gap_adj);
11907 434 : if (tree_int_cst_sgn (vect_dr_behavior (vinfo, dr_info)->step) == -1)
11908 9 : bump_val = -bump_val;
11909 434 : tree bump = wide_int_to_tree (sizetype, bump_val);
11910 434 : dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
11911 : stmt_info, bump);
11912 434 : }
11913 :
11914 504130 : if (ls.slp_perm)
11915 : {
11916 : /* For SLP we know we've seen all possible uses of dr_chain so
11917 : direct vect_transform_slp_perm_load to DCE the unused parts.
11918 : ??? This is a hack to prevent compile-time issues as seen
11919 : in PR101120 and friends. */
11920 48692 : if (costing_p)
11921 : {
11922 31537 : gcc_assert (ls.n_perms != -1U && ls.n_loads != -1U);
11923 31537 : if (ls.n_perms != 0)
11924 31173 : inside_cost = record_stmt_cost (cost_vec, ls.n_perms, vec_perm,
11925 : slp_node, 0, vect_body);
11926 31537 : if (n_adjacent_loads > 0)
11927 31537 : n_adjacent_loads = ls.n_loads;
11928 : }
11929 : else
11930 : {
11931 17155 : unsigned n_perms2, n_loads2;
11932 17155 : bool ok = vect_transform_slp_perm_load (vinfo, slp_node, dr_chain,
11933 : gsi, vf, false, &n_perms2,
11934 : &n_loads2, true);
11935 17155 : gcc_assert (ok && ls.n_perms == n_perms2 && ls.n_loads == n_loads2);
11936 : }
11937 : }
11938 :
11939 504130 : if (costing_p)
11940 : {
11941 343196 : gcc_assert (memory_access_type == VMAT_CONTIGUOUS
11942 : || memory_access_type == VMAT_CONTIGUOUS_REVERSE);
11943 343196 : if (n_adjacent_loads > 0)
11944 343196 : vect_get_load_cost (vinfo, stmt_info, slp_node, n_adjacent_loads,
11945 : alignment_support_scheme, misalignment, false,
11946 : &inside_cost, &prologue_cost, cost_vec, cost_vec,
11947 : true);
11948 343196 : if (dump_enabled_p ())
11949 23169 : dump_printf_loc (MSG_NOTE, vect_location,
11950 : "vect_model_load_cost: inside_cost = %u, "
11951 : "prologue_cost = %u .\n",
11952 : inside_cost, prologue_cost);
11953 : }
11954 :
11955 504130 : return true;
11956 1650645 : }
11957 :
11958 : /* Function vect_is_simple_cond.
11959 :
11960 : Input:
11961 : LOOP - the loop that is being vectorized.
11962 : COND - Condition that is checked for simple use.
11963 :
11964 : Output:
11965 : *COMP_VECTYPE - the vector type for the comparison.
11966 : *DTS - The def types for the arguments of the comparison
11967 :
11968 : Returns whether a COND can be vectorized. Checks whether
11969 : condition operands are supportable using vec_is_simple_use. */
11970 :
11971 : static bool
11972 27854 : vect_is_simple_cond (tree cond, vec_info *vinfo,
11973 : slp_tree slp_node, tree *comp_vectype,
11974 : enum vect_def_type *dts, tree vectype)
11975 : {
11976 27854 : tree lhs, rhs;
11977 27854 : tree vectype1 = NULL_TREE, vectype2 = NULL_TREE;
11978 27854 : slp_tree slp_op;
11979 :
11980 : /* Mask case. */
11981 27854 : if (TREE_CODE (cond) == SSA_NAME
11982 27854 : && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond)))
11983 : {
11984 27842 : if (!vect_is_simple_use (vinfo, slp_node, 0, &cond,
11985 : &slp_op, &dts[0], comp_vectype)
11986 27842 : || !*comp_vectype
11987 55669 : || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype))
11988 : return false;
11989 : return true;
11990 : }
11991 :
11992 12 : if (!COMPARISON_CLASS_P (cond))
11993 : return false;
11994 :
11995 0 : lhs = TREE_OPERAND (cond, 0);
11996 0 : rhs = TREE_OPERAND (cond, 1);
11997 :
11998 0 : if (TREE_CODE (lhs) == SSA_NAME)
11999 : {
12000 0 : if (!vect_is_simple_use (vinfo, slp_node, 0,
12001 : &lhs, &slp_op, &dts[0], &vectype1))
12002 : return false;
12003 : }
12004 0 : else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST
12005 0 : || TREE_CODE (lhs) == FIXED_CST)
12006 0 : dts[0] = vect_constant_def;
12007 : else
12008 : return false;
12009 :
12010 0 : if (TREE_CODE (rhs) == SSA_NAME)
12011 : {
12012 0 : if (!vect_is_simple_use (vinfo, slp_node, 1,
12013 : &rhs, &slp_op, &dts[1], &vectype2))
12014 : return false;
12015 : }
12016 0 : else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST
12017 0 : || TREE_CODE (rhs) == FIXED_CST)
12018 0 : dts[1] = vect_constant_def;
12019 : else
12020 : return false;
12021 :
12022 0 : if (vectype1 && vectype2
12023 0 : && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1),
12024 0 : TYPE_VECTOR_SUBPARTS (vectype2)))
12025 0 : return false;
12026 :
12027 0 : *comp_vectype = vectype1 ? vectype1 : vectype2;
12028 : /* Invariant comparison. */
12029 0 : if (! *comp_vectype)
12030 : {
12031 0 : tree scalar_type = TREE_TYPE (lhs);
12032 0 : if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type))
12033 0 : *comp_vectype = truth_type_for (vectype);
12034 : else
12035 : {
12036 : /* If we can widen the comparison to match vectype do so. */
12037 0 : if (INTEGRAL_TYPE_P (scalar_type)
12038 0 : && !slp_node
12039 0 : && tree_int_cst_lt (TYPE_SIZE (scalar_type),
12040 0 : TYPE_SIZE (TREE_TYPE (vectype))))
12041 0 : scalar_type = build_nonstandard_integer_type
12042 0 : (vector_element_bits (vectype), TYPE_UNSIGNED (scalar_type));
12043 0 : *comp_vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
12044 : slp_node);
12045 : }
12046 : }
12047 :
12048 : return true;
12049 : }
12050 :
12051 : /* vectorizable_condition.
12052 :
12053 : Check if STMT_INFO is conditional modify expression that can be vectorized.
12054 : If COST_VEC is passed, calculate costs but don't change anything,
12055 : otherwise, vectorize STMT_INFO: create a vectorized stmt using
12056 : VEC_COND_EXPR to replace it, and insert it at GSI.
12057 :
12058 : When STMT_INFO is vectorized as a nested cycle, for_reduction is true.
12059 :
12060 : Return true if STMT_INFO is vectorizable in this way. */
12061 :
12062 : static bool
12063 606273 : vectorizable_condition (vec_info *vinfo,
12064 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
12065 : slp_tree slp_node, stmt_vector_for_cost *cost_vec)
12066 : {
12067 606273 : tree scalar_dest = NULL_TREE;
12068 606273 : tree vec_dest = NULL_TREE;
12069 606273 : tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE;
12070 606273 : tree then_clause, else_clause;
12071 606273 : tree comp_vectype = NULL_TREE;
12072 606273 : tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE;
12073 606273 : tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE;
12074 606273 : tree vec_compare;
12075 606273 : tree new_temp;
12076 606273 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
12077 606273 : enum vect_def_type dts[4]
12078 : = {vect_unknown_def_type, vect_unknown_def_type,
12079 : vect_unknown_def_type, vect_unknown_def_type};
12080 606273 : enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR;
12081 606273 : int i;
12082 606273 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
12083 606273 : vec<tree> vec_oprnds0 = vNULL;
12084 606273 : vec<tree> vec_oprnds1 = vNULL;
12085 606273 : vec<tree> vec_oprnds2 = vNULL;
12086 606273 : vec<tree> vec_oprnds3 = vNULL;
12087 606273 : tree vec_cmp_type;
12088 606273 : bool masked = false;
12089 :
12090 606273 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
12091 : return false;
12092 :
12093 : /* Is vectorizable conditional operation? */
12094 930997 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
12095 352549 : if (!stmt)
12096 : return false;
12097 :
12098 352549 : code = gimple_assign_rhs_code (stmt);
12099 352549 : if (code != COND_EXPR)
12100 : return false;
12101 :
12102 27854 : int reduc_index = SLP_TREE_REDUC_IDX (slp_node);
12103 27854 : vect_reduction_type reduction_type = TREE_CODE_REDUCTION;
12104 27854 : bool nested_cycle_p = false;
12105 27854 : bool for_reduction = vect_is_reduction (stmt_info);
12106 27854 : if (for_reduction)
12107 : {
12108 547 : if (SLP_TREE_LANES (slp_node) > 1)
12109 : return false;
12110 : /* ??? With a reduction path we do not get at the reduction info from
12111 : every stmt, use the conservative default setting then. */
12112 627 : if (STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info)))
12113 : {
12114 529 : vect_reduc_info reduc_info
12115 529 : = info_for_reduction (loop_vinfo, slp_node);
12116 529 : reduction_type = VECT_REDUC_INFO_TYPE (reduc_info);
12117 529 : nested_cycle_p = nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo),
12118 : stmt_info);
12119 : }
12120 : }
12121 : else
12122 : {
12123 27307 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
12124 : return false;
12125 : }
12126 :
12127 27854 : tree vectype = SLP_TREE_VECTYPE (slp_node);
12128 27854 : tree vectype1 = NULL_TREE, vectype2 = NULL_TREE;
12129 :
12130 27854 : int vec_num = vect_get_num_copies (vinfo, slp_node);
12131 :
12132 27854 : cond_expr = gimple_assign_rhs1 (stmt);
12133 27854 : gcc_assert (! COMPARISON_CLASS_P (cond_expr));
12134 :
12135 27854 : if (!vect_is_simple_cond (cond_expr, vinfo, slp_node,
12136 : &comp_vectype, &dts[0], vectype)
12137 27854 : || !comp_vectype)
12138 : return false;
12139 :
12140 27827 : unsigned op_adjust = COMPARISON_CLASS_P (cond_expr) ? 1 : 0;
12141 27827 : slp_tree then_slp_node, else_slp_node;
12142 27827 : if (!vect_is_simple_use (vinfo, slp_node, 1 + op_adjust,
12143 : &then_clause, &then_slp_node, &dts[2], &vectype1))
12144 : return false;
12145 27827 : if (!vect_is_simple_use (vinfo, slp_node, 2 + op_adjust,
12146 : &else_clause, &else_slp_node, &dts[3], &vectype2))
12147 : return false;
12148 :
12149 27827 : if (vectype1 && !useless_type_conversion_p (vectype, vectype1))
12150 : return false;
12151 :
12152 27827 : if (vectype2 && !useless_type_conversion_p (vectype, vectype2))
12153 : return false;
12154 :
12155 27827 : masked = !COMPARISON_CLASS_P (cond_expr);
12156 27827 : vec_cmp_type = truth_type_for (comp_vectype);
12157 27827 : if (vec_cmp_type == NULL_TREE
12158 55654 : || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype),
12159 27827 : TYPE_VECTOR_SUBPARTS (vec_cmp_type)))
12160 0 : return false;
12161 :
12162 27827 : cond_code = TREE_CODE (cond_expr);
12163 27827 : if (!masked)
12164 : {
12165 0 : cond_expr0 = TREE_OPERAND (cond_expr, 0);
12166 0 : cond_expr1 = TREE_OPERAND (cond_expr, 1);
12167 : }
12168 :
12169 : /* For conditional reductions, the "then" value needs to be the candidate
12170 : value calculated by this iteration while the "else" value needs to be
12171 : the result carried over from previous iterations. If the COND_EXPR
12172 : is the other way around, we need to swap it. */
12173 27827 : bool must_invert_cmp_result = false;
12174 27827 : if (reduction_type == EXTRACT_LAST_REDUCTION && reduc_index == 1)
12175 : {
12176 0 : if (masked)
12177 0 : must_invert_cmp_result = true;
12178 : else
12179 : {
12180 0 : bool honor_nans = HONOR_NANS (TREE_TYPE (cond_expr0));
12181 0 : tree_code new_code = invert_tree_comparison (cond_code, honor_nans);
12182 0 : if (new_code == ERROR_MARK)
12183 : must_invert_cmp_result = true;
12184 : else
12185 : {
12186 0 : cond_code = new_code;
12187 : /* Make sure we don't accidentally use the old condition. */
12188 0 : cond_expr = NULL_TREE;
12189 : }
12190 : }
12191 : /* ??? The vectorized operand query below doesn't allow swapping
12192 : this way for SLP. */
12193 0 : return false;
12194 : /* std::swap (then_clause, else_clause); */
12195 : }
12196 :
12197 27827 : if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype))
12198 : {
12199 : /* Boolean values may have another representation in vectors
12200 : and therefore we prefer bit operations over comparison for
12201 : them (which also works for scalar masks). We store opcodes
12202 : to use in bitop1 and bitop2. Statement is vectorized as
12203 : BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2)
12204 : depending on bitop1 and bitop2 arity. */
12205 0 : switch (cond_code)
12206 : {
12207 : case GT_EXPR:
12208 : bitop1 = BIT_NOT_EXPR;
12209 : bitop2 = BIT_AND_EXPR;
12210 : break;
12211 0 : case GE_EXPR:
12212 0 : bitop1 = BIT_NOT_EXPR;
12213 0 : bitop2 = BIT_IOR_EXPR;
12214 0 : break;
12215 0 : case LT_EXPR:
12216 0 : bitop1 = BIT_NOT_EXPR;
12217 0 : bitop2 = BIT_AND_EXPR;
12218 0 : std::swap (cond_expr0, cond_expr1);
12219 0 : break;
12220 0 : case LE_EXPR:
12221 0 : bitop1 = BIT_NOT_EXPR;
12222 0 : bitop2 = BIT_IOR_EXPR;
12223 0 : std::swap (cond_expr0, cond_expr1);
12224 0 : break;
12225 0 : case NE_EXPR:
12226 0 : bitop1 = BIT_XOR_EXPR;
12227 0 : break;
12228 0 : case EQ_EXPR:
12229 0 : bitop1 = BIT_XOR_EXPR;
12230 0 : bitop2 = BIT_NOT_EXPR;
12231 0 : break;
12232 : default:
12233 : return false;
12234 : }
12235 : cond_code = SSA_NAME;
12236 : }
12237 :
12238 27827 : if (TREE_CODE_CLASS (cond_code) == tcc_comparison
12239 0 : && reduction_type == EXTRACT_LAST_REDUCTION
12240 27827 : && !expand_vec_cmp_expr_p (comp_vectype, vec_cmp_type, cond_code))
12241 : {
12242 0 : if (dump_enabled_p ())
12243 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
12244 : "reduction comparison operation not supported.\n");
12245 0 : return false;
12246 : }
12247 :
12248 27827 : if (cost_vec)
12249 : {
12250 19346 : if (bitop1 != NOP_EXPR)
12251 : {
12252 0 : machine_mode mode = TYPE_MODE (comp_vectype);
12253 0 : optab optab;
12254 :
12255 0 : optab = optab_for_tree_code (bitop1, comp_vectype, optab_default);
12256 0 : if (!optab || !can_implement_p (optab, mode))
12257 0 : return false;
12258 :
12259 0 : if (bitop2 != NOP_EXPR)
12260 : {
12261 0 : optab = optab_for_tree_code (bitop2, comp_vectype,
12262 : optab_default);
12263 0 : if (!optab || !can_implement_p (optab, mode))
12264 0 : return false;
12265 : }
12266 : }
12267 :
12268 19346 : vect_cost_for_stmt kind = vector_stmt;
12269 19346 : if (reduction_type == EXTRACT_LAST_REDUCTION)
12270 : /* Count one reduction-like operation per vector. */
12271 : kind = vec_to_scalar;
12272 19346 : else if ((masked && !expand_vec_cond_expr_p (vectype, comp_vectype))
12273 19346 : || (!masked
12274 0 : && (!expand_vec_cmp_expr_p (comp_vectype, vec_cmp_type,
12275 : cond_code)
12276 0 : || !expand_vec_cond_expr_p (vectype, vec_cmp_type))))
12277 2 : return false;
12278 :
12279 19344 : if (!vect_maybe_update_slp_op_vectype (SLP_TREE_CHILDREN (slp_node)[0],
12280 : comp_vectype)
12281 19344 : || (op_adjust == 1
12282 0 : && !vect_maybe_update_slp_op_vectype
12283 0 : (SLP_TREE_CHILDREN (slp_node)[1], comp_vectype))
12284 19344 : || !vect_maybe_update_slp_op_vectype (then_slp_node, vectype)
12285 38688 : || !vect_maybe_update_slp_op_vectype (else_slp_node, vectype))
12286 : {
12287 0 : if (dump_enabled_p ())
12288 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
12289 : "incompatible vector types for invariants\n");
12290 0 : return false;
12291 : }
12292 :
12293 19344 : if (loop_vinfo && for_reduction
12294 400 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
12295 : {
12296 65 : if (reduction_type == EXTRACT_LAST_REDUCTION)
12297 : {
12298 0 : if (direct_internal_fn_supported_p (IFN_LEN_FOLD_EXTRACT_LAST,
12299 : vectype, OPTIMIZE_FOR_SPEED))
12300 0 : vect_record_loop_len (loop_vinfo,
12301 : &LOOP_VINFO_LENS (loop_vinfo),
12302 : vec_num, vectype, 1);
12303 : else
12304 0 : vect_record_loop_mask (loop_vinfo,
12305 : &LOOP_VINFO_MASKS (loop_vinfo),
12306 : vec_num, vectype, NULL);
12307 : }
12308 : /* Extra inactive lanes should be safe for vect_nested_cycle. */
12309 65 : else if (!nested_cycle_p)
12310 : {
12311 65 : if (dump_enabled_p ())
12312 8 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
12313 : "conditional reduction prevents the use"
12314 : " of partial vectors.\n");
12315 65 : LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
12316 : }
12317 : }
12318 :
12319 19344 : SLP_TREE_TYPE (slp_node) = condition_vec_info_type;
12320 19344 : vect_model_simple_cost (vinfo, 1, slp_node, cost_vec, kind);
12321 19344 : return true;
12322 : }
12323 :
12324 : /* Transform. */
12325 :
12326 : /* Handle def. */
12327 8481 : scalar_dest = gimple_assign_lhs (stmt);
12328 8481 : if (reduction_type != EXTRACT_LAST_REDUCTION)
12329 8481 : vec_dest = vect_create_destination_var (scalar_dest, vectype);
12330 :
12331 8481 : bool swap_cond_operands = false;
12332 :
12333 : /* See whether another part of the vectorized code applies a loop
12334 : mask to the condition, or to its inverse. */
12335 :
12336 8481 : vec_loop_masks *masks = NULL;
12337 8481 : vec_loop_lens *lens = NULL;
12338 8481 : if (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo))
12339 : {
12340 0 : if (reduction_type == EXTRACT_LAST_REDUCTION)
12341 0 : lens = &LOOP_VINFO_LENS (loop_vinfo);
12342 : }
12343 8481 : else if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
12344 : {
12345 3 : if (reduction_type == EXTRACT_LAST_REDUCTION)
12346 0 : masks = &LOOP_VINFO_MASKS (loop_vinfo);
12347 : else
12348 : {
12349 3 : scalar_cond_masked_key cond (cond_expr, 1);
12350 3 : if (loop_vinfo->scalar_cond_masked_set.contains (cond))
12351 0 : masks = &LOOP_VINFO_MASKS (loop_vinfo);
12352 : else
12353 : {
12354 3 : bool honor_nans = HONOR_NANS (TREE_TYPE (cond.op0));
12355 3 : tree_code orig_code = cond.code;
12356 3 : cond.code = invert_tree_comparison (cond.code, honor_nans);
12357 3 : if (!masked && loop_vinfo->scalar_cond_masked_set.contains (cond))
12358 : {
12359 0 : masks = &LOOP_VINFO_MASKS (loop_vinfo);
12360 0 : cond_code = cond.code;
12361 0 : swap_cond_operands = true;
12362 : }
12363 : else
12364 : {
12365 : /* Try the inverse of the current mask. We check if the
12366 : inverse mask is live and if so we generate a negate of
12367 : the current mask such that we still honor NaNs. */
12368 3 : cond.inverted_p = true;
12369 3 : cond.code = orig_code;
12370 3 : if (loop_vinfo->scalar_cond_masked_set.contains (cond))
12371 : {
12372 0 : masks = &LOOP_VINFO_MASKS (loop_vinfo);
12373 0 : cond_code = cond.code;
12374 0 : swap_cond_operands = true;
12375 0 : must_invert_cmp_result = true;
12376 : }
12377 : }
12378 : }
12379 : }
12380 : }
12381 :
12382 : /* Handle cond expr. */
12383 8481 : if (masked)
12384 8481 : vect_get_vec_defs (vinfo, slp_node,
12385 : cond_expr, &vec_oprnds0,
12386 : then_clause, &vec_oprnds2,
12387 : reduction_type != EXTRACT_LAST_REDUCTION
12388 : ? else_clause : NULL, &vec_oprnds3);
12389 : else
12390 0 : vect_get_vec_defs (vinfo, slp_node,
12391 : cond_expr0, &vec_oprnds0,
12392 : cond_expr1, &vec_oprnds1,
12393 : then_clause, &vec_oprnds2,
12394 : reduction_type != EXTRACT_LAST_REDUCTION
12395 : ? else_clause : NULL, &vec_oprnds3);
12396 :
12397 8481 : if (reduction_type == EXTRACT_LAST_REDUCTION)
12398 0 : vec_else_clause = else_clause;
12399 :
12400 : /* Arguments are ready. Create the new vector stmt. */
12401 19959 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs)
12402 : {
12403 11478 : vec_then_clause = vec_oprnds2[i];
12404 11478 : if (reduction_type != EXTRACT_LAST_REDUCTION)
12405 11478 : vec_else_clause = vec_oprnds3[i];
12406 :
12407 11478 : if (swap_cond_operands)
12408 0 : std::swap (vec_then_clause, vec_else_clause);
12409 :
12410 11478 : if (masked)
12411 : vec_compare = vec_cond_lhs;
12412 : else
12413 : {
12414 0 : vec_cond_rhs = vec_oprnds1[i];
12415 0 : if (bitop1 == NOP_EXPR)
12416 : {
12417 0 : gimple_seq stmts = NULL;
12418 0 : vec_compare = gimple_build (&stmts, cond_code, vec_cmp_type,
12419 : vec_cond_lhs, vec_cond_rhs);
12420 0 : gsi_insert_before (gsi, stmts, GSI_SAME_STMT);
12421 : }
12422 : else
12423 : {
12424 0 : new_temp = make_ssa_name (vec_cmp_type);
12425 0 : gassign *new_stmt;
12426 0 : if (bitop1 == BIT_NOT_EXPR)
12427 0 : new_stmt = gimple_build_assign (new_temp, bitop1,
12428 : vec_cond_rhs);
12429 : else
12430 0 : new_stmt
12431 0 : = gimple_build_assign (new_temp, bitop1, vec_cond_lhs,
12432 : vec_cond_rhs);
12433 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12434 0 : if (bitop2 == NOP_EXPR)
12435 : vec_compare = new_temp;
12436 0 : else if (bitop2 == BIT_NOT_EXPR
12437 0 : && reduction_type != EXTRACT_LAST_REDUCTION)
12438 : {
12439 : /* Instead of doing ~x ? y : z do x ? z : y. */
12440 : vec_compare = new_temp;
12441 : std::swap (vec_then_clause, vec_else_clause);
12442 : }
12443 : else
12444 : {
12445 0 : vec_compare = make_ssa_name (vec_cmp_type);
12446 0 : if (bitop2 == BIT_NOT_EXPR)
12447 0 : new_stmt
12448 0 : = gimple_build_assign (vec_compare, bitop2, new_temp);
12449 : else
12450 0 : new_stmt
12451 0 : = gimple_build_assign (vec_compare, bitop2,
12452 : vec_cond_lhs, new_temp);
12453 0 : vect_finish_stmt_generation (vinfo, stmt_info,
12454 : new_stmt, gsi);
12455 : }
12456 : }
12457 : }
12458 :
12459 : /* If we decided to apply a loop mask to the result of the vector
12460 : comparison, AND the comparison with the mask now. Later passes
12461 : should then be able to reuse the AND results between mulitple
12462 : vector statements.
12463 :
12464 : For example:
12465 : for (int i = 0; i < 100; ++i)
12466 : x[i] = y[i] ? z[i] : 10;
12467 :
12468 : results in following optimized GIMPLE:
12469 :
12470 : mask__35.8_43 = vect__4.7_41 != { 0, ... };
12471 : vec_mask_and_46 = loop_mask_40 & mask__35.8_43;
12472 : _19 = &MEM[base: z_12(D), index: ivtmp_56, step: 4, offset: 0B];
12473 : vect_iftmp.11_47 = .MASK_LOAD (_19, 4B, vec_mask_and_46);
12474 : vect_iftmp.12_52 = VEC_COND_EXPR <vec_mask_and_46,
12475 : vect_iftmp.11_47, { 10, ... }>;
12476 :
12477 : instead of using a masked and unmasked forms of
12478 : vec != { 0, ... } (masked in the MASK_LOAD,
12479 : unmasked in the VEC_COND_EXPR). */
12480 :
12481 : /* Force vec_compare to be an SSA_NAME rather than a comparison,
12482 : in cases where that's necessary. */
12483 :
12484 11478 : tree len = NULL_TREE, bias = NULL_TREE;
12485 11478 : if (masks || lens || reduction_type == EXTRACT_LAST_REDUCTION)
12486 : {
12487 0 : if (!is_gimple_val (vec_compare))
12488 : {
12489 0 : tree vec_compare_name = make_ssa_name (vec_cmp_type);
12490 0 : gassign *new_stmt = gimple_build_assign (vec_compare_name,
12491 : vec_compare);
12492 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12493 0 : vec_compare = vec_compare_name;
12494 : }
12495 :
12496 0 : if (must_invert_cmp_result)
12497 : {
12498 0 : tree vec_compare_name = make_ssa_name (vec_cmp_type);
12499 0 : gassign *new_stmt = gimple_build_assign (vec_compare_name,
12500 : BIT_NOT_EXPR,
12501 : vec_compare);
12502 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12503 0 : vec_compare = vec_compare_name;
12504 : }
12505 :
12506 0 : if (direct_internal_fn_supported_p (IFN_LEN_FOLD_EXTRACT_LAST,
12507 : vectype, OPTIMIZE_FOR_SPEED))
12508 : {
12509 0 : if (lens)
12510 : {
12511 : /* ??? Do we really want the adjusted LEN here? Isn't this
12512 : based on number of elements? */
12513 0 : len = vect_get_loop_len (loop_vinfo, gsi, lens,
12514 : vec_num, vectype, i, 1, true);
12515 0 : signed char biasval
12516 0 : = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo);
12517 0 : bias = build_int_cst (intQI_type_node, biasval);
12518 : }
12519 : else
12520 : {
12521 0 : len = size_int (TYPE_VECTOR_SUBPARTS (vectype));
12522 0 : bias = build_int_cst (intQI_type_node, 0);
12523 : }
12524 : }
12525 0 : if (masks)
12526 : {
12527 0 : tree loop_mask
12528 0 : = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num,
12529 : vectype, i);
12530 0 : tree tmp2 = make_ssa_name (vec_cmp_type);
12531 0 : gassign *g
12532 0 : = gimple_build_assign (tmp2, BIT_AND_EXPR, vec_compare,
12533 : loop_mask);
12534 0 : vect_finish_stmt_generation (vinfo, stmt_info, g, gsi);
12535 0 : vec_compare = tmp2;
12536 : }
12537 : }
12538 :
12539 0 : gimple *new_stmt;
12540 0 : if (reduction_type == EXTRACT_LAST_REDUCTION)
12541 : {
12542 0 : gimple *old_stmt = vect_orig_stmt (stmt_info)->stmt;
12543 0 : tree lhs = gimple_get_lhs (old_stmt);
12544 0 : if ((unsigned)i != vec_oprnds0.length () - 1)
12545 0 : lhs = copy_ssa_name (lhs);
12546 0 : if (len)
12547 0 : new_stmt = gimple_build_call_internal
12548 0 : (IFN_LEN_FOLD_EXTRACT_LAST, 5, vec_else_clause, vec_compare,
12549 : vec_then_clause, len, bias);
12550 : else
12551 0 : new_stmt = gimple_build_call_internal
12552 0 : (IFN_FOLD_EXTRACT_LAST, 3, vec_else_clause, vec_compare,
12553 : vec_then_clause);
12554 0 : gimple_call_set_lhs (new_stmt, lhs);
12555 0 : SSA_NAME_DEF_STMT (lhs) = new_stmt;
12556 0 : if ((unsigned)i != vec_oprnds0.length () - 1)
12557 : {
12558 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12559 0 : vec_else_clause = lhs;
12560 : }
12561 0 : else if (old_stmt == gsi_stmt (*gsi))
12562 0 : vect_finish_replace_stmt (vinfo, stmt_info, new_stmt);
12563 : else
12564 : {
12565 : /* In this case we're moving the definition to later in the
12566 : block. That doesn't matter because the only uses of the
12567 : lhs are in phi statements. */
12568 0 : gimple_stmt_iterator old_gsi = gsi_for_stmt (old_stmt);
12569 0 : gsi_remove (&old_gsi, true);
12570 0 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12571 : }
12572 : }
12573 : else
12574 : {
12575 11478 : new_temp = make_ssa_name (vec_dest);
12576 11478 : new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, vec_compare,
12577 : vec_then_clause, vec_else_clause);
12578 11478 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12579 : }
12580 11478 : slp_node->push_vec_def (new_stmt);
12581 : }
12582 :
12583 8481 : vec_oprnds0.release ();
12584 8481 : vec_oprnds1.release ();
12585 8481 : vec_oprnds2.release ();
12586 8481 : vec_oprnds3.release ();
12587 :
12588 8481 : return true;
12589 : }
12590 :
12591 : /* Helper of vectorizable_comparison.
12592 :
12593 : Check if STMT_INFO is comparison expression CODE that can be vectorized.
12594 : If COST_VEC is passed, calculate costs but don't change anything,
12595 : otherwise, vectorize STMT_INFO: create a vectorized comparison, and insert
12596 : it at GSI.
12597 :
12598 : Return true if STMT_INFO is vectorizable in this way. */
12599 :
12600 : static bool
12601 333155 : vectorizable_comparison_1 (vec_info *vinfo, tree vectype,
12602 : stmt_vec_info stmt_info, tree_code code,
12603 : gimple_stmt_iterator *gsi,
12604 : slp_tree slp_node, stmt_vector_for_cost *cost_vec)
12605 : {
12606 333155 : tree lhs, rhs1, rhs2;
12607 333155 : tree vectype1 = NULL_TREE, vectype2 = NULL_TREE;
12608 333155 : tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE;
12609 333155 : tree new_temp;
12610 333155 : enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type};
12611 333155 : poly_uint64 nunits;
12612 333155 : enum tree_code bitop1 = NOP_EXPR, bitop2 = NOP_EXPR;
12613 333155 : int i;
12614 333155 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
12615 333155 : vec<tree> vec_oprnds0 = vNULL;
12616 333155 : vec<tree> vec_oprnds1 = vNULL;
12617 333155 : tree mask_type;
12618 333155 : tree mask = NULL_TREE;
12619 :
12620 333155 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
12621 : return false;
12622 :
12623 333155 : if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype))
12624 : return false;
12625 :
12626 142884 : mask_type = vectype;
12627 142884 : nunits = TYPE_VECTOR_SUBPARTS (vectype);
12628 :
12629 142884 : if (TREE_CODE_CLASS (code) != tcc_comparison)
12630 : return false;
12631 :
12632 141132 : slp_tree slp_rhs1, slp_rhs2;
12633 141132 : if (!vect_is_simple_use (vinfo, slp_node,
12634 : 0, &rhs1, &slp_rhs1, &dts[0], &vectype1))
12635 : return false;
12636 :
12637 141132 : if (!vect_is_simple_use (vinfo, slp_node,
12638 : 1, &rhs2, &slp_rhs2, &dts[1], &vectype2))
12639 : return false;
12640 :
12641 106407 : if (vectype1 && vectype2
12642 203830 : && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1),
12643 62698 : TYPE_VECTOR_SUBPARTS (vectype2)))
12644 16 : return false;
12645 :
12646 141116 : vectype = vectype1 ? vectype1 : vectype2;
12647 :
12648 : /* Invariant comparison. */
12649 141116 : if (!vectype)
12650 : {
12651 30019 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (rhs1), slp_node);
12652 30019 : if (!vectype || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits))
12653 7 : return false;
12654 : }
12655 111097 : else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype)))
12656 : return false;
12657 :
12658 : /* Can't compare mask and non-mask types. */
12659 106391 : if (vectype1 && vectype2
12660 328617 : && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2)))
12661 : return false;
12662 :
12663 : /* Boolean values may have another representation in vectors
12664 : and therefore we prefer bit operations over comparison for
12665 : them (which also works for scalar masks). We store opcodes
12666 : to use in bitop1 and bitop2. Statement is vectorized as
12667 : BITOP2 (rhs1 BITOP1 rhs2) or
12668 : rhs1 BITOP2 (BITOP1 rhs2)
12669 : depending on bitop1 and bitop2 arity. */
12670 141101 : bool swap_p = false;
12671 141101 : if (VECTOR_BOOLEAN_TYPE_P (vectype))
12672 : {
12673 635 : if (code == GT_EXPR)
12674 : {
12675 : bitop1 = BIT_NOT_EXPR;
12676 : bitop2 = BIT_AND_EXPR;
12677 : }
12678 : else if (code == GE_EXPR)
12679 : {
12680 : bitop1 = BIT_NOT_EXPR;
12681 : bitop2 = BIT_IOR_EXPR;
12682 : }
12683 : else if (code == LT_EXPR)
12684 : {
12685 : bitop1 = BIT_NOT_EXPR;
12686 : bitop2 = BIT_AND_EXPR;
12687 : swap_p = true;
12688 : }
12689 : else if (code == LE_EXPR)
12690 : {
12691 : bitop1 = BIT_NOT_EXPR;
12692 : bitop2 = BIT_IOR_EXPR;
12693 : swap_p = true;
12694 : }
12695 : else
12696 : {
12697 : bitop1 = BIT_XOR_EXPR;
12698 : if (code == EQ_EXPR)
12699 : bitop2 = BIT_NOT_EXPR;
12700 : }
12701 : }
12702 :
12703 141101 : if (cost_vec)
12704 : {
12705 128709 : if (bitop1 == NOP_EXPR)
12706 : {
12707 128206 : if (!expand_vec_cmp_expr_p (vectype, mask_type, code))
12708 : return false;
12709 : }
12710 : else
12711 : {
12712 503 : machine_mode mode = TYPE_MODE (vectype);
12713 503 : optab optab;
12714 :
12715 503 : optab = optab_for_tree_code (bitop1, vectype, optab_default);
12716 503 : if (!optab || !can_implement_p (optab, mode))
12717 0 : return false;
12718 :
12719 503 : if (bitop2 != NOP_EXPR)
12720 : {
12721 91 : optab = optab_for_tree_code (bitop2, vectype, optab_default);
12722 91 : if (!optab || !can_implement_p (optab, mode))
12723 0 : return false;
12724 : }
12725 : }
12726 :
12727 : /* Put types on constant and invariant SLP children. */
12728 121142 : if (!vect_maybe_update_slp_op_vectype (slp_rhs1, vectype)
12729 121142 : || !vect_maybe_update_slp_op_vectype (slp_rhs2, vectype))
12730 : {
12731 2 : if (dump_enabled_p ())
12732 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
12733 : "incompatible vector types for invariants\n");
12734 2 : return false;
12735 : }
12736 :
12737 121140 : vect_model_simple_cost (vinfo, 1 + (bitop2 != NOP_EXPR),
12738 : slp_node, cost_vec);
12739 121140 : return true;
12740 : }
12741 :
12742 : /* Transform. */
12743 :
12744 : /* Handle def. */
12745 12392 : lhs = gimple_get_lhs (STMT_VINFO_STMT (stmt_info));
12746 12392 : if (lhs)
12747 12392 : mask = vect_create_destination_var (lhs, mask_type);
12748 :
12749 12392 : vect_get_vec_defs (vinfo, slp_node, rhs1, &vec_oprnds0, rhs2, &vec_oprnds1);
12750 12392 : if (swap_p)
12751 58 : std::swap (vec_oprnds0, vec_oprnds1);
12752 :
12753 : /* Arguments are ready. Create the new vector stmt. */
12754 31179 : FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1)
12755 : {
12756 18787 : gimple *new_stmt;
12757 18787 : vec_rhs2 = vec_oprnds1[i];
12758 :
12759 18787 : if (lhs)
12760 18787 : new_temp = make_ssa_name (mask);
12761 : else
12762 0 : new_temp = make_temp_ssa_name (mask_type, NULL, "cmp");
12763 18787 : if (bitop1 == NOP_EXPR)
12764 : {
12765 18645 : new_stmt = gimple_build_assign (new_temp, code,
12766 : vec_rhs1, vec_rhs2);
12767 18645 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12768 : }
12769 : else
12770 : {
12771 142 : if (bitop1 == BIT_NOT_EXPR)
12772 84 : new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2);
12773 : else
12774 58 : new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1,
12775 : vec_rhs2);
12776 142 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12777 142 : if (bitop2 != NOP_EXPR)
12778 : {
12779 84 : tree res = make_ssa_name (mask);
12780 84 : if (bitop2 == BIT_NOT_EXPR)
12781 0 : new_stmt = gimple_build_assign (res, bitop2, new_temp);
12782 : else
12783 84 : new_stmt = gimple_build_assign (res, bitop2, vec_rhs1,
12784 : new_temp);
12785 84 : vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
12786 : }
12787 : }
12788 18787 : slp_node->push_vec_def (new_stmt);
12789 : }
12790 :
12791 12392 : vec_oprnds0.release ();
12792 12392 : vec_oprnds1.release ();
12793 :
12794 12392 : return true;
12795 : }
12796 :
12797 : /* vectorizable_comparison.
12798 :
12799 : Check if STMT_INFO is comparison expression that can be vectorized.
12800 : If COST_VEC is passed, calculate costs but don't change anything,
12801 : otherwise, vectorize STMT_INFO: create a vectorized comparison, and insert
12802 : it at GSI.
12803 :
12804 : Return true if STMT_INFO is vectorizable in this way. */
12805 :
12806 : static bool
12807 590840 : vectorizable_comparison (vec_info *vinfo,
12808 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
12809 : slp_tree slp_node, stmt_vector_for_cost *cost_vec)
12810 : {
12811 590840 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
12812 :
12813 590840 : if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
12814 : return false;
12815 :
12816 590840 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
12817 : return false;
12818 :
12819 790463 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
12820 331164 : if (!stmt)
12821 : return false;
12822 :
12823 331164 : enum tree_code code = gimple_assign_rhs_code (stmt);
12824 331164 : tree vectype = SLP_TREE_VECTYPE (slp_node);
12825 331164 : if (!vectorizable_comparison_1 (vinfo, vectype, stmt_info, code, gsi,
12826 : slp_node, cost_vec))
12827 : return false;
12828 :
12829 131541 : if (cost_vec)
12830 119149 : SLP_TREE_TYPE (slp_node) = comparison_vec_info_type;
12831 :
12832 : return true;
12833 : }
12834 :
12835 : /* Check to see if the target supports any of the compare and branch optabs for
12836 : vectors with MODE as these would be required when expanding. */
12837 : static bool
12838 57416 : supports_vector_compare_and_branch (loop_vec_info loop_vinfo, machine_mode mode)
12839 : {
12840 57416 : bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
12841 57416 : bool len_loop_p = LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo);
12842 :
12843 : /* The vectorizer only produces vec_cbranch_any_optab directly. So only
12844 : check for support for that or vec_cbranch_any_optab when masked.
12845 : We can't produce vcond_cbranch_any directly from the vectorizer as we
12846 : want to keep gimple_cond as the GIMPLE representation. But we'll fold
12847 : it in expand. For that reason we require a backend to support the
12848 : unconditional vector cbranch optab if they support the conditional one,
12849 : which is just an optimization on the unconditional one. */
12850 57416 : if (masked_loop_p
12851 57416 : && direct_optab_handler (cond_vec_cbranch_any_optab, mode)
12852 : != CODE_FOR_nothing)
12853 : return true;
12854 57416 : else if (len_loop_p
12855 57416 : && direct_optab_handler (cond_len_vec_cbranch_any_optab, mode)
12856 : != CODE_FOR_nothing)
12857 : return true;
12858 57416 : else if (!masked_loop_p && !len_loop_p
12859 114832 : && direct_optab_handler (vec_cbranch_any_optab, mode)
12860 : != CODE_FOR_nothing)
12861 : return true;
12862 :
12863 : /* The target can implement cbranch to distinguish between boolean vector
12864 : types and data types if they don't have a different mode for both. */
12865 57416 : return direct_optab_handler (cbranch_optab, mode) != CODE_FOR_nothing;
12866 : }
12867 :
12868 : /* Determine the type to use for early break vectorization's scalar IV. If
12869 : no type is possible return false. */
12870 :
12871 : static bool
12872 1991 : vect_compute_type_for_early_break_scalar_iv (loop_vec_info loop_vinfo)
12873 : {
12874 : /* Check if we have a usable scalar IV type for vectorization. */
12875 1991 : tree iters_vf_type = sizetype;
12876 1991 : if (!LOOP_VINFO_NITERS_UNCOUNTED_P (loop_vinfo))
12877 : {
12878 : /* Find the type with the minimum precision we can use
12879 : for the scalar IV. */
12880 1776 : tree cand_type = TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo));
12881 :
12882 : /* Work out how many bits we need to represent the limit. */
12883 1776 : unsigned int min_ni_width
12884 1776 : = vect_min_prec_for_max_niters (loop_vinfo, 1);
12885 :
12886 : /* Check if we're using PFA, if so we need a signed IV and an
12887 : extra bit for the sign. */
12888 1776 : if (TYPE_UNSIGNED (cand_type)
12889 1776 : && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
12890 2844 : && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo))
12891 164 : min_ni_width += 1;
12892 :
12893 1776 : if (TYPE_PRECISION (cand_type) >= min_ni_width)
12894 1701 : iters_vf_type = unsigned_type_for (cand_type);
12895 : else
12896 : {
12897 75 : opt_scalar_int_mode cmp_mode_iter;
12898 75 : tree iv_type = NULL_TREE;
12899 367 : FOR_EACH_MODE_IN_CLASS (cmp_mode_iter, MODE_INT)
12900 : {
12901 367 : auto cmp_mode = cmp_mode_iter.require ();
12902 367 : unsigned int cmp_bits = GET_MODE_BITSIZE (cmp_mode);
12903 367 : if (cmp_bits >= min_ni_width
12904 367 : && targetm.scalar_mode_supported_p (cmp_mode))
12905 : {
12906 75 : iv_type = build_nonstandard_integer_type (cmp_bits, true);
12907 75 : if (iv_type)
12908 : break;
12909 : }
12910 : }
12911 :
12912 75 : if (!iv_type)
12913 : {
12914 0 : if (dump_enabled_p ())
12915 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
12916 : "can't vectorize early exit because the "
12917 : "target doesn't support a scalar type wide "
12918 : "wide enough to hold niters.\n");
12919 0 : return false;
12920 : }
12921 75 : iters_vf_type = iv_type;
12922 : }
12923 : }
12924 :
12925 1991 : LOOP_VINFO_EARLY_BRK_IV_TYPE (loop_vinfo) = iters_vf_type;
12926 1991 : return true;
12927 : }
12928 :
12929 : /* Check to see if the current early break given in STMT_INFO is valid for
12930 : vectorization. */
12931 :
12932 : bool
12933 231855 : vectorizable_early_exit (loop_vec_info loop_vinfo, stmt_vec_info stmt_info,
12934 : gimple_stmt_iterator *gsi,
12935 : slp_tree slp_node, stmt_vector_for_cost *cost_vec)
12936 : {
12937 231855 : if (!is_a <gcond *> (STMT_VINFO_STMT (stmt_info)))
12938 : return false;
12939 :
12940 58980 : if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_condition_def)
12941 : return false;
12942 :
12943 58980 : if (!STMT_VINFO_RELEVANT_P (stmt_info))
12944 : return false;
12945 :
12946 58980 : DUMP_VECT_SCOPE ("vectorizable_early_exit");
12947 :
12948 58980 : auto code = gimple_cond_code (STMT_VINFO_STMT (stmt_info));
12949 :
12950 : /* For SLP we don't want to use the type of the operands of the SLP node, when
12951 : vectorizing using SLP slp_node will be the children of the gcond and we
12952 : want to use the type of the direct children which since the gcond is root
12953 : will be the current node, rather than a child node as vect_is_simple_use
12954 : assumes. */
12955 58980 : tree vectype = SLP_TREE_VECTYPE (slp_node);
12956 58980 : if (!vectype)
12957 : return false;
12958 :
12959 58980 : machine_mode mode = TYPE_MODE (vectype);
12960 58980 : int vec_num = vect_get_num_copies (loop_vinfo, slp_node);
12961 :
12962 58980 : vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
12963 58980 : vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
12964 58980 : bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
12965 58980 : bool len_loop_p = LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo);
12966 :
12967 : /* Now build the new conditional. Pattern gimple_conds get dropped during
12968 : codegen so we must replace the original insn. */
12969 58980 : gimple *orig_stmt = STMT_VINFO_STMT (vect_orig_stmt (stmt_info));
12970 58980 : gcond *cond_stmt = as_a <gcond *>(orig_stmt);
12971 :
12972 58980 : tree vectype_out = vectype;
12973 58980 : auto bb = gimple_bb (cond_stmt);
12974 58980 : edge exit_true_edge = EDGE_SUCC (bb, 0);
12975 58980 : if (exit_true_edge->flags & EDGE_FALSE_VALUE)
12976 628 : exit_true_edge = EDGE_SUCC (bb, 1);
12977 58980 : gcc_assert (exit_true_edge->flags & EDGE_TRUE_VALUE);
12978 :
12979 : /* When vectorizing we assume that if the branch edge is taken that we're
12980 : exiting the loop. This is not however always the case as the compiler will
12981 : rewrite conditions to always be a comparison against 0. To do this it
12982 : sometimes flips the edges. This is fine for scalar, but for vector we
12983 : then have to negate the result of the test, as we're still assuming that if
12984 : you take the branch edge that we found the exit condition. i.e. we need to
12985 : know whether we are generating a `forall` or an `exist` condition. */
12986 117960 : bool flipped = flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo),
12987 58980 : exit_true_edge->dest);
12988 :
12989 : /* See if we support ADDHN and use that for the reduction. */
12990 58980 : internal_fn ifn = IFN_VEC_TRUNC_ADD_HIGH;
12991 58980 : bool addhn_supported_p
12992 58980 : = direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_BOTH);
12993 58980 : tree narrow_type = NULL_TREE;
12994 58980 : if (addhn_supported_p)
12995 : {
12996 : /* Calculate the narrowing type for the result. */
12997 0 : auto halfprec = TYPE_PRECISION (TREE_TYPE (vectype)) / 2;
12998 0 : auto unsignedp = TYPE_UNSIGNED (TREE_TYPE (vectype));
12999 0 : tree itype = build_nonstandard_integer_type (halfprec, unsignedp);
13000 0 : tree tmp_type = build_vector_type (itype, TYPE_VECTOR_SUBPARTS (vectype));
13001 0 : narrow_type = truth_type_for (tmp_type);
13002 :
13003 0 : if (!supports_vector_compare_and_branch (loop_vinfo,
13004 0 : TYPE_MODE (narrow_type)))
13005 : {
13006 0 : if (dump_enabled_p ())
13007 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
13008 : "can't use ADDHN reduction because cbranch for "
13009 : "the narrowed type is not supported by the "
13010 : "target.\n");
13011 : addhn_supported_p = false;
13012 : }
13013 : }
13014 :
13015 : /* Analyze only. */
13016 58980 : if (cost_vec)
13017 : {
13018 57416 : if (!addhn_supported_p
13019 57416 : && !supports_vector_compare_and_branch (loop_vinfo, mode))
13020 : {
13021 55425 : if (dump_enabled_p ())
13022 573 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
13023 : "can't vectorize early exit because the "
13024 : "target doesn't support flag setting vector "
13025 : "comparisons.\n");
13026 55425 : return false;
13027 : }
13028 :
13029 1991 : if (!vectorizable_comparison_1 (loop_vinfo, vectype, stmt_info, code, gsi,
13030 : slp_node, cost_vec))
13031 : return false;
13032 :
13033 1991 : if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
13034 : {
13035 1068 : if (direct_internal_fn_supported_p (IFN_VCOND_MASK_LEN, vectype,
13036 : OPTIMIZE_FOR_SPEED))
13037 0 : vect_record_loop_len (loop_vinfo, lens, vec_num, vectype, 1);
13038 : else
13039 1068 : vect_record_loop_mask (loop_vinfo, masks, vec_num, vectype, NULL);
13040 : }
13041 :
13042 1991 : if (!vect_compute_type_for_early_break_scalar_iv (loop_vinfo))
13043 : return false;
13044 :
13045 : return true;
13046 : }
13047 :
13048 : /* Tranform. */
13049 :
13050 1564 : tree new_temp = NULL_TREE;
13051 1564 : gimple *new_stmt = NULL;
13052 :
13053 1564 : if (dump_enabled_p ())
13054 387 : dump_printf_loc (MSG_NOTE, vect_location, "transform early-exit.\n");
13055 :
13056 : /* For SLP we don't do codegen of the body starting from the gcond, the gconds are
13057 : roots and so by the time we get to them we have already codegened the SLP tree
13058 : and so we shouldn't try to do so again. The arguments have already been
13059 : vectorized. It's not very clean to do this here, But the masking code below is
13060 : complex and this keeps it all in one place to ease fixes and backports. Once we
13061 : drop the non-SLP loop vect or split vectorizable_* this can be simplified. */
13062 :
13063 1564 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
13064 1564 : basic_block cond_bb = gimple_bb (stmt);
13065 1564 : gimple_stmt_iterator cond_gsi = gsi_last_bb (cond_bb);
13066 :
13067 1564 : auto_vec<tree> stmts;
13068 1564 : stmts.safe_splice (SLP_TREE_VEC_DEFS (slp_node));
13069 :
13070 : /* If we're comparing against a previous forall we need to negate the resullts
13071 : before we do the final comparison or reduction. */
13072 1564 : if (flipped)
13073 : {
13074 : /* Rewrite the if(all(mask)) into if (!all(mask)) which is the same as
13075 : if (any(~mask)) by negating the masks and flipping the branches.
13076 :
13077 : 1. For unmasked loops we simply reduce the ~mask.
13078 : 2. For masked loops we reduce (~mask & loop_mask) which is the same as
13079 : doing (mask & loop_mask) ^ loop_mask. */
13080 294 : for (unsigned i = 0; i < stmts.length (); i++)
13081 : {
13082 173 : tree inv_lhs = make_temp_ssa_name (vectype, NULL, "vexit_inv");
13083 173 : auto inv_stmt = gimple_build_assign (inv_lhs, BIT_NOT_EXPR, stmts[i]);
13084 173 : vect_finish_stmt_generation (loop_vinfo, stmt_info, inv_stmt,
13085 : &cond_gsi);
13086 173 : stmts[i] = inv_lhs;
13087 : }
13088 :
13089 121 : EDGE_SUCC (bb, 0)->flags ^= (EDGE_TRUE_VALUE|EDGE_FALSE_VALUE);
13090 121 : EDGE_SUCC (bb, 1)->flags ^= (EDGE_TRUE_VALUE|EDGE_FALSE_VALUE);
13091 : }
13092 :
13093 : /* Determine if we need to reduce the final value. */
13094 1564 : if (stmts.length () > 1)
13095 : {
13096 : /* We build the reductions in a way to maintain as much parallelism as
13097 : possible. */
13098 141 : auto_vec<tree> workset (stmts.length ());
13099 :
13100 : /* Mask the statements as we queue them up. Normally we loop over
13101 : vec_num, but since we inspect the exact results of vectorization
13102 : we don't need to and instead can just use the stmts themselves. */
13103 141 : if (masked_loop_p)
13104 0 : for (unsigned i = 0; i < stmts.length (); i++)
13105 : {
13106 0 : tree stmt_mask
13107 0 : = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num,
13108 : vectype, i);
13109 0 : stmt_mask
13110 0 : = prepare_vec_mask (loop_vinfo, TREE_TYPE (stmt_mask), stmt_mask,
13111 0 : stmts[i], &cond_gsi);
13112 0 : workset.quick_push (stmt_mask);
13113 : }
13114 141 : else if (len_loop_p)
13115 0 : for (unsigned i = 0; i < stmts.length (); i++)
13116 : {
13117 0 : tree len_mask = vect_gen_loop_len_mask (loop_vinfo, gsi, &cond_gsi,
13118 : lens, vec_num,
13119 0 : vectype, stmts[i], i, 1);
13120 :
13121 0 : workset.quick_push (len_mask);
13122 : }
13123 : else
13124 141 : workset.splice (stmts);
13125 :
13126 430 : while (workset.length () > 1)
13127 : {
13128 289 : tree arg0 = workset.pop ();
13129 289 : tree arg1 = workset.pop ();
13130 289 : if (addhn_supported_p && workset.length () == 0)
13131 : {
13132 0 : new_stmt = gimple_build_call_internal (ifn, 2, arg0, arg1);
13133 0 : vectype_out = narrow_type;
13134 0 : new_temp = make_temp_ssa_name (vectype_out, NULL, "vexit_reduc");
13135 0 : gimple_call_set_lhs (as_a <gcall *> (new_stmt), new_temp);
13136 0 : gimple_call_set_nothrow (as_a <gcall *> (new_stmt), true);
13137 : }
13138 : else
13139 : {
13140 289 : new_temp = make_temp_ssa_name (vectype_out, NULL, "vexit_reduc");
13141 289 : new_stmt
13142 289 : = gimple_build_assign (new_temp, BIT_IOR_EXPR, arg0, arg1);
13143 : }
13144 289 : vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt,
13145 : &cond_gsi);
13146 289 : workset.quick_insert (0, new_temp);
13147 : }
13148 141 : }
13149 : else
13150 : {
13151 1423 : new_temp = stmts[0];
13152 1423 : if (masked_loop_p)
13153 : {
13154 0 : tree mask
13155 0 : = vect_get_loop_mask (loop_vinfo, gsi, masks, 1, vectype, 0);
13156 0 : new_temp = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), mask,
13157 : new_temp, &cond_gsi);
13158 : }
13159 1423 : else if (len_loop_p)
13160 0 : new_temp = vect_gen_loop_len_mask (loop_vinfo, gsi, &cond_gsi, lens,
13161 : 1, vectype, new_temp, 0, 1);
13162 : }
13163 :
13164 1564 : gcc_assert (new_temp);
13165 :
13166 1564 : tree cst = build_zero_cst (vectype_out);
13167 1564 : gimple_cond_set_condition (cond_stmt, NE_EXPR, new_temp, cst);
13168 1564 : update_stmt (orig_stmt);
13169 :
13170 : /* ??? */
13171 1564 : SLP_TREE_VEC_DEFS (slp_node).truncate (0);
13172 :
13173 1564 : return true;
13174 1564 : }
13175 :
13176 : /* If SLP_NODE is nonnull, return true if vectorizable_live_operation
13177 : can handle all live statements in the node. Otherwise return true
13178 : if STMT_INFO is not live or if vectorizable_live_operation can handle it.
13179 : VEC_STMT_P is as for vectorizable_live_operation. */
13180 :
13181 : static bool
13182 1423603 : can_vectorize_live_stmts (vec_info *vinfo,
13183 : slp_tree slp_node, slp_instance slp_node_instance,
13184 : bool vec_stmt_p,
13185 : stmt_vector_for_cost *cost_vec)
13186 : {
13187 1423603 : stmt_vec_info slp_stmt_info;
13188 1423603 : unsigned int i;
13189 3211130 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt_info)
13190 : {
13191 1787527 : if (slp_stmt_info
13192 1768060 : && STMT_VINFO_LIVE_P (slp_stmt_info)
13193 1958375 : && !vectorizable_live_operation (vinfo, slp_stmt_info, slp_node,
13194 : slp_node_instance, i,
13195 : vec_stmt_p, cost_vec))
13196 : return false;
13197 : }
13198 :
13199 : return true;
13200 : }
13201 :
13202 : /* Make sure the statement is vectorizable. */
13203 :
13204 : opt_result
13205 2329924 : vect_analyze_stmt (vec_info *vinfo,
13206 : slp_tree node, slp_instance node_instance,
13207 : stmt_vector_for_cost *cost_vec)
13208 : {
13209 2329924 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
13210 2329924 : bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
13211 2329924 : enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info);
13212 2329924 : bool ok;
13213 :
13214 2329924 : if (dump_enabled_p ())
13215 97851 : dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: %G",
13216 : stmt_info->stmt);
13217 :
13218 4414436 : if (gimple_has_volatile_ops (stmt_info->stmt))
13219 : {
13220 : /* ??? This shouldn't really happen, volatile stmts should
13221 : not end up in the SLP graph. */
13222 0 : return opt_result::failure_at (stmt_info->stmt,
13223 : "not vectorized:"
13224 : " stmt has volatile operands: %G\n",
13225 : stmt_info->stmt);
13226 : }
13227 :
13228 : /* Skip stmts that do not need to be vectorized. */
13229 2329924 : if (!STMT_VINFO_RELEVANT_P (stmt_info)
13230 0 : && !STMT_VINFO_LIVE_P (stmt_info))
13231 : {
13232 0 : if (dump_enabled_p ())
13233 0 : dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n");
13234 :
13235 : /* ??? This shouldn't really happen, irrelevant stmts should
13236 : not end up in the SLP graph. */
13237 0 : return opt_result::failure_at (stmt_info->stmt,
13238 : "not vectorized:"
13239 : " irrelevant stmt as SLP node %p "
13240 : "representative.\n",
13241 : (void *)node);
13242 : }
13243 :
13244 2329924 : switch (STMT_VINFO_DEF_TYPE (stmt_info))
13245 : {
13246 : case vect_internal_def:
13247 : case vect_condition_def:
13248 : break;
13249 :
13250 57445 : case vect_reduction_def:
13251 57445 : case vect_nested_cycle:
13252 57445 : gcc_assert (!bb_vinfo
13253 : && (relevance == vect_used_in_outer
13254 : || relevance == vect_used_in_outer_by_reduction
13255 : || relevance == vect_used_by_reduction
13256 : || relevance == vect_unused_in_scope
13257 : || relevance == vect_used_only_live));
13258 : break;
13259 :
13260 287 : case vect_double_reduction_def:
13261 287 : gcc_assert (!bb_vinfo && node);
13262 : break;
13263 :
13264 136784 : case vect_induction_def:
13265 136784 : case vect_first_order_recurrence:
13266 136784 : gcc_assert (!bb_vinfo);
13267 : break;
13268 :
13269 0 : case vect_constant_def:
13270 0 : case vect_external_def:
13271 0 : case vect_unknown_def_type:
13272 0 : default:
13273 0 : gcc_unreachable ();
13274 : }
13275 :
13276 2329924 : tree saved_vectype = STMT_VINFO_VECTYPE (stmt_info);
13277 2329924 : STMT_VINFO_VECTYPE (stmt_info) = NULL_TREE;
13278 :
13279 2329924 : if (STMT_VINFO_RELEVANT_P (stmt_info))
13280 : {
13281 2329924 : gcall *call = dyn_cast <gcall *> (stmt_info->stmt);
13282 2329924 : gcc_assert (SLP_TREE_VECTYPE (node)
13283 : || gimple_code (stmt_info->stmt) == GIMPLE_COND
13284 : || (call && gimple_call_lhs (call) == NULL_TREE));
13285 : }
13286 :
13287 2329924 : ok = true;
13288 2329924 : if (bb_vinfo
13289 1177638 : || (STMT_VINFO_RELEVANT_P (stmt_info)
13290 0 : || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def))
13291 : /* Prefer vectorizable_call over vectorizable_simd_clone_call so
13292 : -mveclibabi= takes preference over library functions with
13293 : the simd attribute. */
13294 2329924 : ok = (vectorizable_call (vinfo, stmt_info, NULL, node, cost_vec)
13295 2324411 : || vectorizable_simd_clone_call (vinfo, stmt_info, NULL, node,
13296 : cost_vec)
13297 2323959 : || vectorizable_conversion (vinfo, stmt_info, NULL, node, cost_vec)
13298 2258385 : || vectorizable_operation (vinfo, stmt_info, NULL, node, cost_vec)
13299 1840414 : || vectorizable_assignment (vinfo, stmt_info, NULL, node, cost_vec)
13300 1783100 : || vectorizable_load (vinfo, stmt_info, NULL, node, cost_vec)
13301 1415468 : || vectorizable_store (vinfo, stmt_info, NULL, node, cost_vec)
13302 643253 : || vectorizable_shift (vinfo, stmt_info, NULL, node, cost_vec)
13303 597792 : || vectorizable_condition (vinfo, stmt_info, NULL, node, cost_vec)
13304 578448 : || vectorizable_comparison (vinfo, stmt_info, NULL, node, cost_vec)
13305 459299 : || (bb_vinfo
13306 124415 : && vectorizable_phi (bb_vinfo, stmt_info, node, cost_vec))
13307 2732244 : || (is_a <loop_vec_info> (vinfo)
13308 334884 : && (vectorizable_lane_reducing (as_a <loop_vec_info> (vinfo),
13309 : stmt_info, node, cost_vec)
13310 334430 : || vectorizable_reduction (as_a <loop_vec_info> (vinfo),
13311 : stmt_info,
13312 : node, node_instance, cost_vec)
13313 279208 : || vectorizable_induction (as_a <loop_vec_info> (vinfo),
13314 : stmt_info, node, cost_vec)
13315 173884 : || vectorizable_lc_phi (as_a <loop_vec_info> (vinfo),
13316 : stmt_info, node)
13317 173123 : || vectorizable_recurr (as_a <loop_vec_info> (vinfo),
13318 : stmt_info, node, cost_vec)
13319 172875 : || vectorizable_early_exit (as_a <loop_vec_info> (vinfo),
13320 : stmt_info, NULL, node,
13321 : cost_vec))));
13322 :
13323 2329924 : STMT_VINFO_VECTYPE (stmt_info) = saved_vectype;
13324 :
13325 2089613 : if (!ok)
13326 240311 : return opt_result::failure_at (stmt_info->stmt,
13327 : "not vectorized:"
13328 : " relevant stmt not supported: %G",
13329 : stmt_info->stmt);
13330 :
13331 : /* Stmts that are (also) "live" (i.e. - that are used out of the loop)
13332 : need extra handling, except for vectorizable reductions. */
13333 2089613 : if (!bb_vinfo
13334 1004763 : && SLP_TREE_TYPE (node) != reduc_vec_info_type
13335 997429 : && (SLP_TREE_TYPE (node) != lc_phi_info_type
13336 761 : || SLP_TREE_DEF_TYPE (node) == vect_internal_def)
13337 997429 : && (!node->ldst_lanes || SLP_TREE_PERMUTE_P (node))
13338 3087042 : && !can_vectorize_live_stmts (as_a <loop_vec_info> (vinfo),
13339 : node, node_instance,
13340 : false, cost_vec))
13341 0 : return opt_result::failure_at (stmt_info->stmt,
13342 : "not vectorized:"
13343 : " live stmt not supported: %G",
13344 : stmt_info->stmt);
13345 :
13346 2089613 : return opt_result::success ();
13347 : }
13348 :
13349 :
13350 : /* Function vect_transform_stmt.
13351 :
13352 : Create a vectorized stmt to replace STMT_INFO, and insert it at GSI. */
13353 :
13354 : bool
13355 969861 : vect_transform_stmt (vec_info *vinfo,
13356 : stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
13357 : slp_tree slp_node, slp_instance slp_node_instance)
13358 : {
13359 969861 : bool is_store = false;
13360 969861 : bool done;
13361 :
13362 969861 : gcc_assert (slp_node);
13363 :
13364 969861 : if (stmt_info)
13365 969020 : STMT_VINFO_VECTYPE (stmt_info) = NULL_TREE;
13366 :
13367 969861 : switch (SLP_TREE_TYPE (slp_node))
13368 : {
13369 22934 : case type_demotion_vec_info_type:
13370 22934 : case type_promotion_vec_info_type:
13371 22934 : case type_conversion_vec_info_type:
13372 22934 : done = vectorizable_conversion (vinfo, stmt_info, gsi, slp_node, NULL);
13373 22934 : gcc_assert (done);
13374 : break;
13375 :
13376 16173 : case induc_vec_info_type:
13377 16173 : done = vectorizable_induction (as_a <loop_vec_info> (vinfo),
13378 : stmt_info, slp_node, NULL);
13379 16173 : gcc_assert (done);
13380 : break;
13381 :
13382 7821 : case shift_vec_info_type:
13383 7821 : done = vectorizable_shift (vinfo, stmt_info, gsi, slp_node, NULL);
13384 7821 : gcc_assert (done);
13385 : break;
13386 :
13387 115465 : case op_vec_info_type:
13388 115465 : done = vectorizable_operation (vinfo, stmt_info, gsi, slp_node, NULL);
13389 115465 : gcc_assert (done);
13390 : break;
13391 :
13392 15020 : case assignment_vec_info_type:
13393 15020 : done = vectorizable_assignment (vinfo, stmt_info, gsi, slp_node, NULL);
13394 15020 : gcc_assert (done);
13395 : break;
13396 :
13397 165901 : case load_vec_info_type:
13398 165901 : done = vectorizable_load (vinfo, stmt_info, gsi, slp_node, NULL);
13399 165901 : gcc_assert (done);
13400 : break;
13401 :
13402 543687 : case store_vec_info_type:
13403 543687 : done = vectorizable_store (vinfo, stmt_info, gsi, slp_node, NULL);
13404 543687 : gcc_assert (done);
13405 : is_store = true;
13406 : break;
13407 :
13408 8481 : case condition_vec_info_type:
13409 8481 : done = vectorizable_condition (vinfo, stmt_info, gsi, slp_node, NULL);
13410 8481 : gcc_assert (done);
13411 : break;
13412 :
13413 12392 : case comparison_vec_info_type:
13414 12392 : done = vectorizable_comparison (vinfo, stmt_info, gsi, slp_node, NULL);
13415 12392 : gcc_assert (done);
13416 : break;
13417 :
13418 4210 : case call_vec_info_type:
13419 4210 : done = vectorizable_call (vinfo, stmt_info, gsi, slp_node, NULL);
13420 4210 : break;
13421 :
13422 357 : case call_simd_clone_vec_info_type:
13423 357 : done = vectorizable_simd_clone_call (vinfo, stmt_info, gsi,
13424 : slp_node, NULL);
13425 357 : break;
13426 :
13427 2575 : case reduc_vec_info_type:
13428 2575 : done = vect_transform_reduction (as_a <loop_vec_info> (vinfo), stmt_info,
13429 : gsi, slp_node);
13430 2575 : gcc_assert (done);
13431 : break;
13432 :
13433 23709 : case cycle_phi_info_type:
13434 23709 : done = vect_transform_cycle_phi (as_a <loop_vec_info> (vinfo), stmt_info,
13435 : slp_node, slp_node_instance);
13436 23709 : gcc_assert (done);
13437 : break;
13438 :
13439 504 : case lc_phi_info_type:
13440 504 : done = vect_transform_lc_phi (as_a <loop_vec_info> (vinfo),
13441 : stmt_info, slp_node);
13442 504 : gcc_assert (done);
13443 : break;
13444 :
13445 40 : case recurr_info_type:
13446 40 : done = vectorizable_recurr (as_a <loop_vec_info> (vinfo),
13447 : stmt_info, slp_node, NULL);
13448 40 : gcc_assert (done);
13449 : break;
13450 :
13451 14115 : case phi_info_type:
13452 14115 : done = vectorizable_phi (as_a <bb_vec_info> (vinfo),
13453 : stmt_info, slp_node, NULL);
13454 14115 : gcc_assert (done);
13455 : break;
13456 :
13457 0 : case loop_exit_ctrl_vec_info_type:
13458 0 : done = vectorizable_early_exit (as_a <loop_vec_info> (vinfo),
13459 : stmt_info, gsi, slp_node, NULL);
13460 0 : gcc_assert (done);
13461 : break;
13462 :
13463 16477 : case permute_info_type:
13464 16477 : done = vectorizable_slp_permutation (vinfo, gsi, slp_node, NULL);
13465 16477 : gcc_assert (done);
13466 : break;
13467 :
13468 0 : default:
13469 0 : if (!STMT_VINFO_LIVE_P (stmt_info))
13470 : {
13471 0 : if (dump_enabled_p ())
13472 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
13473 : "stmt not supported.\n");
13474 0 : gcc_unreachable ();
13475 : }
13476 969861 : done = true;
13477 : }
13478 :
13479 969861 : if (SLP_TREE_TYPE (slp_node) != store_vec_info_type
13480 426174 : && (!slp_node->ldst_lanes || SLP_TREE_PERMUTE_P (slp_node)))
13481 : {
13482 : /* Handle stmts whose DEF is used outside the loop-nest that is
13483 : being vectorized. */
13484 426174 : done = can_vectorize_live_stmts (vinfo, slp_node,
13485 : slp_node_instance, true, NULL);
13486 426174 : gcc_assert (done);
13487 : }
13488 :
13489 969861 : return is_store;
13490 : }
13491 :
13492 :
13493 : /* Remove a group of stores (for SLP or interleaving), free their
13494 : stmt_vec_info. */
13495 :
13496 : void
13497 0 : vect_remove_stores (vec_info *vinfo, stmt_vec_info first_stmt_info)
13498 : {
13499 0 : stmt_vec_info next_stmt_info = first_stmt_info;
13500 :
13501 0 : while (next_stmt_info)
13502 : {
13503 0 : stmt_vec_info tmp = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
13504 0 : next_stmt_info = vect_orig_stmt (next_stmt_info);
13505 : /* Free the attached stmt_vec_info and remove the stmt. */
13506 0 : vinfo->remove_stmt (next_stmt_info);
13507 0 : next_stmt_info = tmp;
13508 : }
13509 0 : }
13510 :
13511 : /* If NUNITS is nonzero, return a vector type that contains NUNITS
13512 : elements of type SCALAR_TYPE, or null if the target doesn't support
13513 : such a type.
13514 :
13515 : If NUNITS is zero, return a vector type that contains elements of
13516 : type SCALAR_TYPE, choosing whichever vector size the target prefers.
13517 :
13518 : If PREVAILING_MODE is VOIDmode, we have not yet chosen a vector mode
13519 : for this vectorization region and want to "autodetect" the best choice.
13520 : Otherwise, PREVAILING_MODE is a previously-chosen vector TYPE_MODE
13521 : and we want the new type to be interoperable with it. PREVAILING_MODE
13522 : in this case can be a scalar integer mode or a vector mode; when it
13523 : is a vector mode, the function acts like a tree-level version of
13524 : related_vector_mode. */
13525 :
13526 : tree
13527 30101309 : get_related_vectype_for_scalar_type (machine_mode prevailing_mode,
13528 : tree scalar_type, poly_uint64 nunits)
13529 : {
13530 30101309 : tree orig_scalar_type = scalar_type;
13531 30101309 : scalar_mode inner_mode;
13532 30101309 : machine_mode simd_mode;
13533 30101309 : tree vectype;
13534 :
13535 30101309 : if ((!INTEGRAL_TYPE_P (scalar_type)
13536 10307679 : && !POINTER_TYPE_P (scalar_type)
13537 1650598 : && !SCALAR_FLOAT_TYPE_P (scalar_type))
13538 39904339 : || (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode)
13539 1146031 : && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)))
13540 507864 : return NULL_TREE;
13541 :
13542 29593445 : unsigned int nbytes = GET_MODE_SIZE (inner_mode);
13543 :
13544 : /* Interoperability between modes requires one to be a constant multiple
13545 : of the other, so that the number of vectors required for each operation
13546 : is a compile-time constant. */
13547 29593445 : if (prevailing_mode != VOIDmode
13548 28467631 : && !constant_multiple_p (nunits * nbytes,
13549 28467631 : GET_MODE_SIZE (prevailing_mode))
13550 31071032 : && !constant_multiple_p (GET_MODE_SIZE (prevailing_mode),
13551 1477587 : nunits * nbytes))
13552 : return NULL_TREE;
13553 :
13554 : /* For vector types of elements whose mode precision doesn't
13555 : match their types precision we use a element type of mode
13556 : precision. The vectorization routines will have to make sure
13557 : they support the proper result truncation/extension.
13558 : We also make sure to build vector types with INTEGER_TYPE
13559 : component type only. */
13560 29593445 : if (INTEGRAL_TYPE_P (scalar_type)
13561 49386993 : && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type)
13562 18324615 : || TREE_CODE (scalar_type) != INTEGER_TYPE))
13563 1677944 : scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode),
13564 1677944 : TYPE_UNSIGNED (scalar_type));
13565 :
13566 : /* We shouldn't end up building VECTOR_TYPEs of non-scalar components.
13567 : When the component mode passes the above test simply use a type
13568 : corresponding to that mode. The theory is that any use that
13569 : would cause problems with this will disable vectorization anyway. */
13570 27915501 : else if (!SCALAR_FLOAT_TYPE_P (scalar_type)
13571 : && !INTEGRAL_TYPE_P (scalar_type))
13572 8657081 : scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1);
13573 :
13574 : /* We can't build a vector type of elements with alignment bigger than
13575 : their size. */
13576 19258420 : else if (nbytes < TYPE_ALIGN_UNIT (scalar_type))
13577 375360 : scalar_type = lang_hooks.types.type_for_mode (inner_mode,
13578 187680 : TYPE_UNSIGNED (scalar_type));
13579 :
13580 : /* If we felt back to using the mode fail if there was
13581 : no scalar type for it. */
13582 29593445 : if (scalar_type == NULL_TREE)
13583 : return NULL_TREE;
13584 :
13585 : /* If no prevailing mode was supplied, use the mode the target prefers.
13586 : Otherwise lookup a vector mode based on the prevailing mode. */
13587 29593445 : if (prevailing_mode == VOIDmode)
13588 : {
13589 1125814 : gcc_assert (known_eq (nunits, 0U));
13590 1125814 : simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode);
13591 1125814 : if (SCALAR_INT_MODE_P (simd_mode))
13592 : {
13593 : /* Traditional behavior is not to take the integer mode
13594 : literally, but simply to use it as a way of determining
13595 : the vector size. It is up to mode_for_vector to decide
13596 : what the TYPE_MODE should be.
13597 :
13598 : Note that nunits == 1 is allowed in order to support single
13599 : element vector types. */
13600 60154 : if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)
13601 545 : || !mode_for_vector (inner_mode, nunits).exists (&simd_mode))
13602 29532 : return NULL_TREE;
13603 : }
13604 : }
13605 28467631 : else if (SCALAR_INT_MODE_P (prevailing_mode)
13606 28467631 : || !related_vector_mode (prevailing_mode,
13607 26487669 : inner_mode, nunits).exists (&simd_mode))
13608 : {
13609 : /* Fall back to using mode_for_vector, mostly in the hope of being
13610 : able to use an integer mode. */
13611 1979962 : if (known_eq (nunits, 0U)
13612 4628613 : && !multiple_p (GET_MODE_SIZE (prevailing_mode), nbytes, &nunits))
13613 : return NULL_TREE;
13614 :
13615 138670 : if (!mode_for_vector (inner_mode, nunits).exists (&simd_mode))
13616 128604 : return NULL_TREE;
13617 : }
13618 :
13619 27594017 : vectype = build_vector_type_for_mode (scalar_type, simd_mode);
13620 :
13621 : /* In cases where the mode was chosen by mode_for_vector, check that
13622 : the target actually supports the chosen mode, or that it at least
13623 : allows the vector mode to be replaced by a like-sized integer. */
13624 55188034 : if (!VECTOR_MODE_P (TYPE_MODE (vectype))
13625 27604341 : && !INTEGRAL_MODE_P (TYPE_MODE (vectype)))
13626 : return NULL_TREE;
13627 :
13628 : /* Re-attach the address-space qualifier if we canonicalized the scalar
13629 : type. */
13630 27585934 : if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype))
13631 5 : return build_qualified_type
13632 5 : (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type)));
13633 :
13634 : return vectype;
13635 : }
13636 :
13637 : /* Function get_vectype_for_scalar_type.
13638 :
13639 : Returns the vector type corresponding to SCALAR_TYPE as supported
13640 : by the target. If GROUP_SIZE is nonzero and we're performing BB
13641 : vectorization, make sure that the number of elements in the vector
13642 : is no bigger than GROUP_SIZE. */
13643 :
13644 : tree
13645 25698445 : get_vectype_for_scalar_type (vec_info *vinfo, tree scalar_type,
13646 : unsigned int group_size)
13647 : {
13648 : /* For BB vectorization, we should always have a group size once we've
13649 : constructed the SLP tree; the only valid uses of zero GROUP_SIZEs
13650 : are tentative requests during things like early data reference
13651 : analysis and pattern recognition. */
13652 25698445 : if (is_a <bb_vec_info> (vinfo))
13653 23434866 : gcc_assert (vinfo->slp_instances.is_empty () || group_size != 0);
13654 : else
13655 : group_size = 0;
13656 :
13657 25698445 : tree vectype = get_related_vectype_for_scalar_type (vinfo->vector_mode,
13658 : scalar_type);
13659 25698445 : if (vectype && vinfo->vector_mode == VOIDmode)
13660 1049591 : vinfo->vector_mode = TYPE_MODE (vectype);
13661 :
13662 : /* Register the natural choice of vector type, before the group size
13663 : has been applied. */
13664 0 : if (vectype)
13665 23344940 : vinfo->used_vector_modes.add (TYPE_MODE (vectype));
13666 :
13667 : /* If the natural choice of vector type doesn't satisfy GROUP_SIZE,
13668 : try again with an explicit number of elements. */
13669 23344940 : if (vectype
13670 23344940 : && group_size
13671 25698445 : && maybe_ge (TYPE_VECTOR_SUBPARTS (vectype), group_size))
13672 : {
13673 : /* Start with the biggest number of units that fits within
13674 : GROUP_SIZE and halve it until we find a valid vector type.
13675 : Usually either the first attempt will succeed or all will
13676 : fail (in the latter case because GROUP_SIZE is too small
13677 : for the target), but it's possible that a target could have
13678 : a hole between supported vector types.
13679 :
13680 : If GROUP_SIZE is not a power of 2, this has the effect of
13681 : trying the largest power of 2 that fits within the group,
13682 : even though the group is not a multiple of that vector size.
13683 : The BB vectorizer will then try to carve up the group into
13684 : smaller pieces. */
13685 3012793 : unsigned int nunits = 1 << floor_log2 (group_size);
13686 3012793 : do
13687 : {
13688 3012793 : vectype = get_related_vectype_for_scalar_type (vinfo->vector_mode,
13689 3012793 : scalar_type, nunits);
13690 3012793 : nunits /= 2;
13691 : }
13692 3012793 : while (nunits > 1 && !vectype);
13693 : }
13694 :
13695 25698445 : return vectype;
13696 : }
13697 :
13698 : /* Return the vector type corresponding to SCALAR_TYPE as supported
13699 : by the target. NODE, if nonnull, is the SLP tree node that will
13700 : use the returned vector type. */
13701 :
13702 : tree
13703 160252 : get_vectype_for_scalar_type (vec_info *vinfo, tree scalar_type, slp_tree node)
13704 : {
13705 160252 : unsigned int group_size = 0;
13706 160252 : if (node)
13707 160252 : group_size = SLP_TREE_LANES (node);
13708 160252 : return get_vectype_for_scalar_type (vinfo, scalar_type, group_size);
13709 : }
13710 :
13711 : /* Function get_mask_type_for_scalar_type.
13712 :
13713 : Returns the mask type corresponding to a result of comparison
13714 : of vectors of specified SCALAR_TYPE as supported by target.
13715 : If GROUP_SIZE is nonzero and we're performing BB vectorization,
13716 : make sure that the number of elements in the vector is no bigger
13717 : than GROUP_SIZE. */
13718 :
13719 : tree
13720 1057935 : get_mask_type_for_scalar_type (vec_info *vinfo, tree scalar_type,
13721 : unsigned int group_size)
13722 : {
13723 1057935 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type, group_size);
13724 :
13725 1057935 : if (!vectype)
13726 : return NULL;
13727 :
13728 1034543 : return truth_type_for (vectype);
13729 : }
13730 :
13731 : /* Function get_mask_type_for_scalar_type.
13732 :
13733 : Returns the mask type corresponding to a result of comparison
13734 : of vectors of specified SCALAR_TYPE as supported by target.
13735 : NODE, if nonnull, is the SLP tree node that will use the returned
13736 : vector type. */
13737 :
13738 : tree
13739 17 : get_mask_type_for_scalar_type (vec_info *vinfo, tree scalar_type,
13740 : slp_tree node)
13741 : {
13742 17 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type, node);
13743 :
13744 17 : if (!vectype)
13745 : return NULL;
13746 :
13747 17 : return truth_type_for (vectype);
13748 : }
13749 :
13750 : /* Function get_same_sized_vectype
13751 :
13752 : Returns a vector type corresponding to SCALAR_TYPE of size
13753 : VECTOR_TYPE if supported by the target. */
13754 :
13755 : tree
13756 143359 : get_same_sized_vectype (tree scalar_type, tree vector_type)
13757 : {
13758 143359 : if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type))
13759 0 : return truth_type_for (vector_type);
13760 :
13761 143359 : poly_uint64 nunits;
13762 286718 : if (!multiple_p (GET_MODE_SIZE (TYPE_MODE (vector_type)),
13763 286718 : GET_MODE_SIZE (TYPE_MODE (scalar_type)), &nunits))
13764 : return NULL_TREE;
13765 :
13766 143359 : return get_related_vectype_for_scalar_type (TYPE_MODE (vector_type),
13767 143359 : scalar_type, nunits);
13768 : }
13769 :
13770 : /* Return true if replacing LOOP_VINFO->vector_mode with VECTOR_MODE
13771 : would not change the chosen vector modes. */
13772 :
13773 : bool
13774 1538518 : vect_chooses_same_modes_p (vec_info *vinfo, machine_mode vector_mode)
13775 : {
13776 1538518 : for (vec_info::mode_set::iterator i = vinfo->used_vector_modes.begin ();
13777 3560944 : i != vinfo->used_vector_modes.end (); ++i)
13778 1806717 : if (!VECTOR_MODE_P (*i)
13779 5420151 : || related_vector_mode (vector_mode, GET_MODE_INNER (*i), 0) != *i)
13780 795504 : return false;
13781 743014 : return true;
13782 : }
13783 :
13784 : /* Return true if replacing VECTOR_MODE with ALT_VECTOR_MODE would not
13785 : change the chosen vector modes for analysis of a loop. */
13786 :
13787 : bool
13788 348262 : vect_chooses_same_modes_p (machine_mode vector_mode,
13789 : machine_mode alt_vector_mode)
13790 : {
13791 50801 : return (VECTOR_MODE_P (vector_mode)
13792 348262 : && VECTOR_MODE_P (alt_vector_mode)
13793 696524 : && (related_vector_mode (vector_mode,
13794 : GET_MODE_INNER (alt_vector_mode))
13795 348262 : == alt_vector_mode)
13796 374292 : && (related_vector_mode (alt_vector_mode,
13797 : GET_MODE_INNER (vector_mode))
13798 13015 : == vector_mode));
13799 : }
13800 :
13801 : /* Function vect_is_simple_use.
13802 :
13803 : Input:
13804 : VINFO - the vect info of the loop or basic block that is being vectorized.
13805 : OPERAND - operand in the loop or bb.
13806 : Output:
13807 : DEF_STMT_INFO_OUT (optional) - information about the defining stmt in
13808 : case OPERAND is an SSA_NAME that is defined in the vectorizable region
13809 : DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME;
13810 : the definition could be anywhere in the function
13811 : DT - the type of definition
13812 :
13813 : Returns whether a stmt with OPERAND can be vectorized.
13814 : For loops, supportable operands are constants, loop invariants, and operands
13815 : that are defined by the current iteration of the loop. Unsupportable
13816 : operands are those that are defined by a previous iteration of the loop (as
13817 : is the case in reduction/induction computations).
13818 : For basic blocks, supportable operands are constants and bb invariants.
13819 : For now, operands defined outside the basic block are not supported. */
13820 :
13821 : bool
13822 39595240 : vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt,
13823 : stmt_vec_info *def_stmt_info_out, gimple **def_stmt_out)
13824 : {
13825 39595240 : if (def_stmt_info_out)
13826 37750362 : *def_stmt_info_out = NULL;
13827 39595240 : if (def_stmt_out)
13828 9234056 : *def_stmt_out = NULL;
13829 39595240 : *dt = vect_unknown_def_type;
13830 :
13831 39595240 : if (dump_enabled_p ())
13832 : {
13833 749070 : dump_printf_loc (MSG_NOTE, vect_location,
13834 : "vect_is_simple_use: operand ");
13835 749070 : if (TREE_CODE (operand) == SSA_NAME
13836 749070 : && !SSA_NAME_IS_DEFAULT_DEF (operand))
13837 687415 : dump_gimple_expr (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (operand), 0);
13838 : else
13839 61655 : dump_generic_expr (MSG_NOTE, TDF_SLIM, operand);
13840 : }
13841 :
13842 39595240 : if (CONSTANT_CLASS_P (operand))
13843 2649491 : *dt = vect_constant_def;
13844 36945749 : else if (is_gimple_min_invariant (operand))
13845 331454 : *dt = vect_external_def;
13846 36614295 : else if (TREE_CODE (operand) != SSA_NAME)
13847 957 : *dt = vect_unknown_def_type;
13848 36613338 : else if (SSA_NAME_IS_DEFAULT_DEF (operand))
13849 525503 : *dt = vect_external_def;
13850 : else
13851 : {
13852 36087835 : gimple *def_stmt = SSA_NAME_DEF_STMT (operand);
13853 36087835 : stmt_vec_info stmt_vinfo = vinfo->lookup_def (operand);
13854 36087835 : if (!stmt_vinfo)
13855 753984 : *dt = vect_external_def;
13856 : else
13857 : {
13858 35333851 : stmt_vinfo = vect_stmt_to_vectorize (stmt_vinfo);
13859 35333851 : def_stmt = stmt_vinfo->stmt;
13860 35333851 : *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo);
13861 35333851 : if (def_stmt_info_out)
13862 33497927 : *def_stmt_info_out = stmt_vinfo;
13863 : }
13864 36087835 : if (def_stmt_out)
13865 9024650 : *def_stmt_out = def_stmt;
13866 : }
13867 :
13868 39595240 : if (dump_enabled_p ())
13869 : {
13870 749070 : dump_printf (MSG_NOTE, ", type of def: ");
13871 749070 : switch (*dt)
13872 : {
13873 0 : case vect_uninitialized_def:
13874 0 : dump_printf (MSG_NOTE, "uninitialized\n");
13875 0 : break;
13876 51066 : case vect_constant_def:
13877 51066 : dump_printf (MSG_NOTE, "constant\n");
13878 51066 : break;
13879 25488 : case vect_external_def:
13880 25488 : dump_printf (MSG_NOTE, "external\n");
13881 25488 : break;
13882 537471 : case vect_internal_def:
13883 537471 : dump_printf (MSG_NOTE, "internal\n");
13884 537471 : break;
13885 105316 : case vect_induction_def:
13886 105316 : dump_printf (MSG_NOTE, "induction\n");
13887 105316 : break;
13888 26380 : case vect_reduction_def:
13889 26380 : dump_printf (MSG_NOTE, "reduction\n");
13890 26380 : break;
13891 482 : case vect_double_reduction_def:
13892 482 : dump_printf (MSG_NOTE, "double reduction\n");
13893 482 : break;
13894 2175 : case vect_nested_cycle:
13895 2175 : dump_printf (MSG_NOTE, "nested cycle\n");
13896 2175 : break;
13897 264 : case vect_first_order_recurrence:
13898 264 : dump_printf (MSG_NOTE, "first order recurrence\n");
13899 264 : break;
13900 0 : case vect_condition_def:
13901 0 : dump_printf (MSG_NOTE, "control flow\n");
13902 0 : break;
13903 428 : case vect_unknown_def_type:
13904 428 : dump_printf (MSG_NOTE, "unknown\n");
13905 428 : break;
13906 : }
13907 : }
13908 :
13909 39595240 : if (*dt == vect_unknown_def_type)
13910 : {
13911 59381 : if (dump_enabled_p ())
13912 428 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
13913 : "Unsupported pattern.\n");
13914 59381 : return false;
13915 : }
13916 :
13917 : return true;
13918 : }
13919 :
13920 : /* Function vect_is_simple_use.
13921 :
13922 : Same as vect_is_simple_use but determines the operand by operand
13923 : position OPERAND from either STMT or SLP_NODE, filling in *OP
13924 : and *SLP_DEF (when SLP_NODE is not NULL). */
13925 :
13926 : bool
13927 3366948 : vect_is_simple_use (vec_info *vinfo, slp_tree slp_node,
13928 : unsigned operand, tree *op, slp_tree *slp_def,
13929 : enum vect_def_type *dt,
13930 : tree *vectype, stmt_vec_info *def_stmt_info_out)
13931 : {
13932 3366948 : slp_tree child = SLP_TREE_CHILDREN (slp_node)[operand];
13933 3366948 : *slp_def = child;
13934 3366948 : *vectype = SLP_TREE_VECTYPE (child);
13935 3366948 : if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
13936 : {
13937 : /* ??? VEC_PERM nodes might be intermediate and their lane value
13938 : have no representative (nor do we build a VEC_PERM stmt for
13939 : the actual operation). Note for two-operator nodes we set
13940 : a representative but leave scalar stmts empty as we'd only
13941 : have one for a subset of lanes. Ideally no caller would
13942 : require *op for internal defs. */
13943 1745242 : if (SLP_TREE_REPRESENTATIVE (child))
13944 : {
13945 1744514 : *op = gimple_get_lhs (SLP_TREE_REPRESENTATIVE (child)->stmt);
13946 1744514 : return vect_is_simple_use (*op, vinfo, dt, def_stmt_info_out);
13947 : }
13948 : else
13949 : {
13950 728 : gcc_assert (SLP_TREE_PERMUTE_P (child));
13951 728 : *op = error_mark_node;
13952 728 : *dt = vect_internal_def;
13953 728 : if (def_stmt_info_out)
13954 0 : *def_stmt_info_out = NULL;
13955 728 : return true;
13956 : }
13957 : }
13958 : else
13959 : {
13960 1621706 : if (def_stmt_info_out)
13961 50196 : *def_stmt_info_out = NULL;
13962 1621706 : *op = SLP_TREE_SCALAR_OPS (child)[0];
13963 1621706 : *dt = SLP_TREE_DEF_TYPE (child);
13964 1621706 : return true;
13965 : }
13966 : }
13967 :
13968 : /* If OP is not NULL and is external or constant update its vector
13969 : type with VECTYPE. Returns true if successful or false if not,
13970 : for example when conflicting vector types are present. */
13971 :
13972 : bool
13973 2955234 : vect_maybe_update_slp_op_vectype (slp_tree op, tree vectype)
13974 : {
13975 2955234 : if (!op || SLP_TREE_DEF_TYPE (op) == vect_internal_def)
13976 : return true;
13977 1068545 : if (SLP_TREE_VECTYPE (op))
13978 66689 : return types_compatible_p (SLP_TREE_VECTYPE (op), vectype);
13979 : /* For external defs refuse to produce VECTOR_BOOLEAN_TYPE_P, those
13980 : should be handled by patters. Allow vect_constant_def for now
13981 : as well as the trivial single-lane uniform vect_external_def case
13982 : both of which we code-generate reasonably. */
13983 1001856 : if (VECTOR_BOOLEAN_TYPE_P (vectype)
13984 1047 : && SLP_TREE_DEF_TYPE (op) == vect_external_def
13985 1002527 : && SLP_TREE_LANES (op) > 1)
13986 : return false;
13987 1001657 : SLP_TREE_VECTYPE (op) = vectype;
13988 1001657 : return true;
13989 : }
13990 :
13991 : /* Function supportable_widening_operation
13992 :
13993 : Check whether an operation represented by the code CODE is a
13994 : widening operation that is supported by the target platform in
13995 : vector form (i.e., when operating on arguments of type VECTYPE_IN
13996 : producing a result of type VECTYPE_OUT).
13997 :
13998 : Widening operations we currently support are NOP (CONVERT), FLOAT,
13999 : FIX_TRUNC and WIDEN_MULT. This function checks if these operations
14000 : are supported by the target platform either directly (via vector
14001 : tree-codes), or via target builtins.
14002 :
14003 : When EVENODD_OK then also lane-swizzling operations are considered.
14004 :
14005 : Output:
14006 : - CODE1 and CODE2 are codes of vector operations to be used when
14007 : vectorizing the operation, if available.
14008 : - MULTI_STEP_CVT determines the number of required intermediate steps in
14009 : case of multi-step conversion (like char->short->int - in that case
14010 : MULTI_STEP_CVT will be 1).
14011 : - INTERM_TYPES contains the intermediate type required to perform the
14012 : widening operation (short in the above example). */
14013 :
14014 : bool
14015 446531 : supportable_widening_operation (code_helper code,
14016 : tree vectype_out, tree vectype_in,
14017 : bool evenodd_ok,
14018 : code_helper *code1,
14019 : code_helper *code2,
14020 : int *multi_step_cvt,
14021 : vec<tree> *interm_types)
14022 : {
14023 446531 : machine_mode vec_mode;
14024 446531 : enum insn_code icode1, icode2;
14025 446531 : optab optab1 = unknown_optab, optab2 = unknown_optab;
14026 446531 : tree vectype = vectype_in;
14027 446531 : tree wide_vectype = vectype_out;
14028 446531 : tree_code c1 = MAX_TREE_CODES, c2 = MAX_TREE_CODES;
14029 446531 : int i;
14030 446531 : tree prev_type, intermediate_type;
14031 446531 : machine_mode intermediate_mode, prev_mode;
14032 446531 : optab optab3, optab4;
14033 :
14034 446531 : *multi_step_cvt = 0;
14035 :
14036 446531 : switch (code.safe_as_tree_code ())
14037 : {
14038 : case MAX_TREE_CODES:
14039 : /* Don't set c1 and c2 if code is not a tree_code. */
14040 : break;
14041 :
14042 172250 : case WIDEN_MULT_EXPR:
14043 : /* The result of a vectorized widening operation usually requires
14044 : two vectors (because the widened results do not fit into one vector).
14045 : The generated vector results would normally be expected to be
14046 : generated in the same order as in the original scalar computation,
14047 : i.e. if 8 results are generated in each vector iteration, they are
14048 : to be organized as follows:
14049 : vect1: [res1,res2,res3,res4],
14050 : vect2: [res5,res6,res7,res8].
14051 :
14052 : However, in the special case that the result of the widening
14053 : operation is used in a reduction computation only, the order doesn't
14054 : matter (because when vectorizing a reduction we change the order of
14055 : the computation). Some targets can take advantage of this and
14056 : generate more efficient code. For example, targets like Altivec,
14057 : that support widen_mult using a sequence of {mult_even,mult_odd}
14058 : generate the following vectors:
14059 : vect1: [res1,res3,res5,res7],
14060 : vect2: [res2,res4,res6,res8].
14061 :
14062 : When vectorizing outer-loops, we execute the inner-loop sequentially
14063 : (each vectorized inner-loop iteration contributes to VF outer-loop
14064 : iterations in parallel). We therefore don't allow to change the
14065 : order of the computation in the inner-loop during outer-loop
14066 : vectorization. */
14067 : /* TODO: Another case in which order doesn't *really* matter is when we
14068 : widen and then contract again, e.g. (short)((int)x * y >> 8).
14069 : Normally, pack_trunc performs an even/odd permute, whereas the
14070 : repack from an even/odd expansion would be an interleave, which
14071 : would be significantly simpler for e.g. AVX2. */
14072 : /* In any case, in order to avoid duplicating the code below, recurse
14073 : on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values
14074 : are properly set up for the caller. If we fail, we'll continue with
14075 : a VEC_WIDEN_MULT_LO/HI_EXPR check. */
14076 172250 : if (evenodd_ok
14077 172250 : && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR,
14078 : vectype_out, vectype_in,
14079 : evenodd_ok, code1,
14080 : code2, multi_step_cvt,
14081 : interm_types))
14082 93193 : return true;
14083 : c1 = VEC_WIDEN_MULT_LO_EXPR;
14084 : c2 = VEC_WIDEN_MULT_HI_EXPR;
14085 : break;
14086 :
14087 : case DOT_PROD_EXPR:
14088 353338 : c1 = DOT_PROD_EXPR;
14089 353338 : c2 = DOT_PROD_EXPR;
14090 : break;
14091 :
14092 0 : case SAD_EXPR:
14093 0 : c1 = SAD_EXPR;
14094 0 : c2 = SAD_EXPR;
14095 0 : break;
14096 :
14097 170220 : case VEC_WIDEN_MULT_EVEN_EXPR:
14098 : /* Support the recursion induced just above. */
14099 170220 : c1 = VEC_WIDEN_MULT_EVEN_EXPR;
14100 170220 : c2 = VEC_WIDEN_MULT_ODD_EXPR;
14101 170220 : break;
14102 :
14103 9305 : case WIDEN_LSHIFT_EXPR:
14104 9305 : c1 = VEC_WIDEN_LSHIFT_LO_EXPR;
14105 9305 : c2 = VEC_WIDEN_LSHIFT_HI_EXPR;
14106 9305 : break;
14107 :
14108 34830 : CASE_CONVERT:
14109 34830 : c1 = VEC_UNPACK_LO_EXPR;
14110 34830 : c2 = VEC_UNPACK_HI_EXPR;
14111 34830 : break;
14112 :
14113 7207 : case FLOAT_EXPR:
14114 7207 : c1 = VEC_UNPACK_FLOAT_LO_EXPR;
14115 7207 : c2 = VEC_UNPACK_FLOAT_HI_EXPR;
14116 7207 : break;
14117 :
14118 110 : case FIX_TRUNC_EXPR:
14119 110 : c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR;
14120 110 : c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR;
14121 110 : break;
14122 :
14123 0 : default:
14124 0 : gcc_unreachable ();
14125 : }
14126 :
14127 353338 : if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR)
14128 : std::swap (c1, c2);
14129 :
14130 353338 : if (code == FIX_TRUNC_EXPR)
14131 : {
14132 : /* The signedness is determined from output operand. */
14133 110 : optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
14134 110 : optab2 = optab_for_tree_code (c2, vectype_out, optab_default);
14135 : }
14136 625036 : else if (CONVERT_EXPR_CODE_P (code.safe_as_tree_code ())
14137 34830 : && VECTOR_BOOLEAN_TYPE_P (wide_vectype)
14138 5777 : && VECTOR_BOOLEAN_TYPE_P (vectype)
14139 5777 : && TYPE_MODE (wide_vectype) == TYPE_MODE (vectype)
14140 300932 : && SCALAR_INT_MODE_P (TYPE_MODE (vectype)))
14141 : {
14142 : /* If the input and result modes are the same, a different optab
14143 : is needed where we pass in the number of units in vectype. */
14144 : optab1 = vec_unpacks_sbool_lo_optab;
14145 : optab2 = vec_unpacks_sbool_hi_optab;
14146 : }
14147 :
14148 353338 : vec_mode = TYPE_MODE (vectype);
14149 353338 : if (widening_fn_p (code))
14150 : {
14151 : /* If this is an internal fn then we must check whether the target
14152 : supports either a low-high split or an even-odd split. */
14153 52609 : internal_fn ifn = as_internal_fn ((combined_fn) code);
14154 :
14155 52609 : internal_fn lo, hi, even, odd;
14156 52609 : lookup_hilo_internal_fn (ifn, &lo, &hi);
14157 52609 : if (BYTES_BIG_ENDIAN)
14158 : std::swap (lo, hi);
14159 52609 : *code1 = as_combined_fn (lo);
14160 52609 : *code2 = as_combined_fn (hi);
14161 52609 : optab1 = direct_internal_fn_optab (lo, {vectype, vectype});
14162 52609 : optab2 = direct_internal_fn_optab (hi, {vectype, vectype});
14163 :
14164 : /* If we don't support low-high, then check for even-odd. */
14165 52609 : if (!optab1
14166 52609 : || (icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing
14167 0 : || !optab2
14168 52609 : || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing)
14169 : {
14170 52609 : lookup_evenodd_internal_fn (ifn, &even, &odd);
14171 52609 : *code1 = as_combined_fn (even);
14172 52609 : *code2 = as_combined_fn (odd);
14173 52609 : optab1 = direct_internal_fn_optab (even, {vectype, vectype});
14174 52609 : optab2 = direct_internal_fn_optab (odd, {vectype, vectype});
14175 : }
14176 : }
14177 300729 : else if (code.is_tree_code ())
14178 : {
14179 300729 : if (code == FIX_TRUNC_EXPR)
14180 : {
14181 : /* The signedness is determined from output operand. */
14182 110 : optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
14183 110 : optab2 = optab_for_tree_code (c2, vectype_out, optab_default);
14184 : }
14185 300619 : else if (CONVERT_EXPR_CODE_P ((tree_code) code.safe_as_tree_code ())
14186 34830 : && VECTOR_BOOLEAN_TYPE_P (wide_vectype)
14187 5777 : && VECTOR_BOOLEAN_TYPE_P (vectype)
14188 5777 : && TYPE_MODE (wide_vectype) == TYPE_MODE (vectype)
14189 300932 : && SCALAR_INT_MODE_P (TYPE_MODE (vectype)))
14190 : {
14191 : /* If the input and result modes are the same, a different optab
14192 : is needed where we pass in the number of units in vectype. */
14193 : optab1 = vec_unpacks_sbool_lo_optab;
14194 : optab2 = vec_unpacks_sbool_hi_optab;
14195 : }
14196 : else
14197 : {
14198 300306 : optab1 = optab_for_tree_code (c1, vectype, optab_default);
14199 300306 : optab2 = optab_for_tree_code (c2, vectype, optab_default);
14200 : }
14201 300729 : *code1 = c1;
14202 300729 : *code2 = c2;
14203 : }
14204 :
14205 353338 : if (!optab1 || !optab2)
14206 : return false;
14207 :
14208 353338 : if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing
14209 353338 : || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing)
14210 209213 : return false;
14211 :
14212 :
14213 144125 : if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype)
14214 144125 : && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype))
14215 : {
14216 133864 : if (!VECTOR_BOOLEAN_TYPE_P (vectype))
14217 : return true;
14218 : /* For scalar masks we may have different boolean
14219 : vector types having the same QImode. Thus we
14220 : add additional check for elements number. */
14221 2935 : if (known_eq (TYPE_VECTOR_SUBPARTS (vectype),
14222 : TYPE_VECTOR_SUBPARTS (wide_vectype) * 2))
14223 : return true;
14224 : }
14225 :
14226 : /* Check if it's a multi-step conversion that can be done using intermediate
14227 : types. */
14228 :
14229 10382 : prev_type = vectype;
14230 10382 : prev_mode = vec_mode;
14231 :
14232 219844 : if (!CONVERT_EXPR_CODE_P (code.safe_as_tree_code ()))
14233 : return false;
14234 :
14235 : /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
14236 : intermediate steps in promotion sequence. We try
14237 : MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
14238 : not. */
14239 10330 : interm_types->create (MAX_INTERM_CVT_STEPS);
14240 11604 : for (i = 0; i < MAX_INTERM_CVT_STEPS; i++)
14241 : {
14242 11604 : intermediate_mode = insn_data[icode1].operand[0].mode;
14243 11604 : if (VECTOR_BOOLEAN_TYPE_P (prev_type))
14244 3714 : intermediate_type
14245 3714 : = vect_halve_mask_nunits (prev_type, intermediate_mode);
14246 7890 : else if (VECTOR_MODE_P (intermediate_mode))
14247 : {
14248 7890 : tree intermediate_element_type
14249 7890 : = lang_hooks.types.type_for_mode (GET_MODE_INNER (intermediate_mode),
14250 7890 : TYPE_UNSIGNED (prev_type));
14251 7890 : intermediate_type
14252 7890 : = build_vector_type_for_mode (intermediate_element_type,
14253 : intermediate_mode);
14254 7890 : }
14255 : else
14256 0 : intermediate_type
14257 0 : = lang_hooks.types.type_for_mode (intermediate_mode,
14258 0 : TYPE_UNSIGNED (prev_type));
14259 :
14260 11604 : if (VECTOR_BOOLEAN_TYPE_P (intermediate_type)
14261 3714 : && VECTOR_BOOLEAN_TYPE_P (wide_vectype)
14262 3714 : && intermediate_mode == TYPE_MODE (wide_vectype)
14263 11761 : && SCALAR_INT_MODE_P (intermediate_mode))
14264 : {
14265 : /* If the input and result modes are the same, a different optab
14266 : is needed where we pass in the number of units in vectype. */
14267 : optab3 = vec_unpacks_sbool_lo_optab;
14268 : optab4 = vec_unpacks_sbool_hi_optab;
14269 : }
14270 : else
14271 : {
14272 11447 : optab3 = optab_for_tree_code (c1, intermediate_type, optab_default);
14273 11447 : optab4 = optab_for_tree_code (c2, intermediate_type, optab_default);
14274 : }
14275 :
14276 11604 : if (!optab3 || !optab4
14277 11604 : || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing
14278 11588 : || insn_data[icode1].operand[0].mode != intermediate_mode
14279 11588 : || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing
14280 11588 : || insn_data[icode2].operand[0].mode != intermediate_mode
14281 11588 : || ((icode1 = optab_handler (optab3, intermediate_mode))
14282 : == CODE_FOR_nothing)
14283 22959 : || ((icode2 = optab_handler (optab4, intermediate_mode))
14284 : == CODE_FOR_nothing))
14285 : break;
14286 :
14287 11355 : interm_types->quick_push (intermediate_type);
14288 11355 : (*multi_step_cvt)++;
14289 :
14290 11355 : if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype)
14291 11355 : && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype))
14292 : {
14293 10117 : if (!VECTOR_BOOLEAN_TYPE_P (vectype))
14294 : return true;
14295 2780 : if (known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type),
14296 : TYPE_VECTOR_SUBPARTS (wide_vectype) * 2))
14297 : return true;
14298 : }
14299 :
14300 1274 : prev_type = intermediate_type;
14301 1274 : prev_mode = intermediate_mode;
14302 : }
14303 :
14304 249 : interm_types->release ();
14305 249 : return false;
14306 : }
14307 :
14308 :
14309 : /* Function supportable_narrowing_operation
14310 :
14311 : Check whether an operation represented by the code CODE is a
14312 : narrowing operation that is supported by the target platform in
14313 : vector form (i.e., when operating on arguments of type VECTYPE_IN
14314 : and producing a result of type VECTYPE_OUT).
14315 :
14316 : Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC
14317 : and FLOAT. This function checks if these operations are supported by
14318 : the target platform directly via vector tree-codes.
14319 :
14320 : Output:
14321 : - CODE1 is the code of a vector operation to be used when
14322 : vectorizing the operation, if available.
14323 : - MULTI_STEP_CVT determines the number of required intermediate steps in
14324 : case of multi-step conversion (like int->short->char - in that case
14325 : MULTI_STEP_CVT will be 1).
14326 : - INTERM_TYPES contains the intermediate type required to perform the
14327 : narrowing operation (short in the above example). */
14328 :
14329 : bool
14330 34827 : supportable_narrowing_operation (code_helper code,
14331 : tree vectype_out, tree vectype_in,
14332 : code_helper *code1, int *multi_step_cvt,
14333 : vec<tree> *interm_types)
14334 : {
14335 34827 : machine_mode vec_mode;
14336 34827 : enum insn_code icode1;
14337 34827 : optab optab1, interm_optab;
14338 34827 : tree vectype = vectype_in;
14339 34827 : tree narrow_vectype = vectype_out;
14340 34827 : enum tree_code c1;
14341 34827 : tree intermediate_type, prev_type;
14342 34827 : machine_mode intermediate_mode, prev_mode;
14343 34827 : int i;
14344 34827 : unsigned HOST_WIDE_INT n_elts;
14345 34827 : bool uns;
14346 :
14347 34827 : if (!code.is_tree_code ())
14348 : return false;
14349 :
14350 34827 : *multi_step_cvt = 0;
14351 34827 : switch ((tree_code) code)
14352 : {
14353 34134 : CASE_CONVERT:
14354 34134 : c1 = VEC_PACK_TRUNC_EXPR;
14355 34134 : if (VECTOR_BOOLEAN_TYPE_P (narrow_vectype)
14356 9544 : && VECTOR_BOOLEAN_TYPE_P (vectype)
14357 9544 : && SCALAR_INT_MODE_P (TYPE_MODE (vectype))
14358 4546 : && TYPE_VECTOR_SUBPARTS (vectype).is_constant (&n_elts)
14359 38680 : && n_elts < BITS_PER_UNIT)
14360 : optab1 = vec_pack_sbool_trunc_optab;
14361 : else
14362 32331 : optab1 = optab_for_tree_code (c1, vectype, optab_default);
14363 : break;
14364 :
14365 483 : case FIX_TRUNC_EXPR:
14366 483 : c1 = VEC_PACK_FIX_TRUNC_EXPR;
14367 : /* The signedness is determined from output operand. */
14368 483 : optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
14369 483 : break;
14370 :
14371 210 : case FLOAT_EXPR:
14372 210 : c1 = VEC_PACK_FLOAT_EXPR;
14373 210 : optab1 = optab_for_tree_code (c1, vectype, optab_default);
14374 210 : break;
14375 :
14376 0 : default:
14377 0 : gcc_unreachable ();
14378 : }
14379 :
14380 34827 : if (!optab1)
14381 : return false;
14382 :
14383 34827 : vec_mode = TYPE_MODE (vectype);
14384 34827 : if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing)
14385 : return false;
14386 :
14387 30797 : *code1 = c1;
14388 :
14389 30797 : if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype))
14390 : {
14391 18719 : if (!VECTOR_BOOLEAN_TYPE_P (vectype))
14392 : return true;
14393 : /* For scalar masks we may have different boolean
14394 : vector types having the same QImode. Thus we
14395 : add additional check for elements number. */
14396 4595 : if (known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2,
14397 : TYPE_VECTOR_SUBPARTS (narrow_vectype)))
14398 : return true;
14399 : }
14400 :
14401 12079 : if (code == FLOAT_EXPR)
14402 : return false;
14403 :
14404 : /* Check if it's a multi-step conversion that can be done using intermediate
14405 : types. */
14406 12079 : prev_mode = vec_mode;
14407 12079 : prev_type = vectype;
14408 12079 : if (code == FIX_TRUNC_EXPR)
14409 87 : uns = TYPE_UNSIGNED (vectype_out);
14410 : else
14411 11992 : uns = TYPE_UNSIGNED (vectype);
14412 :
14413 : /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer
14414 : conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more
14415 : costly than signed. */
14416 12079 : if (code == FIX_TRUNC_EXPR && uns)
14417 : {
14418 28 : enum insn_code icode2;
14419 :
14420 28 : intermediate_type
14421 28 : = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0);
14422 28 : interm_optab
14423 28 : = optab_for_tree_code (c1, intermediate_type, optab_default);
14424 28 : if (interm_optab != unknown_optab
14425 28 : && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing
14426 28 : && insn_data[icode1].operand[0].mode
14427 28 : == insn_data[icode2].operand[0].mode)
14428 : {
14429 : uns = false;
14430 : optab1 = interm_optab;
14431 : icode1 = icode2;
14432 : }
14433 : }
14434 :
14435 : /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
14436 : intermediate steps in promotion sequence. We try
14437 : MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */
14438 12079 : interm_types->create (MAX_INTERM_CVT_STEPS);
14439 26066 : for (i = 0; i < MAX_INTERM_CVT_STEPS; i++)
14440 : {
14441 13987 : intermediate_mode = insn_data[icode1].operand[0].mode;
14442 13987 : if (VECTOR_BOOLEAN_TYPE_P (prev_type))
14443 5915 : intermediate_type
14444 5915 : = vect_double_mask_nunits (prev_type, intermediate_mode);
14445 : else
14446 8072 : intermediate_type
14447 8072 : = lang_hooks.types.type_for_mode (intermediate_mode, uns);
14448 13987 : if (VECTOR_BOOLEAN_TYPE_P (intermediate_type)
14449 5915 : && VECTOR_BOOLEAN_TYPE_P (prev_type)
14450 5915 : && SCALAR_INT_MODE_P (prev_mode)
14451 2681 : && TYPE_VECTOR_SUBPARTS (intermediate_type).is_constant (&n_elts)
14452 16668 : && n_elts < BITS_PER_UNIT)
14453 : interm_optab = vec_pack_sbool_trunc_optab;
14454 : else
14455 13870 : interm_optab
14456 13870 : = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type,
14457 : optab_default);
14458 117 : if (!interm_optab
14459 13987 : || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing)
14460 13987 : || insn_data[icode1].operand[0].mode != intermediate_mode
14461 27857 : || ((icode1 = optab_handler (interm_optab, intermediate_mode))
14462 : == CODE_FOR_nothing))
14463 : break;
14464 :
14465 13094 : interm_types->quick_push (intermediate_type);
14466 13094 : (*multi_step_cvt)++;
14467 :
14468 13094 : if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype))
14469 : {
14470 11186 : if (!VECTOR_BOOLEAN_TYPE_P (vectype))
14471 : return true;
14472 3984 : if (known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2,
14473 : TYPE_VECTOR_SUBPARTS (narrow_vectype)))
14474 : return true;
14475 : }
14476 :
14477 1908 : prev_mode = intermediate_mode;
14478 1908 : prev_type = intermediate_type;
14479 1908 : optab1 = interm_optab;
14480 : }
14481 :
14482 893 : interm_types->release ();
14483 893 : return false;
14484 : }
14485 :
14486 : /* Function supportable_indirect_convert_operation
14487 :
14488 : Check whether an operation represented by the code CODE is single or multi
14489 : operations that are supported by the target platform in
14490 : vector form (i.e., when operating on arguments of type VECTYPE_IN
14491 : producing a result of type VECTYPE_OUT).
14492 :
14493 : Convert operations we currently support directly are FIX_TRUNC and FLOAT.
14494 : This function checks if these operations are supported
14495 : by the target platform directly (via vector tree-codes).
14496 :
14497 : Output:
14498 : - converts contains some pairs to perform the convert operation,
14499 : the pair's first is the intermediate type, and its second is the code of
14500 : a vector operation to be used when converting the operation from the
14501 : previous type to the intermediate type. */
14502 : bool
14503 71418 : supportable_indirect_convert_operation (code_helper code,
14504 : tree vectype_out,
14505 : tree vectype_in,
14506 : vec<std::pair<tree, tree_code> > &converts,
14507 : tree op0, slp_tree slp_op0)
14508 : {
14509 71418 : bool found_mode = false;
14510 71418 : scalar_mode lhs_mode = GET_MODE_INNER (TYPE_MODE (vectype_out));
14511 71418 : scalar_mode rhs_mode = GET_MODE_INNER (TYPE_MODE (vectype_in));
14512 71418 : tree_code tc1, tc2, code1, code2;
14513 :
14514 71418 : tree cvt_type = NULL_TREE;
14515 71418 : poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (vectype_in);
14516 :
14517 71418 : if (supportable_convert_operation ((tree_code) code,
14518 : vectype_out,
14519 : vectype_in,
14520 : &tc1))
14521 : {
14522 16312 : converts.safe_push (std::make_pair (vectype_out, tc1));
14523 16312 : return true;
14524 : }
14525 :
14526 : /* For conversions between float and integer types try whether
14527 : we can use intermediate signed integer types to support the
14528 : conversion. */
14529 110212 : if (GET_MODE_SIZE (lhs_mode) != GET_MODE_SIZE (rhs_mode)
14530 55106 : && (code == FLOAT_EXPR
14531 3134 : || (code == FIX_TRUNC_EXPR && !flag_trapping_math)))
14532 : {
14533 456 : bool demotion = GET_MODE_SIZE (rhs_mode) > GET_MODE_SIZE (lhs_mode);
14534 228 : bool float_expr_p = code == FLOAT_EXPR;
14535 228 : unsigned short target_size;
14536 228 : scalar_mode intermediate_mode;
14537 228 : if (demotion)
14538 : {
14539 84 : intermediate_mode = lhs_mode;
14540 84 : target_size = GET_MODE_SIZE (rhs_mode);
14541 : }
14542 : else
14543 : {
14544 144 : target_size = GET_MODE_SIZE (lhs_mode);
14545 144 : if (!int_mode_for_size
14546 144 : (GET_MODE_BITSIZE (rhs_mode), 0).exists (&intermediate_mode))
14547 132 : return false;
14548 : }
14549 228 : code1 = float_expr_p ? (tree_code) code : NOP_EXPR;
14550 : code2 = float_expr_p ? NOP_EXPR : (tree_code) code;
14551 228 : opt_scalar_mode mode_iter;
14552 382 : FOR_EACH_2XWIDER_MODE (mode_iter, intermediate_mode)
14553 : {
14554 382 : intermediate_mode = mode_iter.require ();
14555 :
14556 764 : if (GET_MODE_SIZE (intermediate_mode) > target_size)
14557 : break;
14558 :
14559 324 : scalar_mode cvt_mode;
14560 324 : if (!int_mode_for_size
14561 324 : (GET_MODE_BITSIZE (intermediate_mode), 0).exists (&cvt_mode))
14562 : break;
14563 :
14564 294 : cvt_type = build_nonstandard_integer_type
14565 294 : (GET_MODE_BITSIZE (cvt_mode), 0);
14566 :
14567 : /* Check if the intermediate type can hold OP0's range.
14568 : When converting from float to integer this is not necessary
14569 : because values that do not fit the (smaller) target type are
14570 : unspecified anyway. */
14571 294 : if (demotion && float_expr_p)
14572 : {
14573 8 : wide_int op_min_value, op_max_value;
14574 : /* For vector form, it looks like op0 doesn't have RANGE_INFO.
14575 : In the future, if it is supported, changes may need to be made
14576 : to this part, such as checking the RANGE of each element
14577 : in the vector. */
14578 8 : if (slp_op0)
14579 : {
14580 4 : tree def;
14581 : /* ??? Merge ranges in case of more than one lane. */
14582 4 : if (SLP_TREE_LANES (slp_op0) != 1
14583 0 : || !(def = vect_get_slp_scalar_def (slp_op0, 0))
14584 4 : || !vect_get_range_info (def,
14585 : &op_min_value, &op_max_value))
14586 : break;
14587 : }
14588 4 : else if (!op0
14589 0 : || TREE_CODE (op0) != SSA_NAME
14590 0 : || !SSA_NAME_RANGE_INFO (op0)
14591 4 : || !vect_get_range_info (op0, &op_min_value,
14592 : &op_max_value))
14593 : break;
14594 :
14595 0 : if (cvt_type == NULL_TREE
14596 0 : || (wi::min_precision (op_max_value, SIGNED)
14597 0 : > TYPE_PRECISION (cvt_type))
14598 0 : || (wi::min_precision (op_min_value, SIGNED)
14599 0 : > TYPE_PRECISION (cvt_type)))
14600 0 : continue;
14601 8 : }
14602 :
14603 286 : cvt_type = get_related_vectype_for_scalar_type (TYPE_MODE (vectype_in),
14604 : cvt_type,
14605 : nelts);
14606 : /* This should only happened for SLP as long as loop vectorizer
14607 : only supports same-sized vector. */
14608 440 : if (cvt_type == NULL_TREE
14609 418 : || maybe_ne (TYPE_VECTOR_SUBPARTS (cvt_type), nelts)
14610 286 : || !supportable_convert_operation ((tree_code) code1,
14611 : vectype_out,
14612 : cvt_type, &tc1)
14613 486 : || !supportable_convert_operation ((tree_code) code2,
14614 : cvt_type,
14615 : vectype_in, &tc2))
14616 154 : continue;
14617 :
14618 : found_mode = true;
14619 : break;
14620 : }
14621 :
14622 228 : if (found_mode)
14623 : {
14624 132 : converts.safe_push (std::make_pair (cvt_type, tc2));
14625 132 : if (TYPE_MODE (cvt_type) != TYPE_MODE (vectype_out))
14626 132 : converts.safe_push (std::make_pair (vectype_out, tc1));
14627 132 : return true;
14628 : }
14629 : }
14630 : return false;
14631 : }
14632 :
14633 : /* Generate and return a vector mask of MASK_TYPE such that
14634 : mask[I] is true iff J + START_INDEX < END_INDEX for all J <= I.
14635 : Add the statements to SEQ. */
14636 :
14637 : tree
14638 0 : vect_gen_while (gimple_seq *seq, tree mask_type, tree start_index,
14639 : tree end_index, const char *name)
14640 : {
14641 0 : tree cmp_type = TREE_TYPE (start_index);
14642 0 : gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT,
14643 : cmp_type, mask_type,
14644 : OPTIMIZE_FOR_SPEED));
14645 0 : gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3,
14646 : start_index, end_index,
14647 : build_zero_cst (mask_type));
14648 0 : tree tmp;
14649 0 : if (name)
14650 0 : tmp = make_temp_ssa_name (mask_type, NULL, name);
14651 : else
14652 0 : tmp = make_ssa_name (mask_type);
14653 0 : gimple_call_set_lhs (call, tmp);
14654 0 : gimple_seq_add_stmt (seq, call);
14655 0 : return tmp;
14656 : }
14657 :
14658 : /* Generate a vector mask of type MASK_TYPE for which index I is false iff
14659 : J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */
14660 :
14661 : tree
14662 0 : vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index,
14663 : tree end_index)
14664 : {
14665 0 : tree tmp = vect_gen_while (seq, mask_type, start_index, end_index);
14666 0 : return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp);
14667 : }
14668 :
14669 : /* Try to compute the vector types required to vectorize STMT_INFO,
14670 : returning true on success and false if vectorization isn't possible.
14671 : If GROUP_SIZE is nonzero and we're performing BB vectorization,
14672 : take sure that the number of elements in the vectors is no bigger
14673 : than GROUP_SIZE.
14674 :
14675 : On success:
14676 :
14677 : - Set *STMT_VECTYPE_OUT to:
14678 : - NULL_TREE if the statement doesn't need to be vectorized;
14679 : - the equivalent of STMT_VINFO_VECTYPE otherwise.
14680 :
14681 : - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum
14682 : number of units needed to vectorize STMT_INFO, or NULL_TREE if the
14683 : statement does not help to determine the overall number of units. */
14684 :
14685 : opt_result
14686 5253029 : vect_get_vector_types_for_stmt (vec_info *vinfo, stmt_vec_info stmt_info,
14687 : tree *stmt_vectype_out,
14688 : tree *nunits_vectype_out,
14689 : unsigned int group_size)
14690 : {
14691 5253029 : gimple *stmt = stmt_info->stmt;
14692 :
14693 : /* For BB vectorization, we should always have a group size once we've
14694 : constructed the SLP tree; the only valid uses of zero GROUP_SIZEs
14695 : are tentative requests during things like early data reference
14696 : analysis and pattern recognition. */
14697 5253029 : if (is_a <bb_vec_info> (vinfo))
14698 4429478 : gcc_assert (vinfo->slp_instances.is_empty () || group_size != 0);
14699 : else
14700 : group_size = 0;
14701 :
14702 5253029 : *stmt_vectype_out = NULL_TREE;
14703 5253029 : *nunits_vectype_out = NULL_TREE;
14704 :
14705 5253029 : if (gimple_get_lhs (stmt) == NULL_TREE
14706 : /* Allow vector conditionals through here. */
14707 1543 : && !is_a <gcond *> (stmt)
14708 : /* MASK_STORE and friends have no lhs, but are ok. */
14709 5256095 : && !(is_gimple_call (stmt)
14710 1543 : && gimple_call_internal_p (stmt)
14711 1523 : && internal_store_fn_p (gimple_call_internal_fn (stmt))))
14712 : {
14713 20 : if (is_a <gcall *> (stmt))
14714 : {
14715 : /* Ignore calls with no lhs. These must be calls to
14716 : #pragma omp simd functions, and what vectorization factor
14717 : it really needs can't be determined until
14718 : vectorizable_simd_clone_call. */
14719 20 : if (dump_enabled_p ())
14720 18 : dump_printf_loc (MSG_NOTE, vect_location,
14721 : "defer to SIMD clone analysis.\n");
14722 20 : return opt_result::success ();
14723 : }
14724 :
14725 0 : return opt_result::failure_at (stmt,
14726 : "not vectorized: irregular stmt: %G", stmt);
14727 : }
14728 :
14729 5253009 : tree vectype;
14730 5253009 : tree scalar_type = NULL_TREE;
14731 5253009 : if (group_size == 0 && STMT_VINFO_VECTYPE (stmt_info))
14732 : {
14733 1349591 : vectype = STMT_VINFO_VECTYPE (stmt_info);
14734 1349591 : if (dump_enabled_p ())
14735 76898 : dump_printf_loc (MSG_NOTE, vect_location,
14736 : "precomputed vectype: %T\n", vectype);
14737 : }
14738 3903418 : else if (vect_use_mask_type_p (stmt_info))
14739 : {
14740 176180 : unsigned int precision = stmt_info->mask_precision;
14741 176180 : scalar_type = build_nonstandard_integer_type (precision, 1);
14742 176180 : vectype = get_mask_type_for_scalar_type (vinfo, scalar_type, group_size);
14743 176180 : if (!vectype)
14744 0 : return opt_result::failure_at (stmt, "not vectorized: unsupported"
14745 : " data-type %T\n", scalar_type);
14746 176180 : if (dump_enabled_p ())
14747 4504 : dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n", vectype);
14748 : }
14749 : else
14750 : {
14751 : /* If we got here with a gcond it means that the target had no available vector
14752 : mode for the scalar type. We can't vectorize so abort. */
14753 3727238 : if (is_a <gcond *> (stmt))
14754 0 : return opt_result::failure_at (stmt,
14755 : "not vectorized:"
14756 : " unsupported data-type for gcond %T\n",
14757 : scalar_type);
14758 :
14759 3727238 : if (data_reference *dr = STMT_VINFO_DATA_REF (stmt_info))
14760 1445074 : scalar_type = TREE_TYPE (DR_REF (dr));
14761 : else
14762 2282164 : scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
14763 :
14764 3727238 : if (dump_enabled_p ())
14765 : {
14766 60872 : if (group_size)
14767 7373 : dump_printf_loc (MSG_NOTE, vect_location,
14768 : "get vectype for scalar type (group size %d):"
14769 : " %T\n", group_size, scalar_type);
14770 : else
14771 53499 : dump_printf_loc (MSG_NOTE, vect_location,
14772 : "get vectype for scalar type: %T\n", scalar_type);
14773 : }
14774 3727238 : vectype = get_vectype_for_scalar_type (vinfo, scalar_type, group_size);
14775 3727238 : if (!vectype)
14776 194857 : return opt_result::failure_at (stmt,
14777 : "not vectorized:"
14778 : " unsupported data-type %T\n",
14779 : scalar_type);
14780 :
14781 3532381 : if (dump_enabled_p ())
14782 60679 : dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n", vectype);
14783 : }
14784 :
14785 3785459 : if (scalar_type && VECTOR_MODE_P (TYPE_MODE (scalar_type)))
14786 0 : return opt_result::failure_at (stmt,
14787 : "not vectorized: vector stmt in loop:%G",
14788 : stmt);
14789 :
14790 5058152 : *stmt_vectype_out = vectype;
14791 :
14792 : /* Don't try to compute scalar types if the stmt produces a boolean
14793 : vector; use the existing vector type instead. */
14794 5058152 : tree nunits_vectype = vectype;
14795 5058152 : if (!VECTOR_BOOLEAN_TYPE_P (vectype))
14796 : {
14797 : /* The number of units is set according to the smallest scalar
14798 : type (or the largest vector size, but we only support one
14799 : vector size per vectorization). */
14800 4576843 : scalar_type = vect_get_smallest_scalar_type (stmt_info,
14801 4576843 : TREE_TYPE (vectype));
14802 4576843 : if (!types_compatible_p (scalar_type, TREE_TYPE (vectype)))
14803 : {
14804 970491 : if (dump_enabled_p ())
14805 9541 : dump_printf_loc (MSG_NOTE, vect_location,
14806 : "get vectype for smallest scalar type: %T\n",
14807 : scalar_type);
14808 970491 : nunits_vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
14809 : group_size);
14810 970491 : if (!nunits_vectype)
14811 10 : return opt_result::failure_at
14812 10 : (stmt, "not vectorized: unsupported data-type %T\n",
14813 : scalar_type);
14814 970481 : if (dump_enabled_p ())
14815 9541 : dump_printf_loc (MSG_NOTE, vect_location, "nunits vectype: %T\n",
14816 : nunits_vectype);
14817 : }
14818 : }
14819 :
14820 5058142 : if (!multiple_p (TYPE_VECTOR_SUBPARTS (nunits_vectype),
14821 5058142 : TYPE_VECTOR_SUBPARTS (*stmt_vectype_out)))
14822 0 : return opt_result::failure_at (stmt,
14823 : "Not vectorized: Incompatible number "
14824 : "of vector subparts between %T and %T\n",
14825 : nunits_vectype, *stmt_vectype_out);
14826 :
14827 5058142 : if (dump_enabled_p ())
14828 : {
14829 142081 : dump_printf_loc (MSG_NOTE, vect_location, "nunits = ");
14830 142081 : dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (nunits_vectype));
14831 142081 : dump_printf (MSG_NOTE, "\n");
14832 : }
14833 :
14834 5058142 : *nunits_vectype_out = nunits_vectype;
14835 5058142 : return opt_result::success ();
14836 : }
14837 :
14838 : /* Generate and return statement sequence that sets vector length LEN that is:
14839 :
14840 : min_of_start_and_end = min (START_INDEX, END_INDEX);
14841 : left_len = END_INDEX - min_of_start_and_end;
14842 : rhs = min (left_len, LEN_LIMIT);
14843 : LEN = rhs;
14844 :
14845 : Note: the cost of the code generated by this function is modeled
14846 : by vect_estimate_min_profitable_iters, so changes here may need
14847 : corresponding changes there. */
14848 :
14849 : gimple_seq
14850 0 : vect_gen_len (tree len, tree start_index, tree end_index, tree len_limit)
14851 : {
14852 0 : gimple_seq stmts = NULL;
14853 0 : tree len_type = TREE_TYPE (len);
14854 0 : gcc_assert (TREE_TYPE (start_index) == len_type);
14855 :
14856 0 : tree min = gimple_build (&stmts, MIN_EXPR, len_type, start_index, end_index);
14857 0 : tree left_len = gimple_build (&stmts, MINUS_EXPR, len_type, end_index, min);
14858 0 : tree rhs = gimple_build (&stmts, MIN_EXPR, len_type, left_len, len_limit);
14859 0 : gimple* stmt = gimple_build_assign (len, rhs);
14860 0 : gimple_seq_add_stmt (&stmts, stmt);
14861 :
14862 0 : return stmts;
14863 : }
14864 :
|
