Line data Source code
1 : /* SLP - Basic Block Vectorization
2 : Copyright (C) 2007-2026 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 : #define INCLUDE_ALGORITHM
24 : #include "system.h"
25 : #include "coretypes.h"
26 : #include "backend.h"
27 : #include "target.h"
28 : #include "rtl.h"
29 : #include "tree.h"
30 : #include "gimple.h"
31 : #include "tree-pass.h"
32 : #include "ssa.h"
33 : #include "optabs-tree.h"
34 : #include "insn-config.h"
35 : #include "recog.h" /* FIXME: for insn_data */
36 : #include "fold-const.h"
37 : #include "stor-layout.h"
38 : #include "gimple-iterator.h"
39 : #include "cfgloop.h"
40 : #include "tree-vectorizer.h"
41 : #include "langhooks.h"
42 : #include "gimple-walk.h"
43 : #include "dbgcnt.h"
44 : #include "tree-vector-builder.h"
45 : #include "vec-perm-indices.h"
46 : #include "gimple-fold.h"
47 : #include "internal-fn.h"
48 : #include "dump-context.h"
49 : #include "cfganal.h"
50 : #include "tree-eh.h"
51 : #include "tree-cfg.h"
52 : #include "alloc-pool.h"
53 : #include "sreal.h"
54 : #include "predict.h"
55 :
56 : #define REDUC_GROUP_FIRST_ELEMENT(S) \
57 : (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
58 :
59 : static bool vect_transform_slp_perm_load_1 (vec_info *, slp_tree,
60 : load_permutation_t &,
61 : const vec<tree> &,
62 : gimple_stmt_iterator *,
63 : poly_uint64, bool, bool,
64 : unsigned *,
65 : unsigned * = nullptr,
66 : bool = false);
67 : static int vectorizable_slp_permutation_1 (vec_info *, gimple_stmt_iterator *,
68 : slp_tree, lane_permutation_t &,
69 : vec<slp_tree> &, bool);
70 : static void vect_print_slp_tree (dump_flags_t, dump_location_t, slp_tree);
71 : static bool vect_slp_can_convert_to_external (const vec<stmt_vec_info> &);
72 :
73 : static object_allocator<_slp_tree> *slp_tree_pool;
74 : static slp_tree slp_first_node;
75 :
76 : void
77 1138792 : vect_slp_init (void)
78 : {
79 1138792 : slp_tree_pool = new object_allocator<_slp_tree> ("SLP nodes");
80 1138792 : }
81 :
82 : void
83 1138792 : vect_slp_fini (void)
84 : {
85 1830668 : while (slp_first_node)
86 691876 : delete slp_first_node;
87 2277584 : delete slp_tree_pool;
88 1138792 : slp_tree_pool = NULL;
89 1138792 : }
90 :
91 : void *
92 8194569 : _slp_tree::operator new (size_t n)
93 : {
94 8194569 : gcc_assert (n == sizeof (_slp_tree));
95 8194569 : return slp_tree_pool->allocate_raw ();
96 : }
97 :
98 : void
99 8194569 : _slp_tree::operator delete (void *node, size_t n)
100 : {
101 8194569 : gcc_assert (n == sizeof (_slp_tree));
102 8194569 : slp_tree_pool->remove_raw (node);
103 8194569 : }
104 :
105 :
106 : /* Initialize a SLP node. */
107 :
108 8194569 : _slp_tree::_slp_tree ()
109 : {
110 8194569 : this->prev_node = NULL;
111 8194569 : if (slp_first_node)
112 7168112 : slp_first_node->prev_node = this;
113 8194569 : this->next_node = slp_first_node;
114 8194569 : slp_first_node = this;
115 8194569 : SLP_TREE_SCALAR_STMTS (this) = vNULL;
116 8194569 : SLP_TREE_SCALAR_OPS (this) = vNULL;
117 8194569 : SLP_TREE_LIVE_LANES (this) = vNULL;
118 8194569 : SLP_TREE_VEC_DEFS (this) = vNULL;
119 8194569 : SLP_TREE_CHILDREN (this) = vNULL;
120 8194569 : SLP_TREE_LOAD_PERMUTATION (this) = vNULL;
121 8194569 : SLP_TREE_LANE_PERMUTATION (this) = vNULL;
122 8194569 : SLP_TREE_DEF_TYPE (this) = vect_uninitialized_def;
123 8194569 : SLP_TREE_CODE (this) = ERROR_MARK;
124 8194569 : SLP_TREE_GS_SCALE (this) = 0;
125 8194569 : SLP_TREE_GS_BASE (this) = NULL_TREE;
126 8194569 : this->ldst_lanes = false;
127 8194569 : this->avoid_stlf_fail = false;
128 8194569 : SLP_TREE_VECTYPE (this) = NULL_TREE;
129 8194569 : SLP_TREE_REPRESENTATIVE (this) = NULL;
130 8194569 : this->cycle_info.id = -1;
131 8194569 : this->cycle_info.reduc_idx = -1;
132 8194569 : SLP_TREE_REF_COUNT (this) = 1;
133 8194569 : this->failed = NULL;
134 8194569 : this->lanes = 0;
135 8194569 : SLP_TREE_TYPE (this) = undef_vec_info_type;
136 8194569 : this->data = NULL;
137 8194569 : this->si = NULL;
138 8194569 : }
139 :
140 : /* Tear down a SLP node. */
141 :
142 8194569 : _slp_tree::~_slp_tree ()
143 : {
144 8194569 : if (this->prev_node)
145 4995305 : this->prev_node->next_node = this->next_node;
146 : else
147 3199264 : slp_first_node = this->next_node;
148 8194569 : if (this->next_node)
149 6190445 : this->next_node->prev_node = this->prev_node;
150 8194569 : SLP_TREE_CHILDREN (this).release ();
151 8194569 : SLP_TREE_SCALAR_STMTS (this).release ();
152 8194569 : SLP_TREE_SCALAR_OPS (this).release ();
153 8194569 : SLP_TREE_LIVE_LANES (this).release ();
154 8194569 : SLP_TREE_VEC_DEFS (this).release ();
155 8194569 : SLP_TREE_LOAD_PERMUTATION (this).release ();
156 8194569 : SLP_TREE_LANE_PERMUTATION (this).release ();
157 8194569 : if (this->failed)
158 2135485 : free (failed);
159 8194569 : if (this->data)
160 1280343 : delete this->data;
161 8194569 : }
162 :
163 : /* Push the single SSA definition in DEF to the vector of vector defs. */
164 :
165 : void
166 537040 : _slp_tree::push_vec_def (gimple *def)
167 : {
168 537040 : if (gphi *phi = dyn_cast <gphi *> (def))
169 60077 : vec_defs.quick_push (gimple_phi_result (phi));
170 : else
171 : {
172 476963 : def_operand_p defop = single_ssa_def_operand (def, SSA_OP_ALL_DEFS);
173 476963 : vec_defs.quick_push (get_def_from_ptr (defop));
174 : }
175 537040 : }
176 :
177 : /* Recursively free the memory allocated for the SLP tree rooted at NODE. */
178 :
179 : void
180 15525630 : vect_free_slp_tree (slp_tree node)
181 : {
182 15525630 : int i;
183 15525630 : slp_tree child;
184 :
185 15525630 : if (--SLP_TREE_REF_COUNT (node) != 0)
186 15525630 : return;
187 :
188 11674966 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
189 4172273 : if (child)
190 3790644 : vect_free_slp_tree (child);
191 :
192 7502693 : delete node;
193 : }
194 :
195 : /* Return a location suitable for dumpings related to the SLP instance. */
196 :
197 : dump_user_location_t
198 3556319 : _slp_instance::location () const
199 : {
200 3556319 : if (!root_stmts.is_empty ())
201 416435 : return root_stmts[0]->stmt;
202 : else
203 3139884 : return SLP_TREE_SCALAR_STMTS (root)[0]->stmt;
204 : }
205 :
206 :
207 : /* Free the memory allocated for the SLP instance. */
208 :
209 : void
210 1631226 : vect_free_slp_instance (slp_instance instance)
211 : {
212 1631226 : vect_free_slp_tree (SLP_INSTANCE_TREE (instance));
213 1631226 : SLP_INSTANCE_LOADS (instance).release ();
214 1631226 : SLP_INSTANCE_ROOT_STMTS (instance).release ();
215 1631226 : SLP_INSTANCE_REMAIN_DEFS (instance).release ();
216 1631226 : instance->subgraph_entries.release ();
217 1631226 : instance->cost_vec.release ();
218 1631226 : free (instance);
219 1631226 : }
220 :
221 :
222 : /* Create a SLP node with NOPS children with CODE, either VEC_PERM_EXPR
223 : for a permute node or else ERROR_MARK. */
224 :
225 : slp_tree
226 97864 : vect_create_new_slp_node (unsigned nops, tree_code code)
227 : {
228 97864 : gcc_assert (code == ERROR_MARK || code == VEC_PERM_EXPR);
229 97864 : slp_tree node = new _slp_tree;
230 97864 : SLP_TREE_SCALAR_STMTS (node) = vNULL;
231 97864 : SLP_TREE_CHILDREN (node).create (nops);
232 97864 : SLP_TREE_DEF_TYPE (node) = vect_internal_def;
233 97864 : SLP_TREE_CODE (node) = code;
234 97864 : return node;
235 : }
236 :
237 : /* Create a SLP node inplace at NODE for SCALAR_STMTS and NOPS children. */
238 :
239 : static slp_tree
240 3981291 : vect_create_new_slp_node (slp_tree node,
241 : vec<stmt_vec_info> scalar_stmts, unsigned nops)
242 : {
243 3981291 : SLP_TREE_SCALAR_STMTS (node) = scalar_stmts;
244 3981291 : SLP_TREE_CHILDREN (node).create (nops);
245 3981291 : SLP_TREE_DEF_TYPE (node) = vect_internal_def;
246 3981291 : SLP_TREE_REPRESENTATIVE (node) = scalar_stmts[0];
247 3981291 : SLP_TREE_LANES (node) = scalar_stmts.length ();
248 3981291 : return node;
249 : }
250 :
251 : /* Create an SLP node for SCALAR_STMTS and NOPS children. */
252 :
253 : static slp_tree
254 8398 : vect_create_new_slp_node (vec<stmt_vec_info> scalar_stmts, unsigned nops)
255 : {
256 8398 : return vect_create_new_slp_node (new _slp_tree, scalar_stmts, nops);
257 : }
258 :
259 : /* Create a vect_external_def SLP node inplace at NODE for scalar
260 : operands OPS. */
261 :
262 : static slp_tree
263 1967666 : vect_create_new_slp_node (slp_tree node, vec<tree> ops)
264 : {
265 1967666 : SLP_TREE_SCALAR_OPS (node) = ops;
266 1967666 : SLP_TREE_DEF_TYPE (node) = vect_external_def;
267 0 : SLP_TREE_LANES (node) = ops.length ();
268 1967666 : return node;
269 : }
270 :
271 : /* Create a vect_external_def SLP node for scalar operands OPS. */
272 :
273 : static slp_tree
274 1967666 : vect_create_new_slp_node (vec<tree> ops)
275 : {
276 1967666 : return vect_create_new_slp_node (new _slp_tree, ops);
277 : }
278 :
279 :
280 : /* This structure is used in creation of an SLP tree. Each instance
281 : corresponds to the same operand in a group of scalar stmts in an SLP
282 : node. */
283 : typedef struct _slp_oprnd_info
284 : {
285 : /* Def-stmts for the operands. */
286 : vec<stmt_vec_info> def_stmts;
287 : /* Operands. */
288 : vec<tree> ops;
289 : /* Information about the first statement, its vector def-type, type, the
290 : operand itself in case it's constant, and an indication if it's a pattern
291 : stmt and gather/scatter info. */
292 : tree first_op_type;
293 : enum vect_def_type first_dt;
294 : bool any_pattern;
295 : bool first_gs_p;
296 : gather_scatter_info first_gs_info;
297 : } *slp_oprnd_info;
298 :
299 :
300 : /* Allocate operands info for NOPS operands, and GROUP_SIZE def-stmts for each
301 : operand. */
302 : static vec<slp_oprnd_info>
303 3665743 : vect_create_oprnd_info (int nops, int group_size)
304 : {
305 3665743 : int i;
306 3665743 : slp_oprnd_info oprnd_info;
307 3665743 : vec<slp_oprnd_info> oprnds_info;
308 :
309 3665743 : oprnds_info.create (nops);
310 13118391 : for (i = 0; i < nops; i++)
311 : {
312 5786905 : oprnd_info = XNEW (struct _slp_oprnd_info);
313 5786905 : oprnd_info->def_stmts.create (group_size);
314 5786905 : oprnd_info->ops.create (group_size);
315 5786905 : oprnd_info->first_dt = vect_uninitialized_def;
316 5786905 : oprnd_info->first_op_type = NULL_TREE;
317 5786905 : oprnd_info->any_pattern = false;
318 5786905 : oprnd_info->first_gs_p = false;
319 5786905 : oprnds_info.quick_push (oprnd_info);
320 : }
321 :
322 3665743 : return oprnds_info;
323 : }
324 :
325 :
326 : /* Free operands info. */
327 :
328 : static void
329 3543126 : vect_free_oprnd_info (vec<slp_oprnd_info> &oprnds_info)
330 : {
331 3543126 : int i;
332 3543126 : slp_oprnd_info oprnd_info;
333 :
334 9198340 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
335 : {
336 5655214 : oprnd_info->def_stmts.release ();
337 5655214 : oprnd_info->ops.release ();
338 5655214 : XDELETE (oprnd_info);
339 : }
340 :
341 3543126 : oprnds_info.release ();
342 3543126 : }
343 :
344 : /* Return the execution frequency of NODE (so that a higher value indicates
345 : a "more important" node when optimizing for speed). */
346 :
347 : static sreal
348 3672517 : vect_slp_node_weight (vec_info *vinfo, slp_tree node)
349 : {
350 3672517 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
351 3672517 : basic_block bb;
352 : /* ??? This is imprecise, VEC_PERM nodes do not have a representative
353 : but are laid out close to their children. */
354 3672517 : if (!stmt_info)
355 142973 : bb = vinfo->bbs[0];
356 : else
357 4000459 : bb = gimple_bb (vect_orig_stmt (stmt_info)->stmt);
358 3672517 : return bb->count.to_sreal_scale (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count);
359 : }
360 :
361 : /* Return true if STMTS contains a pattern statement. */
362 :
363 : static bool
364 29778 : vect_contains_pattern_stmt_p (vec<stmt_vec_info> stmts)
365 : {
366 29778 : stmt_vec_info stmt_info;
367 29778 : unsigned int i;
368 85430 : FOR_EACH_VEC_ELT (stmts, i, stmt_info)
369 63085 : if (stmt_info && is_pattern_stmt_p (stmt_info))
370 : return true;
371 : return false;
372 : }
373 :
374 : /* Return true when all lanes in the external or constant NODE have
375 : the same value. */
376 :
377 : static bool
378 645414 : vect_slp_tree_uniform_p (slp_tree node)
379 : {
380 645414 : gcc_assert (SLP_TREE_DEF_TYPE (node) == vect_constant_def
381 : || SLP_TREE_DEF_TYPE (node) == vect_external_def);
382 :
383 : /* Pre-existing vectors. */
384 645414 : if (SLP_TREE_SCALAR_OPS (node).is_empty ())
385 : return false;
386 :
387 : unsigned i;
388 : tree op, first = NULL_TREE;
389 1480857 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
390 1322441 : if (!first)
391 : first = op;
392 677027 : else if (!operand_equal_p (first, op, 0))
393 : return false;
394 :
395 : return true;
396 : }
397 :
398 : /* Find the place of the data-ref in STMT_INFO in the interleaving chain
399 : that starts from FIRST_STMT_INFO. Return -1 if the data-ref is not a part
400 : of the chain. */
401 :
402 : int
403 750963 : vect_get_place_in_interleaving_chain (stmt_vec_info stmt_info,
404 : stmt_vec_info first_stmt_info)
405 : {
406 750963 : stmt_vec_info next_stmt_info = first_stmt_info;
407 750963 : int result = 0;
408 :
409 750963 : if (first_stmt_info != DR_GROUP_FIRST_ELEMENT (stmt_info))
410 : return -1;
411 :
412 2125562 : do
413 : {
414 2125562 : if (next_stmt_info == stmt_info)
415 : return result;
416 1374599 : next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
417 1374599 : if (next_stmt_info)
418 1374599 : result += DR_GROUP_GAP (next_stmt_info);
419 : }
420 1374599 : while (next_stmt_info);
421 :
422 : return -1;
423 : }
424 :
425 : /* Check whether it is possible to load COUNT elements of type ELT_TYPE
426 : using the method implemented by duplicate_and_interleave. Return true
427 : if so, returning the number of intermediate vectors in *NVECTORS_OUT
428 : (if nonnull) and the type of each intermediate vector in *VECTOR_TYPE_OUT
429 : (if nonnull). */
430 :
431 : bool
432 0 : can_duplicate_and_interleave_p (vec_info *vinfo, unsigned int count,
433 : tree elt_type, unsigned int *nvectors_out,
434 : tree *vector_type_out,
435 : tree *permutes)
436 : {
437 0 : tree base_vector_type = get_vectype_for_scalar_type (vinfo, elt_type, count);
438 0 : if (!base_vector_type || !VECTOR_MODE_P (TYPE_MODE (base_vector_type)))
439 : return false;
440 :
441 0 : machine_mode base_vector_mode = TYPE_MODE (base_vector_type);
442 0 : poly_int64 elt_bytes = count * GET_MODE_UNIT_SIZE (base_vector_mode);
443 0 : unsigned int nvectors = 1;
444 0 : for (;;)
445 : {
446 0 : scalar_int_mode int_mode;
447 0 : poly_int64 elt_bits = elt_bytes * BITS_PER_UNIT;
448 0 : if (int_mode_for_size (elt_bits, 1).exists (&int_mode))
449 : {
450 : /* Get the natural vector type for this SLP group size. */
451 0 : tree int_type = build_nonstandard_integer_type
452 0 : (GET_MODE_BITSIZE (int_mode), 1);
453 0 : tree vector_type
454 0 : = get_vectype_for_scalar_type (vinfo, int_type, count);
455 0 : poly_int64 half_nelts;
456 0 : if (vector_type
457 0 : && VECTOR_MODE_P (TYPE_MODE (vector_type))
458 0 : && known_eq (GET_MODE_SIZE (TYPE_MODE (vector_type)),
459 : GET_MODE_SIZE (base_vector_mode))
460 0 : && multiple_p (GET_MODE_NUNITS (TYPE_MODE (vector_type)),
461 : 2, &half_nelts))
462 : {
463 : /* Try fusing consecutive sequences of COUNT / NVECTORS elements
464 : together into elements of type INT_TYPE and using the result
465 : to build NVECTORS vectors. */
466 0 : poly_uint64 nelts = GET_MODE_NUNITS (TYPE_MODE (vector_type));
467 0 : vec_perm_builder sel1 (nelts, 2, 3);
468 0 : vec_perm_builder sel2 (nelts, 2, 3);
469 :
470 0 : for (unsigned int i = 0; i < 3; ++i)
471 : {
472 0 : sel1.quick_push (i);
473 0 : sel1.quick_push (i + nelts);
474 0 : sel2.quick_push (half_nelts + i);
475 0 : sel2.quick_push (half_nelts + i + nelts);
476 : }
477 0 : vec_perm_indices indices1 (sel1, 2, nelts);
478 0 : vec_perm_indices indices2 (sel2, 2, nelts);
479 0 : machine_mode vmode = TYPE_MODE (vector_type);
480 0 : if (can_vec_perm_const_p (vmode, vmode, indices1)
481 0 : && can_vec_perm_const_p (vmode, vmode, indices2))
482 : {
483 0 : if (nvectors_out)
484 0 : *nvectors_out = nvectors;
485 0 : if (vector_type_out)
486 0 : *vector_type_out = vector_type;
487 0 : if (permutes)
488 : {
489 0 : permutes[0] = vect_gen_perm_mask_checked (vector_type,
490 : indices1);
491 0 : permutes[1] = vect_gen_perm_mask_checked (vector_type,
492 : indices2);
493 : }
494 0 : return true;
495 : }
496 0 : }
497 : }
498 0 : if (!multiple_p (elt_bytes, 2, &elt_bytes))
499 : return false;
500 0 : nvectors *= 2;
501 : /* We need to be able to fuse COUNT / NVECTORS elements together. */
502 0 : if (!multiple_p (count, nvectors))
503 : return false;
504 : }
505 : }
506 :
507 : /* Return true if DTA and DTB match. */
508 :
509 : static bool
510 17747838 : vect_def_types_match (enum vect_def_type dta, enum vect_def_type dtb)
511 : {
512 17747838 : return (dta == dtb
513 416861 : || ((dta == vect_external_def || dta == vect_constant_def)
514 266790 : && (dtb == vect_external_def || dtb == vect_constant_def)));
515 : }
516 :
517 : #define GATHER_SCATTER_OFFSET (-3)
518 :
519 : /* For most SLP statements, there is a one-to-one mapping between
520 : gimple arguments and child nodes. If that is not true for STMT,
521 : return an array that contains:
522 :
523 : - the number of child nodes, followed by
524 : - for each child node, the index of the argument associated with that node.
525 : The special index -1 is the first operand of an embedded comparison and
526 : the special index -2 is the second operand of an embedded comparison.
527 : The special index -3 is the offset of a gather as analyzed by
528 : vect_check_gather_scatter.
529 :
530 : SWAP is as for vect_get_and_check_slp_defs. */
531 :
532 : static const int *
533 25857378 : vect_get_operand_map (const gimple *stmt, bool gather_scatter_p,
534 : unsigned char swap)
535 : {
536 25857378 : static const int no_arg_map[] = { 0 };
537 25857378 : static const int arg0_map[] = { 1, 0 };
538 25857378 : static const int arg2_map[] = { 1, 2 };
539 25857378 : static const int arg2_arg3_map[] = { 2, 2, 3 };
540 25857378 : static const int arg2_arg4_map[] = { 2, 2, 4 };
541 25857378 : static const int arg2_arg5_arg6_map[] = { 3, 2, 5, 6 };
542 25857378 : static const int arg2_arg4_arg5_map[] = { 3, 2, 4, 5 };
543 25857378 : static const int arg3_arg2_map[] = { 2, 3, 2 };
544 25857378 : static const int op00_map[] = { 1, -1 };
545 25857378 : static const int op1_op0_map[] = { 2, 1, 0 };
546 25857378 : static const int off_map[] = { 1, GATHER_SCATTER_OFFSET };
547 25857378 : static const int off_op0_map[] = { 2, GATHER_SCATTER_OFFSET, 0 };
548 25857378 : static const int off_arg2_arg3_map[] = { 3, GATHER_SCATTER_OFFSET, 2, 3 };
549 25857378 : static const int off_arg3_arg2_map[] = { 3, GATHER_SCATTER_OFFSET, 3, 2 };
550 25857378 : static const int mask_call_maps[6][7] = {
551 : { 1, 1, },
552 : { 2, 1, 2, },
553 : { 3, 1, 2, 3, },
554 : { 4, 1, 2, 3, 4, },
555 : { 5, 1, 2, 3, 4, 5, },
556 : { 6, 1, 2, 3, 4, 5, 6 },
557 : };
558 :
559 25857378 : gcc_checking_assert (!swap
560 : || !is_gimple_assign (stmt)
561 : || TREE_CODE_CLASS
562 : (gimple_assign_rhs_code (stmt)) == tcc_comparison
563 : || commutative_tree_code
564 : (gimple_assign_rhs_code (stmt)));
565 :
566 25857378 : if (auto assign = dyn_cast<const gassign *> (stmt))
567 : {
568 24307908 : tree_code code = gimple_assign_rhs_code (assign);
569 24307908 : if (code == COND_EXPR
570 24307908 : && COMPARISON_CLASS_P (gimple_assign_rhs1 (assign)))
571 0 : gcc_unreachable ();
572 24307908 : else if ((TREE_CODE_CLASS (code) == tcc_comparison
573 22752838 : || commutative_tree_code (code))
574 33404136 : && swap)
575 : return op1_op0_map;
576 24264035 : else if (code == VIEW_CONVERT_EXPR)
577 : return op00_map;
578 24254748 : else if (gather_scatter_p)
579 46824 : return (TREE_CODE (gimple_assign_lhs (assign)) != SSA_NAME
580 46824 : ? off_op0_map : off_map);
581 : }
582 1549470 : else if (auto call = dyn_cast<const gcall *> (stmt))
583 : {
584 168120 : if (gimple_call_internal_p (call))
585 92195 : switch (gimple_call_internal_fn (call))
586 : {
587 13928 : case IFN_MASK_LOAD:
588 13928 : return gather_scatter_p ? off_arg2_arg3_map : arg2_arg3_map;
589 :
590 0 : case IFN_GATHER_LOAD:
591 0 : return arg2_map;
592 :
593 0 : case IFN_MASK_GATHER_LOAD:
594 0 : case IFN_MASK_LEN_GATHER_LOAD:
595 0 : return arg2_arg5_arg6_map;
596 :
597 0 : case IFN_SCATTER_STORE:
598 0 : return arg2_arg4_map;
599 :
600 0 : case IFN_MASK_SCATTER_STORE:
601 0 : case IFN_MASK_LEN_SCATTER_STORE:
602 0 : return arg2_arg4_arg5_map;
603 :
604 9200 : case IFN_MASK_STORE:
605 9200 : return gather_scatter_p ? off_arg3_arg2_map : arg3_arg2_map;
606 :
607 996 : case IFN_MASK_CALL:
608 996 : {
609 996 : unsigned nargs = gimple_call_num_args (call);
610 996 : if (nargs >= 2 && nargs <= 7)
611 996 : return mask_call_maps[nargs-2];
612 : else
613 : return nullptr;
614 : }
615 :
616 278 : case IFN_CLZ:
617 278 : case IFN_CTZ:
618 278 : return arg0_map;
619 :
620 7302 : case IFN_GOMP_SIMD_LANE:
621 7302 : return no_arg_map;
622 :
623 : default:
624 : break;
625 : }
626 : }
627 : return nullptr;
628 : }
629 :
630 : static const int *
631 25838814 : vect_get_operand_map (const stmt_vec_info stmt, unsigned char swap = 0)
632 : {
633 0 : return vect_get_operand_map (stmt->stmt, STMT_VINFO_GATHER_SCATTER_P (stmt),
634 0 : swap);
635 : }
636 :
637 : /* Return the SLP node child index for operand OP of STMT. */
638 :
639 : int
640 1398704 : vect_slp_child_index_for_operand (const stmt_vec_info stmt, int op)
641 : {
642 1398704 : const int *opmap = vect_get_operand_map (stmt);
643 1398704 : if (!opmap)
644 : return op;
645 20313 : for (int i = 1; i < 1 + opmap[0]; ++i)
646 20313 : if (opmap[i] == op)
647 11556 : return i - 1;
648 0 : gcc_unreachable ();
649 : }
650 :
651 : /* Helper class for mapping of GIMPLE operands to SLP children. */
652 : /* ??? Add vect_slp_child_index_for_operand here and amend opmaps
653 : with the full reverse mapping and indicating the position of the
654 : first commutative operand index, eliding the swap_p argument from
655 : vect_get_operand_map. Adjust all consumers. */
656 :
657 : struct slp_oprnds {
658 : slp_oprnds (stmt_vec_info);
659 : tree get_op_for_slp_child (stmt_vec_info, unsigned);
660 : const int *opmap;
661 : const unsigned int num_slp_children;
662 : };
663 :
664 4622469 : slp_oprnds::slp_oprnds (stmt_vec_info stmt_info)
665 4622469 : : opmap (vect_get_operand_map (stmt_info)),
666 4622469 : num_slp_children (opmap ? opmap[0] : gimple_num_args (stmt_info->stmt))
667 : {
668 4622469 : }
669 :
670 : /* For SLP child number N get the corresponding tree operand from GIMPLE
671 : statement described by STMT_INFO. */
672 :
673 : tree
674 5158092 : slp_oprnds::get_op_for_slp_child (stmt_vec_info stmt_info, unsigned n)
675 : {
676 5158092 : gcc_assert (n < num_slp_children);
677 5158092 : int opno = opmap ? opmap[n + 1] : (int) n;
678 5158092 : if (opno == GATHER_SCATTER_OFFSET)
679 0 : gcc_unreachable (); // TODO
680 5158092 : else if (opno < 0)
681 2530 : return TREE_OPERAND (gimple_arg (stmt_info->stmt, 0), -1 - opno);
682 : else
683 5155562 : return gimple_arg (stmt_info->stmt, opno);
684 : }
685 :
686 : /* Get the defs for the rhs of STMT (collect them in OPRNDS_INFO), check that
687 : they are of a valid type and that they match the defs of the first stmt of
688 : the SLP group (stored in OPRNDS_INFO). This function tries to match stmts
689 : by swapping operands of STMTS[STMT_NUM] when possible. Non-zero SWAP
690 : indicates swap is required for cond_expr stmts. Specifically, SWAP
691 : is 1 if STMT is cond and operands of comparison need to be swapped;
692 : SWAP is 2 if STMT is cond and code of comparison needs to be inverted.
693 :
694 : If there was a fatal error return -1; if the error could be corrected by
695 : swapping operands of father node of this one, return 1; if everything is
696 : ok return 0. */
697 : static int
698 13712770 : vect_get_and_check_slp_defs (vec_info *vinfo, tree vectype, unsigned char swap,
699 : bool *skip_args,
700 : vec<stmt_vec_info> stmts, unsigned stmt_num,
701 : vec<slp_oprnd_info> *oprnds_info,
702 : bool soft_fail)
703 : {
704 13712770 : stmt_vec_info stmt_info = stmts[stmt_num];
705 13712770 : tree oprnd;
706 13712770 : unsigned int i, number_of_oprnds;
707 13712770 : enum vect_def_type dt = vect_uninitialized_def;
708 13712770 : slp_oprnd_info oprnd_info;
709 13712770 : gather_scatter_info gs_info;
710 13712770 : unsigned int gs_op = -1u;
711 13712770 : unsigned int commutative_op = -1U;
712 13712770 : bool first = stmt_num == 0;
713 :
714 13712770 : if (!stmt_info)
715 : {
716 0 : for (auto oi : *oprnds_info)
717 : {
718 0 : oi->def_stmts.quick_push (NULL);
719 0 : oi->ops.quick_push (NULL_TREE);
720 : }
721 : return 0;
722 : }
723 :
724 13712770 : if (!is_a<gcall *> (stmt_info->stmt)
725 : && !is_a<gassign *> (stmt_info->stmt)
726 : && !is_a<gphi *> (stmt_info->stmt))
727 : return -1;
728 :
729 13712770 : number_of_oprnds = gimple_num_args (stmt_info->stmt);
730 13712770 : const int *map = vect_get_operand_map (stmt_info, swap);
731 13712770 : if (map)
732 80855 : number_of_oprnds = *map++;
733 13712770 : if (gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt))
734 : {
735 53327 : if (gimple_call_internal_p (stmt))
736 : {
737 32950 : internal_fn ifn = gimple_call_internal_fn (stmt);
738 32950 : commutative_op = first_commutative_argument (ifn);
739 32950 : if (internal_gather_scatter_fn_p (ifn))
740 : {
741 0 : vect_describe_gather_scatter_call
742 0 : (stmt_info,
743 0 : first ? &(*oprnds_info)[0]->first_gs_info : &gs_info);
744 0 : if (first)
745 0 : (*oprnds_info)[0]->first_gs_p = true;
746 : gs_op = 0;
747 : }
748 : }
749 : }
750 13659443 : else if (gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt))
751 : {
752 16401313 : if (commutative_tree_code (gimple_assign_rhs_code (stmt)))
753 8527409 : commutative_op = 0;
754 : }
755 :
756 13712770 : bool swapped = (swap != 0);
757 13712770 : bool backedge = false;
758 13712770 : enum vect_def_type *dts = XALLOCAVEC (enum vect_def_type, number_of_oprnds);
759 37451963 : for (i = 0; i < number_of_oprnds; i++)
760 : {
761 23740507 : oprnd_info = (*oprnds_info)[i];
762 23740507 : int opno = map ? map[i] : int (i);
763 23740507 : if (opno == GATHER_SCATTER_OFFSET)
764 : {
765 23847 : gcc_assert (STMT_VINFO_GATHER_SCATTER_P (stmt_info));
766 23847 : if (!is_a <loop_vec_info> (vinfo)
767 23847 : || !vect_check_gather_scatter (stmt_info, vectype,
768 : as_a <loop_vec_info> (vinfo),
769 : first ? &oprnd_info->first_gs_info
770 : : &gs_info))
771 1314 : return -1;
772 :
773 23847 : if (first)
774 : {
775 23573 : oprnd_info->first_gs_p = true;
776 23573 : oprnd = oprnd_info->first_gs_info.offset;
777 : }
778 : else
779 : {
780 274 : gs_op = i;
781 274 : oprnd = gs_info.offset;
782 : }
783 : }
784 23716660 : else if (opno < 0)
785 3167 : oprnd = TREE_OPERAND (gimple_arg (stmt_info->stmt, 0), -1 - opno);
786 : else
787 : {
788 23713493 : oprnd = gimple_arg (stmt_info->stmt, opno);
789 23713493 : if (gphi *stmt = dyn_cast <gphi *> (stmt_info->stmt))
790 : {
791 1271593 : edge e = gimple_phi_arg_edge (stmt, opno);
792 2543186 : backedge = (is_a <bb_vec_info> (vinfo)
793 1967594 : ? e->flags & EDGE_DFS_BACK
794 696001 : : dominated_by_p (CDI_DOMINATORS, e->src,
795 696001 : gimple_bb (stmt_info->stmt)));
796 : }
797 : }
798 :
799 23740507 : stmt_vec_info def_stmt_info;
800 23740507 : if (!vect_is_simple_use (oprnd, vinfo, &dts[i], &def_stmt_info))
801 : {
802 1038 : if (dump_enabled_p ())
803 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
804 : "Build SLP failed: can't analyze def for %T\n",
805 : oprnd);
806 :
807 : return -1;
808 : }
809 :
810 23739469 : if (skip_args[i])
811 : {
812 559077 : oprnd_info->def_stmts.quick_push (NULL);
813 559077 : oprnd_info->ops.quick_push (NULL_TREE);
814 559077 : oprnd_info->first_dt = vect_uninitialized_def;
815 559077 : continue;
816 : }
817 :
818 23180392 : oprnd_info->def_stmts.quick_push (def_stmt_info);
819 23180392 : oprnd_info->ops.quick_push (oprnd);
820 :
821 23180392 : if (def_stmt_info
822 23180392 : && is_pattern_stmt_p (def_stmt_info))
823 : {
824 450593 : if (STMT_VINFO_RELATED_STMT (vect_orig_stmt (def_stmt_info))
825 : != def_stmt_info)
826 320191 : oprnd_info->any_pattern = true;
827 : else
828 : /* If we promote this to external use the original stmt def. */
829 130402 : oprnd_info->ops.last ()
830 260804 : = gimple_get_lhs (vect_orig_stmt (def_stmt_info)->stmt);
831 : }
832 :
833 : /* If there's a extern def on a backedge make sure we can
834 : code-generate at the region start.
835 : ??? This is another case that could be fixed by adjusting
836 : how we split the function but at the moment we'd have conflicting
837 : goals there. */
838 23180392 : if (backedge
839 174424 : && dts[i] == vect_external_def
840 297 : && is_a <bb_vec_info> (vinfo)
841 297 : && TREE_CODE (oprnd) == SSA_NAME
842 276 : && !SSA_NAME_IS_DEFAULT_DEF (oprnd)
843 23180668 : && !dominated_by_p (CDI_DOMINATORS, vinfo->bbs[0],
844 276 : gimple_bb (SSA_NAME_DEF_STMT (oprnd))))
845 : {
846 276 : if (dump_enabled_p ())
847 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
848 : "Build SLP failed: extern def %T only defined "
849 : "on backedge\n", oprnd);
850 : return -1;
851 : }
852 :
853 23180116 : if (first)
854 : {
855 5267020 : tree type = TREE_TYPE (oprnd);
856 5267020 : dt = dts[i];
857 :
858 : /* For the swapping logic below force vect_reduction_def
859 : for the reduction op in a SLP reduction group. */
860 5267020 : if (!STMT_VINFO_DATA_REF (stmt_info)
861 3955130 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
862 5312 : && (int)i == STMT_VINFO_REDUC_IDX (stmt_info)
863 5269633 : && def_stmt_info)
864 2613 : dts[i] = dt = vect_reduction_def;
865 :
866 : /* Check the types of the definition. */
867 5267020 : switch (dt)
868 : {
869 5267020 : case vect_external_def:
870 5267020 : case vect_constant_def:
871 5267020 : case vect_internal_def:
872 5267020 : case vect_reduction_def:
873 5267020 : case vect_double_reduction_def:
874 5267020 : case vect_induction_def:
875 5267020 : case vect_nested_cycle:
876 5267020 : case vect_first_order_recurrence:
877 5267020 : break;
878 :
879 0 : default:
880 : /* FORNOW: Not supported. */
881 0 : if (dump_enabled_p ())
882 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
883 : "Build SLP failed: illegal type of def %T\n",
884 : oprnd);
885 : return -1;
886 : }
887 :
888 5267020 : oprnd_info->first_dt = dt;
889 5267020 : oprnd_info->first_op_type = type;
890 : }
891 : }
892 13711456 : if (first)
893 : return 0;
894 :
895 : /* Now match the operand definition types to that of the first stmt. */
896 27615836 : for (i = 0; i < number_of_oprnds;)
897 : {
898 17904061 : if (skip_args[i])
899 : {
900 44940 : ++i;
901 44940 : continue;
902 : }
903 :
904 17859121 : oprnd_info = (*oprnds_info)[i];
905 17859121 : dt = dts[i];
906 17859121 : stmt_vec_info def_stmt_info = oprnd_info->def_stmts[stmt_num];
907 17859121 : oprnd = oprnd_info->ops[stmt_num];
908 17859121 : tree type = TREE_TYPE (oprnd);
909 :
910 17859121 : if (!types_compatible_p (oprnd_info->first_op_type, type))
911 : {
912 118533 : if (dump_enabled_p ())
913 93 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
914 : "Build SLP failed: different operand types\n");
915 : return 1;
916 : }
917 :
918 17740588 : if ((gs_op == i) != oprnd_info->first_gs_p)
919 : {
920 0 : if (dump_enabled_p ())
921 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
922 : "Build SLP failed: mixed gather and non-gather\n");
923 : return 1;
924 : }
925 17740588 : else if (gs_op == i)
926 : {
927 242 : if (!operand_equal_p (oprnd_info->first_gs_info.base,
928 242 : gs_info.base))
929 : {
930 16 : if (dump_enabled_p ())
931 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
932 : "Build SLP failed: different gather base\n");
933 : return 1;
934 : }
935 226 : if (oprnd_info->first_gs_info.scale != gs_info.scale)
936 : {
937 8 : if (dump_enabled_p ())
938 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
939 : "Build SLP failed: different gather scale\n");
940 : return 1;
941 : }
942 : }
943 :
944 : /* Not first stmt of the group, check that the def-stmt/s match
945 : the def-stmt/s of the first stmt. Allow different definition
946 : types for reduction chains: the first stmt must be a
947 : vect_reduction_def (a phi node), and the rest
948 : end in the reduction chain. */
949 17740564 : if ((!vect_def_types_match (oprnd_info->first_dt, dt)
950 334406 : && !(oprnd_info->first_dt == vect_reduction_def
951 5019 : && !STMT_VINFO_DATA_REF (stmt_info)
952 5019 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
953 4993 : && def_stmt_info
954 4991 : && !STMT_VINFO_DATA_REF (def_stmt_info)
955 4991 : && (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
956 : == REDUC_GROUP_FIRST_ELEMENT (stmt_info))))
957 17411149 : || (!STMT_VINFO_DATA_REF (stmt_info)
958 15629617 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
959 10429 : && ((!def_stmt_info
960 10193 : || STMT_VINFO_DATA_REF (def_stmt_info)
961 18668 : || (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
962 : != REDUC_GROUP_FIRST_ELEMENT (stmt_info)))
963 10429 : != (oprnd_info->first_dt != vect_reduction_def))))
964 : {
965 : /* Try swapping operands if we got a mismatch. For BB
966 : vectorization only in case it will clearly improve things. */
967 332034 : if (i == commutative_op && !swapped
968 329415 : && (!is_a <bb_vec_info> (vinfo)
969 5623 : || (!vect_def_types_match ((*oprnds_info)[i+1]->first_dt,
970 5623 : dts[i+1])
971 1418 : && (vect_def_types_match (oprnd_info->first_dt, dts[i+1])
972 : || vect_def_types_match
973 233 : ((*oprnds_info)[i+1]->first_dt, dts[i])))))
974 : {
975 2619 : if (dump_enabled_p ())
976 198 : dump_printf_loc (MSG_NOTE, vect_location,
977 : "trying swapped operands\n");
978 2619 : std::swap (dts[i], dts[i+1]);
979 2619 : std::swap ((*oprnds_info)[i]->def_stmts[stmt_num],
980 2619 : (*oprnds_info)[i+1]->def_stmts[stmt_num]);
981 2619 : std::swap ((*oprnds_info)[i]->ops[stmt_num],
982 2619 : (*oprnds_info)[i+1]->ops[stmt_num]);
983 : /* After swapping some operands we lost track whether an
984 : operand has any pattern defs so be conservative here. */
985 2619 : if ((*oprnds_info)[i]->any_pattern
986 2619 : || (*oprnds_info)[i+1]->any_pattern)
987 36 : (*oprnds_info)[i]->any_pattern
988 18 : = (*oprnds_info)[i+1]->any_pattern = true;
989 2619 : swapped = true;
990 2619 : continue;
991 : }
992 :
993 326796 : if (is_a <bb_vec_info> (vinfo)
994 309447 : && !oprnd_info->any_pattern
995 309179 : && number_of_oprnds > 1
996 635975 : && !soft_fail)
997 : {
998 : /* Now for commutative ops we should see whether we can
999 : make the other operand matching. */
1000 106098 : if (dump_enabled_p ())
1001 338 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1002 : "treating operand as external\n");
1003 : oprnd_info->first_dt = dt = vect_external_def;
1004 : }
1005 : else
1006 : {
1007 220698 : if (dump_enabled_p ())
1008 410 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1009 : "Build SLP failed: different types\n");
1010 : return 1;
1011 : }
1012 : }
1013 :
1014 : /* Make sure to demote the overall operand to external. */
1015 17411149 : if (dt == vect_external_def)
1016 422807 : oprnd_info->first_dt = vect_external_def;
1017 : /* For a SLP reduction chain we want to duplicate the reduction to
1018 : each of the chain members. That gets us a sane SLP graph (still
1019 : the stmts are not 100% correct wrt the initial values). */
1020 17094440 : else if ((dt == vect_internal_def
1021 17094440 : || dt == vect_reduction_def)
1022 15833770 : && oprnd_info->first_dt == vect_reduction_def
1023 101523 : && !STMT_VINFO_DATA_REF (stmt_info)
1024 101523 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
1025 4991 : && !STMT_VINFO_DATA_REF (def_stmt_info)
1026 17099431 : && (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
1027 : == REDUC_GROUP_FIRST_ELEMENT (stmt_info)))
1028 : {
1029 4991 : oprnd_info->def_stmts[stmt_num] = oprnd_info->def_stmts[0];
1030 4991 : oprnd_info->ops[stmt_num] = oprnd_info->ops[0];
1031 : }
1032 :
1033 17517247 : ++i;
1034 : }
1035 :
1036 : /* Swap operands. */
1037 9711775 : if (swapped)
1038 : {
1039 43610 : if (dump_enabled_p ())
1040 502 : dump_printf_loc (MSG_NOTE, vect_location,
1041 : "swapped operands to match def types in %G",
1042 : stmt_info->stmt);
1043 : }
1044 :
1045 : return 0;
1046 : }
1047 :
1048 : /* Return true if call statements CALL1 and CALL2 are similar enough
1049 : to be combined into the same SLP group. */
1050 :
1051 : bool
1052 65083 : compatible_calls_p (gcall *call1, gcall *call2, bool allow_two_operators)
1053 : {
1054 65083 : unsigned int nargs = gimple_call_num_args (call1);
1055 65083 : if (nargs != gimple_call_num_args (call2))
1056 : return false;
1057 :
1058 47996 : auto cfn1 = gimple_call_combined_fn (call1);
1059 47996 : auto cfn2 = gimple_call_combined_fn (call2);
1060 47996 : if (cfn1 != cfn2
1061 39 : && (!allow_two_operators
1062 39 : || !((cfn1 == CFN_FMA || cfn1 == CFN_FMS)
1063 2 : && (cfn2 == CFN_FMA || cfn2 == CFN_FMS))))
1064 : return false;
1065 :
1066 47959 : if (gimple_call_internal_p (call1))
1067 : {
1068 7286 : if (!types_compatible_p (TREE_TYPE (gimple_call_lhs (call1)),
1069 7286 : TREE_TYPE (gimple_call_lhs (call2))))
1070 : return false;
1071 15010 : for (unsigned int i = 0; i < nargs; ++i)
1072 7724 : if (!types_compatible_p (TREE_TYPE (gimple_call_arg (call1, i)),
1073 7724 : TREE_TYPE (gimple_call_arg (call2, i))))
1074 : return false;
1075 : }
1076 : else
1077 : {
1078 40673 : if (!operand_equal_p (gimple_call_fn (call1),
1079 40673 : gimple_call_fn (call2), 0))
1080 : return false;
1081 :
1082 33834 : if (gimple_call_fntype (call1) != gimple_call_fntype (call2))
1083 : return false;
1084 : }
1085 :
1086 : /* Check that any unvectorized arguments are equal. */
1087 18564 : if (const int *map = vect_get_operand_map (call1, false, false))
1088 : {
1089 15 : unsigned int nkept = *map++;
1090 15 : unsigned int mapi = 0;
1091 57 : for (unsigned int i = 0; i < nargs; ++i)
1092 42 : if (mapi < nkept && map[mapi] == int (i))
1093 27 : mapi += 1;
1094 15 : else if (!operand_equal_p (gimple_call_arg (call1, i),
1095 15 : gimple_call_arg (call2, i)))
1096 : return false;
1097 : }
1098 :
1099 : return true;
1100 : }
1101 :
1102 : /* Verify if the scalar stmts STMTS are isomorphic, require data
1103 : permutation or are of unsupported types of operation.
1104 : Return false if at least one (or all in case of BB vectorization)
1105 : stmt is unvectorizable or the comparison could not be carried out.
1106 : Return true if they all (or at least one in case of BB vectorization)
1107 : are and indicate in MATCHES[] which stmts are not isomorphic to the
1108 : stmt at *START_I.
1109 :
1110 : This function is designed to be invoked repeatedly on the same
1111 : set of STMTS with increasing *START_I.
1112 :
1113 : When *START_I is zero and the function returns true MATCHES[] will be
1114 : initialized with, in case of BB vectorization, unvectorizable stmts
1115 : marked with -2, stmts isomorphic to *START_I as *START_I and
1116 : other vectorizable stmts with -1 (not isomorphic to *START_I).
1117 : *START_I, when initially zero, is updated to the first vectorizable
1118 : statement, so MATCHES[] can have a prefix with entries valued -2.
1119 :
1120 : When *START_I is not zero MATCHES[] is expected to be pre-initialized
1121 : by a former call with *START_I zero and MATCHES[*START_I] should be -1.
1122 : The function will return true and have the stmts isomorphic to
1123 : MATCHES[*START_I] marked with *START_I.
1124 :
1125 : *TWO_OPERATORS indicates whether the group of isomorphic statements
1126 : uses two related operations like PLUS_EXPR and MINUS_EXPR. If
1127 : TWO_OPERATORS is NULL such case is not considered isomorphic.
1128 :
1129 : SWAP[] indicates whether for a stmt to be isomorphic to its group
1130 : leader, has to have its operands swapped (1) or its predicate inverted (2).
1131 : COND_EXPR is possibly isomorphic to another one after swapping its
1132 : operands. Set SWAP[i] to 1 if stmt I is COND_EXPR and isomorphic to
1133 : the first stmt by swapping the two operands of comparison; set SWAP[i]
1134 : to 2 if stmt I is isormorphic to the first stmt by inverting the code
1135 : of comparison. Take A1 >= B1 ? X1 : Y1 as an example, it can be swapped
1136 : to (B1 <= A1 ? X1 : Y1); or be inverted to (A1 < B1) ? Y1 : X1. */
1137 :
1138 : static bool
1139 6271538 : vect_build_slp_tree_3 (vec_info *vinfo, vec<stmt_vec_info> stmts,
1140 : match_elt_t *matches, unsigned char *swap,
1141 : bool *two_operators, tree vectype, unsigned *start_i)
1142 : {
1143 6271538 : stmt_vec_info first_stmt_info = NULL;
1144 6271538 : code_helper first_stmt_code = ERROR_MARK;
1145 6271538 : code_helper alt_stmt_code = ERROR_MARK;
1146 6271538 : code_helper first_cond_code = ERROR_MARK;
1147 6271538 : bool need_same_oprnds = false;
1148 6271538 : tree first_lhs = NULL_TREE;
1149 6271538 : tree first_op1 = NULL_TREE;
1150 6271538 : stmt_vec_info first_load = NULL, prev_first_load = NULL;
1151 6271538 : bool first_stmt_ldst_p = false, first_stmt_ldst_masklen_p = false;
1152 6271538 : bool first_stmt_phi_p = false;
1153 6271538 : int first_reduc_idx = -1;
1154 :
1155 6271538 : basic_block common_bb = NULL;
1156 6271538 : gimple *trapping_stmt = NULL;
1157 6271538 : int first_match = -1;
1158 :
1159 26370927 : for (unsigned i = *start_i; i < stmts.length (); ++i)
1160 : {
1161 20099624 : stmt_vec_info stmt_info = stmts[i];
1162 20099624 : bool ldst_p = false;
1163 20099624 : bool ldst_masklen_p = false;
1164 20099624 : bool phi_p = false;
1165 20099624 : code_helper rhs_code = ERROR_MARK;
1166 :
1167 20099624 : if (*start_i == 0)
1168 : {
1169 18057542 : swap[i] = 0;
1170 18057542 : matches[i] = -1;
1171 : }
1172 2042082 : else if (matches[i] != -1)
1173 20099389 : continue;
1174 :
1175 20009511 : if (!stmt_info)
1176 : {
1177 : /* ??? We shouldn't run into this. */
1178 40928 : gcc_assert (first_match != -1);
1179 40928 : matches[i] = first_match;
1180 40928 : continue;
1181 : }
1182 :
1183 19968583 : gimple *stmt = stmt_info->stmt;
1184 19968583 : if (dump_enabled_p ())
1185 236230 : dump_printf_loc (MSG_NOTE, vect_location, "Build SLP for %G", stmt);
1186 :
1187 : /* Fail to vectorize statements marked as unvectorizable, throw
1188 : or are volatile. */
1189 19968583 : if (!STMT_VINFO_VECTORIZABLE (stmt_info)
1190 19576022 : || stmt_can_throw_internal (cfun, stmt)
1191 38518603 : || gimple_has_volatile_ops (stmt))
1192 : {
1193 409468 : if (dump_enabled_p ())
1194 466 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1195 : "Build SLP failed: unvectorizable statement %G",
1196 : stmt);
1197 409468 : if (is_a <bb_vec_info> (vinfo))
1198 : {
1199 409468 : matches[i] = -2;
1200 409468 : continue;
1201 : }
1202 6271538 : return false;
1203 : }
1204 :
1205 19559115 : gcall *call_stmt = dyn_cast <gcall *> (stmt);
1206 19559115 : tree lhs = gimple_get_lhs (stmt);
1207 19559115 : if (lhs == NULL_TREE && !call_stmt)
1208 : {
1209 34 : if (dump_enabled_p ())
1210 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1211 : "Build SLP failed: not GIMPLE_ASSIGN nor "
1212 : "GIMPLE_CALL %G", stmt);
1213 34 : if (is_a <bb_vec_info> (vinfo))
1214 : {
1215 34 : matches[i] = -2;
1216 34 : continue;
1217 : }
1218 : return false;
1219 : }
1220 :
1221 19559081 : if (call_stmt)
1222 : {
1223 204078 : combined_fn cfn = gimple_call_combined_fn (call_stmt);
1224 204078 : if (cfn != CFN_LAST && cfn != CFN_MASK_CALL)
1225 62511 : rhs_code = cfn;
1226 : else
1227 : rhs_code = CALL_EXPR;
1228 :
1229 204078 : if (cfn == CFN_GATHER_LOAD
1230 204078 : || cfn == CFN_SCATTER_STORE)
1231 : ldst_p = true;
1232 : else if (cfn == CFN_MASK_LOAD
1233 : || cfn == CFN_MASK_GATHER_LOAD
1234 : || cfn == CFN_MASK_LEN_GATHER_LOAD
1235 : || cfn == CFN_MASK_SCATTER_STORE
1236 : || cfn == CFN_MASK_LEN_SCATTER_STORE)
1237 : {
1238 : ldst_p = true;
1239 : ldst_masklen_p = true;
1240 : }
1241 : else if (cfn == CFN_MASK_STORE)
1242 : {
1243 : ldst_p = true;
1244 : ldst_masklen_p = true;
1245 : rhs_code = CFN_MASK_STORE;
1246 : }
1247 : else if (cfn == CFN_GOMP_SIMD_LANE)
1248 : ;
1249 192977 : else if ((cfn != CFN_LAST
1250 : && cfn != CFN_MASK_CALL
1251 51410 : && internal_fn_p (cfn)
1252 38787 : && !vectorizable_internal_fn_p (as_internal_fn (cfn)))
1253 192858 : || gimple_call_tail_p (call_stmt)
1254 192858 : || gimple_call_noreturn_p (call_stmt)
1255 385835 : || gimple_call_chain (call_stmt))
1256 : {
1257 1402 : if (dump_enabled_p ())
1258 8 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1259 : "Build SLP failed: unsupported call type %G",
1260 : (gimple *) call_stmt);
1261 1402 : if (is_a <bb_vec_info> (vinfo))
1262 : {
1263 1402 : matches[i] = -2;
1264 1402 : continue;
1265 : }
1266 : return false;
1267 : }
1268 : }
1269 19355003 : else if (gimple_code (stmt) == GIMPLE_PHI)
1270 : {
1271 19509638 : rhs_code = ERROR_MARK;
1272 : phi_p = true;
1273 : }
1274 : else
1275 : {
1276 18329001 : rhs_code = gimple_assign_rhs_code (stmt);
1277 18329001 : ldst_p = STMT_VINFO_DATA_REF (stmt_info) != nullptr;
1278 : }
1279 :
1280 18329001 : if (!ldst_p
1281 : && !phi_p
1282 13244094 : && rhs_code.is_tree_code ()
1283 13189152 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_binary
1284 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_unary
1285 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_expression
1286 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_comparison
1287 239926 : && rhs_code != VIEW_CONVERT_EXPR
1288 : && rhs_code != CALL_EXPR
1289 : && rhs_code != BIT_FIELD_REF
1290 18329001 : && rhs_code != SSA_NAME)
1291 : {
1292 48041 : if (dump_enabled_p ())
1293 98 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1294 : "Build SLP failed: operation unsupported %G",
1295 : stmt);
1296 48041 : if (is_a <bb_vec_info> (vinfo))
1297 : {
1298 47932 : matches[i] = -2;
1299 47932 : continue;
1300 : }
1301 : return false;
1302 : }
1303 :
1304 : /* Non-grouped store or load. */
1305 19509638 : if (ldst_p
1306 5287583 : && !STMT_VINFO_GROUPED_ACCESS (stmt_info)
1307 1128752 : && DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))
1308 851914 : && rhs_code != CFN_GATHER_LOAD
1309 851914 : && rhs_code != CFN_MASK_GATHER_LOAD
1310 851914 : && rhs_code != CFN_MASK_LEN_GATHER_LOAD
1311 851914 : && rhs_code != CFN_SCATTER_STORE
1312 851914 : && rhs_code != CFN_MASK_SCATTER_STORE
1313 851914 : && rhs_code != CFN_MASK_LEN_SCATTER_STORE
1314 851914 : && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)
1315 : /* Not grouped loads are handled as externals for BB
1316 : vectorization. Treat them as not vectorizable. */
1317 20341892 : && is_a <bb_vec_info> (vinfo))
1318 : {
1319 : /* Not grouped load. */
1320 0 : if (dump_enabled_p ())
1321 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1322 : "Build SLP failed: not grouped load %G", stmt);
1323 0 : matches[i] = -2;
1324 0 : continue;
1325 : }
1326 :
1327 19509638 : if (!ldst_p && rhs_code == BIT_FIELD_REF)
1328 : {
1329 45031 : tree vec = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
1330 45031 : if (!is_a <bb_vec_info> (vinfo)
1331 44905 : || TREE_CODE (vec) != SSA_NAME
1332 : /* When the element types are not compatible we pun the
1333 : source to the target vectype which requires equal size. */
1334 89912 : || ((!VECTOR_TYPE_P (TREE_TYPE (vec))
1335 42153 : || !types_compatible_p (TREE_TYPE (vectype),
1336 42153 : TREE_TYPE (TREE_TYPE (vec))))
1337 3997 : && !operand_equal_p (TYPE_SIZE (vectype),
1338 3997 : TYPE_SIZE (TREE_TYPE (vec)))))
1339 : {
1340 2282 : if (dump_enabled_p ())
1341 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1342 : "Build SLP failed: "
1343 : "BIT_FIELD_REF not supported\n");
1344 2282 : if (is_a <bb_vec_info> (vinfo))
1345 : {
1346 2156 : matches[i] = -2;
1347 2156 : continue;
1348 : }
1349 : return false;
1350 : }
1351 : }
1352 :
1353 : /* Check the operation. */
1354 19507356 : if (first_match == -1)
1355 : {
1356 6170795 : first_match = i;
1357 6170795 : first_stmt_info = stmt_info;
1358 6170795 : common_bb = gimple_bb (stmt_info->stmt);
1359 6170795 : first_lhs = lhs;
1360 6170795 : first_stmt_code = rhs_code;
1361 6170795 : first_stmt_ldst_p = ldst_p;
1362 6170795 : first_stmt_ldst_masklen_p = ldst_masklen_p;
1363 6170795 : first_stmt_phi_p = phi_p;
1364 6170795 : first_reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info);
1365 :
1366 : /* Shift arguments should be equal in all the packed stmts for a
1367 : vector shift with scalar shift operand. */
1368 6170795 : if (rhs_code == LSHIFT_EXPR
1369 6078221 : || rhs_code == RSHIFT_EXPR
1370 5986824 : || rhs_code == LROTATE_EXPR
1371 12157505 : || rhs_code == RROTATE_EXPR)
1372 : {
1373 : /* First see if we have a vector/vector shift. */
1374 184570 : if (!directly_supported_p (rhs_code, vectype, optab_vector))
1375 : {
1376 : /* No vector/vector shift, arrange for a vector/scalar
1377 : SLP layout. */
1378 171187 : need_same_oprnds = true;
1379 171187 : first_op1 = gimple_assign_rhs2 (stmt);
1380 : }
1381 : }
1382 5986225 : else if (rhs_code == WIDEN_LSHIFT_EXPR)
1383 : {
1384 0 : need_same_oprnds = true;
1385 0 : first_op1 = gimple_assign_rhs2 (stmt);
1386 : }
1387 5986225 : else if (rhs_code == CFN_DIV_POW2)
1388 : {
1389 0 : need_same_oprnds = true;
1390 0 : first_op1 = gimple_call_arg (call_stmt, 1);
1391 : }
1392 5986225 : else if (rhs_code == CFN_GOMP_SIMD_LANE)
1393 : {
1394 3651 : need_same_oprnds = true;
1395 3651 : first_op1 = gimple_call_arg (call_stmt, 1);
1396 : }
1397 : }
1398 : else
1399 : {
1400 13336561 : int comm_arg;
1401 13337000 : if (first_reduc_idx != STMT_VINFO_REDUC_IDX (stmt_info)
1402 : /* For SLP reduction groups the index isn't necessarily
1403 : uniform but only that of the first stmt matters. */
1404 2525 : && !(first_reduc_idx != -1
1405 2525 : && STMT_VINFO_REDUC_IDX (stmt_info) != -1
1406 2525 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info))
1407 13336561 : && !(first_reduc_idx != -1
1408 1187 : && STMT_VINFO_REDUC_IDX (stmt_info) != -1
1409 1187 : && (comm_arg = first_commutative_argument
1410 1187 : (rhs_code, TREE_TYPE (lhs))) >= 0
1411 : && (first_reduc_idx
1412 932 : == 2 * comm_arg + 1 - STMT_VINFO_REDUC_IDX (stmt_info))))
1413 : {
1414 439 : if (dump_enabled_p ())
1415 : {
1416 12 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1417 : "Build SLP failed: different reduc_idx "
1418 : "%d instead of %d in %G",
1419 : STMT_VINFO_REDUC_IDX (stmt_info),
1420 : first_reduc_idx, stmt);
1421 : }
1422 : /* Mismatch. */
1423 439 : continue;
1424 : }
1425 13336122 : if (!ldst_p
1426 13336122 : && two_operators
1427 10496394 : && first_stmt_code != rhs_code
1428 14849032 : && alt_stmt_code == ERROR_MARK)
1429 : alt_stmt_code = rhs_code;
1430 14828021 : if ((!ldst_p
1431 10496394 : && first_stmt_code != rhs_code
1432 1512910 : && (first_stmt_code != IMAGPART_EXPR
1433 203 : || rhs_code != REALPART_EXPR)
1434 1512910 : && (first_stmt_code != REALPART_EXPR
1435 738 : || rhs_code != IMAGPART_EXPR)
1436 : /* Handle mismatches in plus/minus by computing both
1437 : and merging the results. */
1438 1512910 : && !((((first_stmt_code == PLUS_EXPR
1439 1402599 : || first_stmt_code == MINUS_EXPR)
1440 135401 : && (alt_stmt_code == PLUS_EXPR
1441 125459 : || alt_stmt_code == MINUS_EXPR))
1442 1480975 : || ((first_stmt_code == CFN_FMA
1443 1480973 : || first_stmt_code == CFN_FMS)
1444 2 : && (alt_stmt_code == CFN_FMA
1445 2 : || alt_stmt_code == CFN_FMS)))
1446 31937 : && rhs_code == alt_stmt_code)
1447 1524580 : && !(first_stmt_code.is_tree_code ()
1448 1382048 : && rhs_code.is_tree_code ()
1449 1264503 : && (TREE_CODE_CLASS (tree_code (first_stmt_code))
1450 : == tcc_comparison)
1451 140262 : && (swap_tree_comparison (tree_code (first_stmt_code))
1452 140262 : == tree_code (rhs_code))
1453 : && (first_reduc_idx == -1
1454 0 : || REDUC_GROUP_FIRST_ELEMENT (stmt_info))))
1455 : || (ldst_p
1456 5679456 : && (STMT_VINFO_GROUPED_ACCESS (stmt_info)
1457 2839728 : != STMT_VINFO_GROUPED_ACCESS (first_stmt_info)))
1458 : || (ldst_p
1459 2786422 : && (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
1460 2786422 : != STMT_VINFO_GATHER_SCATTER_P (first_stmt_info)))
1461 11844361 : || first_stmt_ldst_p != ldst_p
1462 11844231 : || (ldst_p && first_stmt_ldst_masklen_p != ldst_masklen_p)
1463 25180345 : || first_stmt_phi_p != phi_p)
1464 : {
1465 1491899 : if (dump_enabled_p ())
1466 : {
1467 3609 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1468 : "Build SLP failed: different operation "
1469 : "in stmt %G", stmt);
1470 3609 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1471 : "original stmt %G", first_stmt_info->stmt);
1472 : }
1473 : /* Mismatch. */
1474 1491899 : continue;
1475 : }
1476 :
1477 11863632 : if (!ldst_p
1478 9057954 : && first_stmt_code == BIT_FIELD_REF
1479 11874139 : && (TREE_OPERAND (gimple_assign_rhs1 (first_stmt_info->stmt), 0)
1480 29916 : != TREE_OPERAND (gimple_assign_rhs1 (stmt_info->stmt), 0)))
1481 : {
1482 19409 : if (dump_enabled_p ())
1483 116 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1484 : "Build SLP failed: different BIT_FIELD_REF "
1485 : "arguments in %G", stmt);
1486 : /* Mismatch. */
1487 19409 : continue;
1488 : }
1489 :
1490 11824814 : if (call_stmt
1491 88439 : && first_stmt_code != CFN_MASK_LOAD
1492 11912851 : && first_stmt_code != CFN_MASK_STORE)
1493 : {
1494 87726 : if (!is_a <gcall *> (stmts[0]->stmt)
1495 87726 : || !compatible_calls_p (as_a <gcall *> (stmts[0]->stmt),
1496 : call_stmt, true))
1497 : {
1498 69162 : if (dump_enabled_p ())
1499 13 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1500 : "Build SLP failed: different calls in %G",
1501 : stmt);
1502 : /* Mismatch. */
1503 69162 : continue;
1504 : }
1505 : }
1506 :
1507 11798288 : if (phi_p
1508 11755652 : && (gimple_bb (first_stmt_info->stmt)
1509 213259 : != gimple_bb (stmt_info->stmt)))
1510 : {
1511 42636 : if (dump_enabled_p ())
1512 8 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1513 : "Build SLP failed: different BB for PHI %G",
1514 : stmt);
1515 : /* Mismatch. */
1516 42636 : continue;
1517 : }
1518 :
1519 11713016 : if (need_same_oprnds)
1520 : {
1521 62452 : tree other_op1 = gimple_arg (stmt, 1);
1522 62452 : if (!operand_equal_p (first_op1, other_op1, 0))
1523 : {
1524 11032 : if (dump_enabled_p ())
1525 195 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1526 : "Build SLP failed: different shift "
1527 : "arguments in %G", stmt);
1528 : /* Mismatch. */
1529 11032 : continue;
1530 : }
1531 : }
1532 :
1533 11702825 : if (first_lhs
1534 11701984 : && lhs
1535 11701984 : && !types_compatible_p (TREE_TYPE (lhs), TREE_TYPE (first_lhs)))
1536 : {
1537 841 : if (dump_enabled_p ())
1538 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1539 : "Build SLP failed: different vector type "
1540 : "in %G", stmt);
1541 : /* Mismatch. */
1542 841 : continue;
1543 : }
1544 :
1545 : /* We need to ensure all stmts are in the same BB when one stmt could
1546 : trap. */
1547 11701143 : if (trapping_stmt || gimple_could_trap_p (stmt))
1548 : {
1549 1171457 : gcc_assert (!trapping_stmt || common_bb);
1550 1171457 : if (gimple_bb (stmt) != common_bb)
1551 : {
1552 6066 : if (dump_enabled_p ())
1553 13 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1554 : "Build SLP failed: not all stmts in same "
1555 : "BB but possibly trapping operation in %G",
1556 : trapping_stmt ? trapping_stmt : stmt);
1557 6066 : continue;
1558 : }
1559 : }
1560 : }
1561 :
1562 : /* Grouped store or load. */
1563 17865872 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
1564 : {
1565 4100101 : gcc_assert (ldst_p);
1566 4100101 : if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info)))
1567 : {
1568 : /* Store. */
1569 3022043 : gcc_assert (rhs_code == CFN_MASK_STORE
1570 : || REFERENCE_CLASS_P (lhs)
1571 : || DECL_P (lhs));
1572 : }
1573 : else
1574 : {
1575 : /* Load. */
1576 1078058 : first_load = DR_GROUP_FIRST_ELEMENT (stmt_info);
1577 1078058 : if (prev_first_load)
1578 : {
1579 : /* Check that there are no loads from different interleaving
1580 : chains in the same node. */
1581 505623 : if (prev_first_load != first_load)
1582 : {
1583 106286 : if (dump_enabled_p ())
1584 2296 : dump_printf_loc (MSG_MISSED_OPTIMIZATION,
1585 : vect_location,
1586 : "Build SLP failed: different "
1587 : "interleaving chains in one node %G",
1588 : stmt);
1589 : /* Mismatch. */
1590 106286 : continue;
1591 : }
1592 : }
1593 : else
1594 : prev_first_load = first_load;
1595 : }
1596 : }
1597 : /* Non-grouped store or load. */
1598 13765771 : else if (ldst_p)
1599 : {
1600 1128441 : if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))
1601 851603 : && rhs_code != CFN_GATHER_LOAD
1602 : && rhs_code != CFN_MASK_GATHER_LOAD
1603 : && rhs_code != CFN_MASK_LEN_GATHER_LOAD
1604 : && rhs_code != CFN_SCATTER_STORE
1605 : && rhs_code != CFN_MASK_SCATTER_STORE
1606 : && rhs_code != CFN_MASK_LEN_SCATTER_STORE
1607 1980044 : && !STMT_VINFO_GATHER_SCATTER_P (stmt_info))
1608 : {
1609 831962 : gcc_assert (!is_a <bb_vec_info> (vinfo));
1610 : /* For loop vectorization we can handle splats the same we
1611 : handle single element interleaving. Likewise we can handle
1612 : a collection of invariant refs. */
1613 1036161 : if (stmt_info != first_stmt_info
1614 832224 : && !(integer_zerop (DR_STEP (STMT_VINFO_DATA_REF (stmt_info)))
1615 262 : && integer_zerop (DR_STEP (STMT_VINFO_DATA_REF
1616 : (first_stmt_info)))))
1617 : {
1618 : /* Not grouped load. */
1619 204199 : if (dump_enabled_p ())
1620 179 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1621 : "Build SLP failed: not grouped load %G",
1622 : stmt);
1623 204199 : continue;
1624 : }
1625 : }
1626 : }
1627 : /* Not memory operation. */
1628 : else
1629 : {
1630 12637330 : if (rhs_code == COND_EXPR)
1631 : {
1632 71368 : tree cond_expr = gimple_assign_rhs1 (stmt);
1633 71368 : enum tree_code cond_code = TREE_CODE (cond_expr);
1634 71368 : enum tree_code swap_code = ERROR_MARK;
1635 71368 : enum tree_code invert_code = ERROR_MARK;
1636 :
1637 71368 : if (i == (unsigned)first_match)
1638 59186 : first_cond_code = TREE_CODE (cond_expr);
1639 12182 : else if (TREE_CODE_CLASS (cond_code) == tcc_comparison)
1640 : {
1641 0 : bool honor_nans = HONOR_NANS (TREE_OPERAND (cond_expr, 0));
1642 0 : swap_code = swap_tree_comparison (cond_code);
1643 0 : invert_code = invert_tree_comparison (cond_code, honor_nans);
1644 : }
1645 :
1646 71368 : if (first_cond_code == cond_code)
1647 : ;
1648 : /* Isomorphic can be achieved by swapping. */
1649 0 : else if (first_cond_code == swap_code)
1650 0 : swap[i] = 1;
1651 : /* Isomorphic can be achieved by inverting. */
1652 0 : else if (first_cond_code == invert_code)
1653 0 : swap[i] = 2;
1654 : else
1655 : {
1656 0 : if (dump_enabled_p ())
1657 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1658 : "Build SLP failed: different"
1659 : " operation %G", stmt);
1660 : /* Mismatch. */
1661 0 : continue;
1662 : }
1663 : }
1664 :
1665 12637330 : if (i != (unsigned) first_match
1666 8914390 : && first_stmt_code != rhs_code
1667 74468 : && first_stmt_code.is_tree_code ()
1668 74466 : && rhs_code.is_tree_code ()
1669 74466 : && TREE_CODE_CLASS ((tree_code)first_stmt_code) == tcc_comparison
1670 12680402 : && (swap_tree_comparison ((tree_code)first_stmt_code)
1671 43072 : == (tree_code)rhs_code))
1672 43072 : swap[i] = 1;
1673 :
1674 12637330 : if (i != (unsigned) first_match
1675 8914390 : && first_reduc_idx != STMT_VINFO_REDUC_IDX (stmt_info)
1676 1786 : && first_reduc_idx != -1
1677 1786 : && STMT_VINFO_REDUC_IDX (stmt_info) != -1
1678 1786 : && rhs_code.is_tree_code ()
1679 1778 : && commutative_tree_code (tree_code (rhs_code))
1680 12639106 : && first_reduc_idx == 1 - STMT_VINFO_REDUC_IDX (stmt_info))
1681 1776 : swap[i] = 1;
1682 : }
1683 :
1684 : /* We need to ensure all stmts are in the same BB when one stmt could
1685 : trap. Not matching stmts are not relevant, so exclude those. */
1686 17555387 : if (!trapping_stmt && gimple_could_trap_p (stmt))
1687 : trapping_stmt = stmt;
1688 17555387 : if (common_bb != gimple_bb (stmt))
1689 : {
1690 63244 : common_bb = NULL;
1691 63244 : gcc_assert (!trapping_stmt);
1692 : }
1693 :
1694 17555387 : matches[i] = first_match;
1695 : }
1696 :
1697 : /* Record if we allowed two distinct operations for the SLP node. */
1698 6271303 : if (((first_stmt_code == PLUS_EXPR
1699 5749453 : || first_stmt_code == MINUS_EXPR)
1700 632321 : && (alt_stmt_code == PLUS_EXPR
1701 625022 : || alt_stmt_code == MINUS_EXPR))
1702 12525561 : || ((first_stmt_code == CFN_FMA
1703 6254209 : || first_stmt_code == CFN_FMS)
1704 49 : && (alt_stmt_code == CFN_FMA
1705 49 : || alt_stmt_code == CFN_FMS)))
1706 17047 : *two_operators = true;
1707 :
1708 6271303 : if (first_match == -1)
1709 : {
1710 : /* If there was no useful stmt, fail. Can only happen during
1711 : the first sweep. */
1712 100508 : gcc_assert (*start_i == 0);
1713 100508 : *start_i = stmts.length () - 1;
1714 100508 : return false;
1715 : }
1716 :
1717 6170795 : *start_i = first_match;
1718 6170795 : return true;
1719 : }
1720 :
1721 : /* Verify if the scalar stmts STMTS are isomorphic, require data
1722 : permutation or are of unsupported types of operation. Return
1723 : true if they are, otherwise return false and indicate in *MATCHES
1724 : the groups of isomorphic stmts. See vect_build_slp_tree_3 for
1725 : details. *TWO_OPERATORS is for the first isomorphic group,
1726 : knowledge whether following isomorphic groups have one or two operators
1727 : is not retained. */
1728 :
1729 : static bool
1730 6100601 : vect_build_slp_tree_1 (vec_info *vinfo, unsigned char *swap,
1731 : vec<stmt_vec_info> stmts, match_elt_t *matches,
1732 : bool *two_operators, tree *node_vectype)
1733 : {
1734 6100601 : stmt_vec_info first_stmt_info = stmts[0];
1735 6100601 : unsigned int group_size = stmts.length ();
1736 6100601 : tree vectype;
1737 6100601 : if (!vect_get_vector_types_for_stmt (vinfo, first_stmt_info, &vectype,
1738 : group_size))
1739 : {
1740 : /* Fatal mismatch. */
1741 229172 : matches[0] = -1;
1742 229172 : return false;
1743 : }
1744 5871429 : if (is_a <bb_vec_info> (vinfo)
1745 5871429 : && known_le (TYPE_VECTOR_SUBPARTS (vectype), 1U))
1746 : {
1747 348731 : if (dump_enabled_p ())
1748 301 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1749 : "Build SLP failed: not using single lane "
1750 : "vector type %T\n", vectype);
1751 348731 : matches[0] = -1;
1752 348731 : return false;
1753 : }
1754 : /* Check nunits required but continue analysis, producing matches[]
1755 : as if nunits was not an issue. This allows splitting of groups
1756 : to happen. */
1757 5522698 : unsigned HOST_WIDE_INT const_nunits = 0;
1758 5522698 : if (vectype
1759 5522678 : && is_a <bb_vec_info> (vinfo)
1760 11045376 : && !multiple_p (group_size, TYPE_VECTOR_SUBPARTS (vectype)))
1761 : {
1762 209716 : if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits)
1763 209716 : || const_nunits > group_size)
1764 : {
1765 0 : if (dump_enabled_p ())
1766 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1767 : "Build SLP failed: unrolling required "
1768 : "in basic block SLP\n");
1769 : /* Fatal mismatch. */
1770 0 : matches[0] = -1;
1771 0 : return false;
1772 : }
1773 : }
1774 :
1775 5522698 : gcc_assert (vectype || !gimple_get_lhs (first_stmt_info->stmt));
1776 5522698 : *node_vectype = vectype;
1777 :
1778 5522698 : unsigned start_i = 0;
1779 5522698 : if (!vect_build_slp_tree_3 (vinfo, stmts, matches, swap, two_operators,
1780 : vectype, &start_i))
1781 : {
1782 : /* Fatal mismatch. */
1783 100743 : matches[0] = -1;
1784 100743 : return false;
1785 : }
1786 5421955 : gcc_assert (matches[start_i] == (int)start_i);
1787 : /* Discover further isomorphic groups. */
1788 17708496 : for (start_i = start_i + 1; start_i < group_size; ++start_i)
1789 12286541 : if (matches[start_i] == -1)
1790 : {
1791 748840 : unsigned prev_start_i = start_i;
1792 748840 : bool tem_two_operators;
1793 748840 : bool res = vect_build_slp_tree_3 (vinfo, stmts, matches, swap,
1794 : &tem_two_operators,
1795 : vectype, &start_i);
1796 748840 : gcc_assert (res && matches[start_i] == (int)prev_start_i);
1797 : }
1798 :
1799 : bool res = true;
1800 23208210 : for (unsigned i = 0; i < group_size; ++i)
1801 17786255 : if (matches[i] != 0)
1802 1890510 : res = false;
1803 : if (!res)
1804 : return false;
1805 :
1806 : return true;
1807 : }
1808 :
1809 : /* Traits for the hash_set to record failed SLP builds for a stmt set.
1810 : Note we never remove apart from at destruction time so we do not
1811 : need a special value for deleted that differs from empty. */
1812 : struct bst_traits
1813 : {
1814 : typedef vec <stmt_vec_info> value_type;
1815 : typedef vec <stmt_vec_info> compare_type;
1816 : static inline hashval_t hash (value_type);
1817 : static inline bool equal (value_type existing, value_type candidate);
1818 506535097 : static inline bool is_empty (value_type x) { return !x.exists (); }
1819 113595722 : static inline bool is_deleted (value_type x) { return !x.exists (); }
1820 : static const bool empty_zero_p = true;
1821 0 : static inline void mark_empty (value_type &x) { x.release (); }
1822 : static inline void mark_deleted (value_type &x) { x.release (); }
1823 9744003 : static inline void remove (value_type &x) { x.release (); }
1824 : };
1825 : inline hashval_t
1826 98879219 : bst_traits::hash (value_type x)
1827 : {
1828 98879219 : inchash::hash h;
1829 438705729 : for (unsigned i = 0; i < x.length (); ++i)
1830 339826510 : h.add_int (x[i] ? gimple_uid (x[i]->stmt) : -1);
1831 98879219 : return h.end ();
1832 : }
1833 : inline bool
1834 86602193 : bst_traits::equal (value_type existing, value_type candidate)
1835 : {
1836 259806579 : if (existing.length () != candidate.length ())
1837 : return false;
1838 87197819 : for (unsigned i = 0; i < existing.length (); ++i)
1839 82639957 : if (existing[i] != candidate[i])
1840 : return false;
1841 : return true;
1842 : }
1843 :
1844 : typedef hash_map <vec <stmt_vec_info>, slp_tree,
1845 : simple_hashmap_traits <bst_traits, slp_tree> >
1846 : scalar_stmts_to_slp_tree_map_t;
1847 :
1848 : /* Release BST_MAP. */
1849 :
1850 : static void
1851 1897140 : release_scalar_stmts_to_slp_tree_map (scalar_stmts_to_slp_tree_map_t *bst_map)
1852 : {
1853 : /* The map keeps a reference on SLP nodes built, release that. */
1854 11641143 : for (scalar_stmts_to_slp_tree_map_t::iterator it = bst_map->begin ();
1855 11641143 : it != bst_map->end (); ++it)
1856 9744003 : if ((*it).second)
1857 9744003 : vect_free_slp_tree ((*it).second);
1858 3794280 : delete bst_map;
1859 1897140 : }
1860 :
1861 : /* ??? This was std::pair<std::pair<tree_code, vect_def_type>, tree>
1862 : but then vec::insert does memmove and that's not compatible with
1863 : std::pair. */
1864 : struct chain_op_t
1865 : {
1866 4226552 : chain_op_t (tree_code code_, vect_def_type dt_, tree op_)
1867 4226552 : : code (code_), dt (dt_), op (op_) {}
1868 : tree_code code;
1869 : vect_def_type dt;
1870 : tree op;
1871 : };
1872 :
1873 : /* Comparator for sorting associatable chains. */
1874 :
1875 : static int
1876 12395998 : dt_sort_cmp (const void *op1_, const void *op2_, void *)
1877 : {
1878 12395998 : auto *op1 = (const chain_op_t *) op1_;
1879 12395998 : auto *op2 = (const chain_op_t *) op2_;
1880 12395998 : if (op1->dt != op2->dt)
1881 1828057 : return (int)op1->dt - (int)op2->dt;
1882 10567941 : return (int)op1->code - (int)op2->code;
1883 : }
1884 :
1885 : /* Linearize the associatable expression chain at START with the
1886 : associatable operation CODE (where PLUS_EXPR also allows MINUS_EXPR),
1887 : filling CHAIN with the result and using WORKLIST as intermediate storage.
1888 : CODE_STMT and ALT_CODE_STMT are filled with the first stmt using CODE
1889 : or MINUS_EXPR. *CHAIN_STMTS if not NULL is filled with all computation
1890 : stmts, starting with START. When ALLOW_ALT_CODE is false, do not
1891 : follow into MINUS_EXPR when building a PLUS chain (treat MINUS as leaf). */
1892 :
1893 : static void
1894 1822716 : vect_slp_linearize_chain (vec_info *vinfo,
1895 : vec<std::pair<tree_code, gimple *> > &worklist,
1896 : vec<chain_op_t> &chain,
1897 : enum tree_code code, gimple *start,
1898 : gimple *&code_stmt, gimple *&alt_code_stmt,
1899 : vec<gimple *> *chain_stmts,
1900 : bool allow_alt_code = true)
1901 : {
1902 : /* For each lane linearize the addition/subtraction (or other
1903 : uniform associatable operation) expression tree. */
1904 1822716 : worklist.safe_push (std::make_pair (code, start));
1905 4226552 : while (!worklist.is_empty ())
1906 : {
1907 2403836 : auto entry = worklist.pop ();
1908 2403836 : gassign *stmt = as_a <gassign *> (entry.second);
1909 2403836 : enum tree_code in_code = entry.first;
1910 4807672 : enum tree_code this_code = gimple_assign_rhs_code (stmt);
1911 : /* Pick some stmts suitable for SLP_TREE_REPRESENTATIVE. */
1912 2403836 : if (!code_stmt
1913 2403836 : && gimple_assign_rhs_code (stmt) == code)
1914 1515991 : code_stmt = stmt;
1915 887845 : else if (!alt_code_stmt
1916 887845 : && gimple_assign_rhs_code (stmt) == MINUS_EXPR)
1917 361728 : alt_code_stmt = stmt;
1918 2403836 : if (chain_stmts)
1919 2322768 : chain_stmts->safe_push (stmt);
1920 7211508 : for (unsigned opnum = 1; opnum <= 2; ++opnum)
1921 : {
1922 4807672 : tree op = gimple_op (stmt, opnum);
1923 4807672 : vect_def_type dt;
1924 4807672 : stmt_vec_info def_stmt_info;
1925 4807672 : bool res = vect_is_simple_use (op, vinfo, &dt, &def_stmt_info);
1926 4807672 : gcc_assert (res);
1927 4807672 : if (dt == vect_internal_def
1928 4807672 : && is_pattern_stmt_p (def_stmt_info))
1929 9249 : op = gimple_get_lhs (def_stmt_info->stmt);
1930 4807672 : gimple *use_stmt;
1931 4807672 : use_operand_p use_p;
1932 4807672 : if (dt == vect_internal_def
1933 4376677 : && single_imm_use (op, &use_p, &use_stmt)
1934 2824532 : && is_gimple_assign (def_stmt_info->stmt)
1935 7437839 : && (gimple_assign_rhs_code (def_stmt_info->stmt) == code
1936 2049364 : || (allow_alt_code
1937 57371 : && code == PLUS_EXPR
1938 36328 : && (gimple_assign_rhs_code (def_stmt_info->stmt)
1939 : == MINUS_EXPR))))
1940 : {
1941 581120 : tree_code op_def_code = this_code;
1942 581120 : if (op_def_code == MINUS_EXPR && opnum == 1)
1943 53246 : op_def_code = PLUS_EXPR;
1944 581120 : if (in_code == MINUS_EXPR)
1945 222 : op_def_code = op_def_code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR;
1946 581120 : worklist.safe_push (std::make_pair (op_def_code,
1947 581120 : def_stmt_info->stmt));
1948 : }
1949 : else
1950 : {
1951 4226552 : tree_code op_def_code = this_code;
1952 4226552 : if (op_def_code == MINUS_EXPR && opnum == 1)
1953 308601 : op_def_code = PLUS_EXPR;
1954 4226552 : if (in_code == MINUS_EXPR)
1955 4218 : op_def_code = op_def_code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR;
1956 4226552 : chain.safe_push (chain_op_t (op_def_code, dt, op));
1957 : }
1958 : }
1959 : }
1960 1822716 : }
1961 :
1962 : /* Distance from the node currently being discovered to the closest upthread
1963 : commutative operation whose operand-zero discovery may still be fixed by
1964 : retrying with swapped operands, or -1U if there is none. */
1965 :
1966 : static unsigned least_upthread_swappable_op_distance = -1U;
1967 :
1968 : static slp_tree
1969 : vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
1970 : vec<stmt_vec_info> stmts,
1971 : match_elt_t *matches, unsigned *limit,
1972 : unsigned *tree_size,
1973 : scalar_stmts_to_slp_tree_map_t *bst_map);
1974 :
1975 : static slp_tree
1976 6602700 : vect_build_slp_tree (vec_info *vinfo,
1977 : vec<stmt_vec_info> stmts,
1978 : match_elt_t *matches, unsigned *limit, unsigned *tree_size,
1979 : scalar_stmts_to_slp_tree_map_t *bst_map)
1980 : {
1981 6602700 : unsigned int group_size = stmts.length ();
1982 6602700 : if (slp_tree *leader = bst_map->get (stmts))
1983 : {
1984 496314 : if (dump_enabled_p ())
1985 17816 : dump_printf_loc (MSG_NOTE, vect_location, "re-using %sSLP tree %p\n",
1986 17816 : !(*leader)->failed ? "" : "failed ",
1987 : (void *) *leader);
1988 496314 : if (!(*leader)->failed)
1989 : {
1990 451834 : SLP_TREE_REF_COUNT (*leader)++;
1991 451834 : stmts.release ();
1992 451834 : return *leader;
1993 : }
1994 44480 : memcpy (matches, (*leader)->failed, sizeof (match_elt_t) * group_size);
1995 44480 : return NULL;
1996 : }
1997 :
1998 : /* Single-lane SLP doesn't have the chance of run-away, do not account
1999 : it to the limit. */
2000 6106386 : if (stmts.length () > 1)
2001 : {
2002 3398561 : if (*limit == 0)
2003 : {
2004 1345 : if (dump_enabled_p ())
2005 15 : dump_printf_loc (MSG_NOTE, vect_location,
2006 : "SLP discovery limit exceeded\n");
2007 1345 : memset (matches, -1, sizeof (match_elt_t) * group_size);
2008 1345 : return NULL;
2009 : }
2010 3397216 : --*limit;
2011 : }
2012 :
2013 : /* Seed the bst_map with a stub node to be filled by vect_build_slp_tree_2
2014 : so we can pick up backedge destinations during discovery. */
2015 6105041 : slp_tree res = new _slp_tree;
2016 6105041 : SLP_TREE_DEF_TYPE (res) = vect_internal_def;
2017 6105041 : SLP_TREE_SCALAR_STMTS (res) = stmts;
2018 6105041 : bst_map->put (stmts.copy (), res);
2019 :
2020 6105041 : if (dump_enabled_p ())
2021 152241 : dump_printf_loc (MSG_NOTE, vect_location,
2022 : "starting SLP discovery for node %p\n", (void *) res);
2023 :
2024 6105041 : slp_tree res_ = vect_build_slp_tree_2 (vinfo, res, stmts,
2025 : matches, limit, tree_size, bst_map);
2026 6105041 : if (!res_)
2027 : {
2028 2135485 : if (dump_enabled_p ())
2029 8175 : dump_printf_loc (MSG_NOTE, vect_location,
2030 : "SLP discovery for node %p failed\n", (void *) res);
2031 : /* Mark the node invalid so we can detect those when still in use
2032 : as backedge destinations. */
2033 2135485 : SLP_TREE_SCALAR_STMTS (res) = vNULL;
2034 2135485 : SLP_TREE_DEF_TYPE (res) = vect_uninitialized_def;
2035 2135485 : res->failed = XNEWVEC (match_elt_t, group_size);
2036 2135485 : if (flag_checking)
2037 : {
2038 : unsigned i;
2039 3805460 : for (i = 0; i < group_size; ++i)
2040 3805460 : if (matches[i] != 0)
2041 : break;
2042 2135485 : gcc_assert (i < group_size);
2043 : }
2044 2135485 : memcpy (res->failed, matches, sizeof (match_elt_t) * group_size);
2045 : }
2046 : else
2047 : {
2048 3969556 : if (dump_enabled_p ())
2049 144066 : dump_printf_loc (MSG_NOTE, vect_location,
2050 : "SLP discovery for node %p succeeded\n",
2051 : (void *) res);
2052 3969556 : gcc_assert (res_ == res);
2053 : /* Keep a reference for the bst_map use. */
2054 3969556 : SLP_TREE_REF_COUNT (res)++;
2055 : /* For BB vectorization nodes that need splitting should be failed. */
2056 3969556 : gcc_assert (!is_a <bb_vec_info> (vinfo)
2057 : || multiple_p (group_size,
2058 : TYPE_VECTOR_SUBPARTS (res->vectype)));
2059 : }
2060 : return res_;
2061 : }
2062 :
2063 : /* Helper for building an associated SLP node chain. */
2064 :
2065 : static void
2066 162 : vect_slp_build_two_operator_nodes (slp_tree perm, tree vectype,
2067 : slp_tree op0, slp_tree op1,
2068 : stmt_vec_info oper1, stmt_vec_info oper2,
2069 : vec<std::pair<unsigned, unsigned> > lperm)
2070 : {
2071 162 : unsigned group_size = SLP_TREE_LANES (op1);
2072 :
2073 162 : slp_tree child1 = new _slp_tree;
2074 162 : SLP_TREE_DEF_TYPE (child1) = vect_internal_def;
2075 162 : SLP_TREE_VECTYPE (child1) = vectype;
2076 162 : SLP_TREE_LANES (child1) = group_size;
2077 162 : SLP_TREE_CHILDREN (child1).create (2);
2078 162 : SLP_TREE_CHILDREN (child1).quick_push (op0);
2079 162 : SLP_TREE_CHILDREN (child1).quick_push (op1);
2080 162 : SLP_TREE_REPRESENTATIVE (child1) = oper1;
2081 :
2082 162 : slp_tree child2 = new _slp_tree;
2083 162 : SLP_TREE_DEF_TYPE (child2) = vect_internal_def;
2084 162 : SLP_TREE_VECTYPE (child2) = vectype;
2085 162 : SLP_TREE_LANES (child2) = group_size;
2086 162 : SLP_TREE_CHILDREN (child2).create (2);
2087 162 : SLP_TREE_CHILDREN (child2).quick_push (op0);
2088 162 : SLP_TREE_REF_COUNT (op0)++;
2089 162 : SLP_TREE_CHILDREN (child2).quick_push (op1);
2090 162 : SLP_TREE_REF_COUNT (op1)++;
2091 162 : SLP_TREE_REPRESENTATIVE (child2) = oper2;
2092 :
2093 162 : SLP_TREE_DEF_TYPE (perm) = vect_internal_def;
2094 162 : SLP_TREE_CODE (perm) = VEC_PERM_EXPR;
2095 162 : SLP_TREE_VECTYPE (perm) = vectype;
2096 162 : SLP_TREE_LANES (perm) = group_size;
2097 162 : SLP_TREE_REPRESENTATIVE (perm) = NULL;
2098 162 : SLP_TREE_LANE_PERMUTATION (perm) = lperm;
2099 162 : SLP_TREE_CHILDREN (perm).quick_push (child1);
2100 162 : SLP_TREE_CHILDREN (perm).quick_push (child2);
2101 162 : }
2102 :
2103 : /* For isomorphic matches[], indicate a splitting point according to
2104 : CONST_NUNITS. */
2105 :
2106 : static void
2107 123577 : force_split_matches (match_elt_t *matches, unsigned group_size,
2108 : unsigned const_nunits)
2109 : {
2110 123577 : if (dump_enabled_p ())
2111 53 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2112 : "Build SLP failed: unrolling required "
2113 : "in basic block SLP\n");
2114 :
2115 : /* With constant vector elements simulate a mismatch at the
2116 : point we need to split. But indicate the tail is isomorphic. */
2117 123577 : unsigned tail = group_size & (const_nunits - 1);
2118 272241 : for (unsigned i = group_size - tail; i < group_size; ++i)
2119 148664 : matches[i] = (int)(group_size - tail);
2120 123577 : }
2121 :
2122 : /* Recursively build an SLP tree starting from NODE.
2123 : Fail (and return a value not equal to zero) if def-stmts are not
2124 : isomorphic, require data permutation or are of unsupported types of
2125 : operation. Otherwise, return 0.
2126 : The value returned is the depth in the SLP tree where a mismatch
2127 : was found. */
2128 :
2129 : static slp_tree
2130 6105041 : vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
2131 : vec<stmt_vec_info> stmts,
2132 : match_elt_t *matches, unsigned *limit,
2133 : unsigned *tree_size,
2134 : scalar_stmts_to_slp_tree_map_t *bst_map)
2135 : {
2136 6105041 : unsigned int group_size = stmts.length ();
2137 6105041 : unsigned nops, i, this_tree_size = 0;
2138 :
2139 6105041 : matches[0] = -1;
2140 :
2141 6105041 : stmt_vec_info stmt_info = stmts[0];
2142 6105041 : if (!is_a<gcall *> (stmt_info->stmt)
2143 : && !is_a<gassign *> (stmt_info->stmt)
2144 : && !is_a<gphi *> (stmt_info->stmt))
2145 : return NULL;
2146 :
2147 6104871 : nops = gimple_num_args (stmt_info->stmt);
2148 6104871 : if (const int *map = vect_get_operand_map (stmt_info))
2149 36716 : nops = map[0];
2150 :
2151 : /* If the SLP node is a PHI (induction or reduction), terminate
2152 : the recursion. */
2153 6104871 : bool *skip_args = XALLOCAVEC (bool, nops);
2154 6104871 : memset (skip_args, 0, sizeof (bool) * nops);
2155 6104871 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
2156 2906985 : if (gphi *stmt = dyn_cast <gphi *> (stmt_info->stmt))
2157 : {
2158 316989 : tree scalar_type = TREE_TYPE (PHI_RESULT (stmt));
2159 316989 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
2160 : group_size);
2161 316989 : if (!vectype)
2162 : {
2163 4264 : if (dump_enabled_p ())
2164 7 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2165 : "Build SLP failed: unsupported data-type in %G\n",
2166 : stmt_info->stmt);
2167 : return NULL;
2168 : }
2169 :
2170 312725 : vect_def_type def_type = STMT_VINFO_DEF_TYPE (stmt_info);
2171 312725 : if (def_type == vect_induction_def)
2172 : {
2173 : /* Induction PHIs are not cycles but walk the initial
2174 : value. Only for inner loops through, for outer loops
2175 : we need to pick up the value from the actual PHIs
2176 : to more easily support peeling and epilogue vectorization. */
2177 204288 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2178 204288 : if (!nested_in_vect_loop_p (loop, stmt_info))
2179 203459 : skip_args[loop_preheader_edge (loop)->dest_idx] = true;
2180 : else
2181 : loop = loop->inner;
2182 204288 : skip_args[loop_latch_edge (loop)->dest_idx] = true;
2183 : }
2184 108437 : else if (def_type == vect_reduction_def
2185 : || def_type == vect_double_reduction_def
2186 : || def_type == vect_nested_cycle
2187 108437 : || def_type == vect_first_order_recurrence)
2188 : {
2189 : /* Else def types have to match. */
2190 : stmt_vec_info other_info;
2191 : bool all_same = true;
2192 245433 : FOR_EACH_VEC_ELT (stmts, i, other_info)
2193 : {
2194 138281 : if (STMT_VINFO_DEF_TYPE (other_info) != def_type)
2195 6105041 : return NULL;
2196 138275 : if (other_info != stmt_info)
2197 26269 : all_same = false;
2198 : }
2199 107152 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2200 : /* Reduction initial values are not explicitly represented. */
2201 107152 : if (def_type != vect_first_order_recurrence
2202 107152 : && gimple_bb (stmt_info->stmt) == loop->header)
2203 103974 : skip_args[loop_preheader_edge (loop)->dest_idx] = true;
2204 : /* Reduction chain backedge defs are filled manually.
2205 : ??? Need a better way to identify a SLP reduction chain PHI.
2206 : Or a better overall way to SLP match those. */
2207 107152 : if (stmts.length () > 1
2208 107152 : && all_same && def_type == vect_reduction_def)
2209 2416 : skip_args[loop_latch_edge (loop)->dest_idx] = true;
2210 : }
2211 1279 : else if (def_type != vect_internal_def)
2212 : return NULL;
2213 : }
2214 :
2215 :
2216 6100601 : bool two_operators = false;
2217 6100601 : unsigned char *swap = XALLOCAVEC (unsigned char, group_size);
2218 6100601 : tree vectype = NULL_TREE;
2219 6100601 : if (!vect_build_slp_tree_1 (vinfo, swap, stmts, matches, &two_operators,
2220 : &vectype))
2221 : return NULL;
2222 :
2223 4594854 : bool soft_fail
2224 4594854 : = (is_a <bb_vec_info> (vinfo)
2225 4594854 : && !multiple_p (group_size, TYPE_VECTOR_SUBPARTS (vectype)));
2226 :
2227 : /* If the SLP node is a load, terminate the recursion unless masked. */
2228 4594854 : if (STMT_VINFO_DATA_REF (stmt_info)
2229 2229943 : && DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
2230 : {
2231 952171 : if (soft_fail)
2232 : {
2233 944 : force_split_matches (matches, group_size,
2234 944 : TYPE_VECTOR_SUBPARTS (vectype).to_constant ());
2235 944 : return NULL;
2236 : }
2237 951227 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
2238 : gcc_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)));
2239 : else
2240 : {
2241 931879 : (*tree_size)++;
2242 931879 : node = vect_create_new_slp_node (node, stmts, 0);
2243 931879 : SLP_TREE_VECTYPE (node) = vectype;
2244 : /* And compute the load permutation. Whether it is actually
2245 : a permutation depends on the unrolling factor which is
2246 : decided later. */
2247 931879 : vec<unsigned> load_permutation;
2248 931879 : int j;
2249 931879 : stmt_vec_info load_info;
2250 931879 : load_permutation.create (group_size);
2251 931879 : stmt_vec_info first_stmt_info
2252 931879 : = STMT_VINFO_GROUPED_ACCESS (stmt_info)
2253 931879 : ? DR_GROUP_FIRST_ELEMENT (stmt_info) : stmt_info;
2254 931879 : bool any_permute = false;
2255 2255230 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
2256 : {
2257 1323351 : int load_place;
2258 1323351 : if (! load_info)
2259 : {
2260 40688 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2261 : load_place = j;
2262 : else
2263 : load_place = 0;
2264 : }
2265 1282663 : else if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2266 750963 : load_place = vect_get_place_in_interleaving_chain
2267 750963 : (load_info, first_stmt_info);
2268 : else
2269 : /* Recognize the splat case as { 0, 0, ... } but make
2270 : sure to use the appropriate refs for collections
2271 : of invariant refs. */
2272 531700 : load_place = (load_info == stmt_info) ? 0 : j;
2273 791904 : gcc_assert (load_place != -1);
2274 1323351 : any_permute |= load_place != j;
2275 1323351 : load_permutation.quick_push (load_place);
2276 : }
2277 :
2278 931879 : if (gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt))
2279 : {
2280 3353 : gcc_assert (gimple_call_internal_p (stmt, IFN_MASK_LOAD));
2281 3353 : bool has_gaps = false;
2282 3353 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2283 189 : for (stmt_vec_info si = DR_GROUP_NEXT_ELEMENT (first_stmt_info);
2284 846 : si; si = DR_GROUP_NEXT_ELEMENT (si))
2285 657 : if (DR_GROUP_GAP (si) != 1)
2286 80 : has_gaps = true;
2287 : /* We cannot handle permuted masked loads directly, see
2288 : PR114375. We cannot handle strided masked loads or masked
2289 : loads with gaps unless the mask is uniform. */
2290 3353 : if ((STMT_VINFO_GROUPED_ACCESS (stmt_info)
2291 189 : && (DR_GROUP_GAP (first_stmt_info) != 0
2292 129 : || (has_gaps
2293 35 : && STMT_VINFO_SLP_VECT_ONLY (first_stmt_info))))
2294 6626 : || STMT_VINFO_STRIDED_P (stmt_info))
2295 : {
2296 108 : load_permutation.release ();
2297 108 : matches[0] = -1;
2298 928673 : return NULL;
2299 : }
2300 :
2301 : /* For permuted masked loads do an unpermuted masked load of
2302 : the whole group followed by a SLP permute node. */
2303 3245 : if (any_permute
2304 3245 : || (STMT_VINFO_GROUPED_ACCESS (stmt_info)
2305 83 : && DR_GROUP_SIZE (first_stmt_info) != group_size))
2306 : {
2307 : /* Discover the whole unpermuted load. */
2308 39 : vec<stmt_vec_info> stmts2;
2309 39 : unsigned dr_group_size = STMT_VINFO_GROUPED_ACCESS (stmt_info)
2310 68 : ? DR_GROUP_SIZE (first_stmt_info) : 1;
2311 39 : stmts2.create (dr_group_size);
2312 39 : stmts2.quick_grow_cleared (dr_group_size);
2313 39 : unsigned i = 0;
2314 39 : for (stmt_vec_info si = first_stmt_info;
2315 464 : si; si = DR_GROUP_NEXT_ELEMENT (si))
2316 : {
2317 425 : if (si != first_stmt_info)
2318 1586 : for (unsigned k = 1; k < DR_GROUP_GAP (si); ++k)
2319 1200 : stmts2[i++] = NULL;
2320 425 : stmts2[i++] = si;
2321 : }
2322 39 : match_elt_t *matches2
2323 39 : = XALLOCAVEC (match_elt_t, dr_group_size);
2324 39 : slp_tree unperm_load
2325 39 : = vect_build_slp_tree (vinfo, stmts2, matches2, limit,
2326 39 : &this_tree_size, bst_map);
2327 : /* When we are able to do the full masked load emit that
2328 : followed by 'node' being the desired final permutation. */
2329 39 : if (unperm_load)
2330 : {
2331 16 : gcc_assert
2332 : (!SLP_TREE_LOAD_PERMUTATION (unperm_load).exists ());
2333 16 : lane_permutation_t lperm;
2334 16 : lperm.create (group_size);
2335 72 : for (unsigned j = 0; j < load_permutation.length (); ++j)
2336 40 : lperm.quick_push
2337 40 : (std::make_pair (0, load_permutation[j]));
2338 16 : SLP_TREE_CODE (node) = VEC_PERM_EXPR;
2339 16 : SLP_TREE_CHILDREN (node).safe_push (unperm_load);
2340 16 : SLP_TREE_LANE_PERMUTATION (node) = lperm;
2341 16 : SLP_TREE_REPRESENTATIVE (node) = NULL;
2342 16 : load_permutation.release ();
2343 16 : return node;
2344 : }
2345 23 : stmts2.release ();
2346 23 : load_permutation.release ();
2347 23 : matches[0] = -1;
2348 23 : return NULL;
2349 : }
2350 3206 : load_permutation.release ();
2351 : }
2352 : else
2353 : {
2354 928526 : if (!any_permute
2355 798604 : && STMT_VINFO_GROUPED_ACCESS (stmt_info)
2356 1221631 : && group_size == DR_GROUP_SIZE (first_stmt_info))
2357 128893 : load_permutation.release ();
2358 928526 : SLP_TREE_LOAD_PERMUTATION (node) = load_permutation;
2359 928526 : return node;
2360 : }
2361 : }
2362 : }
2363 3642683 : else if (gimple_assign_single_p (stmt_info->stmt)
2364 2562094 : && !gimple_vuse (stmt_info->stmt)
2365 3651469 : && gimple_assign_rhs_code (stmt_info->stmt) == BIT_FIELD_REF)
2366 : {
2367 3184 : if (soft_fail)
2368 : {
2369 16 : force_split_matches (matches, group_size,
2370 16 : TYPE_VECTOR_SUBPARTS (vectype).to_constant ());
2371 16 : return NULL;
2372 : }
2373 : /* vect_build_slp_tree_2 determined all BIT_FIELD_REFs reference
2374 : the same SSA name vector of a compatible type to vectype. */
2375 3168 : vec<std::pair<unsigned, unsigned> > lperm = vNULL;
2376 3168 : tree vec = TREE_OPERAND (gimple_assign_rhs1 (stmt_info->stmt), 0);
2377 3168 : stmt_vec_info estmt_info;
2378 11072 : FOR_EACH_VEC_ELT (stmts, i, estmt_info)
2379 : {
2380 8051 : gassign *estmt = as_a <gassign *> (estmt_info->stmt);
2381 8051 : tree bfref = gimple_assign_rhs1 (estmt);
2382 8051 : HOST_WIDE_INT lane;
2383 8051 : if (!known_eq (bit_field_size (bfref),
2384 : tree_to_poly_uint64 (TYPE_SIZE (TREE_TYPE (vectype))))
2385 15955 : || !constant_multiple_p (bit_field_offset (bfref),
2386 8051 : bit_field_size (bfref), &lane))
2387 : {
2388 147 : lperm.release ();
2389 147 : matches[0] = -1;
2390 147 : return NULL;
2391 : }
2392 7904 : lperm.safe_push (std::make_pair (0, (unsigned)lane));
2393 : }
2394 3021 : slp_tree vnode = vect_create_new_slp_node (vNULL);
2395 3021 : if (operand_equal_p (TYPE_SIZE (vectype), TYPE_SIZE (TREE_TYPE (vec))))
2396 : /* ??? We record vectype here but we hide eventually necessary
2397 : punning and instead rely on code generation to materialize
2398 : VIEW_CONVERT_EXPRs as necessary. We instead should make
2399 : this explicit somehow. */
2400 1325 : SLP_TREE_VECTYPE (vnode) = vectype;
2401 : else
2402 : {
2403 : /* For different size but compatible elements we can still
2404 : use VEC_PERM_EXPR without punning. */
2405 1696 : gcc_assert (VECTOR_TYPE_P (TREE_TYPE (vec))
2406 : && types_compatible_p (TREE_TYPE (vectype),
2407 : TREE_TYPE (TREE_TYPE (vec))));
2408 1696 : SLP_TREE_VECTYPE (vnode) = TREE_TYPE (vec);
2409 : }
2410 3021 : auto nunits = TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (vnode));
2411 3021 : unsigned HOST_WIDE_INT const_nunits;
2412 3021 : if (nunits.is_constant (&const_nunits))
2413 3021 : SLP_TREE_LANES (vnode) = const_nunits;
2414 3021 : SLP_TREE_VEC_DEFS (vnode).safe_push (vec);
2415 : /* We are always building a permutation node even if it is an identity
2416 : permute to shield the rest of the vectorizer from the odd node
2417 : representing an actual vector without any scalar ops.
2418 : ??? We could hide it completely with making the permute node
2419 : external? */
2420 3021 : node = vect_create_new_slp_node (node, stmts, 1);
2421 3021 : SLP_TREE_CODE (node) = VEC_PERM_EXPR;
2422 3021 : SLP_TREE_LANE_PERMUTATION (node) = lperm;
2423 3021 : SLP_TREE_VECTYPE (node) = vectype;
2424 3021 : SLP_TREE_CHILDREN (node).quick_push (vnode);
2425 3021 : SLP_TREE_REPRESENTATIVE (node) = NULL;
2426 3021 : return node;
2427 : }
2428 : /* When discovery reaches an associatable operation see whether we can
2429 : improve that to match up lanes in a way superior to the operand
2430 : swapping code which at most looks at two defs.
2431 : ??? For BB vectorization we cannot do the brute-force search
2432 : for matching as we can succeed by means of builds from scalars
2433 : and have no good way to "cost" one build against another. */
2434 3639499 : else if (is_a <loop_vec_info> (vinfo)
2435 : /* Do not bother for single-lane SLP. */
2436 2064970 : && group_size > 1
2437 : /* ??? We don't handle !vect_internal_def defs below. */
2438 114225 : && STMT_VINFO_DEF_TYPE (stmt_info) == vect_internal_def
2439 : /* ??? Do not associate a reduction, this will wreck REDUC_IDX
2440 : mapping as long as that exists on the stmt_info level. */
2441 88545 : && STMT_VINFO_REDUC_IDX (stmt_info) == -1
2442 79918 : && is_gimple_assign (stmt_info->stmt)
2443 79593 : && (associative_tree_code (gimple_assign_rhs_code (stmt_info->stmt))
2444 52647 : || gimple_assign_rhs_code (stmt_info->stmt) == MINUS_EXPR)
2445 3668390 : && ((FLOAT_TYPE_P (vectype) && flag_associative_math)
2446 16624 : || (INTEGRAL_TYPE_P (TREE_TYPE (vectype))
2447 14085 : && TYPE_OVERFLOW_WRAPS (TREE_TYPE (vectype)))))
2448 : {
2449 : /* See if we have a chain of (mixed) adds or subtracts or other
2450 : associatable ops. */
2451 21756 : enum tree_code code = gimple_assign_rhs_code (stmt_info->stmt);
2452 21756 : if (code == MINUS_EXPR)
2453 842 : code = PLUS_EXPR;
2454 21756 : stmt_vec_info other_op_stmt_info = NULL;
2455 21756 : stmt_vec_info op_stmt_info = NULL;
2456 21756 : unsigned chain_len = 0;
2457 21756 : auto_vec<chain_op_t> chain;
2458 21756 : auto_vec<std::pair<tree_code, gimple *> > worklist;
2459 21756 : auto_vec<vec<chain_op_t> > chains (group_size);
2460 21756 : auto_vec<slp_tree, 4> children;
2461 21756 : bool hard_fail = true;
2462 22739 : for (unsigned lane = 0; lane < group_size; ++lane)
2463 : {
2464 22413 : if (!stmts[lane])
2465 : {
2466 : /* ??? Below we require lane zero is present. */
2467 0 : if (lane == 0)
2468 : {
2469 : hard_fail = false;
2470 21430 : break;
2471 : }
2472 0 : chains.quick_push (vNULL);
2473 0 : continue;
2474 : }
2475 : /* For each lane linearize the addition/subtraction (or other
2476 : uniform associatable operation) expression tree. */
2477 22413 : gimple *op_stmt = NULL, *other_op_stmt = NULL;
2478 22413 : vect_slp_linearize_chain (vinfo, worklist, chain, code,
2479 22413 : stmts[lane]->stmt, op_stmt, other_op_stmt,
2480 : NULL);
2481 22413 : if (!op_stmt_info && op_stmt)
2482 21105 : op_stmt_info = vinfo->lookup_stmt (op_stmt);
2483 22413 : if (!other_op_stmt_info && other_op_stmt)
2484 878 : other_op_stmt_info = vinfo->lookup_stmt (other_op_stmt);
2485 22413 : if (chain.length () == 2)
2486 : {
2487 : /* In a chain of just two elements resort to the regular
2488 : operand swapping scheme. Likewise if we run into a
2489 : length mismatch process regularly as well as we did not
2490 : process the other lanes we cannot report a good hint what
2491 : lanes to try swapping in the parent. */
2492 : hard_fail = false;
2493 : break;
2494 : }
2495 986 : else if (chain_len == 0)
2496 366 : chain_len = chain.length ();
2497 1240 : else if (chain.length () != chain_len)
2498 : {
2499 : /* ??? Here we could slip in magic to compensate with
2500 : neutral operands. */
2501 3 : matches[lane] = -1;
2502 3 : if (lane != group_size - 1)
2503 3 : matches[0] = -1;
2504 : break;
2505 : }
2506 983 : chains.quick_push (chain.copy ());
2507 983 : chain.truncate (0);
2508 : }
2509 43512 : if (chains.length () == group_size)
2510 : {
2511 : /* We cannot yet use SLP_TREE_CODE to communicate the operation. */
2512 326 : if (!op_stmt_info)
2513 : {
2514 3 : hard_fail = false;
2515 3 : goto out;
2516 : }
2517 : /* Now we have a set of chains with the same length. */
2518 : /* 1. pre-sort according to def_type and operation. */
2519 1194 : for (unsigned lane = 0; lane < group_size; ++lane)
2520 1742 : chains[lane].stablesort (dt_sort_cmp, vinfo);
2521 323 : if (dump_enabled_p ())
2522 : {
2523 155 : dump_printf_loc (MSG_NOTE, vect_location,
2524 : "pre-sorted chains of %s\n",
2525 : get_tree_code_name (code));
2526 655 : for (unsigned lane = 0; lane < group_size; ++lane)
2527 : {
2528 500 : if (!stmts[lane])
2529 0 : dump_printf (MSG_NOTE, "--");
2530 : else
2531 2250 : for (unsigned opnum = 0; opnum < chain_len; ++opnum)
2532 3500 : dump_printf (MSG_NOTE, "%s %T ",
2533 1750 : get_tree_code_name (chains[lane][opnum].code),
2534 1750 : chains[lane][opnum].op);
2535 500 : dump_printf (MSG_NOTE, "\n");
2536 : }
2537 : }
2538 : /* 2. try to build children nodes, associating as necessary. */
2539 : /* 2a. prepare and perform early checks to avoid eating into
2540 : discovery limit unnecessarily. */
2541 323 : vect_def_type *dts = XALLOCAVEC (vect_def_type, chain_len);
2542 1361 : for (unsigned n = 0; n < chain_len; ++n)
2543 : {
2544 1038 : vect_def_type dt = chains[0][n].dt;
2545 1038 : unsigned lane;
2546 3949 : for (lane = 0; lane < group_size; ++lane)
2547 5822 : if (stmts[lane] && chains[lane][n].dt != dt)
2548 : {
2549 0 : if (dt == vect_constant_def
2550 0 : && chains[lane][n].dt == vect_external_def)
2551 : dt = vect_external_def;
2552 0 : else if (dt == vect_external_def
2553 0 : && chains[lane][n].dt == vect_constant_def)
2554 : ;
2555 : else
2556 : break;
2557 : }
2558 1038 : if (lane != group_size)
2559 : {
2560 0 : if (dump_enabled_p ())
2561 0 : dump_printf_loc (MSG_NOTE, vect_location,
2562 : "giving up on chain due to mismatched "
2563 : "def types\n");
2564 0 : matches[lane] = -1;
2565 0 : if (lane != group_size - 1)
2566 0 : matches[0] = -1;
2567 0 : goto out;
2568 : }
2569 1038 : dts[n] = dt;
2570 1038 : if (dt == vect_constant_def
2571 1038 : || dt == vect_external_def)
2572 : {
2573 : /* Check whether we can build the invariant. If we can't
2574 : we never will be able to. */
2575 77 : tree type = TREE_TYPE (chains[0][n].op);
2576 1038 : if (!GET_MODE_SIZE (vinfo->vector_mode).is_constant ()
2577 : && (TREE_CODE (type) == BOOLEAN_TYPE
2578 : || !can_duplicate_and_interleave_p (vinfo, group_size,
2579 : type)))
2580 : {
2581 : matches[0] = -1;
2582 : goto out;
2583 : }
2584 : }
2585 961 : else if (dt != vect_internal_def)
2586 : {
2587 : /* Not sure, we might need sth special.
2588 : gcc.dg/vect/pr96854.c,
2589 : gfortran.dg/vect/fast-math-pr37021.f90
2590 : and gfortran.dg/vect/pr61171.f trigger. */
2591 : /* Soft-fail for now. */
2592 0 : hard_fail = false;
2593 0 : goto out;
2594 : }
2595 : }
2596 : /* 2b. do the actual build. */
2597 1307 : for (unsigned n = 0; n < chain_len; ++n)
2598 : {
2599 1003 : vect_def_type dt = dts[n];
2600 1003 : unsigned lane;
2601 1003 : if (dt == vect_constant_def
2602 1003 : || dt == vect_external_def)
2603 : {
2604 77 : vec<tree> ops;
2605 77 : ops.create (group_size);
2606 397 : for (lane = 0; lane < group_size; ++lane)
2607 243 : if (stmts[lane])
2608 243 : ops.quick_push (chains[lane][n].op);
2609 : else
2610 0 : ops.quick_push (NULL_TREE);
2611 77 : slp_tree child = vect_create_new_slp_node (ops);
2612 77 : SLP_TREE_DEF_TYPE (child) = dt;
2613 77 : children.safe_push (child);
2614 : }
2615 : else
2616 : {
2617 926 : vec<stmt_vec_info> op_stmts;
2618 926 : op_stmts.create (group_size);
2619 926 : slp_tree child = NULL;
2620 : /* Brute-force our way. We have to consider a lane
2621 : failing after fixing an earlier fail up in the
2622 : SLP discovery recursion. So track the current
2623 : permute per lane. */
2624 926 : unsigned *perms = XALLOCAVEC (unsigned, group_size);
2625 926 : memset (perms, 0, sizeof (unsigned) * group_size);
2626 1022 : do
2627 : {
2628 1022 : op_stmts.truncate (0);
2629 4900 : for (lane = 0; lane < group_size; ++lane)
2630 2856 : if (stmts[lane])
2631 2856 : op_stmts.quick_push
2632 2856 : (vinfo->lookup_def (chains[lane][n].op));
2633 : else
2634 0 : op_stmts.quick_push (NULL);
2635 1022 : child = vect_build_slp_tree (vinfo, op_stmts,
2636 : matches, limit,
2637 : &this_tree_size, bst_map);
2638 : /* ??? We're likely getting too many fatal mismatches
2639 : here so maybe we want to ignore them (but then we
2640 : have no idea which lanes fatally mismatched). */
2641 : /* ??? Revisit this with matches[] improvements. */
2642 1022 : if (child || matches[0] != 0)
2643 : break;
2644 : /* Swap another lane we have not yet matched up into
2645 : lanes that did not match. If we run out of
2646 : permute possibilities for a lane terminate the
2647 : search. */
2648 291 : bool term = false;
2649 291 : for (lane = 1; lane < group_size; ++lane)
2650 195 : if (matches[lane] != 0)
2651 : {
2652 167 : if (n + perms[lane] + 1 == chain_len)
2653 : {
2654 : term = true;
2655 : break;
2656 : }
2657 148 : if (dump_enabled_p ())
2658 115 : dump_printf_loc (MSG_NOTE, vect_location,
2659 : "swapping operand %d and %d "
2660 : "of lane %d\n",
2661 : n, n + perms[lane] + 1, lane);
2662 296 : std::swap (chains[lane][n],
2663 148 : chains[lane][n + perms[lane] + 1]);
2664 148 : perms[lane]++;
2665 : }
2666 115 : if (term)
2667 : break;
2668 : }
2669 : while (1);
2670 926 : if (!child)
2671 : {
2672 19 : if (dump_enabled_p ())
2673 18 : dump_printf_loc (MSG_NOTE, vect_location,
2674 : "failed to match up op %d\n", n);
2675 19 : op_stmts.release ();
2676 19 : if (lane != group_size - 1)
2677 9 : matches[0] = -1;
2678 : else
2679 10 : matches[lane] = -1;
2680 19 : goto out;
2681 : }
2682 907 : if (dump_enabled_p ())
2683 : {
2684 409 : dump_printf_loc (MSG_NOTE, vect_location,
2685 : "matched up op %d to\n", n);
2686 409 : vect_print_slp_tree (MSG_NOTE, vect_location, child);
2687 : }
2688 907 : children.safe_push (child);
2689 : }
2690 : }
2691 : /* 3. build SLP nodes to combine the chain. */
2692 1102 : for (unsigned lane = 0; lane < group_size; ++lane)
2693 1608 : if (stmts[lane] && chains[lane][0].code != code)
2694 : {
2695 : /* See if there's any alternate all-PLUS entry. */
2696 : unsigned n;
2697 6 : for (n = 1; n < chain_len; ++n)
2698 : {
2699 30 : for (lane = 0; lane < group_size; ++lane)
2700 48 : if (stmts[lane] && chains[lane][n].code != code)
2701 : break;
2702 6 : if (lane == group_size)
2703 : break;
2704 : }
2705 6 : if (n != chain_len)
2706 : {
2707 : /* Swap that in at first position. */
2708 6 : std::swap (children[0], children[n]);
2709 30 : for (lane = 0; lane < group_size; ++lane)
2710 24 : if (stmts[lane])
2711 24 : std::swap (chains[lane][0], chains[lane][n]);
2712 : }
2713 : else
2714 : {
2715 : /* ??? When this triggers and we end up with two
2716 : vect_constant/external_def up-front things break (ICE)
2717 : spectacularly finding an insertion place for the
2718 : all-constant op. We should have a fully
2719 : vect_internal_def operand though(?) so we can swap
2720 : that into first place and then prepend the all-zero
2721 : constant. */
2722 0 : if (dump_enabled_p ())
2723 0 : dump_printf_loc (MSG_NOTE, vect_location,
2724 : "inserting constant zero to compensate "
2725 : "for (partially) negated first "
2726 : "operand\n");
2727 0 : chain_len++;
2728 0 : for (lane = 0; lane < group_size; ++lane)
2729 0 : if (stmts[lane])
2730 0 : chains[lane].safe_insert
2731 0 : (0, chain_op_t (code, vect_constant_def, NULL_TREE));
2732 0 : vec<tree> zero_ops;
2733 0 : zero_ops.create (group_size);
2734 0 : zero_ops.quick_push (build_zero_cst (TREE_TYPE (vectype)));
2735 0 : for (lane = 1; lane < group_size; ++lane)
2736 0 : if (stmts[lane])
2737 0 : zero_ops.quick_push (zero_ops[0]);
2738 : else
2739 0 : zero_ops.quick_push (NULL_TREE);
2740 0 : slp_tree zero = vect_create_new_slp_node (zero_ops);
2741 0 : SLP_TREE_DEF_TYPE (zero) = vect_constant_def;
2742 0 : children.safe_insert (0, zero);
2743 : }
2744 : break;
2745 : }
2746 979 : for (unsigned i = 1; i < children.length (); ++i)
2747 : {
2748 675 : slp_tree op0 = children[i - 1];
2749 675 : slp_tree op1 = children[i];
2750 675 : bool this_two_op = false;
2751 2391 : for (unsigned lane = 0; lane < group_size; ++lane)
2752 3756 : if (stmts[lane] && chains[lane][i].code != chains[0][i].code)
2753 : {
2754 : this_two_op = true;
2755 : break;
2756 : }
2757 675 : slp_tree child;
2758 675 : if (i == children.length () - 1)
2759 304 : child = vect_create_new_slp_node (node, stmts, 2);
2760 : else
2761 371 : child = vect_create_new_slp_node (2, ERROR_MARK);
2762 675 : if (this_two_op)
2763 : {
2764 162 : vec<std::pair<unsigned, unsigned> > lperm;
2765 162 : lperm.create (group_size);
2766 744 : for (unsigned lane = 0; lane < group_size; ++lane)
2767 840 : lperm.quick_push (std::make_pair
2768 420 : (chains[lane][i].code != chains[0][i].code, lane));
2769 324 : vect_slp_build_two_operator_nodes (child, vectype, op0, op1,
2770 162 : (chains[0][i].code == code
2771 : ? op_stmt_info
2772 : : other_op_stmt_info),
2773 162 : (chains[0][i].code == code
2774 : ? other_op_stmt_info
2775 : : op_stmt_info),
2776 : lperm);
2777 : }
2778 : else
2779 : {
2780 513 : SLP_TREE_DEF_TYPE (child) = vect_internal_def;
2781 513 : SLP_TREE_VECTYPE (child) = vectype;
2782 513 : SLP_TREE_LANES (child) = group_size;
2783 513 : SLP_TREE_CHILDREN (child).quick_push (op0);
2784 513 : SLP_TREE_CHILDREN (child).quick_push (op1);
2785 513 : SLP_TREE_REPRESENTATIVE (child)
2786 1026 : = (chains[0][i].code == code
2787 513 : ? op_stmt_info : other_op_stmt_info);
2788 : }
2789 675 : children[i] = child;
2790 : }
2791 304 : *tree_size += this_tree_size + 1;
2792 1452 : while (!chains.is_empty ())
2793 822 : chains.pop ().release ();
2794 : return node;
2795 : }
2796 21430 : out:
2797 21452 : if (dump_enabled_p ())
2798 2825 : dump_printf_loc (MSG_NOTE, vect_location,
2799 : "failed to line up SLP graph by re-associating "
2800 : "operations in lanes%s\n",
2801 : !hard_fail ? " trying regular discovery" : "");
2802 21457 : while (!children.is_empty ())
2803 5 : vect_free_slp_tree (children.pop ());
2804 21613 : while (!chains.is_empty ())
2805 161 : chains.pop ().release ();
2806 : /* Hard-fail, otherwise we might run into quadratic processing of the
2807 : chains starting one stmt into the chain again. */
2808 21452 : if (hard_fail)
2809 : return NULL;
2810 : /* Fall thru to normal processing. */
2811 21756 : }
2812 :
2813 : /* Get at the operands, verifying they are compatible. */
2814 3661727 : vec<slp_oprnd_info> oprnds_info = vect_create_oprnd_info (nops, group_size);
2815 3661727 : slp_oprnd_info oprnd_info;
2816 21034910 : FOR_EACH_VEC_ELT (stmts, i, stmt_info)
2817 : {
2818 27425540 : int res = vect_get_and_check_slp_defs (vinfo, vectype,
2819 13712770 : swap[i], skip_args,
2820 : stmts, i, &oprnds_info,
2821 : soft_fail);
2822 13712770 : if (res != 0)
2823 : /* ??? This puts -1 back into matches[] and the cache. */
2824 340569 : matches[(res == -1) ? 0 : i] = -1;
2825 13712770 : if (matches[0] == -1)
2826 : break;
2827 : }
2828 16980417 : for (i = 0; i < group_size; ++i)
2829 13568847 : if (matches[i] != 0)
2830 : {
2831 250157 : vect_free_oprnd_info (oprnds_info);
2832 250157 : return NULL;
2833 : }
2834 3411570 : swap = NULL;
2835 :
2836 : /* Perform delayed soft-failing only here so we can factor in mismatches
2837 : determined by vect_get_and_check_slp_defs. */
2838 3411570 : if (soft_fail)
2839 : {
2840 122617 : force_split_matches (matches, group_size,
2841 122617 : TYPE_VECTOR_SUBPARTS (vectype).to_constant ());
2842 122617 : return NULL;
2843 : }
2844 :
2845 3288953 : bool has_two_operators_perm = false;
2846 16444765 : auto_vec<unsigned> two_op_perm_indices[2];
2847 3288953 : vec<stmt_vec_info> two_op_scalar_stmts[2] = {vNULL, vNULL};
2848 :
2849 3305185 : if (two_operators && oprnds_info.length () == 2 && group_size > 2)
2850 : {
2851 4016 : unsigned idx = 0;
2852 4016 : hash_map<gimple *, unsigned> seen;
2853 4016 : vec<slp_oprnd_info> new_oprnds_info
2854 4016 : = vect_create_oprnd_info (1, group_size);
2855 4016 : bool success = true;
2856 :
2857 4016 : enum tree_code code = ERROR_MARK;
2858 4016 : if (oprnds_info[0]->def_stmts[0]
2859 4016 : && is_a<gassign *> (oprnds_info[0]->def_stmts[0]->stmt))
2860 3956 : code = gimple_assign_rhs_code (oprnds_info[0]->def_stmts[0]->stmt);
2861 4016 : basic_block bb = nullptr;
2862 :
2863 7718 : for (unsigned j = 0; j < group_size; ++j)
2864 : {
2865 18000 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
2866 : {
2867 14298 : stmt_vec_info stmt_info = oprnd_info->def_stmts[j];
2868 14298 : if (!stmt_info
2869 14131 : || !is_a<gassign *> (stmt_info->stmt)
2870 14131 : || gimple_assign_rhs_code (stmt_info->stmt) != code
2871 25063 : || skip_args[i])
2872 : {
2873 : success = false;
2874 3537 : break;
2875 : }
2876 : /* Avoid mixing lanes with defs in different basic-blocks. */
2877 10765 : if (!bb)
2878 4152 : bb = gimple_bb (vect_orig_stmt (stmt_info)->stmt);
2879 8547 : else if (gimple_bb (vect_orig_stmt (stmt_info)->stmt) != bb)
2880 : {
2881 : success = false;
2882 : break;
2883 : }
2884 :
2885 10761 : bool exists;
2886 10761 : unsigned &stmt_idx
2887 10761 : = seen.get_or_insert (stmt_info->stmt, &exists);
2888 :
2889 10761 : if (!exists)
2890 : {
2891 9380 : new_oprnds_info[0]->def_stmts.safe_push (stmt_info);
2892 9380 : new_oprnds_info[0]->ops.safe_push (oprnd_info->ops[j]);
2893 9380 : stmt_idx = idx;
2894 9380 : idx++;
2895 : }
2896 :
2897 10761 : two_op_perm_indices[i].safe_push (stmt_idx);
2898 : }
2899 :
2900 7239 : if (!success)
2901 : break;
2902 : }
2903 :
2904 4016 : if (success && idx == group_size)
2905 : {
2906 98 : if (dump_enabled_p ())
2907 : {
2908 0 : dump_printf_loc (MSG_NOTE, vect_location,
2909 : "Replace two_operators operands:\n");
2910 :
2911 0 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
2912 : {
2913 0 : dump_printf_loc (MSG_NOTE, vect_location,
2914 : "Operand %u:\n", i);
2915 0 : for (unsigned j = 0; j < group_size; j++)
2916 0 : dump_printf_loc (MSG_NOTE, vect_location, "\tstmt %u %G",
2917 0 : j, oprnd_info->def_stmts[j]->stmt);
2918 : }
2919 :
2920 0 : dump_printf_loc (MSG_NOTE, vect_location,
2921 : "With a single operand:\n");
2922 0 : for (unsigned j = 0; j < group_size; j++)
2923 0 : dump_printf_loc (MSG_NOTE, vect_location, "\tstmt %u %G",
2924 0 : j, new_oprnds_info[0]->def_stmts[j]->stmt);
2925 : }
2926 :
2927 98 : two_op_scalar_stmts[0].safe_splice (oprnds_info[0]->def_stmts);
2928 98 : two_op_scalar_stmts[1].safe_splice (oprnds_info[1]->def_stmts);
2929 :
2930 98 : new_oprnds_info[0]->first_op_type = oprnds_info[0]->first_op_type;
2931 98 : new_oprnds_info[0]->first_dt = oprnds_info[0]->first_dt;
2932 98 : new_oprnds_info[0]->any_pattern = oprnds_info[0]->any_pattern;
2933 98 : new_oprnds_info[0]->first_gs_p = oprnds_info[0]->first_gs_p;
2934 98 : new_oprnds_info[0]->first_gs_info = oprnds_info[0]->first_gs_info;
2935 :
2936 98 : vect_free_oprnd_info (oprnds_info);
2937 98 : oprnds_info = new_oprnds_info;
2938 98 : nops = 1;
2939 98 : has_two_operators_perm = true;
2940 : }
2941 : else
2942 3918 : vect_free_oprnd_info (new_oprnds_info);
2943 4016 : }
2944 :
2945 6577906 : auto_vec<slp_tree, 4> children;
2946 :
2947 3288953 : stmt_info = stmts[0];
2948 :
2949 3288953 : int reduc_idx = -1;
2950 3288953 : int gs_scale = 0;
2951 3288953 : tree gs_base = NULL_TREE;
2952 :
2953 : /* Create SLP_TREE nodes for the definition node/s. */
2954 8461556 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
2955 : {
2956 5284339 : slp_tree child = nullptr;
2957 5284339 : unsigned int j;
2958 5284339 : unsigned old_swap_distance;
2959 5284339 : bool can_swap;
2960 5284339 : bool can_swap_nonmatching;
2961 5284339 : bool *stmt_can_swap;
2962 :
2963 : /* We're skipping certain operands from processing, for example
2964 : outer loop reduction initial defs. */
2965 5284339 : if (skip_args[i])
2966 : {
2967 514137 : children.safe_push (NULL);
2968 5172603 : continue;
2969 : }
2970 :
2971 4770202 : if (oprnd_info->first_dt == vect_uninitialized_def)
2972 : {
2973 : /* COND_EXPR have one too many eventually if the condition
2974 : is a SSA name. */
2975 0 : gcc_assert (i == 3 && nops == 4);
2976 0 : continue;
2977 : }
2978 :
2979 4770202 : if (oprnd_info->first_gs_p)
2980 : {
2981 23523 : gs_scale = oprnd_info->first_gs_info.scale;
2982 23523 : gs_base = oprnd_info->first_gs_info.base;
2983 : }
2984 :
2985 4770202 : if (is_a <bb_vec_info> (vinfo)
2986 1694936 : && oprnd_info->first_dt == vect_internal_def
2987 5674490 : && !oprnd_info->any_pattern)
2988 : {
2989 : /* For BB vectorization, if all defs are the same do not
2990 : bother to continue the build along the single-lane
2991 : graph but use a splat of the scalar value. */
2992 852461 : stmt_vec_info first_def = oprnd_info->def_stmts[0];
2993 920019 : for (j = 1; j < group_size; ++j)
2994 869481 : if (oprnd_info->def_stmts[j] != first_def)
2995 : break;
2996 852461 : if (j == group_size
2997 : /* But avoid doing this for loads where we may be
2998 : able to CSE things, unless the stmt is not
2999 : vectorizable. */
3000 852461 : && (!STMT_VINFO_VECTORIZABLE (first_def)
3001 60644 : || !gimple_vuse (first_def->stmt)))
3002 : {
3003 40924 : if (dump_enabled_p ())
3004 125 : dump_printf_loc (MSG_NOTE, vect_location,
3005 : "Using a splat of the uniform operand %G",
3006 : first_def->stmt);
3007 40924 : oprnd_info->first_dt = vect_external_def;
3008 : }
3009 : }
3010 :
3011 4770202 : if (oprnd_info->first_dt == vect_external_def
3012 4770202 : || oprnd_info->first_dt == vect_constant_def)
3013 : {
3014 1573311 : if (!GET_MODE_SIZE (vinfo->vector_mode).is_constant ())
3015 : {
3016 : tree op0;
3017 : tree uniform_val = op0 = oprnd_info->ops[0];
3018 : for (j = 1; j < oprnd_info->ops.length (); ++j)
3019 : if (oprnd_info->ops[j]
3020 : && !operand_equal_p (uniform_val, oprnd_info->ops[j]))
3021 : {
3022 : uniform_val = NULL_TREE;
3023 : break;
3024 : }
3025 : if (!uniform_val
3026 : && !can_duplicate_and_interleave_p (vinfo,
3027 : oprnd_info->ops.length (),
3028 : TREE_TYPE (op0)))
3029 : {
3030 : matches[j] = -1;
3031 : if (dump_enabled_p ())
3032 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3033 : "Build SLP failed: invalid type of def "
3034 : "for variable-length SLP %T\n", op0);
3035 : goto fail;
3036 : }
3037 : }
3038 1573311 : slp_tree invnode = vect_create_new_slp_node (oprnd_info->ops);
3039 1573311 : SLP_TREE_DEF_TYPE (invnode) = oprnd_info->first_dt;
3040 1573311 : oprnd_info->ops = vNULL;
3041 1573311 : children.safe_push (invnode);
3042 1573311 : continue;
3043 1573311 : }
3044 :
3045 : /* See which SLP operand a reduction chain continues on. We want
3046 : to chain even PHIs but not backedges. */
3047 3196891 : if (STMT_VINFO_REDUC_DEF (oprnd_info->def_stmts[0])
3048 3196891 : || STMT_VINFO_REDUC_IDX (oprnd_info->def_stmts[0]) != -1)
3049 : {
3050 238277 : if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle)
3051 : {
3052 776 : if (oprnd_info->first_dt == vect_double_reduction_def)
3053 388 : reduc_idx = i;
3054 : }
3055 237501 : else if (is_a <gphi *> (stmt_info->stmt)
3056 237501 : && gimple_phi_num_args
3057 101951 : (as_a <gphi *> (stmt_info->stmt)) != 1)
3058 : ;
3059 135943 : else if (STMT_VINFO_REDUC_IDX (stmt_info) == -1
3060 393 : && STMT_VINFO_DEF_TYPE (stmt_info) != vect_double_reduction_def)
3061 : ;
3062 135943 : else if (reduc_idx == -1)
3063 127451 : reduc_idx = i;
3064 : else
3065 : /* For .COND_* reduction operations the else value can be the
3066 : same as one of the operation operands. The other def
3067 : stmts have been moved, so we can't check easily. Check
3068 : it's a call at least. */
3069 8492 : gcc_assert (is_a <gcall *> (stmt_info->stmt));
3070 : }
3071 :
3072 : /* When we have a masked load with uniform mask discover this
3073 : as a single-lane mask with a splat permute. This way we can
3074 : recognize this as a masked load-lane by stripping the splat. */
3075 3196891 : if (is_a <gcall *> (STMT_VINFO_STMT (stmt_info))
3076 57187 : && gimple_call_internal_p (STMT_VINFO_STMT (stmt_info),
3077 : IFN_MASK_LOAD)
3078 4751 : && STMT_VINFO_GROUPED_ACCESS (stmt_info)
3079 3196968 : && ! STMT_VINFO_SLP_VECT_ONLY (DR_GROUP_FIRST_ELEMENT (stmt_info)))
3080 : {
3081 35 : vec<stmt_vec_info> def_stmts2;
3082 35 : def_stmts2.create (1);
3083 35 : def_stmts2.quick_push (oprnd_info->def_stmts[0]);
3084 35 : child = vect_build_slp_tree (vinfo, def_stmts2,
3085 : matches, limit,
3086 : &this_tree_size, bst_map);
3087 35 : if (child)
3088 : {
3089 35 : slp_tree pnode = vect_create_new_slp_node (1, VEC_PERM_EXPR);
3090 35 : SLP_TREE_VECTYPE (pnode) = SLP_TREE_VECTYPE (child);
3091 35 : SLP_TREE_LANES (pnode) = group_size;
3092 35 : SLP_TREE_SCALAR_STMTS (pnode).create (group_size);
3093 35 : SLP_TREE_LANE_PERMUTATION (pnode).create (group_size);
3094 245 : for (unsigned k = 0; k < group_size; ++k)
3095 : {
3096 175 : SLP_TREE_SCALAR_STMTS (pnode)
3097 175 : .quick_push (oprnd_info->def_stmts[0]);
3098 175 : SLP_TREE_LANE_PERMUTATION (pnode)
3099 175 : .quick_push (std::make_pair (0u, 0u));
3100 : }
3101 35 : SLP_TREE_CHILDREN (pnode).quick_push (child);
3102 35 : children.safe_push (pnode);
3103 35 : oprnd_info->def_stmts = vNULL;
3104 35 : continue;
3105 35 : }
3106 : else
3107 0 : def_stmts2.release ();
3108 : }
3109 :
3110 6393712 : can_swap = (i == 0
3111 2361444 : && (nops == 2 || nops == 3)
3112 1530350 : && oprnds_info.length () > 1
3113 1530350 : && oprnds_info[1]->first_dt == vect_internal_def
3114 624038 : && (is_gimple_assign (stmt_info->stmt)
3115 50974 : || is_gimple_call (stmt_info->stmt))
3116 : /* Swapping operands for reductions breaks assumptions
3117 : later on. */
3118 3774713 : && STMT_VINFO_REDUC_IDX (stmt_info) == -1);
3119 3196856 : can_swap_nonmatching = can_swap;
3120 3196856 : stmt_can_swap = NULL;
3121 3196856 : if (can_swap)
3122 : {
3123 522085 : stmt_can_swap = XALLOCAVEC (bool, group_size);
3124 8198588 : for (j = 0; j < group_size; ++j)
3125 : {
3126 7676503 : stmt_can_swap[j] = false;
3127 7676503 : if (!stmts[j])
3128 : /* NULL lanes are gaps and have no stmt to swap. */
3129 0 : stmt_can_swap[j] = true;
3130 7676503 : else if (gassign *stmt = dyn_cast <gassign *> (stmts[j]->stmt))
3131 : {
3132 7671231 : tree_code code = gimple_assign_rhs_code (stmt);
3133 15342462 : stmt_can_swap[j] = (commutative_tree_code (code)
3134 7671231 : || commutative_ternary_tree_code (code));
3135 : }
3136 5272 : else if (gcall *call = dyn_cast <gcall *> (stmts[j]->stmt))
3137 : {
3138 5272 : internal_fn fn = (gimple_call_internal_p (call)
3139 5272 : ? gimple_call_internal_fn (call) : IFN_LAST);
3140 10544 : stmt_can_swap[j] = ((commutative_binary_fn_p (fn)
3141 4724 : || commutative_ternary_fn_p (fn))
3142 5308 : && first_commutative_argument (fn) == 0);
3143 : }
3144 :
3145 7676503 : if (j != 0 && !stmt_can_swap[j])
3146 7676503 : can_swap_nonmatching = false;
3147 : }
3148 : }
3149 :
3150 3196856 : old_swap_distance = least_upthread_swappable_op_distance;
3151 3196856 : if (can_swap_nonmatching)
3152 479777 : least_upthread_swappable_op_distance = 1;
3153 2717079 : else if (least_upthread_swappable_op_distance != -1U)
3154 349940 : least_upthread_swappable_op_distance++;
3155 3196856 : child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts,
3156 : matches, limit,
3157 : &this_tree_size, bst_map);
3158 3196856 : least_upthread_swappable_op_distance = old_swap_distance;
3159 3196856 : if (child != NULL)
3160 : {
3161 2689703 : oprnd_info->def_stmts = vNULL;
3162 2689703 : children.safe_push (child);
3163 2689703 : continue;
3164 : }
3165 :
3166 : /* If the SLP build for operand zero failed and operand zero
3167 : and one can be commuted try that for the scalar stmts
3168 : that failed the match. */
3169 507153 : if (/* A first scalar stmt mismatch signals a fatal mismatch. */
3170 507153 : matches[0] == 0
3171 507153 : && can_swap)
3172 : {
3173 : /* See whether we can swap the matching or the non-matching
3174 : stmt operands. */
3175 : bool swap_not_matching = true;
3176 74220 : do
3177 : {
3178 7114026 : for (j = 0; j < group_size; ++j)
3179 : {
3180 7057902 : if ((matches[j] == 0) != !swap_not_matching)
3181 88512 : continue;
3182 : /* Verify if we can swap operands of this stmt. */
3183 6969390 : if (!stmt_can_swap[j])
3184 : {
3185 18096 : if (!swap_not_matching)
3186 7966 : goto fail;
3187 : swap_not_matching = false;
3188 : break;
3189 : }
3190 : }
3191 : }
3192 66254 : while (j != group_size);
3193 :
3194 : /* Swap mismatched definition stmts. */
3195 56124 : if (dump_enabled_p ())
3196 425 : dump_printf_loc (MSG_NOTE, vect_location,
3197 : "Re-trying with swapped operands of stmts ");
3198 7084743 : for (j = 0; j < group_size; ++j)
3199 7028619 : if ((matches[j] == 0) == !swap_not_matching)
3200 : {
3201 13902180 : std::swap (oprnds_info[0]->def_stmts[j],
3202 6951090 : oprnds_info[1]->def_stmts[j]);
3203 13902180 : std::swap (oprnds_info[0]->ops[j],
3204 6951090 : oprnds_info[1]->ops[j]);
3205 6951090 : if (dump_enabled_p ())
3206 1096 : dump_printf (MSG_NOTE, "%d ", j);
3207 : }
3208 56124 : if (dump_enabled_p ())
3209 425 : dump_printf (MSG_NOTE, "\n");
3210 : /* After swapping some operands we lost track whether an
3211 : operand has any pattern defs so be conservative here. */
3212 109103 : if (oprnds_info[0]->any_pattern || oprnds_info[1]->any_pattern)
3213 4220 : oprnds_info[0]->any_pattern = oprnds_info[1]->any_pattern = true;
3214 : /* And try again with scratch 'matches' ... */
3215 56124 : match_elt_t *tem = XALLOCAVEC (match_elt_t, group_size);
3216 56124 : if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts,
3217 : tem, limit,
3218 : &this_tree_size, bst_map)) != NULL)
3219 : {
3220 7762 : oprnd_info->def_stmts = vNULL;
3221 7762 : children.safe_push (child);
3222 7762 : continue;
3223 : }
3224 : }
3225 499391 : fail:
3226 :
3227 : /* If the SLP build failed and we analyze a basic-block
3228 : simply treat nodes we fail to build as externally defined
3229 : (and thus build vectors from the scalar defs).
3230 : The cost model will reject outright expensive cases.
3231 : ??? This doesn't treat cases where permutation ultimatively
3232 : fails (or we don't try permutation below). Ideally we'd
3233 : even compute a permutation that will end up with the maximum
3234 : SLP tree size... */
3235 499391 : if (is_a <bb_vec_info> (vinfo)
3236 : /* ??? Rejecting patterns this way doesn't work. We'd have to
3237 : do extra work to cancel the pattern so the uses see the
3238 : scalar version. */
3239 : /* Skip building vector operands from scalars while operand
3240 : discovery may still be fixed by retrying with swapped operands. */
3241 440389 : && (least_upthread_swappable_op_distance != 1
3242 : /* A first scalar stmt mismatch signals a fatal mismatch
3243 : that the parent commutative retry cannot recover.
3244 : ??? Possibly revisit this with the matches[] improvements
3245 : as we can have matches[0] == -2 here. */
3246 27612 : || matches[0] != 0)
3247 421394 : && !is_pattern_stmt_p (stmt_info)
3248 893888 : && !oprnd_info->any_pattern)
3249 : {
3250 : /* But if there's a leading vector sized set of matching stmts
3251 : fail here so we can split the group. This matches the condition
3252 : vect_analyze_slp_instance uses. */
3253 : /* ??? We might want to split here and combine the results to support
3254 : multiple vector sizes better. */
3255 611934 : for (j = 0; j < group_size; ++j)
3256 611934 : if (matches[j] != 0)
3257 : break;
3258 394131 : if (!known_ge (j, TYPE_VECTOR_SUBPARTS (vectype))
3259 394097 : && vect_slp_can_convert_to_external (oprnd_info->def_stmts))
3260 : {
3261 387655 : if (dump_enabled_p ())
3262 788 : dump_printf_loc (MSG_NOTE, vect_location,
3263 : "Building vector operands from scalars\n");
3264 387655 : this_tree_size++;
3265 387655 : child = vect_create_new_slp_node (oprnd_info->ops);
3266 387655 : children.safe_push (child);
3267 387655 : oprnd_info->ops = vNULL;
3268 387655 : continue;
3269 : }
3270 : }
3271 :
3272 111736 : gcc_assert (child == NULL);
3273 243800 : FOR_EACH_VEC_ELT (children, j, child)
3274 20328 : if (child)
3275 20328 : vect_free_slp_tree (child);
3276 111736 : vect_free_oprnd_info (oprnds_info);
3277 111736 : return NULL;
3278 : }
3279 :
3280 3177217 : vect_free_oprnd_info (oprnds_info);
3281 :
3282 : /* If we have all children of a child built up from uniform scalars
3283 : or does more than one possibly expensive vector construction then
3284 : just throw that away, causing it built up from scalars.
3285 : The exception is the SLP node for the vector store. */
3286 3177217 : if (is_a <bb_vec_info> (vinfo)
3287 1155187 : && !STMT_VINFO_GROUPED_ACCESS (stmt_info)
3288 : /* ??? Rejecting patterns this way doesn't work. We'd have to
3289 : do extra work to cancel the pattern so the uses see the
3290 : scalar version. */
3291 3657493 : && !is_pattern_stmt_p (stmt_info))
3292 : {
3293 : slp_tree child;
3294 : unsigned j;
3295 : bool all_uniform_p = true;
3296 : unsigned n_vector_builds = 0;
3297 1340792 : FOR_EACH_VEC_ELT (children, j, child)
3298 : {
3299 894020 : if (!child)
3300 : ;
3301 894020 : else if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
3302 : all_uniform_p = false;
3303 639112 : else if (!vect_slp_tree_uniform_p (child))
3304 : {
3305 483396 : all_uniform_p = false;
3306 483396 : if (SLP_TREE_DEF_TYPE (child) == vect_external_def)
3307 442792 : n_vector_builds++;
3308 : }
3309 : }
3310 446772 : if (all_uniform_p
3311 446772 : || n_vector_builds > 1
3312 758926 : || (n_vector_builds == children.length ()
3313 35200 : && is_a <gphi *> (stmt_info->stmt)))
3314 : {
3315 : /* Roll back. */
3316 139528 : matches[0] = -1;
3317 441950 : FOR_EACH_VEC_ELT (children, j, child)
3318 302422 : if (child)
3319 302422 : vect_free_slp_tree (child);
3320 :
3321 139528 : if (dump_enabled_p ())
3322 259 : dump_printf_loc (MSG_NOTE, vect_location,
3323 : "Building parent vector operands from "
3324 : "scalars instead\n");
3325 3288953 : return NULL;
3326 : }
3327 : }
3328 :
3329 3037689 : *tree_size += this_tree_size + 1;
3330 :
3331 3037689 : if (two_operators)
3332 : {
3333 : /* ??? We'd likely want to either cache in bst_map sth like
3334 : { a+b, NULL, a+b, NULL } and { NULL, a-b, NULL, a-b } or
3335 : the true { a+b, a+b, a+b, a+b } ... but there we don't have
3336 : explicit stmts to put in so the keying on 'stmts' doesn't
3337 : work (but we have the same issue with nodes that use 'ops'). */
3338 :
3339 7638 : if (has_two_operators_perm)
3340 : {
3341 46 : slp_tree child = children[0];
3342 46 : children.truncate (0);
3343 184 : for (i = 0; i < 2; i++)
3344 : {
3345 92 : slp_tree pnode
3346 92 : = vect_create_new_slp_node (two_op_scalar_stmts[i], 2);
3347 92 : SLP_TREE_CODE (pnode) = VEC_PERM_EXPR;
3348 92 : SLP_TREE_VECTYPE (pnode) = vectype;
3349 92 : SLP_TREE_CHILDREN (pnode).quick_push (child);
3350 92 : SLP_TREE_CHILDREN (pnode).quick_push (child);
3351 92 : SLP_TREE_REPRESENTATIVE (pnode) = NULL;
3352 92 : lane_permutation_t& perm = SLP_TREE_LANE_PERMUTATION (pnode);
3353 92 : children.safe_push (pnode);
3354 :
3355 808 : for (unsigned j = 0; j < stmts.length (); j++)
3356 624 : perm.safe_push (std::make_pair (0, two_op_perm_indices[i][j]));
3357 : }
3358 :
3359 46 : SLP_TREE_REF_COUNT (child) += 4;
3360 : }
3361 :
3362 7638 : slp_tree one = new _slp_tree;
3363 7638 : slp_tree two = new _slp_tree;
3364 7638 : SLP_TREE_DEF_TYPE (one) = vect_internal_def;
3365 7638 : SLP_TREE_DEF_TYPE (two) = vect_internal_def;
3366 7638 : SLP_TREE_VECTYPE (one) = vectype;
3367 7638 : SLP_TREE_VECTYPE (two) = vectype;
3368 7638 : SLP_TREE_CHILDREN (one).safe_splice (children);
3369 7638 : SLP_TREE_CHILDREN (two).safe_splice (children);
3370 7638 : slp_tree child;
3371 30554 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (two), i, child)
3372 15278 : SLP_TREE_REF_COUNT (child)++;
3373 :
3374 : /* Here we record the original defs since this
3375 : node represents the final lane configuration. */
3376 7638 : node = vect_create_new_slp_node (node, stmts, 2);
3377 7638 : SLP_TREE_VECTYPE (node) = vectype;
3378 7638 : SLP_TREE_CODE (node) = VEC_PERM_EXPR;
3379 7638 : SLP_TREE_CHILDREN (node).quick_push (one);
3380 7638 : SLP_TREE_CHILDREN (node).quick_push (two);
3381 7638 : SLP_TREE_REPRESENTATIVE (node) = NULL;
3382 7638 : enum tree_code code0 = ERROR_MARK;
3383 7638 : enum tree_code ocode = ERROR_MARK;
3384 7638 : if (gassign *stmt = dyn_cast <gassign *> (stmts[0]->stmt))
3385 7636 : code0 = gimple_assign_rhs_code (stmt);
3386 7638 : stmt_vec_info ostmt_info;
3387 7638 : unsigned j = 0;
3388 27309 : FOR_EACH_VEC_ELT (stmts, i, ostmt_info)
3389 : {
3390 19671 : int op = 0;
3391 19671 : if (gassign *ostmt = dyn_cast <gassign *> (ostmt_info->stmt))
3392 : {
3393 19667 : if (gimple_assign_rhs_code (ostmt) != code0)
3394 : {
3395 9855 : ocode = gimple_assign_rhs_code (ostmt);
3396 : op = 1;
3397 : j = i;
3398 : }
3399 : }
3400 : else
3401 : {
3402 8 : if (gimple_call_combined_fn (stmts[0]->stmt)
3403 4 : != gimple_call_combined_fn (ostmt_info->stmt))
3404 : {
3405 2 : op = 1;
3406 2 : j = i;
3407 : }
3408 : }
3409 19671 : SLP_TREE_LANE_PERMUTATION (node).safe_push (std::make_pair (op, i));
3410 : }
3411 7638 : SLP_TREE_CODE (one) = code0;
3412 7638 : SLP_TREE_CODE (two) = ocode;
3413 7638 : SLP_TREE_LANES (one) = stmts.length ();
3414 7638 : SLP_TREE_LANES (two) = stmts.length ();
3415 7638 : SLP_TREE_REPRESENTATIVE (one) = stmts[0];
3416 7638 : SLP_TREE_REPRESENTATIVE (two) = stmts[j];
3417 :
3418 7638 : return node;
3419 : }
3420 :
3421 3030051 : node = vect_create_new_slp_node (node, stmts, nops);
3422 3030051 : SLP_TREE_VECTYPE (node) = vectype;
3423 3030051 : SLP_TREE_CHILDREN (node).splice (children);
3424 3030051 : SLP_TREE_GS_SCALE (node) = gs_scale;
3425 3030051 : SLP_TREE_GS_BASE (node) = gs_base;
3426 3030051 : if (reduc_idx != -1)
3427 : {
3428 119310 : gcc_assert (STMT_VINFO_REDUC_IDX (stmt_info) != -1
3429 : || STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle
3430 : || STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def);
3431 119310 : SLP_TREE_REDUC_IDX (node) = reduc_idx;
3432 119310 : node->cycle_info.id = SLP_TREE_CHILDREN (node)[reduc_idx]->cycle_info.id;
3433 : }
3434 : /* When reaching the reduction PHI, create a vect_reduc_info. */
3435 2910741 : else if ((STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def
3436 2910741 : || STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def)
3437 2910741 : && is_a <gphi *> (STMT_VINFO_STMT (stmt_info)))
3438 : {
3439 103974 : loop_vec_info loop_vinfo = as_a <loop_vec_info> (vinfo);
3440 103974 : gcc_assert (STMT_VINFO_REDUC_IDX (stmt_info) == -1);
3441 103974 : node->cycle_info.id = loop_vinfo->reduc_infos.length ();
3442 103974 : vect_reduc_info reduc_info = new vect_reduc_info_s ();
3443 103974 : loop_vinfo->reduc_infos.safe_push (reduc_info);
3444 103974 : stmt_vec_info reduc_phi = stmt_info;
3445 : /* ??? For double reductions vect_is_simple_reduction stores the
3446 : reduction type and code on the inner loop header PHI. */
3447 103974 : if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def)
3448 : {
3449 388 : use_operand_p use_p;
3450 388 : gimple *use_stmt;
3451 388 : bool res = single_imm_use (gimple_phi_result (stmt_info->stmt),
3452 : &use_p, &use_stmt);
3453 388 : gcc_assert (res);
3454 388 : reduc_phi = loop_vinfo->lookup_stmt (use_stmt);
3455 : }
3456 103974 : VECT_REDUC_INFO_DEF_TYPE (reduc_info) = STMT_VINFO_DEF_TYPE (stmt_info);
3457 103974 : VECT_REDUC_INFO_TYPE (reduc_info) = STMT_VINFO_REDUC_TYPE (reduc_phi);
3458 103974 : VECT_REDUC_INFO_CODE (reduc_info) = STMT_VINFO_REDUC_CODE (reduc_phi);
3459 103974 : VECT_REDUC_INFO_FN (reduc_info) = IFN_LAST;
3460 : }
3461 : return node;
3462 9866859 : }
3463 :
3464 : /* Dump a single SLP tree NODE. */
3465 :
3466 : static void
3467 467515 : vect_print_slp_tree (dump_flags_t dump_kind, dump_location_t loc,
3468 : slp_tree node)
3469 : {
3470 467515 : unsigned i, j;
3471 467515 : slp_tree child;
3472 467515 : stmt_vec_info stmt_info;
3473 467515 : tree op;
3474 :
3475 467515 : dump_metadata_t metadata (dump_kind, loc.get_impl_location ());
3476 467515 : dump_user_location_t user_loc = loc.get_user_location ();
3477 467515 : dump_printf_loc (metadata, user_loc,
3478 : "node%s %p (refcnt=%u)",
3479 467515 : SLP_TREE_DEF_TYPE (node) == vect_external_def
3480 : ? " (external)"
3481 : : (SLP_TREE_DEF_TYPE (node) == vect_constant_def
3482 450829 : ? " (constant)"
3483 : : ""), (void *) node,
3484 : SLP_TREE_REF_COUNT (node));
3485 467515 : if (SLP_TREE_VECTYPE (node))
3486 395546 : dump_printf (metadata, " %T", SLP_TREE_VECTYPE (node));
3487 467515 : dump_printf (metadata, "%s",
3488 467515 : node->avoid_stlf_fail ? " (avoid-stlf-fail)" : "");
3489 467515 : if (node->cycle_info.id != -1 || node->cycle_info.reduc_idx != -1)
3490 24505 : dump_printf (metadata, " cycle %d, link %d", node->cycle_info.id,
3491 : node->cycle_info.reduc_idx);
3492 467515 : if (node->si)
3493 5334 : dump_printf (metadata, " @%G", node->si);
3494 : else
3495 462181 : dump_printf (metadata, "\n");
3496 467515 : if (SLP_TREE_DEF_TYPE (node) == vect_internal_def)
3497 : {
3498 379963 : if (SLP_TREE_PERMUTE_P (node))
3499 17301 : dump_printf_loc (metadata, user_loc, "op: VEC_PERM_EXPR\n");
3500 : else
3501 362662 : dump_printf_loc (metadata, user_loc, "op template: %G",
3502 362662 : SLP_TREE_REPRESENTATIVE (node)->stmt);
3503 : }
3504 467515 : if (SLP_TREE_SCALAR_STMTS (node).exists ())
3505 912996 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
3506 542071 : if (stmt_info)
3507 535949 : dump_printf_loc (metadata, user_loc, "\t%sstmt %u %G",
3508 535949 : SLP_TREE_LIVE_LANES (node).contains (i)
3509 532202 : ? "[l*]" : (STMT_VINFO_LIVE_P (stmt_info)
3510 532202 : ? "[l] " : ""),
3511 : i, stmt_info->stmt);
3512 : else
3513 6122 : dump_printf_loc (metadata, user_loc, "\tstmt %u ---\n", i);
3514 : else
3515 : {
3516 96590 : dump_printf_loc (metadata, user_loc, "\t{ ");
3517 214040 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
3518 117450 : dump_printf (metadata, "%T%s ", op,
3519 117450 : i < SLP_TREE_SCALAR_OPS (node).length () - 1 ? "," : "");
3520 96590 : dump_printf (metadata, "}\n");
3521 : }
3522 467515 : if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
3523 : {
3524 67026 : dump_printf_loc (metadata, user_loc, "\tload permutation {");
3525 153193 : FOR_EACH_VEC_ELT (SLP_TREE_LOAD_PERMUTATION (node), i, j)
3526 86167 : dump_printf (dump_kind, " %u", j);
3527 67026 : dump_printf (dump_kind, " }\n");
3528 : }
3529 467515 : if (SLP_TREE_LANE_PERMUTATION (node).exists ())
3530 : {
3531 17309 : dump_printf_loc (metadata, user_loc, "\tlane permutation {");
3532 64824 : for (i = 0; i < SLP_TREE_LANE_PERMUTATION (node).length (); ++i)
3533 47515 : dump_printf (dump_kind, " %u[%u]",
3534 47515 : SLP_TREE_LANE_PERMUTATION (node)[i].first,
3535 47515 : SLP_TREE_LANE_PERMUTATION (node)[i].second);
3536 17309 : dump_printf (dump_kind, " }%s\n",
3537 17309 : node->ldst_lanes ? " (load-lanes)" : "");
3538 : }
3539 467515 : if (SLP_TREE_CHILDREN (node).is_empty ())
3540 177139 : return;
3541 290376 : dump_printf_loc (metadata, user_loc, "\tchildren");
3542 766157 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3543 475781 : dump_printf (dump_kind, " %p", (void *)child);
3544 290376 : dump_printf (dump_kind, "%s\n",
3545 290376 : node->ldst_lanes && !SLP_TREE_LANE_PERMUTATION (node).exists ()
3546 : ? " (store-lanes)" : "");
3547 : }
3548 :
3549 : DEBUG_FUNCTION void
3550 0 : debug (slp_tree node)
3551 : {
3552 0 : debug_dump_context ctx;
3553 0 : vect_print_slp_tree (MSG_NOTE,
3554 0 : dump_location_t::from_location_t (UNKNOWN_LOCATION),
3555 : node);
3556 0 : }
3557 :
3558 : /* Recursive helper for the dot producer below. */
3559 :
3560 : static void
3561 0 : dot_slp_tree (FILE *f, slp_tree node, hash_set<slp_tree> &visited)
3562 : {
3563 0 : if (visited.add (node))
3564 : return;
3565 :
3566 0 : fprintf (f, "\"%p\" [label=\"", (void *)node);
3567 0 : vect_print_slp_tree (MSG_NOTE,
3568 0 : dump_location_t::from_location_t (UNKNOWN_LOCATION),
3569 : node);
3570 0 : fprintf (f, "\"];\n");
3571 :
3572 :
3573 0 : for (slp_tree child : SLP_TREE_CHILDREN (node))
3574 0 : fprintf (f, "\"%p\" -> \"%p\";", (void *)node, (void *)child);
3575 :
3576 0 : for (slp_tree child : SLP_TREE_CHILDREN (node))
3577 0 : if (child)
3578 0 : dot_slp_tree (f, child, visited);
3579 : }
3580 :
3581 : DEBUG_FUNCTION void
3582 0 : dot_slp_tree (const char *fname, slp_tree node)
3583 : {
3584 0 : FILE *f = fopen (fname, "w");
3585 0 : fprintf (f, "digraph {\n");
3586 0 : fflush (f);
3587 0 : {
3588 0 : debug_dump_context ctx (f);
3589 0 : hash_set<slp_tree> visited;
3590 0 : dot_slp_tree (f, node, visited);
3591 0 : }
3592 0 : fflush (f);
3593 0 : fprintf (f, "}\n");
3594 0 : fclose (f);
3595 0 : }
3596 :
3597 : DEBUG_FUNCTION void
3598 0 : dot_slp_tree (const char *fname, const vec<slp_instance> &slp_instances)
3599 : {
3600 0 : FILE *f = fopen (fname, "w");
3601 0 : fprintf (f, "digraph {\n");
3602 0 : fflush (f);
3603 0 : {
3604 0 : debug_dump_context ctx (f);
3605 0 : hash_set<slp_tree> visited;
3606 0 : for (auto inst : slp_instances)
3607 0 : dot_slp_tree (f, SLP_INSTANCE_TREE (inst), visited);
3608 0 : }
3609 0 : fflush (f);
3610 0 : fprintf (f, "}\n");
3611 0 : fclose (f);
3612 0 : }
3613 :
3614 : /* Dump a slp tree NODE using flags specified in DUMP_KIND. */
3615 :
3616 : static void
3617 504734 : vect_print_slp_graph (dump_flags_t dump_kind, dump_location_t loc,
3618 : slp_tree node, hash_set<slp_tree> &visited)
3619 : {
3620 504734 : unsigned i;
3621 504734 : slp_tree child;
3622 :
3623 504734 : if (visited.add (node))
3624 504734 : return;
3625 :
3626 463684 : vect_print_slp_tree (dump_kind, loc, node);
3627 :
3628 1398020 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3629 470652 : if (child)
3630 426537 : vect_print_slp_graph (dump_kind, loc, child, visited);
3631 : }
3632 :
3633 : static void
3634 47942 : vect_print_slp_graph (dump_flags_t dump_kind, dump_location_t loc,
3635 : slp_tree entry)
3636 : {
3637 47942 : hash_set<slp_tree> visited;
3638 47942 : vect_print_slp_graph (dump_kind, loc, entry, visited);
3639 47942 : }
3640 :
3641 : DEBUG_FUNCTION void
3642 0 : debug (slp_instance instance)
3643 : {
3644 0 : debug_dump_context ctx;
3645 0 : vect_print_slp_graph (MSG_NOTE,
3646 0 : dump_location_t::from_location_t (UNKNOWN_LOCATION),
3647 : SLP_INSTANCE_TREE (instance));
3648 0 : }
3649 :
3650 :
3651 : /* Compute the set of scalar stmts participating in external nodes. */
3652 :
3653 : static void
3654 1699413 : vect_slp_gather_extern_scalar_stmts (vec_info *vinfo, slp_tree node,
3655 : hash_set<slp_tree> &visited,
3656 : hash_set<stmt_vec_info> &estmts)
3657 : {
3658 1699413 : if (visited.add (node))
3659 : return;
3660 :
3661 1638705 : if (SLP_TREE_DEF_TYPE (node) == vect_internal_def)
3662 : {
3663 : slp_tree child;
3664 : int i;
3665 1921298 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3666 981903 : if (child)
3667 981903 : vect_slp_gather_extern_scalar_stmts (vinfo, child, visited, estmts);
3668 : }
3669 : else
3670 3915070 : for (tree def : SLP_TREE_SCALAR_OPS (node))
3671 : {
3672 1819178 : stmt_vec_info def_stmt = vinfo->lookup_def (def);
3673 1819178 : if (def_stmt)
3674 387629 : estmts.add (def_stmt);
3675 : }
3676 : }
3677 :
3678 : /* Mark the original scalar stmt coverage of the vector SLP graph of VINFO
3679 : with STMT_SLP_TYPE == pure_slp. */
3680 :
3681 : static void
3682 247322 : vect_bb_slp_mark_stmts_vectorized (bb_vec_info vinfo)
3683 : {
3684 : /* Gather the scalar stmt leafs of the SLP graph to stop the below DFS
3685 : walk on. */
3686 247322 : hash_set<stmt_vec_info> scalar_stmts_in_externs;
3687 247322 : hash_set<slp_tree> visited;
3688 1459476 : for (auto instance : BB_VINFO_SLP_INSTANCES (vinfo))
3689 717510 : vect_slp_gather_extern_scalar_stmts (vinfo, SLP_INSTANCE_TREE (instance),
3690 : visited, scalar_stmts_in_externs);
3691 :
3692 : /* DFS walk scalar stmts to compute the vectorized coverage indicated
3693 : by STMT_SLP_TYPE (stmt) == pure_slp on the original scalar (non-pattern)
3694 : stmts. */
3695 1459476 : for (auto instance : BB_VINFO_SLP_INSTANCES (vinfo))
3696 : {
3697 910497 : for (auto stmt : SLP_INSTANCE_ROOT_STMTS (instance))
3698 89203 : if (!scalar_stmts_in_externs.contains (stmt))
3699 87955 : STMT_SLP_TYPE (stmt) = pure_slp;
3700 717510 : auto_vec<stmt_vec_info> worklist;
3701 4044342 : for (auto stmt : SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (instance)))
3702 : {
3703 1891812 : stmt = vect_orig_stmt (stmt);
3704 1891812 : if (!scalar_stmts_in_externs.contains (stmt)
3705 1891812 : && STMT_SLP_TYPE (stmt) != pure_slp)
3706 : {
3707 1874966 : STMT_SLP_TYPE (stmt) = pure_slp;
3708 1874966 : worklist.safe_push (stmt);
3709 : }
3710 : }
3711 3773746 : while (!worklist.is_empty ())
3712 : {
3713 2343573 : stmt_vec_info stmt = worklist.pop ();
3714 :
3715 : /* Now walk relevant parts of the SSA use-def graph. */
3716 2343573 : slp_oprnds child_ops (stmt);
3717 7309441 : for (unsigned i = 0; i < child_ops.num_slp_children; ++i)
3718 : {
3719 2622295 : tree op = child_ops.get_op_for_slp_child (stmt, i);
3720 2622295 : stmt_vec_info def = vinfo->lookup_def (op);
3721 2622295 : if (def
3722 939401 : && !scalar_stmts_in_externs.contains (def)
3723 3185796 : && STMT_SLP_TYPE (def) != pure_slp)
3724 : {
3725 468607 : STMT_SLP_TYPE (def) = pure_slp;
3726 468607 : worklist.safe_push (def);
3727 : }
3728 : }
3729 : }
3730 717510 : }
3731 247322 : }
3732 :
3733 : /* Mark the statements of the tree rooted at NODE as relevant (vect_used). */
3734 :
3735 : static void
3736 2604824 : vect_mark_slp_stmts_relevant (slp_tree node, hash_set<slp_tree> &visited)
3737 : {
3738 2604824 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
3739 : return;
3740 :
3741 1565256 : if (visited.add (node))
3742 : return;
3743 :
3744 7299955 : for (auto stmt_info : SLP_TREE_SCALAR_STMTS (node))
3745 3259728 : if (stmt_info)
3746 : {
3747 3259728 : gcc_assert (!STMT_VINFO_RELEVANT (stmt_info)
3748 : || STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope);
3749 3259728 : STMT_VINFO_RELEVANT (stmt_info) = vect_used_in_scope;
3750 : }
3751 :
3752 5666895 : for (auto child: SLP_TREE_CHILDREN (node))
3753 1782944 : if (child)
3754 1782944 : vect_mark_slp_stmts_relevant (child, visited);
3755 : }
3756 :
3757 : static void
3758 821880 : vect_mark_slp_stmts_relevant (slp_tree node)
3759 : {
3760 821880 : hash_set<slp_tree> visited;
3761 821880 : vect_mark_slp_stmts_relevant (node, visited);
3762 821880 : }
3763 :
3764 :
3765 : /* Gather loads in the SLP graph NODE and populate the INST loads array. */
3766 :
3767 : static void
3768 11170807 : vect_gather_slp_loads (vec<slp_tree> &loads, slp_tree node,
3769 : hash_set<slp_tree> &visited)
3770 : {
3771 11170807 : if (!node || visited.add (node))
3772 : return;
3773 :
3774 9340826 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
3775 : return;
3776 :
3777 6907044 : if (!SLP_TREE_PERMUTE_P (node))
3778 : {
3779 6692327 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
3780 6692327 : if (STMT_VINFO_DATA_REF (stmt_info)
3781 2870662 : && DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
3782 1637301 : loads.safe_push (node);
3783 : }
3784 :
3785 6907044 : unsigned i;
3786 6907044 : slp_tree child;
3787 15743397 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3788 8836353 : vect_gather_slp_loads (loads, child, visited);
3789 : }
3790 :
3791 :
3792 : /* Find the last store in SLP INSTANCE. */
3793 :
3794 : stmt_vec_info
3795 2565388 : vect_find_last_scalar_stmt_in_slp (slp_tree node)
3796 : {
3797 2565388 : stmt_vec_info last = NULL;
3798 14515890 : for (auto stmt_vinfo : SLP_TREE_SCALAR_STMTS (node))
3799 6819726 : if (stmt_vinfo)
3800 : {
3801 6819726 : stmt_vinfo = vect_orig_stmt (stmt_vinfo);
3802 6819726 : last = last ? get_later_stmt (stmt_vinfo, last) : stmt_vinfo;
3803 : }
3804 :
3805 2565388 : return last;
3806 : }
3807 :
3808 : /* Find the first stmt in NODE. */
3809 :
3810 : stmt_vec_info
3811 684581 : vect_find_first_scalar_stmt_in_slp (slp_tree node)
3812 : {
3813 684581 : stmt_vec_info first = NULL;
3814 :
3815 3708828 : for (auto stmt_vinfo : SLP_TREE_SCALAR_STMTS (node))
3816 1655085 : if (stmt_vinfo)
3817 : {
3818 1652377 : stmt_vinfo = vect_orig_stmt (stmt_vinfo);
3819 1652377 : if (!first
3820 1652377 : || get_later_stmt (stmt_vinfo, first) == first)
3821 : first = stmt_vinfo;
3822 : }
3823 :
3824 684581 : return first;
3825 : }
3826 :
3827 : /* Splits a group of stores, currently beginning at FIRST_VINFO, into
3828 : two groups: one (still beginning at FIRST_VINFO) of size GROUP1_SIZE
3829 : (also containing the first GROUP1_SIZE stmts, since stores are
3830 : consecutive), the second containing the remainder.
3831 : Return the first stmt in the second group. */
3832 :
3833 : static stmt_vec_info
3834 136724 : vect_split_slp_store_group (stmt_vec_info first_vinfo, unsigned group1_size)
3835 : {
3836 136724 : gcc_assert (DR_GROUP_FIRST_ELEMENT (first_vinfo) == first_vinfo);
3837 136724 : gcc_assert (group1_size > 0);
3838 136724 : int group2_size = DR_GROUP_SIZE (first_vinfo) - group1_size;
3839 136724 : gcc_assert (group2_size > 0);
3840 136724 : DR_GROUP_SIZE (first_vinfo) = group1_size;
3841 :
3842 136724 : stmt_vec_info stmt_info = first_vinfo;
3843 459101 : for (unsigned i = group1_size; i > 1; i--)
3844 : {
3845 322377 : stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info);
3846 322377 : gcc_assert (DR_GROUP_GAP (stmt_info) == 1);
3847 : }
3848 : /* STMT is now the last element of the first group. */
3849 136724 : stmt_vec_info group2 = DR_GROUP_NEXT_ELEMENT (stmt_info);
3850 136724 : DR_GROUP_NEXT_ELEMENT (stmt_info) = 0;
3851 :
3852 136724 : DR_GROUP_SIZE (group2) = group2_size;
3853 406615 : for (stmt_info = group2; stmt_info;
3854 269891 : stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info))
3855 : {
3856 269891 : DR_GROUP_FIRST_ELEMENT (stmt_info) = group2;
3857 269891 : gcc_assert (DR_GROUP_GAP (stmt_info) == 1);
3858 : }
3859 :
3860 : /* For the second group, the DR_GROUP_GAP is that before the original group,
3861 : plus skipping over the first vector. */
3862 136724 : DR_GROUP_GAP (group2) = DR_GROUP_GAP (first_vinfo) + group1_size;
3863 :
3864 : /* DR_GROUP_GAP of the first group now has to skip over the second group too. */
3865 136724 : DR_GROUP_GAP (first_vinfo) += group2_size;
3866 :
3867 136724 : if (dump_enabled_p ())
3868 85 : dump_printf_loc (MSG_NOTE, vect_location, "Split group into %d and %d\n",
3869 : group1_size, group2_size);
3870 :
3871 136724 : return group2;
3872 : }
3873 :
3874 : /* Calculate the unrolling factor for an SLP instance with GROUP_SIZE
3875 : statements and a vector of NUNITS elements. */
3876 :
3877 : static poly_uint64
3878 2869511 : calculate_unrolling_factor (poly_uint64 nunits, unsigned int group_size)
3879 : {
3880 2869511 : return exact_div (common_multiple (nunits, group_size), group_size);
3881 : }
3882 :
3883 : /* Helper that checks to see if a node is a load node. */
3884 :
3885 : static inline bool
3886 103 : vect_is_slp_load_node (slp_tree root)
3887 : {
3888 103 : return (!SLP_TREE_PERMUTE_P (root)
3889 103 : && SLP_TREE_DEF_TYPE (root) == vect_internal_def
3890 97 : && STMT_VINFO_GROUPED_ACCESS (SLP_TREE_REPRESENTATIVE (root))
3891 167 : && DR_IS_READ (STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (root))));
3892 : }
3893 :
3894 :
3895 : /* Helper function of optimize_load_redistribution that performs the operation
3896 : recursively. */
3897 :
3898 : static slp_tree
3899 21801 : optimize_load_redistribution_1 (scalar_stmts_to_slp_tree_map_t *bst_map,
3900 : vec_info *vinfo, unsigned int group_size,
3901 : hash_map<slp_tree, slp_tree> *load_map,
3902 : slp_tree root)
3903 : {
3904 21801 : if (slp_tree *leader = load_map->get (root))
3905 4549 : return *leader;
3906 :
3907 17252 : slp_tree node;
3908 17252 : unsigned i;
3909 :
3910 : /* For now, we don't know anything about externals so do not do anything. */
3911 17252 : if (!root || SLP_TREE_DEF_TYPE (root) != vect_internal_def)
3912 : return NULL;
3913 12116 : else if (SLP_TREE_PERMUTE_P (root))
3914 : {
3915 : /* First convert this node into a load node and add it to the leaves
3916 : list and flatten the permute from a lane to a load one. If it's
3917 : unneeded it will be elided later. */
3918 71 : vec<stmt_vec_info> stmts;
3919 71 : stmts.create (SLP_TREE_LANES (root));
3920 71 : lane_permutation_t lane_perm = SLP_TREE_LANE_PERMUTATION (root);
3921 135 : for (unsigned j = 0; j < lane_perm.length (); j++)
3922 : {
3923 103 : std::pair<unsigned, unsigned> perm = lane_perm[j];
3924 103 : node = SLP_TREE_CHILDREN (root)[perm.first];
3925 :
3926 103 : if (!vect_is_slp_load_node (node)
3927 103 : || SLP_TREE_CHILDREN (node).exists ())
3928 : {
3929 39 : stmts.release ();
3930 39 : goto next;
3931 : }
3932 :
3933 64 : stmts.quick_push (SLP_TREE_SCALAR_STMTS (node)[perm.second]);
3934 : }
3935 :
3936 32 : if (dump_enabled_p ())
3937 0 : dump_printf_loc (MSG_NOTE, vect_location,
3938 : "converting stmts on permute node %p\n",
3939 : (void *) root);
3940 :
3941 32 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_size);
3942 32 : unsigned tree_size = 0, limit = 1;
3943 32 : node = vect_build_slp_tree (vinfo, stmts,
3944 : matches, &limit, &tree_size, bst_map);
3945 32 : if (!node)
3946 0 : stmts.release ();
3947 :
3948 32 : load_map->put (root, node);
3949 32 : return node;
3950 : }
3951 :
3952 12045 : next:
3953 12084 : load_map->put (root, NULL);
3954 :
3955 27523 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (root), i , node)
3956 : {
3957 15439 : slp_tree value
3958 15439 : = optimize_load_redistribution_1 (bst_map, vinfo, group_size, load_map,
3959 : node);
3960 15439 : if (value)
3961 : {
3962 32 : SLP_TREE_REF_COUNT (value)++;
3963 32 : SLP_TREE_CHILDREN (root)[i] = value;
3964 : /* ??? We know the original leafs of the replaced nodes will
3965 : be referenced by bst_map, only the permutes created by
3966 : pattern matching are not. */
3967 32 : if (SLP_TREE_REF_COUNT (node) == 1)
3968 32 : load_map->remove (node);
3969 32 : vect_free_slp_tree (node);
3970 : }
3971 : }
3972 :
3973 : return NULL;
3974 : }
3975 :
3976 : /* Temporary workaround for loads not being CSEd during SLP build. This
3977 : function will traverse the SLP tree rooted in ROOT for INSTANCE and find
3978 : VEC_PERM nodes that blend vectors from multiple nodes that all read from the
3979 : same DR such that the final operation is equal to a permuted load. Such
3980 : NODES are then directly converted into LOADS themselves. The nodes are
3981 : CSEd using BST_MAP. */
3982 :
3983 : static void
3984 4180 : optimize_load_redistribution (scalar_stmts_to_slp_tree_map_t *bst_map,
3985 : vec_info *vinfo, unsigned int group_size,
3986 : hash_map<slp_tree, slp_tree> *load_map,
3987 : slp_tree root)
3988 : {
3989 4180 : slp_tree node;
3990 4180 : unsigned i;
3991 :
3992 10542 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (root), i , node)
3993 : {
3994 6362 : slp_tree value
3995 6362 : = optimize_load_redistribution_1 (bst_map, vinfo, group_size, load_map,
3996 : node);
3997 6362 : if (value)
3998 : {
3999 0 : SLP_TREE_REF_COUNT (value)++;
4000 0 : SLP_TREE_CHILDREN (root)[i] = value;
4001 : /* ??? We know the original leafs of the replaced nodes will
4002 : be referenced by bst_map, only the permutes created by
4003 : pattern matching are not. */
4004 0 : if (SLP_TREE_REF_COUNT (node) == 1)
4005 0 : load_map->remove (node);
4006 0 : vect_free_slp_tree (node);
4007 : }
4008 : }
4009 4180 : }
4010 :
4011 : /* Helper function of vect_match_slp_patterns.
4012 :
4013 : Attempts to match patterns against the slp tree rooted in REF_NODE using
4014 : VINFO. Patterns are matched in post-order traversal.
4015 :
4016 : If matching is successful the value in REF_NODE is updated and returned, if
4017 : not then it is returned unchanged. */
4018 :
4019 : static bool
4020 6486159 : vect_match_slp_patterns_2 (slp_tree *ref_node, vec_info *vinfo,
4021 : slp_tree_to_load_perm_map_t *perm_cache,
4022 : slp_compat_nodes_map_t *compat_cache,
4023 : hash_set<slp_tree> *visited)
4024 : {
4025 6486159 : unsigned i;
4026 6486159 : slp_tree node = *ref_node;
4027 6486159 : bool found_p = false;
4028 6486159 : if (!node || visited->add (node))
4029 : return false;
4030 :
4031 : slp_tree child;
4032 10417232 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
4033 4857006 : found_p |= vect_match_slp_patterns_2 (&SLP_TREE_CHILDREN (node)[i],
4034 : vinfo, perm_cache, compat_cache,
4035 : visited);
4036 :
4037 16680678 : for (unsigned x = 0; x < num__slp_patterns; x++)
4038 : {
4039 11120452 : vect_pattern *pattern
4040 11120452 : = slp_patterns[x] (perm_cache, compat_cache, ref_node);
4041 11120452 : if (pattern)
4042 : {
4043 1114 : pattern->build (vinfo);
4044 1114 : delete pattern;
4045 : found_p = true;
4046 : }
4047 : }
4048 :
4049 : return found_p;
4050 : }
4051 :
4052 : /* Applies pattern matching to the given SLP tree rooted in REF_NODE using
4053 : vec_info VINFO.
4054 :
4055 : The modified tree is returned. Patterns are tried in order and multiple
4056 : patterns may match. */
4057 :
4058 : static bool
4059 1629153 : vect_match_slp_patterns (slp_instance instance, vec_info *vinfo,
4060 : hash_set<slp_tree> *visited,
4061 : slp_tree_to_load_perm_map_t *perm_cache,
4062 : slp_compat_nodes_map_t *compat_cache)
4063 : {
4064 1629153 : DUMP_VECT_SCOPE ("vect_match_slp_patterns");
4065 1629153 : slp_tree *ref_node = &SLP_INSTANCE_TREE (instance);
4066 :
4067 1629153 : if (dump_enabled_p ())
4068 31606 : dump_printf_loc (MSG_NOTE, vect_location,
4069 : "Analyzing SLP tree %p for patterns\n",
4070 31606 : (void *) SLP_INSTANCE_TREE (instance));
4071 :
4072 1629153 : return vect_match_slp_patterns_2 (ref_node, vinfo, perm_cache, compat_cache,
4073 1629153 : visited);
4074 : }
4075 :
4076 : /* STMT_INFO is a store group of size GROUP_SIZE that we are considering
4077 : vectorizing with VECTYPE that might be NULL. MASKED_P indicates whether
4078 : the stores are masked.
4079 : Return true if we could use IFN_STORE_LANES instead and if that appears
4080 : to be the better approach. */
4081 :
4082 : static bool
4083 6160 : vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
4084 : tree vectype, bool masked_p,
4085 : unsigned int group_size,
4086 : unsigned int new_group_size)
4087 : {
4088 6160 : if (!vectype)
4089 : {
4090 6160 : tree scalar_type = TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
4091 6160 : vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
4092 : }
4093 6160 : if (!vectype)
4094 : return false;
4095 : /* Allow the split if one of the two new groups would operate on full
4096 : vectors *within* rather than across one scalar loop iteration.
4097 : This is purely a heuristic, but it should work well for group
4098 : sizes of 3 and 4, where the possible splits are:
4099 :
4100 : 3->2+1: OK if the vector has exactly two elements
4101 : 4->2+2: Likewise
4102 : 4->3+1: Less clear-cut. */
4103 6160 : if (multiple_p (group_size - new_group_size, TYPE_VECTOR_SUBPARTS (vectype))
4104 6160 : || multiple_p (new_group_size, TYPE_VECTOR_SUBPARTS (vectype)))
4105 : return false;
4106 3482 : return vect_store_lanes_supported (vectype, group_size, masked_p) != IFN_LAST;
4107 : }
4108 :
4109 : /* Analyze an SLP instance starting from a group of grouped stores. Call
4110 : vect_build_slp_tree to build a tree of packed stmts if possible.
4111 : Return FALSE if it's impossible to SLP any stmt in the loop. */
4112 :
4113 : static bool
4114 : vect_analyze_slp_instance (vec_info *vinfo,
4115 : scalar_stmts_to_slp_tree_map_t *bst_map,
4116 : stmt_vec_info stmt_info, slp_instance_kind kind,
4117 : unsigned max_tree_size, unsigned *limit,
4118 : bool force_single_lane);
4119 :
4120 : /* Build an interleaving scheme for the store sources RHS_NODES from
4121 : SCALAR_STMTS. */
4122 :
4123 : static slp_tree
4124 8263 : vect_build_slp_store_interleaving (vec<slp_tree> &rhs_nodes,
4125 : vec<stmt_vec_info> &scalar_stmts)
4126 : {
4127 8263 : unsigned int group_size = scalar_stmts.length ();
4128 16526 : slp_tree node = vect_create_new_slp_node (scalar_stmts,
4129 8263 : SLP_TREE_CHILDREN
4130 : (rhs_nodes[0]).length ());
4131 8263 : SLP_TREE_VECTYPE (node) = SLP_TREE_VECTYPE (rhs_nodes[0]);
4132 8263 : for (unsigned l = 0;
4133 16553 : l < SLP_TREE_CHILDREN (rhs_nodes[0]).length (); ++l)
4134 : {
4135 : /* And a permute merging all RHS SLP trees. */
4136 8290 : slp_tree perm = vect_create_new_slp_node (rhs_nodes.length (),
4137 8290 : VEC_PERM_EXPR);
4138 8290 : SLP_TREE_CHILDREN (node).quick_push (perm);
4139 8290 : SLP_TREE_LANE_PERMUTATION (perm).create (group_size);
4140 8290 : SLP_TREE_VECTYPE (perm) = SLP_TREE_VECTYPE (node);
4141 8290 : SLP_TREE_LANES (perm) = group_size;
4142 8290 : SLP_TREE_REPRESENTATIVE (perm) = NULL;
4143 64852 : for (unsigned j = 0; j < rhs_nodes.length (); ++j)
4144 : {
4145 24136 : SLP_TREE_CHILDREN (perm)
4146 24136 : .quick_push (SLP_TREE_CHILDREN (rhs_nodes[j])[l]);
4147 24136 : SLP_TREE_CHILDREN (rhs_nodes[j])[l]->refcnt++;
4148 24136 : for (unsigned k = 0;
4149 50645 : k < SLP_TREE_LANES (rhs_nodes[j]); ++k)
4150 : {
4151 : /* ??? We should populate SLP_TREE_SCALAR_STMTS
4152 : or SLP_TREE_SCALAR_OPS but then we might have
4153 : a mix of both in our children. */
4154 26509 : SLP_TREE_LANE_PERMUTATION (perm)
4155 26509 : .quick_push (std::make_pair (j, k));
4156 : }
4157 : }
4158 :
4159 : /* Now we have a single permute node but we cannot code-generate
4160 : the case with more than two inputs.
4161 : Perform pairwise reduction, reducing the two inputs
4162 : with the least number of lanes to one and then repeat until
4163 : we end up with two inputs. That scheme makes sure we end
4164 : up with permutes satisfying the restriction of requiring at
4165 : most two vector inputs to produce a single vector output
4166 : when the number of lanes is even. */
4167 15846 : while (SLP_TREE_CHILDREN (perm).length () > 2)
4168 : {
4169 : /* When we have three equal sized groups left the pairwise
4170 : reduction does not result in a scheme that avoids using
4171 : three vectors. Instead merge the first two groups
4172 : to the final size with do-not-care elements (chosen
4173 : from the first group) and then merge with the third.
4174 : { A0, B0, x, A1, B1, x, ... }
4175 : -> { A0, B0, C0, A1, B1, C1, ... }
4176 : This handles group size of three (and at least
4177 : power-of-two multiples of that). */
4178 7556 : if (SLP_TREE_CHILDREN (perm).length () == 3
4179 3400 : && (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[0])
4180 3400 : == SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[1]))
4181 7556 : && (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[0])
4182 2519 : == SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[2])))
4183 : {
4184 2213 : int ai = 0;
4185 2213 : int bi = 1;
4186 2213 : slp_tree a = SLP_TREE_CHILDREN (perm)[ai];
4187 2213 : slp_tree b = SLP_TREE_CHILDREN (perm)[bi];
4188 2213 : unsigned n = SLP_TREE_LANES (perm);
4189 :
4190 2213 : slp_tree permab = vect_create_new_slp_node (2, VEC_PERM_EXPR);
4191 2213 : SLP_TREE_LANES (permab) = n;
4192 2213 : SLP_TREE_LANE_PERMUTATION (permab).create (n);
4193 2213 : SLP_TREE_VECTYPE (permab) = SLP_TREE_VECTYPE (perm);
4194 : /* ??? Should be NULL but that's not expected. */
4195 2213 : SLP_TREE_REPRESENTATIVE (permab) = SLP_TREE_REPRESENTATIVE (perm);
4196 2213 : SLP_TREE_CHILDREN (permab).quick_push (a);
4197 6653 : for (unsigned k = 0; k < SLP_TREE_LANES (a); ++k)
4198 2227 : SLP_TREE_LANE_PERMUTATION (permab)
4199 2227 : .quick_push (std::make_pair (0, k));
4200 2213 : SLP_TREE_CHILDREN (permab).quick_push (b);
4201 6653 : for (unsigned k = 0; k < SLP_TREE_LANES (b); ++k)
4202 2227 : SLP_TREE_LANE_PERMUTATION (permab)
4203 2227 : .quick_push (std::make_pair (1, k));
4204 : /* Push the do-not-care lanes. */
4205 4440 : for (unsigned k = 0; k < SLP_TREE_LANES (a); ++k)
4206 2227 : SLP_TREE_LANE_PERMUTATION (permab)
4207 2227 : .quick_push (std::make_pair (0, k));
4208 :
4209 : /* Put the merged node into 'perm', in place of a. */
4210 2213 : SLP_TREE_CHILDREN (perm)[ai] = permab;
4211 : /* Adjust the references to b in the permutation
4212 : of perm and to the later children which we'll
4213 : remove. */
4214 8894 : for (unsigned k = 0; k < SLP_TREE_LANES (perm); ++k)
4215 : {
4216 6681 : std::pair<unsigned, unsigned> &p
4217 6681 : = SLP_TREE_LANE_PERMUTATION (perm)[k];
4218 6681 : if (p.first == (unsigned) bi)
4219 : {
4220 2227 : p.first = ai;
4221 2227 : p.second += SLP_TREE_LANES (a);
4222 : }
4223 4454 : else if (p.first > (unsigned) bi)
4224 2227 : p.first--;
4225 : }
4226 2213 : SLP_TREE_CHILDREN (perm).ordered_remove (bi);
4227 2213 : break;
4228 : }
4229 :
4230 : /* Pick the two nodes with the least number of lanes,
4231 : prefer the earliest candidate and maintain ai < bi. */
4232 : int ai = -1;
4233 : int bi = -1;
4234 50446 : for (unsigned ci = 0; ci < SLP_TREE_CHILDREN (perm).length (); ++ci)
4235 : {
4236 45103 : if (ai == -1)
4237 5343 : ai = ci;
4238 39760 : else if (bi == -1)
4239 5343 : bi = ci;
4240 34417 : else if ((SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ci])
4241 34417 : < SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ai]))
4242 34417 : || (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ci])
4243 28593 : < SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[bi])))
4244 : {
4245 12538 : if (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ai])
4246 6269 : <= SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[bi]))
4247 2892 : bi = ci;
4248 : else
4249 : {
4250 3377 : ai = bi;
4251 3377 : bi = ci;
4252 : }
4253 : }
4254 : }
4255 :
4256 : /* Produce a merge of nodes ai and bi. */
4257 5343 : slp_tree a = SLP_TREE_CHILDREN (perm)[ai];
4258 5343 : slp_tree b = SLP_TREE_CHILDREN (perm)[bi];
4259 5343 : unsigned n = SLP_TREE_LANES (a) + SLP_TREE_LANES (b);
4260 5343 : slp_tree permab = vect_create_new_slp_node (2, VEC_PERM_EXPR);
4261 5343 : SLP_TREE_LANES (permab) = n;
4262 5343 : SLP_TREE_LANE_PERMUTATION (permab).create (n);
4263 5343 : SLP_TREE_VECTYPE (permab) = SLP_TREE_VECTYPE (perm);
4264 : /* ??? Should be NULL but that's not expected. */
4265 5343 : SLP_TREE_REPRESENTATIVE (permab) = SLP_TREE_REPRESENTATIVE (perm);
4266 5343 : SLP_TREE_CHILDREN (permab).quick_push (a);
4267 19553 : for (unsigned k = 0; k < SLP_TREE_LANES (a); ++k)
4268 8867 : SLP_TREE_LANE_PERMUTATION (permab)
4269 8867 : .quick_push (std::make_pair (0, k));
4270 5343 : SLP_TREE_CHILDREN (permab).quick_push (b);
4271 18829 : for (unsigned k = 0; k < SLP_TREE_LANES (b); ++k)
4272 8143 : SLP_TREE_LANE_PERMUTATION (permab)
4273 8143 : .quick_push (std::make_pair (1, k));
4274 :
4275 : /* Put the merged node into 'perm', in place of a. */
4276 5343 : SLP_TREE_CHILDREN (perm)[ai] = permab;
4277 : /* Adjust the references to b in the permutation
4278 : of perm and to the later children which we'll
4279 : remove. */
4280 81265 : for (unsigned k = 0; k < SLP_TREE_LANES (perm); ++k)
4281 : {
4282 75922 : std::pair<unsigned, unsigned> &p
4283 75922 : = SLP_TREE_LANE_PERMUTATION (perm)[k];
4284 75922 : if (p.first == (unsigned) bi)
4285 : {
4286 8143 : p.first = ai;
4287 8143 : p.second += SLP_TREE_LANES (a);
4288 : }
4289 67779 : else if (p.first > (unsigned) bi)
4290 28686 : p.first--;
4291 : }
4292 5343 : SLP_TREE_CHILDREN (perm).ordered_remove (bi);
4293 : }
4294 : }
4295 :
4296 8263 : return node;
4297 : }
4298 :
4299 : /* Analyze an SLP instance starting from SCALAR_STMTS which are a group
4300 : of KIND. Return true if successful. SCALAR_STMTS is owned by this
4301 : function, REMAIN and ROOT_STMT_INFOS ownership is transferred back to
4302 : the caller upon failure. */
4303 :
4304 : static bool
4305 2045723 : vect_build_slp_instance (vec_info *vinfo,
4306 : slp_instance_kind kind,
4307 : vec<stmt_vec_info> &scalar_stmts,
4308 : vec<stmt_vec_info> &root_stmt_infos,
4309 : vec<tree> &remain,
4310 : unsigned max_tree_size, unsigned *limit,
4311 : scalar_stmts_to_slp_tree_map_t *bst_map,
4312 : bool force_single_lane)
4313 : {
4314 : /* If there's no budget left bail out early. */
4315 2045723 : if (*limit == 0)
4316 : {
4317 15992 : scalar_stmts.release ();
4318 15992 : return false;
4319 : }
4320 :
4321 2029731 : if (kind == slp_inst_kind_ctor)
4322 : {
4323 14484 : if (dump_enabled_p ())
4324 98 : dump_printf_loc (MSG_NOTE, vect_location,
4325 : "Analyzing vectorizable constructor: %G\n",
4326 49 : root_stmt_infos[0]->stmt);
4327 : }
4328 2015247 : else if (kind == slp_inst_kind_gcond)
4329 : {
4330 293399 : if (dump_enabled_p ())
4331 5864 : dump_printf_loc (MSG_NOTE, vect_location,
4332 : "Analyzing vectorizable control flow: %G",
4333 2932 : root_stmt_infos[0]->stmt);
4334 : }
4335 1721848 : else if (kind == slp_inst_kind_bb_reduc)
4336 : {
4337 1331567 : if (dump_enabled_p ())
4338 7434 : dump_printf_loc (MSG_NOTE, vect_location,
4339 : "Analyzing vectorizable BB reduction: %G",
4340 3717 : root_stmt_infos[0]->stmt);
4341 : }
4342 :
4343 2029731 : if (dump_enabled_p ())
4344 : {
4345 26653 : dump_printf_loc (MSG_NOTE, vect_location,
4346 : "Starting SLP discovery for\n");
4347 57723 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
4348 62140 : dump_printf_loc (MSG_NOTE, vect_location,
4349 31070 : " %G", scalar_stmts[i]->stmt);
4350 : }
4351 :
4352 : /* Build the tree for the SLP instance. */
4353 2029731 : unsigned int group_size = scalar_stmts.length ();
4354 2029731 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_size);
4355 2029731 : unsigned tree_size = 0;
4356 :
4357 2029731 : slp_tree node = NULL;
4358 2029731 : if (group_size > 1 && force_single_lane)
4359 : {
4360 0 : matches[0] = 0;
4361 0 : matches[1] = -1;
4362 : }
4363 : else
4364 2029731 : node = vect_build_slp_tree (vinfo, scalar_stmts, matches, limit,
4365 : &tree_size, bst_map);
4366 2029731 : if (node != NULL)
4367 : {
4368 : /* Create a new SLP instance. */
4369 824906 : slp_instance new_instance = XNEW (class _slp_instance);
4370 824906 : SLP_INSTANCE_TREE (new_instance) = node;
4371 824906 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
4372 824906 : SLP_INSTANCE_ROOT_STMTS (new_instance) = root_stmt_infos;
4373 824906 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = remain;
4374 824906 : SLP_INSTANCE_KIND (new_instance) = kind;
4375 824906 : new_instance->reduc_phis = NULL;
4376 824906 : new_instance->cost_vec = vNULL;
4377 824906 : new_instance->subgraph_entries = vNULL;
4378 :
4379 824906 : if (dump_enabled_p ())
4380 23359 : dump_printf_loc (MSG_NOTE, vect_location,
4381 : "SLP size %u vs. limit %u.\n",
4382 : tree_size, max_tree_size);
4383 :
4384 824906 : vinfo->slp_instances.safe_push (new_instance);
4385 :
4386 : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
4387 : the number of SLP lanes of the root in a few places.
4388 : Verify that assumption holds. */
4389 824906 : gcc_assert (SLP_TREE_LANES (SLP_INSTANCE_TREE (new_instance))
4390 : == group_size);
4391 :
4392 824906 : if (dump_enabled_p ())
4393 : {
4394 23359 : if (kind == slp_inst_kind_reduc_group)
4395 1488 : dump_printf_loc (MSG_NOTE, vect_location,
4396 : "SLP discovery of size %d reduction group "
4397 : "succeeded\n", group_size);
4398 23359 : dump_printf_loc (MSG_NOTE, vect_location,
4399 : "Final SLP tree for instance %p:\n",
4400 : (void *) new_instance);
4401 23359 : vect_print_slp_graph (MSG_NOTE, vect_location,
4402 : SLP_INSTANCE_TREE (new_instance));
4403 : }
4404 :
4405 824906 : return true;
4406 : }
4407 : /* Failed to SLP. */
4408 :
4409 : /* While we arrive here even with slp_inst_kind_store we should only
4410 : for group_size == 1. The code to split store groups is only in
4411 : vect_analyze_slp_instance now. */
4412 1204825 : gcc_assert (kind != slp_inst_kind_store || group_size == 1);
4413 :
4414 : /* For BB vectorization we get failures only in case of the need of
4415 : unrolling, as otherwise we'll simply get operands built from scalars.
4416 : Iff there is any mismatches in the toplevel stmts those will prevail,
4417 : otherwise we get the non-power-of-two tail of the lanes failed.
4418 : For BB reductions we mainly want to catch the first case so we pick
4419 : a more useful subset of lanes to reduce. Pick the largest matching
4420 : subset of that covers half of the group or more. */
4421 1204825 : if (kind == slp_inst_kind_bb_reduc && matches[0] != -1)
4422 : {
4423 667783 : unsigned *n_matching = XALLOCAVEC (unsigned, group_size);
4424 667783 : memset (n_matching, 0, sizeof (unsigned) * group_size);
4425 2174660 : for (unsigned i = 0; i < group_size; ++i)
4426 1506877 : if (matches[i] != -2 && matches[i] != -1)
4427 1270738 : n_matching[matches[i].v]++;
4428 : unsigned largest_i = 0;
4429 1506877 : for (unsigned i = 1; i < group_size; ++i)
4430 839094 : if (n_matching[i] > n_matching[largest_i])
4431 62226 : largest_i = i;
4432 : /* Pick the largest matching part and put the rest to remain. */
4433 667783 : if (n_matching[largest_i] >= 2
4434 50428 : && n_matching[largest_i] >= group_size / 2)
4435 : {
4436 : /* As we know the matches[] stmts match up, recursing for
4437 : non-power-of-two sizes will just force-fail the tail
4438 : for us at hopefully optimal vector size and succesfully
4439 : finish discovery. */
4440 49030 : vec<stmt_vec_info> scalar_stmts2;
4441 49030 : scalar_stmts2.create (n_matching[largest_i]);
4442 301730 : for (unsigned i = 0; i < group_size; ++i)
4443 203670 : if (matches[i] == (int)largest_i)
4444 142403 : scalar_stmts2.quick_push (scalar_stmts[i]);
4445 : else
4446 61267 : remain.safe_push
4447 64916 : (gimple_get_lhs (vect_orig_stmt (scalar_stmts[i])->stmt));
4448 49030 : if (dump_enabled_p ())
4449 220 : dump_printf_loc (MSG_NOTE, vect_location, "Splitting %d "
4450 : "non-matching lanes to scalar remains\n",
4451 220 : scalar_stmts.length () - scalar_stmts2.length ());
4452 49030 : scalar_stmts.release ();
4453 49030 : return vect_build_slp_instance (vinfo, kind, scalar_stmts2,
4454 : root_stmt_infos, remain,
4455 : max_tree_size, limit, bst_map,
4456 : force_single_lane);
4457 : }
4458 : }
4459 :
4460 : /* Free the allocated memory. */
4461 1155795 : scalar_stmts.release ();
4462 :
4463 : /* Failed to SLP. */
4464 1155795 : if (dump_enabled_p ())
4465 3184 : dump_printf_loc (MSG_NOTE, vect_location, "SLP discovery failed\n");
4466 : return false;
4467 : }
4468 :
4469 : /* Analyze an SLP instance starting from a the start of a reduction chain.
4470 : Call vect_build_slp_tree to build a tree of packed stmts if possible.
4471 : Return FALSE if SLP build fails. */
4472 :
4473 : static bool
4474 73463 : vect_analyze_slp_reduc_chain (loop_vec_info vinfo,
4475 : scalar_stmts_to_slp_tree_map_t *bst_map,
4476 : stmt_vec_info scalar_stmt,
4477 : unsigned max_tree_size, unsigned *limit)
4478 : {
4479 73463 : vec<stmt_vec_info> scalar_stmts = vNULL;
4480 :
4481 73463 : bool fail = false;
4482 : /* ??? We could leave operation code checking to SLP discovery. */
4483 73463 : code_helper code = STMT_VINFO_REDUC_CODE (STMT_VINFO_REDUC_DEF
4484 : (vect_orig_stmt (scalar_stmt)));
4485 73463 : bool first = true;
4486 73463 : stmt_vec_info next_stmt = scalar_stmt;
4487 83662 : do
4488 : {
4489 83662 : stmt_vec_info stmt = next_stmt;
4490 83662 : gimple_match_op op, orig_op;
4491 83662 : if (!gimple_extract_op (STMT_VINFO_STMT (stmt), &op))
4492 0 : gcc_unreachable ();
4493 167324 : tree reduc_def = gimple_arg (STMT_VINFO_STMT (stmt),
4494 83662 : STMT_VINFO_REDUC_IDX (stmt));
4495 83662 : next_stmt = vect_stmt_to_vectorize (vinfo->lookup_def (reduc_def));
4496 83662 : gcc_assert (is_a <gphi *> (STMT_VINFO_STMT (next_stmt))
4497 : || STMT_VINFO_REDUC_IDX (next_stmt) != -1);
4498 89624 : if (!gimple_extract_op (STMT_VINFO_STMT (vect_orig_stmt (stmt)),
4499 : &orig_op))
4500 0 : gcc_unreachable ();
4501 83662 : if (CONVERT_EXPR_CODE_P (op.code)
4502 4995 : && tree_nop_conversion_p (op.type, TREE_TYPE (op.ops[0]))
4503 88645 : && (first
4504 2496 : || is_a <gphi *> (STMT_VINFO_STMT (next_stmt))))
4505 : ;
4506 78739 : else if (code != orig_op.code)
4507 : {
4508 2728 : fail = true;
4509 2728 : break;
4510 : }
4511 : else
4512 76011 : scalar_stmts.safe_push (stmt);
4513 80934 : first = false;
4514 : }
4515 80934 : while (!is_a <gphi *> (STMT_VINFO_STMT (next_stmt)));
4516 73463 : if (fail)
4517 2728 : return false;
4518 :
4519 : /* When the SSA def chain through reduc-idx does not form a natural
4520 : reduction chain try to linearize an associative operation manually. */
4521 70735 : if (scalar_stmts.length () == 1
4522 68065 : && code.is_tree_code ()
4523 61993 : && associative_tree_code ((tree_code)code)
4524 : /* We may not associate if a fold-left reduction is required. */
4525 131315 : && !needs_fold_left_reduction_p (TREE_TYPE (gimple_get_lhs
4526 : (scalar_stmts[0]->stmt)),
4527 : code))
4528 : {
4529 : /* Remember a stmt with the actual reduction operation. */
4530 57307 : stmt_vec_info reduc_scalar_stmt = scalar_stmts[0];
4531 57307 : auto_vec<chain_op_t> chain;
4532 57307 : auto_vec<std::pair<tree_code, gimple *> > worklist;
4533 57307 : gimple *op_stmt = NULL, *other_op_stmt = NULL;
4534 57307 : if (is_a <gassign *> (scalar_stmts[0]->stmt)
4535 : /* We cannot linearize an operation that vect_slp_linearize_chain
4536 : would not put on its worklist. */
4537 57307 : && gimple_assign_rhs_code (scalar_stmts[0]->stmt) == (tree_code)code)
4538 : {
4539 56651 : vect_slp_linearize_chain (vinfo, worklist, chain, (tree_code)code,
4540 56651 : scalar_stmts[0]->stmt, op_stmt,
4541 : other_op_stmt,
4542 : NULL);
4543 :
4544 56651 : scalar_stmts.truncate (0);
4545 56651 : stmt_vec_info tail = NULL;
4546 283550 : for (auto el : chain)
4547 : {
4548 113896 : if (el.dt == vect_external_def
4549 113896 : || el.dt == vect_constant_def
4550 113896 : || el.code != (tree_code) code)
4551 : {
4552 299 : scalar_stmts.release ();
4553 299 : return false;
4554 : }
4555 113597 : stmt_vec_info stmt = vinfo->lookup_def (el.op);
4556 113597 : if (STMT_VINFO_REDUC_IDX (stmt) != -1
4557 111313 : || STMT_VINFO_REDUC_DEF (stmt))
4558 : {
4559 56567 : gcc_assert (tail == NULL);
4560 56567 : tail = stmt;
4561 56567 : continue;
4562 : }
4563 57030 : scalar_stmts.safe_push (stmt);
4564 : }
4565 56352 : gcc_assert (tail);
4566 : }
4567 :
4568 : /* When this linearization didn't produce a chain see if stripping
4569 : a wrapping sign conversion produces one. */
4570 57008 : if (scalar_stmts.length () == 1
4571 57008 : && (code == PLUS_EXPR || code == MULT_EXPR || code == BIT_IOR_EXPR
4572 : || code == BIT_AND_EXPR || code == BIT_XOR_EXPR))
4573 : {
4574 55197 : gimple *stmt = scalar_stmts[0]->stmt;
4575 55197 : if (!is_gimple_assign (stmt)
4576 54036 : || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
4577 4704 : || TREE_CODE (gimple_assign_rhs1 (stmt)) != SSA_NAME
4578 59901 : || !tree_nop_conversion_p (TREE_TYPE (gimple_assign_lhs (stmt)),
4579 4704 : TREE_TYPE (gimple_assign_rhs1 (stmt))))
4580 : {
4581 53417 : scalar_stmts.release ();
4582 53417 : return false;
4583 : }
4584 1780 : stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
4585 1780 : if (!is_gimple_assign (stmt)
4586 1780 : || gimple_assign_rhs_code (stmt) != (tree_code)code)
4587 : {
4588 1762 : scalar_stmts.release ();
4589 1762 : return false;
4590 : }
4591 18 : chain.truncate (0);
4592 18 : vect_slp_linearize_chain (vinfo, worklist, chain, (tree_code)code,
4593 : stmt, op_stmt, other_op_stmt, NULL);
4594 :
4595 18 : scalar_stmts.truncate (0);
4596 18 : stmt_vec_info tail = NULL;
4597 88 : for (auto el : chain)
4598 : {
4599 42 : if (el.dt == vect_external_def
4600 42 : || el.dt == vect_constant_def
4601 42 : || el.code != (tree_code) code)
4602 : {
4603 8 : scalar_stmts.release ();
4604 8 : return false;
4605 : }
4606 34 : stmt_vec_info stmt = vinfo->lookup_def (el.op);
4607 34 : if (STMT_VINFO_REDUC_IDX (stmt) != -1
4608 34 : || STMT_VINFO_REDUC_DEF (stmt))
4609 : {
4610 0 : gcc_assert (tail == NULL);
4611 0 : tail = stmt;
4612 0 : continue;
4613 : }
4614 34 : scalar_stmts.safe_push (stmt);
4615 : }
4616 : /* Unlike the above this does not include the reduction SSA
4617 : cycle. */
4618 10 : gcc_assert (!tail);
4619 : }
4620 :
4621 1821 : if (scalar_stmts.length () < 2)
4622 : {
4623 1697 : scalar_stmts.release ();
4624 1697 : return false;
4625 : }
4626 :
4627 124 : if (dump_enabled_p ())
4628 : {
4629 34 : dump_printf_loc (MSG_NOTE, vect_location,
4630 : "Starting SLP discovery of reduction chain for\n");
4631 140 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
4632 212 : dump_printf_loc (MSG_NOTE, vect_location,
4633 106 : " %G", scalar_stmts[i]->stmt);
4634 : }
4635 :
4636 124 : unsigned int group_size = scalar_stmts.length ();
4637 124 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_size);
4638 124 : unsigned tree_size = 0;
4639 124 : slp_tree node = vect_build_slp_tree (vinfo, scalar_stmts, matches, limit,
4640 124 : &tree_size, bst_map);
4641 124 : if (!node)
4642 : {
4643 46 : scalar_stmts.release ();
4644 46 : return false;
4645 : }
4646 :
4647 78 : unsigned cycle_id = vinfo->reduc_infos.length ();
4648 78 : vect_reduc_info reduc_info = new vect_reduc_info_s ();
4649 78 : vinfo->reduc_infos.safe_push (reduc_info);
4650 78 : VECT_REDUC_INFO_DEF_TYPE (reduc_info) = STMT_VINFO_DEF_TYPE (next_stmt);
4651 78 : VECT_REDUC_INFO_TYPE (reduc_info) = STMT_VINFO_REDUC_TYPE (next_stmt);
4652 78 : VECT_REDUC_INFO_CODE (reduc_info) = STMT_VINFO_REDUC_CODE (next_stmt);
4653 78 : VECT_REDUC_INFO_FN (reduc_info) = IFN_LAST;
4654 78 : reduc_info->is_reduc_chain = true;
4655 :
4656 : /* Build the node for the PHI and possibly the conversions. */
4657 78 : slp_tree phis = vect_create_new_slp_node (2, ERROR_MARK);
4658 78 : SLP_TREE_REPRESENTATIVE (phis) = next_stmt;
4659 78 : phis->cycle_info.id = cycle_id;
4660 78 : SLP_TREE_LANES (phis) = group_size;
4661 78 : if (reduc_scalar_stmt == scalar_stmt)
4662 74 : SLP_TREE_VECTYPE (phis) = SLP_TREE_VECTYPE (node);
4663 : else
4664 4 : SLP_TREE_VECTYPE (phis)
4665 4 : = signed_or_unsigned_type_for (TYPE_UNSIGNED
4666 : (TREE_TYPE (gimple_get_lhs
4667 : (scalar_stmt->stmt))),
4668 : SLP_TREE_VECTYPE (node));
4669 : /* ??? vect_cse_slp_nodes cannot cope with cycles without any
4670 : SLP_TREE_SCALAR_STMTS. */
4671 78 : SLP_TREE_SCALAR_STMTS (phis).create (group_size);
4672 471 : for (unsigned i = 0; i < group_size; ++i)
4673 315 : SLP_TREE_SCALAR_STMTS (phis).quick_push (next_stmt);
4674 :
4675 78 : slp_tree op_input = phis;
4676 78 : if (reduc_scalar_stmt != scalar_stmt)
4677 : {
4678 4 : slp_tree conv = vect_create_new_slp_node (1, ERROR_MARK);
4679 4 : SLP_TREE_REPRESENTATIVE (conv)
4680 4 : = vinfo->lookup_def (gimple_arg (reduc_scalar_stmt->stmt,
4681 4 : STMT_VINFO_REDUC_IDX
4682 : (reduc_scalar_stmt)));
4683 4 : SLP_TREE_CHILDREN (conv).quick_push (phis);
4684 4 : conv->cycle_info.id = cycle_id;
4685 4 : SLP_TREE_REDUC_IDX (conv) = 0;
4686 4 : SLP_TREE_LANES (conv) = group_size;
4687 4 : SLP_TREE_VECTYPE (conv) = SLP_TREE_VECTYPE (node);
4688 4 : SLP_TREE_SCALAR_STMTS (conv) = vNULL;
4689 4 : op_input = conv;
4690 : }
4691 :
4692 78 : slp_tree reduc = vect_create_new_slp_node (2, ERROR_MARK);
4693 78 : SLP_TREE_REPRESENTATIVE (reduc) = reduc_scalar_stmt;
4694 78 : SLP_TREE_CHILDREN (reduc).quick_push (op_input);
4695 78 : SLP_TREE_CHILDREN (reduc).quick_push (node);
4696 78 : reduc->cycle_info.id = cycle_id;
4697 78 : SLP_TREE_REDUC_IDX (reduc) = 0;
4698 78 : SLP_TREE_LANES (reduc) = group_size;
4699 78 : SLP_TREE_VECTYPE (reduc) = SLP_TREE_VECTYPE (node);
4700 : /* ??? For the reduction epilogue we need a live lane. */
4701 78 : SLP_TREE_SCALAR_STMTS (reduc).create (group_size);
4702 78 : SLP_TREE_SCALAR_STMTS (reduc).quick_push (reduc_scalar_stmt);
4703 393 : for (unsigned i = 1; i < group_size; ++i)
4704 237 : SLP_TREE_SCALAR_STMTS (reduc).quick_push (NULL);
4705 :
4706 78 : if (reduc_scalar_stmt != scalar_stmt)
4707 : {
4708 4 : slp_tree conv = vect_create_new_slp_node (1, ERROR_MARK);
4709 4 : SLP_TREE_REPRESENTATIVE (conv) = scalar_stmt;
4710 4 : SLP_TREE_CHILDREN (conv).quick_push (reduc);
4711 4 : conv->cycle_info.id = cycle_id;
4712 4 : SLP_TREE_REDUC_IDX (conv) = 0;
4713 4 : SLP_TREE_LANES (conv) = group_size;
4714 4 : SLP_TREE_VECTYPE (conv) = SLP_TREE_VECTYPE (phis);
4715 : /* ??? For the reduction epilogue we need a live lane. */
4716 4 : SLP_TREE_SCALAR_STMTS (conv).create (group_size);
4717 4 : SLP_TREE_SCALAR_STMTS (conv).quick_push (scalar_stmt);
4718 12 : for (unsigned i = 1; i < group_size; ++i)
4719 4 : SLP_TREE_SCALAR_STMTS (conv).quick_push (NULL);
4720 4 : reduc = conv;
4721 : }
4722 :
4723 78 : edge le = loop_latch_edge (LOOP_VINFO_LOOP (vinfo));
4724 78 : SLP_TREE_CHILDREN (phis).quick_push (NULL);
4725 78 : SLP_TREE_CHILDREN (phis).quick_push (NULL);
4726 78 : SLP_TREE_CHILDREN (phis)[le->dest_idx] = reduc;
4727 78 : SLP_TREE_REF_COUNT (reduc)++;
4728 :
4729 : /* Create a new SLP instance. */
4730 78 : slp_instance new_instance = XNEW (class _slp_instance);
4731 78 : SLP_INSTANCE_TREE (new_instance) = reduc;
4732 78 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
4733 78 : SLP_INSTANCE_ROOT_STMTS (new_instance) = vNULL;
4734 78 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = vNULL;
4735 78 : SLP_INSTANCE_KIND (new_instance) = slp_inst_kind_reduc_chain;
4736 78 : new_instance->reduc_phis = NULL;
4737 78 : new_instance->cost_vec = vNULL;
4738 78 : new_instance->subgraph_entries = vNULL;
4739 :
4740 78 : vinfo->slp_instances.safe_push (new_instance);
4741 :
4742 78 : if (dump_enabled_p ())
4743 : {
4744 24 : dump_printf_loc (MSG_NOTE, vect_location,
4745 : "Final SLP tree for instance %p:\n",
4746 : (void *) new_instance);
4747 24 : vect_print_slp_graph (MSG_NOTE, vect_location,
4748 : SLP_INSTANCE_TREE (new_instance));
4749 : }
4750 :
4751 : return true;
4752 57307 : }
4753 :
4754 13428 : if (scalar_stmts.length () <= 1)
4755 : {
4756 10764 : scalar_stmts.release ();
4757 10764 : return false;
4758 : }
4759 :
4760 2664 : scalar_stmts.reverse ();
4761 2664 : stmt_vec_info reduc_phi_info = next_stmt;
4762 :
4763 : /* Build the tree for the SLP instance. */
4764 2664 : vec<stmt_vec_info> root_stmt_infos = vNULL;
4765 2664 : vec<tree> remain = vNULL;
4766 :
4767 2664 : if (dump_enabled_p ())
4768 : {
4769 219 : dump_printf_loc (MSG_NOTE, vect_location,
4770 : "Starting SLP discovery of reduction chain for\n");
4771 1143 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
4772 1848 : dump_printf_loc (MSG_NOTE, vect_location,
4773 924 : " %G", scalar_stmts[i]->stmt);
4774 : }
4775 :
4776 : /* Build the tree for the SLP instance. */
4777 2664 : unsigned int group_size = scalar_stmts.length ();
4778 2664 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_size);
4779 2664 : unsigned tree_size = 0;
4780 :
4781 : /* ??? We need this only for SLP discovery. */
4782 10528 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
4783 7864 : REDUC_GROUP_FIRST_ELEMENT (scalar_stmts[i]) = scalar_stmts[0];
4784 :
4785 2664 : slp_tree node = vect_build_slp_tree (vinfo, scalar_stmts, matches, limit,
4786 2664 : &tree_size, bst_map);
4787 :
4788 10528 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
4789 7864 : REDUC_GROUP_FIRST_ELEMENT (scalar_stmts[i]) = NULL;
4790 :
4791 2664 : if (node != NULL)
4792 : {
4793 : /* Create a new SLP instance. */
4794 2385 : slp_instance new_instance = XNEW (class _slp_instance);
4795 2385 : SLP_INSTANCE_TREE (new_instance) = node;
4796 2385 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
4797 2385 : SLP_INSTANCE_ROOT_STMTS (new_instance) = root_stmt_infos;
4798 2385 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = remain;
4799 2385 : SLP_INSTANCE_KIND (new_instance) = slp_inst_kind_reduc_chain;
4800 2385 : new_instance->reduc_phis = NULL;
4801 2385 : new_instance->cost_vec = vNULL;
4802 2385 : new_instance->subgraph_entries = vNULL;
4803 :
4804 2385 : vect_reduc_info reduc_info = info_for_reduction (vinfo, node);
4805 2385 : reduc_info->is_reduc_chain = true;
4806 :
4807 2385 : if (dump_enabled_p ())
4808 153 : dump_printf_loc (MSG_NOTE, vect_location,
4809 : "SLP size %u vs. limit %u.\n",
4810 : tree_size, max_tree_size);
4811 :
4812 : /* Fixup SLP reduction chains. If this is a reduction chain with
4813 : a conversion in front amend the SLP tree with a node for that. */
4814 2385 : gimple *scalar_def = STMT_VINFO_REDUC_DEF (reduc_phi_info)->stmt;
4815 2385 : if (is_gimple_assign (scalar_def)
4816 2385 : && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (scalar_def)))
4817 : {
4818 43 : stmt_vec_info conv_info = vect_stmt_to_vectorize
4819 43 : (STMT_VINFO_REDUC_DEF (reduc_phi_info));
4820 43 : scalar_stmts = vNULL;
4821 43 : scalar_stmts.create (group_size);
4822 178 : for (unsigned i = 0; i < group_size; ++i)
4823 92 : scalar_stmts.quick_push (conv_info);
4824 43 : slp_tree conv = vect_create_new_slp_node (scalar_stmts, 1);
4825 43 : SLP_TREE_VECTYPE (conv)
4826 43 : = get_vectype_for_scalar_type (vinfo,
4827 43 : TREE_TYPE
4828 : (gimple_assign_lhs (scalar_def)),
4829 : group_size);
4830 43 : SLP_TREE_REDUC_IDX (conv) = 0;
4831 43 : conv->cycle_info.id = node->cycle_info.id;
4832 43 : SLP_TREE_CHILDREN (conv).quick_push (node);
4833 43 : SLP_INSTANCE_TREE (new_instance) = conv;
4834 : }
4835 : /* Fill the backedge child of the PHI SLP node. The
4836 : general matching code cannot find it because the
4837 : scalar code does not reflect how we vectorize the
4838 : reduction. */
4839 2385 : use_operand_p use_p;
4840 2385 : imm_use_iterator imm_iter;
4841 2385 : class loop *loop = LOOP_VINFO_LOOP (vinfo);
4842 9058 : FOR_EACH_IMM_USE_FAST (use_p, imm_iter,
4843 : gimple_get_lhs (scalar_def))
4844 : /* There are exactly two non-debug uses, the reduction
4845 : PHI and the loop-closed PHI node. */
4846 6673 : if (!is_gimple_debug (USE_STMT (use_p))
4847 6673 : && gimple_bb (USE_STMT (use_p)) == loop->header)
4848 : {
4849 2385 : auto_vec<stmt_vec_info, 64> phis (group_size);
4850 2385 : stmt_vec_info phi_info = vinfo->lookup_stmt (USE_STMT (use_p));
4851 9541 : for (unsigned i = 0; i < group_size; ++i)
4852 7156 : phis.quick_push (phi_info);
4853 2385 : slp_tree *phi_node = bst_map->get (phis);
4854 2385 : unsigned dest_idx = loop_latch_edge (loop)->dest_idx;
4855 4770 : SLP_TREE_CHILDREN (*phi_node)[dest_idx]
4856 2385 : = SLP_INSTANCE_TREE (new_instance);
4857 2385 : SLP_INSTANCE_TREE (new_instance)->refcnt++;
4858 2385 : }
4859 :
4860 2385 : vinfo->slp_instances.safe_push (new_instance);
4861 :
4862 : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
4863 : the number of SLP lanes of the root in a few places.
4864 : Verify that assumption holds. */
4865 2385 : gcc_assert (SLP_TREE_LANES (SLP_INSTANCE_TREE (new_instance))
4866 : == group_size);
4867 :
4868 2385 : if (dump_enabled_p ())
4869 : {
4870 153 : dump_printf_loc (MSG_NOTE, vect_location,
4871 : "Final SLP tree for instance %p:\n",
4872 : (void *) new_instance);
4873 153 : vect_print_slp_graph (MSG_NOTE, vect_location,
4874 : SLP_INSTANCE_TREE (new_instance));
4875 : }
4876 :
4877 2385 : return true;
4878 : }
4879 :
4880 : /* Failed to SLP. */
4881 279 : scalar_stmts.release ();
4882 279 : if (dump_enabled_p ())
4883 66 : dump_printf_loc (MSG_NOTE, vect_location,
4884 : "SLP discovery of reduction chain failed\n");
4885 : return false;
4886 : }
4887 :
4888 : /* Analyze an SLP instance starting from SCALAR_STMTS which are a group
4889 : of KIND. Return true if successful. */
4890 :
4891 : static bool
4892 100407 : vect_analyze_slp_reduction (loop_vec_info vinfo,
4893 : stmt_vec_info scalar_stmt,
4894 : unsigned max_tree_size, unsigned *limit,
4895 : scalar_stmts_to_slp_tree_map_t *bst_map,
4896 : bool force_single_lane)
4897 : {
4898 100407 : slp_instance_kind kind = slp_inst_kind_reduc_group;
4899 :
4900 : /* Try to gather a reduction chain. Only attempt if there's budget left
4901 : since chain analysis may build multi-lane trees that consume limit. */
4902 100407 : if (! force_single_lane
4903 73748 : && *limit != 0
4904 73748 : && STMT_VINFO_DEF_TYPE (scalar_stmt) == vect_reduction_def
4905 173870 : && vect_analyze_slp_reduc_chain (vinfo, bst_map, scalar_stmt,
4906 : max_tree_size, limit))
4907 : return true;
4908 :
4909 97944 : vec<stmt_vec_info> scalar_stmts;
4910 97944 : scalar_stmts.create (1);
4911 97944 : scalar_stmts.quick_push (scalar_stmt);
4912 :
4913 97944 : if (dump_enabled_p ())
4914 : {
4915 3926 : dump_printf_loc (MSG_NOTE, vect_location,
4916 : "Starting SLP discovery for\n");
4917 7852 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
4918 7852 : dump_printf_loc (MSG_NOTE, vect_location,
4919 3926 : " %G", scalar_stmts[i]->stmt);
4920 : }
4921 :
4922 : /* Build the tree for the SLP instance. */
4923 97944 : unsigned int group_size = scalar_stmts.length ();
4924 97944 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_size);
4925 97944 : unsigned tree_size = 0;
4926 :
4927 97944 : slp_tree node = vect_build_slp_tree (vinfo, scalar_stmts, matches, limit,
4928 : &tree_size, bst_map);
4929 97944 : if (node != NULL)
4930 : {
4931 : /* Create a new SLP instance. */
4932 96057 : slp_instance new_instance = XNEW (class _slp_instance);
4933 96057 : SLP_INSTANCE_TREE (new_instance) = node;
4934 96057 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
4935 96057 : SLP_INSTANCE_ROOT_STMTS (new_instance) = vNULL;
4936 96057 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = vNULL;
4937 96057 : SLP_INSTANCE_KIND (new_instance) = kind;
4938 96057 : new_instance->reduc_phis = NULL;
4939 96057 : new_instance->cost_vec = vNULL;
4940 96057 : new_instance->subgraph_entries = vNULL;
4941 :
4942 96057 : if (dump_enabled_p ())
4943 3849 : dump_printf_loc (MSG_NOTE, vect_location,
4944 : "SLP size %u vs. limit %u.\n",
4945 : tree_size, max_tree_size);
4946 :
4947 96057 : vinfo->slp_instances.safe_push (new_instance);
4948 :
4949 : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
4950 : the number of SLP lanes of the root in a few places.
4951 : Verify that assumption holds. */
4952 96057 : gcc_assert (SLP_TREE_LANES (SLP_INSTANCE_TREE (new_instance))
4953 : == group_size);
4954 :
4955 96057 : if (dump_enabled_p ())
4956 : {
4957 3849 : dump_printf_loc (MSG_NOTE, vect_location,
4958 : "Final SLP tree for instance %p:\n",
4959 : (void *) new_instance);
4960 3849 : vect_print_slp_graph (MSG_NOTE, vect_location,
4961 : SLP_INSTANCE_TREE (new_instance));
4962 : }
4963 :
4964 96057 : return true;
4965 : }
4966 : /* Failed to SLP. */
4967 :
4968 : /* Free the allocated memory. */
4969 1887 : scalar_stmts.release ();
4970 :
4971 : /* Failed to SLP. */
4972 1887 : if (dump_enabled_p ())
4973 77 : dump_printf_loc (MSG_NOTE, vect_location, "SLP discovery failed\n");
4974 : return false;
4975 : }
4976 :
4977 : /* Analyze a single SLP reduction group. If successful add a SLP instance
4978 : for it and return true, otherwise return false and have *MATCHES
4979 : populated. */
4980 :
4981 : static bool
4982 24242 : vect_analyze_slp_reduction_group (loop_vec_info loop_vinfo,
4983 : vec<stmt_vec_info> scalar_stmts,
4984 : scalar_stmts_to_slp_tree_map_t *bst_map,
4985 : unsigned max_tree_size, unsigned *limit,
4986 : match_elt_t *matches)
4987 : {
4988 : /* Try to form a reduction group. Size-1 groups are not suitable
4989 : for SLP reduction and should fall back to single-lane reduction. */
4990 24242 : unsigned int group_size = scalar_stmts.length ();
4991 24242 : if (group_size <= 1)
4992 : return false;
4993 17523 : if (!matches)
4994 4547 : matches = XALLOCAVEC (match_elt_t, group_size);
4995 17523 : unsigned tree_size = 0;
4996 17523 : slp_tree node = vect_build_slp_tree (loop_vinfo, scalar_stmts, matches, limit,
4997 : &tree_size, bst_map);
4998 17523 : if (!node)
4999 : return false;
5000 :
5001 : /* Create a new SLP instance. */
5002 2784 : slp_instance new_instance = XNEW (class _slp_instance);
5003 2784 : SLP_INSTANCE_TREE (new_instance) = node;
5004 2784 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
5005 2784 : SLP_INSTANCE_ROOT_STMTS (new_instance) = vNULL;
5006 2784 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = vNULL;
5007 2784 : SLP_INSTANCE_KIND (new_instance) = slp_inst_kind_reduc_group;
5008 2784 : new_instance->reduc_phis = NULL;
5009 2784 : new_instance->cost_vec = vNULL;
5010 2784 : new_instance->subgraph_entries = vNULL;
5011 :
5012 2784 : if (dump_enabled_p ())
5013 203 : dump_printf_loc (MSG_NOTE, vect_location,
5014 : "SLP size %u vs. limit %u.\n",
5015 : tree_size, max_tree_size);
5016 :
5017 2784 : loop_vinfo->slp_instances.safe_push (new_instance);
5018 :
5019 : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
5020 : the number of SLP lanes of the root in a few places.
5021 : Verify that assumption holds. */
5022 2784 : gcc_assert (SLP_TREE_LANES (SLP_INSTANCE_TREE (new_instance))
5023 : == group_size);
5024 :
5025 2784 : if (dump_enabled_p ())
5026 : {
5027 203 : dump_printf_loc (MSG_NOTE, vect_location,
5028 : "SLP discovery of size %d reduction group "
5029 : "succeeded\n", group_size);
5030 203 : dump_printf_loc (MSG_NOTE, vect_location,
5031 : "Final SLP tree for instance %p:\n",
5032 : (void *) new_instance);
5033 203 : vect_print_slp_graph (MSG_NOTE, vect_location,
5034 : SLP_INSTANCE_TREE (new_instance));
5035 : }
5036 :
5037 : return true;
5038 : }
5039 :
5040 : /* Analyze reductions in LOOP_VINFO and populate SLP instances
5041 : accordingly. Returns false if something fails. */
5042 :
5043 : static bool
5044 515947 : vect_analyze_slp_reductions (loop_vec_info loop_vinfo,
5045 : unsigned max_tree_size, unsigned *limit,
5046 : scalar_stmts_to_slp_tree_map_t *bst_map,
5047 : bool force_single_lane)
5048 : {
5049 515947 : if (loop_vinfo->reductions.is_empty ())
5050 : return true;
5051 :
5052 : /* Collect reduction statements we can combine into
5053 : a SLP reduction. */
5054 74590 : vec<stmt_vec_info> scalar_stmts;
5055 74590 : scalar_stmts.create (loop_vinfo->reductions.length ());
5056 330609 : for (auto next_info : loop_vinfo->reductions)
5057 : {
5058 106839 : next_info = vect_stmt_to_vectorize (next_info);
5059 106839 : if ((STMT_VINFO_RELEVANT_P (next_info)
5060 14 : || STMT_VINFO_LIVE_P (next_info))
5061 : /* ??? Make sure we didn't skip a conversion around a
5062 : reduction path. In that case we'd have to reverse
5063 : engineer that conversion stmt following the chain using
5064 : reduc_idx and from the PHI using reduc_def. */
5065 106825 : && (STMT_VINFO_DEF_TYPE (next_info) == vect_reduction_def
5066 106825 : || (STMT_VINFO_DEF_TYPE (next_info)
5067 : == vect_double_reduction_def)))
5068 : {
5069 : /* Do not discover SLP reductions combining lane-reducing
5070 : ops, that will fail later. */
5071 106825 : if (!force_single_lane
5072 106825 : && !lane_reducing_stmt_p (STMT_VINFO_STMT (next_info)))
5073 79451 : scalar_stmts.quick_push (next_info);
5074 : /* Do SLP discovery for single-lane reductions. */
5075 27374 : else if (! vect_analyze_slp_reduction (loop_vinfo, next_info,
5076 : max_tree_size, limit,
5077 : bst_map,
5078 : force_single_lane))
5079 : {
5080 0 : scalar_stmts.release ();
5081 0 : return false;
5082 : }
5083 : }
5084 : }
5085 :
5086 74590 : if (scalar_stmts.length () > 1)
5087 : {
5088 : /* Try to form a reduction group. */
5089 4675 : unsigned int group_size = scalar_stmts.length ();
5090 4675 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_size);
5091 4675 : if (vect_analyze_slp_reduction_group (loop_vinfo, scalar_stmts, bst_map,
5092 : max_tree_size, limit, matches))
5093 1601 : return true;
5094 :
5095 : /* When analysis as a single SLP reduction group failed try to
5096 : form sub-groups by collecting matching lanes. Do not recurse
5097 : that on failure (to limit compile-time costs), but recurse
5098 : for the initial non-matching parts. Everything not covered
5099 : by a sub-group gets single-reduction treatment. */
5100 : /* ??? Improve this with the matches[] improvements. */
5101 3526 : vec<stmt_vec_info> cands = vNULL;
5102 11397 : while (matches[0] == 0)
5103 : {
5104 11266 : cands.truncate (0);
5105 11266 : cands.reserve (group_size, true);
5106 99737 : for (unsigned i = 0; i < group_size; ++i)
5107 77205 : if (matches[i] == 0)
5108 19596 : cands.quick_push (scalar_stmts[i]);
5109 :
5110 : /* Try to form a reduction group. */
5111 11266 : if (vect_analyze_slp_reduction_group (loop_vinfo, cands, bst_map,
5112 : max_tree_size, limit, NULL))
5113 1205 : cands = vNULL;
5114 : else
5115 : {
5116 : /* Do SLP discovery for single-lane reductions. */
5117 47319 : for (auto stmt_info : cands)
5118 17158 : if (! vect_analyze_slp_reduction (loop_vinfo,
5119 : vect_stmt_to_vectorize
5120 : (stmt_info),
5121 : max_tree_size, limit,
5122 : bst_map, force_single_lane))
5123 : {
5124 22 : scalar_stmts.release ();
5125 22 : cands.release ();
5126 22 : return false;
5127 : }
5128 : }
5129 : /* Remove the handled stmts from scalar_stmts and try again,
5130 : possibly repeating the above with updated matches[]. */
5131 11244 : unsigned j = 0;
5132 88387 : for (unsigned i = 0; i < group_size; ++i)
5133 77143 : if (matches[i] != 0)
5134 : {
5135 57582 : scalar_stmts[j] = scalar_stmts[i];
5136 57582 : ++j;
5137 : }
5138 11244 : scalar_stmts.truncate (j);
5139 11375 : group_size = scalar_stmts.length ();
5140 11244 : if (group_size <= 1)
5141 : break;
5142 8301 : if (vect_analyze_slp_reduction_group (loop_vinfo, scalar_stmts,
5143 : bst_map, max_tree_size, limit,
5144 : matches))
5145 : return true;
5146 : }
5147 : }
5148 : /* Do SLP discovery for single-lane reductions. */
5149 272977 : for (auto stmt_info : scalar_stmts)
5150 55875 : if (! vect_analyze_slp_reduction (loop_vinfo,
5151 : vect_stmt_to_vectorize (stmt_info),
5152 : max_tree_size, limit,
5153 : bst_map, force_single_lane))
5154 : {
5155 1865 : scalar_stmts.release ();
5156 1865 : return false;
5157 : }
5158 :
5159 71124 : scalar_stmts.release ();
5160 71124 : return true;
5161 : }
5162 :
5163 : /* Analyze an SLP instance starting from a group of grouped stores. Call
5164 : vect_build_slp_tree to build a tree of packed stmts if possible.
5165 : Return FALSE if it's impossible to SLP any stmt in the group. */
5166 :
5167 : static bool
5168 1096340 : vect_analyze_slp_instance (vec_info *vinfo,
5169 : scalar_stmts_to_slp_tree_map_t *bst_map,
5170 : stmt_vec_info stmt_info,
5171 : slp_instance_kind kind,
5172 : unsigned max_tree_size, unsigned *limit,
5173 : bool force_single_lane)
5174 : {
5175 1096340 : vec<stmt_vec_info> scalar_stmts;
5176 :
5177 1096340 : if (is_a <bb_vec_info> (vinfo))
5178 1066109 : vect_location = stmt_info->stmt;
5179 :
5180 1096340 : gcc_assert (kind == slp_inst_kind_store);
5181 :
5182 : /* Collect the stores and store them in scalar_stmts. */
5183 1096340 : scalar_stmts.create (DR_GROUP_SIZE (stmt_info));
5184 1096340 : stmt_vec_info next_info = stmt_info;
5185 5448049 : while (next_info)
5186 : {
5187 3255369 : scalar_stmts.quick_push (vect_stmt_to_vectorize (next_info));
5188 3255369 : next_info = DR_GROUP_NEXT_ELEMENT (next_info);
5189 : }
5190 :
5191 1096340 : vec<stmt_vec_info> root_stmt_infos = vNULL;
5192 1096340 : vec<tree> remain = vNULL;
5193 :
5194 : /* Build the tree for the SLP instance. */
5195 :
5196 : /* If there's no budget left bail out early. */
5197 1096340 : if (*limit == 0)
5198 : return false;
5199 :
5200 1096320 : if (dump_enabled_p ())
5201 : {
5202 4215 : dump_printf_loc (MSG_NOTE, vect_location,
5203 : "Starting SLP discovery for\n");
5204 24525 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
5205 40620 : dump_printf_loc (MSG_NOTE, vect_location,
5206 20310 : " %G", scalar_stmts[i]->stmt);
5207 : }
5208 :
5209 : /* Build the tree for the SLP instance. */
5210 1096320 : unsigned int group_size = scalar_stmts.length ();
5211 1096320 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_size);
5212 1096320 : unsigned tree_size = 0;
5213 1096320 : unsigned i;
5214 :
5215 1096320 : slp_tree node = NULL;
5216 1096320 : if (group_size > 1 && force_single_lane)
5217 : {
5218 1770 : matches[0] = 0;
5219 1770 : matches[1] = -1;
5220 : }
5221 : else
5222 1094550 : node = vect_build_slp_tree (vinfo, scalar_stmts, matches, limit,
5223 : &tree_size, bst_map);
5224 1096320 : if (node != NULL)
5225 : {
5226 : /* Create a new SLP instance. */
5227 696753 : slp_instance new_instance = XNEW (class _slp_instance);
5228 696753 : SLP_INSTANCE_TREE (new_instance) = node;
5229 696753 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
5230 696753 : SLP_INSTANCE_ROOT_STMTS (new_instance) = root_stmt_infos;
5231 696753 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = remain;
5232 696753 : SLP_INSTANCE_KIND (new_instance) = kind;
5233 696753 : new_instance->reduc_phis = NULL;
5234 696753 : new_instance->cost_vec = vNULL;
5235 696753 : new_instance->subgraph_entries = vNULL;
5236 :
5237 696753 : if (dump_enabled_p ())
5238 3211 : dump_printf_loc (MSG_NOTE, vect_location,
5239 : "SLP size %u vs. limit %u.\n",
5240 : tree_size, max_tree_size);
5241 :
5242 696753 : vinfo->slp_instances.safe_push (new_instance);
5243 :
5244 : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
5245 : the number of SLP lanes of the root in a few places.
5246 : Verify that assumption holds. */
5247 696753 : gcc_assert (SLP_TREE_LANES (SLP_INSTANCE_TREE (new_instance))
5248 : == group_size);
5249 :
5250 696753 : if (dump_enabled_p ())
5251 : {
5252 3211 : dump_printf_loc (MSG_NOTE, vect_location,
5253 : "Final SLP tree for instance %p:\n",
5254 : (void *) new_instance);
5255 3211 : vect_print_slp_graph (MSG_NOTE, vect_location,
5256 : SLP_INSTANCE_TREE (new_instance));
5257 : }
5258 696753 : return true;
5259 : }
5260 : /* Failed to SLP. */
5261 :
5262 : /* Try to break the group up into pieces. */
5263 : /* ??? Improve this with the matches[] improvements. */
5264 399567 : if (*limit > 0 && kind == slp_inst_kind_store)
5265 : {
5266 : /* ??? We could delay all the actual splitting of store-groups
5267 : until after SLP discovery of the original group completed.
5268 : Then we can recurse to vect_build_slp_instance directly. */
5269 1004392 : for (i = 0; i < group_size; i++)
5270 1004392 : if (matches[i] != 0)
5271 : break;
5272 :
5273 : /* For basic block SLP, try to break the group up into multiples of
5274 : a vector size. */
5275 399566 : if (is_a <bb_vec_info> (vinfo)
5276 399566 : && (i > 1 && i < group_size))
5277 : {
5278 : /* Free the allocated memory. */
5279 133476 : scalar_stmts.release ();
5280 :
5281 133476 : tree scalar_type
5282 133476 : = TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
5283 266952 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
5284 133476 : 1 << floor_log2 (i));
5285 133476 : unsigned HOST_WIDE_INT const_nunits;
5286 133476 : if (vectype
5287 133476 : && TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits))
5288 : {
5289 : /* Split into two groups at the first vector boundary. */
5290 133476 : gcc_assert ((const_nunits & (const_nunits - 1)) == 0);
5291 133476 : unsigned group1_size = i & ~(const_nunits - 1);
5292 :
5293 133476 : if (dump_enabled_p ())
5294 81 : dump_printf_loc (MSG_NOTE, vect_location,
5295 : "Splitting SLP group at stmt %u\n", i);
5296 133476 : stmt_vec_info rest = vect_split_slp_store_group (stmt_info,
5297 : group1_size);
5298 133476 : bool res = vect_analyze_slp_instance (vinfo, bst_map, stmt_info,
5299 : kind, max_tree_size,
5300 : limit, false);
5301 : /* Split the rest at the failure point and possibly
5302 : re-analyze the remaining matching part if it has
5303 : at least two lanes. */
5304 133476 : if (group1_size < i
5305 5770 : && (i + 1 < group_size
5306 2558 : || i - group1_size > 1))
5307 : {
5308 3248 : stmt_vec_info rest2 = rest;
5309 3248 : rest = vect_split_slp_store_group (rest, i - group1_size);
5310 3248 : if (i - group1_size > 1)
5311 71 : res |= vect_analyze_slp_instance (vinfo, bst_map, rest2,
5312 : kind, max_tree_size,
5313 : limit, false);
5314 : }
5315 : /* Re-analyze the non-matching tail if it has at least
5316 : two lanes. */
5317 133476 : if (i + 1 < group_size)
5318 22355 : res |= vect_analyze_slp_instance (vinfo, bst_map,
5319 : rest, kind, max_tree_size,
5320 : limit, false);
5321 1096340 : return res;
5322 : }
5323 : }
5324 :
5325 : /* For loop vectorization split the RHS into arbitrary pieces of
5326 : size >= 1. */
5327 266090 : else if (is_a <loop_vec_info> (vinfo)
5328 266090 : && (group_size != 1 && i < group_size))
5329 : {
5330 8378 : gcall *call = dyn_cast <gcall *> (stmt_info->stmt);
5331 28 : bool masked_p = call
5332 28 : && gimple_call_internal_p (call)
5333 28 : && internal_fn_mask_index (gimple_call_internal_fn (call)) != -1;
5334 : /* There are targets that cannot do even/odd interleaving schemes
5335 : so they absolutely need to use load/store-lanes. For now
5336 : force single-lane SLP for them - they would be happy with
5337 : uniform power-of-two lanes (but depending on element size),
5338 : but even if we can use 'i' as indicator we would need to
5339 : backtrack when later lanes fail to discover with the same
5340 : granularity. We cannot turn any of strided or scatter store
5341 : into store-lanes. */
5342 : /* ??? If this is not in sync with what get_load_store_type
5343 : later decides the SLP representation is not good for other
5344 : store vectorization methods. */
5345 8378 : bool want_store_lanes
5346 8378 : = (! STMT_VINFO_GATHER_SCATTER_P (stmt_info)
5347 8378 : && ! STMT_VINFO_STRIDED_P (stmt_info)
5348 6269 : && ! STMT_VINFO_SLP_VECT_ONLY (stmt_info)
5349 6265 : && compare_step_with_zero (vinfo, stmt_info) > 0
5350 14538 : && vect_slp_prefer_store_lanes_p (vinfo, stmt_info, NULL_TREE,
5351 16756 : masked_p, group_size, i));
5352 8378 : if (want_store_lanes || force_single_lane)
5353 : i = 1;
5354 :
5355 : /* A fatal discovery fail doesn't always mean single-lane SLP
5356 : isn't a possibility, so try. */
5357 6608 : if (i == 0)
5358 : i = 1;
5359 :
5360 8378 : if (dump_enabled_p ())
5361 873 : dump_printf_loc (MSG_NOTE, vect_location,
5362 : "Splitting SLP group at stmt %u\n", i);
5363 :
5364 : /* Analyze the stored values and pinch them together with
5365 : a permute node so we can preserve the whole store group. */
5366 8378 : auto_vec<slp_tree> rhs_nodes;
5367 :
5368 8378 : unsigned int rhs_common_nlanes = 0;
5369 8378 : unsigned int start = 0, end = i;
5370 38420 : while (start < group_size)
5371 : {
5372 30157 : gcc_assert (end - start >= 1);
5373 30157 : vec<stmt_vec_info> substmts;
5374 30157 : substmts.create (end - start);
5375 126654 : for (unsigned j = start; j < end; ++j)
5376 66340 : substmts.quick_push (scalar_stmts[j]);
5377 30157 : node = vect_build_slp_tree (vinfo, substmts,
5378 : matches, limit, &tree_size, bst_map);
5379 30157 : if (node)
5380 : {
5381 24073 : rhs_nodes.safe_push (node);
5382 24073 : if (start == 0)
5383 8263 : rhs_common_nlanes = SLP_TREE_LANES (node);
5384 15810 : else if (rhs_common_nlanes != SLP_TREE_LANES (node))
5385 1413 : rhs_common_nlanes = 0;
5386 24073 : start = end;
5387 24073 : if (want_store_lanes || force_single_lane)
5388 5321 : end = start + 1;
5389 : else
5390 : end = group_size;
5391 : }
5392 : else
5393 : {
5394 6084 : substmts.release ();
5395 6084 : if (end - start == 1)
5396 : {
5397 : /* Single-lane discovery failed. Free resources. */
5398 115 : for (auto node : rhs_nodes)
5399 0 : vect_free_slp_tree (node);
5400 115 : scalar_stmts.release ();
5401 115 : if (dump_enabled_p ())
5402 14 : dump_printf_loc (MSG_NOTE, vect_location,
5403 : "SLP discovery failed\n");
5404 115 : return false;
5405 : }
5406 :
5407 : /* ??? It really happens that we soft-fail SLP
5408 : build at a mismatch but the matching part hard-fails
5409 : later. As we know we arrived here with a group
5410 : larger than one try a group of size one! */
5411 5969 : if (matches[0] != 0)
5412 12 : end = start + 1;
5413 : else
5414 13085 : for (unsigned j = start; j < end; j++)
5415 13085 : if (matches[j - start] != 0)
5416 : {
5417 : end = j;
5418 : break;
5419 : }
5420 : }
5421 : }
5422 :
5423 : /* Now re-assess whether we want store lanes in case the
5424 : discovery ended up producing all single-lane RHSs. */
5425 8263 : if (! want_store_lanes
5426 8263 : && rhs_common_nlanes == 1
5427 7183 : && ! STMT_VINFO_GATHER_SCATTER_P (stmt_info)
5428 7183 : && ! STMT_VINFO_STRIDED_P (stmt_info)
5429 5445 : && ! STMT_VINFO_SLP_VECT_ONLY (stmt_info)
5430 5442 : && compare_step_with_zero (vinfo, stmt_info) > 0
5431 13618 : && (vect_store_lanes_supported (SLP_TREE_VECTYPE (rhs_nodes[0]),
5432 : group_size, masked_p)
5433 : != IFN_LAST))
5434 : want_store_lanes = true;
5435 :
5436 : /* Now we assume we can build the root SLP node from all stores. */
5437 8263 : if (want_store_lanes)
5438 : {
5439 : /* For store-lanes feed the store node with all RHS nodes
5440 : in order. */
5441 0 : node = vect_create_new_slp_node (scalar_stmts,
5442 0 : SLP_TREE_CHILDREN
5443 : (rhs_nodes[0]).length ());
5444 0 : SLP_TREE_VECTYPE (node) = SLP_TREE_VECTYPE (rhs_nodes[0]);
5445 0 : node->ldst_lanes = true;
5446 0 : SLP_TREE_CHILDREN (node)
5447 0 : .reserve_exact (SLP_TREE_CHILDREN (rhs_nodes[0]).length ()
5448 0 : + rhs_nodes.length () - 1);
5449 : /* First store value and possibly mask. */
5450 0 : SLP_TREE_CHILDREN (node)
5451 0 : .splice (SLP_TREE_CHILDREN (rhs_nodes[0]));
5452 : /* Rest of the store values. All mask nodes are the same,
5453 : this should be guaranteed by dataref group discovery. */
5454 0 : for (unsigned j = 1; j < rhs_nodes.length (); ++j)
5455 0 : SLP_TREE_CHILDREN (node)
5456 0 : .quick_push (SLP_TREE_CHILDREN (rhs_nodes[j])[0]);
5457 0 : for (slp_tree child : SLP_TREE_CHILDREN (node))
5458 0 : child->refcnt++;
5459 : }
5460 : else
5461 8263 : node = vect_build_slp_store_interleaving (rhs_nodes, scalar_stmts);
5462 :
5463 32336 : while (!rhs_nodes.is_empty ())
5464 24073 : vect_free_slp_tree (rhs_nodes.pop ());
5465 :
5466 : /* Create a new SLP instance. */
5467 8263 : slp_instance new_instance = XNEW (class _slp_instance);
5468 8263 : SLP_INSTANCE_TREE (new_instance) = node;
5469 8263 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
5470 8263 : SLP_INSTANCE_ROOT_STMTS (new_instance) = root_stmt_infos;
5471 8263 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = remain;
5472 8263 : SLP_INSTANCE_KIND (new_instance) = kind;
5473 8263 : new_instance->reduc_phis = NULL;
5474 8263 : new_instance->cost_vec = vNULL;
5475 8263 : new_instance->subgraph_entries = vNULL;
5476 :
5477 8263 : if (dump_enabled_p ())
5478 859 : dump_printf_loc (MSG_NOTE, vect_location,
5479 : "SLP size %u vs. limit %u.\n",
5480 : tree_size, max_tree_size);
5481 :
5482 8263 : vinfo->slp_instances.safe_push (new_instance);
5483 :
5484 : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
5485 : the number of SLP lanes of the root in a few places.
5486 : Verify that assumption holds. */
5487 8263 : gcc_assert (SLP_TREE_LANES (SLP_INSTANCE_TREE (new_instance))
5488 : == group_size);
5489 :
5490 8263 : if (dump_enabled_p ())
5491 : {
5492 859 : dump_printf_loc (MSG_NOTE, vect_location,
5493 : "Final SLP tree for instance %p:\n",
5494 : (void *) new_instance);
5495 859 : vect_print_slp_graph (MSG_NOTE, vect_location,
5496 : SLP_INSTANCE_TREE (new_instance));
5497 : }
5498 : return true;
5499 8378 : }
5500 : else
5501 : /* Free the allocated memory. */
5502 257712 : scalar_stmts.release ();
5503 :
5504 : /* Even though the first vector did not all match, we might be able to SLP
5505 : (some) of the remainder. FORNOW ignore this possibility. */
5506 : }
5507 : else
5508 : /* Free the allocated memory. */
5509 1 : scalar_stmts.release ();
5510 :
5511 : /* Failed to SLP. */
5512 257713 : if (dump_enabled_p ())
5513 50 : dump_printf_loc (MSG_NOTE, vect_location, "SLP discovery failed\n");
5514 : return false;
5515 : }
5516 :
5517 : /* qsort comparator ordering SLP load nodes. */
5518 :
5519 : static int
5520 2690328 : vllp_cmp (const void *a_, const void *b_)
5521 : {
5522 2690328 : const slp_tree a = *(const slp_tree *)a_;
5523 2690328 : const slp_tree b = *(const slp_tree *)b_;
5524 2690328 : stmt_vec_info a0 = SLP_TREE_SCALAR_STMTS (a)[0];
5525 2690328 : stmt_vec_info b0 = SLP_TREE_SCALAR_STMTS (b)[0];
5526 2690328 : if (STMT_VINFO_GROUPED_ACCESS (a0)
5527 1547899 : && STMT_VINFO_GROUPED_ACCESS (b0)
5528 4176179 : && DR_GROUP_FIRST_ELEMENT (a0) == DR_GROUP_FIRST_ELEMENT (b0))
5529 : {
5530 : /* Same group, order after lanes used. */
5531 350037 : if (SLP_TREE_LANES (a) < SLP_TREE_LANES (b))
5532 : return 1;
5533 341064 : else if (SLP_TREE_LANES (a) > SLP_TREE_LANES (b))
5534 : return -1;
5535 : else
5536 : {
5537 : /* Try to order loads using the same lanes together, breaking
5538 : the tie with the lane number that first differs. */
5539 331327 : if (!SLP_TREE_LOAD_PERMUTATION (a).exists ()
5540 331327 : && !SLP_TREE_LOAD_PERMUTATION (b).exists ())
5541 : return 0;
5542 331327 : else if (SLP_TREE_LOAD_PERMUTATION (a).exists ()
5543 331327 : && !SLP_TREE_LOAD_PERMUTATION (b).exists ())
5544 : return 1;
5545 327260 : else if (!SLP_TREE_LOAD_PERMUTATION (a).exists ()
5546 327260 : && SLP_TREE_LOAD_PERMUTATION (b).exists ())
5547 : return -1;
5548 : else
5549 : {
5550 319770 : for (unsigned i = 0; i < SLP_TREE_LANES (a); ++i)
5551 319770 : if (SLP_TREE_LOAD_PERMUTATION (a)[i]
5552 319770 : != SLP_TREE_LOAD_PERMUTATION (b)[i])
5553 : {
5554 : /* In-order lane first, that's what the above case for
5555 : no permutation does. */
5556 318458 : if (SLP_TREE_LOAD_PERMUTATION (a)[i] == i)
5557 : return -1;
5558 195596 : else if (SLP_TREE_LOAD_PERMUTATION (b)[i] == i)
5559 : return 1;
5560 103282 : else if (SLP_TREE_LOAD_PERMUTATION (a)[i]
5561 103282 : < SLP_TREE_LOAD_PERMUTATION (b)[i])
5562 : return -1;
5563 : else
5564 38668 : return 1;
5565 : }
5566 : return 0;
5567 : }
5568 : }
5569 : }
5570 : else /* Different groups or non-groups. */
5571 : {
5572 : /* Order groups as their first element to keep them together. */
5573 2340291 : if (STMT_VINFO_GROUPED_ACCESS (a0))
5574 2340291 : a0 = DR_GROUP_FIRST_ELEMENT (a0);
5575 2340291 : if (STMT_VINFO_GROUPED_ACCESS (b0))
5576 2340291 : b0 = DR_GROUP_FIRST_ELEMENT (b0);
5577 2340291 : if (a0 == b0)
5578 : return 0;
5579 : /* Tie using UID. */
5580 2340171 : else if (gimple_uid (STMT_VINFO_STMT (a0))
5581 2340171 : < gimple_uid (STMT_VINFO_STMT (b0)))
5582 : return -1;
5583 : else
5584 : {
5585 1040005 : gcc_assert (gimple_uid (STMT_VINFO_STMT (a0))
5586 : != gimple_uid (STMT_VINFO_STMT (b0)));
5587 : return 1;
5588 : }
5589 : }
5590 : }
5591 :
5592 : /* Return whether if the load permutation of NODE is consecutive starting
5593 : with value START_VAL in the first element. If START_VAL is not given
5594 : the first element's value is used. */
5595 :
5596 : bool
5597 649513 : vect_load_perm_consecutive_p (slp_tree node, unsigned start_val)
5598 : {
5599 649513 : load_permutation_t perm = SLP_TREE_LOAD_PERMUTATION (node);
5600 :
5601 649513 : if (!perm.exists () || !perm.length ())
5602 : return false;
5603 :
5604 649513 : if (start_val == UINT_MAX)
5605 79957 : start_val = perm[0];
5606 :
5607 1282956 : for (unsigned int i = 0; i < perm.length (); i++)
5608 657139 : if (perm[i] != start_val + (unsigned int) i)
5609 : return false;
5610 :
5611 : return true;
5612 : }
5613 :
5614 : /* Process the set of LOADS that are all from the same dataref group. */
5615 :
5616 : static void
5617 162196 : vect_lower_load_permutations (loop_vec_info loop_vinfo,
5618 : scalar_stmts_to_slp_tree_map_t *bst_map,
5619 : const array_slice<slp_tree> &loads,
5620 : bool force_single_lane)
5621 : {
5622 : /* We at this point want to lower without a fixed VF or vector
5623 : size in mind which means we cannot actually compute whether we
5624 : need three or more vectors for a load permutation yet. So always
5625 : lower. */
5626 162196 : stmt_vec_info first
5627 162196 : = DR_GROUP_FIRST_ELEMENT (SLP_TREE_SCALAR_STMTS (loads[0])[0]);
5628 162196 : unsigned group_lanes = DR_GROUP_SIZE (first);
5629 :
5630 : /* Verify if all load permutations can be implemented with a suitably
5631 : large element load-lanes operation. */
5632 162196 : unsigned ld_lanes_lanes = SLP_TREE_LANES (loads[0]);
5633 162196 : if (STMT_VINFO_STRIDED_P (first)
5634 159718 : || compare_step_with_zero (loop_vinfo, first) <= 0
5635 157086 : || exact_log2 (ld_lanes_lanes) == -1
5636 : /* ??? For now only support the single-lane case as there is
5637 : missing support on the store-lane side and code generation
5638 : isn't up to the task yet. */
5639 154243 : || ld_lanes_lanes != 1
5640 305377 : || vect_load_lanes_supported (SLP_TREE_VECTYPE (loads[0]),
5641 : group_lanes / ld_lanes_lanes,
5642 : false) == IFN_LAST)
5643 : ld_lanes_lanes = 0;
5644 : else
5645 : /* Verify the loads access the same number of lanes aligned to
5646 : ld_lanes_lanes. */
5647 0 : for (slp_tree load : loads)
5648 : {
5649 0 : if (SLP_TREE_LANES (load) != ld_lanes_lanes)
5650 : {
5651 : ld_lanes_lanes = 0;
5652 : break;
5653 : }
5654 0 : unsigned first = SLP_TREE_LOAD_PERMUTATION (load)[0];
5655 0 : if (first % ld_lanes_lanes != 0)
5656 : {
5657 : ld_lanes_lanes = 0;
5658 : break;
5659 : }
5660 0 : if (!vect_load_perm_consecutive_p (load))
5661 : {
5662 : ld_lanes_lanes = 0;
5663 : break;
5664 : }
5665 : }
5666 :
5667 : /* Only a power-of-two number of lanes matches interleaving with N levels.
5668 : ??? An even number of lanes could be reduced to 1<<ceil_log2(N)-1 lanes
5669 : at each step. */
5670 263009 : if (ld_lanes_lanes == 0 && exact_log2 (group_lanes) == -1 && group_lanes != 3)
5671 : return;
5672 :
5673 267060 : for (slp_tree load : loads)
5674 : {
5675 : /* Leave masked or gather loads alone for now. */
5676 188455 : if (!SLP_TREE_CHILDREN (load).is_empty ())
5677 61169 : continue;
5678 :
5679 : /* For single-element interleaving spanning multiple vectors avoid
5680 : lowering, we want to use VMAT_ELEMENTWISE later. */
5681 188449 : if (ld_lanes_lanes == 0
5682 188449 : && SLP_TREE_LANES (load) == 1
5683 168919 : && !DR_GROUP_NEXT_ELEMENT (first)
5684 268331 : && maybe_gt (group_lanes,
5685 : TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (load))))
5686 51387 : return;
5687 :
5688 : /* We want to pattern-match special cases here and keep those
5689 : alone. Candidates are splats and load-lane. */
5690 :
5691 : /* We need to lower only loads of less than half of the groups
5692 : lanes, including duplicate lanes. Note this leaves nodes
5693 : with a non-1:1 load permutation around instead of canonicalizing
5694 : those into a load and a permute node. Removing this early
5695 : check would do such canonicalization. */
5696 137062 : if (SLP_TREE_LANES (load) >= (group_lanes + 1) / 2
5697 57594 : && ld_lanes_lanes == 0)
5698 57594 : continue;
5699 :
5700 : /* Build the permute to get the original load permutation order. */
5701 79468 : bool contiguous = vect_load_perm_consecutive_p (load);
5702 79468 : lane_permutation_t final_perm;
5703 79468 : final_perm.create (SLP_TREE_LANES (load));
5704 239348 : for (unsigned i = 0; i < SLP_TREE_LANES (load); ++i)
5705 160824 : final_perm.quick_push (
5706 80412 : std::make_pair (0, SLP_TREE_LOAD_PERMUTATION (load)[i]));
5707 :
5708 : /* When the load permutation accesses a contiguous unpermuted,
5709 : power-of-two aligned and sized chunk leave the load alone.
5710 : We can likely (re-)load it more efficiently rather than
5711 : extracting it from the larger load.
5712 : ??? Long-term some of the lowering should move to where
5713 : the vector types involved are fixed. */
5714 83037 : if (!force_single_lane
5715 79468 : && ld_lanes_lanes == 0
5716 53591 : && contiguous
5717 53331 : && (SLP_TREE_LANES (load) > 1 || loads.size () == 1)
5718 6575 : && pow2p_hwi (SLP_TREE_LANES (load))
5719 6539 : && pow2p_hwi (group_lanes)
5720 3569 : && SLP_TREE_LOAD_PERMUTATION (load)[0] % SLP_TREE_LANES (load) == 0
5721 83037 : && group_lanes % SLP_TREE_LANES (load) == 0)
5722 : {
5723 3569 : final_perm.release ();
5724 3569 : continue;
5725 : }
5726 :
5727 : /* First build (and possibly re-use) a load node for the
5728 : unpermuted group. Gaps in the middle and on the end are
5729 : represented with NULL stmts. */
5730 75899 : vec<stmt_vec_info> stmts;
5731 75899 : stmts.create (group_lanes);
5732 270993 : for (stmt_vec_info s = first; s; s = DR_GROUP_NEXT_ELEMENT (s))
5733 : {
5734 195094 : if (s != first)
5735 124250 : for (unsigned i = 1; i < DR_GROUP_GAP (s); ++i)
5736 5055 : stmts.quick_push (NULL);
5737 195094 : stmts.quick_push (s);
5738 : }
5739 139731 : for (unsigned i = 0; i < DR_GROUP_GAP (first); ++i)
5740 63832 : stmts.quick_push (NULL);
5741 75899 : match_elt_t *matches = XALLOCAVEC (match_elt_t, group_lanes);
5742 75899 : unsigned limit = 1;
5743 75899 : unsigned tree_size = 0;
5744 75899 : slp_tree l0 = vect_build_slp_tree (loop_vinfo, stmts, matches, &limit,
5745 75899 : &tree_size, bst_map);
5746 75899 : gcc_assert (!SLP_TREE_LOAD_PERMUTATION (l0).exists ());
5747 :
5748 75899 : if (ld_lanes_lanes != 0)
5749 : {
5750 : /* ??? If this is not in sync with what get_load_store_type
5751 : later decides the SLP representation is not good for other
5752 : store vectorization methods. */
5753 0 : l0->ldst_lanes = true;
5754 0 : load->ldst_lanes = true;
5755 : }
5756 :
5757 236343 : while (1)
5758 : {
5759 156121 : unsigned group_lanes = SLP_TREE_LANES (l0);
5760 156121 : if (ld_lanes_lanes != 0
5761 156121 : || SLP_TREE_LANES (load) >= (group_lanes + 1) / 2)
5762 : break;
5763 :
5764 : /* Try to lower by reducing the group to half its size using an
5765 : interleaving scheme. For this try to compute whether all
5766 : elements needed for this load are in even or odd elements of
5767 : an even/odd decomposition with N consecutive elements.
5768 : Thus { e, e, o, o, e, e, o, o } would be an even/odd decomposition
5769 : with N == 2. */
5770 : /* ??? Only an even number of lanes can be handed this way, but the
5771 : fallback below could work for any number. We have to make sure
5772 : to round up in that case. */
5773 80222 : gcc_assert ((group_lanes & 1) == 0 || group_lanes == 3);
5774 12119 : unsigned even = 0, odd = 0;
5775 12119 : if ((group_lanes & 1) == 0)
5776 : {
5777 12119 : even = (1 << ceil_log2 (group_lanes)) - 1;
5778 12119 : odd = even;
5779 49225 : for (auto l : final_perm)
5780 : {
5781 12868 : even &= ~l.second;
5782 12868 : odd &= l.second;
5783 : }
5784 : }
5785 :
5786 : /* Now build an even or odd extraction from the unpermuted load. */
5787 80222 : lane_permutation_t perm;
5788 80222 : perm.create ((group_lanes + 1) / 2);
5789 80222 : unsigned even_level = even ? 1 << ctz_hwi (even) : 0;
5790 80222 : unsigned odd_level = odd ? 1 << ctz_hwi (odd) : 0;
5791 80222 : if (even_level
5792 11141 : && group_lanes % (2 * even_level) == 0
5793 : /* ??? When code generating permutes we do not try to pun
5794 : to larger component modes so level != 1 isn't a natural
5795 : even/odd extract. Prefer one if possible. */
5796 11141 : && (even_level == 1 || !odd_level || odd_level != 1))
5797 : {
5798 : /* { 0, 1, ... 4, 5 ..., } */
5799 39435 : for (unsigned i = 0; i < group_lanes / 2 / even_level; ++i)
5800 62245 : for (unsigned j = 0; j < even_level; ++j)
5801 31318 : perm.quick_push (std::make_pair (0, 2 * i * even_level + j));
5802 : }
5803 69081 : else if (odd_level)
5804 : {
5805 : /* { ..., 2, 3, ... 6, 7 } */
5806 3581 : gcc_assert (group_lanes % (2 * odd_level) == 0);
5807 15445 : for (unsigned i = 0; i < group_lanes / 2 / odd_level; ++i)
5808 23782 : for (unsigned j = 0; j < odd_level; ++j)
5809 11918 : perm.quick_push
5810 11918 : (std::make_pair (0, (2 * i + 1) * odd_level + j));
5811 : }
5812 : else
5813 : {
5814 : /* As fallback extract all used lanes and fill to half the
5815 : group size by repeating the last element.
5816 : ??? This is quite a bad strathegy for re-use - we could
5817 : brute force our way to find more optimal filling lanes to
5818 : maximize re-use when looking at all loads from the group. */
5819 68133 : auto_bitmap l;
5820 272588 : for (auto p : final_perm)
5821 68189 : bitmap_set_bit (l, p.second);
5822 68133 : unsigned i = 0;
5823 68133 : bitmap_iterator bi;
5824 136322 : EXECUTE_IF_SET_IN_BITMAP (l, 0, i, bi)
5825 68189 : perm.quick_push (std::make_pair (0, i));
5826 272684 : while (perm.length () < (group_lanes + 1) / 2)
5827 68209 : perm.quick_push (perm.last ());
5828 68133 : }
5829 :
5830 : /* Update final_perm with the intermediate permute. */
5831 161193 : for (unsigned i = 0; i < final_perm.length (); ++i)
5832 : {
5833 80971 : unsigned l = final_perm[i].second;
5834 80971 : unsigned j;
5835 90542 : for (j = 0; j < perm.length (); ++j)
5836 90542 : if (perm[j].second == l)
5837 : {
5838 80971 : final_perm[i].second = j;
5839 80971 : break;
5840 : }
5841 80971 : gcc_assert (j < perm.length ());
5842 : }
5843 :
5844 : /* And create scalar stmts. */
5845 80222 : vec<stmt_vec_info> perm_stmts;
5846 80222 : perm_stmts.create (perm.length ());
5847 340078 : for (unsigned i = 0; i < perm.length (); ++i)
5848 179634 : perm_stmts.quick_push (SLP_TREE_SCALAR_STMTS (l0)[perm[i].second]);
5849 :
5850 80222 : slp_tree p = vect_create_new_slp_node (1, VEC_PERM_EXPR);
5851 80222 : SLP_TREE_CHILDREN (p).quick_push (l0);
5852 80222 : SLP_TREE_LANE_PERMUTATION (p) = perm;
5853 80222 : SLP_TREE_VECTYPE (p) = SLP_TREE_VECTYPE (load);
5854 80222 : SLP_TREE_LANES (p) = perm.length ();
5855 : /* ??? As we have scalar stmts for this intermediate permute we
5856 : could CSE it via bst_map but we do not want to pick up
5857 : another SLP node with a load permutation. We instead should
5858 : have a "local" CSE map here. */
5859 80222 : SLP_TREE_SCALAR_STMTS (p) = perm_stmts;
5860 :
5861 : /* We now have a node for (group_lanes + 1) / 2 lanes. */
5862 80222 : l0 = p;
5863 80222 : }
5864 :
5865 : /* And finally from the ordered reduction node create the
5866 : permute to shuffle the lanes into the original load-permutation
5867 : order. We replace the original load node with this. */
5868 75899 : SLP_TREE_CODE (load) = VEC_PERM_EXPR;
5869 75899 : SLP_TREE_LOAD_PERMUTATION (load).release ();
5870 75899 : SLP_TREE_LANE_PERMUTATION (load) = final_perm;
5871 75899 : SLP_TREE_CHILDREN (load).create (1);
5872 75899 : SLP_TREE_CHILDREN (load).quick_push (l0);
5873 75899 : SLP_TREE_REPRESENTATIVE (load) = NULL;
5874 : }
5875 : }
5876 :
5877 : /* Transform SLP loads in the SLP graph created by SLP discovery to
5878 : group loads from the same group and lower load permutations that
5879 : are unlikely to be supported into a series of permutes.
5880 : In the degenerate case of having only single-lane SLP instances
5881 : this should result in a series of permute nodes emulating an
5882 : interleaving scheme. */
5883 :
5884 : static void
5885 501509 : vect_lower_load_permutations (loop_vec_info loop_vinfo,
5886 : scalar_stmts_to_slp_tree_map_t *bst_map,
5887 : bool force_single_lane)
5888 : {
5889 : /* Gather and sort loads across all instances. */
5890 501509 : hash_set<slp_tree> visited;
5891 501509 : auto_vec<slp_tree> loads;
5892 2301121 : for (auto inst : loop_vinfo->slp_instances)
5893 798564 : vect_gather_slp_loads (loads, SLP_INSTANCE_TREE (inst), visited);
5894 501509 : if (loads.is_empty ())
5895 93441 : return;
5896 408068 : loads.qsort (vllp_cmp);
5897 :
5898 : /* Now process each dataref group separately. */
5899 408068 : unsigned firsti = 0;
5900 1160895 : for (unsigned i = 1; i < loads.length (); ++i)
5901 : {
5902 344759 : slp_tree first = loads[firsti];
5903 344759 : slp_tree next = loads[i];
5904 344759 : stmt_vec_info a0 = SLP_TREE_SCALAR_STMTS (first)[0];
5905 344759 : stmt_vec_info b0 = SLP_TREE_SCALAR_STMTS (next)[0];
5906 344759 : if (STMT_VINFO_GROUPED_ACCESS (a0)
5907 158976 : && STMT_VINFO_GROUPED_ACCESS (b0)
5908 490541 : && DR_GROUP_FIRST_ELEMENT (a0) == DR_GROUP_FIRST_ELEMENT (b0))
5909 63652 : continue;
5910 : /* Now we have one or multiple SLP loads of the same group from
5911 : firsti to i - 1. */
5912 281107 : if (STMT_VINFO_GROUPED_ACCESS (a0))
5913 95324 : vect_lower_load_permutations (loop_vinfo, bst_map,
5914 95324 : make_array_slice (&loads[firsti],
5915 : i - firsti),
5916 : force_single_lane);
5917 : firsti = i;
5918 : }
5919 816136 : if (firsti < loads.length ()
5920 816136 : && STMT_VINFO_GROUPED_ACCESS (SLP_TREE_SCALAR_STMTS (loads[firsti])[0]))
5921 66872 : vect_lower_load_permutations (loop_vinfo, bst_map,
5922 66872 : make_array_slice (&loads[firsti],
5923 66872 : loads.length () - firsti),
5924 : force_single_lane);
5925 501509 : }
5926 :
5927 : /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP
5928 : trees of packed scalar stmts if SLP is possible. */
5929 :
5930 : opt_result
5931 1180843 : vect_analyze_slp (vec_info *vinfo, unsigned max_tree_size,
5932 : bool force_single_lane)
5933 : {
5934 1180843 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
5935 1180843 : unsigned int i;
5936 1180843 : stmt_vec_info first_element;
5937 1180843 : slp_instance instance;
5938 :
5939 1180843 : DUMP_VECT_SCOPE ("vect_analyze_slp");
5940 :
5941 1180843 : unsigned limit = max_tree_size;
5942 :
5943 1180843 : scalar_stmts_to_slp_tree_map_t *bst_map
5944 1180843 : = new scalar_stmts_to_slp_tree_map_t ();
5945 :
5946 : /* Find SLP sequences starting from groups of grouped stores. */
5947 3302001 : FOR_EACH_VEC_ELT (vinfo->grouped_stores, i, first_element)
5948 940438 : if (! vect_analyze_slp_instance (vinfo, bst_map, first_element,
5949 : slp_inst_kind_store, max_tree_size, &limit,
5950 : force_single_lane)
5951 940438 : && loop_vinfo)
5952 : {
5953 123 : release_scalar_stmts_to_slp_tree_map (bst_map);
5954 123 : return opt_result::failure_at (vect_location, "SLP build failed.\n");
5955 : }
5956 :
5957 : /* For loops also start SLP discovery from non-grouped stores. */
5958 1180720 : if (loop_vinfo)
5959 : {
5960 : data_reference_p dr;
5961 1690933 : FOR_EACH_VEC_ELT (vinfo->shared->datarefs, i, dr)
5962 1174986 : if (DR_IS_WRITE (dr))
5963 : {
5964 380112 : stmt_vec_info stmt_info = vinfo->lookup_dr (dr)->stmt;
5965 : /* Grouped stores are already handled above. */
5966 380112 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
5967 103274 : continue;
5968 276838 : vec<stmt_vec_info> stmts;
5969 276838 : vec<stmt_vec_info> roots = vNULL;
5970 276838 : vec<tree> remain = vNULL;
5971 276838 : stmts.create (1);
5972 276838 : stmts.quick_push (stmt_info);
5973 276838 : if (! vect_build_slp_instance (vinfo, slp_inst_kind_store,
5974 : stmts, roots, remain, max_tree_size,
5975 : &limit, bst_map, force_single_lane))
5976 : {
5977 3592 : release_scalar_stmts_to_slp_tree_map (bst_map);
5978 3592 : return opt_result::failure_at (vect_location,
5979 : "SLP build failed.\n");
5980 : }
5981 : }
5982 :
5983 : stmt_vec_info stmt_info;
5984 515987 : FOR_EACH_VEC_ELT (LOOP_VINFO_ALTERNATE_DEFS (loop_vinfo), i, stmt_info)
5985 : {
5986 20 : vec<stmt_vec_info> stmts;
5987 20 : vec<stmt_vec_info> roots = vNULL;
5988 20 : vec<tree> remain = vNULL;
5989 20 : stmts.create (1);
5990 20 : stmts.quick_push (stmt_info);
5991 20 : if (! vect_build_slp_instance (vinfo, slp_inst_kind_store,
5992 : stmts, roots, remain, max_tree_size,
5993 : &limit, bst_map, force_single_lane))
5994 : {
5995 0 : release_scalar_stmts_to_slp_tree_map (bst_map);
5996 0 : return opt_result::failure_at (vect_location,
5997 : "SLP build failed.\n");
5998 : }
5999 : }
6000 : }
6001 :
6002 1177128 : if (bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo))
6003 : {
6004 1974194 : for (unsigned i = 0; i < bb_vinfo->roots.length (); ++i)
6005 : {
6006 1313013 : vect_location = bb_vinfo->roots[i].roots[0]->stmt;
6007 : /* Apply patterns. */
6008 4302962 : for (unsigned j = 0; j < bb_vinfo->roots[i].stmts.length (); ++j)
6009 5979898 : bb_vinfo->roots[i].stmts[j]
6010 3082937 : = vect_stmt_to_vectorize (bb_vinfo->roots[i].stmts[j]);
6011 1313013 : if (vect_build_slp_instance (bb_vinfo, bb_vinfo->roots[i].kind,
6012 1313013 : bb_vinfo->roots[i].stmts,
6013 1313013 : bb_vinfo->roots[i].roots,
6014 1313013 : bb_vinfo->roots[i].remain,
6015 : max_tree_size, &limit, bst_map, false))
6016 : {
6017 155681 : bb_vinfo->roots[i].roots = vNULL;
6018 155681 : bb_vinfo->roots[i].remain = vNULL;
6019 : }
6020 1313013 : bb_vinfo->roots[i].stmts = vNULL;
6021 : }
6022 : }
6023 :
6024 1177128 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
6025 : {
6026 : /* Find SLP sequences starting from groups of reductions. */
6027 515947 : if (!vect_analyze_slp_reductions (loop_vinfo, max_tree_size, &limit,
6028 : bst_map, force_single_lane))
6029 : {
6030 1887 : release_scalar_stmts_to_slp_tree_map (bst_map);
6031 1887 : return opt_result::failure_at (vect_location, "SLP build failed.\n");
6032 : }
6033 :
6034 : /* Make sure to vectorize only-live stmts, usually inductions. */
6035 2316161 : for (edge e : get_loop_exit_edges (LOOP_VINFO_LOOP (loop_vinfo)))
6036 1496199 : for (auto gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi);
6037 712262 : gsi_next (&gsi))
6038 : {
6039 722218 : gphi *lc_phi = *gsi;
6040 722218 : tree def = gimple_phi_arg_def_from_edge (lc_phi, e);
6041 722218 : stmt_vec_info stmt_info;
6042 722218 : if (TREE_CODE (def) == SSA_NAME
6043 606171 : && !virtual_operand_p (def)
6044 312410 : && (stmt_info = loop_vinfo->lookup_def (def))
6045 279972 : && ((stmt_info = vect_stmt_to_vectorize (stmt_info)), true)
6046 279972 : && STMT_VINFO_RELEVANT (stmt_info) == vect_used_only_live
6047 217783 : && STMT_VINFO_LIVE_P (stmt_info)
6048 217783 : && !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))
6049 835729 : && STMT_VINFO_REDUC_IDX (stmt_info) == -1)
6050 : {
6051 113423 : vec<stmt_vec_info> stmts;
6052 113423 : vec<stmt_vec_info> roots = vNULL;
6053 113423 : vec<tree> remain = vNULL;
6054 113423 : stmts.create (1);
6055 113423 : stmts.quick_push (vect_stmt_to_vectorize (stmt_info));
6056 113423 : if (! vect_build_slp_instance (vinfo,
6057 : slp_inst_kind_reduc_group,
6058 : stmts, roots, remain,
6059 : max_tree_size, &limit,
6060 : bst_map, force_single_lane))
6061 : {
6062 9956 : release_scalar_stmts_to_slp_tree_map (bst_map);
6063 9956 : return opt_result::failure_at (vect_location,
6064 : "SLP build failed.\n");
6065 : }
6066 : }
6067 9956 : }
6068 :
6069 : /* Find SLP sequences starting from gconds. */
6070 1263794 : for (auto cond : LOOP_VINFO_LOOP_CONDS (loop_vinfo))
6071 : {
6072 295087 : auto cond_info = loop_vinfo->lookup_stmt (cond);
6073 :
6074 295087 : cond_info = vect_stmt_to_vectorize (cond_info);
6075 295087 : vec<stmt_vec_info> roots = vNULL;
6076 295087 : roots.safe_push (cond_info);
6077 295087 : gimple *stmt = STMT_VINFO_STMT (cond_info);
6078 295087 : tree args0 = gimple_cond_lhs (stmt);
6079 295087 : tree args1 = gimple_cond_rhs (stmt);
6080 :
6081 : /* These should be enforced by cond lowering, but if it failed
6082 : bail. */
6083 295087 : if (gimple_cond_code (stmt) != NE_EXPR
6084 293995 : || TREE_TYPE (args0) != boolean_type_node
6085 588486 : || !integer_zerop (args1))
6086 : {
6087 1688 : roots.release ();
6088 1688 : release_scalar_stmts_to_slp_tree_map (bst_map);
6089 1688 : return opt_result::failure_at (vect_location,
6090 : "SLP build failed.\n");
6091 : }
6092 :
6093 : /* An argument without a loop def will be codegened from vectorizing the
6094 : root gcond itself. As such we don't need to try to build an SLP tree
6095 : from them. It's highly likely that the resulting SLP tree here if both
6096 : arguments have a def will be incompatible, but we rely on it being split
6097 : later on. */
6098 293399 : auto varg = loop_vinfo->lookup_def (args0);
6099 293399 : vec<stmt_vec_info> stmts;
6100 293399 : vec<tree> remain = vNULL;
6101 293399 : stmts.create (1);
6102 293399 : stmts.quick_push (vect_stmt_to_vectorize (varg));
6103 :
6104 293399 : if (! vect_build_slp_instance (vinfo, slp_inst_kind_gcond,
6105 : stmts, roots, remain,
6106 : max_tree_size, &limit,
6107 : bst_map, force_single_lane))
6108 : {
6109 907 : roots.release ();
6110 907 : release_scalar_stmts_to_slp_tree_map (bst_map);
6111 907 : return opt_result::failure_at (vect_location,
6112 : "SLP build failed.\n");
6113 : }
6114 : }
6115 : }
6116 :
6117 1162690 : hash_set<slp_tree> visited_patterns;
6118 1162690 : slp_tree_to_load_perm_map_t perm_cache;
6119 1162690 : slp_compat_nodes_map_t compat_cache;
6120 :
6121 : /* See if any patterns can be found in the SLP tree. */
6122 1162690 : bool pattern_found = false;
6123 3954533 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
6124 1629153 : pattern_found |= vect_match_slp_patterns (instance, vinfo,
6125 : &visited_patterns, &perm_cache,
6126 : &compat_cache);
6127 :
6128 : /* If any were found optimize permutations of loads. */
6129 1162690 : if (pattern_found)
6130 : {
6131 267 : hash_map<slp_tree, slp_tree> load_map;
6132 4981 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
6133 : {
6134 4180 : slp_tree root = SLP_INSTANCE_TREE (instance);
6135 4180 : optimize_load_redistribution (bst_map, vinfo, SLP_TREE_LANES (root),
6136 : &load_map, root);
6137 : }
6138 267 : }
6139 :
6140 : /* Check whether we should force some SLP instances to use load/store-lanes
6141 : and do so by forcing SLP re-discovery with single lanes. We used
6142 : to cancel SLP when this applied to all instances in a loop but now
6143 : we decide this per SLP instance. It's important to do this only
6144 : after SLP pattern recognition. */
6145 1162690 : if (is_a <loop_vec_info> (vinfo))
6146 1300073 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
6147 798564 : if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_store
6148 301955 : && !SLP_INSTANCE_TREE (instance)->ldst_lanes)
6149 : {
6150 301955 : slp_tree slp_root = SLP_INSTANCE_TREE (instance);
6151 301955 : unsigned int group_size = SLP_TREE_LANES (slp_root);
6152 301955 : tree vectype = SLP_TREE_VECTYPE (slp_root);
6153 :
6154 301955 : stmt_vec_info rep_info = SLP_TREE_REPRESENTATIVE (slp_root);
6155 301955 : gimple *rep = STMT_VINFO_STMT (rep_info);
6156 301955 : bool masked = (is_gimple_call (rep)
6157 2446 : && gimple_call_internal_p (rep)
6158 304381 : && internal_fn_mask_index
6159 2426 : (gimple_call_internal_fn (rep)) != -1);
6160 301935 : if (!STMT_VINFO_GROUPED_ACCESS (rep_info)
6161 30073 : || slp_root->ldst_lanes
6162 332028 : || (vect_store_lanes_supported (vectype, group_size, masked)
6163 : == IFN_LAST))
6164 301955 : continue;
6165 :
6166 0 : auto_vec<slp_tree> loads;
6167 0 : hash_set<slp_tree> visited;
6168 0 : vect_gather_slp_loads (loads, slp_root, visited);
6169 :
6170 : /* Check whether any load in the SLP instance is possibly
6171 : permuted. */
6172 0 : bool loads_permuted = false;
6173 0 : slp_tree load_node;
6174 0 : unsigned j;
6175 0 : FOR_EACH_VEC_ELT (loads, j, load_node)
6176 : {
6177 0 : if (!SLP_TREE_LOAD_PERMUTATION (load_node).exists ())
6178 0 : continue;
6179 0 : for (unsigned k = 0; k < SLP_TREE_LANES (load_node); k++)
6180 0 : if (SLP_TREE_LOAD_PERMUTATION (load_node)[k] != k)
6181 : {
6182 : loads_permuted = true;
6183 : break;
6184 : }
6185 : }
6186 :
6187 : /* If the loads and stores can use load/store-lanes force re-discovery
6188 : with single lanes. */
6189 0 : if (loads_permuted)
6190 : {
6191 0 : bool can_use_lanes = true;
6192 : bool prefer_load_lanes = false;
6193 0 : FOR_EACH_VEC_ELT (loads, j, load_node)
6194 0 : if (STMT_VINFO_GROUPED_ACCESS
6195 : (SLP_TREE_REPRESENTATIVE (load_node)))
6196 : {
6197 0 : stmt_vec_info stmt_vinfo = DR_GROUP_FIRST_ELEMENT
6198 : (SLP_TREE_REPRESENTATIVE (load_node));
6199 0 : rep = STMT_VINFO_STMT (stmt_vinfo);
6200 0 : masked = (is_gimple_call (rep)
6201 0 : && gimple_call_internal_p (rep)
6202 0 : && internal_fn_mask_index
6203 0 : (gimple_call_internal_fn (rep)));
6204 : /* Use SLP for strided accesses (or if we can't
6205 : load-lanes). */
6206 0 : if (STMT_VINFO_STRIDED_P (stmt_vinfo)
6207 0 : || compare_step_with_zero (vinfo, stmt_vinfo) <= 0
6208 0 : || vect_load_lanes_supported
6209 0 : (SLP_TREE_VECTYPE (load_node),
6210 0 : DR_GROUP_SIZE (stmt_vinfo), masked) == IFN_LAST
6211 : /* ??? During SLP re-discovery with a single lane
6212 : a masked grouped load will appear permuted and
6213 : discovery will fail. We have to rework this
6214 : on the discovery side - for now avoid ICEing. */
6215 0 : || masked)
6216 : {
6217 : can_use_lanes = false;
6218 : break;
6219 : }
6220 : /* Make sure that the target would prefer store-lanes
6221 : for at least one of the loads.
6222 :
6223 : ??? Perhaps we should instead require this for
6224 : all loads? */
6225 0 : prefer_load_lanes
6226 : = (prefer_load_lanes
6227 0 : || SLP_TREE_LANES (load_node) == group_size
6228 0 : || (vect_slp_prefer_store_lanes_p
6229 0 : (vinfo, stmt_vinfo,
6230 : SLP_TREE_VECTYPE (load_node), masked,
6231 : group_size, SLP_TREE_LANES (load_node))));
6232 : }
6233 :
6234 0 : if (can_use_lanes && prefer_load_lanes)
6235 : {
6236 0 : if (dump_enabled_p ())
6237 0 : dump_printf_loc (MSG_NOTE, vect_location,
6238 : "SLP instance %p can use load/store-lanes,"
6239 : " re-discovering with single-lanes\n",
6240 : (void *) instance);
6241 :
6242 0 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (slp_root);
6243 :
6244 0 : vect_free_slp_instance (instance);
6245 0 : limit = max_tree_size;
6246 0 : bool res = vect_analyze_slp_instance (vinfo, bst_map,
6247 : stmt_info,
6248 : slp_inst_kind_store,
6249 : max_tree_size, &limit,
6250 : true);
6251 0 : gcc_assert (res);
6252 0 : auto new_inst = LOOP_VINFO_SLP_INSTANCES (vinfo).pop ();
6253 0 : LOOP_VINFO_SLP_INSTANCES (vinfo)[i] = new_inst;
6254 : }
6255 : }
6256 0 : }
6257 :
6258 : /* When we end up with load permutations that we cannot possibly handle,
6259 : like those requiring three vector inputs, lower them using interleaving
6260 : like schemes. */
6261 1162690 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
6262 : {
6263 501509 : vect_lower_load_permutations (loop_vinfo, bst_map, force_single_lane);
6264 501509 : if (dump_enabled_p ())
6265 : {
6266 20777 : dump_printf_loc (MSG_NOTE, vect_location,
6267 : "SLP graph after lowering permutations:\n");
6268 20777 : hash_set<slp_tree> visited;
6269 92384 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
6270 30080 : vect_print_slp_graph (MSG_NOTE, vect_location,
6271 : SLP_INSTANCE_TREE (instance), visited);
6272 20777 : }
6273 : }
6274 :
6275 1162690 : release_scalar_stmts_to_slp_tree_map (bst_map);
6276 :
6277 1162690 : if (pattern_found && dump_enabled_p ())
6278 : {
6279 20 : dump_printf_loc (MSG_NOTE, vect_location,
6280 : "Pattern matched SLP tree\n");
6281 20 : hash_set<slp_tree> visited;
6282 101 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
6283 41 : vect_print_slp_graph (MSG_NOTE, vect_location,
6284 : SLP_INSTANCE_TREE (instance), visited);
6285 20 : }
6286 :
6287 1162690 : return opt_result::success ();
6288 1162690 : }
6289 :
6290 : /* Estimates the cost of inserting layout changes into the SLP graph.
6291 : It can also say that the insertion is impossible. */
6292 :
6293 : struct slpg_layout_cost
6294 : {
6295 11212297 : slpg_layout_cost () = default;
6296 : slpg_layout_cost (sreal, bool);
6297 :
6298 523073 : static slpg_layout_cost impossible () { return { sreal::max (), 0 }; }
6299 5548573 : bool is_possible () const { return depth != sreal::max (); }
6300 :
6301 : bool operator== (const slpg_layout_cost &) const;
6302 : bool operator!= (const slpg_layout_cost &) const;
6303 :
6304 : bool is_better_than (const slpg_layout_cost &, bool) const;
6305 :
6306 : void add_parallel_cost (const slpg_layout_cost &);
6307 : void add_serial_cost (const slpg_layout_cost &);
6308 : void split (unsigned int);
6309 :
6310 : /* The longest sequence of layout changes needed during any traversal
6311 : of the partition dag, weighted by execution frequency.
6312 :
6313 : This is the most important metric when optimizing for speed, since
6314 : it helps to ensure that we keep the number of operations on
6315 : critical paths to a minimum. */
6316 : sreal depth = 0;
6317 :
6318 : /* An estimate of the total number of operations needed. It is weighted by
6319 : execution frequency when optimizing for speed but not when optimizing for
6320 : size. In order to avoid double-counting, a node with a fanout of N will
6321 : distribute 1/N of its total cost to each successor.
6322 :
6323 : This is the most important metric when optimizing for size, since
6324 : it helps to keep the total number of operations to a minimum, */
6325 : sreal total = 0;
6326 : };
6327 :
6328 : /* Construct costs for a node with weight WEIGHT. A higher weight
6329 : indicates more frequent execution. IS_FOR_SIZE is true if we are
6330 : optimizing for size rather than speed. */
6331 :
6332 1309898 : slpg_layout_cost::slpg_layout_cost (sreal weight, bool is_for_size)
6333 1317505 : : depth (weight), total (is_for_size && weight > 0 ? 1 : weight)
6334 : {
6335 1309898 : }
6336 :
6337 : bool
6338 0 : slpg_layout_cost::operator== (const slpg_layout_cost &other) const
6339 : {
6340 0 : return depth == other.depth && total == other.total;
6341 : }
6342 :
6343 : bool
6344 0 : slpg_layout_cost::operator!= (const slpg_layout_cost &other) const
6345 : {
6346 0 : return !operator== (other);
6347 : }
6348 :
6349 : /* Return true if these costs are better than OTHER. IS_FOR_SIZE is
6350 : true if we are optimizing for size rather than speed. */
6351 :
6352 : bool
6353 315660 : slpg_layout_cost::is_better_than (const slpg_layout_cost &other,
6354 : bool is_for_size) const
6355 : {
6356 315660 : if (is_for_size)
6357 : {
6358 3260 : if (total != other.total)
6359 1666 : return total < other.total;
6360 1594 : return depth < other.depth;
6361 : }
6362 : else
6363 : {
6364 312400 : if (depth != other.depth)
6365 127685 : return depth < other.depth;
6366 184715 : return total < other.total;
6367 : }
6368 : }
6369 :
6370 : /* Increase the costs to account for something with cost INPUT_COST
6371 : happening in parallel with the current costs. */
6372 :
6373 : void
6374 378553 : slpg_layout_cost::add_parallel_cost (const slpg_layout_cost &input_cost)
6375 : {
6376 378553 : depth = std::max (depth, input_cost.depth);
6377 378553 : total += input_cost.total;
6378 378553 : }
6379 :
6380 : /* Increase the costs to account for something with cost INPUT_COST
6381 : happening in series with the current costs. */
6382 :
6383 : void
6384 1556756 : slpg_layout_cost::add_serial_cost (const slpg_layout_cost &other)
6385 : {
6386 1556756 : depth += other.depth;
6387 1556756 : total += other.total;
6388 1556756 : }
6389 :
6390 : /* Split the total cost among TIMES successors or predecessors. */
6391 :
6392 : void
6393 1239937 : slpg_layout_cost::split (unsigned int times)
6394 : {
6395 1239937 : if (times > 1)
6396 599771 : total /= times;
6397 1239937 : }
6398 :
6399 : /* Information about one node in the SLP graph, for use during
6400 : vect_optimize_slp_pass. */
6401 :
6402 : struct slpg_vertex
6403 : {
6404 10475599 : slpg_vertex (slp_tree node_) : node (node_) {}
6405 :
6406 : /* The node itself. */
6407 : slp_tree node;
6408 :
6409 : /* Which partition the node belongs to, or -1 if none. Nodes outside of
6410 : partitions are flexible; they can have whichever layout consumers
6411 : want them to have. */
6412 : int partition = -1;
6413 :
6414 : /* The number of nodes that directly use the result of this one
6415 : (i.e. the number of nodes that count this one as a child). */
6416 : unsigned int out_degree = 0;
6417 :
6418 : /* The execution frequency of the node. */
6419 : sreal weight = 0;
6420 :
6421 : /* The total execution frequency of all nodes that directly use the
6422 : result of this one. */
6423 : sreal out_weight = 0;
6424 : };
6425 :
6426 : /* Information about one partition of the SLP graph, for use during
6427 : vect_optimize_slp_pass. */
6428 :
6429 : struct slpg_partition_info
6430 : {
6431 : /* The nodes in the partition occupy indices [NODE_BEGIN, NODE_END)
6432 : of m_partitioned_nodes. */
6433 : unsigned int node_begin = 0;
6434 : unsigned int node_end = 0;
6435 :
6436 : /* Which layout we've chosen to use for this partition, or -1 if
6437 : we haven't picked one yet. */
6438 : int layout = -1;
6439 :
6440 : /* The number of predecessors and successors in the partition dag.
6441 : The predecessors always have lower partition numbers and the
6442 : successors always have higher partition numbers.
6443 :
6444 : Note that the directions of these edges are not necessarily the
6445 : same as in the data flow graph. For example, if an SCC has separate
6446 : partitions for an inner loop and an outer loop, the inner loop's
6447 : partition will have at least two incoming edges from the outer loop's
6448 : partition: one for a live-in value and one for a live-out value.
6449 : In data flow terms, one of these edges would also be from the outer loop
6450 : to the inner loop, but the other would be in the opposite direction. */
6451 : unsigned int in_degree = 0;
6452 : unsigned int out_degree = 0;
6453 : };
6454 :
6455 : /* Information about the costs of using a particular layout for a
6456 : particular partition. It can also say that the combination is
6457 : impossible. */
6458 :
6459 : struct slpg_partition_layout_costs
6460 : {
6461 1631093 : bool is_possible () const { return internal_cost.is_possible (); }
6462 67316 : void mark_impossible () { internal_cost = slpg_layout_cost::impossible (); }
6463 :
6464 : /* The costs inherited from predecessor partitions. */
6465 : slpg_layout_cost in_cost;
6466 :
6467 : /* The inherent cost of the layout within the node itself. For example,
6468 : this is nonzero for a load if choosing a particular layout would require
6469 : the load to permute the loaded elements. It is nonzero for a
6470 : VEC_PERM_EXPR if the permutation cannot be eliminated or converted
6471 : to full-vector moves. */
6472 : slpg_layout_cost internal_cost;
6473 :
6474 : /* The costs inherited from successor partitions. */
6475 : slpg_layout_cost out_cost;
6476 : };
6477 :
6478 : /* This class tries to optimize the layout of vectors in order to avoid
6479 : unnecessary shuffling. At the moment, the set of possible layouts are
6480 : restricted to bijective permutations.
6481 :
6482 : The goal of the pass depends on whether we're optimizing for size or
6483 : for speed. When optimizing for size, the goal is to reduce the overall
6484 : number of layout changes (including layout changes implied by things
6485 : like load permutations). When optimizing for speed, the goal is to
6486 : reduce the maximum latency attributable to layout changes on any
6487 : non-cyclical path through the data flow graph.
6488 :
6489 : For example, when optimizing a loop nest for speed, we will prefer
6490 : to make layout changes outside of a loop rather than inside of a loop,
6491 : and will prefer to make layout changes in parallel rather than serially,
6492 : even if that increases the overall number of layout changes.
6493 :
6494 : The high-level procedure is:
6495 :
6496 : (1) Build a graph in which edges go from uses (parents) to definitions
6497 : (children).
6498 :
6499 : (2) Divide the graph into a dag of strongly-connected components (SCCs).
6500 :
6501 : (3) When optimizing for speed, partition the nodes in each SCC based
6502 : on their containing cfg loop. When optimizing for size, treat
6503 : each SCC as a single partition.
6504 :
6505 : This gives us a dag of partitions. The goal is now to assign a
6506 : layout to each partition.
6507 :
6508 : (4) Construct a set of vector layouts that are worth considering.
6509 : Record which nodes must keep their current layout.
6510 :
6511 : (5) Perform a forward walk over the partition dag (from loads to stores)
6512 : accumulating the "forward" cost of using each layout. When visiting
6513 : each partition, assign a tentative choice of layout to the partition
6514 : and use that choice when calculating the cost of using a different
6515 : layout in successor partitions.
6516 :
6517 : (6) Perform a backward walk over the partition dag (from stores to loads),
6518 : accumulating the "backward" cost of using each layout. When visiting
6519 : each partition, make a final choice of layout for that partition based
6520 : on the accumulated forward costs (from (5)) and backward costs
6521 : (from (6)).
6522 :
6523 : (7) Apply the chosen layouts to the SLP graph.
6524 :
6525 : For example, consider the SLP statements:
6526 :
6527 : S1: a_1 = load
6528 : loop:
6529 : S2: a_2 = PHI<a_1, a_3>
6530 : S3: b_1 = load
6531 : S4: a_3 = a_2 + b_1
6532 : exit:
6533 : S5: a_4 = PHI<a_3>
6534 : S6: store a_4
6535 :
6536 : S2 and S4 form an SCC and are part of the same loop. Every other
6537 : statement is in a singleton SCC. In this example there is a one-to-one
6538 : mapping between SCCs and partitions and the partition dag looks like this;
6539 :
6540 : S1 S3
6541 : \ /
6542 : S2+S4
6543 : |
6544 : S5
6545 : |
6546 : S6
6547 :
6548 : S2, S3 and S4 will have a higher execution frequency than the other
6549 : statements, so when optimizing for speed, the goal is to avoid any
6550 : layout changes:
6551 :
6552 : - within S3
6553 : - within S2+S4
6554 : - on the S3->S2+S4 edge
6555 :
6556 : For example, if S3 was originally a reversing load, the goal of the
6557 : pass is to make it an unreversed load and change the layout on the
6558 : S1->S2+S4 and S2+S4->S5 edges to compensate. (Changing the layout
6559 : on S1->S2+S4 and S5->S6 would also be acceptable.)
6560 :
6561 : The difference between SCCs and partitions becomes important if we
6562 : add an outer loop:
6563 :
6564 : S1: a_1 = ...
6565 : loop1:
6566 : S2: a_2 = PHI<a_1, a_6>
6567 : S3: b_1 = load
6568 : S4: a_3 = a_2 + b_1
6569 : loop2:
6570 : S5: a_4 = PHI<a_3, a_5>
6571 : S6: c_1 = load
6572 : S7: a_5 = a_4 + c_1
6573 : exit2:
6574 : S8: a_6 = PHI<a_5>
6575 : S9: store a_6
6576 : exit1:
6577 :
6578 : Here, S2, S4, S5, S7 and S8 form a single SCC. However, when optimizing
6579 : for speed, we usually do not want restrictions in the outer loop to "infect"
6580 : the decision for the inner loop. For example, if an outer-loop node
6581 : in the SCC contains a statement with a fixed layout, that should not
6582 : prevent the inner loop from using a different layout. Conversely,
6583 : the inner loop should not dictate a layout to the outer loop: if the
6584 : outer loop does a lot of computation, then it may not be efficient to
6585 : do all of that computation in the inner loop's preferred layout.
6586 :
6587 : So when optimizing for speed, we partition the SCC into S2+S4+S8 (outer)
6588 : and S5+S7 (inner). We also try to arrange partitions so that:
6589 :
6590 : - the partition for an outer loop comes before the partition for
6591 : an inner loop
6592 :
6593 : - if a sibling loop A dominates a sibling loop B, A's partition
6594 : comes before B's
6595 :
6596 : This gives the following partition dag for the example above:
6597 :
6598 : S1 S3
6599 : \ /
6600 : S2+S4+S8 S6
6601 : | \\ /
6602 : | S5+S7
6603 : |
6604 : S9
6605 :
6606 : There are two edges from S2+S4+S8 to S5+S7: one for the edge S4->S5 and
6607 : one for a reversal of the edge S7->S8.
6608 :
6609 : The backward walk picks a layout for S5+S7 before S2+S4+S8. The choice
6610 : for S2+S4+S8 therefore has to balance the cost of using the outer loop's
6611 : preferred layout against the cost of changing the layout on entry to the
6612 : inner loop (S4->S5) and on exit from the inner loop (S7->S8 reversed).
6613 :
6614 : Although this works well when optimizing for speed, it has the downside
6615 : when optimizing for size that the choice of layout for S5+S7 is completely
6616 : independent of S9, which lessens the chance of reducing the overall number
6617 : of permutations. We therefore do not partition SCCs when optimizing
6618 : for size.
6619 :
6620 : To give a concrete example of the difference between optimizing
6621 : for size and speed, consider:
6622 :
6623 : a[0] = (b[1] << c[3]) - d[1];
6624 : a[1] = (b[0] << c[2]) - d[0];
6625 : a[2] = (b[3] << c[1]) - d[3];
6626 : a[3] = (b[2] << c[0]) - d[2];
6627 :
6628 : There are three different layouts here: one for a, one for b and d,
6629 : and one for c. When optimizing for speed it is better to permute each
6630 : of b, c and d into the order required by a, since those permutations
6631 : happen in parallel. But when optimizing for size, it is better to:
6632 :
6633 : - permute c into the same order as b
6634 : - do the arithmetic
6635 : - permute the result into the order required by a
6636 :
6637 : This gives 2 permutations rather than 3. */
6638 :
6639 : class vect_optimize_slp_pass
6640 : {
6641 : public:
6642 716297 : vect_optimize_slp_pass (vec_info *vinfo) : m_vinfo (vinfo) {}
6643 : void run ();
6644 :
6645 : private:
6646 : /* Graph building. */
6647 : struct loop *containing_loop (slp_tree);
6648 : bool is_cfg_latch_edge (graph_edge *);
6649 : void build_vertices (hash_set<slp_tree> &, slp_tree);
6650 : void build_vertices ();
6651 : void build_graph ();
6652 :
6653 : /* Partitioning. */
6654 : void create_partitions ();
6655 : template<typename T> void for_each_partition_edge (unsigned int, T);
6656 :
6657 : /* Layout selection. */
6658 : bool is_compatible_layout (slp_tree, unsigned int);
6659 : bool is_compatible_layout (const slpg_partition_info &, unsigned int);
6660 : int change_layout_cost (slp_tree, unsigned int, unsigned int);
6661 : slpg_partition_layout_costs &partition_layout_costs (unsigned int,
6662 : unsigned int);
6663 : void change_vec_perm_layout (slp_tree, lane_permutation_t &,
6664 : int, unsigned int);
6665 : int internal_node_cost (slp_tree, int, unsigned int);
6666 : void start_choosing_layouts ();
6667 : bool legitimize ();
6668 :
6669 : /* Cost propagation. */
6670 : slpg_layout_cost edge_layout_cost (graph_edge *, unsigned int,
6671 : unsigned int, unsigned int);
6672 : slpg_layout_cost total_in_cost (unsigned int);
6673 : slpg_layout_cost forward_cost (graph_edge *, unsigned int, unsigned int);
6674 : slpg_layout_cost backward_cost (graph_edge *, unsigned int, unsigned int);
6675 : void forward_pass ();
6676 : void backward_pass ();
6677 :
6678 : /* Rematerialization. */
6679 : slp_tree get_result_with_layout (slp_tree, unsigned int);
6680 : void materialize ();
6681 :
6682 : /* Clean-up. */
6683 : void remove_redundant_permutations ();
6684 :
6685 : /* Masked load lanes discovery. */
6686 : void decide_masked_load_lanes ();
6687 :
6688 : void dump ();
6689 :
6690 : vec_info *m_vinfo;
6691 :
6692 : /* True if we should optimize the graph for size, false if we should
6693 : optimize it for speed. (It wouldn't be easy to make this decision
6694 : more locally.) */
6695 : bool m_optimize_size;
6696 :
6697 : /* A graph of all SLP nodes, with edges leading from uses to definitions.
6698 : In other words, a node's predecessors are its slp_tree parents and
6699 : a node's successors are its slp_tree children. */
6700 : graph *m_slpg = nullptr;
6701 :
6702 : /* The vertices of M_SLPG, indexed by slp_tree::vertex. */
6703 : auto_vec<slpg_vertex> m_vertices;
6704 :
6705 : /* The list of all leaves of M_SLPG. such as external definitions, constants,
6706 : and loads. */
6707 : auto_vec<int> m_leafs;
6708 :
6709 : /* This array has one entry for every vector layout that we're considering.
6710 : Element 0 is null and indicates "no change". Other entries describe
6711 : permutations that are inherent in the current graph and that we would
6712 : like to reverse if possible.
6713 :
6714 : For example, a permutation { 1, 2, 3, 0 } means that something has
6715 : effectively been permuted in that way, such as a load group
6716 : { a[1], a[2], a[3], a[0] } (viewed as a permutation of a[0:3]).
6717 : We'd then like to apply the reverse permutation { 3, 0, 1, 2 }
6718 : in order to put things "back" in order. */
6719 : auto_vec<vec<unsigned> > m_perms;
6720 :
6721 : /* A partitioning of the nodes for which a layout must be chosen.
6722 : Each partition represents an <SCC, cfg loop> pair; that is,
6723 : nodes in different SCCs belong to different partitions, and nodes
6724 : within an SCC can be further partitioned according to a containing
6725 : cfg loop. Partition <SCC1, L1> comes before <SCC2, L2> if:
6726 :
6727 : - SCC1 != SCC2 and SCC1 is a predecessor of SCC2 in a forward walk
6728 : from leaves (such as loads) to roots (such as stores).
6729 :
6730 : - SCC1 == SCC2 and L1's header strictly dominates L2's header. */
6731 : auto_vec<slpg_partition_info> m_partitions;
6732 :
6733 : /* The list of all nodes for which a layout must be chosen. Nodes for
6734 : partition P come before the nodes for partition P+1. Nodes within a
6735 : partition are in reverse postorder. */
6736 : auto_vec<unsigned int> m_partitioned_nodes;
6737 :
6738 : /* Index P * num-layouts + L contains the cost of using layout L
6739 : for partition P. */
6740 : auto_vec<slpg_partition_layout_costs> m_partition_layout_costs;
6741 :
6742 : /* Index N * num-layouts + L, if nonnull, is a node that provides the
6743 : original output of node N adjusted to have layout L. */
6744 : auto_vec<slp_tree> m_node_layouts;
6745 : };
6746 :
6747 : /* Fill the vertices and leafs vector with all nodes in the SLP graph.
6748 : Also record whether we should optimize anything for speed rather
6749 : than size. */
6750 :
6751 : void
6752 11317809 : vect_optimize_slp_pass::build_vertices (hash_set<slp_tree> &visited,
6753 : slp_tree node)
6754 : {
6755 11317809 : unsigned i;
6756 11317809 : slp_tree child;
6757 :
6758 11317809 : if (visited.add (node))
6759 11317809 : return;
6760 :
6761 10475599 : if (stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node))
6762 : {
6763 8000918 : basic_block bb = gimple_bb (vect_orig_stmt (rep)->stmt);
6764 7059088 : if (optimize_bb_for_speed_p (bb))
6765 6929190 : m_optimize_size = false;
6766 : }
6767 :
6768 10475599 : node->vertex = m_vertices.length ();
6769 10475599 : m_vertices.safe_push (slpg_vertex (node));
6770 :
6771 10475599 : bool leaf = true;
6772 10475599 : bool force_leaf = false;
6773 19636588 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
6774 9160989 : if (child)
6775 : {
6776 8246029 : leaf = false;
6777 8246029 : build_vertices (visited, child);
6778 : }
6779 : else
6780 : force_leaf = true;
6781 : /* Since SLP discovery works along use-def edges all cycles have an
6782 : entry - but there's the exception of cycles where we do not handle
6783 : the entry explicitly (but with a NULL SLP node), like some reductions
6784 : and inductions. Force those SLP PHIs to act as leafs to make them
6785 : backwards reachable. */
6786 10475599 : if (leaf || force_leaf)
6787 5192434 : m_leafs.safe_push (node->vertex);
6788 : }
6789 :
6790 : /* Fill the vertices and leafs vector with all nodes in the SLP graph. */
6791 :
6792 : void
6793 1432594 : vect_optimize_slp_pass::build_vertices ()
6794 : {
6795 1432594 : hash_set<slp_tree> visited;
6796 1432594 : unsigned i;
6797 1432594 : slp_instance instance;
6798 1432594 : m_vertices.truncate (0);
6799 1432594 : m_leafs.truncate (0);
6800 7369562 : FOR_EACH_VEC_ELT (m_vinfo->slp_instances, i, instance)
6801 3071780 : build_vertices (visited, SLP_INSTANCE_TREE (instance));
6802 1432594 : }
6803 :
6804 : /* Apply (reverse) bijectite PERM to VEC. */
6805 :
6806 : template <class T>
6807 : static void
6808 229408 : vect_slp_permute (vec<unsigned> perm,
6809 : vec<T> &vec, bool reverse)
6810 : {
6811 229408 : auto_vec<T, 64> saved;
6812 229408 : saved.create (vec.length ());
6813 761682 : for (unsigned i = 0; i < vec.length (); ++i)
6814 532274 : saved.quick_push (vec[i]);
6815 :
6816 229408 : if (reverse)
6817 : {
6818 1504304 : for (unsigned i = 0; i < vec.length (); ++i)
6819 530230 : vec[perm[i]] = saved[i];
6820 758684 : for (unsigned i = 0; i < vec.length (); ++i)
6821 897685 : gcc_assert (vec[perm[i]] == saved[i]);
6822 : }
6823 : else
6824 : {
6825 5996 : for (unsigned i = 0; i < vec.length (); ++i)
6826 2044 : vec[i] = saved[perm[i]];
6827 231452 : for (unsigned i = 0; i < vec.length (); ++i)
6828 3066 : gcc_assert (vec[i] == saved[perm[i]]);
6829 : }
6830 229408 : }
6831 :
6832 : /* Return the cfg loop that contains NODE. */
6833 :
6834 : struct loop *
6835 4077981 : vect_optimize_slp_pass::containing_loop (slp_tree node)
6836 : {
6837 4077981 : stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node);
6838 : /* ??? This is imprecise, VEC_PERM nodes do not have a representative
6839 : but are laid out close to their children. */
6840 4077981 : if (!rep)
6841 144957 : return m_vinfo->bbs[0]->loop_father;
6842 4417287 : return gimple_bb (vect_orig_stmt (rep)->stmt)->loop_father;
6843 : }
6844 :
6845 : /* Return true if UD (an edge from a use to a definition) is associated
6846 : with a loop latch edge in the cfg. */
6847 :
6848 : bool
6849 8246029 : vect_optimize_slp_pass::is_cfg_latch_edge (graph_edge *ud)
6850 : {
6851 8246029 : slp_tree use = m_vertices[ud->src].node;
6852 8246029 : slp_tree def = m_vertices[ud->dest].node;
6853 8246029 : if ((SLP_TREE_DEF_TYPE (use) != vect_internal_def
6854 8246029 : || SLP_TREE_PERMUTE_P (use))
6855 7925200 : || SLP_TREE_DEF_TYPE (def) != vect_internal_def)
6856 : return false;
6857 :
6858 4798186 : stmt_vec_info use_rep = vect_orig_stmt (SLP_TREE_REPRESENTATIVE (use));
6859 4798186 : return (is_a<gphi *> (use_rep->stmt)
6860 385596 : && bb_loop_header_p (gimple_bb (use_rep->stmt))
6861 5016076 : && containing_loop (def) == containing_loop (use));
6862 : }
6863 :
6864 : /* Build the graph. Mark edges that correspond to cfg loop latch edges with
6865 : a nonnull data field. */
6866 :
6867 : void
6868 1432594 : vect_optimize_slp_pass::build_graph ()
6869 : {
6870 1432594 : m_optimize_size = true;
6871 1432594 : build_vertices ();
6872 :
6873 2865188 : m_slpg = new_graph (m_vertices.length ());
6874 14773381 : for (slpg_vertex &v : m_vertices)
6875 31302842 : for (slp_tree child : SLP_TREE_CHILDREN (v.node))
6876 9160989 : if (child)
6877 : {
6878 8246029 : graph_edge *ud = add_edge (m_slpg, v.node->vertex, child->vertex);
6879 8246029 : if (is_cfg_latch_edge (ud))
6880 208144 : ud->data = this;
6881 : }
6882 1432594 : }
6883 :
6884 : /* Return true if E corresponds to a loop latch edge in the cfg. */
6885 :
6886 : static bool
6887 4226529 : skip_cfg_latch_edges (graph_edge *e)
6888 : {
6889 4226529 : return e->data;
6890 : }
6891 :
6892 : /* Create the node partitions. */
6893 :
6894 : void
6895 716297 : vect_optimize_slp_pass::create_partitions ()
6896 : {
6897 : /* Calculate a postorder of the graph, ignoring edges that correspond
6898 : to natural latch edges in the cfg. Reading the vector from the end
6899 : to the beginning gives the reverse postorder. */
6900 716297 : auto_vec<int> initial_rpo;
6901 1432594 : graphds_dfs (m_slpg, &m_leafs[0], m_leafs.length (), &initial_rpo,
6902 : false, NULL, skip_cfg_latch_edges);
6903 2148891 : gcc_assert (initial_rpo.length () == m_vertices.length ());
6904 :
6905 : /* Calculate the strongly connected components of the graph. */
6906 716297 : auto_vec<int> scc_grouping;
6907 716297 : unsigned int num_sccs = graphds_scc (m_slpg, NULL, NULL, &scc_grouping);
6908 :
6909 : /* Create a new index order in which all nodes from the same SCC are
6910 : consecutive. Use scc_pos to record the index of the first node in
6911 : each SCC. */
6912 716297 : auto_vec<unsigned int> scc_pos (num_sccs);
6913 716297 : int last_component = -1;
6914 716297 : unsigned int node_count = 0;
6915 7386139 : for (unsigned int node_i : scc_grouping)
6916 : {
6917 5237248 : if (last_component != m_slpg->vertices[node_i].component)
6918 : {
6919 5106682 : last_component = m_slpg->vertices[node_i].component;
6920 10213364 : gcc_assert (last_component == int (scc_pos.length ()));
6921 5106682 : scc_pos.quick_push (node_count);
6922 : }
6923 5237248 : node_count += 1;
6924 : }
6925 1432594 : gcc_assert (node_count == initial_rpo.length ()
6926 : && last_component + 1 == int (num_sccs));
6927 :
6928 : /* Use m_partitioned_nodes to group nodes into SCC order, with the nodes
6929 : inside each SCC following the RPO we calculated above. The fact that
6930 : we ignored natural latch edges when calculating the RPO should ensure
6931 : that, for natural loop nests:
6932 :
6933 : - the first node that we encounter in a cfg loop is the loop header phi
6934 : - the loop header phis are in dominance order
6935 :
6936 : Arranging for this is an optimization (see below) rather than a
6937 : correctness issue. Unnatural loops with a tangled mess of backedges
6938 : will still work correctly, but might give poorer results.
6939 :
6940 : Also update scc_pos so that it gives 1 + the index of the last node
6941 : in the SCC. */
6942 716297 : m_partitioned_nodes.safe_grow (node_count);
6943 6669842 : for (unsigned int old_i = initial_rpo.length (); old_i-- > 0;)
6944 : {
6945 5237248 : unsigned int node_i = initial_rpo[old_i];
6946 5237248 : unsigned int new_i = scc_pos[m_slpg->vertices[node_i].component]++;
6947 5237248 : m_partitioned_nodes[new_i] = node_i;
6948 : }
6949 :
6950 : /* When optimizing for speed, partition each SCC based on the containing
6951 : cfg loop. The order we constructed above should ensure that, for natural
6952 : cfg loops, we'll create sub-SCC partitions for outer loops before
6953 : the corresponding sub-SCC partitions for inner loops. Similarly,
6954 : when one sibling loop A dominates another sibling loop B, we should
6955 : create a sub-SCC partition for A before a sub-SCC partition for B.
6956 :
6957 : As above, nothing depends for correctness on whether this achieves
6958 : a natural nesting, but we should get better results when it does. */
6959 1432594 : m_partitions.reserve (m_vertices.length ());
6960 716297 : unsigned int next_partition_i = 0;
6961 716297 : hash_map<struct loop *, int> loop_partitions;
6962 716297 : unsigned int rpo_begin = 0;
6963 716297 : unsigned int num_partitioned_nodes = 0;
6964 7255573 : for (unsigned int rpo_end : scc_pos)
6965 : {
6966 5106682 : loop_partitions.empty ();
6967 5106682 : unsigned int partition_i = next_partition_i;
6968 10343930 : for (unsigned int rpo_i = rpo_begin; rpo_i < rpo_end; ++rpo_i)
6969 : {
6970 : /* Handle externals and constants optimistically throughout.
6971 : But treat existing vectors as fixed since we do not handle
6972 : permuting them. */
6973 5237248 : unsigned int node_i = m_partitioned_nodes[rpo_i];
6974 5237248 : auto &vertex = m_vertices[node_i];
6975 5237248 : if ((SLP_TREE_DEF_TYPE (vertex.node) == vect_external_def
6976 529349 : && !SLP_TREE_VEC_DEFS (vertex.node).exists ())
6977 5240241 : || SLP_TREE_DEF_TYPE (vertex.node) == vect_constant_def)
6978 1564731 : vertex.partition = -1;
6979 : else
6980 : {
6981 3672517 : bool existed;
6982 3672517 : if (m_optimize_size)
6983 30316 : existed = next_partition_i > partition_i;
6984 : else
6985 : {
6986 3642201 : struct loop *loop = containing_loop (vertex.node);
6987 3642201 : auto &entry = loop_partitions.get_or_insert (loop, &existed);
6988 3642201 : if (!existed)
6989 3512862 : entry = next_partition_i;
6990 3642201 : partition_i = entry;
6991 : }
6992 3672517 : if (!existed)
6993 : {
6994 3543086 : m_partitions.quick_push (slpg_partition_info ());
6995 3543086 : next_partition_i += 1;
6996 : }
6997 3672517 : vertex.partition = partition_i;
6998 3672517 : num_partitioned_nodes += 1;
6999 3672517 : m_partitions[partition_i].node_end += 1;
7000 : }
7001 : }
7002 5106682 : rpo_begin = rpo_end;
7003 : }
7004 :
7005 : /* Assign ranges of consecutive node indices to each partition,
7006 : in partition order. Start with node_end being the same as
7007 : node_begin so that the next loop can use it as a counter. */
7008 716297 : unsigned int node_begin = 0;
7009 5691977 : for (auto &partition : m_partitions)
7010 : {
7011 3543086 : partition.node_begin = node_begin;
7012 3543086 : node_begin += partition.node_end;
7013 3543086 : partition.node_end = partition.node_begin;
7014 : }
7015 716297 : gcc_assert (node_begin == num_partitioned_nodes);
7016 :
7017 : /* Finally build the list of nodes in partition order. */
7018 716297 : m_partitioned_nodes.truncate (num_partitioned_nodes);
7019 6669842 : for (unsigned int node_i = 0; node_i < m_vertices.length (); ++node_i)
7020 : {
7021 5237248 : int partition_i = m_vertices[node_i].partition;
7022 5237248 : if (partition_i >= 0)
7023 : {
7024 3672517 : unsigned int order_i = m_partitions[partition_i].node_end++;
7025 3672517 : m_partitioned_nodes[order_i] = node_i;
7026 : }
7027 : }
7028 716297 : }
7029 :
7030 : /* Look for edges from earlier partitions into node NODE_I and edges from
7031 : node NODE_I into later partitions. Call:
7032 :
7033 : FN (ud, other_node_i)
7034 :
7035 : for each such use-to-def edge ud, where other_node_i is the node at the
7036 : other end of the edge. */
7037 :
7038 : template<typename T>
7039 : void
7040 4130490 : vect_optimize_slp_pass::for_each_partition_edge (unsigned int node_i, T fn)
7041 : {
7042 4130490 : int partition_i = m_vertices[node_i].partition;
7043 4130490 : for (graph_edge *pred = m_slpg->vertices[node_i].pred;
7044 7138955 : pred; pred = pred->pred_next)
7045 : {
7046 3008465 : int src_partition_i = m_vertices[pred->src].partition;
7047 3008465 : if (src_partition_i >= 0 && src_partition_i != partition_i)
7048 2678537 : fn (pred, pred->src);
7049 : }
7050 4130490 : for (graph_edge *succ = m_slpg->vertices[node_i].succ;
7051 8872060 : succ; succ = succ->succ_next)
7052 : {
7053 4741570 : int dest_partition_i = m_vertices[succ->dest].partition;
7054 4741570 : if (dest_partition_i >= 0 && dest_partition_i != partition_i)
7055 2714579 : fn (succ, succ->dest);
7056 : }
7057 4130490 : }
7058 :
7059 : /* Return true if layout LAYOUT_I is compatible with the number of SLP lanes
7060 : that NODE would operate on. This test is independent of NODE's actual
7061 : operation. */
7062 :
7063 : bool
7064 1753986 : vect_optimize_slp_pass::is_compatible_layout (slp_tree node,
7065 : unsigned int layout_i)
7066 : {
7067 1753986 : if (layout_i == 0)
7068 : return true;
7069 :
7070 1033190 : if (SLP_TREE_LANES (node) != m_perms[layout_i].length ())
7071 18355 : return false;
7072 :
7073 : return true;
7074 : }
7075 :
7076 : /* Return true if layout LAYOUT_I is compatible with the number of SLP lanes
7077 : that NODE would operate on for each NODE in PARTITION.
7078 : This test is independent of NODE's actual operations. */
7079 :
7080 : bool
7081 25071 : vect_optimize_slp_pass::is_compatible_layout (const slpg_partition_info
7082 : &partition,
7083 : unsigned int layout_i)
7084 : {
7085 50525 : for (unsigned int order_i = partition.node_begin;
7086 50525 : order_i < partition.node_end; ++order_i)
7087 : {
7088 25567 : unsigned int node_i = m_partitioned_nodes[order_i];
7089 25567 : auto &vertex = m_vertices[node_i];
7090 :
7091 : /* The layout is incompatible if it is individually incompatible
7092 : with any node in the partition. */
7093 25567 : if (!is_compatible_layout (vertex.node, layout_i))
7094 : return false;
7095 : }
7096 : return true;
7097 : }
7098 :
7099 : /* Return the cost (in arbitrary units) of going from layout FROM_LAYOUT_I
7100 : to layout TO_LAYOUT_I for a node like NODE. Return -1 if either of the
7101 : layouts is incompatible with NODE or if the change is not possible for
7102 : some other reason.
7103 :
7104 : The properties taken from NODE include the number of lanes and the
7105 : vector type. The actual operation doesn't matter. */
7106 :
7107 : int
7108 739433 : vect_optimize_slp_pass::change_layout_cost (slp_tree node,
7109 : unsigned int from_layout_i,
7110 : unsigned int to_layout_i)
7111 : {
7112 739433 : if (!is_compatible_layout (node, from_layout_i)
7113 739433 : || !is_compatible_layout (node, to_layout_i))
7114 : return -1;
7115 :
7116 738793 : if (from_layout_i == to_layout_i)
7117 : return 0;
7118 :
7119 305920 : auto_vec<slp_tree, 1> children (1);
7120 305920 : children.quick_push (node);
7121 305920 : auto_lane_permutation_t perm (SLP_TREE_LANES (node));
7122 305920 : if (from_layout_i > 0)
7123 863419 : for (unsigned int i : m_perms[from_layout_i])
7124 383785 : perm.quick_push ({ 0, i });
7125 : else
7126 479499 : for (unsigned int i = 0; i < SLP_TREE_LANES (node); ++i)
7127 333457 : perm.quick_push ({ 0, i });
7128 305920 : if (to_layout_i > 0)
7129 146791 : vect_slp_permute (m_perms[to_layout_i], perm, true);
7130 305920 : auto count = vectorizable_slp_permutation_1 (m_vinfo, nullptr, node, perm,
7131 : children, false);
7132 305920 : if (count >= 0)
7133 300547 : return MAX (count, 1);
7134 :
7135 : /* ??? In principle we could try changing via layout 0, giving two
7136 : layout changes rather than 1. Doing that would require
7137 : corresponding support in get_result_with_layout. */
7138 : return -1;
7139 305920 : }
7140 :
7141 : /* Return the costs of assigning layout LAYOUT_I to partition PARTITION_I. */
7142 :
7143 : inline slpg_partition_layout_costs &
7144 1108800 : vect_optimize_slp_pass::partition_layout_costs (unsigned int partition_i,
7145 : unsigned int layout_i)
7146 : {
7147 2217600 : return m_partition_layout_costs[partition_i * m_perms.length () + layout_i];
7148 : }
7149 :
7150 : /* Change PERM in one of two ways:
7151 :
7152 : - if IN_LAYOUT_I < 0, accept input operand I in the layout that has been
7153 : chosen for child I of NODE.
7154 :
7155 : - if IN_LAYOUT >= 0, accept all inputs operands with that layout.
7156 :
7157 : In both cases, arrange for the output to have layout OUT_LAYOUT_I */
7158 :
7159 : void
7160 38180 : vect_optimize_slp_pass::
7161 : change_vec_perm_layout (slp_tree node, lane_permutation_t &perm,
7162 : int in_layout_i, unsigned int out_layout_i)
7163 : {
7164 223810 : for (auto &entry : perm)
7165 : {
7166 109270 : int this_in_layout_i = in_layout_i;
7167 109270 : if (this_in_layout_i < 0)
7168 : {
7169 74683 : slp_tree in_node = SLP_TREE_CHILDREN (node)[entry.first];
7170 74683 : unsigned int in_partition_i = m_vertices[in_node->vertex].partition;
7171 74683 : if (in_partition_i == -1u)
7172 329 : continue;
7173 74354 : this_in_layout_i = m_partitions[in_partition_i].layout;
7174 : }
7175 108941 : if (this_in_layout_i > 0)
7176 25039 : entry.second = m_perms[this_in_layout_i][entry.second];
7177 : }
7178 38180 : if (out_layout_i > 0)
7179 10567 : vect_slp_permute (m_perms[out_layout_i], perm, true);
7180 38180 : }
7181 :
7182 : /* Check whether the target allows NODE to be rearranged so that the node's
7183 : output has layout OUT_LAYOUT_I. Return the cost of the change if so,
7184 : in the same arbitrary units as for change_layout_cost. Return -1 otherwise.
7185 :
7186 : If NODE is a VEC_PERM_EXPR and IN_LAYOUT_I < 0, also check whether
7187 : NODE can adapt to the layout changes that have (perhaps provisionally)
7188 : been chosen for NODE's children, so that no extra permutations are
7189 : needed on either the input or the output of NODE.
7190 :
7191 : If NODE is a VEC_PERM_EXPR and IN_LAYOUT_I >= 0, instead assume
7192 : that all inputs will be forced into layout IN_LAYOUT_I beforehand.
7193 :
7194 : IN_LAYOUT_I has no meaning for other types of node.
7195 :
7196 : Keeping the node as-is is always valid. If the target doesn't appear
7197 : to support the node as-is, but might realistically support other layouts,
7198 : then layout 0 instead has the cost of a worst-case permutation. On the
7199 : one hand, this ensures that every node has at least one valid layout,
7200 : avoiding what would otherwise be an awkward special case. On the other,
7201 : it still encourages the pass to change an invalid pre-existing layout
7202 : choice into a valid one. */
7203 :
7204 : int
7205 245204 : vect_optimize_slp_pass::internal_node_cost (slp_tree node, int in_layout_i,
7206 : unsigned int out_layout_i)
7207 : {
7208 245204 : const int fallback_cost = 1;
7209 :
7210 245204 : if (SLP_TREE_PERMUTE_P (node))
7211 : {
7212 31737 : auto_lane_permutation_t tmp_perm;
7213 31737 : tmp_perm.safe_splice (SLP_TREE_LANE_PERMUTATION (node));
7214 :
7215 : /* Check that the child nodes support the chosen layout. Checking
7216 : the first child is enough, since any second child would have the
7217 : same shape. */
7218 31737 : auto first_child = SLP_TREE_CHILDREN (node)[0];
7219 31737 : if (in_layout_i > 0
7220 31737 : && !is_compatible_layout (first_child, in_layout_i))
7221 : return -1;
7222 :
7223 31102 : change_vec_perm_layout (node, tmp_perm, in_layout_i, out_layout_i);
7224 62204 : int count = vectorizable_slp_permutation_1 (m_vinfo, nullptr,
7225 : node, tmp_perm,
7226 31102 : SLP_TREE_CHILDREN (node),
7227 : false);
7228 31102 : if (count < 0)
7229 : {
7230 2430 : if (in_layout_i == 0 && out_layout_i == 0)
7231 : {
7232 : /* Use the fallback cost if the node could in principle support
7233 : some nonzero layout for both the inputs and the outputs.
7234 : Otherwise assume that the node will be rejected later
7235 : and rebuilt from scalars. */
7236 595 : if (SLP_TREE_LANES (node) == SLP_TREE_LANES (first_child))
7237 : return fallback_cost;
7238 314 : return 0;
7239 : }
7240 : return -1;
7241 : }
7242 :
7243 : /* We currently have no way of telling whether the new layout is cheaper
7244 : or more expensive than the old one. But at least in principle,
7245 : it should be worth making zero permutations (whole-vector shuffles)
7246 : cheaper than real permutations, in case the pass is able to remove
7247 : the latter. */
7248 28672 : return count == 0 ? 0 : 1;
7249 31737 : }
7250 :
7251 213467 : stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node);
7252 213467 : if (rep
7253 211686 : && STMT_VINFO_DATA_REF (rep)
7254 71491 : && DR_IS_READ (STMT_VINFO_DATA_REF (rep))
7255 260788 : && SLP_TREE_LOAD_PERMUTATION (node).exists ())
7256 : {
7257 38837 : auto_load_permutation_t tmp_perm;
7258 38837 : tmp_perm.safe_splice (SLP_TREE_LOAD_PERMUTATION (node));
7259 38837 : if (out_layout_i > 0)
7260 14918 : vect_slp_permute (m_perms[out_layout_i], tmp_perm, true);
7261 :
7262 38837 : poly_uint64 vf = 1;
7263 38837 : if (auto loop_vinfo = dyn_cast<loop_vec_info> (m_vinfo))
7264 12176 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
7265 38837 : unsigned int n_perms;
7266 38837 : if (!vect_transform_slp_perm_load_1 (m_vinfo, node, tmp_perm, vNULL,
7267 : nullptr, vf, true, false, &n_perms))
7268 : {
7269 1991 : auto rep = SLP_TREE_REPRESENTATIVE (node);
7270 1991 : if (out_layout_i == 0)
7271 : {
7272 : /* Use the fallback cost if the load is an N-to-N permutation.
7273 : Otherwise assume that the node will be rejected later
7274 : and rebuilt from scalars. */
7275 1467 : if (STMT_VINFO_GROUPED_ACCESS (rep)
7276 2934 : && (DR_GROUP_SIZE (DR_GROUP_FIRST_ELEMENT (rep))
7277 1467 : == SLP_TREE_LANES (node)))
7278 667 : return fallback_cost;
7279 : return 0;
7280 : }
7281 : return -1;
7282 : }
7283 :
7284 : /* See the comment above the corresponding VEC_PERM_EXPR handling. */
7285 36846 : return n_perms == 0 ? 0 : 1;
7286 38837 : }
7287 :
7288 : return 0;
7289 : }
7290 :
7291 : /* Decide which element layouts we should consider using. Calculate the
7292 : weights associated with inserting layout changes on partition edges.
7293 : Also mark partitions that cannot change layout, by setting their
7294 : layout to zero. */
7295 :
7296 : void
7297 716297 : vect_optimize_slp_pass::start_choosing_layouts ()
7298 : {
7299 : /* Used to assign unique permutation indices. */
7300 716297 : using perm_hash = unbounded_hashmap_traits<
7301 : vec_free_hash_base<int_hash_base<unsigned>>,
7302 : int_hash<int, -1, -2>
7303 : >;
7304 716297 : hash_map<vec<unsigned>, int, perm_hash> layout_ids;
7305 :
7306 : /* Layout 0 is "no change". */
7307 716297 : m_perms.safe_push (vNULL);
7308 :
7309 : /* Create layouts from existing permutations. */
7310 716297 : auto_load_permutation_t tmp_perm;
7311 5821408 : for (unsigned int node_i : m_partitioned_nodes)
7312 : {
7313 : /* Leafs also double as entries to the reverse graph. Allow the
7314 : layout of those to be changed. */
7315 3672517 : auto &vertex = m_vertices[node_i];
7316 3672517 : auto &partition = m_partitions[vertex.partition];
7317 3672517 : if (!m_slpg->vertices[node_i].succ)
7318 939004 : partition.layout = 0;
7319 :
7320 : /* Loads and VEC_PERM_EXPRs are the only things generating permutes. */
7321 3672517 : slp_tree node = vertex.node;
7322 3672517 : stmt_vec_info dr_stmt = SLP_TREE_REPRESENTATIVE (node);
7323 3672517 : slp_tree child;
7324 3672517 : unsigned HOST_WIDE_INT imin, imax = 0;
7325 3672517 : bool any_permute = false;
7326 3672517 : tmp_perm.truncate (0);
7327 3672517 : if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
7328 : {
7329 : /* If splitting out a SLP_TREE_LANE_PERMUTATION can make the node
7330 : unpermuted, record a layout that reverses this permutation.
7331 :
7332 : We would need more work to cope with loads that are internally
7333 : permuted and also have inputs (such as masks for
7334 : IFN_MASK_LOADs). */
7335 633439 : gcc_assert (partition.layout == 0 && !m_slpg->vertices[node_i].succ);
7336 633439 : if (!STMT_VINFO_GROUPED_ACCESS (dr_stmt))
7337 : {
7338 447813 : partition.layout = -1;
7339 3650712 : continue;
7340 : }
7341 185626 : dr_stmt = DR_GROUP_FIRST_ELEMENT (dr_stmt);
7342 185626 : imin = DR_GROUP_SIZE (dr_stmt) + 1;
7343 185626 : tmp_perm.safe_splice (SLP_TREE_LOAD_PERMUTATION (node));
7344 : }
7345 5958063 : else if (SLP_TREE_PERMUTE_P (node)
7346 139975 : && SLP_TREE_CHILDREN (node).length () == 1
7347 120093 : && (child = SLP_TREE_CHILDREN (node)[0])
7348 3179053 : && (TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (child))
7349 120093 : .is_constant (&imin)))
7350 : {
7351 : /* If the child has the same vector size as this node,
7352 : reversing the permutation can make the permutation a no-op.
7353 : In other cases it can change a true permutation into a
7354 : full-vector extract. */
7355 120093 : tmp_perm.reserve (SLP_TREE_LANES (node));
7356 444538 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
7357 204352 : tmp_perm.quick_push (SLP_TREE_LANE_PERMUTATION (node)[j].second);
7358 : }
7359 : else
7360 2918985 : continue;
7361 :
7362 818723 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
7363 : {
7364 513004 : unsigned idx = tmp_perm[j];
7365 513004 : imin = MIN (imin, idx);
7366 513004 : imax = MAX (imax, idx);
7367 513004 : if (idx - tmp_perm[0] != j)
7368 163450 : any_permute = true;
7369 : }
7370 : /* If the span doesn't match we'd disrupt VF computation, avoid
7371 : that for now. */
7372 305719 : if (imax - imin + 1 != SLP_TREE_LANES (node))
7373 91633 : continue;
7374 : /* If there's no permute no need to split one out. In this case
7375 : we can consider turning a load into a permuted load, if that
7376 : turns out to be cheaper than alternatives. */
7377 214086 : if (!any_permute)
7378 : {
7379 192126 : partition.layout = -1;
7380 192126 : continue;
7381 : }
7382 :
7383 : /* For now only handle true permutes, like
7384 : vect_attempt_slp_rearrange_stmts did. This allows us to be lazy
7385 : when permuting constants and invariants keeping the permute
7386 : bijective. */
7387 21960 : auto_sbitmap load_index (SLP_TREE_LANES (node));
7388 21960 : bitmap_clear (load_index);
7389 110534 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
7390 66614 : bitmap_set_bit (load_index, tmp_perm[j] - imin);
7391 : unsigned j;
7392 87731 : for (j = 0; j < SLP_TREE_LANES (node); ++j)
7393 65926 : if (!bitmap_bit_p (load_index, j))
7394 : break;
7395 21960 : if (j != SLP_TREE_LANES (node))
7396 155 : continue;
7397 :
7398 21805 : vec<unsigned> perm = vNULL;
7399 21805 : perm.safe_grow (SLP_TREE_LANES (node), true);
7400 109257 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
7401 65647 : perm[j] = tmp_perm[j] - imin;
7402 :
7403 43610 : if (int (m_perms.length ()) >= param_vect_max_layout_candidates)
7404 : {
7405 : /* Continue to use existing layouts, but don't add any more. */
7406 0 : int *entry = layout_ids.get (perm);
7407 0 : partition.layout = entry ? *entry : 0;
7408 0 : perm.release ();
7409 : }
7410 : else
7411 : {
7412 21805 : bool existed;
7413 21805 : int &layout_i = layout_ids.get_or_insert (perm, &existed);
7414 21805 : if (existed)
7415 7771 : perm.release ();
7416 : else
7417 : {
7418 14034 : layout_i = m_perms.length ();
7419 14034 : m_perms.safe_push (perm);
7420 : }
7421 21805 : partition.layout = layout_i;
7422 : }
7423 21960 : }
7424 :
7425 : /* Initially assume that every layout is possible and has zero cost
7426 : in every partition. */
7427 716297 : m_partition_layout_costs.safe_grow_cleared (m_partitions.length ()
7428 1432594 : * m_perms.length ());
7429 :
7430 : /* We have to mark outgoing permutations facing non-associating-reduction
7431 : graph entries that are not represented as to be materialized.
7432 : slp_inst_kind_bb_reduc currently only covers associatable reductions. */
7433 3684781 : for (slp_instance instance : m_vinfo->slp_instances)
7434 1535890 : if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_ctor)
7435 : {
7436 7085 : unsigned int node_i = SLP_INSTANCE_TREE (instance)->vertex;
7437 7085 : m_partitions[m_vertices[node_i].partition].layout = 0;
7438 : }
7439 1528805 : else if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_reduc_chain)
7440 : {
7441 2348 : stmt_vec_info stmt_info
7442 2348 : = SLP_TREE_REPRESENTATIVE (SLP_INSTANCE_TREE (instance));
7443 2348 : vect_reduc_info reduc_info
7444 2348 : = info_for_reduction (as_a <loop_vec_info> (m_vinfo),
7445 : SLP_INSTANCE_TREE (instance));
7446 2348 : if (needs_fold_left_reduction_p (TREE_TYPE
7447 : (gimple_get_lhs (stmt_info->stmt)),
7448 : VECT_REDUC_INFO_CODE (reduc_info)))
7449 : {
7450 124 : unsigned int node_i = SLP_INSTANCE_TREE (instance)->vertex;
7451 124 : m_partitions[m_vertices[node_i].partition].layout = 0;
7452 : }
7453 : }
7454 :
7455 : /* Check which layouts each node and partition can handle. Calculate the
7456 : weights associated with inserting layout changes on edges. */
7457 5821408 : for (unsigned int node_i : m_partitioned_nodes)
7458 : {
7459 3672517 : auto &vertex = m_vertices[node_i];
7460 3672517 : auto &partition = m_partitions[vertex.partition];
7461 3672517 : slp_tree node = vertex.node;
7462 :
7463 3672517 : vertex.weight = vect_slp_node_weight (m_vinfo, node);
7464 :
7465 3672517 : if (stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node))
7466 : {
7467 : /* We do not handle stores with a permutation, so all
7468 : incoming permutations must have been materialized.
7469 :
7470 : We also don't handle masked grouped loads, which lack a
7471 : permutation vector. In this case the memory locations
7472 : form an implicit second input to the loads, on top of the
7473 : explicit mask input, and the memory input's layout cannot
7474 : be changed.
7475 :
7476 : On the other hand, we do support permuting gather loads and
7477 : masked gather loads, where each scalar load is independent
7478 : of the others. This can be useful if the address/index input
7479 : benefits from permutation. */
7480 3529544 : if (STMT_VINFO_DATA_REF (rep)
7481 1701816 : && STMT_VINFO_GROUPED_ACCESS (rep)
7482 4532025 : && !SLP_TREE_LOAD_PERMUTATION (node).exists ())
7483 816855 : partition.layout = 0;
7484 :
7485 : /* We cannot change the layout of an operation that is
7486 : not independent on lanes. Note this is an explicit
7487 : negative list since that's much shorter than the respective
7488 : positive one but it's critical to keep maintaining it. */
7489 3529544 : if (is_gimple_call (STMT_VINFO_STMT (rep)))
7490 32635 : switch (gimple_call_combined_fn (STMT_VINFO_STMT (rep)))
7491 : {
7492 1098 : case CFN_COMPLEX_ADD_ROT90:
7493 1098 : case CFN_COMPLEX_ADD_ROT270:
7494 1098 : case CFN_COMPLEX_MUL:
7495 1098 : case CFN_COMPLEX_MUL_CONJ:
7496 1098 : case CFN_VEC_ADDSUB:
7497 1098 : case CFN_VEC_FMADDSUB:
7498 1098 : case CFN_VEC_FMSUBADD:
7499 1098 : partition.layout = 0;
7500 : default:;
7501 : }
7502 : }
7503 :
7504 8294323 : auto process_edge = [&](graph_edge *ud, unsigned int other_node_i)
7505 : {
7506 4621806 : auto &other_vertex = m_vertices[other_node_i];
7507 :
7508 : /* Count the number of edges from earlier partitions and the number
7509 : of edges to later partitions. */
7510 4621806 : if (other_vertex.partition < vertex.partition)
7511 2310903 : partition.in_degree += 1;
7512 : else
7513 2310903 : partition.out_degree += 1;
7514 :
7515 : /* If the current node uses the result of OTHER_NODE_I, accumulate
7516 : the effects of that. */
7517 4621806 : if (ud->src == int (node_i))
7518 : {
7519 2310903 : other_vertex.out_weight += vertex.weight;
7520 2310903 : other_vertex.out_degree += 1;
7521 : }
7522 8294323 : };
7523 3672517 : for_each_partition_edge (node_i, process_edge);
7524 : }
7525 716297 : }
7526 :
7527 : /* Return the incoming costs for node NODE_I, assuming that each input keeps
7528 : its current (provisional) choice of layout. The inputs do not necessarily
7529 : have the same layout as each other. */
7530 :
7531 : slpg_layout_cost
7532 4284 : vect_optimize_slp_pass::total_in_cost (unsigned int node_i)
7533 : {
7534 4284 : auto &vertex = m_vertices[node_i];
7535 4284 : slpg_layout_cost cost;
7536 14349 : auto add_cost = [&](graph_edge *, unsigned int other_node_i)
7537 : {
7538 10065 : auto &other_vertex = m_vertices[other_node_i];
7539 10065 : if (other_vertex.partition < vertex.partition)
7540 : {
7541 6718 : auto &other_partition = m_partitions[other_vertex.partition];
7542 13436 : auto &other_costs = partition_layout_costs (other_vertex.partition,
7543 6718 : other_partition.layout);
7544 6718 : slpg_layout_cost this_cost = other_costs.in_cost;
7545 6718 : this_cost.add_serial_cost (other_costs.internal_cost);
7546 6718 : this_cost.split (other_partition.out_degree);
7547 6718 : cost.add_parallel_cost (this_cost);
7548 : }
7549 14349 : };
7550 4284 : for_each_partition_edge (node_i, add_cost);
7551 4284 : return cost;
7552 : }
7553 :
7554 : /* Return the cost of switching between layout LAYOUT1_I (at node NODE1_I)
7555 : and layout LAYOUT2_I on cross-partition use-to-def edge UD. Return
7556 : slpg_layout_cost::impossible () if the change isn't possible. */
7557 :
7558 : slpg_layout_cost
7559 739433 : vect_optimize_slp_pass::
7560 : edge_layout_cost (graph_edge *ud, unsigned int node1_i, unsigned int layout1_i,
7561 : unsigned int layout2_i)
7562 : {
7563 739433 : auto &def_vertex = m_vertices[ud->dest];
7564 739433 : auto &use_vertex = m_vertices[ud->src];
7565 739433 : auto def_layout_i = ud->dest == int (node1_i) ? layout1_i : layout2_i;
7566 739433 : auto use_layout_i = ud->dest == int (node1_i) ? layout2_i : layout1_i;
7567 739433 : auto factor = change_layout_cost (def_vertex.node, def_layout_i,
7568 : use_layout_i);
7569 739433 : if (factor < 0)
7570 6013 : return slpg_layout_cost::impossible ();
7571 :
7572 : /* We have a choice of putting the layout change at the site of the
7573 : definition or at the site of the use. Prefer the former when
7574 : optimizing for size or when the execution frequency of the
7575 : definition is no greater than the combined execution frequencies of
7576 : the uses. When putting the layout change at the site of the definition,
7577 : divvy up the cost among all consumers. */
7578 733420 : if (m_optimize_size || def_vertex.weight <= def_vertex.out_weight)
7579 : {
7580 690785 : slpg_layout_cost cost = { def_vertex.weight * factor, m_optimize_size };
7581 690785 : cost.split (def_vertex.out_degree);
7582 690785 : return cost;
7583 : }
7584 42635 : return { use_vertex.weight * factor, m_optimize_size };
7585 : }
7586 :
7587 : /* UD represents a use-def link between FROM_NODE_I and a node in a later
7588 : partition; FROM_NODE_I could be the definition node or the use node.
7589 : The node at the other end of the link wants to use layout TO_LAYOUT_I.
7590 : Return the cost of any necessary fix-ups on edge UD, or return
7591 : slpg_layout_cost::impossible () if the change isn't possible.
7592 :
7593 : At this point, FROM_NODE_I's partition has chosen the cheapest
7594 : layout based on the information available so far, but this choice
7595 : is only provisional. */
7596 :
7597 : slpg_layout_cost
7598 197502 : vect_optimize_slp_pass::forward_cost (graph_edge *ud, unsigned int from_node_i,
7599 : unsigned int to_layout_i)
7600 : {
7601 197502 : auto &from_vertex = m_vertices[from_node_i];
7602 197502 : unsigned int from_partition_i = from_vertex.partition;
7603 197502 : slpg_partition_info &from_partition = m_partitions[from_partition_i];
7604 197502 : gcc_assert (from_partition.layout >= 0);
7605 :
7606 : /* First calculate the cost on the assumption that FROM_PARTITION sticks
7607 : with its current layout preference. */
7608 197502 : slpg_layout_cost cost = slpg_layout_cost::impossible ();
7609 197502 : auto edge_cost = edge_layout_cost (ud, from_node_i,
7610 197502 : from_partition.layout, to_layout_i);
7611 197502 : if (edge_cost.is_possible ())
7612 : {
7613 388684 : auto &from_costs = partition_layout_costs (from_partition_i,
7614 194342 : from_partition.layout);
7615 194342 : cost = from_costs.in_cost;
7616 194342 : cost.add_serial_cost (from_costs.internal_cost);
7617 194342 : cost.split (from_partition.out_degree);
7618 194342 : cost.add_serial_cost (edge_cost);
7619 : }
7620 3160 : else if (from_partition.layout == 0)
7621 : /* We must allow the source partition to have layout 0 as a fallback,
7622 : in case all other options turn out to be impossible. */
7623 3160 : return cost;
7624 :
7625 : /* Take the minimum of that cost and the cost that applies if
7626 : FROM_PARTITION instead switches to TO_LAYOUT_I. */
7627 194342 : auto &direct_layout_costs = partition_layout_costs (from_partition_i,
7628 : to_layout_i);
7629 194342 : if (direct_layout_costs.is_possible ())
7630 : {
7631 170599 : slpg_layout_cost direct_cost = direct_layout_costs.in_cost;
7632 170599 : direct_cost.add_serial_cost (direct_layout_costs.internal_cost);
7633 170599 : direct_cost.split (from_partition.out_degree);
7634 170599 : if (!cost.is_possible ()
7635 170599 : || direct_cost.is_better_than (cost, m_optimize_size))
7636 33731 : cost = direct_cost;
7637 : }
7638 :
7639 194342 : return cost;
7640 : }
7641 :
7642 : /* UD represents a use-def link between TO_NODE_I and a node in an earlier
7643 : partition; TO_NODE_I could be the definition node or the use node.
7644 : The node at the other end of the link wants to use layout FROM_LAYOUT_I;
7645 : return the cost of any necessary fix-ups on edge UD, or
7646 : slpg_layout_cost::impossible () if the choice cannot be made.
7647 :
7648 : At this point, TO_NODE_I's partition has a fixed choice of layout. */
7649 :
7650 : slpg_layout_cost
7651 177493 : vect_optimize_slp_pass::backward_cost (graph_edge *ud, unsigned int to_node_i,
7652 : unsigned int from_layout_i)
7653 : {
7654 177493 : auto &to_vertex = m_vertices[to_node_i];
7655 177493 : unsigned int to_partition_i = to_vertex.partition;
7656 177493 : slpg_partition_info &to_partition = m_partitions[to_partition_i];
7657 177493 : gcc_assert (to_partition.layout >= 0);
7658 :
7659 : /* If TO_NODE_I is a VEC_PERM_EXPR consumer, see whether it can be
7660 : adjusted for this input having layout FROM_LAYOUT_I. Assume that
7661 : any other inputs keep their current choice of layout. */
7662 177493 : auto &to_costs = partition_layout_costs (to_partition_i,
7663 : to_partition.layout);
7664 177493 : if (ud->src == int (to_node_i)
7665 177211 : && SLP_TREE_PERMUTE_P (to_vertex.node))
7666 : {
7667 11468 : auto &from_partition = m_partitions[m_vertices[ud->dest].partition];
7668 11468 : auto old_layout = from_partition.layout;
7669 11468 : from_partition.layout = from_layout_i;
7670 22936 : int factor = internal_node_cost (to_vertex.node, -1,
7671 11468 : to_partition.layout);
7672 11468 : from_partition.layout = old_layout;
7673 11468 : if (factor >= 0)
7674 : {
7675 10610 : slpg_layout_cost cost = to_costs.out_cost;
7676 10610 : cost.add_serial_cost ({ to_vertex.weight * factor,
7677 : m_optimize_size });
7678 10610 : cost.split (to_partition.in_degree);
7679 10610 : return cost;
7680 : }
7681 : }
7682 :
7683 : /* Compute the cost if we insert any necessary layout change on edge UD. */
7684 166883 : auto edge_cost = edge_layout_cost (ud, to_node_i,
7685 166883 : to_partition.layout, from_layout_i);
7686 166883 : if (edge_cost.is_possible ())
7687 : {
7688 166883 : slpg_layout_cost cost = to_costs.out_cost;
7689 166883 : cost.add_serial_cost (to_costs.internal_cost);
7690 166883 : cost.split (to_partition.in_degree);
7691 166883 : cost.add_serial_cost (edge_cost);
7692 166883 : return cost;
7693 : }
7694 :
7695 0 : return slpg_layout_cost::impossible ();
7696 : }
7697 :
7698 : /* Make a forward pass through the partitions, accumulating input costs.
7699 : Make a tentative (provisional) choice of layout for each partition,
7700 : ensuring that this choice still allows later partitions to keep
7701 : their original layout. */
7702 :
7703 : void
7704 6532 : vect_optimize_slp_pass::forward_pass ()
7705 : {
7706 132653 : for (unsigned int partition_i = 0; partition_i < m_partitions.length ();
7707 : ++partition_i)
7708 : {
7709 126121 : auto &partition = m_partitions[partition_i];
7710 :
7711 : /* If the partition consists of a single VEC_PERM_EXPR, precompute
7712 : the incoming cost that would apply if every predecessor partition
7713 : keeps its current layout. This is used within the loop below. */
7714 126121 : slpg_layout_cost in_cost;
7715 126121 : slp_tree single_node = nullptr;
7716 126121 : if (partition.node_end == partition.node_begin + 1)
7717 : {
7718 119687 : unsigned int node_i = m_partitioned_nodes[partition.node_begin];
7719 119687 : single_node = m_vertices[node_i].node;
7720 119687 : if (SLP_TREE_PERMUTE_P (single_node))
7721 4284 : in_cost = total_in_cost (node_i);
7722 : }
7723 :
7724 : /* Go through the possible layouts. Decide which ones are valid
7725 : for this partition and record which of the valid layouts has
7726 : the lowest cost. */
7727 126121 : unsigned int min_layout_i = 0;
7728 126121 : slpg_layout_cost min_layout_cost = slpg_layout_cost::impossible ();
7729 388693 : for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
7730 : {
7731 262572 : auto &layout_costs = partition_layout_costs (partition_i, layout_i);
7732 262572 : if (!layout_costs.is_possible ())
7733 67316 : continue;
7734 :
7735 : /* If the recorded layout is already 0 then the layout cannot
7736 : change. */
7737 262572 : if (partition.layout == 0 && layout_i != 0)
7738 : {
7739 46012 : layout_costs.mark_impossible ();
7740 46012 : continue;
7741 : }
7742 :
7743 216560 : bool is_possible = true;
7744 441369 : for (unsigned int order_i = partition.node_begin;
7745 441369 : order_i < partition.node_end; ++order_i)
7746 : {
7747 243235 : unsigned int node_i = m_partitioned_nodes[order_i];
7748 243235 : auto &vertex = m_vertices[node_i];
7749 :
7750 : /* Reject the layout if it is individually incompatible
7751 : with any node in the partition. */
7752 243235 : if (!is_compatible_layout (vertex.node, layout_i))
7753 : {
7754 16967 : is_possible = false;
7755 18426 : break;
7756 : }
7757 :
7758 604955 : auto add_cost = [&](graph_edge *ud, unsigned int other_node_i)
7759 : {
7760 378687 : auto &other_vertex = m_vertices[other_node_i];
7761 378687 : if (other_vertex.partition < vertex.partition)
7762 : {
7763 : /* Accumulate the incoming costs from earlier
7764 : partitions, plus the cost of any layout changes
7765 : on UD itself. */
7766 197502 : auto cost = forward_cost (ud, other_node_i, layout_i);
7767 197502 : if (!cost.is_possible ())
7768 3160 : is_possible = false;
7769 : else
7770 194342 : layout_costs.in_cost.add_parallel_cost (cost);
7771 : }
7772 : else
7773 : /* Reject the layout if it would make layout 0 impossible
7774 : for later partitions. This amounts to testing that the
7775 : target supports reversing the layout change on edges
7776 : to later partitions.
7777 :
7778 : In principle, it might be possible to push a layout
7779 : change all the way down a graph, so that it never
7780 : needs to be reversed and so that the target doesn't
7781 : need to support the reverse operation. But it would
7782 : be awkward to bail out if we hit a partition that
7783 : does not support the new layout, especially since
7784 : we are not dealing with a lattice. */
7785 181185 : is_possible &= edge_layout_cost (ud, other_node_i, 0,
7786 181185 : layout_i).is_possible ();
7787 604955 : };
7788 226268 : for_each_partition_edge (node_i, add_cost);
7789 :
7790 : /* Accumulate the cost of using LAYOUT_I within NODE,
7791 : both for the inputs and the outputs. */
7792 226268 : int factor = internal_node_cost (vertex.node, layout_i,
7793 : layout_i);
7794 226268 : if (factor < 0)
7795 : {
7796 1459 : is_possible = false;
7797 1459 : break;
7798 : }
7799 224809 : else if (factor)
7800 36004 : layout_costs.internal_cost.add_serial_cost
7801 36004 : ({ vertex.weight * factor, m_optimize_size });
7802 : }
7803 216560 : if (!is_possible)
7804 : {
7805 21304 : layout_costs.mark_impossible ();
7806 21304 : continue;
7807 : }
7808 :
7809 : /* Combine the incoming and partition-internal costs. */
7810 195256 : slpg_layout_cost combined_cost = layout_costs.in_cost;
7811 195256 : combined_cost.add_serial_cost (layout_costs.internal_cost);
7812 :
7813 : /* If this partition consists of a single VEC_PERM_EXPR, see
7814 : if the VEC_PERM_EXPR can be changed to support output layout
7815 : LAYOUT_I while keeping all the provisional choices of input
7816 : layout. */
7817 195256 : if (single_node && SLP_TREE_PERMUTE_P (single_node))
7818 : {
7819 7468 : int factor = internal_node_cost (single_node, -1, layout_i);
7820 7468 : if (factor >= 0)
7821 : {
7822 6791 : auto weight = m_vertices[single_node->vertex].weight;
7823 6791 : slpg_layout_cost internal_cost
7824 6791 : = { weight * factor, m_optimize_size };
7825 :
7826 6791 : slpg_layout_cost alt_cost = in_cost;
7827 6791 : alt_cost.add_serial_cost (internal_cost);
7828 6791 : if (alt_cost.is_better_than (combined_cost, m_optimize_size))
7829 : {
7830 1622 : combined_cost = alt_cost;
7831 1622 : layout_costs.in_cost = in_cost;
7832 1622 : layout_costs.internal_cost = internal_cost;
7833 : }
7834 : }
7835 : }
7836 :
7837 : /* Record the layout with the lowest cost. Prefer layout 0 in
7838 : the event of a tie between it and another layout. */
7839 195256 : if (!min_layout_cost.is_possible ()
7840 69135 : || combined_cost.is_better_than (min_layout_cost,
7841 : m_optimize_size))
7842 : {
7843 143347 : min_layout_i = layout_i;
7844 143347 : min_layout_cost = combined_cost;
7845 : }
7846 : }
7847 :
7848 : /* This loop's handling of earlier partitions should ensure that
7849 : choosing the original layout for the current partition is no
7850 : less valid than it was in the original graph, even with the
7851 : provisional layout choices for those earlier partitions. */
7852 126121 : gcc_assert (min_layout_cost.is_possible ());
7853 126121 : partition.layout = min_layout_i;
7854 : }
7855 6532 : }
7856 :
7857 : /* Make a backward pass through the partitions, accumulating output costs.
7858 : Make a final choice of layout for each partition. */
7859 :
7860 : void
7861 6532 : vect_optimize_slp_pass::backward_pass ()
7862 : {
7863 139185 : for (unsigned int partition_i = m_partitions.length (); partition_i-- > 0;)
7864 : {
7865 126121 : auto &partition = m_partitions[partition_i];
7866 :
7867 126121 : unsigned int min_layout_i = 0;
7868 126121 : slpg_layout_cost min_layout_cost = slpg_layout_cost::impossible ();
7869 388693 : for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
7870 : {
7871 262572 : auto &layout_costs = partition_layout_costs (partition_i, layout_i);
7872 262572 : if (!layout_costs.is_possible ())
7873 67316 : continue;
7874 :
7875 : /* Accumulate the costs from successor partitions. */
7876 195256 : bool is_possible = true;
7877 417162 : for (unsigned int order_i = partition.node_begin;
7878 417162 : order_i < partition.node_end; ++order_i)
7879 : {
7880 221906 : unsigned int node_i = m_partitioned_nodes[order_i];
7881 221906 : auto &vertex = m_vertices[node_i];
7882 593262 : auto add_cost = [&](graph_edge *ud, unsigned int other_node_i)
7883 : {
7884 371356 : auto &other_vertex = m_vertices[other_node_i];
7885 371356 : auto &other_partition = m_partitions[other_vertex.partition];
7886 371356 : if (other_vertex.partition > vertex.partition)
7887 : {
7888 : /* Accumulate the incoming costs from later
7889 : partitions, plus the cost of any layout changes
7890 : on UD itself. */
7891 177493 : auto cost = backward_cost (ud, other_node_i, layout_i);
7892 177493 : if (!cost.is_possible ())
7893 0 : is_possible = false;
7894 : else
7895 177493 : layout_costs.out_cost.add_parallel_cost (cost);
7896 : }
7897 : else
7898 : /* Make sure that earlier partitions can (if necessary
7899 : or beneficial) keep the layout that they chose in
7900 : the forward pass. This ensures that there is at
7901 : least one valid choice of layout. */
7902 193863 : is_possible &= edge_layout_cost (ud, other_node_i,
7903 193863 : other_partition.layout,
7904 193863 : layout_i).is_possible ();
7905 593262 : };
7906 221906 : for_each_partition_edge (node_i, add_cost);
7907 : }
7908 195256 : if (!is_possible)
7909 : {
7910 0 : layout_costs.mark_impossible ();
7911 0 : continue;
7912 : }
7913 :
7914 : /* Locally combine the costs from the forward and backward passes.
7915 : (This combined cost is not passed on, since that would lead
7916 : to double counting.) */
7917 195256 : slpg_layout_cost combined_cost = layout_costs.in_cost;
7918 195256 : combined_cost.add_serial_cost (layout_costs.internal_cost);
7919 195256 : combined_cost.add_serial_cost (layout_costs.out_cost);
7920 :
7921 : /* Record the layout with the lowest cost. Prefer layout 0 in
7922 : the event of a tie between it and another layout. */
7923 195256 : if (!min_layout_cost.is_possible ()
7924 69135 : || combined_cost.is_better_than (min_layout_cost,
7925 : m_optimize_size))
7926 : {
7927 137082 : min_layout_i = layout_i;
7928 137082 : min_layout_cost = combined_cost;
7929 : }
7930 : }
7931 :
7932 126121 : gcc_assert (min_layout_cost.is_possible ());
7933 126121 : partition.layout = min_layout_i;
7934 : }
7935 6532 : }
7936 :
7937 : /* Return a node that applies layout TO_LAYOUT_I to the original form of NODE.
7938 : NODE already has the layout that was selected for its partition. */
7939 :
7940 : slp_tree
7941 177356 : vect_optimize_slp_pass::get_result_with_layout (slp_tree node,
7942 : unsigned int to_layout_i)
7943 : {
7944 177356 : unsigned int result_i = node->vertex * m_perms.length () + to_layout_i;
7945 177356 : slp_tree result = m_node_layouts[result_i];
7946 177356 : if (result)
7947 : return result;
7948 :
7949 176480 : if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
7950 176480 : || (SLP_TREE_DEF_TYPE (node) == vect_external_def
7951 : /* We can't permute vector defs in place. */
7952 21412 : && SLP_TREE_VEC_DEFS (node).is_empty ()))
7953 : {
7954 : /* If the vector is uniform or unchanged, there's nothing to do. */
7955 44965 : if (to_layout_i == 0 || vect_slp_tree_uniform_p (node))
7956 : result = node;
7957 : else
7958 : {
7959 3602 : auto scalar_ops = SLP_TREE_SCALAR_OPS (node).copy ();
7960 3602 : result = vect_create_new_slp_node (scalar_ops);
7961 3602 : vect_slp_permute (m_perms[to_layout_i], scalar_ops, true);
7962 : }
7963 : }
7964 : else
7965 : {
7966 131515 : unsigned int partition_i = m_vertices[node->vertex].partition;
7967 131515 : unsigned int from_layout_i = m_partitions[partition_i].layout;
7968 131515 : if (from_layout_i == to_layout_i)
7969 130321 : return node;
7970 :
7971 : /* If NODE is itself a VEC_PERM_EXPR, try to create a parallel
7972 : permutation instead of a serial one. Leave the new permutation
7973 : in TMP_PERM on success. */
7974 1194 : auto_lane_permutation_t tmp_perm;
7975 1194 : unsigned int num_inputs = 1;
7976 1194 : if (SLP_TREE_PERMUTE_P (node))
7977 : {
7978 104 : tmp_perm.safe_splice (SLP_TREE_LANE_PERMUTATION (node));
7979 104 : if (from_layout_i != 0)
7980 44 : vect_slp_permute (m_perms[from_layout_i], tmp_perm, false);
7981 104 : if (to_layout_i != 0)
7982 64 : vect_slp_permute (m_perms[to_layout_i], tmp_perm, true);
7983 104 : if (vectorizable_slp_permutation_1 (m_vinfo, nullptr, node,
7984 : tmp_perm,
7985 104 : SLP_TREE_CHILDREN (node),
7986 : false) >= 0)
7987 104 : num_inputs = SLP_TREE_CHILDREN (node).length ();
7988 : else
7989 0 : tmp_perm.truncate (0);
7990 : }
7991 :
7992 1194 : if (dump_enabled_p ())
7993 : {
7994 70 : if (tmp_perm.length () > 0)
7995 6 : dump_printf_loc (MSG_NOTE, vect_location,
7996 : "duplicating permutation node %p with"
7997 : " layout %d\n",
7998 : (void *) node, to_layout_i);
7999 : else
8000 64 : dump_printf_loc (MSG_NOTE, vect_location,
8001 : "inserting permutation node in place of %p\n",
8002 : (void *) node);
8003 : }
8004 :
8005 1194 : unsigned int num_lanes = SLP_TREE_LANES (node);
8006 1194 : result = vect_create_new_slp_node (num_inputs, VEC_PERM_EXPR);
8007 1194 : if (SLP_TREE_SCALAR_STMTS (node).exists ())
8008 : {
8009 1194 : auto &stmts = SLP_TREE_SCALAR_STMTS (result);
8010 1194 : stmts.safe_splice (SLP_TREE_SCALAR_STMTS (node));
8011 1194 : if (from_layout_i != 0)
8012 477 : vect_slp_permute (m_perms[from_layout_i], stmts, false);
8013 1194 : if (to_layout_i != 0)
8014 721 : vect_slp_permute (m_perms[to_layout_i], stmts, true);
8015 : }
8016 1194 : SLP_TREE_LANES (result) = num_lanes;
8017 1194 : SLP_TREE_VECTYPE (result) = SLP_TREE_VECTYPE (node);
8018 1194 : result->vertex = -1;
8019 :
8020 1194 : auto &lane_perm = SLP_TREE_LANE_PERMUTATION (result);
8021 1194 : if (tmp_perm.length ())
8022 : {
8023 104 : lane_perm.safe_splice (tmp_perm);
8024 104 : SLP_TREE_CHILDREN (result).safe_splice (SLP_TREE_CHILDREN (node));
8025 : }
8026 : else
8027 : {
8028 1090 : lane_perm.create (num_lanes);
8029 4448 : for (unsigned j = 0; j < num_lanes; ++j)
8030 2268 : lane_perm.quick_push ({ 0, j });
8031 1090 : if (from_layout_i != 0)
8032 433 : vect_slp_permute (m_perms[from_layout_i], lane_perm, false);
8033 1090 : if (to_layout_i != 0)
8034 657 : vect_slp_permute (m_perms[to_layout_i], lane_perm, true);
8035 1090 : SLP_TREE_CHILDREN (result).safe_push (node);
8036 : }
8037 4848 : for (slp_tree child : SLP_TREE_CHILDREN (result))
8038 1266 : child->refcnt++;
8039 1194 : }
8040 46159 : m_node_layouts[result_i] = result;
8041 46159 : return result;
8042 : }
8043 :
8044 : /* Apply the chosen vector layouts to the SLP graph. */
8045 :
8046 : void
8047 13467 : vect_optimize_slp_pass::materialize ()
8048 : {
8049 : /* We no longer need the costs, so avoid having two O(N * P) arrays
8050 : live at the same time. */
8051 13467 : m_partition_layout_costs.release ();
8052 40401 : m_node_layouts.safe_grow_cleared (m_vertices.length () * m_perms.length ());
8053 :
8054 26934 : auto_sbitmap fully_folded (m_vertices.length ());
8055 13467 : bitmap_clear (fully_folded);
8056 196517 : for (unsigned int node_i : m_partitioned_nodes)
8057 : {
8058 156116 : auto &vertex = m_vertices[node_i];
8059 156116 : slp_tree node = vertex.node;
8060 156116 : int layout_i = m_partitions[vertex.partition].layout;
8061 156116 : gcc_assert (layout_i >= 0);
8062 :
8063 : /* Rearrange the scalar statements to match the chosen layout. */
8064 156116 : if (layout_i > 0)
8065 25963 : vect_slp_permute (m_perms[layout_i],
8066 25963 : SLP_TREE_SCALAR_STMTS (node), true);
8067 :
8068 : /* Update load and lane permutations. */
8069 156116 : if (SLP_TREE_PERMUTE_P (node))
8070 : {
8071 : /* First try to absorb the input vector layouts. If that fails,
8072 : force the inputs to have layout LAYOUT_I too. We checked that
8073 : that was possible before deciding to use nonzero output layouts.
8074 : (Note that at this stage we don't really have any guarantee that
8075 : the target supports the original VEC_PERM_EXPR.) */
8076 6483 : auto &perm = SLP_TREE_LANE_PERMUTATION (node);
8077 6483 : auto_lane_permutation_t tmp_perm;
8078 6483 : tmp_perm.safe_splice (perm);
8079 6483 : change_vec_perm_layout (node, tmp_perm, -1, layout_i);
8080 6483 : if (vectorizable_slp_permutation_1 (m_vinfo, nullptr, node,
8081 : tmp_perm,
8082 6483 : SLP_TREE_CHILDREN (node),
8083 : false) >= 0)
8084 : {
8085 5888 : if (dump_enabled_p ()
8086 6816 : && !std::equal (tmp_perm.begin (), tmp_perm.end (),
8087 : perm.begin ()))
8088 58 : dump_printf_loc (MSG_NOTE, vect_location,
8089 : "absorbing input layouts into %p\n",
8090 : (void *) node);
8091 33303 : std::copy (tmp_perm.begin (), tmp_perm.end (), perm.begin ());
8092 5888 : bitmap_set_bit (fully_folded, node_i);
8093 : }
8094 : else
8095 : {
8096 : /* Not MSG_MISSED because it would make no sense to users. */
8097 595 : if (dump_enabled_p ())
8098 46 : dump_printf_loc (MSG_NOTE, vect_location,
8099 : "failed to absorb input layouts into %p\n",
8100 : (void *) node);
8101 595 : change_vec_perm_layout (nullptr, perm, layout_i, layout_i);
8102 : }
8103 6483 : }
8104 : else
8105 : {
8106 149633 : gcc_assert (!SLP_TREE_LANE_PERMUTATION (node).exists ());
8107 149633 : auto &load_perm = SLP_TREE_LOAD_PERMUTATION (node);
8108 149633 : if (layout_i > 0)
8109 : /* ??? When we handle non-bijective permutes the idea
8110 : is that we can force the load-permutation to be
8111 : { min, min + 1, min + 2, ... max }. But then the
8112 : scalar defs might no longer match the lane content
8113 : which means wrong-code with live lane vectorization.
8114 : So we possibly have to have NULL entries for those. */
8115 25171 : vect_slp_permute (m_perms[layout_i], load_perm, true);
8116 : }
8117 : }
8118 :
8119 : /* Do this before any nodes disappear, since it involves a walk
8120 : over the leaves. */
8121 13467 : remove_redundant_permutations ();
8122 :
8123 : /* Replace each child with a correctly laid-out version. */
8124 196517 : for (unsigned int node_i : m_partitioned_nodes)
8125 : {
8126 : /* Skip nodes that have already been handled above. */
8127 156116 : if (bitmap_bit_p (fully_folded, node_i))
8128 5888 : continue;
8129 :
8130 150228 : auto &vertex = m_vertices[node_i];
8131 150228 : int in_layout_i = m_partitions[vertex.partition].layout;
8132 150228 : gcc_assert (in_layout_i >= 0);
8133 :
8134 : unsigned j;
8135 : slp_tree child;
8136 447588 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (vertex.node), j, child)
8137 : {
8138 183403 : if (!child)
8139 6047 : continue;
8140 :
8141 177356 : slp_tree new_child = get_result_with_layout (child, in_layout_i);
8142 177356 : if (new_child != child)
8143 : {
8144 5379 : vect_free_slp_tree (child);
8145 5379 : SLP_TREE_CHILDREN (vertex.node)[j] = new_child;
8146 5379 : new_child->refcnt += 1;
8147 : }
8148 : }
8149 : }
8150 13467 : }
8151 :
8152 : /* Elide load permutations that are not necessary. Such permutations might
8153 : be pre-existing, rather than created by the layout optimizations. */
8154 :
8155 : void
8156 716297 : vect_optimize_slp_pass::remove_redundant_permutations ()
8157 : {
8158 4745108 : for (unsigned int node_i : m_leafs)
8159 : {
8160 2596217 : slp_tree node = m_vertices[node_i].node;
8161 2596217 : if (!SLP_TREE_LOAD_PERMUTATION (node).exists ())
8162 1962778 : continue;
8163 :
8164 : /* In basic block vectorization we allow any subchain of an interleaving
8165 : chain.
8166 : FORNOW: not in loop SLP because of realignment complications. */
8167 633439 : if (is_a <bb_vec_info> (m_vinfo))
8168 : {
8169 188823 : bool subchain_p = true;
8170 : stmt_vec_info next_load_info = NULL;
8171 : stmt_vec_info load_info;
8172 : unsigned j;
8173 188823 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
8174 : {
8175 156063 : if (j != 0
8176 156063 : && (next_load_info != load_info
8177 69285 : || ! load_info
8178 69285 : || DR_GROUP_GAP (load_info) != 1))
8179 : {
8180 : subchain_p = false;
8181 : break;
8182 : }
8183 124940 : next_load_info = DR_GROUP_NEXT_ELEMENT (load_info);
8184 : }
8185 63883 : if (subchain_p)
8186 : {
8187 32760 : SLP_TREE_LOAD_PERMUTATION (node).release ();
8188 32760 : continue;
8189 : }
8190 : }
8191 : else
8192 : {
8193 569556 : loop_vec_info loop_vinfo = as_a<loop_vec_info> (m_vinfo);
8194 569556 : bool this_load_permuted = !vect_load_perm_consecutive_p (node, 0);
8195 : /* When this isn't a grouped access we know it's single element
8196 : and contiguous. */
8197 569556 : if (!STMT_VINFO_GROUPED_ACCESS (SLP_TREE_SCALAR_STMTS (node)[0]))
8198 : {
8199 447813 : if (!this_load_permuted
8200 447813 : && (known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1U)
8201 447040 : || SLP_TREE_LANES (node) == 1))
8202 447052 : SLP_TREE_LOAD_PERMUTATION (node).release ();
8203 447813 : continue;
8204 : }
8205 121743 : stmt_vec_info first_stmt_info
8206 121743 : = DR_GROUP_FIRST_ELEMENT (SLP_TREE_SCALAR_STMTS (node)[0]);
8207 122276 : if (!this_load_permuted
8208 : /* The load requires permutation when unrolling exposes
8209 : a gap either because the group is larger than the SLP
8210 : group-size or because there is a gap between the groups. */
8211 121743 : && (known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1U)
8212 99057 : || ((SLP_TREE_LANES (node) == DR_GROUP_SIZE (first_stmt_info))
8213 147 : && DR_GROUP_GAP (first_stmt_info) == 0)))
8214 : {
8215 533 : SLP_TREE_LOAD_PERMUTATION (node).release ();
8216 533 : continue;
8217 : }
8218 : }
8219 : }
8220 716297 : }
8221 :
8222 : /* Print the partition graph and layout information to the dump file. */
8223 :
8224 : void
8225 687 : vect_optimize_slp_pass::dump ()
8226 : {
8227 687 : dump_printf_loc (MSG_NOTE, vect_location,
8228 : "SLP optimize permutations:\n");
8229 1387 : for (unsigned int layout_i = 1; layout_i < m_perms.length (); ++layout_i)
8230 : {
8231 700 : dump_printf_loc (MSG_NOTE, vect_location, " %d: { ", layout_i);
8232 700 : const char *sep = "";
8233 5953 : for (unsigned int idx : m_perms[layout_i])
8234 : {
8235 3853 : dump_printf (MSG_NOTE, "%s%d", sep, idx);
8236 3853 : sep = ", ";
8237 : }
8238 700 : dump_printf (MSG_NOTE, " }\n");
8239 : }
8240 687 : dump_printf_loc (MSG_NOTE, vect_location,
8241 : "SLP optimize partitions:\n");
8242 5968 : for (unsigned int partition_i = 0; partition_i < m_partitions.length ();
8243 : ++partition_i)
8244 : {
8245 5281 : auto &partition = m_partitions[partition_i];
8246 5281 : dump_printf_loc (MSG_NOTE, vect_location, " -------------\n");
8247 5281 : dump_printf_loc (MSG_NOTE, vect_location,
8248 : " partition %d (layout %d):\n",
8249 : partition_i, partition.layout);
8250 5281 : dump_printf_loc (MSG_NOTE, vect_location, " nodes:\n");
8251 10796 : for (unsigned int order_i = partition.node_begin;
8252 10796 : order_i < partition.node_end; ++order_i)
8253 : {
8254 5515 : auto &vertex = m_vertices[m_partitioned_nodes[order_i]];
8255 11030 : dump_printf_loc (MSG_NOTE, vect_location, " - %p:\n",
8256 5515 : (void *) vertex.node);
8257 5515 : dump_printf_loc (MSG_NOTE, vect_location,
8258 : " weight: %f\n",
8259 : vertex.weight.to_double ());
8260 5515 : if (vertex.out_degree)
8261 4326 : dump_printf_loc (MSG_NOTE, vect_location,
8262 : " out weight: %f (degree %d)\n",
8263 : vertex.out_weight.to_double (),
8264 : vertex.out_degree);
8265 5515 : if (SLP_TREE_PERMUTE_P (vertex.node))
8266 510 : dump_printf_loc (MSG_NOTE, vect_location,
8267 : " op: VEC_PERM_EXPR\n");
8268 5005 : else if (auto rep = SLP_TREE_REPRESENTATIVE (vertex.node))
8269 4987 : dump_printf_loc (MSG_NOTE, vect_location,
8270 : " op template: %G", rep->stmt);
8271 : }
8272 5281 : dump_printf_loc (MSG_NOTE, vect_location, " edges:\n");
8273 10796 : for (unsigned int order_i = partition.node_begin;
8274 10796 : order_i < partition.node_end; ++order_i)
8275 : {
8276 5515 : unsigned int node_i = m_partitioned_nodes[order_i];
8277 5515 : auto &vertex = m_vertices[node_i];
8278 16717 : auto print_edge = [&](graph_edge *, unsigned int other_node_i)
8279 : {
8280 11202 : auto &other_vertex = m_vertices[other_node_i];
8281 11202 : if (other_vertex.partition < vertex.partition)
8282 5601 : dump_printf_loc (MSG_NOTE, vect_location,
8283 : " - %p [%d] --> %p\n",
8284 5601 : (void *) other_vertex.node,
8285 : other_vertex.partition,
8286 5601 : (void *) vertex.node);
8287 : else
8288 5601 : dump_printf_loc (MSG_NOTE, vect_location,
8289 : " - %p --> [%d] %p\n",
8290 5601 : (void *) vertex.node,
8291 : other_vertex.partition,
8292 5601 : (void *) other_vertex.node);
8293 16717 : };
8294 5515 : for_each_partition_edge (node_i, print_edge);
8295 : }
8296 :
8297 16042 : for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
8298 : {
8299 10761 : auto &layout_costs = partition_layout_costs (partition_i, layout_i);
8300 10761 : if (layout_costs.is_possible ())
8301 : {
8302 8908 : dump_printf_loc (MSG_NOTE, vect_location,
8303 : " layout %d:%s\n", layout_i,
8304 8908 : partition.layout == int (layout_i)
8305 : ? " (*)" : "");
8306 8908 : slpg_layout_cost combined_cost = layout_costs.in_cost;
8307 8908 : combined_cost.add_serial_cost (layout_costs.internal_cost);
8308 8908 : combined_cost.add_serial_cost (layout_costs.out_cost);
8309 : #define TEMPLATE "{depth: %f, total: %f}"
8310 8908 : dump_printf_loc (MSG_NOTE, vect_location,
8311 : " " TEMPLATE "\n",
8312 : layout_costs.in_cost.depth.to_double (),
8313 : layout_costs.in_cost.total.to_double ());
8314 8908 : dump_printf_loc (MSG_NOTE, vect_location,
8315 : " + " TEMPLATE "\n",
8316 : layout_costs.internal_cost.depth.to_double (),
8317 : layout_costs.internal_cost.total.to_double ());
8318 8908 : dump_printf_loc (MSG_NOTE, vect_location,
8319 : " + " TEMPLATE "\n",
8320 : layout_costs.out_cost.depth.to_double (),
8321 : layout_costs.out_cost.total.to_double ());
8322 8908 : dump_printf_loc (MSG_NOTE, vect_location,
8323 : " = " TEMPLATE "\n",
8324 : combined_cost.depth.to_double (),
8325 : combined_cost.total.to_double ());
8326 : #undef TEMPLATE
8327 : }
8328 : else
8329 1853 : dump_printf_loc (MSG_NOTE, vect_location,
8330 : " layout %d: rejected\n", layout_i);
8331 : }
8332 : }
8333 687 : }
8334 :
8335 : /* Masked load lanes discovery. */
8336 :
8337 : void
8338 716297 : vect_optimize_slp_pass::decide_masked_load_lanes ()
8339 : {
8340 7387242 : for (auto v : m_vertices)
8341 : {
8342 5238351 : slp_tree node = v.node;
8343 5238351 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
8344 3670622 : || SLP_TREE_PERMUTE_P (node))
8345 1708807 : continue;
8346 3529544 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
8347 1701816 : if (! STMT_VINFO_GROUPED_ACCESS (stmt_info)
8348 : /* The mask has to be uniform. */
8349 1002481 : || STMT_VINFO_SLP_VECT_ONLY (stmt_info)
8350 1002403 : || ! is_a <gcall *> (STMT_VINFO_STMT (stmt_info))
8351 3529629 : || ! gimple_call_internal_p (STMT_VINFO_STMT (stmt_info),
8352 : IFN_MASK_LOAD))
8353 3529511 : continue;
8354 33 : stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
8355 66 : if (STMT_VINFO_STRIDED_P (stmt_info)
8356 33 : || compare_step_with_zero (m_vinfo, stmt_info) <= 0
8357 63 : || vect_load_lanes_supported (SLP_TREE_VECTYPE (node),
8358 30 : DR_GROUP_SIZE (stmt_info),
8359 : true) == IFN_LAST)
8360 33 : continue;
8361 :
8362 : /* Uniform masks need to be suitably represented. */
8363 0 : slp_tree mask = SLP_TREE_CHILDREN (node)[0];
8364 0 : if (!SLP_TREE_PERMUTE_P (mask)
8365 0 : || SLP_TREE_CHILDREN (mask).length () != 1)
8366 0 : continue;
8367 0 : bool match = true;
8368 0 : for (auto perm : SLP_TREE_LANE_PERMUTATION (mask))
8369 0 : if (perm.first != 0 || perm.second != 0)
8370 : {
8371 : match = false;
8372 : break;
8373 : }
8374 0 : if (!match)
8375 0 : continue;
8376 :
8377 : /* Now see if the consumer side matches. */
8378 0 : for (graph_edge *pred = m_slpg->vertices[node->vertex].pred;
8379 0 : pred; pred = pred->pred_next)
8380 : {
8381 0 : slp_tree pred_node = m_vertices[pred->src].node;
8382 : /* All consumers should be a permute with a single outgoing lane. */
8383 0 : if (!SLP_TREE_PERMUTE_P (pred_node)
8384 0 : || SLP_TREE_LANES (pred_node) != 1)
8385 : {
8386 : match = false;
8387 : break;
8388 : }
8389 0 : gcc_assert (SLP_TREE_CHILDREN (pred_node).length () == 1);
8390 : }
8391 0 : if (!match)
8392 0 : continue;
8393 : /* Now we can mark the nodes as to use load lanes. */
8394 0 : node->ldst_lanes = true;
8395 0 : for (graph_edge *pred = m_slpg->vertices[node->vertex].pred;
8396 0 : pred; pred = pred->pred_next)
8397 0 : m_vertices[pred->src].node->ldst_lanes = true;
8398 : /* The catch is we have to massage the mask. We have arranged
8399 : analyzed uniform masks to be represented by a splat VEC_PERM
8400 : which we can now simply elide as we cannot easily re-do SLP
8401 : discovery here. */
8402 0 : slp_tree new_mask = SLP_TREE_CHILDREN (mask)[0];
8403 0 : SLP_TREE_REF_COUNT (new_mask)++;
8404 0 : SLP_TREE_CHILDREN (node)[0] = new_mask;
8405 0 : vect_free_slp_tree (mask);
8406 : }
8407 716297 : }
8408 :
8409 : /* Perform legitimizing attempts. This is intended to improve the
8410 : situation when layout 0 is not valid which is a situation the cost
8411 : based propagation does not handle well.
8412 : Return true if further layout optimization is possible, false if
8413 : the layout configuration should be considered final. */
8414 :
8415 : bool
8416 13467 : vect_optimize_slp_pass::legitimize ()
8417 : {
8418 : /* Perform a very simple legitimizing attempt by attempting to choose
8419 : a single layout for all partitions that will make all permutations
8420 : a noop. That should also be the optimal layout choice in case
8421 : layout zero is legitimate.
8422 : ??? Disconnected components of the SLP graph could have distinct
8423 : single layouts. */
8424 13467 : int single_layout_i = -1;
8425 13467 : unsigned deferred_up_to = -1U;
8426 41788 : for (unsigned partition_i = 0; partition_i < m_partitions.length ();
8427 : ++partition_i)
8428 : {
8429 34836 : auto &partition = m_partitions[partition_i];
8430 34836 : if (single_layout_i == -1)
8431 : {
8432 17423 : single_layout_i = partition.layout;
8433 17423 : deferred_up_to = partition_i;
8434 : }
8435 17413 : else if (partition.layout == single_layout_i || partition.layout == -1)
8436 : ;
8437 : else
8438 : single_layout_i = 0;
8439 31619 : if (single_layout_i == 0)
8440 : return true;
8441 :
8442 28417 : if (single_layout_i != -1
8443 28417 : && !is_compatible_layout (partition, single_layout_i))
8444 : return true;
8445 : }
8446 :
8447 6952 : if (single_layout_i <= 0)
8448 : return true;
8449 :
8450 7545 : for (unsigned partition_i = 0; partition_i < deferred_up_to; ++partition_i)
8451 610 : if (!is_compatible_layout (m_partitions[partition_i],
8452 : single_layout_i))
8453 : return true;
8454 :
8455 21362 : for (unsigned partition_i = 0; partition_i < m_partitions.length ();
8456 : ++partition_i)
8457 : {
8458 14427 : auto &partition = m_partitions[partition_i];
8459 14427 : partition.layout = single_layout_i;
8460 : }
8461 :
8462 : return false;
8463 : }
8464 :
8465 : /* Main entry point for the SLP graph optimization pass. */
8466 :
8467 : void
8468 716297 : vect_optimize_slp_pass::run ()
8469 : {
8470 716297 : build_graph ();
8471 716297 : create_partitions ();
8472 716297 : start_choosing_layouts ();
8473 716297 : if (m_perms.length () > 1)
8474 : {
8475 13467 : if (legitimize ())
8476 : {
8477 6532 : forward_pass ();
8478 6532 : backward_pass ();
8479 : }
8480 13467 : if (dump_enabled_p ())
8481 687 : dump ();
8482 13467 : materialize ();
8483 54435 : while (!m_perms.is_empty ())
8484 27501 : m_perms.pop ().release ();
8485 : }
8486 : else
8487 702830 : remove_redundant_permutations ();
8488 716297 : free_graph (m_slpg);
8489 716297 : build_graph ();
8490 716297 : decide_masked_load_lanes ();
8491 716297 : free_graph (m_slpg);
8492 716297 : }
8493 :
8494 : /* Apply CSE to NODE and its children using BST_MAP. */
8495 :
8496 : static void
8497 5655294 : vect_cse_slp_nodes (scalar_stmts_to_slp_tree_map_t *bst_map, slp_tree& node)
8498 : {
8499 5655294 : bool put_p = false;
8500 5655294 : if (SLP_TREE_DEF_TYPE (node) == vect_internal_def
8501 : /* Besides some VEC_PERM_EXPR, two-operator nodes also
8502 : lack scalar stmts and thus CSE doesn't work via bst_map. Ideally
8503 : we'd have sth that works for all internal and external nodes. */
8504 5655294 : && !SLP_TREE_SCALAR_STMTS (node).is_empty ())
8505 : {
8506 4059163 : slp_tree *leader = bst_map->get (SLP_TREE_SCALAR_STMTS (node));
8507 4059163 : if (leader)
8508 : {
8509 : /* We've visited this node already. */
8510 420201 : if (!*leader || *leader == node)
8511 : return;
8512 :
8513 4393 : if (dump_enabled_p ())
8514 915 : dump_printf_loc (MSG_NOTE, vect_location,
8515 : "re-using SLP tree %p for %p\n",
8516 : (void *)*leader, (void *)node);
8517 4393 : vect_free_slp_tree (node);
8518 4393 : (*leader)->refcnt += 1;
8519 4393 : node = *leader;
8520 4393 : return;
8521 : }
8522 :
8523 : /* Avoid creating a cycle by populating the map only after recursion. */
8524 3638962 : bst_map->put (SLP_TREE_SCALAR_STMTS (node).copy (), nullptr);
8525 3638962 : node->refcnt += 1;
8526 3638962 : put_p = true;
8527 : /* And recurse. */
8528 : }
8529 :
8530 15638027 : for (slp_tree &child : SLP_TREE_CHILDREN (node))
8531 4576878 : if (child)
8532 4119404 : vect_cse_slp_nodes (bst_map, child);
8533 :
8534 : /* Now record the node for CSE in other siblings. */
8535 5235093 : if (put_p)
8536 3638962 : *bst_map->get (SLP_TREE_SCALAR_STMTS (node)) = node;
8537 : }
8538 :
8539 : /* Associate stmts with possible starts of a subset of lanes of NODE
8540 : in PART_STARTS. */
8541 :
8542 : static void
8543 2149344 : vect_cse_gather_part_starts (hash_set<slp_tree> &visited,
8544 : vec<vec<slp_tree>> part_starts, slp_tree node)
8545 : {
8546 : /* CSEing external nodes complicates scheduling since we materialize
8547 : those at the latest position, so avoid that. */
8548 2149344 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
8549 2149344 : || visited.add (node))
8550 : return;
8551 :
8552 : /* Besides some VEC_PERM_EXPR, two-operator nodes also lack scalar stmts
8553 : and thus CSE doesn't work. For now gather two-lane aligned starts
8554 : of nodes with a multiple of two number of lanes. */
8555 1153727 : if (!SLP_TREE_SCALAR_STMTS (node).is_empty ()
8556 1146183 : && SLP_TREE_LANES (node) > 2
8557 149849 : && (SLP_TREE_LANES (node) & 1) == 0)
8558 : {
8559 149849 : auto_vec<unsigned, 8> uids;
8560 623601 : for (unsigned i = 0; i < SLP_TREE_LANES (node); i += 2)
8561 : {
8562 473752 : stmt_vec_info s = SLP_TREE_SCALAR_STMTS (node)[i];
8563 473752 : if (!s)
8564 0 : continue;
8565 473752 : unsigned uid = gimple_uid (s->stmt);
8566 473752 : if (!uids.contains (uid))
8567 : {
8568 469757 : uids.safe_push (uid);
8569 469757 : part_starts[uid].safe_push (node);
8570 : }
8571 : }
8572 149849 : }
8573 :
8574 4492864 : for (slp_tree &child : SLP_TREE_CHILDREN (node))
8575 1318755 : if (child)
8576 1318755 : vect_cse_gather_part_starts (visited, part_starts, child);
8577 : }
8578 :
8579 : /* Apply CSE to NODE and its children using lowparts of nodes in BST_MAP. */
8580 :
8581 : static void
8582 2145669 : vect_cse_slp_node_parts (hash_set<slp_tree> &visited,
8583 : const vec<vec<slp_tree>> part_starts,
8584 : vec<slp_tree> &drops, slp_tree node)
8585 : {
8586 2145669 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
8587 2145669 : || visited.add (node))
8588 : return;
8589 :
8590 : /* Besides some VEC_PERM_EXPR, two-operator nodes also
8591 : lack scalar stmts and thus CSE doesn't work. */
8592 1152039 : unsigned HOST_WIDE_INT c;
8593 1152039 : if (!SLP_TREE_SCALAR_STMTS (node).is_empty ()
8594 1144499 : && SLP_TREE_SCALAR_STMTS (node)[0]
8595 : /* Avoid touching loads which need care with load permutations
8596 : and specialities like load-lane representations. */
8597 1144499 : && (SLP_TREE_PERMUTE_P (node)
8598 1137055 : || !STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (node))))
8599 36206 : for (slp_tree cand
8600 427801 : : part_starts[gimple_uid (SLP_TREE_SCALAR_STMTS (node)[0]->stmt)])
8601 : /* ??? There is a possible ordering/optimality problem in that
8602 : the CSE then can keep a wider feeding live even though it itself
8603 : becomes dead by means of CSE. Which might be solvable by doing
8604 : the CSE in a wide-to-narrow order. */
8605 36673 : if (SLP_TREE_LANES (cand) > SLP_TREE_LANES (node)
8606 : /* We can do high/lo extracts and full vector copies. */
8607 : && constant_multiple_p
8608 4458 : (TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (cand)),
8609 37163 : TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (node)), &c)
8610 41131 : && c <= 2)
8611 : {
8612 3968 : unsigned HOST_WIDE_INT s;
8613 3968 : bool const_p
8614 3968 : = TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (node)).is_constant (&s);
8615 3968 : unsigned i;
8616 14230 : for (i = 0; i <= SLP_TREE_LANES (cand) - SLP_TREE_LANES (node);)
8617 : {
8618 : unsigned j;
8619 15195 : for (j = 0; j < SLP_TREE_LANES (node); ++j)
8620 14728 : if (!SLP_TREE_SCALAR_STMTS (node)[j]
8621 14728 : || (SLP_TREE_SCALAR_STMTS (cand)[i+j]
8622 14728 : != SLP_TREE_SCALAR_STMTS (node)[j]))
8623 : break;
8624 10729 : if (j == SLP_TREE_LANES (node))
8625 : break;
8626 10262 : if (!const_p)
8627 : {
8628 : i = SLP_TREE_LANES (cand);
8629 : break;
8630 : }
8631 : /* We can extract only aligned on node vector type boundary. */
8632 10262 : i += s;
8633 : }
8634 3968 : if (i > SLP_TREE_LANES (cand) - SLP_TREE_LANES (node))
8635 3501 : continue;
8636 : /* Found node within cand at i. Put a permute in place
8637 : of it, selecting the subset from cand. */
8638 467 : if (dump_enabled_p ())
8639 43 : dump_printf (MSG_NOTE, "CSEd node %p as %spart of node %p\n",
8640 : (void *)node, i == 0 ? "low" : "high", (void *)cand);
8641 2298 : for (slp_tree child : SLP_TREE_CHILDREN (node))
8642 : /* Delay SLP tree release since we might still reference a node
8643 : from the part_starts map. */
8644 897 : drops.safe_push (child);
8645 467 : SLP_TREE_CHILDREN (node).truncate (1);
8646 467 : SLP_TREE_REF_COUNT (cand)++;
8647 467 : SLP_TREE_CHILDREN (node)[0] = cand;
8648 467 : SLP_TREE_CODE (node) = VEC_PERM_EXPR;
8649 467 : SLP_TREE_REPRESENTATIVE (node) = NULL;
8650 467 : SLP_TREE_LANE_PERMUTATION (node).create (SLP_TREE_LANES (node));
8651 2000 : for (unsigned j = i; j < i + SLP_TREE_LANES (node); ++j)
8652 1066 : SLP_TREE_LANE_PERMUTATION (node).quick_push (std::make_pair (0, j));
8653 2145669 : return;
8654 : }
8655 :
8656 4483290 : for (slp_tree &child : SLP_TREE_CHILDREN (node))
8657 1315080 : if (child)
8658 1315080 : vect_cse_slp_node_parts (visited, part_starts, drops, child);
8659 : }
8660 :
8661 : /* Optimize the SLP graph of VINFO. */
8662 :
8663 : void
8664 1096083 : vect_optimize_slp (vec_info *vinfo)
8665 : {
8666 1096083 : if (vinfo->slp_instances.is_empty ())
8667 814688 : return;
8668 716297 : vect_optimize_slp_pass (vinfo).run ();
8669 :
8670 : /* Apply CSE again to nodes after permute optimization. */
8671 716297 : scalar_stmts_to_slp_tree_map_t *bst_map
8672 716297 : = new scalar_stmts_to_slp_tree_map_t ();
8673 :
8674 3684781 : for (auto inst : vinfo->slp_instances)
8675 1535890 : vect_cse_slp_nodes (bst_map, SLP_INSTANCE_TREE (inst));
8676 :
8677 716297 : release_scalar_stmts_to_slp_tree_map (bst_map);
8678 :
8679 716297 : if (!is_a <bb_vec_info> (vinfo))
8680 : return;
8681 :
8682 : /* Attempt to merge SLP sub-graphs that intersect in low or highparts of
8683 : each other. Build the reverse mapping from stmt to SLP node for
8684 : lanes starting at the low or high part.
8685 : ??? In the future we can extend this to do a two-step permute
8686 : and extract or extract and permute to put the high/low part in
8687 : place on the original vector or permute the hogh/low part to
8688 : match up the target lane order. */
8689 281395 : hash_set<slp_tree> visited;
8690 281395 : vec<vec<slp_tree>> start_for_part;
8691 562790 : start_for_part.create (vinfo->stmt_vec_infos.length () + 1);
8692 562790 : start_for_part.quick_grow_cleared (vinfo->stmt_vec_infos.length () + 1);
8693 1674774 : for (auto inst : vinfo->slp_instances)
8694 830589 : vect_cse_gather_part_starts (visited,
8695 : start_for_part, SLP_INSTANCE_TREE (inst));
8696 :
8697 : /* Now replace low/highpart copies with extracting permutes. */
8698 281395 : auto_vec<slp_tree> drops;
8699 281395 : visited.empty ();
8700 1674774 : for (auto inst : vinfo->slp_instances)
8701 830589 : vect_cse_slp_node_parts (visited, start_for_part, drops,
8702 : SLP_INSTANCE_TREE (inst));
8703 :
8704 : /* Now perform delayed releases of nodes. */
8705 282786 : for (slp_tree node : drops)
8706 897 : vect_free_slp_tree (node);
8707 :
8708 24176618 : for (auto v : start_for_part)
8709 23332433 : v.release ();
8710 281395 : start_for_part.release ();
8711 281395 : }
8712 :
8713 : /* Gather loads reachable from the individual SLP graph entries. */
8714 :
8715 : void
8716 1096083 : vect_gather_slp_loads (vec_info *vinfo)
8717 : {
8718 1096083 : unsigned i;
8719 1096083 : slp_instance instance;
8720 2631973 : FOR_EACH_VEC_ELT (vinfo->slp_instances, i, instance)
8721 : {
8722 1535890 : hash_set<slp_tree> visited;
8723 1535890 : vect_gather_slp_loads (SLP_INSTANCE_LOADS (instance),
8724 : SLP_INSTANCE_TREE (instance), visited);
8725 1535890 : }
8726 1096083 : }
8727 :
8728 : /* For NODE update VF based on the number of lanes and the vector types
8729 : used. */
8730 :
8731 : static void
8732 4488549 : vect_update_slp_vf_for_node (slp_tree node, poly_uint64 &vf,
8733 : hash_set<slp_tree> &visited)
8734 : {
8735 4488549 : if (!node || SLP_TREE_DEF_TYPE (node) != vect_internal_def)
8736 1622728 : return;
8737 3240288 : if (visited.add (node))
8738 : return;
8739 :
8740 10901580 : for (slp_tree child : SLP_TREE_CHILDREN (node))
8741 3689985 : vect_update_slp_vf_for_node (child, vf, visited);
8742 :
8743 : /* We do not visit SLP nodes for constants or externals - those neither
8744 : have a vector type set yet (vectorizable_* does this).
8745 : Note that when we stop using fixed size vectors externs and constants
8746 : shouldn't influence the (minimum) vectorization factor, instead
8747 : vectorizable_* should honor the vectorization factor when trying to
8748 : assign vector types to constants and externals and cause iteration
8749 : to a higher vectorization factor when required. */
8750 2865841 : tree vectype = SLP_TREE_VECTYPE (node);
8751 2865841 : if (!vectype)
8752 : /* OMP SIMD calls w/o LHS have no SLP_TREE_VECTYPE set. */
8753 : return;
8754 2865821 : poly_uint64 node_vf
8755 2865821 : = calculate_unrolling_factor (TYPE_VECTOR_SUBPARTS (vectype),
8756 : SLP_TREE_LANES (node));
8757 2865821 : vf = force_common_multiple (vf, node_vf);
8758 :
8759 : /* For permute nodes that are fed from externs or constants we have to
8760 : consider their number of lanes as well. Likewise for store-lanes. */
8761 2865821 : if (SLP_TREE_PERMUTE_P (node) || node->ldst_lanes)
8762 718313 : for (slp_tree child : SLP_TREE_CHILDREN (node))
8763 193622 : if (SLP_TREE_DEF_TYPE (child) != vect_internal_def)
8764 : {
8765 3690 : poly_uint64 child_vf
8766 3690 : = calculate_unrolling_factor (TYPE_VECTOR_SUBPARTS (vectype),
8767 : SLP_TREE_LANES (child));
8768 3690 : vf = force_common_multiple (vf, child_vf);
8769 : }
8770 : }
8771 :
8772 : /* For each possible SLP instance decide whether to SLP it and calculate overall
8773 : unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at
8774 : least one instance. */
8775 :
8776 : bool
8777 501509 : vect_make_slp_decision (loop_vec_info loop_vinfo)
8778 : {
8779 501509 : unsigned int i;
8780 501509 : poly_uint64 unrolling_factor = 1;
8781 501509 : const vec<slp_instance> &slp_instances
8782 : = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
8783 501509 : slp_instance instance;
8784 501509 : int decided_to_slp = 0;
8785 :
8786 501509 : DUMP_VECT_SCOPE ("vect_make_slp_decision");
8787 :
8788 501509 : hash_set<slp_tree> visited;
8789 1801582 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
8790 : {
8791 798564 : slp_tree root = SLP_INSTANCE_TREE (instance);
8792 :
8793 : /* All unroll factors have the form:
8794 :
8795 : GET_MODE_SIZE (vinfo->vector_mode) * X
8796 :
8797 : for some rational X, so they must have a common multiple. */
8798 798564 : vect_update_slp_vf_for_node (root, unrolling_factor, visited);
8799 :
8800 : /* If all instances ended up with vector(1) T roots make sure to
8801 : not vectorize. RVV for example relies on loop vectorization
8802 : when some instances are essentially kept scalar. See PR121048. */
8803 798564 : if (SLP_TREE_VECTYPE (root)
8804 798564 : && known_gt (TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (root)), 1U))
8805 648657 : decided_to_slp++;
8806 : }
8807 :
8808 501509 : LOOP_VINFO_VECT_FACTOR (loop_vinfo) = unrolling_factor;
8809 :
8810 501509 : if (decided_to_slp && dump_enabled_p ())
8811 : {
8812 19664 : dump_printf_loc (MSG_NOTE, vect_location,
8813 : "Decided to SLP %d instances. Unrolling factor ",
8814 : decided_to_slp);
8815 19664 : dump_dec (MSG_NOTE, unrolling_factor);
8816 19664 : dump_printf (MSG_NOTE, "\n");
8817 : }
8818 :
8819 501509 : return (decided_to_slp > 0);
8820 501509 : }
8821 :
8822 : /* Initialize a bb_vec_info struct for the statements in BBS basic blocks. */
8823 :
8824 2285514 : _bb_vec_info::_bb_vec_info (vec<basic_block> _bbs, vec_info_shared *shared)
8825 : : vec_info (vec_info::bb, shared),
8826 2285514 : roots (vNULL)
8827 : {
8828 : /* The region we are operating on. bbs[0] is the entry, excluding
8829 : its PHI nodes. In the future we might want to track an explicit
8830 : entry edge to cover bbs[0] PHI nodes and have a region entry
8831 : insert location. */
8832 2285514 : bbs = _bbs.address ();
8833 2285514 : nbbs = _bbs.length ();
8834 :
8835 18297430 : for (unsigned i = 0; i < nbbs; ++i)
8836 : {
8837 16011916 : if (i != 0)
8838 20816991 : for (gphi_iterator si = gsi_start_phis (bbs[i]); !gsi_end_p (si);
8839 7090589 : gsi_next (&si))
8840 : {
8841 7090589 : gphi *phi = si.phi ();
8842 7090589 : gimple_set_uid (phi, 0);
8843 7090589 : add_stmt (phi);
8844 : }
8845 32023832 : for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]);
8846 147758515 : !gsi_end_p (gsi); gsi_next (&gsi))
8847 : {
8848 131746599 : gimple *stmt = gsi_stmt (gsi);
8849 131746599 : gimple_set_uid (stmt, 0);
8850 131746599 : if (is_gimple_debug (stmt) || is_a <glabel *> (stmt))
8851 85429894 : continue;
8852 46316705 : add_stmt (stmt);
8853 : }
8854 : }
8855 2285514 : }
8856 :
8857 :
8858 : /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the
8859 : stmts in the basic block. */
8860 :
8861 2285514 : _bb_vec_info::~_bb_vec_info ()
8862 : {
8863 : /* Reset region marker. */
8864 18297430 : for (unsigned i = 0; i < nbbs; ++i)
8865 : {
8866 16011916 : if (i != 0)
8867 20833772 : for (gphi_iterator si = gsi_start_phis (bbs[i]); !gsi_end_p (si);
8868 7107370 : gsi_next (&si))
8869 : {
8870 7107370 : gphi *phi = si.phi ();
8871 7107370 : gimple_set_uid (phi, -1);
8872 : }
8873 32023832 : for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]);
8874 147752280 : !gsi_end_p (gsi); gsi_next (&gsi))
8875 : {
8876 131740364 : gimple *stmt = gsi_stmt (gsi);
8877 131740364 : gimple_set_uid (stmt, -1);
8878 : }
8879 : }
8880 :
8881 3598527 : for (unsigned i = 0; i < roots.length (); ++i)
8882 : {
8883 1313013 : roots[i].stmts.release ();
8884 1313013 : roots[i].roots.release ();
8885 1313013 : roots[i].remain.release ();
8886 : }
8887 2285514 : roots.release ();
8888 2285514 : }
8889 :
8890 : /* Subroutine of vect_slp_analyze_node_operations. Handle the root of NODE,
8891 : given then that child nodes have already been processed, and that
8892 : their def types currently match their SLP node's def type. */
8893 :
8894 : static bool
8895 2722803 : vect_slp_analyze_node_operations_1 (vec_info *vinfo, slp_tree node,
8896 : slp_instance node_instance,
8897 : stmt_vector_for_cost *cost_vec)
8898 : {
8899 : /* Handle purely internal nodes. */
8900 2722803 : if (SLP_TREE_PERMUTE_P (node))
8901 : {
8902 82995 : gcc_checking_assert (!SLP_TREE_REPRESENTATIVE (node));
8903 82995 : if (!vectorizable_slp_permutation (vinfo, NULL, node, cost_vec))
8904 : return false;
8905 :
8906 : stmt_vec_info slp_stmt_info;
8907 : unsigned int i;
8908 204446 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, slp_stmt_info)
8909 : {
8910 122982 : if (slp_stmt_info
8911 117283 : && STMT_VINFO_LIVE_P (slp_stmt_info)
8912 122982 : && !vectorizable_live_operation (vinfo, slp_stmt_info, node,
8913 : node_instance, i,
8914 : false, cost_vec))
8915 : return false;
8916 : }
8917 81464 : SLP_TREE_TYPE (node) = permute_info_type;
8918 81464 : return true;
8919 : }
8920 :
8921 2639808 : return vect_analyze_stmt (vinfo, node, node_instance, cost_vec);
8922 : }
8923 :
8924 : static int
8925 1918494 : sort_ints (const void *a_, const void *b_)
8926 : {
8927 1918494 : int a = *(const int *)a_;
8928 1918494 : int b = *(const int *)b_;
8929 1918494 : return a - b;
8930 : }
8931 :
8932 : /* Verify if we can externalize a set of internal defs. */
8933 :
8934 : static bool
8935 409771 : vect_slp_can_convert_to_external (const vec<stmt_vec_info> &stmts)
8936 : {
8937 : /* Constant generation uses get_later_stmt which can only handle
8938 : defs from the same BB or a set of defs that can be ordered
8939 : with a dominance query. */
8940 409771 : basic_block bb = NULL;
8941 409771 : bool all_same = true;
8942 409771 : auto_vec<int> bbs;
8943 819542 : bbs.reserve_exact (stmts.length ());
8944 2192569 : for (stmt_vec_info stmt : stmts)
8945 : {
8946 963256 : if (!stmt)
8947 : return false;
8948 963256 : else if (!bb)
8949 409771 : bb = gimple_bb (stmt->stmt);
8950 553485 : else if (gimple_bb (stmt->stmt) != bb)
8951 187040 : all_same = false;
8952 963256 : bbs.quick_push (gimple_bb (stmt->stmt)->index);
8953 : }
8954 409771 : if (all_same)
8955 : return true;
8956 :
8957 : /* Produce a vector of unique BB indexes for the defs. */
8958 140897 : bbs.qsort (sort_ints);
8959 140897 : unsigned i, j;
8960 339412 : for (i = 1, j = 1; i < bbs.length (); ++i)
8961 198515 : if (bbs[i] != bbs[j-1])
8962 150326 : bbs[j++] = bbs[i];
8963 140897 : gcc_assert (j >= 2);
8964 140897 : bbs.truncate (j);
8965 :
8966 281794 : if (bbs.length () == 2)
8967 137494 : return (dominated_by_p (CDI_DOMINATORS,
8968 137494 : BASIC_BLOCK_FOR_FN (cfun, bbs[0]),
8969 137494 : BASIC_BLOCK_FOR_FN (cfun, bbs[1]))
8970 268632 : || dominated_by_p (CDI_DOMINATORS,
8971 131138 : BASIC_BLOCK_FOR_FN (cfun, bbs[1]),
8972 131138 : BASIC_BLOCK_FOR_FN (cfun, bbs[0])));
8973 :
8974 : /* ??? For more than two BBs we can sort the vector and verify the
8975 : result is a total order. But we can't use vec::qsort with a
8976 : compare function using a dominance query since there's no way to
8977 : signal failure and any fallback for an unordered pair would
8978 : fail qsort_chk later.
8979 : For now simply hope that ordering after BB index provides the
8980 : best candidate total order. If required we can implement our
8981 : own mergesort or export an entry without checking. */
8982 425948 : for (unsigned i = 1; i < bbs.length (); ++i)
8983 12808 : if (!dominated_by_p (CDI_DOMINATORS,
8984 12808 : BASIC_BLOCK_FOR_FN (cfun, bbs[i]),
8985 12808 : BASIC_BLOCK_FOR_FN (cfun, bbs[i-1])))
8986 : return false;
8987 :
8988 : return true;
8989 409771 : }
8990 :
8991 : /* Try to build NODE from scalars, returning true on success.
8992 : NODE_INSTANCE is the SLP instance that contains NODE. */
8993 :
8994 : static bool
8995 640205 : vect_slp_convert_to_external (vec_info *vinfo, slp_tree node,
8996 : slp_instance node_instance)
8997 : {
8998 640205 : stmt_vec_info stmt_info;
8999 640205 : unsigned int i;
9000 :
9001 640205 : if (!is_a <bb_vec_info> (vinfo)
9002 89645 : || node == SLP_INSTANCE_TREE (node_instance)
9003 29819 : || !SLP_TREE_SCALAR_STMTS (node).exists ()
9004 29778 : || vect_contains_pattern_stmt_p (SLP_TREE_SCALAR_STMTS (node))
9005 : /* Force the mask use to be built from scalars instead. */
9006 22345 : || VECTOR_BOOLEAN_TYPE_P (SLP_TREE_VECTYPE (node))
9007 662287 : || !vect_slp_can_convert_to_external (SLP_TREE_SCALAR_STMTS (node)))
9008 : return false;
9009 :
9010 22082 : if (dump_enabled_p ())
9011 82 : dump_printf_loc (MSG_NOTE, vect_location,
9012 : "Building vector operands of %p from scalars instead\n",
9013 : (void *) node);
9014 :
9015 : /* Don't remove and free the child nodes here, since they could be
9016 : referenced by other structures. The analysis and scheduling phases
9017 : (need to) ignore child nodes of anything that isn't vect_internal_def. */
9018 22082 : unsigned int group_size = SLP_TREE_LANES (node);
9019 22082 : SLP_TREE_DEF_TYPE (node) = vect_external_def;
9020 : /* Invariants get their vector type from the uses. */
9021 22082 : SLP_TREE_VECTYPE (node) = NULL_TREE;
9022 22082 : SLP_TREE_SCALAR_OPS (node).safe_grow (group_size, true);
9023 22082 : SLP_TREE_LOAD_PERMUTATION (node).release ();
9024 99048 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
9025 : {
9026 54884 : tree lhs = gimple_get_lhs (vect_orig_stmt (stmt_info)->stmt);
9027 54884 : SLP_TREE_SCALAR_OPS (node)[i] = lhs;
9028 : }
9029 : return true;
9030 : }
9031 :
9032 : /* Return true if all elements of the slice are the same. */
9033 : bool
9034 449055 : vect_scalar_ops_slice::all_same_p () const
9035 : {
9036 500805 : for (unsigned int i = 1; i < length; ++i)
9037 413700 : if (!operand_equal_p (op (0), op (i)))
9038 : return false;
9039 : return true;
9040 : }
9041 :
9042 : hashval_t
9043 424573 : vect_scalar_ops_slice_hash::hash (const value_type &s)
9044 : {
9045 424573 : hashval_t hash = 0;
9046 1625201 : for (unsigned i = 0; i < s.length; ++i)
9047 1200628 : hash = iterative_hash_expr (s.op (i), hash);
9048 424573 : return hash;
9049 : }
9050 :
9051 : bool
9052 233034 : vect_scalar_ops_slice_hash::equal (const value_type &s1,
9053 : const compare_type &s2)
9054 : {
9055 233034 : if (s1.length != s2.length)
9056 : return false;
9057 401321 : for (unsigned i = 0; i < s1.length; ++i)
9058 350167 : if (!operand_equal_p (s1.op (i), s2.op (i)))
9059 : return false;
9060 : return true;
9061 : }
9062 :
9063 : /* Like vect_get_num_copies but N copies of the vector might have
9064 : excess elements in the last vector. Stores the number of excess
9065 : elements in the last vector in *EXCESS_ELTS.
9066 : Returns false if *NVECTORS or *EXCESS_ELTS cannot be computed. */
9067 :
9068 : bool
9069 1597250 : vect_get_num_copies_for_invariant (vec_info *vinfo, slp_tree node,
9070 : unsigned *nvectors,
9071 : unsigned *excess_elts)
9072 : {
9073 1597250 : poly_uint64 vf;
9074 :
9075 1597250 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
9076 374260 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
9077 : else
9078 1222990 : vf = 1;
9079 1597250 : vf *= SLP_TREE_LANES (node);
9080 :
9081 1597250 : tree vectype = SLP_TREE_VECTYPE (node);
9082 1597250 : uint64_t rem;
9083 1597250 : if (can_div_away_from_zero_p (vf, TYPE_VECTOR_SUBPARTS (vectype),
9084 : nvectors)
9085 1597250 : && (TYPE_VECTOR_SUBPARTS (vectype) * *nvectors - vf) .is_constant (&rem))
9086 : {
9087 1597250 : *excess_elts = rem;
9088 1597250 : return true;
9089 : }
9090 : return false;
9091 : }
9092 :
9093 : /* Compute the prologue cost for invariant or constant operands represented
9094 : by NODE. */
9095 :
9096 : static void
9097 1094129 : vect_prologue_cost_for_slp (slp_tree node, unsigned nvectors,
9098 : stmt_vector_for_cost *cost_vec)
9099 : {
9100 : /* There's a special case of an existing vector, that costs nothing. */
9101 1094129 : if (SLP_TREE_SCALAR_OPS (node).length () == 0
9102 1094129 : && !SLP_TREE_VEC_DEFS (node).is_empty ())
9103 2146 : return;
9104 : /* Without looking at the actual initializer a vector of
9105 : constants can be implemented as load from the constant pool.
9106 : When all elements are the same we can use a splat. */
9107 1091983 : tree vectype = SLP_TREE_VECTYPE (node);
9108 1091983 : unsigned group_size = SLP_TREE_LANES (node);
9109 1091983 : unsigned HOST_WIDE_INT const_nunits;
9110 1091983 : unsigned nelt_limit;
9111 1091983 : auto ops = &SLP_TREE_SCALAR_OPS (node);
9112 1091983 : auto_vec<unsigned int> starts (nvectors);
9113 1091983 : if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits)
9114 1091983 : && ! multiple_p (const_nunits, group_size))
9115 : {
9116 67227 : nelt_limit = const_nunits;
9117 67227 : hash_set<vect_scalar_ops_slice_hash> vector_ops;
9118 345287 : for (unsigned int i = 0; i < nvectors; ++i)
9119 210833 : if (!vector_ops.add ({ ops, i * nelt_limit, nelt_limit }))
9120 159679 : starts.quick_push (i * nelt_limit);
9121 67227 : }
9122 : else
9123 : {
9124 : /* If either the vector has variable length or the vectors
9125 : are composed of repeated whole groups we only need to
9126 : cost construction once. All vectors will be the same. */
9127 1024756 : nelt_limit = group_size;
9128 1024756 : starts.quick_push (0);
9129 : }
9130 : /* ??? We're just tracking whether vectors in a single node are the same.
9131 : Ideally we'd do something more global. */
9132 4460384 : for (unsigned int start : starts)
9133 : {
9134 1184435 : vect_cost_for_stmt kind;
9135 1184435 : if (SLP_TREE_DEF_TYPE (node) == vect_constant_def)
9136 : kind = vector_load;
9137 449055 : else if (vect_scalar_ops_slice { ops, start, nelt_limit }.all_same_p ())
9138 : kind = scalar_to_vec;
9139 : else
9140 361950 : kind = vec_construct;
9141 1184435 : record_stmt_cost (cost_vec, 1, kind, nullptr, node,
9142 : vectype, 0, vect_prologue);
9143 : }
9144 1091983 : }
9145 :
9146 : /* Analyze statements contained in SLP tree NODE after recursively analyzing
9147 : the subtree. NODE_INSTANCE contains NODE and VINFO contains INSTANCE.
9148 :
9149 : Return true if the operations are supported. */
9150 :
9151 : static bool
9152 5053948 : vect_slp_analyze_node_operations (vec_info *vinfo, slp_tree node,
9153 : slp_instance node_instance,
9154 : hash_set<slp_tree> &visited_set,
9155 : vec<slp_tree> &visited_vec,
9156 : stmt_vector_for_cost *cost_vec)
9157 : {
9158 5053948 : int i, j;
9159 5053948 : slp_tree child;
9160 :
9161 : /* Costing and analysis of invariants is delayed. */
9162 5053948 : if (!node
9163 4660003 : || SLP_TREE_DEF_TYPE (node) == vect_constant_def
9164 3868456 : || SLP_TREE_DEF_TYPE (node) == vect_external_def)
9165 : return true;
9166 :
9167 3383763 : if (SLP_TREE_DEF_TYPE (node) == vect_uninitialized_def)
9168 : {
9169 5 : if (dump_enabled_p ())
9170 0 : dump_printf_loc (MSG_NOTE, vect_location,
9171 : "Failed cyclic SLP reference in %p\n", (void *) node);
9172 : return false;
9173 : }
9174 3383758 : gcc_assert (SLP_TREE_DEF_TYPE (node) == vect_internal_def);
9175 :
9176 : /* If we already analyzed the exact same set of scalar stmts we're done.
9177 : We share the generated vector stmts for those. */
9178 3383758 : if (visited_set.add (node))
9179 : return true;
9180 3084518 : visited_vec.safe_push (node);
9181 :
9182 3084518 : bool res = true;
9183 3084518 : unsigned visited_rec_start = visited_vec.length ();
9184 3084518 : unsigned cost_vec_rec_start = cost_vec->length ();
9185 3084518 : bool seen_non_constant_child = false;
9186 8928629 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
9187 : {
9188 3685583 : res = vect_slp_analyze_node_operations (vinfo, child, node_instance,
9189 : visited_set, visited_vec,
9190 : cost_vec);
9191 3685583 : if (!res)
9192 : break;
9193 3324082 : if (child && SLP_TREE_DEF_TYPE (child) != vect_constant_def)
9194 3324082 : seen_non_constant_child = true;
9195 : }
9196 : /* We're having difficulties scheduling nodes with just constant
9197 : operands and no scalar stmts since we then cannot compute a stmt
9198 : insertion place. */
9199 3084518 : if (res
9200 3084518 : && !seen_non_constant_child
9201 3084518 : && SLP_TREE_SCALAR_STMTS (node).is_empty ())
9202 : {
9203 214 : if (dump_enabled_p ())
9204 6 : dump_printf_loc (MSG_NOTE, vect_location,
9205 : "Cannot vectorize all-constant op node %p\n",
9206 : (void *) node);
9207 : res = false;
9208 : }
9209 :
9210 3084310 : if (res)
9211 2722803 : res = vect_slp_analyze_node_operations_1 (vinfo, node, node_instance,
9212 : cost_vec);
9213 : /* If analysis failed we have to pop all recursive visited nodes
9214 : plus ourselves. */
9215 3084518 : if (!res)
9216 : {
9217 3158768 : while (visited_vec.length () >= visited_rec_start)
9218 939179 : visited_set.remove (visited_vec.pop ());
9219 640205 : cost_vec->truncate (cost_vec_rec_start);
9220 : }
9221 :
9222 : /* When the node can be vectorized cost invariant nodes it references.
9223 : This is not done in DFS order to allow the referring node
9224 : vectorizable_* calls to nail down the invariant nodes vector type
9225 : and possibly unshare it if it needs a different vector type than
9226 : other referrers. */
9227 3084518 : if (res)
9228 9392736 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), j, child)
9229 2969583 : if (child
9230 2648655 : && (SLP_TREE_DEF_TYPE (child) == vect_constant_def
9231 2648655 : || SLP_TREE_DEF_TYPE (child) == vect_external_def)
9232 : /* Perform usual caching, note code-generation still
9233 : code-gens these nodes multiple times but we expect
9234 : to CSE them later. */
9235 4104399 : && !visited_set.add (child))
9236 : {
9237 1131599 : visited_vec.safe_push (child);
9238 : /* ??? After auditing more code paths make a "default"
9239 : and push the vector type from NODE to all children
9240 : if it is not already set. */
9241 : /* Compute the number of vectors to be generated. */
9242 1131599 : tree vector_type = SLP_TREE_VECTYPE (child);
9243 1131599 : if (!vector_type)
9244 : {
9245 : /* Masked loads can have an undefined (default SSA definition)
9246 : else operand. We do not need to cost it. */
9247 37470 : vec<tree> ops = SLP_TREE_SCALAR_OPS (child);
9248 38901 : if (SLP_TREE_TYPE (node) == load_vec_info_type
9249 38901 : && ((ops.length ()
9250 1431 : && TREE_CODE (ops[0]) == SSA_NAME
9251 0 : && SSA_NAME_IS_DEFAULT_DEF (ops[0])
9252 0 : && VAR_P (SSA_NAME_VAR (ops[0])))
9253 1431 : || SLP_TREE_DEF_TYPE (child) == vect_constant_def))
9254 1431 : continue;
9255 :
9256 : /* For shifts with a scalar argument we don't need
9257 : to cost or code-generate anything.
9258 : ??? Represent this more explicitly. */
9259 36039 : gcc_assert (SLP_TREE_TYPE (node) == shift_vec_info_type
9260 : && j == 1);
9261 36039 : continue;
9262 36039 : }
9263 :
9264 : /* Make sure we can generate them and then cost them. */
9265 1094129 : unsigned nvectors, excess_elts;
9266 1094129 : if (!vect_get_num_copies_for_invariant (vinfo, child, &nvectors,
9267 : &excess_elts))
9268 0 : return false;
9269 1094129 : vect_prologue_cost_for_slp (child, nvectors, cost_vec);
9270 : }
9271 :
9272 : /* If this node or any of its children can't be vectorized, try pruning
9273 : the tree here rather than felling the whole thing. */
9274 640205 : if (!res && vect_slp_convert_to_external (vinfo, node, node_instance))
9275 : {
9276 : /* We'll need to revisit this for invariant costing and number
9277 : of vectorized stmt setting. */
9278 : res = true;
9279 : }
9280 :
9281 : return res;
9282 : }
9283 :
9284 : /* Mark lanes of NODE that are live outside of the basic-block vectorized
9285 : region and that can be vectorized using vectorizable_live_operation
9286 : with STMT_VINFO_LIVE_P. Not handled live operations will cause the
9287 : scalar code computing it to be retained. */
9288 :
9289 : static void
9290 997441 : vect_bb_slp_mark_live_stmts (bb_vec_info bb_vinfo, slp_tree node,
9291 : slp_instance instance,
9292 : stmt_vector_for_cost *cost_vec,
9293 : hash_set<stmt_vec_info> &svisited,
9294 : hash_set<slp_tree> &visited)
9295 : {
9296 997441 : if (visited.add (node))
9297 58023 : return;
9298 :
9299 : unsigned i;
9300 : stmt_vec_info stmt_info;
9301 : gimple *last_stmt = NULL;
9302 3372726 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
9303 : {
9304 2433308 : if (!stmt_info || svisited.contains (stmt_info))
9305 93611 : continue;
9306 2398534 : stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info);
9307 2398534 : if (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)
9308 35698 : && STMT_VINFO_RELATED_STMT (orig_stmt_info) != stmt_info)
9309 : /* Only the pattern root stmt computes the original scalar value. */
9310 29737 : continue;
9311 2368797 : if (!PURE_SLP_STMT (orig_stmt_info))
9312 : /* Iff the stmt is not part of the vector coverage because it or
9313 : uses of it are used by SLP graph leafs as extern input there is
9314 : no point in trying to live code-generate from a vector stmt as
9315 : the scalar stmt will survive anyway. */
9316 29100 : continue;
9317 2339697 : bool mark_visited = true;
9318 2339697 : gimple *orig_stmt = orig_stmt_info->stmt;
9319 2339697 : ssa_op_iter op_iter;
9320 2339697 : def_operand_p def_p;
9321 5245773 : FOR_EACH_PHI_OR_STMT_DEF (def_p, orig_stmt, op_iter, SSA_OP_DEF)
9322 : {
9323 : /* We have to verify whether we can insert the lane extract
9324 : before all uses. The following is a conservative approximation.
9325 : We cannot put this into vectorizable_live_operation because
9326 : iterating over all use stmts from inside a FOR_EACH_IMM_USE_STMT
9327 : doesn't work.
9328 : Note that while the fact that we emit code for loads at the
9329 : first load should make this a non-problem leafs we construct
9330 : from scalars are vectorized after the last scalar def.
9331 : ??? If we'd actually compute the insert location during
9332 : analysis we could use sth less conservative than the last
9333 : scalar stmt in the node for the dominance check. */
9334 : /* ??? What remains is "live" uses in vector CTORs in the same
9335 : SLP graph which is where those uses can end up code-generated
9336 : right after their definition instead of close to their original
9337 : use. But that would restrict us to code-generate lane-extracts
9338 : from the latest stmt in a node. So we compensate for this
9339 : during code-generation, simply not replacing uses for those
9340 : hopefully rare cases. */
9341 566379 : imm_use_iterator use_iter;
9342 :
9343 566379 : bool live_p = false;
9344 566379 : bool can_insert = true;
9345 566379 : use_operand_p use_p;
9346 1699664 : FOR_EACH_IMM_USE_FAST (use_p, use_iter, DEF_FROM_PTR (def_p))
9347 : {
9348 1141374 : gimple *use_stmt = USE_STMT (use_p);
9349 1141374 : stmt_vec_info use_stmt_info;
9350 1141374 : if (!(!is_gimple_debug (use_stmt)
9351 878633 : && (!(use_stmt_info = bb_vinfo->lookup_stmt (use_stmt))
9352 868103 : || !PURE_SLP_STMT (use_stmt_info))))
9353 931460 : continue;
9354 209914 : live_p = true;
9355 209914 : if (!last_stmt)
9356 65103 : last_stmt
9357 65365 : = (node->si ? node->si
9358 262 : : vect_find_last_scalar_stmt_in_slp (node)->stmt);
9359 209914 : if (is_a <gphi *> (use_stmt))
9360 : {
9361 72076 : if (!dominated_by_p (CDI_DOMINATORS,
9362 36038 : phi_arg_edge_from_use (use_p)->src,
9363 36038 : gimple_bb (last_stmt)))
9364 : can_insert = false;
9365 : }
9366 : /* As we instert after last_stmt it may not be the use_stmt
9367 : itself. */
9368 173876 : else if (last_stmt == use_stmt
9369 173876 : || !vect_stmt_dominates_stmt_p (last_stmt, use_stmt))
9370 : can_insert = false;
9371 : if (!can_insert)
9372 : {
9373 8089 : if (dump_enabled_p ())
9374 80 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9375 : "Cannot determine insertion place for "
9376 : "lane extract of %T at node %p\n",
9377 : DEF_FROM_PTR (def_p), (void *)node);
9378 : can_insert = false;
9379 : break;
9380 : }
9381 566379 : }
9382 566379 : if (live_p && can_insert)
9383 : {
9384 : /* Only record a live stmt when we can replace all uses. We
9385 : record from which SLP tree we vectorize the uses, so we'll
9386 : cost once and can deal with the case that not all SLP nodes
9387 : may be suitable for code-generation of all live uses.
9388 : ??? But we never split up the work between multiple SLP
9389 : nodes. */
9390 104993 : STMT_VINFO_LIVE_P (stmt_info) = true;
9391 104993 : if (!vectorizable_live_operation (bb_vinfo, stmt_info, node,
9392 : instance, i, false, cost_vec))
9393 : {
9394 0 : STMT_VINFO_LIVE_P (stmt_info) = false;
9395 0 : mark_visited = false;
9396 : }
9397 : }
9398 : }
9399 2339697 : if (mark_visited)
9400 2339697 : svisited.add (stmt_info);
9401 : }
9402 :
9403 : slp_tree child;
9404 2747493 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
9405 981921 : if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
9406 279931 : vect_bb_slp_mark_live_stmts (bb_vinfo, child, instance, cost_vec,
9407 : svisited, visited);
9408 : }
9409 :
9410 : /* Traverse all slp instances of BB_VINFO, and mark lanes of every node that
9411 : are live outside of the basic-block vectorized region and that can be
9412 : vectorized using vectorizable_live_operation with STMT_VINFO_LIVE_P. */
9413 :
9414 : static void
9415 247322 : vect_bb_slp_mark_live_stmts (bb_vec_info bb_vinfo)
9416 : {
9417 247322 : if (bb_vinfo->slp_instances.is_empty ())
9418 0 : return;
9419 :
9420 247322 : hash_set<slp_tree> visited;
9421 247322 : hash_set<stmt_vec_info> svisited;
9422 1459476 : for (slp_instance instance : bb_vinfo->slp_instances)
9423 : {
9424 717510 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
9425 51892 : STMT_VINFO_LIVE_P (SLP_INSTANCE_ROOT_STMTS (instance)[0]) = true;
9426 717510 : vect_location = instance->location ();
9427 717510 : vect_bb_slp_mark_live_stmts (bb_vinfo, SLP_INSTANCE_TREE (instance),
9428 : instance, &instance->cost_vec,
9429 : svisited, visited);
9430 : }
9431 247322 : }
9432 :
9433 : /* Determine whether we can vectorize the reduction epilogue for INSTANCE. */
9434 :
9435 : static bool
9436 90113 : vectorizable_bb_reduc_epilogue (slp_instance instance,
9437 : stmt_vector_for_cost *cost_vec)
9438 : {
9439 90113 : gassign *stmt = as_a <gassign *> (instance->root_stmts[0]->stmt);
9440 90113 : enum tree_code reduc_code = gimple_assign_rhs_code (stmt);
9441 90113 : if (reduc_code == MINUS_EXPR)
9442 0 : reduc_code = PLUS_EXPR;
9443 90113 : internal_fn reduc_fn;
9444 90113 : tree vectype = SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (instance));
9445 90113 : if (!vectype
9446 90101 : || !useless_type_conversion_p (TREE_TYPE (gimple_assign_lhs (stmt)),
9447 90101 : TREE_TYPE (vectype))
9448 53212 : || (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), 2u)
9449 993 : && (!reduction_fn_for_scalar_code (reduc_code, &reduc_fn)
9450 993 : || reduc_fn == IFN_LAST
9451 993 : || !direct_internal_fn_supported_p (reduc_fn, vectype,
9452 : OPTIMIZE_FOR_BOTH)))
9453 : /* Two-element reductions do not need special-handling for fold-left,
9454 : other cases are not yet implemented. remain_defs also have to
9455 : be included here. */
9456 180262 : || (needs_fold_left_reduction_p (TREE_TYPE (vectype), reduc_code)
9457 5948 : && (!instance->remain_defs.is_empty ()
9458 1878 : || SLP_TREE_LANES (SLP_INSTANCE_TREE (instance)) != 2)))
9459 : {
9460 38986 : if (dump_enabled_p ())
9461 60 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9462 : "not vectorized: basic block reduction epilogue "
9463 : "operation unsupported.\n");
9464 : return false;
9465 : }
9466 :
9467 : /* There's no way to cost a horizontal vector reduction via REDUC_FN so
9468 : cost log2 vector operations plus shuffles and one extraction. */
9469 51127 : unsigned steps = floor_log2 (vect_nunits_for_cost (vectype));
9470 51127 : record_stmt_cost (cost_vec, steps, vector_stmt, instance->root_stmts[0],
9471 : vectype, 0, vect_body);
9472 51127 : record_stmt_cost (cost_vec, steps, vec_perm, instance->root_stmts[0],
9473 : vectype, 0, vect_body);
9474 51127 : record_stmt_cost (cost_vec, 1, vec_to_scalar, instance->root_stmts[0],
9475 : vectype, 0, vect_body);
9476 :
9477 : /* Since we replace all stmts of a possibly longer scalar reduction
9478 : chain account for the extra scalar stmts for that. */
9479 51127 : if (!instance->remain_defs.is_empty ())
9480 31990 : record_stmt_cost (cost_vec, instance->remain_defs.length (), scalar_stmt,
9481 15995 : instance->root_stmts[0], 0, vect_body);
9482 : return true;
9483 : }
9484 :
9485 : /* Prune from ROOTS all stmts that are computed as part of lanes of NODE
9486 : and recurse to children. */
9487 :
9488 : static void
9489 362246 : vect_slp_prune_covered_roots (slp_tree node, hash_set<stmt_vec_info> &roots,
9490 : hash_set<slp_tree> &visited)
9491 : {
9492 362246 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
9493 362246 : || visited.add (node))
9494 : return;
9495 :
9496 1014792 : for (auto stmt : SLP_TREE_SCALAR_STMTS (node))
9497 450898 : if (stmt)
9498 482333 : roots.remove (vect_orig_stmt (stmt));
9499 :
9500 766290 : for (auto child : SLP_TREE_CHILDREN (node))
9501 255786 : if (child)
9502 254244 : vect_slp_prune_covered_roots (child, roots, visited);
9503 : }
9504 :
9505 : /* Hand over COST_VEC to the target COSTS grouped by SLP node. */
9506 :
9507 : static void
9508 1029856 : add_slp_costs (vector_costs *costs, stmt_vector_for_cost& cost_vec)
9509 : {
9510 3828528 : for (unsigned start = 0; start < cost_vec.length ();)
9511 : {
9512 2798672 : unsigned end = start + 1;
9513 3460377 : while (end < cost_vec.length ()
9514 5899621 : && cost_vec[start].node == cost_vec[end].node)
9515 661705 : end++;
9516 2798672 : if (cost_vec[start].node)
9517 2756324 : costs->add_slp_cost (cost_vec[start].node,
9518 2756324 : array_slice<stmt_info_for_cost>
9519 2756324 : (cost_vec.begin () + start, end - start));
9520 : else
9521 42348 : costs->vector_costs::add_slp_cost (cost_vec[start].node,
9522 42348 : array_slice<stmt_info_for_cost>
9523 42348 : (cost_vec.begin () + start, end - start));
9524 2798672 : start = end;
9525 : }
9526 1029856 : }
9527 :
9528 : /* Analyze statements in SLP instances of VINFO. Return true if the
9529 : operations are supported. */
9530 :
9531 : bool
9532 696476 : vect_slp_analyze_operations (vec_info *vinfo)
9533 : {
9534 696476 : slp_instance instance;
9535 696476 : int i;
9536 :
9537 696476 : DUMP_VECT_SCOPE ("vect_slp_analyze_operations");
9538 :
9539 696476 : hash_set<slp_tree> visited;
9540 2501152 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); )
9541 : {
9542 1368365 : auto_vec<slp_tree> visited_vec;
9543 1368365 : stmt_vector_for_cost cost_vec;
9544 1368365 : cost_vec.create (2);
9545 1368365 : if (is_a <bb_vec_info> (vinfo))
9546 821880 : vect_location = instance->location ();
9547 1368365 : if (!vect_slp_analyze_node_operations (vinfo,
9548 : SLP_INSTANCE_TREE (instance),
9549 : instance, visited, visited_vec,
9550 : &cost_vec)
9551 : /* CTOR instances require vectorized defs for the SLP tree root. */
9552 1111738 : || (SLP_INSTANCE_KIND (instance) == slp_inst_kind_ctor
9553 6315 : && (SLP_TREE_DEF_TYPE (SLP_INSTANCE_TREE (instance))
9554 : != vect_internal_def
9555 : /* Make sure we vectorized with the expected type. */
9556 6315 : || !useless_type_conversion_p
9557 6315 : (TREE_TYPE (TREE_TYPE (gimple_assign_rhs1
9558 : (instance->root_stmts[0]->stmt))),
9559 6315 : TREE_TYPE (SLP_TREE_VECTYPE
9560 : (SLP_INSTANCE_TREE (instance))))))
9561 : /* Check we can vectorize the reduction. */
9562 1111723 : || (SLP_INSTANCE_KIND (instance) == slp_inst_kind_bb_reduc
9563 90113 : && !vectorizable_bb_reduc_epilogue (instance, &cost_vec))
9564 : /* Check we can vectorize the gcond. */
9565 2441102 : || (SLP_INSTANCE_KIND (instance) == slp_inst_kind_gcond
9566 66034 : && !vectorizable_early_exit (as_a <loop_vec_info> (vinfo),
9567 66034 : SLP_INSTANCE_ROOT_STMTS (instance)[0],
9568 : NULL,
9569 : SLP_INSTANCE_TREE (instance),
9570 : &cost_vec)))
9571 : {
9572 358992 : cost_vec.release ();
9573 358992 : slp_tree node = SLP_INSTANCE_TREE (instance);
9574 358992 : stmt_vec_info stmt_info;
9575 358992 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
9576 268599 : stmt_info = SLP_INSTANCE_ROOT_STMTS (instance)[0];
9577 90393 : else if (!SLP_TREE_SCALAR_STMTS (node).is_empty ()
9578 90393 : && SLP_TREE_SCALAR_STMTS (node)[0])
9579 : stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
9580 : else
9581 0 : stmt_info = SLP_TREE_REPRESENTATIVE (node);
9582 358992 : if (is_a <loop_vec_info> (vinfo))
9583 : {
9584 260165 : if (dump_enabled_p ())
9585 6899 : dump_printf_loc (MSG_NOTE, vect_location,
9586 : "unsupported SLP instance starting from: %G",
9587 : stmt_info->stmt);
9588 260165 : return false;
9589 : }
9590 98827 : if (dump_enabled_p ())
9591 120 : dump_printf_loc (MSG_NOTE, vect_location,
9592 : "removing SLP instance operations starting from: %G",
9593 : stmt_info->stmt);
9594 235558 : while (!visited_vec.is_empty ())
9595 : {
9596 136731 : slp_tree node = visited_vec.pop ();
9597 136731 : SLP_TREE_TYPE (node) = undef_vec_info_type;
9598 136731 : if (node->data)
9599 : {
9600 8545 : delete node->data;
9601 8545 : node->data = nullptr;
9602 : }
9603 136731 : visited.remove (node);
9604 : }
9605 98827 : vect_free_slp_instance (instance);
9606 98827 : vinfo->slp_instances.ordered_remove (i);
9607 : }
9608 : else
9609 : {
9610 1009373 : i++;
9611 1009373 : if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
9612 : {
9613 286320 : add_slp_costs (loop_vinfo->vector_costs, cost_vec);
9614 286320 : cost_vec.release ();
9615 : }
9616 : else
9617 : /* For BB vectorization remember the SLP graph entry
9618 : cost for later. */
9619 723053 : instance->cost_vec = cost_vec;
9620 : }
9621 1368365 : }
9622 :
9623 : /* Now look for SLP instances with a root that are covered by other
9624 : instances and remove them. */
9625 436311 : hash_set<stmt_vec_info> roots;
9626 1810824 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); ++i)
9627 998287 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
9628 60085 : roots.add (SLP_INSTANCE_ROOT_STMTS (instance)[0]);
9629 436311 : if (!roots.is_empty ())
9630 : {
9631 24420 : visited.empty ();
9632 132422 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); ++i)
9633 108002 : vect_slp_prune_covered_roots (SLP_INSTANCE_TREE (instance), roots,
9634 : visited);
9635 132422 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); )
9636 108002 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ()
9637 60085 : && !roots.contains (SLP_INSTANCE_ROOT_STMTS (instance)[0]))
9638 : {
9639 5503 : stmt_vec_info root = SLP_INSTANCE_ROOT_STMTS (instance)[0];
9640 5503 : if (dump_enabled_p ())
9641 287 : dump_printf_loc (MSG_NOTE, vect_location,
9642 : "removing SLP instance operations starting "
9643 : "from: %G", root->stmt);
9644 5503 : vect_free_slp_instance (instance);
9645 5503 : vinfo->slp_instances.ordered_remove (i);
9646 : }
9647 : else
9648 102499 : ++i;
9649 : }
9650 :
9651 872622 : return !vinfo->slp_instances.is_empty ();
9652 1132787 : }
9653 :
9654 : /* Get the SLP instance leader from INSTANCE_LEADER thereby transitively
9655 : closing the eventual chain. */
9656 :
9657 : static slp_instance
9658 806856 : get_ultimate_leader (slp_instance instance,
9659 : hash_map<slp_instance, slp_instance> &instance_leader)
9660 : {
9661 806856 : auto_vec<slp_instance *, 8> chain;
9662 806856 : slp_instance *tem;
9663 915864 : while (*(tem = instance_leader.get (instance)) != instance)
9664 : {
9665 109008 : chain.safe_push (tem);
9666 109008 : instance = *tem;
9667 : }
9668 915864 : while (!chain.is_empty ())
9669 109008 : *chain.pop () = instance;
9670 806856 : return instance;
9671 806856 : }
9672 :
9673 : namespace {
9674 : /* Subroutine of vect_bb_partition_graph_r. Map KEY to INSTANCE in
9675 : KEY_TO_INSTANCE, making INSTANCE the leader of any previous mapping
9676 : for KEY. Return true if KEY was already in KEY_TO_INSTANCE.
9677 :
9678 : INSTANCE_LEADER is as for get_ultimate_leader. */
9679 :
9680 : template<typename T>
9681 : bool
9682 3567685 : vect_map_to_instance (slp_instance instance, T key,
9683 : hash_map<T, slp_instance> &key_to_instance,
9684 : hash_map<slp_instance, slp_instance> &instance_leader)
9685 : {
9686 : bool existed_p;
9687 3567685 : slp_instance &key_instance = key_to_instance.get_or_insert (key, &existed_p);
9688 3567685 : if (!existed_p)
9689 : ;
9690 242402 : else if (key_instance != instance)
9691 : {
9692 : /* If we're running into a previously marked key make us the
9693 : leader of the current ultimate leader. This keeps the
9694 : leader chain acyclic and works even when the current instance
9695 : connects two previously independent graph parts. */
9696 89346 : slp_instance key_leader
9697 89346 : = get_ultimate_leader (key_instance, instance_leader);
9698 89346 : if (key_leader != instance)
9699 26948 : instance_leader.put (key_leader, instance);
9700 : }
9701 3567685 : key_instance = instance;
9702 3567685 : return existed_p;
9703 : }
9704 : }
9705 :
9706 : /* Worker of vect_bb_partition_graph, recurse on NODE. */
9707 :
9708 : static void
9709 997441 : vect_bb_partition_graph_r (bb_vec_info bb_vinfo,
9710 : slp_instance instance, slp_tree node,
9711 : hash_map<stmt_vec_info, slp_instance> &stmt_to_instance,
9712 : hash_map<slp_tree, slp_instance> &node_to_instance,
9713 : hash_map<slp_instance, slp_instance> &instance_leader)
9714 : {
9715 5551151 : for (auto stmt_info : SLP_TREE_SCALAR_STMTS (node))
9716 2570244 : if (stmt_info)
9717 2570244 : vect_map_to_instance (instance, stmt_info, stmt_to_instance,
9718 : instance_leader);
9719 :
9720 997441 : if (vect_map_to_instance (instance, node, node_to_instance,
9721 : instance_leader))
9722 : return;
9723 :
9724 3573647 : for (auto child : SLP_TREE_CHILDREN (node))
9725 981921 : if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
9726 279931 : vect_bb_partition_graph_r (bb_vinfo, instance, child, stmt_to_instance,
9727 : node_to_instance, instance_leader);
9728 : }
9729 :
9730 : /* Partition the SLP graph into pieces that can be costed independently. */
9731 :
9732 : static void
9733 247322 : vect_bb_partition_graph (bb_vec_info bb_vinfo)
9734 : {
9735 247322 : DUMP_VECT_SCOPE ("vect_bb_partition_graph");
9736 :
9737 : /* First walk the SLP graph assigning each involved scalar stmt a
9738 : corresponding SLP graph entry and upon visiting a previously
9739 : marked stmt, make the stmts leader the current SLP graph entry. */
9740 247322 : hash_map<stmt_vec_info, slp_instance> stmt_to_instance;
9741 247322 : hash_map<slp_tree, slp_instance> node_to_instance;
9742 247322 : hash_map<slp_instance, slp_instance> instance_leader;
9743 247322 : slp_instance instance;
9744 1212154 : for (unsigned i = 0; bb_vinfo->slp_instances.iterate (i, &instance); ++i)
9745 : {
9746 717510 : instance_leader.put (instance, instance);
9747 717510 : vect_bb_partition_graph_r (bb_vinfo,
9748 : instance, SLP_INSTANCE_TREE (instance),
9749 : stmt_to_instance, node_to_instance,
9750 : instance_leader);
9751 : }
9752 :
9753 : /* Then collect entries to each independent subgraph. */
9754 1212154 : for (unsigned i = 0; bb_vinfo->slp_instances.iterate (i, &instance); ++i)
9755 : {
9756 717510 : slp_instance leader = get_ultimate_leader (instance, instance_leader);
9757 717510 : leader->subgraph_entries.safe_push (instance);
9758 717510 : if (dump_enabled_p ()
9759 717510 : && leader != instance)
9760 99 : dump_printf_loc (MSG_NOTE, vect_location,
9761 : "instance %p is leader of %p\n",
9762 : (void *) leader, (void *) instance);
9763 : }
9764 247322 : }
9765 :
9766 : /* Compute the scalar cost of the SLP node NODE and its children
9767 : and return it. Do not account defs that are marked in LIFE and
9768 : update LIFE according to uses of NODE. */
9769 :
9770 : static void
9771 713425 : vect_bb_slp_scalar_cost (bb_vec_info vinfo,
9772 : vec<stmt_vec_info> &worklist,
9773 : stmt_vector_for_cost *cost_vec,
9774 : hash_set<stmt_vec_info> &visited)
9775 : {
9776 3323394 : while (!worklist.is_empty ())
9777 : {
9778 2609969 : stmt_vec_info stmt = worklist.pop ();
9779 2932617 : if (!PURE_SLP_STMT (stmt))
9780 331073 : continue;
9781 :
9782 : /* When the stmt is live but not actually vectorized we have
9783 : to keep the feeding scalar defs. */
9784 2292932 : if (!STMT_VINFO_LIVE_P (vect_stmt_to_vectorize (stmt)))
9785 : {
9786 2184651 : bool live_p = false;
9787 2184651 : ssa_op_iter op_iter;
9788 2184651 : def_operand_p def_p;
9789 4803697 : FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt->stmt, op_iter, SSA_OP_DEF)
9790 : {
9791 434395 : imm_use_iterator use_iter;
9792 434395 : gimple *use_stmt;
9793 1112603 : FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p))
9794 678208 : if (!is_gimple_debug (use_stmt))
9795 : {
9796 511025 : stmt_vec_info use_stmt_info = vinfo->lookup_stmt (use_stmt);
9797 511025 : if (!use_stmt_info || !PURE_SLP_STMT (use_stmt_info))
9798 : {
9799 15578 : if (dump_enabled_p ())
9800 : {
9801 74 : dump_printf_loc (MSG_NOTE, vect_location,
9802 : "stmt considered live: %G",
9803 : stmt->stmt);
9804 74 : dump_printf_loc (MSG_NOTE, vect_location,
9805 : "because of use in: %G",
9806 : use_stmt);
9807 : }
9808 : live_p = true;
9809 : }
9810 434395 : }
9811 : }
9812 2184651 : if (live_p)
9813 8425 : continue;
9814 : }
9815 :
9816 : /* The following assert verifies that vect_bb_partition_graph
9817 : partitions the SLP graph in a way that each scalar stmt of
9818 : the coverage of the SLP graph belongs to exactly one subgraph.
9819 : ??? This is currently not guaranteed since the function
9820 : works purely on SLP_TREE_SCALAR_STMTS, resulting in the assert
9821 : tripping or scalar stmts costed multiple times, making vectorization
9822 : more profitable than it really is. */
9823 : /* gcc_checking_assert (!gimple_visited_p (stmt->stmt)); */
9824 :
9825 2278896 : if (vect_nop_conversion_p (stmt))
9826 : ;
9827 : /* For single-argument PHIs assume coalescing which means zero
9828 : cost for the scalar and the vector PHIs. This avoids
9829 : artificially favoring the vector path (but may pessimize it
9830 : in some cases). */
9831 2248029 : else if (is_a <gphi *> (stmt->stmt)
9832 2248029 : && gimple_phi_num_args (as_a <gphi *> (stmt->stmt)) == 1)
9833 : ;
9834 : else
9835 : {
9836 2237515 : vect_cost_for_stmt kind;
9837 2237515 : if (STMT_VINFO_DATA_REF (stmt))
9838 : {
9839 2012889 : data_reference_p dr = STMT_VINFO_DATA_REF (stmt);
9840 2012889 : tree base = get_base_address (DR_REF (dr));
9841 : /* When the scalar access is to a non-global not
9842 : address-taken decl that is not BLKmode assume we can
9843 : access it with a single non-load/store instruction. */
9844 2012889 : if (DECL_P (base)
9845 1546018 : && !is_global_var (base)
9846 1468407 : && !TREE_ADDRESSABLE (base)
9847 2571164 : && DECL_MODE (base) != BLKmode)
9848 : kind = scalar_stmt;
9849 1868754 : else if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt)))
9850 : kind = scalar_load;
9851 : else
9852 1623062 : kind = scalar_store;
9853 : }
9854 : else
9855 : kind = scalar_stmt;
9856 : /* Cost each scalar stmt only once. */
9857 2237515 : gimple_set_visited (stmt->stmt, true);
9858 2237515 : record_stmt_cost (cost_vec, 1, kind, stmt, NULL_TREE, 0, vect_body);
9859 : }
9860 :
9861 : /* Now walk relevant parts of the SSA use-def graph. */
9862 2278896 : slp_oprnds child_ops (stmt);
9863 7093589 : for (unsigned i = 0; i < child_ops.num_slp_children; ++i)
9864 : {
9865 2535797 : tree op = child_ops.get_op_for_slp_child (stmt, i);
9866 2535797 : stmt_vec_info def = vinfo->lookup_def (op);
9867 2535797 : if (def && !visited.add (def))
9868 758903 : worklist.safe_push (def);
9869 : }
9870 : }
9871 713425 : }
9872 :
9873 :
9874 : /* Comparator for the loop-index sorted cost vectors. */
9875 :
9876 : static int
9877 20055598 : li_cost_vec_cmp (const void *a_, const void *b_, void *)
9878 : {
9879 20055598 : auto *a = (const std::pair<unsigned, stmt_info_for_cost *> *)a_;
9880 20055598 : auto *b = (const std::pair<unsigned, stmt_info_for_cost *> *)b_;
9881 20055598 : if (a->first < b->first)
9882 : return -1;
9883 18625603 : else if (a->first == b->first)
9884 17483316 : return 0;
9885 : return 1;
9886 : }
9887 :
9888 : /* Check if vectorization of the basic block is profitable for the
9889 : subgraph denoted by SLP_INSTANCES. */
9890 :
9891 : static bool
9892 686679 : vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo,
9893 : vec<slp_instance> slp_instances,
9894 : loop_p orig_loop)
9895 : {
9896 686679 : slp_instance instance;
9897 686679 : int i;
9898 686679 : unsigned int vec_inside_cost = 0, vec_outside_cost = 0, scalar_cost = 0;
9899 686679 : unsigned int vec_prologue_cost = 0, vec_epilogue_cost = 0;
9900 :
9901 686679 : if (dump_enabled_p ())
9902 : {
9903 125 : dump_printf_loc (MSG_NOTE, vect_location, "Costing subgraph:\n");
9904 259 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
9905 134 : dump_printf_loc (MSG_NOTE, vect_location, " entry instance %p -> "
9906 : "node %p\n", (void *)instance,
9907 134 : (void *)SLP_INSTANCE_TREE (instance));
9908 125 : hash_set<slp_tree> visited;
9909 509 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
9910 134 : vect_print_slp_graph (MSG_NOTE, vect_location,
9911 : SLP_INSTANCE_TREE (instance), visited);
9912 125 : }
9913 :
9914 : /* Then DFS walk scalar stmts, performing costing and handling
9915 : still live scalar stmts via the previously computed vector coverage. */
9916 686679 : stmt_vector_for_cost scalar_costs = vNULL;
9917 686679 : stmt_vector_for_cost vector_costs = vNULL;
9918 686679 : hash_set<slp_tree> visited;
9919 686679 : hash_set<stmt_vec_info> svisited;
9920 2086783 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
9921 : {
9922 713425 : auto_vec<stmt_vec_info> worklist;
9923 713425 : if (SLP_INSTANCE_ROOT_STMTS (instance).exists ())
9924 102246 : record_stmt_cost (&scalar_costs,
9925 51123 : SLP_INSTANCE_ROOT_STMTS (instance).length (),
9926 : scalar_stmt,
9927 51123 : SLP_INSTANCE_ROOT_STMTS (instance)[0], 0, vect_body);
9928 4006649 : for (auto stmt : SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (instance)))
9929 : {
9930 1866374 : stmt = vect_orig_stmt (stmt);
9931 1866374 : if (!svisited.add (stmt))
9932 1851066 : worklist.safe_push (stmt);
9933 : }
9934 713425 : vect_bb_slp_scalar_cost (bb_vinfo, worklist, &scalar_costs, svisited);
9935 713425 : vector_costs.safe_splice (instance->cost_vec);
9936 713425 : instance->cost_vec.release ();
9937 713425 : }
9938 :
9939 686679 : if (dump_enabled_p ())
9940 125 : dump_printf_loc (MSG_NOTE, vect_location, "Cost model analysis: \n");
9941 :
9942 : /* When costing non-loop vectorization we need to consider each covered
9943 : loop independently and make sure vectorization is profitable. For
9944 : now we assume a loop may be not entered or executed an arbitrary
9945 : number of iterations (??? static information can provide more
9946 : precise info here) which means we can simply cost each containing
9947 : loops stmts separately. */
9948 :
9949 : /* First produce cost vectors sorted by loop index. */
9950 686679 : auto_vec<std::pair<unsigned, stmt_info_for_cost *> >
9951 1373358 : li_scalar_costs (scalar_costs.length ());
9952 686679 : auto_vec<std::pair<unsigned, stmt_info_for_cost *> >
9953 1373345 : li_vector_costs (vector_costs.length ());
9954 686679 : stmt_info_for_cost *cost;
9955 3661996 : FOR_EACH_VEC_ELT (scalar_costs, i, cost)
9956 : {
9957 2288638 : unsigned l = gimple_bb (cost->stmt_info->stmt)->loop_father->num;
9958 2288638 : li_scalar_costs.quick_push (std::make_pair (l, cost));
9959 : }
9960 : /* Use a random used loop as fallback in case the first vector_costs
9961 : entry does not have a location associated with it. */
9962 686679 : unsigned l = li_scalar_costs[0].first;
9963 2637251 : FOR_EACH_VEC_ELT (vector_costs, i, cost)
9964 : {
9965 : /* Use SLP node placement according to the computed schedule. */
9966 1950572 : if (cost->node && cost->node->si)
9967 1214010 : l = gimple_bb (cost->node->si)->loop_father->num;
9968 : /* For schedules at region boundary use the region entry loop. */
9969 736562 : else if (cost->node)
9970 588077 : l = bb_vinfo->bbs[0]->loop_father->num;
9971 : /* SLP instance root stmts do not have an associated SLP node. */
9972 148485 : else if (cost->stmt_info)
9973 148485 : l = gimple_bb (cost->stmt_info->stmt)->loop_father->num;
9974 : /* And since vect_prologue_cost_for_slp can end up costing with
9975 : neither, inherit from the previous node. */
9976 1950572 : li_vector_costs.quick_push (std::make_pair (l, cost));
9977 : }
9978 686679 : li_scalar_costs.stablesort (li_cost_vec_cmp, NULL);
9979 686679 : li_vector_costs.stablesort (li_cost_vec_cmp, NULL);
9980 :
9981 686679 : unsigned total_vec_outside_cost = 0;
9982 686679 : unsigned total_vec_inside_cost = 0;
9983 686679 : unsigned total_scalar_cost = 0;
9984 :
9985 : /* Now cost the portions individually. */
9986 686679 : unsigned vi = 0;
9987 686679 : unsigned si = 0;
9988 686679 : bool profitable = true;
9989 1374685 : while (si < li_scalar_costs.length ()
9990 2120044 : || vi < li_vector_costs.length ())
9991 : {
9992 745372 : unsigned sl
9993 1489417 : = si < li_scalar_costs.length () ? li_scalar_costs[si].first : -1U;
9994 745372 : unsigned vl
9995 1490676 : = vi < li_vector_costs.length () ? li_vector_costs[vi].first : -1U;
9996 :
9997 745372 : class vector_costs *scalar_target_cost_data = nullptr;
9998 745372 : scalar_cost = 0;
9999 745372 : if (sl <= vl)
10000 : {
10001 702016 : if (dump_enabled_p ())
10002 145 : dump_printf_loc (MSG_NOTE, vect_location,
10003 : "Scalar cost for part in loop %d\n", sl);
10004 702016 : scalar_target_cost_data = init_cost (bb_vinfo, true);
10005 2288638 : do
10006 : {
10007 2288638 : add_stmt_cost (scalar_target_cost_data,
10008 2288638 : li_scalar_costs[si].second);
10009 2288638 : si++;
10010 : }
10011 2288638 : while (si < li_scalar_costs.length ()
10012 4592613 : && li_scalar_costs[si].first == sl);
10013 702016 : scalar_target_cost_data->finish_cost (nullptr);
10014 702016 : scalar_cost = scalar_target_cost_data->body_cost ();
10015 702016 : total_scalar_cost += scalar_cost;
10016 702016 : if (sl < vl)
10017 : {
10018 1836 : if (dump_enabled_p ())
10019 2 : dump_printf_loc (MSG_NOTE, vect_location,
10020 : "Scalar %d loop part does not "
10021 : "have corresponding vector part\n", sl);
10022 1836 : delete scalar_target_cost_data;
10023 648374 : continue;
10024 : }
10025 : }
10026 :
10027 : /* Complete the target-specific vector cost calculation. */
10028 743536 : if (dump_enabled_p ())
10029 147 : dump_printf_loc (MSG_NOTE, vect_location,
10030 : "Vector cost for part in loop %d\n", vl);
10031 743536 : class vector_costs *vect_target_cost_data = init_cost (bb_vinfo, false);
10032 743536 : auto_vec<stmt_info_for_cost> tem;
10033 1950572 : do
10034 : {
10035 1950572 : tem.safe_push (*li_vector_costs[vi].second);
10036 1950572 : vi++;
10037 : }
10038 1950572 : while (vi < li_vector_costs.length ()
10039 3958014 : && li_vector_costs[vi].first == vl);
10040 743536 : add_slp_costs (vect_target_cost_data, tem);
10041 743536 : vect_target_cost_data->finish_cost (scalar_target_cost_data);
10042 743536 : vec_prologue_cost = vect_target_cost_data->prologue_cost ();
10043 743536 : vec_inside_cost = vect_target_cost_data->body_cost ();
10044 743536 : vec_epilogue_cost = vect_target_cost_data->epilogue_cost ();
10045 743536 : if (scalar_target_cost_data)
10046 700180 : delete scalar_target_cost_data;
10047 743536 : delete vect_target_cost_data;
10048 :
10049 743536 : vec_outside_cost = vec_prologue_cost + vec_epilogue_cost;
10050 :
10051 743536 : total_vec_outside_cost += vec_outside_cost;
10052 743536 : total_vec_inside_cost += vec_inside_cost;
10053 :
10054 743536 : if (sl > vl && dump_enabled_p ())
10055 4 : dump_printf_loc (MSG_NOTE, vect_location,
10056 : "Vector %d loop part does not "
10057 : "have corresponding scalar part\n", vl);
10058 :
10059 : /* When this is vector costs for the region entry delay costing
10060 : and instead only require the total costs to be profitable. */
10061 743536 : if (vl == (unsigned) bb_vinfo->bbs[0]->loop_father->num)
10062 646538 : continue;
10063 :
10064 96998 : if (dump_enabled_p ())
10065 : {
10066 50 : dump_printf_loc (MSG_NOTE, vect_location,
10067 : "Cost model analysis for part in loop %d:\n",
10068 50 : std::min (sl, vl));
10069 50 : dump_printf (MSG_NOTE, " Vector cost: %d\n",
10070 : vec_inside_cost + vec_outside_cost);
10071 50 : dump_printf (MSG_NOTE, " Scalar cost: %d\n", scalar_cost);
10072 : }
10073 :
10074 : /* Vectorization is profitable if its cost is more than the cost of scalar
10075 : version. Note that we err on the vector side for equal cost because
10076 : the cost estimate is otherwise quite pessimistic (constant uses are
10077 : free on the scalar side but cost a load on the vector side for
10078 : example). */
10079 96998 : if (vec_outside_cost + vec_inside_cost > scalar_cost)
10080 42198 : profitable = false;
10081 743536 : }
10082 :
10083 686679 : if (dump_enabled_p ())
10084 : {
10085 125 : dump_printf_loc (MSG_NOTE, vect_location,
10086 : "Cost model analysis for whole subgraph:\n");
10087 125 : dump_printf (MSG_NOTE, " Vector cost: %d\n",
10088 : total_vec_inside_cost + total_vec_outside_cost);
10089 125 : dump_printf (MSG_NOTE, " Scalar cost: %d\n", total_scalar_cost);
10090 : }
10091 :
10092 : /* For the case where the outermost region had no scalar cost require
10093 : overall profitability. */
10094 686679 : if (total_vec_outside_cost + total_vec_inside_cost > total_scalar_cost)
10095 195157 : profitable = false;
10096 :
10097 : /* Unset visited flag. This is delayed when the subgraph is profitable
10098 : and we process the loop for remaining unvectorized if-converted code. */
10099 686679 : if (!orig_loop || !profitable)
10100 2973082 : FOR_EACH_VEC_ELT (scalar_costs, i, cost)
10101 2286563 : gimple_set_visited (cost->stmt_info->stmt, false);
10102 :
10103 686679 : scalar_costs.release ();
10104 686679 : vector_costs.release ();
10105 :
10106 686679 : return profitable;
10107 686679 : }
10108 :
10109 : /* qsort comparator for lane defs. */
10110 :
10111 : static int
10112 128 : vld_cmp (const void *a_, const void *b_)
10113 : {
10114 128 : auto *a = (const std::pair<unsigned, tree> *)a_;
10115 128 : auto *b = (const std::pair<unsigned, tree> *)b_;
10116 128 : return a->first - b->first;
10117 : }
10118 :
10119 : /* Return true if USE_STMT is a vector lane insert into VEC and set
10120 : *THIS_LANE to the lane number that is set. */
10121 :
10122 : static bool
10123 295 : vect_slp_is_lane_insert (gimple *use_stmt, tree vec, unsigned *this_lane)
10124 : {
10125 295 : gassign *use_ass = dyn_cast <gassign *> (use_stmt);
10126 110 : if (!use_ass
10127 110 : || gimple_assign_rhs_code (use_ass) != BIT_INSERT_EXPR
10128 35 : || (vec
10129 35 : ? gimple_assign_rhs1 (use_ass) != vec
10130 26 : : ((vec = gimple_assign_rhs1 (use_ass)), false))
10131 61 : || !useless_type_conversion_p (TREE_TYPE (TREE_TYPE (vec)),
10132 61 : TREE_TYPE (gimple_assign_rhs2 (use_ass)))
10133 61 : || !constant_multiple_p
10134 61 : (tree_to_poly_uint64 (gimple_assign_rhs3 (use_ass)),
10135 356 : tree_to_poly_uint64 (TYPE_SIZE (TREE_TYPE (TREE_TYPE (vec)))),
10136 : this_lane))
10137 : return false;
10138 : return true;
10139 : }
10140 :
10141 : /* Find any vectorizable constructors and add them to the grouped_store
10142 : array. */
10143 :
10144 : static void
10145 2285514 : vect_slp_check_for_roots (bb_vec_info bb_vinfo)
10146 : {
10147 18297430 : for (unsigned i = 0; i < bb_vinfo->nbbs; ++i)
10148 32023832 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb_vinfo->bbs[i]);
10149 147758515 : !gsi_end_p (gsi); gsi_next (&gsi))
10150 : {
10151 131746599 : gassign *assign = dyn_cast<gassign *> (gsi_stmt (gsi));
10152 : /* This can be used to start SLP discovery for early breaks for BB early breaks
10153 : when we get that far. */
10154 131746599 : if (!assign)
10155 101206953 : continue;
10156 :
10157 32660029 : tree rhs = gimple_assign_rhs1 (assign);
10158 32660029 : enum tree_code code = gimple_assign_rhs_code (assign);
10159 32660029 : use_operand_p use_p;
10160 32660029 : gimple *use_stmt;
10161 32660029 : if (code == CONSTRUCTOR)
10162 : {
10163 1738289 : if (!VECTOR_TYPE_P (TREE_TYPE (rhs))
10164 1784247 : || maybe_ne (TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)),
10165 100453 : CONSTRUCTOR_NELTS (rhs))
10166 45958 : || VECTOR_TYPE_P (TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value))
10167 1784227 : || uniform_vector_p (rhs))
10168 1723813 : continue;
10169 :
10170 : unsigned j;
10171 : tree val;
10172 71600 : FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (rhs), j, val)
10173 57124 : if (TREE_CODE (val) != SSA_NAME
10174 57124 : || !bb_vinfo->lookup_def (val))
10175 : break;
10176 35884 : if (j != CONSTRUCTOR_NELTS (rhs))
10177 3466 : continue;
10178 :
10179 14476 : vec<stmt_vec_info> roots = vNULL;
10180 14476 : roots.safe_push (bb_vinfo->lookup_stmt (assign));
10181 14476 : vec<stmt_vec_info> stmts;
10182 14476 : stmts.create (CONSTRUCTOR_NELTS (rhs));
10183 80678 : FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (rhs), j, val)
10184 51726 : stmts.quick_push
10185 51726 : (vect_stmt_to_vectorize (bb_vinfo->lookup_def (val)));
10186 14476 : bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_ctor,
10187 14476 : stmts, roots));
10188 : }
10189 30921740 : else if (code == BIT_INSERT_EXPR
10190 1037 : && VECTOR_TYPE_P (TREE_TYPE (rhs))
10191 711 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)).is_constant ()
10192 711 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)).to_constant () > 1
10193 708 : && integer_zerop (gimple_assign_rhs3 (assign))
10194 392 : && useless_type_conversion_p
10195 392 : (TREE_TYPE (TREE_TYPE (rhs)),
10196 392 : TREE_TYPE (gimple_assign_rhs2 (assign)))
10197 30922464 : && bb_vinfo->lookup_def (gimple_assign_rhs2 (assign)))
10198 : {
10199 : /* We start to match on insert to lane zero but since the
10200 : inserts need not be ordered we'd have to search both
10201 : the def and the use chains. */
10202 257 : tree vectype = TREE_TYPE (rhs);
10203 257 : unsigned nlanes = TYPE_VECTOR_SUBPARTS (vectype).to_constant ();
10204 257 : auto_vec<std::pair<unsigned, tree> > lane_defs (nlanes);
10205 257 : auto_sbitmap lanes (nlanes);
10206 257 : bitmap_clear (lanes);
10207 257 : bitmap_set_bit (lanes, 0);
10208 257 : tree def = gimple_assign_lhs (assign);
10209 257 : lane_defs.quick_push
10210 257 : (std::make_pair (0, gimple_assign_rhs2 (assign)));
10211 257 : unsigned lanes_found = 1;
10212 : /* Start with the use chains, the last stmt will be the root. */
10213 257 : stmt_vec_info last = bb_vinfo->lookup_stmt (assign);
10214 257 : vec<stmt_vec_info> roots = vNULL;
10215 257 : roots.safe_push (last);
10216 268 : do
10217 : {
10218 268 : use_operand_p use_p;
10219 268 : gimple *use_stmt;
10220 268 : if (!single_imm_use (def, &use_p, &use_stmt))
10221 : break;
10222 256 : unsigned this_lane;
10223 256 : if (!bb_vinfo->lookup_stmt (use_stmt)
10224 256 : || !vect_slp_is_lane_insert (use_stmt, def, &this_lane)
10225 291 : || !bb_vinfo->lookup_def (gimple_assign_rhs2 (use_stmt)))
10226 : break;
10227 35 : if (bitmap_bit_p (lanes, this_lane))
10228 : break;
10229 15 : lanes_found++;
10230 15 : bitmap_set_bit (lanes, this_lane);
10231 15 : gassign *use_ass = as_a <gassign *> (use_stmt);
10232 15 : lane_defs.quick_push (std::make_pair
10233 15 : (this_lane, gimple_assign_rhs2 (use_ass)));
10234 15 : last = bb_vinfo->lookup_stmt (use_ass);
10235 15 : roots.safe_push (last);
10236 15 : def = gimple_assign_lhs (use_ass);
10237 : }
10238 15 : while (lanes_found < nlanes);
10239 257 : if (roots.length () > 1)
10240 7 : std::swap(roots[0], roots[roots.length () - 1]);
10241 257 : if (lanes_found < nlanes)
10242 : {
10243 : /* Now search the def chain. */
10244 253 : def = gimple_assign_rhs1 (assign);
10245 255 : do
10246 : {
10247 255 : if (TREE_CODE (def) != SSA_NAME
10248 255 : || !has_single_use (def))
10249 : break;
10250 59 : gimple *def_stmt = SSA_NAME_DEF_STMT (def);
10251 59 : unsigned this_lane;
10252 59 : if (!bb_vinfo->lookup_stmt (def_stmt)
10253 39 : || !vect_slp_is_lane_insert (def_stmt,
10254 : NULL_TREE, &this_lane)
10255 85 : || !bb_vinfo->lookup_def (gimple_assign_rhs2 (def_stmt)))
10256 : break;
10257 26 : if (bitmap_bit_p (lanes, this_lane))
10258 : break;
10259 6 : lanes_found++;
10260 6 : bitmap_set_bit (lanes, this_lane);
10261 12 : lane_defs.quick_push (std::make_pair
10262 6 : (this_lane,
10263 6 : gimple_assign_rhs2 (def_stmt)));
10264 6 : roots.safe_push (bb_vinfo->lookup_stmt (def_stmt));
10265 6 : def = gimple_assign_rhs1 (def_stmt);
10266 : }
10267 6 : while (lanes_found < nlanes);
10268 : }
10269 257 : if (lanes_found == nlanes)
10270 : {
10271 : /* Sort lane_defs after the lane index and register the root. */
10272 8 : lane_defs.qsort (vld_cmp);
10273 8 : vec<stmt_vec_info> stmts;
10274 8 : stmts.create (nlanes);
10275 44 : for (unsigned i = 0; i < nlanes; ++i)
10276 28 : stmts.quick_push (bb_vinfo->lookup_def (lane_defs[i].second));
10277 8 : bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_ctor,
10278 8 : stmts, roots));
10279 : }
10280 : else
10281 249 : roots.release ();
10282 257 : }
10283 30921483 : else if (!VECTOR_TYPE_P (TREE_TYPE (rhs))
10284 29884365 : && (associative_tree_code (code) || code == MINUS_EXPR)
10285 : /* Ops with constants at the tail can be stripped here. */
10286 6573867 : && TREE_CODE (rhs) == SSA_NAME
10287 6511987 : && TREE_CODE (gimple_assign_rhs2 (assign)) == SSA_NAME
10288 : /* Should be the chain end. */
10289 33720303 : && (!single_imm_use (gimple_assign_lhs (assign),
10290 : &use_p, &use_stmt)
10291 2253042 : || !is_gimple_assign (use_stmt)
10292 1627925 : || (gimple_assign_rhs_code (use_stmt) != code
10293 1084681 : && ((code != PLUS_EXPR && code != MINUS_EXPR)
10294 619996 : || (gimple_assign_rhs_code (use_stmt)
10295 619996 : != (code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR))))))
10296 : {
10297 : /* We start the match at the end of a possible association
10298 : chain. */
10299 2140204 : auto_vec<chain_op_t> chain;
10300 2140204 : auto_vec<std::pair<tree_code, gimple *> > worklist;
10301 2140204 : auto_vec<gimple *> chain_stmts;
10302 2140204 : gimple *code_stmt = NULL, *alt_code_stmt = NULL;
10303 2140204 : if (code == MINUS_EXPR)
10304 360559 : code = PLUS_EXPR;
10305 2140204 : internal_fn reduc_fn;
10306 2536774 : if (!reduction_fn_for_scalar_code (code, &reduc_fn)
10307 2140204 : || reduc_fn == IFN_LAST)
10308 396570 : continue;
10309 1743634 : vect_slp_linearize_chain (bb_vinfo, worklist, chain, code, assign,
10310 : /* ??? */
10311 : code_stmt, alt_code_stmt, &chain_stmts,
10312 : false);
10313 3487268 : if (chain.length () > 1)
10314 : {
10315 : /* Sort the chain according to def_type and operation. */
10316 1743634 : chain.sort (dt_sort_cmp, bb_vinfo);
10317 : /* ??? Now we'd want to strip externals and constants
10318 : but record those to be handled in the epilogue. */
10319 : /* ??? For now do not allow mixing ops or externs/constants. */
10320 1743634 : bool invalid = false;
10321 1743634 : unsigned remain_cnt = 0;
10322 7179900 : for (unsigned i = 0; i < chain.length (); ++i)
10323 : {
10324 4053191 : if (chain[i].code != code)
10325 : {
10326 : invalid = true;
10327 : break;
10328 : }
10329 3692632 : if (chain[i].dt != vect_internal_def
10330 : /* Avoid stmts where the def is not the LHS, like
10331 : ASMs. */
10332 7033895 : || (gimple_get_lhs (bb_vinfo->lookup_def
10333 3341263 : (chain[i].op)->stmt)
10334 3341263 : != chain[i].op))
10335 354465 : remain_cnt++;
10336 : }
10337 1743634 : if (!invalid && chain.length () - remain_cnt > 1)
10338 : {
10339 1298529 : vec<stmt_vec_info> stmts;
10340 1298529 : vec<tree> remain = vNULL;
10341 1298529 : stmts.create (chain.length ());
10342 1298529 : if (remain_cnt > 0)
10343 47561 : remain.create (remain_cnt);
10344 4325325 : for (unsigned i = 0; i < chain.length (); ++i)
10345 : {
10346 3026796 : stmt_vec_info stmt_info;
10347 3026796 : if (chain[i].dt == vect_internal_def
10348 3026796 : && ((stmt_info = bb_vinfo->lookup_def (chain[i].op)),
10349 2938411 : gimple_get_lhs (stmt_info->stmt) == chain[i].op))
10350 2938195 : stmts.quick_push (stmt_info);
10351 : else
10352 88601 : remain.quick_push (chain[i].op);
10353 : }
10354 1298529 : vec<stmt_vec_info> roots;
10355 1298529 : roots.create (chain_stmts.length ());
10356 4325325 : for (unsigned i = 0; i < chain_stmts.length (); ++i)
10357 1728267 : roots.quick_push (bb_vinfo->lookup_stmt (chain_stmts[i]));
10358 1298529 : bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_bb_reduc,
10359 1298529 : stmts, roots, remain));
10360 : }
10361 : }
10362 2140204 : }
10363 : }
10364 2285514 : }
10365 :
10366 : /* Walk the grouped store chains and replace entries with their
10367 : pattern variant if any. */
10368 :
10369 : static void
10370 661181 : vect_fixup_store_groups_with_patterns (vec_info *vinfo)
10371 : {
10372 661181 : stmt_vec_info first_element;
10373 661181 : unsigned i;
10374 :
10375 1571388 : FOR_EACH_VEC_ELT (vinfo->grouped_stores, i, first_element)
10376 : {
10377 : /* We also have CTORs in this array. */
10378 910207 : if (!STMT_VINFO_GROUPED_ACCESS (first_element))
10379 0 : continue;
10380 910207 : if (STMT_VINFO_IN_PATTERN_P (first_element))
10381 : {
10382 252 : stmt_vec_info orig = first_element;
10383 252 : first_element = STMT_VINFO_RELATED_STMT (first_element);
10384 252 : DR_GROUP_FIRST_ELEMENT (first_element) = first_element;
10385 252 : DR_GROUP_SIZE (first_element) = DR_GROUP_SIZE (orig);
10386 252 : DR_GROUP_GAP (first_element) = DR_GROUP_GAP (orig);
10387 252 : DR_GROUP_NEXT_ELEMENT (first_element) = DR_GROUP_NEXT_ELEMENT (orig);
10388 252 : vinfo->grouped_stores[i] = first_element;
10389 : }
10390 910207 : stmt_vec_info prev = first_element;
10391 2559177 : while (DR_GROUP_NEXT_ELEMENT (prev))
10392 : {
10393 1648970 : stmt_vec_info elt = DR_GROUP_NEXT_ELEMENT (prev);
10394 1648970 : if (STMT_VINFO_IN_PATTERN_P (elt))
10395 : {
10396 849 : stmt_vec_info orig = elt;
10397 849 : elt = STMT_VINFO_RELATED_STMT (elt);
10398 849 : DR_GROUP_NEXT_ELEMENT (prev) = elt;
10399 849 : DR_GROUP_GAP (elt) = DR_GROUP_GAP (orig);
10400 849 : DR_GROUP_NEXT_ELEMENT (elt) = DR_GROUP_NEXT_ELEMENT (orig);
10401 : }
10402 1648970 : DR_GROUP_FIRST_ELEMENT (elt) = first_element;
10403 1648970 : prev = elt;
10404 : }
10405 : }
10406 661181 : }
10407 :
10408 : /* Check if the region described by BB_VINFO can be vectorized, returning
10409 : true if so. When returning false, set FATAL to true if the same failure
10410 : would prevent vectorization at other vector sizes, false if it is still
10411 : worth trying other sizes. N_STMTS is the number of statements in the
10412 : region. */
10413 :
10414 : static bool
10415 2285514 : vect_slp_analyze_bb_1 (bb_vec_info bb_vinfo, int n_stmts, bool &fatal,
10416 : vec<int> *dataref_groups)
10417 : {
10418 2285514 : DUMP_VECT_SCOPE ("vect_slp_analyze_bb");
10419 :
10420 2285514 : slp_instance instance;
10421 2285514 : int i;
10422 :
10423 : /* The first group of checks is independent of the vector size. */
10424 2285514 : fatal = true;
10425 :
10426 : /* Analyze the data references. */
10427 :
10428 2285514 : if (!vect_analyze_data_refs (bb_vinfo, NULL))
10429 : {
10430 0 : if (dump_enabled_p ())
10431 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10432 : "not vectorized: unhandled data-ref in basic "
10433 : "block.\n");
10434 : return false;
10435 : }
10436 :
10437 2285514 : if (!vect_analyze_data_ref_accesses (bb_vinfo, dataref_groups))
10438 : {
10439 0 : if (dump_enabled_p ())
10440 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10441 : "not vectorized: unhandled data access in "
10442 : "basic block.\n");
10443 : return false;
10444 : }
10445 :
10446 2285514 : vect_slp_check_for_roots (bb_vinfo);
10447 :
10448 : /* If there are no grouped stores and no constructors in the region
10449 : there is no need to continue with pattern recog as vect_analyze_slp
10450 : will fail anyway. */
10451 2285514 : if (bb_vinfo->grouped_stores.is_empty ()
10452 1935187 : && bb_vinfo->roots.is_empty ())
10453 : {
10454 1624333 : if (dump_enabled_p ())
10455 1020 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10456 : "not vectorized: no grouped stores in "
10457 : "basic block.\n");
10458 : return false;
10459 : }
10460 :
10461 : /* While the rest of the analysis below depends on it in some way. */
10462 661181 : fatal = false;
10463 :
10464 661181 : vect_pattern_recog (bb_vinfo);
10465 :
10466 : /* Update store groups from pattern processing. */
10467 661181 : vect_fixup_store_groups_with_patterns (bb_vinfo);
10468 :
10469 : /* Check the SLP opportunities in the basic block, analyze and build SLP
10470 : trees. */
10471 661181 : if (!vect_analyze_slp (bb_vinfo, n_stmts, false))
10472 : {
10473 0 : if (dump_enabled_p ())
10474 : {
10475 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10476 : "Failed to SLP the basic block.\n");
10477 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10478 : "not vectorized: failed to find SLP opportunities "
10479 : "in basic block.\n");
10480 : }
10481 : return false;
10482 : }
10483 :
10484 : /* Optimize permutations. */
10485 661181 : vect_optimize_slp (bb_vinfo);
10486 :
10487 : /* Gather the loads reachable from the SLP graph entries. */
10488 661181 : vect_gather_slp_loads (bb_vinfo);
10489 :
10490 661181 : vect_record_base_alignments (bb_vinfo);
10491 :
10492 : /* Analyze and verify the alignment of data references and the
10493 : dependence in the SLP instances. */
10494 2152951 : for (i = 0; BB_VINFO_SLP_INSTANCES (bb_vinfo).iterate (i, &instance); )
10495 : {
10496 830589 : vect_location = instance->location ();
10497 830589 : if (! vect_slp_analyze_instance_alignment (bb_vinfo, instance)
10498 830589 : || ! vect_slp_analyze_instance_dependence (bb_vinfo, instance))
10499 : {
10500 8709 : slp_tree node = SLP_INSTANCE_TREE (instance);
10501 8709 : stmt_vec_info stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
10502 8709 : if (dump_enabled_p ())
10503 4 : dump_printf_loc (MSG_NOTE, vect_location,
10504 : "removing SLP instance operations starting from: %G",
10505 : stmt_info->stmt);
10506 8709 : vect_free_slp_instance (instance);
10507 8709 : BB_VINFO_SLP_INSTANCES (bb_vinfo).ordered_remove (i);
10508 8709 : continue;
10509 8709 : }
10510 :
10511 : /* Mark all the statements that we want to vectorize as relevant. */
10512 821880 : vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance));
10513 :
10514 821880 : i++;
10515 : }
10516 2564985 : if (! BB_VINFO_SLP_INSTANCES (bb_vinfo).length ())
10517 : return false;
10518 :
10519 279471 : if (!vect_slp_analyze_operations (bb_vinfo))
10520 : {
10521 32149 : if (dump_enabled_p ())
10522 73 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10523 : "not vectorized: bad operation in basic block.\n");
10524 : return false;
10525 : }
10526 :
10527 : /* Compute vector stmt placement. */
10528 247322 : if (!vect_schedule_slp (bb_vinfo, BB_VINFO_SLP_INSTANCES (bb_vinfo), true))
10529 : {
10530 0 : if (dump_enabled_p ())
10531 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10532 : "not vectorized: cannot schedule SLP graph\n");
10533 : return false;
10534 : }
10535 :
10536 : /* Mark all the statements that we vectorize. */
10537 247322 : vect_bb_slp_mark_stmts_vectorized (bb_vinfo);
10538 :
10539 : /* Compute vectorizable live stmts. */
10540 247322 : vect_bb_slp_mark_live_stmts (bb_vinfo);
10541 :
10542 247322 : vect_bb_partition_graph (bb_vinfo);
10543 :
10544 247322 : return true;
10545 : }
10546 :
10547 : /* Subroutine of vect_slp_bb. Try to vectorize the statements for all
10548 : basic blocks in BBS, returning true on success.
10549 : The region has N_STMTS statements and has the datarefs given by DATAREFS. */
10550 :
10551 : static bool
10552 1934395 : vect_slp_region (vec<basic_block> bbs, vec<data_reference_p> datarefs,
10553 : vec<int> *dataref_groups, unsigned int n_stmts,
10554 : loop_p orig_loop)
10555 : {
10556 1934395 : bb_vec_info bb_vinfo;
10557 1934395 : auto_vector_modes vector_modes;
10558 :
10559 : /* Autodetect first vector size we try. */
10560 1934395 : machine_mode next_vector_mode = VOIDmode;
10561 1934395 : targetm.vectorize.autovectorize_vector_modes (&vector_modes, false);
10562 1934395 : unsigned int mode_i = 0;
10563 :
10564 1934395 : vec_info_shared shared;
10565 :
10566 1934395 : machine_mode autodetected_vector_mode = VOIDmode;
10567 2636633 : while (1)
10568 : {
10569 2285514 : bool vectorized = false;
10570 2285514 : bool fatal = false;
10571 2285514 : bb_vinfo = new _bb_vec_info (bbs, &shared);
10572 :
10573 2285514 : bool first_time_p = shared.datarefs.is_empty ();
10574 2285514 : BB_VINFO_DATAREFS (bb_vinfo) = datarefs;
10575 2285514 : if (first_time_p)
10576 1958803 : bb_vinfo->shared->save_datarefs ();
10577 : else
10578 326711 : bb_vinfo->shared->check_datarefs ();
10579 2285514 : bb_vinfo->vector_mode = next_vector_mode;
10580 :
10581 2285514 : if (vect_slp_analyze_bb_1 (bb_vinfo, n_stmts, fatal, dataref_groups))
10582 : {
10583 247322 : if (dump_enabled_p ())
10584 : {
10585 1664 : dump_printf_loc (MSG_NOTE, vect_location,
10586 : "***** Analysis succeeded with vector mode"
10587 832 : " %s\n", GET_MODE_NAME (bb_vinfo->vector_mode));
10588 832 : dump_printf_loc (MSG_NOTE, vect_location, "SLPing BB part\n");
10589 : }
10590 :
10591 247322 : bb_vinfo->shared->check_datarefs ();
10592 :
10593 247322 : bool force_clear = false;
10594 247322 : auto_vec<slp_instance> profitable_subgraphs;
10595 1459476 : for (slp_instance instance : BB_VINFO_SLP_INSTANCES (bb_vinfo))
10596 : {
10597 717510 : if (instance->subgraph_entries.is_empty ())
10598 222675 : continue;
10599 :
10600 690562 : dump_user_location_t saved_vect_location = vect_location;
10601 690562 : vect_location = instance->location ();
10602 690562 : if (!unlimited_cost_model (NULL)
10603 686688 : && !param_vect_allow_possibly_not_worthwhile_vectorizations
10604 1377241 : && !vect_bb_vectorization_profitable_p
10605 686679 : (bb_vinfo, instance->subgraph_entries, orig_loop))
10606 : {
10607 195727 : if (dump_enabled_p ())
10608 51 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10609 : "not vectorized: vectorization is not "
10610 : "profitable.\n");
10611 195727 : vect_location = saved_vect_location;
10612 195727 : continue;
10613 : }
10614 :
10615 494835 : vect_location = saved_vect_location;
10616 494835 : if (!dbg_cnt (vect_slp))
10617 : {
10618 0 : force_clear = true;
10619 0 : continue;
10620 : }
10621 :
10622 494835 : profitable_subgraphs.safe_push (instance);
10623 : }
10624 :
10625 : /* When we're vectorizing an if-converted loop body make sure
10626 : we vectorized all if-converted code. */
10627 411863 : if ((!profitable_subgraphs.is_empty () || force_clear) && orig_loop)
10628 : {
10629 159 : gcc_assert (bb_vinfo->nbbs == 1);
10630 318 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb_vinfo->bbs[0]);
10631 6235 : !gsi_end_p (gsi); gsi_next (&gsi))
10632 : {
10633 : /* The costing above left us with DCEable vectorized scalar
10634 : stmts having the visited flag set on profitable
10635 : subgraphs. Do the delayed clearing of the flag here. */
10636 6076 : if (gimple_visited_p (gsi_stmt (gsi)))
10637 : {
10638 1936 : gimple_set_visited (gsi_stmt (gsi), false);
10639 1936 : continue;
10640 : }
10641 4140 : if (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED)
10642 813 : continue;
10643 :
10644 9115 : if (gassign *ass = dyn_cast <gassign *> (gsi_stmt (gsi)))
10645 3773 : if (gimple_assign_rhs_code (ass) == COND_EXPR)
10646 : {
10647 175 : if (!profitable_subgraphs.is_empty ()
10648 74 : && dump_enabled_p ())
10649 0 : dump_printf_loc (MSG_NOTE, vect_location,
10650 : "not profitable because of "
10651 : "unprofitable if-converted scalar "
10652 : "code\n");
10653 101 : profitable_subgraphs.truncate (0);
10654 : }
10655 : }
10656 : }
10657 :
10658 : /* Finally schedule the profitable subgraphs. */
10659 1071146 : for (slp_instance instance : profitable_subgraphs)
10660 : {
10661 494742 : if (!vectorized && dump_enabled_p ())
10662 793 : dump_printf_loc (MSG_NOTE, vect_location,
10663 : "Basic block will be vectorized "
10664 : "using SLP\n");
10665 494742 : vectorized = true;
10666 :
10667 : /* Dump before scheduling as store vectorization will remove
10668 : the original stores and mess with the instance tree
10669 : so querying its location will eventually ICE. */
10670 494742 : if (flag_checking)
10671 1991206 : for (slp_instance sub : instance->subgraph_entries)
10672 506980 : gcc_assert (SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (sub)));
10673 494742 : unsigned HOST_WIDE_INT bytes;
10674 494742 : if (dump_enabled_p ())
10675 3853 : for (slp_instance sub : instance->subgraph_entries)
10676 : {
10677 1036 : tree vtype = SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (sub));
10678 2072 : if (GET_MODE_SIZE (TYPE_MODE (vtype)).is_constant (&bytes))
10679 1036 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS,
10680 1036 : sub->location (),
10681 : "basic block part vectorized using %wu "
10682 : "byte vectors\n", bytes);
10683 : else
10684 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS,
10685 : sub->location (),
10686 : "basic block part vectorized using "
10687 : "variable length vectors\n");
10688 : }
10689 :
10690 494742 : dump_user_location_t saved_vect_location = vect_location;
10691 494742 : vect_location = instance->location ();
10692 :
10693 494742 : vect_schedule_slp (bb_vinfo, instance->subgraph_entries, false);
10694 :
10695 494742 : vect_location = saved_vect_location;
10696 : }
10697 :
10698 : /* Generate the invariant statements. */
10699 247322 : if (!gimple_seq_empty_p (bb_vinfo->inv_pattern_def_seq))
10700 : {
10701 27 : if (dump_enabled_p ())
10702 0 : dump_printf_loc (MSG_NOTE, vect_location,
10703 : "------>generating invariant statements\n");
10704 :
10705 27 : bb_vinfo->insert_seq_on_entry (NULL,
10706 : bb_vinfo->inv_pattern_def_seq);
10707 : }
10708 247322 : }
10709 : else
10710 : {
10711 2038192 : if (dump_enabled_p ())
10712 1327 : dump_printf_loc (MSG_NOTE, vect_location,
10713 : "***** Analysis failed with vector mode %s\n",
10714 1327 : GET_MODE_NAME (bb_vinfo->vector_mode));
10715 : }
10716 :
10717 2285514 : if (mode_i == 0)
10718 1934395 : autodetected_vector_mode = bb_vinfo->vector_mode;
10719 :
10720 2285514 : if (!fatal)
10721 3302517 : while (mode_i < vector_modes.length ()
10722 1888361 : && vect_chooses_same_modes_p (bb_vinfo, vector_modes[mode_i]))
10723 : {
10724 355822 : if (dump_enabled_p ())
10725 1804 : dump_printf_loc (MSG_NOTE, vect_location,
10726 : "***** The result for vector mode %s would"
10727 : " be the same\n",
10728 902 : GET_MODE_NAME (vector_modes[mode_i]));
10729 355822 : mode_i += 1;
10730 : }
10731 :
10732 2285514 : delete bb_vinfo;
10733 :
10734 2285514 : if (mode_i < vector_modes.length ()
10735 2092682 : && VECTOR_MODE_P (autodetected_vector_mode)
10736 2082764 : && (related_vector_mode (vector_modes[mode_i],
10737 : GET_MODE_INNER (autodetected_vector_mode))
10738 1041382 : == autodetected_vector_mode)
10739 4378196 : && (related_vector_mode (autodetected_vector_mode,
10740 525699 : GET_MODE_INNER (vector_modes[mode_i]))
10741 1051398 : == vector_modes[mode_i]))
10742 : {
10743 525699 : if (dump_enabled_p ())
10744 201 : dump_printf_loc (MSG_NOTE, vect_location,
10745 : "***** Skipping vector mode %s, which would"
10746 : " repeat the analysis for %s\n",
10747 201 : GET_MODE_NAME (vector_modes[mode_i]),
10748 201 : GET_MODE_NAME (autodetected_vector_mode));
10749 525699 : mode_i += 1;
10750 : }
10751 :
10752 2285514 : if (vectorized
10753 2121047 : || mode_i == vector_modes.length ()
10754 1928265 : || autodetected_vector_mode == VOIDmode
10755 : /* If vect_slp_analyze_bb_1 signaled that analysis for all
10756 : vector sizes will fail do not bother iterating. */
10757 3162479 : || fatal)
10758 3868790 : return vectorized;
10759 :
10760 : /* Try the next biggest vector size. */
10761 351119 : next_vector_mode = vector_modes[mode_i++];
10762 351119 : if (dump_enabled_p ())
10763 239 : dump_printf_loc (MSG_NOTE, vect_location,
10764 : "***** Re-trying analysis with vector mode %s\n",
10765 239 : GET_MODE_NAME (next_vector_mode));
10766 351119 : }
10767 1934395 : }
10768 :
10769 :
10770 : /* Main entry for the BB vectorizer. Analyze and transform BBS, returns
10771 : true if anything in the basic-block was vectorized. */
10772 :
10773 : static bool
10774 1934395 : vect_slp_bbs (const vec<basic_block> &bbs, loop_p orig_loop)
10775 : {
10776 1934395 : vec<data_reference_p> datarefs = vNULL;
10777 1934395 : auto_vec<int> dataref_groups;
10778 1934395 : int insns = 0;
10779 1934395 : int current_group = 0;
10780 :
10781 12827877 : for (unsigned i = 0; i < bbs.length (); i++)
10782 : {
10783 10893482 : basic_block bb = bbs[i];
10784 95539218 : for (gimple_stmt_iterator gsi = gsi_after_labels (bb); !gsi_end_p (gsi);
10785 84645736 : gsi_next (&gsi))
10786 : {
10787 84645736 : gimple *stmt = gsi_stmt (gsi);
10788 84645736 : if (is_gimple_debug (stmt))
10789 53907894 : continue;
10790 :
10791 30737842 : insns++;
10792 :
10793 30737842 : if (gimple_location (stmt) != UNKNOWN_LOCATION)
10794 27574763 : vect_location = stmt;
10795 :
10796 30737842 : if (!vect_find_stmt_data_reference (NULL, stmt, &datarefs,
10797 : &dataref_groups, current_group))
10798 5242328 : ++current_group;
10799 : }
10800 : /* New BBs always start a new DR group. */
10801 10893482 : ++current_group;
10802 : }
10803 :
10804 1934395 : return vect_slp_region (bbs, datarefs, &dataref_groups, insns, orig_loop);
10805 1934395 : }
10806 :
10807 : /* Special entry for the BB vectorizer. Analyze and transform a single
10808 : if-converted BB with ORIG_LOOPs body being the not if-converted
10809 : representation. Returns true if anything in the basic-block was
10810 : vectorized. */
10811 :
10812 : bool
10813 16800 : vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop)
10814 : {
10815 16800 : auto_vec<basic_block> bbs;
10816 16800 : bbs.safe_push (bb);
10817 16800 : return vect_slp_bbs (bbs, orig_loop);
10818 16800 : }
10819 :
10820 : /* Main entry for the BB vectorizer. Analyze and transform BB, returns
10821 : true if anything in the basic-block was vectorized. */
10822 :
10823 : bool
10824 927011 : vect_slp_function (function *fun)
10825 : {
10826 927011 : bool r = false;
10827 927011 : int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (fun));
10828 927011 : auto_bitmap exit_bbs;
10829 927011 : bitmap_set_bit (exit_bbs, EXIT_BLOCK);
10830 927011 : edge entry = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (fun));
10831 927011 : unsigned n = rev_post_order_and_mark_dfs_back_seme (fun, entry, exit_bbs,
10832 927011 : true, rpo, NULL);
10833 :
10834 : /* For the moment split the function into pieces to avoid making
10835 : the iteration on the vector mode moot. Split at points we know
10836 : to not handle well which is CFG merges (SLP discovery doesn't
10837 : handle non-loop-header PHIs) and loop exits. Since pattern
10838 : recog requires reverse iteration to visit uses before defs
10839 : simply chop RPO into pieces. */
10840 927011 : auto_vec<basic_block> bbs;
10841 11830266 : for (unsigned i = 0; i < n; i++)
10842 : {
10843 10903255 : basic_block bb = BASIC_BLOCK_FOR_FN (fun, rpo[i]);
10844 10903255 : bool split = false;
10845 :
10846 : /* Split when a BB is not dominated by the first block. */
10847 20569038 : if (!bbs.is_empty ()
10848 9665783 : && !dominated_by_p (CDI_DOMINATORS, bb, bbs[0]))
10849 : {
10850 696625 : if (dump_enabled_p ())
10851 146 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10852 : "splitting region at dominance boundary bb%d\n",
10853 : bb->index);
10854 : split = true;
10855 : }
10856 : /* Split when the loop determined by the first block
10857 : is exited. This is because we eventually insert
10858 : invariants at region begin. */
10859 19175788 : else if (!bbs.is_empty ()
10860 8969158 : && bbs[0]->loop_father != bb->loop_father
10861 2322007 : && !flow_loop_nested_p (bbs[0]->loop_father, bb->loop_father))
10862 : {
10863 3859 : if (dump_enabled_p ())
10864 3 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10865 : "splitting region at loop %d exit at bb%d\n",
10866 3 : bbs[0]->loop_father->num, bb->index);
10867 : split = true;
10868 : }
10869 10202771 : else if (!bbs.is_empty ()
10870 8965299 : && bb->loop_father->header == bb
10871 476740 : && bb->loop_father->dont_vectorize)
10872 : {
10873 6212 : if (dump_enabled_p ())
10874 72 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10875 : "splitting region at dont-vectorize loop %d "
10876 : "entry at bb%d\n",
10877 : bb->loop_father->num, bb->index);
10878 : split = true;
10879 : }
10880 :
10881 221 : if (split && !bbs.is_empty ())
10882 : {
10883 706696 : r |= vect_slp_bbs (bbs, NULL);
10884 706696 : bbs.truncate (0);
10885 : }
10886 :
10887 10903255 : if (bbs.is_empty ())
10888 : {
10889 : /* We need to be able to insert at the head of the region which
10890 : we cannot for region starting with a returns-twice call. */
10891 1944168 : if (gcall *first = safe_dyn_cast <gcall *> (first_stmt (bb)))
10892 409668 : if (gimple_call_flags (first) & ECF_RETURNS_TWICE)
10893 : {
10894 306 : if (dump_enabled_p ())
10895 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10896 : "skipping bb%d as start of region as it "
10897 : "starts with returns-twice call\n",
10898 : bb->index);
10899 26573 : continue;
10900 : }
10901 : /* If the loop this BB belongs to is marked as not to be vectorized
10902 : honor that also for BB vectorization. */
10903 1943862 : if (bb->loop_father->dont_vectorize)
10904 26267 : continue;
10905 : }
10906 :
10907 10876682 : bbs.safe_push (bb);
10908 :
10909 : /* When we have a stmt ending this block and defining a
10910 : value we have to insert on edges when inserting after it for
10911 : a vector containing its definition. Avoid this for now. */
10912 21753364 : if (gimple *last = *gsi_last_bb (bb))
10913 8852161 : if (gimple_get_lhs (last)
10914 8852161 : && is_ctrl_altering_stmt (last))
10915 : {
10916 283895 : if (dump_enabled_p ())
10917 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10918 : "splitting region at control altering "
10919 : "definition %G", last);
10920 283895 : r |= vect_slp_bbs (bbs, NULL);
10921 283895 : bbs.truncate (0);
10922 : }
10923 : }
10924 :
10925 927011 : if (!bbs.is_empty ())
10926 927004 : r |= vect_slp_bbs (bbs, NULL);
10927 :
10928 927011 : free (rpo);
10929 :
10930 927011 : return r;
10931 927011 : }
10932 :
10933 : /* Build a variable-length vector in which the elements in ELTS are repeated
10934 : to a fill NRESULTS vectors of type VECTOR_TYPE. Store the vectors in
10935 : RESULTS and add any new instructions to SEQ.
10936 :
10937 : The approach we use is:
10938 :
10939 : (1) Find a vector mode VM with integer elements of mode IM.
10940 :
10941 : (2) Replace ELTS[0:NELTS] with ELTS'[0:NELTS'], where each element of
10942 : ELTS' has mode IM. This involves creating NELTS' VIEW_CONVERT_EXPRs
10943 : from small vectors to IM.
10944 :
10945 : (3) Duplicate each ELTS'[I] into a vector of mode VM.
10946 :
10947 : (4) Use a tree of interleaving VEC_PERM_EXPRs to create VMs with the
10948 : correct byte contents.
10949 :
10950 : (5) Use VIEW_CONVERT_EXPR to cast the final VMs to the required type.
10951 :
10952 : We try to find the largest IM for which this sequence works, in order
10953 : to cut down on the number of interleaves. */
10954 :
10955 : void
10956 0 : duplicate_and_interleave (vec_info *vinfo, gimple_seq *seq, tree vector_type,
10957 : const vec<tree> &elts, unsigned int nresults,
10958 : vec<tree> &results)
10959 : {
10960 0 : unsigned int nelts = elts.length ();
10961 0 : tree element_type = TREE_TYPE (vector_type);
10962 :
10963 : /* (1) Find a vector mode VM with integer elements of mode IM. */
10964 0 : unsigned int nvectors = 1;
10965 0 : tree new_vector_type;
10966 0 : tree permutes[2];
10967 0 : if (!can_duplicate_and_interleave_p (vinfo, nelts, element_type,
10968 : &nvectors, &new_vector_type,
10969 : permutes))
10970 0 : gcc_unreachable ();
10971 :
10972 : /* Get a vector type that holds ELTS[0:NELTS/NELTS']. */
10973 0 : unsigned int partial_nelts = nelts / nvectors;
10974 0 : tree partial_vector_type = build_vector_type (element_type, partial_nelts);
10975 :
10976 0 : tree_vector_builder partial_elts;
10977 0 : auto_vec<tree, 32> pieces (nvectors * 2);
10978 0 : pieces.quick_grow_cleared (nvectors * 2);
10979 0 : for (unsigned int i = 0; i < nvectors; ++i)
10980 : {
10981 : /* (2) Replace ELTS[0:NELTS] with ELTS'[0:NELTS'], where each element of
10982 : ELTS' has mode IM. */
10983 0 : partial_elts.new_vector (partial_vector_type, partial_nelts, 1);
10984 0 : for (unsigned int j = 0; j < partial_nelts; ++j)
10985 0 : partial_elts.quick_push (elts[i * partial_nelts + j]);
10986 0 : tree t = gimple_build_vector (seq, &partial_elts);
10987 0 : t = gimple_build (seq, VIEW_CONVERT_EXPR,
10988 0 : TREE_TYPE (new_vector_type), t);
10989 :
10990 : /* (3) Duplicate each ELTS'[I] into a vector of mode VM. */
10991 0 : pieces[i] = gimple_build_vector_from_val (seq, new_vector_type, t);
10992 : }
10993 :
10994 : /* (4) Use a tree of VEC_PERM_EXPRs to create a single VM with the
10995 : correct byte contents.
10996 :
10997 : Conceptually, we need to repeat the following operation log2(nvectors)
10998 : times, where hi_start = nvectors / 2:
10999 :
11000 : out[i * 2] = VEC_PERM_EXPR (in[i], in[i + hi_start], lo_permute);
11001 : out[i * 2 + 1] = VEC_PERM_EXPR (in[i], in[i + hi_start], hi_permute);
11002 :
11003 : However, if each input repeats every N elements and the VF is
11004 : a multiple of N * 2, the HI result is the same as the LO result.
11005 : This will be true for the first N1 iterations of the outer loop,
11006 : followed by N2 iterations for which both the LO and HI results
11007 : are needed. I.e.:
11008 :
11009 : N1 + N2 = log2(nvectors)
11010 :
11011 : Each "N1 iteration" doubles the number of redundant vectors and the
11012 : effect of the process as a whole is to have a sequence of nvectors/2**N1
11013 : vectors that repeats 2**N1 times. Rather than generate these redundant
11014 : vectors, we halve the number of vectors for each N1 iteration. */
11015 : unsigned int in_start = 0;
11016 : unsigned int out_start = nvectors;
11017 : unsigned int new_nvectors = nvectors;
11018 0 : for (unsigned int in_repeat = 1; in_repeat < nvectors; in_repeat *= 2)
11019 : {
11020 0 : unsigned int hi_start = new_nvectors / 2;
11021 0 : unsigned int out_i = 0;
11022 0 : for (unsigned int in_i = 0; in_i < new_nvectors; ++in_i)
11023 : {
11024 0 : if ((in_i & 1) != 0
11025 0 : && multiple_p (TYPE_VECTOR_SUBPARTS (new_vector_type),
11026 : 2 * in_repeat))
11027 0 : continue;
11028 :
11029 0 : tree output = make_ssa_name (new_vector_type);
11030 0 : tree input1 = pieces[in_start + (in_i / 2)];
11031 0 : tree input2 = pieces[in_start + (in_i / 2) + hi_start];
11032 0 : gassign *stmt = gimple_build_assign (output, VEC_PERM_EXPR,
11033 : input1, input2,
11034 : permutes[in_i & 1]);
11035 0 : gimple_seq_add_stmt (seq, stmt);
11036 0 : pieces[out_start + out_i] = output;
11037 0 : out_i += 1;
11038 : }
11039 0 : std::swap (in_start, out_start);
11040 0 : new_nvectors = out_i;
11041 : }
11042 :
11043 : /* (5) Use VIEW_CONVERT_EXPR to cast the final VM to the required type. */
11044 0 : results.reserve (nresults);
11045 0 : for (unsigned int i = 0; i < nresults; ++i)
11046 0 : if (i < new_nvectors)
11047 0 : results.quick_push (gimple_build (seq, VIEW_CONVERT_EXPR, vector_type,
11048 0 : pieces[in_start + i]));
11049 : else
11050 0 : results.quick_push (results[i - new_nvectors]);
11051 0 : }
11052 :
11053 :
11054 : /* For constant and loop invariant defs in OP_NODE this function creates
11055 : vector defs that will be used in the vectorized stmts and stores them
11056 : to SLP_TREE_VEC_DEFS of OP_NODE. */
11057 :
11058 : static void
11059 502381 : vect_create_constant_vectors (vec_info *vinfo, slp_tree op_node)
11060 : {
11061 502381 : unsigned HOST_WIDE_INT nunits;
11062 502381 : tree vec_cst;
11063 502381 : unsigned j, number_of_places_left_in_vector;
11064 502381 : tree vector_type;
11065 502381 : tree vop;
11066 502381 : int group_size = op_node->ops.length ();
11067 502381 : unsigned int vec_num, i;
11068 502381 : unsigned number_of_copies = 1;
11069 502381 : bool constant_p;
11070 502381 : gimple_seq ctor_seq = NULL;
11071 502381 : auto_vec<tree, 16> permute_results;
11072 :
11073 : /* We always want SLP_TREE_VECTYPE (op_node) here correctly set. */
11074 502381 : vector_type = SLP_TREE_VECTYPE (op_node);
11075 :
11076 502381 : unsigned int number_of_vectors, excess_elts;
11077 502381 : bool res = vect_get_num_copies_for_invariant (vinfo, op_node,
11078 : &number_of_vectors,
11079 : &excess_elts);
11080 502381 : gcc_assert (res);
11081 502381 : SLP_TREE_VEC_DEFS (op_node).create (number_of_vectors);
11082 502381 : auto_vec<tree> voprnds (number_of_vectors);
11083 :
11084 : /* NUMBER_OF_COPIES is the number of times we need to use the same values in
11085 : created vectors. It is greater than 1 if unrolling is performed.
11086 :
11087 : For example, we have two scalar operands, s1 and s2 (e.g., group of
11088 : strided accesses of size two), while NUNITS is four (i.e., four scalars
11089 : of this type can be packed in a vector). The output vector will contain
11090 : two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
11091 : will be 2).
11092 :
11093 : If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
11094 : containing the operands.
11095 :
11096 : For example, NUNITS is four as before, and the group size is 8
11097 : (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
11098 : {s5, s6, s7, s8}. */
11099 :
11100 : /* When using duplicate_and_interleave, we just need one element for
11101 : each scalar statement. */
11102 502381 : if (!TYPE_VECTOR_SUBPARTS (vector_type).is_constant (&nunits))
11103 : nunits = group_size;
11104 :
11105 502381 : number_of_copies = (nunits * number_of_vectors - excess_elts) / group_size;
11106 :
11107 502381 : constant_p = true;
11108 502381 : tree uniform_elt = NULL_TREE;
11109 502381 : tree_vector_builder elts (vector_type, nunits, 1);
11110 502381 : elts.quick_grow (nunits);
11111 : /* Zero-pad the last vector if necessary. */
11112 502381 : number_of_places_left_in_vector = nunits - excess_elts;
11113 502639 : for (i = nunits; i > number_of_places_left_in_vector; --i)
11114 258 : elts[i-1] = build_zero_cst (TREE_TYPE (vector_type));
11115 : stmt_vec_info insert_after = NULL;
11116 1489423 : for (j = 0; j < number_of_copies; j++)
11117 : {
11118 987042 : tree op;
11119 3794428 : for (i = group_size - 1; op_node->ops.iterate (i, &op); i--)
11120 : {
11121 : /* Create 'vect_ = {op0,op1,...,opn}'. */
11122 1820344 : tree orig_op = op;
11123 1820344 : if (number_of_places_left_in_vector == nunits)
11124 : uniform_elt = op;
11125 1184866 : else if (uniform_elt && operand_equal_p (uniform_elt, op))
11126 749217 : op = elts[number_of_places_left_in_vector];
11127 : else
11128 : uniform_elt = NULL_TREE;
11129 1820344 : number_of_places_left_in_vector--;
11130 1820344 : if (!types_compatible_p (TREE_TYPE (vector_type), TREE_TYPE (op)))
11131 : {
11132 283395 : if (CONSTANT_CLASS_P (op))
11133 : {
11134 102452 : if (VECTOR_BOOLEAN_TYPE_P (vector_type))
11135 : {
11136 : /* Can't use VIEW_CONVERT_EXPR for booleans because
11137 : of possibly different sizes of scalar value and
11138 : vector element. */
11139 66 : if (integer_zerop (op))
11140 66 : op = build_int_cst (TREE_TYPE (vector_type), 0);
11141 0 : else if (integer_onep (op))
11142 0 : op = build_all_ones_cst (TREE_TYPE (vector_type));
11143 : else
11144 0 : gcc_unreachable ();
11145 : }
11146 : else
11147 102386 : op = fold_unary (VIEW_CONVERT_EXPR,
11148 : TREE_TYPE (vector_type), op);
11149 102452 : gcc_assert (op && CONSTANT_CLASS_P (op));
11150 : }
11151 : else
11152 : {
11153 180943 : tree new_temp = make_ssa_name (TREE_TYPE (vector_type));
11154 180943 : gimple *init_stmt;
11155 180943 : if (VECTOR_BOOLEAN_TYPE_P (vector_type))
11156 : {
11157 427 : tree true_val
11158 427 : = build_all_ones_cst (TREE_TYPE (vector_type));
11159 427 : tree false_val
11160 427 : = build_zero_cst (TREE_TYPE (vector_type));
11161 427 : gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (op)));
11162 427 : init_stmt = gimple_build_assign (new_temp, COND_EXPR,
11163 : op, true_val,
11164 : false_val);
11165 : }
11166 : else
11167 : {
11168 180516 : op = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type),
11169 : op);
11170 180516 : init_stmt
11171 180516 : = gimple_build_assign (new_temp, VIEW_CONVERT_EXPR,
11172 : op);
11173 : }
11174 180943 : gimple_seq_add_stmt (&ctor_seq, init_stmt);
11175 180943 : op = new_temp;
11176 : }
11177 : }
11178 1820344 : elts[number_of_places_left_in_vector] = op;
11179 1820344 : if (!CONSTANT_CLASS_P (op))
11180 326988 : constant_p = false;
11181 : /* For BB vectorization we have to compute an insert location
11182 : when a def is inside the analyzed region since we cannot
11183 : simply insert at the BB start in this case. */
11184 1820344 : stmt_vec_info opdef;
11185 1820344 : if (TREE_CODE (orig_op) == SSA_NAME
11186 189348 : && !SSA_NAME_IS_DEFAULT_DEF (orig_op)
11187 168990 : && is_a <bb_vec_info> (vinfo)
11188 1931555 : && (opdef = vinfo->lookup_def (orig_op)))
11189 : {
11190 91030 : if (!insert_after)
11191 : insert_after = opdef;
11192 : else
11193 49939 : insert_after = get_later_stmt (insert_after, opdef);
11194 : }
11195 :
11196 1820344 : if (number_of_places_left_in_vector == 0)
11197 : {
11198 635507 : auto type_nunits = TYPE_VECTOR_SUBPARTS (vector_type);
11199 635507 : if (uniform_elt)
11200 657920 : vec_cst = gimple_build_vector_from_val (&ctor_seq, vector_type,
11201 328960 : elts[0]);
11202 613094 : else if (constant_p
11203 613094 : ? multiple_p (type_nunits, nunits)
11204 113958 : : known_eq (type_nunits, nunits))
11205 306547 : vec_cst = gimple_build_vector (&ctor_seq, &elts);
11206 : else
11207 : {
11208 0 : if (permute_results.is_empty ())
11209 0 : duplicate_and_interleave (vinfo, &ctor_seq, vector_type,
11210 : elts, number_of_vectors,
11211 : permute_results);
11212 0 : vec_cst = permute_results[number_of_vectors - j - 1];
11213 : }
11214 635507 : if (!gimple_seq_empty_p (ctor_seq))
11215 : {
11216 141908 : if (insert_after)
11217 : {
11218 41091 : gimple_stmt_iterator gsi;
11219 41091 : if (gimple_code (insert_after->stmt) == GIMPLE_PHI)
11220 : {
11221 763 : gsi = gsi_after_labels (gimple_bb (insert_after->stmt));
11222 763 : gsi_insert_seq_before (&gsi, ctor_seq,
11223 : GSI_CONTINUE_LINKING);
11224 : }
11225 40328 : else if (!stmt_ends_bb_p (insert_after->stmt))
11226 : {
11227 40328 : gsi = gsi_for_stmt (insert_after->stmt);
11228 40328 : gsi_insert_seq_after (&gsi, ctor_seq,
11229 : GSI_CONTINUE_LINKING);
11230 : }
11231 : else
11232 : {
11233 : /* When we want to insert after a def where the
11234 : defining stmt throws then insert on the fallthru
11235 : edge. */
11236 0 : edge e = find_fallthru_edge
11237 0 : (gimple_bb (insert_after->stmt)->succs);
11238 0 : basic_block new_bb
11239 0 : = gsi_insert_seq_on_edge_immediate (e, ctor_seq);
11240 0 : gcc_assert (!new_bb);
11241 : }
11242 : }
11243 : else
11244 100817 : vinfo->insert_seq_on_entry (NULL, ctor_seq);
11245 141908 : ctor_seq = NULL;
11246 : }
11247 635507 : voprnds.quick_push (vec_cst);
11248 635507 : insert_after = NULL;
11249 635507 : number_of_places_left_in_vector = nunits;
11250 635507 : constant_p = true;
11251 635507 : elts.new_vector (vector_type, nunits, 1);
11252 635507 : elts.quick_grow (nunits);
11253 : }
11254 : }
11255 : }
11256 :
11257 502381 : gcc_assert (number_of_places_left_in_vector == nunits);
11258 :
11259 : /* Since the vectors are created in the reverse order, we should invert
11260 : them. */
11261 502381 : vec_num = voprnds.length ();
11262 502381 : gcc_checking_assert (vec_num > 0);
11263 :
11264 1137888 : for (j = vec_num; j != 0; j--)
11265 : {
11266 635507 : vop = voprnds[j - 1];
11267 635507 : SLP_TREE_VEC_DEFS (op_node).quick_push (vop);
11268 : }
11269 :
11270 : /* In case that VF is greater than the unrolling factor needed for the SLP
11271 : group of stmts, NUMBER_OF_VECTORS to be created is greater than
11272 : NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
11273 : to replicate the vectors. */
11274 1004762 : while (number_of_vectors > SLP_TREE_VEC_DEFS (op_node).length ())
11275 0 : for (i = 0; SLP_TREE_VEC_DEFS (op_node).iterate (i, &vop) && i < vec_num;
11276 : i++)
11277 0 : SLP_TREE_VEC_DEFS (op_node).quick_push (vop);
11278 502381 : }
11279 :
11280 : /* Get the scalar definition of the Nth lane from SLP_NODE or NULL_TREE
11281 : if there is no definition for it in the scalar IL or it is not known. */
11282 :
11283 : tree
11284 2617 : vect_get_slp_scalar_def (slp_tree slp_node, unsigned n)
11285 : {
11286 2617 : if (SLP_TREE_DEF_TYPE (slp_node) == vect_internal_def)
11287 : {
11288 2605 : if (!SLP_TREE_SCALAR_STMTS (slp_node).exists ())
11289 : return NULL_TREE;
11290 2605 : stmt_vec_info def = SLP_TREE_SCALAR_STMTS (slp_node)[n];
11291 2605 : if (!def)
11292 : return NULL_TREE;
11293 2605 : return gimple_get_lhs (STMT_VINFO_STMT (def));
11294 : }
11295 : else
11296 12 : return SLP_TREE_SCALAR_OPS (slp_node)[n];
11297 : }
11298 :
11299 : /* Get the Ith vectorized definition from SLP_NODE. */
11300 :
11301 : tree
11302 148409 : vect_get_slp_vect_def (slp_tree slp_node, unsigned i)
11303 : {
11304 148409 : return SLP_TREE_VEC_DEFS (slp_node)[i];
11305 : }
11306 :
11307 : /* Get the vectorized definitions of SLP_NODE in *VEC_DEFS. */
11308 :
11309 : void
11310 957443 : vect_get_slp_defs (slp_tree slp_node, vec<tree> *vec_defs)
11311 : {
11312 1914886 : vec_defs->create (SLP_TREE_VEC_DEFS (slp_node).length ());
11313 957443 : vec_defs->splice (SLP_TREE_VEC_DEFS (slp_node));
11314 957443 : }
11315 :
11316 : /* Get N vectorized definitions for SLP_NODE. */
11317 :
11318 : void
11319 3051 : vect_get_slp_defs (vec_info *,
11320 : slp_tree slp_node, vec<vec<tree> > *vec_oprnds, unsigned n)
11321 : {
11322 3051 : if (n == -1U)
11323 3051 : n = SLP_TREE_CHILDREN (slp_node).length ();
11324 :
11325 10920 : for (unsigned i = 0; i < n; ++i)
11326 : {
11327 7869 : slp_tree child = SLP_TREE_CHILDREN (slp_node)[i];
11328 7869 : vec<tree> vec_defs = vNULL;
11329 7869 : vect_get_slp_defs (child, &vec_defs);
11330 7869 : vec_oprnds->quick_push (vec_defs);
11331 : }
11332 3051 : }
11333 :
11334 : /* A subroutine of vect_transform_slp_perm_load with two extra arguments:
11335 : - PERM gives the permutation that the caller wants to use for NODE,
11336 : which might be different from SLP_LOAD_PERMUTATION.
11337 : - DUMP_P controls whether the function dumps information. */
11338 :
11339 : static bool
11340 138262 : vect_transform_slp_perm_load_1 (vec_info *vinfo, slp_tree node,
11341 : load_permutation_t &perm,
11342 : const vec<tree> &dr_chain,
11343 : gimple_stmt_iterator *gsi, poly_uint64 vf,
11344 : bool analyze_only, bool dump_p,
11345 : unsigned *n_perms, unsigned int *n_loads,
11346 : bool dce_chain)
11347 : {
11348 138262 : stmt_vec_info stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
11349 138262 : int vec_index = 0;
11350 138262 : tree vectype = SLP_TREE_VECTYPE (node);
11351 138262 : unsigned int group_size = SLP_TREE_LANES (node);
11352 138262 : unsigned int mask_element;
11353 138262 : unsigned dr_group_size;
11354 138262 : machine_mode mode;
11355 :
11356 138262 : if (!STMT_VINFO_GROUPED_ACCESS (stmt_info))
11357 : {
11358 : /* We have both splats of the same non-grouped load and groups
11359 : of distinct invariant loads entering here. */
11360 1513 : unsigned max_idx = 0;
11361 8373 : for (auto idx : perm)
11362 3834 : max_idx = idx > max_idx ? idx : max_idx;
11363 1513 : dr_group_size = max_idx + 1;
11364 : }
11365 : else
11366 : {
11367 136749 : stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
11368 136749 : dr_group_size = DR_GROUP_SIZE (stmt_info);
11369 : }
11370 :
11371 138262 : mode = TYPE_MODE (vectype);
11372 138262 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
11373 138262 : unsigned int nstmts = vect_get_num_copies (vinfo, node);
11374 :
11375 : /* Initialize the vect stmts of NODE to properly insert the generated
11376 : stmts later. */
11377 138262 : if (! analyze_only)
11378 60758 : for (unsigned i = SLP_TREE_VEC_DEFS (node).length (); i < nstmts; i++)
11379 23310 : SLP_TREE_VEC_DEFS (node).quick_push (NULL_TREE);
11380 :
11381 : /* Generate permutation masks for every NODE. Number of masks for each NODE
11382 : is equal to GROUP_SIZE.
11383 : E.g., we have a group of three nodes with three loads from the same
11384 : location in each node, and the vector size is 4. I.e., we have a
11385 : a0b0c0a1b1c1... sequence and we need to create the following vectors:
11386 : for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3
11387 : for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3
11388 : ...
11389 :
11390 : The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}.
11391 : The last mask is illegal since we assume two operands for permute
11392 : operation, and the mask element values can't be outside that range.
11393 : Hence, the last mask must be converted into {2,5,5,5}.
11394 : For the first two permutations we need the first and the second input
11395 : vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation
11396 : we need the second and the third vectors: {b1,c1,a2,b2} and
11397 : {c2,a3,b3,c3}. */
11398 :
11399 138262 : int vect_stmts_counter = 0;
11400 138262 : unsigned int index = 0;
11401 138262 : int first_vec_index = -1;
11402 138262 : int second_vec_index = -1;
11403 138262 : bool noop_p = true;
11404 138262 : *n_perms = 0;
11405 :
11406 138262 : vec_perm_builder mask;
11407 138262 : unsigned int nelts_to_build;
11408 138262 : unsigned int nvectors_per_build;
11409 138262 : unsigned int in_nlanes;
11410 138262 : bool repeating_p = (group_size == dr_group_size
11411 174240 : && multiple_p (nunits, group_size));
11412 138262 : if (repeating_p)
11413 : {
11414 : /* A single vector contains a whole number of copies of the node, so:
11415 : (a) all permutes can use the same mask; and
11416 : (b) the permutes only need a single vector input. */
11417 33520 : mask.new_vector (nunits, group_size, 3);
11418 33520 : nelts_to_build = mask.encoded_nelts ();
11419 : /* It's possible to obtain zero nstmts during analyze_only, so make
11420 : it at least one to ensure the later computation for n_perms
11421 : proceed. */
11422 33520 : nvectors_per_build = nstmts > 0 ? nstmts : 1;
11423 33520 : in_nlanes = dr_group_size * 3;
11424 : }
11425 : else
11426 : {
11427 : /* We need to construct a separate mask for each vector statement. */
11428 104742 : unsigned HOST_WIDE_INT const_nunits, const_vf;
11429 104742 : if (!nunits.is_constant (&const_nunits)
11430 104742 : || !vf.is_constant (&const_vf))
11431 : return false;
11432 104742 : mask.new_vector (const_nunits, const_nunits, 1);
11433 104742 : nelts_to_build = const_vf * group_size;
11434 104742 : nvectors_per_build = 1;
11435 104742 : in_nlanes = const_vf * dr_group_size;
11436 : }
11437 138262 : auto_sbitmap used_in_lanes (in_nlanes);
11438 138262 : bitmap_clear (used_in_lanes);
11439 138262 : auto_bitmap used_defs;
11440 :
11441 138262 : unsigned int count = mask.encoded_nelts ();
11442 138262 : mask.quick_grow (count);
11443 138262 : vec_perm_indices indices;
11444 :
11445 727980 : for (unsigned int j = 0; j < nelts_to_build; j++)
11446 : {
11447 600549 : unsigned int iter_num = j / group_size;
11448 600549 : unsigned int stmt_num = j % group_size;
11449 600549 : unsigned int i = (iter_num * dr_group_size + perm[stmt_num]);
11450 600549 : bitmap_set_bit (used_in_lanes, i);
11451 600549 : if (repeating_p)
11452 : {
11453 : first_vec_index = 0;
11454 : mask_element = i;
11455 : }
11456 : else
11457 : {
11458 : /* Enforced before the loop when !repeating_p. */
11459 382143 : unsigned int const_nunits = nunits.to_constant ();
11460 382143 : vec_index = i / const_nunits;
11461 382143 : mask_element = i % const_nunits;
11462 382143 : if (vec_index == first_vec_index
11463 382143 : || first_vec_index == -1)
11464 : {
11465 : first_vec_index = vec_index;
11466 : }
11467 153253 : else if (vec_index == second_vec_index
11468 153253 : || second_vec_index == -1)
11469 : {
11470 146955 : second_vec_index = vec_index;
11471 146955 : mask_element += const_nunits;
11472 : }
11473 : else
11474 : {
11475 6298 : if (dump_p)
11476 280 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
11477 : "permutation requires at "
11478 : "least three vectors %G",
11479 : stmt_info->stmt);
11480 6298 : gcc_assert (analyze_only);
11481 : return false;
11482 : }
11483 :
11484 375845 : gcc_assert (mask_element < 2 * const_nunits);
11485 : }
11486 :
11487 594251 : if (mask_element != index)
11488 381914 : noop_p = false;
11489 594251 : mask[index++] = mask_element;
11490 :
11491 594251 : if (index == count)
11492 : {
11493 162838 : if (!noop_p)
11494 : {
11495 132750 : indices.new_vector (mask, second_vec_index == -1 ? 1 : 2, nunits);
11496 132750 : if (!can_vec_perm_const_p (mode, mode, indices))
11497 : {
11498 4533 : if (dump_p)
11499 : {
11500 90 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
11501 : "unsupported vect permute { ");
11502 876 : for (i = 0; i < count; ++i)
11503 : {
11504 786 : dump_dec (MSG_MISSED_OPTIMIZATION, mask[i]);
11505 786 : dump_printf (MSG_MISSED_OPTIMIZATION, " ");
11506 : }
11507 90 : dump_printf (MSG_MISSED_OPTIMIZATION, "}\n");
11508 : }
11509 4533 : gcc_assert (analyze_only);
11510 : return false;
11511 : }
11512 :
11513 128217 : tree mask_vec = NULL_TREE;
11514 128217 : if (!analyze_only)
11515 21567 : mask_vec = vect_gen_perm_mask_checked (vectype, indices);
11516 :
11517 128217 : if (second_vec_index == -1)
11518 36979 : second_vec_index = first_vec_index;
11519 :
11520 259299 : for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
11521 : {
11522 131082 : ++*n_perms;
11523 131082 : if (analyze_only)
11524 109232 : continue;
11525 : /* Generate the permute statement if necessary. */
11526 21850 : tree first_vec = dr_chain[first_vec_index + ri];
11527 21850 : tree second_vec = dr_chain[second_vec_index + ri];
11528 21850 : gassign *stmt = as_a<gassign *> (stmt_info->stmt);
11529 21850 : tree perm_dest
11530 21850 : = vect_create_destination_var (gimple_assign_lhs (stmt),
11531 : vectype);
11532 21850 : perm_dest = make_ssa_name (perm_dest);
11533 21850 : gimple *perm_stmt
11534 21850 : = gimple_build_assign (perm_dest, VEC_PERM_EXPR, first_vec,
11535 : second_vec, mask_vec);
11536 21850 : vect_finish_stmt_generation (vinfo, stmt_info, perm_stmt,
11537 : gsi);
11538 21850 : if (dce_chain)
11539 : {
11540 20927 : bitmap_set_bit (used_defs, first_vec_index + ri);
11541 20927 : bitmap_set_bit (used_defs, second_vec_index + ri);
11542 : }
11543 :
11544 : /* Store the vector statement in NODE. */
11545 21850 : SLP_TREE_VEC_DEFS (node)[vect_stmts_counter++] = perm_dest;
11546 : }
11547 : }
11548 30088 : else if (!analyze_only)
11549 : {
11550 2920 : for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
11551 : {
11552 1460 : tree first_vec = dr_chain[first_vec_index + ri];
11553 : /* If mask was NULL_TREE generate the requested
11554 : identity transform. */
11555 1460 : if (dce_chain)
11556 1453 : bitmap_set_bit (used_defs, first_vec_index + ri);
11557 :
11558 : /* Store the vector statement in NODE. */
11559 1460 : SLP_TREE_VEC_DEFS (node)[vect_stmts_counter++] = first_vec;
11560 : }
11561 : }
11562 :
11563 : index = 0;
11564 : first_vec_index = -1;
11565 : second_vec_index = -1;
11566 : noop_p = true;
11567 : }
11568 : }
11569 :
11570 127431 : if (n_loads)
11571 : {
11572 89792 : if (repeating_p)
11573 10885 : *n_loads = nstmts;
11574 : else
11575 : {
11576 : /* Enforced above when !repeating_p. */
11577 78907 : unsigned int const_nunits = nunits.to_constant ();
11578 78907 : *n_loads = 0;
11579 78907 : bool load_seen = false;
11580 1059222 : for (unsigned i = 0; i < in_nlanes; ++i)
11581 : {
11582 980315 : if (i % const_nunits == 0)
11583 : {
11584 418979 : if (load_seen)
11585 124751 : *n_loads += 1;
11586 : load_seen = false;
11587 : }
11588 980315 : if (bitmap_bit_p (used_in_lanes, i))
11589 272643 : load_seen = true;
11590 : }
11591 78907 : if (load_seen)
11592 52517 : *n_loads += 1;
11593 : }
11594 : }
11595 :
11596 127431 : if (dce_chain)
11597 214757 : for (unsigned i = 0; i < dr_chain.length (); ++i)
11598 76495 : if (!bitmap_bit_p (used_defs, i))
11599 : {
11600 41956 : tree def = dr_chain[i];
11601 42685 : do
11602 : {
11603 42685 : gimple *stmt = SSA_NAME_DEF_STMT (def);
11604 42685 : if (is_gimple_assign (stmt)
11605 42685 : && (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
11606 42685 : || gimple_assign_rhs_code (stmt) == CONSTRUCTOR))
11607 5330 : def = single_ssa_tree_operand (stmt, SSA_OP_USE);
11608 : else
11609 : def = NULL;
11610 42685 : gimple_stmt_iterator rgsi = gsi_for_stmt (stmt);
11611 42685 : gsi_remove (&rgsi, true);
11612 42685 : release_defs (stmt);
11613 : }
11614 42685 : while (def);
11615 : }
11616 :
11617 : return true;
11618 138262 : }
11619 :
11620 : /* Generate vector permute statements from a list of loads in DR_CHAIN.
11621 : If ANALYZE_ONLY is TRUE, only check that it is possible to create valid
11622 : permute statements for the SLP node NODE. Store the number of vector
11623 : permute instructions in *N_PERMS and the number of vector load
11624 : instructions in *N_LOADS. If DCE_CHAIN is true, remove all definitions
11625 : that were not needed. */
11626 :
11627 : bool
11628 99425 : vect_transform_slp_perm_load (vec_info *vinfo,
11629 : slp_tree node, const vec<tree> &dr_chain,
11630 : gimple_stmt_iterator *gsi, poly_uint64 vf,
11631 : bool analyze_only, unsigned *n_perms,
11632 : unsigned int *n_loads, bool dce_chain)
11633 : {
11634 99425 : return vect_transform_slp_perm_load_1 (vinfo, node,
11635 99425 : SLP_TREE_LOAD_PERMUTATION (node),
11636 : dr_chain, gsi, vf, analyze_only,
11637 : dump_enabled_p (), n_perms, n_loads,
11638 99425 : dce_chain);
11639 : }
11640 :
11641 : /* Produce the next vector result for SLP permutation NODE by adding a vector
11642 : statement at GSI. If MASK_VEC is nonnull, add:
11643 :
11644 : <new SSA name> = VEC_PERM_EXPR <FIRST_DEF, SECOND_DEF, MASK_VEC>
11645 :
11646 : otherwise add:
11647 :
11648 : <new SSA name> = VEC_PERM_EXPR <FIRST_DEF, SECOND_DEF,
11649 : { N, N+1, N+2, ... }>
11650 :
11651 : where N == IDENTITY_OFFSET which is either zero or equal to the
11652 : number of elements of the result. */
11653 :
11654 : static void
11655 32492 : vect_add_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
11656 : slp_tree node, tree first_def, tree second_def,
11657 : tree mask_vec, poly_uint64 identity_offset)
11658 : {
11659 32492 : tree vectype = SLP_TREE_VECTYPE (node);
11660 :
11661 : /* ??? We SLP match existing vector element extracts but
11662 : allow punning which we need to re-instantiate at uses
11663 : but have no good way of explicitly representing. */
11664 32492 : if (operand_equal_p (TYPE_SIZE (TREE_TYPE (first_def)), TYPE_SIZE (vectype))
11665 32492 : && !types_compatible_p (TREE_TYPE (first_def), vectype))
11666 : {
11667 20 : gassign *conv_stmt
11668 20 : = gimple_build_assign (make_ssa_name (vectype),
11669 : build1 (VIEW_CONVERT_EXPR, vectype, first_def));
11670 20 : vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
11671 20 : first_def = gimple_assign_lhs (conv_stmt);
11672 : }
11673 32492 : gassign *perm_stmt;
11674 32492 : if (mask_vec)
11675 : {
11676 28421 : if (operand_equal_p (TYPE_SIZE (TREE_TYPE (first_def)),
11677 28421 : TYPE_SIZE (vectype))
11678 28421 : && !types_compatible_p (TREE_TYPE (second_def), vectype))
11679 : {
11680 8 : gassign *conv_stmt
11681 8 : = gimple_build_assign (make_ssa_name (vectype),
11682 : build1 (VIEW_CONVERT_EXPR,
11683 : vectype, second_def));
11684 8 : vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
11685 8 : second_def = gimple_assign_lhs (conv_stmt);
11686 : }
11687 28421 : tree perm_dest = make_ssa_name (vectype);
11688 28421 : perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
11689 : first_def, second_def,
11690 : mask_vec);
11691 : }
11692 : else
11693 : {
11694 4071 : auto def_nunits = TYPE_VECTOR_SUBPARTS (TREE_TYPE (first_def));
11695 4071 : unsigned HOST_WIDE_INT vecno;
11696 4071 : poly_uint64 eltno;
11697 4071 : if (!can_div_trunc_p (poly_uint64 (identity_offset), def_nunits,
11698 : &vecno, &eltno))
11699 : gcc_unreachable ();
11700 4071 : tree def = vecno & 1 ? second_def : first_def;
11701 4071 : if (!types_compatible_p (TREE_TYPE (def), vectype))
11702 : {
11703 : /* For identity permutes we still need to handle the case
11704 : of offsetted extracts or concats. */
11705 406 : tree perm_dest = make_ssa_name (vectype);
11706 406 : unsigned HOST_WIDE_INT c;
11707 406 : if (known_le (TYPE_VECTOR_SUBPARTS (vectype), def_nunits))
11708 : {
11709 402 : unsigned HOST_WIDE_INT elsz
11710 402 : = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (TREE_TYPE (def))));
11711 804 : tree lowpart = build3 (BIT_FIELD_REF, vectype, def,
11712 402 : TYPE_SIZE (vectype),
11713 402 : bitsize_int (eltno * elsz));
11714 402 : perm_stmt = gimple_build_assign (perm_dest, lowpart);
11715 : }
11716 4 : else if (constant_multiple_p (TYPE_VECTOR_SUBPARTS (vectype),
11717 4 : def_nunits, &c) && c == 2)
11718 : {
11719 4 : gcc_assert (known_eq (identity_offset, 0U));
11720 4 : tree ctor = build_constructor_va (vectype, 2,
11721 : NULL_TREE, first_def,
11722 : NULL_TREE, second_def);
11723 4 : perm_stmt = gimple_build_assign (perm_dest, ctor);
11724 : }
11725 : else
11726 0 : gcc_unreachable ();
11727 : }
11728 : else
11729 : {
11730 3665 : gcc_assert (known_eq (eltno, 0U));
11731 3665 : node->push_vec_def (def);
11732 3665 : return;
11733 : }
11734 : }
11735 28827 : vect_finish_stmt_generation (vinfo, NULL, perm_stmt, gsi);
11736 : /* Store the vector statement in NODE. */
11737 28827 : node->push_vec_def (perm_stmt);
11738 : }
11739 :
11740 : /* Subroutine of vectorizable_slp_permutation. Check whether the target
11741 : can perform permutation PERM on the (1 or 2) input nodes in CHILDREN.
11742 : If GSI is nonnull, emit the permutation there.
11743 :
11744 : When GSI is null, the only purpose of NODE is to give properties
11745 : of the result, such as the vector type and number of SLP lanes.
11746 : The node does not need to be a VEC_PERM_EXPR.
11747 :
11748 : If the target supports the operation, return the number of individual
11749 : VEC_PERM_EXPRs needed, otherwise return -1. Print information to the
11750 : dump file if DUMP_P is true. */
11751 :
11752 : static int
11753 443593 : vectorizable_slp_permutation_1 (vec_info *vinfo, gimple_stmt_iterator *gsi,
11754 : slp_tree node, lane_permutation_t &perm,
11755 : vec<slp_tree> &children, bool dump_p)
11756 : {
11757 443593 : tree vectype = SLP_TREE_VECTYPE (node);
11758 :
11759 : /* ??? We currently only support all same vector input types
11760 : while the SLP IL should really do a concat + select and thus accept
11761 : arbitrary mismatches. */
11762 443593 : slp_tree child;
11763 443593 : unsigned i;
11764 443593 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
11765 443593 : bool repeating_p = multiple_p (nunits, SLP_TREE_LANES (node));
11766 : /* True if we're permuting a single input of 2N vectors down
11767 : to N vectors. This case doesn't generalize beyond 2 since
11768 : VEC_PERM_EXPR only takes 2 inputs. */
11769 443593 : bool pack_p = false;
11770 : /* If we're permuting inputs of N vectors each into X*N outputs,
11771 : this is the value of X, otherwise it is 1. */
11772 443593 : unsigned int unpack_factor = 1;
11773 443593 : tree op_vectype = NULL_TREE;
11774 445209 : FOR_EACH_VEC_ELT (children, i, child)
11775 445115 : if (SLP_TREE_VECTYPE (child))
11776 : {
11777 : op_vectype = SLP_TREE_VECTYPE (child);
11778 : break;
11779 : }
11780 443593 : if (!op_vectype)
11781 94 : op_vectype = vectype;
11782 932907 : FOR_EACH_VEC_ELT (children, i, child)
11783 : {
11784 489314 : if ((SLP_TREE_DEF_TYPE (child) != vect_internal_def
11785 17749 : && !vect_maybe_update_slp_op_vectype (child, op_vectype))
11786 489314 : || !types_compatible_p (SLP_TREE_VECTYPE (child), op_vectype)
11787 978628 : || !types_compatible_p (TREE_TYPE (vectype), TREE_TYPE (op_vectype)))
11788 : {
11789 0 : if (dump_p)
11790 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
11791 : "Unsupported vector types in lane permutation\n");
11792 443593 : return -1;
11793 : }
11794 489314 : auto op_nunits = TYPE_VECTOR_SUBPARTS (op_vectype);
11795 489314 : unsigned int this_unpack_factor;
11796 : /* Detect permutations of external, pre-existing vectors. The external
11797 : node's SLP_TREE_LANES stores the total number of units in the vector,
11798 : or zero if the vector has variable length.
11799 :
11800 : We are expected to keep the original VEC_PERM_EXPR for such cases.
11801 : There is no repetition to model. */
11802 489314 : if (SLP_TREE_DEF_TYPE (child) == vect_external_def
11803 489314 : && SLP_TREE_SCALAR_OPS (child).is_empty ())
11804 : repeating_p = false;
11805 : /* Check whether the input has twice as many lanes per vector. */
11806 474315 : else if (children.length () == 1
11807 474315 : && known_eq (SLP_TREE_LANES (child) * nunits,
11808 : SLP_TREE_LANES (node) * op_nunits * 2))
11809 : pack_p = true;
11810 : /* Check whether the output has N times as many lanes per vector. */
11811 489314 : else if (constant_multiple_p (SLP_TREE_LANES (node) * op_nunits,
11812 430135 : SLP_TREE_LANES (child) * nunits,
11813 : &this_unpack_factor)
11814 395196 : && (i == 0 || unpack_factor == this_unpack_factor))
11815 : unpack_factor = this_unpack_factor;
11816 : else
11817 : repeating_p = false;
11818 : }
11819 :
11820 887186 : gcc_assert (perm.length () == SLP_TREE_LANES (node));
11821 :
11822 : /* Load-lanes permute. This permute only acts as a forwarder to
11823 : select the correct vector def of the load-lanes load which
11824 : has the permuted vectors in its vector defs like
11825 : { v0, w0, r0, v1, w1, r1 ... } for a ld3. All costs are
11826 : accounted for in the costing for the actual load so we
11827 : return zero here. */
11828 443593 : if (node->ldst_lanes)
11829 : {
11830 0 : gcc_assert (children.length () == 1);
11831 0 : if (!gsi)
11832 : /* This is a trivial op always supported. */
11833 : return 0;
11834 0 : slp_tree child = children[0];
11835 0 : unsigned vec_idx = (SLP_TREE_LANE_PERMUTATION (node)[0].second
11836 0 : / SLP_TREE_LANES (node));
11837 0 : unsigned vec_num = SLP_TREE_LANES (child) / SLP_TREE_LANES (node);
11838 0 : unsigned nvectors = vect_get_num_copies (vinfo, node);
11839 0 : for (unsigned i = 0; i < nvectors; ++i)
11840 : {
11841 0 : tree def = SLP_TREE_VEC_DEFS (child)[i * vec_num + vec_idx];
11842 0 : node->push_vec_def (def);
11843 : }
11844 : return 0;
11845 : }
11846 :
11847 : /* Set REPEATING_P to true if the permutations are cyclical wrt UNPACK_FACTOR
11848 : and if we can generate the vectors in a vector-length agnostic way.
11849 : This requires UNPACK_STEP == NUNITS / UNPACK_FACTOR to be known at
11850 : compile time.
11851 :
11852 : The significance of UNPACK_STEP is that, when PACK_P is false,
11853 : output vector I operates on a window of UNPACK_STEP elements from each
11854 : input, starting at lane UNPACK_STEP * (I % UNPACK_FACTOR). For example,
11855 : when UNPACK_FACTOR is 2, the first output vector operates on lanes
11856 : [0, NUNITS / 2 - 1] of each input vector and the second output vector
11857 : operates on lanes [NUNITS / 2, NUNITS - 1] of each input vector.
11858 :
11859 : When REPEATING_P is true, NOUTPUTS holds the total number of outputs
11860 : that we actually need to generate. */
11861 443593 : uint64_t noutputs = 0;
11862 443593 : poly_uint64 unpack_step = 0;
11863 443593 : loop_vec_info linfo = dyn_cast <loop_vec_info> (vinfo);
11864 183554 : if (!linfo
11865 482734 : || !multiple_p (nunits, unpack_factor, &unpack_step)
11866 182604 : || !constant_multiple_p (LOOP_VINFO_VECT_FACTOR (linfo)
11867 182604 : * SLP_TREE_LANES (node), nunits, &noutputs))
11868 : repeating_p = false;
11869 :
11870 : /* We can handle the conditions described for REPEATING_P above for
11871 : both variable- and constant-length vectors. The fallback requires
11872 : us to generate every element of every permute vector explicitly,
11873 : which is only possible for constant-length permute vectors.
11874 :
11875 : Set:
11876 :
11877 : - NPATTERNS and NELTS_PER_PATTERN to the encoding of the permute
11878 : mask vectors that we want to build.
11879 :
11880 : - NCOPIES to the number of copies of PERM that we need in order
11881 : to build the necessary permute mask vectors. */
11882 182604 : uint64_t npatterns;
11883 182604 : unsigned nelts_per_pattern;
11884 182604 : uint64_t ncopies;
11885 182604 : if (repeating_p)
11886 : {
11887 : /* We need permute mask vectors that have the form:
11888 :
11889 : { X1, ..., Xn, X1 + n, ..., Xn + n, X1 + 2n, ..., Xn + 2n, ... }
11890 :
11891 : In other words, the original n-element permute in PERM is
11892 : "unrolled" to fill a full vector. The stepped vector encoding
11893 : that we use for permutes requires 3n elements. */
11894 143463 : npatterns = SLP_TREE_LANES (node);
11895 143463 : nelts_per_pattern = ncopies = 3;
11896 : }
11897 : else
11898 : {
11899 : /* Calculate every element of every permute mask vector explicitly,
11900 : instead of relying on the pattern described above. */
11901 300130 : if (!nunits.is_constant (&npatterns)
11902 300130 : || !TYPE_VECTOR_SUBPARTS (op_vectype).is_constant ())
11903 : {
11904 : if (dump_p)
11905 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
11906 : "unsupported permutation %p on variable-length"
11907 : " vectors\n", (void *) node);
11908 : return -1;
11909 : }
11910 300130 : nelts_per_pattern = ncopies = 1;
11911 300130 : if (linfo && !LOOP_VINFO_VECT_FACTOR (linfo).is_constant (&ncopies))
11912 : {
11913 : if (dump_p)
11914 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
11915 : "unsupported permutation %p for variable VF\n",
11916 : (void *) node);
11917 : return -1;
11918 : }
11919 : pack_p = false;
11920 : unpack_factor = 1;
11921 : }
11922 443593 : unsigned olanes = unpack_factor * ncopies * SLP_TREE_LANES (node);
11923 443593 : gcc_assert (repeating_p || multiple_p (olanes, nunits));
11924 :
11925 : /* Compute the { { SLP operand, vector index}, lane } permutation sequence
11926 : from the { SLP operand, scalar lane } permutation as recorded in the
11927 : SLP node as intermediate step. This part should already work
11928 : with SLP children with arbitrary number of lanes. */
11929 443593 : auto_vec<std::pair<std::pair<unsigned, unsigned>, poly_uint64>> vperm;
11930 443593 : auto_vec<poly_uint64> active_lane;
11931 443593 : vperm.create (olanes);
11932 443593 : active_lane.safe_grow_cleared (children.length (), true);
11933 1339292 : for (unsigned int ui = 0; ui < unpack_factor; ++ui)
11934 : {
11935 1916844 : for (unsigned j = 0; j < children.length (); ++j)
11936 506316 : active_lane[j] = ui * unpack_step;
11937 1306280 : for (unsigned i = 0; i < ncopies; ++i)
11938 : {
11939 5389330 : for (unsigned pi = 0; pi < perm.length (); ++pi)
11940 : {
11941 1840491 : std::pair<unsigned, unsigned> p = perm[pi];
11942 1840491 : tree vtype = SLP_TREE_VECTYPE (children[p.first]);
11943 1840491 : if (repeating_p)
11944 838959 : vperm.quick_push ({{p.first, 0},
11945 838959 : p.second + active_lane[p.first]});
11946 : else
11947 : {
11948 : /* We checked above that the vectors are constant-length. */
11949 1001532 : unsigned vnunits = TYPE_VECTOR_SUBPARTS (vtype)
11950 1001532 : .to_constant ();
11951 1001532 : unsigned lane = active_lane[p.first].to_constant ();
11952 1001532 : unsigned vi = (lane + p.second) / vnunits;
11953 1001532 : unsigned vl = (lane + p.second) % vnunits;
11954 1001532 : vperm.quick_push ({{p.first, vi}, vl});
11955 : }
11956 : }
11957 : /* Advance to the next group. */
11958 1828285 : for (unsigned j = 0; j < children.length (); ++j)
11959 974111 : active_lane[j] += SLP_TREE_LANES (children[j]);
11960 : }
11961 : }
11962 :
11963 443593 : if (dump_p)
11964 : {
11965 9032 : dump_printf_loc (MSG_NOTE, vect_location,
11966 : "vectorizing permutation %p", (void *)node);
11967 32685 : for (unsigned i = 0; i < perm.length (); ++i)
11968 23653 : dump_printf (MSG_NOTE, " op%u[%u]", perm[i].first, perm[i].second);
11969 9032 : if (repeating_p)
11970 7574 : dump_printf (MSG_NOTE, " (repeat %d)", SLP_TREE_LANES (node));
11971 9032 : dump_printf (MSG_NOTE, "\n");
11972 9032 : dump_printf_loc (MSG_NOTE, vect_location, "as");
11973 90765 : for (unsigned i = 0; i < vperm.length (); ++i)
11974 : {
11975 81733 : if (i != 0
11976 81733 : && (repeating_p
11977 55351 : ? multiple_p (i, npatterns)
11978 81733 : : multiple_p (i, TYPE_VECTOR_SUBPARTS (vectype))))
11979 24385 : dump_printf (MSG_NOTE, ",");
11980 81733 : dump_printf (MSG_NOTE, " vops%u[%u][",
11981 81733 : vperm[i].first.first, vperm[i].first.second);
11982 81733 : dump_dec (MSG_NOTE, vperm[i].second);
11983 81733 : dump_printf (MSG_NOTE, "]");
11984 : }
11985 9032 : dump_printf (MSG_NOTE, "\n");
11986 : }
11987 :
11988 : /* We can only handle two-vector permutes, everything else should
11989 : be lowered on the SLP level. The following is closely inspired
11990 : by vect_transform_slp_perm_load and is supposed to eventually
11991 : replace it.
11992 : ??? As intermediate step do code-gen in the SLP tree representation
11993 : somehow? */
11994 443593 : std::pair<unsigned, unsigned> first_vec = std::make_pair (-1U, -1U);
11995 443593 : std::pair<unsigned, unsigned> second_vec = std::make_pair (-1U, -1U);
11996 443593 : unsigned int index = 0;
11997 443593 : poly_uint64 mask_element;
11998 443593 : vec_perm_builder mask;
11999 443593 : mask.new_vector (nunits, npatterns, nelts_per_pattern);
12000 443593 : unsigned int count = mask.encoded_nelts ();
12001 443593 : mask.quick_grow (count);
12002 443593 : vec_perm_indices indices;
12003 443593 : unsigned nperms = 0;
12004 : /* When REPEATING_P is true, we only have UNPACK_FACTOR unique permute
12005 : vectors to check during analysis, but we need to generate NOUTPUTS
12006 : vectors during transformation. */
12007 443593 : unsigned total_nelts = olanes;
12008 443593 : unsigned process_nelts = olanes;
12009 443593 : if (repeating_p)
12010 : {
12011 143463 : total_nelts = (total_nelts / unpack_factor) * noutputs;
12012 143463 : if (gsi)
12013 9917 : process_nelts = total_nelts;
12014 : }
12015 443593 : unsigned last_ei = (total_nelts - 1) % process_nelts;
12016 2288322 : for (unsigned i = 0; i < process_nelts; ++i)
12017 : {
12018 : /* VI is the input vector index when generating code for REPEATING_P. */
12019 1854658 : unsigned vi = i / olanes * (pack_p ? 2 : 1);
12020 1854658 : unsigned ei = i % olanes;
12021 1854658 : mask_element = vperm[ei].second;
12022 1854658 : if (pack_p)
12023 : {
12024 : /* In this case, we have N outputs and the single child provides 2N
12025 : inputs. Output X permutes inputs 2X and 2X+1.
12026 :
12027 : The mask indices are taken directly from the SLP permutation node.
12028 : Index X selects from the first vector if (X / NUNITS) % 2 == 0;
12029 : X selects from the second vector otherwise. These conditions
12030 : are only known at compile time for constant-length vectors. */
12031 : first_vec = std::make_pair (0, 0);
12032 : second_vec = std::make_pair (0, 1);
12033 : }
12034 1684924 : else if (first_vec.first == -1U
12035 1684924 : || first_vec == vperm[ei].first)
12036 1487033 : first_vec = vperm[ei].first;
12037 197891 : else if (second_vec.first == -1U
12038 197891 : || second_vec == vperm[ei].first)
12039 : {
12040 197487 : second_vec = vperm[ei].first;
12041 197487 : mask_element += nunits;
12042 : }
12043 : else
12044 : {
12045 404 : if (dump_p)
12046 7 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
12047 : "permutation requires at "
12048 : "least three vectors\n");
12049 404 : gcc_assert (!gsi);
12050 : return -1;
12051 : }
12052 :
12053 1854254 : mask[index++] = mask_element;
12054 :
12055 1854254 : if (index == count)
12056 : {
12057 588796 : indices.new_vector (mask, second_vec.first == -1U ? 1 : 2,
12058 : TYPE_VECTOR_SUBPARTS (op_vectype));
12059 588796 : bool identity_p = (indices.series_p (0, 1, mask[0], 1)
12060 915521 : && constant_multiple_p (mask[0], nunits));
12061 588796 : machine_mode vmode = TYPE_MODE (vectype);
12062 588796 : machine_mode op_vmode = TYPE_MODE (op_vectype);
12063 588796 : unsigned HOST_WIDE_INT c;
12064 588796 : if ((!identity_p
12065 539438 : && !can_vec_perm_const_p (vmode, op_vmode, indices))
12066 588796 : || (identity_p
12067 49358 : && !known_le (nunits,
12068 : TYPE_VECTOR_SUBPARTS (op_vectype))
12069 8 : && (!constant_multiple_p (nunits,
12070 9533 : TYPE_VECTOR_SUBPARTS (op_vectype),
12071 8 : &c) || c != 2)))
12072 : {
12073 9525 : if (dump_p)
12074 : {
12075 154 : dump_printf_loc (MSG_MISSED_OPTIMIZATION,
12076 : vect_location,
12077 : "unsupported vect permute { ");
12078 1596 : for (i = 0; i < count; ++i)
12079 : {
12080 1442 : dump_dec (MSG_MISSED_OPTIMIZATION, mask[i]);
12081 1442 : dump_printf (MSG_MISSED_OPTIMIZATION, " ");
12082 : }
12083 154 : dump_printf (MSG_MISSED_OPTIMIZATION, "}\n");
12084 : }
12085 9525 : gcc_assert (!gsi);
12086 443593 : return -1;
12087 : }
12088 :
12089 579271 : if (!identity_p)
12090 529913 : nperms += CEIL (total_nelts, process_nelts) - (ei > last_ei);
12091 579271 : if (gsi)
12092 : {
12093 32492 : if (second_vec.first == -1U)
12094 7938 : second_vec = first_vec;
12095 :
12096 32492 : slp_tree
12097 32492 : first_node = children[first_vec.first],
12098 32492 : second_node = children[second_vec.first];
12099 :
12100 32492 : tree mask_vec = NULL_TREE;
12101 32492 : if (!identity_p)
12102 28421 : mask_vec = vect_gen_perm_mask_checked (vectype, indices);
12103 :
12104 32492 : tree first_def
12105 32492 : = vect_get_slp_vect_def (first_node, first_vec.second + vi);
12106 32492 : tree second_def
12107 32492 : = vect_get_slp_vect_def (second_node, second_vec.second + vi);
12108 32492 : vect_add_slp_permutation (vinfo, gsi, node, first_def,
12109 32492 : second_def, mask_vec, mask[0]);
12110 : }
12111 :
12112 : index = 0;
12113 : first_vec = std::make_pair (-1U, -1U);
12114 : second_vec = std::make_pair (-1U, -1U);
12115 : }
12116 : }
12117 :
12118 433664 : return nperms;
12119 443593 : }
12120 :
12121 : /* Vectorize the SLP permutations in NODE as specified
12122 : in SLP_TREE_LANE_PERMUTATION which is a vector of pairs of SLP
12123 : child number and lane number.
12124 : Interleaving of two two-lane two-child SLP subtrees (not supported):
12125 : [ { 0, 0 }, { 1, 0 }, { 0, 1 }, { 1, 1 } ]
12126 : A blend of two four-lane two-child SLP subtrees:
12127 : [ { 0, 0 }, { 1, 1 }, { 0, 2 }, { 1, 3 } ]
12128 : Highpart of a four-lane one-child SLP subtree (not supported):
12129 : [ { 0, 2 }, { 0, 3 } ]
12130 : Where currently only a subset is supported by code generating below. */
12131 :
12132 : bool
12133 99984 : vectorizable_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
12134 : slp_tree node, stmt_vector_for_cost *cost_vec)
12135 : {
12136 99984 : tree vectype = SLP_TREE_VECTYPE (node);
12137 99984 : lane_permutation_t &perm = SLP_TREE_LANE_PERMUTATION (node);
12138 99984 : int nperms = vectorizable_slp_permutation_1 (vinfo, gsi, node, perm,
12139 99984 : SLP_TREE_CHILDREN (node),
12140 : dump_enabled_p ());
12141 99984 : if (nperms < 0)
12142 : return false;
12143 :
12144 98453 : if (!gsi && nperms != 0)
12145 74671 : record_stmt_cost (cost_vec, nperms, vec_perm, node, vectype, 0, vect_body);
12146 :
12147 : return true;
12148 : }
12149 :
12150 : /* Vectorize SLP NODE. Only compute the vector insertion places when
12151 : PLACE_ONLY is true. When placing, return false if there is no possible
12152 : schedule. */
12153 :
12154 : static bool
12155 3147937 : vect_schedule_slp_node (vec_info *vinfo,
12156 : slp_tree node, slp_instance instance, bool place_only)
12157 : {
12158 3147937 : int i;
12159 3147937 : slp_tree child;
12160 :
12161 : /* Vectorize externals and constants. */
12162 3147937 : if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
12163 3147937 : || SLP_TREE_DEF_TYPE (node) == vect_external_def)
12164 : {
12165 1210071 : if (place_only)
12166 : {
12167 699320 : if (SLP_TREE_DEF_TYPE (node) == vect_constant_def)
12168 : return true;
12169 319147 : gimple *last_stmt = NULL;
12170 319147 : vec<tree> &defs = (!SLP_TREE_SCALAR_OPS (node).is_empty ()
12171 : ? SLP_TREE_SCALAR_OPS (node)
12172 319147 : : SLP_TREE_VEC_DEFS (node));
12173 1708352 : for (tree def : defs)
12174 : /* If the stmt is not inside the region do not
12175 : use it as possible insertion point. */
12176 750911 : if (auto stmt_info = vinfo->lookup_def (def))
12177 : {
12178 388523 : gimple *stmt = stmt_info->stmt;
12179 388523 : if (!last_stmt)
12180 : last_stmt = stmt;
12181 218502 : else if (vect_stmt_dominates_stmt_p (last_stmt, stmt))
12182 : last_stmt = stmt;
12183 65299 : else if (vect_stmt_dominates_stmt_p (stmt, last_stmt))
12184 : ;
12185 : else
12186 0 : gcc_unreachable ();
12187 : }
12188 319147 : node->si = last_stmt;
12189 319147 : return true;
12190 : }
12191 :
12192 : /* ??? vectorizable_shift can end up using a scalar operand which is
12193 : currently denoted as !SLP_TREE_VECTYPE. No need to vectorize the
12194 : node in this case. */
12195 510751 : if (!SLP_TREE_VECTYPE (node))
12196 : return true;
12197 :
12198 : /* There are two reasons vector defs might already exist. The first
12199 : is that we are vectorizing an existing vector def. The second is
12200 : when performing BB vectorization shared constant/external nodes
12201 : are not split apart during partitioning so during the code-gen
12202 : DFS walk we can end up visiting them twice. */
12203 503327 : if (! SLP_TREE_VEC_DEFS (node).exists ())
12204 502381 : vect_create_constant_vectors (vinfo, node);
12205 : return true;
12206 : }
12207 :
12208 1937866 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
12209 :
12210 1937866 : gcc_assert (SLP_TREE_VEC_DEFS (node).is_empty ());
12211 1937866 : if (!place_only && SLP_TREE_VECTYPE (node))
12212 998398 : SLP_TREE_VEC_DEFS (node).create (vect_get_num_copies (vinfo, node));
12213 :
12214 1937866 : gimple *last_stmt;
12215 1937866 : gimple_stmt_iterator si;
12216 : /* ??? When !place_only we'd like to re-use place_only computed info,
12217 : but this is a bit awkward due to using gsi_insert_before and the
12218 : requirement to insert after vector defs. So we compute last_stmt
12219 : during pre-scheduling and si during scheduling. */
12220 1937866 : if (!SLP_TREE_PERMUTE_P (node) && STMT_VINFO_DATA_REF (stmt_info))
12221 : {
12222 : /* Vectorized loads go before the first scalar load to make it
12223 : ready early, vectorized stores go before the last scalar
12224 : stmt which is where all uses are ready.
12225 : In theory, if we delay dependence checking until after
12226 : placing, we can schedule at other points, but then
12227 : dependence checking would need to honor that. On the
12228 : other hand dependence checking could request a different
12229 : scheduling point as well, if dependences require that. */
12230 1507947 : stmt_vec_info last_stmt_info = NULL;
12231 1507947 : if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
12232 284920 : last_stmt_info = vect_find_first_scalar_stmt_in_slp (node);
12233 : else /* DR_IS_WRITE */
12234 : {
12235 1223027 : last_stmt_info = vect_find_last_scalar_stmt_in_slp (node);
12236 1223027 : if (place_only)
12237 1331248 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
12238 : {
12239 665626 : if (child->si
12240 665626 : && !vect_stmt_dominates_stmt_p (child->si,
12241 : last_stmt_info->stmt))
12242 : return false;
12243 : }
12244 : }
12245 1507943 : last_stmt = last_stmt_info->stmt;
12246 1507943 : si = gsi_for_stmt (last_stmt);
12247 1507943 : }
12248 429919 : else if (!SLP_TREE_PERMUTE_P (node)
12249 408155 : && (SLP_TREE_TYPE (node) == cycle_phi_info_type
12250 : || SLP_TREE_TYPE (node) == induc_vec_info_type
12251 : || SLP_TREE_TYPE (node) == phi_info_type))
12252 : {
12253 : /* For PHI node vectorization we do not use the insertion iterator. */
12254 108649 : last_stmt = SLP_TREE_SCALAR_STMTS (node)[0]->stmt;
12255 108649 : if (place_only)
12256 181872 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
12257 : {
12258 128703 : if (child->si
12259 224383 : && !dominated_by_p (CDI_DOMINATORS,
12260 : gimple_phi_arg_edge
12261 95680 : (as_a <gphi *> (last_stmt), i)->src,
12262 95680 : gimple_bb (child->si)))
12263 : return false;
12264 : }
12265 108645 : si = gsi_none ();
12266 : }
12267 : else
12268 : {
12269 : /* Emit other stmts after the children vectorized defs which is
12270 : earliest possible. */
12271 : last_stmt = NULL;
12272 888898 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
12273 567630 : if (place_only)
12274 : {
12275 187657 : gimple *vstmt = child->si;
12276 187657 : if (!vstmt)
12277 : {
12278 : /* vect_constant_def and defs at region boundary do not
12279 : constrain placement. */
12280 : gcc_assert (SLP_TREE_DEF_TYPE (child) == vect_constant_def
12281 : /* ??? Region boundary is not representated
12282 : by a NULL stmt. */
12283 : || true);
12284 : }
12285 146890 : else if (!last_stmt)
12286 : last_stmt = vstmt;
12287 42458 : else if (vect_stmt_dominates_stmt_p (last_stmt, vstmt))
12288 : last_stmt = vstmt;
12289 10854 : else if (vect_stmt_dominates_stmt_p (vstmt, last_stmt))
12290 : ;
12291 : else
12292 : /* Non-trapping stmts from different BBs might be combined,
12293 : and if we later CSE a low/high part we can run into this. */
12294 : return false;
12295 : }
12296 379973 : else if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
12297 : {
12298 : /* For fold-left reductions we are retaining the scalar
12299 : reduction PHI but we still have SLP_TREE_NUM_VEC_STMTS
12300 : set so the representation isn't perfect. Resort to the
12301 : last scalar def here. */
12302 303119 : if (SLP_TREE_VEC_DEFS (child).is_empty ())
12303 : {
12304 956 : gcc_assert (SLP_TREE_TYPE (child) == cycle_phi_info_type);
12305 956 : gphi *phi = as_a <gphi *>
12306 956 : (vect_find_last_scalar_stmt_in_slp (child)->stmt);
12307 956 : if (!last_stmt)
12308 : last_stmt = phi;
12309 726 : else if (vect_stmt_dominates_stmt_p (last_stmt, phi))
12310 : last_stmt = phi;
12311 715 : else if (vect_stmt_dominates_stmt_p (phi, last_stmt))
12312 : ;
12313 : else
12314 0 : gcc_unreachable ();
12315 : }
12316 : /* We are emitting all vectorized stmts in the same place and
12317 : the last one is the last.
12318 : ??? Unless we have a load permutation applied and that
12319 : figures to re-use an earlier generated load. */
12320 303119 : unsigned j;
12321 303119 : tree vdef;
12322 715127 : FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (child), j, vdef)
12323 412008 : if (TREE_CODE (vdef) == SSA_NAME
12324 412008 : && !SSA_NAME_IS_DEFAULT_DEF (vdef))
12325 : {
12326 411956 : gimple *vstmt = SSA_NAME_DEF_STMT (vdef);
12327 411956 : if (!last_stmt)
12328 : last_stmt = vstmt;
12329 210319 : else if (vect_stmt_dominates_stmt_p (last_stmt, vstmt))
12330 : last_stmt = vstmt;
12331 47488 : else if (vect_stmt_dominates_stmt_p (vstmt, last_stmt))
12332 : ;
12333 : else
12334 0 : gcc_unreachable ();
12335 : }
12336 : }
12337 76854 : else if (!SLP_TREE_VECTYPE (child))
12338 : {
12339 : /* For externals we use unvectorized at all scalar defs. */
12340 : unsigned j;
12341 : tree def;
12342 16599 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (child), j, def)
12343 : /* If the stmt is not inside the region do not
12344 : use it as possible insertion point. */
12345 9803 : if (auto stmt_info = vinfo->lookup_def (def))
12346 : {
12347 250 : gimple *stmt = stmt_info->stmt;
12348 250 : if (!last_stmt)
12349 : last_stmt = stmt;
12350 230 : else if (vect_stmt_dominates_stmt_p (last_stmt, stmt))
12351 : last_stmt = stmt;
12352 26 : else if (vect_stmt_dominates_stmt_p (stmt, last_stmt))
12353 : ;
12354 : else
12355 0 : gcc_unreachable ();
12356 : }
12357 : }
12358 : else
12359 : {
12360 : /* For externals we have to look at all defs since their
12361 : insertion place is decided per vector. But beware
12362 : of pre-existing vectors where we need to make sure
12363 : we do not insert before the region boundary. */
12364 139985 : if (SLP_TREE_SCALAR_OPS (child).is_empty ()
12365 716 : && !vinfo->lookup_def (SLP_TREE_VEC_DEFS (child)[0]))
12366 : ;
12367 : else
12368 : {
12369 : unsigned j;
12370 : tree vdef;
12371 736032 : FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (child), j, vdef)
12372 98477 : if (TREE_CODE (vdef) == SSA_NAME
12373 98477 : && !SSA_NAME_IS_DEFAULT_DEF (vdef))
12374 : {
12375 22392 : gimple *vstmt = SSA_NAME_DEF_STMT (vdef);
12376 22392 : if (!last_stmt)
12377 : last_stmt = vstmt;
12378 12175 : else if (vect_stmt_dominates_stmt_p (last_stmt, vstmt))
12379 : last_stmt = vstmt;
12380 9599 : else if (vect_stmt_dominates_stmt_p (vstmt, last_stmt))
12381 : ;
12382 : else
12383 0 : gcc_unreachable ();
12384 : }
12385 : }
12386 : }
12387 :
12388 : /* We split regions to vectorize at control altering stmts
12389 : with a definition so this can only be an external. */
12390 321268 : gcc_checking_assert (!last_stmt
12391 : || !is_ctrl_altering_stmt (last_stmt));
12392 :
12393 321268 : if (is_a <bb_vec_info> (vinfo)
12394 130561 : && !SLP_TREE_PERMUTE_P (node)
12395 123942 : && (!last_stmt
12396 121093 : || !dominated_by_p (CDI_DOMINATORS, gimple_bb (last_stmt),
12397 121093 : gimple_bb (stmt_info->stmt)))
12398 346826 : && gimple_could_trap_p (stmt_info->stmt))
12399 : {
12400 : /* We've constrained possibly trapping operations to all come
12401 : from the same basic-block, if vectorized defs would allow earlier
12402 : scheduling still force vectorized stmts to the original block.
12403 : This is only necessary for BB vectorization since for loop vect
12404 : all operations are in a single BB and scalar stmt based
12405 : placement doesn't play well with epilogue vectorization. */
12406 675 : if (last_stmt
12407 1258 : && !dominated_by_p (CDI_DOMINATORS,
12408 583 : gimple_bb (stmt_info->stmt),
12409 583 : gimple_bb (last_stmt)))
12410 : {
12411 4 : gcc_assert (place_only);
12412 : return false;
12413 : }
12414 671 : si = gsi_after_labels (gimple_bb (stmt_info->stmt));
12415 671 : last_stmt = gsi_stmt (si);
12416 : }
12417 : /* When there is no in-region child def to guide placement, insert
12418 : at region boundary. */
12419 320593 : else if (!last_stmt)
12420 : {
12421 4642 : si = gsi_after_labels (vinfo->bbs[0]);
12422 : /* last_stmt NULL marks the region start. */
12423 : }
12424 315951 : else if (is_a <gphi *> (last_stmt))
12425 31278 : si = gsi_after_labels (gimple_bb (last_stmt));
12426 : else
12427 : {
12428 284673 : si = gsi_for_stmt (last_stmt);
12429 : /* We use gsi_insert_before, so when last_stmt is a vector
12430 : def we have to advance (or use gsi_insert_after). */
12431 284673 : gsi_next (&si);
12432 :
12433 284673 : if (auto loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
12434 : {
12435 : /* We'll have to fix this up for loop vect. */
12436 174775 : gcc_assert (!place_only);
12437 : /* Avoid scheduling stmts to random places in the CFG, any
12438 : stmt dominance check we performed is possibly wrong as UIDs
12439 : are not initialized for all of the function for loop
12440 : vectorization. Instead append to the loop preheader. */
12441 174775 : if ((LOOP_VINFO_LOOP (loop_vinfo)->header
12442 174775 : != gimple_bb (last_stmt))
12443 177729 : && dominated_by_p (CDI_DOMINATORS,
12444 : LOOP_VINFO_LOOP (loop_vinfo)->header,
12445 2954 : gimple_bb (last_stmt)))
12446 1126 : si = gsi_end_bb (loop_preheader_edge
12447 563 : (LOOP_VINFO_LOOP (loop_vinfo))->src);
12448 : /* Avoid scheduling internal defs outside of the loop when
12449 : we might have only implicitly tracked loop mask/len defs. */
12450 76 : if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
12451 174775 : || LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo))
12452 : {
12453 76 : gimple_stmt_iterator si2
12454 76 : = gsi_after_labels (LOOP_VINFO_LOOP (loop_vinfo)->header);
12455 76 : if ((gsi_end_p (si2)
12456 0 : && (LOOP_VINFO_LOOP (loop_vinfo)->header
12457 0 : != gimple_bb (last_stmt))
12458 0 : && dominated_by_p (CDI_DOMINATORS,
12459 : LOOP_VINFO_LOOP (loop_vinfo)->header,
12460 0 : gimple_bb (last_stmt)))
12461 76 : || (!gsi_end_p (si2)
12462 76 : && last_stmt != *si2
12463 73 : && vect_stmt_dominates_stmt_p (last_stmt, *si2)))
12464 3 : si = si2;
12465 : }
12466 : }
12467 : }
12468 : }
12469 :
12470 1937852 : if (place_only)
12471 : {
12472 939448 : if (dump_enabled_p () && last_stmt)
12473 3943 : dump_printf_loc (MSG_NOTE, vect_location,
12474 : "placing node %p at %G:", (void *)node, last_stmt);
12475 : /* Verify we either get a stmt anchor or region start. */
12476 939448 : gcc_assert ((last_stmt && gimple_bb (last_stmt))
12477 : || (!last_stmt && gsi_bb (si)));
12478 939448 : node->si = last_stmt;
12479 939448 : return true;
12480 : }
12481 :
12482 : /* ??? Asserting vect_stmt_dominates_stmt_p (gsi_stmt (si), node->si)
12483 : does not work because in some cases we advance si from last_stmt (as
12484 : we want to insert after vector stmts) and because vector stmts of
12485 : children have been inserted possibly at the same location constraint,
12486 : moving si even further. */
12487 998404 : if (flag_checking && node->si && gimple_bb (node->si) && !gsi_end_p (si))
12488 : {
12489 636587 : auto gsi2 = si;
12490 636587 : while (1)
12491 : {
12492 636587 : if (vect_stmt_dominates_stmt_p (gsi_stmt (gsi2), node->si))
12493 : break;
12494 : /* As we have possibly advanced si it might now point to the
12495 : scalar stmt immediately following node->si. That's OK. */
12496 30224 : if (gsi_stmt (gsi2) != gsi_stmt (si)
12497 30224 : && gimple_uid (gsi_stmt (gsi2)) != 0)
12498 0 : gcc_unreachable ();
12499 30224 : gsi_prev (&gsi2);
12500 30224 : if (gsi_end_p (gsi2))
12501 : {
12502 618 : if (is_a <gphi *> (node->si)
12503 618 : && gimple_bb (node->si) == gsi_bb (si))
12504 : break;
12505 0 : gcc_unreachable ();
12506 : }
12507 : }
12508 : }
12509 :
12510 998404 : if (dump_enabled_p ())
12511 : {
12512 72819 : if (stmt_info)
12513 69397 : dump_printf_loc (MSG_NOTE, vect_location,
12514 : "------>vectorizing SLP node starting from: %G",
12515 : stmt_info->stmt);
12516 : else
12517 : {
12518 3422 : dump_printf_loc (MSG_NOTE, vect_location,
12519 : "------>vectorizing SLP node:\n");
12520 3422 : vect_print_slp_tree (MSG_NOTE, vect_location, node);
12521 : }
12522 : }
12523 998404 : vect_transform_stmt (vinfo, stmt_info, &si, node, instance);
12524 998404 : return true;
12525 : }
12526 :
12527 : /* Replace scalar calls from SLP node NODE with setting of their lhs to zero.
12528 : For loop vectorization this is done in vectorizable_call, but for SLP
12529 : it needs to be deferred until end of vect_schedule_slp, because multiple
12530 : SLP instances may refer to the same scalar stmt. */
12531 :
12532 : static void
12533 606036 : vect_remove_slp_scalar_calls (vec_info *vinfo,
12534 : slp_tree node, hash_set<slp_tree> &visited)
12535 : {
12536 606036 : gimple *new_stmt;
12537 606036 : gimple_stmt_iterator gsi;
12538 606036 : tree lhs;
12539 :
12540 606036 : if (!node || SLP_TREE_DEF_TYPE (node) != vect_internal_def)
12541 189834 : return;
12542 :
12543 460138 : if (visited.add (node))
12544 : return;
12545 :
12546 1551417 : for (auto child : SLP_TREE_CHILDREN (node))
12547 515351 : vect_remove_slp_scalar_calls (vinfo, child, visited);
12548 :
12549 1729935 : for (auto stmt_info : SLP_TREE_SCALAR_STMTS (node))
12550 : {
12551 489939 : if (!stmt_info)
12552 4006 : continue;
12553 485933 : stmt_info = vect_orig_stmt (stmt_info);
12554 485933 : gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt);
12555 5166 : if (!stmt || gimple_bb (stmt) == NULL)
12556 480809 : continue;
12557 5124 : lhs = gimple_call_lhs (stmt);
12558 5124 : if (lhs)
12559 4553 : new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs)));
12560 : else
12561 571 : new_stmt = gimple_build_nop ();
12562 5124 : unlink_stmt_vdef (stmt_info->stmt);
12563 5124 : gsi = gsi_for_stmt (stmt);
12564 5124 : vinfo->replace_stmt (&gsi, stmt_info, new_stmt);
12565 5124 : if (lhs)
12566 4553 : SSA_NAME_DEF_STMT (lhs) = new_stmt;
12567 : }
12568 : }
12569 :
12570 : static void
12571 90685 : vect_remove_slp_scalar_calls (vec_info *vinfo, slp_tree node)
12572 : {
12573 90685 : hash_set<slp_tree> visited;
12574 90685 : vect_remove_slp_scalar_calls (vinfo, node, visited);
12575 90685 : }
12576 :
12577 : /* Vectorize the instance root. */
12578 :
12579 : void
12580 14746 : vectorize_slp_instance_root_stmt (vec_info *vinfo, slp_tree node, slp_instance instance)
12581 : {
12582 14746 : if (instance->kind == slp_inst_kind_ctor)
12583 : {
12584 5661 : tree new_def;
12585 5661 : if (SLP_TREE_VEC_DEFS (node).length () == 1)
12586 : {
12587 5616 : new_def = SLP_TREE_VEC_DEFS (node)[0];
12588 5616 : tree root_lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
12589 5616 : if (!useless_type_conversion_p (TREE_TYPE (root_lhs),
12590 5616 : TREE_TYPE (new_def)))
12591 0 : new_def = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (root_lhs),
12592 : new_def);
12593 : }
12594 : else
12595 : {
12596 45 : gcc_assert (SLP_TREE_VEC_DEFS (node).length () > 1);
12597 45 : tree child_def;
12598 45 : int j;
12599 45 : vec<constructor_elt, va_gc> *v;
12600 45 : vec_alloc (v, SLP_TREE_VEC_DEFS (node).length ());
12601 :
12602 : /* A CTOR can handle V16HI composition from VNx8HI so we
12603 : do not need to convert vector elements if the types
12604 : do not match. */
12605 180 : FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (node), j, child_def)
12606 90 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, child_def);
12607 45 : tree rtype
12608 45 : = TREE_TYPE (gimple_assign_rhs1 (instance->root_stmts[0]->stmt));
12609 45 : new_def = build_constructor (rtype, v);
12610 : }
12611 :
12612 5661 : gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
12613 5661 : gimple_assign_set_rhs_from_tree (&rgsi, new_def);
12614 5661 : update_stmt (gsi_stmt (rgsi));
12615 5661 : return;
12616 : }
12617 9085 : else if (instance->kind == slp_inst_kind_bb_reduc)
12618 : {
12619 : /* Largely inspired by reduction chain epilogue handling in
12620 : vect_create_epilog_for_reduction. */
12621 7457 : vec<tree> vec_defs = vNULL;
12622 7457 : vect_get_slp_defs (node, &vec_defs);
12623 7457 : enum tree_code reduc_code
12624 7457 : = gimple_assign_rhs_code (instance->root_stmts[0]->stmt);
12625 : /* ??? We actually have to reflect signs somewhere. */
12626 7457 : if (reduc_code == MINUS_EXPR)
12627 0 : reduc_code = PLUS_EXPR;
12628 7457 : gimple_seq epilogue = NULL;
12629 : /* We may end up with more than one vector result, reduce them
12630 : to one vector. */
12631 7457 : tree vec_def = vec_defs[0];
12632 7457 : tree vectype = TREE_TYPE (vec_def);
12633 7457 : tree compute_vectype = vectype;
12634 7457 : bool pun_for_overflow_p = (ANY_INTEGRAL_TYPE_P (vectype)
12635 6305 : && TYPE_OVERFLOW_UNDEFINED (vectype)
12636 11943 : && operation_can_overflow (reduc_code));
12637 3922 : if (pun_for_overflow_p)
12638 : {
12639 3922 : compute_vectype = unsigned_type_for (vectype);
12640 3922 : vec_def = gimple_build (&epilogue, VIEW_CONVERT_EXPR,
12641 : compute_vectype, vec_def);
12642 : }
12643 10764 : for (unsigned i = 1; i < vec_defs.length (); ++i)
12644 : {
12645 3307 : tree def = vec_defs[i];
12646 3307 : if (pun_for_overflow_p)
12647 3111 : def = gimple_build (&epilogue, VIEW_CONVERT_EXPR,
12648 : compute_vectype, def);
12649 3307 : vec_def = gimple_build (&epilogue, reduc_code, compute_vectype,
12650 : vec_def, def);
12651 : }
12652 7457 : vec_defs.release ();
12653 : /* ??? Support other schemes than direct internal fn or two
12654 : element vectors. */
12655 7457 : tree scalar_def;
12656 7457 : internal_fn reduc_fn;
12657 9143 : if (!reduction_fn_for_scalar_code (reduc_code, &reduc_fn)
12658 7457 : || reduc_fn == IFN_LAST
12659 14914 : || !direct_internal_fn_supported_p (reduc_fn, compute_vectype,
12660 : OPTIMIZE_FOR_BOTH))
12661 : {
12662 1686 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (compute_vectype), 2u));
12663 1686 : tree tem0 = gimple_build (&epilogue, BIT_FIELD_REF,
12664 1686 : TREE_TYPE (compute_vectype), vec_def,
12665 1686 : TYPE_SIZE (TREE_TYPE (compute_vectype)),
12666 1686 : bitsize_zero_node);
12667 1686 : tree tem1 = gimple_build (&epilogue, BIT_FIELD_REF,
12668 1686 : TREE_TYPE (compute_vectype), vec_def,
12669 1686 : TYPE_SIZE (TREE_TYPE (compute_vectype)),
12670 1686 : TYPE_SIZE (TREE_TYPE (compute_vectype)));
12671 1686 : scalar_def = gimple_build (&epilogue, reduc_code,
12672 1686 : TREE_TYPE (compute_vectype), tem0, tem1);
12673 : }
12674 : else
12675 5771 : scalar_def = gimple_build (&epilogue, as_combined_fn (reduc_fn),
12676 5771 : TREE_TYPE (compute_vectype), vec_def);
12677 7457 : if (!SLP_INSTANCE_REMAIN_DEFS (instance).is_empty ())
12678 : {
12679 4352 : tree rem_def = NULL_TREE;
12680 17528 : for (auto def : SLP_INSTANCE_REMAIN_DEFS (instance))
12681 : {
12682 13176 : def = gimple_convert (&epilogue, TREE_TYPE (scalar_def), def);
12683 13176 : if (!rem_def)
12684 : rem_def = def;
12685 : else
12686 8824 : rem_def = gimple_build (&epilogue, reduc_code,
12687 8824 : TREE_TYPE (scalar_def),
12688 : rem_def, def);
12689 : }
12690 4352 : scalar_def = gimple_build (&epilogue, reduc_code,
12691 4352 : TREE_TYPE (scalar_def),
12692 : scalar_def, rem_def);
12693 : }
12694 7457 : scalar_def = gimple_convert (&epilogue,
12695 7457 : TREE_TYPE (vectype), scalar_def);
12696 7457 : gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
12697 7457 : gsi_insert_seq_before (&rgsi, epilogue, GSI_SAME_STMT);
12698 7457 : gimple_assign_set_rhs_from_tree (&rgsi, scalar_def);
12699 7457 : update_stmt (gsi_stmt (rgsi));
12700 7457 : return;
12701 : }
12702 1628 : else if (instance->kind == slp_inst_kind_gcond)
12703 : {
12704 : /* Only support a single root for now as we can't codegen CFG yet and so we
12705 : can't support lane > 1 at this time. */
12706 1628 : gcc_assert (instance->root_stmts.length () == 1);
12707 1628 : auto root_stmt_info = instance->root_stmts[0];
12708 1628 : auto last_stmt = STMT_VINFO_STMT (vect_orig_stmt (root_stmt_info));
12709 1628 : gimple_stmt_iterator rgsi = gsi_for_stmt (last_stmt);
12710 1628 : gcc_assert (!SLP_TREE_VEC_DEFS (node).is_empty ());
12711 1628 : bool res = vectorizable_early_exit (as_a <loop_vec_info> (vinfo),
12712 : root_stmt_info, &rgsi, node, NULL);
12713 1628 : gcc_assert (res);
12714 1628 : return;
12715 : }
12716 : else
12717 0 : gcc_unreachable ();
12718 : }
12719 :
12720 : struct slp_scc_info
12721 : {
12722 : bool on_stack;
12723 : bool res;
12724 : int dfs;
12725 : int lowlink;
12726 : };
12727 :
12728 : /* Schedule the SLP INSTANCE doing a DFS walk and collecting SCCs.
12729 : When PLACE_ONLY, return false if there is no possible schedule. */
12730 :
12731 : static bool
12732 3147937 : vect_schedule_scc (vec_info *vinfo, slp_tree node, slp_instance instance,
12733 : hash_map<slp_tree, slp_scc_info> &scc_info,
12734 : int &maxdfs, vec<slp_tree> &stack, bool place_only)
12735 : {
12736 3147937 : bool existed_p;
12737 3147937 : slp_scc_info *info = &scc_info.get_or_insert (node, &existed_p);
12738 3147937 : gcc_assert (!existed_p);
12739 3147937 : info->dfs = maxdfs;
12740 3147937 : info->lowlink = maxdfs;
12741 3147937 : maxdfs++;
12742 :
12743 : /* Leaf. */
12744 3147937 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
12745 : {
12746 1210071 : info->on_stack = false;
12747 1210071 : bool res = vect_schedule_slp_node (vinfo, node, instance, place_only);
12748 1210071 : gcc_assert (res);
12749 1210071 : info->res = res;
12750 1210071 : return res;
12751 : }
12752 :
12753 1937866 : info->on_stack = true;
12754 1937866 : info->res = true;
12755 1937866 : stack.safe_push (node);
12756 :
12757 1937866 : bool res = true;
12758 1937866 : unsigned i;
12759 1937866 : slp_tree child;
12760 : /* DFS recurse. */
12761 7570556 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
12762 : {
12763 2042209 : if (!child)
12764 55726 : continue;
12765 1986483 : slp_scc_info *child_info = scc_info.get (child);
12766 1986483 : if (!child_info)
12767 : {
12768 1836604 : res &= vect_schedule_scc (vinfo, child, instance, scc_info,
12769 : maxdfs, stack, place_only);
12770 : /* Recursion might have re-allocated the node. */
12771 1836604 : info = scc_info.get (node);
12772 1836604 : child_info = scc_info.get (child);
12773 1836604 : info->lowlink = MIN (info->lowlink, child_info->lowlink);
12774 : }
12775 149879 : else if (child_info->on_stack)
12776 35792 : info->lowlink = MIN (info->lowlink, child_info->dfs);
12777 : else
12778 114087 : res &= child_info->res;
12779 : }
12780 1937866 : if (info->lowlink != info->dfs)
12781 : return res;
12782 :
12783 1898230 : auto_vec<slp_tree, 4> phis_to_fixup;
12784 :
12785 : /* Singleton. */
12786 1898230 : if (stack.last () == node)
12787 : {
12788 1869106 : stack.pop ();
12789 1869106 : info->on_stack = false;
12790 1869106 : res &= vect_schedule_slp_node (vinfo, node, instance, place_only);
12791 1869106 : info->res = res;
12792 1869106 : if (!SLP_TREE_PERMUTE_P (node)
12793 1869106 : && is_a <gphi *> (SLP_TREE_REPRESENTATIVE (node)->stmt))
12794 73579 : phis_to_fixup.quick_push (node);
12795 : }
12796 : else
12797 : {
12798 : /* SCC. */
12799 29124 : int last_idx = stack.length () - 1;
12800 68760 : while (stack[last_idx] != node)
12801 39636 : last_idx--;
12802 : /* We can break the cycle at PHIs who have at least one child
12803 : code generated. Then we could re-start the DFS walk until
12804 : all nodes in the SCC are covered (we might have new entries
12805 : for only back-reachable nodes). But it's simpler to just
12806 : iterate and schedule those that are ready. */
12807 29124 : unsigned todo = stack.length () - last_idx;
12808 29124 : auto_vec<slp_tree, 4> saved_scc (todo);
12809 87372 : memcpy (saved_scc.address (), stack.address ()[last_idx + 1],
12810 : sizeof (slp_tree) * todo);
12811 29959 : do
12812 : {
12813 132934 : for (int idx = stack.length () - 1; idx >= last_idx; --idx)
12814 : {
12815 73016 : slp_tree entry = stack[idx];
12816 73016 : if (!entry)
12817 2561 : continue;
12818 70455 : bool phi = (!SLP_TREE_PERMUTE_P (entry)
12819 70455 : && is_a <gphi *> (SLP_TREE_REPRESENTATIVE (entry)->stmt));
12820 70455 : bool ready = !phi;
12821 214280 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (entry), i, child)
12822 133586 : if (!child)
12823 : {
12824 23081 : gcc_assert (phi);
12825 : ready = true;
12826 : break;
12827 : }
12828 110505 : else if (scc_info.get (child)->on_stack)
12829 : {
12830 31008 : if (!phi)
12831 : {
12832 : ready = false;
12833 : break;
12834 : }
12835 : }
12836 : else
12837 : {
12838 79497 : if (phi)
12839 : {
12840 : ready = true;
12841 : break;
12842 : }
12843 : }
12844 47374 : if (ready)
12845 : {
12846 68760 : res &= vect_schedule_slp_node (vinfo, entry, instance,
12847 : place_only);
12848 68760 : scc_info.get (entry)->on_stack = false;
12849 68760 : stack[idx] = NULL;
12850 68760 : todo--;
12851 68760 : if (phi)
12852 35628 : phis_to_fixup.safe_push (entry);
12853 : }
12854 : }
12855 : }
12856 29959 : while (todo != 0);
12857 :
12858 : /* Push the scheduling result to all of the SCC. */
12859 87372 : for (slp_tree entry : saved_scc)
12860 0 : scc_info.get (entry)->res = res;
12861 :
12862 : /* Pop the SCC. */
12863 29124 : stack.truncate (last_idx);
12864 29124 : }
12865 :
12866 1898230 : if (place_only)
12867 : return res;
12868 :
12869 : /* Now fixup the backedge def of the vectorized PHIs in this SCC. */
12870 : slp_tree phi_node;
12871 1022381 : FOR_EACH_VEC_ELT (phis_to_fixup, i, phi_node)
12872 : {
12873 56034 : gphi *phi = as_a <gphi *> (SLP_TREE_REPRESENTATIVE (phi_node)->stmt);
12874 56034 : edge_iterator ei;
12875 56034 : edge e;
12876 176462 : FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
12877 : {
12878 120428 : unsigned dest_idx = e->dest_idx;
12879 120428 : child = SLP_TREE_CHILDREN (phi_node)[dest_idx];
12880 120428 : if (!child || SLP_TREE_DEF_TYPE (child) != vect_internal_def)
12881 67256 : continue;
12882 53172 : unsigned n = SLP_TREE_VEC_DEFS (phi_node).length ();
12883 : /* Simply fill all args. */
12884 53172 : if (STMT_VINFO_DEF_TYPE (SLP_TREE_REPRESENTATIVE (phi_node))
12885 : != vect_first_order_recurrence)
12886 112342 : for (unsigned i = 0; i < n; ++i)
12887 : {
12888 59213 : tree phidef = SLP_TREE_VEC_DEFS (phi_node)[i];
12889 59213 : gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (phidef));
12890 59213 : add_phi_arg (phi, vect_get_slp_vect_def (child, i),
12891 : e, gimple_phi_arg_location (phi, dest_idx));
12892 : }
12893 : else
12894 : {
12895 : /* Unless it is a first order recurrence which needs
12896 : args filled in for both the PHI node and the permutes. */
12897 43 : gimple *perm
12898 43 : = SSA_NAME_DEF_STMT (SLP_TREE_VEC_DEFS (phi_node)[0]);
12899 43 : gimple *rphi = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (perm));
12900 43 : add_phi_arg (as_a <gphi *> (rphi),
12901 : vect_get_slp_vect_def (child, n - 1),
12902 : e, gimple_phi_arg_location (phi, dest_idx));
12903 166 : for (unsigned i = 0; i < n; ++i)
12904 : {
12905 80 : gimple *perm
12906 80 : = SSA_NAME_DEF_STMT (SLP_TREE_VEC_DEFS (phi_node)[i]);
12907 80 : if (i > 0)
12908 37 : gimple_assign_set_rhs1 (perm,
12909 : vect_get_slp_vect_def (child, i - 1));
12910 80 : gimple_assign_set_rhs2 (perm,
12911 : vect_get_slp_vect_def (child, i));
12912 80 : update_stmt (perm);
12913 : }
12914 : }
12915 : }
12916 : }
12917 :
12918 966347 : gcc_assert (res);
12919 : return true;
12920 1898230 : }
12921 :
12922 : /* Generate vector code for SLP_INSTANCES in the loop/basic block. Perform
12923 : vector stmt placement only when PLACE_ONLY is true, removing SLP graph
12924 : entries that cannot be scheduled. If placing, return false if a schedule
12925 : cannot be computed for any entry. */
12926 :
12927 : bool
12928 804364 : vect_schedule_slp (vec_info *vinfo, vec<slp_instance> &slp_instances,
12929 : bool place_only)
12930 : {
12931 804364 : slp_instance instance;
12932 804364 : unsigned int i;
12933 :
12934 804364 : hash_map<slp_tree, slp_scc_info> scc_info;
12935 804364 : int maxdfs = 0;
12936 2119579 : for (i = 0; slp_instances.iterate (i, &instance); )
12937 : {
12938 1315215 : slp_tree node = SLP_INSTANCE_TREE (instance);
12939 1315215 : if (!place_only && dump_enabled_p ())
12940 : {
12941 16284 : dump_printf_loc (MSG_NOTE, vect_location,
12942 : "Vectorizing SLP tree:\n");
12943 : /* ??? Dump all? */
12944 16284 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
12945 522 : dump_printf_loc (MSG_NOTE, vect_location, "Root stmt: %G",
12946 522 : SLP_INSTANCE_ROOT_STMTS (instance)[0]->stmt);
12947 16284 : vect_print_slp_graph (MSG_NOTE, vect_location,
12948 : SLP_INSTANCE_TREE (instance));
12949 : }
12950 : /* Schedule the tree of INSTANCE, scheduling SCCs in a way to
12951 : have a PHI be the node breaking the cycle. */
12952 1315215 : bool res;
12953 1315215 : if (slp_scc_info *info = scc_info.get (node))
12954 3882 : res = info->res;
12955 : else
12956 : {
12957 1311333 : auto_vec<slp_tree> stack;
12958 1311333 : res = vect_schedule_scc (vinfo, node, instance, scc_info,
12959 : maxdfs, stack, place_only);
12960 1311333 : }
12961 :
12962 1315215 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
12963 : {
12964 66670 : if (place_only)
12965 : {
12966 51924 : gimple *root_stmt = instance->root_stmts[0]->stmt;
12967 51924 : res &= (!node->si
12968 : /* As we instert after node->si it may not be the
12969 : root_stmt itself. */
12970 51924 : || (node->si != root_stmt
12971 51031 : && vect_stmt_dominates_stmt_p (node->si, root_stmt)));
12972 : }
12973 : else
12974 14746 : vectorize_slp_instance_root_stmt (vinfo, node, instance);
12975 : }
12976 :
12977 1315215 : if (!place_only && dump_enabled_p ())
12978 16284 : dump_printf_loc (MSG_NOTE, vect_location,
12979 : "vectorizing stmts using SLP.\n");
12980 :
12981 1315215 : if (!res)
12982 : {
12983 40 : gcc_assert (place_only);
12984 40 : if (dump_enabled_p ())
12985 26 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
12986 : "not vectorized: cannot schedule SLP graph "
12987 : "entry %p\n", (void *)instance);
12988 40 : vect_free_slp_instance (instance);
12989 40 : slp_instances.ordered_remove (i);
12990 40 : continue;
12991 : }
12992 1315175 : ++i;
12993 : }
12994 :
12995 804364 : if (place_only)
12996 494644 : return !slp_instances.is_empty ();
12997 :
12998 1959071 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
12999 : {
13000 597665 : slp_tree root = SLP_INSTANCE_TREE (instance);
13001 597665 : stmt_vec_info store_info;
13002 597665 : unsigned int j;
13003 :
13004 : /* Remove scalar call stmts. Do not do this for basic-block
13005 : vectorization as not all uses may be vectorized.
13006 : ??? Why should this be necessary? DCE should be able to
13007 : remove the stmts itself.
13008 : ??? For BB vectorization we can as well remove scalar
13009 : stmts starting from the SLP tree root if they have no
13010 : uses. */
13011 597665 : if (is_a <loop_vec_info> (vinfo))
13012 90685 : vect_remove_slp_scalar_calls (vinfo, root);
13013 :
13014 : /* Remove vectorized stores original scalar stmts. */
13015 2661692 : for (j = 0; SLP_TREE_SCALAR_STMTS (root).iterate (j, &store_info); j++)
13016 : {
13017 1506626 : if (!store_info
13018 1506612 : || !STMT_VINFO_DATA_REF (store_info)
13019 1477029 : || !DR_IS_WRITE (STMT_VINFO_DATA_REF (store_info)))
13020 : break;
13021 :
13022 1466362 : store_info = vect_orig_stmt (store_info);
13023 : /* Free the attached stmt_vec_info and remove the stmt. */
13024 1466362 : vinfo->remove_stmt (store_info);
13025 :
13026 : /* Invalidate SLP_TREE_REPRESENTATIVE in case we released it
13027 : to not crash in vect_free_slp_tree later. */
13028 1466362 : if (SLP_TREE_REPRESENTATIVE (root) == store_info)
13029 557063 : SLP_TREE_REPRESENTATIVE (root) = NULL;
13030 : }
13031 : }
13032 :
13033 : return true;
13034 804364 : }
|
